1 /* $NetBSD: uvm_map.c,v 1.287 2010/02/08 19:02:33 joerg 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 1.287 2010/02/08 19:02:33 joerg 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 #include <uvm/uvm_readahead.h> 99 100 #if defined(DDB) || defined(DEBUGPRINT) 101 #include <uvm/uvm_ddb.h> 102 #endif 103 104 #if !defined(UVMMAP_COUNTERS) 105 106 #define UVMMAP_EVCNT_DEFINE(name) /* nothing */ 107 #define UVMMAP_EVCNT_INCR(ev) /* nothing */ 108 #define UVMMAP_EVCNT_DECR(ev) /* nothing */ 109 110 #else /* defined(UVMMAP_NOCOUNTERS) */ 111 112 #include <sys/evcnt.h> 113 #define UVMMAP_EVCNT_DEFINE(name) \ 114 struct evcnt uvmmap_evcnt_##name = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, \ 115 "uvmmap", #name); \ 116 EVCNT_ATTACH_STATIC(uvmmap_evcnt_##name); 117 #define UVMMAP_EVCNT_INCR(ev) uvmmap_evcnt_##ev.ev_count++ 118 #define UVMMAP_EVCNT_DECR(ev) uvmmap_evcnt_##ev.ev_count-- 119 120 #endif /* defined(UVMMAP_NOCOUNTERS) */ 121 122 UVMMAP_EVCNT_DEFINE(ubackmerge) 123 UVMMAP_EVCNT_DEFINE(uforwmerge) 124 UVMMAP_EVCNT_DEFINE(ubimerge) 125 UVMMAP_EVCNT_DEFINE(unomerge) 126 UVMMAP_EVCNT_DEFINE(kbackmerge) 127 UVMMAP_EVCNT_DEFINE(kforwmerge) 128 UVMMAP_EVCNT_DEFINE(kbimerge) 129 UVMMAP_EVCNT_DEFINE(knomerge) 130 UVMMAP_EVCNT_DEFINE(map_call) 131 UVMMAP_EVCNT_DEFINE(mlk_call) 132 UVMMAP_EVCNT_DEFINE(mlk_hint) 133 UVMMAP_EVCNT_DEFINE(mlk_list) 134 UVMMAP_EVCNT_DEFINE(mlk_tree) 135 UVMMAP_EVCNT_DEFINE(mlk_treeloop) 136 UVMMAP_EVCNT_DEFINE(mlk_listloop) 137 138 UVMMAP_EVCNT_DEFINE(uke_alloc) 139 UVMMAP_EVCNT_DEFINE(uke_free) 140 UVMMAP_EVCNT_DEFINE(ukh_alloc) 141 UVMMAP_EVCNT_DEFINE(ukh_free) 142 143 const char vmmapbsy[] = "vmmapbsy"; 144 145 /* 146 * cache for vmspace structures. 147 */ 148 149 static struct pool_cache uvm_vmspace_cache; 150 151 /* 152 * cache for dynamically-allocated map entries. 153 */ 154 155 static struct pool_cache uvm_map_entry_cache; 156 157 MALLOC_DEFINE(M_VMMAP, "VM map", "VM map structures"); 158 MALLOC_DEFINE(M_VMPMAP, "VM pmap", "VM pmap"); 159 160 #ifdef PMAP_GROWKERNEL 161 /* 162 * This global represents the end of the kernel virtual address 163 * space. If we want to exceed this, we must grow the kernel 164 * virtual address space dynamically. 165 * 166 * Note, this variable is locked by kernel_map's lock. 167 */ 168 vaddr_t uvm_maxkaddr; 169 #endif 170 171 /* 172 * macros 173 */ 174 175 /* 176 * VM_MAP_USE_KMAPENT: determine if uvm_kmapent_alloc/free is used 177 * for the vm_map. 178 */ 179 extern struct vm_map *pager_map; /* XXX */ 180 #define VM_MAP_USE_KMAPENT_FLAGS(flags) \ 181 (((flags) & VM_MAP_INTRSAFE) != 0) 182 #define VM_MAP_USE_KMAPENT(map) \ 183 (VM_MAP_USE_KMAPENT_FLAGS((map)->flags) || (map) == kernel_map) 184 185 /* 186 * UVM_ET_ISCOMPATIBLE: check some requirements for map entry merging 187 */ 188 189 #define UVM_ET_ISCOMPATIBLE(ent, type, uobj, meflags, \ 190 prot, maxprot, inh, adv, wire) \ 191 ((ent)->etype == (type) && \ 192 (((ent)->flags ^ (meflags)) & (UVM_MAP_NOMERGE | UVM_MAP_QUANTUM)) \ 193 == 0 && \ 194 (ent)->object.uvm_obj == (uobj) && \ 195 (ent)->protection == (prot) && \ 196 (ent)->max_protection == (maxprot) && \ 197 (ent)->inheritance == (inh) && \ 198 (ent)->advice == (adv) && \ 199 (ent)->wired_count == (wire)) 200 201 /* 202 * uvm_map_entry_link: insert entry into a map 203 * 204 * => map must be locked 205 */ 206 #define uvm_map_entry_link(map, after_where, entry) do { \ 207 uvm_mapent_check(entry); \ 208 (map)->nentries++; \ 209 (entry)->prev = (after_where); \ 210 (entry)->next = (after_where)->next; \ 211 (entry)->prev->next = (entry); \ 212 (entry)->next->prev = (entry); \ 213 uvm_rb_insert((map), (entry)); \ 214 } while (/*CONSTCOND*/ 0) 215 216 /* 217 * uvm_map_entry_unlink: remove entry from a map 218 * 219 * => map must be locked 220 */ 221 #define uvm_map_entry_unlink(map, entry) do { \ 222 KASSERT((entry) != (map)->first_free); \ 223 KASSERT((entry) != (map)->hint); \ 224 uvm_mapent_check(entry); \ 225 (map)->nentries--; \ 226 (entry)->next->prev = (entry)->prev; \ 227 (entry)->prev->next = (entry)->next; \ 228 uvm_rb_remove((map), (entry)); \ 229 } while (/*CONSTCOND*/ 0) 230 231 /* 232 * SAVE_HINT: saves the specified entry as the hint for future lookups. 233 * 234 * => map need not be locked. 235 */ 236 #define SAVE_HINT(map, check, value) do { \ 237 if ((map)->hint == (check)) \ 238 (map)->hint = (value); \ 239 } while (/*CONSTCOND*/ 0) 240 241 /* 242 * clear_hints: ensure that hints don't point to the entry. 243 * 244 * => map must be write-locked. 245 */ 246 static void 247 clear_hints(struct vm_map *map, struct vm_map_entry *ent) 248 { 249 250 SAVE_HINT(map, ent, ent->prev); 251 if (map->first_free == ent) { 252 map->first_free = ent->prev; 253 } 254 } 255 256 /* 257 * VM_MAP_RANGE_CHECK: check and correct range 258 * 259 * => map must at least be read locked 260 */ 261 262 #define VM_MAP_RANGE_CHECK(map, start, end) do { \ 263 if (start < vm_map_min(map)) \ 264 start = vm_map_min(map); \ 265 if (end > vm_map_max(map)) \ 266 end = vm_map_max(map); \ 267 if (start > end) \ 268 start = end; \ 269 } while (/*CONSTCOND*/ 0) 270 271 /* 272 * local prototypes 273 */ 274 275 static struct vm_map_entry * 276 uvm_mapent_alloc(struct vm_map *, int); 277 static struct vm_map_entry * 278 uvm_mapent_alloc_split(struct vm_map *, 279 const struct vm_map_entry *, int, 280 struct uvm_mapent_reservation *); 281 static void uvm_mapent_copy(struct vm_map_entry *, struct vm_map_entry *); 282 static void uvm_mapent_free(struct vm_map_entry *); 283 #if defined(DEBUG) 284 static void _uvm_mapent_check(const struct vm_map_entry *, const char *, 285 int); 286 #define uvm_mapent_check(map) _uvm_mapent_check(map, __FILE__, __LINE__) 287 #else /* defined(DEBUG) */ 288 #define uvm_mapent_check(e) /* nothing */ 289 #endif /* defined(DEBUG) */ 290 static struct vm_map_entry * 291 uvm_kmapent_alloc(struct vm_map *, int); 292 static void uvm_kmapent_free(struct vm_map_entry *); 293 static vsize_t uvm_kmapent_overhead(vsize_t); 294 295 static void uvm_map_entry_unwire(struct vm_map *, struct vm_map_entry *); 296 static void uvm_map_reference_amap(struct vm_map_entry *, int); 297 static int uvm_map_space_avail(vaddr_t *, vsize_t, voff_t, vsize_t, int, 298 struct vm_map_entry *); 299 static void uvm_map_unreference_amap(struct vm_map_entry *, int); 300 301 int _uvm_map_sanity(struct vm_map *); 302 int _uvm_tree_sanity(struct vm_map *); 303 static vsize_t uvm_rb_maxgap(const struct vm_map_entry *); 304 305 CTASSERT(offsetof(struct vm_map_entry, rb_node) == 0); 306 #define ROOT_ENTRY(map) ((struct vm_map_entry *)(map)->rb_tree.rbt_root) 307 #define LEFT_ENTRY(entry) ((struct vm_map_entry *)(entry)->rb_node.rb_left) 308 #define RIGHT_ENTRY(entry) ((struct vm_map_entry *)(entry)->rb_node.rb_right) 309 #define PARENT_ENTRY(map, entry) \ 310 (ROOT_ENTRY(map) == (entry) \ 311 ? NULL \ 312 : (struct vm_map_entry *)RB_FATHER(&(entry)->rb_node)) 313 314 static int 315 uvm_map_compare_nodes(const struct rb_node *nparent, 316 const struct rb_node *nkey) 317 { 318 const struct vm_map_entry *eparent = (const void *) nparent; 319 const struct vm_map_entry *ekey = (const void *) nkey; 320 321 KASSERT(eparent->start < ekey->start || eparent->start >= ekey->end); 322 KASSERT(ekey->start < eparent->start || ekey->start >= eparent->end); 323 324 if (ekey->start < eparent->start) 325 return -1; 326 if (ekey->start >= eparent->end) 327 return 1; 328 return 0; 329 } 330 331 static int 332 uvm_map_compare_key(const struct rb_node *nparent, const void *vkey) 333 { 334 const struct vm_map_entry *eparent = (const void *) nparent; 335 const vaddr_t va = *(const vaddr_t *) vkey; 336 337 if (va < eparent->start) 338 return -1; 339 if (va >= eparent->end) 340 return 1; 341 return 0; 342 } 343 344 static const struct rb_tree_ops uvm_map_tree_ops = { 345 .rbto_compare_nodes = uvm_map_compare_nodes, 346 .rbto_compare_key = uvm_map_compare_key, 347 }; 348 349 static inline vsize_t 350 uvm_rb_gap(const struct vm_map_entry *entry) 351 { 352 KASSERT(entry->next != NULL); 353 return entry->next->start - entry->end; 354 } 355 356 static vsize_t 357 uvm_rb_maxgap(const struct vm_map_entry *entry) 358 { 359 struct vm_map_entry *child; 360 vsize_t maxgap = entry->gap; 361 362 /* 363 * We need maxgap to be the largest gap of us or any of our 364 * descendents. Since each of our children's maxgap is the 365 * cached value of their largest gap of themselves or their 366 * descendents, we can just use that value and avoid recursing 367 * down the tree to calculate it. 368 */ 369 if ((child = LEFT_ENTRY(entry)) != NULL && maxgap < child->maxgap) 370 maxgap = child->maxgap; 371 372 if ((child = RIGHT_ENTRY(entry)) != NULL && maxgap < child->maxgap) 373 maxgap = child->maxgap; 374 375 return maxgap; 376 } 377 378 static void 379 uvm_rb_fixup(struct vm_map *map, struct vm_map_entry *entry) 380 { 381 struct vm_map_entry *parent; 382 383 KASSERT(entry->gap == uvm_rb_gap(entry)); 384 entry->maxgap = uvm_rb_maxgap(entry); 385 386 while ((parent = PARENT_ENTRY(map, entry)) != NULL) { 387 struct vm_map_entry *brother; 388 vsize_t maxgap = parent->gap; 389 390 KDASSERT(parent->gap == uvm_rb_gap(parent)); 391 if (maxgap < entry->maxgap) 392 maxgap = entry->maxgap; 393 /* 394 * Since we work our towards the root, we know entry's maxgap 395 * value is ok but its brothers may now be out-of-date due 396 * rebalancing. So refresh it. 397 */ 398 brother = (struct vm_map_entry *)parent->rb_node.rb_nodes[RB_POSITION(&entry->rb_node) ^ RB_DIR_OTHER]; 399 if (brother != NULL) { 400 KDASSERT(brother->gap == uvm_rb_gap(brother)); 401 brother->maxgap = uvm_rb_maxgap(brother); 402 if (maxgap < brother->maxgap) 403 maxgap = brother->maxgap; 404 } 405 406 parent->maxgap = maxgap; 407 entry = parent; 408 } 409 } 410 411 static void 412 uvm_rb_insert(struct vm_map *map, struct vm_map_entry *entry) 413 { 414 entry->gap = entry->maxgap = uvm_rb_gap(entry); 415 if (entry->prev != &map->header) 416 entry->prev->gap = uvm_rb_gap(entry->prev); 417 418 if (!rb_tree_insert_node(&map->rb_tree, &entry->rb_node)) 419 panic("uvm_rb_insert: map %p: duplicate entry?", map); 420 421 /* 422 * If the previous entry is not our immediate left child, then it's an 423 * ancestor and will be fixed up on the way to the root. We don't 424 * have to check entry->prev against &map->header since &map->header 425 * will never be in the tree. 426 */ 427 uvm_rb_fixup(map, 428 LEFT_ENTRY(entry) == entry->prev ? entry->prev : entry); 429 } 430 431 static void 432 uvm_rb_remove(struct vm_map *map, struct vm_map_entry *entry) 433 { 434 struct vm_map_entry *prev_parent = NULL, *next_parent = NULL; 435 436 /* 437 * If we are removing an interior node, then an adjacent node will 438 * be used to replace its position in the tree. Therefore we will 439 * need to fixup the tree starting at the parent of the replacement 440 * node. So record their parents for later use. 441 */ 442 if (entry->prev != &map->header) 443 prev_parent = PARENT_ENTRY(map, entry->prev); 444 if (entry->next != &map->header) 445 next_parent = PARENT_ENTRY(map, entry->next); 446 447 rb_tree_remove_node(&map->rb_tree, &entry->rb_node); 448 449 /* 450 * If the previous node has a new parent, fixup the tree starting 451 * at the previous node's old parent. 452 */ 453 if (entry->prev != &map->header) { 454 /* 455 * Update the previous entry's gap due to our absence. 456 */ 457 entry->prev->gap = uvm_rb_gap(entry->prev); 458 uvm_rb_fixup(map, entry->prev); 459 if (prev_parent != NULL 460 && prev_parent != entry 461 && prev_parent != PARENT_ENTRY(map, entry->prev)) 462 uvm_rb_fixup(map, prev_parent); 463 } 464 465 /* 466 * If the next node has a new parent, fixup the tree starting 467 * at the next node's old parent. 468 */ 469 if (entry->next != &map->header) { 470 uvm_rb_fixup(map, entry->next); 471 if (next_parent != NULL 472 && next_parent != entry 473 && next_parent != PARENT_ENTRY(map, entry->next)) 474 uvm_rb_fixup(map, next_parent); 475 } 476 } 477 478 #if defined(DEBUG) 479 int uvm_debug_check_map = 0; 480 int uvm_debug_check_rbtree = 0; 481 #define uvm_map_check(map, name) \ 482 _uvm_map_check((map), (name), __FILE__, __LINE__) 483 static void 484 _uvm_map_check(struct vm_map *map, const char *name, 485 const char *file, int line) 486 { 487 488 if ((uvm_debug_check_map && _uvm_map_sanity(map)) || 489 (uvm_debug_check_rbtree && _uvm_tree_sanity(map))) { 490 panic("uvm_map_check failed: \"%s\" map=%p (%s:%d)", 491 name, map, file, line); 492 } 493 } 494 #else /* defined(DEBUG) */ 495 #define uvm_map_check(map, name) /* nothing */ 496 #endif /* defined(DEBUG) */ 497 498 #if defined(DEBUG) || defined(DDB) 499 int 500 _uvm_map_sanity(struct vm_map *map) 501 { 502 bool first_free_found = false; 503 bool hint_found = false; 504 const struct vm_map_entry *e; 505 struct vm_map_entry *hint = map->hint; 506 507 e = &map->header; 508 for (;;) { 509 if (map->first_free == e) { 510 first_free_found = true; 511 } else if (!first_free_found && e->next->start > e->end) { 512 printf("first_free %p should be %p\n", 513 map->first_free, e); 514 return -1; 515 } 516 if (hint == e) { 517 hint_found = true; 518 } 519 520 e = e->next; 521 if (e == &map->header) { 522 break; 523 } 524 } 525 if (!first_free_found) { 526 printf("stale first_free\n"); 527 return -1; 528 } 529 if (!hint_found) { 530 printf("stale hint\n"); 531 return -1; 532 } 533 return 0; 534 } 535 536 int 537 _uvm_tree_sanity(struct vm_map *map) 538 { 539 struct vm_map_entry *tmp, *trtmp; 540 int n = 0, i = 1; 541 542 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 543 if (tmp->gap != uvm_rb_gap(tmp)) { 544 printf("%d/%d gap %lx != %lx %s\n", 545 n + 1, map->nentries, 546 (ulong)tmp->gap, (ulong)uvm_rb_gap(tmp), 547 tmp->next == &map->header ? "(last)" : ""); 548 goto error; 549 } 550 /* 551 * If any entries are out of order, tmp->gap will be unsigned 552 * and will likely exceed the size of the map. 553 */ 554 if (tmp->gap >= vm_map_max(map) - vm_map_min(map)) { 555 printf("too large gap %zu\n", (size_t)tmp->gap); 556 goto error; 557 } 558 n++; 559 } 560 561 if (n != map->nentries) { 562 printf("nentries: %d vs %d\n", n, map->nentries); 563 goto error; 564 } 565 566 trtmp = NULL; 567 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 568 if (tmp->maxgap != uvm_rb_maxgap(tmp)) { 569 printf("maxgap %lx != %lx\n", 570 (ulong)tmp->maxgap, 571 (ulong)uvm_rb_maxgap(tmp)); 572 goto error; 573 } 574 if (trtmp != NULL && trtmp->start >= tmp->start) { 575 printf("corrupt: 0x%"PRIxVADDR"x >= 0x%"PRIxVADDR"x\n", 576 trtmp->start, tmp->start); 577 goto error; 578 } 579 580 trtmp = tmp; 581 } 582 583 for (tmp = map->header.next; tmp != &map->header; 584 tmp = tmp->next, i++) { 585 trtmp = (void *) rb_tree_iterate(&map->rb_tree, &tmp->rb_node, 586 RB_DIR_LEFT); 587 if (trtmp == NULL) 588 trtmp = &map->header; 589 if (tmp->prev != trtmp) { 590 printf("lookup: %d: %p->prev=%p: %p\n", 591 i, tmp, tmp->prev, trtmp); 592 goto error; 593 } 594 trtmp = (void *) rb_tree_iterate(&map->rb_tree, &tmp->rb_node, 595 RB_DIR_RIGHT); 596 if (trtmp == NULL) 597 trtmp = &map->header; 598 if (tmp->next != trtmp) { 599 printf("lookup: %d: %p->next=%p: %p\n", 600 i, tmp, tmp->next, trtmp); 601 goto error; 602 } 603 trtmp = (void *)rb_tree_find_node(&map->rb_tree, &tmp->start); 604 if (trtmp != tmp) { 605 printf("lookup: %d: %p - %p: %p\n", i, tmp, trtmp, 606 PARENT_ENTRY(map, tmp)); 607 goto error; 608 } 609 } 610 611 return (0); 612 error: 613 return (-1); 614 } 615 #endif /* defined(DEBUG) || defined(DDB) */ 616 617 #ifdef DIAGNOSTIC 618 static struct vm_map *uvm_kmapent_map(struct vm_map_entry *); 619 #endif 620 621 /* 622 * vm_map_lock: acquire an exclusive (write) lock on a map. 623 * 624 * => Note that "intrsafe" maps use only exclusive, spin locks. 625 * 626 * => The locking protocol provides for guaranteed upgrade from shared -> 627 * exclusive by whichever thread currently has the map marked busy. 628 * See "LOCKING PROTOCOL NOTES" in uvm_map.h. This is horrible; among 629 * other problems, it defeats any fairness guarantees provided by RW 630 * locks. 631 */ 632 633 void 634 vm_map_lock(struct vm_map *map) 635 { 636 637 if ((map->flags & VM_MAP_INTRSAFE) != 0) { 638 mutex_spin_enter(&map->mutex); 639 return; 640 } 641 642 for (;;) { 643 rw_enter(&map->lock, RW_WRITER); 644 if (map->busy == NULL) 645 break; 646 if (map->busy == curlwp) 647 break; 648 mutex_enter(&map->misc_lock); 649 rw_exit(&map->lock); 650 if (map->busy != NULL) 651 cv_wait(&map->cv, &map->misc_lock); 652 mutex_exit(&map->misc_lock); 653 } 654 655 map->timestamp++; 656 } 657 658 /* 659 * vm_map_lock_try: try to lock a map, failing if it is already locked. 660 */ 661 662 bool 663 vm_map_lock_try(struct vm_map *map) 664 { 665 666 if ((map->flags & VM_MAP_INTRSAFE) != 0) 667 return mutex_tryenter(&map->mutex); 668 if (!rw_tryenter(&map->lock, RW_WRITER)) 669 return false; 670 if (map->busy != NULL) { 671 rw_exit(&map->lock); 672 return false; 673 } 674 675 map->timestamp++; 676 return true; 677 } 678 679 /* 680 * vm_map_unlock: release an exclusive lock on a map. 681 */ 682 683 void 684 vm_map_unlock(struct vm_map *map) 685 { 686 687 if ((map->flags & VM_MAP_INTRSAFE) != 0) 688 mutex_spin_exit(&map->mutex); 689 else { 690 KASSERT(rw_write_held(&map->lock)); 691 KASSERT(map->busy == NULL || map->busy == curlwp); 692 rw_exit(&map->lock); 693 } 694 } 695 696 /* 697 * vm_map_unbusy: mark the map as unbusy, and wake any waiters that 698 * want an exclusive lock. 699 */ 700 701 void 702 vm_map_unbusy(struct vm_map *map) 703 { 704 705 KASSERT(map->busy == curlwp); 706 707 /* 708 * Safe to clear 'busy' and 'waiters' with only a read lock held: 709 * 710 * o they can only be set with a write lock held 711 * o writers are blocked out with a read or write hold 712 * o at any time, only one thread owns the set of values 713 */ 714 mutex_enter(&map->misc_lock); 715 map->busy = NULL; 716 cv_broadcast(&map->cv); 717 mutex_exit(&map->misc_lock); 718 } 719 720 /* 721 * vm_map_lock_read: acquire a shared (read) lock on a map. 722 */ 723 724 void 725 vm_map_lock_read(struct vm_map *map) 726 { 727 728 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 729 730 rw_enter(&map->lock, RW_READER); 731 } 732 733 /* 734 * vm_map_unlock_read: release a shared lock on a map. 735 */ 736 737 void 738 vm_map_unlock_read(struct vm_map *map) 739 { 740 741 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 742 743 rw_exit(&map->lock); 744 } 745 746 /* 747 * vm_map_busy: mark a map as busy. 748 * 749 * => the caller must hold the map write locked 750 */ 751 752 void 753 vm_map_busy(struct vm_map *map) 754 { 755 756 KASSERT(rw_write_held(&map->lock)); 757 KASSERT(map->busy == NULL); 758 759 map->busy = curlwp; 760 } 761 762 /* 763 * vm_map_locked_p: return true if the map is write locked. 764 * 765 * => only for debug purposes like KASSERTs. 766 * => should not be used to verify that a map is not locked. 767 */ 768 769 bool 770 vm_map_locked_p(struct vm_map *map) 771 { 772 773 if ((map->flags & VM_MAP_INTRSAFE) != 0) { 774 return mutex_owned(&map->mutex); 775 } else { 776 return rw_write_held(&map->lock); 777 } 778 } 779 780 /* 781 * uvm_mapent_alloc: allocate a map entry 782 */ 783 784 static struct vm_map_entry * 785 uvm_mapent_alloc(struct vm_map *map, int flags) 786 { 787 struct vm_map_entry *me; 788 int pflags = (flags & UVM_FLAG_NOWAIT) ? PR_NOWAIT : PR_WAITOK; 789 UVMHIST_FUNC("uvm_mapent_alloc"); UVMHIST_CALLED(maphist); 790 791 if (VM_MAP_USE_KMAPENT(map)) { 792 me = uvm_kmapent_alloc(map, flags); 793 } else { 794 me = pool_cache_get(&uvm_map_entry_cache, pflags); 795 if (__predict_false(me == NULL)) 796 return NULL; 797 me->flags = 0; 798 } 799 800 UVMHIST_LOG(maphist, "<- new entry=0x%x [kentry=%d]", me, 801 ((map->flags & VM_MAP_INTRSAFE) != 0 || map == kernel_map), 0, 0); 802 return (me); 803 } 804 805 /* 806 * uvm_mapent_alloc_split: allocate a map entry for clipping. 807 * 808 * => map must be locked by caller if UVM_MAP_QUANTUM is set. 809 */ 810 811 static struct vm_map_entry * 812 uvm_mapent_alloc_split(struct vm_map *map, 813 const struct vm_map_entry *old_entry, int flags, 814 struct uvm_mapent_reservation *umr) 815 { 816 struct vm_map_entry *me; 817 818 KASSERT(!VM_MAP_USE_KMAPENT(map) || 819 (old_entry->flags & UVM_MAP_QUANTUM) || !UMR_EMPTY(umr)); 820 821 if (old_entry->flags & UVM_MAP_QUANTUM) { 822 struct vm_map_kernel *vmk = vm_map_to_kernel(map); 823 824 KASSERT(vm_map_locked_p(map)); 825 me = vmk->vmk_merged_entries; 826 KASSERT(me); 827 vmk->vmk_merged_entries = me->next; 828 KASSERT(me->flags & UVM_MAP_QUANTUM); 829 } else { 830 me = uvm_mapent_alloc(map, flags); 831 } 832 833 return me; 834 } 835 836 /* 837 * uvm_mapent_free: free map entry 838 */ 839 840 static void 841 uvm_mapent_free(struct vm_map_entry *me) 842 { 843 UVMHIST_FUNC("uvm_mapent_free"); UVMHIST_CALLED(maphist); 844 845 UVMHIST_LOG(maphist,"<- freeing map entry=0x%x [flags=%d]", 846 me, me->flags, 0, 0); 847 if (me->flags & UVM_MAP_KERNEL) { 848 uvm_kmapent_free(me); 849 } else { 850 pool_cache_put(&uvm_map_entry_cache, me); 851 } 852 } 853 854 /* 855 * uvm_mapent_free_merged: free merged map entry 856 * 857 * => keep the entry if needed. 858 * => caller shouldn't hold map locked if VM_MAP_USE_KMAPENT(map) is true. 859 * => map should be locked if UVM_MAP_QUANTUM is set. 860 */ 861 862 static void 863 uvm_mapent_free_merged(struct vm_map *map, struct vm_map_entry *me) 864 { 865 866 KASSERT(!(me->flags & UVM_MAP_KERNEL) || uvm_kmapent_map(me) == map); 867 868 if (me->flags & UVM_MAP_QUANTUM) { 869 /* 870 * keep this entry for later splitting. 871 */ 872 struct vm_map_kernel *vmk; 873 874 KASSERT(vm_map_locked_p(map)); 875 KASSERT(VM_MAP_IS_KERNEL(map)); 876 KASSERT(!VM_MAP_USE_KMAPENT(map) || 877 (me->flags & UVM_MAP_KERNEL)); 878 879 vmk = vm_map_to_kernel(map); 880 me->next = vmk->vmk_merged_entries; 881 vmk->vmk_merged_entries = me; 882 } else { 883 uvm_mapent_free(me); 884 } 885 } 886 887 /* 888 * uvm_mapent_copy: copy a map entry, preserving flags 889 */ 890 891 static inline void 892 uvm_mapent_copy(struct vm_map_entry *src, struct vm_map_entry *dst) 893 { 894 895 memcpy(dst, src, ((char *)&src->uvm_map_entry_stop_copy) - 896 ((char *)src)); 897 } 898 899 /* 900 * uvm_mapent_overhead: calculate maximum kva overhead necessary for 901 * map entries. 902 * 903 * => size and flags are the same as uvm_km_suballoc's ones. 904 */ 905 906 vsize_t 907 uvm_mapent_overhead(vsize_t size, int flags) 908 { 909 910 if (VM_MAP_USE_KMAPENT_FLAGS(flags)) { 911 return uvm_kmapent_overhead(size); 912 } 913 return 0; 914 } 915 916 #if defined(DEBUG) 917 static void 918 _uvm_mapent_check(const struct vm_map_entry *entry, const char *file, int line) 919 { 920 921 if (entry->start >= entry->end) { 922 goto bad; 923 } 924 if (UVM_ET_ISOBJ(entry)) { 925 if (entry->object.uvm_obj == NULL) { 926 goto bad; 927 } 928 } else if (UVM_ET_ISSUBMAP(entry)) { 929 if (entry->object.sub_map == NULL) { 930 goto bad; 931 } 932 } else { 933 if (entry->object.uvm_obj != NULL || 934 entry->object.sub_map != NULL) { 935 goto bad; 936 } 937 } 938 if (!UVM_ET_ISOBJ(entry)) { 939 if (entry->offset != 0) { 940 goto bad; 941 } 942 } 943 944 return; 945 946 bad: 947 panic("%s: bad entry %p (%s:%d)", __func__, entry, file, line); 948 } 949 #endif /* defined(DEBUG) */ 950 951 /* 952 * uvm_map_entry_unwire: unwire a map entry 953 * 954 * => map should be locked by caller 955 */ 956 957 static inline void 958 uvm_map_entry_unwire(struct vm_map *map, struct vm_map_entry *entry) 959 { 960 961 entry->wired_count = 0; 962 uvm_fault_unwire_locked(map, entry->start, entry->end); 963 } 964 965 966 /* 967 * wrapper for calling amap_ref() 968 */ 969 static inline void 970 uvm_map_reference_amap(struct vm_map_entry *entry, int flags) 971 { 972 973 amap_ref(entry->aref.ar_amap, entry->aref.ar_pageoff, 974 (entry->end - entry->start) >> PAGE_SHIFT, flags); 975 } 976 977 978 /* 979 * wrapper for calling amap_unref() 980 */ 981 static inline void 982 uvm_map_unreference_amap(struct vm_map_entry *entry, int flags) 983 { 984 985 amap_unref(entry->aref.ar_amap, entry->aref.ar_pageoff, 986 (entry->end - entry->start) >> PAGE_SHIFT, flags); 987 } 988 989 990 /* 991 * uvm_map_init: init mapping system at boot time. 992 */ 993 994 void 995 uvm_map_init(void) 996 { 997 #if defined(UVMHIST) 998 static struct uvm_history_ent maphistbuf[100]; 999 static struct uvm_history_ent pdhistbuf[100]; 1000 #endif 1001 1002 /* 1003 * first, init logging system. 1004 */ 1005 1006 UVMHIST_FUNC("uvm_map_init"); 1007 UVMHIST_INIT_STATIC(maphist, maphistbuf); 1008 UVMHIST_INIT_STATIC(pdhist, pdhistbuf); 1009 UVMHIST_CALLED(maphist); 1010 UVMHIST_LOG(maphist,"<starting uvm map system>", 0, 0, 0, 0); 1011 1012 /* 1013 * initialize the global lock for kernel map entry. 1014 */ 1015 1016 mutex_init(&uvm_kentry_lock, MUTEX_DRIVER, IPL_VM); 1017 1018 /* 1019 * initialize caches. 1020 */ 1021 1022 pool_cache_bootstrap(&uvm_map_entry_cache, sizeof(struct vm_map_entry), 1023 0, 0, 0, "vmmpepl", NULL, IPL_NONE, NULL, NULL, NULL); 1024 pool_cache_bootstrap(&uvm_vmspace_cache, sizeof(struct vmspace), 1025 0, 0, 0, "vmsppl", NULL, IPL_NONE, NULL, NULL, NULL); 1026 } 1027 1028 /* 1029 * clippers 1030 */ 1031 1032 /* 1033 * uvm_mapent_splitadj: adjust map entries for splitting, after uvm_mapent_copy. 1034 */ 1035 1036 static void 1037 uvm_mapent_splitadj(struct vm_map_entry *entry1, struct vm_map_entry *entry2, 1038 vaddr_t splitat) 1039 { 1040 vaddr_t adj; 1041 1042 KASSERT(entry1->start < splitat); 1043 KASSERT(splitat < entry1->end); 1044 1045 adj = splitat - entry1->start; 1046 entry1->end = entry2->start = splitat; 1047 1048 if (entry1->aref.ar_amap) { 1049 amap_splitref(&entry1->aref, &entry2->aref, adj); 1050 } 1051 if (UVM_ET_ISSUBMAP(entry1)) { 1052 /* ... unlikely to happen, but play it safe */ 1053 uvm_map_reference(entry1->object.sub_map); 1054 } else if (UVM_ET_ISOBJ(entry1)) { 1055 KASSERT(entry1->object.uvm_obj != NULL); /* suppress coverity */ 1056 entry2->offset += adj; 1057 if (entry1->object.uvm_obj->pgops && 1058 entry1->object.uvm_obj->pgops->pgo_reference) 1059 entry1->object.uvm_obj->pgops->pgo_reference( 1060 entry1->object.uvm_obj); 1061 } 1062 } 1063 1064 /* 1065 * uvm_map_clip_start: ensure that the entry begins at or after 1066 * the starting address, if it doesn't we split the entry. 1067 * 1068 * => caller should use UVM_MAP_CLIP_START macro rather than calling 1069 * this directly 1070 * => map must be locked by caller 1071 */ 1072 1073 void 1074 uvm_map_clip_start(struct vm_map *map, struct vm_map_entry *entry, 1075 vaddr_t start, struct uvm_mapent_reservation *umr) 1076 { 1077 struct vm_map_entry *new_entry; 1078 1079 /* uvm_map_simplify_entry(map, entry); */ /* XXX */ 1080 1081 uvm_map_check(map, "clip_start entry"); 1082 uvm_mapent_check(entry); 1083 1084 /* 1085 * Split off the front portion. note that we must insert the new 1086 * entry BEFORE this one, so that this entry has the specified 1087 * starting address. 1088 */ 1089 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 1090 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1091 uvm_mapent_splitadj(new_entry, entry, start); 1092 uvm_map_entry_link(map, entry->prev, new_entry); 1093 1094 uvm_map_check(map, "clip_start leave"); 1095 } 1096 1097 /* 1098 * uvm_map_clip_end: ensure that the entry ends at or before 1099 * the ending address, if it does't we split the reference 1100 * 1101 * => caller should use UVM_MAP_CLIP_END macro rather than calling 1102 * this directly 1103 * => map must be locked by caller 1104 */ 1105 1106 void 1107 uvm_map_clip_end(struct vm_map *map, struct vm_map_entry *entry, vaddr_t end, 1108 struct uvm_mapent_reservation *umr) 1109 { 1110 struct vm_map_entry *new_entry; 1111 1112 uvm_map_check(map, "clip_end entry"); 1113 uvm_mapent_check(entry); 1114 1115 /* 1116 * Create a new entry and insert it 1117 * AFTER the specified entry 1118 */ 1119 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 1120 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1121 uvm_mapent_splitadj(entry, new_entry, end); 1122 uvm_map_entry_link(map, entry, new_entry); 1123 1124 uvm_map_check(map, "clip_end leave"); 1125 } 1126 1127 static void 1128 vm_map_drain(struct vm_map *map, uvm_flag_t flags) 1129 { 1130 1131 if (!VM_MAP_IS_KERNEL(map)) { 1132 return; 1133 } 1134 1135 uvm_km_va_drain(map, flags); 1136 } 1137 1138 /* 1139 * M A P - m a i n e n t r y p o i n t 1140 */ 1141 /* 1142 * uvm_map: establish a valid mapping in a map 1143 * 1144 * => assume startp is page aligned. 1145 * => assume size is a multiple of PAGE_SIZE. 1146 * => assume sys_mmap provides enough of a "hint" to have us skip 1147 * over text/data/bss area. 1148 * => map must be unlocked (we will lock it) 1149 * => <uobj,uoffset> value meanings (4 cases): 1150 * [1] <NULL,uoffset> == uoffset is a hint for PMAP_PREFER 1151 * [2] <NULL,UVM_UNKNOWN_OFFSET> == don't PMAP_PREFER 1152 * [3] <uobj,uoffset> == normal mapping 1153 * [4] <uobj,UVM_UNKNOWN_OFFSET> == uvm_map finds offset based on VA 1154 * 1155 * case [4] is for kernel mappings where we don't know the offset until 1156 * we've found a virtual address. note that kernel object offsets are 1157 * always relative to vm_map_min(kernel_map). 1158 * 1159 * => if `align' is non-zero, we align the virtual address to the specified 1160 * alignment. 1161 * this is provided as a mechanism for large pages. 1162 * 1163 * => XXXCDC: need way to map in external amap? 1164 */ 1165 1166 int 1167 uvm_map(struct vm_map *map, vaddr_t *startp /* IN/OUT */, vsize_t size, 1168 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags) 1169 { 1170 struct uvm_map_args args; 1171 struct vm_map_entry *new_entry; 1172 int error; 1173 1174 KASSERT((flags & UVM_FLAG_QUANTUM) == 0 || VM_MAP_IS_KERNEL(map)); 1175 KASSERT((size & PAGE_MASK) == 0); 1176 1177 /* 1178 * for pager_map, allocate the new entry first to avoid sleeping 1179 * for memory while we have the map locked. 1180 * 1181 * Also, because we allocate entries for in-kernel maps 1182 * a bit differently (cf. uvm_kmapent_alloc/free), we need to 1183 * allocate them before locking the map. 1184 */ 1185 1186 new_entry = NULL; 1187 if (VM_MAP_USE_KMAPENT(map) || (flags & UVM_FLAG_QUANTUM) || 1188 map == pager_map) { 1189 new_entry = uvm_mapent_alloc(map, (flags & UVM_FLAG_NOWAIT)); 1190 if (__predict_false(new_entry == NULL)) 1191 return ENOMEM; 1192 if (flags & UVM_FLAG_QUANTUM) 1193 new_entry->flags |= UVM_MAP_QUANTUM; 1194 } 1195 if (map == pager_map) 1196 flags |= UVM_FLAG_NOMERGE; 1197 1198 error = uvm_map_prepare(map, *startp, size, uobj, uoffset, align, 1199 flags, &args); 1200 if (!error) { 1201 error = uvm_map_enter(map, &args, new_entry); 1202 *startp = args.uma_start; 1203 } else if (new_entry) { 1204 uvm_mapent_free(new_entry); 1205 } 1206 1207 #if defined(DEBUG) 1208 if (!error && VM_MAP_IS_KERNEL(map)) { 1209 uvm_km_check_empty(map, *startp, *startp + size); 1210 } 1211 #endif /* defined(DEBUG) */ 1212 1213 return error; 1214 } 1215 1216 int 1217 uvm_map_prepare(struct vm_map *map, vaddr_t start, vsize_t size, 1218 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags, 1219 struct uvm_map_args *args) 1220 { 1221 struct vm_map_entry *prev_entry; 1222 vm_prot_t prot = UVM_PROTECTION(flags); 1223 vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1224 1225 UVMHIST_FUNC("uvm_map_prepare"); 1226 UVMHIST_CALLED(maphist); 1227 1228 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1229 map, start, size, flags); 1230 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1231 1232 /* 1233 * detect a popular device driver bug. 1234 */ 1235 1236 KASSERT(doing_shutdown || curlwp != NULL || 1237 (map->flags & VM_MAP_INTRSAFE)); 1238 1239 /* 1240 * zero-sized mapping doesn't make any sense. 1241 */ 1242 KASSERT(size > 0); 1243 1244 KASSERT((~flags & (UVM_FLAG_NOWAIT | UVM_FLAG_WAITVA)) != 0); 1245 1246 uvm_map_check(map, "map entry"); 1247 1248 /* 1249 * check sanity of protection code 1250 */ 1251 1252 if ((prot & maxprot) != prot) { 1253 UVMHIST_LOG(maphist, "<- prot. failure: prot=0x%x, max=0x%x", 1254 prot, maxprot,0,0); 1255 return EACCES; 1256 } 1257 1258 /* 1259 * figure out where to put new VM range 1260 */ 1261 1262 retry: 1263 if (vm_map_lock_try(map) == false) { 1264 if ((flags & UVM_FLAG_TRYLOCK) != 0 && 1265 (map->flags & VM_MAP_INTRSAFE) == 0) { 1266 return EAGAIN; 1267 } 1268 vm_map_lock(map); /* could sleep here */ 1269 } 1270 prev_entry = uvm_map_findspace(map, start, size, &start, 1271 uobj, uoffset, align, flags); 1272 if (prev_entry == NULL) { 1273 unsigned int timestamp; 1274 1275 timestamp = map->timestamp; 1276 UVMHIST_LOG(maphist,"waiting va timestamp=0x%x", 1277 timestamp,0,0,0); 1278 map->flags |= VM_MAP_WANTVA; 1279 vm_map_unlock(map); 1280 1281 /* 1282 * try to reclaim kva and wait until someone does unmap. 1283 * fragile locking here, so we awaken every second to 1284 * recheck the condition. 1285 */ 1286 1287 vm_map_drain(map, flags); 1288 1289 mutex_enter(&map->misc_lock); 1290 while ((map->flags & VM_MAP_WANTVA) != 0 && 1291 map->timestamp == timestamp) { 1292 if ((flags & UVM_FLAG_WAITVA) == 0) { 1293 mutex_exit(&map->misc_lock); 1294 UVMHIST_LOG(maphist, 1295 "<- uvm_map_findspace failed!", 0,0,0,0); 1296 return ENOMEM; 1297 } else { 1298 cv_timedwait(&map->cv, &map->misc_lock, hz); 1299 } 1300 } 1301 mutex_exit(&map->misc_lock); 1302 goto retry; 1303 } 1304 1305 #ifdef PMAP_GROWKERNEL 1306 /* 1307 * If the kernel pmap can't map the requested space, 1308 * then allocate more resources for it. 1309 */ 1310 if (map == kernel_map && uvm_maxkaddr < (start + size)) 1311 uvm_maxkaddr = pmap_growkernel(start + size); 1312 #endif 1313 1314 UVMMAP_EVCNT_INCR(map_call); 1315 1316 /* 1317 * if uobj is null, then uoffset is either a VAC hint for PMAP_PREFER 1318 * [typically from uvm_map_reserve] or it is UVM_UNKNOWN_OFFSET. in 1319 * either case we want to zero it before storing it in the map entry 1320 * (because it looks strange and confusing when debugging...) 1321 * 1322 * if uobj is not null 1323 * if uoffset is not UVM_UNKNOWN_OFFSET then we have a normal mapping 1324 * and we do not need to change uoffset. 1325 * if uoffset is UVM_UNKNOWN_OFFSET then we need to find the offset 1326 * now (based on the starting address of the map). this case is 1327 * for kernel object mappings where we don't know the offset until 1328 * the virtual address is found (with uvm_map_findspace). the 1329 * offset is the distance we are from the start of the map. 1330 */ 1331 1332 if (uobj == NULL) { 1333 uoffset = 0; 1334 } else { 1335 if (uoffset == UVM_UNKNOWN_OFFSET) { 1336 KASSERT(UVM_OBJ_IS_KERN_OBJECT(uobj)); 1337 uoffset = start - vm_map_min(kernel_map); 1338 } 1339 } 1340 1341 args->uma_flags = flags; 1342 args->uma_prev = prev_entry; 1343 args->uma_start = start; 1344 args->uma_size = size; 1345 args->uma_uobj = uobj; 1346 args->uma_uoffset = uoffset; 1347 1348 UVMHIST_LOG(maphist, "<- done!", 0,0,0,0); 1349 return 0; 1350 } 1351 1352 int 1353 uvm_map_enter(struct vm_map *map, const struct uvm_map_args *args, 1354 struct vm_map_entry *new_entry) 1355 { 1356 struct vm_map_entry *prev_entry = args->uma_prev; 1357 struct vm_map_entry *dead = NULL; 1358 1359 const uvm_flag_t flags = args->uma_flags; 1360 const vm_prot_t prot = UVM_PROTECTION(flags); 1361 const vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1362 const vm_inherit_t inherit = UVM_INHERIT(flags); 1363 const int amapwaitflag = (flags & UVM_FLAG_NOWAIT) ? 1364 AMAP_EXTEND_NOWAIT : 0; 1365 const int advice = UVM_ADVICE(flags); 1366 const int meflagval = (flags & UVM_FLAG_QUANTUM) ? 1367 UVM_MAP_QUANTUM : 0; 1368 1369 vaddr_t start = args->uma_start; 1370 vsize_t size = args->uma_size; 1371 struct uvm_object *uobj = args->uma_uobj; 1372 voff_t uoffset = args->uma_uoffset; 1373 1374 const int kmap = (vm_map_pmap(map) == pmap_kernel()); 1375 int merged = 0; 1376 int error; 1377 int newetype; 1378 1379 UVMHIST_FUNC("uvm_map_enter"); 1380 UVMHIST_CALLED(maphist); 1381 1382 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1383 map, start, size, flags); 1384 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1385 1386 KASSERT(map->hint == prev_entry); /* bimerge case assumes this */ 1387 1388 if (flags & UVM_FLAG_QUANTUM) { 1389 KASSERT(new_entry); 1390 KASSERT(new_entry->flags & UVM_MAP_QUANTUM); 1391 } 1392 1393 if (uobj) 1394 newetype = UVM_ET_OBJ; 1395 else 1396 newetype = 0; 1397 1398 if (flags & UVM_FLAG_COPYONW) { 1399 newetype |= UVM_ET_COPYONWRITE; 1400 if ((flags & UVM_FLAG_OVERLAY) == 0) 1401 newetype |= UVM_ET_NEEDSCOPY; 1402 } 1403 1404 /* 1405 * try and insert in map by extending previous entry, if possible. 1406 * XXX: we don't try and pull back the next entry. might be useful 1407 * for a stack, but we are currently allocating our stack in advance. 1408 */ 1409 1410 if (flags & UVM_FLAG_NOMERGE) 1411 goto nomerge; 1412 1413 if (prev_entry->end == start && 1414 prev_entry != &map->header && 1415 UVM_ET_ISCOMPATIBLE(prev_entry, newetype, uobj, meflagval, 1416 prot, maxprot, inherit, advice, 0)) { 1417 1418 if (uobj && prev_entry->offset + 1419 (prev_entry->end - prev_entry->start) != uoffset) 1420 goto forwardmerge; 1421 1422 /* 1423 * can't extend a shared amap. note: no need to lock amap to 1424 * look at refs since we don't care about its exact value. 1425 * if it is one (i.e. we have only reference) it will stay there 1426 */ 1427 1428 if (prev_entry->aref.ar_amap && 1429 amap_refs(prev_entry->aref.ar_amap) != 1) { 1430 goto forwardmerge; 1431 } 1432 1433 if (prev_entry->aref.ar_amap) { 1434 error = amap_extend(prev_entry, size, 1435 amapwaitflag | AMAP_EXTEND_FORWARDS); 1436 if (error) 1437 goto nomerge; 1438 } 1439 1440 if (kmap) { 1441 UVMMAP_EVCNT_INCR(kbackmerge); 1442 } else { 1443 UVMMAP_EVCNT_INCR(ubackmerge); 1444 } 1445 UVMHIST_LOG(maphist," starting back merge", 0, 0, 0, 0); 1446 1447 /* 1448 * drop our reference to uobj since we are extending a reference 1449 * that we already have (the ref count can not drop to zero). 1450 */ 1451 1452 if (uobj && uobj->pgops->pgo_detach) 1453 uobj->pgops->pgo_detach(uobj); 1454 1455 /* 1456 * Now that we've merged the entries, note that we've grown 1457 * and our gap has shrunk. Then fix the tree. 1458 */ 1459 prev_entry->end += size; 1460 prev_entry->gap -= size; 1461 uvm_rb_fixup(map, prev_entry); 1462 1463 uvm_map_check(map, "map backmerged"); 1464 1465 UVMHIST_LOG(maphist,"<- done (via backmerge)!", 0, 0, 0, 0); 1466 merged++; 1467 } 1468 1469 forwardmerge: 1470 if (prev_entry->next->start == (start + size) && 1471 prev_entry->next != &map->header && 1472 UVM_ET_ISCOMPATIBLE(prev_entry->next, newetype, uobj, meflagval, 1473 prot, maxprot, inherit, advice, 0)) { 1474 1475 if (uobj && prev_entry->next->offset != uoffset + size) 1476 goto nomerge; 1477 1478 /* 1479 * can't extend a shared amap. note: no need to lock amap to 1480 * look at refs since we don't care about its exact value. 1481 * if it is one (i.e. we have only reference) it will stay there. 1482 * 1483 * note that we also can't merge two amaps, so if we 1484 * merged with the previous entry which has an amap, 1485 * and the next entry also has an amap, we give up. 1486 * 1487 * Interesting cases: 1488 * amap, new, amap -> give up second merge (single fwd extend) 1489 * amap, new, none -> double forward extend (extend again here) 1490 * none, new, amap -> double backward extend (done here) 1491 * uobj, new, amap -> single backward extend (done here) 1492 * 1493 * XXX should we attempt to deal with someone refilling 1494 * the deallocated region between two entries that are 1495 * backed by the same amap (ie, arefs is 2, "prev" and 1496 * "next" refer to it, and adding this allocation will 1497 * close the hole, thus restoring arefs to 1 and 1498 * deallocating the "next" vm_map_entry)? -- @@@ 1499 */ 1500 1501 if (prev_entry->next->aref.ar_amap && 1502 (amap_refs(prev_entry->next->aref.ar_amap) != 1 || 1503 (merged && prev_entry->aref.ar_amap))) { 1504 goto nomerge; 1505 } 1506 1507 if (merged) { 1508 /* 1509 * Try to extend the amap of the previous entry to 1510 * cover the next entry as well. If it doesn't work 1511 * just skip on, don't actually give up, since we've 1512 * already completed the back merge. 1513 */ 1514 if (prev_entry->aref.ar_amap) { 1515 if (amap_extend(prev_entry, 1516 prev_entry->next->end - 1517 prev_entry->next->start, 1518 amapwaitflag | AMAP_EXTEND_FORWARDS)) 1519 goto nomerge; 1520 } 1521 1522 /* 1523 * Try to extend the amap of the *next* entry 1524 * back to cover the new allocation *and* the 1525 * previous entry as well (the previous merge 1526 * didn't have an amap already otherwise we 1527 * wouldn't be checking here for an amap). If 1528 * it doesn't work just skip on, again, don't 1529 * actually give up, since we've already 1530 * completed the back merge. 1531 */ 1532 else if (prev_entry->next->aref.ar_amap) { 1533 if (amap_extend(prev_entry->next, 1534 prev_entry->end - 1535 prev_entry->start, 1536 amapwaitflag | AMAP_EXTEND_BACKWARDS)) 1537 goto nomerge; 1538 } 1539 } else { 1540 /* 1541 * Pull the next entry's amap backwards to cover this 1542 * new allocation. 1543 */ 1544 if (prev_entry->next->aref.ar_amap) { 1545 error = amap_extend(prev_entry->next, size, 1546 amapwaitflag | AMAP_EXTEND_BACKWARDS); 1547 if (error) 1548 goto nomerge; 1549 } 1550 } 1551 1552 if (merged) { 1553 if (kmap) { 1554 UVMMAP_EVCNT_DECR(kbackmerge); 1555 UVMMAP_EVCNT_INCR(kbimerge); 1556 } else { 1557 UVMMAP_EVCNT_DECR(ubackmerge); 1558 UVMMAP_EVCNT_INCR(ubimerge); 1559 } 1560 } else { 1561 if (kmap) { 1562 UVMMAP_EVCNT_INCR(kforwmerge); 1563 } else { 1564 UVMMAP_EVCNT_INCR(uforwmerge); 1565 } 1566 } 1567 UVMHIST_LOG(maphist," starting forward merge", 0, 0, 0, 0); 1568 1569 /* 1570 * drop our reference to uobj since we are extending a reference 1571 * that we already have (the ref count can not drop to zero). 1572 * (if merged, we've already detached) 1573 */ 1574 if (uobj && uobj->pgops->pgo_detach && !merged) 1575 uobj->pgops->pgo_detach(uobj); 1576 1577 if (merged) { 1578 dead = prev_entry->next; 1579 prev_entry->end = dead->end; 1580 uvm_map_entry_unlink(map, dead); 1581 if (dead->aref.ar_amap != NULL) { 1582 prev_entry->aref = dead->aref; 1583 dead->aref.ar_amap = NULL; 1584 } 1585 } else { 1586 prev_entry->next->start -= size; 1587 if (prev_entry != &map->header) { 1588 prev_entry->gap -= size; 1589 KASSERT(prev_entry->gap == uvm_rb_gap(prev_entry)); 1590 uvm_rb_fixup(map, prev_entry); 1591 } 1592 if (uobj) 1593 prev_entry->next->offset = uoffset; 1594 } 1595 1596 uvm_map_check(map, "map forwardmerged"); 1597 1598 UVMHIST_LOG(maphist,"<- done forwardmerge", 0, 0, 0, 0); 1599 merged++; 1600 } 1601 1602 nomerge: 1603 if (!merged) { 1604 UVMHIST_LOG(maphist," allocating new map entry", 0, 0, 0, 0); 1605 if (kmap) { 1606 UVMMAP_EVCNT_INCR(knomerge); 1607 } else { 1608 UVMMAP_EVCNT_INCR(unomerge); 1609 } 1610 1611 /* 1612 * allocate new entry and link it in. 1613 */ 1614 1615 if (new_entry == NULL) { 1616 new_entry = uvm_mapent_alloc(map, 1617 (flags & UVM_FLAG_NOWAIT)); 1618 if (__predict_false(new_entry == NULL)) { 1619 error = ENOMEM; 1620 goto done; 1621 } 1622 } 1623 new_entry->start = start; 1624 new_entry->end = new_entry->start + size; 1625 new_entry->object.uvm_obj = uobj; 1626 new_entry->offset = uoffset; 1627 1628 new_entry->etype = newetype; 1629 1630 if (flags & UVM_FLAG_NOMERGE) { 1631 new_entry->flags |= UVM_MAP_NOMERGE; 1632 } 1633 1634 new_entry->protection = prot; 1635 new_entry->max_protection = maxprot; 1636 new_entry->inheritance = inherit; 1637 new_entry->wired_count = 0; 1638 new_entry->advice = advice; 1639 if (flags & UVM_FLAG_OVERLAY) { 1640 1641 /* 1642 * to_add: for BSS we overallocate a little since we 1643 * are likely to extend 1644 */ 1645 1646 vaddr_t to_add = (flags & UVM_FLAG_AMAPPAD) ? 1647 UVM_AMAP_CHUNK << PAGE_SHIFT : 0; 1648 struct vm_amap *amap = amap_alloc(size, to_add, 1649 (flags & UVM_FLAG_NOWAIT)); 1650 if (__predict_false(amap == NULL)) { 1651 error = ENOMEM; 1652 goto done; 1653 } 1654 new_entry->aref.ar_pageoff = 0; 1655 new_entry->aref.ar_amap = amap; 1656 } else { 1657 new_entry->aref.ar_pageoff = 0; 1658 new_entry->aref.ar_amap = NULL; 1659 } 1660 uvm_map_entry_link(map, prev_entry, new_entry); 1661 1662 /* 1663 * Update the free space hint 1664 */ 1665 1666 if ((map->first_free == prev_entry) && 1667 (prev_entry->end >= new_entry->start)) 1668 map->first_free = new_entry; 1669 1670 new_entry = NULL; 1671 } 1672 1673 map->size += size; 1674 1675 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 1676 1677 error = 0; 1678 done: 1679 if ((flags & UVM_FLAG_QUANTUM) == 0) { 1680 /* 1681 * vmk_merged_entries is locked by the map's lock. 1682 */ 1683 vm_map_unlock(map); 1684 } 1685 if (new_entry && error == 0) { 1686 KDASSERT(merged); 1687 uvm_mapent_free_merged(map, new_entry); 1688 new_entry = NULL; 1689 } 1690 if (dead) { 1691 KDASSERT(merged); 1692 uvm_mapent_free_merged(map, dead); 1693 } 1694 if ((flags & UVM_FLAG_QUANTUM) != 0) { 1695 vm_map_unlock(map); 1696 } 1697 if (new_entry != NULL) { 1698 uvm_mapent_free(new_entry); 1699 } 1700 return error; 1701 } 1702 1703 /* 1704 * uvm_map_lookup_entry_bytree: lookup an entry in tree 1705 */ 1706 1707 static inline bool 1708 uvm_map_lookup_entry_bytree(struct vm_map *map, vaddr_t address, 1709 struct vm_map_entry **entry /* OUT */) 1710 { 1711 struct vm_map_entry *prev = &map->header; 1712 struct vm_map_entry *cur = ROOT_ENTRY(map); 1713 1714 while (cur) { 1715 UVMMAP_EVCNT_INCR(mlk_treeloop); 1716 if (address >= cur->start) { 1717 if (address < cur->end) { 1718 *entry = cur; 1719 return true; 1720 } 1721 prev = cur; 1722 cur = RIGHT_ENTRY(cur); 1723 } else 1724 cur = LEFT_ENTRY(cur); 1725 } 1726 *entry = prev; 1727 return false; 1728 } 1729 1730 /* 1731 * uvm_map_lookup_entry: find map entry at or before an address 1732 * 1733 * => map must at least be read-locked by caller 1734 * => entry is returned in "entry" 1735 * => return value is true if address is in the returned entry 1736 */ 1737 1738 bool 1739 uvm_map_lookup_entry(struct vm_map *map, vaddr_t address, 1740 struct vm_map_entry **entry /* OUT */) 1741 { 1742 struct vm_map_entry *cur; 1743 bool use_tree = false; 1744 UVMHIST_FUNC("uvm_map_lookup_entry"); 1745 UVMHIST_CALLED(maphist); 1746 1747 UVMHIST_LOG(maphist,"(map=0x%x,addr=0x%x,ent=0x%x)", 1748 map, address, entry, 0); 1749 1750 /* 1751 * start looking either from the head of the 1752 * list, or from the hint. 1753 */ 1754 1755 cur = map->hint; 1756 1757 if (cur == &map->header) 1758 cur = cur->next; 1759 1760 UVMMAP_EVCNT_INCR(mlk_call); 1761 if (address >= cur->start) { 1762 1763 /* 1764 * go from hint to end of list. 1765 * 1766 * but first, make a quick check to see if 1767 * we are already looking at the entry we 1768 * want (which is usually the case). 1769 * note also that we don't need to save the hint 1770 * here... it is the same hint (unless we are 1771 * at the header, in which case the hint didn't 1772 * buy us anything anyway). 1773 */ 1774 1775 if (cur != &map->header && cur->end > address) { 1776 UVMMAP_EVCNT_INCR(mlk_hint); 1777 *entry = cur; 1778 UVMHIST_LOG(maphist,"<- got it via hint (0x%x)", 1779 cur, 0, 0, 0); 1780 uvm_mapent_check(*entry); 1781 return (true); 1782 } 1783 1784 if (map->nentries > 15) 1785 use_tree = true; 1786 } else { 1787 1788 /* 1789 * invalid hint. use tree. 1790 */ 1791 use_tree = true; 1792 } 1793 1794 uvm_map_check(map, __func__); 1795 1796 if (use_tree) { 1797 /* 1798 * Simple lookup in the tree. Happens when the hint is 1799 * invalid, or nentries reach a threshold. 1800 */ 1801 UVMMAP_EVCNT_INCR(mlk_tree); 1802 if (uvm_map_lookup_entry_bytree(map, address, entry)) { 1803 goto got; 1804 } else { 1805 goto failed; 1806 } 1807 } 1808 1809 /* 1810 * search linearly 1811 */ 1812 1813 UVMMAP_EVCNT_INCR(mlk_list); 1814 while (cur != &map->header) { 1815 UVMMAP_EVCNT_INCR(mlk_listloop); 1816 if (cur->end > address) { 1817 if (address >= cur->start) { 1818 /* 1819 * save this lookup for future 1820 * hints, and return 1821 */ 1822 1823 *entry = cur; 1824 got: 1825 SAVE_HINT(map, map->hint, *entry); 1826 UVMHIST_LOG(maphist,"<- search got it (0x%x)", 1827 cur, 0, 0, 0); 1828 KDASSERT((*entry)->start <= address); 1829 KDASSERT(address < (*entry)->end); 1830 uvm_mapent_check(*entry); 1831 return (true); 1832 } 1833 break; 1834 } 1835 cur = cur->next; 1836 } 1837 *entry = cur->prev; 1838 failed: 1839 SAVE_HINT(map, map->hint, *entry); 1840 UVMHIST_LOG(maphist,"<- failed!",0,0,0,0); 1841 KDASSERT((*entry) == &map->header || (*entry)->end <= address); 1842 KDASSERT((*entry)->next == &map->header || 1843 address < (*entry)->next->start); 1844 return (false); 1845 } 1846 1847 /* 1848 * See if the range between start and start + length fits in the gap 1849 * entry->next->start and entry->end. Returns 1 if fits, 0 if doesn't 1850 * fit, and -1 address wraps around. 1851 */ 1852 static int 1853 uvm_map_space_avail(vaddr_t *start, vsize_t length, voff_t uoffset, 1854 vsize_t align, int topdown, struct vm_map_entry *entry) 1855 { 1856 vaddr_t end; 1857 1858 #ifdef PMAP_PREFER 1859 /* 1860 * push start address forward as needed to avoid VAC alias problems. 1861 * we only do this if a valid offset is specified. 1862 */ 1863 1864 if (uoffset != UVM_UNKNOWN_OFFSET) 1865 PMAP_PREFER(uoffset, start, length, topdown); 1866 #endif 1867 if (align != 0) { 1868 if ((*start & (align - 1)) != 0) { 1869 if (topdown) 1870 *start &= ~(align - 1); 1871 else 1872 *start = roundup(*start, align); 1873 } 1874 /* 1875 * XXX Should we PMAP_PREFER() here again? 1876 * eh...i think we're okay 1877 */ 1878 } 1879 1880 /* 1881 * Find the end of the proposed new region. Be sure we didn't 1882 * wrap around the address; if so, we lose. Otherwise, if the 1883 * proposed new region fits before the next entry, we win. 1884 */ 1885 1886 end = *start + length; 1887 if (end < *start) 1888 return (-1); 1889 1890 if (entry->next->start >= end && *start >= entry->end) 1891 return (1); 1892 1893 return (0); 1894 } 1895 1896 /* 1897 * uvm_map_findspace: find "length" sized space in "map". 1898 * 1899 * => "hint" is a hint about where we want it, unless UVM_FLAG_FIXED is 1900 * set in "flags" (in which case we insist on using "hint"). 1901 * => "result" is VA returned 1902 * => uobj/uoffset are to be used to handle VAC alignment, if required 1903 * => if "align" is non-zero, we attempt to align to that value. 1904 * => caller must at least have read-locked map 1905 * => returns NULL on failure, or pointer to prev. map entry if success 1906 * => note this is a cross between the old vm_map_findspace and vm_map_find 1907 */ 1908 1909 struct vm_map_entry * 1910 uvm_map_findspace(struct vm_map *map, vaddr_t hint, vsize_t length, 1911 vaddr_t *result /* OUT */, struct uvm_object *uobj, voff_t uoffset, 1912 vsize_t align, int flags) 1913 { 1914 struct vm_map_entry *entry; 1915 struct vm_map_entry *child, *prev, *tmp; 1916 vaddr_t orig_hint; 1917 const int topdown = map->flags & VM_MAP_TOPDOWN; 1918 UVMHIST_FUNC("uvm_map_findspace"); 1919 UVMHIST_CALLED(maphist); 1920 1921 UVMHIST_LOG(maphist, "(map=0x%x, hint=0x%x, len=%d, flags=0x%x)", 1922 map, hint, length, flags); 1923 KASSERT((align & (align - 1)) == 0); 1924 KASSERT((flags & UVM_FLAG_FIXED) == 0 || align == 0); 1925 1926 uvm_map_check(map, "map_findspace entry"); 1927 1928 /* 1929 * remember the original hint. if we are aligning, then we 1930 * may have to try again with no alignment constraint if 1931 * we fail the first time. 1932 */ 1933 1934 orig_hint = hint; 1935 if (hint < vm_map_min(map)) { /* check ranges ... */ 1936 if (flags & UVM_FLAG_FIXED) { 1937 UVMHIST_LOG(maphist,"<- VA below map range",0,0,0,0); 1938 return (NULL); 1939 } 1940 hint = vm_map_min(map); 1941 } 1942 if (hint > vm_map_max(map)) { 1943 UVMHIST_LOG(maphist,"<- VA 0x%x > range [0x%x->0x%x]", 1944 hint, vm_map_min(map), vm_map_max(map), 0); 1945 return (NULL); 1946 } 1947 1948 /* 1949 * Look for the first possible address; if there's already 1950 * something at this address, we have to start after it. 1951 */ 1952 1953 /* 1954 * @@@: there are four, no, eight cases to consider. 1955 * 1956 * 0: found, fixed, bottom up -> fail 1957 * 1: found, fixed, top down -> fail 1958 * 2: found, not fixed, bottom up -> start after entry->end, 1959 * loop up 1960 * 3: found, not fixed, top down -> start before entry->start, 1961 * loop down 1962 * 4: not found, fixed, bottom up -> check entry->next->start, fail 1963 * 5: not found, fixed, top down -> check entry->next->start, fail 1964 * 6: not found, not fixed, bottom up -> check entry->next->start, 1965 * loop up 1966 * 7: not found, not fixed, top down -> check entry->next->start, 1967 * loop down 1968 * 1969 * as you can see, it reduces to roughly five cases, and that 1970 * adding top down mapping only adds one unique case (without 1971 * it, there would be four cases). 1972 */ 1973 1974 if ((flags & UVM_FLAG_FIXED) == 0 && hint == vm_map_min(map)) { 1975 entry = map->first_free; 1976 } else { 1977 if (uvm_map_lookup_entry(map, hint, &entry)) { 1978 /* "hint" address already in use ... */ 1979 if (flags & UVM_FLAG_FIXED) { 1980 UVMHIST_LOG(maphist, "<- fixed & VA in use", 1981 0, 0, 0, 0); 1982 return (NULL); 1983 } 1984 if (topdown) 1985 /* Start from lower gap. */ 1986 entry = entry->prev; 1987 } else if (flags & UVM_FLAG_FIXED) { 1988 if (entry->next->start >= hint + length && 1989 hint + length > hint) 1990 goto found; 1991 1992 /* "hint" address is gap but too small */ 1993 UVMHIST_LOG(maphist, "<- fixed mapping failed", 1994 0, 0, 0, 0); 1995 return (NULL); /* only one shot at it ... */ 1996 } else { 1997 /* 1998 * See if given hint fits in this gap. 1999 */ 2000 switch (uvm_map_space_avail(&hint, length, 2001 uoffset, align, topdown, entry)) { 2002 case 1: 2003 goto found; 2004 case -1: 2005 goto wraparound; 2006 } 2007 2008 if (topdown) { 2009 /* 2010 * Still there is a chance to fit 2011 * if hint > entry->end. 2012 */ 2013 } else { 2014 /* Start from higher gap. */ 2015 entry = entry->next; 2016 if (entry == &map->header) 2017 goto notfound; 2018 goto nextgap; 2019 } 2020 } 2021 } 2022 2023 /* 2024 * Note that all UVM_FLAGS_FIXED case is already handled. 2025 */ 2026 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2027 2028 /* Try to find the space in the red-black tree */ 2029 2030 /* Check slot before any entry */ 2031 hint = topdown ? entry->next->start - length : entry->end; 2032 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2033 topdown, entry)) { 2034 case 1: 2035 goto found; 2036 case -1: 2037 goto wraparound; 2038 } 2039 2040 nextgap: 2041 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2042 /* If there is not enough space in the whole tree, we fail */ 2043 tmp = ROOT_ENTRY(map); 2044 if (tmp == NULL || tmp->maxgap < length) 2045 goto notfound; 2046 2047 prev = NULL; /* previous candidate */ 2048 2049 /* Find an entry close to hint that has enough space */ 2050 for (; tmp;) { 2051 KASSERT(tmp->next->start == tmp->end + tmp->gap); 2052 if (topdown) { 2053 if (tmp->next->start < hint + length && 2054 (prev == NULL || tmp->end > prev->end)) { 2055 if (tmp->gap >= length) 2056 prev = tmp; 2057 else if ((child = LEFT_ENTRY(tmp)) != NULL 2058 && child->maxgap >= length) 2059 prev = tmp; 2060 } 2061 } else { 2062 if (tmp->end >= hint && 2063 (prev == NULL || tmp->end < prev->end)) { 2064 if (tmp->gap >= length) 2065 prev = tmp; 2066 else if ((child = RIGHT_ENTRY(tmp)) != NULL 2067 && child->maxgap >= length) 2068 prev = tmp; 2069 } 2070 } 2071 if (tmp->next->start < hint + length) 2072 child = RIGHT_ENTRY(tmp); 2073 else if (tmp->end > hint) 2074 child = LEFT_ENTRY(tmp); 2075 else { 2076 if (tmp->gap >= length) 2077 break; 2078 if (topdown) 2079 child = LEFT_ENTRY(tmp); 2080 else 2081 child = RIGHT_ENTRY(tmp); 2082 } 2083 if (child == NULL || child->maxgap < length) 2084 break; 2085 tmp = child; 2086 } 2087 2088 if (tmp != NULL && tmp->start < hint && hint < tmp->next->start) { 2089 /* 2090 * Check if the entry that we found satifies the 2091 * space requirement 2092 */ 2093 if (topdown) { 2094 if (hint > tmp->next->start - length) 2095 hint = tmp->next->start - length; 2096 } else { 2097 if (hint < tmp->end) 2098 hint = tmp->end; 2099 } 2100 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2101 topdown, tmp)) { 2102 case 1: 2103 entry = tmp; 2104 goto found; 2105 case -1: 2106 goto wraparound; 2107 } 2108 if (tmp->gap >= length) 2109 goto listsearch; 2110 } 2111 if (prev == NULL) 2112 goto notfound; 2113 2114 if (topdown) { 2115 KASSERT(orig_hint >= prev->next->start - length || 2116 prev->next->start - length > prev->next->start); 2117 hint = prev->next->start - length; 2118 } else { 2119 KASSERT(orig_hint <= prev->end); 2120 hint = prev->end; 2121 } 2122 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2123 topdown, prev)) { 2124 case 1: 2125 entry = prev; 2126 goto found; 2127 case -1: 2128 goto wraparound; 2129 } 2130 if (prev->gap >= length) 2131 goto listsearch; 2132 2133 if (topdown) 2134 tmp = LEFT_ENTRY(prev); 2135 else 2136 tmp = RIGHT_ENTRY(prev); 2137 for (;;) { 2138 KASSERT(tmp && tmp->maxgap >= length); 2139 if (topdown) 2140 child = RIGHT_ENTRY(tmp); 2141 else 2142 child = LEFT_ENTRY(tmp); 2143 if (child && child->maxgap >= length) { 2144 tmp = child; 2145 continue; 2146 } 2147 if (tmp->gap >= length) 2148 break; 2149 if (topdown) 2150 tmp = LEFT_ENTRY(tmp); 2151 else 2152 tmp = RIGHT_ENTRY(tmp); 2153 } 2154 2155 if (topdown) { 2156 KASSERT(orig_hint >= tmp->next->start - length || 2157 tmp->next->start - length > tmp->next->start); 2158 hint = tmp->next->start - length; 2159 } else { 2160 KASSERT(orig_hint <= tmp->end); 2161 hint = tmp->end; 2162 } 2163 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2164 topdown, tmp)) { 2165 case 1: 2166 entry = tmp; 2167 goto found; 2168 case -1: 2169 goto wraparound; 2170 } 2171 2172 /* 2173 * The tree fails to find an entry because of offset or alignment 2174 * restrictions. Search the list instead. 2175 */ 2176 listsearch: 2177 /* 2178 * Look through the rest of the map, trying to fit a new region in 2179 * the gap between existing regions, or after the very last region. 2180 * note: entry->end = base VA of current gap, 2181 * entry->next->start = VA of end of current gap 2182 */ 2183 2184 for (;;) { 2185 /* Update hint for current gap. */ 2186 hint = topdown ? entry->next->start - length : entry->end; 2187 2188 /* See if it fits. */ 2189 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2190 topdown, entry)) { 2191 case 1: 2192 goto found; 2193 case -1: 2194 goto wraparound; 2195 } 2196 2197 /* Advance to next/previous gap */ 2198 if (topdown) { 2199 if (entry == &map->header) { 2200 UVMHIST_LOG(maphist, "<- failed (off start)", 2201 0,0,0,0); 2202 goto notfound; 2203 } 2204 entry = entry->prev; 2205 } else { 2206 entry = entry->next; 2207 if (entry == &map->header) { 2208 UVMHIST_LOG(maphist, "<- failed (off end)", 2209 0,0,0,0); 2210 goto notfound; 2211 } 2212 } 2213 } 2214 2215 found: 2216 SAVE_HINT(map, map->hint, entry); 2217 *result = hint; 2218 UVMHIST_LOG(maphist,"<- got it! (result=0x%x)", hint, 0,0,0); 2219 KASSERT( topdown || hint >= orig_hint); 2220 KASSERT(!topdown || hint <= orig_hint); 2221 KASSERT(entry->end <= hint); 2222 KASSERT(hint + length <= entry->next->start); 2223 return (entry); 2224 2225 wraparound: 2226 UVMHIST_LOG(maphist, "<- failed (wrap around)", 0,0,0,0); 2227 2228 return (NULL); 2229 2230 notfound: 2231 UVMHIST_LOG(maphist, "<- failed (notfound)", 0,0,0,0); 2232 2233 return (NULL); 2234 } 2235 2236 /* 2237 * U N M A P - m a i n h e l p e r f u n c t i o n s 2238 */ 2239 2240 /* 2241 * uvm_unmap_remove: remove mappings from a vm_map (from "start" up to "stop") 2242 * 2243 * => caller must check alignment and size 2244 * => map must be locked by caller 2245 * => we return a list of map entries that we've remove from the map 2246 * in "entry_list" 2247 */ 2248 2249 void 2250 uvm_unmap_remove(struct vm_map *map, vaddr_t start, vaddr_t end, 2251 struct vm_map_entry **entry_list /* OUT */, 2252 struct uvm_mapent_reservation *umr, int flags) 2253 { 2254 struct vm_map_entry *entry, *first_entry, *next; 2255 vaddr_t len; 2256 UVMHIST_FUNC("uvm_unmap_remove"); UVMHIST_CALLED(maphist); 2257 2258 UVMHIST_LOG(maphist,"(map=0x%x, start=0x%x, end=0x%x)", 2259 map, start, end, 0); 2260 VM_MAP_RANGE_CHECK(map, start, end); 2261 2262 uvm_map_check(map, "unmap_remove entry"); 2263 2264 /* 2265 * find first entry 2266 */ 2267 2268 if (uvm_map_lookup_entry(map, start, &first_entry) == true) { 2269 /* clip and go... */ 2270 entry = first_entry; 2271 UVM_MAP_CLIP_START(map, entry, start, umr); 2272 /* critical! prevents stale hint */ 2273 SAVE_HINT(map, entry, entry->prev); 2274 } else { 2275 entry = first_entry->next; 2276 } 2277 2278 /* 2279 * Save the free space hint 2280 */ 2281 2282 if (map->first_free != &map->header && map->first_free->start >= start) 2283 map->first_free = entry->prev; 2284 2285 /* 2286 * note: we now re-use first_entry for a different task. we remove 2287 * a number of map entries from the map and save them in a linked 2288 * list headed by "first_entry". once we remove them from the map 2289 * the caller should unlock the map and drop the references to the 2290 * backing objects [c.f. uvm_unmap_detach]. the object is to 2291 * separate unmapping from reference dropping. why? 2292 * [1] the map has to be locked for unmapping 2293 * [2] the map need not be locked for reference dropping 2294 * [3] dropping references may trigger pager I/O, and if we hit 2295 * a pager that does synchronous I/O we may have to wait for it. 2296 * [4] we would like all waiting for I/O to occur with maps unlocked 2297 * so that we don't block other threads. 2298 */ 2299 2300 first_entry = NULL; 2301 *entry_list = NULL; 2302 2303 /* 2304 * break up the area into map entry sized regions and unmap. note 2305 * that all mappings have to be removed before we can even consider 2306 * dropping references to amaps or VM objects (otherwise we could end 2307 * up with a mapping to a page on the free list which would be very bad) 2308 */ 2309 2310 while ((entry != &map->header) && (entry->start < end)) { 2311 KASSERT((entry->flags & UVM_MAP_FIRST) == 0); 2312 2313 UVM_MAP_CLIP_END(map, entry, end, umr); 2314 next = entry->next; 2315 len = entry->end - entry->start; 2316 2317 /* 2318 * unwire before removing addresses from the pmap; otherwise 2319 * unwiring will put the entries back into the pmap (XXX). 2320 */ 2321 2322 if (VM_MAPENT_ISWIRED(entry)) { 2323 uvm_map_entry_unwire(map, entry); 2324 } 2325 if (flags & UVM_FLAG_VAONLY) { 2326 2327 /* nothing */ 2328 2329 } else if ((map->flags & VM_MAP_PAGEABLE) == 0) { 2330 2331 /* 2332 * if the map is non-pageable, any pages mapped there 2333 * must be wired and entered with pmap_kenter_pa(), 2334 * and we should free any such pages immediately. 2335 * this is mostly used for kmem_map. 2336 */ 2337 2338 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2339 uvm_km_pgremove_intrsafe(map, entry->start, 2340 entry->end); 2341 pmap_kremove(entry->start, len); 2342 } 2343 } else if (UVM_ET_ISOBJ(entry) && 2344 UVM_OBJ_IS_KERN_OBJECT(entry->object.uvm_obj)) { 2345 KASSERT(vm_map_pmap(map) == pmap_kernel()); 2346 2347 /* 2348 * note: kernel object mappings are currently used in 2349 * two ways: 2350 * [1] "normal" mappings of pages in the kernel object 2351 * [2] uvm_km_valloc'd allocations in which we 2352 * pmap_enter in some non-kernel-object page 2353 * (e.g. vmapbuf). 2354 * 2355 * for case [1], we need to remove the mapping from 2356 * the pmap and then remove the page from the kernel 2357 * object (because, once pages in a kernel object are 2358 * unmapped they are no longer needed, unlike, say, 2359 * a vnode where you might want the data to persist 2360 * until flushed out of a queue). 2361 * 2362 * for case [2], we need to remove the mapping from 2363 * the pmap. there shouldn't be any pages at the 2364 * specified offset in the kernel object [but it 2365 * doesn't hurt to call uvm_km_pgremove just to be 2366 * safe?] 2367 * 2368 * uvm_km_pgremove currently does the following: 2369 * for pages in the kernel object in range: 2370 * - drops the swap slot 2371 * - uvm_pagefree the page 2372 */ 2373 2374 /* 2375 * remove mappings from pmap and drop the pages 2376 * from the object. offsets are always relative 2377 * to vm_map_min(kernel_map). 2378 */ 2379 2380 pmap_remove(pmap_kernel(), entry->start, 2381 entry->start + len); 2382 uvm_km_pgremove(entry->start, entry->end); 2383 2384 /* 2385 * null out kernel_object reference, we've just 2386 * dropped it 2387 */ 2388 2389 entry->etype &= ~UVM_ET_OBJ; 2390 entry->object.uvm_obj = NULL; 2391 } else if (UVM_ET_ISOBJ(entry) || entry->aref.ar_amap) { 2392 2393 /* 2394 * remove mappings the standard way. 2395 */ 2396 2397 pmap_remove(map->pmap, entry->start, entry->end); 2398 } 2399 2400 #if defined(DEBUG) 2401 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2402 2403 /* 2404 * check if there's remaining mapping, 2405 * which is a bug in caller. 2406 */ 2407 2408 vaddr_t va; 2409 for (va = entry->start; va < entry->end; 2410 va += PAGE_SIZE) { 2411 if (pmap_extract(vm_map_pmap(map), va, NULL)) { 2412 panic("uvm_unmap_remove: has mapping"); 2413 } 2414 } 2415 2416 if (VM_MAP_IS_KERNEL(map)) { 2417 uvm_km_check_empty(map, entry->start, 2418 entry->end); 2419 } 2420 } 2421 #endif /* defined(DEBUG) */ 2422 2423 /* 2424 * remove entry from map and put it on our list of entries 2425 * that we've nuked. then go to next entry. 2426 */ 2427 2428 UVMHIST_LOG(maphist, " removed map entry 0x%x", entry, 0, 0,0); 2429 2430 /* critical! prevents stale hint */ 2431 SAVE_HINT(map, entry, entry->prev); 2432 2433 uvm_map_entry_unlink(map, entry); 2434 KASSERT(map->size >= len); 2435 map->size -= len; 2436 entry->prev = NULL; 2437 entry->next = first_entry; 2438 first_entry = entry; 2439 entry = next; 2440 } 2441 if ((map->flags & VM_MAP_DYING) == 0) { 2442 pmap_update(vm_map_pmap(map)); 2443 } 2444 2445 uvm_map_check(map, "unmap_remove leave"); 2446 2447 /* 2448 * now we've cleaned up the map and are ready for the caller to drop 2449 * references to the mapped objects. 2450 */ 2451 2452 *entry_list = first_entry; 2453 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 2454 2455 if (map->flags & VM_MAP_WANTVA) { 2456 mutex_enter(&map->misc_lock); 2457 map->flags &= ~VM_MAP_WANTVA; 2458 cv_broadcast(&map->cv); 2459 mutex_exit(&map->misc_lock); 2460 } 2461 } 2462 2463 /* 2464 * uvm_unmap_detach: drop references in a chain of map entries 2465 * 2466 * => we will free the map entries as we traverse the list. 2467 */ 2468 2469 void 2470 uvm_unmap_detach(struct vm_map_entry *first_entry, int flags) 2471 { 2472 struct vm_map_entry *next_entry; 2473 UVMHIST_FUNC("uvm_unmap_detach"); UVMHIST_CALLED(maphist); 2474 2475 while (first_entry) { 2476 KASSERT(!VM_MAPENT_ISWIRED(first_entry)); 2477 UVMHIST_LOG(maphist, 2478 " detach 0x%x: amap=0x%x, obj=0x%x, submap?=%d", 2479 first_entry, first_entry->aref.ar_amap, 2480 first_entry->object.uvm_obj, 2481 UVM_ET_ISSUBMAP(first_entry)); 2482 2483 /* 2484 * drop reference to amap, if we've got one 2485 */ 2486 2487 if (first_entry->aref.ar_amap) 2488 uvm_map_unreference_amap(first_entry, flags); 2489 2490 /* 2491 * drop reference to our backing object, if we've got one 2492 */ 2493 2494 KASSERT(!UVM_ET_ISSUBMAP(first_entry)); 2495 if (UVM_ET_ISOBJ(first_entry) && 2496 first_entry->object.uvm_obj->pgops->pgo_detach) { 2497 (*first_entry->object.uvm_obj->pgops->pgo_detach) 2498 (first_entry->object.uvm_obj); 2499 } 2500 next_entry = first_entry->next; 2501 uvm_mapent_free(first_entry); 2502 first_entry = next_entry; 2503 } 2504 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 2505 } 2506 2507 /* 2508 * E X T R A C T I O N F U N C T I O N S 2509 */ 2510 2511 /* 2512 * uvm_map_reserve: reserve space in a vm_map for future use. 2513 * 2514 * => we reserve space in a map by putting a dummy map entry in the 2515 * map (dummy means obj=NULL, amap=NULL, prot=VM_PROT_NONE) 2516 * => map should be unlocked (we will write lock it) 2517 * => we return true if we were able to reserve space 2518 * => XXXCDC: should be inline? 2519 */ 2520 2521 int 2522 uvm_map_reserve(struct vm_map *map, vsize_t size, 2523 vaddr_t offset /* hint for pmap_prefer */, 2524 vsize_t align /* alignment */, 2525 vaddr_t *raddr /* IN:hint, OUT: reserved VA */, 2526 uvm_flag_t flags /* UVM_FLAG_FIXED or 0 */) 2527 { 2528 UVMHIST_FUNC("uvm_map_reserve"); UVMHIST_CALLED(maphist); 2529 2530 UVMHIST_LOG(maphist, "(map=0x%x, size=0x%x, offset=0x%x,addr=0x%x)", 2531 map,size,offset,raddr); 2532 2533 size = round_page(size); 2534 2535 /* 2536 * reserve some virtual space. 2537 */ 2538 2539 if (uvm_map(map, raddr, size, NULL, offset, align, 2540 UVM_MAPFLAG(UVM_PROT_NONE, UVM_PROT_NONE, UVM_INH_NONE, 2541 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE|flags)) != 0) { 2542 UVMHIST_LOG(maphist, "<- done (no VM)", 0,0,0,0); 2543 return (false); 2544 } 2545 2546 UVMHIST_LOG(maphist, "<- done (*raddr=0x%x)", *raddr,0,0,0); 2547 return (true); 2548 } 2549 2550 /* 2551 * uvm_map_replace: replace a reserved (blank) area of memory with 2552 * real mappings. 2553 * 2554 * => caller must WRITE-LOCK the map 2555 * => we return true if replacement was a success 2556 * => we expect the newents chain to have nnewents entrys on it and 2557 * we expect newents->prev to point to the last entry on the list 2558 * => note newents is allowed to be NULL 2559 */ 2560 2561 static int 2562 uvm_map_replace(struct vm_map *map, vaddr_t start, vaddr_t end, 2563 struct vm_map_entry *newents, int nnewents, vsize_t nsize, 2564 struct vm_map_entry **oldentryp) 2565 { 2566 struct vm_map_entry *oldent, *last; 2567 2568 uvm_map_check(map, "map_replace entry"); 2569 2570 /* 2571 * first find the blank map entry at the specified address 2572 */ 2573 2574 if (!uvm_map_lookup_entry(map, start, &oldent)) { 2575 return (false); 2576 } 2577 2578 /* 2579 * check to make sure we have a proper blank entry 2580 */ 2581 2582 if (end < oldent->end && !VM_MAP_USE_KMAPENT(map)) { 2583 UVM_MAP_CLIP_END(map, oldent, end, NULL); 2584 } 2585 if (oldent->start != start || oldent->end != end || 2586 oldent->object.uvm_obj != NULL || oldent->aref.ar_amap != NULL) { 2587 return (false); 2588 } 2589 2590 #ifdef DIAGNOSTIC 2591 2592 /* 2593 * sanity check the newents chain 2594 */ 2595 2596 { 2597 struct vm_map_entry *tmpent = newents; 2598 int nent = 0; 2599 vsize_t sz = 0; 2600 vaddr_t cur = start; 2601 2602 while (tmpent) { 2603 nent++; 2604 sz += tmpent->end - tmpent->start; 2605 if (tmpent->start < cur) 2606 panic("uvm_map_replace1"); 2607 if (tmpent->start >= tmpent->end || tmpent->end > end) { 2608 panic("uvm_map_replace2: " 2609 "tmpent->start=0x%"PRIxVADDR 2610 ", tmpent->end=0x%"PRIxVADDR 2611 ", end=0x%"PRIxVADDR, 2612 tmpent->start, tmpent->end, end); 2613 } 2614 cur = tmpent->end; 2615 if (tmpent->next) { 2616 if (tmpent->next->prev != tmpent) 2617 panic("uvm_map_replace3"); 2618 } else { 2619 if (newents->prev != tmpent) 2620 panic("uvm_map_replace4"); 2621 } 2622 tmpent = tmpent->next; 2623 } 2624 if (nent != nnewents) 2625 panic("uvm_map_replace5"); 2626 if (sz != nsize) 2627 panic("uvm_map_replace6"); 2628 } 2629 #endif 2630 2631 /* 2632 * map entry is a valid blank! replace it. (this does all the 2633 * work of map entry link/unlink...). 2634 */ 2635 2636 if (newents) { 2637 last = newents->prev; 2638 2639 /* critical: flush stale hints out of map */ 2640 SAVE_HINT(map, map->hint, newents); 2641 if (map->first_free == oldent) 2642 map->first_free = last; 2643 2644 last->next = oldent->next; 2645 last->next->prev = last; 2646 2647 /* Fix RB tree */ 2648 uvm_rb_remove(map, oldent); 2649 2650 newents->prev = oldent->prev; 2651 newents->prev->next = newents; 2652 map->nentries = map->nentries + (nnewents - 1); 2653 2654 /* Fixup the RB tree */ 2655 { 2656 int i; 2657 struct vm_map_entry *tmp; 2658 2659 tmp = newents; 2660 for (i = 0; i < nnewents && tmp; i++) { 2661 uvm_rb_insert(map, tmp); 2662 tmp = tmp->next; 2663 } 2664 } 2665 } else { 2666 /* NULL list of new entries: just remove the old one */ 2667 clear_hints(map, oldent); 2668 uvm_map_entry_unlink(map, oldent); 2669 } 2670 map->size -= end - start - nsize; 2671 2672 uvm_map_check(map, "map_replace leave"); 2673 2674 /* 2675 * now we can free the old blank entry and return. 2676 */ 2677 2678 *oldentryp = oldent; 2679 return (true); 2680 } 2681 2682 /* 2683 * uvm_map_extract: extract a mapping from a map and put it somewhere 2684 * (maybe removing the old mapping) 2685 * 2686 * => maps should be unlocked (we will write lock them) 2687 * => returns 0 on success, error code otherwise 2688 * => start must be page aligned 2689 * => len must be page sized 2690 * => flags: 2691 * UVM_EXTRACT_REMOVE: remove mappings from srcmap 2692 * UVM_EXTRACT_CONTIG: abort if unmapped area (advisory only) 2693 * UVM_EXTRACT_QREF: for a temporary extraction do quick obj refs 2694 * UVM_EXTRACT_FIXPROT: set prot to maxprot as we go 2695 * >>>NOTE: if you set REMOVE, you are not allowed to use CONTIG or QREF!<<< 2696 * >>>NOTE: QREF's must be unmapped via the QREF path, thus should only 2697 * be used from within the kernel in a kernel level map <<< 2698 */ 2699 2700 int 2701 uvm_map_extract(struct vm_map *srcmap, vaddr_t start, vsize_t len, 2702 struct vm_map *dstmap, vaddr_t *dstaddrp, int flags) 2703 { 2704 vaddr_t dstaddr, end, newend, oldoffset, fudge, orig_fudge; 2705 struct vm_map_entry *chain, *endchain, *entry, *orig_entry, *newentry, 2706 *deadentry, *oldentry; 2707 struct vm_map_entry *resentry = NULL; /* a dummy reservation entry */ 2708 vsize_t elen; 2709 int nchain, error, copy_ok; 2710 vsize_t nsize; 2711 UVMHIST_FUNC("uvm_map_extract"); UVMHIST_CALLED(maphist); 2712 2713 UVMHIST_LOG(maphist,"(srcmap=0x%x,start=0x%x, len=0x%x", srcmap, start, 2714 len,0); 2715 UVMHIST_LOG(maphist," ...,dstmap=0x%x, flags=0x%x)", dstmap,flags,0,0); 2716 2717 /* 2718 * step 0: sanity check: start must be on a page boundary, length 2719 * must be page sized. can't ask for CONTIG/QREF if you asked for 2720 * REMOVE. 2721 */ 2722 2723 KASSERT((start & PAGE_MASK) == 0 && (len & PAGE_MASK) == 0); 2724 KASSERT((flags & UVM_EXTRACT_REMOVE) == 0 || 2725 (flags & (UVM_EXTRACT_CONTIG|UVM_EXTRACT_QREF)) == 0); 2726 2727 /* 2728 * step 1: reserve space in the target map for the extracted area 2729 */ 2730 2731 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 2732 dstaddr = vm_map_min(dstmap); 2733 if (!uvm_map_reserve(dstmap, len, start, 0, &dstaddr, 0)) 2734 return (ENOMEM); 2735 *dstaddrp = dstaddr; /* pass address back to caller */ 2736 UVMHIST_LOG(maphist, " dstaddr=0x%x", dstaddr,0,0,0); 2737 } else { 2738 dstaddr = *dstaddrp; 2739 } 2740 2741 /* 2742 * step 2: setup for the extraction process loop by init'ing the 2743 * map entry chain, locking src map, and looking up the first useful 2744 * entry in the map. 2745 */ 2746 2747 end = start + len; 2748 newend = dstaddr + len; 2749 chain = endchain = NULL; 2750 nchain = 0; 2751 nsize = 0; 2752 vm_map_lock(srcmap); 2753 2754 if (uvm_map_lookup_entry(srcmap, start, &entry)) { 2755 2756 /* "start" is within an entry */ 2757 if (flags & UVM_EXTRACT_QREF) { 2758 2759 /* 2760 * for quick references we don't clip the entry, so 2761 * the entry may map space "before" the starting 2762 * virtual address... this is the "fudge" factor 2763 * (which can be non-zero only the first time 2764 * through the "while" loop in step 3). 2765 */ 2766 2767 fudge = start - entry->start; 2768 } else { 2769 2770 /* 2771 * normal reference: we clip the map to fit (thus 2772 * fudge is zero) 2773 */ 2774 2775 UVM_MAP_CLIP_START(srcmap, entry, start, NULL); 2776 SAVE_HINT(srcmap, srcmap->hint, entry->prev); 2777 fudge = 0; 2778 } 2779 } else { 2780 2781 /* "start" is not within an entry ... skip to next entry */ 2782 if (flags & UVM_EXTRACT_CONTIG) { 2783 error = EINVAL; 2784 goto bad; /* definite hole here ... */ 2785 } 2786 2787 entry = entry->next; 2788 fudge = 0; 2789 } 2790 2791 /* save values from srcmap for step 6 */ 2792 orig_entry = entry; 2793 orig_fudge = fudge; 2794 2795 /* 2796 * step 3: now start looping through the map entries, extracting 2797 * as we go. 2798 */ 2799 2800 while (entry->start < end && entry != &srcmap->header) { 2801 2802 /* if we are not doing a quick reference, clip it */ 2803 if ((flags & UVM_EXTRACT_QREF) == 0) 2804 UVM_MAP_CLIP_END(srcmap, entry, end, NULL); 2805 2806 /* clear needs_copy (allow chunking) */ 2807 if (UVM_ET_ISNEEDSCOPY(entry)) { 2808 amap_copy(srcmap, entry, 2809 AMAP_COPY_NOWAIT|AMAP_COPY_NOMERGE, start, end); 2810 if (UVM_ET_ISNEEDSCOPY(entry)) { /* failed? */ 2811 error = ENOMEM; 2812 goto bad; 2813 } 2814 2815 /* amap_copy could clip (during chunk)! update fudge */ 2816 if (fudge) { 2817 fudge = start - entry->start; 2818 orig_fudge = fudge; 2819 } 2820 } 2821 2822 /* calculate the offset of this from "start" */ 2823 oldoffset = (entry->start + fudge) - start; 2824 2825 /* allocate a new map entry */ 2826 newentry = uvm_mapent_alloc(dstmap, 0); 2827 if (newentry == NULL) { 2828 error = ENOMEM; 2829 goto bad; 2830 } 2831 2832 /* set up new map entry */ 2833 newentry->next = NULL; 2834 newentry->prev = endchain; 2835 newentry->start = dstaddr + oldoffset; 2836 newentry->end = 2837 newentry->start + (entry->end - (entry->start + fudge)); 2838 if (newentry->end > newend || newentry->end < newentry->start) 2839 newentry->end = newend; 2840 newentry->object.uvm_obj = entry->object.uvm_obj; 2841 if (newentry->object.uvm_obj) { 2842 if (newentry->object.uvm_obj->pgops->pgo_reference) 2843 newentry->object.uvm_obj->pgops-> 2844 pgo_reference(newentry->object.uvm_obj); 2845 newentry->offset = entry->offset + fudge; 2846 } else { 2847 newentry->offset = 0; 2848 } 2849 newentry->etype = entry->etype; 2850 newentry->protection = (flags & UVM_EXTRACT_FIXPROT) ? 2851 entry->max_protection : entry->protection; 2852 newentry->max_protection = entry->max_protection; 2853 newentry->inheritance = entry->inheritance; 2854 newentry->wired_count = 0; 2855 newentry->aref.ar_amap = entry->aref.ar_amap; 2856 if (newentry->aref.ar_amap) { 2857 newentry->aref.ar_pageoff = 2858 entry->aref.ar_pageoff + (fudge >> PAGE_SHIFT); 2859 uvm_map_reference_amap(newentry, AMAP_SHARED | 2860 ((flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0)); 2861 } else { 2862 newentry->aref.ar_pageoff = 0; 2863 } 2864 newentry->advice = entry->advice; 2865 if ((flags & UVM_EXTRACT_QREF) != 0) { 2866 newentry->flags |= UVM_MAP_NOMERGE; 2867 } 2868 2869 /* now link it on the chain */ 2870 nchain++; 2871 nsize += newentry->end - newentry->start; 2872 if (endchain == NULL) { 2873 chain = endchain = newentry; 2874 } else { 2875 endchain->next = newentry; 2876 endchain = newentry; 2877 } 2878 2879 /* end of 'while' loop! */ 2880 if ((flags & UVM_EXTRACT_CONTIG) && entry->end < end && 2881 (entry->next == &srcmap->header || 2882 entry->next->start != entry->end)) { 2883 error = EINVAL; 2884 goto bad; 2885 } 2886 entry = entry->next; 2887 fudge = 0; 2888 } 2889 2890 /* 2891 * step 4: close off chain (in format expected by uvm_map_replace) 2892 */ 2893 2894 if (chain) 2895 chain->prev = endchain; 2896 2897 /* 2898 * step 5: attempt to lock the dest map so we can pmap_copy. 2899 * note usage of copy_ok: 2900 * 1 => dstmap locked, pmap_copy ok, and we "replace" here (step 5) 2901 * 0 => dstmap unlocked, NO pmap_copy, and we will "replace" in step 7 2902 */ 2903 2904 if (srcmap == dstmap || vm_map_lock_try(dstmap) == true) { 2905 copy_ok = 1; 2906 if (!uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2907 nchain, nsize, &resentry)) { 2908 if (srcmap != dstmap) 2909 vm_map_unlock(dstmap); 2910 error = EIO; 2911 goto bad; 2912 } 2913 } else { 2914 copy_ok = 0; 2915 /* replace defered until step 7 */ 2916 } 2917 2918 /* 2919 * step 6: traverse the srcmap a second time to do the following: 2920 * - if we got a lock on the dstmap do pmap_copy 2921 * - if UVM_EXTRACT_REMOVE remove the entries 2922 * we make use of orig_entry and orig_fudge (saved in step 2) 2923 */ 2924 2925 if (copy_ok || (flags & UVM_EXTRACT_REMOVE)) { 2926 2927 /* purge possible stale hints from srcmap */ 2928 if (flags & UVM_EXTRACT_REMOVE) { 2929 SAVE_HINT(srcmap, srcmap->hint, orig_entry->prev); 2930 if (srcmap->first_free != &srcmap->header && 2931 srcmap->first_free->start >= start) 2932 srcmap->first_free = orig_entry->prev; 2933 } 2934 2935 entry = orig_entry; 2936 fudge = orig_fudge; 2937 deadentry = NULL; /* for UVM_EXTRACT_REMOVE */ 2938 2939 while (entry->start < end && entry != &srcmap->header) { 2940 if (copy_ok) { 2941 oldoffset = (entry->start + fudge) - start; 2942 elen = MIN(end, entry->end) - 2943 (entry->start + fudge); 2944 pmap_copy(dstmap->pmap, srcmap->pmap, 2945 dstaddr + oldoffset, elen, 2946 entry->start + fudge); 2947 } 2948 2949 /* we advance "entry" in the following if statement */ 2950 if (flags & UVM_EXTRACT_REMOVE) { 2951 pmap_remove(srcmap->pmap, entry->start, 2952 entry->end); 2953 oldentry = entry; /* save entry */ 2954 entry = entry->next; /* advance */ 2955 uvm_map_entry_unlink(srcmap, oldentry); 2956 /* add to dead list */ 2957 oldentry->next = deadentry; 2958 deadentry = oldentry; 2959 } else { 2960 entry = entry->next; /* advance */ 2961 } 2962 2963 /* end of 'while' loop */ 2964 fudge = 0; 2965 } 2966 pmap_update(srcmap->pmap); 2967 2968 /* 2969 * unlock dstmap. we will dispose of deadentry in 2970 * step 7 if needed 2971 */ 2972 2973 if (copy_ok && srcmap != dstmap) 2974 vm_map_unlock(dstmap); 2975 2976 } else { 2977 deadentry = NULL; 2978 } 2979 2980 /* 2981 * step 7: we are done with the source map, unlock. if copy_ok 2982 * is 0 then we have not replaced the dummy mapping in dstmap yet 2983 * and we need to do so now. 2984 */ 2985 2986 vm_map_unlock(srcmap); 2987 if ((flags & UVM_EXTRACT_REMOVE) && deadentry) 2988 uvm_unmap_detach(deadentry, 0); /* dispose of old entries */ 2989 2990 /* now do the replacement if we didn't do it in step 5 */ 2991 if (copy_ok == 0) { 2992 vm_map_lock(dstmap); 2993 error = uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2994 nchain, nsize, &resentry); 2995 vm_map_unlock(dstmap); 2996 2997 if (error == false) { 2998 error = EIO; 2999 goto bad2; 3000 } 3001 } 3002 3003 if (resentry != NULL) 3004 uvm_mapent_free(resentry); 3005 3006 return (0); 3007 3008 /* 3009 * bad: failure recovery 3010 */ 3011 bad: 3012 vm_map_unlock(srcmap); 3013 bad2: /* src already unlocked */ 3014 if (chain) 3015 uvm_unmap_detach(chain, 3016 (flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0); 3017 3018 if (resentry != NULL) 3019 uvm_mapent_free(resentry); 3020 3021 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 3022 uvm_unmap(dstmap, dstaddr, dstaddr+len); /* ??? */ 3023 } 3024 return (error); 3025 } 3026 3027 /* end of extraction functions */ 3028 3029 /* 3030 * uvm_map_submap: punch down part of a map into a submap 3031 * 3032 * => only the kernel_map is allowed to be submapped 3033 * => the purpose of submapping is to break up the locking granularity 3034 * of a larger map 3035 * => the range specified must have been mapped previously with a uvm_map() 3036 * call [with uobj==NULL] to create a blank map entry in the main map. 3037 * [And it had better still be blank!] 3038 * => maps which contain submaps should never be copied or forked. 3039 * => to remove a submap, use uvm_unmap() on the main map 3040 * and then uvm_map_deallocate() the submap. 3041 * => main map must be unlocked. 3042 * => submap must have been init'd and have a zero reference count. 3043 * [need not be locked as we don't actually reference it] 3044 */ 3045 3046 int 3047 uvm_map_submap(struct vm_map *map, vaddr_t start, vaddr_t end, 3048 struct vm_map *submap) 3049 { 3050 struct vm_map_entry *entry; 3051 struct uvm_mapent_reservation umr; 3052 int error; 3053 3054 uvm_mapent_reserve(map, &umr, 2, 0); 3055 3056 vm_map_lock(map); 3057 VM_MAP_RANGE_CHECK(map, start, end); 3058 3059 if (uvm_map_lookup_entry(map, start, &entry)) { 3060 UVM_MAP_CLIP_START(map, entry, start, &umr); 3061 UVM_MAP_CLIP_END(map, entry, end, &umr); /* to be safe */ 3062 } else { 3063 entry = NULL; 3064 } 3065 3066 if (entry != NULL && 3067 entry->start == start && entry->end == end && 3068 entry->object.uvm_obj == NULL && entry->aref.ar_amap == NULL && 3069 !UVM_ET_ISCOPYONWRITE(entry) && !UVM_ET_ISNEEDSCOPY(entry)) { 3070 entry->etype |= UVM_ET_SUBMAP; 3071 entry->object.sub_map = submap; 3072 entry->offset = 0; 3073 uvm_map_reference(submap); 3074 error = 0; 3075 } else { 3076 error = EINVAL; 3077 } 3078 vm_map_unlock(map); 3079 3080 uvm_mapent_unreserve(map, &umr); 3081 3082 return error; 3083 } 3084 3085 /* 3086 * uvm_map_setup_kernel: init in-kernel map 3087 * 3088 * => map must not be in service yet. 3089 */ 3090 3091 void 3092 uvm_map_setup_kernel(struct vm_map_kernel *map, 3093 vaddr_t vmin, vaddr_t vmax, int flags) 3094 { 3095 3096 uvm_map_setup(&map->vmk_map, vmin, vmax, flags); 3097 callback_head_init(&map->vmk_reclaim_callback, IPL_VM); 3098 LIST_INIT(&map->vmk_kentry_free); 3099 map->vmk_merged_entries = NULL; 3100 } 3101 3102 3103 /* 3104 * uvm_map_protect: change map protection 3105 * 3106 * => set_max means set max_protection. 3107 * => map must be unlocked. 3108 */ 3109 3110 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 3111 ~VM_PROT_WRITE : VM_PROT_ALL) 3112 3113 int 3114 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end, 3115 vm_prot_t new_prot, bool set_max) 3116 { 3117 struct vm_map_entry *current, *entry; 3118 int error = 0; 3119 UVMHIST_FUNC("uvm_map_protect"); UVMHIST_CALLED(maphist); 3120 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_prot=0x%x)", 3121 map, start, end, new_prot); 3122 3123 vm_map_lock(map); 3124 VM_MAP_RANGE_CHECK(map, start, end); 3125 if (uvm_map_lookup_entry(map, start, &entry)) { 3126 UVM_MAP_CLIP_START(map, entry, start, NULL); 3127 } else { 3128 entry = entry->next; 3129 } 3130 3131 /* 3132 * make a first pass to check for protection violations. 3133 */ 3134 3135 current = entry; 3136 while ((current != &map->header) && (current->start < end)) { 3137 if (UVM_ET_ISSUBMAP(current)) { 3138 error = EINVAL; 3139 goto out; 3140 } 3141 if ((new_prot & current->max_protection) != new_prot) { 3142 error = EACCES; 3143 goto out; 3144 } 3145 /* 3146 * Don't allow VM_PROT_EXECUTE to be set on entries that 3147 * point to vnodes that are associated with a NOEXEC file 3148 * system. 3149 */ 3150 if (UVM_ET_ISOBJ(current) && 3151 UVM_OBJ_IS_VNODE(current->object.uvm_obj)) { 3152 struct vnode *vp = 3153 (struct vnode *) current->object.uvm_obj; 3154 3155 if ((new_prot & VM_PROT_EXECUTE) != 0 && 3156 (vp->v_mount->mnt_flag & MNT_NOEXEC) != 0) { 3157 error = EACCES; 3158 goto out; 3159 } 3160 } 3161 3162 current = current->next; 3163 } 3164 3165 /* go back and fix up protections (no need to clip this time). */ 3166 3167 current = entry; 3168 while ((current != &map->header) && (current->start < end)) { 3169 vm_prot_t old_prot; 3170 3171 UVM_MAP_CLIP_END(map, current, end, NULL); 3172 old_prot = current->protection; 3173 if (set_max) 3174 current->protection = 3175 (current->max_protection = new_prot) & old_prot; 3176 else 3177 current->protection = new_prot; 3178 3179 /* 3180 * update physical map if necessary. worry about copy-on-write 3181 * here -- CHECK THIS XXX 3182 */ 3183 3184 if (current->protection != old_prot) { 3185 /* update pmap! */ 3186 pmap_protect(map->pmap, current->start, current->end, 3187 current->protection & MASK(entry)); 3188 3189 /* 3190 * If this entry points at a vnode, and the 3191 * protection includes VM_PROT_EXECUTE, mark 3192 * the vnode as VEXECMAP. 3193 */ 3194 if (UVM_ET_ISOBJ(current)) { 3195 struct uvm_object *uobj = 3196 current->object.uvm_obj; 3197 3198 if (UVM_OBJ_IS_VNODE(uobj) && 3199 (current->protection & VM_PROT_EXECUTE)) { 3200 vn_markexec((struct vnode *) uobj); 3201 } 3202 } 3203 } 3204 3205 /* 3206 * If the map is configured to lock any future mappings, 3207 * wire this entry now if the old protection was VM_PROT_NONE 3208 * and the new protection is not VM_PROT_NONE. 3209 */ 3210 3211 if ((map->flags & VM_MAP_WIREFUTURE) != 0 && 3212 VM_MAPENT_ISWIRED(entry) == 0 && 3213 old_prot == VM_PROT_NONE && 3214 new_prot != VM_PROT_NONE) { 3215 if (uvm_map_pageable(map, entry->start, 3216 entry->end, false, 3217 UVM_LK_ENTER|UVM_LK_EXIT) != 0) { 3218 3219 /* 3220 * If locking the entry fails, remember the 3221 * error if it's the first one. Note we 3222 * still continue setting the protection in 3223 * the map, but will return the error 3224 * condition regardless. 3225 * 3226 * XXX Ignore what the actual error is, 3227 * XXX just call it a resource shortage 3228 * XXX so that it doesn't get confused 3229 * XXX what uvm_map_protect() itself would 3230 * XXX normally return. 3231 */ 3232 3233 error = ENOMEM; 3234 } 3235 } 3236 current = current->next; 3237 } 3238 pmap_update(map->pmap); 3239 3240 out: 3241 vm_map_unlock(map); 3242 3243 UVMHIST_LOG(maphist, "<- done, error=%d",error,0,0,0); 3244 return error; 3245 } 3246 3247 #undef MASK 3248 3249 /* 3250 * uvm_map_inherit: set inheritance code for range of addrs in map. 3251 * 3252 * => map must be unlocked 3253 * => note that the inherit code is used during a "fork". see fork 3254 * code for details. 3255 */ 3256 3257 int 3258 uvm_map_inherit(struct vm_map *map, vaddr_t start, vaddr_t end, 3259 vm_inherit_t new_inheritance) 3260 { 3261 struct vm_map_entry *entry, *temp_entry; 3262 UVMHIST_FUNC("uvm_map_inherit"); UVMHIST_CALLED(maphist); 3263 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_inh=0x%x)", 3264 map, start, end, new_inheritance); 3265 3266 switch (new_inheritance) { 3267 case MAP_INHERIT_NONE: 3268 case MAP_INHERIT_COPY: 3269 case MAP_INHERIT_SHARE: 3270 break; 3271 default: 3272 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3273 return EINVAL; 3274 } 3275 3276 vm_map_lock(map); 3277 VM_MAP_RANGE_CHECK(map, start, end); 3278 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3279 entry = temp_entry; 3280 UVM_MAP_CLIP_START(map, entry, start, NULL); 3281 } else { 3282 entry = temp_entry->next; 3283 } 3284 while ((entry != &map->header) && (entry->start < end)) { 3285 UVM_MAP_CLIP_END(map, entry, end, NULL); 3286 entry->inheritance = new_inheritance; 3287 entry = entry->next; 3288 } 3289 vm_map_unlock(map); 3290 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3291 return 0; 3292 } 3293 3294 /* 3295 * uvm_map_advice: set advice code for range of addrs in map. 3296 * 3297 * => map must be unlocked 3298 */ 3299 3300 int 3301 uvm_map_advice(struct vm_map *map, vaddr_t start, vaddr_t end, int new_advice) 3302 { 3303 struct vm_map_entry *entry, *temp_entry; 3304 UVMHIST_FUNC("uvm_map_advice"); UVMHIST_CALLED(maphist); 3305 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_adv=0x%x)", 3306 map, start, end, new_advice); 3307 3308 vm_map_lock(map); 3309 VM_MAP_RANGE_CHECK(map, start, end); 3310 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3311 entry = temp_entry; 3312 UVM_MAP_CLIP_START(map, entry, start, NULL); 3313 } else { 3314 entry = temp_entry->next; 3315 } 3316 3317 /* 3318 * XXXJRT: disallow holes? 3319 */ 3320 3321 while ((entry != &map->header) && (entry->start < end)) { 3322 UVM_MAP_CLIP_END(map, entry, end, NULL); 3323 3324 switch (new_advice) { 3325 case MADV_NORMAL: 3326 case MADV_RANDOM: 3327 case MADV_SEQUENTIAL: 3328 /* nothing special here */ 3329 break; 3330 3331 default: 3332 vm_map_unlock(map); 3333 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3334 return EINVAL; 3335 } 3336 entry->advice = new_advice; 3337 entry = entry->next; 3338 } 3339 3340 vm_map_unlock(map); 3341 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3342 return 0; 3343 } 3344 3345 /* 3346 * uvm_map_willneed: apply MADV_WILLNEED 3347 */ 3348 3349 int 3350 uvm_map_willneed(struct vm_map *map, vaddr_t start, vaddr_t end) 3351 { 3352 struct vm_map_entry *entry; 3353 UVMHIST_FUNC("uvm_map_willneed"); UVMHIST_CALLED(maphist); 3354 UVMHIST_LOG(maphist,"(map=0x%lx,start=0x%lx,end=0x%lx)", 3355 map, start, end, 0); 3356 3357 vm_map_lock_read(map); 3358 VM_MAP_RANGE_CHECK(map, start, end); 3359 if (!uvm_map_lookup_entry(map, start, &entry)) { 3360 entry = entry->next; 3361 } 3362 while (entry->start < end) { 3363 struct vm_amap * const amap = entry->aref.ar_amap; 3364 struct uvm_object * const uobj = entry->object.uvm_obj; 3365 3366 KASSERT(entry != &map->header); 3367 KASSERT(start < entry->end); 3368 /* 3369 * XXX IMPLEMENT ME. 3370 * Should invent a "weak" mode for uvm_fault() 3371 * which would only do the PGO_LOCKED pgo_get(). 3372 * 3373 * for now, we handle only the easy but common case. 3374 */ 3375 if (UVM_ET_ISOBJ(entry) && amap == NULL && uobj != NULL) { 3376 off_t offset; 3377 off_t size; 3378 3379 offset = entry->offset; 3380 if (start < entry->start) { 3381 offset += entry->start - start; 3382 } 3383 size = entry->offset + (entry->end - entry->start); 3384 if (entry->end < end) { 3385 size -= end - entry->end; 3386 } 3387 uvm_readahead(uobj, offset, size); 3388 } 3389 entry = entry->next; 3390 } 3391 vm_map_unlock_read(map); 3392 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3393 return 0; 3394 } 3395 3396 /* 3397 * uvm_map_pageable: sets the pageability of a range in a map. 3398 * 3399 * => wires map entries. should not be used for transient page locking. 3400 * for that, use uvm_fault_wire()/uvm_fault_unwire() (see uvm_vslock()). 3401 * => regions specified as not pageable require lock-down (wired) memory 3402 * and page tables. 3403 * => map must never be read-locked 3404 * => if islocked is true, map is already write-locked 3405 * => we always unlock the map, since we must downgrade to a read-lock 3406 * to call uvm_fault_wire() 3407 * => XXXCDC: check this and try and clean it up. 3408 */ 3409 3410 int 3411 uvm_map_pageable(struct vm_map *map, vaddr_t start, vaddr_t end, 3412 bool new_pageable, int lockflags) 3413 { 3414 struct vm_map_entry *entry, *start_entry, *failed_entry; 3415 int rv; 3416 #ifdef DIAGNOSTIC 3417 u_int timestamp_save; 3418 #endif 3419 UVMHIST_FUNC("uvm_map_pageable"); UVMHIST_CALLED(maphist); 3420 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_pageable=0x%x)", 3421 map, start, end, new_pageable); 3422 KASSERT(map->flags & VM_MAP_PAGEABLE); 3423 3424 if ((lockflags & UVM_LK_ENTER) == 0) 3425 vm_map_lock(map); 3426 VM_MAP_RANGE_CHECK(map, start, end); 3427 3428 /* 3429 * only one pageability change may take place at one time, since 3430 * uvm_fault_wire assumes it will be called only once for each 3431 * wiring/unwiring. therefore, we have to make sure we're actually 3432 * changing the pageability for the entire region. we do so before 3433 * making any changes. 3434 */ 3435 3436 if (uvm_map_lookup_entry(map, start, &start_entry) == false) { 3437 if ((lockflags & UVM_LK_EXIT) == 0) 3438 vm_map_unlock(map); 3439 3440 UVMHIST_LOG(maphist,"<- done (fault)",0,0,0,0); 3441 return EFAULT; 3442 } 3443 entry = start_entry; 3444 3445 /* 3446 * handle wiring and unwiring separately. 3447 */ 3448 3449 if (new_pageable) { /* unwire */ 3450 UVM_MAP_CLIP_START(map, entry, start, NULL); 3451 3452 /* 3453 * unwiring. first ensure that the range to be unwired is 3454 * really wired down and that there are no holes. 3455 */ 3456 3457 while ((entry != &map->header) && (entry->start < end)) { 3458 if (entry->wired_count == 0 || 3459 (entry->end < end && 3460 (entry->next == &map->header || 3461 entry->next->start > entry->end))) { 3462 if ((lockflags & UVM_LK_EXIT) == 0) 3463 vm_map_unlock(map); 3464 UVMHIST_LOG(maphist, "<- done (INVAL)",0,0,0,0); 3465 return EINVAL; 3466 } 3467 entry = entry->next; 3468 } 3469 3470 /* 3471 * POSIX 1003.1b - a single munlock call unlocks a region, 3472 * regardless of the number of mlock calls made on that 3473 * region. 3474 */ 3475 3476 entry = start_entry; 3477 while ((entry != &map->header) && (entry->start < end)) { 3478 UVM_MAP_CLIP_END(map, entry, end, NULL); 3479 if (VM_MAPENT_ISWIRED(entry)) 3480 uvm_map_entry_unwire(map, entry); 3481 entry = entry->next; 3482 } 3483 if ((lockflags & UVM_LK_EXIT) == 0) 3484 vm_map_unlock(map); 3485 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3486 return 0; 3487 } 3488 3489 /* 3490 * wire case: in two passes [XXXCDC: ugly block of code here] 3491 * 3492 * 1: holding the write lock, we create any anonymous maps that need 3493 * to be created. then we clip each map entry to the region to 3494 * be wired and increment its wiring count. 3495 * 3496 * 2: we downgrade to a read lock, and call uvm_fault_wire to fault 3497 * in the pages for any newly wired area (wired_count == 1). 3498 * 3499 * downgrading to a read lock for uvm_fault_wire avoids a possible 3500 * deadlock with another thread that may have faulted on one of 3501 * the pages to be wired (it would mark the page busy, blocking 3502 * us, then in turn block on the map lock that we hold). because 3503 * of problems in the recursive lock package, we cannot upgrade 3504 * to a write lock in vm_map_lookup. thus, any actions that 3505 * require the write lock must be done beforehand. because we 3506 * keep the read lock on the map, the copy-on-write status of the 3507 * entries we modify here cannot change. 3508 */ 3509 3510 while ((entry != &map->header) && (entry->start < end)) { 3511 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3512 3513 /* 3514 * perform actions of vm_map_lookup that need the 3515 * write lock on the map: create an anonymous map 3516 * for a copy-on-write region, or an anonymous map 3517 * for a zero-fill region. (XXXCDC: submap case 3518 * ok?) 3519 */ 3520 3521 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3522 if (UVM_ET_ISNEEDSCOPY(entry) && 3523 ((entry->max_protection & VM_PROT_WRITE) || 3524 (entry->object.uvm_obj == NULL))) { 3525 amap_copy(map, entry, 0, start, end); 3526 /* XXXCDC: wait OK? */ 3527 } 3528 } 3529 } 3530 UVM_MAP_CLIP_START(map, entry, start, NULL); 3531 UVM_MAP_CLIP_END(map, entry, end, NULL); 3532 entry->wired_count++; 3533 3534 /* 3535 * Check for holes 3536 */ 3537 3538 if (entry->protection == VM_PROT_NONE || 3539 (entry->end < end && 3540 (entry->next == &map->header || 3541 entry->next->start > entry->end))) { 3542 3543 /* 3544 * found one. amap creation actions do not need to 3545 * be undone, but the wired counts need to be restored. 3546 */ 3547 3548 while (entry != &map->header && entry->end > start) { 3549 entry->wired_count--; 3550 entry = entry->prev; 3551 } 3552 if ((lockflags & UVM_LK_EXIT) == 0) 3553 vm_map_unlock(map); 3554 UVMHIST_LOG(maphist,"<- done (INVALID WIRE)",0,0,0,0); 3555 return EINVAL; 3556 } 3557 entry = entry->next; 3558 } 3559 3560 /* 3561 * Pass 2. 3562 */ 3563 3564 #ifdef DIAGNOSTIC 3565 timestamp_save = map->timestamp; 3566 #endif 3567 vm_map_busy(map); 3568 vm_map_unlock(map); 3569 3570 rv = 0; 3571 entry = start_entry; 3572 while (entry != &map->header && entry->start < end) { 3573 if (entry->wired_count == 1) { 3574 rv = uvm_fault_wire(map, entry->start, entry->end, 3575 entry->max_protection, 1); 3576 if (rv) { 3577 3578 /* 3579 * wiring failed. break out of the loop. 3580 * we'll clean up the map below, once we 3581 * have a write lock again. 3582 */ 3583 3584 break; 3585 } 3586 } 3587 entry = entry->next; 3588 } 3589 3590 if (rv) { /* failed? */ 3591 3592 /* 3593 * Get back to an exclusive (write) lock. 3594 */ 3595 3596 vm_map_lock(map); 3597 vm_map_unbusy(map); 3598 3599 #ifdef DIAGNOSTIC 3600 if (timestamp_save + 1 != map->timestamp) 3601 panic("uvm_map_pageable: stale map"); 3602 #endif 3603 3604 /* 3605 * first drop the wiring count on all the entries 3606 * which haven't actually been wired yet. 3607 */ 3608 3609 failed_entry = entry; 3610 while (entry != &map->header && entry->start < end) { 3611 entry->wired_count--; 3612 entry = entry->next; 3613 } 3614 3615 /* 3616 * now, unwire all the entries that were successfully 3617 * wired above. 3618 */ 3619 3620 entry = start_entry; 3621 while (entry != failed_entry) { 3622 entry->wired_count--; 3623 if (VM_MAPENT_ISWIRED(entry) == 0) 3624 uvm_map_entry_unwire(map, entry); 3625 entry = entry->next; 3626 } 3627 if ((lockflags & UVM_LK_EXIT) == 0) 3628 vm_map_unlock(map); 3629 UVMHIST_LOG(maphist, "<- done (RV=%d)", rv,0,0,0); 3630 return (rv); 3631 } 3632 3633 if ((lockflags & UVM_LK_EXIT) == 0) { 3634 vm_map_unbusy(map); 3635 } else { 3636 3637 /* 3638 * Get back to an exclusive (write) lock. 3639 */ 3640 3641 vm_map_lock(map); 3642 vm_map_unbusy(map); 3643 } 3644 3645 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3646 return 0; 3647 } 3648 3649 /* 3650 * uvm_map_pageable_all: special case of uvm_map_pageable - affects 3651 * all mapped regions. 3652 * 3653 * => map must not be locked. 3654 * => if no flags are specified, all regions are unwired. 3655 * => XXXJRT: has some of the same problems as uvm_map_pageable() above. 3656 */ 3657 3658 int 3659 uvm_map_pageable_all(struct vm_map *map, int flags, vsize_t limit) 3660 { 3661 struct vm_map_entry *entry, *failed_entry; 3662 vsize_t size; 3663 int rv; 3664 #ifdef DIAGNOSTIC 3665 u_int timestamp_save; 3666 #endif 3667 UVMHIST_FUNC("uvm_map_pageable_all"); UVMHIST_CALLED(maphist); 3668 UVMHIST_LOG(maphist,"(map=0x%x,flags=0x%x)", map, flags, 0, 0); 3669 3670 KASSERT(map->flags & VM_MAP_PAGEABLE); 3671 3672 vm_map_lock(map); 3673 3674 /* 3675 * handle wiring and unwiring separately. 3676 */ 3677 3678 if (flags == 0) { /* unwire */ 3679 3680 /* 3681 * POSIX 1003.1b -- munlockall unlocks all regions, 3682 * regardless of how many times mlockall has been called. 3683 */ 3684 3685 for (entry = map->header.next; entry != &map->header; 3686 entry = entry->next) { 3687 if (VM_MAPENT_ISWIRED(entry)) 3688 uvm_map_entry_unwire(map, entry); 3689 } 3690 map->flags &= ~VM_MAP_WIREFUTURE; 3691 vm_map_unlock(map); 3692 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3693 return 0; 3694 } 3695 3696 if (flags & MCL_FUTURE) { 3697 3698 /* 3699 * must wire all future mappings; remember this. 3700 */ 3701 3702 map->flags |= VM_MAP_WIREFUTURE; 3703 } 3704 3705 if ((flags & MCL_CURRENT) == 0) { 3706 3707 /* 3708 * no more work to do! 3709 */ 3710 3711 UVMHIST_LOG(maphist,"<- done (OK no wire)",0,0,0,0); 3712 vm_map_unlock(map); 3713 return 0; 3714 } 3715 3716 /* 3717 * wire case: in three passes [XXXCDC: ugly block of code here] 3718 * 3719 * 1: holding the write lock, count all pages mapped by non-wired 3720 * entries. if this would cause us to go over our limit, we fail. 3721 * 3722 * 2: still holding the write lock, we create any anonymous maps that 3723 * need to be created. then we increment its wiring count. 3724 * 3725 * 3: we downgrade to a read lock, and call uvm_fault_wire to fault 3726 * in the pages for any newly wired area (wired_count == 1). 3727 * 3728 * downgrading to a read lock for uvm_fault_wire avoids a possible 3729 * deadlock with another thread that may have faulted on one of 3730 * the pages to be wired (it would mark the page busy, blocking 3731 * us, then in turn block on the map lock that we hold). because 3732 * of problems in the recursive lock package, we cannot upgrade 3733 * to a write lock in vm_map_lookup. thus, any actions that 3734 * require the write lock must be done beforehand. because we 3735 * keep the read lock on the map, the copy-on-write status of the 3736 * entries we modify here cannot change. 3737 */ 3738 3739 for (size = 0, entry = map->header.next; entry != &map->header; 3740 entry = entry->next) { 3741 if (entry->protection != VM_PROT_NONE && 3742 VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3743 size += entry->end - entry->start; 3744 } 3745 } 3746 3747 if (atop(size) + uvmexp.wired > uvmexp.wiredmax) { 3748 vm_map_unlock(map); 3749 return ENOMEM; 3750 } 3751 3752 if (limit != 0 && 3753 (size + ptoa(pmap_wired_count(vm_map_pmap(map))) > limit)) { 3754 vm_map_unlock(map); 3755 return ENOMEM; 3756 } 3757 3758 /* 3759 * Pass 2. 3760 */ 3761 3762 for (entry = map->header.next; entry != &map->header; 3763 entry = entry->next) { 3764 if (entry->protection == VM_PROT_NONE) 3765 continue; 3766 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3767 3768 /* 3769 * perform actions of vm_map_lookup that need the 3770 * write lock on the map: create an anonymous map 3771 * for a copy-on-write region, or an anonymous map 3772 * for a zero-fill region. (XXXCDC: submap case 3773 * ok?) 3774 */ 3775 3776 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3777 if (UVM_ET_ISNEEDSCOPY(entry) && 3778 ((entry->max_protection & VM_PROT_WRITE) || 3779 (entry->object.uvm_obj == NULL))) { 3780 amap_copy(map, entry, 0, entry->start, 3781 entry->end); 3782 /* XXXCDC: wait OK? */ 3783 } 3784 } 3785 } 3786 entry->wired_count++; 3787 } 3788 3789 /* 3790 * Pass 3. 3791 */ 3792 3793 #ifdef DIAGNOSTIC 3794 timestamp_save = map->timestamp; 3795 #endif 3796 vm_map_busy(map); 3797 vm_map_unlock(map); 3798 3799 rv = 0; 3800 for (entry = map->header.next; entry != &map->header; 3801 entry = entry->next) { 3802 if (entry->wired_count == 1) { 3803 rv = uvm_fault_wire(map, entry->start, entry->end, 3804 entry->max_protection, 1); 3805 if (rv) { 3806 3807 /* 3808 * wiring failed. break out of the loop. 3809 * we'll clean up the map below, once we 3810 * have a write lock again. 3811 */ 3812 3813 break; 3814 } 3815 } 3816 } 3817 3818 if (rv) { 3819 3820 /* 3821 * Get back an exclusive (write) lock. 3822 */ 3823 3824 vm_map_lock(map); 3825 vm_map_unbusy(map); 3826 3827 #ifdef DIAGNOSTIC 3828 if (timestamp_save + 1 != map->timestamp) 3829 panic("uvm_map_pageable_all: stale map"); 3830 #endif 3831 3832 /* 3833 * first drop the wiring count on all the entries 3834 * which haven't actually been wired yet. 3835 * 3836 * Skip VM_PROT_NONE entries like we did above. 3837 */ 3838 3839 failed_entry = entry; 3840 for (/* nothing */; entry != &map->header; 3841 entry = entry->next) { 3842 if (entry->protection == VM_PROT_NONE) 3843 continue; 3844 entry->wired_count--; 3845 } 3846 3847 /* 3848 * now, unwire all the entries that were successfully 3849 * wired above. 3850 * 3851 * Skip VM_PROT_NONE entries like we did above. 3852 */ 3853 3854 for (entry = map->header.next; entry != failed_entry; 3855 entry = entry->next) { 3856 if (entry->protection == VM_PROT_NONE) 3857 continue; 3858 entry->wired_count--; 3859 if (VM_MAPENT_ISWIRED(entry)) 3860 uvm_map_entry_unwire(map, entry); 3861 } 3862 vm_map_unlock(map); 3863 UVMHIST_LOG(maphist,"<- done (RV=%d)", rv,0,0,0); 3864 return (rv); 3865 } 3866 3867 vm_map_unbusy(map); 3868 3869 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3870 return 0; 3871 } 3872 3873 /* 3874 * uvm_map_clean: clean out a map range 3875 * 3876 * => valid flags: 3877 * if (flags & PGO_CLEANIT): dirty pages are cleaned first 3878 * if (flags & PGO_SYNCIO): dirty pages are written synchronously 3879 * if (flags & PGO_DEACTIVATE): any cached pages are deactivated after clean 3880 * if (flags & PGO_FREE): any cached pages are freed after clean 3881 * => returns an error if any part of the specified range isn't mapped 3882 * => never a need to flush amap layer since the anonymous memory has 3883 * no permanent home, but may deactivate pages there 3884 * => called from sys_msync() and sys_madvise() 3885 * => caller must not write-lock map (read OK). 3886 * => we may sleep while cleaning if SYNCIO [with map read-locked] 3887 */ 3888 3889 int 3890 uvm_map_clean(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 3891 { 3892 struct vm_map_entry *current, *entry; 3893 struct uvm_object *uobj; 3894 struct vm_amap *amap; 3895 struct vm_anon *anon; 3896 struct vm_page *pg; 3897 vaddr_t offset; 3898 vsize_t size; 3899 voff_t uoff; 3900 int error, refs; 3901 UVMHIST_FUNC("uvm_map_clean"); UVMHIST_CALLED(maphist); 3902 3903 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,flags=0x%x)", 3904 map, start, end, flags); 3905 KASSERT((flags & (PGO_FREE|PGO_DEACTIVATE)) != 3906 (PGO_FREE|PGO_DEACTIVATE)); 3907 3908 vm_map_lock_read(map); 3909 VM_MAP_RANGE_CHECK(map, start, end); 3910 if (uvm_map_lookup_entry(map, start, &entry) == false) { 3911 vm_map_unlock_read(map); 3912 return EFAULT; 3913 } 3914 3915 /* 3916 * Make a first pass to check for holes and wiring problems. 3917 */ 3918 3919 for (current = entry; current->start < end; current = current->next) { 3920 if (UVM_ET_ISSUBMAP(current)) { 3921 vm_map_unlock_read(map); 3922 return EINVAL; 3923 } 3924 if ((flags & PGO_FREE) != 0 && VM_MAPENT_ISWIRED(entry)) { 3925 vm_map_unlock_read(map); 3926 return EBUSY; 3927 } 3928 if (end <= current->end) { 3929 break; 3930 } 3931 if (current->end != current->next->start) { 3932 vm_map_unlock_read(map); 3933 return EFAULT; 3934 } 3935 } 3936 3937 error = 0; 3938 for (current = entry; start < end; current = current->next) { 3939 amap = current->aref.ar_amap; /* upper layer */ 3940 uobj = current->object.uvm_obj; /* lower layer */ 3941 KASSERT(start >= current->start); 3942 3943 /* 3944 * No amap cleaning necessary if: 3945 * 3946 * (1) There's no amap. 3947 * 3948 * (2) We're not deactivating or freeing pages. 3949 */ 3950 3951 if (amap == NULL || (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) 3952 goto flush_object; 3953 3954 amap_lock(amap); 3955 offset = start - current->start; 3956 size = MIN(end, current->end) - start; 3957 for ( ; size != 0; size -= PAGE_SIZE, offset += PAGE_SIZE) { 3958 anon = amap_lookup(¤t->aref, offset); 3959 if (anon == NULL) 3960 continue; 3961 3962 mutex_enter(&anon->an_lock); 3963 pg = anon->an_page; 3964 if (pg == NULL) { 3965 mutex_exit(&anon->an_lock); 3966 continue; 3967 } 3968 3969 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 3970 3971 /* 3972 * In these first 3 cases, we just deactivate the page. 3973 */ 3974 3975 case PGO_CLEANIT|PGO_FREE: 3976 case PGO_CLEANIT|PGO_DEACTIVATE: 3977 case PGO_DEACTIVATE: 3978 deactivate_it: 3979 /* 3980 * skip the page if it's loaned or wired, 3981 * since it shouldn't be on a paging queue 3982 * at all in these cases. 3983 */ 3984 3985 mutex_enter(&uvm_pageqlock); 3986 if (pg->loan_count != 0 || 3987 pg->wire_count != 0) { 3988 mutex_exit(&uvm_pageqlock); 3989 mutex_exit(&anon->an_lock); 3990 continue; 3991 } 3992 KASSERT(pg->uanon == anon); 3993 uvm_pagedeactivate(pg); 3994 mutex_exit(&uvm_pageqlock); 3995 mutex_exit(&anon->an_lock); 3996 continue; 3997 3998 case PGO_FREE: 3999 4000 /* 4001 * If there are multiple references to 4002 * the amap, just deactivate the page. 4003 */ 4004 4005 if (amap_refs(amap) > 1) 4006 goto deactivate_it; 4007 4008 /* skip the page if it's wired */ 4009 if (pg->wire_count != 0) { 4010 mutex_exit(&anon->an_lock); 4011 continue; 4012 } 4013 amap_unadd(¤t->aref, offset); 4014 refs = --anon->an_ref; 4015 mutex_exit(&anon->an_lock); 4016 if (refs == 0) 4017 uvm_anfree(anon); 4018 continue; 4019 } 4020 } 4021 amap_unlock(amap); 4022 4023 flush_object: 4024 /* 4025 * flush pages if we've got a valid backing object. 4026 * note that we must always clean object pages before 4027 * freeing them since otherwise we could reveal stale 4028 * data from files. 4029 */ 4030 4031 uoff = current->offset + (start - current->start); 4032 size = MIN(end, current->end) - start; 4033 if (uobj != NULL) { 4034 mutex_enter(&uobj->vmobjlock); 4035 if (uobj->pgops->pgo_put != NULL) 4036 error = (uobj->pgops->pgo_put)(uobj, uoff, 4037 uoff + size, flags | PGO_CLEANIT); 4038 else 4039 error = 0; 4040 } 4041 start += size; 4042 } 4043 vm_map_unlock_read(map); 4044 return (error); 4045 } 4046 4047 4048 /* 4049 * uvm_map_checkprot: check protection in map 4050 * 4051 * => must allow specified protection in a fully allocated region. 4052 * => map must be read or write locked by caller. 4053 */ 4054 4055 bool 4056 uvm_map_checkprot(struct vm_map *map, vaddr_t start, vaddr_t end, 4057 vm_prot_t protection) 4058 { 4059 struct vm_map_entry *entry; 4060 struct vm_map_entry *tmp_entry; 4061 4062 if (!uvm_map_lookup_entry(map, start, &tmp_entry)) { 4063 return (false); 4064 } 4065 entry = tmp_entry; 4066 while (start < end) { 4067 if (entry == &map->header) { 4068 return (false); 4069 } 4070 4071 /* 4072 * no holes allowed 4073 */ 4074 4075 if (start < entry->start) { 4076 return (false); 4077 } 4078 4079 /* 4080 * check protection associated with entry 4081 */ 4082 4083 if ((entry->protection & protection) != protection) { 4084 return (false); 4085 } 4086 start = entry->end; 4087 entry = entry->next; 4088 } 4089 return (true); 4090 } 4091 4092 /* 4093 * uvmspace_alloc: allocate a vmspace structure. 4094 * 4095 * - structure includes vm_map and pmap 4096 * - XXX: no locking on this structure 4097 * - refcnt set to 1, rest must be init'd by caller 4098 */ 4099 struct vmspace * 4100 uvmspace_alloc(vaddr_t vmin, vaddr_t vmax) 4101 { 4102 struct vmspace *vm; 4103 UVMHIST_FUNC("uvmspace_alloc"); UVMHIST_CALLED(maphist); 4104 4105 vm = pool_cache_get(&uvm_vmspace_cache, PR_WAITOK); 4106 uvmspace_init(vm, NULL, vmin, vmax); 4107 UVMHIST_LOG(maphist,"<- done (vm=0x%x)", vm,0,0,0); 4108 return (vm); 4109 } 4110 4111 /* 4112 * uvmspace_init: initialize a vmspace structure. 4113 * 4114 * - XXX: no locking on this structure 4115 * - refcnt set to 1, rest must be init'd by caller 4116 */ 4117 void 4118 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t vmin, vaddr_t vmax) 4119 { 4120 UVMHIST_FUNC("uvmspace_init"); UVMHIST_CALLED(maphist); 4121 4122 memset(vm, 0, sizeof(*vm)); 4123 uvm_map_setup(&vm->vm_map, vmin, vmax, VM_MAP_PAGEABLE 4124 #ifdef __USING_TOPDOWN_VM 4125 | VM_MAP_TOPDOWN 4126 #endif 4127 ); 4128 if (pmap) 4129 pmap_reference(pmap); 4130 else 4131 pmap = pmap_create(); 4132 vm->vm_map.pmap = pmap; 4133 vm->vm_refcnt = 1; 4134 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4135 } 4136 4137 /* 4138 * uvmspace_share: share a vmspace between two processes 4139 * 4140 * - used for vfork, threads(?) 4141 */ 4142 4143 void 4144 uvmspace_share(struct proc *p1, struct proc *p2) 4145 { 4146 4147 uvmspace_addref(p1->p_vmspace); 4148 p2->p_vmspace = p1->p_vmspace; 4149 } 4150 4151 #if 0 4152 4153 /* 4154 * uvmspace_unshare: ensure that process "p" has its own, unshared, vmspace 4155 * 4156 * - XXX: no locking on vmspace 4157 */ 4158 4159 void 4160 uvmspace_unshare(struct lwp *l) 4161 { 4162 struct proc *p = l->l_proc; 4163 struct vmspace *nvm, *ovm = p->p_vmspace; 4164 4165 if (ovm->vm_refcnt == 1) 4166 /* nothing to do: vmspace isn't shared in the first place */ 4167 return; 4168 4169 /* make a new vmspace, still holding old one */ 4170 nvm = uvmspace_fork(ovm); 4171 4172 kpreempt_disable(); 4173 pmap_deactivate(l); /* unbind old vmspace */ 4174 p->p_vmspace = nvm; 4175 pmap_activate(l); /* switch to new vmspace */ 4176 kpreempt_enable(); 4177 4178 uvmspace_free(ovm); /* drop reference to old vmspace */ 4179 } 4180 4181 #endif 4182 4183 /* 4184 * uvmspace_exec: the process wants to exec a new program 4185 */ 4186 4187 void 4188 uvmspace_exec(struct lwp *l, vaddr_t start, vaddr_t end) 4189 { 4190 struct proc *p = l->l_proc; 4191 struct vmspace *nvm, *ovm = p->p_vmspace; 4192 struct vm_map *map = &ovm->vm_map; 4193 4194 #ifdef __sparc__ 4195 /* XXX cgd 960926: the sparc #ifdef should be a MD hook */ 4196 kill_user_windows(l); /* before stack addresses go away */ 4197 #endif 4198 4199 /* 4200 * see if more than one process is using this vmspace... 4201 */ 4202 4203 if (ovm->vm_refcnt == 1) { 4204 4205 /* 4206 * if p is the only process using its vmspace then we can safely 4207 * recycle that vmspace for the program that is being exec'd. 4208 */ 4209 4210 #ifdef SYSVSHM 4211 /* 4212 * SYSV SHM semantics require us to kill all segments on an exec 4213 */ 4214 4215 if (ovm->vm_shm) 4216 shmexit(ovm); 4217 #endif 4218 4219 /* 4220 * POSIX 1003.1b -- "lock future mappings" is revoked 4221 * when a process execs another program image. 4222 */ 4223 4224 map->flags &= ~VM_MAP_WIREFUTURE; 4225 4226 /* 4227 * now unmap the old program 4228 */ 4229 4230 pmap_remove_all(map->pmap); 4231 uvm_unmap(map, vm_map_min(map), vm_map_max(map)); 4232 KASSERT(map->header.prev == &map->header); 4233 KASSERT(map->nentries == 0); 4234 4235 /* 4236 * resize the map 4237 */ 4238 4239 vm_map_setmin(map, start); 4240 vm_map_setmax(map, end); 4241 } else { 4242 4243 /* 4244 * p's vmspace is being shared, so we can't reuse it for p since 4245 * it is still being used for others. allocate a new vmspace 4246 * for p 4247 */ 4248 4249 nvm = uvmspace_alloc(start, end); 4250 4251 /* 4252 * install new vmspace and drop our ref to the old one. 4253 */ 4254 4255 kpreempt_disable(); 4256 pmap_deactivate(l); 4257 p->p_vmspace = nvm; 4258 pmap_activate(l); 4259 kpreempt_enable(); 4260 4261 uvmspace_free(ovm); 4262 } 4263 } 4264 4265 /* 4266 * uvmspace_addref: add a referece to a vmspace. 4267 */ 4268 4269 void 4270 uvmspace_addref(struct vmspace *vm) 4271 { 4272 struct vm_map *map = &vm->vm_map; 4273 4274 KASSERT((map->flags & VM_MAP_DYING) == 0); 4275 4276 mutex_enter(&map->misc_lock); 4277 KASSERT(vm->vm_refcnt > 0); 4278 vm->vm_refcnt++; 4279 mutex_exit(&map->misc_lock); 4280 } 4281 4282 /* 4283 * uvmspace_free: free a vmspace data structure 4284 */ 4285 4286 void 4287 uvmspace_free(struct vmspace *vm) 4288 { 4289 struct vm_map_entry *dead_entries; 4290 struct vm_map *map = &vm->vm_map; 4291 int n; 4292 4293 UVMHIST_FUNC("uvmspace_free"); UVMHIST_CALLED(maphist); 4294 4295 UVMHIST_LOG(maphist,"(vm=0x%x) ref=%d", vm, vm->vm_refcnt,0,0); 4296 mutex_enter(&map->misc_lock); 4297 n = --vm->vm_refcnt; 4298 mutex_exit(&map->misc_lock); 4299 if (n > 0) 4300 return; 4301 4302 /* 4303 * at this point, there should be no other references to the map. 4304 * delete all of the mappings, then destroy the pmap. 4305 */ 4306 4307 map->flags |= VM_MAP_DYING; 4308 pmap_remove_all(map->pmap); 4309 #ifdef SYSVSHM 4310 /* Get rid of any SYSV shared memory segments. */ 4311 if (vm->vm_shm != NULL) 4312 shmexit(vm); 4313 #endif 4314 if (map->nentries) { 4315 uvm_unmap_remove(map, vm_map_min(map), vm_map_max(map), 4316 &dead_entries, NULL, 0); 4317 if (dead_entries != NULL) 4318 uvm_unmap_detach(dead_entries, 0); 4319 } 4320 KASSERT(map->nentries == 0); 4321 KASSERT(map->size == 0); 4322 mutex_destroy(&map->misc_lock); 4323 mutex_destroy(&map->mutex); 4324 rw_destroy(&map->lock); 4325 cv_destroy(&map->cv); 4326 pmap_destroy(map->pmap); 4327 pool_cache_put(&uvm_vmspace_cache, vm); 4328 } 4329 4330 /* 4331 * F O R K - m a i n e n t r y p o i n t 4332 */ 4333 /* 4334 * uvmspace_fork: fork a process' main map 4335 * 4336 * => create a new vmspace for child process from parent. 4337 * => parent's map must not be locked. 4338 */ 4339 4340 struct vmspace * 4341 uvmspace_fork(struct vmspace *vm1) 4342 { 4343 struct vmspace *vm2; 4344 struct vm_map *old_map = &vm1->vm_map; 4345 struct vm_map *new_map; 4346 struct vm_map_entry *old_entry; 4347 struct vm_map_entry *new_entry; 4348 UVMHIST_FUNC("uvmspace_fork"); UVMHIST_CALLED(maphist); 4349 4350 vm_map_lock(old_map); 4351 4352 vm2 = uvmspace_alloc(vm_map_min(old_map), vm_map_max(old_map)); 4353 memcpy(&vm2->vm_startcopy, &vm1->vm_startcopy, 4354 (char *) (vm1 + 1) - (char *) &vm1->vm_startcopy); 4355 new_map = &vm2->vm_map; /* XXX */ 4356 4357 old_entry = old_map->header.next; 4358 new_map->size = old_map->size; 4359 4360 /* 4361 * go entry-by-entry 4362 */ 4363 4364 while (old_entry != &old_map->header) { 4365 4366 /* 4367 * first, some sanity checks on the old entry 4368 */ 4369 4370 KASSERT(!UVM_ET_ISSUBMAP(old_entry)); 4371 KASSERT(UVM_ET_ISCOPYONWRITE(old_entry) || 4372 !UVM_ET_ISNEEDSCOPY(old_entry)); 4373 4374 switch (old_entry->inheritance) { 4375 case MAP_INHERIT_NONE: 4376 4377 /* 4378 * drop the mapping, modify size 4379 */ 4380 new_map->size -= old_entry->end - old_entry->start; 4381 break; 4382 4383 case MAP_INHERIT_SHARE: 4384 4385 /* 4386 * share the mapping: this means we want the old and 4387 * new entries to share amaps and backing objects. 4388 */ 4389 /* 4390 * if the old_entry needs a new amap (due to prev fork) 4391 * then we need to allocate it now so that we have 4392 * something we own to share with the new_entry. [in 4393 * other words, we need to clear needs_copy] 4394 */ 4395 4396 if (UVM_ET_ISNEEDSCOPY(old_entry)) { 4397 /* get our own amap, clears needs_copy */ 4398 amap_copy(old_map, old_entry, AMAP_COPY_NOCHUNK, 4399 0, 0); 4400 /* XXXCDC: WAITOK??? */ 4401 } 4402 4403 new_entry = uvm_mapent_alloc(new_map, 0); 4404 /* old_entry -> new_entry */ 4405 uvm_mapent_copy(old_entry, new_entry); 4406 4407 /* new pmap has nothing wired in it */ 4408 new_entry->wired_count = 0; 4409 4410 /* 4411 * gain reference to object backing the map (can't 4412 * be a submap, already checked this case). 4413 */ 4414 4415 if (new_entry->aref.ar_amap) 4416 uvm_map_reference_amap(new_entry, AMAP_SHARED); 4417 4418 if (new_entry->object.uvm_obj && 4419 new_entry->object.uvm_obj->pgops->pgo_reference) 4420 new_entry->object.uvm_obj-> 4421 pgops->pgo_reference( 4422 new_entry->object.uvm_obj); 4423 4424 /* insert entry at end of new_map's entry list */ 4425 uvm_map_entry_link(new_map, new_map->header.prev, 4426 new_entry); 4427 4428 break; 4429 4430 case MAP_INHERIT_COPY: 4431 4432 /* 4433 * copy-on-write the mapping (using mmap's 4434 * MAP_PRIVATE semantics) 4435 * 4436 * allocate new_entry, adjust reference counts. 4437 * (note that new references are read-only). 4438 */ 4439 4440 new_entry = uvm_mapent_alloc(new_map, 0); 4441 /* old_entry -> new_entry */ 4442 uvm_mapent_copy(old_entry, new_entry); 4443 4444 if (new_entry->aref.ar_amap) 4445 uvm_map_reference_amap(new_entry, 0); 4446 4447 if (new_entry->object.uvm_obj && 4448 new_entry->object.uvm_obj->pgops->pgo_reference) 4449 new_entry->object.uvm_obj->pgops->pgo_reference 4450 (new_entry->object.uvm_obj); 4451 4452 /* new pmap has nothing wired in it */ 4453 new_entry->wired_count = 0; 4454 4455 new_entry->etype |= 4456 (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY); 4457 uvm_map_entry_link(new_map, new_map->header.prev, 4458 new_entry); 4459 4460 /* 4461 * the new entry will need an amap. it will either 4462 * need to be copied from the old entry or created 4463 * from scratch (if the old entry does not have an 4464 * amap). can we defer this process until later 4465 * (by setting "needs_copy") or do we need to copy 4466 * the amap now? 4467 * 4468 * we must copy the amap now if any of the following 4469 * conditions hold: 4470 * 1. the old entry has an amap and that amap is 4471 * being shared. this means that the old (parent) 4472 * process is sharing the amap with another 4473 * process. if we do not clear needs_copy here 4474 * we will end up in a situation where both the 4475 * parent and child process are refering to the 4476 * same amap with "needs_copy" set. if the 4477 * parent write-faults, the fault routine will 4478 * clear "needs_copy" in the parent by allocating 4479 * a new amap. this is wrong because the 4480 * parent is supposed to be sharing the old amap 4481 * and the new amap will break that. 4482 * 4483 * 2. if the old entry has an amap and a non-zero 4484 * wire count then we are going to have to call 4485 * amap_cow_now to avoid page faults in the 4486 * parent process. since amap_cow_now requires 4487 * "needs_copy" to be clear we might as well 4488 * clear it here as well. 4489 * 4490 */ 4491 4492 if (old_entry->aref.ar_amap != NULL) { 4493 if ((amap_flags(old_entry->aref.ar_amap) & 4494 AMAP_SHARED) != 0 || 4495 VM_MAPENT_ISWIRED(old_entry)) { 4496 4497 amap_copy(new_map, new_entry, 4498 AMAP_COPY_NOCHUNK, 0, 0); 4499 /* XXXCDC: M_WAITOK ... ok? */ 4500 } 4501 } 4502 4503 /* 4504 * if the parent's entry is wired down, then the 4505 * parent process does not want page faults on 4506 * access to that memory. this means that we 4507 * cannot do copy-on-write because we can't write 4508 * protect the old entry. in this case we 4509 * resolve all copy-on-write faults now, using 4510 * amap_cow_now. note that we have already 4511 * allocated any needed amap (above). 4512 */ 4513 4514 if (VM_MAPENT_ISWIRED(old_entry)) { 4515 4516 /* 4517 * resolve all copy-on-write faults now 4518 * (note that there is nothing to do if 4519 * the old mapping does not have an amap). 4520 */ 4521 if (old_entry->aref.ar_amap) 4522 amap_cow_now(new_map, new_entry); 4523 4524 } else { 4525 4526 /* 4527 * setup mappings to trigger copy-on-write faults 4528 * we must write-protect the parent if it has 4529 * an amap and it is not already "needs_copy"... 4530 * if it is already "needs_copy" then the parent 4531 * has already been write-protected by a previous 4532 * fork operation. 4533 */ 4534 4535 if (old_entry->aref.ar_amap && 4536 !UVM_ET_ISNEEDSCOPY(old_entry)) { 4537 if (old_entry->max_protection & VM_PROT_WRITE) { 4538 pmap_protect(old_map->pmap, 4539 old_entry->start, 4540 old_entry->end, 4541 old_entry->protection & 4542 ~VM_PROT_WRITE); 4543 } 4544 old_entry->etype |= UVM_ET_NEEDSCOPY; 4545 } 4546 } 4547 break; 4548 } /* end of switch statement */ 4549 old_entry = old_entry->next; 4550 } 4551 4552 pmap_update(old_map->pmap); 4553 vm_map_unlock(old_map); 4554 4555 #ifdef SYSVSHM 4556 if (vm1->vm_shm) 4557 shmfork(vm1, vm2); 4558 #endif 4559 4560 #ifdef PMAP_FORK 4561 pmap_fork(vm1->vm_map.pmap, vm2->vm_map.pmap); 4562 #endif 4563 4564 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4565 return (vm2); 4566 } 4567 4568 4569 /* 4570 * in-kernel map entry allocation. 4571 */ 4572 4573 struct uvm_kmapent_hdr { 4574 LIST_ENTRY(uvm_kmapent_hdr) ukh_listq; 4575 int ukh_nused; 4576 struct vm_map_entry *ukh_freelist; 4577 struct vm_map *ukh_map; 4578 struct vm_map_entry ukh_entries[0]; 4579 }; 4580 4581 #define UVM_KMAPENT_CHUNK \ 4582 ((PAGE_SIZE - sizeof(struct uvm_kmapent_hdr)) \ 4583 / sizeof(struct vm_map_entry)) 4584 4585 #define UVM_KHDR_FIND(entry) \ 4586 ((struct uvm_kmapent_hdr *)(((vaddr_t)entry) & ~PAGE_MASK)) 4587 4588 4589 #ifdef DIAGNOSTIC 4590 static struct vm_map * 4591 uvm_kmapent_map(struct vm_map_entry *entry) 4592 { 4593 const struct uvm_kmapent_hdr *ukh; 4594 4595 ukh = UVM_KHDR_FIND(entry); 4596 return ukh->ukh_map; 4597 } 4598 #endif 4599 4600 static inline struct vm_map_entry * 4601 uvm_kmapent_get(struct uvm_kmapent_hdr *ukh) 4602 { 4603 struct vm_map_entry *entry; 4604 4605 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4606 KASSERT(ukh->ukh_nused >= 0); 4607 4608 entry = ukh->ukh_freelist; 4609 if (entry) { 4610 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4611 == UVM_MAP_KERNEL); 4612 ukh->ukh_freelist = entry->next; 4613 ukh->ukh_nused++; 4614 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4615 } else { 4616 KASSERT(ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4617 } 4618 4619 return entry; 4620 } 4621 4622 static inline void 4623 uvm_kmapent_put(struct uvm_kmapent_hdr *ukh, struct vm_map_entry *entry) 4624 { 4625 4626 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4627 == UVM_MAP_KERNEL); 4628 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4629 KASSERT(ukh->ukh_nused > 0); 4630 KASSERT(ukh->ukh_freelist != NULL || 4631 ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4632 KASSERT(ukh->ukh_freelist == NULL || 4633 ukh->ukh_nused < UVM_KMAPENT_CHUNK); 4634 4635 ukh->ukh_nused--; 4636 entry->next = ukh->ukh_freelist; 4637 ukh->ukh_freelist = entry; 4638 } 4639 4640 /* 4641 * uvm_kmapent_alloc: allocate a map entry for in-kernel map 4642 */ 4643 4644 static struct vm_map_entry * 4645 uvm_kmapent_alloc(struct vm_map *map, int flags) 4646 { 4647 struct vm_page *pg; 4648 struct uvm_kmapent_hdr *ukh; 4649 struct vm_map_entry *entry; 4650 #ifndef PMAP_MAP_POOLPAGE 4651 struct uvm_map_args args; 4652 uvm_flag_t mapflags = UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, 4653 UVM_INH_NONE, UVM_ADV_RANDOM, flags | UVM_FLAG_NOMERGE); 4654 int error; 4655 #endif 4656 vaddr_t va; 4657 int i; 4658 4659 KDASSERT(UVM_KMAPENT_CHUNK > 2); 4660 KDASSERT(kernel_map != NULL); 4661 KASSERT(vm_map_pmap(map) == pmap_kernel()); 4662 4663 UVMMAP_EVCNT_INCR(uke_alloc); 4664 entry = NULL; 4665 again: 4666 /* 4667 * try to grab an entry from freelist. 4668 */ 4669 mutex_spin_enter(&uvm_kentry_lock); 4670 ukh = LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free); 4671 if (ukh) { 4672 entry = uvm_kmapent_get(ukh); 4673 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK) 4674 LIST_REMOVE(ukh, ukh_listq); 4675 } 4676 mutex_spin_exit(&uvm_kentry_lock); 4677 4678 if (entry) 4679 return entry; 4680 4681 /* 4682 * there's no free entry for this vm_map. 4683 * now we need to allocate some vm_map_entry. 4684 * for simplicity, always allocate one page chunk of them at once. 4685 */ 4686 4687 pg = uvm_pagealloc(NULL, 0, NULL, 4688 (flags & UVM_KMF_NOWAIT) != 0 ? UVM_PGA_USERESERVE : 0); 4689 if (__predict_false(pg == NULL)) { 4690 if (flags & UVM_FLAG_NOWAIT) 4691 return NULL; 4692 uvm_wait("kme_alloc"); 4693 goto again; 4694 } 4695 4696 #ifdef PMAP_MAP_POOLPAGE 4697 va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg)); 4698 KASSERT(va != 0); 4699 #else 4700 error = uvm_map_prepare(map, 0, PAGE_SIZE, NULL, UVM_UNKNOWN_OFFSET, 4701 0, mapflags, &args); 4702 if (error) { 4703 uvm_pagefree(pg); 4704 return NULL; 4705 } 4706 4707 va = args.uma_start; 4708 4709 pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), 4710 VM_PROT_READ|VM_PROT_WRITE|PMAP_KMPAGE, 0); 4711 pmap_update(vm_map_pmap(map)); 4712 4713 #endif 4714 ukh = (void *)va; 4715 4716 /* 4717 * use the last entry for ukh itsself. 4718 */ 4719 4720 i = UVM_KMAPENT_CHUNK - 1; 4721 #ifndef PMAP_MAP_POOLPAGE 4722 entry = &ukh->ukh_entries[i--]; 4723 entry->flags = UVM_MAP_KERNEL | UVM_MAP_KMAPENT; 4724 error = uvm_map_enter(map, &args, entry); 4725 KASSERT(error == 0); 4726 #endif 4727 4728 ukh->ukh_nused = UVM_KMAPENT_CHUNK; 4729 ukh->ukh_map = map; 4730 ukh->ukh_freelist = NULL; 4731 for (; i >= 1; i--) { 4732 struct vm_map_entry *xentry = &ukh->ukh_entries[i]; 4733 4734 xentry->flags = UVM_MAP_KERNEL; 4735 uvm_kmapent_put(ukh, xentry); 4736 } 4737 #ifdef PMAP_MAP_POOLPAGE 4738 KASSERT(ukh->ukh_nused == 1); 4739 #else 4740 KASSERT(ukh->ukh_nused == 2); 4741 #endif 4742 4743 mutex_spin_enter(&uvm_kentry_lock); 4744 LIST_INSERT_HEAD(&vm_map_to_kernel(map)->vmk_kentry_free, 4745 ukh, ukh_listq); 4746 mutex_spin_exit(&uvm_kentry_lock); 4747 4748 /* 4749 * return first entry. 4750 */ 4751 4752 entry = &ukh->ukh_entries[0]; 4753 entry->flags = UVM_MAP_KERNEL; 4754 UVMMAP_EVCNT_INCR(ukh_alloc); 4755 4756 return entry; 4757 } 4758 4759 /* 4760 * uvm_mapent_free: free map entry for in-kernel map 4761 */ 4762 4763 static void 4764 uvm_kmapent_free(struct vm_map_entry *entry) 4765 { 4766 struct uvm_kmapent_hdr *ukh; 4767 struct vm_page *pg; 4768 struct vm_map *map; 4769 #ifndef PMAP_UNMAP_POOLPAGE 4770 struct pmap *pmap; 4771 struct vm_map_entry *deadentry; 4772 #endif 4773 vaddr_t va; 4774 paddr_t pa; 4775 4776 UVMMAP_EVCNT_INCR(uke_free); 4777 ukh = UVM_KHDR_FIND(entry); 4778 map = ukh->ukh_map; 4779 4780 mutex_spin_enter(&uvm_kentry_lock); 4781 uvm_kmapent_put(ukh, entry); 4782 #ifdef PMAP_UNMAP_POOLPAGE 4783 if (ukh->ukh_nused > 0) { 4784 #else 4785 if (ukh->ukh_nused > 1) { 4786 #endif 4787 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK - 1) 4788 LIST_INSERT_HEAD( 4789 &vm_map_to_kernel(map)->vmk_kentry_free, 4790 ukh, ukh_listq); 4791 mutex_spin_exit(&uvm_kentry_lock); 4792 return; 4793 } 4794 4795 /* 4796 * now we can free this ukh. 4797 * 4798 * however, keep an empty ukh to avoid ping-pong. 4799 */ 4800 4801 if (LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free) == ukh && 4802 LIST_NEXT(ukh, ukh_listq) == NULL) { 4803 mutex_spin_exit(&uvm_kentry_lock); 4804 return; 4805 } 4806 LIST_REMOVE(ukh, ukh_listq); 4807 mutex_spin_exit(&uvm_kentry_lock); 4808 4809 va = (vaddr_t)ukh; 4810 4811 #ifdef PMAP_UNMAP_POOLPAGE 4812 KASSERT(ukh->ukh_nused == 0); 4813 pa = PMAP_UNMAP_POOLPAGE(va); 4814 KASSERT(pa != 0); 4815 #else 4816 KASSERT(ukh->ukh_nused == 1); 4817 4818 /* 4819 * remove map entry for ukh itsself. 4820 */ 4821 4822 KASSERT((va & PAGE_MASK) == 0); 4823 vm_map_lock(map); 4824 uvm_unmap_remove(map, va, va + PAGE_SIZE, &deadentry, NULL, 0); 4825 KASSERT(deadentry->flags & UVM_MAP_KERNEL); 4826 KASSERT(deadentry->flags & UVM_MAP_KMAPENT); 4827 KASSERT(deadentry->next == NULL); 4828 KASSERT(deadentry == &ukh->ukh_entries[UVM_KMAPENT_CHUNK - 1]); 4829 4830 /* 4831 * unmap the page from pmap and free it. 4832 */ 4833 4834 pmap = vm_map_pmap(map); 4835 KASSERT(pmap == pmap_kernel()); 4836 if (!pmap_extract(pmap, va, &pa)) 4837 panic("%s: no mapping", __func__); 4838 pmap_kremove(va, PAGE_SIZE); 4839 pmap_update(vm_map_pmap(map)); 4840 vm_map_unlock(map); 4841 #endif /* !PMAP_UNMAP_POOLPAGE */ 4842 pg = PHYS_TO_VM_PAGE(pa); 4843 uvm_pagefree(pg); 4844 UVMMAP_EVCNT_INCR(ukh_free); 4845 } 4846 4847 static vsize_t 4848 uvm_kmapent_overhead(vsize_t size) 4849 { 4850 4851 /* 4852 * - the max number of unmerged entries is howmany(size, PAGE_SIZE) 4853 * as the min allocation unit is PAGE_SIZE. 4854 * - UVM_KMAPENT_CHUNK "kmapent"s are allocated from a page. 4855 * one of them are used to map the page itself. 4856 */ 4857 4858 return howmany(howmany(size, PAGE_SIZE), (UVM_KMAPENT_CHUNK - 1)) * 4859 PAGE_SIZE; 4860 } 4861 4862 /* 4863 * map entry reservation 4864 */ 4865 4866 /* 4867 * uvm_mapent_reserve: reserve map entries for clipping before locking map. 4868 * 4869 * => needed when unmapping entries allocated without UVM_FLAG_QUANTUM. 4870 * => caller shouldn't hold map locked. 4871 */ 4872 int 4873 uvm_mapent_reserve(struct vm_map *map, struct uvm_mapent_reservation *umr, 4874 int nentries, int flags) 4875 { 4876 4877 umr->umr_nentries = 0; 4878 4879 if ((flags & UVM_FLAG_QUANTUM) != 0) 4880 return 0; 4881 4882 if (!VM_MAP_USE_KMAPENT(map)) 4883 return 0; 4884 4885 while (nentries--) { 4886 struct vm_map_entry *ent; 4887 ent = uvm_kmapent_alloc(map, flags); 4888 if (!ent) { 4889 uvm_mapent_unreserve(map, umr); 4890 return ENOMEM; 4891 } 4892 UMR_PUTENTRY(umr, ent); 4893 } 4894 4895 return 0; 4896 } 4897 4898 /* 4899 * uvm_mapent_unreserve: 4900 * 4901 * => caller shouldn't hold map locked. 4902 * => never fail or sleep. 4903 */ 4904 void 4905 uvm_mapent_unreserve(struct vm_map *map, struct uvm_mapent_reservation *umr) 4906 { 4907 4908 while (!UMR_EMPTY(umr)) 4909 uvm_kmapent_free(UMR_GETENTRY(umr)); 4910 } 4911 4912 /* 4913 * uvm_mapent_trymerge: try to merge an entry with its neighbors. 4914 * 4915 * => called with map locked. 4916 * => return non zero if successfully merged. 4917 */ 4918 4919 int 4920 uvm_mapent_trymerge(struct vm_map *map, struct vm_map_entry *entry, int flags) 4921 { 4922 struct uvm_object *uobj; 4923 struct vm_map_entry *next; 4924 struct vm_map_entry *prev; 4925 vsize_t size; 4926 int merged = 0; 4927 bool copying; 4928 int newetype; 4929 4930 if (VM_MAP_USE_KMAPENT(map)) { 4931 return 0; 4932 } 4933 if (entry->aref.ar_amap != NULL) { 4934 return 0; 4935 } 4936 if ((entry->flags & UVM_MAP_NOMERGE) != 0) { 4937 return 0; 4938 } 4939 4940 uobj = entry->object.uvm_obj; 4941 size = entry->end - entry->start; 4942 copying = (flags & UVM_MERGE_COPYING) != 0; 4943 newetype = copying ? (entry->etype & ~UVM_ET_NEEDSCOPY) : entry->etype; 4944 4945 next = entry->next; 4946 if (next != &map->header && 4947 next->start == entry->end && 4948 ((copying && next->aref.ar_amap != NULL && 4949 amap_refs(next->aref.ar_amap) == 1) || 4950 (!copying && next->aref.ar_amap == NULL)) && 4951 UVM_ET_ISCOMPATIBLE(next, newetype, 4952 uobj, entry->flags, entry->protection, 4953 entry->max_protection, entry->inheritance, entry->advice, 4954 entry->wired_count) && 4955 (uobj == NULL || entry->offset + size == next->offset)) { 4956 int error; 4957 4958 if (copying) { 4959 error = amap_extend(next, size, 4960 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_BACKWARDS); 4961 } else { 4962 error = 0; 4963 } 4964 if (error == 0) { 4965 if (uobj) { 4966 if (uobj->pgops->pgo_detach) { 4967 uobj->pgops->pgo_detach(uobj); 4968 } 4969 } 4970 4971 entry->end = next->end; 4972 clear_hints(map, next); 4973 uvm_map_entry_unlink(map, next); 4974 if (copying) { 4975 entry->aref = next->aref; 4976 entry->etype &= ~UVM_ET_NEEDSCOPY; 4977 } 4978 uvm_map_check(map, "trymerge forwardmerge"); 4979 uvm_mapent_free_merged(map, next); 4980 merged++; 4981 } 4982 } 4983 4984 prev = entry->prev; 4985 if (prev != &map->header && 4986 prev->end == entry->start && 4987 ((copying && !merged && prev->aref.ar_amap != NULL && 4988 amap_refs(prev->aref.ar_amap) == 1) || 4989 (!copying && prev->aref.ar_amap == NULL)) && 4990 UVM_ET_ISCOMPATIBLE(prev, newetype, 4991 uobj, entry->flags, entry->protection, 4992 entry->max_protection, entry->inheritance, entry->advice, 4993 entry->wired_count) && 4994 (uobj == NULL || 4995 prev->offset + prev->end - prev->start == entry->offset)) { 4996 int error; 4997 4998 if (copying) { 4999 error = amap_extend(prev, size, 5000 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_FORWARDS); 5001 } else { 5002 error = 0; 5003 } 5004 if (error == 0) { 5005 if (uobj) { 5006 if (uobj->pgops->pgo_detach) { 5007 uobj->pgops->pgo_detach(uobj); 5008 } 5009 entry->offset = prev->offset; 5010 } 5011 5012 entry->start = prev->start; 5013 clear_hints(map, prev); 5014 uvm_map_entry_unlink(map, prev); 5015 if (copying) { 5016 entry->aref = prev->aref; 5017 entry->etype &= ~UVM_ET_NEEDSCOPY; 5018 } 5019 uvm_map_check(map, "trymerge backmerge"); 5020 uvm_mapent_free_merged(map, prev); 5021 merged++; 5022 } 5023 } 5024 5025 return merged; 5026 } 5027 5028 /* 5029 * uvm_map_create: create map 5030 */ 5031 5032 struct vm_map * 5033 uvm_map_create(pmap_t pmap, vaddr_t vmin, vaddr_t vmax, int flags) 5034 { 5035 struct vm_map *result; 5036 5037 result = malloc(sizeof(struct vm_map), M_VMMAP, M_WAITOK); 5038 uvm_map_setup(result, vmin, vmax, flags); 5039 result->pmap = pmap; 5040 return(result); 5041 } 5042 5043 /* 5044 * uvm_map_setup: init map 5045 * 5046 * => map must not be in service yet. 5047 */ 5048 5049 void 5050 uvm_map_setup(struct vm_map *map, vaddr_t vmin, vaddr_t vmax, int flags) 5051 { 5052 int ipl; 5053 5054 rb_tree_init(&map->rb_tree, &uvm_map_tree_ops); 5055 map->header.next = map->header.prev = &map->header; 5056 map->nentries = 0; 5057 map->size = 0; 5058 map->ref_count = 1; 5059 vm_map_setmin(map, vmin); 5060 vm_map_setmax(map, vmax); 5061 map->flags = flags; 5062 map->first_free = &map->header; 5063 map->hint = &map->header; 5064 map->timestamp = 0; 5065 map->busy = NULL; 5066 5067 if ((flags & VM_MAP_INTRSAFE) != 0) { 5068 ipl = IPL_VM; 5069 } else { 5070 ipl = IPL_NONE; 5071 } 5072 5073 rw_init(&map->lock); 5074 cv_init(&map->cv, "vm_map"); 5075 mutex_init(&map->misc_lock, MUTEX_DRIVER, ipl); 5076 mutex_init(&map->mutex, MUTEX_DRIVER, ipl); 5077 } 5078 5079 5080 /* 5081 * U N M A P - m a i n e n t r y p o i n t 5082 */ 5083 5084 /* 5085 * uvm_unmap1: remove mappings from a vm_map (from "start" up to "stop") 5086 * 5087 * => caller must check alignment and size 5088 * => map must be unlocked (we will lock it) 5089 * => flags is UVM_FLAG_QUANTUM or 0. 5090 */ 5091 5092 void 5093 uvm_unmap1(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 5094 { 5095 struct vm_map_entry *dead_entries; 5096 struct uvm_mapent_reservation umr; 5097 UVMHIST_FUNC("uvm_unmap"); UVMHIST_CALLED(maphist); 5098 5099 UVMHIST_LOG(maphist, " (map=0x%x, start=0x%x, end=0x%x)", 5100 map, start, end, 0); 5101 if (map == kernel_map) { 5102 LOCKDEBUG_MEM_CHECK((void *)start, end - start); 5103 } 5104 /* 5105 * work now done by helper functions. wipe the pmap's and then 5106 * detach from the dead entries... 5107 */ 5108 uvm_mapent_reserve(map, &umr, 2, flags); 5109 vm_map_lock(map); 5110 uvm_unmap_remove(map, start, end, &dead_entries, &umr, flags); 5111 vm_map_unlock(map); 5112 uvm_mapent_unreserve(map, &umr); 5113 5114 if (dead_entries != NULL) 5115 uvm_unmap_detach(dead_entries, 0); 5116 5117 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 5118 } 5119 5120 5121 /* 5122 * uvm_map_reference: add reference to a map 5123 * 5124 * => map need not be locked (we use misc_lock). 5125 */ 5126 5127 void 5128 uvm_map_reference(struct vm_map *map) 5129 { 5130 mutex_enter(&map->misc_lock); 5131 map->ref_count++; 5132 mutex_exit(&map->misc_lock); 5133 } 5134 5135 struct vm_map_kernel * 5136 vm_map_to_kernel(struct vm_map *map) 5137 { 5138 5139 KASSERT(VM_MAP_IS_KERNEL(map)); 5140 5141 return (struct vm_map_kernel *)map; 5142 } 5143 5144 bool 5145 vm_map_starved_p(struct vm_map *map) 5146 { 5147 5148 if ((map->flags & VM_MAP_WANTVA) != 0) { 5149 return true; 5150 } 5151 /* XXX */ 5152 if ((vm_map_max(map) - vm_map_min(map)) / 16 * 15 < map->size) { 5153 return true; 5154 } 5155 return false; 5156 } 5157 5158 #if defined(DDB) || defined(DEBUGPRINT) 5159 5160 /* 5161 * uvm_map_printit: actually prints the map 5162 */ 5163 5164 void 5165 uvm_map_printit(struct vm_map *map, bool full, 5166 void (*pr)(const char *, ...)) 5167 { 5168 struct vm_map_entry *entry; 5169 5170 (*pr)("MAP %p: [0x%lx->0x%lx]\n", map, vm_map_min(map), 5171 vm_map_max(map)); 5172 (*pr)("\t#ent=%d, sz=%d, ref=%d, version=%d, flags=0x%x\n", 5173 map->nentries, map->size, map->ref_count, map->timestamp, 5174 map->flags); 5175 (*pr)("\tpmap=%p(resident=%ld, wired=%ld)\n", map->pmap, 5176 pmap_resident_count(map->pmap), pmap_wired_count(map->pmap)); 5177 if (!full) 5178 return; 5179 for (entry = map->header.next; entry != &map->header; 5180 entry = entry->next) { 5181 (*pr)(" - %p: 0x%lx->0x%lx: obj=%p/0x%llx, amap=%p/%d\n", 5182 entry, entry->start, entry->end, entry->object.uvm_obj, 5183 (long long)entry->offset, entry->aref.ar_amap, 5184 entry->aref.ar_pageoff); 5185 (*pr)( 5186 "\tsubmap=%c, cow=%c, nc=%c, prot(max)=%d/%d, inh=%d, " 5187 "wc=%d, adv=%d\n", 5188 (entry->etype & UVM_ET_SUBMAP) ? 'T' : 'F', 5189 (entry->etype & UVM_ET_COPYONWRITE) ? 'T' : 'F', 5190 (entry->etype & UVM_ET_NEEDSCOPY) ? 'T' : 'F', 5191 entry->protection, entry->max_protection, 5192 entry->inheritance, entry->wired_count, entry->advice); 5193 } 5194 } 5195 5196 void 5197 uvm_whatis(uintptr_t addr, void (*pr)(const char *, ...)) 5198 { 5199 struct vm_map *map; 5200 5201 for (map = kernel_map;;) { 5202 struct vm_map_entry *entry; 5203 5204 if (!uvm_map_lookup_entry_bytree(map, (vaddr_t)addr, &entry)) { 5205 break; 5206 } 5207 (*pr)("%p is %p+%zu from VMMAP %p\n", 5208 (void *)addr, (void *)entry->start, 5209 (size_t)(addr - (uintptr_t)entry->start), map); 5210 if (!UVM_ET_ISSUBMAP(entry)) { 5211 break; 5212 } 5213 map = entry->object.sub_map; 5214 } 5215 } 5216 5217 #endif /* DDB || DEBUGPRINT */ 5218