uvm_map.c revision 1.275
1 /* $NetBSD: uvm_map.c,v 1.275 2009/08/01 16:35:51 yamt 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 (at) 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 1.275 2009/08/01 16:35:51 yamt 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%lx >= 0x%lx\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 return 0; 1349 } 1350 1351 int 1352 uvm_map_enter(struct vm_map *map, const struct uvm_map_args *args, 1353 struct vm_map_entry *new_entry) 1354 { 1355 struct vm_map_entry *prev_entry = args->uma_prev; 1356 struct vm_map_entry *dead = NULL; 1357 1358 const uvm_flag_t flags = args->uma_flags; 1359 const vm_prot_t prot = UVM_PROTECTION(flags); 1360 const vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1361 const vm_inherit_t inherit = UVM_INHERIT(flags); 1362 const int amapwaitflag = (flags & UVM_FLAG_NOWAIT) ? 1363 AMAP_EXTEND_NOWAIT : 0; 1364 const int advice = UVM_ADVICE(flags); 1365 const int meflagval = (flags & UVM_FLAG_QUANTUM) ? 1366 UVM_MAP_QUANTUM : 0; 1367 1368 vaddr_t start = args->uma_start; 1369 vsize_t size = args->uma_size; 1370 struct uvm_object *uobj = args->uma_uobj; 1371 voff_t uoffset = args->uma_uoffset; 1372 1373 const int kmap = (vm_map_pmap(map) == pmap_kernel()); 1374 int merged = 0; 1375 int error; 1376 int newetype; 1377 1378 UVMHIST_FUNC("uvm_map_enter"); 1379 UVMHIST_CALLED(maphist); 1380 1381 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1382 map, start, size, flags); 1383 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1384 1385 KASSERT(map->hint == prev_entry); /* bimerge case assumes this */ 1386 1387 if (flags & UVM_FLAG_QUANTUM) { 1388 KASSERT(new_entry); 1389 KASSERT(new_entry->flags & UVM_MAP_QUANTUM); 1390 } 1391 1392 if (uobj) 1393 newetype = UVM_ET_OBJ; 1394 else 1395 newetype = 0; 1396 1397 if (flags & UVM_FLAG_COPYONW) { 1398 newetype |= UVM_ET_COPYONWRITE; 1399 if ((flags & UVM_FLAG_OVERLAY) == 0) 1400 newetype |= UVM_ET_NEEDSCOPY; 1401 } 1402 1403 /* 1404 * try and insert in map by extending previous entry, if possible. 1405 * XXX: we don't try and pull back the next entry. might be useful 1406 * for a stack, but we are currently allocating our stack in advance. 1407 */ 1408 1409 if (flags & UVM_FLAG_NOMERGE) 1410 goto nomerge; 1411 1412 if (prev_entry->end == start && 1413 prev_entry != &map->header && 1414 UVM_ET_ISCOMPATIBLE(prev_entry, newetype, uobj, meflagval, 1415 prot, maxprot, inherit, advice, 0)) { 1416 1417 if (uobj && prev_entry->offset + 1418 (prev_entry->end - prev_entry->start) != uoffset) 1419 goto forwardmerge; 1420 1421 /* 1422 * can't extend a shared amap. note: no need to lock amap to 1423 * look at refs since we don't care about its exact value. 1424 * if it is one (i.e. we have only reference) it will stay there 1425 */ 1426 1427 if (prev_entry->aref.ar_amap && 1428 amap_refs(prev_entry->aref.ar_amap) != 1) { 1429 goto forwardmerge; 1430 } 1431 1432 if (prev_entry->aref.ar_amap) { 1433 error = amap_extend(prev_entry, size, 1434 amapwaitflag | AMAP_EXTEND_FORWARDS); 1435 if (error) 1436 goto nomerge; 1437 } 1438 1439 if (kmap) { 1440 UVMMAP_EVCNT_INCR(kbackmerge); 1441 } else { 1442 UVMMAP_EVCNT_INCR(ubackmerge); 1443 } 1444 UVMHIST_LOG(maphist," starting back merge", 0, 0, 0, 0); 1445 1446 /* 1447 * drop our reference to uobj since we are extending a reference 1448 * that we already have (the ref count can not drop to zero). 1449 */ 1450 1451 if (uobj && uobj->pgops->pgo_detach) 1452 uobj->pgops->pgo_detach(uobj); 1453 1454 /* 1455 * Now that we've merged the entries, note that we've grown 1456 * and our gap has shrunk. Then fix the tree. 1457 */ 1458 prev_entry->end += size; 1459 prev_entry->gap -= size; 1460 uvm_rb_fixup(map, prev_entry); 1461 1462 uvm_map_check(map, "map backmerged"); 1463 1464 UVMHIST_LOG(maphist,"<- done (via backmerge)!", 0, 0, 0, 0); 1465 merged++; 1466 } 1467 1468 forwardmerge: 1469 if (prev_entry->next->start == (start + size) && 1470 prev_entry->next != &map->header && 1471 UVM_ET_ISCOMPATIBLE(prev_entry->next, newetype, uobj, meflagval, 1472 prot, maxprot, inherit, advice, 0)) { 1473 1474 if (uobj && prev_entry->next->offset != uoffset + size) 1475 goto nomerge; 1476 1477 /* 1478 * can't extend a shared amap. note: no need to lock amap to 1479 * look at refs since we don't care about its exact value. 1480 * if it is one (i.e. we have only reference) it will stay there. 1481 * 1482 * note that we also can't merge two amaps, so if we 1483 * merged with the previous entry which has an amap, 1484 * and the next entry also has an amap, we give up. 1485 * 1486 * Interesting cases: 1487 * amap, new, amap -> give up second merge (single fwd extend) 1488 * amap, new, none -> double forward extend (extend again here) 1489 * none, new, amap -> double backward extend (done here) 1490 * uobj, new, amap -> single backward extend (done here) 1491 * 1492 * XXX should we attempt to deal with someone refilling 1493 * the deallocated region between two entries that are 1494 * backed by the same amap (ie, arefs is 2, "prev" and 1495 * "next" refer to it, and adding this allocation will 1496 * close the hole, thus restoring arefs to 1 and 1497 * deallocating the "next" vm_map_entry)? -- @@@ 1498 */ 1499 1500 if (prev_entry->next->aref.ar_amap && 1501 (amap_refs(prev_entry->next->aref.ar_amap) != 1 || 1502 (merged && prev_entry->aref.ar_amap))) { 1503 goto nomerge; 1504 } 1505 1506 if (merged) { 1507 /* 1508 * Try to extend the amap of the previous entry to 1509 * cover the next entry as well. If it doesn't work 1510 * just skip on, don't actually give up, since we've 1511 * already completed the back merge. 1512 */ 1513 if (prev_entry->aref.ar_amap) { 1514 if (amap_extend(prev_entry, 1515 prev_entry->next->end - 1516 prev_entry->next->start, 1517 amapwaitflag | AMAP_EXTEND_FORWARDS)) 1518 goto nomerge; 1519 } 1520 1521 /* 1522 * Try to extend the amap of the *next* entry 1523 * back to cover the new allocation *and* the 1524 * previous entry as well (the previous merge 1525 * didn't have an amap already otherwise we 1526 * wouldn't be checking here for an amap). If 1527 * it doesn't work just skip on, again, don't 1528 * actually give up, since we've already 1529 * completed the back merge. 1530 */ 1531 else if (prev_entry->next->aref.ar_amap) { 1532 if (amap_extend(prev_entry->next, 1533 prev_entry->end - 1534 prev_entry->start, 1535 amapwaitflag | AMAP_EXTEND_BACKWARDS)) 1536 goto nomerge; 1537 } 1538 } else { 1539 /* 1540 * Pull the next entry's amap backwards to cover this 1541 * new allocation. 1542 */ 1543 if (prev_entry->next->aref.ar_amap) { 1544 error = amap_extend(prev_entry->next, size, 1545 amapwaitflag | AMAP_EXTEND_BACKWARDS); 1546 if (error) 1547 goto nomerge; 1548 } 1549 } 1550 1551 if (merged) { 1552 if (kmap) { 1553 UVMMAP_EVCNT_DECR(kbackmerge); 1554 UVMMAP_EVCNT_INCR(kbimerge); 1555 } else { 1556 UVMMAP_EVCNT_DECR(ubackmerge); 1557 UVMMAP_EVCNT_INCR(ubimerge); 1558 } 1559 } else { 1560 if (kmap) { 1561 UVMMAP_EVCNT_INCR(kforwmerge); 1562 } else { 1563 UVMMAP_EVCNT_INCR(uforwmerge); 1564 } 1565 } 1566 UVMHIST_LOG(maphist," starting forward merge", 0, 0, 0, 0); 1567 1568 /* 1569 * drop our reference to uobj since we are extending a reference 1570 * that we already have (the ref count can not drop to zero). 1571 * (if merged, we've already detached) 1572 */ 1573 if (uobj && uobj->pgops->pgo_detach && !merged) 1574 uobj->pgops->pgo_detach(uobj); 1575 1576 if (merged) { 1577 dead = prev_entry->next; 1578 prev_entry->end = dead->end; 1579 uvm_map_entry_unlink(map, dead); 1580 if (dead->aref.ar_amap != NULL) { 1581 prev_entry->aref = dead->aref; 1582 dead->aref.ar_amap = NULL; 1583 } 1584 } else { 1585 prev_entry->next->start -= size; 1586 if (prev_entry != &map->header) { 1587 prev_entry->gap -= size; 1588 KASSERT(prev_entry->gap == uvm_rb_gap(prev_entry)); 1589 uvm_rb_fixup(map, prev_entry); 1590 } 1591 if (uobj) 1592 prev_entry->next->offset = uoffset; 1593 } 1594 1595 uvm_map_check(map, "map forwardmerged"); 1596 1597 UVMHIST_LOG(maphist,"<- done forwardmerge", 0, 0, 0, 0); 1598 merged++; 1599 } 1600 1601 nomerge: 1602 if (!merged) { 1603 UVMHIST_LOG(maphist," allocating new map entry", 0, 0, 0, 0); 1604 if (kmap) { 1605 UVMMAP_EVCNT_INCR(knomerge); 1606 } else { 1607 UVMMAP_EVCNT_INCR(unomerge); 1608 } 1609 1610 /* 1611 * allocate new entry and link it in. 1612 */ 1613 1614 if (new_entry == NULL) { 1615 new_entry = uvm_mapent_alloc(map, 1616 (flags & UVM_FLAG_NOWAIT)); 1617 if (__predict_false(new_entry == NULL)) { 1618 error = ENOMEM; 1619 goto done; 1620 } 1621 } 1622 new_entry->start = start; 1623 new_entry->end = new_entry->start + size; 1624 new_entry->object.uvm_obj = uobj; 1625 new_entry->offset = uoffset; 1626 1627 new_entry->etype = newetype; 1628 1629 if (flags & UVM_FLAG_NOMERGE) { 1630 new_entry->flags |= UVM_MAP_NOMERGE; 1631 } 1632 1633 new_entry->protection = prot; 1634 new_entry->max_protection = maxprot; 1635 new_entry->inheritance = inherit; 1636 new_entry->wired_count = 0; 1637 new_entry->advice = advice; 1638 if (flags & UVM_FLAG_OVERLAY) { 1639 1640 /* 1641 * to_add: for BSS we overallocate a little since we 1642 * are likely to extend 1643 */ 1644 1645 vaddr_t to_add = (flags & UVM_FLAG_AMAPPAD) ? 1646 UVM_AMAP_CHUNK << PAGE_SHIFT : 0; 1647 struct vm_amap *amap = amap_alloc(size, to_add, 1648 (flags & UVM_FLAG_NOWAIT)); 1649 if (__predict_false(amap == NULL)) { 1650 error = ENOMEM; 1651 goto done; 1652 } 1653 new_entry->aref.ar_pageoff = 0; 1654 new_entry->aref.ar_amap = amap; 1655 } else { 1656 new_entry->aref.ar_pageoff = 0; 1657 new_entry->aref.ar_amap = NULL; 1658 } 1659 uvm_map_entry_link(map, prev_entry, new_entry); 1660 1661 /* 1662 * Update the free space hint 1663 */ 1664 1665 if ((map->first_free == prev_entry) && 1666 (prev_entry->end >= new_entry->start)) 1667 map->first_free = new_entry; 1668 1669 new_entry = NULL; 1670 } 1671 1672 map->size += size; 1673 1674 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 1675 1676 error = 0; 1677 done: 1678 if ((flags & UVM_FLAG_QUANTUM) == 0) { 1679 /* 1680 * vmk_merged_entries is locked by the map's lock. 1681 */ 1682 vm_map_unlock(map); 1683 } 1684 if (new_entry && error == 0) { 1685 KDASSERT(merged); 1686 uvm_mapent_free_merged(map, new_entry); 1687 new_entry = NULL; 1688 } 1689 if (dead) { 1690 KDASSERT(merged); 1691 uvm_mapent_free_merged(map, dead); 1692 } 1693 if ((flags & UVM_FLAG_QUANTUM) != 0) { 1694 vm_map_unlock(map); 1695 } 1696 if (new_entry != NULL) { 1697 uvm_mapent_free(new_entry); 1698 } 1699 return error; 1700 } 1701 1702 /* 1703 * uvm_map_lookup_entry_bytree: lookup an entry in tree 1704 */ 1705 1706 static inline bool 1707 uvm_map_lookup_entry_bytree(struct vm_map *map, vaddr_t address, 1708 struct vm_map_entry **entry /* OUT */) 1709 { 1710 struct vm_map_entry *prev = &map->header; 1711 struct vm_map_entry *cur = ROOT_ENTRY(map); 1712 1713 while (cur) { 1714 UVMMAP_EVCNT_INCR(mlk_treeloop); 1715 if (address >= cur->start) { 1716 if (address < cur->end) { 1717 *entry = cur; 1718 return true; 1719 } 1720 prev = cur; 1721 cur = RIGHT_ENTRY(cur); 1722 } else 1723 cur = LEFT_ENTRY(cur); 1724 } 1725 *entry = prev; 1726 return false; 1727 } 1728 1729 /* 1730 * uvm_map_lookup_entry: find map entry at or before an address 1731 * 1732 * => map must at least be read-locked by caller 1733 * => entry is returned in "entry" 1734 * => return value is true if address is in the returned entry 1735 */ 1736 1737 bool 1738 uvm_map_lookup_entry(struct vm_map *map, vaddr_t address, 1739 struct vm_map_entry **entry /* OUT */) 1740 { 1741 struct vm_map_entry *cur; 1742 bool use_tree = false; 1743 UVMHIST_FUNC("uvm_map_lookup_entry"); 1744 UVMHIST_CALLED(maphist); 1745 1746 UVMHIST_LOG(maphist,"(map=0x%x,addr=0x%x,ent=0x%x)", 1747 map, address, entry, 0); 1748 1749 /* 1750 * start looking either from the head of the 1751 * list, or from the hint. 1752 */ 1753 1754 cur = map->hint; 1755 1756 if (cur == &map->header) 1757 cur = cur->next; 1758 1759 UVMMAP_EVCNT_INCR(mlk_call); 1760 if (address >= cur->start) { 1761 1762 /* 1763 * go from hint to end of list. 1764 * 1765 * but first, make a quick check to see if 1766 * we are already looking at the entry we 1767 * want (which is usually the case). 1768 * note also that we don't need to save the hint 1769 * here... it is the same hint (unless we are 1770 * at the header, in which case the hint didn't 1771 * buy us anything anyway). 1772 */ 1773 1774 if (cur != &map->header && cur->end > address) { 1775 UVMMAP_EVCNT_INCR(mlk_hint); 1776 *entry = cur; 1777 UVMHIST_LOG(maphist,"<- got it via hint (0x%x)", 1778 cur, 0, 0, 0); 1779 uvm_mapent_check(*entry); 1780 return (true); 1781 } 1782 1783 if (map->nentries > 15) 1784 use_tree = true; 1785 } else { 1786 1787 /* 1788 * invalid hint. use tree. 1789 */ 1790 use_tree = true; 1791 } 1792 1793 uvm_map_check(map, __func__); 1794 1795 if (use_tree) { 1796 /* 1797 * Simple lookup in the tree. Happens when the hint is 1798 * invalid, or nentries reach a threshold. 1799 */ 1800 UVMMAP_EVCNT_INCR(mlk_tree); 1801 if (uvm_map_lookup_entry_bytree(map, address, entry)) { 1802 goto got; 1803 } else { 1804 goto failed; 1805 } 1806 } 1807 1808 /* 1809 * search linearly 1810 */ 1811 1812 UVMMAP_EVCNT_INCR(mlk_list); 1813 while (cur != &map->header) { 1814 UVMMAP_EVCNT_INCR(mlk_listloop); 1815 if (cur->end > address) { 1816 if (address >= cur->start) { 1817 /* 1818 * save this lookup for future 1819 * hints, and return 1820 */ 1821 1822 *entry = cur; 1823 got: 1824 SAVE_HINT(map, map->hint, *entry); 1825 UVMHIST_LOG(maphist,"<- search got it (0x%x)", 1826 cur, 0, 0, 0); 1827 KDASSERT((*entry)->start <= address); 1828 KDASSERT(address < (*entry)->end); 1829 uvm_mapent_check(*entry); 1830 return (true); 1831 } 1832 break; 1833 } 1834 cur = cur->next; 1835 } 1836 *entry = cur->prev; 1837 failed: 1838 SAVE_HINT(map, map->hint, *entry); 1839 UVMHIST_LOG(maphist,"<- failed!",0,0,0,0); 1840 KDASSERT((*entry) == &map->header || (*entry)->end <= address); 1841 KDASSERT((*entry)->next == &map->header || 1842 address < (*entry)->next->start); 1843 return (false); 1844 } 1845 1846 /* 1847 * See if the range between start and start + length fits in the gap 1848 * entry->next->start and entry->end. Returns 1 if fits, 0 if doesn't 1849 * fit, and -1 address wraps around. 1850 */ 1851 static int 1852 uvm_map_space_avail(vaddr_t *start, vsize_t length, voff_t uoffset, 1853 vsize_t align, int topdown, struct vm_map_entry *entry) 1854 { 1855 vaddr_t end; 1856 1857 #ifdef PMAP_PREFER 1858 /* 1859 * push start address forward as needed to avoid VAC alias problems. 1860 * we only do this if a valid offset is specified. 1861 */ 1862 1863 if (uoffset != UVM_UNKNOWN_OFFSET) 1864 PMAP_PREFER(uoffset, start, length, topdown); 1865 #endif 1866 if (align != 0) { 1867 if ((*start & (align - 1)) != 0) { 1868 if (topdown) 1869 *start &= ~(align - 1); 1870 else 1871 *start = roundup(*start, align); 1872 } 1873 /* 1874 * XXX Should we PMAP_PREFER() here again? 1875 * eh...i think we're okay 1876 */ 1877 } 1878 1879 /* 1880 * Find the end of the proposed new region. Be sure we didn't 1881 * wrap around the address; if so, we lose. Otherwise, if the 1882 * proposed new region fits before the next entry, we win. 1883 */ 1884 1885 end = *start + length; 1886 if (end < *start) 1887 return (-1); 1888 1889 if (entry->next->start >= end && *start >= entry->end) 1890 return (1); 1891 1892 return (0); 1893 } 1894 1895 /* 1896 * uvm_map_findspace: find "length" sized space in "map". 1897 * 1898 * => "hint" is a hint about where we want it, unless UVM_FLAG_FIXED is 1899 * set in "flags" (in which case we insist on using "hint"). 1900 * => "result" is VA returned 1901 * => uobj/uoffset are to be used to handle VAC alignment, if required 1902 * => if "align" is non-zero, we attempt to align to that value. 1903 * => caller must at least have read-locked map 1904 * => returns NULL on failure, or pointer to prev. map entry if success 1905 * => note this is a cross between the old vm_map_findspace and vm_map_find 1906 */ 1907 1908 struct vm_map_entry * 1909 uvm_map_findspace(struct vm_map *map, vaddr_t hint, vsize_t length, 1910 vaddr_t *result /* OUT */, struct uvm_object *uobj, voff_t uoffset, 1911 vsize_t align, int flags) 1912 { 1913 struct vm_map_entry *entry; 1914 struct vm_map_entry *child, *prev, *tmp; 1915 vaddr_t orig_hint; 1916 const int topdown = map->flags & VM_MAP_TOPDOWN; 1917 UVMHIST_FUNC("uvm_map_findspace"); 1918 UVMHIST_CALLED(maphist); 1919 1920 UVMHIST_LOG(maphist, "(map=0x%x, hint=0x%x, len=%d, flags=0x%x)", 1921 map, hint, length, flags); 1922 KASSERT((align & (align - 1)) == 0); 1923 KASSERT((flags & UVM_FLAG_FIXED) == 0 || align == 0); 1924 1925 uvm_map_check(map, "map_findspace entry"); 1926 1927 /* 1928 * remember the original hint. if we are aligning, then we 1929 * may have to try again with no alignment constraint if 1930 * we fail the first time. 1931 */ 1932 1933 orig_hint = hint; 1934 if (hint < vm_map_min(map)) { /* check ranges ... */ 1935 if (flags & UVM_FLAG_FIXED) { 1936 UVMHIST_LOG(maphist,"<- VA below map range",0,0,0,0); 1937 return (NULL); 1938 } 1939 hint = vm_map_min(map); 1940 } 1941 if (hint > vm_map_max(map)) { 1942 UVMHIST_LOG(maphist,"<- VA 0x%x > range [0x%x->0x%x]", 1943 hint, vm_map_min(map), vm_map_max(map), 0); 1944 return (NULL); 1945 } 1946 1947 /* 1948 * Look for the first possible address; if there's already 1949 * something at this address, we have to start after it. 1950 */ 1951 1952 /* 1953 * @@@: there are four, no, eight cases to consider. 1954 * 1955 * 0: found, fixed, bottom up -> fail 1956 * 1: found, fixed, top down -> fail 1957 * 2: found, not fixed, bottom up -> start after entry->end, 1958 * loop up 1959 * 3: found, not fixed, top down -> start before entry->start, 1960 * loop down 1961 * 4: not found, fixed, bottom up -> check entry->next->start, fail 1962 * 5: not found, fixed, top down -> check entry->next->start, fail 1963 * 6: not found, not fixed, bottom up -> check entry->next->start, 1964 * loop up 1965 * 7: not found, not fixed, top down -> check entry->next->start, 1966 * loop down 1967 * 1968 * as you can see, it reduces to roughly five cases, and that 1969 * adding top down mapping only adds one unique case (without 1970 * it, there would be four cases). 1971 */ 1972 1973 if ((flags & UVM_FLAG_FIXED) == 0 && hint == vm_map_min(map)) { 1974 entry = map->first_free; 1975 } else { 1976 if (uvm_map_lookup_entry(map, hint, &entry)) { 1977 /* "hint" address already in use ... */ 1978 if (flags & UVM_FLAG_FIXED) { 1979 UVMHIST_LOG(maphist, "<- fixed & VA in use", 1980 0, 0, 0, 0); 1981 return (NULL); 1982 } 1983 if (topdown) 1984 /* Start from lower gap. */ 1985 entry = entry->prev; 1986 } else if (flags & UVM_FLAG_FIXED) { 1987 if (entry->next->start >= hint + length && 1988 hint + length > hint) 1989 goto found; 1990 1991 /* "hint" address is gap but too small */ 1992 UVMHIST_LOG(maphist, "<- fixed mapping failed", 1993 0, 0, 0, 0); 1994 return (NULL); /* only one shot at it ... */ 1995 } else { 1996 /* 1997 * See if given hint fits in this gap. 1998 */ 1999 switch (uvm_map_space_avail(&hint, length, 2000 uoffset, align, topdown, entry)) { 2001 case 1: 2002 goto found; 2003 case -1: 2004 goto wraparound; 2005 } 2006 2007 if (topdown) { 2008 /* 2009 * Still there is a chance to fit 2010 * if hint > entry->end. 2011 */ 2012 } else { 2013 /* Start from higher gap. */ 2014 entry = entry->next; 2015 if (entry == &map->header) 2016 goto notfound; 2017 goto nextgap; 2018 } 2019 } 2020 } 2021 2022 /* 2023 * Note that all UVM_FLAGS_FIXED case is already handled. 2024 */ 2025 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2026 2027 /* Try to find the space in the red-black tree */ 2028 2029 /* Check slot before any entry */ 2030 hint = topdown ? entry->next->start - length : entry->end; 2031 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2032 topdown, entry)) { 2033 case 1: 2034 goto found; 2035 case -1: 2036 goto wraparound; 2037 } 2038 2039 nextgap: 2040 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2041 /* If there is not enough space in the whole tree, we fail */ 2042 tmp = ROOT_ENTRY(map); 2043 if (tmp == NULL || tmp->maxgap < length) 2044 goto notfound; 2045 2046 prev = NULL; /* previous candidate */ 2047 2048 /* Find an entry close to hint that has enough space */ 2049 for (; tmp;) { 2050 KASSERT(tmp->next->start == tmp->end + tmp->gap); 2051 if (topdown) { 2052 if (tmp->next->start < hint + length && 2053 (prev == NULL || tmp->end > prev->end)) { 2054 if (tmp->gap >= length) 2055 prev = tmp; 2056 else if ((child = LEFT_ENTRY(tmp)) != NULL 2057 && child->maxgap >= length) 2058 prev = tmp; 2059 } 2060 } else { 2061 if (tmp->end >= hint && 2062 (prev == NULL || tmp->end < prev->end)) { 2063 if (tmp->gap >= length) 2064 prev = tmp; 2065 else if ((child = RIGHT_ENTRY(tmp)) != NULL 2066 && child->maxgap >= length) 2067 prev = tmp; 2068 } 2069 } 2070 if (tmp->next->start < hint + length) 2071 child = RIGHT_ENTRY(tmp); 2072 else if (tmp->end > hint) 2073 child = LEFT_ENTRY(tmp); 2074 else { 2075 if (tmp->gap >= length) 2076 break; 2077 if (topdown) 2078 child = LEFT_ENTRY(tmp); 2079 else 2080 child = RIGHT_ENTRY(tmp); 2081 } 2082 if (child == NULL || child->maxgap < length) 2083 break; 2084 tmp = child; 2085 } 2086 2087 if (tmp != NULL && tmp->start < hint && hint < tmp->next->start) { 2088 /* 2089 * Check if the entry that we found satifies the 2090 * space requirement 2091 */ 2092 if (topdown) { 2093 if (hint > tmp->next->start - length) 2094 hint = tmp->next->start - length; 2095 } else { 2096 if (hint < tmp->end) 2097 hint = tmp->end; 2098 } 2099 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2100 topdown, tmp)) { 2101 case 1: 2102 entry = tmp; 2103 goto found; 2104 case -1: 2105 goto wraparound; 2106 } 2107 if (tmp->gap >= length) 2108 goto listsearch; 2109 } 2110 if (prev == NULL) 2111 goto notfound; 2112 2113 if (topdown) { 2114 KASSERT(orig_hint >= prev->next->start - length || 2115 prev->next->start - length > prev->next->start); 2116 hint = prev->next->start - length; 2117 } else { 2118 KASSERT(orig_hint <= prev->end); 2119 hint = prev->end; 2120 } 2121 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2122 topdown, prev)) { 2123 case 1: 2124 entry = prev; 2125 goto found; 2126 case -1: 2127 goto wraparound; 2128 } 2129 if (prev->gap >= length) 2130 goto listsearch; 2131 2132 if (topdown) 2133 tmp = LEFT_ENTRY(prev); 2134 else 2135 tmp = RIGHT_ENTRY(prev); 2136 for (;;) { 2137 KASSERT(tmp && tmp->maxgap >= length); 2138 if (topdown) 2139 child = RIGHT_ENTRY(tmp); 2140 else 2141 child = LEFT_ENTRY(tmp); 2142 if (child && child->maxgap >= length) { 2143 tmp = child; 2144 continue; 2145 } 2146 if (tmp->gap >= length) 2147 break; 2148 if (topdown) 2149 tmp = LEFT_ENTRY(tmp); 2150 else 2151 tmp = RIGHT_ENTRY(tmp); 2152 } 2153 2154 if (topdown) { 2155 KASSERT(orig_hint >= tmp->next->start - length || 2156 tmp->next->start - length > tmp->next->start); 2157 hint = tmp->next->start - length; 2158 } else { 2159 KASSERT(orig_hint <= tmp->end); 2160 hint = tmp->end; 2161 } 2162 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2163 topdown, tmp)) { 2164 case 1: 2165 entry = tmp; 2166 goto found; 2167 case -1: 2168 goto wraparound; 2169 } 2170 2171 /* 2172 * The tree fails to find an entry because of offset or alignment 2173 * restrictions. Search the list instead. 2174 */ 2175 listsearch: 2176 /* 2177 * Look through the rest of the map, trying to fit a new region in 2178 * the gap between existing regions, or after the very last region. 2179 * note: entry->end = base VA of current gap, 2180 * entry->next->start = VA of end of current gap 2181 */ 2182 2183 for (;;) { 2184 /* Update hint for current gap. */ 2185 hint = topdown ? entry->next->start - length : entry->end; 2186 2187 /* See if it fits. */ 2188 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2189 topdown, entry)) { 2190 case 1: 2191 goto found; 2192 case -1: 2193 goto wraparound; 2194 } 2195 2196 /* Advance to next/previous gap */ 2197 if (topdown) { 2198 if (entry == &map->header) { 2199 UVMHIST_LOG(maphist, "<- failed (off start)", 2200 0,0,0,0); 2201 goto notfound; 2202 } 2203 entry = entry->prev; 2204 } else { 2205 entry = entry->next; 2206 if (entry == &map->header) { 2207 UVMHIST_LOG(maphist, "<- failed (off end)", 2208 0,0,0,0); 2209 goto notfound; 2210 } 2211 } 2212 } 2213 2214 found: 2215 SAVE_HINT(map, map->hint, entry); 2216 *result = hint; 2217 UVMHIST_LOG(maphist,"<- got it! (result=0x%x)", hint, 0,0,0); 2218 KASSERT( topdown || hint >= orig_hint); 2219 KASSERT(!topdown || hint <= orig_hint); 2220 KASSERT(entry->end <= hint); 2221 KASSERT(hint + length <= entry->next->start); 2222 return (entry); 2223 2224 wraparound: 2225 UVMHIST_LOG(maphist, "<- failed (wrap around)", 0,0,0,0); 2226 2227 return (NULL); 2228 2229 notfound: 2230 UVMHIST_LOG(maphist, "<- failed (notfound)", 0,0,0,0); 2231 2232 return (NULL); 2233 } 2234 2235 /* 2236 * U N M A P - m a i n h e l p e r f u n c t i o n s 2237 */ 2238 2239 /* 2240 * uvm_unmap_remove: remove mappings from a vm_map (from "start" up to "stop") 2241 * 2242 * => caller must check alignment and size 2243 * => map must be locked by caller 2244 * => we return a list of map entries that we've remove from the map 2245 * in "entry_list" 2246 */ 2247 2248 void 2249 uvm_unmap_remove(struct vm_map *map, vaddr_t start, vaddr_t end, 2250 struct vm_map_entry **entry_list /* OUT */, 2251 struct uvm_mapent_reservation *umr, int flags) 2252 { 2253 struct vm_map_entry *entry, *first_entry, *next; 2254 vaddr_t len; 2255 UVMHIST_FUNC("uvm_unmap_remove"); UVMHIST_CALLED(maphist); 2256 2257 UVMHIST_LOG(maphist,"(map=0x%x, start=0x%x, end=0x%x)", 2258 map, start, end, 0); 2259 VM_MAP_RANGE_CHECK(map, start, end); 2260 2261 uvm_map_check(map, "unmap_remove entry"); 2262 2263 /* 2264 * find first entry 2265 */ 2266 2267 if (uvm_map_lookup_entry(map, start, &first_entry) == true) { 2268 /* clip and go... */ 2269 entry = first_entry; 2270 UVM_MAP_CLIP_START(map, entry, start, umr); 2271 /* critical! prevents stale hint */ 2272 SAVE_HINT(map, entry, entry->prev); 2273 } else { 2274 entry = first_entry->next; 2275 } 2276 2277 /* 2278 * Save the free space hint 2279 */ 2280 2281 if (map->first_free != &map->header && map->first_free->start >= start) 2282 map->first_free = entry->prev; 2283 2284 /* 2285 * note: we now re-use first_entry for a different task. we remove 2286 * a number of map entries from the map and save them in a linked 2287 * list headed by "first_entry". once we remove them from the map 2288 * the caller should unlock the map and drop the references to the 2289 * backing objects [c.f. uvm_unmap_detach]. the object is to 2290 * separate unmapping from reference dropping. why? 2291 * [1] the map has to be locked for unmapping 2292 * [2] the map need not be locked for reference dropping 2293 * [3] dropping references may trigger pager I/O, and if we hit 2294 * a pager that does synchronous I/O we may have to wait for it. 2295 * [4] we would like all waiting for I/O to occur with maps unlocked 2296 * so that we don't block other threads. 2297 */ 2298 2299 first_entry = NULL; 2300 *entry_list = NULL; 2301 2302 /* 2303 * break up the area into map entry sized regions and unmap. note 2304 * that all mappings have to be removed before we can even consider 2305 * dropping references to amaps or VM objects (otherwise we could end 2306 * up with a mapping to a page on the free list which would be very bad) 2307 */ 2308 2309 while ((entry != &map->header) && (entry->start < end)) { 2310 KASSERT((entry->flags & UVM_MAP_FIRST) == 0); 2311 2312 UVM_MAP_CLIP_END(map, entry, end, umr); 2313 next = entry->next; 2314 len = entry->end - entry->start; 2315 2316 /* 2317 * unwire before removing addresses from the pmap; otherwise 2318 * unwiring will put the entries back into the pmap (XXX). 2319 */ 2320 2321 if (VM_MAPENT_ISWIRED(entry)) { 2322 uvm_map_entry_unwire(map, entry); 2323 } 2324 if (flags & UVM_FLAG_VAONLY) { 2325 2326 /* nothing */ 2327 2328 } else if ((map->flags & VM_MAP_PAGEABLE) == 0) { 2329 2330 /* 2331 * if the map is non-pageable, any pages mapped there 2332 * must be wired and entered with pmap_kenter_pa(), 2333 * and we should free any such pages immediately. 2334 * this is mostly used for kmem_map and mb_map. 2335 */ 2336 2337 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2338 uvm_km_pgremove_intrsafe(map, entry->start, 2339 entry->end); 2340 pmap_kremove(entry->start, len); 2341 } 2342 } else if (UVM_ET_ISOBJ(entry) && 2343 UVM_OBJ_IS_KERN_OBJECT(entry->object.uvm_obj)) { 2344 KASSERT(vm_map_pmap(map) == pmap_kernel()); 2345 2346 /* 2347 * note: kernel object mappings are currently used in 2348 * two ways: 2349 * [1] "normal" mappings of pages in the kernel object 2350 * [2] uvm_km_valloc'd allocations in which we 2351 * pmap_enter in some non-kernel-object page 2352 * (e.g. vmapbuf). 2353 * 2354 * for case [1], we need to remove the mapping from 2355 * the pmap and then remove the page from the kernel 2356 * object (because, once pages in a kernel object are 2357 * unmapped they are no longer needed, unlike, say, 2358 * a vnode where you might want the data to persist 2359 * until flushed out of a queue). 2360 * 2361 * for case [2], we need to remove the mapping from 2362 * the pmap. there shouldn't be any pages at the 2363 * specified offset in the kernel object [but it 2364 * doesn't hurt to call uvm_km_pgremove just to be 2365 * safe?] 2366 * 2367 * uvm_km_pgremove currently does the following: 2368 * for pages in the kernel object in range: 2369 * - drops the swap slot 2370 * - uvm_pagefree the page 2371 */ 2372 2373 /* 2374 * remove mappings from pmap and drop the pages 2375 * from the object. offsets are always relative 2376 * to vm_map_min(kernel_map). 2377 */ 2378 2379 pmap_remove(pmap_kernel(), entry->start, 2380 entry->start + len); 2381 uvm_km_pgremove(entry->start, entry->end); 2382 2383 /* 2384 * null out kernel_object reference, we've just 2385 * dropped it 2386 */ 2387 2388 entry->etype &= ~UVM_ET_OBJ; 2389 entry->object.uvm_obj = NULL; 2390 } else if (UVM_ET_ISOBJ(entry) || entry->aref.ar_amap) { 2391 2392 /* 2393 * remove mappings the standard way. 2394 */ 2395 2396 pmap_remove(map->pmap, entry->start, entry->end); 2397 } 2398 2399 #if defined(DEBUG) 2400 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2401 2402 /* 2403 * check if there's remaining mapping, 2404 * which is a bug in caller. 2405 */ 2406 2407 vaddr_t va; 2408 for (va = entry->start; va < entry->end; 2409 va += PAGE_SIZE) { 2410 if (pmap_extract(vm_map_pmap(map), va, NULL)) { 2411 panic("uvm_unmap_remove: has mapping"); 2412 } 2413 } 2414 2415 if (VM_MAP_IS_KERNEL(map)) { 2416 uvm_km_check_empty(map, entry->start, 2417 entry->end); 2418 } 2419 } 2420 #endif /* defined(DEBUG) */ 2421 2422 /* 2423 * remove entry from map and put it on our list of entries 2424 * that we've nuked. then go to next entry. 2425 */ 2426 2427 UVMHIST_LOG(maphist, " removed map entry 0x%x", entry, 0, 0,0); 2428 2429 /* critical! prevents stale hint */ 2430 SAVE_HINT(map, entry, entry->prev); 2431 2432 uvm_map_entry_unlink(map, entry); 2433 KASSERT(map->size >= len); 2434 map->size -= len; 2435 entry->prev = NULL; 2436 entry->next = first_entry; 2437 first_entry = entry; 2438 entry = next; 2439 } 2440 if ((map->flags & VM_MAP_DYING) == 0) { 2441 pmap_update(vm_map_pmap(map)); 2442 } 2443 2444 uvm_map_check(map, "unmap_remove leave"); 2445 2446 /* 2447 * now we've cleaned up the map and are ready for the caller to drop 2448 * references to the mapped objects. 2449 */ 2450 2451 *entry_list = first_entry; 2452 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 2453 2454 if (map->flags & VM_MAP_WANTVA) { 2455 mutex_enter(&map->misc_lock); 2456 map->flags &= ~VM_MAP_WANTVA; 2457 cv_broadcast(&map->cv); 2458 mutex_exit(&map->misc_lock); 2459 } 2460 } 2461 2462 /* 2463 * uvm_unmap_detach: drop references in a chain of map entries 2464 * 2465 * => we will free the map entries as we traverse the list. 2466 */ 2467 2468 void 2469 uvm_unmap_detach(struct vm_map_entry *first_entry, int flags) 2470 { 2471 struct vm_map_entry *next_entry; 2472 UVMHIST_FUNC("uvm_unmap_detach"); UVMHIST_CALLED(maphist); 2473 2474 while (first_entry) { 2475 KASSERT(!VM_MAPENT_ISWIRED(first_entry)); 2476 UVMHIST_LOG(maphist, 2477 " detach 0x%x: amap=0x%x, obj=0x%x, submap?=%d", 2478 first_entry, first_entry->aref.ar_amap, 2479 first_entry->object.uvm_obj, 2480 UVM_ET_ISSUBMAP(first_entry)); 2481 2482 /* 2483 * drop reference to amap, if we've got one 2484 */ 2485 2486 if (first_entry->aref.ar_amap) 2487 uvm_map_unreference_amap(first_entry, flags); 2488 2489 /* 2490 * drop reference to our backing object, if we've got one 2491 */ 2492 2493 KASSERT(!UVM_ET_ISSUBMAP(first_entry)); 2494 if (UVM_ET_ISOBJ(first_entry) && 2495 first_entry->object.uvm_obj->pgops->pgo_detach) { 2496 (*first_entry->object.uvm_obj->pgops->pgo_detach) 2497 (first_entry->object.uvm_obj); 2498 } 2499 next_entry = first_entry->next; 2500 uvm_mapent_free(first_entry); 2501 first_entry = next_entry; 2502 } 2503 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 2504 } 2505 2506 /* 2507 * E X T R A C T I O N F U N C T I O N S 2508 */ 2509 2510 /* 2511 * uvm_map_reserve: reserve space in a vm_map for future use. 2512 * 2513 * => we reserve space in a map by putting a dummy map entry in the 2514 * map (dummy means obj=NULL, amap=NULL, prot=VM_PROT_NONE) 2515 * => map should be unlocked (we will write lock it) 2516 * => we return true if we were able to reserve space 2517 * => XXXCDC: should be inline? 2518 */ 2519 2520 int 2521 uvm_map_reserve(struct vm_map *map, vsize_t size, 2522 vaddr_t offset /* hint for pmap_prefer */, 2523 vsize_t align /* alignment */, 2524 vaddr_t *raddr /* IN:hint, OUT: reserved VA */, 2525 uvm_flag_t flags /* UVM_FLAG_FIXED or 0 */) 2526 { 2527 UVMHIST_FUNC("uvm_map_reserve"); UVMHIST_CALLED(maphist); 2528 2529 UVMHIST_LOG(maphist, "(map=0x%x, size=0x%x, offset=0x%x,addr=0x%x)", 2530 map,size,offset,raddr); 2531 2532 size = round_page(size); 2533 2534 /* 2535 * reserve some virtual space. 2536 */ 2537 2538 if (uvm_map(map, raddr, size, NULL, offset, align, 2539 UVM_MAPFLAG(UVM_PROT_NONE, UVM_PROT_NONE, UVM_INH_NONE, 2540 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE|flags)) != 0) { 2541 UVMHIST_LOG(maphist, "<- done (no VM)", 0,0,0,0); 2542 return (false); 2543 } 2544 2545 UVMHIST_LOG(maphist, "<- done (*raddr=0x%x)", *raddr,0,0,0); 2546 return (true); 2547 } 2548 2549 /* 2550 * uvm_map_replace: replace a reserved (blank) area of memory with 2551 * real mappings. 2552 * 2553 * => caller must WRITE-LOCK the map 2554 * => we return true if replacement was a success 2555 * => we expect the newents chain to have nnewents entrys on it and 2556 * we expect newents->prev to point to the last entry on the list 2557 * => note newents is allowed to be NULL 2558 */ 2559 2560 static int 2561 uvm_map_replace(struct vm_map *map, vaddr_t start, vaddr_t end, 2562 struct vm_map_entry *newents, int nnewents, vsize_t nsize, 2563 struct vm_map_entry **oldentryp) 2564 { 2565 struct vm_map_entry *oldent, *last; 2566 2567 uvm_map_check(map, "map_replace entry"); 2568 2569 /* 2570 * first find the blank map entry at the specified address 2571 */ 2572 2573 if (!uvm_map_lookup_entry(map, start, &oldent)) { 2574 return (false); 2575 } 2576 2577 /* 2578 * check to make sure we have a proper blank entry 2579 */ 2580 2581 if (end < oldent->end && !VM_MAP_USE_KMAPENT(map)) { 2582 UVM_MAP_CLIP_END(map, oldent, end, NULL); 2583 } 2584 if (oldent->start != start || oldent->end != end || 2585 oldent->object.uvm_obj != NULL || oldent->aref.ar_amap != NULL) { 2586 return (false); 2587 } 2588 2589 #ifdef DIAGNOSTIC 2590 2591 /* 2592 * sanity check the newents chain 2593 */ 2594 2595 { 2596 struct vm_map_entry *tmpent = newents; 2597 int nent = 0; 2598 vsize_t sz = 0; 2599 vaddr_t cur = start; 2600 2601 while (tmpent) { 2602 nent++; 2603 sz += tmpent->end - tmpent->start; 2604 if (tmpent->start < cur) 2605 panic("uvm_map_replace1"); 2606 if (tmpent->start >= tmpent->end || tmpent->end > end) { 2607 printf("tmpent->start=0x%lx, tmpent->end=0x%lx, end=0x%lx\n", 2608 tmpent->start, tmpent->end, end); 2609 panic("uvm_map_replace2"); 2610 } 2611 cur = tmpent->end; 2612 if (tmpent->next) { 2613 if (tmpent->next->prev != tmpent) 2614 panic("uvm_map_replace3"); 2615 } else { 2616 if (newents->prev != tmpent) 2617 panic("uvm_map_replace4"); 2618 } 2619 tmpent = tmpent->next; 2620 } 2621 if (nent != nnewents) 2622 panic("uvm_map_replace5"); 2623 if (sz != nsize) 2624 panic("uvm_map_replace6"); 2625 } 2626 #endif 2627 2628 /* 2629 * map entry is a valid blank! replace it. (this does all the 2630 * work of map entry link/unlink...). 2631 */ 2632 2633 if (newents) { 2634 last = newents->prev; 2635 2636 /* critical: flush stale hints out of map */ 2637 SAVE_HINT(map, map->hint, newents); 2638 if (map->first_free == oldent) 2639 map->first_free = last; 2640 2641 last->next = oldent->next; 2642 last->next->prev = last; 2643 2644 /* Fix RB tree */ 2645 uvm_rb_remove(map, oldent); 2646 2647 newents->prev = oldent->prev; 2648 newents->prev->next = newents; 2649 map->nentries = map->nentries + (nnewents - 1); 2650 2651 /* Fixup the RB tree */ 2652 { 2653 int i; 2654 struct vm_map_entry *tmp; 2655 2656 tmp = newents; 2657 for (i = 0; i < nnewents && tmp; i++) { 2658 uvm_rb_insert(map, tmp); 2659 tmp = tmp->next; 2660 } 2661 } 2662 } else { 2663 /* NULL list of new entries: just remove the old one */ 2664 clear_hints(map, oldent); 2665 uvm_map_entry_unlink(map, oldent); 2666 } 2667 map->size -= end - start - nsize; 2668 2669 uvm_map_check(map, "map_replace leave"); 2670 2671 /* 2672 * now we can free the old blank entry and return. 2673 */ 2674 2675 *oldentryp = oldent; 2676 return (true); 2677 } 2678 2679 /* 2680 * uvm_map_extract: extract a mapping from a map and put it somewhere 2681 * (maybe removing the old mapping) 2682 * 2683 * => maps should be unlocked (we will write lock them) 2684 * => returns 0 on success, error code otherwise 2685 * => start must be page aligned 2686 * => len must be page sized 2687 * => flags: 2688 * UVM_EXTRACT_REMOVE: remove mappings from srcmap 2689 * UVM_EXTRACT_CONTIG: abort if unmapped area (advisory only) 2690 * UVM_EXTRACT_QREF: for a temporary extraction do quick obj refs 2691 * UVM_EXTRACT_FIXPROT: set prot to maxprot as we go 2692 * >>>NOTE: if you set REMOVE, you are not allowed to use CONTIG or QREF!<<< 2693 * >>>NOTE: QREF's must be unmapped via the QREF path, thus should only 2694 * be used from within the kernel in a kernel level map <<< 2695 */ 2696 2697 int 2698 uvm_map_extract(struct vm_map *srcmap, vaddr_t start, vsize_t len, 2699 struct vm_map *dstmap, vaddr_t *dstaddrp, int flags) 2700 { 2701 vaddr_t dstaddr, end, newend, oldoffset, fudge, orig_fudge; 2702 struct vm_map_entry *chain, *endchain, *entry, *orig_entry, *newentry, 2703 *deadentry, *oldentry; 2704 struct vm_map_entry *resentry = NULL; /* a dummy reservation entry */ 2705 vsize_t elen; 2706 int nchain, error, copy_ok; 2707 vsize_t nsize; 2708 UVMHIST_FUNC("uvm_map_extract"); UVMHIST_CALLED(maphist); 2709 2710 UVMHIST_LOG(maphist,"(srcmap=0x%x,start=0x%x, len=0x%x", srcmap, start, 2711 len,0); 2712 UVMHIST_LOG(maphist," ...,dstmap=0x%x, flags=0x%x)", dstmap,flags,0,0); 2713 2714 /* 2715 * step 0: sanity check: start must be on a page boundary, length 2716 * must be page sized. can't ask for CONTIG/QREF if you asked for 2717 * REMOVE. 2718 */ 2719 2720 KASSERT((start & PAGE_MASK) == 0 && (len & PAGE_MASK) == 0); 2721 KASSERT((flags & UVM_EXTRACT_REMOVE) == 0 || 2722 (flags & (UVM_EXTRACT_CONTIG|UVM_EXTRACT_QREF)) == 0); 2723 2724 /* 2725 * step 1: reserve space in the target map for the extracted area 2726 */ 2727 2728 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 2729 dstaddr = vm_map_min(dstmap); 2730 if (!uvm_map_reserve(dstmap, len, start, 0, &dstaddr, 0)) 2731 return (ENOMEM); 2732 *dstaddrp = dstaddr; /* pass address back to caller */ 2733 UVMHIST_LOG(maphist, " dstaddr=0x%x", dstaddr,0,0,0); 2734 } else { 2735 dstaddr = *dstaddrp; 2736 } 2737 2738 /* 2739 * step 2: setup for the extraction process loop by init'ing the 2740 * map entry chain, locking src map, and looking up the first useful 2741 * entry in the map. 2742 */ 2743 2744 end = start + len; 2745 newend = dstaddr + len; 2746 chain = endchain = NULL; 2747 nchain = 0; 2748 nsize = 0; 2749 vm_map_lock(srcmap); 2750 2751 if (uvm_map_lookup_entry(srcmap, start, &entry)) { 2752 2753 /* "start" is within an entry */ 2754 if (flags & UVM_EXTRACT_QREF) { 2755 2756 /* 2757 * for quick references we don't clip the entry, so 2758 * the entry may map space "before" the starting 2759 * virtual address... this is the "fudge" factor 2760 * (which can be non-zero only the first time 2761 * through the "while" loop in step 3). 2762 */ 2763 2764 fudge = start - entry->start; 2765 } else { 2766 2767 /* 2768 * normal reference: we clip the map to fit (thus 2769 * fudge is zero) 2770 */ 2771 2772 UVM_MAP_CLIP_START(srcmap, entry, start, NULL); 2773 SAVE_HINT(srcmap, srcmap->hint, entry->prev); 2774 fudge = 0; 2775 } 2776 } else { 2777 2778 /* "start" is not within an entry ... skip to next entry */ 2779 if (flags & UVM_EXTRACT_CONTIG) { 2780 error = EINVAL; 2781 goto bad; /* definite hole here ... */ 2782 } 2783 2784 entry = entry->next; 2785 fudge = 0; 2786 } 2787 2788 /* save values from srcmap for step 6 */ 2789 orig_entry = entry; 2790 orig_fudge = fudge; 2791 2792 /* 2793 * step 3: now start looping through the map entries, extracting 2794 * as we go. 2795 */ 2796 2797 while (entry->start < end && entry != &srcmap->header) { 2798 2799 /* if we are not doing a quick reference, clip it */ 2800 if ((flags & UVM_EXTRACT_QREF) == 0) 2801 UVM_MAP_CLIP_END(srcmap, entry, end, NULL); 2802 2803 /* clear needs_copy (allow chunking) */ 2804 if (UVM_ET_ISNEEDSCOPY(entry)) { 2805 amap_copy(srcmap, entry, 2806 AMAP_COPY_NOWAIT|AMAP_COPY_NOMERGE, start, end); 2807 if (UVM_ET_ISNEEDSCOPY(entry)) { /* failed? */ 2808 error = ENOMEM; 2809 goto bad; 2810 } 2811 2812 /* amap_copy could clip (during chunk)! update fudge */ 2813 if (fudge) { 2814 fudge = start - entry->start; 2815 orig_fudge = fudge; 2816 } 2817 } 2818 2819 /* calculate the offset of this from "start" */ 2820 oldoffset = (entry->start + fudge) - start; 2821 2822 /* allocate a new map entry */ 2823 newentry = uvm_mapent_alloc(dstmap, 0); 2824 if (newentry == NULL) { 2825 error = ENOMEM; 2826 goto bad; 2827 } 2828 2829 /* set up new map entry */ 2830 newentry->next = NULL; 2831 newentry->prev = endchain; 2832 newentry->start = dstaddr + oldoffset; 2833 newentry->end = 2834 newentry->start + (entry->end - (entry->start + fudge)); 2835 if (newentry->end > newend || newentry->end < newentry->start) 2836 newentry->end = newend; 2837 newentry->object.uvm_obj = entry->object.uvm_obj; 2838 if (newentry->object.uvm_obj) { 2839 if (newentry->object.uvm_obj->pgops->pgo_reference) 2840 newentry->object.uvm_obj->pgops-> 2841 pgo_reference(newentry->object.uvm_obj); 2842 newentry->offset = entry->offset + fudge; 2843 } else { 2844 newentry->offset = 0; 2845 } 2846 newentry->etype = entry->etype; 2847 newentry->protection = (flags & UVM_EXTRACT_FIXPROT) ? 2848 entry->max_protection : entry->protection; 2849 newentry->max_protection = entry->max_protection; 2850 newentry->inheritance = entry->inheritance; 2851 newentry->wired_count = 0; 2852 newentry->aref.ar_amap = entry->aref.ar_amap; 2853 if (newentry->aref.ar_amap) { 2854 newentry->aref.ar_pageoff = 2855 entry->aref.ar_pageoff + (fudge >> PAGE_SHIFT); 2856 uvm_map_reference_amap(newentry, AMAP_SHARED | 2857 ((flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0)); 2858 } else { 2859 newentry->aref.ar_pageoff = 0; 2860 } 2861 newentry->advice = entry->advice; 2862 if ((flags & UVM_EXTRACT_QREF) != 0) { 2863 newentry->flags |= UVM_MAP_NOMERGE; 2864 } 2865 2866 /* now link it on the chain */ 2867 nchain++; 2868 nsize += newentry->end - newentry->start; 2869 if (endchain == NULL) { 2870 chain = endchain = newentry; 2871 } else { 2872 endchain->next = newentry; 2873 endchain = newentry; 2874 } 2875 2876 /* end of 'while' loop! */ 2877 if ((flags & UVM_EXTRACT_CONTIG) && entry->end < end && 2878 (entry->next == &srcmap->header || 2879 entry->next->start != entry->end)) { 2880 error = EINVAL; 2881 goto bad; 2882 } 2883 entry = entry->next; 2884 fudge = 0; 2885 } 2886 2887 /* 2888 * step 4: close off chain (in format expected by uvm_map_replace) 2889 */ 2890 2891 if (chain) 2892 chain->prev = endchain; 2893 2894 /* 2895 * step 5: attempt to lock the dest map so we can pmap_copy. 2896 * note usage of copy_ok: 2897 * 1 => dstmap locked, pmap_copy ok, and we "replace" here (step 5) 2898 * 0 => dstmap unlocked, NO pmap_copy, and we will "replace" in step 7 2899 */ 2900 2901 if (srcmap == dstmap || vm_map_lock_try(dstmap) == true) { 2902 copy_ok = 1; 2903 if (!uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2904 nchain, nsize, &resentry)) { 2905 if (srcmap != dstmap) 2906 vm_map_unlock(dstmap); 2907 error = EIO; 2908 goto bad; 2909 } 2910 } else { 2911 copy_ok = 0; 2912 /* replace defered until step 7 */ 2913 } 2914 2915 /* 2916 * step 6: traverse the srcmap a second time to do the following: 2917 * - if we got a lock on the dstmap do pmap_copy 2918 * - if UVM_EXTRACT_REMOVE remove the entries 2919 * we make use of orig_entry and orig_fudge (saved in step 2) 2920 */ 2921 2922 if (copy_ok || (flags & UVM_EXTRACT_REMOVE)) { 2923 2924 /* purge possible stale hints from srcmap */ 2925 if (flags & UVM_EXTRACT_REMOVE) { 2926 SAVE_HINT(srcmap, srcmap->hint, orig_entry->prev); 2927 if (srcmap->first_free != &srcmap->header && 2928 srcmap->first_free->start >= start) 2929 srcmap->first_free = orig_entry->prev; 2930 } 2931 2932 entry = orig_entry; 2933 fudge = orig_fudge; 2934 deadentry = NULL; /* for UVM_EXTRACT_REMOVE */ 2935 2936 while (entry->start < end && entry != &srcmap->header) { 2937 if (copy_ok) { 2938 oldoffset = (entry->start + fudge) - start; 2939 elen = MIN(end, entry->end) - 2940 (entry->start + fudge); 2941 pmap_copy(dstmap->pmap, srcmap->pmap, 2942 dstaddr + oldoffset, elen, 2943 entry->start + fudge); 2944 } 2945 2946 /* we advance "entry" in the following if statement */ 2947 if (flags & UVM_EXTRACT_REMOVE) { 2948 pmap_remove(srcmap->pmap, entry->start, 2949 entry->end); 2950 oldentry = entry; /* save entry */ 2951 entry = entry->next; /* advance */ 2952 uvm_map_entry_unlink(srcmap, oldentry); 2953 /* add to dead list */ 2954 oldentry->next = deadentry; 2955 deadentry = oldentry; 2956 } else { 2957 entry = entry->next; /* advance */ 2958 } 2959 2960 /* end of 'while' loop */ 2961 fudge = 0; 2962 } 2963 pmap_update(srcmap->pmap); 2964 2965 /* 2966 * unlock dstmap. we will dispose of deadentry in 2967 * step 7 if needed 2968 */ 2969 2970 if (copy_ok && srcmap != dstmap) 2971 vm_map_unlock(dstmap); 2972 2973 } else { 2974 deadentry = NULL; 2975 } 2976 2977 /* 2978 * step 7: we are done with the source map, unlock. if copy_ok 2979 * is 0 then we have not replaced the dummy mapping in dstmap yet 2980 * and we need to do so now. 2981 */ 2982 2983 vm_map_unlock(srcmap); 2984 if ((flags & UVM_EXTRACT_REMOVE) && deadentry) 2985 uvm_unmap_detach(deadentry, 0); /* dispose of old entries */ 2986 2987 /* now do the replacement if we didn't do it in step 5 */ 2988 if (copy_ok == 0) { 2989 vm_map_lock(dstmap); 2990 error = uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2991 nchain, nsize, &resentry); 2992 vm_map_unlock(dstmap); 2993 2994 if (error == false) { 2995 error = EIO; 2996 goto bad2; 2997 } 2998 } 2999 3000 if (resentry != NULL) 3001 uvm_mapent_free(resentry); 3002 3003 return (0); 3004 3005 /* 3006 * bad: failure recovery 3007 */ 3008 bad: 3009 vm_map_unlock(srcmap); 3010 bad2: /* src already unlocked */ 3011 if (chain) 3012 uvm_unmap_detach(chain, 3013 (flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0); 3014 3015 if (resentry != NULL) 3016 uvm_mapent_free(resentry); 3017 3018 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 3019 uvm_unmap(dstmap, dstaddr, dstaddr+len); /* ??? */ 3020 } 3021 return (error); 3022 } 3023 3024 /* end of extraction functions */ 3025 3026 /* 3027 * uvm_map_submap: punch down part of a map into a submap 3028 * 3029 * => only the kernel_map is allowed to be submapped 3030 * => the purpose of submapping is to break up the locking granularity 3031 * of a larger map 3032 * => the range specified must have been mapped previously with a uvm_map() 3033 * call [with uobj==NULL] to create a blank map entry in the main map. 3034 * [And it had better still be blank!] 3035 * => maps which contain submaps should never be copied or forked. 3036 * => to remove a submap, use uvm_unmap() on the main map 3037 * and then uvm_map_deallocate() the submap. 3038 * => main map must be unlocked. 3039 * => submap must have been init'd and have a zero reference count. 3040 * [need not be locked as we don't actually reference it] 3041 */ 3042 3043 int 3044 uvm_map_submap(struct vm_map *map, vaddr_t start, vaddr_t end, 3045 struct vm_map *submap) 3046 { 3047 struct vm_map_entry *entry; 3048 struct uvm_mapent_reservation umr; 3049 int error; 3050 3051 uvm_mapent_reserve(map, &umr, 2, 0); 3052 3053 vm_map_lock(map); 3054 VM_MAP_RANGE_CHECK(map, start, end); 3055 3056 if (uvm_map_lookup_entry(map, start, &entry)) { 3057 UVM_MAP_CLIP_START(map, entry, start, &umr); 3058 UVM_MAP_CLIP_END(map, entry, end, &umr); /* to be safe */ 3059 } else { 3060 entry = NULL; 3061 } 3062 3063 if (entry != NULL && 3064 entry->start == start && entry->end == end && 3065 entry->object.uvm_obj == NULL && entry->aref.ar_amap == NULL && 3066 !UVM_ET_ISCOPYONWRITE(entry) && !UVM_ET_ISNEEDSCOPY(entry)) { 3067 entry->etype |= UVM_ET_SUBMAP; 3068 entry->object.sub_map = submap; 3069 entry->offset = 0; 3070 uvm_map_reference(submap); 3071 error = 0; 3072 } else { 3073 error = EINVAL; 3074 } 3075 vm_map_unlock(map); 3076 3077 uvm_mapent_unreserve(map, &umr); 3078 3079 return error; 3080 } 3081 3082 /* 3083 * uvm_map_setup_kernel: init in-kernel map 3084 * 3085 * => map must not be in service yet. 3086 */ 3087 3088 void 3089 uvm_map_setup_kernel(struct vm_map_kernel *map, 3090 vaddr_t vmin, vaddr_t vmax, int flags) 3091 { 3092 3093 uvm_map_setup(&map->vmk_map, vmin, vmax, flags); 3094 callback_head_init(&map->vmk_reclaim_callback, IPL_VM); 3095 LIST_INIT(&map->vmk_kentry_free); 3096 map->vmk_merged_entries = NULL; 3097 } 3098 3099 3100 /* 3101 * uvm_map_protect: change map protection 3102 * 3103 * => set_max means set max_protection. 3104 * => map must be unlocked. 3105 */ 3106 3107 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 3108 ~VM_PROT_WRITE : VM_PROT_ALL) 3109 3110 int 3111 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end, 3112 vm_prot_t new_prot, bool set_max) 3113 { 3114 struct vm_map_entry *current, *entry; 3115 int error = 0; 3116 UVMHIST_FUNC("uvm_map_protect"); UVMHIST_CALLED(maphist); 3117 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_prot=0x%x)", 3118 map, start, end, new_prot); 3119 3120 vm_map_lock(map); 3121 VM_MAP_RANGE_CHECK(map, start, end); 3122 if (uvm_map_lookup_entry(map, start, &entry)) { 3123 UVM_MAP_CLIP_START(map, entry, start, NULL); 3124 } else { 3125 entry = entry->next; 3126 } 3127 3128 /* 3129 * make a first pass to check for protection violations. 3130 */ 3131 3132 current = entry; 3133 while ((current != &map->header) && (current->start < end)) { 3134 if (UVM_ET_ISSUBMAP(current)) { 3135 error = EINVAL; 3136 goto out; 3137 } 3138 if ((new_prot & current->max_protection) != new_prot) { 3139 error = EACCES; 3140 goto out; 3141 } 3142 /* 3143 * Don't allow VM_PROT_EXECUTE to be set on entries that 3144 * point to vnodes that are associated with a NOEXEC file 3145 * system. 3146 */ 3147 if (UVM_ET_ISOBJ(current) && 3148 UVM_OBJ_IS_VNODE(current->object.uvm_obj)) { 3149 struct vnode *vp = 3150 (struct vnode *) current->object.uvm_obj; 3151 3152 if ((new_prot & VM_PROT_EXECUTE) != 0 && 3153 (vp->v_mount->mnt_flag & MNT_NOEXEC) != 0) { 3154 error = EACCES; 3155 goto out; 3156 } 3157 } 3158 3159 current = current->next; 3160 } 3161 3162 /* go back and fix up protections (no need to clip this time). */ 3163 3164 current = entry; 3165 while ((current != &map->header) && (current->start < end)) { 3166 vm_prot_t old_prot; 3167 3168 UVM_MAP_CLIP_END(map, current, end, NULL); 3169 old_prot = current->protection; 3170 if (set_max) 3171 current->protection = 3172 (current->max_protection = new_prot) & old_prot; 3173 else 3174 current->protection = new_prot; 3175 3176 /* 3177 * update physical map if necessary. worry about copy-on-write 3178 * here -- CHECK THIS XXX 3179 */ 3180 3181 if (current->protection != old_prot) { 3182 /* update pmap! */ 3183 pmap_protect(map->pmap, current->start, current->end, 3184 current->protection & MASK(entry)); 3185 3186 /* 3187 * If this entry points at a vnode, and the 3188 * protection includes VM_PROT_EXECUTE, mark 3189 * the vnode as VEXECMAP. 3190 */ 3191 if (UVM_ET_ISOBJ(current)) { 3192 struct uvm_object *uobj = 3193 current->object.uvm_obj; 3194 3195 if (UVM_OBJ_IS_VNODE(uobj) && 3196 (current->protection & VM_PROT_EXECUTE)) { 3197 vn_markexec((struct vnode *) uobj); 3198 } 3199 } 3200 } 3201 3202 /* 3203 * If the map is configured to lock any future mappings, 3204 * wire this entry now if the old protection was VM_PROT_NONE 3205 * and the new protection is not VM_PROT_NONE. 3206 */ 3207 3208 if ((map->flags & VM_MAP_WIREFUTURE) != 0 && 3209 VM_MAPENT_ISWIRED(entry) == 0 && 3210 old_prot == VM_PROT_NONE && 3211 new_prot != VM_PROT_NONE) { 3212 if (uvm_map_pageable(map, entry->start, 3213 entry->end, false, 3214 UVM_LK_ENTER|UVM_LK_EXIT) != 0) { 3215 3216 /* 3217 * If locking the entry fails, remember the 3218 * error if it's the first one. Note we 3219 * still continue setting the protection in 3220 * the map, but will return the error 3221 * condition regardless. 3222 * 3223 * XXX Ignore what the actual error is, 3224 * XXX just call it a resource shortage 3225 * XXX so that it doesn't get confused 3226 * XXX what uvm_map_protect() itself would 3227 * XXX normally return. 3228 */ 3229 3230 error = ENOMEM; 3231 } 3232 } 3233 current = current->next; 3234 } 3235 pmap_update(map->pmap); 3236 3237 out: 3238 vm_map_unlock(map); 3239 3240 UVMHIST_LOG(maphist, "<- done, error=%d",error,0,0,0); 3241 return error; 3242 } 3243 3244 #undef MASK 3245 3246 /* 3247 * uvm_map_inherit: set inheritance code for range of addrs in map. 3248 * 3249 * => map must be unlocked 3250 * => note that the inherit code is used during a "fork". see fork 3251 * code for details. 3252 */ 3253 3254 int 3255 uvm_map_inherit(struct vm_map *map, vaddr_t start, vaddr_t end, 3256 vm_inherit_t new_inheritance) 3257 { 3258 struct vm_map_entry *entry, *temp_entry; 3259 UVMHIST_FUNC("uvm_map_inherit"); UVMHIST_CALLED(maphist); 3260 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_inh=0x%x)", 3261 map, start, end, new_inheritance); 3262 3263 switch (new_inheritance) { 3264 case MAP_INHERIT_NONE: 3265 case MAP_INHERIT_COPY: 3266 case MAP_INHERIT_SHARE: 3267 break; 3268 default: 3269 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3270 return EINVAL; 3271 } 3272 3273 vm_map_lock(map); 3274 VM_MAP_RANGE_CHECK(map, start, end); 3275 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3276 entry = temp_entry; 3277 UVM_MAP_CLIP_START(map, entry, start, NULL); 3278 } else { 3279 entry = temp_entry->next; 3280 } 3281 while ((entry != &map->header) && (entry->start < end)) { 3282 UVM_MAP_CLIP_END(map, entry, end, NULL); 3283 entry->inheritance = new_inheritance; 3284 entry = entry->next; 3285 } 3286 vm_map_unlock(map); 3287 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3288 return 0; 3289 } 3290 3291 /* 3292 * uvm_map_advice: set advice code for range of addrs in map. 3293 * 3294 * => map must be unlocked 3295 */ 3296 3297 int 3298 uvm_map_advice(struct vm_map *map, vaddr_t start, vaddr_t end, int new_advice) 3299 { 3300 struct vm_map_entry *entry, *temp_entry; 3301 UVMHIST_FUNC("uvm_map_advice"); UVMHIST_CALLED(maphist); 3302 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_adv=0x%x)", 3303 map, start, end, new_advice); 3304 3305 vm_map_lock(map); 3306 VM_MAP_RANGE_CHECK(map, start, end); 3307 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3308 entry = temp_entry; 3309 UVM_MAP_CLIP_START(map, entry, start, NULL); 3310 } else { 3311 entry = temp_entry->next; 3312 } 3313 3314 /* 3315 * XXXJRT: disallow holes? 3316 */ 3317 3318 while ((entry != &map->header) && (entry->start < end)) { 3319 UVM_MAP_CLIP_END(map, entry, end, NULL); 3320 3321 switch (new_advice) { 3322 case MADV_NORMAL: 3323 case MADV_RANDOM: 3324 case MADV_SEQUENTIAL: 3325 /* nothing special here */ 3326 break; 3327 3328 default: 3329 vm_map_unlock(map); 3330 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3331 return EINVAL; 3332 } 3333 entry->advice = new_advice; 3334 entry = entry->next; 3335 } 3336 3337 vm_map_unlock(map); 3338 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3339 return 0; 3340 } 3341 3342 /* 3343 * uvm_map_willneed: apply MADV_WILLNEED 3344 */ 3345 3346 int 3347 uvm_map_willneed(struct vm_map *map, vaddr_t start, vaddr_t end) 3348 { 3349 struct vm_map_entry *entry; 3350 UVMHIST_FUNC("uvm_map_willneed"); UVMHIST_CALLED(maphist); 3351 UVMHIST_LOG(maphist,"(map=0x%lx,start=0x%lx,end=0x%lx)", 3352 map, start, end, 0); 3353 3354 vm_map_lock_read(map); 3355 VM_MAP_RANGE_CHECK(map, start, end); 3356 if (!uvm_map_lookup_entry(map, start, &entry)) { 3357 entry = entry->next; 3358 } 3359 while (entry->start < end) { 3360 struct vm_amap * const amap = entry->aref.ar_amap; 3361 struct uvm_object * const uobj = entry->object.uvm_obj; 3362 3363 KASSERT(entry != &map->header); 3364 KASSERT(start < entry->end); 3365 /* 3366 * XXX IMPLEMENT ME. 3367 * Should invent a "weak" mode for uvm_fault() 3368 * which would only do the PGO_LOCKED pgo_get(). 3369 * 3370 * for now, we handle only the easy but common case. 3371 */ 3372 if (UVM_ET_ISOBJ(entry) && amap == NULL && uobj != NULL) { 3373 off_t offset; 3374 off_t size; 3375 3376 offset = entry->offset; 3377 if (start < entry->start) { 3378 offset += entry->start - start; 3379 } 3380 size = entry->offset + (entry->end - entry->start); 3381 if (entry->end < end) { 3382 size -= end - entry->end; 3383 } 3384 uvm_readahead(uobj, offset, size); 3385 } 3386 entry = entry->next; 3387 } 3388 vm_map_unlock_read(map); 3389 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3390 return 0; 3391 } 3392 3393 /* 3394 * uvm_map_pageable: sets the pageability of a range in a map. 3395 * 3396 * => wires map entries. should not be used for transient page locking. 3397 * for that, use uvm_fault_wire()/uvm_fault_unwire() (see uvm_vslock()). 3398 * => regions specified as not pageable require lock-down (wired) memory 3399 * and page tables. 3400 * => map must never be read-locked 3401 * => if islocked is true, map is already write-locked 3402 * => we always unlock the map, since we must downgrade to a read-lock 3403 * to call uvm_fault_wire() 3404 * => XXXCDC: check this and try and clean it up. 3405 */ 3406 3407 int 3408 uvm_map_pageable(struct vm_map *map, vaddr_t start, vaddr_t end, 3409 bool new_pageable, int lockflags) 3410 { 3411 struct vm_map_entry *entry, *start_entry, *failed_entry; 3412 int rv; 3413 #ifdef DIAGNOSTIC 3414 u_int timestamp_save; 3415 #endif 3416 UVMHIST_FUNC("uvm_map_pageable"); UVMHIST_CALLED(maphist); 3417 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_pageable=0x%x)", 3418 map, start, end, new_pageable); 3419 KASSERT(map->flags & VM_MAP_PAGEABLE); 3420 3421 if ((lockflags & UVM_LK_ENTER) == 0) 3422 vm_map_lock(map); 3423 VM_MAP_RANGE_CHECK(map, start, end); 3424 3425 /* 3426 * only one pageability change may take place at one time, since 3427 * uvm_fault_wire assumes it will be called only once for each 3428 * wiring/unwiring. therefore, we have to make sure we're actually 3429 * changing the pageability for the entire region. we do so before 3430 * making any changes. 3431 */ 3432 3433 if (uvm_map_lookup_entry(map, start, &start_entry) == false) { 3434 if ((lockflags & UVM_LK_EXIT) == 0) 3435 vm_map_unlock(map); 3436 3437 UVMHIST_LOG(maphist,"<- done (fault)",0,0,0,0); 3438 return EFAULT; 3439 } 3440 entry = start_entry; 3441 3442 /* 3443 * handle wiring and unwiring separately. 3444 */ 3445 3446 if (new_pageable) { /* unwire */ 3447 UVM_MAP_CLIP_START(map, entry, start, NULL); 3448 3449 /* 3450 * unwiring. first ensure that the range to be unwired is 3451 * really wired down and that there are no holes. 3452 */ 3453 3454 while ((entry != &map->header) && (entry->start < end)) { 3455 if (entry->wired_count == 0 || 3456 (entry->end < end && 3457 (entry->next == &map->header || 3458 entry->next->start > entry->end))) { 3459 if ((lockflags & UVM_LK_EXIT) == 0) 3460 vm_map_unlock(map); 3461 UVMHIST_LOG(maphist, "<- done (INVAL)",0,0,0,0); 3462 return EINVAL; 3463 } 3464 entry = entry->next; 3465 } 3466 3467 /* 3468 * POSIX 1003.1b - a single munlock call unlocks a region, 3469 * regardless of the number of mlock calls made on that 3470 * region. 3471 */ 3472 3473 entry = start_entry; 3474 while ((entry != &map->header) && (entry->start < end)) { 3475 UVM_MAP_CLIP_END(map, entry, end, NULL); 3476 if (VM_MAPENT_ISWIRED(entry)) 3477 uvm_map_entry_unwire(map, entry); 3478 entry = entry->next; 3479 } 3480 if ((lockflags & UVM_LK_EXIT) == 0) 3481 vm_map_unlock(map); 3482 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3483 return 0; 3484 } 3485 3486 /* 3487 * wire case: in two passes [XXXCDC: ugly block of code here] 3488 * 3489 * 1: holding the write lock, we create any anonymous maps that need 3490 * to be created. then we clip each map entry to the region to 3491 * be wired and increment its wiring count. 3492 * 3493 * 2: we downgrade to a read lock, and call uvm_fault_wire to fault 3494 * in the pages for any newly wired area (wired_count == 1). 3495 * 3496 * downgrading to a read lock for uvm_fault_wire avoids a possible 3497 * deadlock with another thread that may have faulted on one of 3498 * the pages to be wired (it would mark the page busy, blocking 3499 * us, then in turn block on the map lock that we hold). because 3500 * of problems in the recursive lock package, we cannot upgrade 3501 * to a write lock in vm_map_lookup. thus, any actions that 3502 * require the write lock must be done beforehand. because we 3503 * keep the read lock on the map, the copy-on-write status of the 3504 * entries we modify here cannot change. 3505 */ 3506 3507 while ((entry != &map->header) && (entry->start < end)) { 3508 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3509 3510 /* 3511 * perform actions of vm_map_lookup that need the 3512 * write lock on the map: create an anonymous map 3513 * for a copy-on-write region, or an anonymous map 3514 * for a zero-fill region. (XXXCDC: submap case 3515 * ok?) 3516 */ 3517 3518 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3519 if (UVM_ET_ISNEEDSCOPY(entry) && 3520 ((entry->max_protection & VM_PROT_WRITE) || 3521 (entry->object.uvm_obj == NULL))) { 3522 amap_copy(map, entry, 0, start, end); 3523 /* XXXCDC: wait OK? */ 3524 } 3525 } 3526 } 3527 UVM_MAP_CLIP_START(map, entry, start, NULL); 3528 UVM_MAP_CLIP_END(map, entry, end, NULL); 3529 entry->wired_count++; 3530 3531 /* 3532 * Check for holes 3533 */ 3534 3535 if (entry->protection == VM_PROT_NONE || 3536 (entry->end < end && 3537 (entry->next == &map->header || 3538 entry->next->start > entry->end))) { 3539 3540 /* 3541 * found one. amap creation actions do not need to 3542 * be undone, but the wired counts need to be restored. 3543 */ 3544 3545 while (entry != &map->header && entry->end > start) { 3546 entry->wired_count--; 3547 entry = entry->prev; 3548 } 3549 if ((lockflags & UVM_LK_EXIT) == 0) 3550 vm_map_unlock(map); 3551 UVMHIST_LOG(maphist,"<- done (INVALID WIRE)",0,0,0,0); 3552 return EINVAL; 3553 } 3554 entry = entry->next; 3555 } 3556 3557 /* 3558 * Pass 2. 3559 */ 3560 3561 #ifdef DIAGNOSTIC 3562 timestamp_save = map->timestamp; 3563 #endif 3564 vm_map_busy(map); 3565 vm_map_unlock(map); 3566 3567 rv = 0; 3568 entry = start_entry; 3569 while (entry != &map->header && entry->start < end) { 3570 if (entry->wired_count == 1) { 3571 rv = uvm_fault_wire(map, entry->start, entry->end, 3572 entry->max_protection, 1); 3573 if (rv) { 3574 3575 /* 3576 * wiring failed. break out of the loop. 3577 * we'll clean up the map below, once we 3578 * have a write lock again. 3579 */ 3580 3581 break; 3582 } 3583 } 3584 entry = entry->next; 3585 } 3586 3587 if (rv) { /* failed? */ 3588 3589 /* 3590 * Get back to an exclusive (write) lock. 3591 */ 3592 3593 vm_map_lock(map); 3594 vm_map_unbusy(map); 3595 3596 #ifdef DIAGNOSTIC 3597 if (timestamp_save + 1 != map->timestamp) 3598 panic("uvm_map_pageable: stale map"); 3599 #endif 3600 3601 /* 3602 * first drop the wiring count on all the entries 3603 * which haven't actually been wired yet. 3604 */ 3605 3606 failed_entry = entry; 3607 while (entry != &map->header && entry->start < end) { 3608 entry->wired_count--; 3609 entry = entry->next; 3610 } 3611 3612 /* 3613 * now, unwire all the entries that were successfully 3614 * wired above. 3615 */ 3616 3617 entry = start_entry; 3618 while (entry != failed_entry) { 3619 entry->wired_count--; 3620 if (VM_MAPENT_ISWIRED(entry) == 0) 3621 uvm_map_entry_unwire(map, entry); 3622 entry = entry->next; 3623 } 3624 if ((lockflags & UVM_LK_EXIT) == 0) 3625 vm_map_unlock(map); 3626 UVMHIST_LOG(maphist, "<- done (RV=%d)", rv,0,0,0); 3627 return (rv); 3628 } 3629 3630 if ((lockflags & UVM_LK_EXIT) == 0) { 3631 vm_map_unbusy(map); 3632 } else { 3633 3634 /* 3635 * Get back to an exclusive (write) lock. 3636 */ 3637 3638 vm_map_lock(map); 3639 vm_map_unbusy(map); 3640 } 3641 3642 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3643 return 0; 3644 } 3645 3646 /* 3647 * uvm_map_pageable_all: special case of uvm_map_pageable - affects 3648 * all mapped regions. 3649 * 3650 * => map must not be locked. 3651 * => if no flags are specified, all regions are unwired. 3652 * => XXXJRT: has some of the same problems as uvm_map_pageable() above. 3653 */ 3654 3655 int 3656 uvm_map_pageable_all(struct vm_map *map, int flags, vsize_t limit) 3657 { 3658 struct vm_map_entry *entry, *failed_entry; 3659 vsize_t size; 3660 int rv; 3661 #ifdef DIAGNOSTIC 3662 u_int timestamp_save; 3663 #endif 3664 UVMHIST_FUNC("uvm_map_pageable_all"); UVMHIST_CALLED(maphist); 3665 UVMHIST_LOG(maphist,"(map=0x%x,flags=0x%x)", map, flags, 0, 0); 3666 3667 KASSERT(map->flags & VM_MAP_PAGEABLE); 3668 3669 vm_map_lock(map); 3670 3671 /* 3672 * handle wiring and unwiring separately. 3673 */ 3674 3675 if (flags == 0) { /* unwire */ 3676 3677 /* 3678 * POSIX 1003.1b -- munlockall unlocks all regions, 3679 * regardless of how many times mlockall has been called. 3680 */ 3681 3682 for (entry = map->header.next; entry != &map->header; 3683 entry = entry->next) { 3684 if (VM_MAPENT_ISWIRED(entry)) 3685 uvm_map_entry_unwire(map, entry); 3686 } 3687 map->flags &= ~VM_MAP_WIREFUTURE; 3688 vm_map_unlock(map); 3689 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3690 return 0; 3691 } 3692 3693 if (flags & MCL_FUTURE) { 3694 3695 /* 3696 * must wire all future mappings; remember this. 3697 */ 3698 3699 map->flags |= VM_MAP_WIREFUTURE; 3700 } 3701 3702 if ((flags & MCL_CURRENT) == 0) { 3703 3704 /* 3705 * no more work to do! 3706 */ 3707 3708 UVMHIST_LOG(maphist,"<- done (OK no wire)",0,0,0,0); 3709 vm_map_unlock(map); 3710 return 0; 3711 } 3712 3713 /* 3714 * wire case: in three passes [XXXCDC: ugly block of code here] 3715 * 3716 * 1: holding the write lock, count all pages mapped by non-wired 3717 * entries. if this would cause us to go over our limit, we fail. 3718 * 3719 * 2: still holding the write lock, we create any anonymous maps that 3720 * need to be created. then we increment its wiring count. 3721 * 3722 * 3: we downgrade to a read lock, and call uvm_fault_wire to fault 3723 * in the pages for any newly wired area (wired_count == 1). 3724 * 3725 * downgrading to a read lock for uvm_fault_wire avoids a possible 3726 * deadlock with another thread that may have faulted on one of 3727 * the pages to be wired (it would mark the page busy, blocking 3728 * us, then in turn block on the map lock that we hold). because 3729 * of problems in the recursive lock package, we cannot upgrade 3730 * to a write lock in vm_map_lookup. thus, any actions that 3731 * require the write lock must be done beforehand. because we 3732 * keep the read lock on the map, the copy-on-write status of the 3733 * entries we modify here cannot change. 3734 */ 3735 3736 for (size = 0, entry = map->header.next; entry != &map->header; 3737 entry = entry->next) { 3738 if (entry->protection != VM_PROT_NONE && 3739 VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3740 size += entry->end - entry->start; 3741 } 3742 } 3743 3744 if (atop(size) + uvmexp.wired > uvmexp.wiredmax) { 3745 vm_map_unlock(map); 3746 return ENOMEM; 3747 } 3748 3749 if (limit != 0 && 3750 (size + ptoa(pmap_wired_count(vm_map_pmap(map))) > limit)) { 3751 vm_map_unlock(map); 3752 return ENOMEM; 3753 } 3754 3755 /* 3756 * Pass 2. 3757 */ 3758 3759 for (entry = map->header.next; entry != &map->header; 3760 entry = entry->next) { 3761 if (entry->protection == VM_PROT_NONE) 3762 continue; 3763 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3764 3765 /* 3766 * perform actions of vm_map_lookup that need the 3767 * write lock on the map: create an anonymous map 3768 * for a copy-on-write region, or an anonymous map 3769 * for a zero-fill region. (XXXCDC: submap case 3770 * ok?) 3771 */ 3772 3773 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3774 if (UVM_ET_ISNEEDSCOPY(entry) && 3775 ((entry->max_protection & VM_PROT_WRITE) || 3776 (entry->object.uvm_obj == NULL))) { 3777 amap_copy(map, entry, 0, entry->start, 3778 entry->end); 3779 /* XXXCDC: wait OK? */ 3780 } 3781 } 3782 } 3783 entry->wired_count++; 3784 } 3785 3786 /* 3787 * Pass 3. 3788 */ 3789 3790 #ifdef DIAGNOSTIC 3791 timestamp_save = map->timestamp; 3792 #endif 3793 vm_map_busy(map); 3794 vm_map_unlock(map); 3795 3796 rv = 0; 3797 for (entry = map->header.next; entry != &map->header; 3798 entry = entry->next) { 3799 if (entry->wired_count == 1) { 3800 rv = uvm_fault_wire(map, entry->start, entry->end, 3801 entry->max_protection, 1); 3802 if (rv) { 3803 3804 /* 3805 * wiring failed. break out of the loop. 3806 * we'll clean up the map below, once we 3807 * have a write lock again. 3808 */ 3809 3810 break; 3811 } 3812 } 3813 } 3814 3815 if (rv) { 3816 3817 /* 3818 * Get back an exclusive (write) lock. 3819 */ 3820 3821 vm_map_lock(map); 3822 vm_map_unbusy(map); 3823 3824 #ifdef DIAGNOSTIC 3825 if (timestamp_save + 1 != map->timestamp) 3826 panic("uvm_map_pageable_all: stale map"); 3827 #endif 3828 3829 /* 3830 * first drop the wiring count on all the entries 3831 * which haven't actually been wired yet. 3832 * 3833 * Skip VM_PROT_NONE entries like we did above. 3834 */ 3835 3836 failed_entry = entry; 3837 for (/* nothing */; entry != &map->header; 3838 entry = entry->next) { 3839 if (entry->protection == VM_PROT_NONE) 3840 continue; 3841 entry->wired_count--; 3842 } 3843 3844 /* 3845 * now, unwire all the entries that were successfully 3846 * wired above. 3847 * 3848 * Skip VM_PROT_NONE entries like we did above. 3849 */ 3850 3851 for (entry = map->header.next; entry != failed_entry; 3852 entry = entry->next) { 3853 if (entry->protection == VM_PROT_NONE) 3854 continue; 3855 entry->wired_count--; 3856 if (VM_MAPENT_ISWIRED(entry)) 3857 uvm_map_entry_unwire(map, entry); 3858 } 3859 vm_map_unlock(map); 3860 UVMHIST_LOG(maphist,"<- done (RV=%d)", rv,0,0,0); 3861 return (rv); 3862 } 3863 3864 vm_map_unbusy(map); 3865 3866 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3867 return 0; 3868 } 3869 3870 /* 3871 * uvm_map_clean: clean out a map range 3872 * 3873 * => valid flags: 3874 * if (flags & PGO_CLEANIT): dirty pages are cleaned first 3875 * if (flags & PGO_SYNCIO): dirty pages are written synchronously 3876 * if (flags & PGO_DEACTIVATE): any cached pages are deactivated after clean 3877 * if (flags & PGO_FREE): any cached pages are freed after clean 3878 * => returns an error if any part of the specified range isn't mapped 3879 * => never a need to flush amap layer since the anonymous memory has 3880 * no permanent home, but may deactivate pages there 3881 * => called from sys_msync() and sys_madvise() 3882 * => caller must not write-lock map (read OK). 3883 * => we may sleep while cleaning if SYNCIO [with map read-locked] 3884 */ 3885 3886 int 3887 uvm_map_clean(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 3888 { 3889 struct vm_map_entry *current, *entry; 3890 struct uvm_object *uobj; 3891 struct vm_amap *amap; 3892 struct vm_anon *anon; 3893 struct vm_page *pg; 3894 vaddr_t offset; 3895 vsize_t size; 3896 voff_t uoff; 3897 int error, refs; 3898 UVMHIST_FUNC("uvm_map_clean"); UVMHIST_CALLED(maphist); 3899 3900 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,flags=0x%x)", 3901 map, start, end, flags); 3902 KASSERT((flags & (PGO_FREE|PGO_DEACTIVATE)) != 3903 (PGO_FREE|PGO_DEACTIVATE)); 3904 3905 vm_map_lock_read(map); 3906 VM_MAP_RANGE_CHECK(map, start, end); 3907 if (uvm_map_lookup_entry(map, start, &entry) == false) { 3908 vm_map_unlock_read(map); 3909 return EFAULT; 3910 } 3911 3912 /* 3913 * Make a first pass to check for holes and wiring problems. 3914 */ 3915 3916 for (current = entry; current->start < end; current = current->next) { 3917 if (UVM_ET_ISSUBMAP(current)) { 3918 vm_map_unlock_read(map); 3919 return EINVAL; 3920 } 3921 if ((flags & PGO_FREE) != 0 && VM_MAPENT_ISWIRED(entry)) { 3922 vm_map_unlock_read(map); 3923 return EBUSY; 3924 } 3925 if (end <= current->end) { 3926 break; 3927 } 3928 if (current->end != current->next->start) { 3929 vm_map_unlock_read(map); 3930 return EFAULT; 3931 } 3932 } 3933 3934 error = 0; 3935 for (current = entry; start < end; current = current->next) { 3936 amap = current->aref.ar_amap; /* top layer */ 3937 uobj = current->object.uvm_obj; /* bottom layer */ 3938 KASSERT(start >= current->start); 3939 3940 /* 3941 * No amap cleaning necessary if: 3942 * 3943 * (1) There's no amap. 3944 * 3945 * (2) We're not deactivating or freeing pages. 3946 */ 3947 3948 if (amap == NULL || (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) 3949 goto flush_object; 3950 3951 amap_lock(amap); 3952 offset = start - current->start; 3953 size = MIN(end, current->end) - start; 3954 for ( ; size != 0; size -= PAGE_SIZE, offset += PAGE_SIZE) { 3955 anon = amap_lookup(¤t->aref, offset); 3956 if (anon == NULL) 3957 continue; 3958 3959 mutex_enter(&anon->an_lock); 3960 pg = anon->an_page; 3961 if (pg == NULL) { 3962 mutex_exit(&anon->an_lock); 3963 continue; 3964 } 3965 3966 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 3967 3968 /* 3969 * In these first 3 cases, we just deactivate the page. 3970 */ 3971 3972 case PGO_CLEANIT|PGO_FREE: 3973 case PGO_CLEANIT|PGO_DEACTIVATE: 3974 case PGO_DEACTIVATE: 3975 deactivate_it: 3976 /* 3977 * skip the page if it's loaned or wired, 3978 * since it shouldn't be on a paging queue 3979 * at all in these cases. 3980 */ 3981 3982 mutex_enter(&uvm_pageqlock); 3983 if (pg->loan_count != 0 || 3984 pg->wire_count != 0) { 3985 mutex_exit(&uvm_pageqlock); 3986 mutex_exit(&anon->an_lock); 3987 continue; 3988 } 3989 KASSERT(pg->uanon == anon); 3990 uvm_pagedeactivate(pg); 3991 mutex_exit(&uvm_pageqlock); 3992 mutex_exit(&anon->an_lock); 3993 continue; 3994 3995 case PGO_FREE: 3996 3997 /* 3998 * If there are multiple references to 3999 * the amap, just deactivate the page. 4000 */ 4001 4002 if (amap_refs(amap) > 1) 4003 goto deactivate_it; 4004 4005 /* skip the page if it's wired */ 4006 if (pg->wire_count != 0) { 4007 mutex_exit(&anon->an_lock); 4008 continue; 4009 } 4010 amap_unadd(¤t->aref, offset); 4011 refs = --anon->an_ref; 4012 mutex_exit(&anon->an_lock); 4013 if (refs == 0) 4014 uvm_anfree(anon); 4015 continue; 4016 } 4017 } 4018 amap_unlock(amap); 4019 4020 flush_object: 4021 /* 4022 * flush pages if we've got a valid backing object. 4023 * note that we must always clean object pages before 4024 * freeing them since otherwise we could reveal stale 4025 * data from files. 4026 */ 4027 4028 uoff = current->offset + (start - current->start); 4029 size = MIN(end, current->end) - start; 4030 if (uobj != NULL) { 4031 mutex_enter(&uobj->vmobjlock); 4032 if (uobj->pgops->pgo_put != NULL) 4033 error = (uobj->pgops->pgo_put)(uobj, uoff, 4034 uoff + size, flags | PGO_CLEANIT); 4035 else 4036 error = 0; 4037 } 4038 start += size; 4039 } 4040 vm_map_unlock_read(map); 4041 return (error); 4042 } 4043 4044 4045 /* 4046 * uvm_map_checkprot: check protection in map 4047 * 4048 * => must allow specified protection in a fully allocated region. 4049 * => map must be read or write locked by caller. 4050 */ 4051 4052 bool 4053 uvm_map_checkprot(struct vm_map *map, vaddr_t start, vaddr_t end, 4054 vm_prot_t protection) 4055 { 4056 struct vm_map_entry *entry; 4057 struct vm_map_entry *tmp_entry; 4058 4059 if (!uvm_map_lookup_entry(map, start, &tmp_entry)) { 4060 return (false); 4061 } 4062 entry = tmp_entry; 4063 while (start < end) { 4064 if (entry == &map->header) { 4065 return (false); 4066 } 4067 4068 /* 4069 * no holes allowed 4070 */ 4071 4072 if (start < entry->start) { 4073 return (false); 4074 } 4075 4076 /* 4077 * check protection associated with entry 4078 */ 4079 4080 if ((entry->protection & protection) != protection) { 4081 return (false); 4082 } 4083 start = entry->end; 4084 entry = entry->next; 4085 } 4086 return (true); 4087 } 4088 4089 /* 4090 * uvmspace_alloc: allocate a vmspace structure. 4091 * 4092 * - structure includes vm_map and pmap 4093 * - XXX: no locking on this structure 4094 * - refcnt set to 1, rest must be init'd by caller 4095 */ 4096 struct vmspace * 4097 uvmspace_alloc(vaddr_t vmin, vaddr_t vmax) 4098 { 4099 struct vmspace *vm; 4100 UVMHIST_FUNC("uvmspace_alloc"); UVMHIST_CALLED(maphist); 4101 4102 vm = pool_cache_get(&uvm_vmspace_cache, PR_WAITOK); 4103 uvmspace_init(vm, NULL, vmin, vmax); 4104 UVMHIST_LOG(maphist,"<- done (vm=0x%x)", vm,0,0,0); 4105 return (vm); 4106 } 4107 4108 /* 4109 * uvmspace_init: initialize a vmspace structure. 4110 * 4111 * - XXX: no locking on this structure 4112 * - refcnt set to 1, rest must be init'd by caller 4113 */ 4114 void 4115 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t vmin, vaddr_t vmax) 4116 { 4117 UVMHIST_FUNC("uvmspace_init"); UVMHIST_CALLED(maphist); 4118 4119 memset(vm, 0, sizeof(*vm)); 4120 uvm_map_setup(&vm->vm_map, vmin, vmax, VM_MAP_PAGEABLE 4121 #ifdef __USING_TOPDOWN_VM 4122 | VM_MAP_TOPDOWN 4123 #endif 4124 ); 4125 if (pmap) 4126 pmap_reference(pmap); 4127 else 4128 pmap = pmap_create(); 4129 vm->vm_map.pmap = pmap; 4130 vm->vm_refcnt = 1; 4131 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4132 } 4133 4134 /* 4135 * uvmspace_share: share a vmspace between two processes 4136 * 4137 * - used for vfork, threads(?) 4138 */ 4139 4140 void 4141 uvmspace_share(struct proc *p1, struct proc *p2) 4142 { 4143 4144 uvmspace_addref(p1->p_vmspace); 4145 p2->p_vmspace = p1->p_vmspace; 4146 } 4147 4148 /* 4149 * uvmspace_unshare: ensure that process "p" has its own, unshared, vmspace 4150 * 4151 * - XXX: no locking on vmspace 4152 */ 4153 4154 void 4155 uvmspace_unshare(struct lwp *l) 4156 { 4157 struct proc *p = l->l_proc; 4158 struct vmspace *nvm, *ovm = p->p_vmspace; 4159 4160 if (ovm->vm_refcnt == 1) 4161 /* nothing to do: vmspace isn't shared in the first place */ 4162 return; 4163 4164 /* make a new vmspace, still holding old one */ 4165 nvm = uvmspace_fork(ovm); 4166 4167 kpreempt_disable(); 4168 pmap_deactivate(l); /* unbind old vmspace */ 4169 p->p_vmspace = nvm; 4170 pmap_activate(l); /* switch to new vmspace */ 4171 kpreempt_enable(); 4172 4173 uvmspace_free(ovm); /* drop reference to old vmspace */ 4174 } 4175 4176 /* 4177 * uvmspace_exec: the process wants to exec a new program 4178 */ 4179 4180 void 4181 uvmspace_exec(struct lwp *l, vaddr_t start, vaddr_t end) 4182 { 4183 struct proc *p = l->l_proc; 4184 struct vmspace *nvm, *ovm = p->p_vmspace; 4185 struct vm_map *map = &ovm->vm_map; 4186 4187 #ifdef __sparc__ 4188 /* XXX cgd 960926: the sparc #ifdef should be a MD hook */ 4189 kill_user_windows(l); /* before stack addresses go away */ 4190 #endif 4191 4192 /* 4193 * see if more than one process is using this vmspace... 4194 */ 4195 4196 if (ovm->vm_refcnt == 1) { 4197 4198 /* 4199 * if p is the only process using its vmspace then we can safely 4200 * recycle that vmspace for the program that is being exec'd. 4201 */ 4202 4203 #ifdef SYSVSHM 4204 /* 4205 * SYSV SHM semantics require us to kill all segments on an exec 4206 */ 4207 4208 if (ovm->vm_shm) 4209 shmexit(ovm); 4210 #endif 4211 4212 /* 4213 * POSIX 1003.1b -- "lock future mappings" is revoked 4214 * when a process execs another program image. 4215 */ 4216 4217 map->flags &= ~VM_MAP_WIREFUTURE; 4218 4219 /* 4220 * now unmap the old program 4221 */ 4222 4223 pmap_remove_all(map->pmap); 4224 uvm_unmap(map, vm_map_min(map), vm_map_max(map)); 4225 KASSERT(map->header.prev == &map->header); 4226 KASSERT(map->nentries == 0); 4227 4228 /* 4229 * resize the map 4230 */ 4231 4232 vm_map_setmin(map, start); 4233 vm_map_setmax(map, end); 4234 } else { 4235 4236 /* 4237 * p's vmspace is being shared, so we can't reuse it for p since 4238 * it is still being used for others. allocate a new vmspace 4239 * for p 4240 */ 4241 4242 nvm = uvmspace_alloc(start, end); 4243 4244 /* 4245 * install new vmspace and drop our ref to the old one. 4246 */ 4247 4248 kpreempt_disable(); 4249 pmap_deactivate(l); 4250 p->p_vmspace = nvm; 4251 pmap_activate(l); 4252 kpreempt_enable(); 4253 4254 uvmspace_free(ovm); 4255 } 4256 } 4257 4258 /* 4259 * uvmspace_addref: add a referece to a vmspace. 4260 */ 4261 4262 void 4263 uvmspace_addref(struct vmspace *vm) 4264 { 4265 struct vm_map *map = &vm->vm_map; 4266 4267 KASSERT((map->flags & VM_MAP_DYING) == 0); 4268 4269 mutex_enter(&map->misc_lock); 4270 KASSERT(vm->vm_refcnt > 0); 4271 vm->vm_refcnt++; 4272 mutex_exit(&map->misc_lock); 4273 } 4274 4275 /* 4276 * uvmspace_free: free a vmspace data structure 4277 */ 4278 4279 void 4280 uvmspace_free(struct vmspace *vm) 4281 { 4282 struct vm_map_entry *dead_entries; 4283 struct vm_map *map = &vm->vm_map; 4284 int n; 4285 4286 UVMHIST_FUNC("uvmspace_free"); UVMHIST_CALLED(maphist); 4287 4288 UVMHIST_LOG(maphist,"(vm=0x%x) ref=%d", vm, vm->vm_refcnt,0,0); 4289 mutex_enter(&map->misc_lock); 4290 n = --vm->vm_refcnt; 4291 mutex_exit(&map->misc_lock); 4292 if (n > 0) 4293 return; 4294 4295 /* 4296 * at this point, there should be no other references to the map. 4297 * delete all of the mappings, then destroy the pmap. 4298 */ 4299 4300 map->flags |= VM_MAP_DYING; 4301 pmap_remove_all(map->pmap); 4302 #ifdef SYSVSHM 4303 /* Get rid of any SYSV shared memory segments. */ 4304 if (vm->vm_shm != NULL) 4305 shmexit(vm); 4306 #endif 4307 if (map->nentries) { 4308 uvm_unmap_remove(map, vm_map_min(map), vm_map_max(map), 4309 &dead_entries, NULL, 0); 4310 if (dead_entries != NULL) 4311 uvm_unmap_detach(dead_entries, 0); 4312 } 4313 KASSERT(map->nentries == 0); 4314 KASSERT(map->size == 0); 4315 mutex_destroy(&map->misc_lock); 4316 mutex_destroy(&map->mutex); 4317 rw_destroy(&map->lock); 4318 cv_destroy(&map->cv); 4319 pmap_destroy(map->pmap); 4320 pool_cache_put(&uvm_vmspace_cache, vm); 4321 } 4322 4323 /* 4324 * F O R K - m a i n e n t r y p o i n t 4325 */ 4326 /* 4327 * uvmspace_fork: fork a process' main map 4328 * 4329 * => create a new vmspace for child process from parent. 4330 * => parent's map must not be locked. 4331 */ 4332 4333 struct vmspace * 4334 uvmspace_fork(struct vmspace *vm1) 4335 { 4336 struct vmspace *vm2; 4337 struct vm_map *old_map = &vm1->vm_map; 4338 struct vm_map *new_map; 4339 struct vm_map_entry *old_entry; 4340 struct vm_map_entry *new_entry; 4341 UVMHIST_FUNC("uvmspace_fork"); UVMHIST_CALLED(maphist); 4342 4343 vm_map_lock(old_map); 4344 4345 vm2 = uvmspace_alloc(vm_map_min(old_map), vm_map_max(old_map)); 4346 memcpy(&vm2->vm_startcopy, &vm1->vm_startcopy, 4347 (char *) (vm1 + 1) - (char *) &vm1->vm_startcopy); 4348 new_map = &vm2->vm_map; /* XXX */ 4349 4350 old_entry = old_map->header.next; 4351 new_map->size = old_map->size; 4352 4353 /* 4354 * go entry-by-entry 4355 */ 4356 4357 while (old_entry != &old_map->header) { 4358 4359 /* 4360 * first, some sanity checks on the old entry 4361 */ 4362 4363 KASSERT(!UVM_ET_ISSUBMAP(old_entry)); 4364 KASSERT(UVM_ET_ISCOPYONWRITE(old_entry) || 4365 !UVM_ET_ISNEEDSCOPY(old_entry)); 4366 4367 switch (old_entry->inheritance) { 4368 case MAP_INHERIT_NONE: 4369 4370 /* 4371 * drop the mapping, modify size 4372 */ 4373 new_map->size -= old_entry->end - old_entry->start; 4374 break; 4375 4376 case MAP_INHERIT_SHARE: 4377 4378 /* 4379 * share the mapping: this means we want the old and 4380 * new entries to share amaps and backing objects. 4381 */ 4382 /* 4383 * if the old_entry needs a new amap (due to prev fork) 4384 * then we need to allocate it now so that we have 4385 * something we own to share with the new_entry. [in 4386 * other words, we need to clear needs_copy] 4387 */ 4388 4389 if (UVM_ET_ISNEEDSCOPY(old_entry)) { 4390 /* get our own amap, clears needs_copy */ 4391 amap_copy(old_map, old_entry, AMAP_COPY_NOCHUNK, 4392 0, 0); 4393 /* XXXCDC: WAITOK??? */ 4394 } 4395 4396 new_entry = uvm_mapent_alloc(new_map, 0); 4397 /* old_entry -> new_entry */ 4398 uvm_mapent_copy(old_entry, new_entry); 4399 4400 /* new pmap has nothing wired in it */ 4401 new_entry->wired_count = 0; 4402 4403 /* 4404 * gain reference to object backing the map (can't 4405 * be a submap, already checked this case). 4406 */ 4407 4408 if (new_entry->aref.ar_amap) 4409 uvm_map_reference_amap(new_entry, AMAP_SHARED); 4410 4411 if (new_entry->object.uvm_obj && 4412 new_entry->object.uvm_obj->pgops->pgo_reference) 4413 new_entry->object.uvm_obj-> 4414 pgops->pgo_reference( 4415 new_entry->object.uvm_obj); 4416 4417 /* insert entry at end of new_map's entry list */ 4418 uvm_map_entry_link(new_map, new_map->header.prev, 4419 new_entry); 4420 4421 break; 4422 4423 case MAP_INHERIT_COPY: 4424 4425 /* 4426 * copy-on-write the mapping (using mmap's 4427 * MAP_PRIVATE semantics) 4428 * 4429 * allocate new_entry, adjust reference counts. 4430 * (note that new references are read-only). 4431 */ 4432 4433 new_entry = uvm_mapent_alloc(new_map, 0); 4434 /* old_entry -> new_entry */ 4435 uvm_mapent_copy(old_entry, new_entry); 4436 4437 if (new_entry->aref.ar_amap) 4438 uvm_map_reference_amap(new_entry, 0); 4439 4440 if (new_entry->object.uvm_obj && 4441 new_entry->object.uvm_obj->pgops->pgo_reference) 4442 new_entry->object.uvm_obj->pgops->pgo_reference 4443 (new_entry->object.uvm_obj); 4444 4445 /* new pmap has nothing wired in it */ 4446 new_entry->wired_count = 0; 4447 4448 new_entry->etype |= 4449 (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY); 4450 uvm_map_entry_link(new_map, new_map->header.prev, 4451 new_entry); 4452 4453 /* 4454 * the new entry will need an amap. it will either 4455 * need to be copied from the old entry or created 4456 * from scratch (if the old entry does not have an 4457 * amap). can we defer this process until later 4458 * (by setting "needs_copy") or do we need to copy 4459 * the amap now? 4460 * 4461 * we must copy the amap now if any of the following 4462 * conditions hold: 4463 * 1. the old entry has an amap and that amap is 4464 * being shared. this means that the old (parent) 4465 * process is sharing the amap with another 4466 * process. if we do not clear needs_copy here 4467 * we will end up in a situation where both the 4468 * parent and child process are refering to the 4469 * same amap with "needs_copy" set. if the 4470 * parent write-faults, the fault routine will 4471 * clear "needs_copy" in the parent by allocating 4472 * a new amap. this is wrong because the 4473 * parent is supposed to be sharing the old amap 4474 * and the new amap will break that. 4475 * 4476 * 2. if the old entry has an amap and a non-zero 4477 * wire count then we are going to have to call 4478 * amap_cow_now to avoid page faults in the 4479 * parent process. since amap_cow_now requires 4480 * "needs_copy" to be clear we might as well 4481 * clear it here as well. 4482 * 4483 */ 4484 4485 if (old_entry->aref.ar_amap != NULL) { 4486 if ((amap_flags(old_entry->aref.ar_amap) & 4487 AMAP_SHARED) != 0 || 4488 VM_MAPENT_ISWIRED(old_entry)) { 4489 4490 amap_copy(new_map, new_entry, 4491 AMAP_COPY_NOCHUNK, 0, 0); 4492 /* XXXCDC: M_WAITOK ... ok? */ 4493 } 4494 } 4495 4496 /* 4497 * if the parent's entry is wired down, then the 4498 * parent process does not want page faults on 4499 * access to that memory. this means that we 4500 * cannot do copy-on-write because we can't write 4501 * protect the old entry. in this case we 4502 * resolve all copy-on-write faults now, using 4503 * amap_cow_now. note that we have already 4504 * allocated any needed amap (above). 4505 */ 4506 4507 if (VM_MAPENT_ISWIRED(old_entry)) { 4508 4509 /* 4510 * resolve all copy-on-write faults now 4511 * (note that there is nothing to do if 4512 * the old mapping does not have an amap). 4513 */ 4514 if (old_entry->aref.ar_amap) 4515 amap_cow_now(new_map, new_entry); 4516 4517 } else { 4518 4519 /* 4520 * setup mappings to trigger copy-on-write faults 4521 * we must write-protect the parent if it has 4522 * an amap and it is not already "needs_copy"... 4523 * if it is already "needs_copy" then the parent 4524 * has already been write-protected by a previous 4525 * fork operation. 4526 */ 4527 4528 if (old_entry->aref.ar_amap && 4529 !UVM_ET_ISNEEDSCOPY(old_entry)) { 4530 if (old_entry->max_protection & VM_PROT_WRITE) { 4531 pmap_protect(old_map->pmap, 4532 old_entry->start, 4533 old_entry->end, 4534 old_entry->protection & 4535 ~VM_PROT_WRITE); 4536 } 4537 old_entry->etype |= UVM_ET_NEEDSCOPY; 4538 } 4539 } 4540 break; 4541 } /* end of switch statement */ 4542 old_entry = old_entry->next; 4543 } 4544 4545 pmap_update(old_map->pmap); 4546 vm_map_unlock(old_map); 4547 4548 #ifdef SYSVSHM 4549 if (vm1->vm_shm) 4550 shmfork(vm1, vm2); 4551 #endif 4552 4553 #ifdef PMAP_FORK 4554 pmap_fork(vm1->vm_map.pmap, vm2->vm_map.pmap); 4555 #endif 4556 4557 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4558 return (vm2); 4559 } 4560 4561 4562 /* 4563 * in-kernel map entry allocation. 4564 */ 4565 4566 struct uvm_kmapent_hdr { 4567 LIST_ENTRY(uvm_kmapent_hdr) ukh_listq; 4568 int ukh_nused; 4569 struct vm_map_entry *ukh_freelist; 4570 struct vm_map *ukh_map; 4571 struct vm_map_entry ukh_entries[0]; 4572 }; 4573 4574 #define UVM_KMAPENT_CHUNK \ 4575 ((PAGE_SIZE - sizeof(struct uvm_kmapent_hdr)) \ 4576 / sizeof(struct vm_map_entry)) 4577 4578 #define UVM_KHDR_FIND(entry) \ 4579 ((struct uvm_kmapent_hdr *)(((vaddr_t)entry) & ~PAGE_MASK)) 4580 4581 4582 #ifdef DIAGNOSTIC 4583 static struct vm_map * 4584 uvm_kmapent_map(struct vm_map_entry *entry) 4585 { 4586 const struct uvm_kmapent_hdr *ukh; 4587 4588 ukh = UVM_KHDR_FIND(entry); 4589 return ukh->ukh_map; 4590 } 4591 #endif 4592 4593 static inline struct vm_map_entry * 4594 uvm_kmapent_get(struct uvm_kmapent_hdr *ukh) 4595 { 4596 struct vm_map_entry *entry; 4597 4598 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4599 KASSERT(ukh->ukh_nused >= 0); 4600 4601 entry = ukh->ukh_freelist; 4602 if (entry) { 4603 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4604 == UVM_MAP_KERNEL); 4605 ukh->ukh_freelist = entry->next; 4606 ukh->ukh_nused++; 4607 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4608 } else { 4609 KASSERT(ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4610 } 4611 4612 return entry; 4613 } 4614 4615 static inline void 4616 uvm_kmapent_put(struct uvm_kmapent_hdr *ukh, struct vm_map_entry *entry) 4617 { 4618 4619 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4620 == UVM_MAP_KERNEL); 4621 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4622 KASSERT(ukh->ukh_nused > 0); 4623 KASSERT(ukh->ukh_freelist != NULL || 4624 ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4625 KASSERT(ukh->ukh_freelist == NULL || 4626 ukh->ukh_nused < UVM_KMAPENT_CHUNK); 4627 4628 ukh->ukh_nused--; 4629 entry->next = ukh->ukh_freelist; 4630 ukh->ukh_freelist = entry; 4631 } 4632 4633 /* 4634 * uvm_kmapent_alloc: allocate a map entry for in-kernel map 4635 */ 4636 4637 static struct vm_map_entry * 4638 uvm_kmapent_alloc(struct vm_map *map, int flags) 4639 { 4640 struct vm_page *pg; 4641 struct uvm_map_args args; 4642 struct uvm_kmapent_hdr *ukh; 4643 struct vm_map_entry *entry; 4644 uvm_flag_t mapflags = UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, 4645 UVM_INH_NONE, UVM_ADV_RANDOM, flags | UVM_FLAG_NOMERGE); 4646 vaddr_t va; 4647 int error; 4648 int i; 4649 4650 KDASSERT(UVM_KMAPENT_CHUNK > 2); 4651 KDASSERT(kernel_map != NULL); 4652 KASSERT(vm_map_pmap(map) == pmap_kernel()); 4653 4654 UVMMAP_EVCNT_INCR(uke_alloc); 4655 entry = NULL; 4656 again: 4657 /* 4658 * try to grab an entry from freelist. 4659 */ 4660 mutex_spin_enter(&uvm_kentry_lock); 4661 ukh = LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free); 4662 if (ukh) { 4663 entry = uvm_kmapent_get(ukh); 4664 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK) 4665 LIST_REMOVE(ukh, ukh_listq); 4666 } 4667 mutex_spin_exit(&uvm_kentry_lock); 4668 4669 if (entry) 4670 return entry; 4671 4672 /* 4673 * there's no free entry for this vm_map. 4674 * now we need to allocate some vm_map_entry. 4675 * for simplicity, always allocate one page chunk of them at once. 4676 */ 4677 4678 pg = uvm_pagealloc(NULL, 0, NULL, 4679 (flags & UVM_KMF_NOWAIT) != 0 ? UVM_PGA_USERESERVE : 0); 4680 if (__predict_false(pg == NULL)) { 4681 if (flags & UVM_FLAG_NOWAIT) 4682 return NULL; 4683 uvm_wait("kme_alloc"); 4684 goto again; 4685 } 4686 4687 error = uvm_map_prepare(map, 0, PAGE_SIZE, NULL, UVM_UNKNOWN_OFFSET, 4688 0, mapflags, &args); 4689 if (error) { 4690 uvm_pagefree(pg); 4691 return NULL; 4692 } 4693 4694 va = args.uma_start; 4695 4696 pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), 4697 VM_PROT_READ|VM_PROT_WRITE|PMAP_KMPAGE); 4698 pmap_update(vm_map_pmap(map)); 4699 4700 ukh = (void *)va; 4701 4702 /* 4703 * use the first entry for ukh itsself. 4704 */ 4705 4706 entry = &ukh->ukh_entries[0]; 4707 entry->flags = UVM_MAP_KERNEL | UVM_MAP_KMAPENT; 4708 error = uvm_map_enter(map, &args, entry); 4709 KASSERT(error == 0); 4710 4711 ukh->ukh_nused = UVM_KMAPENT_CHUNK; 4712 ukh->ukh_map = map; 4713 ukh->ukh_freelist = NULL; 4714 for (i = UVM_KMAPENT_CHUNK - 1; i >= 2; i--) { 4715 struct vm_map_entry *xentry = &ukh->ukh_entries[i]; 4716 4717 xentry->flags = UVM_MAP_KERNEL; 4718 uvm_kmapent_put(ukh, xentry); 4719 } 4720 KASSERT(ukh->ukh_nused == 2); 4721 4722 mutex_spin_enter(&uvm_kentry_lock); 4723 LIST_INSERT_HEAD(&vm_map_to_kernel(map)->vmk_kentry_free, 4724 ukh, ukh_listq); 4725 mutex_spin_exit(&uvm_kentry_lock); 4726 4727 /* 4728 * return second entry. 4729 */ 4730 4731 entry = &ukh->ukh_entries[1]; 4732 entry->flags = UVM_MAP_KERNEL; 4733 UVMMAP_EVCNT_INCR(ukh_alloc); 4734 return entry; 4735 } 4736 4737 /* 4738 * uvm_mapent_free: free map entry for in-kernel map 4739 */ 4740 4741 static void 4742 uvm_kmapent_free(struct vm_map_entry *entry) 4743 { 4744 struct uvm_kmapent_hdr *ukh; 4745 struct vm_page *pg; 4746 struct vm_map *map; 4747 struct pmap *pmap; 4748 vaddr_t va; 4749 paddr_t pa; 4750 struct vm_map_entry *deadentry; 4751 4752 UVMMAP_EVCNT_INCR(uke_free); 4753 ukh = UVM_KHDR_FIND(entry); 4754 map = ukh->ukh_map; 4755 4756 mutex_spin_enter(&uvm_kentry_lock); 4757 uvm_kmapent_put(ukh, entry); 4758 if (ukh->ukh_nused > 1) { 4759 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK - 1) 4760 LIST_INSERT_HEAD( 4761 &vm_map_to_kernel(map)->vmk_kentry_free, 4762 ukh, ukh_listq); 4763 mutex_spin_exit(&uvm_kentry_lock); 4764 return; 4765 } 4766 4767 /* 4768 * now we can free this ukh. 4769 * 4770 * however, keep an empty ukh to avoid ping-pong. 4771 */ 4772 4773 if (LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free) == ukh && 4774 LIST_NEXT(ukh, ukh_listq) == NULL) { 4775 mutex_spin_exit(&uvm_kentry_lock); 4776 return; 4777 } 4778 LIST_REMOVE(ukh, ukh_listq); 4779 mutex_spin_exit(&uvm_kentry_lock); 4780 4781 KASSERT(ukh->ukh_nused == 1); 4782 4783 /* 4784 * remove map entry for ukh itsself. 4785 */ 4786 4787 va = (vaddr_t)ukh; 4788 KASSERT((va & PAGE_MASK) == 0); 4789 vm_map_lock(map); 4790 uvm_unmap_remove(map, va, va + PAGE_SIZE, &deadentry, NULL, 0); 4791 KASSERT(deadentry->flags & UVM_MAP_KERNEL); 4792 KASSERT(deadentry->flags & UVM_MAP_KMAPENT); 4793 KASSERT(deadentry->next == NULL); 4794 KASSERT(deadentry == &ukh->ukh_entries[0]); 4795 4796 /* 4797 * unmap the page from pmap and free it. 4798 */ 4799 4800 pmap = vm_map_pmap(map); 4801 KASSERT(pmap == pmap_kernel()); 4802 if (!pmap_extract(pmap, va, &pa)) 4803 panic("%s: no mapping", __func__); 4804 pmap_kremove(va, PAGE_SIZE); 4805 pmap_update(vm_map_pmap(map)); 4806 vm_map_unlock(map); 4807 pg = PHYS_TO_VM_PAGE(pa); 4808 uvm_pagefree(pg); 4809 UVMMAP_EVCNT_INCR(ukh_free); 4810 } 4811 4812 static vsize_t 4813 uvm_kmapent_overhead(vsize_t size) 4814 { 4815 4816 /* 4817 * - the max number of unmerged entries is howmany(size, PAGE_SIZE) 4818 * as the min allocation unit is PAGE_SIZE. 4819 * - UVM_KMAPENT_CHUNK "kmapent"s are allocated from a page. 4820 * one of them are used to map the page itself. 4821 */ 4822 4823 return howmany(howmany(size, PAGE_SIZE), (UVM_KMAPENT_CHUNK - 1)) * 4824 PAGE_SIZE; 4825 } 4826 4827 /* 4828 * map entry reservation 4829 */ 4830 4831 /* 4832 * uvm_mapent_reserve: reserve map entries for clipping before locking map. 4833 * 4834 * => needed when unmapping entries allocated without UVM_FLAG_QUANTUM. 4835 * => caller shouldn't hold map locked. 4836 */ 4837 int 4838 uvm_mapent_reserve(struct vm_map *map, struct uvm_mapent_reservation *umr, 4839 int nentries, int flags) 4840 { 4841 4842 umr->umr_nentries = 0; 4843 4844 if ((flags & UVM_FLAG_QUANTUM) != 0) 4845 return 0; 4846 4847 if (!VM_MAP_USE_KMAPENT(map)) 4848 return 0; 4849 4850 while (nentries--) { 4851 struct vm_map_entry *ent; 4852 ent = uvm_kmapent_alloc(map, flags); 4853 if (!ent) { 4854 uvm_mapent_unreserve(map, umr); 4855 return ENOMEM; 4856 } 4857 UMR_PUTENTRY(umr, ent); 4858 } 4859 4860 return 0; 4861 } 4862 4863 /* 4864 * uvm_mapent_unreserve: 4865 * 4866 * => caller shouldn't hold map locked. 4867 * => never fail or sleep. 4868 */ 4869 void 4870 uvm_mapent_unreserve(struct vm_map *map, struct uvm_mapent_reservation *umr) 4871 { 4872 4873 while (!UMR_EMPTY(umr)) 4874 uvm_kmapent_free(UMR_GETENTRY(umr)); 4875 } 4876 4877 /* 4878 * uvm_mapent_trymerge: try to merge an entry with its neighbors. 4879 * 4880 * => called with map locked. 4881 * => return non zero if successfully merged. 4882 */ 4883 4884 int 4885 uvm_mapent_trymerge(struct vm_map *map, struct vm_map_entry *entry, int flags) 4886 { 4887 struct uvm_object *uobj; 4888 struct vm_map_entry *next; 4889 struct vm_map_entry *prev; 4890 vsize_t size; 4891 int merged = 0; 4892 bool copying; 4893 int newetype; 4894 4895 if (VM_MAP_USE_KMAPENT(map)) { 4896 return 0; 4897 } 4898 if (entry->aref.ar_amap != NULL) { 4899 return 0; 4900 } 4901 if ((entry->flags & UVM_MAP_NOMERGE) != 0) { 4902 return 0; 4903 } 4904 4905 uobj = entry->object.uvm_obj; 4906 size = entry->end - entry->start; 4907 copying = (flags & UVM_MERGE_COPYING) != 0; 4908 newetype = copying ? (entry->etype & ~UVM_ET_NEEDSCOPY) : entry->etype; 4909 4910 next = entry->next; 4911 if (next != &map->header && 4912 next->start == entry->end && 4913 ((copying && next->aref.ar_amap != NULL && 4914 amap_refs(next->aref.ar_amap) == 1) || 4915 (!copying && next->aref.ar_amap == NULL)) && 4916 UVM_ET_ISCOMPATIBLE(next, newetype, 4917 uobj, entry->flags, entry->protection, 4918 entry->max_protection, entry->inheritance, entry->advice, 4919 entry->wired_count) && 4920 (uobj == NULL || entry->offset + size == next->offset)) { 4921 int error; 4922 4923 if (copying) { 4924 error = amap_extend(next, size, 4925 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_BACKWARDS); 4926 } else { 4927 error = 0; 4928 } 4929 if (error == 0) { 4930 if (uobj) { 4931 if (uobj->pgops->pgo_detach) { 4932 uobj->pgops->pgo_detach(uobj); 4933 } 4934 } 4935 4936 entry->end = next->end; 4937 clear_hints(map, next); 4938 uvm_map_entry_unlink(map, next); 4939 if (copying) { 4940 entry->aref = next->aref; 4941 entry->etype &= ~UVM_ET_NEEDSCOPY; 4942 } 4943 uvm_map_check(map, "trymerge forwardmerge"); 4944 uvm_mapent_free_merged(map, next); 4945 merged++; 4946 } 4947 } 4948 4949 prev = entry->prev; 4950 if (prev != &map->header && 4951 prev->end == entry->start && 4952 ((copying && !merged && prev->aref.ar_amap != NULL && 4953 amap_refs(prev->aref.ar_amap) == 1) || 4954 (!copying && prev->aref.ar_amap == NULL)) && 4955 UVM_ET_ISCOMPATIBLE(prev, newetype, 4956 uobj, entry->flags, entry->protection, 4957 entry->max_protection, entry->inheritance, entry->advice, 4958 entry->wired_count) && 4959 (uobj == NULL || 4960 prev->offset + prev->end - prev->start == entry->offset)) { 4961 int error; 4962 4963 if (copying) { 4964 error = amap_extend(prev, size, 4965 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_FORWARDS); 4966 } else { 4967 error = 0; 4968 } 4969 if (error == 0) { 4970 if (uobj) { 4971 if (uobj->pgops->pgo_detach) { 4972 uobj->pgops->pgo_detach(uobj); 4973 } 4974 entry->offset = prev->offset; 4975 } 4976 4977 entry->start = prev->start; 4978 clear_hints(map, prev); 4979 uvm_map_entry_unlink(map, prev); 4980 if (copying) { 4981 entry->aref = prev->aref; 4982 entry->etype &= ~UVM_ET_NEEDSCOPY; 4983 } 4984 uvm_map_check(map, "trymerge backmerge"); 4985 uvm_mapent_free_merged(map, prev); 4986 merged++; 4987 } 4988 } 4989 4990 return merged; 4991 } 4992 4993 #if defined(DDB) || defined(DEBUGPRINT) 4994 4995 /* 4996 * DDB hooks 4997 */ 4998 4999 /* 5000 * uvm_map_printit: actually prints the map 5001 */ 5002 5003 void 5004 uvm_map_printit(struct vm_map *map, bool full, 5005 void (*pr)(const char *, ...)) 5006 { 5007 struct vm_map_entry *entry; 5008 5009 (*pr)("MAP %p: [0x%lx->0x%lx]\n", map, vm_map_min(map), 5010 vm_map_max(map)); 5011 (*pr)("\t#ent=%d, sz=%d, ref=%d, version=%d, flags=0x%x\n", 5012 map->nentries, map->size, map->ref_count, map->timestamp, 5013 map->flags); 5014 (*pr)("\tpmap=%p(resident=%ld, wired=%ld)\n", map->pmap, 5015 pmap_resident_count(map->pmap), pmap_wired_count(map->pmap)); 5016 if (!full) 5017 return; 5018 for (entry = map->header.next; entry != &map->header; 5019 entry = entry->next) { 5020 (*pr)(" - %p: 0x%lx->0x%lx: obj=%p/0x%llx, amap=%p/%d\n", 5021 entry, entry->start, entry->end, entry->object.uvm_obj, 5022 (long long)entry->offset, entry->aref.ar_amap, 5023 entry->aref.ar_pageoff); 5024 (*pr)( 5025 "\tsubmap=%c, cow=%c, nc=%c, prot(max)=%d/%d, inh=%d, " 5026 "wc=%d, adv=%d\n", 5027 (entry->etype & UVM_ET_SUBMAP) ? 'T' : 'F', 5028 (entry->etype & UVM_ET_COPYONWRITE) ? 'T' : 'F', 5029 (entry->etype & UVM_ET_NEEDSCOPY) ? 'T' : 'F', 5030 entry->protection, entry->max_protection, 5031 entry->inheritance, entry->wired_count, entry->advice); 5032 } 5033 } 5034 5035 /* 5036 * uvm_object_printit: actually prints the object 5037 */ 5038 5039 void 5040 uvm_object_printit(struct uvm_object *uobj, bool full, 5041 void (*pr)(const char *, ...)) 5042 { 5043 struct vm_page *pg; 5044 int cnt = 0; 5045 5046 (*pr)("OBJECT %p: locked=%d, pgops=%p, npages=%d, ", 5047 uobj, mutex_owned(&uobj->vmobjlock), uobj->pgops, uobj->uo_npages); 5048 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 5049 (*pr)("refs=<SYSTEM>\n"); 5050 else 5051 (*pr)("refs=%d\n", uobj->uo_refs); 5052 5053 if (!full) { 5054 return; 5055 } 5056 (*pr)(" PAGES <pg,offset>:\n "); 5057 TAILQ_FOREACH(pg, &uobj->memq, listq.queue) { 5058 cnt++; 5059 (*pr)("<%p,0x%llx> ", pg, (long long)pg->offset); 5060 if ((cnt % 3) == 0) { 5061 (*pr)("\n "); 5062 } 5063 } 5064 if ((cnt % 3) != 0) { 5065 (*pr)("\n"); 5066 } 5067 } 5068 5069 /* 5070 * uvm_page_printit: actually print the page 5071 */ 5072 5073 static const char page_flagbits[] = UVM_PGFLAGBITS; 5074 static const char page_pqflagbits[] = UVM_PQFLAGBITS; 5075 5076 void 5077 uvm_page_printit(struct vm_page *pg, bool full, 5078 void (*pr)(const char *, ...)) 5079 { 5080 struct vm_page *tpg; 5081 struct uvm_object *uobj; 5082 struct pgflist *pgl; 5083 char pgbuf[128]; 5084 char pqbuf[128]; 5085 5086 (*pr)("PAGE %p:\n", pg); 5087 snprintb(pgbuf, sizeof(pgbuf), page_flagbits, pg->flags); 5088 snprintb(pqbuf, sizeof(pqbuf), page_pqflagbits, pg->pqflags); 5089 (*pr)(" flags=%s, pqflags=%s, wire_count=%d, pa=0x%lx\n", 5090 pgbuf, pqbuf, pg->wire_count, (long)VM_PAGE_TO_PHYS(pg)); 5091 (*pr)(" uobject=%p, uanon=%p, offset=0x%llx loan_count=%d\n", 5092 pg->uobject, pg->uanon, (long long)pg->offset, pg->loan_count); 5093 #if defined(UVM_PAGE_TRKOWN) 5094 if (pg->flags & PG_BUSY) 5095 (*pr)(" owning process = %d, tag=%s\n", 5096 pg->owner, pg->owner_tag); 5097 else 5098 (*pr)(" page not busy, no owner\n"); 5099 #else 5100 (*pr)(" [page ownership tracking disabled]\n"); 5101 #endif 5102 5103 if (!full) 5104 return; 5105 5106 /* cross-verify object/anon */ 5107 if ((pg->pqflags & PQ_FREE) == 0) { 5108 if (pg->pqflags & PQ_ANON) { 5109 if (pg->uanon == NULL || pg->uanon->an_page != pg) 5110 (*pr)(" >>> ANON DOES NOT POINT HERE <<< (%p)\n", 5111 (pg->uanon) ? pg->uanon->an_page : NULL); 5112 else 5113 (*pr)(" anon backpointer is OK\n"); 5114 } else { 5115 uobj = pg->uobject; 5116 if (uobj) { 5117 (*pr)(" checking object list\n"); 5118 TAILQ_FOREACH(tpg, &uobj->memq, listq.queue) { 5119 if (tpg == pg) { 5120 break; 5121 } 5122 } 5123 if (tpg) 5124 (*pr)(" page found on object list\n"); 5125 else 5126 (*pr)(" >>> PAGE NOT FOUND ON OBJECT LIST! <<<\n"); 5127 } 5128 } 5129 } 5130 5131 /* cross-verify page queue */ 5132 if (pg->pqflags & PQ_FREE) { 5133 int fl = uvm_page_lookup_freelist(pg); 5134 int color = VM_PGCOLOR_BUCKET(pg); 5135 pgl = &uvm.page_free[fl].pgfl_buckets[color].pgfl_queues[ 5136 ((pg)->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN]; 5137 } else { 5138 pgl = NULL; 5139 } 5140 5141 if (pgl) { 5142 (*pr)(" checking pageq list\n"); 5143 LIST_FOREACH(tpg, pgl, pageq.list) { 5144 if (tpg == pg) { 5145 break; 5146 } 5147 } 5148 if (tpg) 5149 (*pr)(" page found on pageq list\n"); 5150 else 5151 (*pr)(" >>> PAGE NOT FOUND ON PAGEQ LIST! <<<\n"); 5152 } 5153 } 5154 5155 /* 5156 * uvm_pages_printthem - print a summary of all managed pages 5157 */ 5158 5159 void 5160 uvm_page_printall(void (*pr)(const char *, ...)) 5161 { 5162 unsigned i; 5163 struct vm_page *pg; 5164 5165 (*pr)("%18s %4s %4s %18s %18s" 5166 #ifdef UVM_PAGE_TRKOWN 5167 " OWNER" 5168 #endif 5169 "\n", "PAGE", "FLAG", "PQ", "UOBJECT", "UANON"); 5170 for (i = 0; i < vm_nphysseg; i++) { 5171 for (pg = vm_physmem[i].pgs; pg <= vm_physmem[i].lastpg; pg++) { 5172 (*pr)("%18p %04x %04x %18p %18p", 5173 pg, pg->flags, pg->pqflags, pg->uobject, 5174 pg->uanon); 5175 #ifdef UVM_PAGE_TRKOWN 5176 if (pg->flags & PG_BUSY) 5177 (*pr)(" %d [%s]", pg->owner, pg->owner_tag); 5178 #endif 5179 (*pr)("\n"); 5180 } 5181 } 5182 } 5183 5184 #endif 5185 5186 /* 5187 * uvm_map_create: create map 5188 */ 5189 5190 struct vm_map * 5191 uvm_map_create(pmap_t pmap, vaddr_t vmin, vaddr_t vmax, int flags) 5192 { 5193 struct vm_map *result; 5194 5195 result = malloc(sizeof(struct vm_map), M_VMMAP, M_WAITOK); 5196 uvm_map_setup(result, vmin, vmax, flags); 5197 result->pmap = pmap; 5198 return(result); 5199 } 5200 5201 /* 5202 * uvm_map_setup: init map 5203 * 5204 * => map must not be in service yet. 5205 */ 5206 5207 void 5208 uvm_map_setup(struct vm_map *map, vaddr_t vmin, vaddr_t vmax, int flags) 5209 { 5210 int ipl; 5211 5212 rb_tree_init(&map->rb_tree, &uvm_map_tree_ops); 5213 map->header.next = map->header.prev = &map->header; 5214 map->nentries = 0; 5215 map->size = 0; 5216 map->ref_count = 1; 5217 vm_map_setmin(map, vmin); 5218 vm_map_setmax(map, vmax); 5219 map->flags = flags; 5220 map->first_free = &map->header; 5221 map->hint = &map->header; 5222 map->timestamp = 0; 5223 map->busy = NULL; 5224 5225 if ((flags & VM_MAP_INTRSAFE) != 0) { 5226 ipl = IPL_VM; 5227 } else { 5228 ipl = IPL_NONE; 5229 } 5230 5231 rw_init(&map->lock); 5232 cv_init(&map->cv, "vm_map"); 5233 mutex_init(&map->misc_lock, MUTEX_DRIVER, ipl); 5234 mutex_init(&map->mutex, MUTEX_DRIVER, ipl); 5235 } 5236 5237 5238 /* 5239 * U N M A P - m a i n e n t r y p o i n t 5240 */ 5241 5242 /* 5243 * uvm_unmap1: remove mappings from a vm_map (from "start" up to "stop") 5244 * 5245 * => caller must check alignment and size 5246 * => map must be unlocked (we will lock it) 5247 * => flags is UVM_FLAG_QUANTUM or 0. 5248 */ 5249 5250 void 5251 uvm_unmap1(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 5252 { 5253 struct vm_map_entry *dead_entries; 5254 struct uvm_mapent_reservation umr; 5255 UVMHIST_FUNC("uvm_unmap"); UVMHIST_CALLED(maphist); 5256 5257 UVMHIST_LOG(maphist, " (map=0x%x, start=0x%x, end=0x%x)", 5258 map, start, end, 0); 5259 if (map == kernel_map) { 5260 LOCKDEBUG_MEM_CHECK((void *)start, end - start); 5261 } 5262 /* 5263 * work now done by helper functions. wipe the pmap's and then 5264 * detach from the dead entries... 5265 */ 5266 uvm_mapent_reserve(map, &umr, 2, flags); 5267 vm_map_lock(map); 5268 uvm_unmap_remove(map, start, end, &dead_entries, &umr, flags); 5269 vm_map_unlock(map); 5270 uvm_mapent_unreserve(map, &umr); 5271 5272 if (dead_entries != NULL) 5273 uvm_unmap_detach(dead_entries, 0); 5274 5275 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 5276 } 5277 5278 5279 /* 5280 * uvm_map_reference: add reference to a map 5281 * 5282 * => map need not be locked (we use misc_lock). 5283 */ 5284 5285 void 5286 uvm_map_reference(struct vm_map *map) 5287 { 5288 mutex_enter(&map->misc_lock); 5289 map->ref_count++; 5290 mutex_exit(&map->misc_lock); 5291 } 5292 5293 struct vm_map_kernel * 5294 vm_map_to_kernel(struct vm_map *map) 5295 { 5296 5297 KASSERT(VM_MAP_IS_KERNEL(map)); 5298 5299 return (struct vm_map_kernel *)map; 5300 } 5301 5302 bool 5303 vm_map_starved_p(struct vm_map *map) 5304 { 5305 5306 if ((map->flags & VM_MAP_WANTVA) != 0) { 5307 return true; 5308 } 5309 /* XXX */ 5310 if ((vm_map_max(map) - vm_map_min(map)) / 16 * 15 < map->size) { 5311 return true; 5312 } 5313 return false; 5314 } 5315 5316 #if defined(DDB) || defined(DEBUGPRINT) 5317 void 5318 uvm_whatis(uintptr_t addr, void (*pr)(const char *, ...)) 5319 { 5320 struct vm_map *map; 5321 5322 for (map = kernel_map;;) { 5323 struct vm_map_entry *entry; 5324 5325 if (!uvm_map_lookup_entry_bytree(map, (vaddr_t)addr, &entry)) { 5326 break; 5327 } 5328 (*pr)("%p is %p+%zu from VMMAP %p\n", 5329 (void *)addr, (void *)entry->start, 5330 (size_t)(addr - (uintptr_t)entry->start), map); 5331 if (!UVM_ET_ISSUBMAP(entry)) { 5332 break; 5333 } 5334 map = entry->object.sub_map; 5335 } 5336 } 5337 #endif /* defined(DDB) || defined(DEBUGPRINT) */ 5338
Indexes created Mon Sep 22 13:09:51 GMT 2025