1 /* $NetBSD: uvm_map.c,v 1.391 2021/11/25 09:40:45 skrll Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * Copyright (c) 1991, 1993, The Regents of the University of California. 6 * 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vm_map.c 8.3 (Berkeley) 1/12/94 37 * from: Id: uvm_map.c,v 1.1.2.27 1998/02/07 01:16:54 chs Exp 38 * 39 * 40 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 41 * All rights reserved. 42 * 43 * Permission to use, copy, modify and distribute this software and 44 * its documentation is hereby granted, provided that both the copyright 45 * notice and this permission notice appear in all copies of the 46 * software, derivative works or modified versions, and any portions 47 * thereof, and that both notices appear in supporting documentation. 48 * 49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 50 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 52 * 53 * Carnegie Mellon requests users of this software to return to 54 * 55 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 56 * School of Computer Science 57 * Carnegie Mellon University 58 * Pittsburgh PA 15213-3890 59 * 60 * any improvements or extensions that they make and grant Carnegie the 61 * rights to redistribute these changes. 62 */ 63 64 /* 65 * uvm_map.c: uvm map operations 66 */ 67 68 #include <sys/cdefs.h> 69 __KERNEL_RCSID(0, "$NetBSD: uvm_map.c,v 1.391 2021/11/25 09:40:45 skrll Exp $"); 70 71 #include "opt_ddb.h" 72 #include "opt_pax.h" 73 #include "opt_uvmhist.h" 74 #include "opt_uvm.h" 75 #include "opt_sysv.h" 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/mman.h> 80 #include <sys/proc.h> 81 #include <sys/pool.h> 82 #include <sys/kernel.h> 83 #include <sys/mount.h> 84 #include <sys/pax.h> 85 #include <sys/vnode.h> 86 #include <sys/filedesc.h> 87 #include <sys/lockdebug.h> 88 #include <sys/atomic.h> 89 #include <sys/sysctl.h> 90 #ifndef __USER_VA0_IS_SAFE 91 #include <sys/kauth.h> 92 #include "opt_user_va0_disable_default.h" 93 #endif 94 95 #include <sys/shm.h> 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 #ifdef UVMHIST 105 #ifndef UVMHIST_MAPHIST_SIZE 106 #define UVMHIST_MAPHIST_SIZE 100 107 #endif 108 static struct kern_history_ent maphistbuf[UVMHIST_MAPHIST_SIZE]; 109 UVMHIST_DEFINE(maphist) = UVMHIST_INITIALIZER(maphist, maphistbuf); 110 #endif 111 112 #if !defined(UVMMAP_COUNTERS) 113 114 #define UVMMAP_EVCNT_DEFINE(name) /* nothing */ 115 #define UVMMAP_EVCNT_INCR(ev) /* nothing */ 116 #define UVMMAP_EVCNT_DECR(ev) /* nothing */ 117 118 #else /* defined(UVMMAP_NOCOUNTERS) */ 119 120 #include <sys/evcnt.h> 121 #define UVMMAP_EVCNT_DEFINE(name) \ 122 struct evcnt uvmmap_evcnt_##name = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, \ 123 "uvmmap", #name); \ 124 EVCNT_ATTACH_STATIC(uvmmap_evcnt_##name); 125 #define UVMMAP_EVCNT_INCR(ev) uvmmap_evcnt_##ev.ev_count++ 126 #define UVMMAP_EVCNT_DECR(ev) uvmmap_evcnt_##ev.ev_count-- 127 128 #endif /* defined(UVMMAP_NOCOUNTERS) */ 129 130 UVMMAP_EVCNT_DEFINE(ubackmerge) 131 UVMMAP_EVCNT_DEFINE(uforwmerge) 132 UVMMAP_EVCNT_DEFINE(ubimerge) 133 UVMMAP_EVCNT_DEFINE(unomerge) 134 UVMMAP_EVCNT_DEFINE(kbackmerge) 135 UVMMAP_EVCNT_DEFINE(kforwmerge) 136 UVMMAP_EVCNT_DEFINE(kbimerge) 137 UVMMAP_EVCNT_DEFINE(knomerge) 138 UVMMAP_EVCNT_DEFINE(map_call) 139 UVMMAP_EVCNT_DEFINE(mlk_call) 140 UVMMAP_EVCNT_DEFINE(mlk_hint) 141 UVMMAP_EVCNT_DEFINE(mlk_tree) 142 UVMMAP_EVCNT_DEFINE(mlk_treeloop) 143 144 const char vmmapbsy[] = "vmmapbsy"; 145 146 /* 147 * cache for vmspace structures. 148 */ 149 150 static struct pool_cache uvm_vmspace_cache; 151 152 /* 153 * cache for dynamically-allocated map entries. 154 */ 155 156 static struct pool_cache uvm_map_entry_cache; 157 158 #ifdef PMAP_GROWKERNEL 159 /* 160 * This global represents the end of the kernel virtual address 161 * space. If we want to exceed this, we must grow the kernel 162 * virtual address space dynamically. 163 * 164 * Note, this variable is locked by kernel_map's lock. 165 */ 166 vaddr_t uvm_maxkaddr; 167 #endif 168 169 #ifndef __USER_VA0_IS_SAFE 170 #ifndef __USER_VA0_DISABLE_DEFAULT 171 #define __USER_VA0_DISABLE_DEFAULT 1 172 #endif 173 #ifdef USER_VA0_DISABLE_DEFAULT /* kernel config option overrides */ 174 #undef __USER_VA0_DISABLE_DEFAULT 175 #define __USER_VA0_DISABLE_DEFAULT USER_VA0_DISABLE_DEFAULT 176 #endif 177 int user_va0_disable = __USER_VA0_DISABLE_DEFAULT; 178 #endif 179 180 /* 181 * macros 182 */ 183 184 /* 185 * uvm_map_align_va: round down or up virtual address 186 */ 187 static __inline void 188 uvm_map_align_va(vaddr_t *vap, vsize_t align, int topdown) 189 { 190 191 KASSERT(powerof2(align)); 192 193 if (align != 0 && (*vap & (align - 1)) != 0) { 194 if (topdown) 195 *vap = rounddown2(*vap, align); 196 else 197 *vap = roundup2(*vap, align); 198 } 199 } 200 201 /* 202 * UVM_ET_ISCOMPATIBLE: check some requirements for map entry merging 203 */ 204 extern struct vm_map *pager_map; 205 206 #define UVM_ET_ISCOMPATIBLE(ent, type, uobj, meflags, \ 207 prot, maxprot, inh, adv, wire) \ 208 ((ent)->etype == (type) && \ 209 (((ent)->flags ^ (meflags)) & (UVM_MAP_NOMERGE)) == 0 && \ 210 (ent)->object.uvm_obj == (uobj) && \ 211 (ent)->protection == (prot) && \ 212 (ent)->max_protection == (maxprot) && \ 213 (ent)->inheritance == (inh) && \ 214 (ent)->advice == (adv) && \ 215 (ent)->wired_count == (wire)) 216 217 /* 218 * uvm_map_entry_link: insert entry into a map 219 * 220 * => map must be locked 221 */ 222 #define uvm_map_entry_link(map, after_where, entry) do { \ 223 uvm_mapent_check(entry); \ 224 (map)->nentries++; \ 225 (entry)->prev = (after_where); \ 226 (entry)->next = (after_where)->next; \ 227 (entry)->prev->next = (entry); \ 228 (entry)->next->prev = (entry); \ 229 uvm_rb_insert((map), (entry)); \ 230 } while (/*CONSTCOND*/ 0) 231 232 /* 233 * uvm_map_entry_unlink: remove entry from a map 234 * 235 * => map must be locked 236 */ 237 #define uvm_map_entry_unlink(map, entry) do { \ 238 KASSERT((entry) != (map)->first_free); \ 239 KASSERT((entry) != (map)->hint); \ 240 uvm_mapent_check(entry); \ 241 (map)->nentries--; \ 242 (entry)->next->prev = (entry)->prev; \ 243 (entry)->prev->next = (entry)->next; \ 244 uvm_rb_remove((map), (entry)); \ 245 } while (/*CONSTCOND*/ 0) 246 247 /* 248 * SAVE_HINT: saves the specified entry as the hint for future lookups. 249 * 250 * => map need not be locked. 251 */ 252 #define SAVE_HINT(map, check, value) do { \ 253 if ((map)->hint == (check)) \ 254 (map)->hint = (value); \ 255 } while (/*CONSTCOND*/ 0) 256 257 /* 258 * clear_hints: ensure that hints don't point to the entry. 259 * 260 * => map must be write-locked. 261 */ 262 static void 263 clear_hints(struct vm_map *map, struct vm_map_entry *ent) 264 { 265 266 SAVE_HINT(map, ent, ent->prev); 267 if (map->first_free == ent) { 268 map->first_free = ent->prev; 269 } 270 } 271 272 /* 273 * VM_MAP_RANGE_CHECK: check and correct range 274 * 275 * => map must at least be read locked 276 */ 277 278 #define VM_MAP_RANGE_CHECK(map, start, end) do { \ 279 if (start < vm_map_min(map)) \ 280 start = vm_map_min(map); \ 281 if (end > vm_map_max(map)) \ 282 end = vm_map_max(map); \ 283 if (start > end) \ 284 start = end; \ 285 } while (/*CONSTCOND*/ 0) 286 287 /* 288 * local prototypes 289 */ 290 291 static struct vm_map_entry * 292 uvm_mapent_alloc(struct vm_map *, int); 293 static void uvm_mapent_copy(struct vm_map_entry *, struct vm_map_entry *); 294 static void uvm_mapent_free(struct vm_map_entry *); 295 #if defined(DEBUG) 296 static void _uvm_mapent_check(const struct vm_map_entry *, int); 297 #define uvm_mapent_check(map) _uvm_mapent_check(map, __LINE__) 298 #else /* defined(DEBUG) */ 299 #define uvm_mapent_check(e) /* nothing */ 300 #endif /* defined(DEBUG) */ 301 302 static void uvm_map_entry_unwire(struct vm_map *, struct vm_map_entry *); 303 static void uvm_map_reference_amap(struct vm_map_entry *, int); 304 static int uvm_map_space_avail(vaddr_t *, vsize_t, voff_t, vsize_t, int, 305 int, struct vm_map_entry *); 306 static void uvm_map_unreference_amap(struct vm_map_entry *, int); 307 308 int _uvm_map_sanity(struct vm_map *); 309 int _uvm_tree_sanity(struct vm_map *); 310 static vsize_t uvm_rb_maxgap(const struct vm_map_entry *); 311 312 #define ROOT_ENTRY(map) ((struct vm_map_entry *)(map)->rb_tree.rbt_root) 313 #define LEFT_ENTRY(entry) ((struct vm_map_entry *)(entry)->rb_node.rb_left) 314 #define RIGHT_ENTRY(entry) ((struct vm_map_entry *)(entry)->rb_node.rb_right) 315 #define PARENT_ENTRY(map, entry) \ 316 (ROOT_ENTRY(map) == (entry) \ 317 ? NULL : (struct vm_map_entry *)RB_FATHER(&(entry)->rb_node)) 318 319 /* 320 * These get filled in if/when SYSVSHM shared memory code is loaded 321 * 322 * We do this with function pointers rather the #ifdef SYSVSHM so the 323 * SYSVSHM code can be loaded and unloaded 324 */ 325 void (*uvm_shmexit)(struct vmspace *) = NULL; 326 void (*uvm_shmfork)(struct vmspace *, struct vmspace *) = NULL; 327 328 static int 329 uvm_map_compare_nodes(void *ctx, const void *nparent, const void *nkey) 330 { 331 const struct vm_map_entry *eparent = nparent; 332 const struct vm_map_entry *ekey = nkey; 333 334 KASSERT(eparent->start < ekey->start || eparent->start >= ekey->end); 335 KASSERT(ekey->start < eparent->start || ekey->start >= eparent->end); 336 337 if (eparent->start < ekey->start) 338 return -1; 339 if (eparent->end >= ekey->start) 340 return 1; 341 return 0; 342 } 343 344 static int 345 uvm_map_compare_key(void *ctx, const void *nparent, const void *vkey) 346 { 347 const struct vm_map_entry *eparent = nparent; 348 const vaddr_t va = *(const vaddr_t *) vkey; 349 350 if (eparent->start < va) 351 return -1; 352 if (eparent->end >= va) 353 return 1; 354 return 0; 355 } 356 357 static const rb_tree_ops_t uvm_map_tree_ops = { 358 .rbto_compare_nodes = uvm_map_compare_nodes, 359 .rbto_compare_key = uvm_map_compare_key, 360 .rbto_node_offset = offsetof(struct vm_map_entry, rb_node), 361 .rbto_context = NULL 362 }; 363 364 /* 365 * uvm_rb_gap: return the gap size between our entry and next entry. 366 */ 367 static inline vsize_t 368 uvm_rb_gap(const struct vm_map_entry *entry) 369 { 370 371 KASSERT(entry->next != NULL); 372 return entry->next->start - entry->end; 373 } 374 375 static vsize_t 376 uvm_rb_maxgap(const struct vm_map_entry *entry) 377 { 378 struct vm_map_entry *child; 379 vsize_t maxgap = entry->gap; 380 381 /* 382 * We need maxgap to be the largest gap of us or any of our 383 * descendents. Since each of our children's maxgap is the 384 * cached value of their largest gap of themselves or their 385 * descendents, we can just use that value and avoid recursing 386 * down the tree to calculate it. 387 */ 388 if ((child = LEFT_ENTRY(entry)) != NULL && maxgap < child->maxgap) 389 maxgap = child->maxgap; 390 391 if ((child = RIGHT_ENTRY(entry)) != NULL && maxgap < child->maxgap) 392 maxgap = child->maxgap; 393 394 return maxgap; 395 } 396 397 static void 398 uvm_rb_fixup(struct vm_map *map, struct vm_map_entry *entry) 399 { 400 struct vm_map_entry *parent; 401 402 KASSERT(entry->gap == uvm_rb_gap(entry)); 403 entry->maxgap = uvm_rb_maxgap(entry); 404 405 while ((parent = PARENT_ENTRY(map, entry)) != NULL) { 406 struct vm_map_entry *brother; 407 vsize_t maxgap = parent->gap; 408 unsigned int which; 409 410 KDASSERT(parent->gap == uvm_rb_gap(parent)); 411 if (maxgap < entry->maxgap) 412 maxgap = entry->maxgap; 413 /* 414 * Since we work towards the root, we know entry's maxgap 415 * value is OK, but its brothers may now be out-of-date due 416 * to rebalancing. So refresh it. 417 */ 418 which = RB_POSITION(&entry->rb_node) ^ RB_DIR_OTHER; 419 brother = (struct vm_map_entry *)parent->rb_node.rb_nodes[which]; 420 if (brother != NULL) { 421 KDASSERT(brother->gap == uvm_rb_gap(brother)); 422 brother->maxgap = uvm_rb_maxgap(brother); 423 if (maxgap < brother->maxgap) 424 maxgap = brother->maxgap; 425 } 426 427 parent->maxgap = maxgap; 428 entry = parent; 429 } 430 } 431 432 static void 433 uvm_rb_insert(struct vm_map *map, struct vm_map_entry *entry) 434 { 435 struct vm_map_entry *ret __diagused; 436 437 entry->gap = entry->maxgap = uvm_rb_gap(entry); 438 if (entry->prev != &map->header) 439 entry->prev->gap = uvm_rb_gap(entry->prev); 440 441 ret = rb_tree_insert_node(&map->rb_tree, entry); 442 KASSERTMSG(ret == entry, 443 "uvm_rb_insert: map %p: duplicate entry %p", map, ret); 444 445 /* 446 * If the previous entry is not our immediate left child, then it's an 447 * ancestor and will be fixed up on the way to the root. We don't 448 * have to check entry->prev against &map->header since &map->header 449 * will never be in the tree. 450 */ 451 uvm_rb_fixup(map, 452 LEFT_ENTRY(entry) == entry->prev ? entry->prev : entry); 453 } 454 455 static void 456 uvm_rb_remove(struct vm_map *map, struct vm_map_entry *entry) 457 { 458 struct vm_map_entry *prev_parent = NULL, *next_parent = NULL; 459 460 /* 461 * If we are removing an interior node, then an adjacent node will 462 * be used to replace its position in the tree. Therefore we will 463 * need to fixup the tree starting at the parent of the replacement 464 * node. So record their parents for later use. 465 */ 466 if (entry->prev != &map->header) 467 prev_parent = PARENT_ENTRY(map, entry->prev); 468 if (entry->next != &map->header) 469 next_parent = PARENT_ENTRY(map, entry->next); 470 471 rb_tree_remove_node(&map->rb_tree, entry); 472 473 /* 474 * If the previous node has a new parent, fixup the tree starting 475 * at the previous node's old parent. 476 */ 477 if (entry->prev != &map->header) { 478 /* 479 * Update the previous entry's gap due to our absence. 480 */ 481 entry->prev->gap = uvm_rb_gap(entry->prev); 482 uvm_rb_fixup(map, entry->prev); 483 if (prev_parent != NULL 484 && prev_parent != entry 485 && prev_parent != PARENT_ENTRY(map, entry->prev)) 486 uvm_rb_fixup(map, prev_parent); 487 } 488 489 /* 490 * If the next node has a new parent, fixup the tree starting 491 * at the next node's old parent. 492 */ 493 if (entry->next != &map->header) { 494 uvm_rb_fixup(map, entry->next); 495 if (next_parent != NULL 496 && next_parent != entry 497 && next_parent != PARENT_ENTRY(map, entry->next)) 498 uvm_rb_fixup(map, next_parent); 499 } 500 } 501 502 #if defined(DEBUG) 503 int uvm_debug_check_map = 0; 504 int uvm_debug_check_rbtree = 0; 505 #define uvm_map_check(map, name) \ 506 _uvm_map_check((map), (name), __FILE__, __LINE__) 507 static void 508 _uvm_map_check(struct vm_map *map, const char *name, 509 const char *file, int line) 510 { 511 512 if ((uvm_debug_check_map && _uvm_map_sanity(map)) || 513 (uvm_debug_check_rbtree && _uvm_tree_sanity(map))) { 514 panic("uvm_map_check failed: \"%s\" map=%p (%s:%d)", 515 name, map, file, line); 516 } 517 } 518 #else /* defined(DEBUG) */ 519 #define uvm_map_check(map, name) /* nothing */ 520 #endif /* defined(DEBUG) */ 521 522 #if defined(DEBUG) || defined(DDB) 523 int 524 _uvm_map_sanity(struct vm_map *map) 525 { 526 bool first_free_found = false; 527 bool hint_found = false; 528 const struct vm_map_entry *e; 529 struct vm_map_entry *hint = map->hint; 530 531 e = &map->header; 532 for (;;) { 533 if (map->first_free == e) { 534 first_free_found = true; 535 } else if (!first_free_found && e->next->start > e->end) { 536 printf("first_free %p should be %p\n", 537 map->first_free, e); 538 return -1; 539 } 540 if (hint == e) { 541 hint_found = true; 542 } 543 544 e = e->next; 545 if (e == &map->header) { 546 break; 547 } 548 } 549 if (!first_free_found) { 550 printf("stale first_free\n"); 551 return -1; 552 } 553 if (!hint_found) { 554 printf("stale hint\n"); 555 return -1; 556 } 557 return 0; 558 } 559 560 int 561 _uvm_tree_sanity(struct vm_map *map) 562 { 563 struct vm_map_entry *tmp, *trtmp; 564 int n = 0, i = 1; 565 566 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 567 if (tmp->gap != uvm_rb_gap(tmp)) { 568 printf("%d/%d gap %#lx != %#lx %s\n", 569 n + 1, map->nentries, 570 (ulong)tmp->gap, (ulong)uvm_rb_gap(tmp), 571 tmp->next == &map->header ? "(last)" : ""); 572 goto error; 573 } 574 /* 575 * If any entries are out of order, tmp->gap will be unsigned 576 * and will likely exceed the size of the map. 577 */ 578 if (tmp->gap >= vm_map_max(map) - vm_map_min(map)) { 579 printf("too large gap %zu\n", (size_t)tmp->gap); 580 goto error; 581 } 582 n++; 583 } 584 585 if (n != map->nentries) { 586 printf("nentries: %d vs %d\n", n, map->nentries); 587 goto error; 588 } 589 590 trtmp = NULL; 591 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 592 if (tmp->maxgap != uvm_rb_maxgap(tmp)) { 593 printf("maxgap %#lx != %#lx\n", 594 (ulong)tmp->maxgap, 595 (ulong)uvm_rb_maxgap(tmp)); 596 goto error; 597 } 598 if (trtmp != NULL && trtmp->start >= tmp->start) { 599 printf("corrupt: 0x%"PRIxVADDR"x >= 0x%"PRIxVADDR"x\n", 600 trtmp->start, tmp->start); 601 goto error; 602 } 603 604 trtmp = tmp; 605 } 606 607 for (tmp = map->header.next; tmp != &map->header; 608 tmp = tmp->next, i++) { 609 trtmp = rb_tree_iterate(&map->rb_tree, tmp, RB_DIR_LEFT); 610 if (trtmp == NULL) 611 trtmp = &map->header; 612 if (tmp->prev != trtmp) { 613 printf("lookup: %d: %p->prev=%p: %p\n", 614 i, tmp, tmp->prev, trtmp); 615 goto error; 616 } 617 trtmp = rb_tree_iterate(&map->rb_tree, tmp, RB_DIR_RIGHT); 618 if (trtmp == NULL) 619 trtmp = &map->header; 620 if (tmp->next != trtmp) { 621 printf("lookup: %d: %p->next=%p: %p\n", 622 i, tmp, tmp->next, trtmp); 623 goto error; 624 } 625 trtmp = rb_tree_find_node(&map->rb_tree, &tmp->start); 626 if (trtmp != tmp) { 627 printf("lookup: %d: %p - %p: %p\n", i, tmp, trtmp, 628 PARENT_ENTRY(map, tmp)); 629 goto error; 630 } 631 } 632 633 return (0); 634 error: 635 return (-1); 636 } 637 #endif /* defined(DEBUG) || defined(DDB) */ 638 639 /* 640 * vm_map_lock: acquire an exclusive (write) lock on a map. 641 * 642 * => The locking protocol provides for guaranteed upgrade from shared -> 643 * exclusive by whichever thread currently has the map marked busy. 644 * See "LOCKING PROTOCOL NOTES" in uvm_map.h. This is horrible; among 645 * other problems, it defeats any fairness guarantees provided by RW 646 * locks. 647 */ 648 649 void 650 vm_map_lock(struct vm_map *map) 651 { 652 653 for (;;) { 654 rw_enter(&map->lock, RW_WRITER); 655 if (map->busy == NULL || map->busy == curlwp) { 656 break; 657 } 658 mutex_enter(&map->misc_lock); 659 rw_exit(&map->lock); 660 if (map->busy != NULL) { 661 cv_wait(&map->cv, &map->misc_lock); 662 } 663 mutex_exit(&map->misc_lock); 664 } 665 map->timestamp++; 666 } 667 668 /* 669 * vm_map_lock_try: try to lock a map, failing if it is already locked. 670 */ 671 672 bool 673 vm_map_lock_try(struct vm_map *map) 674 { 675 676 if (!rw_tryenter(&map->lock, RW_WRITER)) { 677 return false; 678 } 679 if (map->busy != NULL) { 680 rw_exit(&map->lock); 681 return false; 682 } 683 map->timestamp++; 684 return true; 685 } 686 687 /* 688 * vm_map_unlock: release an exclusive lock on a map. 689 */ 690 691 void 692 vm_map_unlock(struct vm_map *map) 693 { 694 695 KASSERT(rw_write_held(&map->lock)); 696 KASSERT(map->busy == NULL || map->busy == curlwp); 697 rw_exit(&map->lock); 698 } 699 700 /* 701 * vm_map_unbusy: mark the map as unbusy, and wake any waiters that 702 * want an exclusive lock. 703 */ 704 705 void 706 vm_map_unbusy(struct vm_map *map) 707 { 708 709 KASSERT(map->busy == curlwp); 710 711 /* 712 * Safe to clear 'busy' and 'waiters' with only a read lock held: 713 * 714 * o they can only be set with a write lock held 715 * o writers are blocked out with a read or write hold 716 * o at any time, only one thread owns the set of values 717 */ 718 mutex_enter(&map->misc_lock); 719 map->busy = NULL; 720 cv_broadcast(&map->cv); 721 mutex_exit(&map->misc_lock); 722 } 723 724 /* 725 * vm_map_lock_read: acquire a shared (read) lock on a map. 726 */ 727 728 void 729 vm_map_lock_read(struct vm_map *map) 730 { 731 732 rw_enter(&map->lock, RW_READER); 733 } 734 735 /* 736 * vm_map_unlock_read: release a shared lock on a map. 737 */ 738 739 void 740 vm_map_unlock_read(struct vm_map *map) 741 { 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 return rw_write_held(&map->lock); 774 } 775 776 /* 777 * uvm_mapent_alloc: allocate a map entry 778 */ 779 780 static struct vm_map_entry * 781 uvm_mapent_alloc(struct vm_map *map, int flags) 782 { 783 struct vm_map_entry *me; 784 int pflags = (flags & UVM_FLAG_NOWAIT) ? PR_NOWAIT : PR_WAITOK; 785 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 786 787 me = pool_cache_get(&uvm_map_entry_cache, pflags); 788 if (__predict_false(me == NULL)) { 789 return NULL; 790 } 791 me->flags = 0; 792 793 UVMHIST_LOG(maphist, "<- new entry=%#jx [kentry=%jd]", (uintptr_t)me, 794 (map == kernel_map), 0, 0); 795 return me; 796 } 797 798 /* 799 * uvm_mapent_free: free map entry 800 */ 801 802 static void 803 uvm_mapent_free(struct vm_map_entry *me) 804 { 805 UVMHIST_FUNC(__func__); 806 UVMHIST_CALLARGS(maphist,"<- freeing map entry=%#jx [flags=%#jx]", 807 (uintptr_t)me, me->flags, 0, 0); 808 pool_cache_put(&uvm_map_entry_cache, me); 809 } 810 811 /* 812 * uvm_mapent_copy: copy a map entry, preserving flags 813 */ 814 815 static inline void 816 uvm_mapent_copy(struct vm_map_entry *src, struct vm_map_entry *dst) 817 { 818 819 memcpy(dst, src, sizeof(*dst)); 820 dst->flags = 0; 821 } 822 823 #if defined(DEBUG) 824 static void 825 _uvm_mapent_check(const struct vm_map_entry *entry, int line) 826 { 827 828 if (entry->start >= entry->end) { 829 goto bad; 830 } 831 if (UVM_ET_ISOBJ(entry)) { 832 if (entry->object.uvm_obj == NULL) { 833 goto bad; 834 } 835 } else if (UVM_ET_ISSUBMAP(entry)) { 836 if (entry->object.sub_map == NULL) { 837 goto bad; 838 } 839 } else { 840 if (entry->object.uvm_obj != NULL || 841 entry->object.sub_map != NULL) { 842 goto bad; 843 } 844 } 845 if (!UVM_ET_ISOBJ(entry)) { 846 if (entry->offset != 0) { 847 goto bad; 848 } 849 } 850 851 return; 852 853 bad: 854 panic("%s: bad entry %p, line %d", __func__, entry, line); 855 } 856 #endif /* defined(DEBUG) */ 857 858 /* 859 * uvm_map_entry_unwire: unwire a map entry 860 * 861 * => map should be locked by caller 862 */ 863 864 static inline void 865 uvm_map_entry_unwire(struct vm_map *map, struct vm_map_entry *entry) 866 { 867 868 entry->wired_count = 0; 869 uvm_fault_unwire_locked(map, entry->start, entry->end); 870 } 871 872 873 /* 874 * wrapper for calling amap_ref() 875 */ 876 static inline void 877 uvm_map_reference_amap(struct vm_map_entry *entry, int flags) 878 { 879 880 amap_ref(entry->aref.ar_amap, entry->aref.ar_pageoff, 881 (entry->end - entry->start) >> PAGE_SHIFT, flags); 882 } 883 884 885 /* 886 * wrapper for calling amap_unref() 887 */ 888 static inline void 889 uvm_map_unreference_amap(struct vm_map_entry *entry, int flags) 890 { 891 892 amap_unref(entry->aref.ar_amap, entry->aref.ar_pageoff, 893 (entry->end - entry->start) >> PAGE_SHIFT, flags); 894 } 895 896 897 /* 898 * uvm_map_init: init mapping system at boot time. 899 */ 900 901 void 902 uvm_map_init(void) 903 { 904 /* 905 * first, init logging system. 906 */ 907 908 UVMHIST_FUNC(__func__); 909 UVMHIST_LINK_STATIC(maphist); 910 UVMHIST_LINK_STATIC(pdhist); 911 UVMHIST_CALLED(maphist); 912 UVMHIST_LOG(maphist,"<starting uvm map system>", 0, 0, 0, 0); 913 914 /* 915 * initialize the global lock for kernel map entry. 916 */ 917 918 mutex_init(&uvm_kentry_lock, MUTEX_DRIVER, IPL_VM); 919 } 920 921 /* 922 * uvm_map_init_caches: init mapping system caches. 923 */ 924 void 925 uvm_map_init_caches(void) 926 { 927 /* 928 * initialize caches. 929 */ 930 931 pool_cache_bootstrap(&uvm_map_entry_cache, sizeof(struct vm_map_entry), 932 coherency_unit, 0, PR_LARGECACHE, "vmmpepl", NULL, IPL_NONE, NULL, 933 NULL, NULL); 934 pool_cache_bootstrap(&uvm_vmspace_cache, sizeof(struct vmspace), 935 0, 0, 0, "vmsppl", NULL, IPL_NONE, NULL, NULL, NULL); 936 } 937 938 /* 939 * clippers 940 */ 941 942 /* 943 * uvm_mapent_splitadj: adjust map entries for splitting, after uvm_mapent_copy. 944 */ 945 946 static void 947 uvm_mapent_splitadj(struct vm_map_entry *entry1, struct vm_map_entry *entry2, 948 vaddr_t splitat) 949 { 950 vaddr_t adj; 951 952 KASSERT(entry1->start < splitat); 953 KASSERT(splitat < entry1->end); 954 955 adj = splitat - entry1->start; 956 entry1->end = entry2->start = splitat; 957 958 if (entry1->aref.ar_amap) { 959 amap_splitref(&entry1->aref, &entry2->aref, adj); 960 } 961 if (UVM_ET_ISSUBMAP(entry1)) { 962 /* ... unlikely to happen, but play it safe */ 963 uvm_map_reference(entry1->object.sub_map); 964 } else if (UVM_ET_ISOBJ(entry1)) { 965 KASSERT(entry1->object.uvm_obj != NULL); /* suppress coverity */ 966 entry2->offset += adj; 967 if (entry1->object.uvm_obj->pgops && 968 entry1->object.uvm_obj->pgops->pgo_reference) 969 entry1->object.uvm_obj->pgops->pgo_reference( 970 entry1->object.uvm_obj); 971 } 972 } 973 974 /* 975 * uvm_map_clip_start: ensure that the entry begins at or after 976 * the starting address, if it doesn't we split the entry. 977 * 978 * => caller should use UVM_MAP_CLIP_START macro rather than calling 979 * this directly 980 * => map must be locked by caller 981 */ 982 983 void 984 uvm_map_clip_start(struct vm_map *map, struct vm_map_entry *entry, 985 vaddr_t start) 986 { 987 struct vm_map_entry *new_entry; 988 989 /* uvm_map_simplify_entry(map, entry); */ /* XXX */ 990 991 uvm_map_check(map, "clip_start entry"); 992 uvm_mapent_check(entry); 993 994 /* 995 * Split off the front portion. note that we must insert the new 996 * entry BEFORE this one, so that this entry has the specified 997 * starting address. 998 */ 999 new_entry = uvm_mapent_alloc(map, 0); 1000 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1001 uvm_mapent_splitadj(new_entry, entry, start); 1002 uvm_map_entry_link(map, entry->prev, new_entry); 1003 1004 uvm_map_check(map, "clip_start leave"); 1005 } 1006 1007 /* 1008 * uvm_map_clip_end: ensure that the entry ends at or before 1009 * the ending address, if it does't we split the reference 1010 * 1011 * => caller should use UVM_MAP_CLIP_END macro rather than calling 1012 * this directly 1013 * => map must be locked by caller 1014 */ 1015 1016 void 1017 uvm_map_clip_end(struct vm_map *map, struct vm_map_entry *entry, vaddr_t end) 1018 { 1019 struct vm_map_entry *new_entry; 1020 1021 uvm_map_check(map, "clip_end entry"); 1022 uvm_mapent_check(entry); 1023 1024 /* 1025 * Create a new entry and insert it 1026 * AFTER the specified entry 1027 */ 1028 new_entry = uvm_mapent_alloc(map, 0); 1029 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1030 uvm_mapent_splitadj(entry, new_entry, end); 1031 uvm_map_entry_link(map, entry, new_entry); 1032 1033 uvm_map_check(map, "clip_end leave"); 1034 } 1035 1036 /* 1037 * M A P - m a i n e n t r y p o i n t 1038 */ 1039 /* 1040 * uvm_map: establish a valid mapping in a map 1041 * 1042 * => assume startp is page aligned. 1043 * => assume size is a multiple of PAGE_SIZE. 1044 * => assume sys_mmap provides enough of a "hint" to have us skip 1045 * over text/data/bss area. 1046 * => map must be unlocked (we will lock it) 1047 * => <uobj,uoffset> value meanings (4 cases): 1048 * [1] <NULL,uoffset> == uoffset is a hint for PMAP_PREFER 1049 * [2] <NULL,UVM_UNKNOWN_OFFSET> == don't PMAP_PREFER 1050 * [3] <uobj,uoffset> == normal mapping 1051 * [4] <uobj,UVM_UNKNOWN_OFFSET> == uvm_map finds offset based on VA 1052 * 1053 * case [4] is for kernel mappings where we don't know the offset until 1054 * we've found a virtual address. note that kernel object offsets are 1055 * always relative to vm_map_min(kernel_map). 1056 * 1057 * => if `align' is non-zero, we align the virtual address to the specified 1058 * alignment. 1059 * this is provided as a mechanism for large pages. 1060 * 1061 * => XXXCDC: need way to map in external amap? 1062 */ 1063 1064 int 1065 uvm_map(struct vm_map *map, vaddr_t *startp /* IN/OUT */, vsize_t size, 1066 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags) 1067 { 1068 struct uvm_map_args args; 1069 struct vm_map_entry *new_entry; 1070 int error; 1071 1072 KASSERT((size & PAGE_MASK) == 0); 1073 KASSERT((flags & UVM_FLAG_FIXED) == 0 || align == 0); 1074 1075 /* 1076 * for pager_map, allocate the new entry first to avoid sleeping 1077 * for memory while we have the map locked. 1078 */ 1079 1080 new_entry = NULL; 1081 if (map == pager_map) { 1082 new_entry = uvm_mapent_alloc(map, (flags & UVM_FLAG_NOWAIT)); 1083 if (__predict_false(new_entry == NULL)) 1084 return ENOMEM; 1085 } 1086 if (map == pager_map) 1087 flags |= UVM_FLAG_NOMERGE; 1088 1089 error = uvm_map_prepare(map, *startp, size, uobj, uoffset, align, 1090 flags, &args); 1091 if (!error) { 1092 error = uvm_map_enter(map, &args, new_entry); 1093 *startp = args.uma_start; 1094 } else if (new_entry) { 1095 uvm_mapent_free(new_entry); 1096 } 1097 1098 #if defined(DEBUG) 1099 if (!error && VM_MAP_IS_KERNEL(map) && (flags & UVM_FLAG_NOWAIT) == 0) { 1100 uvm_km_check_empty(map, *startp, *startp + size); 1101 } 1102 #endif /* defined(DEBUG) */ 1103 1104 return error; 1105 } 1106 1107 /* 1108 * uvm_map_prepare: 1109 * 1110 * called with map unlocked. 1111 * on success, returns the map locked. 1112 */ 1113 1114 int 1115 uvm_map_prepare(struct vm_map *map, vaddr_t start, vsize_t size, 1116 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags, 1117 struct uvm_map_args *args) 1118 { 1119 struct vm_map_entry *prev_entry; 1120 vm_prot_t prot = UVM_PROTECTION(flags); 1121 vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1122 1123 UVMHIST_FUNC(__func__); 1124 UVMHIST_CALLARGS(maphist, "(map=%#jx, start=%#jx, size=%jx, flags=%#jx)", 1125 (uintptr_t)map, start, size, flags); 1126 UVMHIST_LOG(maphist, " uobj/offset %#jx/%jd", (uintptr_t)uobj, 1127 uoffset,0,0); 1128 1129 /* 1130 * detect a popular device driver bug. 1131 */ 1132 1133 KASSERT(doing_shutdown || curlwp != NULL); 1134 1135 /* 1136 * zero-sized mapping doesn't make any sense. 1137 */ 1138 KASSERT(size > 0); 1139 1140 KASSERT((~flags & (UVM_FLAG_NOWAIT | UVM_FLAG_WAITVA)) != 0); 1141 1142 uvm_map_check(map, "map entry"); 1143 1144 /* 1145 * check sanity of protection code 1146 */ 1147 1148 if ((prot & maxprot) != prot) { 1149 UVMHIST_LOG(maphist, "<- prot. failure: prot=%#jx, max=%#jx", 1150 prot, maxprot,0,0); 1151 return EACCES; 1152 } 1153 1154 /* 1155 * figure out where to put new VM range 1156 */ 1157 retry: 1158 if (vm_map_lock_try(map) == false) { 1159 if ((flags & UVM_FLAG_TRYLOCK) != 0) { 1160 return EAGAIN; 1161 } 1162 vm_map_lock(map); /* could sleep here */ 1163 } 1164 if (flags & UVM_FLAG_UNMAP) { 1165 KASSERT(flags & UVM_FLAG_FIXED); 1166 KASSERT((flags & UVM_FLAG_NOWAIT) == 0); 1167 1168 /* 1169 * Set prev_entry to what it will need to be after any existing 1170 * entries are removed later in uvm_map_enter(). 1171 */ 1172 1173 if (uvm_map_lookup_entry(map, start, &prev_entry)) { 1174 if (start == prev_entry->start) 1175 prev_entry = prev_entry->prev; 1176 else 1177 UVM_MAP_CLIP_END(map, prev_entry, start); 1178 SAVE_HINT(map, map->hint, prev_entry); 1179 } 1180 } else { 1181 prev_entry = uvm_map_findspace(map, start, size, &start, 1182 uobj, uoffset, align, flags); 1183 } 1184 if (prev_entry == NULL) { 1185 unsigned int timestamp; 1186 1187 timestamp = map->timestamp; 1188 UVMHIST_LOG(maphist,"waiting va timestamp=%#jx", 1189 timestamp,0,0,0); 1190 map->flags |= VM_MAP_WANTVA; 1191 vm_map_unlock(map); 1192 1193 /* 1194 * try to reclaim kva and wait until someone does unmap. 1195 * fragile locking here, so we awaken every second to 1196 * recheck the condition. 1197 */ 1198 1199 mutex_enter(&map->misc_lock); 1200 while ((map->flags & VM_MAP_WANTVA) != 0 && 1201 map->timestamp == timestamp) { 1202 if ((flags & UVM_FLAG_WAITVA) == 0) { 1203 mutex_exit(&map->misc_lock); 1204 UVMHIST_LOG(maphist, 1205 "<- uvm_map_findspace failed!", 0,0,0,0); 1206 return ENOMEM; 1207 } else { 1208 cv_timedwait(&map->cv, &map->misc_lock, hz); 1209 } 1210 } 1211 mutex_exit(&map->misc_lock); 1212 goto retry; 1213 } 1214 1215 #ifdef PMAP_GROWKERNEL 1216 /* 1217 * If the kernel pmap can't map the requested space, 1218 * then allocate more resources for it. 1219 */ 1220 if (map == kernel_map && uvm_maxkaddr < (start + size)) 1221 uvm_maxkaddr = pmap_growkernel(start + size); 1222 #endif 1223 1224 UVMMAP_EVCNT_INCR(map_call); 1225 1226 /* 1227 * if uobj is null, then uoffset is either a VAC hint for PMAP_PREFER 1228 * [typically from uvm_map_reserve] or it is UVM_UNKNOWN_OFFSET. in 1229 * either case we want to zero it before storing it in the map entry 1230 * (because it looks strange and confusing when debugging...) 1231 * 1232 * if uobj is not null 1233 * if uoffset is not UVM_UNKNOWN_OFFSET then we have a normal mapping 1234 * and we do not need to change uoffset. 1235 * if uoffset is UVM_UNKNOWN_OFFSET then we need to find the offset 1236 * now (based on the starting address of the map). this case is 1237 * for kernel object mappings where we don't know the offset until 1238 * the virtual address is found (with uvm_map_findspace). the 1239 * offset is the distance we are from the start of the map. 1240 */ 1241 1242 if (uobj == NULL) { 1243 uoffset = 0; 1244 } else { 1245 if (uoffset == UVM_UNKNOWN_OFFSET) { 1246 KASSERT(UVM_OBJ_IS_KERN_OBJECT(uobj)); 1247 uoffset = start - vm_map_min(kernel_map); 1248 } 1249 } 1250 1251 args->uma_flags = flags; 1252 args->uma_prev = prev_entry; 1253 args->uma_start = start; 1254 args->uma_size = size; 1255 args->uma_uobj = uobj; 1256 args->uma_uoffset = uoffset; 1257 1258 UVMHIST_LOG(maphist, "<- done!", 0,0,0,0); 1259 return 0; 1260 } 1261 1262 /* 1263 * uvm_map_enter: 1264 * 1265 * called with map locked. 1266 * unlock the map before returning. 1267 */ 1268 1269 int 1270 uvm_map_enter(struct vm_map *map, const struct uvm_map_args *args, 1271 struct vm_map_entry *new_entry) 1272 { 1273 struct vm_map_entry *prev_entry = args->uma_prev; 1274 struct vm_map_entry *dead = NULL, *dead_entries = NULL; 1275 1276 const uvm_flag_t flags = args->uma_flags; 1277 const vm_prot_t prot = UVM_PROTECTION(flags); 1278 const vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1279 const vm_inherit_t inherit = UVM_INHERIT(flags); 1280 const int amapwaitflag = (flags & UVM_FLAG_NOWAIT) ? 1281 AMAP_EXTEND_NOWAIT : 0; 1282 const int advice = UVM_ADVICE(flags); 1283 1284 vaddr_t start = args->uma_start; 1285 vsize_t size = args->uma_size; 1286 struct uvm_object *uobj = args->uma_uobj; 1287 voff_t uoffset = args->uma_uoffset; 1288 1289 const int kmap = (vm_map_pmap(map) == pmap_kernel()); 1290 int merged = 0; 1291 int error; 1292 int newetype; 1293 1294 UVMHIST_FUNC(__func__); 1295 UVMHIST_CALLARGS(maphist, "(map=%#jx, start=%#jx, size=%ju, flags=%#jx)", 1296 (uintptr_t)map, start, size, flags); 1297 UVMHIST_LOG(maphist, " uobj/offset %#jx/%jd", (uintptr_t)uobj, 1298 uoffset,0,0); 1299 1300 KASSERT(map->hint == prev_entry); /* bimerge case assumes this */ 1301 KASSERT(vm_map_locked_p(map)); 1302 KASSERT((flags & (UVM_FLAG_NOWAIT | UVM_FLAG_UNMAP)) != 1303 (UVM_FLAG_NOWAIT | UVM_FLAG_UNMAP)); 1304 1305 if (uobj) 1306 newetype = UVM_ET_OBJ; 1307 else 1308 newetype = 0; 1309 1310 if (flags & UVM_FLAG_COPYONW) { 1311 newetype |= UVM_ET_COPYONWRITE; 1312 if ((flags & UVM_FLAG_OVERLAY) == 0) 1313 newetype |= UVM_ET_NEEDSCOPY; 1314 } 1315 1316 /* 1317 * For mappings with unmap, remove any old entries now. Adding the new 1318 * entry cannot fail because that can only happen if UVM_FLAG_NOWAIT 1319 * is set, and we do not support nowait and unmap together. 1320 */ 1321 1322 if (flags & UVM_FLAG_UNMAP) { 1323 KASSERT(flags & UVM_FLAG_FIXED); 1324 uvm_unmap_remove(map, start, start + size, &dead_entries, 0); 1325 #ifdef DEBUG 1326 struct vm_map_entry *tmp_entry __diagused; 1327 bool rv __diagused; 1328 1329 rv = uvm_map_lookup_entry(map, start, &tmp_entry); 1330 KASSERT(!rv); 1331 KASSERTMSG(prev_entry == tmp_entry, 1332 "args %p prev_entry %p tmp_entry %p", 1333 args, prev_entry, tmp_entry); 1334 #endif 1335 SAVE_HINT(map, map->hint, prev_entry); 1336 } 1337 1338 /* 1339 * try and insert in map by extending previous entry, if possible. 1340 * XXX: we don't try and pull back the next entry. might be useful 1341 * for a stack, but we are currently allocating our stack in advance. 1342 */ 1343 1344 if (flags & UVM_FLAG_NOMERGE) 1345 goto nomerge; 1346 1347 if (prev_entry->end == start && 1348 prev_entry != &map->header && 1349 UVM_ET_ISCOMPATIBLE(prev_entry, newetype, uobj, 0, 1350 prot, maxprot, inherit, advice, 0)) { 1351 1352 if (uobj && prev_entry->offset + 1353 (prev_entry->end - prev_entry->start) != uoffset) 1354 goto forwardmerge; 1355 1356 /* 1357 * can't extend a shared amap. note: no need to lock amap to 1358 * look at refs since we don't care about its exact value. 1359 * if it is one (i.e. we have only reference) it will stay there 1360 */ 1361 1362 if (prev_entry->aref.ar_amap && 1363 amap_refs(prev_entry->aref.ar_amap) != 1) { 1364 goto forwardmerge; 1365 } 1366 1367 if (prev_entry->aref.ar_amap) { 1368 error = amap_extend(prev_entry, size, 1369 amapwaitflag | AMAP_EXTEND_FORWARDS); 1370 if (error) 1371 goto nomerge; 1372 } 1373 1374 if (kmap) { 1375 UVMMAP_EVCNT_INCR(kbackmerge); 1376 } else { 1377 UVMMAP_EVCNT_INCR(ubackmerge); 1378 } 1379 UVMHIST_LOG(maphist," starting back merge", 0, 0, 0, 0); 1380 1381 /* 1382 * drop our reference to uobj since we are extending a reference 1383 * that we already have (the ref count can not drop to zero). 1384 */ 1385 1386 if (uobj && uobj->pgops->pgo_detach) 1387 uobj->pgops->pgo_detach(uobj); 1388 1389 /* 1390 * Now that we've merged the entries, note that we've grown 1391 * and our gap has shrunk. Then fix the tree. 1392 */ 1393 prev_entry->end += size; 1394 prev_entry->gap -= size; 1395 uvm_rb_fixup(map, prev_entry); 1396 1397 uvm_map_check(map, "map backmerged"); 1398 1399 UVMHIST_LOG(maphist,"<- done (via backmerge)!", 0, 0, 0, 0); 1400 merged++; 1401 } 1402 1403 forwardmerge: 1404 if (prev_entry->next->start == (start + size) && 1405 prev_entry->next != &map->header && 1406 UVM_ET_ISCOMPATIBLE(prev_entry->next, newetype, uobj, 0, 1407 prot, maxprot, inherit, advice, 0)) { 1408 1409 if (uobj && prev_entry->next->offset != uoffset + size) 1410 goto nomerge; 1411 1412 /* 1413 * can't extend a shared amap. note: no need to lock amap to 1414 * look at refs since we don't care about its exact value. 1415 * if it is one (i.e. we have only reference) it will stay there. 1416 * 1417 * note that we also can't merge two amaps, so if we 1418 * merged with the previous entry which has an amap, 1419 * and the next entry also has an amap, we give up. 1420 * 1421 * Interesting cases: 1422 * amap, new, amap -> give up second merge (single fwd extend) 1423 * amap, new, none -> double forward extend (extend again here) 1424 * none, new, amap -> double backward extend (done here) 1425 * uobj, new, amap -> single backward extend (done here) 1426 * 1427 * XXX should we attempt to deal with someone refilling 1428 * the deallocated region between two entries that are 1429 * backed by the same amap (ie, arefs is 2, "prev" and 1430 * "next" refer to it, and adding this allocation will 1431 * close the hole, thus restoring arefs to 1 and 1432 * deallocating the "next" vm_map_entry)? -- @@@ 1433 */ 1434 1435 if (prev_entry->next->aref.ar_amap && 1436 (amap_refs(prev_entry->next->aref.ar_amap) != 1 || 1437 (merged && prev_entry->aref.ar_amap))) { 1438 goto nomerge; 1439 } 1440 1441 if (merged) { 1442 /* 1443 * Try to extend the amap of the previous entry to 1444 * cover the next entry as well. If it doesn't work 1445 * just skip on, don't actually give up, since we've 1446 * already completed the back merge. 1447 */ 1448 if (prev_entry->aref.ar_amap) { 1449 if (amap_extend(prev_entry, 1450 prev_entry->next->end - 1451 prev_entry->next->start, 1452 amapwaitflag | AMAP_EXTEND_FORWARDS)) 1453 goto nomerge; 1454 } 1455 1456 /* 1457 * Try to extend the amap of the *next* entry 1458 * back to cover the new allocation *and* the 1459 * previous entry as well (the previous merge 1460 * didn't have an amap already otherwise we 1461 * wouldn't be checking here for an amap). If 1462 * it doesn't work just skip on, again, don't 1463 * actually give up, since we've already 1464 * completed the back merge. 1465 */ 1466 else if (prev_entry->next->aref.ar_amap) { 1467 if (amap_extend(prev_entry->next, 1468 prev_entry->end - 1469 prev_entry->start, 1470 amapwaitflag | AMAP_EXTEND_BACKWARDS)) 1471 goto nomerge; 1472 } 1473 } else { 1474 /* 1475 * Pull the next entry's amap backwards to cover this 1476 * new allocation. 1477 */ 1478 if (prev_entry->next->aref.ar_amap) { 1479 error = amap_extend(prev_entry->next, size, 1480 amapwaitflag | AMAP_EXTEND_BACKWARDS); 1481 if (error) 1482 goto nomerge; 1483 } 1484 } 1485 1486 if (merged) { 1487 if (kmap) { 1488 UVMMAP_EVCNT_DECR(kbackmerge); 1489 UVMMAP_EVCNT_INCR(kbimerge); 1490 } else { 1491 UVMMAP_EVCNT_DECR(ubackmerge); 1492 UVMMAP_EVCNT_INCR(ubimerge); 1493 } 1494 } else { 1495 if (kmap) { 1496 UVMMAP_EVCNT_INCR(kforwmerge); 1497 } else { 1498 UVMMAP_EVCNT_INCR(uforwmerge); 1499 } 1500 } 1501 UVMHIST_LOG(maphist," starting forward merge", 0, 0, 0, 0); 1502 1503 /* 1504 * drop our reference to uobj since we are extending a reference 1505 * that we already have (the ref count can not drop to zero). 1506 */ 1507 if (uobj && uobj->pgops->pgo_detach) 1508 uobj->pgops->pgo_detach(uobj); 1509 1510 if (merged) { 1511 dead = prev_entry->next; 1512 prev_entry->end = dead->end; 1513 uvm_map_entry_unlink(map, dead); 1514 if (dead->aref.ar_amap != NULL) { 1515 prev_entry->aref = dead->aref; 1516 dead->aref.ar_amap = NULL; 1517 } 1518 } else { 1519 prev_entry->next->start -= size; 1520 if (prev_entry != &map->header) { 1521 prev_entry->gap -= size; 1522 KASSERT(prev_entry->gap == uvm_rb_gap(prev_entry)); 1523 uvm_rb_fixup(map, prev_entry); 1524 } 1525 if (uobj) 1526 prev_entry->next->offset = uoffset; 1527 } 1528 1529 uvm_map_check(map, "map forwardmerged"); 1530 1531 UVMHIST_LOG(maphist,"<- done forwardmerge", 0, 0, 0, 0); 1532 merged++; 1533 } 1534 1535 nomerge: 1536 if (!merged) { 1537 UVMHIST_LOG(maphist," allocating new map entry", 0, 0, 0, 0); 1538 if (kmap) { 1539 UVMMAP_EVCNT_INCR(knomerge); 1540 } else { 1541 UVMMAP_EVCNT_INCR(unomerge); 1542 } 1543 1544 /* 1545 * allocate new entry and link it in. 1546 */ 1547 1548 if (new_entry == NULL) { 1549 new_entry = uvm_mapent_alloc(map, 1550 (flags & UVM_FLAG_NOWAIT)); 1551 if (__predict_false(new_entry == NULL)) { 1552 error = ENOMEM; 1553 goto done; 1554 } 1555 } 1556 new_entry->start = start; 1557 new_entry->end = new_entry->start + size; 1558 new_entry->object.uvm_obj = uobj; 1559 new_entry->offset = uoffset; 1560 1561 new_entry->etype = newetype; 1562 1563 if (flags & UVM_FLAG_NOMERGE) { 1564 new_entry->flags |= UVM_MAP_NOMERGE; 1565 } 1566 1567 new_entry->protection = prot; 1568 new_entry->max_protection = maxprot; 1569 new_entry->inheritance = inherit; 1570 new_entry->wired_count = 0; 1571 new_entry->advice = advice; 1572 if (flags & UVM_FLAG_OVERLAY) { 1573 1574 /* 1575 * to_add: for BSS we overallocate a little since we 1576 * are likely to extend 1577 */ 1578 1579 vaddr_t to_add = (flags & UVM_FLAG_AMAPPAD) ? 1580 UVM_AMAP_CHUNK << PAGE_SHIFT : 0; 1581 struct vm_amap *amap = amap_alloc(size, to_add, 1582 (flags & UVM_FLAG_NOWAIT)); 1583 if (__predict_false(amap == NULL)) { 1584 error = ENOMEM; 1585 goto done; 1586 } 1587 new_entry->aref.ar_pageoff = 0; 1588 new_entry->aref.ar_amap = amap; 1589 } else { 1590 new_entry->aref.ar_pageoff = 0; 1591 new_entry->aref.ar_amap = NULL; 1592 } 1593 uvm_map_entry_link(map, prev_entry, new_entry); 1594 1595 /* 1596 * Update the free space hint 1597 */ 1598 1599 if ((map->first_free == prev_entry) && 1600 (prev_entry->end >= new_entry->start)) 1601 map->first_free = new_entry; 1602 1603 new_entry = NULL; 1604 } 1605 1606 map->size += size; 1607 1608 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 1609 1610 error = 0; 1611 1612 done: 1613 vm_map_unlock(map); 1614 1615 if (new_entry) { 1616 uvm_mapent_free(new_entry); 1617 } 1618 if (dead) { 1619 KDASSERT(merged); 1620 uvm_mapent_free(dead); 1621 } 1622 if (dead_entries) 1623 uvm_unmap_detach(dead_entries, 0); 1624 1625 return error; 1626 } 1627 1628 /* 1629 * uvm_map_lookup_entry_bytree: lookup an entry in tree 1630 */ 1631 1632 static inline bool 1633 uvm_map_lookup_entry_bytree(struct vm_map *map, vaddr_t address, 1634 struct vm_map_entry **entry /* OUT */) 1635 { 1636 struct vm_map_entry *prev = &map->header; 1637 struct vm_map_entry *cur = ROOT_ENTRY(map); 1638 1639 while (cur) { 1640 UVMMAP_EVCNT_INCR(mlk_treeloop); 1641 if (address >= cur->start) { 1642 if (address < cur->end) { 1643 *entry = cur; 1644 return true; 1645 } 1646 prev = cur; 1647 cur = RIGHT_ENTRY(cur); 1648 } else 1649 cur = LEFT_ENTRY(cur); 1650 } 1651 *entry = prev; 1652 return false; 1653 } 1654 1655 /* 1656 * uvm_map_lookup_entry: find map entry at or before an address 1657 * 1658 * => map must at least be read-locked by caller 1659 * => entry is returned in "entry" 1660 * => return value is true if address is in the returned entry 1661 */ 1662 1663 bool 1664 uvm_map_lookup_entry(struct vm_map *map, vaddr_t address, 1665 struct vm_map_entry **entry /* OUT */) 1666 { 1667 struct vm_map_entry *cur; 1668 UVMHIST_FUNC(__func__); 1669 UVMHIST_CALLARGS(maphist,"(map=%#jx,addr=%#jx,ent=%#jx)", 1670 (uintptr_t)map, address, (uintptr_t)entry, 0); 1671 1672 /* 1673 * make a quick check to see if we are already looking at 1674 * the entry we want (which is usually the case). note also 1675 * that we don't need to save the hint here... it is the 1676 * same hint (unless we are at the header, in which case the 1677 * hint didn't buy us anything anyway). 1678 */ 1679 1680 cur = map->hint; 1681 UVMMAP_EVCNT_INCR(mlk_call); 1682 if (cur != &map->header && 1683 address >= cur->start && cur->end > address) { 1684 UVMMAP_EVCNT_INCR(mlk_hint); 1685 *entry = cur; 1686 UVMHIST_LOG(maphist,"<- got it via hint (%#jx)", 1687 (uintptr_t)cur, 0, 0, 0); 1688 uvm_mapent_check(*entry); 1689 return (true); 1690 } 1691 uvm_map_check(map, __func__); 1692 1693 /* 1694 * lookup in the tree. 1695 */ 1696 1697 UVMMAP_EVCNT_INCR(mlk_tree); 1698 if (__predict_true(uvm_map_lookup_entry_bytree(map, address, entry))) { 1699 SAVE_HINT(map, map->hint, *entry); 1700 UVMHIST_LOG(maphist,"<- search got it (%#jx)", 1701 (uintptr_t)cur, 0, 0, 0); 1702 KDASSERT((*entry)->start <= address); 1703 KDASSERT(address < (*entry)->end); 1704 uvm_mapent_check(*entry); 1705 return (true); 1706 } 1707 1708 SAVE_HINT(map, map->hint, *entry); 1709 UVMHIST_LOG(maphist,"<- failed!",0,0,0,0); 1710 KDASSERT((*entry) == &map->header || (*entry)->end <= address); 1711 KDASSERT((*entry)->next == &map->header || 1712 address < (*entry)->next->start); 1713 return (false); 1714 } 1715 1716 /* 1717 * See if the range between start and start + length fits in the gap 1718 * entry->next->start and entry->end. Returns 1 if fits, 0 if doesn't 1719 * fit, and -1 address wraps around. 1720 */ 1721 static int 1722 uvm_map_space_avail(vaddr_t *start, vsize_t length, voff_t uoffset, 1723 vsize_t align, int flags, int topdown, struct vm_map_entry *entry) 1724 { 1725 vaddr_t end; 1726 1727 #ifdef PMAP_PREFER 1728 /* 1729 * push start address forward as needed to avoid VAC alias problems. 1730 * we only do this if a valid offset is specified. 1731 */ 1732 1733 if (uoffset != UVM_UNKNOWN_OFFSET) 1734 PMAP_PREFER(uoffset, start, length, topdown); 1735 #endif 1736 if ((flags & UVM_FLAG_COLORMATCH) != 0) { 1737 KASSERT(align < uvmexp.ncolors); 1738 if (uvmexp.ncolors > 1) { 1739 const u_int colormask = uvmexp.colormask; 1740 const u_int colorsize = colormask + 1; 1741 vaddr_t hint = atop(*start); 1742 const u_int color = hint & colormask; 1743 if (color != align) { 1744 hint -= color; /* adjust to color boundary */ 1745 KASSERT((hint & colormask) == 0); 1746 if (topdown) { 1747 if (align > color) 1748 hint -= colorsize; 1749 } else { 1750 if (align < color) 1751 hint += colorsize; 1752 } 1753 *start = ptoa(hint + align); /* adjust to color */ 1754 } 1755 } 1756 } else { 1757 KASSERT(powerof2(align)); 1758 uvm_map_align_va(start, align, topdown); 1759 /* 1760 * XXX Should we PMAP_PREFER() here again? 1761 * eh...i think we're okay 1762 */ 1763 } 1764 1765 /* 1766 * Find the end of the proposed new region. Be sure we didn't 1767 * wrap around the address; if so, we lose. Otherwise, if the 1768 * proposed new region fits before the next entry, we win. 1769 */ 1770 1771 end = *start + length; 1772 if (end < *start) 1773 return (-1); 1774 1775 if (entry->next->start >= end && *start >= entry->end) 1776 return (1); 1777 1778 return (0); 1779 } 1780 1781 /* 1782 * uvm_map_findspace: find "length" sized space in "map". 1783 * 1784 * => "hint" is a hint about where we want it, unless UVM_FLAG_FIXED is 1785 * set in "flags" (in which case we insist on using "hint"). 1786 * => "result" is VA returned 1787 * => uobj/uoffset are to be used to handle VAC alignment, if required 1788 * => if "align" is non-zero, we attempt to align to that value. 1789 * => caller must at least have read-locked map 1790 * => returns NULL on failure, or pointer to prev. map entry if success 1791 * => note this is a cross between the old vm_map_findspace and vm_map_find 1792 */ 1793 1794 struct vm_map_entry * 1795 uvm_map_findspace(struct vm_map *map, vaddr_t hint, vsize_t length, 1796 vaddr_t *result /* OUT */, struct uvm_object *uobj, voff_t uoffset, 1797 vsize_t align, int flags) 1798 { 1799 struct vm_map_entry *entry; 1800 struct vm_map_entry *child, *prev, *tmp; 1801 vaddr_t orig_hint __diagused; 1802 const int topdown = map->flags & VM_MAP_TOPDOWN; 1803 UVMHIST_FUNC(__func__); 1804 UVMHIST_CALLARGS(maphist, "(map=%#jx, hint=%#jx, len=%ju, flags=%#jx...", 1805 (uintptr_t)map, hint, length, flags); 1806 UVMHIST_LOG(maphist, " uobj=%#jx, uoffset=%#jx, align=%#jx)", 1807 (uintptr_t)uobj, uoffset, align, 0); 1808 1809 KASSERT((flags & UVM_FLAG_COLORMATCH) != 0 || powerof2(align)); 1810 KASSERT((flags & UVM_FLAG_COLORMATCH) == 0 || align < uvmexp.ncolors); 1811 KASSERT((flags & UVM_FLAG_FIXED) == 0 || align == 0); 1812 1813 uvm_map_check(map, "map_findspace entry"); 1814 1815 /* 1816 * remember the original hint. if we are aligning, then we 1817 * may have to try again with no alignment constraint if 1818 * we fail the first time. 1819 */ 1820 1821 orig_hint = hint; 1822 if (hint < vm_map_min(map)) { /* check ranges ... */ 1823 if (flags & UVM_FLAG_FIXED) { 1824 UVMHIST_LOG(maphist,"<- VA below map range",0,0,0,0); 1825 return (NULL); 1826 } 1827 hint = vm_map_min(map); 1828 } 1829 if (hint > vm_map_max(map)) { 1830 UVMHIST_LOG(maphist,"<- VA %#jx > range [%#jx->%#jx]", 1831 hint, vm_map_min(map), vm_map_max(map), 0); 1832 return (NULL); 1833 } 1834 1835 UVMHIST_LOG(maphist,"<- VA %#jx vs range [%#jx->%#jx]", 1836 hint, vm_map_min(map), vm_map_max(map), 0); 1837 1838 /* 1839 * hint may not be aligned properly; we need round up or down it 1840 * before proceeding further. 1841 */ 1842 if ((flags & UVM_FLAG_COLORMATCH) == 0) 1843 uvm_map_align_va(&hint, align, topdown); 1844 1845 UVMHIST_LOG(maphist,"<- VA %#jx vs range [%#jx->%#jx]", 1846 hint, vm_map_min(map), vm_map_max(map), 0); 1847 /* 1848 * Look for the first possible address; if there's already 1849 * something at this address, we have to start after it. 1850 */ 1851 1852 /* 1853 * @@@: there are four, no, eight cases to consider. 1854 * 1855 * 0: found, fixed, bottom up -> fail 1856 * 1: found, fixed, top down -> fail 1857 * 2: found, not fixed, bottom up -> start after entry->end, 1858 * loop up 1859 * 3: found, not fixed, top down -> start before entry->start, 1860 * loop down 1861 * 4: not found, fixed, bottom up -> check entry->next->start, fail 1862 * 5: not found, fixed, top down -> check entry->next->start, fail 1863 * 6: not found, not fixed, bottom up -> check entry->next->start, 1864 * loop up 1865 * 7: not found, not fixed, top down -> check entry->next->start, 1866 * loop down 1867 * 1868 * as you can see, it reduces to roughly five cases, and that 1869 * adding top down mapping only adds one unique case (without 1870 * it, there would be four cases). 1871 */ 1872 1873 if ((flags & UVM_FLAG_FIXED) == 0 && hint == vm_map_min(map)) { 1874 entry = map->first_free; 1875 } else { 1876 if (uvm_map_lookup_entry(map, hint, &entry)) { 1877 /* "hint" address already in use ... */ 1878 if (flags & UVM_FLAG_FIXED) { 1879 UVMHIST_LOG(maphist, "<- fixed & VA in use", 1880 0, 0, 0, 0); 1881 return (NULL); 1882 } 1883 if (topdown) 1884 /* Start from lower gap. */ 1885 entry = entry->prev; 1886 } else if (flags & UVM_FLAG_FIXED) { 1887 if (entry->next->start >= hint + length && 1888 hint + length > hint) 1889 goto found; 1890 1891 /* "hint" address is gap but too small */ 1892 UVMHIST_LOG(maphist, "<- fixed mapping failed", 1893 0, 0, 0, 0); 1894 return (NULL); /* only one shot at it ... */ 1895 } else { 1896 /* 1897 * See if given hint fits in this gap. 1898 */ 1899 switch (uvm_map_space_avail(&hint, length, 1900 uoffset, align, flags, topdown, entry)) { 1901 case 1: 1902 goto found; 1903 case -1: 1904 goto wraparound; 1905 } 1906 1907 if (topdown) { 1908 /* 1909 * Still there is a chance to fit 1910 * if hint > entry->end. 1911 */ 1912 } else { 1913 /* Start from higher gap. */ 1914 entry = entry->next; 1915 if (entry == &map->header) 1916 goto notfound; 1917 goto nextgap; 1918 } 1919 } 1920 } 1921 1922 /* 1923 * Note that all UVM_FLAGS_FIXED case is already handled. 1924 */ 1925 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 1926 1927 /* Try to find the space in the red-black tree */ 1928 1929 /* Check slot before any entry */ 1930 hint = topdown ? entry->next->start - length : entry->end; 1931 switch (uvm_map_space_avail(&hint, length, uoffset, align, flags, 1932 topdown, entry)) { 1933 case 1: 1934 goto found; 1935 case -1: 1936 goto wraparound; 1937 } 1938 1939 nextgap: 1940 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 1941 /* If there is not enough space in the whole tree, we fail */ 1942 tmp = ROOT_ENTRY(map); 1943 if (tmp == NULL || tmp->maxgap < length) 1944 goto notfound; 1945 1946 prev = NULL; /* previous candidate */ 1947 1948 /* Find an entry close to hint that has enough space */ 1949 for (; tmp;) { 1950 KASSERT(tmp->next->start == tmp->end + tmp->gap); 1951 if (topdown) { 1952 if (tmp->next->start < hint + length && 1953 (prev == NULL || tmp->end > prev->end)) { 1954 if (tmp->gap >= length) 1955 prev = tmp; 1956 else if ((child = LEFT_ENTRY(tmp)) != NULL 1957 && child->maxgap >= length) 1958 prev = tmp; 1959 } 1960 } else { 1961 if (tmp->end >= hint && 1962 (prev == NULL || tmp->end < prev->end)) { 1963 if (tmp->gap >= length) 1964 prev = tmp; 1965 else if ((child = RIGHT_ENTRY(tmp)) != NULL 1966 && child->maxgap >= length) 1967 prev = tmp; 1968 } 1969 } 1970 if (tmp->next->start < hint + length) 1971 child = RIGHT_ENTRY(tmp); 1972 else if (tmp->end > hint) 1973 child = LEFT_ENTRY(tmp); 1974 else { 1975 if (tmp->gap >= length) 1976 break; 1977 if (topdown) 1978 child = LEFT_ENTRY(tmp); 1979 else 1980 child = RIGHT_ENTRY(tmp); 1981 } 1982 if (child == NULL || child->maxgap < length) 1983 break; 1984 tmp = child; 1985 } 1986 1987 if (tmp != NULL && tmp->start < hint && hint < tmp->next->start) { 1988 /* 1989 * Check if the entry that we found satifies the 1990 * space requirement 1991 */ 1992 if (topdown) { 1993 if (hint > tmp->next->start - length) 1994 hint = tmp->next->start - length; 1995 } else { 1996 if (hint < tmp->end) 1997 hint = tmp->end; 1998 } 1999 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2000 flags, topdown, tmp)) { 2001 case 1: 2002 entry = tmp; 2003 goto found; 2004 case -1: 2005 goto wraparound; 2006 } 2007 if (tmp->gap >= length) 2008 goto listsearch; 2009 } 2010 if (prev == NULL) 2011 goto notfound; 2012 2013 if (topdown) { 2014 KASSERT(orig_hint >= prev->next->start - length || 2015 prev->next->start - length > prev->next->start); 2016 hint = prev->next->start - length; 2017 } else { 2018 KASSERT(orig_hint <= prev->end); 2019 hint = prev->end; 2020 } 2021 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2022 flags, topdown, prev)) { 2023 case 1: 2024 entry = prev; 2025 goto found; 2026 case -1: 2027 goto wraparound; 2028 } 2029 if (prev->gap >= length) 2030 goto listsearch; 2031 2032 if (topdown) 2033 tmp = LEFT_ENTRY(prev); 2034 else 2035 tmp = RIGHT_ENTRY(prev); 2036 for (;;) { 2037 KASSERT(tmp && tmp->maxgap >= length); 2038 if (topdown) 2039 child = RIGHT_ENTRY(tmp); 2040 else 2041 child = LEFT_ENTRY(tmp); 2042 if (child && child->maxgap >= length) { 2043 tmp = child; 2044 continue; 2045 } 2046 if (tmp->gap >= length) 2047 break; 2048 if (topdown) 2049 tmp = LEFT_ENTRY(tmp); 2050 else 2051 tmp = RIGHT_ENTRY(tmp); 2052 } 2053 2054 if (topdown) { 2055 KASSERT(orig_hint >= tmp->next->start - length || 2056 tmp->next->start - length > tmp->next->start); 2057 hint = tmp->next->start - length; 2058 } else { 2059 KASSERT(orig_hint <= tmp->end); 2060 hint = tmp->end; 2061 } 2062 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2063 flags, topdown, tmp)) { 2064 case 1: 2065 entry = tmp; 2066 goto found; 2067 case -1: 2068 goto wraparound; 2069 } 2070 2071 /* 2072 * The tree fails to find an entry because of offset or alignment 2073 * restrictions. Search the list instead. 2074 */ 2075 listsearch: 2076 /* 2077 * Look through the rest of the map, trying to fit a new region in 2078 * the gap between existing regions, or after the very last region. 2079 * note: entry->end = base VA of current gap, 2080 * entry->next->start = VA of end of current gap 2081 */ 2082 2083 for (;;) { 2084 /* Update hint for current gap. */ 2085 hint = topdown ? entry->next->start - length : entry->end; 2086 2087 /* See if it fits. */ 2088 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2089 flags, topdown, entry)) { 2090 case 1: 2091 goto found; 2092 case -1: 2093 goto wraparound; 2094 } 2095 2096 /* Advance to next/previous gap */ 2097 if (topdown) { 2098 if (entry == &map->header) { 2099 UVMHIST_LOG(maphist, "<- failed (off start)", 2100 0,0,0,0); 2101 goto notfound; 2102 } 2103 entry = entry->prev; 2104 } else { 2105 entry = entry->next; 2106 if (entry == &map->header) { 2107 UVMHIST_LOG(maphist, "<- failed (off end)", 2108 0,0,0,0); 2109 goto notfound; 2110 } 2111 } 2112 } 2113 2114 found: 2115 SAVE_HINT(map, map->hint, entry); 2116 *result = hint; 2117 UVMHIST_LOG(maphist,"<- got it! (result=%#jx)", hint, 0,0,0); 2118 KASSERTMSG( topdown || hint >= orig_hint, "hint: %#jx, orig_hint: %#jx", 2119 (uintmax_t)hint, (uintmax_t)orig_hint); 2120 KASSERTMSG(!topdown || hint <= orig_hint, "hint: %#jx, orig_hint: %#jx", 2121 (uintmax_t)hint, (uintmax_t)orig_hint); 2122 KASSERT(entry->end <= hint); 2123 KASSERT(hint + length <= entry->next->start); 2124 return (entry); 2125 2126 wraparound: 2127 UVMHIST_LOG(maphist, "<- failed (wrap around)", 0,0,0,0); 2128 2129 return (NULL); 2130 2131 notfound: 2132 UVMHIST_LOG(maphist, "<- failed (notfound)", 0,0,0,0); 2133 2134 return (NULL); 2135 } 2136 2137 /* 2138 * U N M A P - m a i n h e l p e r f u n c t i o n s 2139 */ 2140 2141 /* 2142 * uvm_unmap_remove: remove mappings from a vm_map (from "start" up to "stop") 2143 * 2144 * => caller must check alignment and size 2145 * => map must be locked by caller 2146 * => we return a list of map entries that we've remove from the map 2147 * in "entry_list" 2148 */ 2149 2150 void 2151 uvm_unmap_remove(struct vm_map *map, vaddr_t start, vaddr_t end, 2152 struct vm_map_entry **entry_list /* OUT */, int flags) 2153 { 2154 struct vm_map_entry *entry, *first_entry, *next; 2155 vaddr_t len; 2156 UVMHIST_FUNC(__func__); 2157 UVMHIST_CALLARGS(maphist,"(map=%#jx, start=%#jx, end=%#jx)", 2158 (uintptr_t)map, start, end, 0); 2159 VM_MAP_RANGE_CHECK(map, start, end); 2160 2161 uvm_map_check(map, "unmap_remove entry"); 2162 2163 /* 2164 * find first entry 2165 */ 2166 2167 if (uvm_map_lookup_entry(map, start, &first_entry) == true) { 2168 /* clip and go... */ 2169 entry = first_entry; 2170 UVM_MAP_CLIP_START(map, entry, start); 2171 /* critical! prevents stale hint */ 2172 SAVE_HINT(map, entry, entry->prev); 2173 } else { 2174 entry = first_entry->next; 2175 } 2176 2177 /* 2178 * save the free space hint 2179 */ 2180 2181 if (map->first_free != &map->header && map->first_free->start >= start) 2182 map->first_free = entry->prev; 2183 2184 /* 2185 * note: we now re-use first_entry for a different task. we remove 2186 * a number of map entries from the map and save them in a linked 2187 * list headed by "first_entry". once we remove them from the map 2188 * the caller should unlock the map and drop the references to the 2189 * backing objects [c.f. uvm_unmap_detach]. the object is to 2190 * separate unmapping from reference dropping. why? 2191 * [1] the map has to be locked for unmapping 2192 * [2] the map need not be locked for reference dropping 2193 * [3] dropping references may trigger pager I/O, and if we hit 2194 * a pager that does synchronous I/O we may have to wait for it. 2195 * [4] we would like all waiting for I/O to occur with maps unlocked 2196 * so that we don't block other threads. 2197 */ 2198 2199 first_entry = NULL; 2200 *entry_list = NULL; 2201 2202 /* 2203 * break up the area into map entry sized regions and unmap. note 2204 * that all mappings have to be removed before we can even consider 2205 * dropping references to amaps or VM objects (otherwise we could end 2206 * up with a mapping to a page on the free list which would be very bad) 2207 */ 2208 2209 while ((entry != &map->header) && (entry->start < end)) { 2210 KASSERT((entry->flags & UVM_MAP_STATIC) == 0); 2211 2212 UVM_MAP_CLIP_END(map, entry, end); 2213 next = entry->next; 2214 len = entry->end - entry->start; 2215 2216 /* 2217 * unwire before removing addresses from the pmap; otherwise 2218 * unwiring will put the entries back into the pmap (XXX). 2219 */ 2220 2221 if (VM_MAPENT_ISWIRED(entry)) { 2222 uvm_map_entry_unwire(map, entry); 2223 } 2224 if (flags & UVM_FLAG_VAONLY) { 2225 2226 /* nothing */ 2227 2228 } else if ((map->flags & VM_MAP_PAGEABLE) == 0) { 2229 2230 /* 2231 * if the map is non-pageable, any pages mapped there 2232 * must be wired and entered with pmap_kenter_pa(), 2233 * and we should free any such pages immediately. 2234 * this is mostly used for kmem_map. 2235 */ 2236 KASSERT(vm_map_pmap(map) == pmap_kernel()); 2237 2238 uvm_km_pgremove_intrsafe(map, entry->start, entry->end); 2239 } else if (UVM_ET_ISOBJ(entry) && 2240 UVM_OBJ_IS_KERN_OBJECT(entry->object.uvm_obj)) { 2241 panic("%s: kernel object %p %p\n", 2242 __func__, map, entry); 2243 } else if (UVM_ET_ISOBJ(entry) || entry->aref.ar_amap) { 2244 /* 2245 * remove mappings the standard way. lock object 2246 * and/or amap to ensure vm_page state does not 2247 * change while in pmap_remove(). 2248 */ 2249 2250 #ifdef __HAVE_UNLOCKED_PMAP /* XXX temporary */ 2251 uvm_map_lock_entry(entry, RW_WRITER); 2252 #else 2253 uvm_map_lock_entry(entry, RW_READER); 2254 #endif 2255 pmap_remove(map->pmap, entry->start, entry->end); 2256 2257 /* 2258 * note: if map is dying, leave pmap_update() for 2259 * later. if the map is to be reused (exec) then 2260 * pmap_update() will be called. if the map is 2261 * being disposed of (exit) then pmap_destroy() 2262 * will be called. 2263 */ 2264 2265 if ((map->flags & VM_MAP_DYING) == 0) { 2266 pmap_update(vm_map_pmap(map)); 2267 } else { 2268 KASSERT(vm_map_pmap(map) != pmap_kernel()); 2269 } 2270 2271 uvm_map_unlock_entry(entry); 2272 } 2273 2274 #if defined(UVMDEBUG) 2275 /* 2276 * check if there's remaining mapping, 2277 * which is a bug in caller. 2278 */ 2279 2280 vaddr_t va; 2281 for (va = entry->start; va < entry->end; 2282 va += PAGE_SIZE) { 2283 if (pmap_extract(vm_map_pmap(map), va, NULL)) { 2284 panic("%s: %#"PRIxVADDR" has mapping", 2285 __func__, va); 2286 } 2287 } 2288 2289 if (VM_MAP_IS_KERNEL(map) && (flags & UVM_FLAG_NOWAIT) == 0) { 2290 uvm_km_check_empty(map, entry->start, 2291 entry->end); 2292 } 2293 #endif /* defined(UVMDEBUG) */ 2294 2295 /* 2296 * remove entry from map and put it on our list of entries 2297 * that we've nuked. then go to next entry. 2298 */ 2299 2300 UVMHIST_LOG(maphist, " removed map entry %#jx", 2301 (uintptr_t)entry, 0, 0, 0); 2302 2303 /* critical! prevents stale hint */ 2304 SAVE_HINT(map, entry, entry->prev); 2305 2306 uvm_map_entry_unlink(map, entry); 2307 KASSERT(map->size >= len); 2308 map->size -= len; 2309 entry->prev = NULL; 2310 entry->next = first_entry; 2311 first_entry = entry; 2312 entry = next; 2313 } 2314 2315 uvm_map_check(map, "unmap_remove leave"); 2316 2317 /* 2318 * now we've cleaned up the map and are ready for the caller to drop 2319 * references to the mapped objects. 2320 */ 2321 2322 *entry_list = first_entry; 2323 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 2324 2325 if (map->flags & VM_MAP_WANTVA) { 2326 mutex_enter(&map->misc_lock); 2327 map->flags &= ~VM_MAP_WANTVA; 2328 cv_broadcast(&map->cv); 2329 mutex_exit(&map->misc_lock); 2330 } 2331 } 2332 2333 /* 2334 * uvm_unmap_detach: drop references in a chain of map entries 2335 * 2336 * => we will free the map entries as we traverse the list. 2337 */ 2338 2339 void 2340 uvm_unmap_detach(struct vm_map_entry *first_entry, int flags) 2341 { 2342 struct vm_map_entry *next_entry; 2343 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 2344 2345 while (first_entry) { 2346 KASSERT(!VM_MAPENT_ISWIRED(first_entry)); 2347 UVMHIST_LOG(maphist, 2348 " detach %#jx: amap=%#jx, obj=%#jx, submap?=%jd", 2349 (uintptr_t)first_entry, 2350 (uintptr_t)first_entry->aref.ar_amap, 2351 (uintptr_t)first_entry->object.uvm_obj, 2352 UVM_ET_ISSUBMAP(first_entry)); 2353 2354 /* 2355 * drop reference to amap, if we've got one 2356 */ 2357 2358 if (first_entry->aref.ar_amap) 2359 uvm_map_unreference_amap(first_entry, flags); 2360 2361 /* 2362 * drop reference to our backing object, if we've got one 2363 */ 2364 2365 KASSERT(!UVM_ET_ISSUBMAP(first_entry)); 2366 if (UVM_ET_ISOBJ(first_entry) && 2367 first_entry->object.uvm_obj->pgops->pgo_detach) { 2368 (*first_entry->object.uvm_obj->pgops->pgo_detach) 2369 (first_entry->object.uvm_obj); 2370 } 2371 next_entry = first_entry->next; 2372 uvm_mapent_free(first_entry); 2373 first_entry = next_entry; 2374 } 2375 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 2376 } 2377 2378 /* 2379 * E X T R A C T I O N F U N C T I O N S 2380 */ 2381 2382 /* 2383 * uvm_map_reserve: reserve space in a vm_map for future use. 2384 * 2385 * => we reserve space in a map by putting a dummy map entry in the 2386 * map (dummy means obj=NULL, amap=NULL, prot=VM_PROT_NONE) 2387 * => map should be unlocked (we will write lock it) 2388 * => we return true if we were able to reserve space 2389 * => XXXCDC: should be inline? 2390 */ 2391 2392 int 2393 uvm_map_reserve(struct vm_map *map, vsize_t size, 2394 vaddr_t offset /* hint for pmap_prefer */, 2395 vsize_t align /* alignment */, 2396 vaddr_t *raddr /* IN:hint, OUT: reserved VA */, 2397 uvm_flag_t flags /* UVM_FLAG_FIXED or UVM_FLAG_COLORMATCH or 0 */) 2398 { 2399 UVMHIST_FUNC(__func__); 2400 UVMHIST_CALLARGS(maphist, "(map=%#jx, size=%#jx, offset=%#jx, addr=%#jx)", 2401 (uintptr_t)map, size, offset, (uintptr_t)raddr); 2402 2403 size = round_page(size); 2404 2405 /* 2406 * reserve some virtual space. 2407 */ 2408 2409 if (uvm_map(map, raddr, size, NULL, offset, align, 2410 UVM_MAPFLAG(UVM_PROT_NONE, UVM_PROT_NONE, UVM_INH_NONE, 2411 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE|flags)) != 0) { 2412 UVMHIST_LOG(maphist, "<- done (no VM)", 0,0,0,0); 2413 return (false); 2414 } 2415 2416 UVMHIST_LOG(maphist, "<- done (*raddr=%#jx)", *raddr,0,0,0); 2417 return (true); 2418 } 2419 2420 /* 2421 * uvm_map_replace: replace a reserved (blank) area of memory with 2422 * real mappings. 2423 * 2424 * => caller must WRITE-LOCK the map 2425 * => we return true if replacement was a success 2426 * => we expect the newents chain to have nnewents entrys on it and 2427 * we expect newents->prev to point to the last entry on the list 2428 * => note newents is allowed to be NULL 2429 */ 2430 2431 static int 2432 uvm_map_replace(struct vm_map *map, vaddr_t start, vaddr_t end, 2433 struct vm_map_entry *newents, int nnewents, vsize_t nsize, 2434 struct vm_map_entry **oldentryp) 2435 { 2436 struct vm_map_entry *oldent, *last; 2437 2438 uvm_map_check(map, "map_replace entry"); 2439 2440 /* 2441 * first find the blank map entry at the specified address 2442 */ 2443 2444 if (!uvm_map_lookup_entry(map, start, &oldent)) { 2445 return (false); 2446 } 2447 2448 /* 2449 * check to make sure we have a proper blank entry 2450 */ 2451 2452 if (end < oldent->end) { 2453 UVM_MAP_CLIP_END(map, oldent, end); 2454 } 2455 if (oldent->start != start || oldent->end != end || 2456 oldent->object.uvm_obj != NULL || oldent->aref.ar_amap != NULL) { 2457 return (false); 2458 } 2459 2460 #ifdef DIAGNOSTIC 2461 2462 /* 2463 * sanity check the newents chain 2464 */ 2465 2466 { 2467 struct vm_map_entry *tmpent = newents; 2468 int nent = 0; 2469 vsize_t sz = 0; 2470 vaddr_t cur = start; 2471 2472 while (tmpent) { 2473 nent++; 2474 sz += tmpent->end - tmpent->start; 2475 if (tmpent->start < cur) 2476 panic("uvm_map_replace1"); 2477 if (tmpent->start >= tmpent->end || tmpent->end > end) { 2478 panic("uvm_map_replace2: " 2479 "tmpent->start=%#"PRIxVADDR 2480 ", tmpent->end=%#"PRIxVADDR 2481 ", end=%#"PRIxVADDR, 2482 tmpent->start, tmpent->end, end); 2483 } 2484 cur = tmpent->end; 2485 if (tmpent->next) { 2486 if (tmpent->next->prev != tmpent) 2487 panic("uvm_map_replace3"); 2488 } else { 2489 if (newents->prev != tmpent) 2490 panic("uvm_map_replace4"); 2491 } 2492 tmpent = tmpent->next; 2493 } 2494 if (nent != nnewents) 2495 panic("uvm_map_replace5"); 2496 if (sz != nsize) 2497 panic("uvm_map_replace6"); 2498 } 2499 #endif 2500 2501 /* 2502 * map entry is a valid blank! replace it. (this does all the 2503 * work of map entry link/unlink...). 2504 */ 2505 2506 if (newents) { 2507 last = newents->prev; 2508 2509 /* critical: flush stale hints out of map */ 2510 SAVE_HINT(map, map->hint, newents); 2511 if (map->first_free == oldent) 2512 map->first_free = last; 2513 2514 last->next = oldent->next; 2515 last->next->prev = last; 2516 2517 /* Fix RB tree */ 2518 uvm_rb_remove(map, oldent); 2519 2520 newents->prev = oldent->prev; 2521 newents->prev->next = newents; 2522 map->nentries = map->nentries + (nnewents - 1); 2523 2524 /* Fixup the RB tree */ 2525 { 2526 int i; 2527 struct vm_map_entry *tmp; 2528 2529 tmp = newents; 2530 for (i = 0; i < nnewents && tmp; i++) { 2531 uvm_rb_insert(map, tmp); 2532 tmp = tmp->next; 2533 } 2534 } 2535 } else { 2536 /* NULL list of new entries: just remove the old one */ 2537 clear_hints(map, oldent); 2538 uvm_map_entry_unlink(map, oldent); 2539 } 2540 map->size -= end - start - nsize; 2541 2542 uvm_map_check(map, "map_replace leave"); 2543 2544 /* 2545 * now we can free the old blank entry and return. 2546 */ 2547 2548 *oldentryp = oldent; 2549 return (true); 2550 } 2551 2552 /* 2553 * uvm_map_extract: extract a mapping from a map and put it somewhere 2554 * (maybe removing the old mapping) 2555 * 2556 * => maps should be unlocked (we will write lock them) 2557 * => returns 0 on success, error code otherwise 2558 * => start must be page aligned 2559 * => len must be page sized 2560 * => flags: 2561 * UVM_EXTRACT_REMOVE: remove mappings from srcmap 2562 * UVM_EXTRACT_CONTIG: abort if unmapped area (advisory only) 2563 * UVM_EXTRACT_QREF: for a temporary extraction do quick obj refs 2564 * UVM_EXTRACT_FIXPROT: set prot to maxprot as we go 2565 * UVM_EXTRACT_PROT_ALL: set prot to UVM_PROT_ALL as we go 2566 * >>>NOTE: if you set REMOVE, you are not allowed to use CONTIG or QREF!<<< 2567 * >>>NOTE: QREF's must be unmapped via the QREF path, thus should only 2568 * be used from within the kernel in a kernel level map <<< 2569 */ 2570 2571 int 2572 uvm_map_extract(struct vm_map *srcmap, vaddr_t start, vsize_t len, 2573 struct vm_map *dstmap, vaddr_t *dstaddrp, int flags) 2574 { 2575 vaddr_t dstaddr, end, newend, oldoffset, fudge, orig_fudge; 2576 struct vm_map_entry *chain, *endchain, *entry, *orig_entry, *newentry, 2577 *deadentry, *oldentry; 2578 struct vm_map_entry *resentry = NULL; /* a dummy reservation entry */ 2579 vsize_t elen __unused; 2580 int nchain, error, copy_ok; 2581 vsize_t nsize; 2582 UVMHIST_FUNC(__func__); 2583 UVMHIST_CALLARGS(maphist,"(srcmap=%#jx,start=%#jx, len=%#jx", 2584 (uintptr_t)srcmap, start, len, 0); 2585 UVMHIST_LOG(maphist," ...,dstmap=%#jx, flags=%#jx)", 2586 (uintptr_t)dstmap, flags, 0, 0); 2587 2588 /* 2589 * step 0: sanity check: start must be on a page boundary, length 2590 * must be page sized. can't ask for CONTIG/QREF if you asked for 2591 * REMOVE. 2592 */ 2593 2594 KASSERT((start & PAGE_MASK) == 0 && (len & PAGE_MASK) == 0); 2595 KASSERT((flags & UVM_EXTRACT_REMOVE) == 0 || 2596 (flags & (UVM_EXTRACT_CONTIG|UVM_EXTRACT_QREF)) == 0); 2597 2598 /* 2599 * step 1: reserve space in the target map for the extracted area 2600 */ 2601 2602 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 2603 dstaddr = vm_map_min(dstmap); 2604 if (!uvm_map_reserve(dstmap, len, start, 2605 atop(start) & uvmexp.colormask, &dstaddr, 2606 UVM_FLAG_COLORMATCH)) 2607 return (ENOMEM); 2608 KASSERT((atop(start ^ dstaddr) & uvmexp.colormask) == 0); 2609 *dstaddrp = dstaddr; /* pass address back to caller */ 2610 UVMHIST_LOG(maphist, " dstaddr=%#jx", dstaddr,0,0,0); 2611 } else { 2612 dstaddr = *dstaddrp; 2613 } 2614 2615 /* 2616 * step 2: setup for the extraction process loop by init'ing the 2617 * map entry chain, locking src map, and looking up the first useful 2618 * entry in the map. 2619 */ 2620 2621 end = start + len; 2622 newend = dstaddr + len; 2623 chain = endchain = NULL; 2624 nchain = 0; 2625 nsize = 0; 2626 vm_map_lock(srcmap); 2627 2628 if (uvm_map_lookup_entry(srcmap, start, &entry)) { 2629 2630 /* "start" is within an entry */ 2631 if (flags & UVM_EXTRACT_QREF) { 2632 2633 /* 2634 * for quick references we don't clip the entry, so 2635 * the entry may map space "before" the starting 2636 * virtual address... this is the "fudge" factor 2637 * (which can be non-zero only the first time 2638 * through the "while" loop in step 3). 2639 */ 2640 2641 fudge = start - entry->start; 2642 } else { 2643 2644 /* 2645 * normal reference: we clip the map to fit (thus 2646 * fudge is zero) 2647 */ 2648 2649 UVM_MAP_CLIP_START(srcmap, entry, start); 2650 SAVE_HINT(srcmap, srcmap->hint, entry->prev); 2651 fudge = 0; 2652 } 2653 } else { 2654 2655 /* "start" is not within an entry ... skip to next entry */ 2656 if (flags & UVM_EXTRACT_CONTIG) { 2657 error = EINVAL; 2658 goto bad; /* definite hole here ... */ 2659 } 2660 2661 entry = entry->next; 2662 fudge = 0; 2663 } 2664 2665 /* save values from srcmap for step 6 */ 2666 orig_entry = entry; 2667 orig_fudge = fudge; 2668 2669 /* 2670 * step 3: now start looping through the map entries, extracting 2671 * as we go. 2672 */ 2673 2674 while (entry->start < end && entry != &srcmap->header) { 2675 2676 /* if we are not doing a quick reference, clip it */ 2677 if ((flags & UVM_EXTRACT_QREF) == 0) 2678 UVM_MAP_CLIP_END(srcmap, entry, end); 2679 2680 /* clear needs_copy (allow chunking) */ 2681 if (UVM_ET_ISNEEDSCOPY(entry)) { 2682 amap_copy(srcmap, entry, 2683 AMAP_COPY_NOWAIT|AMAP_COPY_NOMERGE, start, end); 2684 if (UVM_ET_ISNEEDSCOPY(entry)) { /* failed? */ 2685 error = ENOMEM; 2686 goto bad; 2687 } 2688 2689 /* amap_copy could clip (during chunk)! update fudge */ 2690 if (fudge) { 2691 fudge = start - entry->start; 2692 orig_fudge = fudge; 2693 } 2694 } 2695 2696 /* calculate the offset of this from "start" */ 2697 oldoffset = (entry->start + fudge) - start; 2698 2699 /* allocate a new map entry */ 2700 newentry = uvm_mapent_alloc(dstmap, 0); 2701 if (newentry == NULL) { 2702 error = ENOMEM; 2703 goto bad; 2704 } 2705 2706 /* set up new map entry */ 2707 newentry->next = NULL; 2708 newentry->prev = endchain; 2709 newentry->start = dstaddr + oldoffset; 2710 newentry->end = 2711 newentry->start + (entry->end - (entry->start + fudge)); 2712 if (newentry->end > newend || newentry->end < newentry->start) 2713 newentry->end = newend; 2714 newentry->object.uvm_obj = entry->object.uvm_obj; 2715 if (newentry->object.uvm_obj) { 2716 if (newentry->object.uvm_obj->pgops->pgo_reference) 2717 newentry->object.uvm_obj->pgops-> 2718 pgo_reference(newentry->object.uvm_obj); 2719 newentry->offset = entry->offset + fudge; 2720 } else { 2721 newentry->offset = 0; 2722 } 2723 newentry->etype = entry->etype; 2724 if (flags & UVM_EXTRACT_PROT_ALL) { 2725 newentry->protection = newentry->max_protection = 2726 UVM_PROT_ALL; 2727 } else { 2728 newentry->protection = (flags & UVM_EXTRACT_FIXPROT) ? 2729 entry->max_protection : entry->protection; 2730 newentry->max_protection = entry->max_protection; 2731 } 2732 newentry->inheritance = entry->inheritance; 2733 newentry->wired_count = 0; 2734 newentry->aref.ar_amap = entry->aref.ar_amap; 2735 if (newentry->aref.ar_amap) { 2736 newentry->aref.ar_pageoff = 2737 entry->aref.ar_pageoff + (fudge >> PAGE_SHIFT); 2738 uvm_map_reference_amap(newentry, AMAP_SHARED | 2739 ((flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0)); 2740 } else { 2741 newentry->aref.ar_pageoff = 0; 2742 } 2743 newentry->advice = entry->advice; 2744 if ((flags & UVM_EXTRACT_QREF) != 0) { 2745 newentry->flags |= UVM_MAP_NOMERGE; 2746 } 2747 2748 /* now link it on the chain */ 2749 nchain++; 2750 nsize += newentry->end - newentry->start; 2751 if (endchain == NULL) { 2752 chain = endchain = newentry; 2753 } else { 2754 endchain->next = newentry; 2755 endchain = newentry; 2756 } 2757 2758 /* end of 'while' loop! */ 2759 if ((flags & UVM_EXTRACT_CONTIG) && entry->end < end && 2760 (entry->next == &srcmap->header || 2761 entry->next->start != entry->end)) { 2762 error = EINVAL; 2763 goto bad; 2764 } 2765 entry = entry->next; 2766 fudge = 0; 2767 } 2768 2769 /* 2770 * step 4: close off chain (in format expected by uvm_map_replace) 2771 */ 2772 2773 if (chain) 2774 chain->prev = endchain; 2775 2776 /* 2777 * step 5: attempt to lock the dest map so we can pmap_copy. 2778 * note usage of copy_ok: 2779 * 1 => dstmap locked, pmap_copy ok, and we "replace" here (step 5) 2780 * 0 => dstmap unlocked, NO pmap_copy, and we will "replace" in step 7 2781 */ 2782 2783 if (srcmap == dstmap || vm_map_lock_try(dstmap) == true) { 2784 copy_ok = 1; 2785 if (!uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2786 nchain, nsize, &resentry)) { 2787 if (srcmap != dstmap) 2788 vm_map_unlock(dstmap); 2789 error = EIO; 2790 goto bad; 2791 } 2792 } else { 2793 copy_ok = 0; 2794 /* replace defered until step 7 */ 2795 } 2796 2797 /* 2798 * step 6: traverse the srcmap a second time to do the following: 2799 * - if we got a lock on the dstmap do pmap_copy 2800 * - if UVM_EXTRACT_REMOVE remove the entries 2801 * we make use of orig_entry and orig_fudge (saved in step 2) 2802 */ 2803 2804 if (copy_ok || (flags & UVM_EXTRACT_REMOVE)) { 2805 2806 /* purge possible stale hints from srcmap */ 2807 if (flags & UVM_EXTRACT_REMOVE) { 2808 SAVE_HINT(srcmap, srcmap->hint, orig_entry->prev); 2809 if (srcmap->first_free != &srcmap->header && 2810 srcmap->first_free->start >= start) 2811 srcmap->first_free = orig_entry->prev; 2812 } 2813 2814 entry = orig_entry; 2815 fudge = orig_fudge; 2816 deadentry = NULL; /* for UVM_EXTRACT_REMOVE */ 2817 2818 while (entry->start < end && entry != &srcmap->header) { 2819 if (copy_ok) { 2820 oldoffset = (entry->start + fudge) - start; 2821 elen = MIN(end, entry->end) - 2822 (entry->start + fudge); 2823 pmap_copy(dstmap->pmap, srcmap->pmap, 2824 dstaddr + oldoffset, elen, 2825 entry->start + fudge); 2826 } 2827 2828 /* we advance "entry" in the following if statement */ 2829 if (flags & UVM_EXTRACT_REMOVE) { 2830 #ifdef __HAVE_UNLOCKED_PMAP /* XXX temporary */ 2831 uvm_map_lock_entry(entry, RW_WRITER); 2832 #else 2833 uvm_map_lock_entry(entry, RW_READER); 2834 #endif 2835 pmap_remove(srcmap->pmap, entry->start, 2836 entry->end); 2837 uvm_map_unlock_entry(entry); 2838 oldentry = entry; /* save entry */ 2839 entry = entry->next; /* advance */ 2840 uvm_map_entry_unlink(srcmap, oldentry); 2841 /* add to dead list */ 2842 oldentry->next = deadentry; 2843 deadentry = oldentry; 2844 } else { 2845 entry = entry->next; /* advance */ 2846 } 2847 2848 /* end of 'while' loop */ 2849 fudge = 0; 2850 } 2851 pmap_update(srcmap->pmap); 2852 2853 /* 2854 * unlock dstmap. we will dispose of deadentry in 2855 * step 7 if needed 2856 */ 2857 2858 if (copy_ok && srcmap != dstmap) 2859 vm_map_unlock(dstmap); 2860 2861 } else { 2862 deadentry = NULL; 2863 } 2864 2865 /* 2866 * step 7: we are done with the source map, unlock. if copy_ok 2867 * is 0 then we have not replaced the dummy mapping in dstmap yet 2868 * and we need to do so now. 2869 */ 2870 2871 vm_map_unlock(srcmap); 2872 if ((flags & UVM_EXTRACT_REMOVE) && deadentry) 2873 uvm_unmap_detach(deadentry, 0); /* dispose of old entries */ 2874 2875 /* now do the replacement if we didn't do it in step 5 */ 2876 if (copy_ok == 0) { 2877 vm_map_lock(dstmap); 2878 error = uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2879 nchain, nsize, &resentry); 2880 vm_map_unlock(dstmap); 2881 2882 if (error == false) { 2883 error = EIO; 2884 goto bad2; 2885 } 2886 } 2887 2888 if (resentry != NULL) 2889 uvm_mapent_free(resentry); 2890 2891 return (0); 2892 2893 /* 2894 * bad: failure recovery 2895 */ 2896 bad: 2897 vm_map_unlock(srcmap); 2898 bad2: /* src already unlocked */ 2899 if (chain) 2900 uvm_unmap_detach(chain, 2901 (flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0); 2902 2903 if (resentry != NULL) 2904 uvm_mapent_free(resentry); 2905 2906 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 2907 uvm_unmap(dstmap, dstaddr, dstaddr+len); /* ??? */ 2908 } 2909 return (error); 2910 } 2911 2912 /* end of extraction functions */ 2913 2914 /* 2915 * uvm_map_submap: punch down part of a map into a submap 2916 * 2917 * => only the kernel_map is allowed to be submapped 2918 * => the purpose of submapping is to break up the locking granularity 2919 * of a larger map 2920 * => the range specified must have been mapped previously with a uvm_map() 2921 * call [with uobj==NULL] to create a blank map entry in the main map. 2922 * [And it had better still be blank!] 2923 * => maps which contain submaps should never be copied or forked. 2924 * => to remove a submap, use uvm_unmap() on the main map 2925 * and then uvm_map_deallocate() the submap. 2926 * => main map must be unlocked. 2927 * => submap must have been init'd and have a zero reference count. 2928 * [need not be locked as we don't actually reference it] 2929 */ 2930 2931 int 2932 uvm_map_submap(struct vm_map *map, vaddr_t start, vaddr_t end, 2933 struct vm_map *submap) 2934 { 2935 struct vm_map_entry *entry; 2936 int error; 2937 2938 vm_map_lock(map); 2939 VM_MAP_RANGE_CHECK(map, start, end); 2940 2941 if (uvm_map_lookup_entry(map, start, &entry)) { 2942 UVM_MAP_CLIP_START(map, entry, start); 2943 UVM_MAP_CLIP_END(map, entry, end); /* to be safe */ 2944 } else { 2945 entry = NULL; 2946 } 2947 2948 if (entry != NULL && 2949 entry->start == start && entry->end == end && 2950 entry->object.uvm_obj == NULL && entry->aref.ar_amap == NULL && 2951 !UVM_ET_ISCOPYONWRITE(entry) && !UVM_ET_ISNEEDSCOPY(entry)) { 2952 entry->etype |= UVM_ET_SUBMAP; 2953 entry->object.sub_map = submap; 2954 entry->offset = 0; 2955 uvm_map_reference(submap); 2956 error = 0; 2957 } else { 2958 error = EINVAL; 2959 } 2960 vm_map_unlock(map); 2961 2962 return error; 2963 } 2964 2965 /* 2966 * uvm_map_protect_user: change map protection on behalf of the user. 2967 * Enforces PAX settings as necessary. 2968 */ 2969 int 2970 uvm_map_protect_user(struct lwp *l, vaddr_t start, vaddr_t end, 2971 vm_prot_t new_prot) 2972 { 2973 int error; 2974 2975 if ((error = PAX_MPROTECT_VALIDATE(l, new_prot))) 2976 return error; 2977 2978 return uvm_map_protect(&l->l_proc->p_vmspace->vm_map, start, end, 2979 new_prot, false); 2980 } 2981 2982 2983 /* 2984 * uvm_map_protect: change map protection 2985 * 2986 * => set_max means set max_protection. 2987 * => map must be unlocked. 2988 */ 2989 2990 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 2991 ~VM_PROT_WRITE : VM_PROT_ALL) 2992 2993 int 2994 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end, 2995 vm_prot_t new_prot, bool set_max) 2996 { 2997 struct vm_map_entry *current, *entry; 2998 int error = 0; 2999 UVMHIST_FUNC(__func__); 3000 UVMHIST_CALLARGS(maphist,"(map=%#jx,start=%#jx,end=%#jx,new_prot=%#jx)", 3001 (uintptr_t)map, start, end, new_prot); 3002 3003 vm_map_lock(map); 3004 VM_MAP_RANGE_CHECK(map, start, end); 3005 if (uvm_map_lookup_entry(map, start, &entry)) { 3006 UVM_MAP_CLIP_START(map, entry, start); 3007 } else { 3008 entry = entry->next; 3009 } 3010 3011 /* 3012 * make a first pass to check for protection violations. 3013 */ 3014 3015 current = entry; 3016 while ((current != &map->header) && (current->start < end)) { 3017 if (UVM_ET_ISSUBMAP(current)) { 3018 error = EINVAL; 3019 goto out; 3020 } 3021 if ((new_prot & current->max_protection) != new_prot) { 3022 error = EACCES; 3023 goto out; 3024 } 3025 /* 3026 * Don't allow VM_PROT_EXECUTE to be set on entries that 3027 * point to vnodes that are associated with a NOEXEC file 3028 * system. 3029 */ 3030 if (UVM_ET_ISOBJ(current) && 3031 UVM_OBJ_IS_VNODE(current->object.uvm_obj)) { 3032 struct vnode *vp = 3033 (struct vnode *) current->object.uvm_obj; 3034 3035 if ((new_prot & VM_PROT_EXECUTE) != 0 && 3036 (vp->v_mount->mnt_flag & MNT_NOEXEC) != 0) { 3037 error = EACCES; 3038 goto out; 3039 } 3040 } 3041 3042 current = current->next; 3043 } 3044 3045 /* go back and fix up protections (no need to clip this time). */ 3046 3047 current = entry; 3048 while ((current != &map->header) && (current->start < end)) { 3049 vm_prot_t old_prot; 3050 3051 UVM_MAP_CLIP_END(map, current, end); 3052 old_prot = current->protection; 3053 if (set_max) 3054 current->protection = 3055 (current->max_protection = new_prot) & old_prot; 3056 else 3057 current->protection = new_prot; 3058 3059 /* 3060 * update physical map if necessary. worry about copy-on-write 3061 * here -- CHECK THIS XXX 3062 */ 3063 3064 if (current->protection != old_prot) { 3065 /* update pmap! */ 3066 #ifdef __HAVE_UNLOCKED_PMAP /* XXX temporary */ 3067 uvm_map_lock_entry(current, RW_WRITER); 3068 #else 3069 uvm_map_lock_entry(current, RW_READER); 3070 #endif 3071 pmap_protect(map->pmap, current->start, current->end, 3072 current->protection & MASK(current)); 3073 uvm_map_unlock_entry(current); 3074 3075 /* 3076 * If this entry points at a vnode, and the 3077 * protection includes VM_PROT_EXECUTE, mark 3078 * the vnode as VEXECMAP. 3079 */ 3080 if (UVM_ET_ISOBJ(current)) { 3081 struct uvm_object *uobj = 3082 current->object.uvm_obj; 3083 3084 if (UVM_OBJ_IS_VNODE(uobj) && 3085 (current->protection & VM_PROT_EXECUTE)) { 3086 vn_markexec((struct vnode *) uobj); 3087 } 3088 } 3089 } 3090 3091 /* 3092 * If the map is configured to lock any future mappings, 3093 * wire this entry now if the old protection was VM_PROT_NONE 3094 * and the new protection is not VM_PROT_NONE. 3095 */ 3096 3097 if ((map->flags & VM_MAP_WIREFUTURE) != 0 && 3098 VM_MAPENT_ISWIRED(current) == 0 && 3099 old_prot == VM_PROT_NONE && 3100 new_prot != VM_PROT_NONE) { 3101 3102 /* 3103 * We must call pmap_update() here because the 3104 * pmap_protect() call above might have removed some 3105 * pmap entries and uvm_map_pageable() might create 3106 * some new pmap entries that rely on the prior 3107 * removals being completely finished. 3108 */ 3109 3110 pmap_update(map->pmap); 3111 3112 if (uvm_map_pageable(map, current->start, 3113 current->end, false, 3114 UVM_LK_ENTER|UVM_LK_EXIT) != 0) { 3115 3116 /* 3117 * If locking the entry fails, remember the 3118 * error if it's the first one. Note we 3119 * still continue setting the protection in 3120 * the map, but will return the error 3121 * condition regardless. 3122 * 3123 * XXX Ignore what the actual error is, 3124 * XXX just call it a resource shortage 3125 * XXX so that it doesn't get confused 3126 * XXX what uvm_map_protect() itself would 3127 * XXX normally return. 3128 */ 3129 3130 error = ENOMEM; 3131 } 3132 } 3133 current = current->next; 3134 } 3135 pmap_update(map->pmap); 3136 3137 out: 3138 vm_map_unlock(map); 3139 3140 UVMHIST_LOG(maphist, "<- done, error=%jd",error,0,0,0); 3141 return error; 3142 } 3143 3144 #undef MASK 3145 3146 /* 3147 * uvm_map_inherit: set inheritance code for range of addrs in map. 3148 * 3149 * => map must be unlocked 3150 * => note that the inherit code is used during a "fork". see fork 3151 * code for details. 3152 */ 3153 3154 int 3155 uvm_map_inherit(struct vm_map *map, vaddr_t start, vaddr_t end, 3156 vm_inherit_t new_inheritance) 3157 { 3158 struct vm_map_entry *entry, *temp_entry; 3159 UVMHIST_FUNC(__func__); 3160 UVMHIST_CALLARGS(maphist,"(map=%#jx,start=%#jx,end=%#jx,new_inh=%#jx)", 3161 (uintptr_t)map, start, end, new_inheritance); 3162 3163 switch (new_inheritance) { 3164 case MAP_INHERIT_NONE: 3165 case MAP_INHERIT_COPY: 3166 case MAP_INHERIT_SHARE: 3167 case MAP_INHERIT_ZERO: 3168 break; 3169 default: 3170 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3171 return EINVAL; 3172 } 3173 3174 vm_map_lock(map); 3175 VM_MAP_RANGE_CHECK(map, start, end); 3176 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3177 entry = temp_entry; 3178 UVM_MAP_CLIP_START(map, entry, start); 3179 } else { 3180 entry = temp_entry->next; 3181 } 3182 while ((entry != &map->header) && (entry->start < end)) { 3183 UVM_MAP_CLIP_END(map, entry, end); 3184 entry->inheritance = new_inheritance; 3185 entry = entry->next; 3186 } 3187 vm_map_unlock(map); 3188 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3189 return 0; 3190 } 3191 3192 /* 3193 * uvm_map_advice: set advice code for range of addrs in map. 3194 * 3195 * => map must be unlocked 3196 */ 3197 3198 int 3199 uvm_map_advice(struct vm_map *map, vaddr_t start, vaddr_t end, int new_advice) 3200 { 3201 struct vm_map_entry *entry, *temp_entry; 3202 UVMHIST_FUNC(__func__); 3203 UVMHIST_CALLARGS(maphist,"(map=%#jx,start=%#jx,end=%#jx,new_adv=%#jx)", 3204 (uintptr_t)map, start, end, new_advice); 3205 3206 vm_map_lock(map); 3207 VM_MAP_RANGE_CHECK(map, start, end); 3208 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3209 entry = temp_entry; 3210 UVM_MAP_CLIP_START(map, entry, start); 3211 } else { 3212 entry = temp_entry->next; 3213 } 3214 3215 /* 3216 * XXXJRT: disallow holes? 3217 */ 3218 3219 while ((entry != &map->header) && (entry->start < end)) { 3220 UVM_MAP_CLIP_END(map, entry, end); 3221 3222 switch (new_advice) { 3223 case MADV_NORMAL: 3224 case MADV_RANDOM: 3225 case MADV_SEQUENTIAL: 3226 /* nothing special here */ 3227 break; 3228 3229 default: 3230 vm_map_unlock(map); 3231 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3232 return EINVAL; 3233 } 3234 entry->advice = new_advice; 3235 entry = entry->next; 3236 } 3237 3238 vm_map_unlock(map); 3239 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3240 return 0; 3241 } 3242 3243 /* 3244 * uvm_map_willneed: apply MADV_WILLNEED 3245 */ 3246 3247 int 3248 uvm_map_willneed(struct vm_map *map, vaddr_t start, vaddr_t end) 3249 { 3250 struct vm_map_entry *entry; 3251 UVMHIST_FUNC(__func__); 3252 UVMHIST_CALLARGS(maphist,"(map=%#jx,start=%#jx,end=%#jx)", 3253 (uintptr_t)map, start, end, 0); 3254 3255 vm_map_lock_read(map); 3256 VM_MAP_RANGE_CHECK(map, start, end); 3257 if (!uvm_map_lookup_entry(map, start, &entry)) { 3258 entry = entry->next; 3259 } 3260 while (entry->start < end) { 3261 struct vm_amap * const amap = entry->aref.ar_amap; 3262 struct uvm_object * const uobj = entry->object.uvm_obj; 3263 3264 KASSERT(entry != &map->header); 3265 KASSERT(start < entry->end); 3266 /* 3267 * For now, we handle only the easy but commonly-requested case. 3268 * ie. start prefetching of backing uobj pages. 3269 * 3270 * XXX It might be useful to pmap_enter() the already-in-core 3271 * pages by inventing a "weak" mode for uvm_fault() which would 3272 * only do the PGO_LOCKED pgo_get(). 3273 */ 3274 if (UVM_ET_ISOBJ(entry) && amap == NULL && uobj != NULL) { 3275 off_t offset; 3276 off_t size; 3277 3278 offset = entry->offset; 3279 if (start < entry->start) { 3280 offset += entry->start - start; 3281 } 3282 size = entry->offset + (entry->end - entry->start); 3283 if (entry->end < end) { 3284 size -= end - entry->end; 3285 } 3286 uvm_readahead(uobj, offset, size); 3287 } 3288 entry = entry->next; 3289 } 3290 vm_map_unlock_read(map); 3291 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3292 return 0; 3293 } 3294 3295 /* 3296 * uvm_map_pageable: sets the pageability of a range in a map. 3297 * 3298 * => wires map entries. should not be used for transient page locking. 3299 * for that, use uvm_fault_wire()/uvm_fault_unwire() (see uvm_vslock()). 3300 * => regions specified as not pageable require lock-down (wired) memory 3301 * and page tables. 3302 * => map must never be read-locked 3303 * => if islocked is true, map is already write-locked 3304 * => we always unlock the map, since we must downgrade to a read-lock 3305 * to call uvm_fault_wire() 3306 * => XXXCDC: check this and try and clean it up. 3307 */ 3308 3309 int 3310 uvm_map_pageable(struct vm_map *map, vaddr_t start, vaddr_t end, 3311 bool new_pageable, int lockflags) 3312 { 3313 struct vm_map_entry *entry, *start_entry, *failed_entry; 3314 int rv; 3315 #ifdef DIAGNOSTIC 3316 u_int timestamp_save; 3317 #endif 3318 UVMHIST_FUNC(__func__); 3319 UVMHIST_CALLARGS(maphist,"(map=%#jx,start=%#jx,end=%#jx,new_pageable=%ju)", 3320 (uintptr_t)map, start, end, new_pageable); 3321 KASSERT(map->flags & VM_MAP_PAGEABLE); 3322 3323 if ((lockflags & UVM_LK_ENTER) == 0) 3324 vm_map_lock(map); 3325 VM_MAP_RANGE_CHECK(map, start, end); 3326 3327 /* 3328 * only one pageability change may take place at one time, since 3329 * uvm_fault_wire assumes it will be called only once for each 3330 * wiring/unwiring. therefore, we have to make sure we're actually 3331 * changing the pageability for the entire region. we do so before 3332 * making any changes. 3333 */ 3334 3335 if (uvm_map_lookup_entry(map, start, &start_entry) == false) { 3336 if ((lockflags & UVM_LK_EXIT) == 0) 3337 vm_map_unlock(map); 3338 3339 UVMHIST_LOG(maphist,"<- done (fault)",0,0,0,0); 3340 return EFAULT; 3341 } 3342 entry = start_entry; 3343 3344 if (start == end) { /* nothing required */ 3345 if ((lockflags & UVM_LK_EXIT) == 0) 3346 vm_map_unlock(map); 3347 3348 UVMHIST_LOG(maphist,"<- done (nothing)",0,0,0,0); 3349 return 0; 3350 } 3351 3352 /* 3353 * handle wiring and unwiring separately. 3354 */ 3355 3356 if (new_pageable) { /* unwire */ 3357 UVM_MAP_CLIP_START(map, entry, start); 3358 3359 /* 3360 * unwiring. first ensure that the range to be unwired is 3361 * really wired down and that there are no holes. 3362 */ 3363 3364 while ((entry != &map->header) && (entry->start < end)) { 3365 if (entry->wired_count == 0 || 3366 (entry->end < end && 3367 (entry->next == &map->header || 3368 entry->next->start > entry->end))) { 3369 if ((lockflags & UVM_LK_EXIT) == 0) 3370 vm_map_unlock(map); 3371 UVMHIST_LOG(maphist, "<- done (INVAL)",0,0,0,0); 3372 return EINVAL; 3373 } 3374 entry = entry->next; 3375 } 3376 3377 /* 3378 * POSIX 1003.1b - a single munlock call unlocks a region, 3379 * regardless of the number of mlock calls made on that 3380 * region. 3381 */ 3382 3383 entry = start_entry; 3384 while ((entry != &map->header) && (entry->start < end)) { 3385 UVM_MAP_CLIP_END(map, entry, end); 3386 if (VM_MAPENT_ISWIRED(entry)) 3387 uvm_map_entry_unwire(map, entry); 3388 entry = entry->next; 3389 } 3390 if ((lockflags & UVM_LK_EXIT) == 0) 3391 vm_map_unlock(map); 3392 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3393 return 0; 3394 } 3395 3396 /* 3397 * wire case: in two passes [XXXCDC: ugly block of code here] 3398 * 3399 * 1: holding the write lock, we create any anonymous maps that need 3400 * to be created. then we clip each map entry to the region to 3401 * be wired and increment its wiring count. 3402 * 3403 * 2: we downgrade to a read lock, and call uvm_fault_wire to fault 3404 * in the pages for any newly wired area (wired_count == 1). 3405 * 3406 * downgrading to a read lock for uvm_fault_wire avoids a possible 3407 * deadlock with another thread that may have faulted on one of 3408 * the pages to be wired (it would mark the page busy, blocking 3409 * us, then in turn block on the map lock that we hold). because 3410 * of problems in the recursive lock package, we cannot upgrade 3411 * to a write lock in vm_map_lookup. thus, any actions that 3412 * require the write lock must be done beforehand. because we 3413 * keep the read lock on the map, the copy-on-write status of the 3414 * entries we modify here cannot change. 3415 */ 3416 3417 while ((entry != &map->header) && (entry->start < end)) { 3418 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3419 3420 /* 3421 * perform actions of vm_map_lookup that need the 3422 * write lock on the map: create an anonymous map 3423 * for a copy-on-write region, or an anonymous map 3424 * for a zero-fill region. (XXXCDC: submap case 3425 * ok?) 3426 */ 3427 3428 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3429 if (UVM_ET_ISNEEDSCOPY(entry) && 3430 ((entry->max_protection & VM_PROT_WRITE) || 3431 (entry->object.uvm_obj == NULL))) { 3432 amap_copy(map, entry, 0, start, end); 3433 /* XXXCDC: wait OK? */ 3434 } 3435 } 3436 } 3437 UVM_MAP_CLIP_START(map, entry, start); 3438 UVM_MAP_CLIP_END(map, entry, end); 3439 entry->wired_count++; 3440 3441 /* 3442 * Check for holes 3443 */ 3444 3445 if (entry->protection == VM_PROT_NONE || 3446 (entry->end < end && 3447 (entry->next == &map->header || 3448 entry->next->start > entry->end))) { 3449 3450 /* 3451 * found one. amap creation actions do not need to 3452 * be undone, but the wired counts need to be restored. 3453 */ 3454 3455 while (entry != &map->header && entry->end > start) { 3456 entry->wired_count--; 3457 entry = entry->prev; 3458 } 3459 if ((lockflags & UVM_LK_EXIT) == 0) 3460 vm_map_unlock(map); 3461 UVMHIST_LOG(maphist,"<- done (INVALID WIRE)",0,0,0,0); 3462 return EINVAL; 3463 } 3464 entry = entry->next; 3465 } 3466 3467 /* 3468 * Pass 2. 3469 */ 3470 3471 #ifdef DIAGNOSTIC 3472 timestamp_save = map->timestamp; 3473 #endif 3474 vm_map_busy(map); 3475 vm_map_unlock(map); 3476 3477 rv = 0; 3478 entry = start_entry; 3479 while (entry != &map->header && entry->start < end) { 3480 if (entry->wired_count == 1) { 3481 rv = uvm_fault_wire(map, entry->start, entry->end, 3482 entry->max_protection, 1); 3483 if (rv) { 3484 3485 /* 3486 * wiring failed. break out of the loop. 3487 * we'll clean up the map below, once we 3488 * have a write lock again. 3489 */ 3490 3491 break; 3492 } 3493 } 3494 entry = entry->next; 3495 } 3496 3497 if (rv) { /* failed? */ 3498 3499 /* 3500 * Get back to an exclusive (write) lock. 3501 */ 3502 3503 vm_map_lock(map); 3504 vm_map_unbusy(map); 3505 3506 #ifdef DIAGNOSTIC 3507 if (timestamp_save + 1 != map->timestamp) 3508 panic("uvm_map_pageable: stale map"); 3509 #endif 3510 3511 /* 3512 * first drop the wiring count on all the entries 3513 * which haven't actually been wired yet. 3514 */ 3515 3516 failed_entry = entry; 3517 while (entry != &map->header && entry->start < end) { 3518 entry->wired_count--; 3519 entry = entry->next; 3520 } 3521 3522 /* 3523 * now, unwire all the entries that were successfully 3524 * wired above. 3525 */ 3526 3527 entry = start_entry; 3528 while (entry != failed_entry) { 3529 entry->wired_count--; 3530 if (VM_MAPENT_ISWIRED(entry) == 0) 3531 uvm_map_entry_unwire(map, entry); 3532 entry = entry->next; 3533 } 3534 if ((lockflags & UVM_LK_EXIT) == 0) 3535 vm_map_unlock(map); 3536 UVMHIST_LOG(maphist, "<- done (RV=%jd)", rv,0,0,0); 3537 return (rv); 3538 } 3539 3540 if ((lockflags & UVM_LK_EXIT) == 0) { 3541 vm_map_unbusy(map); 3542 } else { 3543 3544 /* 3545 * Get back to an exclusive (write) lock. 3546 */ 3547 3548 vm_map_lock(map); 3549 vm_map_unbusy(map); 3550 } 3551 3552 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3553 return 0; 3554 } 3555 3556 /* 3557 * uvm_map_pageable_all: special case of uvm_map_pageable - affects 3558 * all mapped regions. 3559 * 3560 * => map must not be locked. 3561 * => if no flags are specified, all regions are unwired. 3562 * => XXXJRT: has some of the same problems as uvm_map_pageable() above. 3563 */ 3564 3565 int 3566 uvm_map_pageable_all(struct vm_map *map, int flags, vsize_t limit) 3567 { 3568 struct vm_map_entry *entry, *failed_entry; 3569 vsize_t size; 3570 int rv; 3571 #ifdef DIAGNOSTIC 3572 u_int timestamp_save; 3573 #endif 3574 UVMHIST_FUNC(__func__); 3575 UVMHIST_CALLARGS(maphist,"(map=%#jx,flags=%#jx)", (uintptr_t)map, flags, 3576 0, 0); 3577 3578 KASSERT(map->flags & VM_MAP_PAGEABLE); 3579 3580 vm_map_lock(map); 3581 3582 /* 3583 * handle wiring and unwiring separately. 3584 */ 3585 3586 if (flags == 0) { /* unwire */ 3587 3588 /* 3589 * POSIX 1003.1b -- munlockall unlocks all regions, 3590 * regardless of how many times mlockall has been called. 3591 */ 3592 3593 for (entry = map->header.next; entry != &map->header; 3594 entry = entry->next) { 3595 if (VM_MAPENT_ISWIRED(entry)) 3596 uvm_map_entry_unwire(map, entry); 3597 } 3598 map->flags &= ~VM_MAP_WIREFUTURE; 3599 vm_map_unlock(map); 3600 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3601 return 0; 3602 } 3603 3604 if (flags & MCL_FUTURE) { 3605 3606 /* 3607 * must wire all future mappings; remember this. 3608 */ 3609 3610 map->flags |= VM_MAP_WIREFUTURE; 3611 } 3612 3613 if ((flags & MCL_CURRENT) == 0) { 3614 3615 /* 3616 * no more work to do! 3617 */ 3618 3619 UVMHIST_LOG(maphist,"<- done (OK no wire)",0,0,0,0); 3620 vm_map_unlock(map); 3621 return 0; 3622 } 3623 3624 /* 3625 * wire case: in three passes [XXXCDC: ugly block of code here] 3626 * 3627 * 1: holding the write lock, count all pages mapped by non-wired 3628 * entries. if this would cause us to go over our limit, we fail. 3629 * 3630 * 2: still holding the write lock, we create any anonymous maps that 3631 * need to be created. then we increment its wiring count. 3632 * 3633 * 3: we downgrade to a read lock, and call uvm_fault_wire to fault 3634 * in the pages for any newly wired area (wired_count == 1). 3635 * 3636 * downgrading to a read lock for uvm_fault_wire avoids a possible 3637 * deadlock with another thread that may have faulted on one of 3638 * the pages to be wired (it would mark the page busy, blocking 3639 * us, then in turn block on the map lock that we hold). because 3640 * of problems in the recursive lock package, we cannot upgrade 3641 * to a write lock in vm_map_lookup. thus, any actions that 3642 * require the write lock must be done beforehand. because we 3643 * keep the read lock on the map, the copy-on-write status of the 3644 * entries we modify here cannot change. 3645 */ 3646 3647 for (size = 0, entry = map->header.next; entry != &map->header; 3648 entry = entry->next) { 3649 if (entry->protection != VM_PROT_NONE && 3650 VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3651 size += entry->end - entry->start; 3652 } 3653 } 3654 3655 if (atop(size) + uvmexp.wired > uvmexp.wiredmax) { 3656 vm_map_unlock(map); 3657 return ENOMEM; 3658 } 3659 3660 if (limit != 0 && 3661 (size + ptoa(pmap_wired_count(vm_map_pmap(map))) > limit)) { 3662 vm_map_unlock(map); 3663 return ENOMEM; 3664 } 3665 3666 /* 3667 * Pass 2. 3668 */ 3669 3670 for (entry = map->header.next; entry != &map->header; 3671 entry = entry->next) { 3672 if (entry->protection == VM_PROT_NONE) 3673 continue; 3674 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3675 3676 /* 3677 * perform actions of vm_map_lookup that need the 3678 * write lock on the map: create an anonymous map 3679 * for a copy-on-write region, or an anonymous map 3680 * for a zero-fill region. (XXXCDC: submap case 3681 * ok?) 3682 */ 3683 3684 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3685 if (UVM_ET_ISNEEDSCOPY(entry) && 3686 ((entry->max_protection & VM_PROT_WRITE) || 3687 (entry->object.uvm_obj == NULL))) { 3688 amap_copy(map, entry, 0, entry->start, 3689 entry->end); 3690 /* XXXCDC: wait OK? */ 3691 } 3692 } 3693 } 3694 entry->wired_count++; 3695 } 3696 3697 /* 3698 * Pass 3. 3699 */ 3700 3701 #ifdef DIAGNOSTIC 3702 timestamp_save = map->timestamp; 3703 #endif 3704 vm_map_busy(map); 3705 vm_map_unlock(map); 3706 3707 rv = 0; 3708 for (entry = map->header.next; entry != &map->header; 3709 entry = entry->next) { 3710 if (entry->wired_count == 1) { 3711 rv = uvm_fault_wire(map, entry->start, entry->end, 3712 entry->max_protection, 1); 3713 if (rv) { 3714 3715 /* 3716 * wiring failed. break out of the loop. 3717 * we'll clean up the map below, once we 3718 * have a write lock again. 3719 */ 3720 3721 break; 3722 } 3723 } 3724 } 3725 3726 if (rv) { 3727 3728 /* 3729 * Get back an exclusive (write) lock. 3730 */ 3731 3732 vm_map_lock(map); 3733 vm_map_unbusy(map); 3734 3735 #ifdef DIAGNOSTIC 3736 if (timestamp_save + 1 != map->timestamp) 3737 panic("uvm_map_pageable_all: stale map"); 3738 #endif 3739 3740 /* 3741 * first drop the wiring count on all the entries 3742 * which haven't actually been wired yet. 3743 * 3744 * Skip VM_PROT_NONE entries like we did above. 3745 */ 3746 3747 failed_entry = entry; 3748 for (/* nothing */; entry != &map->header; 3749 entry = entry->next) { 3750 if (entry->protection == VM_PROT_NONE) 3751 continue; 3752 entry->wired_count--; 3753 } 3754 3755 /* 3756 * now, unwire all the entries that were successfully 3757 * wired above. 3758 * 3759 * Skip VM_PROT_NONE entries like we did above. 3760 */ 3761 3762 for (entry = map->header.next; entry != failed_entry; 3763 entry = entry->next) { 3764 if (entry->protection == VM_PROT_NONE) 3765 continue; 3766 entry->wired_count--; 3767 if (VM_MAPENT_ISWIRED(entry)) 3768 uvm_map_entry_unwire(map, entry); 3769 } 3770 vm_map_unlock(map); 3771 UVMHIST_LOG(maphist,"<- done (RV=%jd)", rv,0,0,0); 3772 return (rv); 3773 } 3774 3775 vm_map_unbusy(map); 3776 3777 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3778 return 0; 3779 } 3780 3781 /* 3782 * uvm_map_clean: clean out a map range 3783 * 3784 * => valid flags: 3785 * if (flags & PGO_CLEANIT): dirty pages are cleaned first 3786 * if (flags & PGO_SYNCIO): dirty pages are written synchronously 3787 * if (flags & PGO_DEACTIVATE): any cached pages are deactivated after clean 3788 * if (flags & PGO_FREE): any cached pages are freed after clean 3789 * => returns an error if any part of the specified range isn't mapped 3790 * => never a need to flush amap layer since the anonymous memory has 3791 * no permanent home, but may deactivate pages there 3792 * => called from sys_msync() and sys_madvise() 3793 * => caller must not write-lock map (read OK). 3794 * => we may sleep while cleaning if SYNCIO [with map read-locked] 3795 */ 3796 3797 int 3798 uvm_map_clean(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 3799 { 3800 struct vm_map_entry *current, *entry; 3801 struct uvm_object *uobj; 3802 struct vm_amap *amap; 3803 struct vm_anon *anon; 3804 struct vm_page *pg; 3805 vaddr_t offset; 3806 vsize_t size; 3807 voff_t uoff; 3808 int error, refs; 3809 UVMHIST_FUNC(__func__); 3810 UVMHIST_CALLARGS(maphist,"(map=%#jx,start=%#jx,end=%#jx,flags=%#jx)", 3811 (uintptr_t)map, start, end, flags); 3812 3813 KASSERT((flags & (PGO_FREE|PGO_DEACTIVATE)) != 3814 (PGO_FREE|PGO_DEACTIVATE)); 3815 3816 vm_map_lock_read(map); 3817 VM_MAP_RANGE_CHECK(map, start, end); 3818 if (uvm_map_lookup_entry(map, start, &entry) == false) { 3819 vm_map_unlock_read(map); 3820 return EFAULT; 3821 } 3822 3823 /* 3824 * Make a first pass to check for holes and wiring problems. 3825 */ 3826 3827 for (current = entry; current->start < end; current = current->next) { 3828 if (UVM_ET_ISSUBMAP(current)) { 3829 vm_map_unlock_read(map); 3830 return EINVAL; 3831 } 3832 if ((flags & PGO_FREE) != 0 && VM_MAPENT_ISWIRED(entry)) { 3833 vm_map_unlock_read(map); 3834 return EBUSY; 3835 } 3836 if (end <= current->end) { 3837 break; 3838 } 3839 if (current->end != current->next->start) { 3840 vm_map_unlock_read(map); 3841 return EFAULT; 3842 } 3843 } 3844 3845 error = 0; 3846 for (current = entry; start < end; current = current->next) { 3847 amap = current->aref.ar_amap; /* upper layer */ 3848 uobj = current->object.uvm_obj; /* lower layer */ 3849 KASSERT(start >= current->start); 3850 3851 /* 3852 * No amap cleaning necessary if: 3853 * 3854 * (1) There's no amap. 3855 * 3856 * (2) We're not deactivating or freeing pages. 3857 */ 3858 3859 if (amap == NULL || (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) 3860 goto flush_object; 3861 3862 offset = start - current->start; 3863 size = MIN(end, current->end) - start; 3864 3865 amap_lock(amap, RW_WRITER); 3866 for ( ; size != 0; size -= PAGE_SIZE, offset += PAGE_SIZE) { 3867 anon = amap_lookup(¤t->aref, offset); 3868 if (anon == NULL) 3869 continue; 3870 3871 KASSERT(anon->an_lock == amap->am_lock); 3872 pg = anon->an_page; 3873 if (pg == NULL) { 3874 continue; 3875 } 3876 if (pg->flags & PG_BUSY) { 3877 continue; 3878 } 3879 3880 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 3881 3882 /* 3883 * In these first 3 cases, we just deactivate the page. 3884 */ 3885 3886 case PGO_CLEANIT|PGO_FREE: 3887 case PGO_CLEANIT|PGO_DEACTIVATE: 3888 case PGO_DEACTIVATE: 3889 deactivate_it: 3890 /* 3891 * skip the page if it's loaned or wired, 3892 * since it shouldn't be on a paging queue 3893 * at all in these cases. 3894 */ 3895 3896 if (pg->loan_count != 0 || 3897 pg->wire_count != 0) { 3898 continue; 3899 } 3900 KASSERT(pg->uanon == anon); 3901 uvm_pagelock(pg); 3902 uvm_pagedeactivate(pg); 3903 uvm_pageunlock(pg); 3904 continue; 3905 3906 case PGO_FREE: 3907 3908 /* 3909 * If there are multiple references to 3910 * the amap, just deactivate the page. 3911 */ 3912 3913 if (amap_refs(amap) > 1) 3914 goto deactivate_it; 3915 3916 /* skip the page if it's wired */ 3917 if (pg->wire_count != 0) { 3918 continue; 3919 } 3920 amap_unadd(¤t->aref, offset); 3921 refs = --anon->an_ref; 3922 if (refs == 0) { 3923 uvm_anfree(anon); 3924 } 3925 continue; 3926 } 3927 } 3928 amap_unlock(amap); 3929 3930 flush_object: 3931 /* 3932 * flush pages if we've got a valid backing object. 3933 * note that we must always clean object pages before 3934 * freeing them since otherwise we could reveal stale 3935 * data from files. 3936 */ 3937 3938 uoff = current->offset + (start - current->start); 3939 size = MIN(end, current->end) - start; 3940 if (uobj != NULL) { 3941 rw_enter(uobj->vmobjlock, RW_WRITER); 3942 if (uobj->pgops->pgo_put != NULL) 3943 error = (uobj->pgops->pgo_put)(uobj, uoff, 3944 uoff + size, flags | PGO_CLEANIT); 3945 else 3946 error = 0; 3947 } 3948 start += size; 3949 } 3950 vm_map_unlock_read(map); 3951 return (error); 3952 } 3953 3954 3955 /* 3956 * uvm_map_checkprot: check protection in map 3957 * 3958 * => must allow specified protection in a fully allocated region. 3959 * => map must be read or write locked by caller. 3960 */ 3961 3962 bool 3963 uvm_map_checkprot(struct vm_map *map, vaddr_t start, vaddr_t end, 3964 vm_prot_t protection) 3965 { 3966 struct vm_map_entry *entry; 3967 struct vm_map_entry *tmp_entry; 3968 3969 if (!uvm_map_lookup_entry(map, start, &tmp_entry)) { 3970 return (false); 3971 } 3972 entry = tmp_entry; 3973 while (start < end) { 3974 if (entry == &map->header) { 3975 return (false); 3976 } 3977 3978 /* 3979 * no holes allowed 3980 */ 3981 3982 if (start < entry->start) { 3983 return (false); 3984 } 3985 3986 /* 3987 * check protection associated with entry 3988 */ 3989 3990 if ((entry->protection & protection) != protection) { 3991 return (false); 3992 } 3993 start = entry->end; 3994 entry = entry->next; 3995 } 3996 return (true); 3997 } 3998 3999 /* 4000 * uvmspace_alloc: allocate a vmspace structure. 4001 * 4002 * - structure includes vm_map and pmap 4003 * - XXX: no locking on this structure 4004 * - refcnt set to 1, rest must be init'd by caller 4005 */ 4006 struct vmspace * 4007 uvmspace_alloc(vaddr_t vmin, vaddr_t vmax, bool topdown) 4008 { 4009 struct vmspace *vm; 4010 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 4011 4012 vm = pool_cache_get(&uvm_vmspace_cache, PR_WAITOK); 4013 uvmspace_init(vm, NULL, vmin, vmax, topdown); 4014 UVMHIST_LOG(maphist,"<- done (vm=%#jx)", (uintptr_t)vm, 0, 0, 0); 4015 return (vm); 4016 } 4017 4018 /* 4019 * uvmspace_init: initialize a vmspace structure. 4020 * 4021 * - XXX: no locking on this structure 4022 * - refcnt set to 1, rest must be init'd by caller 4023 */ 4024 void 4025 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t vmin, 4026 vaddr_t vmax, bool topdown) 4027 { 4028 UVMHIST_FUNC(__func__); 4029 UVMHIST_CALLARGS(maphist, "(vm=%#jx, pmap=%#jx, vmin=%#jx, vmax=%#jx", 4030 (uintptr_t)vm, (uintptr_t)pmap, vmin, vmax); 4031 UVMHIST_LOG(maphist, " topdown=%ju)", topdown, 0, 0, 0); 4032 4033 memset(vm, 0, sizeof(*vm)); 4034 uvm_map_setup(&vm->vm_map, vmin, vmax, VM_MAP_PAGEABLE 4035 | (topdown ? VM_MAP_TOPDOWN : 0) 4036 ); 4037 if (pmap) 4038 pmap_reference(pmap); 4039 else 4040 pmap = pmap_create(); 4041 vm->vm_map.pmap = pmap; 4042 vm->vm_refcnt = 1; 4043 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4044 } 4045 4046 /* 4047 * uvmspace_share: share a vmspace between two processes 4048 * 4049 * - used for vfork, threads(?) 4050 */ 4051 4052 void 4053 uvmspace_share(struct proc *p1, struct proc *p2) 4054 { 4055 4056 uvmspace_addref(p1->p_vmspace); 4057 p2->p_vmspace = p1->p_vmspace; 4058 } 4059 4060 #if 0 4061 4062 /* 4063 * uvmspace_unshare: ensure that process "p" has its own, unshared, vmspace 4064 * 4065 * - XXX: no locking on vmspace 4066 */ 4067 4068 void 4069 uvmspace_unshare(struct lwp *l) 4070 { 4071 struct proc *p = l->l_proc; 4072 struct vmspace *nvm, *ovm = p->p_vmspace; 4073 4074 if (ovm->vm_refcnt == 1) 4075 /* nothing to do: vmspace isn't shared in the first place */ 4076 return; 4077 4078 /* make a new vmspace, still holding old one */ 4079 nvm = uvmspace_fork(ovm); 4080 4081 kpreempt_disable(); 4082 pmap_deactivate(l); /* unbind old vmspace */ 4083 p->p_vmspace = nvm; 4084 pmap_activate(l); /* switch to new vmspace */ 4085 kpreempt_enable(); 4086 4087 uvmspace_free(ovm); /* drop reference to old vmspace */ 4088 } 4089 4090 #endif 4091 4092 4093 /* 4094 * uvmspace_spawn: a new process has been spawned and needs a vmspace 4095 */ 4096 4097 void 4098 uvmspace_spawn(struct lwp *l, vaddr_t start, vaddr_t end, bool topdown) 4099 { 4100 struct proc *p = l->l_proc; 4101 struct vmspace *nvm; 4102 4103 #ifdef __HAVE_CPU_VMSPACE_EXEC 4104 cpu_vmspace_exec(l, start, end); 4105 #endif 4106 4107 nvm = uvmspace_alloc(start, end, topdown); 4108 kpreempt_disable(); 4109 p->p_vmspace = nvm; 4110 pmap_activate(l); 4111 kpreempt_enable(); 4112 } 4113 4114 /* 4115 * uvmspace_exec: the process wants to exec a new program 4116 */ 4117 4118 void 4119 uvmspace_exec(struct lwp *l, vaddr_t start, vaddr_t end, bool topdown) 4120 { 4121 struct proc *p = l->l_proc; 4122 struct vmspace *nvm, *ovm = p->p_vmspace; 4123 struct vm_map *map; 4124 int flags; 4125 4126 KASSERT(ovm != NULL); 4127 #ifdef __HAVE_CPU_VMSPACE_EXEC 4128 cpu_vmspace_exec(l, start, end); 4129 #endif 4130 4131 map = &ovm->vm_map; 4132 /* 4133 * see if more than one process is using this vmspace... 4134 */ 4135 4136 if (ovm->vm_refcnt == 1 4137 && topdown == ((ovm->vm_map.flags & VM_MAP_TOPDOWN) != 0)) { 4138 4139 /* 4140 * if p is the only process using its vmspace then we can safely 4141 * recycle that vmspace for the program that is being exec'd. 4142 * But only if TOPDOWN matches the requested value for the new 4143 * vm space! 4144 */ 4145 4146 /* 4147 * SYSV SHM semantics require us to kill all segments on an exec 4148 */ 4149 if (uvm_shmexit && ovm->vm_shm) 4150 (*uvm_shmexit)(ovm); 4151 4152 /* 4153 * POSIX 1003.1b -- "lock future mappings" is revoked 4154 * when a process execs another program image. 4155 */ 4156 4157 map->flags &= ~VM_MAP_WIREFUTURE; 4158 4159 /* 4160 * now unmap the old program. 4161 * 4162 * XXX set VM_MAP_DYING for the duration, so pmap_update() 4163 * is not called until the pmap has been totally cleared out 4164 * after pmap_remove_all(), or it can confuse some pmap 4165 * implementations. it would be nice to handle this by 4166 * deferring the pmap_update() while it is known the address 4167 * space is not visible to any user LWP other than curlwp, 4168 * but there isn't an elegant way of inferring that right 4169 * now. 4170 */ 4171 4172 flags = pmap_remove_all(map->pmap) ? UVM_FLAG_VAONLY : 0; 4173 map->flags |= VM_MAP_DYING; 4174 uvm_unmap1(map, vm_map_min(map), vm_map_max(map), flags); 4175 map->flags &= ~VM_MAP_DYING; 4176 pmap_update(map->pmap); 4177 KASSERT(map->header.prev == &map->header); 4178 KASSERT(map->nentries == 0); 4179 4180 /* 4181 * resize the map 4182 */ 4183 4184 vm_map_setmin(map, start); 4185 vm_map_setmax(map, end); 4186 } else { 4187 4188 /* 4189 * p's vmspace is being shared, so we can't reuse it for p since 4190 * it is still being used for others. allocate a new vmspace 4191 * for p 4192 */ 4193 4194 nvm = uvmspace_alloc(start, end, topdown); 4195 4196 /* 4197 * install new vmspace and drop our ref to the old one. 4198 */ 4199 4200 kpreempt_disable(); 4201 pmap_deactivate(l); 4202 p->p_vmspace = nvm; 4203 pmap_activate(l); 4204 kpreempt_enable(); 4205 4206 uvmspace_free(ovm); 4207 } 4208 } 4209 4210 /* 4211 * uvmspace_addref: add a reference to a vmspace. 4212 */ 4213 4214 void 4215 uvmspace_addref(struct vmspace *vm) 4216 { 4217 4218 KASSERT((vm->vm_map.flags & VM_MAP_DYING) == 0); 4219 KASSERT(vm->vm_refcnt > 0); 4220 atomic_inc_uint(&vm->vm_refcnt); 4221 } 4222 4223 /* 4224 * uvmspace_free: free a vmspace data structure 4225 */ 4226 4227 void 4228 uvmspace_free(struct vmspace *vm) 4229 { 4230 struct vm_map_entry *dead_entries; 4231 struct vm_map *map = &vm->vm_map; 4232 int flags; 4233 4234 UVMHIST_FUNC(__func__); 4235 UVMHIST_CALLARGS(maphist,"(vm=%#jx) ref=%jd", (uintptr_t)vm, 4236 vm->vm_refcnt, 0, 0); 4237 if (atomic_dec_uint_nv(&vm->vm_refcnt) > 0) 4238 return; 4239 4240 /* 4241 * at this point, there should be no other references to the map. 4242 * delete all of the mappings, then destroy the pmap. 4243 */ 4244 4245 map->flags |= VM_MAP_DYING; 4246 flags = pmap_remove_all(map->pmap) ? UVM_FLAG_VAONLY : 0; 4247 4248 /* Get rid of any SYSV shared memory segments. */ 4249 if (uvm_shmexit && vm->vm_shm != NULL) 4250 (*uvm_shmexit)(vm); 4251 4252 if (map->nentries) { 4253 uvm_unmap_remove(map, vm_map_min(map), vm_map_max(map), 4254 &dead_entries, flags); 4255 if (dead_entries != NULL) 4256 uvm_unmap_detach(dead_entries, 0); 4257 } 4258 KASSERT(map->nentries == 0); 4259 KASSERT(map->size == 0); 4260 4261 mutex_destroy(&map->misc_lock); 4262 rw_destroy(&map->lock); 4263 cv_destroy(&map->cv); 4264 pmap_destroy(map->pmap); 4265 pool_cache_put(&uvm_vmspace_cache, vm); 4266 } 4267 4268 static struct vm_map_entry * 4269 uvm_mapent_clone(struct vm_map *new_map, struct vm_map_entry *old_entry, 4270 int flags) 4271 { 4272 struct vm_map_entry *new_entry; 4273 4274 new_entry = uvm_mapent_alloc(new_map, 0); 4275 /* old_entry -> new_entry */ 4276 uvm_mapent_copy(old_entry, new_entry); 4277 4278 /* new pmap has nothing wired in it */ 4279 new_entry->wired_count = 0; 4280 4281 /* 4282 * gain reference to object backing the map (can't 4283 * be a submap, already checked this case). 4284 */ 4285 4286 if (new_entry->aref.ar_amap) 4287 uvm_map_reference_amap(new_entry, flags); 4288 4289 if (new_entry->object.uvm_obj && 4290 new_entry->object.uvm_obj->pgops->pgo_reference) 4291 new_entry->object.uvm_obj->pgops->pgo_reference( 4292 new_entry->object.uvm_obj); 4293 4294 /* insert entry at end of new_map's entry list */ 4295 uvm_map_entry_link(new_map, new_map->header.prev, 4296 new_entry); 4297 4298 return new_entry; 4299 } 4300 4301 /* 4302 * share the mapping: this means we want the old and 4303 * new entries to share amaps and backing objects. 4304 */ 4305 static void 4306 uvm_mapent_forkshared(struct vm_map *new_map, struct vm_map *old_map, 4307 struct vm_map_entry *old_entry) 4308 { 4309 /* 4310 * if the old_entry needs a new amap (due to prev fork) 4311 * then we need to allocate it now so that we have 4312 * something we own to share with the new_entry. [in 4313 * other words, we need to clear needs_copy] 4314 */ 4315 4316 if (UVM_ET_ISNEEDSCOPY(old_entry)) { 4317 /* get our own amap, clears needs_copy */ 4318 amap_copy(old_map, old_entry, AMAP_COPY_NOCHUNK, 4319 0, 0); 4320 /* XXXCDC: WAITOK??? */ 4321 } 4322 4323 uvm_mapent_clone(new_map, old_entry, AMAP_SHARED); 4324 } 4325 4326 4327 static void 4328 uvm_mapent_forkcopy(struct vm_map *new_map, struct vm_map *old_map, 4329 struct vm_map_entry *old_entry) 4330 { 4331 struct vm_map_entry *new_entry; 4332 4333 /* 4334 * copy-on-write the mapping (using mmap's 4335 * MAP_PRIVATE semantics) 4336 * 4337 * allocate new_entry, adjust reference counts. 4338 * (note that new references are read-only). 4339 */ 4340 4341 new_entry = uvm_mapent_clone(new_map, old_entry, 0); 4342 4343 new_entry->etype |= 4344 (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY); 4345 4346 /* 4347 * the new entry will need an amap. it will either 4348 * need to be copied from the old entry or created 4349 * from scratch (if the old entry does not have an 4350 * amap). can we defer this process until later 4351 * (by setting "needs_copy") or do we need to copy 4352 * the amap now? 4353 * 4354 * we must copy the amap now if any of the following 4355 * conditions hold: 4356 * 1. the old entry has an amap and that amap is 4357 * being shared. this means that the old (parent) 4358 * process is sharing the amap with another 4359 * process. if we do not clear needs_copy here 4360 * we will end up in a situation where both the 4361 * parent and child process are refering to the 4362 * same amap with "needs_copy" set. if the 4363 * parent write-faults, the fault routine will 4364 * clear "needs_copy" in the parent by allocating 4365 * a new amap. this is wrong because the 4366 * parent is supposed to be sharing the old amap 4367 * and the new amap will break that. 4368 * 4369 * 2. if the old entry has an amap and a non-zero 4370 * wire count then we are going to have to call 4371 * amap_cow_now to avoid page faults in the 4372 * parent process. since amap_cow_now requires 4373 * "needs_copy" to be clear we might as well 4374 * clear it here as well. 4375 * 4376 */ 4377 4378 if (old_entry->aref.ar_amap != NULL) { 4379 if ((amap_flags(old_entry->aref.ar_amap) & AMAP_SHARED) != 0 || 4380 VM_MAPENT_ISWIRED(old_entry)) { 4381 4382 amap_copy(new_map, new_entry, 4383 AMAP_COPY_NOCHUNK, 0, 0); 4384 /* XXXCDC: M_WAITOK ... ok? */ 4385 } 4386 } 4387 4388 /* 4389 * if the parent's entry is wired down, then the 4390 * parent process does not want page faults on 4391 * access to that memory. this means that we 4392 * cannot do copy-on-write because we can't write 4393 * protect the old entry. in this case we 4394 * resolve all copy-on-write faults now, using 4395 * amap_cow_now. note that we have already 4396 * allocated any needed amap (above). 4397 */ 4398 4399 if (VM_MAPENT_ISWIRED(old_entry)) { 4400 4401 /* 4402 * resolve all copy-on-write faults now 4403 * (note that there is nothing to do if 4404 * the old mapping does not have an amap). 4405 */ 4406 if (old_entry->aref.ar_amap) 4407 amap_cow_now(new_map, new_entry); 4408 4409 } else { 4410 /* 4411 * setup mappings to trigger copy-on-write faults 4412 * we must write-protect the parent if it has 4413 * an amap and it is not already "needs_copy"... 4414 * if it is already "needs_copy" then the parent 4415 * has already been write-protected by a previous 4416 * fork operation. 4417 */ 4418 if (old_entry->aref.ar_amap && 4419 !UVM_ET_ISNEEDSCOPY(old_entry)) { 4420 if (old_entry->max_protection & VM_PROT_WRITE) { 4421 #ifdef __HAVE_UNLOCKED_PMAP /* XXX temporary */ 4422 uvm_map_lock_entry(old_entry, RW_WRITER); 4423 #else 4424 uvm_map_lock_entry(old_entry, RW_READER); 4425 #endif 4426 pmap_protect(old_map->pmap, 4427 old_entry->start, old_entry->end, 4428 old_entry->protection & ~VM_PROT_WRITE); 4429 uvm_map_unlock_entry(old_entry); 4430 } 4431 old_entry->etype |= UVM_ET_NEEDSCOPY; 4432 } 4433 } 4434 } 4435 4436 /* 4437 * zero the mapping: the new entry will be zero initialized 4438 */ 4439 static void 4440 uvm_mapent_forkzero(struct vm_map *new_map, struct vm_map *old_map, 4441 struct vm_map_entry *old_entry) 4442 { 4443 struct vm_map_entry *new_entry; 4444 4445 new_entry = uvm_mapent_clone(new_map, old_entry, 0); 4446 4447 new_entry->etype |= 4448 (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY); 4449 4450 if (new_entry->aref.ar_amap) { 4451 uvm_map_unreference_amap(new_entry, 0); 4452 new_entry->aref.ar_pageoff = 0; 4453 new_entry->aref.ar_amap = NULL; 4454 } 4455 4456 if (UVM_ET_ISOBJ(new_entry)) { 4457 if (new_entry->object.uvm_obj->pgops->pgo_detach) 4458 new_entry->object.uvm_obj->pgops->pgo_detach( 4459 new_entry->object.uvm_obj); 4460 new_entry->object.uvm_obj = NULL; 4461 new_entry->offset = 0; 4462 new_entry->etype &= ~UVM_ET_OBJ; 4463 } 4464 } 4465 4466 /* 4467 * F O R K - m a i n e n t r y p o i n t 4468 */ 4469 /* 4470 * uvmspace_fork: fork a process' main map 4471 * 4472 * => create a new vmspace for child process from parent. 4473 * => parent's map must not be locked. 4474 */ 4475 4476 struct vmspace * 4477 uvmspace_fork(struct vmspace *vm1) 4478 { 4479 struct vmspace *vm2; 4480 struct vm_map *old_map = &vm1->vm_map; 4481 struct vm_map *new_map; 4482 struct vm_map_entry *old_entry; 4483 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 4484 4485 vm_map_lock(old_map); 4486 4487 vm2 = uvmspace_alloc(vm_map_min(old_map), vm_map_max(old_map), 4488 vm1->vm_map.flags & VM_MAP_TOPDOWN); 4489 memcpy(&vm2->vm_startcopy, &vm1->vm_startcopy, 4490 (char *) (vm1 + 1) - (char *) &vm1->vm_startcopy); 4491 new_map = &vm2->vm_map; /* XXX */ 4492 4493 old_entry = old_map->header.next; 4494 new_map->size = old_map->size; 4495 4496 /* 4497 * go entry-by-entry 4498 */ 4499 4500 while (old_entry != &old_map->header) { 4501 4502 /* 4503 * first, some sanity checks on the old entry 4504 */ 4505 4506 KASSERT(!UVM_ET_ISSUBMAP(old_entry)); 4507 KASSERT(UVM_ET_ISCOPYONWRITE(old_entry) || 4508 !UVM_ET_ISNEEDSCOPY(old_entry)); 4509 4510 switch (old_entry->inheritance) { 4511 case MAP_INHERIT_NONE: 4512 /* 4513 * drop the mapping, modify size 4514 */ 4515 new_map->size -= old_entry->end - old_entry->start; 4516 break; 4517 4518 case MAP_INHERIT_SHARE: 4519 uvm_mapent_forkshared(new_map, old_map, old_entry); 4520 break; 4521 4522 case MAP_INHERIT_COPY: 4523 uvm_mapent_forkcopy(new_map, old_map, old_entry); 4524 break; 4525 4526 case MAP_INHERIT_ZERO: 4527 uvm_mapent_forkzero(new_map, old_map, old_entry); 4528 break; 4529 default: 4530 KASSERT(0); 4531 break; 4532 } 4533 old_entry = old_entry->next; 4534 } 4535 4536 pmap_update(old_map->pmap); 4537 vm_map_unlock(old_map); 4538 4539 if (uvm_shmfork && vm1->vm_shm) 4540 (*uvm_shmfork)(vm1, vm2); 4541 4542 #ifdef PMAP_FORK 4543 pmap_fork(vm1->vm_map.pmap, vm2->vm_map.pmap); 4544 #endif 4545 4546 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4547 return (vm2); 4548 } 4549 4550 4551 /* 4552 * uvm_mapent_trymerge: try to merge an entry with its neighbors. 4553 * 4554 * => called with map locked. 4555 * => return non zero if successfully merged. 4556 */ 4557 4558 int 4559 uvm_mapent_trymerge(struct vm_map *map, struct vm_map_entry *entry, int flags) 4560 { 4561 struct uvm_object *uobj; 4562 struct vm_map_entry *next; 4563 struct vm_map_entry *prev; 4564 vsize_t size; 4565 int merged = 0; 4566 bool copying; 4567 int newetype; 4568 4569 if (entry->aref.ar_amap != NULL) { 4570 return 0; 4571 } 4572 if ((entry->flags & UVM_MAP_NOMERGE) != 0) { 4573 return 0; 4574 } 4575 4576 uobj = entry->object.uvm_obj; 4577 size = entry->end - entry->start; 4578 copying = (flags & UVM_MERGE_COPYING) != 0; 4579 newetype = copying ? (entry->etype & ~UVM_ET_NEEDSCOPY) : entry->etype; 4580 4581 next = entry->next; 4582 if (next != &map->header && 4583 next->start == entry->end && 4584 ((copying && next->aref.ar_amap != NULL && 4585 amap_refs(next->aref.ar_amap) == 1) || 4586 (!copying && next->aref.ar_amap == NULL)) && 4587 UVM_ET_ISCOMPATIBLE(next, newetype, 4588 uobj, entry->flags, entry->protection, 4589 entry->max_protection, entry->inheritance, entry->advice, 4590 entry->wired_count) && 4591 (uobj == NULL || entry->offset + size == next->offset)) { 4592 int error; 4593 4594 if (copying) { 4595 error = amap_extend(next, size, 4596 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_BACKWARDS); 4597 } else { 4598 error = 0; 4599 } 4600 if (error == 0) { 4601 if (uobj) { 4602 if (uobj->pgops->pgo_detach) { 4603 uobj->pgops->pgo_detach(uobj); 4604 } 4605 } 4606 4607 entry->end = next->end; 4608 clear_hints(map, next); 4609 uvm_map_entry_unlink(map, next); 4610 if (copying) { 4611 entry->aref = next->aref; 4612 entry->etype &= ~UVM_ET_NEEDSCOPY; 4613 } 4614 uvm_map_check(map, "trymerge forwardmerge"); 4615 uvm_mapent_free(next); 4616 merged++; 4617 } 4618 } 4619 4620 prev = entry->prev; 4621 if (prev != &map->header && 4622 prev->end == entry->start && 4623 ((copying && !merged && prev->aref.ar_amap != NULL && 4624 amap_refs(prev->aref.ar_amap) == 1) || 4625 (!copying && prev->aref.ar_amap == NULL)) && 4626 UVM_ET_ISCOMPATIBLE(prev, newetype, 4627 uobj, entry->flags, entry->protection, 4628 entry->max_protection, entry->inheritance, entry->advice, 4629 entry->wired_count) && 4630 (uobj == NULL || 4631 prev->offset + prev->end - prev->start == entry->offset)) { 4632 int error; 4633 4634 if (copying) { 4635 error = amap_extend(prev, size, 4636 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_FORWARDS); 4637 } else { 4638 error = 0; 4639 } 4640 if (error == 0) { 4641 if (uobj) { 4642 if (uobj->pgops->pgo_detach) { 4643 uobj->pgops->pgo_detach(uobj); 4644 } 4645 entry->offset = prev->offset; 4646 } 4647 4648 entry->start = prev->start; 4649 clear_hints(map, prev); 4650 uvm_map_entry_unlink(map, prev); 4651 if (copying) { 4652 entry->aref = prev->aref; 4653 entry->etype &= ~UVM_ET_NEEDSCOPY; 4654 } 4655 uvm_map_check(map, "trymerge backmerge"); 4656 uvm_mapent_free(prev); 4657 merged++; 4658 } 4659 } 4660 4661 return merged; 4662 } 4663 4664 /* 4665 * uvm_map_setup: init map 4666 * 4667 * => map must not be in service yet. 4668 */ 4669 4670 void 4671 uvm_map_setup(struct vm_map *map, vaddr_t vmin, vaddr_t vmax, int flags) 4672 { 4673 4674 rb_tree_init(&map->rb_tree, &uvm_map_tree_ops); 4675 map->header.next = map->header.prev = &map->header; 4676 map->nentries = 0; 4677 map->size = 0; 4678 map->ref_count = 1; 4679 vm_map_setmin(map, vmin); 4680 vm_map_setmax(map, vmax); 4681 map->flags = flags; 4682 map->first_free = &map->header; 4683 map->hint = &map->header; 4684 map->timestamp = 0; 4685 map->busy = NULL; 4686 4687 rw_init(&map->lock); 4688 cv_init(&map->cv, "vm_map"); 4689 mutex_init(&map->misc_lock, MUTEX_DRIVER, IPL_NONE); 4690 } 4691 4692 /* 4693 * U N M A P - m a i n e n t r y p o i n t 4694 */ 4695 4696 /* 4697 * uvm_unmap1: remove mappings from a vm_map (from "start" up to "stop") 4698 * 4699 * => caller must check alignment and size 4700 * => map must be unlocked (we will lock it) 4701 * => flags is UVM_FLAG_QUANTUM or 0. 4702 */ 4703 4704 void 4705 uvm_unmap1(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 4706 { 4707 struct vm_map_entry *dead_entries; 4708 UVMHIST_FUNC(__func__); 4709 UVMHIST_CALLARGS(maphist, " (map=%#jx, start=%#jx, end=%#jx)", 4710 (uintptr_t)map, start, end, 0); 4711 4712 KASSERTMSG(start < end, 4713 "%s: map %p: start %#jx < end %#jx", __func__, map, 4714 (uintmax_t)start, (uintmax_t)end); 4715 if (map == kernel_map) { 4716 LOCKDEBUG_MEM_CHECK((void *)start, end - start); 4717 } 4718 4719 /* 4720 * work now done by helper functions. wipe the pmap's and then 4721 * detach from the dead entries... 4722 */ 4723 vm_map_lock(map); 4724 uvm_unmap_remove(map, start, end, &dead_entries, flags); 4725 vm_map_unlock(map); 4726 4727 if (dead_entries != NULL) 4728 uvm_unmap_detach(dead_entries, 0); 4729 4730 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 4731 } 4732 4733 4734 /* 4735 * uvm_map_reference: add reference to a map 4736 * 4737 * => map need not be locked 4738 */ 4739 4740 void 4741 uvm_map_reference(struct vm_map *map) 4742 { 4743 4744 atomic_inc_uint(&map->ref_count); 4745 } 4746 4747 void 4748 uvm_map_lock_entry(struct vm_map_entry *entry, krw_t op) 4749 { 4750 4751 if (entry->aref.ar_amap != NULL) { 4752 amap_lock(entry->aref.ar_amap, op); 4753 } 4754 if (UVM_ET_ISOBJ(entry)) { 4755 rw_enter(entry->object.uvm_obj->vmobjlock, op); 4756 } 4757 } 4758 4759 void 4760 uvm_map_unlock_entry(struct vm_map_entry *entry) 4761 { 4762 4763 if (UVM_ET_ISOBJ(entry)) { 4764 rw_exit(entry->object.uvm_obj->vmobjlock); 4765 } 4766 if (entry->aref.ar_amap != NULL) { 4767 amap_unlock(entry->aref.ar_amap); 4768 } 4769 } 4770 4771 #define UVM_VOADDR_TYPE_MASK 0x3UL 4772 #define UVM_VOADDR_TYPE_UOBJ 0x1UL 4773 #define UVM_VOADDR_TYPE_ANON 0x2UL 4774 #define UVM_VOADDR_OBJECT_MASK ~UVM_VOADDR_TYPE_MASK 4775 4776 #define UVM_VOADDR_GET_TYPE(voa) \ 4777 ((voa)->object & UVM_VOADDR_TYPE_MASK) 4778 #define UVM_VOADDR_GET_OBJECT(voa) \ 4779 ((voa)->object & UVM_VOADDR_OBJECT_MASK) 4780 #define UVM_VOADDR_SET_OBJECT(voa, obj, type) \ 4781 do { \ 4782 KASSERT(((uintptr_t)(obj) & UVM_VOADDR_TYPE_MASK) == 0); \ 4783 (voa)->object = ((uintptr_t)(obj)) | (type); \ 4784 } while (/*CONSTCOND*/0) 4785 4786 #define UVM_VOADDR_GET_UOBJ(voa) \ 4787 ((struct uvm_object *)UVM_VOADDR_GET_OBJECT(voa)) 4788 #define UVM_VOADDR_SET_UOBJ(voa, uobj) \ 4789 UVM_VOADDR_SET_OBJECT(voa, uobj, UVM_VOADDR_TYPE_UOBJ) 4790 4791 #define UVM_VOADDR_GET_ANON(voa) \ 4792 ((struct vm_anon *)UVM_VOADDR_GET_OBJECT(voa)) 4793 #define UVM_VOADDR_SET_ANON(voa, anon) \ 4794 UVM_VOADDR_SET_OBJECT(voa, anon, UVM_VOADDR_TYPE_ANON) 4795 4796 /* 4797 * uvm_voaddr_acquire: returns the virtual object address corresponding 4798 * to the specified virtual address. 4799 * 4800 * => resolves COW so the true page identity is tracked. 4801 * 4802 * => acquires a reference on the page's owner (uvm_object or vm_anon) 4803 */ 4804 bool 4805 uvm_voaddr_acquire(struct vm_map * const map, vaddr_t const va, 4806 struct uvm_voaddr * const voaddr) 4807 { 4808 struct vm_map_entry *entry; 4809 struct vm_anon *anon = NULL; 4810 bool result = false; 4811 bool exclusive = false; 4812 void (*unlock_fn)(struct vm_map *); 4813 4814 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 4815 UVMHIST_LOG(maphist,"(map=%#jx,va=%#jx)", (uintptr_t)map, va, 0, 0); 4816 4817 const vaddr_t start = trunc_page(va); 4818 const vaddr_t end = round_page(va+1); 4819 4820 lookup_again: 4821 if (__predict_false(exclusive)) { 4822 vm_map_lock(map); 4823 unlock_fn = vm_map_unlock; 4824 } else { 4825 vm_map_lock_read(map); 4826 unlock_fn = vm_map_unlock_read; 4827 } 4828 4829 if (__predict_false(!uvm_map_lookup_entry(map, start, &entry))) { 4830 unlock_fn(map); 4831 UVMHIST_LOG(maphist,"<- done (no entry)",0,0,0,0); 4832 return false; 4833 } 4834 4835 if (__predict_false(entry->protection == VM_PROT_NONE)) { 4836 unlock_fn(map); 4837 UVMHIST_LOG(maphist,"<- done (PROT_NONE)",0,0,0,0); 4838 return false; 4839 } 4840 4841 /* 4842 * We have a fast path for the common case of "no COW resolution 4843 * needed" whereby we have taken a read lock on the map and if 4844 * we don't encounter any need to create a vm_anon then great! 4845 * But if we do, we loop around again, instead taking an exclusive 4846 * lock so that we can perform the fault. 4847 * 4848 * In the event that we have to resolve the fault, we do nearly the 4849 * same work as uvm_map_pageable() does: 4850 * 4851 * 1: holding the write lock, we create any anonymous maps that need 4852 * to be created. however, we do NOT need to clip the map entries 4853 * in this case. 4854 * 4855 * 2: we downgrade to a read lock, and call uvm_fault_wire to fault 4856 * in the page (assuming the entry is not already wired). this 4857 * is done because we need the vm_anon to be present. 4858 */ 4859 if (__predict_true(!VM_MAPENT_ISWIRED(entry))) { 4860 4861 bool need_fault = false; 4862 4863 /* 4864 * perform the action of vm_map_lookup that need the 4865 * write lock on the map: create an anonymous map for 4866 * a copy-on-write region, or an anonymous map for 4867 * a zero-fill region. 4868 */ 4869 if (__predict_false(UVM_ET_ISSUBMAP(entry))) { 4870 unlock_fn(map); 4871 UVMHIST_LOG(maphist,"<- done (submap)",0,0,0,0); 4872 return false; 4873 } 4874 if (__predict_false(UVM_ET_ISNEEDSCOPY(entry) && 4875 ((entry->max_protection & VM_PROT_WRITE) || 4876 (entry->object.uvm_obj == NULL)))) { 4877 if (!exclusive) { 4878 /* need to take the slow path */ 4879 KASSERT(unlock_fn == vm_map_unlock_read); 4880 vm_map_unlock_read(map); 4881 exclusive = true; 4882 goto lookup_again; 4883 } 4884 need_fault = true; 4885 amap_copy(map, entry, 0, start, end); 4886 /* XXXCDC: wait OK? */ 4887 } 4888 4889 /* 4890 * do a quick check to see if the fault has already 4891 * been resolved to the upper layer. 4892 */ 4893 if (__predict_true(entry->aref.ar_amap != NULL && 4894 need_fault == false)) { 4895 amap_lock(entry->aref.ar_amap, RW_WRITER); 4896 anon = amap_lookup(&entry->aref, start - entry->start); 4897 if (__predict_true(anon != NULL)) { 4898 /* amap unlocked below */ 4899 goto found_anon; 4900 } 4901 amap_unlock(entry->aref.ar_amap); 4902 need_fault = true; 4903 } 4904 4905 /* 4906 * we predict this test as false because if we reach 4907 * this point, then we are likely dealing with a 4908 * shared memory region backed by a uvm_object, in 4909 * which case a fault to create the vm_anon is not 4910 * necessary. 4911 */ 4912 if (__predict_false(need_fault)) { 4913 if (exclusive) { 4914 vm_map_busy(map); 4915 vm_map_unlock(map); 4916 unlock_fn = vm_map_unbusy; 4917 } 4918 4919 if (uvm_fault_wire(map, start, end, 4920 entry->max_protection, 1)) { 4921 /* wiring failed */ 4922 unlock_fn(map); 4923 UVMHIST_LOG(maphist,"<- done (wire failed)", 4924 0,0,0,0); 4925 return false; 4926 } 4927 4928 /* 4929 * now that we have resolved the fault, we can unwire 4930 * the page. 4931 */ 4932 if (exclusive) { 4933 vm_map_lock(map); 4934 vm_map_unbusy(map); 4935 unlock_fn = vm_map_unlock; 4936 } 4937 4938 uvm_fault_unwire_locked(map, start, end); 4939 } 4940 } 4941 4942 /* check the upper layer */ 4943 if (entry->aref.ar_amap) { 4944 amap_lock(entry->aref.ar_amap, RW_WRITER); 4945 anon = amap_lookup(&entry->aref, start - entry->start); 4946 if (anon) { 4947 found_anon: KASSERT(anon->an_lock == entry->aref.ar_amap->am_lock); 4948 anon->an_ref++; 4949 rw_obj_hold(anon->an_lock); 4950 KASSERT(anon->an_ref != 0); 4951 UVM_VOADDR_SET_ANON(voaddr, anon); 4952 voaddr->offset = va & PAGE_MASK; 4953 result = true; 4954 } 4955 amap_unlock(entry->aref.ar_amap); 4956 } 4957 4958 /* check the lower layer */ 4959 if (!result && UVM_ET_ISOBJ(entry)) { 4960 struct uvm_object *uobj = entry->object.uvm_obj; 4961 4962 KASSERT(uobj != NULL); 4963 (*uobj->pgops->pgo_reference)(uobj); 4964 UVM_VOADDR_SET_UOBJ(voaddr, uobj); 4965 voaddr->offset = entry->offset + (va - entry->start); 4966 result = true; 4967 } 4968 4969 unlock_fn(map); 4970 4971 if (result) { 4972 UVMHIST_LOG(maphist, 4973 "<- done OK (type=%jd,owner=%#jx,offset=%#jx)", 4974 UVM_VOADDR_GET_TYPE(voaddr), 4975 UVM_VOADDR_GET_OBJECT(voaddr), 4976 voaddr->offset, 0); 4977 } else { 4978 UVMHIST_LOG(maphist,"<- done (failed)",0,0,0,0); 4979 } 4980 4981 return result; 4982 } 4983 4984 /* 4985 * uvm_voaddr_release: release the references held by the 4986 * vitual object address. 4987 */ 4988 void 4989 uvm_voaddr_release(struct uvm_voaddr * const voaddr) 4990 { 4991 4992 switch (UVM_VOADDR_GET_TYPE(voaddr)) { 4993 case UVM_VOADDR_TYPE_UOBJ: { 4994 struct uvm_object * const uobj = UVM_VOADDR_GET_UOBJ(voaddr); 4995 4996 KASSERT(uobj != NULL); 4997 KASSERT(uobj->pgops->pgo_detach != NULL); 4998 (*uobj->pgops->pgo_detach)(uobj); 4999 break; 5000 } 5001 case UVM_VOADDR_TYPE_ANON: { 5002 struct vm_anon * const anon = UVM_VOADDR_GET_ANON(voaddr); 5003 krwlock_t *lock; 5004 5005 KASSERT(anon != NULL); 5006 rw_enter((lock = anon->an_lock), RW_WRITER); 5007 KASSERT(anon->an_ref > 0); 5008 if (--anon->an_ref == 0) { 5009 uvm_anfree(anon); 5010 } 5011 rw_exit(lock); 5012 rw_obj_free(lock); 5013 break; 5014 } 5015 default: 5016 panic("uvm_voaddr_release: bad type"); 5017 } 5018 memset(voaddr, 0, sizeof(*voaddr)); 5019 } 5020 5021 /* 5022 * uvm_voaddr_compare: compare two uvm_voaddr objects. 5023 * 5024 * => memcmp() semantics 5025 */ 5026 int 5027 uvm_voaddr_compare(const struct uvm_voaddr * const voaddr1, 5028 const struct uvm_voaddr * const voaddr2) 5029 { 5030 const uintptr_t type1 = UVM_VOADDR_GET_TYPE(voaddr1); 5031 const uintptr_t type2 = UVM_VOADDR_GET_TYPE(voaddr2); 5032 5033 KASSERT(type1 == UVM_VOADDR_TYPE_UOBJ || 5034 type1 == UVM_VOADDR_TYPE_ANON); 5035 5036 KASSERT(type2 == UVM_VOADDR_TYPE_UOBJ || 5037 type2 == UVM_VOADDR_TYPE_ANON); 5038 5039 if (type1 < type2) 5040 return -1; 5041 if (type1 > type2) 5042 return 1; 5043 5044 const uintptr_t addr1 = UVM_VOADDR_GET_OBJECT(voaddr1); 5045 const uintptr_t addr2 = UVM_VOADDR_GET_OBJECT(voaddr2); 5046 5047 if (addr1 < addr2) 5048 return -1; 5049 if (addr1 > addr2) 5050 return 1; 5051 5052 if (voaddr1->offset < voaddr2->offset) 5053 return -1; 5054 if (voaddr1->offset > voaddr2->offset) 5055 return 1; 5056 5057 return 0; 5058 } 5059 5060 #if defined(DDB) || defined(DEBUGPRINT) 5061 5062 /* 5063 * uvm_map_printit: actually prints the map 5064 */ 5065 5066 void 5067 uvm_map_printit(struct vm_map *map, bool full, 5068 void (*pr)(const char *, ...)) 5069 { 5070 struct vm_map_entry *entry; 5071 5072 (*pr)("MAP %p: [%#lx->%#lx]\n", map, vm_map_min(map), 5073 vm_map_max(map)); 5074 (*pr)("\t#ent=%d, sz=%d, ref=%d, version=%d, flags=%#x\n", 5075 map->nentries, map->size, map->ref_count, map->timestamp, 5076 map->flags); 5077 (*pr)("\tpmap=%p(resident=%ld, wired=%ld)\n", map->pmap, 5078 pmap_resident_count(map->pmap), pmap_wired_count(map->pmap)); 5079 if (!full) 5080 return; 5081 for (entry = map->header.next; entry != &map->header; 5082 entry = entry->next) { 5083 (*pr)(" - %p: %#lx->%#lx: obj=%p/%#llx, amap=%p/%d\n", 5084 entry, entry->start, entry->end, entry->object.uvm_obj, 5085 (long long)entry->offset, entry->aref.ar_amap, 5086 entry->aref.ar_pageoff); 5087 (*pr)( 5088 "\tsubmap=%c, cow=%c, nc=%c, prot(max)=%d/%d, inh=%d, " 5089 "wc=%d, adv=%d\n", 5090 (entry->etype & UVM_ET_SUBMAP) ? 'T' : 'F', 5091 (entry->etype & UVM_ET_COPYONWRITE) ? 'T' : 'F', 5092 (entry->etype & UVM_ET_NEEDSCOPY) ? 'T' : 'F', 5093 entry->protection, entry->max_protection, 5094 entry->inheritance, entry->wired_count, entry->advice); 5095 } 5096 } 5097 5098 void 5099 uvm_whatis(uintptr_t addr, void (*pr)(const char *, ...)) 5100 { 5101 struct vm_map *map; 5102 5103 for (map = kernel_map;;) { 5104 struct vm_map_entry *entry; 5105 5106 if (!uvm_map_lookup_entry_bytree(map, (vaddr_t)addr, &entry)) { 5107 break; 5108 } 5109 (*pr)("%p is %p+%zu from VMMAP %p\n", 5110 (void *)addr, (void *)entry->start, 5111 (size_t)(addr - (uintptr_t)entry->start), map); 5112 if (!UVM_ET_ISSUBMAP(entry)) { 5113 break; 5114 } 5115 map = entry->object.sub_map; 5116 } 5117 } 5118 5119 #endif /* DDB || DEBUGPRINT */ 5120 5121 #ifndef __USER_VA0_IS_SAFE 5122 static int 5123 sysctl_user_va0_disable(SYSCTLFN_ARGS) 5124 { 5125 struct sysctlnode node; 5126 int t, error; 5127 5128 node = *rnode; 5129 node.sysctl_data = &t; 5130 t = user_va0_disable; 5131 error = sysctl_lookup(SYSCTLFN_CALL(&node)); 5132 if (error || newp == NULL) 5133 return (error); 5134 5135 if (!t && user_va0_disable && 5136 kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_MAP_VA_ZERO, 0, 5137 NULL, NULL, NULL)) 5138 return EPERM; 5139 5140 user_va0_disable = !!t; 5141 return 0; 5142 } 5143 #endif 5144 5145 static int 5146 fill_vmentry(struct lwp *l, struct proc *p, struct kinfo_vmentry *kve, 5147 struct vm_map *m, struct vm_map_entry *e) 5148 { 5149 #ifndef _RUMPKERNEL 5150 int error; 5151 5152 memset(kve, 0, sizeof(*kve)); 5153 KASSERT(e != NULL); 5154 if (UVM_ET_ISOBJ(e)) { 5155 struct uvm_object *uobj = e->object.uvm_obj; 5156 KASSERT(uobj != NULL); 5157 kve->kve_ref_count = uobj->uo_refs; 5158 kve->kve_count = uobj->uo_npages; 5159 if (UVM_OBJ_IS_VNODE(uobj)) { 5160 struct vattr va; 5161 struct vnode *vp = (struct vnode *)uobj; 5162 vn_lock(vp, LK_SHARED | LK_RETRY); 5163 error = VOP_GETATTR(vp, &va, l->l_cred); 5164 VOP_UNLOCK(vp); 5165 kve->kve_type = KVME_TYPE_VNODE; 5166 if (error == 0) { 5167 kve->kve_vn_size = vp->v_size; 5168 kve->kve_vn_type = (int)vp->v_type; 5169 kve->kve_vn_mode = va.va_mode; 5170 kve->kve_vn_rdev = va.va_rdev; 5171 kve->kve_vn_fileid = va.va_fileid; 5172 kve->kve_vn_fsid = va.va_fsid; 5173 error = vnode_to_path(kve->kve_path, 5174 sizeof(kve->kve_path) / 2, vp, l, p); 5175 } 5176 } else if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 5177 kve->kve_type = KVME_TYPE_KERN; 5178 } else if (UVM_OBJ_IS_DEVICE(uobj)) { 5179 kve->kve_type = KVME_TYPE_DEVICE; 5180 } else if (UVM_OBJ_IS_AOBJ(uobj)) { 5181 kve->kve_type = KVME_TYPE_ANON; 5182 } else { 5183 kve->kve_type = KVME_TYPE_OBJECT; 5184 } 5185 } else if (UVM_ET_ISSUBMAP(e)) { 5186 struct vm_map *map = e->object.sub_map; 5187 KASSERT(map != NULL); 5188 kve->kve_ref_count = map->ref_count; 5189 kve->kve_count = map->nentries; 5190 kve->kve_type = KVME_TYPE_SUBMAP; 5191 } else 5192 kve->kve_type = KVME_TYPE_UNKNOWN; 5193 5194 kve->kve_start = e->start; 5195 kve->kve_end = e->end; 5196 kve->kve_offset = e->offset; 5197 kve->kve_wired_count = e->wired_count; 5198 kve->kve_inheritance = e->inheritance; 5199 kve->kve_attributes = 0; /* unused */ 5200 kve->kve_advice = e->advice; 5201 #define PROT(p) (((p) & VM_PROT_READ) ? KVME_PROT_READ : 0) | \ 5202 (((p) & VM_PROT_WRITE) ? KVME_PROT_WRITE : 0) | \ 5203 (((p) & VM_PROT_EXECUTE) ? KVME_PROT_EXEC : 0) 5204 kve->kve_protection = PROT(e->protection); 5205 kve->kve_max_protection = PROT(e->max_protection); 5206 kve->kve_flags |= (e->etype & UVM_ET_COPYONWRITE) 5207 ? KVME_FLAG_COW : 0; 5208 kve->kve_flags |= (e->etype & UVM_ET_NEEDSCOPY) 5209 ? KVME_FLAG_NEEDS_COPY : 0; 5210 kve->kve_flags |= (m->flags & VM_MAP_TOPDOWN) 5211 ? KVME_FLAG_GROWS_DOWN : KVME_FLAG_GROWS_UP; 5212 kve->kve_flags |= (m->flags & VM_MAP_PAGEABLE) 5213 ? KVME_FLAG_PAGEABLE : 0; 5214 #endif 5215 return 0; 5216 } 5217 5218 static int 5219 fill_vmentries(struct lwp *l, pid_t pid, u_int elem_size, void *oldp, 5220 size_t *oldlenp) 5221 { 5222 int error; 5223 struct proc *p; 5224 struct kinfo_vmentry *vme; 5225 struct vmspace *vm; 5226 struct vm_map *map; 5227 struct vm_map_entry *entry; 5228 char *dp; 5229 size_t count, vmesize; 5230 5231 if (elem_size == 0 || elem_size > 2 * sizeof(*vme)) 5232 return EINVAL; 5233 5234 if (oldp) { 5235 if (*oldlenp > 10UL * 1024UL * 1024UL) 5236 return E2BIG; 5237 count = *oldlenp / elem_size; 5238 if (count == 0) 5239 return ENOMEM; 5240 vmesize = count * sizeof(*vme); 5241 } else 5242 vmesize = 0; 5243 5244 if ((error = proc_find_locked(l, &p, pid)) != 0) 5245 return error; 5246 5247 vme = NULL; 5248 count = 0; 5249 5250 if ((error = proc_vmspace_getref(p, &vm)) != 0) 5251 goto out; 5252 5253 map = &vm->vm_map; 5254 vm_map_lock_read(map); 5255 5256 dp = oldp; 5257 if (oldp) 5258 vme = kmem_alloc(vmesize, KM_SLEEP); 5259 for (entry = map->header.next; entry != &map->header; 5260 entry = entry->next) { 5261 if (oldp && (dp - (char *)oldp) < vmesize) { 5262 error = fill_vmentry(l, p, &vme[count], map, entry); 5263 if (error) 5264 goto out; 5265 dp += elem_size; 5266 } 5267 count++; 5268 } 5269 vm_map_unlock_read(map); 5270 uvmspace_free(vm); 5271 5272 out: 5273 if (pid != -1) 5274 mutex_exit(p->p_lock); 5275 if (error == 0) { 5276 const u_int esize = uimin(sizeof(*vme), elem_size); 5277 dp = oldp; 5278 for (size_t i = 0; i < count; i++) { 5279 if (oldp && (dp - (char *)oldp) < vmesize) { 5280 error = sysctl_copyout(l, &vme[i], dp, esize); 5281 if (error) 5282 break; 5283 dp += elem_size; 5284 } else 5285 break; 5286 } 5287 count *= elem_size; 5288 if (oldp != NULL && *oldlenp < count) 5289 error = ENOSPC; 5290 *oldlenp = count; 5291 } 5292 if (vme) 5293 kmem_free(vme, vmesize); 5294 return error; 5295 } 5296 5297 static int 5298 sysctl_vmproc(SYSCTLFN_ARGS) 5299 { 5300 int error; 5301 5302 if (namelen == 1 && name[0] == CTL_QUERY) 5303 return (sysctl_query(SYSCTLFN_CALL(rnode))); 5304 5305 if (namelen == 0) 5306 return EINVAL; 5307 5308 switch (name[0]) { 5309 case VM_PROC_MAP: 5310 if (namelen != 3) 5311 return EINVAL; 5312 sysctl_unlock(); 5313 error = fill_vmentries(l, name[1], name[2], oldp, oldlenp); 5314 sysctl_relock(); 5315 return error; 5316 default: 5317 return EINVAL; 5318 } 5319 } 5320 5321 SYSCTL_SETUP(sysctl_uvmmap_setup, "sysctl uvmmap setup") 5322 { 5323 5324 sysctl_createv(clog, 0, NULL, NULL, 5325 CTLFLAG_PERMANENT, 5326 CTLTYPE_STRUCT, "proc", 5327 SYSCTL_DESCR("Process vm information"), 5328 sysctl_vmproc, 0, NULL, 0, 5329 CTL_VM, VM_PROC, CTL_EOL); 5330 #ifndef __USER_VA0_IS_SAFE 5331 sysctl_createv(clog, 0, NULL, NULL, 5332 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 5333 CTLTYPE_INT, "user_va0_disable", 5334 SYSCTL_DESCR("Disable VA 0"), 5335 sysctl_user_va0_disable, 0, &user_va0_disable, 0, 5336 CTL_VM, CTL_CREATE, CTL_EOL); 5337 #endif 5338 } 5339