1 /* $NetBSD: uvm_map.c,v 1.188 2005/04/07 06:44:15 dbj Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * Copyright (c) 1991, 1993, The Regents of the University of California. 6 * 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by Charles D. Cranor, 23 * Washington University, the University of California, Berkeley and 24 * its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * @(#)vm_map.c 8.3 (Berkeley) 1/12/94 42 * from: Id: uvm_map.c,v 1.1.2.27 1998/02/07 01:16:54 chs Exp 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Permission to use, copy, modify and distribute this software and 49 * its documentation is hereby granted, provided that both the copyright 50 * notice and this permission notice appear in all copies of the 51 * software, derivative works or modified versions, and any portions 52 * thereof, and that both notices appear in supporting documentation. 53 * 54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 57 * 58 * Carnegie Mellon requests users of this software to return to 59 * 60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 61 * School of Computer Science 62 * Carnegie Mellon University 63 * Pittsburgh PA 15213-3890 64 * 65 * any improvements or extensions that they make and grant Carnegie the 66 * rights to redistribute these changes. 67 */ 68 69 /* 70 * uvm_map.c: uvm map operations 71 */ 72 73 #include <sys/cdefs.h> 74 __KERNEL_RCSID(0, "$NetBSD: uvm_map.c,v 1.188 2005/04/07 06:44:15 dbj Exp $"); 75 76 #include "opt_ddb.h" 77 #include "opt_uvmhist.h" 78 #include "opt_uvm.h" 79 #include "opt_sysv.h" 80 81 #include <sys/param.h> 82 #include <sys/systm.h> 83 #include <sys/mman.h> 84 #include <sys/proc.h> 85 #include <sys/malloc.h> 86 #include <sys/pool.h> 87 #include <sys/kernel.h> 88 #include <sys/mount.h> 89 #include <sys/vnode.h> 90 91 #ifdef SYSVSHM 92 #include <sys/shm.h> 93 #endif 94 95 #define UVM_MAP 96 #include <uvm/uvm.h> 97 #undef RB_AUGMENT 98 #define RB_AUGMENT(x) uvm_rb_augment(x) 99 100 #ifdef DDB 101 #include <uvm/uvm_ddb.h> 102 #endif 103 104 #ifndef UVMMAP_NOCOUNTERS 105 #include <sys/device.h> 106 struct evcnt map_ubackmerge = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 107 "uvmmap", "ubackmerge"); 108 struct evcnt map_uforwmerge = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 109 "uvmmap", "uforwmerge"); 110 struct evcnt map_ubimerge = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 111 "uvmmap", "ubimerge"); 112 struct evcnt map_unomerge = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 113 "uvmmap", "unomerge"); 114 struct evcnt map_kbackmerge = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 115 "uvmmap", "kbackmerge"); 116 struct evcnt map_kforwmerge = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 117 "uvmmap", "kforwmerge"); 118 struct evcnt map_kbimerge = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 119 "uvmmap", "kbimerge"); 120 struct evcnt map_knomerge = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 121 "uvmmap", "knomerge"); 122 struct evcnt uvm_map_call = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 123 "uvmmap", "map_call"); 124 struct evcnt uvm_mlk_call = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 125 "uvmmap", "mlk_call"); 126 struct evcnt uvm_mlk_hint = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, 127 "uvmmap", "mlk_hint"); 128 129 EVCNT_ATTACH_STATIC(map_ubackmerge); 130 EVCNT_ATTACH_STATIC(map_uforwmerge); 131 EVCNT_ATTACH_STATIC(map_ubimerge); 132 EVCNT_ATTACH_STATIC(map_unomerge); 133 EVCNT_ATTACH_STATIC(map_kbackmerge); 134 EVCNT_ATTACH_STATIC(map_kforwmerge); 135 EVCNT_ATTACH_STATIC(map_kbimerge); 136 EVCNT_ATTACH_STATIC(map_knomerge); 137 EVCNT_ATTACH_STATIC(uvm_map_call); 138 EVCNT_ATTACH_STATIC(uvm_mlk_call); 139 EVCNT_ATTACH_STATIC(uvm_mlk_hint); 140 141 #define UVMCNT_INCR(ev) ev.ev_count++ 142 #define UVMCNT_DECR(ev) ev.ev_count-- 143 #else 144 #define UVMCNT_INCR(ev) 145 #define UVMCNT_DECR(ev) 146 #endif 147 148 const char vmmapbsy[] = "vmmapbsy"; 149 150 /* 151 * pool for vmspace structures. 152 */ 153 154 POOL_INIT(uvm_vmspace_pool, sizeof(struct vmspace), 0, 0, 0, "vmsppl", 155 &pool_allocator_nointr); 156 157 /* 158 * pool for dynamically-allocated map entries. 159 */ 160 161 POOL_INIT(uvm_map_entry_pool, sizeof(struct vm_map_entry), 0, 0, 0, "vmmpepl", 162 &pool_allocator_nointr); 163 164 MALLOC_DEFINE(M_VMMAP, "VM map", "VM map structures"); 165 MALLOC_DEFINE(M_VMPMAP, "VM pmap", "VM pmap"); 166 167 #ifdef PMAP_GROWKERNEL 168 /* 169 * This global represents the end of the kernel virtual address 170 * space. If we want to exceed this, we must grow the kernel 171 * virtual address space dynamically. 172 * 173 * Note, this variable is locked by kernel_map's lock. 174 */ 175 vaddr_t uvm_maxkaddr; 176 #endif 177 178 /* 179 * macros 180 */ 181 182 /* 183 * VM_MAP_USE_KMAPENT: determine if uvm_kmapent_alloc/free is used 184 * for the vm_map. 185 */ 186 extern struct vm_map *pager_map; /* XXX */ 187 #define VM_MAP_USE_KMAPENT(map) \ 188 (((map)->flags & VM_MAP_INTRSAFE) || (map) == kernel_map) 189 190 /* 191 * uvm_map_entry_link: insert entry into a map 192 * 193 * => map must be locked 194 */ 195 #define uvm_map_entry_link(map, after_where, entry) do { \ 196 KASSERT(entry->start < entry->end); \ 197 (map)->nentries++; \ 198 (entry)->prev = (after_where); \ 199 (entry)->next = (after_where)->next; \ 200 (entry)->prev->next = (entry); \ 201 (entry)->next->prev = (entry); \ 202 uvm_rb_insert((map), (entry)); \ 203 } while (/*CONSTCOND*/ 0) 204 205 /* 206 * uvm_map_entry_unlink: remove entry from a map 207 * 208 * => map must be locked 209 */ 210 #define uvm_map_entry_unlink(map, entry) do { \ 211 (map)->nentries--; \ 212 (entry)->next->prev = (entry)->prev; \ 213 (entry)->prev->next = (entry)->next; \ 214 uvm_rb_remove((map), (entry)); \ 215 } while (/*CONSTCOND*/ 0) 216 217 /* 218 * SAVE_HINT: saves the specified entry as the hint for future lookups. 219 * 220 * => map need not be locked (protected by hint_lock). 221 */ 222 #define SAVE_HINT(map,check,value) do { \ 223 simple_lock(&(map)->hint_lock); \ 224 if ((map)->hint == (check)) \ 225 (map)->hint = (value); \ 226 simple_unlock(&(map)->hint_lock); \ 227 } while (/*CONSTCOND*/ 0) 228 229 /* 230 * VM_MAP_RANGE_CHECK: check and correct range 231 * 232 * => map must at least be read locked 233 */ 234 235 #define VM_MAP_RANGE_CHECK(map, start, end) do { \ 236 if (start < vm_map_min(map)) \ 237 start = vm_map_min(map); \ 238 if (end > vm_map_max(map)) \ 239 end = vm_map_max(map); \ 240 if (start > end) \ 241 start = end; \ 242 } while (/*CONSTCOND*/ 0) 243 244 /* 245 * local prototypes 246 */ 247 248 static struct vm_map_entry * 249 uvm_mapent_alloc(struct vm_map *, int); 250 static struct vm_map_entry * 251 uvm_mapent_alloc_split(struct vm_map *, 252 const struct vm_map_entry *, int, 253 struct uvm_mapent_reservation *); 254 static void uvm_mapent_copy(struct vm_map_entry *, struct vm_map_entry *); 255 static void uvm_mapent_free(struct vm_map_entry *); 256 static struct vm_map_entry * 257 uvm_kmapent_alloc(struct vm_map *, int); 258 static void uvm_kmapent_free(struct vm_map_entry *); 259 static void uvm_map_entry_unwire(struct vm_map *, struct vm_map_entry *); 260 static void uvm_map_reference_amap(struct vm_map_entry *, int); 261 static int uvm_map_space_avail(vaddr_t *, vsize_t, voff_t, vsize_t, int, 262 struct vm_map_entry *); 263 static void uvm_map_unreference_amap(struct vm_map_entry *, int); 264 265 int _uvm_tree_sanity(struct vm_map *, const char *); 266 static vsize_t uvm_rb_subtree_space(const struct vm_map_entry *); 267 268 static __inline int 269 uvm_compare(const struct vm_map_entry *a, const struct vm_map_entry *b) 270 { 271 272 if (a->start < b->start) 273 return (-1); 274 else if (a->start > b->start) 275 return (1); 276 277 return (0); 278 } 279 280 static __inline void 281 uvm_rb_augment(struct vm_map_entry *entry) 282 { 283 284 entry->space = uvm_rb_subtree_space(entry); 285 } 286 287 RB_PROTOTYPE(uvm_tree, vm_map_entry, rb_entry, uvm_compare); 288 289 RB_GENERATE(uvm_tree, vm_map_entry, rb_entry, uvm_compare); 290 291 static __inline vsize_t 292 uvm_rb_space(const struct vm_map *map, const struct vm_map_entry *entry) 293 { 294 /* XXX map is not used */ 295 296 KASSERT(entry->next != NULL); 297 return entry->next->start - entry->end; 298 } 299 300 static vsize_t 301 uvm_rb_subtree_space(const struct vm_map_entry *entry) 302 { 303 vaddr_t space, tmp; 304 305 space = entry->ownspace; 306 if (RB_LEFT(entry, rb_entry)) { 307 tmp = RB_LEFT(entry, rb_entry)->space; 308 if (tmp > space) 309 space = tmp; 310 } 311 312 if (RB_RIGHT(entry, rb_entry)) { 313 tmp = RB_RIGHT(entry, rb_entry)->space; 314 if (tmp > space) 315 space = tmp; 316 } 317 318 return (space); 319 } 320 321 static __inline void 322 uvm_rb_fixup(struct vm_map *map, struct vm_map_entry *entry) 323 { 324 /* We need to traverse to the very top */ 325 do { 326 entry->ownspace = uvm_rb_space(map, entry); 327 entry->space = uvm_rb_subtree_space(entry); 328 } while ((entry = RB_PARENT(entry, rb_entry)) != NULL); 329 } 330 331 static __inline void 332 uvm_rb_insert(struct vm_map *map, struct vm_map_entry *entry) 333 { 334 vaddr_t space = uvm_rb_space(map, entry); 335 struct vm_map_entry *tmp; 336 337 entry->ownspace = entry->space = space; 338 tmp = RB_INSERT(uvm_tree, &(map)->rbhead, entry); 339 #ifdef DIAGNOSTIC 340 if (tmp != NULL) 341 panic("uvm_rb_insert: duplicate entry?"); 342 #endif 343 uvm_rb_fixup(map, entry); 344 if (entry->prev != &map->header) 345 uvm_rb_fixup(map, entry->prev); 346 } 347 348 static __inline void 349 uvm_rb_remove(struct vm_map *map, struct vm_map_entry *entry) 350 { 351 struct vm_map_entry *parent; 352 353 parent = RB_PARENT(entry, rb_entry); 354 RB_REMOVE(uvm_tree, &(map)->rbhead, entry); 355 if (entry->prev != &map->header) 356 uvm_rb_fixup(map, entry->prev); 357 if (parent) 358 uvm_rb_fixup(map, parent); 359 } 360 361 #ifdef DEBUG 362 int uvm_debug_check_rbtree = 0; 363 #define uvm_tree_sanity(x,y) \ 364 if (uvm_debug_check_rbtree) \ 365 _uvm_tree_sanity(x,y) 366 #else 367 #define uvm_tree_sanity(x,y) 368 #endif 369 370 int 371 _uvm_tree_sanity(struct vm_map *map, const char *name) 372 { 373 struct vm_map_entry *tmp, *trtmp; 374 int n = 0, i = 1; 375 376 RB_FOREACH(tmp, uvm_tree, &map->rbhead) { 377 if (tmp->ownspace != uvm_rb_space(map, tmp)) { 378 printf("%s: %d/%d ownspace %lx != %lx %s\n", 379 name, n + 1, map->nentries, 380 (ulong)tmp->ownspace, (ulong)uvm_rb_space(map, tmp), 381 tmp->next == &map->header ? "(last)" : ""); 382 goto error; 383 } 384 } 385 trtmp = NULL; 386 RB_FOREACH(tmp, uvm_tree, &map->rbhead) { 387 if (tmp->space != uvm_rb_subtree_space(tmp)) { 388 printf("%s: space %lx != %lx\n", 389 name, (ulong)tmp->space, 390 (ulong)uvm_rb_subtree_space(tmp)); 391 goto error; 392 } 393 if (trtmp != NULL && trtmp->start >= tmp->start) { 394 printf("%s: corrupt: 0x%lx >= 0x%lx\n", 395 name, trtmp->start, tmp->start); 396 goto error; 397 } 398 n++; 399 400 trtmp = tmp; 401 } 402 403 if (n != map->nentries) { 404 printf("%s: nentries: %d vs %d\n", 405 name, n, map->nentries); 406 goto error; 407 } 408 409 for (tmp = map->header.next; tmp && tmp != &map->header; 410 tmp = tmp->next, i++) { 411 trtmp = RB_FIND(uvm_tree, &map->rbhead, tmp); 412 if (trtmp != tmp) { 413 printf("%s: lookup: %d: %p - %p: %p\n", 414 name, i, tmp, trtmp, 415 RB_PARENT(tmp, rb_entry)); 416 goto error; 417 } 418 } 419 420 return (0); 421 error: 422 #ifdef DDB 423 /* handy breakpoint location for error case */ 424 __asm(".globl treesanity_label\ntreesanity_label:"); 425 #endif 426 return (-1); 427 } 428 429 /* 430 * local inlines 431 */ 432 433 static __inline struct vm_map *uvm_kmapent_map(struct vm_map_entry *); 434 435 /* 436 * uvm_mapent_alloc: allocate a map entry 437 */ 438 439 static __inline struct vm_map_entry * 440 uvm_mapent_alloc(struct vm_map *map, int flags) 441 { 442 struct vm_map_entry *me; 443 int pflags = (flags & UVM_FLAG_NOWAIT) ? PR_NOWAIT : PR_WAITOK; 444 UVMHIST_FUNC("uvm_mapent_alloc"); UVMHIST_CALLED(maphist); 445 446 if (VM_MAP_USE_KMAPENT(map)) { 447 me = uvm_kmapent_alloc(map, flags); 448 } else { 449 me = pool_get(&uvm_map_entry_pool, pflags); 450 if (__predict_false(me == NULL)) 451 return NULL; 452 me->flags = 0; 453 } 454 455 UVMHIST_LOG(maphist, "<- new entry=0x%x [kentry=%d]", me, 456 ((map->flags & VM_MAP_INTRSAFE) != 0 || map == kernel_map), 0, 0); 457 return (me); 458 } 459 460 /* 461 * uvm_mapent_alloc_split: allocate a map entry for clipping. 462 */ 463 464 static __inline struct vm_map_entry * 465 uvm_mapent_alloc_split(struct vm_map *map, 466 const struct vm_map_entry *old_entry, int flags, 467 struct uvm_mapent_reservation *umr) 468 { 469 struct vm_map_entry *me; 470 471 KASSERT(!VM_MAP_USE_KMAPENT(map) || 472 (old_entry->flags & UVM_MAP_QUANTUM) || !UMR_EMPTY(umr)); 473 474 if (old_entry->flags & UVM_MAP_QUANTUM) { 475 int s; 476 struct vm_map_kernel *vmk = vm_map_to_kernel(map); 477 478 s = splvm(); 479 simple_lock(&uvm.kentry_lock); 480 me = vmk->vmk_merged_entries; 481 KASSERT(me); 482 vmk->vmk_merged_entries = me->next; 483 simple_unlock(&uvm.kentry_lock); 484 splx(s); 485 KASSERT(me->flags & UVM_MAP_QUANTUM); 486 } else { 487 me = uvm_mapent_alloc(map, flags); 488 } 489 490 return me; 491 } 492 493 /* 494 * uvm_mapent_free: free map entry 495 */ 496 497 static __inline void 498 uvm_mapent_free(struct vm_map_entry *me) 499 { 500 UVMHIST_FUNC("uvm_mapent_free"); UVMHIST_CALLED(maphist); 501 502 UVMHIST_LOG(maphist,"<- freeing map entry=0x%x [flags=%d]", 503 me, me->flags, 0, 0); 504 if (me->flags & UVM_MAP_KERNEL) { 505 uvm_kmapent_free(me); 506 } else { 507 pool_put(&uvm_map_entry_pool, me); 508 } 509 } 510 511 /* 512 * uvm_mapent_free_merge: free merged map entry 513 * 514 * => keep the entry if needed. 515 * => caller shouldn't hold map locked. 516 */ 517 518 static __inline void 519 uvm_mapent_free_merged(struct vm_map *map, struct vm_map_entry *me) 520 { 521 522 KASSERT(!(me->flags & UVM_MAP_KERNEL) || uvm_kmapent_map(me) == map); 523 524 if (me->flags & UVM_MAP_QUANTUM) { 525 /* 526 * keep this entry for later splitting. 527 */ 528 struct vm_map_kernel *vmk; 529 int s; 530 531 KASSERT(VM_MAP_IS_KERNEL(map)); 532 KASSERT(!VM_MAP_USE_KMAPENT(map) || 533 (me->flags & UVM_MAP_KERNEL)); 534 535 vmk = vm_map_to_kernel(map); 536 s = splvm(); 537 simple_lock(&uvm.kentry_lock); 538 me->next = vmk->vmk_merged_entries; 539 vmk->vmk_merged_entries = me; 540 simple_unlock(&uvm.kentry_lock); 541 splx(s); 542 } else { 543 uvm_mapent_free(me); 544 } 545 } 546 547 /* 548 * uvm_mapent_copy: copy a map entry, preserving flags 549 */ 550 551 static __inline void 552 uvm_mapent_copy(struct vm_map_entry *src, struct vm_map_entry *dst) 553 { 554 555 memcpy(dst, src, ((char *)&src->uvm_map_entry_stop_copy) - 556 ((char *)src)); 557 } 558 559 /* 560 * uvm_map_entry_unwire: unwire a map entry 561 * 562 * => map should be locked by caller 563 */ 564 565 static __inline void 566 uvm_map_entry_unwire(struct vm_map *map, struct vm_map_entry *entry) 567 { 568 569 entry->wired_count = 0; 570 uvm_fault_unwire_locked(map, entry->start, entry->end); 571 } 572 573 574 /* 575 * wrapper for calling amap_ref() 576 */ 577 static __inline void 578 uvm_map_reference_amap(struct vm_map_entry *entry, int flags) 579 { 580 581 amap_ref(entry->aref.ar_amap, entry->aref.ar_pageoff, 582 (entry->end - entry->start) >> PAGE_SHIFT, flags); 583 } 584 585 586 /* 587 * wrapper for calling amap_unref() 588 */ 589 static __inline void 590 uvm_map_unreference_amap(struct vm_map_entry *entry, int flags) 591 { 592 593 amap_unref(entry->aref.ar_amap, entry->aref.ar_pageoff, 594 (entry->end - entry->start) >> PAGE_SHIFT, flags); 595 } 596 597 598 /* 599 * uvm_map_init: init mapping system at boot time. note that we allocate 600 * and init the static pool of struct vm_map_entry *'s for the kernel here. 601 */ 602 603 void 604 uvm_map_init(void) 605 { 606 #if defined(UVMHIST) 607 static struct uvm_history_ent maphistbuf[100]; 608 static struct uvm_history_ent pdhistbuf[100]; 609 #endif 610 611 /* 612 * first, init logging system. 613 */ 614 615 UVMHIST_FUNC("uvm_map_init"); 616 UVMHIST_INIT_STATIC(maphist, maphistbuf); 617 UVMHIST_INIT_STATIC(pdhist, pdhistbuf); 618 UVMHIST_CALLED(maphist); 619 UVMHIST_LOG(maphist,"<starting uvm map system>", 0, 0, 0, 0); 620 621 /* 622 * initialize the global lock for kernel map entry. 623 * 624 * XXX is it worth to have per-map lock instead? 625 */ 626 627 simple_lock_init(&uvm.kentry_lock); 628 } 629 630 /* 631 * clippers 632 */ 633 634 /* 635 * uvm_map_clip_start: ensure that the entry begins at or after 636 * the starting address, if it doesn't we split the entry. 637 * 638 * => caller should use UVM_MAP_CLIP_START macro rather than calling 639 * this directly 640 * => map must be locked by caller 641 */ 642 643 void 644 uvm_map_clip_start(struct vm_map *map, struct vm_map_entry *entry, 645 vaddr_t start, struct uvm_mapent_reservation *umr) 646 { 647 struct vm_map_entry *new_entry; 648 vaddr_t new_adj; 649 650 /* uvm_map_simplify_entry(map, entry); */ /* XXX */ 651 652 uvm_tree_sanity(map, "clip_start entry"); 653 654 /* 655 * Split off the front portion. note that we must insert the new 656 * entry BEFORE this one, so that this entry has the specified 657 * starting address. 658 */ 659 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 660 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 661 662 new_entry->end = start; 663 new_adj = start - new_entry->start; 664 if (entry->object.uvm_obj) 665 entry->offset += new_adj; /* shift start over */ 666 667 /* Does not change order for the RB tree */ 668 entry->start = start; 669 670 if (new_entry->aref.ar_amap) { 671 amap_splitref(&new_entry->aref, &entry->aref, new_adj); 672 } 673 674 uvm_map_entry_link(map, entry->prev, new_entry); 675 676 if (UVM_ET_ISSUBMAP(entry)) { 677 /* ... unlikely to happen, but play it safe */ 678 uvm_map_reference(new_entry->object.sub_map); 679 } else { 680 if (UVM_ET_ISOBJ(entry) && 681 entry->object.uvm_obj->pgops && 682 entry->object.uvm_obj->pgops->pgo_reference) 683 entry->object.uvm_obj->pgops->pgo_reference( 684 entry->object.uvm_obj); 685 } 686 687 uvm_tree_sanity(map, "clip_start leave"); 688 } 689 690 /* 691 * uvm_map_clip_end: ensure that the entry ends at or before 692 * the ending address, if it does't we split the reference 693 * 694 * => caller should use UVM_MAP_CLIP_END macro rather than calling 695 * this directly 696 * => map must be locked by caller 697 */ 698 699 void 700 uvm_map_clip_end(struct vm_map *map, struct vm_map_entry *entry, vaddr_t end, 701 struct uvm_mapent_reservation *umr) 702 { 703 struct vm_map_entry * new_entry; 704 vaddr_t new_adj; /* #bytes we move start forward */ 705 706 uvm_tree_sanity(map, "clip_end entry"); 707 708 /* 709 * Create a new entry and insert it 710 * AFTER the specified entry 711 */ 712 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 713 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 714 715 new_entry->start = entry->end = end; 716 new_adj = end - entry->start; 717 if (new_entry->object.uvm_obj) 718 new_entry->offset += new_adj; 719 720 if (entry->aref.ar_amap) 721 amap_splitref(&entry->aref, &new_entry->aref, new_adj); 722 723 uvm_rb_fixup(map, entry); 724 725 uvm_map_entry_link(map, entry, new_entry); 726 727 if (UVM_ET_ISSUBMAP(entry)) { 728 /* ... unlikely to happen, but play it safe */ 729 uvm_map_reference(new_entry->object.sub_map); 730 } else { 731 if (UVM_ET_ISOBJ(entry) && 732 entry->object.uvm_obj->pgops && 733 entry->object.uvm_obj->pgops->pgo_reference) 734 entry->object.uvm_obj->pgops->pgo_reference( 735 entry->object.uvm_obj); 736 } 737 738 uvm_tree_sanity(map, "clip_end leave"); 739 } 740 741 742 /* 743 * M A P - m a i n e n t r y p o i n t 744 */ 745 /* 746 * uvm_map: establish a valid mapping in a map 747 * 748 * => assume startp is page aligned. 749 * => assume size is a multiple of PAGE_SIZE. 750 * => assume sys_mmap provides enough of a "hint" to have us skip 751 * over text/data/bss area. 752 * => map must be unlocked (we will lock it) 753 * => <uobj,uoffset> value meanings (4 cases): 754 * [1] <NULL,uoffset> == uoffset is a hint for PMAP_PREFER 755 * [2] <NULL,UVM_UNKNOWN_OFFSET> == don't PMAP_PREFER 756 * [3] <uobj,uoffset> == normal mapping 757 * [4] <uobj,UVM_UNKNOWN_OFFSET> == uvm_map finds offset based on VA 758 * 759 * case [4] is for kernel mappings where we don't know the offset until 760 * we've found a virtual address. note that kernel object offsets are 761 * always relative to vm_map_min(kernel_map). 762 * 763 * => if `align' is non-zero, we align the virtual address to the specified 764 * alignment. 765 * this is provided as a mechanism for large pages. 766 * 767 * => XXXCDC: need way to map in external amap? 768 */ 769 770 int 771 uvm_map(struct vm_map *map, vaddr_t *startp /* IN/OUT */, vsize_t size, 772 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags) 773 { 774 struct uvm_map_args args; 775 struct vm_map_entry *new_entry; 776 int error; 777 778 KASSERT((flags & UVM_FLAG_QUANTUM) == 0 || VM_MAP_IS_KERNEL(map)); 779 KASSERT((size & PAGE_MASK) == 0); 780 781 /* 782 * for pager_map, allocate the new entry first to avoid sleeping 783 * for memory while we have the map locked. 784 * 785 * besides, because we allocates entries for in-kernel maps 786 * a bit differently (cf. uvm_kmapent_alloc/free), we need to 787 * allocate them before locking the map. 788 */ 789 790 new_entry = NULL; 791 if (VM_MAP_USE_KMAPENT(map) || (flags & UVM_FLAG_QUANTUM) || 792 map == pager_map) { 793 new_entry = uvm_mapent_alloc(map, (flags & UVM_FLAG_NOWAIT)); 794 if (__predict_false(new_entry == NULL)) 795 return ENOMEM; 796 if (flags & UVM_FLAG_QUANTUM) 797 new_entry->flags |= UVM_MAP_QUANTUM; 798 } 799 if (map == pager_map) 800 flags |= UVM_FLAG_NOMERGE; 801 802 error = uvm_map_prepare(map, *startp, size, uobj, uoffset, align, 803 flags, &args); 804 if (!error) { 805 error = uvm_map_enter(map, &args, new_entry); 806 *startp = args.uma_start; 807 } 808 809 #if defined(DEBUG) 810 if (!error && VM_MAP_IS_KERNEL(map)) { 811 uvm_km_check_empty(*startp, *startp + size, 812 (map->flags & VM_MAP_INTRSAFE) != 0); 813 } 814 #endif /* defined(DEBUG) */ 815 816 return error; 817 } 818 819 int 820 uvm_map_prepare(struct vm_map *map, vaddr_t start, vsize_t size, 821 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags, 822 struct uvm_map_args *args) 823 { 824 struct vm_map_entry *prev_entry; 825 vm_prot_t prot = UVM_PROTECTION(flags); 826 vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 827 828 UVMHIST_FUNC("uvm_map_prepare"); 829 UVMHIST_CALLED(maphist); 830 831 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 832 map, start, size, flags); 833 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 834 835 /* 836 * detect a popular device driver bug. 837 */ 838 839 KASSERT(doing_shutdown || curlwp != NULL || 840 (map->flags & VM_MAP_INTRSAFE)); 841 842 /* 843 * zero-sized mapping doesn't make any sense. 844 */ 845 KASSERT(size > 0); 846 847 KASSERT((~flags & (UVM_FLAG_NOWAIT | UVM_FLAG_WAITVA)) != 0); 848 849 uvm_tree_sanity(map, "map entry"); 850 851 /* 852 * check sanity of protection code 853 */ 854 855 if ((prot & maxprot) != prot) { 856 UVMHIST_LOG(maphist, "<- prot. failure: prot=0x%x, max=0x%x", 857 prot, maxprot,0,0); 858 return EACCES; 859 } 860 861 /* 862 * figure out where to put new VM range 863 */ 864 865 retry: 866 if (vm_map_lock_try(map) == FALSE) { 867 if (flags & UVM_FLAG_TRYLOCK) { 868 return EAGAIN; 869 } 870 vm_map_lock(map); /* could sleep here */ 871 } 872 if ((prev_entry = uvm_map_findspace(map, start, size, &start, 873 uobj, uoffset, align, flags)) == NULL) { 874 unsigned int timestamp; 875 876 if ((flags & UVM_FLAG_WAITVA) == 0) { 877 UVMHIST_LOG(maphist,"<- uvm_map_findspace failed!", 878 0,0,0,0); 879 vm_map_unlock(map); 880 return ENOMEM; 881 } 882 timestamp = map->timestamp; 883 UVMHIST_LOG(maphist,"waiting va timestamp=0x%x", 884 timestamp,0,0,0); 885 simple_lock(&map->flags_lock); 886 map->flags |= VM_MAP_WANTVA; 887 simple_unlock(&map->flags_lock); 888 vm_map_unlock(map); 889 890 /* 891 * wait until someone does unmap. 892 * XXX fragile locking 893 */ 894 895 simple_lock(&map->flags_lock); 896 while ((map->flags & VM_MAP_WANTVA) != 0 && 897 map->timestamp == timestamp) { 898 ltsleep(&map->header, PVM, "vmmapva", 0, 899 &map->flags_lock); 900 } 901 simple_unlock(&map->flags_lock); 902 goto retry; 903 } 904 905 #ifdef PMAP_GROWKERNEL 906 /* 907 * If the kernel pmap can't map the requested space, 908 * then allocate more resources for it. 909 */ 910 if (map == kernel_map && uvm_maxkaddr < (start + size)) 911 uvm_maxkaddr = pmap_growkernel(start + size); 912 #endif 913 914 UVMCNT_INCR(uvm_map_call); 915 916 /* 917 * if uobj is null, then uoffset is either a VAC hint for PMAP_PREFER 918 * [typically from uvm_map_reserve] or it is UVM_UNKNOWN_OFFSET. in 919 * either case we want to zero it before storing it in the map entry 920 * (because it looks strange and confusing when debugging...) 921 * 922 * if uobj is not null 923 * if uoffset is not UVM_UNKNOWN_OFFSET then we have a normal mapping 924 * and we do not need to change uoffset. 925 * if uoffset is UVM_UNKNOWN_OFFSET then we need to find the offset 926 * now (based on the starting address of the map). this case is 927 * for kernel object mappings where we don't know the offset until 928 * the virtual address is found (with uvm_map_findspace). the 929 * offset is the distance we are from the start of the map. 930 */ 931 932 if (uobj == NULL) { 933 uoffset = 0; 934 } else { 935 if (uoffset == UVM_UNKNOWN_OFFSET) { 936 KASSERT(UVM_OBJ_IS_KERN_OBJECT(uobj)); 937 uoffset = start - vm_map_min(kernel_map); 938 } 939 } 940 941 args->uma_flags = flags; 942 args->uma_prev = prev_entry; 943 args->uma_start = start; 944 args->uma_size = size; 945 args->uma_uobj = uobj; 946 args->uma_uoffset = uoffset; 947 948 return 0; 949 } 950 951 int 952 uvm_map_enter(struct vm_map *map, const struct uvm_map_args *args, 953 struct vm_map_entry *new_entry) 954 { 955 struct vm_map_entry *prev_entry = args->uma_prev; 956 struct vm_map_entry *dead = NULL; 957 958 const uvm_flag_t flags = args->uma_flags; 959 const vm_prot_t prot = UVM_PROTECTION(flags); 960 const vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 961 const vm_inherit_t inherit = UVM_INHERIT(flags); 962 const int amapwaitflag = (flags & UVM_FLAG_NOWAIT) ? 963 AMAP_EXTEND_NOWAIT : 0; 964 const int advice = UVM_ADVICE(flags); 965 const int meflagmask = UVM_MAP_NOMERGE | UVM_MAP_QUANTUM; 966 const int meflagval = (flags & UVM_FLAG_QUANTUM) ? 967 UVM_MAP_QUANTUM : 0; 968 969 vaddr_t start = args->uma_start; 970 vsize_t size = args->uma_size; 971 struct uvm_object *uobj = args->uma_uobj; 972 voff_t uoffset = args->uma_uoffset; 973 974 const int kmap = (vm_map_pmap(map) == pmap_kernel()); 975 int merged = 0; 976 int error; 977 int newetype; 978 979 UVMHIST_FUNC("uvm_map_enter"); 980 UVMHIST_CALLED(maphist); 981 982 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 983 map, start, size, flags); 984 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 985 986 if (flags & UVM_FLAG_QUANTUM) { 987 KASSERT(new_entry); 988 KASSERT(new_entry->flags & UVM_MAP_QUANTUM); 989 } 990 991 if (uobj) 992 newetype = UVM_ET_OBJ; 993 else 994 newetype = 0; 995 996 if (flags & UVM_FLAG_COPYONW) { 997 newetype |= UVM_ET_COPYONWRITE; 998 if ((flags & UVM_FLAG_OVERLAY) == 0) 999 newetype |= UVM_ET_NEEDSCOPY; 1000 } 1001 1002 /* 1003 * try and insert in map by extending previous entry, if possible. 1004 * XXX: we don't try and pull back the next entry. might be useful 1005 * for a stack, but we are currently allocating our stack in advance. 1006 */ 1007 1008 if (flags & UVM_FLAG_NOMERGE) 1009 goto nomerge; 1010 1011 if (prev_entry->etype == newetype && 1012 prev_entry->end == start && 1013 prev_entry != &map->header && 1014 prev_entry->object.uvm_obj == uobj) { 1015 1016 if ((prev_entry->flags & meflagmask) != meflagval) 1017 goto forwardmerge; 1018 1019 if (uobj && prev_entry->offset + 1020 (prev_entry->end - prev_entry->start) != uoffset) 1021 goto forwardmerge; 1022 1023 if (prev_entry->protection != prot || 1024 prev_entry->max_protection != maxprot) 1025 goto forwardmerge; 1026 1027 if (prev_entry->inheritance != inherit || 1028 prev_entry->advice != advice) 1029 goto forwardmerge; 1030 1031 /* wiring status must match (new area is unwired) */ 1032 if (VM_MAPENT_ISWIRED(prev_entry)) 1033 goto forwardmerge; 1034 1035 /* 1036 * can't extend a shared amap. note: no need to lock amap to 1037 * look at refs since we don't care about its exact value. 1038 * if it is one (i.e. we have only reference) it will stay there 1039 */ 1040 1041 if (prev_entry->aref.ar_amap && 1042 amap_refs(prev_entry->aref.ar_amap) != 1) { 1043 goto forwardmerge; 1044 } 1045 1046 if (prev_entry->aref.ar_amap) { 1047 error = amap_extend(prev_entry, size, 1048 amapwaitflag | AMAP_EXTEND_FORWARDS); 1049 if (error) 1050 goto done; 1051 } 1052 1053 if (kmap) 1054 UVMCNT_INCR(map_kbackmerge); 1055 else 1056 UVMCNT_INCR(map_ubackmerge); 1057 UVMHIST_LOG(maphist," starting back merge", 0, 0, 0, 0); 1058 1059 /* 1060 * drop our reference to uobj since we are extending a reference 1061 * that we already have (the ref count can not drop to zero). 1062 */ 1063 1064 if (uobj && uobj->pgops->pgo_detach) 1065 uobj->pgops->pgo_detach(uobj); 1066 1067 prev_entry->end += size; 1068 uvm_rb_fixup(map, prev_entry); 1069 1070 uvm_tree_sanity(map, "map backmerged"); 1071 1072 UVMHIST_LOG(maphist,"<- done (via backmerge)!", 0, 0, 0, 0); 1073 merged++; 1074 } 1075 1076 forwardmerge: 1077 if (prev_entry->next->etype == newetype && 1078 prev_entry->next->start == (start + size) && 1079 prev_entry->next != &map->header && 1080 prev_entry->next->object.uvm_obj == uobj) { 1081 1082 if ((prev_entry->next->flags & meflagmask) != meflagval) 1083 goto nomerge; 1084 1085 if (uobj && prev_entry->next->offset != uoffset + size) 1086 goto nomerge; 1087 1088 if (prev_entry->next->protection != prot || 1089 prev_entry->next->max_protection != maxprot) 1090 goto nomerge; 1091 1092 if (prev_entry->next->inheritance != inherit || 1093 prev_entry->next->advice != advice) 1094 goto nomerge; 1095 1096 /* wiring status must match (new area is unwired) */ 1097 if (VM_MAPENT_ISWIRED(prev_entry->next)) 1098 goto nomerge; 1099 1100 /* 1101 * can't extend a shared amap. note: no need to lock amap to 1102 * look at refs since we don't care about its exact value. 1103 * if it is one (i.e. we have only reference) it will stay there. 1104 * 1105 * note that we also can't merge two amaps, so if we 1106 * merged with the previous entry which has an amap, 1107 * and the next entry also has an amap, we give up. 1108 * 1109 * Interesting cases: 1110 * amap, new, amap -> give up second merge (single fwd extend) 1111 * amap, new, none -> double forward extend (extend again here) 1112 * none, new, amap -> double backward extend (done here) 1113 * uobj, new, amap -> single backward extend (done here) 1114 * 1115 * XXX should we attempt to deal with someone refilling 1116 * the deallocated region between two entries that are 1117 * backed by the same amap (ie, arefs is 2, "prev" and 1118 * "next" refer to it, and adding this allocation will 1119 * close the hole, thus restoring arefs to 1 and 1120 * deallocating the "next" vm_map_entry)? -- @@@ 1121 */ 1122 1123 if (prev_entry->next->aref.ar_amap && 1124 (amap_refs(prev_entry->next->aref.ar_amap) != 1 || 1125 (merged && prev_entry->aref.ar_amap))) { 1126 goto nomerge; 1127 } 1128 1129 if (merged) { 1130 /* 1131 * Try to extend the amap of the previous entry to 1132 * cover the next entry as well. If it doesn't work 1133 * just skip on, don't actually give up, since we've 1134 * already completed the back merge. 1135 */ 1136 if (prev_entry->aref.ar_amap) { 1137 if (amap_extend(prev_entry, 1138 prev_entry->next->end - 1139 prev_entry->next->start, 1140 amapwaitflag | AMAP_EXTEND_FORWARDS)) 1141 goto nomerge; 1142 } 1143 1144 /* 1145 * Try to extend the amap of the *next* entry 1146 * back to cover the new allocation *and* the 1147 * previous entry as well (the previous merge 1148 * didn't have an amap already otherwise we 1149 * wouldn't be checking here for an amap). If 1150 * it doesn't work just skip on, again, don't 1151 * actually give up, since we've already 1152 * completed the back merge. 1153 */ 1154 else if (prev_entry->next->aref.ar_amap) { 1155 if (amap_extend(prev_entry->next, 1156 prev_entry->end - 1157 prev_entry->start, 1158 amapwaitflag | AMAP_EXTEND_BACKWARDS)) 1159 goto nomerge; 1160 } 1161 } else { 1162 /* 1163 * Pull the next entry's amap backwards to cover this 1164 * new allocation. 1165 */ 1166 if (prev_entry->next->aref.ar_amap) { 1167 error = amap_extend(prev_entry->next, size, 1168 amapwaitflag | AMAP_EXTEND_BACKWARDS); 1169 if (error) 1170 goto done; 1171 } 1172 } 1173 1174 if (merged) { 1175 if (kmap) { 1176 UVMCNT_DECR(map_kbackmerge); 1177 UVMCNT_INCR(map_kbimerge); 1178 } else { 1179 UVMCNT_DECR(map_ubackmerge); 1180 UVMCNT_INCR(map_ubimerge); 1181 } 1182 } else { 1183 if (kmap) 1184 UVMCNT_INCR(map_kforwmerge); 1185 else 1186 UVMCNT_INCR(map_uforwmerge); 1187 } 1188 UVMHIST_LOG(maphist," starting forward merge", 0, 0, 0, 0); 1189 1190 /* 1191 * drop our reference to uobj since we are extending a reference 1192 * that we already have (the ref count can not drop to zero). 1193 * (if merged, we've already detached) 1194 */ 1195 if (uobj && uobj->pgops->pgo_detach && !merged) 1196 uobj->pgops->pgo_detach(uobj); 1197 1198 if (merged) { 1199 dead = prev_entry->next; 1200 prev_entry->end = dead->end; 1201 uvm_map_entry_unlink(map, dead); 1202 if (dead->aref.ar_amap != NULL) { 1203 prev_entry->aref = dead->aref; 1204 dead->aref.ar_amap = NULL; 1205 } 1206 } else { 1207 prev_entry->next->start -= size; 1208 if (prev_entry != &map->header) 1209 uvm_rb_fixup(map, prev_entry); 1210 if (uobj) 1211 prev_entry->next->offset = uoffset; 1212 } 1213 1214 uvm_tree_sanity(map, "map forwardmerged"); 1215 1216 UVMHIST_LOG(maphist,"<- done forwardmerge", 0, 0, 0, 0); 1217 merged++; 1218 } 1219 1220 nomerge: 1221 if (!merged) { 1222 UVMHIST_LOG(maphist," allocating new map entry", 0, 0, 0, 0); 1223 if (kmap) 1224 UVMCNT_INCR(map_knomerge); 1225 else 1226 UVMCNT_INCR(map_unomerge); 1227 1228 /* 1229 * allocate new entry and link it in. 1230 */ 1231 1232 if (new_entry == NULL) { 1233 new_entry = uvm_mapent_alloc(map, 1234 (flags & UVM_FLAG_NOWAIT)); 1235 if (__predict_false(new_entry == NULL)) { 1236 error = ENOMEM; 1237 goto done; 1238 } 1239 } 1240 new_entry->start = start; 1241 new_entry->end = new_entry->start + size; 1242 new_entry->object.uvm_obj = uobj; 1243 new_entry->offset = uoffset; 1244 1245 new_entry->etype = newetype; 1246 1247 if (flags & UVM_FLAG_NOMERGE) { 1248 new_entry->flags |= UVM_MAP_NOMERGE; 1249 } 1250 1251 new_entry->protection = prot; 1252 new_entry->max_protection = maxprot; 1253 new_entry->inheritance = inherit; 1254 new_entry->wired_count = 0; 1255 new_entry->advice = advice; 1256 if (flags & UVM_FLAG_OVERLAY) { 1257 1258 /* 1259 * to_add: for BSS we overallocate a little since we 1260 * are likely to extend 1261 */ 1262 1263 vaddr_t to_add = (flags & UVM_FLAG_AMAPPAD) ? 1264 UVM_AMAP_CHUNK << PAGE_SHIFT : 0; 1265 struct vm_amap *amap = amap_alloc(size, to_add, 1266 (flags & UVM_FLAG_NOWAIT) ? M_NOWAIT : M_WAITOK); 1267 if (__predict_false(amap == NULL)) { 1268 error = ENOMEM; 1269 goto done; 1270 } 1271 new_entry->aref.ar_pageoff = 0; 1272 new_entry->aref.ar_amap = amap; 1273 } else { 1274 new_entry->aref.ar_pageoff = 0; 1275 new_entry->aref.ar_amap = NULL; 1276 } 1277 uvm_map_entry_link(map, prev_entry, new_entry); 1278 1279 /* 1280 * Update the free space hint 1281 */ 1282 1283 if ((map->first_free == prev_entry) && 1284 (prev_entry->end >= new_entry->start)) 1285 map->first_free = new_entry; 1286 1287 new_entry = NULL; 1288 } 1289 1290 map->size += size; 1291 1292 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 1293 1294 error = 0; 1295 done: 1296 vm_map_unlock(map); 1297 if (new_entry) { 1298 if (error == 0) { 1299 KDASSERT(merged); 1300 uvm_mapent_free_merged(map, new_entry); 1301 } else { 1302 uvm_mapent_free(new_entry); 1303 } 1304 } 1305 if (dead) { 1306 KDASSERT(merged); 1307 uvm_mapent_free_merged(map, dead); 1308 } 1309 return error; 1310 } 1311 1312 /* 1313 * uvm_map_lookup_entry: find map entry at or before an address 1314 * 1315 * => map must at least be read-locked by caller 1316 * => entry is returned in "entry" 1317 * => return value is true if address is in the returned entry 1318 */ 1319 1320 boolean_t 1321 uvm_map_lookup_entry(struct vm_map *map, vaddr_t address, 1322 struct vm_map_entry **entry /* OUT */) 1323 { 1324 struct vm_map_entry *cur; 1325 boolean_t use_tree = FALSE; 1326 UVMHIST_FUNC("uvm_map_lookup_entry"); 1327 UVMHIST_CALLED(maphist); 1328 1329 UVMHIST_LOG(maphist,"(map=0x%x,addr=0x%x,ent=0x%x)", 1330 map, address, entry, 0); 1331 1332 /* 1333 * start looking either from the head of the 1334 * list, or from the hint. 1335 */ 1336 1337 simple_lock(&map->hint_lock); 1338 cur = map->hint; 1339 simple_unlock(&map->hint_lock); 1340 1341 if (cur == &map->header) 1342 cur = cur->next; 1343 1344 UVMCNT_INCR(uvm_mlk_call); 1345 if (address >= cur->start) { 1346 1347 /* 1348 * go from hint to end of list. 1349 * 1350 * but first, make a quick check to see if 1351 * we are already looking at the entry we 1352 * want (which is usually the case). 1353 * note also that we don't need to save the hint 1354 * here... it is the same hint (unless we are 1355 * at the header, in which case the hint didn't 1356 * buy us anything anyway). 1357 */ 1358 1359 if (cur != &map->header && cur->end > address) { 1360 UVMCNT_INCR(uvm_mlk_hint); 1361 *entry = cur; 1362 UVMHIST_LOG(maphist,"<- got it via hint (0x%x)", 1363 cur, 0, 0, 0); 1364 return (TRUE); 1365 } 1366 1367 if (map->nentries > 30) 1368 use_tree = TRUE; 1369 } else { 1370 1371 /* 1372 * invalid hint. use tree. 1373 */ 1374 use_tree = TRUE; 1375 } 1376 1377 uvm_tree_sanity(map, __func__); 1378 1379 if (use_tree) { 1380 struct vm_map_entry *prev = &map->header; 1381 cur = RB_ROOT(&map->rbhead); 1382 1383 /* 1384 * Simple lookup in the tree. Happens when the hint is 1385 * invalid, or nentries reach a threshold. 1386 */ 1387 while (cur) { 1388 if (address >= cur->start) { 1389 if (address < cur->end) { 1390 *entry = cur; 1391 goto got; 1392 } 1393 prev = cur; 1394 cur = RB_RIGHT(cur, rb_entry); 1395 } else 1396 cur = RB_LEFT(cur, rb_entry); 1397 } 1398 *entry = prev; 1399 goto failed; 1400 } 1401 1402 /* 1403 * search linearly 1404 */ 1405 1406 while (cur != &map->header) { 1407 if (cur->end > address) { 1408 if (address >= cur->start) { 1409 /* 1410 * save this lookup for future 1411 * hints, and return 1412 */ 1413 1414 *entry = cur; 1415 got: 1416 SAVE_HINT(map, map->hint, *entry); 1417 UVMHIST_LOG(maphist,"<- search got it (0x%x)", 1418 cur, 0, 0, 0); 1419 KDASSERT((*entry)->start <= address); 1420 KDASSERT(address < (*entry)->end); 1421 return (TRUE); 1422 } 1423 break; 1424 } 1425 cur = cur->next; 1426 } 1427 *entry = cur->prev; 1428 failed: 1429 SAVE_HINT(map, map->hint, *entry); 1430 UVMHIST_LOG(maphist,"<- failed!",0,0,0,0); 1431 KDASSERT((*entry) == &map->header || (*entry)->end <= address); 1432 KDASSERT((*entry)->next == &map->header || 1433 address < (*entry)->next->start); 1434 return (FALSE); 1435 } 1436 1437 /* 1438 * See if the range between start and start + length fits in the gap 1439 * entry->next->start and entry->end. Returns 1 if fits, 0 if doesn't 1440 * fit, and -1 address wraps around. 1441 */ 1442 static __inline int 1443 uvm_map_space_avail(vaddr_t *start, vsize_t length, voff_t uoffset, 1444 vsize_t align, int topdown, struct vm_map_entry *entry) 1445 { 1446 vaddr_t end; 1447 1448 #ifdef PMAP_PREFER 1449 /* 1450 * push start address forward as needed to avoid VAC alias problems. 1451 * we only do this if a valid offset is specified. 1452 */ 1453 1454 if (uoffset != UVM_UNKNOWN_OFFSET) 1455 PMAP_PREFER(uoffset, start, length, topdown); 1456 #endif 1457 if (align != 0) { 1458 if ((*start & (align - 1)) != 0) { 1459 if (topdown) 1460 *start &= ~(align - 1); 1461 else 1462 *start = roundup(*start, align); 1463 } 1464 /* 1465 * XXX Should we PMAP_PREFER() here again? 1466 * eh...i think we're okay 1467 */ 1468 } 1469 1470 /* 1471 * Find the end of the proposed new region. Be sure we didn't 1472 * wrap around the address; if so, we lose. Otherwise, if the 1473 * proposed new region fits before the next entry, we win. 1474 */ 1475 1476 end = *start + length; 1477 if (end < *start) 1478 return (-1); 1479 1480 if (entry->next->start >= end && *start >= entry->end) 1481 return (1); 1482 1483 return (0); 1484 } 1485 1486 /* 1487 * uvm_map_findspace: find "length" sized space in "map". 1488 * 1489 * => "hint" is a hint about where we want it, unless UVM_FLAG_FIXED is 1490 * set in "flags" (in which case we insist on using "hint"). 1491 * => "result" is VA returned 1492 * => uobj/uoffset are to be used to handle VAC alignment, if required 1493 * => if "align" is non-zero, we attempt to align to that value. 1494 * => caller must at least have read-locked map 1495 * => returns NULL on failure, or pointer to prev. map entry if success 1496 * => note this is a cross between the old vm_map_findspace and vm_map_find 1497 */ 1498 1499 struct vm_map_entry * 1500 uvm_map_findspace(struct vm_map *map, vaddr_t hint, vsize_t length, 1501 vaddr_t *result /* OUT */, struct uvm_object *uobj, voff_t uoffset, 1502 vsize_t align, int flags) 1503 { 1504 struct vm_map_entry *entry; 1505 struct vm_map_entry *child, *prev, *tmp; 1506 vaddr_t orig_hint; 1507 const int topdown = map->flags & VM_MAP_TOPDOWN; 1508 UVMHIST_FUNC("uvm_map_findspace"); 1509 UVMHIST_CALLED(maphist); 1510 1511 UVMHIST_LOG(maphist, "(map=0x%x, hint=0x%x, len=%d, flags=0x%x)", 1512 map, hint, length, flags); 1513 KASSERT((align & (align - 1)) == 0); 1514 KASSERT((flags & UVM_FLAG_FIXED) == 0 || align == 0); 1515 1516 uvm_tree_sanity(map, "map_findspace entry"); 1517 1518 /* 1519 * remember the original hint. if we are aligning, then we 1520 * may have to try again with no alignment constraint if 1521 * we fail the first time. 1522 */ 1523 1524 orig_hint = hint; 1525 if (hint < vm_map_min(map)) { /* check ranges ... */ 1526 if (flags & UVM_FLAG_FIXED) { 1527 UVMHIST_LOG(maphist,"<- VA below map range",0,0,0,0); 1528 return (NULL); 1529 } 1530 hint = vm_map_min(map); 1531 } 1532 if (hint > vm_map_max(map)) { 1533 UVMHIST_LOG(maphist,"<- VA 0x%x > range [0x%x->0x%x]", 1534 hint, vm_map_min(map), vm_map_max(map), 0); 1535 return (NULL); 1536 } 1537 1538 /* 1539 * Look for the first possible address; if there's already 1540 * something at this address, we have to start after it. 1541 */ 1542 1543 /* 1544 * @@@: there are four, no, eight cases to consider. 1545 * 1546 * 0: found, fixed, bottom up -> fail 1547 * 1: found, fixed, top down -> fail 1548 * 2: found, not fixed, bottom up -> start after entry->end, 1549 * loop up 1550 * 3: found, not fixed, top down -> start before entry->start, 1551 * loop down 1552 * 4: not found, fixed, bottom up -> check entry->next->start, fail 1553 * 5: not found, fixed, top down -> check entry->next->start, fail 1554 * 6: not found, not fixed, bottom up -> check entry->next->start, 1555 * loop up 1556 * 7: not found, not fixed, top down -> check entry->next->start, 1557 * loop down 1558 * 1559 * as you can see, it reduces to roughly five cases, and that 1560 * adding top down mapping only adds one unique case (without 1561 * it, there would be four cases). 1562 */ 1563 1564 if ((flags & UVM_FLAG_FIXED) == 0 && hint == vm_map_min(map)) { 1565 entry = map->first_free; 1566 } else { 1567 if (uvm_map_lookup_entry(map, hint, &entry)) { 1568 /* "hint" address already in use ... */ 1569 if (flags & UVM_FLAG_FIXED) { 1570 UVMHIST_LOG(maphist, "<- fixed & VA in use", 1571 0, 0, 0, 0); 1572 return (NULL); 1573 } 1574 if (topdown) 1575 /* Start from lower gap. */ 1576 entry = entry->prev; 1577 } else if (flags & UVM_FLAG_FIXED) { 1578 if (entry->next->start >= hint + length && 1579 hint + length > hint) 1580 goto found; 1581 1582 /* "hint" address is gap but too small */ 1583 UVMHIST_LOG(maphist, "<- fixed mapping failed", 1584 0, 0, 0, 0); 1585 return (NULL); /* only one shot at it ... */ 1586 } else { 1587 /* 1588 * See if given hint fits in this gap. 1589 */ 1590 switch (uvm_map_space_avail(&hint, length, 1591 uoffset, align, topdown, entry)) { 1592 case 1: 1593 goto found; 1594 case -1: 1595 goto wraparound; 1596 } 1597 1598 if (topdown) { 1599 /* 1600 * Still there is a chance to fit 1601 * if hint > entry->end. 1602 */ 1603 } else { 1604 /* Start from higher gap. */ 1605 entry = entry->next; 1606 if (entry == &map->header) 1607 goto notfound; 1608 goto nextgap; 1609 } 1610 } 1611 } 1612 1613 /* 1614 * Note that all UVM_FLAGS_FIXED case is already handled. 1615 */ 1616 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 1617 1618 /* Try to find the space in the red-black tree */ 1619 1620 /* Check slot before any entry */ 1621 hint = topdown ? entry->next->start - length : entry->end; 1622 switch (uvm_map_space_avail(&hint, length, uoffset, align, 1623 topdown, entry)) { 1624 case 1: 1625 goto found; 1626 case -1: 1627 goto wraparound; 1628 } 1629 1630 nextgap: 1631 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 1632 /* If there is not enough space in the whole tree, we fail */ 1633 tmp = RB_ROOT(&map->rbhead); 1634 if (tmp == NULL || tmp->space < length) 1635 goto notfound; 1636 1637 prev = NULL; /* previous candidate */ 1638 1639 /* Find an entry close to hint that has enough space */ 1640 for (; tmp;) { 1641 KASSERT(tmp->next->start == tmp->end + tmp->ownspace); 1642 if (topdown) { 1643 if (tmp->next->start < hint + length && 1644 (prev == NULL || tmp->end > prev->end)) { 1645 if (tmp->ownspace >= length) 1646 prev = tmp; 1647 else if ((child = RB_LEFT(tmp, rb_entry)) 1648 != NULL && child->space >= length) 1649 prev = tmp; 1650 } 1651 } else { 1652 if (tmp->end >= hint && 1653 (prev == NULL || tmp->end < prev->end)) { 1654 if (tmp->ownspace >= length) 1655 prev = tmp; 1656 else if ((child = RB_RIGHT(tmp, rb_entry)) 1657 != NULL && child->space >= length) 1658 prev = tmp; 1659 } 1660 } 1661 if (tmp->next->start < hint + length) 1662 child = RB_RIGHT(tmp, rb_entry); 1663 else if (tmp->end > hint) 1664 child = RB_LEFT(tmp, rb_entry); 1665 else { 1666 if (tmp->ownspace >= length) 1667 break; 1668 if (topdown) 1669 child = RB_LEFT(tmp, rb_entry); 1670 else 1671 child = RB_RIGHT(tmp, rb_entry); 1672 } 1673 if (child == NULL || child->space < length) 1674 break; 1675 tmp = child; 1676 } 1677 1678 if (tmp != NULL && tmp->start < hint && hint < tmp->next->start) { 1679 /* 1680 * Check if the entry that we found satifies the 1681 * space requirement 1682 */ 1683 if (topdown) { 1684 if (hint > tmp->next->start - length) 1685 hint = tmp->next->start - length; 1686 } else { 1687 if (hint < tmp->end) 1688 hint = tmp->end; 1689 } 1690 switch (uvm_map_space_avail(&hint, length, uoffset, align, 1691 topdown, tmp)) { 1692 case 1: 1693 entry = tmp; 1694 goto found; 1695 case -1: 1696 goto wraparound; 1697 } 1698 if (tmp->ownspace >= length) 1699 goto listsearch; 1700 } 1701 if (prev == NULL) 1702 goto notfound; 1703 1704 if (topdown) { 1705 KASSERT(orig_hint >= prev->next->start - length || 1706 prev->next->start - length > prev->next->start); 1707 hint = prev->next->start - length; 1708 } else { 1709 KASSERT(orig_hint <= prev->end); 1710 hint = prev->end; 1711 } 1712 switch (uvm_map_space_avail(&hint, length, uoffset, align, 1713 topdown, prev)) { 1714 case 1: 1715 entry = prev; 1716 goto found; 1717 case -1: 1718 goto wraparound; 1719 } 1720 if (prev->ownspace >= length) 1721 goto listsearch; 1722 1723 if (topdown) 1724 tmp = RB_LEFT(prev, rb_entry); 1725 else 1726 tmp = RB_RIGHT(prev, rb_entry); 1727 for (;;) { 1728 KASSERT(tmp && tmp->space >= length); 1729 if (topdown) 1730 child = RB_RIGHT(tmp, rb_entry); 1731 else 1732 child = RB_LEFT(tmp, rb_entry); 1733 if (child && child->space >= length) { 1734 tmp = child; 1735 continue; 1736 } 1737 if (tmp->ownspace >= length) 1738 break; 1739 if (topdown) 1740 tmp = RB_LEFT(tmp, rb_entry); 1741 else 1742 tmp = RB_RIGHT(tmp, rb_entry); 1743 } 1744 1745 if (topdown) { 1746 KASSERT(orig_hint >= tmp->next->start - length || 1747 tmp->next->start - length > tmp->next->start); 1748 hint = tmp->next->start - length; 1749 } else { 1750 KASSERT(orig_hint <= tmp->end); 1751 hint = tmp->end; 1752 } 1753 switch (uvm_map_space_avail(&hint, length, uoffset, align, 1754 topdown, tmp)) { 1755 case 1: 1756 entry = tmp; 1757 goto found; 1758 case -1: 1759 goto wraparound; 1760 } 1761 1762 /* 1763 * The tree fails to find an entry because of offset or alignment 1764 * restrictions. Search the list instead. 1765 */ 1766 listsearch: 1767 /* 1768 * Look through the rest of the map, trying to fit a new region in 1769 * the gap between existing regions, or after the very last region. 1770 * note: entry->end = base VA of current gap, 1771 * entry->next->start = VA of end of current gap 1772 */ 1773 1774 for (;;) { 1775 /* Update hint for current gap. */ 1776 hint = topdown ? entry->next->start - length : entry->end; 1777 1778 /* See if it fits. */ 1779 switch (uvm_map_space_avail(&hint, length, uoffset, align, 1780 topdown, entry)) { 1781 case 1: 1782 goto found; 1783 case -1: 1784 goto wraparound; 1785 } 1786 1787 /* Advance to next/previous gap */ 1788 if (topdown) { 1789 if (entry == &map->header) { 1790 UVMHIST_LOG(maphist, "<- failed (off start)", 1791 0,0,0,0); 1792 goto notfound; 1793 } 1794 entry = entry->prev; 1795 } else { 1796 entry = entry->next; 1797 if (entry == &map->header) { 1798 UVMHIST_LOG(maphist, "<- failed (off end)", 1799 0,0,0,0); 1800 goto notfound; 1801 } 1802 } 1803 } 1804 1805 found: 1806 SAVE_HINT(map, map->hint, entry); 1807 *result = hint; 1808 UVMHIST_LOG(maphist,"<- got it! (result=0x%x)", hint, 0,0,0); 1809 KASSERT( topdown || hint >= orig_hint); 1810 KASSERT(!topdown || hint <= orig_hint); 1811 KASSERT(entry->end <= hint); 1812 KASSERT(hint + length <= entry->next->start); 1813 return (entry); 1814 1815 wraparound: 1816 UVMHIST_LOG(maphist, "<- failed (wrap around)", 0,0,0,0); 1817 1818 return (NULL); 1819 1820 notfound: 1821 UVMHIST_LOG(maphist, "<- failed (notfound)", 0,0,0,0); 1822 1823 return (NULL); 1824 } 1825 1826 /* 1827 * U N M A P - m a i n h e l p e r f u n c t i o n s 1828 */ 1829 1830 /* 1831 * uvm_unmap_remove: remove mappings from a vm_map (from "start" up to "stop") 1832 * 1833 * => caller must check alignment and size 1834 * => map must be locked by caller 1835 * => we return a list of map entries that we've remove from the map 1836 * in "entry_list" 1837 */ 1838 1839 void 1840 uvm_unmap_remove(struct vm_map *map, vaddr_t start, vaddr_t end, 1841 struct vm_map_entry **entry_list /* OUT */, 1842 struct uvm_mapent_reservation *umr, int flags) 1843 { 1844 struct vm_map_entry *entry, *first_entry, *next; 1845 vaddr_t len; 1846 UVMHIST_FUNC("uvm_unmap_remove"); UVMHIST_CALLED(maphist); 1847 1848 UVMHIST_LOG(maphist,"(map=0x%x, start=0x%x, end=0x%x)", 1849 map, start, end, 0); 1850 VM_MAP_RANGE_CHECK(map, start, end); 1851 1852 uvm_tree_sanity(map, "unmap_remove entry"); 1853 1854 /* 1855 * find first entry 1856 */ 1857 1858 if (uvm_map_lookup_entry(map, start, &first_entry) == TRUE) { 1859 /* clip and go... */ 1860 entry = first_entry; 1861 UVM_MAP_CLIP_START(map, entry, start, umr); 1862 /* critical! prevents stale hint */ 1863 SAVE_HINT(map, entry, entry->prev); 1864 } else { 1865 entry = first_entry->next; 1866 } 1867 1868 /* 1869 * Save the free space hint 1870 */ 1871 1872 if (map->first_free->start >= start) 1873 map->first_free = entry->prev; 1874 1875 /* 1876 * note: we now re-use first_entry for a different task. we remove 1877 * a number of map entries from the map and save them in a linked 1878 * list headed by "first_entry". once we remove them from the map 1879 * the caller should unlock the map and drop the references to the 1880 * backing objects [c.f. uvm_unmap_detach]. the object is to 1881 * separate unmapping from reference dropping. why? 1882 * [1] the map has to be locked for unmapping 1883 * [2] the map need not be locked for reference dropping 1884 * [3] dropping references may trigger pager I/O, and if we hit 1885 * a pager that does synchronous I/O we may have to wait for it. 1886 * [4] we would like all waiting for I/O to occur with maps unlocked 1887 * so that we don't block other threads. 1888 */ 1889 1890 first_entry = NULL; 1891 *entry_list = NULL; 1892 1893 /* 1894 * break up the area into map entry sized regions and unmap. note 1895 * that all mappings have to be removed before we can even consider 1896 * dropping references to amaps or VM objects (otherwise we could end 1897 * up with a mapping to a page on the free list which would be very bad) 1898 */ 1899 1900 while ((entry != &map->header) && (entry->start < end)) { 1901 KASSERT((entry->flags & UVM_MAP_FIRST) == 0); 1902 1903 UVM_MAP_CLIP_END(map, entry, end, umr); 1904 next = entry->next; 1905 len = entry->end - entry->start; 1906 1907 /* 1908 * unwire before removing addresses from the pmap; otherwise 1909 * unwiring will put the entries back into the pmap (XXX). 1910 */ 1911 1912 if (VM_MAPENT_ISWIRED(entry)) { 1913 uvm_map_entry_unwire(map, entry); 1914 } 1915 if (flags & UVM_FLAG_VAONLY) { 1916 1917 /* nothing */ 1918 1919 } else if ((map->flags & VM_MAP_PAGEABLE) == 0) { 1920 1921 /* 1922 * if the map is non-pageable, any pages mapped there 1923 * must be wired and entered with pmap_kenter_pa(), 1924 * and we should free any such pages immediately. 1925 * this is mostly used for kmem_map and mb_map. 1926 */ 1927 1928 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 1929 uvm_km_pgremove_intrsafe(entry->start, 1930 entry->end); 1931 pmap_kremove(entry->start, len); 1932 } 1933 } else if (UVM_ET_ISOBJ(entry) && 1934 UVM_OBJ_IS_KERN_OBJECT(entry->object.uvm_obj)) { 1935 KASSERT(vm_map_pmap(map) == pmap_kernel()); 1936 1937 /* 1938 * note: kernel object mappings are currently used in 1939 * two ways: 1940 * [1] "normal" mappings of pages in the kernel object 1941 * [2] uvm_km_valloc'd allocations in which we 1942 * pmap_enter in some non-kernel-object page 1943 * (e.g. vmapbuf). 1944 * 1945 * for case [1], we need to remove the mapping from 1946 * the pmap and then remove the page from the kernel 1947 * object (because, once pages in a kernel object are 1948 * unmapped they are no longer needed, unlike, say, 1949 * a vnode where you might want the data to persist 1950 * until flushed out of a queue). 1951 * 1952 * for case [2], we need to remove the mapping from 1953 * the pmap. there shouldn't be any pages at the 1954 * specified offset in the kernel object [but it 1955 * doesn't hurt to call uvm_km_pgremove just to be 1956 * safe?] 1957 * 1958 * uvm_km_pgremove currently does the following: 1959 * for pages in the kernel object in range: 1960 * - drops the swap slot 1961 * - uvm_pagefree the page 1962 */ 1963 1964 /* 1965 * remove mappings from pmap and drop the pages 1966 * from the object. offsets are always relative 1967 * to vm_map_min(kernel_map). 1968 */ 1969 1970 pmap_remove(pmap_kernel(), entry->start, 1971 entry->start + len); 1972 uvm_km_pgremove(entry->start, entry->end); 1973 1974 /* 1975 * null out kernel_object reference, we've just 1976 * dropped it 1977 */ 1978 1979 entry->etype &= ~UVM_ET_OBJ; 1980 entry->object.uvm_obj = NULL; 1981 } else if (UVM_ET_ISOBJ(entry) || entry->aref.ar_amap) { 1982 1983 /* 1984 * remove mappings the standard way. 1985 */ 1986 1987 pmap_remove(map->pmap, entry->start, entry->end); 1988 } 1989 1990 #if defined(DEBUG) 1991 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 1992 1993 /* 1994 * check if there's remaining mapping, 1995 * which is a bug in caller. 1996 */ 1997 1998 vaddr_t va; 1999 for (va = entry->start; va < entry->end; 2000 va += PAGE_SIZE) { 2001 if (pmap_extract(vm_map_pmap(map), va, NULL)) { 2002 panic("uvm_unmap_remove: has mapping"); 2003 } 2004 } 2005 2006 if (VM_MAP_IS_KERNEL(map)) { 2007 uvm_km_check_empty(entry->start, entry->end, 2008 (map->flags & VM_MAP_INTRSAFE) != 0); 2009 } 2010 } 2011 #endif /* defined(DEBUG) */ 2012 2013 /* 2014 * remove entry from map and put it on our list of entries 2015 * that we've nuked. then go to next entry. 2016 */ 2017 2018 UVMHIST_LOG(maphist, " removed map entry 0x%x", entry, 0, 0,0); 2019 2020 /* critical! prevents stale hint */ 2021 SAVE_HINT(map, entry, entry->prev); 2022 2023 uvm_map_entry_unlink(map, entry); 2024 KASSERT(map->size >= len); 2025 map->size -= len; 2026 entry->prev = NULL; 2027 entry->next = first_entry; 2028 first_entry = entry; 2029 entry = next; 2030 } 2031 if ((map->flags & VM_MAP_DYING) == 0) { 2032 pmap_update(vm_map_pmap(map)); 2033 } 2034 2035 uvm_tree_sanity(map, "unmap_remove leave"); 2036 2037 /* 2038 * now we've cleaned up the map and are ready for the caller to drop 2039 * references to the mapped objects. 2040 */ 2041 2042 *entry_list = first_entry; 2043 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 2044 2045 simple_lock(&map->flags_lock); 2046 if (map->flags & VM_MAP_WANTVA) { 2047 map->flags &= ~VM_MAP_WANTVA; 2048 wakeup(&map->header); 2049 } 2050 simple_unlock(&map->flags_lock); 2051 } 2052 2053 /* 2054 * uvm_unmap_detach: drop references in a chain of map entries 2055 * 2056 * => we will free the map entries as we traverse the list. 2057 */ 2058 2059 void 2060 uvm_unmap_detach(struct vm_map_entry *first_entry, int flags) 2061 { 2062 struct vm_map_entry *next_entry; 2063 UVMHIST_FUNC("uvm_unmap_detach"); UVMHIST_CALLED(maphist); 2064 2065 while (first_entry) { 2066 KASSERT(!VM_MAPENT_ISWIRED(first_entry)); 2067 UVMHIST_LOG(maphist, 2068 " detach 0x%x: amap=0x%x, obj=0x%x, submap?=%d", 2069 first_entry, first_entry->aref.ar_amap, 2070 first_entry->object.uvm_obj, 2071 UVM_ET_ISSUBMAP(first_entry)); 2072 2073 /* 2074 * drop reference to amap, if we've got one 2075 */ 2076 2077 if (first_entry->aref.ar_amap) 2078 uvm_map_unreference_amap(first_entry, flags); 2079 2080 /* 2081 * drop reference to our backing object, if we've got one 2082 */ 2083 2084 KASSERT(!UVM_ET_ISSUBMAP(first_entry)); 2085 if (UVM_ET_ISOBJ(first_entry) && 2086 first_entry->object.uvm_obj->pgops->pgo_detach) { 2087 (*first_entry->object.uvm_obj->pgops->pgo_detach) 2088 (first_entry->object.uvm_obj); 2089 } 2090 next_entry = first_entry->next; 2091 uvm_mapent_free(first_entry); 2092 first_entry = next_entry; 2093 } 2094 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 2095 } 2096 2097 /* 2098 * E X T R A C T I O N F U N C T I O N S 2099 */ 2100 2101 /* 2102 * uvm_map_reserve: reserve space in a vm_map for future use. 2103 * 2104 * => we reserve space in a map by putting a dummy map entry in the 2105 * map (dummy means obj=NULL, amap=NULL, prot=VM_PROT_NONE) 2106 * => map should be unlocked (we will write lock it) 2107 * => we return true if we were able to reserve space 2108 * => XXXCDC: should be inline? 2109 */ 2110 2111 int 2112 uvm_map_reserve(struct vm_map *map, vsize_t size, 2113 vaddr_t offset /* hint for pmap_prefer */, 2114 vsize_t align /* alignment hint */, 2115 vaddr_t *raddr /* IN:hint, OUT: reserved VA */) 2116 { 2117 UVMHIST_FUNC("uvm_map_reserve"); UVMHIST_CALLED(maphist); 2118 2119 UVMHIST_LOG(maphist, "(map=0x%x, size=0x%x, offset=0x%x,addr=0x%x)", 2120 map,size,offset,raddr); 2121 2122 size = round_page(size); 2123 if (*raddr < vm_map_min(map)) 2124 *raddr = vm_map_min(map); /* hint */ 2125 2126 /* 2127 * reserve some virtual space. 2128 */ 2129 2130 if (uvm_map(map, raddr, size, NULL, offset, 0, 2131 UVM_MAPFLAG(UVM_PROT_NONE, UVM_PROT_NONE, UVM_INH_NONE, 2132 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE)) != 0) { 2133 UVMHIST_LOG(maphist, "<- done (no VM)", 0,0,0,0); 2134 return (FALSE); 2135 } 2136 2137 UVMHIST_LOG(maphist, "<- done (*raddr=0x%x)", *raddr,0,0,0); 2138 return (TRUE); 2139 } 2140 2141 /* 2142 * uvm_map_replace: replace a reserved (blank) area of memory with 2143 * real mappings. 2144 * 2145 * => caller must WRITE-LOCK the map 2146 * => we return TRUE if replacement was a success 2147 * => we expect the newents chain to have nnewents entrys on it and 2148 * we expect newents->prev to point to the last entry on the list 2149 * => note newents is allowed to be NULL 2150 */ 2151 2152 int 2153 uvm_map_replace(struct vm_map *map, vaddr_t start, vaddr_t end, 2154 struct vm_map_entry *newents, int nnewents) 2155 { 2156 struct vm_map_entry *oldent, *last; 2157 2158 uvm_tree_sanity(map, "map_replace entry"); 2159 2160 /* 2161 * first find the blank map entry at the specified address 2162 */ 2163 2164 if (!uvm_map_lookup_entry(map, start, &oldent)) { 2165 return (FALSE); 2166 } 2167 2168 /* 2169 * check to make sure we have a proper blank entry 2170 */ 2171 2172 if (oldent->start != start || oldent->end != end || 2173 oldent->object.uvm_obj != NULL || oldent->aref.ar_amap != NULL) { 2174 return (FALSE); 2175 } 2176 2177 #ifdef DIAGNOSTIC 2178 2179 /* 2180 * sanity check the newents chain 2181 */ 2182 2183 { 2184 struct vm_map_entry *tmpent = newents; 2185 int nent = 0; 2186 vaddr_t cur = start; 2187 2188 while (tmpent) { 2189 nent++; 2190 if (tmpent->start < cur) 2191 panic("uvm_map_replace1"); 2192 if (tmpent->start > tmpent->end || tmpent->end > end) { 2193 printf("tmpent->start=0x%lx, tmpent->end=0x%lx, end=0x%lx\n", 2194 tmpent->start, tmpent->end, end); 2195 panic("uvm_map_replace2"); 2196 } 2197 cur = tmpent->end; 2198 if (tmpent->next) { 2199 if (tmpent->next->prev != tmpent) 2200 panic("uvm_map_replace3"); 2201 } else { 2202 if (newents->prev != tmpent) 2203 panic("uvm_map_replace4"); 2204 } 2205 tmpent = tmpent->next; 2206 } 2207 if (nent != nnewents) 2208 panic("uvm_map_replace5"); 2209 } 2210 #endif 2211 2212 /* 2213 * map entry is a valid blank! replace it. (this does all the 2214 * work of map entry link/unlink...). 2215 */ 2216 2217 if (newents) { 2218 last = newents->prev; 2219 2220 /* critical: flush stale hints out of map */ 2221 SAVE_HINT(map, map->hint, newents); 2222 if (map->first_free == oldent) 2223 map->first_free = last; 2224 2225 last->next = oldent->next; 2226 last->next->prev = last; 2227 2228 /* Fix RB tree */ 2229 uvm_rb_remove(map, oldent); 2230 2231 newents->prev = oldent->prev; 2232 newents->prev->next = newents; 2233 map->nentries = map->nentries + (nnewents - 1); 2234 2235 /* Fixup the RB tree */ 2236 { 2237 int i; 2238 struct vm_map_entry *tmp; 2239 2240 tmp = newents; 2241 for (i = 0; i < nnewents && tmp; i++) { 2242 uvm_rb_insert(map, tmp); 2243 tmp = tmp->next; 2244 } 2245 } 2246 } else { 2247 2248 /* critical: flush stale hints out of map */ 2249 SAVE_HINT(map, map->hint, oldent->prev); 2250 if (map->first_free == oldent) 2251 map->first_free = oldent->prev; 2252 2253 /* NULL list of new entries: just remove the old one */ 2254 uvm_map_entry_unlink(map, oldent); 2255 } 2256 2257 uvm_tree_sanity(map, "map_replace leave"); 2258 2259 /* 2260 * now we can free the old blank entry, unlock the map and return. 2261 */ 2262 2263 uvm_mapent_free(oldent); 2264 return (TRUE); 2265 } 2266 2267 /* 2268 * uvm_map_extract: extract a mapping from a map and put it somewhere 2269 * (maybe removing the old mapping) 2270 * 2271 * => maps should be unlocked (we will write lock them) 2272 * => returns 0 on success, error code otherwise 2273 * => start must be page aligned 2274 * => len must be page sized 2275 * => flags: 2276 * UVM_EXTRACT_REMOVE: remove mappings from srcmap 2277 * UVM_EXTRACT_CONTIG: abort if unmapped area (advisory only) 2278 * UVM_EXTRACT_QREF: for a temporary extraction do quick obj refs 2279 * UVM_EXTRACT_FIXPROT: set prot to maxprot as we go 2280 * >>>NOTE: if you set REMOVE, you are not allowed to use CONTIG or QREF!<<< 2281 * >>>NOTE: QREF's must be unmapped via the QREF path, thus should only 2282 * be used from within the kernel in a kernel level map <<< 2283 */ 2284 2285 int 2286 uvm_map_extract(struct vm_map *srcmap, vaddr_t start, vsize_t len, 2287 struct vm_map *dstmap, vaddr_t *dstaddrp, int flags) 2288 { 2289 vaddr_t dstaddr, end, newend, oldoffset, fudge, orig_fudge; 2290 struct vm_map_entry *chain, *endchain, *entry, *orig_entry, *newentry, 2291 *deadentry, *oldentry; 2292 vsize_t elen; 2293 int nchain, error, copy_ok; 2294 UVMHIST_FUNC("uvm_map_extract"); UVMHIST_CALLED(maphist); 2295 2296 UVMHIST_LOG(maphist,"(srcmap=0x%x,start=0x%x, len=0x%x", srcmap, start, 2297 len,0); 2298 UVMHIST_LOG(maphist," ...,dstmap=0x%x, flags=0x%x)", dstmap,flags,0,0); 2299 2300 uvm_tree_sanity(srcmap, "map_extract src enter"); 2301 uvm_tree_sanity(dstmap, "map_extract dst enter"); 2302 2303 /* 2304 * step 0: sanity check: start must be on a page boundary, length 2305 * must be page sized. can't ask for CONTIG/QREF if you asked for 2306 * REMOVE. 2307 */ 2308 2309 KASSERT((start & PAGE_MASK) == 0 && (len & PAGE_MASK) == 0); 2310 KASSERT((flags & UVM_EXTRACT_REMOVE) == 0 || 2311 (flags & (UVM_EXTRACT_CONTIG|UVM_EXTRACT_QREF)) == 0); 2312 2313 /* 2314 * step 1: reserve space in the target map for the extracted area 2315 */ 2316 2317 dstaddr = vm_map_min(dstmap); 2318 if (uvm_map_reserve(dstmap, len, start, 0, &dstaddr) == FALSE) 2319 return (ENOMEM); 2320 *dstaddrp = dstaddr; /* pass address back to caller */ 2321 UVMHIST_LOG(maphist, " dstaddr=0x%x", dstaddr,0,0,0); 2322 2323 /* 2324 * step 2: setup for the extraction process loop by init'ing the 2325 * map entry chain, locking src map, and looking up the first useful 2326 * entry in the map. 2327 */ 2328 2329 end = start + len; 2330 newend = dstaddr + len; 2331 chain = endchain = NULL; 2332 nchain = 0; 2333 vm_map_lock(srcmap); 2334 2335 if (uvm_map_lookup_entry(srcmap, start, &entry)) { 2336 2337 /* "start" is within an entry */ 2338 if (flags & UVM_EXTRACT_QREF) { 2339 2340 /* 2341 * for quick references we don't clip the entry, so 2342 * the entry may map space "before" the starting 2343 * virtual address... this is the "fudge" factor 2344 * (which can be non-zero only the first time 2345 * through the "while" loop in step 3). 2346 */ 2347 2348 fudge = start - entry->start; 2349 } else { 2350 2351 /* 2352 * normal reference: we clip the map to fit (thus 2353 * fudge is zero) 2354 */ 2355 2356 UVM_MAP_CLIP_START(srcmap, entry, start, NULL); 2357 SAVE_HINT(srcmap, srcmap->hint, entry->prev); 2358 fudge = 0; 2359 } 2360 } else { 2361 2362 /* "start" is not within an entry ... skip to next entry */ 2363 if (flags & UVM_EXTRACT_CONTIG) { 2364 error = EINVAL; 2365 goto bad; /* definite hole here ... */ 2366 } 2367 2368 entry = entry->next; 2369 fudge = 0; 2370 } 2371 2372 /* save values from srcmap for step 6 */ 2373 orig_entry = entry; 2374 orig_fudge = fudge; 2375 2376 /* 2377 * step 3: now start looping through the map entries, extracting 2378 * as we go. 2379 */ 2380 2381 while (entry->start < end && entry != &srcmap->header) { 2382 2383 /* if we are not doing a quick reference, clip it */ 2384 if ((flags & UVM_EXTRACT_QREF) == 0) 2385 UVM_MAP_CLIP_END(srcmap, entry, end, NULL); 2386 2387 /* clear needs_copy (allow chunking) */ 2388 if (UVM_ET_ISNEEDSCOPY(entry)) { 2389 amap_copy(srcmap, entry, M_NOWAIT, TRUE, start, end); 2390 if (UVM_ET_ISNEEDSCOPY(entry)) { /* failed? */ 2391 error = ENOMEM; 2392 goto bad; 2393 } 2394 2395 /* amap_copy could clip (during chunk)! update fudge */ 2396 if (fudge) { 2397 fudge = start - entry->start; 2398 orig_fudge = fudge; 2399 } 2400 } 2401 2402 /* calculate the offset of this from "start" */ 2403 oldoffset = (entry->start + fudge) - start; 2404 2405 /* allocate a new map entry */ 2406 newentry = uvm_mapent_alloc(dstmap, 0); 2407 if (newentry == NULL) { 2408 error = ENOMEM; 2409 goto bad; 2410 } 2411 2412 /* set up new map entry */ 2413 newentry->next = NULL; 2414 newentry->prev = endchain; 2415 newentry->start = dstaddr + oldoffset; 2416 newentry->end = 2417 newentry->start + (entry->end - (entry->start + fudge)); 2418 if (newentry->end > newend || newentry->end < newentry->start) 2419 newentry->end = newend; 2420 newentry->object.uvm_obj = entry->object.uvm_obj; 2421 if (newentry->object.uvm_obj) { 2422 if (newentry->object.uvm_obj->pgops->pgo_reference) 2423 newentry->object.uvm_obj->pgops-> 2424 pgo_reference(newentry->object.uvm_obj); 2425 newentry->offset = entry->offset + fudge; 2426 } else { 2427 newentry->offset = 0; 2428 } 2429 newentry->etype = entry->etype; 2430 newentry->protection = (flags & UVM_EXTRACT_FIXPROT) ? 2431 entry->max_protection : entry->protection; 2432 newentry->max_protection = entry->max_protection; 2433 newentry->inheritance = entry->inheritance; 2434 newentry->wired_count = 0; 2435 newentry->aref.ar_amap = entry->aref.ar_amap; 2436 if (newentry->aref.ar_amap) { 2437 newentry->aref.ar_pageoff = 2438 entry->aref.ar_pageoff + (fudge >> PAGE_SHIFT); 2439 uvm_map_reference_amap(newentry, AMAP_SHARED | 2440 ((flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0)); 2441 } else { 2442 newentry->aref.ar_pageoff = 0; 2443 } 2444 newentry->advice = entry->advice; 2445 2446 /* now link it on the chain */ 2447 nchain++; 2448 if (endchain == NULL) { 2449 chain = endchain = newentry; 2450 } else { 2451 endchain->next = newentry; 2452 endchain = newentry; 2453 } 2454 2455 /* end of 'while' loop! */ 2456 if ((flags & UVM_EXTRACT_CONTIG) && entry->end < end && 2457 (entry->next == &srcmap->header || 2458 entry->next->start != entry->end)) { 2459 error = EINVAL; 2460 goto bad; 2461 } 2462 entry = entry->next; 2463 fudge = 0; 2464 } 2465 2466 /* 2467 * step 4: close off chain (in format expected by uvm_map_replace) 2468 */ 2469 2470 if (chain) 2471 chain->prev = endchain; 2472 2473 /* 2474 * step 5: attempt to lock the dest map so we can pmap_copy. 2475 * note usage of copy_ok: 2476 * 1 => dstmap locked, pmap_copy ok, and we "replace" here (step 5) 2477 * 0 => dstmap unlocked, NO pmap_copy, and we will "replace" in step 7 2478 */ 2479 2480 if (srcmap == dstmap || vm_map_lock_try(dstmap) == TRUE) { 2481 copy_ok = 1; 2482 if (!uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2483 nchain)) { 2484 if (srcmap != dstmap) 2485 vm_map_unlock(dstmap); 2486 error = EIO; 2487 goto bad; 2488 } 2489 } else { 2490 copy_ok = 0; 2491 /* replace defered until step 7 */ 2492 } 2493 2494 /* 2495 * step 6: traverse the srcmap a second time to do the following: 2496 * - if we got a lock on the dstmap do pmap_copy 2497 * - if UVM_EXTRACT_REMOVE remove the entries 2498 * we make use of orig_entry and orig_fudge (saved in step 2) 2499 */ 2500 2501 if (copy_ok || (flags & UVM_EXTRACT_REMOVE)) { 2502 2503 /* purge possible stale hints from srcmap */ 2504 if (flags & UVM_EXTRACT_REMOVE) { 2505 SAVE_HINT(srcmap, srcmap->hint, orig_entry->prev); 2506 if (srcmap->first_free->start >= start) 2507 srcmap->first_free = orig_entry->prev; 2508 } 2509 2510 entry = orig_entry; 2511 fudge = orig_fudge; 2512 deadentry = NULL; /* for UVM_EXTRACT_REMOVE */ 2513 2514 while (entry->start < end && entry != &srcmap->header) { 2515 if (copy_ok) { 2516 oldoffset = (entry->start + fudge) - start; 2517 elen = MIN(end, entry->end) - 2518 (entry->start + fudge); 2519 pmap_copy(dstmap->pmap, srcmap->pmap, 2520 dstaddr + oldoffset, elen, 2521 entry->start + fudge); 2522 } 2523 2524 /* we advance "entry" in the following if statement */ 2525 if (flags & UVM_EXTRACT_REMOVE) { 2526 pmap_remove(srcmap->pmap, entry->start, 2527 entry->end); 2528 oldentry = entry; /* save entry */ 2529 entry = entry->next; /* advance */ 2530 uvm_map_entry_unlink(srcmap, oldentry); 2531 /* add to dead list */ 2532 oldentry->next = deadentry; 2533 deadentry = oldentry; 2534 } else { 2535 entry = entry->next; /* advance */ 2536 } 2537 2538 /* end of 'while' loop */ 2539 fudge = 0; 2540 } 2541 pmap_update(srcmap->pmap); 2542 2543 /* 2544 * unlock dstmap. we will dispose of deadentry in 2545 * step 7 if needed 2546 */ 2547 2548 if (copy_ok && srcmap != dstmap) 2549 vm_map_unlock(dstmap); 2550 2551 } else { 2552 deadentry = NULL; 2553 } 2554 2555 /* 2556 * step 7: we are done with the source map, unlock. if copy_ok 2557 * is 0 then we have not replaced the dummy mapping in dstmap yet 2558 * and we need to do so now. 2559 */ 2560 2561 vm_map_unlock(srcmap); 2562 if ((flags & UVM_EXTRACT_REMOVE) && deadentry) 2563 uvm_unmap_detach(deadentry, 0); /* dispose of old entries */ 2564 2565 /* now do the replacement if we didn't do it in step 5 */ 2566 if (copy_ok == 0) { 2567 vm_map_lock(dstmap); 2568 error = uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2569 nchain); 2570 vm_map_unlock(dstmap); 2571 2572 if (error == FALSE) { 2573 error = EIO; 2574 goto bad2; 2575 } 2576 } 2577 2578 uvm_tree_sanity(srcmap, "map_extract src leave"); 2579 uvm_tree_sanity(dstmap, "map_extract dst leave"); 2580 2581 return (0); 2582 2583 /* 2584 * bad: failure recovery 2585 */ 2586 bad: 2587 vm_map_unlock(srcmap); 2588 bad2: /* src already unlocked */ 2589 if (chain) 2590 uvm_unmap_detach(chain, 2591 (flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0); 2592 2593 uvm_tree_sanity(srcmap, "map_extract src err leave"); 2594 uvm_tree_sanity(dstmap, "map_extract dst err leave"); 2595 2596 uvm_unmap(dstmap, dstaddr, dstaddr+len); /* ??? */ 2597 return (error); 2598 } 2599 2600 /* end of extraction functions */ 2601 2602 /* 2603 * uvm_map_submap: punch down part of a map into a submap 2604 * 2605 * => only the kernel_map is allowed to be submapped 2606 * => the purpose of submapping is to break up the locking granularity 2607 * of a larger map 2608 * => the range specified must have been mapped previously with a uvm_map() 2609 * call [with uobj==NULL] to create a blank map entry in the main map. 2610 * [And it had better still be blank!] 2611 * => maps which contain submaps should never be copied or forked. 2612 * => to remove a submap, use uvm_unmap() on the main map 2613 * and then uvm_map_deallocate() the submap. 2614 * => main map must be unlocked. 2615 * => submap must have been init'd and have a zero reference count. 2616 * [need not be locked as we don't actually reference it] 2617 */ 2618 2619 int 2620 uvm_map_submap(struct vm_map *map, vaddr_t start, vaddr_t end, 2621 struct vm_map *submap) 2622 { 2623 struct vm_map_entry *entry; 2624 struct uvm_mapent_reservation umr; 2625 int error; 2626 2627 uvm_mapent_reserve(map, &umr, 2, 0); 2628 2629 vm_map_lock(map); 2630 VM_MAP_RANGE_CHECK(map, start, end); 2631 2632 if (uvm_map_lookup_entry(map, start, &entry)) { 2633 UVM_MAP_CLIP_START(map, entry, start, &umr); 2634 UVM_MAP_CLIP_END(map, entry, end, &umr); /* to be safe */ 2635 } else { 2636 entry = NULL; 2637 } 2638 2639 if (entry != NULL && 2640 entry->start == start && entry->end == end && 2641 entry->object.uvm_obj == NULL && entry->aref.ar_amap == NULL && 2642 !UVM_ET_ISCOPYONWRITE(entry) && !UVM_ET_ISNEEDSCOPY(entry)) { 2643 entry->etype |= UVM_ET_SUBMAP; 2644 entry->object.sub_map = submap; 2645 entry->offset = 0; 2646 uvm_map_reference(submap); 2647 error = 0; 2648 } else { 2649 error = EINVAL; 2650 } 2651 vm_map_unlock(map); 2652 2653 uvm_mapent_unreserve(map, &umr); 2654 2655 return error; 2656 } 2657 2658 /* 2659 * uvm_map_setup_kernel: init in-kernel map 2660 * 2661 * => map must not be in service yet. 2662 */ 2663 2664 void 2665 uvm_map_setup_kernel(struct vm_map_kernel *map, 2666 vaddr_t min, vaddr_t max, int flags) 2667 { 2668 2669 uvm_map_setup(&map->vmk_map, min, max, flags); 2670 2671 LIST_INIT(&map->vmk_kentry_free); 2672 map->vmk_merged_entries = NULL; 2673 } 2674 2675 2676 /* 2677 * uvm_map_protect: change map protection 2678 * 2679 * => set_max means set max_protection. 2680 * => map must be unlocked. 2681 */ 2682 2683 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 2684 ~VM_PROT_WRITE : VM_PROT_ALL) 2685 2686 int 2687 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end, 2688 vm_prot_t new_prot, boolean_t set_max) 2689 { 2690 struct vm_map_entry *current, *entry; 2691 int error = 0; 2692 UVMHIST_FUNC("uvm_map_protect"); UVMHIST_CALLED(maphist); 2693 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_prot=0x%x)", 2694 map, start, end, new_prot); 2695 2696 vm_map_lock(map); 2697 VM_MAP_RANGE_CHECK(map, start, end); 2698 if (uvm_map_lookup_entry(map, start, &entry)) { 2699 UVM_MAP_CLIP_START(map, entry, start, NULL); 2700 } else { 2701 entry = entry->next; 2702 } 2703 2704 /* 2705 * make a first pass to check for protection violations. 2706 */ 2707 2708 current = entry; 2709 while ((current != &map->header) && (current->start < end)) { 2710 if (UVM_ET_ISSUBMAP(current)) { 2711 error = EINVAL; 2712 goto out; 2713 } 2714 if ((new_prot & current->max_protection) != new_prot) { 2715 error = EACCES; 2716 goto out; 2717 } 2718 /* 2719 * Don't allow VM_PROT_EXECUTE to be set on entries that 2720 * point to vnodes that are associated with a NOEXEC file 2721 * system. 2722 */ 2723 if (UVM_ET_ISOBJ(current) && 2724 UVM_OBJ_IS_VNODE(current->object.uvm_obj)) { 2725 struct vnode *vp = 2726 (struct vnode *) current->object.uvm_obj; 2727 2728 if ((new_prot & VM_PROT_EXECUTE) != 0 && 2729 (vp->v_mount->mnt_flag & MNT_NOEXEC) != 0) { 2730 error = EACCES; 2731 goto out; 2732 } 2733 } 2734 current = current->next; 2735 } 2736 2737 /* go back and fix up protections (no need to clip this time). */ 2738 2739 current = entry; 2740 while ((current != &map->header) && (current->start < end)) { 2741 vm_prot_t old_prot; 2742 2743 UVM_MAP_CLIP_END(map, current, end, NULL); 2744 old_prot = current->protection; 2745 if (set_max) 2746 current->protection = 2747 (current->max_protection = new_prot) & old_prot; 2748 else 2749 current->protection = new_prot; 2750 2751 /* 2752 * update physical map if necessary. worry about copy-on-write 2753 * here -- CHECK THIS XXX 2754 */ 2755 2756 if (current->protection != old_prot) { 2757 /* update pmap! */ 2758 pmap_protect(map->pmap, current->start, current->end, 2759 current->protection & MASK(entry)); 2760 2761 /* 2762 * If this entry points at a vnode, and the 2763 * protection includes VM_PROT_EXECUTE, mark 2764 * the vnode as VEXECMAP. 2765 */ 2766 if (UVM_ET_ISOBJ(current)) { 2767 struct uvm_object *uobj = 2768 current->object.uvm_obj; 2769 2770 if (UVM_OBJ_IS_VNODE(uobj) && 2771 (current->protection & VM_PROT_EXECUTE)) 2772 vn_markexec((struct vnode *) uobj); 2773 } 2774 } 2775 2776 /* 2777 * If the map is configured to lock any future mappings, 2778 * wire this entry now if the old protection was VM_PROT_NONE 2779 * and the new protection is not VM_PROT_NONE. 2780 */ 2781 2782 if ((map->flags & VM_MAP_WIREFUTURE) != 0 && 2783 VM_MAPENT_ISWIRED(entry) == 0 && 2784 old_prot == VM_PROT_NONE && 2785 new_prot != VM_PROT_NONE) { 2786 if (uvm_map_pageable(map, entry->start, 2787 entry->end, FALSE, 2788 UVM_LK_ENTER|UVM_LK_EXIT) != 0) { 2789 2790 /* 2791 * If locking the entry fails, remember the 2792 * error if it's the first one. Note we 2793 * still continue setting the protection in 2794 * the map, but will return the error 2795 * condition regardless. 2796 * 2797 * XXX Ignore what the actual error is, 2798 * XXX just call it a resource shortage 2799 * XXX so that it doesn't get confused 2800 * XXX what uvm_map_protect() itself would 2801 * XXX normally return. 2802 */ 2803 2804 error = ENOMEM; 2805 } 2806 } 2807 current = current->next; 2808 } 2809 pmap_update(map->pmap); 2810 2811 out: 2812 vm_map_unlock(map); 2813 2814 UVMHIST_LOG(maphist, "<- done, error=%d",error,0,0,0); 2815 return error; 2816 } 2817 2818 #undef MASK 2819 2820 /* 2821 * uvm_map_inherit: set inheritance code for range of addrs in map. 2822 * 2823 * => map must be unlocked 2824 * => note that the inherit code is used during a "fork". see fork 2825 * code for details. 2826 */ 2827 2828 int 2829 uvm_map_inherit(struct vm_map *map, vaddr_t start, vaddr_t end, 2830 vm_inherit_t new_inheritance) 2831 { 2832 struct vm_map_entry *entry, *temp_entry; 2833 UVMHIST_FUNC("uvm_map_inherit"); UVMHIST_CALLED(maphist); 2834 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_inh=0x%x)", 2835 map, start, end, new_inheritance); 2836 2837 switch (new_inheritance) { 2838 case MAP_INHERIT_NONE: 2839 case MAP_INHERIT_COPY: 2840 case MAP_INHERIT_SHARE: 2841 break; 2842 default: 2843 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 2844 return EINVAL; 2845 } 2846 2847 vm_map_lock(map); 2848 VM_MAP_RANGE_CHECK(map, start, end); 2849 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 2850 entry = temp_entry; 2851 UVM_MAP_CLIP_START(map, entry, start, NULL); 2852 } else { 2853 entry = temp_entry->next; 2854 } 2855 while ((entry != &map->header) && (entry->start < end)) { 2856 UVM_MAP_CLIP_END(map, entry, end, NULL); 2857 entry->inheritance = new_inheritance; 2858 entry = entry->next; 2859 } 2860 vm_map_unlock(map); 2861 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 2862 return 0; 2863 } 2864 2865 /* 2866 * uvm_map_advice: set advice code for range of addrs in map. 2867 * 2868 * => map must be unlocked 2869 */ 2870 2871 int 2872 uvm_map_advice(struct vm_map *map, vaddr_t start, vaddr_t end, int new_advice) 2873 { 2874 struct vm_map_entry *entry, *temp_entry; 2875 UVMHIST_FUNC("uvm_map_advice"); UVMHIST_CALLED(maphist); 2876 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_adv=0x%x)", 2877 map, start, end, new_advice); 2878 2879 vm_map_lock(map); 2880 VM_MAP_RANGE_CHECK(map, start, end); 2881 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 2882 entry = temp_entry; 2883 UVM_MAP_CLIP_START(map, entry, start, NULL); 2884 } else { 2885 entry = temp_entry->next; 2886 } 2887 2888 /* 2889 * XXXJRT: disallow holes? 2890 */ 2891 2892 while ((entry != &map->header) && (entry->start < end)) { 2893 UVM_MAP_CLIP_END(map, entry, end, NULL); 2894 2895 switch (new_advice) { 2896 case MADV_NORMAL: 2897 case MADV_RANDOM: 2898 case MADV_SEQUENTIAL: 2899 /* nothing special here */ 2900 break; 2901 2902 default: 2903 vm_map_unlock(map); 2904 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 2905 return EINVAL; 2906 } 2907 entry->advice = new_advice; 2908 entry = entry->next; 2909 } 2910 2911 vm_map_unlock(map); 2912 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 2913 return 0; 2914 } 2915 2916 /* 2917 * uvm_map_pageable: sets the pageability of a range in a map. 2918 * 2919 * => wires map entries. should not be used for transient page locking. 2920 * for that, use uvm_fault_wire()/uvm_fault_unwire() (see uvm_vslock()). 2921 * => regions sepcified as not pageable require lock-down (wired) memory 2922 * and page tables. 2923 * => map must never be read-locked 2924 * => if islocked is TRUE, map is already write-locked 2925 * => we always unlock the map, since we must downgrade to a read-lock 2926 * to call uvm_fault_wire() 2927 * => XXXCDC: check this and try and clean it up. 2928 */ 2929 2930 int 2931 uvm_map_pageable(struct vm_map *map, vaddr_t start, vaddr_t end, 2932 boolean_t new_pageable, int lockflags) 2933 { 2934 struct vm_map_entry *entry, *start_entry, *failed_entry; 2935 int rv; 2936 #ifdef DIAGNOSTIC 2937 u_int timestamp_save; 2938 #endif 2939 UVMHIST_FUNC("uvm_map_pageable"); UVMHIST_CALLED(maphist); 2940 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_pageable=0x%x)", 2941 map, start, end, new_pageable); 2942 KASSERT(map->flags & VM_MAP_PAGEABLE); 2943 2944 if ((lockflags & UVM_LK_ENTER) == 0) 2945 vm_map_lock(map); 2946 VM_MAP_RANGE_CHECK(map, start, end); 2947 2948 /* 2949 * only one pageability change may take place at one time, since 2950 * uvm_fault_wire assumes it will be called only once for each 2951 * wiring/unwiring. therefore, we have to make sure we're actually 2952 * changing the pageability for the entire region. we do so before 2953 * making any changes. 2954 */ 2955 2956 if (uvm_map_lookup_entry(map, start, &start_entry) == FALSE) { 2957 if ((lockflags & UVM_LK_EXIT) == 0) 2958 vm_map_unlock(map); 2959 2960 UVMHIST_LOG(maphist,"<- done (fault)",0,0,0,0); 2961 return EFAULT; 2962 } 2963 entry = start_entry; 2964 2965 /* 2966 * handle wiring and unwiring separately. 2967 */ 2968 2969 if (new_pageable) { /* unwire */ 2970 UVM_MAP_CLIP_START(map, entry, start, NULL); 2971 2972 /* 2973 * unwiring. first ensure that the range to be unwired is 2974 * really wired down and that there are no holes. 2975 */ 2976 2977 while ((entry != &map->header) && (entry->start < end)) { 2978 if (entry->wired_count == 0 || 2979 (entry->end < end && 2980 (entry->next == &map->header || 2981 entry->next->start > entry->end))) { 2982 if ((lockflags & UVM_LK_EXIT) == 0) 2983 vm_map_unlock(map); 2984 UVMHIST_LOG(maphist, "<- done (INVAL)",0,0,0,0); 2985 return EINVAL; 2986 } 2987 entry = entry->next; 2988 } 2989 2990 /* 2991 * POSIX 1003.1b - a single munlock call unlocks a region, 2992 * regardless of the number of mlock calls made on that 2993 * region. 2994 */ 2995 2996 entry = start_entry; 2997 while ((entry != &map->header) && (entry->start < end)) { 2998 UVM_MAP_CLIP_END(map, entry, end, NULL); 2999 if (VM_MAPENT_ISWIRED(entry)) 3000 uvm_map_entry_unwire(map, entry); 3001 entry = entry->next; 3002 } 3003 if ((lockflags & UVM_LK_EXIT) == 0) 3004 vm_map_unlock(map); 3005 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3006 return 0; 3007 } 3008 3009 /* 3010 * wire case: in two passes [XXXCDC: ugly block of code here] 3011 * 3012 * 1: holding the write lock, we create any anonymous maps that need 3013 * to be created. then we clip each map entry to the region to 3014 * be wired and increment its wiring count. 3015 * 3016 * 2: we downgrade to a read lock, and call uvm_fault_wire to fault 3017 * in the pages for any newly wired area (wired_count == 1). 3018 * 3019 * downgrading to a read lock for uvm_fault_wire avoids a possible 3020 * deadlock with another thread that may have faulted on one of 3021 * the pages to be wired (it would mark the page busy, blocking 3022 * us, then in turn block on the map lock that we hold). because 3023 * of problems in the recursive lock package, we cannot upgrade 3024 * to a write lock in vm_map_lookup. thus, any actions that 3025 * require the write lock must be done beforehand. because we 3026 * keep the read lock on the map, the copy-on-write status of the 3027 * entries we modify here cannot change. 3028 */ 3029 3030 while ((entry != &map->header) && (entry->start < end)) { 3031 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3032 3033 /* 3034 * perform actions of vm_map_lookup that need the 3035 * write lock on the map: create an anonymous map 3036 * for a copy-on-write region, or an anonymous map 3037 * for a zero-fill region. (XXXCDC: submap case 3038 * ok?) 3039 */ 3040 3041 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3042 if (UVM_ET_ISNEEDSCOPY(entry) && 3043 ((entry->max_protection & VM_PROT_WRITE) || 3044 (entry->object.uvm_obj == NULL))) { 3045 amap_copy(map, entry, M_WAITOK, TRUE, 3046 start, end); 3047 /* XXXCDC: wait OK? */ 3048 } 3049 } 3050 } 3051 UVM_MAP_CLIP_START(map, entry, start, NULL); 3052 UVM_MAP_CLIP_END(map, entry, end, NULL); 3053 entry->wired_count++; 3054 3055 /* 3056 * Check for holes 3057 */ 3058 3059 if (entry->protection == VM_PROT_NONE || 3060 (entry->end < end && 3061 (entry->next == &map->header || 3062 entry->next->start > entry->end))) { 3063 3064 /* 3065 * found one. amap creation actions do not need to 3066 * be undone, but the wired counts need to be restored. 3067 */ 3068 3069 while (entry != &map->header && entry->end > start) { 3070 entry->wired_count--; 3071 entry = entry->prev; 3072 } 3073 if ((lockflags & UVM_LK_EXIT) == 0) 3074 vm_map_unlock(map); 3075 UVMHIST_LOG(maphist,"<- done (INVALID WIRE)",0,0,0,0); 3076 return EINVAL; 3077 } 3078 entry = entry->next; 3079 } 3080 3081 /* 3082 * Pass 2. 3083 */ 3084 3085 #ifdef DIAGNOSTIC 3086 timestamp_save = map->timestamp; 3087 #endif 3088 vm_map_busy(map); 3089 vm_map_downgrade(map); 3090 3091 rv = 0; 3092 entry = start_entry; 3093 while (entry != &map->header && entry->start < end) { 3094 if (entry->wired_count == 1) { 3095 rv = uvm_fault_wire(map, entry->start, entry->end, 3096 VM_FAULT_WIREMAX, entry->max_protection); 3097 if (rv) { 3098 3099 /* 3100 * wiring failed. break out of the loop. 3101 * we'll clean up the map below, once we 3102 * have a write lock again. 3103 */ 3104 3105 break; 3106 } 3107 } 3108 entry = entry->next; 3109 } 3110 3111 if (rv) { /* failed? */ 3112 3113 /* 3114 * Get back to an exclusive (write) lock. 3115 */ 3116 3117 vm_map_upgrade(map); 3118 vm_map_unbusy(map); 3119 3120 #ifdef DIAGNOSTIC 3121 if (timestamp_save != map->timestamp) 3122 panic("uvm_map_pageable: stale map"); 3123 #endif 3124 3125 /* 3126 * first drop the wiring count on all the entries 3127 * which haven't actually been wired yet. 3128 */ 3129 3130 failed_entry = entry; 3131 while (entry != &map->header && entry->start < end) { 3132 entry->wired_count--; 3133 entry = entry->next; 3134 } 3135 3136 /* 3137 * now, unwire all the entries that were successfully 3138 * wired above. 3139 */ 3140 3141 entry = start_entry; 3142 while (entry != failed_entry) { 3143 entry->wired_count--; 3144 if (VM_MAPENT_ISWIRED(entry) == 0) 3145 uvm_map_entry_unwire(map, entry); 3146 entry = entry->next; 3147 } 3148 if ((lockflags & UVM_LK_EXIT) == 0) 3149 vm_map_unlock(map); 3150 UVMHIST_LOG(maphist, "<- done (RV=%d)", rv,0,0,0); 3151 return (rv); 3152 } 3153 3154 /* We are holding a read lock here. */ 3155 if ((lockflags & UVM_LK_EXIT) == 0) { 3156 vm_map_unbusy(map); 3157 vm_map_unlock_read(map); 3158 } else { 3159 3160 /* 3161 * Get back to an exclusive (write) lock. 3162 */ 3163 3164 vm_map_upgrade(map); 3165 vm_map_unbusy(map); 3166 } 3167 3168 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3169 return 0; 3170 } 3171 3172 /* 3173 * uvm_map_pageable_all: special case of uvm_map_pageable - affects 3174 * all mapped regions. 3175 * 3176 * => map must not be locked. 3177 * => if no flags are specified, all regions are unwired. 3178 * => XXXJRT: has some of the same problems as uvm_map_pageable() above. 3179 */ 3180 3181 int 3182 uvm_map_pageable_all(struct vm_map *map, int flags, vsize_t limit) 3183 { 3184 struct vm_map_entry *entry, *failed_entry; 3185 vsize_t size; 3186 int rv; 3187 #ifdef DIAGNOSTIC 3188 u_int timestamp_save; 3189 #endif 3190 UVMHIST_FUNC("uvm_map_pageable_all"); UVMHIST_CALLED(maphist); 3191 UVMHIST_LOG(maphist,"(map=0x%x,flags=0x%x)", map, flags, 0, 0); 3192 3193 KASSERT(map->flags & VM_MAP_PAGEABLE); 3194 3195 vm_map_lock(map); 3196 3197 /* 3198 * handle wiring and unwiring separately. 3199 */ 3200 3201 if (flags == 0) { /* unwire */ 3202 3203 /* 3204 * POSIX 1003.1b -- munlockall unlocks all regions, 3205 * regardless of how many times mlockall has been called. 3206 */ 3207 3208 for (entry = map->header.next; entry != &map->header; 3209 entry = entry->next) { 3210 if (VM_MAPENT_ISWIRED(entry)) 3211 uvm_map_entry_unwire(map, entry); 3212 } 3213 vm_map_modflags(map, 0, VM_MAP_WIREFUTURE); 3214 vm_map_unlock(map); 3215 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3216 return 0; 3217 } 3218 3219 if (flags & MCL_FUTURE) { 3220 3221 /* 3222 * must wire all future mappings; remember this. 3223 */ 3224 3225 vm_map_modflags(map, VM_MAP_WIREFUTURE, 0); 3226 } 3227 3228 if ((flags & MCL_CURRENT) == 0) { 3229 3230 /* 3231 * no more work to do! 3232 */ 3233 3234 UVMHIST_LOG(maphist,"<- done (OK no wire)",0,0,0,0); 3235 vm_map_unlock(map); 3236 return 0; 3237 } 3238 3239 /* 3240 * wire case: in three passes [XXXCDC: ugly block of code here] 3241 * 3242 * 1: holding the write lock, count all pages mapped by non-wired 3243 * entries. if this would cause us to go over our limit, we fail. 3244 * 3245 * 2: still holding the write lock, we create any anonymous maps that 3246 * need to be created. then we increment its wiring count. 3247 * 3248 * 3: we downgrade to a read lock, and call uvm_fault_wire to fault 3249 * in the pages for any newly wired area (wired_count == 1). 3250 * 3251 * downgrading to a read lock for uvm_fault_wire avoids a possible 3252 * deadlock with another thread that may have faulted on one of 3253 * the pages to be wired (it would mark the page busy, blocking 3254 * us, then in turn block on the map lock that we hold). because 3255 * of problems in the recursive lock package, we cannot upgrade 3256 * to a write lock in vm_map_lookup. thus, any actions that 3257 * require the write lock must be done beforehand. because we 3258 * keep the read lock on the map, the copy-on-write status of the 3259 * entries we modify here cannot change. 3260 */ 3261 3262 for (size = 0, entry = map->header.next; entry != &map->header; 3263 entry = entry->next) { 3264 if (entry->protection != VM_PROT_NONE && 3265 VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3266 size += entry->end - entry->start; 3267 } 3268 } 3269 3270 if (atop(size) + uvmexp.wired > uvmexp.wiredmax) { 3271 vm_map_unlock(map); 3272 return ENOMEM; 3273 } 3274 3275 if (limit != 0 && 3276 (size + ptoa(pmap_wired_count(vm_map_pmap(map))) > limit)) { 3277 vm_map_unlock(map); 3278 return ENOMEM; 3279 } 3280 3281 /* 3282 * Pass 2. 3283 */ 3284 3285 for (entry = map->header.next; entry != &map->header; 3286 entry = entry->next) { 3287 if (entry->protection == VM_PROT_NONE) 3288 continue; 3289 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3290 3291 /* 3292 * perform actions of vm_map_lookup that need the 3293 * write lock on the map: create an anonymous map 3294 * for a copy-on-write region, or an anonymous map 3295 * for a zero-fill region. (XXXCDC: submap case 3296 * ok?) 3297 */ 3298 3299 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3300 if (UVM_ET_ISNEEDSCOPY(entry) && 3301 ((entry->max_protection & VM_PROT_WRITE) || 3302 (entry->object.uvm_obj == NULL))) { 3303 amap_copy(map, entry, M_WAITOK, TRUE, 3304 entry->start, entry->end); 3305 /* XXXCDC: wait OK? */ 3306 } 3307 } 3308 } 3309 entry->wired_count++; 3310 } 3311 3312 /* 3313 * Pass 3. 3314 */ 3315 3316 #ifdef DIAGNOSTIC 3317 timestamp_save = map->timestamp; 3318 #endif 3319 vm_map_busy(map); 3320 vm_map_downgrade(map); 3321 3322 rv = 0; 3323 for (entry = map->header.next; entry != &map->header; 3324 entry = entry->next) { 3325 if (entry->wired_count == 1) { 3326 rv = uvm_fault_wire(map, entry->start, entry->end, 3327 VM_FAULT_WIREMAX, entry->max_protection); 3328 if (rv) { 3329 3330 /* 3331 * wiring failed. break out of the loop. 3332 * we'll clean up the map below, once we 3333 * have a write lock again. 3334 */ 3335 3336 break; 3337 } 3338 } 3339 } 3340 3341 if (rv) { 3342 3343 /* 3344 * Get back an exclusive (write) lock. 3345 */ 3346 3347 vm_map_upgrade(map); 3348 vm_map_unbusy(map); 3349 3350 #ifdef DIAGNOSTIC 3351 if (timestamp_save != map->timestamp) 3352 panic("uvm_map_pageable_all: stale map"); 3353 #endif 3354 3355 /* 3356 * first drop the wiring count on all the entries 3357 * which haven't actually been wired yet. 3358 * 3359 * Skip VM_PROT_NONE entries like we did above. 3360 */ 3361 3362 failed_entry = entry; 3363 for (/* nothing */; entry != &map->header; 3364 entry = entry->next) { 3365 if (entry->protection == VM_PROT_NONE) 3366 continue; 3367 entry->wired_count--; 3368 } 3369 3370 /* 3371 * now, unwire all the entries that were successfully 3372 * wired above. 3373 * 3374 * Skip VM_PROT_NONE entries like we did above. 3375 */ 3376 3377 for (entry = map->header.next; entry != failed_entry; 3378 entry = entry->next) { 3379 if (entry->protection == VM_PROT_NONE) 3380 continue; 3381 entry->wired_count--; 3382 if (VM_MAPENT_ISWIRED(entry)) 3383 uvm_map_entry_unwire(map, entry); 3384 } 3385 vm_map_unlock(map); 3386 UVMHIST_LOG(maphist,"<- done (RV=%d)", rv,0,0,0); 3387 return (rv); 3388 } 3389 3390 /* We are holding a read lock here. */ 3391 vm_map_unbusy(map); 3392 vm_map_unlock_read(map); 3393 3394 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3395 return 0; 3396 } 3397 3398 /* 3399 * uvm_map_clean: clean out a map range 3400 * 3401 * => valid flags: 3402 * if (flags & PGO_CLEANIT): dirty pages are cleaned first 3403 * if (flags & PGO_SYNCIO): dirty pages are written synchronously 3404 * if (flags & PGO_DEACTIVATE): any cached pages are deactivated after clean 3405 * if (flags & PGO_FREE): any cached pages are freed after clean 3406 * => returns an error if any part of the specified range isn't mapped 3407 * => never a need to flush amap layer since the anonymous memory has 3408 * no permanent home, but may deactivate pages there 3409 * => called from sys_msync() and sys_madvise() 3410 * => caller must not write-lock map (read OK). 3411 * => we may sleep while cleaning if SYNCIO [with map read-locked] 3412 */ 3413 3414 int 3415 uvm_map_clean(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 3416 { 3417 struct vm_map_entry *current, *entry; 3418 struct uvm_object *uobj; 3419 struct vm_amap *amap; 3420 struct vm_anon *anon; 3421 struct vm_page *pg; 3422 vaddr_t offset; 3423 vsize_t size; 3424 voff_t uoff; 3425 int error, refs; 3426 UVMHIST_FUNC("uvm_map_clean"); UVMHIST_CALLED(maphist); 3427 3428 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,flags=0x%x)", 3429 map, start, end, flags); 3430 KASSERT((flags & (PGO_FREE|PGO_DEACTIVATE)) != 3431 (PGO_FREE|PGO_DEACTIVATE)); 3432 3433 vm_map_lock_read(map); 3434 VM_MAP_RANGE_CHECK(map, start, end); 3435 if (uvm_map_lookup_entry(map, start, &entry) == FALSE) { 3436 vm_map_unlock_read(map); 3437 return EFAULT; 3438 } 3439 3440 /* 3441 * Make a first pass to check for holes and wiring problems. 3442 */ 3443 3444 for (current = entry; current->start < end; current = current->next) { 3445 if (UVM_ET_ISSUBMAP(current)) { 3446 vm_map_unlock_read(map); 3447 return EINVAL; 3448 } 3449 if ((flags & PGO_FREE) != 0 && VM_MAPENT_ISWIRED(entry)) { 3450 vm_map_unlock_read(map); 3451 return EBUSY; 3452 } 3453 if (end <= current->end) { 3454 break; 3455 } 3456 if (current->end != current->next->start) { 3457 vm_map_unlock_read(map); 3458 return EFAULT; 3459 } 3460 } 3461 3462 error = 0; 3463 for (current = entry; start < end; current = current->next) { 3464 amap = current->aref.ar_amap; /* top layer */ 3465 uobj = current->object.uvm_obj; /* bottom layer */ 3466 KASSERT(start >= current->start); 3467 3468 /* 3469 * No amap cleaning necessary if: 3470 * 3471 * (1) There's no amap. 3472 * 3473 * (2) We're not deactivating or freeing pages. 3474 */ 3475 3476 if (amap == NULL || (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) 3477 goto flush_object; 3478 3479 amap_lock(amap); 3480 offset = start - current->start; 3481 size = MIN(end, current->end) - start; 3482 for ( ; size != 0; size -= PAGE_SIZE, offset += PAGE_SIZE) { 3483 anon = amap_lookup(¤t->aref, offset); 3484 if (anon == NULL) 3485 continue; 3486 3487 simple_lock(&anon->an_lock); 3488 pg = anon->u.an_page; 3489 if (pg == NULL) { 3490 simple_unlock(&anon->an_lock); 3491 continue; 3492 } 3493 3494 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 3495 3496 /* 3497 * In these first 3 cases, we just deactivate the page. 3498 */ 3499 3500 case PGO_CLEANIT|PGO_FREE: 3501 case PGO_CLEANIT|PGO_DEACTIVATE: 3502 case PGO_DEACTIVATE: 3503 deactivate_it: 3504 /* 3505 * skip the page if it's loaned or wired, 3506 * since it shouldn't be on a paging queue 3507 * at all in these cases. 3508 */ 3509 3510 uvm_lock_pageq(); 3511 if (pg->loan_count != 0 || 3512 pg->wire_count != 0) { 3513 uvm_unlock_pageq(); 3514 simple_unlock(&anon->an_lock); 3515 continue; 3516 } 3517 KASSERT(pg->uanon == anon); 3518 pmap_clear_reference(pg); 3519 uvm_pagedeactivate(pg); 3520 uvm_unlock_pageq(); 3521 simple_unlock(&anon->an_lock); 3522 continue; 3523 3524 case PGO_FREE: 3525 3526 /* 3527 * If there are multiple references to 3528 * the amap, just deactivate the page. 3529 */ 3530 3531 if (amap_refs(amap) > 1) 3532 goto deactivate_it; 3533 3534 /* skip the page if it's wired */ 3535 if (pg->wire_count != 0) { 3536 simple_unlock(&anon->an_lock); 3537 continue; 3538 } 3539 amap_unadd(¤t->aref, offset); 3540 refs = --anon->an_ref; 3541 simple_unlock(&anon->an_lock); 3542 if (refs == 0) 3543 uvm_anfree(anon); 3544 continue; 3545 } 3546 } 3547 amap_unlock(amap); 3548 3549 flush_object: 3550 /* 3551 * flush pages if we've got a valid backing object. 3552 * note that we must always clean object pages before 3553 * freeing them since otherwise we could reveal stale 3554 * data from files. 3555 */ 3556 3557 uoff = current->offset + (start - current->start); 3558 size = MIN(end, current->end) - start; 3559 if (uobj != NULL) { 3560 simple_lock(&uobj->vmobjlock); 3561 if (uobj->pgops->pgo_put != NULL) 3562 error = (uobj->pgops->pgo_put)(uobj, uoff, 3563 uoff + size, flags | PGO_CLEANIT); 3564 else 3565 error = 0; 3566 } 3567 start += size; 3568 } 3569 vm_map_unlock_read(map); 3570 return (error); 3571 } 3572 3573 3574 /* 3575 * uvm_map_checkprot: check protection in map 3576 * 3577 * => must allow specified protection in a fully allocated region. 3578 * => map must be read or write locked by caller. 3579 */ 3580 3581 boolean_t 3582 uvm_map_checkprot(struct vm_map *map, vaddr_t start, vaddr_t end, 3583 vm_prot_t protection) 3584 { 3585 struct vm_map_entry *entry; 3586 struct vm_map_entry *tmp_entry; 3587 3588 if (!uvm_map_lookup_entry(map, start, &tmp_entry)) { 3589 return (FALSE); 3590 } 3591 entry = tmp_entry; 3592 while (start < end) { 3593 if (entry == &map->header) { 3594 return (FALSE); 3595 } 3596 3597 /* 3598 * no holes allowed 3599 */ 3600 3601 if (start < entry->start) { 3602 return (FALSE); 3603 } 3604 3605 /* 3606 * check protection associated with entry 3607 */ 3608 3609 if ((entry->protection & protection) != protection) { 3610 return (FALSE); 3611 } 3612 start = entry->end; 3613 entry = entry->next; 3614 } 3615 return (TRUE); 3616 } 3617 3618 /* 3619 * uvmspace_alloc: allocate a vmspace structure. 3620 * 3621 * - structure includes vm_map and pmap 3622 * - XXX: no locking on this structure 3623 * - refcnt set to 1, rest must be init'd by caller 3624 */ 3625 struct vmspace * 3626 uvmspace_alloc(vaddr_t min, vaddr_t max) 3627 { 3628 struct vmspace *vm; 3629 UVMHIST_FUNC("uvmspace_alloc"); UVMHIST_CALLED(maphist); 3630 3631 vm = pool_get(&uvm_vmspace_pool, PR_WAITOK); 3632 uvmspace_init(vm, NULL, min, max); 3633 UVMHIST_LOG(maphist,"<- done (vm=0x%x)", vm,0,0,0); 3634 return (vm); 3635 } 3636 3637 /* 3638 * uvmspace_init: initialize a vmspace structure. 3639 * 3640 * - XXX: no locking on this structure 3641 * - refcnt set to 1, rest must be init'd by caller 3642 */ 3643 void 3644 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t min, vaddr_t max) 3645 { 3646 UVMHIST_FUNC("uvmspace_init"); UVMHIST_CALLED(maphist); 3647 3648 memset(vm, 0, sizeof(*vm)); 3649 uvm_map_setup(&vm->vm_map, min, max, VM_MAP_PAGEABLE 3650 #ifdef __USING_TOPDOWN_VM 3651 | VM_MAP_TOPDOWN 3652 #endif 3653 ); 3654 if (pmap) 3655 pmap_reference(pmap); 3656 else 3657 pmap = pmap_create(); 3658 vm->vm_map.pmap = pmap; 3659 vm->vm_refcnt = 1; 3660 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 3661 } 3662 3663 /* 3664 * uvmspace_share: share a vmspace between two processes 3665 * 3666 * - used for vfork, threads(?) 3667 */ 3668 3669 void 3670 uvmspace_share(struct proc *p1, struct proc *p2) 3671 { 3672 struct simplelock *slock = &p1->p_vmspace->vm_map.ref_lock; 3673 3674 p2->p_vmspace = p1->p_vmspace; 3675 simple_lock(slock); 3676 p1->p_vmspace->vm_refcnt++; 3677 simple_unlock(slock); 3678 } 3679 3680 /* 3681 * uvmspace_unshare: ensure that process "p" has its own, unshared, vmspace 3682 * 3683 * - XXX: no locking on vmspace 3684 */ 3685 3686 void 3687 uvmspace_unshare(struct lwp *l) 3688 { 3689 struct proc *p = l->l_proc; 3690 struct vmspace *nvm, *ovm = p->p_vmspace; 3691 3692 if (ovm->vm_refcnt == 1) 3693 /* nothing to do: vmspace isn't shared in the first place */ 3694 return; 3695 3696 /* make a new vmspace, still holding old one */ 3697 nvm = uvmspace_fork(ovm); 3698 3699 pmap_deactivate(l); /* unbind old vmspace */ 3700 p->p_vmspace = nvm; 3701 pmap_activate(l); /* switch to new vmspace */ 3702 3703 uvmspace_free(ovm); /* drop reference to old vmspace */ 3704 } 3705 3706 /* 3707 * uvmspace_exec: the process wants to exec a new program 3708 */ 3709 3710 void 3711 uvmspace_exec(struct lwp *l, vaddr_t start, vaddr_t end) 3712 { 3713 struct proc *p = l->l_proc; 3714 struct vmspace *nvm, *ovm = p->p_vmspace; 3715 struct vm_map *map = &ovm->vm_map; 3716 3717 #ifdef __sparc__ 3718 /* XXX cgd 960926: the sparc #ifdef should be a MD hook */ 3719 kill_user_windows(l); /* before stack addresses go away */ 3720 #endif 3721 3722 /* 3723 * see if more than one process is using this vmspace... 3724 */ 3725 3726 if (ovm->vm_refcnt == 1) { 3727 3728 /* 3729 * if p is the only process using its vmspace then we can safely 3730 * recycle that vmspace for the program that is being exec'd. 3731 */ 3732 3733 #ifdef SYSVSHM 3734 /* 3735 * SYSV SHM semantics require us to kill all segments on an exec 3736 */ 3737 3738 if (ovm->vm_shm) 3739 shmexit(ovm); 3740 #endif 3741 3742 /* 3743 * POSIX 1003.1b -- "lock future mappings" is revoked 3744 * when a process execs another program image. 3745 */ 3746 3747 vm_map_modflags(map, 0, VM_MAP_WIREFUTURE); 3748 3749 /* 3750 * now unmap the old program 3751 */ 3752 3753 pmap_remove_all(map->pmap); 3754 uvm_unmap(map, vm_map_min(map), vm_map_max(map)); 3755 KASSERT(map->header.prev == &map->header); 3756 KASSERT(map->nentries == 0); 3757 3758 /* 3759 * resize the map 3760 */ 3761 3762 vm_map_setmin(map, start); 3763 vm_map_setmax(map, end); 3764 } else { 3765 3766 /* 3767 * p's vmspace is being shared, so we can't reuse it for p since 3768 * it is still being used for others. allocate a new vmspace 3769 * for p 3770 */ 3771 3772 nvm = uvmspace_alloc(start, end); 3773 3774 /* 3775 * install new vmspace and drop our ref to the old one. 3776 */ 3777 3778 pmap_deactivate(l); 3779 p->p_vmspace = nvm; 3780 pmap_activate(l); 3781 3782 uvmspace_free(ovm); 3783 } 3784 } 3785 3786 /* 3787 * uvmspace_free: free a vmspace data structure 3788 */ 3789 3790 void 3791 uvmspace_free(struct vmspace *vm) 3792 { 3793 struct vm_map_entry *dead_entries; 3794 struct vm_map *map = &vm->vm_map; 3795 int n; 3796 3797 UVMHIST_FUNC("uvmspace_free"); UVMHIST_CALLED(maphist); 3798 3799 UVMHIST_LOG(maphist,"(vm=0x%x) ref=%d", vm, vm->vm_refcnt,0,0); 3800 simple_lock(&map->ref_lock); 3801 n = --vm->vm_refcnt; 3802 simple_unlock(&map->ref_lock); 3803 if (n > 0) 3804 return; 3805 3806 /* 3807 * at this point, there should be no other references to the map. 3808 * delete all of the mappings, then destroy the pmap. 3809 */ 3810 3811 map->flags |= VM_MAP_DYING; 3812 pmap_remove_all(map->pmap); 3813 #ifdef SYSVSHM 3814 /* Get rid of any SYSV shared memory segments. */ 3815 if (vm->vm_shm != NULL) 3816 shmexit(vm); 3817 #endif 3818 if (map->nentries) { 3819 uvm_unmap_remove(map, vm_map_min(map), vm_map_max(map), 3820 &dead_entries, NULL, 0); 3821 if (dead_entries != NULL) 3822 uvm_unmap_detach(dead_entries, 0); 3823 } 3824 KASSERT(map->nentries == 0); 3825 KASSERT(map->size == 0); 3826 pmap_destroy(map->pmap); 3827 pool_put(&uvm_vmspace_pool, vm); 3828 } 3829 3830 /* 3831 * F O R K - m a i n e n t r y p o i n t 3832 */ 3833 /* 3834 * uvmspace_fork: fork a process' main map 3835 * 3836 * => create a new vmspace for child process from parent. 3837 * => parent's map must not be locked. 3838 */ 3839 3840 struct vmspace * 3841 uvmspace_fork(struct vmspace *vm1) 3842 { 3843 struct vmspace *vm2; 3844 struct vm_map *old_map = &vm1->vm_map; 3845 struct vm_map *new_map; 3846 struct vm_map_entry *old_entry; 3847 struct vm_map_entry *new_entry; 3848 UVMHIST_FUNC("uvmspace_fork"); UVMHIST_CALLED(maphist); 3849 3850 vm_map_lock(old_map); 3851 3852 vm2 = uvmspace_alloc(vm_map_min(old_map), vm_map_max(old_map)); 3853 memcpy(&vm2->vm_startcopy, &vm1->vm_startcopy, 3854 (caddr_t) (vm1 + 1) - (caddr_t) &vm1->vm_startcopy); 3855 new_map = &vm2->vm_map; /* XXX */ 3856 3857 old_entry = old_map->header.next; 3858 new_map->size = old_map->size; 3859 3860 /* 3861 * go entry-by-entry 3862 */ 3863 3864 while (old_entry != &old_map->header) { 3865 3866 /* 3867 * first, some sanity checks on the old entry 3868 */ 3869 3870 KASSERT(!UVM_ET_ISSUBMAP(old_entry)); 3871 KASSERT(UVM_ET_ISCOPYONWRITE(old_entry) || 3872 !UVM_ET_ISNEEDSCOPY(old_entry)); 3873 3874 switch (old_entry->inheritance) { 3875 case MAP_INHERIT_NONE: 3876 3877 /* 3878 * drop the mapping, modify size 3879 */ 3880 new_map->size -= old_entry->end - old_entry->start; 3881 break; 3882 3883 case MAP_INHERIT_SHARE: 3884 3885 /* 3886 * share the mapping: this means we want the old and 3887 * new entries to share amaps and backing objects. 3888 */ 3889 /* 3890 * if the old_entry needs a new amap (due to prev fork) 3891 * then we need to allocate it now so that we have 3892 * something we own to share with the new_entry. [in 3893 * other words, we need to clear needs_copy] 3894 */ 3895 3896 if (UVM_ET_ISNEEDSCOPY(old_entry)) { 3897 /* get our own amap, clears needs_copy */ 3898 amap_copy(old_map, old_entry, M_WAITOK, FALSE, 3899 0, 0); 3900 /* XXXCDC: WAITOK??? */ 3901 } 3902 3903 new_entry = uvm_mapent_alloc(new_map, 0); 3904 /* old_entry -> new_entry */ 3905 uvm_mapent_copy(old_entry, new_entry); 3906 3907 /* new pmap has nothing wired in it */ 3908 new_entry->wired_count = 0; 3909 3910 /* 3911 * gain reference to object backing the map (can't 3912 * be a submap, already checked this case). 3913 */ 3914 3915 if (new_entry->aref.ar_amap) 3916 uvm_map_reference_amap(new_entry, AMAP_SHARED); 3917 3918 if (new_entry->object.uvm_obj && 3919 new_entry->object.uvm_obj->pgops->pgo_reference) 3920 new_entry->object.uvm_obj-> 3921 pgops->pgo_reference( 3922 new_entry->object.uvm_obj); 3923 3924 /* insert entry at end of new_map's entry list */ 3925 uvm_map_entry_link(new_map, new_map->header.prev, 3926 new_entry); 3927 3928 break; 3929 3930 case MAP_INHERIT_COPY: 3931 3932 /* 3933 * copy-on-write the mapping (using mmap's 3934 * MAP_PRIVATE semantics) 3935 * 3936 * allocate new_entry, adjust reference counts. 3937 * (note that new references are read-only). 3938 */ 3939 3940 new_entry = uvm_mapent_alloc(new_map, 0); 3941 /* old_entry -> new_entry */ 3942 uvm_mapent_copy(old_entry, new_entry); 3943 3944 if (new_entry->aref.ar_amap) 3945 uvm_map_reference_amap(new_entry, 0); 3946 3947 if (new_entry->object.uvm_obj && 3948 new_entry->object.uvm_obj->pgops->pgo_reference) 3949 new_entry->object.uvm_obj->pgops->pgo_reference 3950 (new_entry->object.uvm_obj); 3951 3952 /* new pmap has nothing wired in it */ 3953 new_entry->wired_count = 0; 3954 3955 new_entry->etype |= 3956 (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY); 3957 uvm_map_entry_link(new_map, new_map->header.prev, 3958 new_entry); 3959 3960 /* 3961 * the new entry will need an amap. it will either 3962 * need to be copied from the old entry or created 3963 * from scratch (if the old entry does not have an 3964 * amap). can we defer this process until later 3965 * (by setting "needs_copy") or do we need to copy 3966 * the amap now? 3967 * 3968 * we must copy the amap now if any of the following 3969 * conditions hold: 3970 * 1. the old entry has an amap and that amap is 3971 * being shared. this means that the old (parent) 3972 * process is sharing the amap with another 3973 * process. if we do not clear needs_copy here 3974 * we will end up in a situation where both the 3975 * parent and child process are refering to the 3976 * same amap with "needs_copy" set. if the 3977 * parent write-faults, the fault routine will 3978 * clear "needs_copy" in the parent by allocating 3979 * a new amap. this is wrong because the 3980 * parent is supposed to be sharing the old amap 3981 * and the new amap will break that. 3982 * 3983 * 2. if the old entry has an amap and a non-zero 3984 * wire count then we are going to have to call 3985 * amap_cow_now to avoid page faults in the 3986 * parent process. since amap_cow_now requires 3987 * "needs_copy" to be clear we might as well 3988 * clear it here as well. 3989 * 3990 */ 3991 3992 if (old_entry->aref.ar_amap != NULL) { 3993 if ((amap_flags(old_entry->aref.ar_amap) & 3994 AMAP_SHARED) != 0 || 3995 VM_MAPENT_ISWIRED(old_entry)) { 3996 3997 amap_copy(new_map, new_entry, M_WAITOK, 3998 FALSE, 0, 0); 3999 /* XXXCDC: M_WAITOK ... ok? */ 4000 } 4001 } 4002 4003 /* 4004 * if the parent's entry is wired down, then the 4005 * parent process does not want page faults on 4006 * access to that memory. this means that we 4007 * cannot do copy-on-write because we can't write 4008 * protect the old entry. in this case we 4009 * resolve all copy-on-write faults now, using 4010 * amap_cow_now. note that we have already 4011 * allocated any needed amap (above). 4012 */ 4013 4014 if (VM_MAPENT_ISWIRED(old_entry)) { 4015 4016 /* 4017 * resolve all copy-on-write faults now 4018 * (note that there is nothing to do if 4019 * the old mapping does not have an amap). 4020 */ 4021 if (old_entry->aref.ar_amap) 4022 amap_cow_now(new_map, new_entry); 4023 4024 } else { 4025 4026 /* 4027 * setup mappings to trigger copy-on-write faults 4028 * we must write-protect the parent if it has 4029 * an amap and it is not already "needs_copy"... 4030 * if it is already "needs_copy" then the parent 4031 * has already been write-protected by a previous 4032 * fork operation. 4033 */ 4034 4035 if (old_entry->aref.ar_amap && 4036 !UVM_ET_ISNEEDSCOPY(old_entry)) { 4037 if (old_entry->max_protection & VM_PROT_WRITE) { 4038 pmap_protect(old_map->pmap, 4039 old_entry->start, 4040 old_entry->end, 4041 old_entry->protection & 4042 ~VM_PROT_WRITE); 4043 pmap_update(old_map->pmap); 4044 } 4045 old_entry->etype |= UVM_ET_NEEDSCOPY; 4046 } 4047 } 4048 break; 4049 } /* end of switch statement */ 4050 old_entry = old_entry->next; 4051 } 4052 4053 vm_map_unlock(old_map); 4054 4055 #ifdef SYSVSHM 4056 if (vm1->vm_shm) 4057 shmfork(vm1, vm2); 4058 #endif 4059 4060 #ifdef PMAP_FORK 4061 pmap_fork(vm1->vm_map.pmap, vm2->vm_map.pmap); 4062 #endif 4063 4064 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4065 return (vm2); 4066 } 4067 4068 4069 /* 4070 * in-kernel map entry allocation. 4071 */ 4072 4073 int ukh_alloc, ukh_free; 4074 int uke_alloc, uke_free; 4075 4076 struct uvm_kmapent_hdr { 4077 LIST_ENTRY(uvm_kmapent_hdr) ukh_listq; 4078 int ukh_nused; 4079 struct vm_map_entry *ukh_freelist; 4080 struct vm_map *ukh_map; 4081 struct vm_map_entry ukh_entries[0]; 4082 }; 4083 4084 #define UVM_KMAPENT_CHUNK \ 4085 ((PAGE_SIZE - sizeof(struct uvm_kmapent_hdr)) \ 4086 / sizeof(struct vm_map_entry)) 4087 4088 #define UVM_KHDR_FIND(entry) \ 4089 ((struct uvm_kmapent_hdr *)(((vaddr_t)entry) & ~PAGE_MASK)) 4090 4091 static __inline struct vm_map_entry *uvm_kmapent_get(struct uvm_kmapent_hdr *); 4092 static __inline void uvm_kmapent_put(struct uvm_kmapent_hdr *, 4093 struct vm_map_entry *); 4094 4095 static __inline struct vm_map * 4096 uvm_kmapent_map(struct vm_map_entry *entry) 4097 { 4098 const struct uvm_kmapent_hdr *ukh; 4099 4100 ukh = UVM_KHDR_FIND(entry); 4101 return ukh->ukh_map; 4102 } 4103 4104 static __inline struct vm_map_entry * 4105 uvm_kmapent_get(struct uvm_kmapent_hdr *ukh) 4106 { 4107 struct vm_map_entry *entry; 4108 4109 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4110 KASSERT(ukh->ukh_nused >= 0); 4111 4112 entry = ukh->ukh_freelist; 4113 if (entry) { 4114 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4115 == UVM_MAP_KERNEL); 4116 ukh->ukh_freelist = entry->next; 4117 ukh->ukh_nused++; 4118 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4119 } else { 4120 KASSERT(ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4121 } 4122 4123 return entry; 4124 } 4125 4126 static __inline void 4127 uvm_kmapent_put(struct uvm_kmapent_hdr *ukh, struct vm_map_entry *entry) 4128 { 4129 4130 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4131 == UVM_MAP_KERNEL); 4132 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4133 KASSERT(ukh->ukh_nused > 0); 4134 KASSERT(ukh->ukh_freelist != NULL || 4135 ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4136 KASSERT(ukh->ukh_freelist == NULL || 4137 ukh->ukh_nused < UVM_KMAPENT_CHUNK); 4138 4139 ukh->ukh_nused--; 4140 entry->next = ukh->ukh_freelist; 4141 ukh->ukh_freelist = entry; 4142 } 4143 4144 /* 4145 * uvm_kmapent_alloc: allocate a map entry for in-kernel map 4146 */ 4147 4148 static struct vm_map_entry * 4149 uvm_kmapent_alloc(struct vm_map *map, int flags) 4150 { 4151 struct vm_page *pg; 4152 struct uvm_map_args args; 4153 struct uvm_kmapent_hdr *ukh; 4154 struct vm_map_entry *entry; 4155 uvm_flag_t mapflags = UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, 4156 UVM_INH_NONE, UVM_ADV_RANDOM, flags | UVM_FLAG_NOMERGE); 4157 vaddr_t va; 4158 int error; 4159 int i; 4160 int s; 4161 4162 KDASSERT(UVM_KMAPENT_CHUNK > 2); 4163 KDASSERT(kernel_map != NULL); 4164 KASSERT(vm_map_pmap(map) == pmap_kernel()); 4165 4166 uke_alloc++; 4167 entry = NULL; 4168 again: 4169 /* 4170 * try to grab an entry from freelist. 4171 */ 4172 s = splvm(); 4173 simple_lock(&uvm.kentry_lock); 4174 ukh = LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free); 4175 if (ukh) { 4176 entry = uvm_kmapent_get(ukh); 4177 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK) 4178 LIST_REMOVE(ukh, ukh_listq); 4179 } 4180 simple_unlock(&uvm.kentry_lock); 4181 splx(s); 4182 4183 if (entry) 4184 return entry; 4185 4186 /* 4187 * there's no free entry for this vm_map. 4188 * now we need to allocate some vm_map_entry. 4189 * for simplicity, always allocate one page chunk of them at once. 4190 */ 4191 4192 pg = uvm_pagealloc(NULL, 0, NULL, 0); 4193 if (__predict_false(pg == NULL)) { 4194 if (flags & UVM_FLAG_NOWAIT) 4195 return NULL; 4196 uvm_wait("kme_alloc"); 4197 goto again; 4198 } 4199 4200 error = uvm_map_prepare(map, 0, PAGE_SIZE, NULL, 0, 0, mapflags, &args); 4201 if (error) { 4202 uvm_pagefree(pg); 4203 return NULL; 4204 } 4205 4206 va = args.uma_start; 4207 4208 pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), VM_PROT_READ|VM_PROT_WRITE); 4209 pmap_update(vm_map_pmap(map)); 4210 4211 ukh = (void *)va; 4212 4213 /* 4214 * use the first entry for ukh itsself. 4215 */ 4216 4217 entry = &ukh->ukh_entries[0]; 4218 entry->flags = UVM_MAP_KERNEL | UVM_MAP_KMAPENT; 4219 error = uvm_map_enter(map, &args, entry); 4220 KASSERT(error == 0); 4221 4222 ukh->ukh_nused = UVM_KMAPENT_CHUNK; 4223 ukh->ukh_map = map; 4224 ukh->ukh_freelist = NULL; 4225 for (i = UVM_KMAPENT_CHUNK - 1; i >= 2; i--) { 4226 struct vm_map_entry *entry = &ukh->ukh_entries[i]; 4227 4228 entry->flags = UVM_MAP_KERNEL; 4229 uvm_kmapent_put(ukh, entry); 4230 } 4231 KASSERT(ukh->ukh_nused == 2); 4232 4233 s = splvm(); 4234 simple_lock(&uvm.kentry_lock); 4235 LIST_INSERT_HEAD(&vm_map_to_kernel(map)->vmk_kentry_free, 4236 ukh, ukh_listq); 4237 simple_unlock(&uvm.kentry_lock); 4238 splx(s); 4239 4240 /* 4241 * return second entry. 4242 */ 4243 4244 entry = &ukh->ukh_entries[1]; 4245 entry->flags = UVM_MAP_KERNEL; 4246 ukh_alloc++; 4247 return entry; 4248 } 4249 4250 /* 4251 * uvm_mapent_free: free map entry for in-kernel map 4252 */ 4253 4254 static void 4255 uvm_kmapent_free(struct vm_map_entry *entry) 4256 { 4257 struct uvm_kmapent_hdr *ukh; 4258 struct vm_page *pg; 4259 struct vm_map *map; 4260 struct pmap *pmap; 4261 vaddr_t va; 4262 paddr_t pa; 4263 struct vm_map_entry *deadentry; 4264 int s; 4265 4266 uke_free++; 4267 ukh = UVM_KHDR_FIND(entry); 4268 map = ukh->ukh_map; 4269 4270 s = splvm(); 4271 simple_lock(&uvm.kentry_lock); 4272 uvm_kmapent_put(ukh, entry); 4273 if (ukh->ukh_nused > 1) { 4274 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK - 1) 4275 LIST_INSERT_HEAD( 4276 &vm_map_to_kernel(map)->vmk_kentry_free, 4277 ukh, ukh_listq); 4278 simple_unlock(&uvm.kentry_lock); 4279 splx(s); 4280 return; 4281 } 4282 4283 /* 4284 * now we can free this ukh. 4285 * 4286 * however, keep an empty ukh to avoid ping-pong. 4287 */ 4288 4289 if (LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free) == ukh && 4290 LIST_NEXT(ukh, ukh_listq) == NULL) { 4291 simple_unlock(&uvm.kentry_lock); 4292 splx(s); 4293 return; 4294 } 4295 LIST_REMOVE(ukh, ukh_listq); 4296 simple_unlock(&uvm.kentry_lock); 4297 splx(s); 4298 4299 KASSERT(ukh->ukh_nused == 1); 4300 4301 /* 4302 * remove map entry for ukh itsself. 4303 */ 4304 4305 va = (vaddr_t)ukh; 4306 KASSERT((va & PAGE_MASK) == 0); 4307 uvm_unmap_remove(map, va, va + PAGE_SIZE, &deadentry, NULL, 0); 4308 KASSERT(deadentry->flags & UVM_MAP_KERNEL); 4309 KASSERT(deadentry->flags & UVM_MAP_KMAPENT); 4310 KASSERT(deadentry->next == NULL); 4311 KASSERT(deadentry == &ukh->ukh_entries[0]); 4312 4313 /* 4314 * unmap the page from pmap and free it. 4315 */ 4316 4317 pmap = vm_map_pmap(map); 4318 KASSERT(pmap == pmap_kernel()); 4319 if (!pmap_extract(pmap, va, &pa)) 4320 panic("%s: no mapping", __func__); 4321 pmap_kremove(va, PAGE_SIZE); 4322 pg = PHYS_TO_VM_PAGE(pa); 4323 uvm_pagefree(pg); 4324 ukh_free++; 4325 } 4326 4327 /* 4328 * map entry reservation 4329 */ 4330 4331 /* 4332 * uvm_mapent_reserve: reserve map entries for clipping before locking map. 4333 * 4334 * => needed when unmapping entries allocated without UVM_FLAG_QUANTUM. 4335 * => caller shouldn't hold map locked. 4336 */ 4337 int 4338 uvm_mapent_reserve(struct vm_map *map, struct uvm_mapent_reservation *umr, 4339 int nentries, int flags) 4340 { 4341 4342 umr->umr_nentries = 0; 4343 4344 if ((flags & UVM_FLAG_QUANTUM) != 0) 4345 return 0; 4346 4347 if (!VM_MAP_USE_KMAPENT(map)) 4348 return 0; 4349 4350 while (nentries--) { 4351 struct vm_map_entry *ent; 4352 ent = uvm_kmapent_alloc(map, flags); 4353 if (!ent) { 4354 uvm_mapent_unreserve(map, umr); 4355 return ENOMEM; 4356 } 4357 UMR_PUTENTRY(umr, ent); 4358 } 4359 4360 return 0; 4361 } 4362 4363 /* 4364 * uvm_mapent_unreserve: 4365 * 4366 * => caller shouldn't hold map locked. 4367 * => never fail or sleep. 4368 */ 4369 void 4370 uvm_mapent_unreserve(struct vm_map *map, struct uvm_mapent_reservation *umr) 4371 { 4372 4373 while (!UMR_EMPTY(umr)) 4374 uvm_kmapent_free(UMR_GETENTRY(umr)); 4375 } 4376 4377 #if defined(DDB) 4378 4379 /* 4380 * DDB hooks 4381 */ 4382 4383 /* 4384 * uvm_map_printit: actually prints the map 4385 */ 4386 4387 void 4388 uvm_map_printit(struct vm_map *map, boolean_t full, 4389 void (*pr)(const char *, ...)) 4390 { 4391 struct vm_map_entry *entry; 4392 4393 (*pr)("MAP %p: [0x%lx->0x%lx]\n", map, vm_map_min(map), 4394 vm_map_max(map)); 4395 (*pr)("\t#ent=%d, sz=%d, ref=%d, version=%d, flags=0x%x\n", 4396 map->nentries, map->size, map->ref_count, map->timestamp, 4397 map->flags); 4398 (*pr)("\tpmap=%p(resident=%ld, wired=%ld)\n", map->pmap, 4399 pmap_resident_count(map->pmap), pmap_wired_count(map->pmap)); 4400 if (!full) 4401 return; 4402 for (entry = map->header.next; entry != &map->header; 4403 entry = entry->next) { 4404 (*pr)(" - %p: 0x%lx->0x%lx: obj=%p/0x%llx, amap=%p/%d\n", 4405 entry, entry->start, entry->end, entry->object.uvm_obj, 4406 (long long)entry->offset, entry->aref.ar_amap, 4407 entry->aref.ar_pageoff); 4408 (*pr)( 4409 "\tsubmap=%c, cow=%c, nc=%c, prot(max)=%d/%d, inh=%d, " 4410 "wc=%d, adv=%d\n", 4411 (entry->etype & UVM_ET_SUBMAP) ? 'T' : 'F', 4412 (entry->etype & UVM_ET_COPYONWRITE) ? 'T' : 'F', 4413 (entry->etype & UVM_ET_NEEDSCOPY) ? 'T' : 'F', 4414 entry->protection, entry->max_protection, 4415 entry->inheritance, entry->wired_count, entry->advice); 4416 } 4417 } 4418 4419 /* 4420 * uvm_object_printit: actually prints the object 4421 */ 4422 4423 void 4424 uvm_object_printit(struct uvm_object *uobj, boolean_t full, 4425 void (*pr)(const char *, ...)) 4426 { 4427 struct vm_page *pg; 4428 int cnt = 0; 4429 4430 (*pr)("OBJECT %p: locked=%d, pgops=%p, npages=%d, ", 4431 uobj, uobj->vmobjlock.lock_data, uobj->pgops, uobj->uo_npages); 4432 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 4433 (*pr)("refs=<SYSTEM>\n"); 4434 else 4435 (*pr)("refs=%d\n", uobj->uo_refs); 4436 4437 if (!full) { 4438 return; 4439 } 4440 (*pr)(" PAGES <pg,offset>:\n "); 4441 TAILQ_FOREACH(pg, &uobj->memq, listq) { 4442 cnt++; 4443 (*pr)("<%p,0x%llx> ", pg, (long long)pg->offset); 4444 if ((cnt % 3) == 0) { 4445 (*pr)("\n "); 4446 } 4447 } 4448 if ((cnt % 3) != 0) { 4449 (*pr)("\n"); 4450 } 4451 } 4452 4453 /* 4454 * uvm_page_printit: actually print the page 4455 */ 4456 4457 static const char page_flagbits[] = 4458 "\20\1BUSY\2WANTED\3TABLED\4CLEAN\5PAGEOUT\6RELEASED\7FAKE\10RDONLY" 4459 "\11ZERO\15PAGER1"; 4460 static const char page_pqflagbits[] = 4461 "\20\1FREE\2INACTIVE\3ACTIVE\5ANON\6AOBJ"; 4462 4463 void 4464 uvm_page_printit(struct vm_page *pg, boolean_t full, 4465 void (*pr)(const char *, ...)) 4466 { 4467 struct vm_page *tpg; 4468 struct uvm_object *uobj; 4469 struct pglist *pgl; 4470 char pgbuf[128]; 4471 char pqbuf[128]; 4472 4473 (*pr)("PAGE %p:\n", pg); 4474 bitmask_snprintf(pg->flags, page_flagbits, pgbuf, sizeof(pgbuf)); 4475 bitmask_snprintf(pg->pqflags, page_pqflagbits, pqbuf, sizeof(pqbuf)); 4476 (*pr)(" flags=%s, pqflags=%s, wire_count=%d, pa=0x%lx\n", 4477 pgbuf, pqbuf, pg->wire_count, (long)VM_PAGE_TO_PHYS(pg)); 4478 (*pr)(" uobject=%p, uanon=%p, offset=0x%llx loan_count=%d\n", 4479 pg->uobject, pg->uanon, (long long)pg->offset, pg->loan_count); 4480 #if defined(UVM_PAGE_TRKOWN) 4481 if (pg->flags & PG_BUSY) 4482 (*pr)(" owning process = %d, tag=%s\n", 4483 pg->owner, pg->owner_tag); 4484 else 4485 (*pr)(" page not busy, no owner\n"); 4486 #else 4487 (*pr)(" [page ownership tracking disabled]\n"); 4488 #endif 4489 4490 if (!full) 4491 return; 4492 4493 /* cross-verify object/anon */ 4494 if ((pg->pqflags & PQ_FREE) == 0) { 4495 if (pg->pqflags & PQ_ANON) { 4496 if (pg->uanon == NULL || pg->uanon->u.an_page != pg) 4497 (*pr)(" >>> ANON DOES NOT POINT HERE <<< (%p)\n", 4498 (pg->uanon) ? pg->uanon->u.an_page : NULL); 4499 else 4500 (*pr)(" anon backpointer is OK\n"); 4501 } else { 4502 uobj = pg->uobject; 4503 if (uobj) { 4504 (*pr)(" checking object list\n"); 4505 TAILQ_FOREACH(tpg, &uobj->memq, listq) { 4506 if (tpg == pg) { 4507 break; 4508 } 4509 } 4510 if (tpg) 4511 (*pr)(" page found on object list\n"); 4512 else 4513 (*pr)(" >>> PAGE NOT FOUND ON OBJECT LIST! <<<\n"); 4514 } 4515 } 4516 } 4517 4518 /* cross-verify page queue */ 4519 if (pg->pqflags & PQ_FREE) { 4520 int fl = uvm_page_lookup_freelist(pg); 4521 int color = VM_PGCOLOR_BUCKET(pg); 4522 pgl = &uvm.page_free[fl].pgfl_buckets[color].pgfl_queues[ 4523 ((pg)->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN]; 4524 } else if (pg->pqflags & PQ_INACTIVE) { 4525 pgl = &uvm.page_inactive; 4526 } else if (pg->pqflags & PQ_ACTIVE) { 4527 pgl = &uvm.page_active; 4528 } else { 4529 pgl = NULL; 4530 } 4531 4532 if (pgl) { 4533 (*pr)(" checking pageq list\n"); 4534 TAILQ_FOREACH(tpg, pgl, pageq) { 4535 if (tpg == pg) { 4536 break; 4537 } 4538 } 4539 if (tpg) 4540 (*pr)(" page found on pageq list\n"); 4541 else 4542 (*pr)(" >>> PAGE NOT FOUND ON PAGEQ LIST! <<<\n"); 4543 } 4544 } 4545 #endif 4546