1 /* 2 * mm/mmap.c 3 * 4 * Written by obz. 5 * 6 * Address space accounting code <alan@redhat.com> 7 */ 8 9 #include <linux/slab.h> 10 #include <linux/shm.h> 11 #include <linux/mman.h> 12 #include <linux/pagemap.h> 13 #include <linux/swap.h> 14 #include <linux/syscalls.h> 15 #include <linux/init.h> 16 #include <linux/file.h> 17 #include <linux/fs.h> 18 #include <linux/personality.h> 19 #include <linux/security.h> 20 #include <linux/hugetlb.h> 21 #include <linux/profile.h> 22 #include <linux/module.h> 23 #include <linux/mount.h> 24 #include <linux/mempolicy.h> 25 #include <linux/rmap.h> 26 27 #include <asm/uaccess.h> 28 #include <asm/cacheflush.h> 29 #include <asm/tlb.h> 30 31 /* 32 * WARNING: the debugging will use recursive algorithms so never enable this 33 * unless you know what you are doing. 34 */ 35 #undef DEBUG_MM_RB 36 37 /* description of effects of mapping type and prot in current implementation. 38 * this is due to the limited x86 page protection hardware. The expected 39 * behavior is in parens: 40 * 41 * map_type prot 42 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 43 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 44 * w: (no) no w: (no) no w: (yes) yes w: (no) no 45 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 46 * 47 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 48 * w: (no) no w: (no) no w: (copy) copy w: (no) no 49 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 50 * 51 */ 52 pgprot_t protection_map[16] = { 53 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 54 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 55 }; 56 57 int sysctl_overcommit_memory = 0; /* default is heuristic overcommit */ 58 int sysctl_overcommit_ratio = 50; /* default is 50% */ 59 int sysctl_max_map_count = DEFAULT_MAX_MAP_COUNT; 60 atomic_t vm_committed_space = ATOMIC_INIT(0); 61 62 EXPORT_SYMBOL(sysctl_overcommit_memory); 63 EXPORT_SYMBOL(sysctl_overcommit_ratio); 64 EXPORT_SYMBOL(sysctl_max_map_count); 65 EXPORT_SYMBOL(vm_committed_space); 66 67 /* 68 * Requires inode->i_mapping->i_mmap_lock 69 */ 70 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 71 struct file *file, struct address_space *mapping) 72 { 73 if (vma->vm_flags & VM_DENYWRITE) 74 atomic_inc(&file->f_dentry->d_inode->i_writecount); 75 if (vma->vm_flags & VM_SHARED) 76 mapping->i_mmap_writable--; 77 78 flush_dcache_mmap_lock(mapping); 79 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 80 list_del_init(&vma->shared.vm_set.list); 81 else 82 vma_prio_tree_remove(vma, &mapping->i_mmap); 83 flush_dcache_mmap_unlock(mapping); 84 } 85 86 /* 87 * Remove one vm structure and free it. 88 */ 89 static void remove_vm_struct(struct vm_area_struct *vma) 90 { 91 struct file *file = vma->vm_file; 92 93 if (file) { 94 struct address_space *mapping = file->f_mapping; 95 spin_lock(&mapping->i_mmap_lock); 96 __remove_shared_vm_struct(vma, file, mapping); 97 spin_unlock(&mapping->i_mmap_lock); 98 } 99 if (vma->vm_ops && vma->vm_ops->close) 100 vma->vm_ops->close(vma); 101 if (file) 102 fput(file); 103 anon_vma_unlink(vma); 104 mpol_free(vma_policy(vma)); 105 kmem_cache_free(vm_area_cachep, vma); 106 } 107 108 /* 109 * sys_brk() for the most part doesn't need the global kernel 110 * lock, except when an application is doing something nasty 111 * like trying to un-brk an area that has already been mapped 112 * to a regular file. in this case, the unmapping will need 113 * to invoke file system routines that need the global lock. 114 */ 115 asmlinkage unsigned long sys_brk(unsigned long brk) 116 { 117 unsigned long rlim, retval; 118 unsigned long newbrk, oldbrk; 119 struct mm_struct *mm = current->mm; 120 121 down_write(&mm->mmap_sem); 122 123 if (brk < mm->end_code) 124 goto out; 125 newbrk = PAGE_ALIGN(brk); 126 oldbrk = PAGE_ALIGN(mm->brk); 127 if (oldbrk == newbrk) 128 goto set_brk; 129 130 /* Always allow shrinking brk. */ 131 if (brk <= mm->brk) { 132 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 133 goto set_brk; 134 goto out; 135 } 136 137 /* Check against rlimit.. */ 138 rlim = current->rlim[RLIMIT_DATA].rlim_cur; 139 if (rlim < RLIM_INFINITY && brk - mm->start_data > rlim) 140 goto out; 141 142 /* Check against existing mmap mappings. */ 143 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 144 goto out; 145 146 /* Ok, looks good - let it rip. */ 147 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 148 goto out; 149 set_brk: 150 mm->brk = brk; 151 out: 152 retval = mm->brk; 153 up_write(&mm->mmap_sem); 154 return retval; 155 } 156 157 #ifdef DEBUG_MM_RB 158 static int browse_rb(struct rb_root *root) 159 { 160 int i = 0, j; 161 struct rb_node *nd, *pn = NULL; 162 unsigned long prev = 0, pend = 0; 163 164 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 165 struct vm_area_struct *vma; 166 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 167 if (vma->vm_start < prev) 168 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1; 169 if (vma->vm_start < pend) 170 printk("vm_start %lx pend %lx\n", vma->vm_start, pend); 171 if (vma->vm_start > vma->vm_end) 172 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start); 173 i++; 174 pn = nd; 175 } 176 j = 0; 177 for (nd = pn; nd; nd = rb_prev(nd)) { 178 j++; 179 } 180 if (i != j) 181 printk("backwards %d, forwards %d\n", j, i), i = 0; 182 return i; 183 } 184 185 void validate_mm(struct mm_struct *mm) 186 { 187 int bug = 0; 188 int i = 0; 189 struct vm_area_struct *tmp = mm->mmap; 190 while (tmp) { 191 tmp = tmp->vm_next; 192 i++; 193 } 194 if (i != mm->map_count) 195 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1; 196 i = browse_rb(&mm->mm_rb); 197 if (i != mm->map_count) 198 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1; 199 if (bug) 200 BUG(); 201 } 202 #else 203 #define validate_mm(mm) do { } while (0) 204 #endif 205 206 static struct vm_area_struct * 207 find_vma_prepare(struct mm_struct *mm, unsigned long addr, 208 struct vm_area_struct **pprev, struct rb_node ***rb_link, 209 struct rb_node ** rb_parent) 210 { 211 struct vm_area_struct * vma; 212 struct rb_node ** __rb_link, * __rb_parent, * rb_prev; 213 214 __rb_link = &mm->mm_rb.rb_node; 215 rb_prev = __rb_parent = NULL; 216 vma = NULL; 217 218 while (*__rb_link) { 219 struct vm_area_struct *vma_tmp; 220 221 __rb_parent = *__rb_link; 222 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 223 224 if (vma_tmp->vm_end > addr) { 225 vma = vma_tmp; 226 if (vma_tmp->vm_start <= addr) 227 return vma; 228 __rb_link = &__rb_parent->rb_left; 229 } else { 230 rb_prev = __rb_parent; 231 __rb_link = &__rb_parent->rb_right; 232 } 233 } 234 235 *pprev = NULL; 236 if (rb_prev) 237 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 238 *rb_link = __rb_link; 239 *rb_parent = __rb_parent; 240 return vma; 241 } 242 243 static inline void 244 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma, 245 struct vm_area_struct *prev, struct rb_node *rb_parent) 246 { 247 if (prev) { 248 vma->vm_next = prev->vm_next; 249 prev->vm_next = vma; 250 } else { 251 mm->mmap = vma; 252 if (rb_parent) 253 vma->vm_next = rb_entry(rb_parent, 254 struct vm_area_struct, vm_rb); 255 else 256 vma->vm_next = NULL; 257 } 258 } 259 260 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 261 struct rb_node **rb_link, struct rb_node *rb_parent) 262 { 263 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 264 rb_insert_color(&vma->vm_rb, &mm->mm_rb); 265 } 266 267 static inline void __vma_link_file(struct vm_area_struct *vma) 268 { 269 struct file * file; 270 271 file = vma->vm_file; 272 if (file) { 273 struct address_space *mapping = file->f_mapping; 274 275 if (vma->vm_flags & VM_DENYWRITE) 276 atomic_dec(&file->f_dentry->d_inode->i_writecount); 277 if (vma->vm_flags & VM_SHARED) 278 mapping->i_mmap_writable++; 279 280 flush_dcache_mmap_lock(mapping); 281 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 282 list_add_tail(&vma->shared.vm_set.list, 283 &mapping->i_mmap_nonlinear); 284 else 285 vma_prio_tree_insert(vma, &mapping->i_mmap); 286 flush_dcache_mmap_unlock(mapping); 287 } 288 } 289 290 static void 291 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 292 struct vm_area_struct *prev, struct rb_node **rb_link, 293 struct rb_node *rb_parent) 294 { 295 __vma_link_list(mm, vma, prev, rb_parent); 296 __vma_link_rb(mm, vma, rb_link, rb_parent); 297 __anon_vma_link(vma); 298 } 299 300 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 301 struct vm_area_struct *prev, struct rb_node **rb_link, 302 struct rb_node *rb_parent) 303 { 304 struct address_space *mapping = NULL; 305 306 if (vma->vm_file) 307 mapping = vma->vm_file->f_mapping; 308 309 if (mapping) 310 spin_lock(&mapping->i_mmap_lock); 311 anon_vma_lock(vma); 312 313 __vma_link(mm, vma, prev, rb_link, rb_parent); 314 __vma_link_file(vma); 315 316 anon_vma_unlock(vma); 317 if (mapping) 318 spin_unlock(&mapping->i_mmap_lock); 319 320 mm->map_count++; 321 validate_mm(mm); 322 } 323 324 /* 325 * Helper for vma_adjust in the split_vma insert case: 326 * insert vm structure into list and rbtree and anon_vma, 327 * but it has already been inserted into prio_tree earlier. 328 */ 329 static void 330 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma) 331 { 332 struct vm_area_struct * __vma, * prev; 333 struct rb_node ** rb_link, * rb_parent; 334 335 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent); 336 if (__vma && __vma->vm_start < vma->vm_end) 337 BUG(); 338 __vma_link(mm, vma, prev, rb_link, rb_parent); 339 mm->map_count++; 340 } 341 342 static inline void 343 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 344 struct vm_area_struct *prev) 345 { 346 prev->vm_next = vma->vm_next; 347 rb_erase(&vma->vm_rb, &mm->mm_rb); 348 if (mm->mmap_cache == vma) 349 mm->mmap_cache = prev; 350 } 351 352 /* 353 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 354 * is already present in an i_mmap tree without adjusting the tree. 355 * The following helper function should be used when such adjustments 356 * are necessary. The "insert" vma (if any) is to be inserted 357 * before we drop the necessary locks. 358 */ 359 void vma_adjust(struct vm_area_struct *vma, unsigned long start, 360 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 361 { 362 struct mm_struct *mm = vma->vm_mm; 363 struct vm_area_struct *next = vma->vm_next; 364 struct vm_area_struct *importer = NULL; 365 struct address_space *mapping = NULL; 366 struct prio_tree_root *root = NULL; 367 struct file *file = vma->vm_file; 368 struct anon_vma *anon_vma = NULL; 369 long adjust_next = 0; 370 int remove_next = 0; 371 372 if (next && !insert) { 373 if (end >= next->vm_end) { 374 /* 375 * vma expands, overlapping all the next, and 376 * perhaps the one after too (mprotect case 6). 377 */ 378 again: remove_next = 1 + (end > next->vm_end); 379 end = next->vm_end; 380 anon_vma = next->anon_vma; 381 } else if (end > next->vm_start) { 382 /* 383 * vma expands, overlapping part of the next: 384 * mprotect case 5 shifting the boundary up. 385 */ 386 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 387 anon_vma = next->anon_vma; 388 importer = vma; 389 } else if (end < vma->vm_end) { 390 /* 391 * vma shrinks, and !insert tells it's not 392 * split_vma inserting another: so it must be 393 * mprotect case 4 shifting the boundary down. 394 */ 395 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT); 396 anon_vma = next->anon_vma; 397 importer = next; 398 } 399 } 400 401 if (file) { 402 mapping = file->f_mapping; 403 if (!(vma->vm_flags & VM_NONLINEAR)) 404 root = &mapping->i_mmap; 405 spin_lock(&mapping->i_mmap_lock); 406 if (insert) { 407 /* 408 * Put into prio_tree now, so instantiated pages 409 * are visible to arm/parisc __flush_dcache_page 410 * throughout; but we cannot insert into address 411 * space until vma start or end is updated. 412 */ 413 __vma_link_file(insert); 414 } 415 } 416 417 /* 418 * When changing only vma->vm_end, we don't really need 419 * anon_vma lock: but is that case worth optimizing out? 420 */ 421 if (vma->anon_vma) 422 anon_vma = vma->anon_vma; 423 if (anon_vma) { 424 spin_lock(&anon_vma->lock); 425 /* 426 * Easily overlooked: when mprotect shifts the boundary, 427 * make sure the expanding vma has anon_vma set if the 428 * shrinking vma had, to cover any anon pages imported. 429 */ 430 if (importer && !importer->anon_vma) { 431 importer->anon_vma = anon_vma; 432 __anon_vma_link(importer); 433 } 434 } 435 436 if (root) { 437 flush_dcache_mmap_lock(mapping); 438 vma_prio_tree_remove(vma, root); 439 if (adjust_next) 440 vma_prio_tree_remove(next, root); 441 } 442 443 vma->vm_start = start; 444 vma->vm_end = end; 445 vma->vm_pgoff = pgoff; 446 if (adjust_next) { 447 next->vm_start += adjust_next << PAGE_SHIFT; 448 next->vm_pgoff += adjust_next; 449 } 450 451 if (root) { 452 if (adjust_next) { 453 vma_prio_tree_init(next); 454 vma_prio_tree_insert(next, root); 455 } 456 vma_prio_tree_init(vma); 457 vma_prio_tree_insert(vma, root); 458 flush_dcache_mmap_unlock(mapping); 459 } 460 461 if (remove_next) { 462 /* 463 * vma_merge has merged next into vma, and needs 464 * us to remove next before dropping the locks. 465 */ 466 __vma_unlink(mm, next, vma); 467 if (file) 468 __remove_shared_vm_struct(next, file, mapping); 469 if (next->anon_vma) 470 __anon_vma_merge(vma, next); 471 } else if (insert) { 472 /* 473 * split_vma has split insert from vma, and needs 474 * us to insert it before dropping the locks 475 * (it may either follow vma or precede it). 476 */ 477 __insert_vm_struct(mm, insert); 478 } 479 480 if (anon_vma) 481 spin_unlock(&anon_vma->lock); 482 if (mapping) 483 spin_unlock(&mapping->i_mmap_lock); 484 485 if (remove_next) { 486 if (file) 487 fput(file); 488 mm->map_count--; 489 mpol_free(vma_policy(next)); 490 kmem_cache_free(vm_area_cachep, next); 491 /* 492 * In mprotect's case 6 (see comments on vma_merge), 493 * we must remove another next too. It would clutter 494 * up the code too much to do both in one go. 495 */ 496 if (remove_next == 2) { 497 next = vma->vm_next; 498 goto again; 499 } 500 } 501 502 validate_mm(mm); 503 } 504 505 /* 506 * If the vma has a ->close operation then the driver probably needs to release 507 * per-vma resources, so we don't attempt to merge those. 508 */ 509 #define VM_SPECIAL (VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_RESERVED) 510 511 static inline int is_mergeable_vma(struct vm_area_struct *vma, 512 struct file *file, unsigned long vm_flags) 513 { 514 if (vma->vm_flags != vm_flags) 515 return 0; 516 if (vma->vm_file != file) 517 return 0; 518 if (vma->vm_ops && vma->vm_ops->close) 519 return 0; 520 return 1; 521 } 522 523 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 524 struct anon_vma *anon_vma2) 525 { 526 return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2); 527 } 528 529 /* 530 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 531 * in front of (at a lower virtual address and file offset than) the vma. 532 * 533 * We cannot merge two vmas if they have differently assigned (non-NULL) 534 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 535 * 536 * We don't check here for the merged mmap wrapping around the end of pagecache 537 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 538 * wrap, nor mmaps which cover the final page at index -1UL. 539 */ 540 static int 541 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 542 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 543 { 544 if (is_mergeable_vma(vma, file, vm_flags) && 545 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) { 546 if (vma->vm_pgoff == vm_pgoff) 547 return 1; 548 } 549 return 0; 550 } 551 552 /* 553 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 554 * beyond (at a higher virtual address and file offset than) the vma. 555 * 556 * We cannot merge two vmas if they have differently assigned (non-NULL) 557 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 558 */ 559 static int 560 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 561 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 562 { 563 if (is_mergeable_vma(vma, file, vm_flags) && 564 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) { 565 pgoff_t vm_pglen; 566 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; 567 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 568 return 1; 569 } 570 return 0; 571 } 572 573 /* 574 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 575 * whether that can be merged with its predecessor or its successor. 576 * Or both (it neatly fills a hole). 577 * 578 * In most cases - when called for mmap, brk or mremap - [addr,end) is 579 * certain not to be mapped by the time vma_merge is called; but when 580 * called for mprotect, it is certain to be already mapped (either at 581 * an offset within prev, or at the start of next), and the flags of 582 * this area are about to be changed to vm_flags - and the no-change 583 * case has already been eliminated. 584 * 585 * The following mprotect cases have to be considered, where AAAA is 586 * the area passed down from mprotect_fixup, never extending beyond one 587 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 588 * 589 * AAAA AAAA AAAA AAAA 590 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 591 * cannot merge might become might become might become 592 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 593 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 594 * mremap move: PPPPNNNNNNNN 8 595 * AAAA 596 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 597 * might become case 1 below case 2 below case 3 below 598 * 599 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 600 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 601 */ 602 struct vm_area_struct *vma_merge(struct mm_struct *mm, 603 struct vm_area_struct *prev, unsigned long addr, 604 unsigned long end, unsigned long vm_flags, 605 struct anon_vma *anon_vma, struct file *file, 606 pgoff_t pgoff, struct mempolicy *policy) 607 { 608 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 609 struct vm_area_struct *area, *next; 610 611 /* 612 * We later require that vma->vm_flags == vm_flags, 613 * so this tests vma->vm_flags & VM_SPECIAL, too. 614 */ 615 if (vm_flags & VM_SPECIAL) 616 return NULL; 617 618 if (prev) 619 next = prev->vm_next; 620 else 621 next = mm->mmap; 622 area = next; 623 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 624 next = next->vm_next; 625 626 /* 627 * Can it merge with the predecessor? 628 */ 629 if (prev && prev->vm_end == addr && 630 mpol_equal(vma_policy(prev), policy) && 631 can_vma_merge_after(prev, vm_flags, 632 anon_vma, file, pgoff)) { 633 /* 634 * OK, it can. Can we now merge in the successor as well? 635 */ 636 if (next && end == next->vm_start && 637 mpol_equal(policy, vma_policy(next)) && 638 can_vma_merge_before(next, vm_flags, 639 anon_vma, file, pgoff+pglen) && 640 is_mergeable_anon_vma(prev->anon_vma, 641 next->anon_vma)) { 642 /* cases 1, 6 */ 643 vma_adjust(prev, prev->vm_start, 644 next->vm_end, prev->vm_pgoff, NULL); 645 } else /* cases 2, 5, 7 */ 646 vma_adjust(prev, prev->vm_start, 647 end, prev->vm_pgoff, NULL); 648 return prev; 649 } 650 651 /* 652 * Can this new request be merged in front of next? 653 */ 654 if (next && end == next->vm_start && 655 mpol_equal(policy, vma_policy(next)) && 656 can_vma_merge_before(next, vm_flags, 657 anon_vma, file, pgoff+pglen)) { 658 if (prev && addr < prev->vm_end) /* case 4 */ 659 vma_adjust(prev, prev->vm_start, 660 addr, prev->vm_pgoff, NULL); 661 else /* cases 3, 8 */ 662 vma_adjust(area, addr, next->vm_end, 663 next->vm_pgoff - pglen, NULL); 664 return area; 665 } 666 667 return NULL; 668 } 669 670 /* 671 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 672 * neighbouring vmas for a suitable anon_vma, before it goes off 673 * to allocate a new anon_vma. It checks because a repetitive 674 * sequence of mprotects and faults may otherwise lead to distinct 675 * anon_vmas being allocated, preventing vma merge in subsequent 676 * mprotect. 677 */ 678 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 679 { 680 struct vm_area_struct *near; 681 unsigned long vm_flags; 682 683 near = vma->vm_next; 684 if (!near) 685 goto try_prev; 686 687 /* 688 * Since only mprotect tries to remerge vmas, match flags 689 * which might be mprotected into each other later on. 690 * Neither mlock nor madvise tries to remerge at present, 691 * so leave their flags as obstructing a merge. 692 */ 693 vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC); 694 vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC); 695 696 if (near->anon_vma && vma->vm_end == near->vm_start && 697 mpol_equal(vma_policy(vma), vma_policy(near)) && 698 can_vma_merge_before(near, vm_flags, 699 NULL, vma->vm_file, vma->vm_pgoff + 700 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT))) 701 return near->anon_vma; 702 try_prev: 703 /* 704 * It is potentially slow to have to call find_vma_prev here. 705 * But it's only on the first write fault on the vma, not 706 * every time, and we could devise a way to avoid it later 707 * (e.g. stash info in next's anon_vma_node when assigning 708 * an anon_vma, or when trying vma_merge). Another time. 709 */ 710 if (find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma) 711 BUG(); 712 if (!near) 713 goto none; 714 715 vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC); 716 vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC); 717 718 if (near->anon_vma && near->vm_end == vma->vm_start && 719 mpol_equal(vma_policy(near), vma_policy(vma)) && 720 can_vma_merge_after(near, vm_flags, 721 NULL, vma->vm_file, vma->vm_pgoff)) 722 return near->anon_vma; 723 none: 724 /* 725 * There's no absolute need to look only at touching neighbours: 726 * we could search further afield for "compatible" anon_vmas. 727 * But it would probably just be a waste of time searching, 728 * or lead to too many vmas hanging off the same anon_vma. 729 * We're trying to allow mprotect remerging later on, 730 * not trying to minimize memory used for anon_vmas. 731 */ 732 return NULL; 733 } 734 735 /* 736 * The caller must hold down_write(current->mm->mmap_sem). 737 */ 738 739 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr, 740 unsigned long len, unsigned long prot, 741 unsigned long flags, unsigned long pgoff) 742 { 743 struct mm_struct * mm = current->mm; 744 struct vm_area_struct * vma, * prev; 745 struct inode *inode; 746 unsigned int vm_flags; 747 int correct_wcount = 0; 748 int error; 749 struct rb_node ** rb_link, * rb_parent; 750 int accountable = 1; 751 unsigned long charged = 0; 752 753 /* 754 * Does the application expect PROT_READ to imply PROT_EXEC: 755 */ 756 if (unlikely((prot & PROT_READ) && 757 (current->personality & READ_IMPLIES_EXEC))) 758 prot |= PROT_EXEC; 759 760 if (file) { 761 if (is_file_hugepages(file)) 762 accountable = 0; 763 764 if (!file->f_op || !file->f_op->mmap) 765 return -ENODEV; 766 767 if ((prot & PROT_EXEC) && 768 (file->f_vfsmnt->mnt_flags & MNT_NOEXEC)) 769 return -EPERM; 770 } 771 772 if (!len) 773 return addr; 774 775 /* Careful about overflows.. */ 776 len = PAGE_ALIGN(len); 777 if (!len || len > TASK_SIZE) 778 return -EINVAL; 779 780 /* offset overflow? */ 781 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 782 return -EINVAL; 783 784 /* Too many mappings? */ 785 if (mm->map_count > sysctl_max_map_count) 786 return -ENOMEM; 787 788 /* Obtain the address to map to. we verify (or select) it and ensure 789 * that it represents a valid section of the address space. 790 */ 791 addr = get_unmapped_area(file, addr, len, pgoff, flags); 792 if (addr & ~PAGE_MASK) 793 return addr; 794 795 /* Do simple checking here so the lower-level routines won't have 796 * to. we assume access permissions have been handled by the open 797 * of the memory object, so we don't do any here. 798 */ 799 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 800 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 801 802 if (flags & MAP_LOCKED) { 803 if (!capable(CAP_IPC_LOCK)) 804 return -EPERM; 805 vm_flags |= VM_LOCKED; 806 } 807 /* mlock MCL_FUTURE? */ 808 if (vm_flags & VM_LOCKED) { 809 unsigned long locked = mm->locked_vm << PAGE_SHIFT; 810 locked += len; 811 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur) 812 return -EAGAIN; 813 } 814 815 inode = file ? file->f_dentry->d_inode : NULL; 816 817 if (file) { 818 switch (flags & MAP_TYPE) { 819 case MAP_SHARED: 820 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 821 return -EACCES; 822 823 /* 824 * Make sure we don't allow writing to an append-only 825 * file.. 826 */ 827 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 828 return -EACCES; 829 830 /* 831 * Make sure there are no mandatory locks on the file. 832 */ 833 if (locks_verify_locked(inode)) 834 return -EAGAIN; 835 836 vm_flags |= VM_SHARED | VM_MAYSHARE; 837 if (!(file->f_mode & FMODE_WRITE)) 838 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 839 840 /* fall through */ 841 case MAP_PRIVATE: 842 if (!(file->f_mode & FMODE_READ)) 843 return -EACCES; 844 break; 845 846 default: 847 return -EINVAL; 848 } 849 } else { 850 switch (flags & MAP_TYPE) { 851 case MAP_SHARED: 852 vm_flags |= VM_SHARED | VM_MAYSHARE; 853 break; 854 case MAP_PRIVATE: 855 /* 856 * Set pgoff according to addr for anon_vma. 857 */ 858 pgoff = addr >> PAGE_SHIFT; 859 break; 860 default: 861 return -EINVAL; 862 } 863 } 864 865 error = security_file_mmap(file, prot, flags); 866 if (error) 867 return error; 868 869 /* Clear old maps */ 870 error = -ENOMEM; 871 munmap_back: 872 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 873 if (vma && vma->vm_start < addr + len) { 874 if (do_munmap(mm, addr, len)) 875 return -ENOMEM; 876 goto munmap_back; 877 } 878 879 /* Check against address space limit. */ 880 if ((mm->total_vm << PAGE_SHIFT) + len 881 > current->rlim[RLIMIT_AS].rlim_cur) 882 return -ENOMEM; 883 884 if (accountable && (!(flags & MAP_NORESERVE) || 885 sysctl_overcommit_memory > 1)) { 886 if (vm_flags & VM_SHARED) { 887 /* Check memory availability in shmem_file_setup? */ 888 vm_flags |= VM_ACCOUNT; 889 } else if (vm_flags & VM_WRITE) { 890 /* 891 * Private writable mapping: check memory availability 892 */ 893 charged = len >> PAGE_SHIFT; 894 if (security_vm_enough_memory(charged)) 895 return -ENOMEM; 896 vm_flags |= VM_ACCOUNT; 897 } 898 } 899 900 /* 901 * Can we just expand an old private anonymous mapping? 902 * The VM_SHARED test is necessary because shmem_zero_setup 903 * will create the file object for a shared anonymous map below. 904 */ 905 if (!file && !(vm_flags & VM_SHARED) && 906 vma_merge(mm, prev, addr, addr + len, vm_flags, 907 NULL, NULL, pgoff, NULL)) 908 goto out; 909 910 /* 911 * Determine the object being mapped and call the appropriate 912 * specific mapper. the address has already been validated, but 913 * not unmapped, but the maps are removed from the list. 914 */ 915 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL); 916 if (!vma) { 917 error = -ENOMEM; 918 goto unacct_error; 919 } 920 memset(vma, 0, sizeof(*vma)); 921 922 vma->vm_mm = mm; 923 vma->vm_start = addr; 924 vma->vm_end = addr + len; 925 vma->vm_flags = vm_flags; 926 vma->vm_page_prot = protection_map[vm_flags & 0x0f]; 927 vma->vm_pgoff = pgoff; 928 929 if (file) { 930 error = -EINVAL; 931 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 932 goto free_vma; 933 if (vm_flags & VM_DENYWRITE) { 934 error = deny_write_access(file); 935 if (error) 936 goto free_vma; 937 correct_wcount = 1; 938 } 939 vma->vm_file = file; 940 get_file(file); 941 error = file->f_op->mmap(file, vma); 942 if (error) 943 goto unmap_and_free_vma; 944 } else if (vm_flags & VM_SHARED) { 945 error = shmem_zero_setup(vma); 946 if (error) 947 goto free_vma; 948 } 949 950 /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform 951 * shmem_zero_setup (perhaps called through /dev/zero's ->mmap) 952 * that memory reservation must be checked; but that reservation 953 * belongs to shared memory object, not to vma: so now clear it. 954 */ 955 if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT)) 956 vma->vm_flags &= ~VM_ACCOUNT; 957 958 /* Can addr have changed?? 959 * 960 * Answer: Yes, several device drivers can do it in their 961 * f_op->mmap method. -DaveM 962 */ 963 addr = vma->vm_start; 964 965 if (!file || !vma_merge(mm, prev, addr, vma->vm_end, 966 vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) { 967 vma_link(mm, vma, prev, rb_link, rb_parent); 968 if (correct_wcount) 969 atomic_inc(&inode->i_writecount); 970 } else { 971 if (file) { 972 if (correct_wcount) 973 atomic_inc(&inode->i_writecount); 974 fput(file); 975 } 976 mpol_free(vma_policy(vma)); 977 kmem_cache_free(vm_area_cachep, vma); 978 } 979 out: 980 mm->total_vm += len >> PAGE_SHIFT; 981 if (vm_flags & VM_LOCKED) { 982 mm->locked_vm += len >> PAGE_SHIFT; 983 make_pages_present(addr, addr + len); 984 } 985 if (flags & MAP_POPULATE) { 986 up_write(&mm->mmap_sem); 987 sys_remap_file_pages(addr, len, 0, 988 pgoff, flags & MAP_NONBLOCK); 989 down_write(&mm->mmap_sem); 990 } 991 return addr; 992 993 unmap_and_free_vma: 994 if (correct_wcount) 995 atomic_inc(&inode->i_writecount); 996 vma->vm_file = NULL; 997 fput(file); 998 999 /* Undo any partial mapping done by a device driver. */ 1000 zap_page_range(vma, vma->vm_start, vma->vm_end - vma->vm_start, NULL); 1001 free_vma: 1002 kmem_cache_free(vm_area_cachep, vma); 1003 unacct_error: 1004 if (charged) 1005 vm_unacct_memory(charged); 1006 return error; 1007 } 1008 1009 EXPORT_SYMBOL(do_mmap_pgoff); 1010 1011 /* Get an address range which is currently unmapped. 1012 * For shmat() with addr=0. 1013 * 1014 * Ugly calling convention alert: 1015 * Return value with the low bits set means error value, 1016 * ie 1017 * if (ret & ~PAGE_MASK) 1018 * error = ret; 1019 * 1020 * This function "knows" that -ENOMEM has the bits set. 1021 */ 1022 #ifndef HAVE_ARCH_UNMAPPED_AREA 1023 static inline unsigned long 1024 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1025 unsigned long len, unsigned long pgoff, unsigned long flags) 1026 { 1027 struct mm_struct *mm = current->mm; 1028 struct vm_area_struct *vma; 1029 unsigned long start_addr; 1030 1031 if (len > TASK_SIZE) 1032 return -ENOMEM; 1033 1034 if (addr) { 1035 addr = PAGE_ALIGN(addr); 1036 vma = find_vma(mm, addr); 1037 if (TASK_SIZE - len >= addr && 1038 (!vma || addr + len <= vma->vm_start)) 1039 return addr; 1040 } 1041 start_addr = addr = mm->free_area_cache; 1042 1043 full_search: 1044 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1045 /* At this point: (!vma || addr < vma->vm_end). */ 1046 if (TASK_SIZE - len < addr) { 1047 /* 1048 * Start a new search - just in case we missed 1049 * some holes. 1050 */ 1051 if (start_addr != TASK_UNMAPPED_BASE) { 1052 start_addr = addr = TASK_UNMAPPED_BASE; 1053 goto full_search; 1054 } 1055 return -ENOMEM; 1056 } 1057 if (!vma || addr + len <= vma->vm_start) { 1058 /* 1059 * Remember the place where we stopped the search: 1060 */ 1061 mm->free_area_cache = addr + len; 1062 return addr; 1063 } 1064 addr = vma->vm_end; 1065 } 1066 } 1067 #else 1068 extern unsigned long 1069 arch_get_unmapped_area(struct file *, unsigned long, unsigned long, 1070 unsigned long, unsigned long); 1071 #endif 1072 1073 unsigned long 1074 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1075 unsigned long pgoff, unsigned long flags) 1076 { 1077 unsigned long ret; 1078 1079 if (!(flags & MAP_FIXED)) { 1080 unsigned long (*get_area)(struct file *, unsigned long, unsigned long, unsigned long, unsigned long); 1081 1082 get_area = arch_get_unmapped_area; 1083 if (file && file->f_op && file->f_op->get_unmapped_area) 1084 get_area = file->f_op->get_unmapped_area; 1085 addr = get_area(file, addr, len, pgoff, flags); 1086 if (IS_ERR_VALUE(addr)) 1087 return addr; 1088 } 1089 1090 1091 if (addr > TASK_SIZE - len) 1092 return -ENOMEM; 1093 if (addr & ~PAGE_MASK) 1094 return -EINVAL; 1095 if (file && is_file_hugepages(file)) { 1096 /* 1097 * Check if the given range is hugepage aligned, and 1098 * can be made suitable for hugepages. 1099 */ 1100 ret = prepare_hugepage_range(addr, len); 1101 } else { 1102 /* 1103 * Ensure that a normal request is not falling in a 1104 * reserved hugepage range. For some archs like IA-64, 1105 * there is a separate region for hugepages. 1106 */ 1107 ret = is_hugepage_only_range(current->mm, addr, len); 1108 } 1109 if (ret) 1110 return -EINVAL; 1111 return addr; 1112 } 1113 1114 EXPORT_SYMBOL(get_unmapped_area); 1115 1116 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 1117 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr) 1118 { 1119 struct vm_area_struct *vma = NULL; 1120 1121 if (mm) { 1122 /* Check the cache first. */ 1123 /* (Cache hit rate is typically around 35%.) */ 1124 vma = mm->mmap_cache; 1125 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) { 1126 struct rb_node * rb_node; 1127 1128 rb_node = mm->mm_rb.rb_node; 1129 vma = NULL; 1130 1131 while (rb_node) { 1132 struct vm_area_struct * vma_tmp; 1133 1134 vma_tmp = rb_entry(rb_node, 1135 struct vm_area_struct, vm_rb); 1136 1137 if (vma_tmp->vm_end > addr) { 1138 vma = vma_tmp; 1139 if (vma_tmp->vm_start <= addr) 1140 break; 1141 rb_node = rb_node->rb_left; 1142 } else 1143 rb_node = rb_node->rb_right; 1144 } 1145 if (vma) 1146 mm->mmap_cache = vma; 1147 } 1148 } 1149 return vma; 1150 } 1151 1152 EXPORT_SYMBOL(find_vma); 1153 1154 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */ 1155 struct vm_area_struct * 1156 find_vma_prev(struct mm_struct *mm, unsigned long addr, 1157 struct vm_area_struct **pprev) 1158 { 1159 struct vm_area_struct *vma = NULL, *prev = NULL; 1160 struct rb_node * rb_node; 1161 if (!mm) 1162 goto out; 1163 1164 /* Guard against addr being lower than the first VMA */ 1165 vma = mm->mmap; 1166 1167 /* Go through the RB tree quickly. */ 1168 rb_node = mm->mm_rb.rb_node; 1169 1170 while (rb_node) { 1171 struct vm_area_struct *vma_tmp; 1172 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 1173 1174 if (addr < vma_tmp->vm_end) { 1175 rb_node = rb_node->rb_left; 1176 } else { 1177 prev = vma_tmp; 1178 if (!prev->vm_next || (addr < prev->vm_next->vm_end)) 1179 break; 1180 rb_node = rb_node->rb_right; 1181 } 1182 } 1183 1184 out: 1185 *pprev = prev; 1186 return prev ? prev->vm_next : vma; 1187 } 1188 1189 /* 1190 * Verify that the stack growth is acceptable and 1191 * update accounting. This is shared with both the 1192 * grow-up and grow-down cases. 1193 */ 1194 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow) 1195 { 1196 struct mm_struct *mm = vma->vm_mm; 1197 struct rlimit *rlim = current->rlim; 1198 1199 /* address space limit tests */ 1200 if (mm->total_vm + grow > rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT) 1201 return -ENOMEM; 1202 1203 /* Stack limit test */ 1204 if (size > rlim[RLIMIT_STACK].rlim_cur) 1205 return -ENOMEM; 1206 1207 /* 1208 * Overcommit.. This must be the final test, as it will 1209 * update security statistics. 1210 */ 1211 if (security_vm_enough_memory(grow)) 1212 return -ENOMEM; 1213 1214 /* Ok, everything looks good - let it rip */ 1215 mm->total_vm += grow; 1216 if (vma->vm_flags & VM_LOCKED) 1217 mm->locked_vm += grow; 1218 return 0; 1219 } 1220 1221 #ifdef CONFIG_STACK_GROWSUP 1222 /* 1223 * vma is the first one with address > vma->vm_end. Have to extend vma. 1224 */ 1225 int expand_stack(struct vm_area_struct * vma, unsigned long address) 1226 { 1227 int error; 1228 1229 if (!(vma->vm_flags & VM_GROWSUP)) 1230 return -EFAULT; 1231 1232 /* 1233 * We must make sure the anon_vma is allocated 1234 * so that the anon_vma locking is not a noop. 1235 */ 1236 if (unlikely(anon_vma_prepare(vma))) 1237 return -ENOMEM; 1238 anon_vma_lock(vma); 1239 1240 /* 1241 * vma->vm_start/vm_end cannot change under us because the caller 1242 * is required to hold the mmap_sem in read mode. We need the 1243 * anon_vma lock to serialize against concurrent expand_stacks. 1244 */ 1245 address += 4 + PAGE_SIZE - 1; 1246 address &= PAGE_MASK; 1247 error = 0; 1248 1249 /* Somebody else might have raced and expanded it already */ 1250 if (address > vma->vm_end) { 1251 unsigned long size, grow; 1252 1253 size = address - vma->vm_start; 1254 grow = (address - vma->vm_end) >> PAGE_SHIFT; 1255 1256 error = acct_stack_growth(vma, size, grow); 1257 if (!error) 1258 vma->vm_end = address; 1259 } 1260 anon_vma_unlock(vma); 1261 return error; 1262 } 1263 1264 struct vm_area_struct * 1265 find_extend_vma(struct mm_struct *mm, unsigned long addr) 1266 { 1267 struct vm_area_struct *vma, *prev; 1268 1269 addr &= PAGE_MASK; 1270 vma = find_vma_prev(mm, addr, &prev); 1271 if (vma && (vma->vm_start <= addr)) 1272 return vma; 1273 if (!prev || expand_stack(prev, addr)) 1274 return NULL; 1275 if (prev->vm_flags & VM_LOCKED) { 1276 make_pages_present(addr, prev->vm_end); 1277 } 1278 return prev; 1279 } 1280 #else 1281 /* 1282 * vma is the first one with address < vma->vm_start. Have to extend vma. 1283 */ 1284 int expand_stack(struct vm_area_struct *vma, unsigned long address) 1285 { 1286 int error; 1287 1288 /* 1289 * We must make sure the anon_vma is allocated 1290 * so that the anon_vma locking is not a noop. 1291 */ 1292 if (unlikely(anon_vma_prepare(vma))) 1293 return -ENOMEM; 1294 anon_vma_lock(vma); 1295 1296 /* 1297 * vma->vm_start/vm_end cannot change under us because the caller 1298 * is required to hold the mmap_sem in read mode. We need the 1299 * anon_vma lock to serialize against concurrent expand_stacks. 1300 */ 1301 address &= PAGE_MASK; 1302 error = 0; 1303 1304 /* Somebody else might have raced and expanded it already */ 1305 if (address < vma->vm_start) { 1306 unsigned long size, grow; 1307 1308 size = vma->vm_end - address; 1309 grow = (vma->vm_start - address) >> PAGE_SHIFT; 1310 1311 error = acct_stack_growth(vma, size, grow); 1312 if (!error) { 1313 vma->vm_start = address; 1314 vma->vm_pgoff -= grow; 1315 } 1316 } 1317 anon_vma_unlock(vma); 1318 return error; 1319 } 1320 1321 struct vm_area_struct * 1322 find_extend_vma(struct mm_struct * mm, unsigned long addr) 1323 { 1324 struct vm_area_struct * vma; 1325 unsigned long start; 1326 1327 addr &= PAGE_MASK; 1328 vma = find_vma(mm,addr); 1329 if (!vma) 1330 return NULL; 1331 if (vma->vm_start <= addr) 1332 return vma; 1333 if (!(vma->vm_flags & VM_GROWSDOWN)) 1334 return NULL; 1335 start = vma->vm_start; 1336 if (expand_stack(vma, addr)) 1337 return NULL; 1338 if (vma->vm_flags & VM_LOCKED) { 1339 make_pages_present(addr, start); 1340 } 1341 return vma; 1342 } 1343 #endif 1344 1345 /* 1346 * Try to free as many page directory entries as we can, 1347 * without having to work very hard at actually scanning 1348 * the page tables themselves. 1349 * 1350 * Right now we try to free page tables if we have a nice 1351 * PGDIR-aligned area that got free'd up. We could be more 1352 * granular if we want to, but this is fast and simple, 1353 * and covers the bad cases. 1354 * 1355 * "prev", if it exists, points to a vma before the one 1356 * we just free'd - but there's no telling how much before. 1357 */ 1358 static void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *prev, 1359 unsigned long start, unsigned long end) 1360 { 1361 unsigned long first = start & PGDIR_MASK; 1362 unsigned long last = end + PGDIR_SIZE - 1; 1363 unsigned long start_index, end_index; 1364 struct mm_struct *mm = tlb->mm; 1365 1366 if (!prev) { 1367 prev = mm->mmap; 1368 if (!prev) 1369 goto no_mmaps; 1370 if (prev->vm_end > start) { 1371 if (last > prev->vm_start) 1372 last = prev->vm_start; 1373 goto no_mmaps; 1374 } 1375 } 1376 for (;;) { 1377 struct vm_area_struct *next = prev->vm_next; 1378 1379 if (next) { 1380 if (next->vm_start < start) { 1381 prev = next; 1382 continue; 1383 } 1384 if (last > next->vm_start) 1385 last = next->vm_start; 1386 } 1387 if (prev->vm_end > first) 1388 first = prev->vm_end + PGDIR_SIZE - 1; 1389 break; 1390 } 1391 no_mmaps: 1392 if (last < first) /* for arches with discontiguous pgd indices */ 1393 return; 1394 /* 1395 * If the PGD bits are not consecutive in the virtual address, the 1396 * old method of shifting the VA >> by PGDIR_SHIFT doesn't work. 1397 */ 1398 start_index = pgd_index(first); 1399 if (start_index < FIRST_USER_PGD_NR) 1400 start_index = FIRST_USER_PGD_NR; 1401 end_index = pgd_index(last); 1402 if (end_index > start_index) { 1403 clear_page_tables(tlb, start_index, end_index - start_index); 1404 flush_tlb_pgtables(mm, first & PGDIR_MASK, last & PGDIR_MASK); 1405 } 1406 } 1407 1408 /* Normal function to fix up a mapping 1409 * This function is the default for when an area has no specific 1410 * function. This may be used as part of a more specific routine. 1411 * 1412 * By the time this function is called, the area struct has been 1413 * removed from the process mapping list. 1414 */ 1415 static void unmap_vma(struct mm_struct *mm, struct vm_area_struct *area) 1416 { 1417 size_t len = area->vm_end - area->vm_start; 1418 1419 area->vm_mm->total_vm -= len >> PAGE_SHIFT; 1420 if (area->vm_flags & VM_LOCKED) 1421 area->vm_mm->locked_vm -= len >> PAGE_SHIFT; 1422 /* 1423 * Is this a new hole at the lowest possible address? 1424 */ 1425 if (area->vm_start >= TASK_UNMAPPED_BASE && 1426 area->vm_start < area->vm_mm->free_area_cache) 1427 area->vm_mm->free_area_cache = area->vm_start; 1428 1429 remove_vm_struct(area); 1430 } 1431 1432 /* 1433 * Update the VMA and inode share lists. 1434 * 1435 * Ok - we have the memory areas we should free on the 'free' list, 1436 * so release them, and do the vma updates. 1437 */ 1438 static void unmap_vma_list(struct mm_struct *mm, 1439 struct vm_area_struct *mpnt) 1440 { 1441 do { 1442 struct vm_area_struct *next = mpnt->vm_next; 1443 unmap_vma(mm, mpnt); 1444 mpnt = next; 1445 } while (mpnt != NULL); 1446 validate_mm(mm); 1447 } 1448 1449 /* 1450 * Get rid of page table information in the indicated region. 1451 * 1452 * Called with the page table lock held. 1453 */ 1454 static void unmap_region(struct mm_struct *mm, 1455 struct vm_area_struct *vma, 1456 struct vm_area_struct *prev, 1457 unsigned long start, 1458 unsigned long end) 1459 { 1460 struct mmu_gather *tlb; 1461 unsigned long nr_accounted = 0; 1462 1463 lru_add_drain(); 1464 tlb = tlb_gather_mmu(mm, 0); 1465 unmap_vmas(&tlb, mm, vma, start, end, &nr_accounted, NULL); 1466 vm_unacct_memory(nr_accounted); 1467 1468 if (is_hugepage_only_range(mm, start, end - start)) 1469 hugetlb_free_pgtables(tlb, prev, start, end); 1470 else 1471 free_pgtables(tlb, prev, start, end); 1472 tlb_finish_mmu(tlb, start, end); 1473 } 1474 1475 /* 1476 * Create a list of vma's touched by the unmap, removing them from the mm's 1477 * vma list as we go.. 1478 */ 1479 static void 1480 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 1481 struct vm_area_struct *prev, unsigned long end) 1482 { 1483 struct vm_area_struct **insertion_point; 1484 struct vm_area_struct *tail_vma = NULL; 1485 1486 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 1487 do { 1488 rb_erase(&vma->vm_rb, &mm->mm_rb); 1489 mm->map_count--; 1490 tail_vma = vma; 1491 vma = vma->vm_next; 1492 } while (vma && vma->vm_start < end); 1493 *insertion_point = vma; 1494 tail_vma->vm_next = NULL; 1495 mm->mmap_cache = NULL; /* Kill the cache. */ 1496 } 1497 1498 /* 1499 * Split a vma into two pieces at address 'addr', a new vma is allocated 1500 * either for the first part or the the tail. 1501 */ 1502 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma, 1503 unsigned long addr, int new_below) 1504 { 1505 struct mempolicy *pol; 1506 struct vm_area_struct *new; 1507 1508 if (mm->map_count >= sysctl_max_map_count) 1509 return -ENOMEM; 1510 1511 new = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL); 1512 if (!new) 1513 return -ENOMEM; 1514 1515 /* most fields are the same, copy all, and then fixup */ 1516 *new = *vma; 1517 vma_prio_tree_init(new); 1518 1519 if (new_below) 1520 new->vm_end = addr; 1521 else { 1522 new->vm_start = addr; 1523 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 1524 } 1525 1526 pol = mpol_copy(vma_policy(vma)); 1527 if (IS_ERR(pol)) { 1528 kmem_cache_free(vm_area_cachep, new); 1529 return PTR_ERR(pol); 1530 } 1531 vma_set_policy(new, pol); 1532 1533 if (new->vm_file) 1534 get_file(new->vm_file); 1535 1536 if (new->vm_ops && new->vm_ops->open) 1537 new->vm_ops->open(new); 1538 1539 if (new_below) 1540 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 1541 ((addr - new->vm_start) >> PAGE_SHIFT), new); 1542 else 1543 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 1544 1545 return 0; 1546 } 1547 1548 /* Munmap is split into 2 main parts -- this part which finds 1549 * what needs doing, and the areas themselves, which do the 1550 * work. This now handles partial unmappings. 1551 * Jeremy Fitzhardinge <jeremy@goop.org> 1552 */ 1553 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 1554 { 1555 unsigned long end; 1556 struct vm_area_struct *mpnt, *prev, *last; 1557 1558 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start) 1559 return -EINVAL; 1560 1561 if ((len = PAGE_ALIGN(len)) == 0) 1562 return -EINVAL; 1563 1564 /* Find the first overlapping VMA */ 1565 mpnt = find_vma_prev(mm, start, &prev); 1566 if (!mpnt) 1567 return 0; 1568 /* we have start < mpnt->vm_end */ 1569 1570 if (is_vm_hugetlb_page(mpnt)) { 1571 int ret = is_aligned_hugepage_range(start, len); 1572 1573 if (ret) 1574 return ret; 1575 } 1576 1577 /* if it doesn't overlap, we have nothing.. */ 1578 end = start + len; 1579 if (mpnt->vm_start >= end) 1580 return 0; 1581 1582 /* Something will probably happen, so notify. */ 1583 if (mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) 1584 profile_exec_unmap(mm); 1585 1586 /* 1587 * If we need to split any vma, do it now to save pain later. 1588 * 1589 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 1590 * unmapped vm_area_struct will remain in use: so lower split_vma 1591 * places tmp vma above, and higher split_vma places tmp vma below. 1592 */ 1593 if (start > mpnt->vm_start) { 1594 if (split_vma(mm, mpnt, start, 0)) 1595 return -ENOMEM; 1596 prev = mpnt; 1597 } 1598 1599 /* Does it split the last one? */ 1600 last = find_vma(mm, end); 1601 if (last && end > last->vm_start) { 1602 if (split_vma(mm, last, end, 1)) 1603 return -ENOMEM; 1604 } 1605 mpnt = prev? prev->vm_next: mm->mmap; 1606 1607 /* 1608 * Remove the vma's, and unmap the actual pages 1609 */ 1610 detach_vmas_to_be_unmapped(mm, mpnt, prev, end); 1611 spin_lock(&mm->page_table_lock); 1612 unmap_region(mm, mpnt, prev, start, end); 1613 spin_unlock(&mm->page_table_lock); 1614 1615 /* Fix up all other VM information */ 1616 unmap_vma_list(mm, mpnt); 1617 1618 return 0; 1619 } 1620 1621 EXPORT_SYMBOL(do_munmap); 1622 1623 asmlinkage long sys_munmap(unsigned long addr, size_t len) 1624 { 1625 int ret; 1626 struct mm_struct *mm = current->mm; 1627 1628 down_write(&mm->mmap_sem); 1629 ret = do_munmap(mm, addr, len); 1630 up_write(&mm->mmap_sem); 1631 return ret; 1632 } 1633 1634 /* 1635 * this is really a simplified "do_mmap". it only handles 1636 * anonymous maps. eventually we may be able to do some 1637 * brk-specific accounting here. 1638 */ 1639 unsigned long do_brk(unsigned long addr, unsigned long len) 1640 { 1641 struct mm_struct * mm = current->mm; 1642 struct vm_area_struct * vma, * prev; 1643 unsigned long flags; 1644 struct rb_node ** rb_link, * rb_parent; 1645 pgoff_t pgoff = addr >> PAGE_SHIFT; 1646 1647 len = PAGE_ALIGN(len); 1648 if (!len) 1649 return addr; 1650 1651 if ((addr + len) > TASK_SIZE || (addr + len) < addr) 1652 return -EINVAL; 1653 1654 /* 1655 * mlock MCL_FUTURE? 1656 */ 1657 if (mm->def_flags & VM_LOCKED) { 1658 unsigned long locked = mm->locked_vm << PAGE_SHIFT; 1659 locked += len; 1660 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur) 1661 return -EAGAIN; 1662 } 1663 1664 /* 1665 * mm->mmap_sem is required to protect against another thread 1666 * changing the mappings in case we sleep. 1667 */ 1668 WARN_ON(down_read_trylock(&mm->mmap_sem)); 1669 1670 /* 1671 * Clear old maps. this also does some error checking for us 1672 */ 1673 munmap_back: 1674 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 1675 if (vma && vma->vm_start < addr + len) { 1676 if (do_munmap(mm, addr, len)) 1677 return -ENOMEM; 1678 goto munmap_back; 1679 } 1680 1681 /* Check against address space limits *after* clearing old maps... */ 1682 if ((mm->total_vm << PAGE_SHIFT) + len 1683 > current->rlim[RLIMIT_AS].rlim_cur) 1684 return -ENOMEM; 1685 1686 if (mm->map_count > sysctl_max_map_count) 1687 return -ENOMEM; 1688 1689 if (security_vm_enough_memory(len >> PAGE_SHIFT)) 1690 return -ENOMEM; 1691 1692 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 1693 1694 /* Can we just expand an old private anonymous mapping? */ 1695 if (vma_merge(mm, prev, addr, addr + len, flags, 1696 NULL, NULL, pgoff, NULL)) 1697 goto out; 1698 1699 /* 1700 * create a vma struct for an anonymous mapping 1701 */ 1702 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL); 1703 if (!vma) { 1704 vm_unacct_memory(len >> PAGE_SHIFT); 1705 return -ENOMEM; 1706 } 1707 memset(vma, 0, sizeof(*vma)); 1708 1709 vma->vm_mm = mm; 1710 vma->vm_start = addr; 1711 vma->vm_end = addr + len; 1712 vma->vm_pgoff = pgoff; 1713 vma->vm_flags = flags; 1714 vma->vm_page_prot = protection_map[flags & 0x0f]; 1715 vma_link(mm, vma, prev, rb_link, rb_parent); 1716 out: 1717 mm->total_vm += len >> PAGE_SHIFT; 1718 if (flags & VM_LOCKED) { 1719 mm->locked_vm += len >> PAGE_SHIFT; 1720 make_pages_present(addr, addr + len); 1721 } 1722 return addr; 1723 } 1724 1725 EXPORT_SYMBOL(do_brk); 1726 1727 /* Release all mmaps. */ 1728 void exit_mmap(struct mm_struct *mm) 1729 { 1730 struct mmu_gather *tlb; 1731 struct vm_area_struct *vma; 1732 unsigned long nr_accounted = 0; 1733 1734 profile_exit_mmap(mm); 1735 1736 lru_add_drain(); 1737 1738 spin_lock(&mm->page_table_lock); 1739 1740 tlb = tlb_gather_mmu(mm, 1); 1741 flush_cache_mm(mm); 1742 /* Use ~0UL here to ensure all VMAs in the mm are unmapped */ 1743 mm->map_count -= unmap_vmas(&tlb, mm, mm->mmap, 0, 1744 ~0UL, &nr_accounted, NULL); 1745 vm_unacct_memory(nr_accounted); 1746 BUG_ON(mm->map_count); /* This is just debugging */ 1747 clear_page_tables(tlb, FIRST_USER_PGD_NR, USER_PTRS_PER_PGD); 1748 tlb_finish_mmu(tlb, 0, MM_VM_SIZE(mm)); 1749 1750 vma = mm->mmap; 1751 mm->mmap = mm->mmap_cache = NULL; 1752 mm->mm_rb = RB_ROOT; 1753 mm->rss = 0; 1754 mm->total_vm = 0; 1755 mm->locked_vm = 0; 1756 1757 spin_unlock(&mm->page_table_lock); 1758 1759 /* 1760 * Walk the list again, actually closing and freeing it 1761 * without holding any MM locks. 1762 */ 1763 while (vma) { 1764 struct vm_area_struct *next = vma->vm_next; 1765 remove_vm_struct(vma); 1766 vma = next; 1767 } 1768 } 1769 1770 /* Insert vm structure into process list sorted by address 1771 * and into the inode's i_mmap tree. If vm_file is non-NULL 1772 * then i_mmap_lock is taken here. 1773 */ 1774 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma) 1775 { 1776 struct vm_area_struct * __vma, * prev; 1777 struct rb_node ** rb_link, * rb_parent; 1778 1779 /* 1780 * The vm_pgoff of a purely anonymous vma should be irrelevant 1781 * until its first write fault, when page's anon_vma and index 1782 * are set. But now set the vm_pgoff it will almost certainly 1783 * end up with (unless mremap moves it elsewhere before that 1784 * first wfault), so /proc/pid/maps tells a consistent story. 1785 * 1786 * By setting it to reflect the virtual start address of the 1787 * vma, merges and splits can happen in a seamless way, just 1788 * using the existing file pgoff checks and manipulations. 1789 * Similarly in do_mmap_pgoff and in do_brk. 1790 */ 1791 if (!vma->vm_file) { 1792 BUG_ON(vma->anon_vma); 1793 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 1794 } 1795 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent); 1796 if (__vma && __vma->vm_start < vma->vm_end) 1797 return -ENOMEM; 1798 vma_link(mm, vma, prev, rb_link, rb_parent); 1799 return 0; 1800 } 1801 1802 /* 1803 * Copy the vma structure to a new location in the same mm, 1804 * prior to moving page table entries, to effect an mremap move. 1805 */ 1806 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 1807 unsigned long addr, unsigned long len, pgoff_t pgoff) 1808 { 1809 struct vm_area_struct *vma = *vmap; 1810 unsigned long vma_start = vma->vm_start; 1811 struct mm_struct *mm = vma->vm_mm; 1812 struct vm_area_struct *new_vma, *prev; 1813 struct rb_node **rb_link, *rb_parent; 1814 struct mempolicy *pol; 1815 1816 /* 1817 * If anonymous vma has not yet been faulted, update new pgoff 1818 * to match new location, to increase its chance of merging. 1819 */ 1820 if (!vma->vm_file && !vma->anon_vma) 1821 pgoff = addr >> PAGE_SHIFT; 1822 1823 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 1824 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 1825 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma)); 1826 if (new_vma) { 1827 /* 1828 * Source vma may have been merged into new_vma 1829 */ 1830 if (vma_start >= new_vma->vm_start && 1831 vma_start < new_vma->vm_end) 1832 *vmap = new_vma; 1833 } else { 1834 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL); 1835 if (new_vma) { 1836 *new_vma = *vma; 1837 vma_prio_tree_init(new_vma); 1838 pol = mpol_copy(vma_policy(vma)); 1839 if (IS_ERR(pol)) { 1840 kmem_cache_free(vm_area_cachep, new_vma); 1841 return NULL; 1842 } 1843 vma_set_policy(new_vma, pol); 1844 new_vma->vm_start = addr; 1845 new_vma->vm_end = addr + len; 1846 new_vma->vm_pgoff = pgoff; 1847 if (new_vma->vm_file) 1848 get_file(new_vma->vm_file); 1849 if (new_vma->vm_ops && new_vma->vm_ops->open) 1850 new_vma->vm_ops->open(new_vma); 1851 vma_link(mm, new_vma, prev, rb_link, rb_parent); 1852 } 1853 } 1854 return new_vma; 1855 } 1856
This page was automatically generated by LXR 0.3.1. • Linux is a registered trademark of Linus Torvalds