1 /* 2 * mm/mmap.c 3 * 4 * Written by obz. 5 * 6 * Address space accounting code <alan@lxorguk.ukuu.org.uk> 7 */ 8 9 #include <linux/slab.h> 10 #include <linux/backing-dev.h> 11 #include <linux/mm.h> 12 #include <linux/shm.h> 13 #include <linux/mman.h> 14 #include <linux/pagemap.h> 15 #include <linux/swap.h> 16 #include <linux/syscalls.h> 17 #include <linux/capability.h> 18 #include <linux/init.h> 19 #include <linux/file.h> 20 #include <linux/fs.h> 21 #include <linux/personality.h> 22 #include <linux/security.h> 23 #include <linux/hugetlb.h> 24 #include <linux/profile.h> 25 #include <linux/module.h> 26 #include <linux/mount.h> 27 #include <linux/mempolicy.h> 28 #include <linux/rmap.h> 29 #include <linux/mmu_notifier.h> 30 #include <linux/perf_event.h> 31 32 #include <asm/uaccess.h> 33 #include <asm/cacheflush.h> 34 #include <asm/tlb.h> 35 #include <asm/mmu_context.h> 36 37 #include "internal.h" 38 39 #ifndef arch_mmap_check 40 #define arch_mmap_check(addr, len, flags) (0) 41 #endif 42 43 #ifndef arch_rebalance_pgtables 44 #define arch_rebalance_pgtables(addr, len) (addr) 45 #endif 46 47 static void unmap_region(struct mm_struct *mm, 48 struct vm_area_struct *vma, struct vm_area_struct *prev, 49 unsigned long start, unsigned long end); 50 51 /* 52 * WARNING: the debugging will use recursive algorithms so never enable this 53 * unless you know what you are doing. 54 */ 55 #undef DEBUG_MM_RB 56 57 /* description of effects of mapping type and prot in current implementation. 58 * this is due to the limited x86 page protection hardware. The expected 59 * behavior is in parens: 60 * 61 * map_type prot 62 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 63 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 64 * w: (no) no w: (no) no w: (yes) yes w: (no) no 65 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 66 * 67 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 68 * w: (no) no w: (no) no w: (copy) copy w: (no) no 69 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 70 * 71 */ 72 pgprot_t protection_map[16] = { 73 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 74 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 75 }; 76 77 pgprot_t vm_get_page_prot(unsigned long vm_flags) 78 { 79 return __pgprot(pgprot_val(protection_map[vm_flags & 80 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) | 81 pgprot_val(arch_vm_get_page_prot(vm_flags))); 82 } 83 EXPORT_SYMBOL(vm_get_page_prot); 84 85 int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */ 86 int sysctl_overcommit_ratio = 50; /* default is 50% */ 87 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT; 88 struct percpu_counter vm_committed_as; 89 90 /* 91 * Check that a process has enough memory to allocate a new virtual 92 * mapping. 0 means there is enough memory for the allocation to 93 * succeed and -ENOMEM implies there is not. 94 * 95 * We currently support three overcommit policies, which are set via the 96 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting 97 * 98 * Strict overcommit modes added 2002 Feb 26 by Alan Cox. 99 * Additional code 2002 Jul 20 by Robert Love. 100 * 101 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise. 102 * 103 * Note this is a helper function intended to be used by LSMs which 104 * wish to use this logic. 105 */ 106 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin) 107 { 108 unsigned long free, allowed; 109 110 vm_acct_memory(pages); 111 112 /* 113 * Sometimes we want to use more memory than we have 114 */ 115 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS) 116 return 0; 117 118 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) { 119 unsigned long n; 120 121 free = global_page_state(NR_FILE_PAGES); 122 free += nr_swap_pages; 123 124 /* 125 * Any slabs which are created with the 126 * SLAB_RECLAIM_ACCOUNT flag claim to have contents 127 * which are reclaimable, under pressure. The dentry 128 * cache and most inode caches should fall into this 129 */ 130 free += global_page_state(NR_SLAB_RECLAIMABLE); 131 132 /* 133 * Leave the last 3% for root 134 */ 135 if (!cap_sys_admin) 136 free -= free / 32; 137 138 if (free > pages) 139 return 0; 140 141 /* 142 * nr_free_pages() is very expensive on large systems, 143 * only call if we're about to fail. 144 */ 145 n = nr_free_pages(); 146 147 /* 148 * Leave reserved pages. The pages are not for anonymous pages. 149 */ 150 if (n <= totalreserve_pages) 151 goto error; 152 else 153 n -= totalreserve_pages; 154 155 /* 156 * Leave the last 3% for root 157 */ 158 if (!cap_sys_admin) 159 n -= n / 32; 160 free += n; 161 162 if (free > pages) 163 return 0; 164 165 goto error; 166 } 167 168 allowed = (totalram_pages - hugetlb_total_pages()) 169 * sysctl_overcommit_ratio / 100; 170 /* 171 * Leave the last 3% for root 172 */ 173 if (!cap_sys_admin) 174 allowed -= allowed / 32; 175 allowed += total_swap_pages; 176 177 /* Don't let a single process grow too big: 178 leave 3% of the size of this process for other processes */ 179 if (mm) 180 allowed -= mm->total_vm / 32; 181 182 if (percpu_counter_read_positive(&vm_committed_as) < allowed) 183 return 0; 184 error: 185 vm_unacct_memory(pages); 186 187 return -ENOMEM; 188 } 189 190 /* 191 * Requires inode->i_mapping->i_mmap_lock 192 */ 193 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 194 struct file *file, struct address_space *mapping) 195 { 196 if (vma->vm_flags & VM_DENYWRITE) 197 atomic_inc(&file->f_path.dentry->d_inode->i_writecount); 198 if (vma->vm_flags & VM_SHARED) 199 mapping->i_mmap_writable--; 200 201 flush_dcache_mmap_lock(mapping); 202 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 203 list_del_init(&vma->shared.vm_set.list); 204 else 205 vma_prio_tree_remove(vma, &mapping->i_mmap); 206 flush_dcache_mmap_unlock(mapping); 207 } 208 209 /* 210 * Unlink a file-based vm structure from its prio_tree, to hide 211 * vma from rmap and vmtruncate before freeing its page tables. 212 */ 213 void unlink_file_vma(struct vm_area_struct *vma) 214 { 215 struct file *file = vma->vm_file; 216 217 if (file) { 218 struct address_space *mapping = file->f_mapping; 219 spin_lock(&mapping->i_mmap_lock); 220 __remove_shared_vm_struct(vma, file, mapping); 221 spin_unlock(&mapping->i_mmap_lock); 222 } 223 } 224 225 /* 226 * Close a vm structure and free it, returning the next. 227 */ 228 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 229 { 230 struct vm_area_struct *next = vma->vm_next; 231 232 might_sleep(); 233 if (vma->vm_ops && vma->vm_ops->close) 234 vma->vm_ops->close(vma); 235 if (vma->vm_file) { 236 fput(vma->vm_file); 237 if (vma->vm_flags & VM_EXECUTABLE) 238 removed_exe_file_vma(vma->vm_mm); 239 } 240 mpol_put(vma_policy(vma)); 241 kmem_cache_free(vm_area_cachep, vma); 242 return next; 243 } 244 245 SYSCALL_DEFINE1(brk, unsigned long, brk) 246 { 247 unsigned long rlim, retval; 248 unsigned long newbrk, oldbrk; 249 struct mm_struct *mm = current->mm; 250 unsigned long min_brk; 251 252 down_write(&mm->mmap_sem); 253 254 #ifdef CONFIG_COMPAT_BRK 255 min_brk = mm->end_code; 256 #else 257 min_brk = mm->start_brk; 258 #endif 259 if (brk < min_brk) 260 goto out; 261 262 /* 263 * Check against rlimit here. If this check is done later after the test 264 * of oldbrk with newbrk then it can escape the test and let the data 265 * segment grow beyond its set limit the in case where the limit is 266 * not page aligned -Ram Gupta 267 */ 268 rlim = rlimit(RLIMIT_DATA); 269 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) + 270 (mm->end_data - mm->start_data) > rlim) 271 goto out; 272 273 newbrk = PAGE_ALIGN(brk); 274 oldbrk = PAGE_ALIGN(mm->brk); 275 if (oldbrk == newbrk) 276 goto set_brk; 277 278 /* Always allow shrinking brk. */ 279 if (brk <= mm->brk) { 280 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 281 goto set_brk; 282 goto out; 283 } 284 285 /* Check against existing mmap mappings. */ 286 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 287 goto out; 288 289 /* Ok, looks good - let it rip. */ 290 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 291 goto out; 292 set_brk: 293 mm->brk = brk; 294 out: 295 retval = mm->brk; 296 up_write(&mm->mmap_sem); 297 return retval; 298 } 299 300 #ifdef DEBUG_MM_RB 301 static int browse_rb(struct rb_root *root) 302 { 303 int i = 0, j; 304 struct rb_node *nd, *pn = NULL; 305 unsigned long prev = 0, pend = 0; 306 307 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 308 struct vm_area_struct *vma; 309 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 310 if (vma->vm_start < prev) 311 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1; 312 if (vma->vm_start < pend) 313 printk("vm_start %lx pend %lx\n", vma->vm_start, pend); 314 if (vma->vm_start > vma->vm_end) 315 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start); 316 i++; 317 pn = nd; 318 prev = vma->vm_start; 319 pend = vma->vm_end; 320 } 321 j = 0; 322 for (nd = pn; nd; nd = rb_prev(nd)) { 323 j++; 324 } 325 if (i != j) 326 printk("backwards %d, forwards %d\n", j, i), i = 0; 327 return i; 328 } 329 330 void validate_mm(struct mm_struct *mm) 331 { 332 int bug = 0; 333 int i = 0; 334 struct vm_area_struct *tmp = mm->mmap; 335 while (tmp) { 336 tmp = tmp->vm_next; 337 i++; 338 } 339 if (i != mm->map_count) 340 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1; 341 i = browse_rb(&mm->mm_rb); 342 if (i != mm->map_count) 343 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1; 344 BUG_ON(bug); 345 } 346 #else 347 #define validate_mm(mm) do { } while (0) 348 #endif 349 350 static struct vm_area_struct * 351 find_vma_prepare(struct mm_struct *mm, unsigned long addr, 352 struct vm_area_struct **pprev, struct rb_node ***rb_link, 353 struct rb_node ** rb_parent) 354 { 355 struct vm_area_struct * vma; 356 struct rb_node ** __rb_link, * __rb_parent, * rb_prev; 357 358 __rb_link = &mm->mm_rb.rb_node; 359 rb_prev = __rb_parent = NULL; 360 vma = NULL; 361 362 while (*__rb_link) { 363 struct vm_area_struct *vma_tmp; 364 365 __rb_parent = *__rb_link; 366 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 367 368 if (vma_tmp->vm_end > addr) { 369 vma = vma_tmp; 370 if (vma_tmp->vm_start <= addr) 371 break; 372 __rb_link = &__rb_parent->rb_left; 373 } else { 374 rb_prev = __rb_parent; 375 __rb_link = &__rb_parent->rb_right; 376 } 377 } 378 379 *pprev = NULL; 380 if (rb_prev) 381 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 382 *rb_link = __rb_link; 383 *rb_parent = __rb_parent; 384 return vma; 385 } 386 387 static inline void 388 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma, 389 struct vm_area_struct *prev, struct rb_node *rb_parent) 390 { 391 if (prev) { 392 vma->vm_next = prev->vm_next; 393 prev->vm_next = vma; 394 } else { 395 mm->mmap = vma; 396 if (rb_parent) 397 vma->vm_next = rb_entry(rb_parent, 398 struct vm_area_struct, vm_rb); 399 else 400 vma->vm_next = NULL; 401 } 402 } 403 404 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 405 struct rb_node **rb_link, struct rb_node *rb_parent) 406 { 407 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 408 rb_insert_color(&vma->vm_rb, &mm->mm_rb); 409 } 410 411 static void __vma_link_file(struct vm_area_struct *vma) 412 { 413 struct file *file; 414 415 file = vma->vm_file; 416 if (file) { 417 struct address_space *mapping = file->f_mapping; 418 419 if (vma->vm_flags & VM_DENYWRITE) 420 atomic_dec(&file->f_path.dentry->d_inode->i_writecount); 421 if (vma->vm_flags & VM_SHARED) 422 mapping->i_mmap_writable++; 423 424 flush_dcache_mmap_lock(mapping); 425 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 426 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear); 427 else 428 vma_prio_tree_insert(vma, &mapping->i_mmap); 429 flush_dcache_mmap_unlock(mapping); 430 } 431 } 432 433 static void 434 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 435 struct vm_area_struct *prev, struct rb_node **rb_link, 436 struct rb_node *rb_parent) 437 { 438 __vma_link_list(mm, vma, prev, rb_parent); 439 __vma_link_rb(mm, vma, rb_link, rb_parent); 440 } 441 442 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 443 struct vm_area_struct *prev, struct rb_node **rb_link, 444 struct rb_node *rb_parent) 445 { 446 struct address_space *mapping = NULL; 447 448 if (vma->vm_file) 449 mapping = vma->vm_file->f_mapping; 450 451 if (mapping) { 452 spin_lock(&mapping->i_mmap_lock); 453 vma->vm_truncate_count = mapping->truncate_count; 454 } 455 anon_vma_lock(vma); 456 457 __vma_link(mm, vma, prev, rb_link, rb_parent); 458 __vma_link_file(vma); 459 460 anon_vma_unlock(vma); 461 if (mapping) 462 spin_unlock(&mapping->i_mmap_lock); 463 464 mm->map_count++; 465 validate_mm(mm); 466 } 467 468 /* 469 * Helper for vma_adjust in the split_vma insert case: 470 * insert vm structure into list and rbtree and anon_vma, 471 * but it has already been inserted into prio_tree earlier. 472 */ 473 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 474 { 475 struct vm_area_struct *__vma, *prev; 476 struct rb_node **rb_link, *rb_parent; 477 478 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent); 479 BUG_ON(__vma && __vma->vm_start < vma->vm_end); 480 __vma_link(mm, vma, prev, rb_link, rb_parent); 481 mm->map_count++; 482 } 483 484 static inline void 485 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 486 struct vm_area_struct *prev) 487 { 488 prev->vm_next = vma->vm_next; 489 rb_erase(&vma->vm_rb, &mm->mm_rb); 490 if (mm->mmap_cache == vma) 491 mm->mmap_cache = prev; 492 } 493 494 /* 495 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 496 * is already present in an i_mmap tree without adjusting the tree. 497 * The following helper function should be used when such adjustments 498 * are necessary. The "insert" vma (if any) is to be inserted 499 * before we drop the necessary locks. 500 */ 501 int vma_adjust(struct vm_area_struct *vma, unsigned long start, 502 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 503 { 504 struct mm_struct *mm = vma->vm_mm; 505 struct vm_area_struct *next = vma->vm_next; 506 struct vm_area_struct *importer = NULL; 507 struct address_space *mapping = NULL; 508 struct prio_tree_root *root = NULL; 509 struct file *file = vma->vm_file; 510 struct anon_vma *anon_vma = NULL; 511 long adjust_next = 0; 512 int remove_next = 0; 513 514 if (next && !insert) { 515 if (end >= next->vm_end) { 516 /* 517 * vma expands, overlapping all the next, and 518 * perhaps the one after too (mprotect case 6). 519 */ 520 again: remove_next = 1 + (end > next->vm_end); 521 end = next->vm_end; 522 anon_vma = next->anon_vma; 523 importer = vma; 524 } else if (end > next->vm_start) { 525 /* 526 * vma expands, overlapping part of the next: 527 * mprotect case 5 shifting the boundary up. 528 */ 529 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 530 anon_vma = next->anon_vma; 531 importer = vma; 532 } else if (end < vma->vm_end) { 533 /* 534 * vma shrinks, and !insert tells it's not 535 * split_vma inserting another: so it must be 536 * mprotect case 4 shifting the boundary down. 537 */ 538 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT); 539 anon_vma = next->anon_vma; 540 importer = next; 541 } 542 } 543 544 /* 545 * When changing only vma->vm_end, we don't really need anon_vma lock. 546 */ 547 if (vma->anon_vma && (insert || importer || start != vma->vm_start)) 548 anon_vma = vma->anon_vma; 549 if (anon_vma) { 550 /* 551 * Easily overlooked: when mprotect shifts the boundary, 552 * make sure the expanding vma has anon_vma set if the 553 * shrinking vma had, to cover any anon pages imported. 554 */ 555 if (importer && !importer->anon_vma) { 556 /* Block reverse map lookups until things are set up. */ 557 if (anon_vma_clone(importer, vma)) { 558 return -ENOMEM; 559 } 560 importer->anon_vma = anon_vma; 561 } 562 } 563 564 if (file) { 565 mapping = file->f_mapping; 566 if (!(vma->vm_flags & VM_NONLINEAR)) 567 root = &mapping->i_mmap; 568 spin_lock(&mapping->i_mmap_lock); 569 if (importer && 570 vma->vm_truncate_count != next->vm_truncate_count) { 571 /* 572 * unmap_mapping_range might be in progress: 573 * ensure that the expanding vma is rescanned. 574 */ 575 importer->vm_truncate_count = 0; 576 } 577 if (insert) { 578 insert->vm_truncate_count = vma->vm_truncate_count; 579 /* 580 * Put into prio_tree now, so instantiated pages 581 * are visible to arm/parisc __flush_dcache_page 582 * throughout; but we cannot insert into address 583 * space until vma start or end is updated. 584 */ 585 __vma_link_file(insert); 586 } 587 } 588 589 if (root) { 590 flush_dcache_mmap_lock(mapping); 591 vma_prio_tree_remove(vma, root); 592 if (adjust_next) 593 vma_prio_tree_remove(next, root); 594 } 595 596 vma->vm_start = start; 597 vma->vm_end = end; 598 vma->vm_pgoff = pgoff; 599 if (adjust_next) { 600 next->vm_start += adjust_next << PAGE_SHIFT; 601 next->vm_pgoff += adjust_next; 602 } 603 604 if (root) { 605 if (adjust_next) 606 vma_prio_tree_insert(next, root); 607 vma_prio_tree_insert(vma, root); 608 flush_dcache_mmap_unlock(mapping); 609 } 610 611 if (remove_next) { 612 /* 613 * vma_merge has merged next into vma, and needs 614 * us to remove next before dropping the locks. 615 */ 616 __vma_unlink(mm, next, vma); 617 if (file) 618 __remove_shared_vm_struct(next, file, mapping); 619 } else if (insert) { 620 /* 621 * split_vma has split insert from vma, and needs 622 * us to insert it before dropping the locks 623 * (it may either follow vma or precede it). 624 */ 625 __insert_vm_struct(mm, insert); 626 } 627 628 if (mapping) 629 spin_unlock(&mapping->i_mmap_lock); 630 631 if (remove_next) { 632 if (file) { 633 fput(file); 634 if (next->vm_flags & VM_EXECUTABLE) 635 removed_exe_file_vma(mm); 636 } 637 if (next->anon_vma) 638 anon_vma_merge(vma, next); 639 mm->map_count--; 640 mpol_put(vma_policy(next)); 641 kmem_cache_free(vm_area_cachep, next); 642 /* 643 * In mprotect's case 6 (see comments on vma_merge), 644 * we must remove another next too. It would clutter 645 * up the code too much to do both in one go. 646 */ 647 if (remove_next == 2) { 648 next = vma->vm_next; 649 goto again; 650 } 651 } 652 653 validate_mm(mm); 654 655 return 0; 656 } 657 658 /* 659 * If the vma has a ->close operation then the driver probably needs to release 660 * per-vma resources, so we don't attempt to merge those. 661 */ 662 static inline int is_mergeable_vma(struct vm_area_struct *vma, 663 struct file *file, unsigned long vm_flags) 664 { 665 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */ 666 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR) 667 return 0; 668 if (vma->vm_file != file) 669 return 0; 670 if (vma->vm_ops && vma->vm_ops->close) 671 return 0; 672 return 1; 673 } 674 675 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 676 struct anon_vma *anon_vma2) 677 { 678 return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2); 679 } 680 681 /* 682 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 683 * in front of (at a lower virtual address and file offset than) the vma. 684 * 685 * We cannot merge two vmas if they have differently assigned (non-NULL) 686 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 687 * 688 * We don't check here for the merged mmap wrapping around the end of pagecache 689 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 690 * wrap, nor mmaps which cover the final page at index -1UL. 691 */ 692 static int 693 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 694 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 695 { 696 if (is_mergeable_vma(vma, file, vm_flags) && 697 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) { 698 if (vma->vm_pgoff == vm_pgoff) 699 return 1; 700 } 701 return 0; 702 } 703 704 /* 705 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 706 * beyond (at a higher virtual address and file offset than) the vma. 707 * 708 * We cannot merge two vmas if they have differently assigned (non-NULL) 709 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 710 */ 711 static int 712 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 713 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 714 { 715 if (is_mergeable_vma(vma, file, vm_flags) && 716 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) { 717 pgoff_t vm_pglen; 718 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; 719 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 720 return 1; 721 } 722 return 0; 723 } 724 725 /* 726 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 727 * whether that can be merged with its predecessor or its successor. 728 * Or both (it neatly fills a hole). 729 * 730 * In most cases - when called for mmap, brk or mremap - [addr,end) is 731 * certain not to be mapped by the time vma_merge is called; but when 732 * called for mprotect, it is certain to be already mapped (either at 733 * an offset within prev, or at the start of next), and the flags of 734 * this area are about to be changed to vm_flags - and the no-change 735 * case has already been eliminated. 736 * 737 * The following mprotect cases have to be considered, where AAAA is 738 * the area passed down from mprotect_fixup, never extending beyond one 739 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 740 * 741 * AAAA AAAA AAAA AAAA 742 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 743 * cannot merge might become might become might become 744 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 745 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 746 * mremap move: PPPPNNNNNNNN 8 747 * AAAA 748 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 749 * might become case 1 below case 2 below case 3 below 750 * 751 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 752 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 753 */ 754 struct vm_area_struct *vma_merge(struct mm_struct *mm, 755 struct vm_area_struct *prev, unsigned long addr, 756 unsigned long end, unsigned long vm_flags, 757 struct anon_vma *anon_vma, struct file *file, 758 pgoff_t pgoff, struct mempolicy *policy) 759 { 760 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 761 struct vm_area_struct *area, *next; 762 int err; 763 764 /* 765 * We later require that vma->vm_flags == vm_flags, 766 * so this tests vma->vm_flags & VM_SPECIAL, too. 767 */ 768 if (vm_flags & VM_SPECIAL) 769 return NULL; 770 771 if (prev) 772 next = prev->vm_next; 773 else 774 next = mm->mmap; 775 area = next; 776 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 777 next = next->vm_next; 778 779 /* 780 * Can it merge with the predecessor? 781 */ 782 if (prev && prev->vm_end == addr && 783 mpol_equal(vma_policy(prev), policy) && 784 can_vma_merge_after(prev, vm_flags, 785 anon_vma, file, pgoff)) { 786 /* 787 * OK, it can. Can we now merge in the successor as well? 788 */ 789 if (next && end == next->vm_start && 790 mpol_equal(policy, vma_policy(next)) && 791 can_vma_merge_before(next, vm_flags, 792 anon_vma, file, pgoff+pglen) && 793 is_mergeable_anon_vma(prev->anon_vma, 794 next->anon_vma)) { 795 /* cases 1, 6 */ 796 err = vma_adjust(prev, prev->vm_start, 797 next->vm_end, prev->vm_pgoff, NULL); 798 } else /* cases 2, 5, 7 */ 799 err = vma_adjust(prev, prev->vm_start, 800 end, prev->vm_pgoff, NULL); 801 if (err) 802 return NULL; 803 return prev; 804 } 805 806 /* 807 * Can this new request be merged in front of next? 808 */ 809 if (next && end == next->vm_start && 810 mpol_equal(policy, vma_policy(next)) && 811 can_vma_merge_before(next, vm_flags, 812 anon_vma, file, pgoff+pglen)) { 813 if (prev && addr < prev->vm_end) /* case 4 */ 814 err = vma_adjust(prev, prev->vm_start, 815 addr, prev->vm_pgoff, NULL); 816 else /* cases 3, 8 */ 817 err = vma_adjust(area, addr, next->vm_end, 818 next->vm_pgoff - pglen, NULL); 819 if (err) 820 return NULL; 821 return area; 822 } 823 824 return NULL; 825 } 826 827 /* 828 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 829 * neighbouring vmas for a suitable anon_vma, before it goes off 830 * to allocate a new anon_vma. It checks because a repetitive 831 * sequence of mprotects and faults may otherwise lead to distinct 832 * anon_vmas being allocated, preventing vma merge in subsequent 833 * mprotect. 834 */ 835 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 836 { 837 struct vm_area_struct *near; 838 unsigned long vm_flags; 839 840 near = vma->vm_next; 841 if (!near) 842 goto try_prev; 843 844 /* 845 * Since only mprotect tries to remerge vmas, match flags 846 * which might be mprotected into each other later on. 847 * Neither mlock nor madvise tries to remerge at present, 848 * so leave their flags as obstructing a merge. 849 */ 850 vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC); 851 vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC); 852 853 if (near->anon_vma && vma->vm_end == near->vm_start && 854 mpol_equal(vma_policy(vma), vma_policy(near)) && 855 can_vma_merge_before(near, vm_flags, 856 NULL, vma->vm_file, vma->vm_pgoff + 857 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT))) 858 return near->anon_vma; 859 try_prev: 860 /* 861 * It is potentially slow to have to call find_vma_prev here. 862 * But it's only on the first write fault on the vma, not 863 * every time, and we could devise a way to avoid it later 864 * (e.g. stash info in next's anon_vma_node when assigning 865 * an anon_vma, or when trying vma_merge). Another time. 866 */ 867 BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma); 868 if (!near) 869 goto none; 870 871 vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC); 872 vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC); 873 874 if (near->anon_vma && near->vm_end == vma->vm_start && 875 mpol_equal(vma_policy(near), vma_policy(vma)) && 876 can_vma_merge_after(near, vm_flags, 877 NULL, vma->vm_file, vma->vm_pgoff)) 878 return near->anon_vma; 879 none: 880 /* 881 * There's no absolute need to look only at touching neighbours: 882 * we could search further afield for "compatible" anon_vmas. 883 * But it would probably just be a waste of time searching, 884 * or lead to too many vmas hanging off the same anon_vma. 885 * We're trying to allow mprotect remerging later on, 886 * not trying to minimize memory used for anon_vmas. 887 */ 888 return NULL; 889 } 890 891 #ifdef CONFIG_PROC_FS 892 void vm_stat_account(struct mm_struct *mm, unsigned long flags, 893 struct file *file, long pages) 894 { 895 const unsigned long stack_flags 896 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN); 897 898 if (file) { 899 mm->shared_vm += pages; 900 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC) 901 mm->exec_vm += pages; 902 } else if (flags & stack_flags) 903 mm->stack_vm += pages; 904 if (flags & (VM_RESERVED|VM_IO)) 905 mm->reserved_vm += pages; 906 } 907 #endif /* CONFIG_PROC_FS */ 908 909 /* 910 * The caller must hold down_write(¤t->mm->mmap_sem). 911 */ 912 913 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr, 914 unsigned long len, unsigned long prot, 915 unsigned long flags, unsigned long pgoff) 916 { 917 struct mm_struct * mm = current->mm; 918 struct inode *inode; 919 unsigned int vm_flags; 920 int error; 921 unsigned long reqprot = prot; 922 923 /* 924 * Does the application expect PROT_READ to imply PROT_EXEC? 925 * 926 * (the exception is when the underlying filesystem is noexec 927 * mounted, in which case we dont add PROT_EXEC.) 928 */ 929 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 930 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC))) 931 prot |= PROT_EXEC; 932 933 if (!len) 934 return -EINVAL; 935 936 if (!(flags & MAP_FIXED)) 937 addr = round_hint_to_min(addr); 938 939 /* Careful about overflows.. */ 940 len = PAGE_ALIGN(len); 941 if (!len) 942 return -ENOMEM; 943 944 /* offset overflow? */ 945 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 946 return -EOVERFLOW; 947 948 /* Too many mappings? */ 949 if (mm->map_count > sysctl_max_map_count) 950 return -ENOMEM; 951 952 /* Obtain the address to map to. we verify (or select) it and ensure 953 * that it represents a valid section of the address space. 954 */ 955 addr = get_unmapped_area(file, addr, len, pgoff, flags); 956 if (addr & ~PAGE_MASK) 957 return addr; 958 959 /* Do simple checking here so the lower-level routines won't have 960 * to. we assume access permissions have been handled by the open 961 * of the memory object, so we don't do any here. 962 */ 963 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 964 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 965 966 if (flags & MAP_LOCKED) 967 if (!can_do_mlock()) 968 return -EPERM; 969 970 /* mlock MCL_FUTURE? */ 971 if (vm_flags & VM_LOCKED) { 972 unsigned long locked, lock_limit; 973 locked = len >> PAGE_SHIFT; 974 locked += mm->locked_vm; 975 lock_limit = rlimit(RLIMIT_MEMLOCK); 976 lock_limit >>= PAGE_SHIFT; 977 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 978 return -EAGAIN; 979 } 980 981 inode = file ? file->f_path.dentry->d_inode : NULL; 982 983 if (file) { 984 switch (flags & MAP_TYPE) { 985 case MAP_SHARED: 986 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 987 return -EACCES; 988 989 /* 990 * Make sure we don't allow writing to an append-only 991 * file.. 992 */ 993 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 994 return -EACCES; 995 996 /* 997 * Make sure there are no mandatory locks on the file. 998 */ 999 if (locks_verify_locked(inode)) 1000 return -EAGAIN; 1001 1002 vm_flags |= VM_SHARED | VM_MAYSHARE; 1003 if (!(file->f_mode & FMODE_WRITE)) 1004 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1005 1006 /* fall through */ 1007 case MAP_PRIVATE: 1008 if (!(file->f_mode & FMODE_READ)) 1009 return -EACCES; 1010 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) { 1011 if (vm_flags & VM_EXEC) 1012 return -EPERM; 1013 vm_flags &= ~VM_MAYEXEC; 1014 } 1015 1016 if (!file->f_op || !file->f_op->mmap) 1017 return -ENODEV; 1018 break; 1019 1020 default: 1021 return -EINVAL; 1022 } 1023 } else { 1024 switch (flags & MAP_TYPE) { 1025 case MAP_SHARED: 1026 /* 1027 * Ignore pgoff. 1028 */ 1029 pgoff = 0; 1030 vm_flags |= VM_SHARED | VM_MAYSHARE; 1031 break; 1032 case MAP_PRIVATE: 1033 /* 1034 * Set pgoff according to addr for anon_vma. 1035 */ 1036 pgoff = addr >> PAGE_SHIFT; 1037 break; 1038 default: 1039 return -EINVAL; 1040 } 1041 } 1042 1043 error = security_file_mmap(file, reqprot, prot, flags, addr, 0); 1044 if (error) 1045 return error; 1046 1047 return mmap_region(file, addr, len, flags, vm_flags, pgoff); 1048 } 1049 EXPORT_SYMBOL(do_mmap_pgoff); 1050 1051 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1052 unsigned long, prot, unsigned long, flags, 1053 unsigned long, fd, unsigned long, pgoff) 1054 { 1055 struct file *file = NULL; 1056 unsigned long retval = -EBADF; 1057 1058 if (!(flags & MAP_ANONYMOUS)) { 1059 if (unlikely(flags & MAP_HUGETLB)) 1060 return -EINVAL; 1061 file = fget(fd); 1062 if (!file) 1063 goto out; 1064 } else if (flags & MAP_HUGETLB) { 1065 struct user_struct *user = NULL; 1066 /* 1067 * VM_NORESERVE is used because the reservations will be 1068 * taken when vm_ops->mmap() is called 1069 * A dummy user value is used because we are not locking 1070 * memory so no accounting is necessary 1071 */ 1072 len = ALIGN(len, huge_page_size(&default_hstate)); 1073 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE, 1074 &user, HUGETLB_ANONHUGE_INODE); 1075 if (IS_ERR(file)) 1076 return PTR_ERR(file); 1077 } 1078 1079 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1080 1081 down_write(¤t->mm->mmap_sem); 1082 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1083 up_write(¤t->mm->mmap_sem); 1084 1085 if (file) 1086 fput(file); 1087 out: 1088 return retval; 1089 } 1090 1091 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1092 struct mmap_arg_struct { 1093 unsigned long addr; 1094 unsigned long len; 1095 unsigned long prot; 1096 unsigned long flags; 1097 unsigned long fd; 1098 unsigned long offset; 1099 }; 1100 1101 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1102 { 1103 struct mmap_arg_struct a; 1104 1105 if (copy_from_user(&a, arg, sizeof(a))) 1106 return -EFAULT; 1107 if (a.offset & ~PAGE_MASK) 1108 return -EINVAL; 1109 1110 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1111 a.offset >> PAGE_SHIFT); 1112 } 1113 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1114 1115 /* 1116 * Some shared mappigns will want the pages marked read-only 1117 * to track write events. If so, we'll downgrade vm_page_prot 1118 * to the private version (using protection_map[] without the 1119 * VM_SHARED bit). 1120 */ 1121 int vma_wants_writenotify(struct vm_area_struct *vma) 1122 { 1123 unsigned int vm_flags = vma->vm_flags; 1124 1125 /* If it was private or non-writable, the write bit is already clear */ 1126 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1127 return 0; 1128 1129 /* The backer wishes to know when pages are first written to? */ 1130 if (vma->vm_ops && vma->vm_ops->page_mkwrite) 1131 return 1; 1132 1133 /* The open routine did something to the protections already? */ 1134 if (pgprot_val(vma->vm_page_prot) != 1135 pgprot_val(vm_get_page_prot(vm_flags))) 1136 return 0; 1137 1138 /* Specialty mapping? */ 1139 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE)) 1140 return 0; 1141 1142 /* Can the mapping track the dirty pages? */ 1143 return vma->vm_file && vma->vm_file->f_mapping && 1144 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1145 } 1146 1147 /* 1148 * We account for memory if it's a private writeable mapping, 1149 * not hugepages and VM_NORESERVE wasn't set. 1150 */ 1151 static inline int accountable_mapping(struct file *file, unsigned int vm_flags) 1152 { 1153 /* 1154 * hugetlb has its own accounting separate from the core VM 1155 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1156 */ 1157 if (file && is_file_hugepages(file)) 1158 return 0; 1159 1160 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1161 } 1162 1163 unsigned long mmap_region(struct file *file, unsigned long addr, 1164 unsigned long len, unsigned long flags, 1165 unsigned int vm_flags, unsigned long pgoff) 1166 { 1167 struct mm_struct *mm = current->mm; 1168 struct vm_area_struct *vma, *prev; 1169 int correct_wcount = 0; 1170 int error; 1171 struct rb_node **rb_link, *rb_parent; 1172 unsigned long charged = 0; 1173 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL; 1174 1175 /* Clear old maps */ 1176 error = -ENOMEM; 1177 munmap_back: 1178 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 1179 if (vma && vma->vm_start < addr + len) { 1180 if (do_munmap(mm, addr, len)) 1181 return -ENOMEM; 1182 goto munmap_back; 1183 } 1184 1185 /* Check against address space limit. */ 1186 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 1187 return -ENOMEM; 1188 1189 /* 1190 * Set 'VM_NORESERVE' if we should not account for the 1191 * memory use of this mapping. 1192 */ 1193 if ((flags & MAP_NORESERVE)) { 1194 /* We honor MAP_NORESERVE if allowed to overcommit */ 1195 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1196 vm_flags |= VM_NORESERVE; 1197 1198 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1199 if (file && is_file_hugepages(file)) 1200 vm_flags |= VM_NORESERVE; 1201 } 1202 1203 /* 1204 * Private writable mapping: check memory availability 1205 */ 1206 if (accountable_mapping(file, vm_flags)) { 1207 charged = len >> PAGE_SHIFT; 1208 if (security_vm_enough_memory(charged)) 1209 return -ENOMEM; 1210 vm_flags |= VM_ACCOUNT; 1211 } 1212 1213 /* 1214 * Can we just expand an old mapping? 1215 */ 1216 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL); 1217 if (vma) 1218 goto out; 1219 1220 /* 1221 * Determine the object being mapped and call the appropriate 1222 * specific mapper. the address has already been validated, but 1223 * not unmapped, but the maps are removed from the list. 1224 */ 1225 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1226 if (!vma) { 1227 error = -ENOMEM; 1228 goto unacct_error; 1229 } 1230 1231 vma->vm_mm = mm; 1232 vma->vm_start = addr; 1233 vma->vm_end = addr + len; 1234 vma->vm_flags = vm_flags; 1235 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1236 vma->vm_pgoff = pgoff; 1237 INIT_LIST_HEAD(&vma->anon_vma_chain); 1238 1239 if (file) { 1240 error = -EINVAL; 1241 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1242 goto free_vma; 1243 if (vm_flags & VM_DENYWRITE) { 1244 error = deny_write_access(file); 1245 if (error) 1246 goto free_vma; 1247 correct_wcount = 1; 1248 } 1249 vma->vm_file = file; 1250 get_file(file); 1251 error = file->f_op->mmap(file, vma); 1252 if (error) 1253 goto unmap_and_free_vma; 1254 if (vm_flags & VM_EXECUTABLE) 1255 added_exe_file_vma(mm); 1256 1257 /* Can addr have changed?? 1258 * 1259 * Answer: Yes, several device drivers can do it in their 1260 * f_op->mmap method. -DaveM 1261 */ 1262 addr = vma->vm_start; 1263 pgoff = vma->vm_pgoff; 1264 vm_flags = vma->vm_flags; 1265 } else if (vm_flags & VM_SHARED) { 1266 error = shmem_zero_setup(vma); 1267 if (error) 1268 goto free_vma; 1269 } 1270 1271 if (vma_wants_writenotify(vma)) { 1272 pgprot_t pprot = vma->vm_page_prot; 1273 1274 /* Can vma->vm_page_prot have changed?? 1275 * 1276 * Answer: Yes, drivers may have changed it in their 1277 * f_op->mmap method. 1278 * 1279 * Ensures that vmas marked as uncached stay that way. 1280 */ 1281 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED); 1282 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot))) 1283 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 1284 } 1285 1286 vma_link(mm, vma, prev, rb_link, rb_parent); 1287 file = vma->vm_file; 1288 1289 /* Once vma denies write, undo our temporary denial count */ 1290 if (correct_wcount) 1291 atomic_inc(&inode->i_writecount); 1292 out: 1293 perf_event_mmap(vma); 1294 1295 mm->total_vm += len >> PAGE_SHIFT; 1296 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT); 1297 if (vm_flags & VM_LOCKED) { 1298 if (!mlock_vma_pages_range(vma, addr, addr + len)) 1299 mm->locked_vm += (len >> PAGE_SHIFT); 1300 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK)) 1301 make_pages_present(addr, addr + len); 1302 return addr; 1303 1304 unmap_and_free_vma: 1305 if (correct_wcount) 1306 atomic_inc(&inode->i_writecount); 1307 vma->vm_file = NULL; 1308 fput(file); 1309 1310 /* Undo any partial mapping done by a device driver. */ 1311 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1312 charged = 0; 1313 free_vma: 1314 kmem_cache_free(vm_area_cachep, vma); 1315 unacct_error: 1316 if (charged) 1317 vm_unacct_memory(charged); 1318 return error; 1319 } 1320 1321 /* Get an address range which is currently unmapped. 1322 * For shmat() with addr=0. 1323 * 1324 * Ugly calling convention alert: 1325 * Return value with the low bits set means error value, 1326 * ie 1327 * if (ret & ~PAGE_MASK) 1328 * error = ret; 1329 * 1330 * This function "knows" that -ENOMEM has the bits set. 1331 */ 1332 #ifndef HAVE_ARCH_UNMAPPED_AREA 1333 unsigned long 1334 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1335 unsigned long len, unsigned long pgoff, unsigned long flags) 1336 { 1337 struct mm_struct *mm = current->mm; 1338 struct vm_area_struct *vma; 1339 unsigned long start_addr; 1340 1341 if (len > TASK_SIZE) 1342 return -ENOMEM; 1343 1344 if (flags & MAP_FIXED) 1345 return addr; 1346 1347 if (addr) { 1348 addr = PAGE_ALIGN(addr); 1349 vma = find_vma(mm, addr); 1350 if (TASK_SIZE - len >= addr && 1351 (!vma || addr + len <= vma->vm_start)) 1352 return addr; 1353 } 1354 if (len > mm->cached_hole_size) { 1355 start_addr = addr = mm->free_area_cache; 1356 } else { 1357 start_addr = addr = TASK_UNMAPPED_BASE; 1358 mm->cached_hole_size = 0; 1359 } 1360 1361 full_search: 1362 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1363 /* At this point: (!vma || addr < vma->vm_end). */ 1364 if (TASK_SIZE - len < addr) { 1365 /* 1366 * Start a new search - just in case we missed 1367 * some holes. 1368 */ 1369 if (start_addr != TASK_UNMAPPED_BASE) { 1370 addr = TASK_UNMAPPED_BASE; 1371 start_addr = addr; 1372 mm->cached_hole_size = 0; 1373 goto full_search; 1374 } 1375 return -ENOMEM; 1376 } 1377 if (!vma || addr + len <= vma->vm_start) { 1378 /* 1379 * Remember the place where we stopped the search: 1380 */ 1381 mm->free_area_cache = addr + len; 1382 return addr; 1383 } 1384 if (addr + mm->cached_hole_size < vma->vm_start) 1385 mm->cached_hole_size = vma->vm_start - addr; 1386 addr = vma->vm_end; 1387 } 1388 } 1389 #endif 1390 1391 void arch_unmap_area(struct mm_struct *mm, unsigned long addr) 1392 { 1393 /* 1394 * Is this a new hole at the lowest possible address? 1395 */ 1396 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) { 1397 mm->free_area_cache = addr; 1398 mm->cached_hole_size = ~0UL; 1399 } 1400 } 1401 1402 /* 1403 * This mmap-allocator allocates new areas top-down from below the 1404 * stack's low limit (the base): 1405 */ 1406 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1407 unsigned long 1408 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1409 const unsigned long len, const unsigned long pgoff, 1410 const unsigned long flags) 1411 { 1412 struct vm_area_struct *vma; 1413 struct mm_struct *mm = current->mm; 1414 unsigned long addr = addr0; 1415 1416 /* requested length too big for entire address space */ 1417 if (len > TASK_SIZE) 1418 return -ENOMEM; 1419 1420 if (flags & MAP_FIXED) 1421 return addr; 1422 1423 /* requesting a specific address */ 1424 if (addr) { 1425 addr = PAGE_ALIGN(addr); 1426 vma = find_vma(mm, addr); 1427 if (TASK_SIZE - len >= addr && 1428 (!vma || addr + len <= vma->vm_start)) 1429 return addr; 1430 } 1431 1432 /* check if free_area_cache is useful for us */ 1433 if (len <= mm->cached_hole_size) { 1434 mm->cached_hole_size = 0; 1435 mm->free_area_cache = mm->mmap_base; 1436 } 1437 1438 /* either no address requested or can't fit in requested address hole */ 1439 addr = mm->free_area_cache; 1440 1441 /* make sure it can fit in the remaining address space */ 1442 if (addr > len) { 1443 vma = find_vma(mm, addr-len); 1444 if (!vma || addr <= vma->vm_start) 1445 /* remember the address as a hint for next time */ 1446 return (mm->free_area_cache = addr-len); 1447 } 1448 1449 if (mm->mmap_base < len) 1450 goto bottomup; 1451 1452 addr = mm->mmap_base-len; 1453 1454 do { 1455 /* 1456 * Lookup failure means no vma is above this address, 1457 * else if new region fits below vma->vm_start, 1458 * return with success: 1459 */ 1460 vma = find_vma(mm, addr); 1461 if (!vma || addr+len <= vma->vm_start) 1462 /* remember the address as a hint for next time */ 1463 return (mm->free_area_cache = addr); 1464 1465 /* remember the largest hole we saw so far */ 1466 if (addr + mm->cached_hole_size < vma->vm_start) 1467 mm->cached_hole_size = vma->vm_start - addr; 1468 1469 /* try just below the current vma->vm_start */ 1470 addr = vma->vm_start-len; 1471 } while (len < vma->vm_start); 1472 1473 bottomup: 1474 /* 1475 * A failed mmap() very likely causes application failure, 1476 * so fall back to the bottom-up function here. This scenario 1477 * can happen with large stack limits and large mmap() 1478 * allocations. 1479 */ 1480 mm->cached_hole_size = ~0UL; 1481 mm->free_area_cache = TASK_UNMAPPED_BASE; 1482 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags); 1483 /* 1484 * Restore the topdown base: 1485 */ 1486 mm->free_area_cache = mm->mmap_base; 1487 mm->cached_hole_size = ~0UL; 1488 1489 return addr; 1490 } 1491 #endif 1492 1493 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr) 1494 { 1495 /* 1496 * Is this a new hole at the highest possible address? 1497 */ 1498 if (addr > mm->free_area_cache) 1499 mm->free_area_cache = addr; 1500 1501 /* dont allow allocations above current base */ 1502 if (mm->free_area_cache > mm->mmap_base) 1503 mm->free_area_cache = mm->mmap_base; 1504 } 1505 1506 unsigned long 1507 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1508 unsigned long pgoff, unsigned long flags) 1509 { 1510 unsigned long (*get_area)(struct file *, unsigned long, 1511 unsigned long, unsigned long, unsigned long); 1512 1513 unsigned long error = arch_mmap_check(addr, len, flags); 1514 if (error) 1515 return error; 1516 1517 /* Careful about overflows.. */ 1518 if (len > TASK_SIZE) 1519 return -ENOMEM; 1520 1521 get_area = current->mm->get_unmapped_area; 1522 if (file && file->f_op && file->f_op->get_unmapped_area) 1523 get_area = file->f_op->get_unmapped_area; 1524 addr = get_area(file, addr, len, pgoff, flags); 1525 if (IS_ERR_VALUE(addr)) 1526 return addr; 1527 1528 if (addr > TASK_SIZE - len) 1529 return -ENOMEM; 1530 if (addr & ~PAGE_MASK) 1531 return -EINVAL; 1532 1533 return arch_rebalance_pgtables(addr, len); 1534 } 1535 1536 EXPORT_SYMBOL(get_unmapped_area); 1537 1538 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 1539 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 1540 { 1541 struct vm_area_struct *vma = NULL; 1542 1543 if (mm) { 1544 /* Check the cache first. */ 1545 /* (Cache hit rate is typically around 35%.) */ 1546 vma = mm->mmap_cache; 1547 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) { 1548 struct rb_node * rb_node; 1549 1550 rb_node = mm->mm_rb.rb_node; 1551 vma = NULL; 1552 1553 while (rb_node) { 1554 struct vm_area_struct * vma_tmp; 1555 1556 vma_tmp = rb_entry(rb_node, 1557 struct vm_area_struct, vm_rb); 1558 1559 if (vma_tmp->vm_end > addr) { 1560 vma = vma_tmp; 1561 if (vma_tmp->vm_start <= addr) 1562 break; 1563 rb_node = rb_node->rb_left; 1564 } else 1565 rb_node = rb_node->rb_right; 1566 } 1567 if (vma) 1568 mm->mmap_cache = vma; 1569 } 1570 } 1571 return vma; 1572 } 1573 1574 EXPORT_SYMBOL(find_vma); 1575 1576 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */ 1577 struct vm_area_struct * 1578 find_vma_prev(struct mm_struct *mm, unsigned long addr, 1579 struct vm_area_struct **pprev) 1580 { 1581 struct vm_area_struct *vma = NULL, *prev = NULL; 1582 struct rb_node *rb_node; 1583 if (!mm) 1584 goto out; 1585 1586 /* Guard against addr being lower than the first VMA */ 1587 vma = mm->mmap; 1588 1589 /* Go through the RB tree quickly. */ 1590 rb_node = mm->mm_rb.rb_node; 1591 1592 while (rb_node) { 1593 struct vm_area_struct *vma_tmp; 1594 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 1595 1596 if (addr < vma_tmp->vm_end) { 1597 rb_node = rb_node->rb_left; 1598 } else { 1599 prev = vma_tmp; 1600 if (!prev->vm_next || (addr < prev->vm_next->vm_end)) 1601 break; 1602 rb_node = rb_node->rb_right; 1603 } 1604 } 1605 1606 out: 1607 *pprev = prev; 1608 return prev ? prev->vm_next : vma; 1609 } 1610 1611 /* 1612 * Verify that the stack growth is acceptable and 1613 * update accounting. This is shared with both the 1614 * grow-up and grow-down cases. 1615 */ 1616 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 1617 { 1618 struct mm_struct *mm = vma->vm_mm; 1619 struct rlimit *rlim = current->signal->rlim; 1620 unsigned long new_start; 1621 1622 /* address space limit tests */ 1623 if (!may_expand_vm(mm, grow)) 1624 return -ENOMEM; 1625 1626 /* Stack limit test */ 1627 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 1628 return -ENOMEM; 1629 1630 /* mlock limit tests */ 1631 if (vma->vm_flags & VM_LOCKED) { 1632 unsigned long locked; 1633 unsigned long limit; 1634 locked = mm->locked_vm + grow; 1635 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 1636 limit >>= PAGE_SHIFT; 1637 if (locked > limit && !capable(CAP_IPC_LOCK)) 1638 return -ENOMEM; 1639 } 1640 1641 /* Check to ensure the stack will not grow into a hugetlb-only region */ 1642 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 1643 vma->vm_end - size; 1644 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 1645 return -EFAULT; 1646 1647 /* 1648 * Overcommit.. This must be the final test, as it will 1649 * update security statistics. 1650 */ 1651 if (security_vm_enough_memory_mm(mm, grow)) 1652 return -ENOMEM; 1653 1654 /* Ok, everything looks good - let it rip */ 1655 mm->total_vm += grow; 1656 if (vma->vm_flags & VM_LOCKED) 1657 mm->locked_vm += grow; 1658 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow); 1659 return 0; 1660 } 1661 1662 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 1663 /* 1664 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 1665 * vma is the last one with address > vma->vm_end. Have to extend vma. 1666 */ 1667 #ifndef CONFIG_IA64 1668 static 1669 #endif 1670 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 1671 { 1672 int error; 1673 1674 if (!(vma->vm_flags & VM_GROWSUP)) 1675 return -EFAULT; 1676 1677 /* 1678 * We must make sure the anon_vma is allocated 1679 * so that the anon_vma locking is not a noop. 1680 */ 1681 if (unlikely(anon_vma_prepare(vma))) 1682 return -ENOMEM; 1683 anon_vma_lock(vma); 1684 1685 /* 1686 * vma->vm_start/vm_end cannot change under us because the caller 1687 * is required to hold the mmap_sem in read mode. We need the 1688 * anon_vma lock to serialize against concurrent expand_stacks. 1689 * Also guard against wrapping around to address 0. 1690 */ 1691 if (address < PAGE_ALIGN(address+4)) 1692 address = PAGE_ALIGN(address+4); 1693 else { 1694 anon_vma_unlock(vma); 1695 return -ENOMEM; 1696 } 1697 error = 0; 1698 1699 /* Somebody else might have raced and expanded it already */ 1700 if (address > vma->vm_end) { 1701 unsigned long size, grow; 1702 1703 size = address - vma->vm_start; 1704 grow = (address - vma->vm_end) >> PAGE_SHIFT; 1705 1706 error = acct_stack_growth(vma, size, grow); 1707 if (!error) 1708 vma->vm_end = address; 1709 } 1710 anon_vma_unlock(vma); 1711 return error; 1712 } 1713 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 1714 1715 /* 1716 * vma is the first one with address < vma->vm_start. Have to extend vma. 1717 */ 1718 static int expand_downwards(struct vm_area_struct *vma, 1719 unsigned long address) 1720 { 1721 int error; 1722 1723 /* 1724 * We must make sure the anon_vma is allocated 1725 * so that the anon_vma locking is not a noop. 1726 */ 1727 if (unlikely(anon_vma_prepare(vma))) 1728 return -ENOMEM; 1729 1730 address &= PAGE_MASK; 1731 error = security_file_mmap(NULL, 0, 0, 0, address, 1); 1732 if (error) 1733 return error; 1734 1735 anon_vma_lock(vma); 1736 1737 /* 1738 * vma->vm_start/vm_end cannot change under us because the caller 1739 * is required to hold the mmap_sem in read mode. We need the 1740 * anon_vma lock to serialize against concurrent expand_stacks. 1741 */ 1742 1743 /* Somebody else might have raced and expanded it already */ 1744 if (address < vma->vm_start) { 1745 unsigned long size, grow; 1746 1747 size = vma->vm_end - address; 1748 grow = (vma->vm_start - address) >> PAGE_SHIFT; 1749 1750 error = acct_stack_growth(vma, size, grow); 1751 if (!error) { 1752 vma->vm_start = address; 1753 vma->vm_pgoff -= grow; 1754 } 1755 } 1756 anon_vma_unlock(vma); 1757 return error; 1758 } 1759 1760 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address) 1761 { 1762 return expand_downwards(vma, address); 1763 } 1764 1765 #ifdef CONFIG_STACK_GROWSUP 1766 int expand_stack(struct vm_area_struct *vma, unsigned long address) 1767 { 1768 return expand_upwards(vma, address); 1769 } 1770 1771 struct vm_area_struct * 1772 find_extend_vma(struct mm_struct *mm, unsigned long addr) 1773 { 1774 struct vm_area_struct *vma, *prev; 1775 1776 addr &= PAGE_MASK; 1777 vma = find_vma_prev(mm, addr, &prev); 1778 if (vma && (vma->vm_start <= addr)) 1779 return vma; 1780 if (!prev || expand_stack(prev, addr)) 1781 return NULL; 1782 if (prev->vm_flags & VM_LOCKED) { 1783 mlock_vma_pages_range(prev, addr, prev->vm_end); 1784 } 1785 return prev; 1786 } 1787 #else 1788 int expand_stack(struct vm_area_struct *vma, unsigned long address) 1789 { 1790 return expand_downwards(vma, address); 1791 } 1792 1793 struct vm_area_struct * 1794 find_extend_vma(struct mm_struct * mm, unsigned long addr) 1795 { 1796 struct vm_area_struct * vma; 1797 unsigned long start; 1798 1799 addr &= PAGE_MASK; 1800 vma = find_vma(mm,addr); 1801 if (!vma) 1802 return NULL; 1803 if (vma->vm_start <= addr) 1804 return vma; 1805 if (!(vma->vm_flags & VM_GROWSDOWN)) 1806 return NULL; 1807 start = vma->vm_start; 1808 if (expand_stack(vma, addr)) 1809 return NULL; 1810 if (vma->vm_flags & VM_LOCKED) { 1811 mlock_vma_pages_range(vma, addr, start); 1812 } 1813 return vma; 1814 } 1815 #endif 1816 1817 /* 1818 * Ok - we have the memory areas we should free on the vma list, 1819 * so release them, and do the vma updates. 1820 * 1821 * Called with the mm semaphore held. 1822 */ 1823 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 1824 { 1825 /* Update high watermark before we lower total_vm */ 1826 update_hiwater_vm(mm); 1827 do { 1828 long nrpages = vma_pages(vma); 1829 1830 mm->total_vm -= nrpages; 1831 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages); 1832 vma = remove_vma(vma); 1833 } while (vma); 1834 validate_mm(mm); 1835 } 1836 1837 /* 1838 * Get rid of page table information in the indicated region. 1839 * 1840 * Called with the mm semaphore held. 1841 */ 1842 static void unmap_region(struct mm_struct *mm, 1843 struct vm_area_struct *vma, struct vm_area_struct *prev, 1844 unsigned long start, unsigned long end) 1845 { 1846 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap; 1847 struct mmu_gather *tlb; 1848 unsigned long nr_accounted = 0; 1849 1850 lru_add_drain(); 1851 tlb = tlb_gather_mmu(mm, 0); 1852 update_hiwater_rss(mm); 1853 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL); 1854 vm_unacct_memory(nr_accounted); 1855 free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS, 1856 next? next->vm_start: 0); 1857 tlb_finish_mmu(tlb, start, end); 1858 } 1859 1860 /* 1861 * Create a list of vma's touched by the unmap, removing them from the mm's 1862 * vma list as we go.. 1863 */ 1864 static void 1865 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 1866 struct vm_area_struct *prev, unsigned long end) 1867 { 1868 struct vm_area_struct **insertion_point; 1869 struct vm_area_struct *tail_vma = NULL; 1870 unsigned long addr; 1871 1872 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 1873 do { 1874 rb_erase(&vma->vm_rb, &mm->mm_rb); 1875 mm->map_count--; 1876 tail_vma = vma; 1877 vma = vma->vm_next; 1878 } while (vma && vma->vm_start < end); 1879 *insertion_point = vma; 1880 tail_vma->vm_next = NULL; 1881 if (mm->unmap_area == arch_unmap_area) 1882 addr = prev ? prev->vm_end : mm->mmap_base; 1883 else 1884 addr = vma ? vma->vm_start : mm->mmap_base; 1885 mm->unmap_area(mm, addr); 1886 mm->mmap_cache = NULL; /* Kill the cache. */ 1887 } 1888 1889 /* 1890 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the 1891 * munmap path where it doesn't make sense to fail. 1892 */ 1893 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma, 1894 unsigned long addr, int new_below) 1895 { 1896 struct mempolicy *pol; 1897 struct vm_area_struct *new; 1898 int err = -ENOMEM; 1899 1900 if (is_vm_hugetlb_page(vma) && (addr & 1901 ~(huge_page_mask(hstate_vma(vma))))) 1902 return -EINVAL; 1903 1904 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 1905 if (!new) 1906 goto out_err; 1907 1908 /* most fields are the same, copy all, and then fixup */ 1909 *new = *vma; 1910 1911 INIT_LIST_HEAD(&new->anon_vma_chain); 1912 1913 if (new_below) 1914 new->vm_end = addr; 1915 else { 1916 new->vm_start = addr; 1917 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 1918 } 1919 1920 pol = mpol_dup(vma_policy(vma)); 1921 if (IS_ERR(pol)) { 1922 err = PTR_ERR(pol); 1923 goto out_free_vma; 1924 } 1925 vma_set_policy(new, pol); 1926 1927 if (anon_vma_clone(new, vma)) 1928 goto out_free_mpol; 1929 1930 if (new->vm_file) { 1931 get_file(new->vm_file); 1932 if (vma->vm_flags & VM_EXECUTABLE) 1933 added_exe_file_vma(mm); 1934 } 1935 1936 if (new->vm_ops && new->vm_ops->open) 1937 new->vm_ops->open(new); 1938 1939 if (new_below) 1940 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 1941 ((addr - new->vm_start) >> PAGE_SHIFT), new); 1942 else 1943 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 1944 1945 /* Success. */ 1946 if (!err) 1947 return 0; 1948 1949 /* Clean everything up if vma_adjust failed. */ 1950 new->vm_ops->close(new); 1951 if (new->vm_file) { 1952 if (vma->vm_flags & VM_EXECUTABLE) 1953 removed_exe_file_vma(mm); 1954 fput(new->vm_file); 1955 } 1956 out_free_mpol: 1957 mpol_put(pol); 1958 out_free_vma: 1959 kmem_cache_free(vm_area_cachep, new); 1960 out_err: 1961 return err; 1962 } 1963 1964 /* 1965 * Split a vma into two pieces at address 'addr', a new vma is allocated 1966 * either for the first part or the tail. 1967 */ 1968 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 1969 unsigned long addr, int new_below) 1970 { 1971 if (mm->map_count >= sysctl_max_map_count) 1972 return -ENOMEM; 1973 1974 return __split_vma(mm, vma, addr, new_below); 1975 } 1976 1977 /* Munmap is split into 2 main parts -- this part which finds 1978 * what needs doing, and the areas themselves, which do the 1979 * work. This now handles partial unmappings. 1980 * Jeremy Fitzhardinge <jeremy@goop.org> 1981 */ 1982 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 1983 { 1984 unsigned long end; 1985 struct vm_area_struct *vma, *prev, *last; 1986 1987 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start) 1988 return -EINVAL; 1989 1990 if ((len = PAGE_ALIGN(len)) == 0) 1991 return -EINVAL; 1992 1993 /* Find the first overlapping VMA */ 1994 vma = find_vma_prev(mm, start, &prev); 1995 if (!vma) 1996 return 0; 1997 /* we have start < vma->vm_end */ 1998 1999 /* if it doesn't overlap, we have nothing.. */ 2000 end = start + len; 2001 if (vma->vm_start >= end) 2002 return 0; 2003 2004 /* 2005 * If we need to split any vma, do it now to save pain later. 2006 * 2007 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2008 * unmapped vm_area_struct will remain in use: so lower split_vma 2009 * places tmp vma above, and higher split_vma places tmp vma below. 2010 */ 2011 if (start > vma->vm_start) { 2012 int error; 2013 2014 /* 2015 * Make sure that map_count on return from munmap() will 2016 * not exceed its limit; but let map_count go just above 2017 * its limit temporarily, to help free resources as expected. 2018 */ 2019 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2020 return -ENOMEM; 2021 2022 error = __split_vma(mm, vma, start, 0); 2023 if (error) 2024 return error; 2025 prev = vma; 2026 } 2027 2028 /* Does it split the last one? */ 2029 last = find_vma(mm, end); 2030 if (last && end > last->vm_start) { 2031 int error = __split_vma(mm, last, end, 1); 2032 if (error) 2033 return error; 2034 } 2035 vma = prev? prev->vm_next: mm->mmap; 2036 2037 /* 2038 * unlock any mlock()ed ranges before detaching vmas 2039 */ 2040 if (mm->locked_vm) { 2041 struct vm_area_struct *tmp = vma; 2042 while (tmp && tmp->vm_start < end) { 2043 if (tmp->vm_flags & VM_LOCKED) { 2044 mm->locked_vm -= vma_pages(tmp); 2045 munlock_vma_pages_all(tmp); 2046 } 2047 tmp = tmp->vm_next; 2048 } 2049 } 2050 2051 /* 2052 * Remove the vma's, and unmap the actual pages 2053 */ 2054 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2055 unmap_region(mm, vma, prev, start, end); 2056 2057 /* Fix up all other VM information */ 2058 remove_vma_list(mm, vma); 2059 2060 return 0; 2061 } 2062 2063 EXPORT_SYMBOL(do_munmap); 2064 2065 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2066 { 2067 int ret; 2068 struct mm_struct *mm = current->mm; 2069 2070 profile_munmap(addr); 2071 2072 down_write(&mm->mmap_sem); 2073 ret = do_munmap(mm, addr, len); 2074 up_write(&mm->mmap_sem); 2075 return ret; 2076 } 2077 2078 static inline void verify_mm_writelocked(struct mm_struct *mm) 2079 { 2080 #ifdef CONFIG_DEBUG_VM 2081 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2082 WARN_ON(1); 2083 up_read(&mm->mmap_sem); 2084 } 2085 #endif 2086 } 2087 2088 /* 2089 * this is really a simplified "do_mmap". it only handles 2090 * anonymous maps. eventually we may be able to do some 2091 * brk-specific accounting here. 2092 */ 2093 unsigned long do_brk(unsigned long addr, unsigned long len) 2094 { 2095 struct mm_struct * mm = current->mm; 2096 struct vm_area_struct * vma, * prev; 2097 unsigned long flags; 2098 struct rb_node ** rb_link, * rb_parent; 2099 pgoff_t pgoff = addr >> PAGE_SHIFT; 2100 int error; 2101 2102 len = PAGE_ALIGN(len); 2103 if (!len) 2104 return addr; 2105 2106 error = security_file_mmap(NULL, 0, 0, 0, addr, 1); 2107 if (error) 2108 return error; 2109 2110 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2111 2112 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2113 if (error & ~PAGE_MASK) 2114 return error; 2115 2116 /* 2117 * mlock MCL_FUTURE? 2118 */ 2119 if (mm->def_flags & VM_LOCKED) { 2120 unsigned long locked, lock_limit; 2121 locked = len >> PAGE_SHIFT; 2122 locked += mm->locked_vm; 2123 lock_limit = rlimit(RLIMIT_MEMLOCK); 2124 lock_limit >>= PAGE_SHIFT; 2125 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 2126 return -EAGAIN; 2127 } 2128 2129 /* 2130 * mm->mmap_sem is required to protect against another thread 2131 * changing the mappings in case we sleep. 2132 */ 2133 verify_mm_writelocked(mm); 2134 2135 /* 2136 * Clear old maps. this also does some error checking for us 2137 */ 2138 munmap_back: 2139 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 2140 if (vma && vma->vm_start < addr + len) { 2141 if (do_munmap(mm, addr, len)) 2142 return -ENOMEM; 2143 goto munmap_back; 2144 } 2145 2146 /* Check against address space limits *after* clearing old maps... */ 2147 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 2148 return -ENOMEM; 2149 2150 if (mm->map_count > sysctl_max_map_count) 2151 return -ENOMEM; 2152 2153 if (security_vm_enough_memory(len >> PAGE_SHIFT)) 2154 return -ENOMEM; 2155 2156 /* Can we just expand an old private anonymous mapping? */ 2157 vma = vma_merge(mm, prev, addr, addr + len, flags, 2158 NULL, NULL, pgoff, NULL); 2159 if (vma) 2160 goto out; 2161 2162 /* 2163 * create a vma struct for an anonymous mapping 2164 */ 2165 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2166 if (!vma) { 2167 vm_unacct_memory(len >> PAGE_SHIFT); 2168 return -ENOMEM; 2169 } 2170 2171 INIT_LIST_HEAD(&vma->anon_vma_chain); 2172 vma->vm_mm = mm; 2173 vma->vm_start = addr; 2174 vma->vm_end = addr + len; 2175 vma->vm_pgoff = pgoff; 2176 vma->vm_flags = flags; 2177 vma->vm_page_prot = vm_get_page_prot(flags); 2178 vma_link(mm, vma, prev, rb_link, rb_parent); 2179 out: 2180 mm->total_vm += len >> PAGE_SHIFT; 2181 if (flags & VM_LOCKED) { 2182 if (!mlock_vma_pages_range(vma, addr, addr + len)) 2183 mm->locked_vm += (len >> PAGE_SHIFT); 2184 } 2185 return addr; 2186 } 2187 2188 EXPORT_SYMBOL(do_brk); 2189 2190 /* Release all mmaps. */ 2191 void exit_mmap(struct mm_struct *mm) 2192 { 2193 struct mmu_gather *tlb; 2194 struct vm_area_struct *vma; 2195 unsigned long nr_accounted = 0; 2196 unsigned long end; 2197 2198 /* mm's last user has gone, and its about to be pulled down */ 2199 mmu_notifier_release(mm); 2200 2201 if (mm->locked_vm) { 2202 vma = mm->mmap; 2203 while (vma) { 2204 if (vma->vm_flags & VM_LOCKED) 2205 munlock_vma_pages_all(vma); 2206 vma = vma->vm_next; 2207 } 2208 } 2209 2210 arch_exit_mmap(mm); 2211 2212 vma = mm->mmap; 2213 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2214 return; 2215 2216 lru_add_drain(); 2217 flush_cache_mm(mm); 2218 tlb = tlb_gather_mmu(mm, 1); 2219 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2220 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2221 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL); 2222 vm_unacct_memory(nr_accounted); 2223 2224 free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0); 2225 tlb_finish_mmu(tlb, 0, end); 2226 2227 /* 2228 * Walk the list again, actually closing and freeing it, 2229 * with preemption enabled, without holding any MM locks. 2230 */ 2231 while (vma) 2232 vma = remove_vma(vma); 2233 2234 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT); 2235 } 2236 2237 /* Insert vm structure into process list sorted by address 2238 * and into the inode's i_mmap tree. If vm_file is non-NULL 2239 * then i_mmap_lock is taken here. 2240 */ 2241 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma) 2242 { 2243 struct vm_area_struct * __vma, * prev; 2244 struct rb_node ** rb_link, * rb_parent; 2245 2246 /* 2247 * The vm_pgoff of a purely anonymous vma should be irrelevant 2248 * until its first write fault, when page's anon_vma and index 2249 * are set. But now set the vm_pgoff it will almost certainly 2250 * end up with (unless mremap moves it elsewhere before that 2251 * first wfault), so /proc/pid/maps tells a consistent story. 2252 * 2253 * By setting it to reflect the virtual start address of the 2254 * vma, merges and splits can happen in a seamless way, just 2255 * using the existing file pgoff checks and manipulations. 2256 * Similarly in do_mmap_pgoff and in do_brk. 2257 */ 2258 if (!vma->vm_file) { 2259 BUG_ON(vma->anon_vma); 2260 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2261 } 2262 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent); 2263 if (__vma && __vma->vm_start < vma->vm_end) 2264 return -ENOMEM; 2265 if ((vma->vm_flags & VM_ACCOUNT) && 2266 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2267 return -ENOMEM; 2268 vma_link(mm, vma, prev, rb_link, rb_parent); 2269 return 0; 2270 } 2271 2272 /* 2273 * Copy the vma structure to a new location in the same mm, 2274 * prior to moving page table entries, to effect an mremap move. 2275 */ 2276 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2277 unsigned long addr, unsigned long len, pgoff_t pgoff) 2278 { 2279 struct vm_area_struct *vma = *vmap; 2280 unsigned long vma_start = vma->vm_start; 2281 struct mm_struct *mm = vma->vm_mm; 2282 struct vm_area_struct *new_vma, *prev; 2283 struct rb_node **rb_link, *rb_parent; 2284 struct mempolicy *pol; 2285 2286 /* 2287 * If anonymous vma has not yet been faulted, update new pgoff 2288 * to match new location, to increase its chance of merging. 2289 */ 2290 if (!vma->vm_file && !vma->anon_vma) 2291 pgoff = addr >> PAGE_SHIFT; 2292 2293 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 2294 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2295 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma)); 2296 if (new_vma) { 2297 /* 2298 * Source vma may have been merged into new_vma 2299 */ 2300 if (vma_start >= new_vma->vm_start && 2301 vma_start < new_vma->vm_end) 2302 *vmap = new_vma; 2303 } else { 2304 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2305 if (new_vma) { 2306 *new_vma = *vma; 2307 pol = mpol_dup(vma_policy(vma)); 2308 if (IS_ERR(pol)) 2309 goto out_free_vma; 2310 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 2311 if (anon_vma_clone(new_vma, vma)) 2312 goto out_free_mempol; 2313 vma_set_policy(new_vma, pol); 2314 new_vma->vm_start = addr; 2315 new_vma->vm_end = addr + len; 2316 new_vma->vm_pgoff = pgoff; 2317 if (new_vma->vm_file) { 2318 get_file(new_vma->vm_file); 2319 if (vma->vm_flags & VM_EXECUTABLE) 2320 added_exe_file_vma(mm); 2321 } 2322 if (new_vma->vm_ops && new_vma->vm_ops->open) 2323 new_vma->vm_ops->open(new_vma); 2324 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2325 } 2326 } 2327 return new_vma; 2328 2329 out_free_mempol: 2330 mpol_put(pol); 2331 out_free_vma: 2332 kmem_cache_free(vm_area_cachep, new_vma); 2333 return NULL; 2334 } 2335 2336 /* 2337 * Return true if the calling process may expand its vm space by the passed 2338 * number of pages 2339 */ 2340 int may_expand_vm(struct mm_struct *mm, unsigned long npages) 2341 { 2342 unsigned long cur = mm->total_vm; /* pages */ 2343 unsigned long lim; 2344 2345 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT; 2346 2347 if (cur + npages > lim) 2348 return 0; 2349 return 1; 2350 } 2351 2352 2353 static int special_mapping_fault(struct vm_area_struct *vma, 2354 struct vm_fault *vmf) 2355 { 2356 pgoff_t pgoff; 2357 struct page **pages; 2358 2359 /* 2360 * special mappings have no vm_file, and in that case, the mm 2361 * uses vm_pgoff internally. So we have to subtract it from here. 2362 * We are allowed to do this because we are the mm; do not copy 2363 * this code into drivers! 2364 */ 2365 pgoff = vmf->pgoff - vma->vm_pgoff; 2366 2367 for (pages = vma->vm_private_data; pgoff && *pages; ++pages) 2368 pgoff--; 2369 2370 if (*pages) { 2371 struct page *page = *pages; 2372 get_page(page); 2373 vmf->page = page; 2374 return 0; 2375 } 2376 2377 return VM_FAULT_SIGBUS; 2378 } 2379 2380 /* 2381 * Having a close hook prevents vma merging regardless of flags. 2382 */ 2383 static void special_mapping_close(struct vm_area_struct *vma) 2384 { 2385 } 2386 2387 static const struct vm_operations_struct special_mapping_vmops = { 2388 .close = special_mapping_close, 2389 .fault = special_mapping_fault, 2390 }; 2391 2392 /* 2393 * Called with mm->mmap_sem held for writing. 2394 * Insert a new vma covering the given region, with the given flags. 2395 * Its pages are supplied by the given array of struct page *. 2396 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 2397 * The region past the last page supplied will always produce SIGBUS. 2398 * The array pointer and the pages it points to are assumed to stay alive 2399 * for as long as this mapping might exist. 2400 */ 2401 int install_special_mapping(struct mm_struct *mm, 2402 unsigned long addr, unsigned long len, 2403 unsigned long vm_flags, struct page **pages) 2404 { 2405 struct vm_area_struct *vma; 2406 2407 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2408 if (unlikely(vma == NULL)) 2409 return -ENOMEM; 2410 2411 INIT_LIST_HEAD(&vma->anon_vma_chain); 2412 vma->vm_mm = mm; 2413 vma->vm_start = addr; 2414 vma->vm_end = addr + len; 2415 2416 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND; 2417 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 2418 2419 vma->vm_ops = &special_mapping_vmops; 2420 vma->vm_private_data = pages; 2421 2422 if (unlikely(insert_vm_struct(mm, vma))) { 2423 kmem_cache_free(vm_area_cachep, vma); 2424 return -ENOMEM; 2425 } 2426 2427 mm->total_vm += len >> PAGE_SHIFT; 2428 2429 perf_event_mmap(vma); 2430 2431 return 0; 2432 } 2433 2434 static DEFINE_MUTEX(mm_all_locks_mutex); 2435 2436 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 2437 { 2438 if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) { 2439 /* 2440 * The LSB of head.next can't change from under us 2441 * because we hold the mm_all_locks_mutex. 2442 */ 2443 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem); 2444 /* 2445 * We can safely modify head.next after taking the 2446 * anon_vma->lock. If some other vma in this mm shares 2447 * the same anon_vma we won't take it again. 2448 * 2449 * No need of atomic instructions here, head.next 2450 * can't change from under us thanks to the 2451 * anon_vma->lock. 2452 */ 2453 if (__test_and_set_bit(0, (unsigned long *) 2454 &anon_vma->head.next)) 2455 BUG(); 2456 } 2457 } 2458 2459 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 2460 { 2461 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 2462 /* 2463 * AS_MM_ALL_LOCKS can't change from under us because 2464 * we hold the mm_all_locks_mutex. 2465 * 2466 * Operations on ->flags have to be atomic because 2467 * even if AS_MM_ALL_LOCKS is stable thanks to the 2468 * mm_all_locks_mutex, there may be other cpus 2469 * changing other bitflags in parallel to us. 2470 */ 2471 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 2472 BUG(); 2473 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem); 2474 } 2475 } 2476 2477 /* 2478 * This operation locks against the VM for all pte/vma/mm related 2479 * operations that could ever happen on a certain mm. This includes 2480 * vmtruncate, try_to_unmap, and all page faults. 2481 * 2482 * The caller must take the mmap_sem in write mode before calling 2483 * mm_take_all_locks(). The caller isn't allowed to release the 2484 * mmap_sem until mm_drop_all_locks() returns. 2485 * 2486 * mmap_sem in write mode is required in order to block all operations 2487 * that could modify pagetables and free pages without need of 2488 * altering the vma layout (for example populate_range() with 2489 * nonlinear vmas). It's also needed in write mode to avoid new 2490 * anon_vmas to be associated with existing vmas. 2491 * 2492 * A single task can't take more than one mm_take_all_locks() in a row 2493 * or it would deadlock. 2494 * 2495 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in 2496 * mapping->flags avoid to take the same lock twice, if more than one 2497 * vma in this mm is backed by the same anon_vma or address_space. 2498 * 2499 * We can take all the locks in random order because the VM code 2500 * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never 2501 * takes more than one of them in a row. Secondly we're protected 2502 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex. 2503 * 2504 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 2505 * that may have to take thousand of locks. 2506 * 2507 * mm_take_all_locks() can fail if it's interrupted by signals. 2508 */ 2509 int mm_take_all_locks(struct mm_struct *mm) 2510 { 2511 struct vm_area_struct *vma; 2512 struct anon_vma_chain *avc; 2513 int ret = -EINTR; 2514 2515 BUG_ON(down_read_trylock(&mm->mmap_sem)); 2516 2517 mutex_lock(&mm_all_locks_mutex); 2518 2519 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2520 if (signal_pending(current)) 2521 goto out_unlock; 2522 if (vma->vm_file && vma->vm_file->f_mapping) 2523 vm_lock_mapping(mm, vma->vm_file->f_mapping); 2524 } 2525 2526 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2527 if (signal_pending(current)) 2528 goto out_unlock; 2529 if (vma->anon_vma) 2530 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 2531 vm_lock_anon_vma(mm, avc->anon_vma); 2532 } 2533 2534 ret = 0; 2535 2536 out_unlock: 2537 if (ret) 2538 mm_drop_all_locks(mm); 2539 2540 return ret; 2541 } 2542 2543 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 2544 { 2545 if (test_bit(0, (unsigned long *) &anon_vma->head.next)) { 2546 /* 2547 * The LSB of head.next can't change to 0 from under 2548 * us because we hold the mm_all_locks_mutex. 2549 * 2550 * We must however clear the bitflag before unlocking 2551 * the vma so the users using the anon_vma->head will 2552 * never see our bitflag. 2553 * 2554 * No need of atomic instructions here, head.next 2555 * can't change from under us until we release the 2556 * anon_vma->lock. 2557 */ 2558 if (!__test_and_clear_bit(0, (unsigned long *) 2559 &anon_vma->head.next)) 2560 BUG(); 2561 spin_unlock(&anon_vma->lock); 2562 } 2563 } 2564 2565 static void vm_unlock_mapping(struct address_space *mapping) 2566 { 2567 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 2568 /* 2569 * AS_MM_ALL_LOCKS can't change to 0 from under us 2570 * because we hold the mm_all_locks_mutex. 2571 */ 2572 spin_unlock(&mapping->i_mmap_lock); 2573 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 2574 &mapping->flags)) 2575 BUG(); 2576 } 2577 } 2578 2579 /* 2580 * The mmap_sem cannot be released by the caller until 2581 * mm_drop_all_locks() returns. 2582 */ 2583 void mm_drop_all_locks(struct mm_struct *mm) 2584 { 2585 struct vm_area_struct *vma; 2586 struct anon_vma_chain *avc; 2587 2588 BUG_ON(down_read_trylock(&mm->mmap_sem)); 2589 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 2590 2591 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2592 if (vma->anon_vma) 2593 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 2594 vm_unlock_anon_vma(avc->anon_vma); 2595 if (vma->vm_file && vma->vm_file->f_mapping) 2596 vm_unlock_mapping(vma->vm_file->f_mapping); 2597 } 2598 2599 mutex_unlock(&mm_all_locks_mutex); 2600 } 2601 2602 /* 2603 * initialise the VMA slab 2604 */ 2605 void __init mmap_init(void) 2606 { 2607 int ret; 2608 2609 ret = percpu_counter_init(&vm_committed_as, 0); 2610 VM_BUG_ON(ret); 2611 } 2612