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 47static 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 */ 72pgprot_t protection_map[16] = { 73 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 74 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 75}; 76 77pgprot_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} 83EXPORT_SYMBOL(vm_get_page_prot); 84 85int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */ 86int sysctl_overcommit_ratio = 50; /* default is 50% */ 87int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT; 88struct 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 */ 106int __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; 184error: 185 vm_unacct_memory(pages); 186 187 return -ENOMEM; 188} 189 190/* 191 * Requires inode->i_mapping->i_mmap_lock 192 */ 193static 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 */ 213void 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 */ 228static 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 245SYSCALL_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; 292set_brk: 293 mm->brk = brk; 294out: 295 retval = mm->brk; 296 up_write(&mm->mmap_sem); 297 return retval; 298} 299 300#ifdef DEBUG_MM_RB 301static 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 330void 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 350static struct vm_area_struct * 351find_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 387static 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 struct vm_area_struct *next; 392 393 vma->vm_prev = prev; 394 if (prev) { 395 next = prev->vm_next; 396 prev->vm_next = vma; 397 } else { 398 mm->mmap = vma; 399 if (rb_parent) 400 next = rb_entry(rb_parent, 401 struct vm_area_struct, vm_rb); 402 else 403 next = NULL; 404 } 405 vma->vm_next = next; 406 if (next) 407 next->vm_prev = vma; 408} 409 410void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 411 struct rb_node **rb_link, struct rb_node *rb_parent) 412{ 413 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 414 rb_insert_color(&vma->vm_rb, &mm->mm_rb); 415} 416 417static void __vma_link_file(struct vm_area_struct *vma) 418{ 419 struct file *file; 420 421 file = vma->vm_file; 422 if (file) { 423 struct address_space *mapping = file->f_mapping; 424 425 if (vma->vm_flags & VM_DENYWRITE) 426 atomic_dec(&file->f_path.dentry->d_inode->i_writecount); 427 if (vma->vm_flags & VM_SHARED) 428 mapping->i_mmap_writable++; 429 430 flush_dcache_mmap_lock(mapping); 431 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 432 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear); 433 else 434 vma_prio_tree_insert(vma, &mapping->i_mmap); 435 flush_dcache_mmap_unlock(mapping); 436 } 437} 438 439static void 440__vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 441 struct vm_area_struct *prev, struct rb_node **rb_link, 442 struct rb_node *rb_parent) 443{ 444 __vma_link_list(mm, vma, prev, rb_parent); 445 __vma_link_rb(mm, vma, rb_link, rb_parent); 446} 447 448static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 449 struct vm_area_struct *prev, struct rb_node **rb_link, 450 struct rb_node *rb_parent) 451{ 452 struct address_space *mapping = NULL; 453 454 if (vma->vm_file) 455 mapping = vma->vm_file->f_mapping; 456 457 if (mapping) { 458 spin_lock(&mapping->i_mmap_lock); 459 vma->vm_truncate_count = mapping->truncate_count; 460 } 461 462 __vma_link(mm, vma, prev, rb_link, rb_parent); 463 __vma_link_file(vma); 464 465 if (mapping) 466 spin_unlock(&mapping->i_mmap_lock); 467 468 mm->map_count++; 469 validate_mm(mm); 470} 471 472/* 473 * Helper for vma_adjust in the split_vma insert case: 474 * insert vm structure into list and rbtree and anon_vma, 475 * but it has already been inserted into prio_tree earlier. 476 */ 477static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 478{ 479 struct vm_area_struct *__vma, *prev; 480 struct rb_node **rb_link, *rb_parent; 481 482 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent); 483 BUG_ON(__vma && __vma->vm_start < vma->vm_end); 484 __vma_link(mm, vma, prev, rb_link, rb_parent); 485 mm->map_count++; 486} 487 488static inline void 489__vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 490 struct vm_area_struct *prev) 491{ 492 struct vm_area_struct *next = vma->vm_next; 493 494 prev->vm_next = next; 495 if (next) 496 next->vm_prev = prev; 497 rb_erase(&vma->vm_rb, &mm->mm_rb); 498 if (mm->mmap_cache == vma) 499 mm->mmap_cache = prev; 500} 501 502/* 503 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 504 * is already present in an i_mmap tree without adjusting the tree. 505 * The following helper function should be used when such adjustments 506 * are necessary. The "insert" vma (if any) is to be inserted 507 * before we drop the necessary locks. 508 */ 509int vma_adjust(struct vm_area_struct *vma, unsigned long start, 510 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 511{ 512 struct mm_struct *mm = vma->vm_mm; 513 struct vm_area_struct *next = vma->vm_next; 514 struct vm_area_struct *importer = NULL; 515 struct address_space *mapping = NULL; 516 struct prio_tree_root *root = NULL; 517 struct anon_vma *anon_vma = NULL; 518 struct file *file = vma->vm_file; 519 long adjust_next = 0; 520 int remove_next = 0; 521 522 if (next && !insert) { 523 struct vm_area_struct *exporter = NULL; 524 525 if (end >= next->vm_end) { 526 /* 527 * vma expands, overlapping all the next, and 528 * perhaps the one after too (mprotect case 6). 529 */ 530again: remove_next = 1 + (end > next->vm_end); 531 end = next->vm_end; 532 exporter = next; 533 importer = vma; 534 } else if (end > next->vm_start) { 535 /* 536 * vma expands, overlapping part of the next: 537 * mprotect case 5 shifting the boundary up. 538 */ 539 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 540 exporter = next; 541 importer = vma; 542 } else if (end < vma->vm_end) { 543 /* 544 * vma shrinks, and !insert tells it's not 545 * split_vma inserting another: so it must be 546 * mprotect case 4 shifting the boundary down. 547 */ 548 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT); 549 exporter = vma; 550 importer = next; 551 } 552 553 /* 554 * Easily overlooked: when mprotect shifts the boundary, 555 * make sure the expanding vma has anon_vma set if the 556 * shrinking vma had, to cover any anon pages imported. 557 */ 558 if (exporter && exporter->anon_vma && !importer->anon_vma) { 559 if (anon_vma_clone(importer, exporter)) 560 return -ENOMEM; 561 importer->anon_vma = exporter->anon_vma; 562 } 563 } 564 565 if (file) { 566 mapping = file->f_mapping; 567 if (!(vma->vm_flags & VM_NONLINEAR)) 568 root = &mapping->i_mmap; 569 spin_lock(&mapping->i_mmap_lock); 570 if (importer && 571 vma->vm_truncate_count != next->vm_truncate_count) { 572 /* 573 * unmap_mapping_range might be in progress: 574 * ensure that the expanding vma is rescanned. 575 */ 576 importer->vm_truncate_count = 0; 577 } 578 if (insert) { 579 insert->vm_truncate_count = vma->vm_truncate_count; 580 /* 581 * Put into prio_tree now, so instantiated pages 582 * are visible to arm/parisc __flush_dcache_page 583 * throughout; but we cannot insert into address 584 * space until vma start or end is updated. 585 */ 586 __vma_link_file(insert); 587 } 588 } 589 590 /* 591 * When changing only vma->vm_end, we don't really need anon_vma 592 * lock. This is a fairly rare case by itself, but the anon_vma 593 * lock may be shared between many sibling processes. Skipping 594 * the lock for brk adjustments makes a difference sometimes. 595 */ 596 if (vma->anon_vma && (insert || importer || start != vma->vm_start)) { 597 anon_vma = vma->anon_vma; 598 anon_vma_lock(anon_vma); 599 } 600 601 if (root) { 602 flush_dcache_mmap_lock(mapping); 603 vma_prio_tree_remove(vma, root); 604 if (adjust_next) 605 vma_prio_tree_remove(next, root); 606 } 607 608 vma->vm_start = start; 609 vma->vm_end = end; 610 vma->vm_pgoff = pgoff; 611 if (adjust_next) { 612 next->vm_start += adjust_next << PAGE_SHIFT; 613 next->vm_pgoff += adjust_next; 614 } 615 616 if (root) { 617 if (adjust_next) 618 vma_prio_tree_insert(next, root); 619 vma_prio_tree_insert(vma, root); 620 flush_dcache_mmap_unlock(mapping); 621 } 622 623 if (remove_next) { 624 /* 625 * vma_merge has merged next into vma, and needs 626 * us to remove next before dropping the locks. 627 */ 628 __vma_unlink(mm, next, vma); 629 if (file) 630 __remove_shared_vm_struct(next, file, mapping); 631 } else if (insert) { 632 /* 633 * split_vma has split insert from vma, and needs 634 * us to insert it before dropping the locks 635 * (it may either follow vma or precede it). 636 */ 637 __insert_vm_struct(mm, insert); 638 } 639 640 if (anon_vma) 641 anon_vma_unlock(anon_vma); 642 if (mapping) 643 spin_unlock(&mapping->i_mmap_lock); 644 645 if (remove_next) { 646 if (file) { 647 fput(file); 648 if (next->vm_flags & VM_EXECUTABLE) 649 removed_exe_file_vma(mm); 650 } 651 if (next->anon_vma) 652 anon_vma_merge(vma, next); 653 mm->map_count--; 654 mpol_put(vma_policy(next)); 655 kmem_cache_free(vm_area_cachep, next); 656 /* 657 * In mprotect's case 6 (see comments on vma_merge), 658 * we must remove another next too. It would clutter 659 * up the code too much to do both in one go. 660 */ 661 if (remove_next == 2) { 662 next = vma->vm_next; 663 goto again; 664 } 665 } 666 667 validate_mm(mm); 668 669 return 0; 670} 671 672/* 673 * If the vma has a ->close operation then the driver probably needs to release 674 * per-vma resources, so we don't attempt to merge those. 675 */ 676static inline int is_mergeable_vma(struct vm_area_struct *vma, 677 struct file *file, unsigned long vm_flags) 678{ 679 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */ 680 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR) 681 return 0; 682 if (vma->vm_file != file) 683 return 0; 684 if (vma->vm_ops && vma->vm_ops->close) 685 return 0; 686 return 1; 687} 688 689static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 690 struct anon_vma *anon_vma2) 691{ 692 return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2); 693} 694 695/* 696 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 697 * in front of (at a lower virtual address and file offset than) the vma. 698 * 699 * We cannot merge two vmas if they have differently assigned (non-NULL) 700 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 701 * 702 * We don't check here for the merged mmap wrapping around the end of pagecache 703 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 704 * wrap, nor mmaps which cover the final page at index -1UL. 705 */ 706static int 707can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 708 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 709{ 710 if (is_mergeable_vma(vma, file, vm_flags) && 711 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) { 712 if (vma->vm_pgoff == vm_pgoff) 713 return 1; 714 } 715 return 0; 716} 717 718/* 719 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 720 * beyond (at a higher virtual address and file offset than) the vma. 721 * 722 * We cannot merge two vmas if they have differently assigned (non-NULL) 723 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 724 */ 725static int 726can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 727 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 728{ 729 if (is_mergeable_vma(vma, file, vm_flags) && 730 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) { 731 pgoff_t vm_pglen; 732 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; 733 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 734 return 1; 735 } 736 return 0; 737} 738 739struct vm_area_struct *vma_merge(struct mm_struct *mm, 740 struct vm_area_struct *prev, unsigned long addr, 741 unsigned long end, unsigned long vm_flags, 742 struct anon_vma *anon_vma, struct file *file, 743 pgoff_t pgoff, struct mempolicy *policy) 744{ 745 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 746 struct vm_area_struct *area, *next; 747 int err; 748 749 /* 750 * We later require that vma->vm_flags == vm_flags, 751 * so this tests vma->vm_flags & VM_SPECIAL, too. 752 */ 753 if (vm_flags & VM_SPECIAL) 754 return NULL; 755 756 if (prev) 757 next = prev->vm_next; 758 else 759 next = mm->mmap; 760 area = next; 761 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 762 next = next->vm_next; 763 764 /* 765 * Can it merge with the predecessor? 766 */ 767 if (prev && prev->vm_end == addr && 768 mpol_equal(vma_policy(prev), policy) && 769 can_vma_merge_after(prev, vm_flags, 770 anon_vma, file, pgoff)) { 771 /* 772 * OK, it can. Can we now merge in the successor as well? 773 */ 774 if (next && end == next->vm_start && 775 mpol_equal(policy, vma_policy(next)) && 776 can_vma_merge_before(next, vm_flags, 777 anon_vma, file, pgoff+pglen) && 778 is_mergeable_anon_vma(prev->anon_vma, 779 next->anon_vma)) { 780 /* cases 1, 6 */ 781 err = vma_adjust(prev, prev->vm_start, 782 next->vm_end, prev->vm_pgoff, NULL); 783 } else /* cases 2, 5, 7 */ 784 err = vma_adjust(prev, prev->vm_start, 785 end, prev->vm_pgoff, NULL); 786 if (err) 787 return NULL; 788 return prev; 789 } 790 791 /* 792 * Can this new request be merged in front of next? 793 */ 794 if (next && end == next->vm_start && 795 mpol_equal(policy, vma_policy(next)) && 796 can_vma_merge_before(next, vm_flags, 797 anon_vma, file, pgoff+pglen)) { 798 if (prev && addr < prev->vm_end) /* case 4 */ 799 err = vma_adjust(prev, prev->vm_start, 800 addr, prev->vm_pgoff, NULL); 801 else /* cases 3, 8 */ 802 err = vma_adjust(area, addr, next->vm_end, 803 next->vm_pgoff - pglen, NULL); 804 if (err) 805 return NULL; 806 return area; 807 } 808 809 return NULL; 810} 811 812/* 813 * Rough compatbility check to quickly see if it's even worth looking 814 * at sharing an anon_vma. 815 * 816 * They need to have the same vm_file, and the flags can only differ 817 * in things that mprotect may change. 818 * 819 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 820 * we can merge the two vma's. For example, we refuse to merge a vma if 821 * there is a vm_ops->close() function, because that indicates that the 822 * driver is doing some kind of reference counting. But that doesn't 823 * really matter for the anon_vma sharing case. 824 */ 825static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 826{ 827 return a->vm_end == b->vm_start && 828 mpol_equal(vma_policy(a), vma_policy(b)) && 829 a->vm_file == b->vm_file && 830 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) && 831 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 832} 833 834/* 835 * Do some basic sanity checking to see if we can re-use the anon_vma 836 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 837 * the same as 'old', the other will be the new one that is trying 838 * to share the anon_vma. 839 * 840 * NOTE! This runs with mm_sem held for reading, so it is possible that 841 * the anon_vma of 'old' is concurrently in the process of being set up 842 * by another page fault trying to merge _that_. But that's ok: if it 843 * is being set up, that automatically means that it will be a singleton 844 * acceptable for merging, so we can do all of this optimistically. But 845 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer. 846 * 847 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 848 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 849 * is to return an anon_vma that is "complex" due to having gone through 850 * a fork). 851 * 852 * We also make sure that the two vma's are compatible (adjacent, 853 * and with the same memory policies). That's all stable, even with just 854 * a read lock on the mm_sem. 855 */ 856static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 857{ 858 if (anon_vma_compatible(a, b)) { 859 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma); 860 861 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 862 return anon_vma; 863 } 864 return NULL; 865} 866 867/* 868 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 869 * neighbouring vmas for a suitable anon_vma, before it goes off 870 * to allocate a new anon_vma. It checks because a repetitive 871 * sequence of mprotects and faults may otherwise lead to distinct 872 * anon_vmas being allocated, preventing vma merge in subsequent 873 * mprotect. 874 */ 875struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 876{ 877 struct anon_vma *anon_vma; 878 struct vm_area_struct *near; 879 880 near = vma->vm_next; 881 if (!near) 882 goto try_prev; 883 884 anon_vma = reusable_anon_vma(near, vma, near); 885 if (anon_vma) 886 return anon_vma; 887try_prev: 888 /* 889 * It is potentially slow to have to call find_vma_prev here. 890 * But it's only on the first write fault on the vma, not 891 * every time, and we could devise a way to avoid it later 892 * (e.g. stash info in next's anon_vma_node when assigning 893 * an anon_vma, or when trying vma_merge). Another time. 894 */ 895 BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma); 896 if (!near) 897 goto none; 898 899 anon_vma = reusable_anon_vma(near, near, vma); 900 if (anon_vma) 901 return anon_vma; 902none: 903 /* 904 * There's no absolute need to look only at touching neighbours: 905 * we could search further afield for "compatible" anon_vmas. 906 * But it would probably just be a waste of time searching, 907 * or lead to too many vmas hanging off the same anon_vma. 908 * We're trying to allow mprotect remerging later on, 909 * not trying to minimize memory used for anon_vmas. 910 */ 911 return NULL; 912} 913 914#ifdef CONFIG_PROC_FS 915void vm_stat_account(struct mm_struct *mm, unsigned long flags, 916 struct file *file, long pages) 917{ 918 const unsigned long stack_flags 919 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN); 920 921 if (file) { 922 mm->shared_vm += pages; 923 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC) 924 mm->exec_vm += pages; 925 } else if (flags & stack_flags) 926 mm->stack_vm += pages; 927 if (flags & (VM_RESERVED|VM_IO)) 928 mm->reserved_vm += pages; 929} 930#endif /* CONFIG_PROC_FS */ 931 932/* 933 * The caller must hold down_write(¤t->mm->mmap_sem). 934 */ 935 936unsigned long do_mmap_pgoff(struct file *file, unsigned long addr, 937 unsigned long len, unsigned long prot, 938 unsigned long flags, unsigned long pgoff) 939{ 940 struct mm_struct * mm = current->mm; 941 struct inode *inode; 942 unsigned int vm_flags; 943 int error; 944 unsigned long reqprot = prot; 945 946 /* 947 * Does the application expect PROT_READ to imply PROT_EXEC? 948 * 949 * (the exception is when the underlying filesystem is noexec 950 * mounted, in which case we dont add PROT_EXEC.) 951 */ 952 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 953 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC))) 954 prot |= PROT_EXEC; 955 956 if (!len) 957 return -EINVAL; 958 959 if (!(flags & MAP_FIXED)) 960 addr = round_hint_to_min(addr); 961 962 /* Careful about overflows.. */ 963 len = PAGE_ALIGN(len); 964 if (!len) 965 return -ENOMEM; 966 967 /* offset overflow? */ 968 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 969 return -EOVERFLOW; 970 971 /* Too many mappings? */ 972 if (mm->map_count > sysctl_max_map_count) 973 return -ENOMEM; 974 975 /* Obtain the address to map to. we verify (or select) it and ensure 976 * that it represents a valid section of the address space. 977 */ 978 addr = get_unmapped_area(file, addr, len, pgoff, flags); 979 if (addr & ~PAGE_MASK) 980 return addr; 981 982 /* Do simple checking here so the lower-level routines won't have 983 * to. we assume access permissions have been handled by the open 984 * of the memory object, so we don't do any here. 985 */ 986 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 987 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 988 989 if (flags & MAP_LOCKED) 990 if (!can_do_mlock()) 991 return -EPERM; 992 993 /* mlock MCL_FUTURE? */ 994 if (vm_flags & VM_LOCKED) { 995 unsigned long locked, lock_limit; 996 locked = len >> PAGE_SHIFT; 997 locked += mm->locked_vm; 998 lock_limit = rlimit(RLIMIT_MEMLOCK); 999 lock_limit >>= PAGE_SHIFT; 1000 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1001 return -EAGAIN; 1002 } 1003 1004 inode = file ? file->f_path.dentry->d_inode : NULL; 1005 1006 if (file) { 1007 switch (flags & MAP_TYPE) { 1008 case MAP_SHARED: 1009 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 1010 return -EACCES; 1011 1012 /* 1013 * Make sure we don't allow writing to an append-only 1014 * file.. 1015 */ 1016 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1017 return -EACCES; 1018 1019 /* 1020 * Make sure there are no mandatory locks on the file. 1021 */ 1022 if (locks_verify_locked(inode)) 1023 return -EAGAIN; 1024 1025 vm_flags |= VM_SHARED | VM_MAYSHARE; 1026 if (!(file->f_mode & FMODE_WRITE)) 1027 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1028 1029 /* fall through */ 1030 case MAP_PRIVATE: 1031 if (!(file->f_mode & FMODE_READ)) 1032 return -EACCES; 1033 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) { 1034 if (vm_flags & VM_EXEC) 1035 return -EPERM; 1036 vm_flags &= ~VM_MAYEXEC; 1037 } 1038 1039 if (!file->f_op || !file->f_op->mmap) 1040 return -ENODEV; 1041 break; 1042 1043 default: 1044 return -EINVAL; 1045 } 1046 } else { 1047 switch (flags & MAP_TYPE) { 1048 case MAP_SHARED: 1049 /* 1050 * Ignore pgoff. 1051 */ 1052 pgoff = 0; 1053 vm_flags |= VM_SHARED | VM_MAYSHARE; 1054 break; 1055 case MAP_PRIVATE: 1056 /* 1057 * Set pgoff according to addr for anon_vma. 1058 */ 1059 pgoff = addr >> PAGE_SHIFT; 1060 break; 1061 default: 1062 return -EINVAL; 1063 } 1064 } 1065 1066 error = security_file_mmap(file, reqprot, prot, flags, addr, 0); 1067 if (error) 1068 return error; 1069 1070 return mmap_region(file, addr, len, flags, vm_flags, pgoff); 1071} 1072EXPORT_SYMBOL(do_mmap_pgoff); 1073 1074SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1075 unsigned long, prot, unsigned long, flags, 1076 unsigned long, fd, unsigned long, pgoff) 1077{ 1078 struct file *file = NULL; 1079 unsigned long retval = -EBADF; 1080 1081 if (!(flags & MAP_ANONYMOUS)) { 1082 if (unlikely(flags & MAP_HUGETLB)) 1083 return -EINVAL; 1084 file = fget(fd); 1085 if (!file) 1086 goto out; 1087 } else if (flags & MAP_HUGETLB) { 1088 struct user_struct *user = NULL; 1089 /* 1090 * VM_NORESERVE is used because the reservations will be 1091 * taken when vm_ops->mmap() is called 1092 * A dummy user value is used because we are not locking 1093 * memory so no accounting is necessary 1094 */ 1095 len = ALIGN(len, huge_page_size(&default_hstate)); 1096 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE, 1097 &user, HUGETLB_ANONHUGE_INODE); 1098 if (IS_ERR(file)) 1099 return PTR_ERR(file); 1100 } 1101 1102 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1103 1104 down_write(¤t->mm->mmap_sem); 1105 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1106 up_write(¤t->mm->mmap_sem); 1107 1108 if (file) 1109 fput(file); 1110out: 1111 return retval; 1112} 1113 1114#ifdef __ARCH_WANT_SYS_OLD_MMAP 1115struct mmap_arg_struct { 1116 unsigned long addr; 1117 unsigned long len; 1118 unsigned long prot; 1119 unsigned long flags; 1120 unsigned long fd; 1121 unsigned long offset; 1122}; 1123 1124SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1125{ 1126 struct mmap_arg_struct a; 1127 1128 if (copy_from_user(&a, arg, sizeof(a))) 1129 return -EFAULT; 1130 if (a.offset & ~PAGE_MASK) 1131 return -EINVAL; 1132 1133 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1134 a.offset >> PAGE_SHIFT); 1135} 1136#endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1137 1138/* 1139 * Some shared mappigns will want the pages marked read-only 1140 * to track write events. If so, we'll downgrade vm_page_prot 1141 * to the private version (using protection_map[] without the 1142 * VM_SHARED bit). 1143 */ 1144int vma_wants_writenotify(struct vm_area_struct *vma) 1145{ 1146 unsigned int vm_flags = vma->vm_flags; 1147 1148 /* If it was private or non-writable, the write bit is already clear */ 1149 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1150 return 0; 1151 1152 /* The backer wishes to know when pages are first written to? */ 1153 if (vma->vm_ops && vma->vm_ops->page_mkwrite) 1154 return 1; 1155 1156 /* The open routine did something to the protections already? */ 1157 if (pgprot_val(vma->vm_page_prot) != 1158 pgprot_val(vm_get_page_prot(vm_flags))) 1159 return 0; 1160 1161 /* Specialty mapping? */ 1162 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE)) 1163 return 0; 1164 1165 /* Can the mapping track the dirty pages? */ 1166 return vma->vm_file && vma->vm_file->f_mapping && 1167 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1168} 1169 1170/* 1171 * We account for memory if it's a private writeable mapping, 1172 * not hugepages and VM_NORESERVE wasn't set. 1173 */ 1174static inline int accountable_mapping(struct file *file, unsigned int vm_flags) 1175{ 1176 /* 1177 * hugetlb has its own accounting separate from the core VM 1178 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1179 */ 1180 if (file && is_file_hugepages(file)) 1181 return 0; 1182 1183 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1184} 1185 1186unsigned long mmap_region(struct file *file, unsigned long addr, 1187 unsigned long len, unsigned long flags, 1188 unsigned int vm_flags, unsigned long pgoff) 1189{ 1190 struct mm_struct *mm = current->mm; 1191 struct vm_area_struct *vma, *prev; 1192 int correct_wcount = 0; 1193 int error; 1194 struct rb_node **rb_link, *rb_parent; 1195 unsigned long charged = 0; 1196 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL; 1197 1198 /* Clear old maps */ 1199 error = -ENOMEM; 1200munmap_back: 1201 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 1202 if (vma && vma->vm_start < addr + len) { 1203 if (do_munmap(mm, addr, len)) 1204 return -ENOMEM; 1205 goto munmap_back; 1206 } 1207 1208 /* Check against address space limit. */ 1209 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 1210 return -ENOMEM; 1211 1212 /* 1213 * Set 'VM_NORESERVE' if we should not account for the 1214 * memory use of this mapping. 1215 */ 1216 if ((flags & MAP_NORESERVE)) { 1217 /* We honor MAP_NORESERVE if allowed to overcommit */ 1218 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1219 vm_flags |= VM_NORESERVE; 1220 1221 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1222 if (file && is_file_hugepages(file)) 1223 vm_flags |= VM_NORESERVE; 1224 } 1225 1226 /* 1227 * Private writable mapping: check memory availability 1228 */ 1229 if (accountable_mapping(file, vm_flags)) { 1230 charged = len >> PAGE_SHIFT; 1231 if (security_vm_enough_memory(charged)) 1232 return -ENOMEM; 1233 vm_flags |= VM_ACCOUNT; 1234 } 1235 1236 /* 1237 * Can we just expand an old mapping? 1238 */ 1239 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL); 1240 if (vma) 1241 goto out; 1242 1243 /* 1244 * Determine the object being mapped and call the appropriate 1245 * specific mapper. the address has already been validated, but 1246 * not unmapped, but the maps are removed from the list. 1247 */ 1248 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1249 if (!vma) { 1250 error = -ENOMEM; 1251 goto unacct_error; 1252 } 1253 1254 vma->vm_mm = mm; 1255 vma->vm_start = addr; 1256 vma->vm_end = addr + len; 1257 vma->vm_flags = vm_flags; 1258 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1259 vma->vm_pgoff = pgoff; 1260 INIT_LIST_HEAD(&vma->anon_vma_chain); 1261 1262 if (file) { 1263 error = -EINVAL; 1264 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1265 goto free_vma; 1266 if (vm_flags & VM_DENYWRITE) { 1267 error = deny_write_access(file); 1268 if (error) 1269 goto free_vma; 1270 correct_wcount = 1; 1271 } 1272 vma->vm_file = file; 1273 get_file(file); 1274 error = file->f_op->mmap(file, vma); 1275 if (error) 1276 goto unmap_and_free_vma; 1277 if (vm_flags & VM_EXECUTABLE) 1278 added_exe_file_vma(mm); 1279 1280 /* Can addr have changed?? 1281 * 1282 * Answer: Yes, several device drivers can do it in their 1283 * f_op->mmap method. -DaveM 1284 */ 1285 addr = vma->vm_start; 1286 pgoff = vma->vm_pgoff; 1287 vm_flags = vma->vm_flags; 1288 } else if (vm_flags & VM_SHARED) { 1289 error = shmem_zero_setup(vma); 1290 if (error) 1291 goto free_vma; 1292 } 1293 1294 if (vma_wants_writenotify(vma)) { 1295 pgprot_t pprot = vma->vm_page_prot; 1296 1297 /* Can vma->vm_page_prot have changed?? 1298 * 1299 * Answer: Yes, drivers may have changed it in their 1300 * f_op->mmap method. 1301 * 1302 * Ensures that vmas marked as uncached stay that way. 1303 */ 1304 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED); 1305 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot))) 1306 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 1307 } 1308 1309 vma_link(mm, vma, prev, rb_link, rb_parent); 1310 file = vma->vm_file; 1311 1312 /* Once vma denies write, undo our temporary denial count */ 1313 if (correct_wcount) 1314 atomic_inc(&inode->i_writecount); 1315out: 1316 perf_event_mmap(vma); 1317 1318 mm->total_vm += len >> PAGE_SHIFT; 1319 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT); 1320 if (vm_flags & VM_LOCKED) { 1321 if (!mlock_vma_pages_range(vma, addr, addr + len)) 1322 mm->locked_vm += (len >> PAGE_SHIFT); 1323 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK)) 1324 make_pages_present(addr, addr + len); 1325 return addr; 1326 1327unmap_and_free_vma: 1328 if (correct_wcount) 1329 atomic_inc(&inode->i_writecount); 1330 vma->vm_file = NULL; 1331 fput(file); 1332 1333 /* Undo any partial mapping done by a device driver. */ 1334 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1335 charged = 0; 1336free_vma: 1337 kmem_cache_free(vm_area_cachep, vma); 1338unacct_error: 1339 if (charged) 1340 vm_unacct_memory(charged); 1341 return error; 1342} 1343 1344/* Get an address range which is currently unmapped. 1345 * For shmat() with addr=0. 1346 * 1347 * Ugly calling convention alert: 1348 * Return value with the low bits set means error value, 1349 * ie 1350 * if (ret & ~PAGE_MASK) 1351 * error = ret; 1352 * 1353 * This function "knows" that -ENOMEM has the bits set. 1354 */ 1355#ifndef HAVE_ARCH_UNMAPPED_AREA 1356unsigned long 1357arch_get_unmapped_area(struct file *filp, unsigned long addr, 1358 unsigned long len, unsigned long pgoff, unsigned long flags) 1359{ 1360 struct mm_struct *mm = current->mm; 1361 struct vm_area_struct *vma; 1362 unsigned long start_addr; 1363 1364 if (len > TASK_SIZE) 1365 return -ENOMEM; 1366 1367 if (flags & MAP_FIXED) 1368 return addr; 1369 1370 if (addr) { 1371 addr = PAGE_ALIGN(addr); 1372 vma = find_vma(mm, addr); 1373 if (TASK_SIZE - len >= addr && 1374 (!vma || addr + len <= vma->vm_start)) 1375 return addr; 1376 } 1377 if (len > mm->cached_hole_size) { 1378 start_addr = addr = mm->free_area_cache; 1379 } else { 1380 start_addr = addr = TASK_UNMAPPED_BASE; 1381 mm->cached_hole_size = 0; 1382 } 1383 1384full_search: 1385 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1386 /* At this point: (!vma || addr < vma->vm_end). */ 1387 if (TASK_SIZE - len < addr) { 1388 /* 1389 * Start a new search - just in case we missed 1390 * some holes. 1391 */ 1392 if (start_addr != TASK_UNMAPPED_BASE) { 1393 addr = TASK_UNMAPPED_BASE; 1394 start_addr = addr; 1395 mm->cached_hole_size = 0; 1396 goto full_search; 1397 } 1398 return -ENOMEM; 1399 } 1400 if (!vma || addr + len <= vma->vm_start) { 1401 /* 1402 * Remember the place where we stopped the search: 1403 */ 1404 mm->free_area_cache = addr + len; 1405 return addr; 1406 } 1407 if (addr + mm->cached_hole_size < vma->vm_start) 1408 mm->cached_hole_size = vma->vm_start - addr; 1409 addr = vma->vm_end; 1410 } 1411} 1412#endif 1413 1414void arch_unmap_area(struct mm_struct *mm, unsigned long addr) 1415{ 1416 /* 1417 * Is this a new hole at the lowest possible address? 1418 */ 1419 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) { 1420 mm->free_area_cache = addr; 1421 mm->cached_hole_size = ~0UL; 1422 } 1423} 1424 1425/* 1426 * This mmap-allocator allocates new areas top-down from below the 1427 * stack's low limit (the base): 1428 */ 1429#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1430unsigned long 1431arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1432 const unsigned long len, const unsigned long pgoff, 1433 const unsigned long flags) 1434{ 1435 struct vm_area_struct *vma; 1436 struct mm_struct *mm = current->mm; 1437 unsigned long addr = addr0; 1438 1439 /* requested length too big for entire address space */ 1440 if (len > TASK_SIZE) 1441 return -ENOMEM; 1442 1443 if (flags & MAP_FIXED) 1444 return addr; 1445 1446 /* requesting a specific address */ 1447 if (addr) { 1448 addr = PAGE_ALIGN(addr); 1449 vma = find_vma(mm, addr); 1450 if (TASK_SIZE - len >= addr && 1451 (!vma || addr + len <= vma->vm_start)) 1452 return addr; 1453 } 1454 1455 /* check if free_area_cache is useful for us */ 1456 if (len <= mm->cached_hole_size) { 1457 mm->cached_hole_size = 0; 1458 mm->free_area_cache = mm->mmap_base; 1459 } 1460 1461 /* either no address requested or can't fit in requested address hole */ 1462 addr = mm->free_area_cache; 1463 1464 /* make sure it can fit in the remaining address space */ 1465 if (addr > len) { 1466 vma = find_vma(mm, addr-len); 1467 if (!vma || addr <= vma->vm_start) 1468 /* remember the address as a hint for next time */ 1469 return (mm->free_area_cache = addr-len); 1470 } 1471 1472 if (mm->mmap_base < len) 1473 goto bottomup; 1474 1475 addr = mm->mmap_base-len; 1476 1477 do { 1478 /* 1479 * Lookup failure means no vma is above this address, 1480 * else if new region fits below vma->vm_start, 1481 * return with success: 1482 */ 1483 vma = find_vma(mm, addr); 1484 if (!vma || addr+len <= vma->vm_start) 1485 /* remember the address as a hint for next time */ 1486 return (mm->free_area_cache = addr); 1487 1488 /* remember the largest hole we saw so far */ 1489 if (addr + mm->cached_hole_size < vma->vm_start) 1490 mm->cached_hole_size = vma->vm_start - addr; 1491 1492 /* try just below the current vma->vm_start */ 1493 addr = vma->vm_start-len; 1494 } while (len < vma->vm_start); 1495 1496bottomup: 1497 /* 1498 * A failed mmap() very likely causes application failure, 1499 * so fall back to the bottom-up function here. This scenario 1500 * can happen with large stack limits and large mmap() 1501 * allocations. 1502 */ 1503 mm->cached_hole_size = ~0UL; 1504 mm->free_area_cache = TASK_UNMAPPED_BASE; 1505 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags); 1506 /* 1507 * Restore the topdown base: 1508 */ 1509 mm->free_area_cache = mm->mmap_base; 1510 mm->cached_hole_size = ~0UL; 1511 1512 return addr; 1513} 1514#endif 1515 1516void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr) 1517{ 1518 /* 1519 * Is this a new hole at the highest possible address? 1520 */ 1521 if (addr > mm->free_area_cache) 1522 mm->free_area_cache = addr; 1523 1524 /* dont allow allocations above current base */ 1525 if (mm->free_area_cache > mm->mmap_base) 1526 mm->free_area_cache = mm->mmap_base; 1527} 1528 1529unsigned long 1530get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1531 unsigned long pgoff, unsigned long flags) 1532{ 1533 unsigned long (*get_area)(struct file *, unsigned long, 1534 unsigned long, unsigned long, unsigned long); 1535 1536 unsigned long error = arch_mmap_check(addr, len, flags); 1537 if (error) 1538 return error; 1539 1540 /* Careful about overflows.. */ 1541 if (len > TASK_SIZE) 1542 return -ENOMEM; 1543 1544 get_area = current->mm->get_unmapped_area; 1545 if (file && file->f_op && file->f_op->get_unmapped_area) 1546 get_area = file->f_op->get_unmapped_area; 1547 addr = get_area(file, addr, len, pgoff, flags); 1548 if (IS_ERR_VALUE(addr)) 1549 return addr; 1550 1551 if (addr > TASK_SIZE - len) 1552 return -ENOMEM; 1553 if (addr & ~PAGE_MASK) 1554 return -EINVAL; 1555 1556 return arch_rebalance_pgtables(addr, len); 1557} 1558 1559EXPORT_SYMBOL(get_unmapped_area); 1560 1561/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 1562struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 1563{ 1564 struct vm_area_struct *vma = NULL; 1565 1566 if (mm) { 1567 /* Check the cache first. */ 1568 /* (Cache hit rate is typically around 35%.) */ 1569 vma = mm->mmap_cache; 1570 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) { 1571 struct rb_node * rb_node; 1572 1573 rb_node = mm->mm_rb.rb_node; 1574 vma = NULL; 1575 1576 while (rb_node) { 1577 struct vm_area_struct * vma_tmp; 1578 1579 vma_tmp = rb_entry(rb_node, 1580 struct vm_area_struct, vm_rb); 1581 1582 if (vma_tmp->vm_end > addr) { 1583 vma = vma_tmp; 1584 if (vma_tmp->vm_start <= addr) 1585 break; 1586 rb_node = rb_node->rb_left; 1587 } else 1588 rb_node = rb_node->rb_right; 1589 } 1590 if (vma) 1591 mm->mmap_cache = vma; 1592 } 1593 } 1594 return vma; 1595} 1596 1597EXPORT_SYMBOL(find_vma); 1598 1599/* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */ 1600struct vm_area_struct * 1601find_vma_prev(struct mm_struct *mm, unsigned long addr, 1602 struct vm_area_struct **pprev) 1603{ 1604 struct vm_area_struct *vma = NULL, *prev = NULL; 1605 struct rb_node *rb_node; 1606 if (!mm) 1607 goto out; 1608 1609 /* Guard against addr being lower than the first VMA */ 1610 vma = mm->mmap; 1611 1612 /* Go through the RB tree quickly. */ 1613 rb_node = mm->mm_rb.rb_node; 1614 1615 while (rb_node) { 1616 struct vm_area_struct *vma_tmp; 1617 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 1618 1619 if (addr < vma_tmp->vm_end) { 1620 rb_node = rb_node->rb_left; 1621 } else { 1622 prev = vma_tmp; 1623 if (!prev->vm_next || (addr < prev->vm_next->vm_end)) 1624 break; 1625 rb_node = rb_node->rb_right; 1626 } 1627 } 1628 1629out: 1630 *pprev = prev; 1631 return prev ? prev->vm_next : vma; 1632} 1633 1634/* 1635 * Verify that the stack growth is acceptable and 1636 * update accounting. This is shared with both the 1637 * grow-up and grow-down cases. 1638 */ 1639static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 1640{ 1641 struct mm_struct *mm = vma->vm_mm; 1642 struct rlimit *rlim = current->signal->rlim; 1643 unsigned long new_start; 1644 1645 /* address space limit tests */ 1646 if (!may_expand_vm(mm, grow)) 1647 return -ENOMEM; 1648 1649 /* Stack limit test */ 1650 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 1651 return -ENOMEM; 1652 1653 /* mlock limit tests */ 1654 if (vma->vm_flags & VM_LOCKED) { 1655 unsigned long locked; 1656 unsigned long limit; 1657 locked = mm->locked_vm + grow; 1658 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 1659 limit >>= PAGE_SHIFT; 1660 if (locked > limit && !capable(CAP_IPC_LOCK)) 1661 return -ENOMEM; 1662 } 1663 1664 /* Check to ensure the stack will not grow into a hugetlb-only region */ 1665 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 1666 vma->vm_end - size; 1667 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 1668 return -EFAULT; 1669 1670 /* 1671 * Overcommit.. This must be the final test, as it will 1672 * update security statistics. 1673 */ 1674 if (security_vm_enough_memory_mm(mm, grow)) 1675 return -ENOMEM; 1676 1677 /* Ok, everything looks good - let it rip */ 1678 mm->total_vm += grow; 1679 if (vma->vm_flags & VM_LOCKED) 1680 mm->locked_vm += grow; 1681 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow); 1682 return 0; 1683} 1684 1685#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 1686/* 1687 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 1688 * vma is the last one with address > vma->vm_end. Have to extend vma. 1689 */ 1690int expand_upwards(struct vm_area_struct *vma, unsigned long address) 1691{ 1692 int error; 1693 1694 if (!(vma->vm_flags & VM_GROWSUP)) 1695 return -EFAULT; 1696 1697 /* 1698 * We must make sure the anon_vma is allocated 1699 * so that the anon_vma locking is not a noop. 1700 */ 1701 if (unlikely(anon_vma_prepare(vma))) 1702 return -ENOMEM; 1703 vma_lock_anon_vma(vma); 1704 1705 /* 1706 * vma->vm_start/vm_end cannot change under us because the caller 1707 * is required to hold the mmap_sem in read mode. We need the 1708 * anon_vma lock to serialize against concurrent expand_stacks. 1709 * Also guard against wrapping around to address 0. 1710 */ 1711 if (address < PAGE_ALIGN(address+4)) 1712 address = PAGE_ALIGN(address+4); 1713 else { 1714 vma_unlock_anon_vma(vma); 1715 return -ENOMEM; 1716 } 1717 error = 0; 1718 1719 /* Somebody else might have raced and expanded it already */ 1720 if (address > vma->vm_end) { 1721 unsigned long size, grow; 1722 1723 size = address - vma->vm_start; 1724 grow = (address - vma->vm_end) >> PAGE_SHIFT; 1725 1726 error = acct_stack_growth(vma, size, grow); 1727 if (!error) { 1728 vma->vm_end = address; 1729 perf_event_mmap(vma); 1730 } 1731 } 1732 vma_unlock_anon_vma(vma); 1733 return error; 1734} 1735#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 1736 1737/* 1738 * vma is the first one with address < vma->vm_start. Have to extend vma. 1739 */ 1740static int expand_downwards(struct vm_area_struct *vma, 1741 unsigned long address) 1742{ 1743 int error; 1744 1745 /* 1746 * We must make sure the anon_vma is allocated 1747 * so that the anon_vma locking is not a noop. 1748 */ 1749 if (unlikely(anon_vma_prepare(vma))) 1750 return -ENOMEM; 1751 1752 address &= PAGE_MASK; 1753 error = security_file_mmap(NULL, 0, 0, 0, address, 1); 1754 if (error) 1755 return error; 1756 1757 vma_lock_anon_vma(vma); 1758 1759 /* 1760 * vma->vm_start/vm_end cannot change under us because the caller 1761 * is required to hold the mmap_sem in read mode. We need the 1762 * anon_vma lock to serialize against concurrent expand_stacks. 1763 */ 1764 1765 /* Somebody else might have raced and expanded it already */ 1766 if (address < vma->vm_start) { 1767 unsigned long size, grow; 1768 1769 size = vma->vm_end - address; 1770 grow = (vma->vm_start - address) >> PAGE_SHIFT; 1771 1772 error = acct_stack_growth(vma, size, grow); 1773 if (!error) { 1774 vma->vm_start = address; 1775 vma->vm_pgoff -= grow; 1776 perf_event_mmap(vma); 1777 } 1778 } 1779 vma_unlock_anon_vma(vma); 1780 return error; 1781} 1782 1783int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address) 1784{ 1785 return expand_downwards(vma, address); 1786} 1787 1788#ifdef CONFIG_STACK_GROWSUP 1789int expand_stack(struct vm_area_struct *vma, unsigned long address) 1790{ 1791 return expand_upwards(vma, address); 1792} 1793 1794struct vm_area_struct * 1795find_extend_vma(struct mm_struct *mm, unsigned long addr) 1796{ 1797 struct vm_area_struct *vma, *prev; 1798 1799 addr &= PAGE_MASK; 1800 vma = find_vma_prev(mm, addr, &prev); 1801 if (vma && (vma->vm_start <= addr)) 1802 return vma; 1803 if (!prev || expand_stack(prev, addr)) 1804 return NULL; 1805 if (prev->vm_flags & VM_LOCKED) { 1806 mlock_vma_pages_range(prev, addr, prev->vm_end); 1807 } 1808 return prev; 1809} 1810#else 1811int expand_stack(struct vm_area_struct *vma, unsigned long address) 1812{ 1813 return expand_downwards(vma, address); 1814} 1815 1816struct vm_area_struct * 1817find_extend_vma(struct mm_struct * mm, unsigned long addr) 1818{ 1819 struct vm_area_struct * vma; 1820 unsigned long start; 1821 1822 addr &= PAGE_MASK; 1823 vma = find_vma(mm,addr); 1824 if (!vma) 1825 return NULL; 1826 if (vma->vm_start <= addr) 1827 return vma; 1828 if (!(vma->vm_flags & VM_GROWSDOWN)) 1829 return NULL; 1830 start = vma->vm_start; 1831 if (expand_stack(vma, addr)) 1832 return NULL; 1833 if (vma->vm_flags & VM_LOCKED) { 1834 mlock_vma_pages_range(vma, addr, start); 1835 } 1836 return vma; 1837} 1838#endif 1839 1840/* 1841 * Ok - we have the memory areas we should free on the vma list, 1842 * so release them, and do the vma updates. 1843 * 1844 * Called with the mm semaphore held. 1845 */ 1846static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 1847{ 1848 /* Update high watermark before we lower total_vm */ 1849 update_hiwater_vm(mm); 1850 do { 1851 long nrpages = vma_pages(vma); 1852 1853 mm->total_vm -= nrpages; 1854 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages); 1855 vma = remove_vma(vma); 1856 } while (vma); 1857 validate_mm(mm); 1858} 1859 1860/* 1861 * Get rid of page table information in the indicated region. 1862 * 1863 * Called with the mm semaphore held. 1864 */ 1865static void unmap_region(struct mm_struct *mm, 1866 struct vm_area_struct *vma, struct vm_area_struct *prev, 1867 unsigned long start, unsigned long end) 1868{ 1869 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap; 1870 struct mmu_gather *tlb; 1871 unsigned long nr_accounted = 0; 1872 1873 lru_add_drain(); 1874 tlb = tlb_gather_mmu(mm, 0); 1875 update_hiwater_rss(mm); 1876 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL); 1877 vm_unacct_memory(nr_accounted); 1878 free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS, 1879 next? next->vm_start: 0); 1880 tlb_finish_mmu(tlb, start, end); 1881} 1882 1883/* 1884 * Create a list of vma's touched by the unmap, removing them from the mm's 1885 * vma list as we go.. 1886 */ 1887static void 1888detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 1889 struct vm_area_struct *prev, unsigned long end) 1890{ 1891 struct vm_area_struct **insertion_point; 1892 struct vm_area_struct *tail_vma = NULL; 1893 unsigned long addr; 1894 1895 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 1896 vma->vm_prev = NULL; 1897 do { 1898 rb_erase(&vma->vm_rb, &mm->mm_rb); 1899 mm->map_count--; 1900 tail_vma = vma; 1901 vma = vma->vm_next; 1902 } while (vma && vma->vm_start < end); 1903 *insertion_point = vma; 1904 if (vma) 1905 vma->vm_prev = prev; 1906 tail_vma->vm_next = NULL; 1907 if (mm->unmap_area == arch_unmap_area) 1908 addr = prev ? prev->vm_end : mm->mmap_base; 1909 else 1910 addr = vma ? vma->vm_start : mm->mmap_base; 1911 mm->unmap_area(mm, addr); 1912 mm->mmap_cache = NULL; /* Kill the cache. */ 1913} 1914 1915/* 1916 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the 1917 * munmap path where it doesn't make sense to fail. 1918 */ 1919static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma, 1920 unsigned long addr, int new_below) 1921{ 1922 struct mempolicy *pol; 1923 struct vm_area_struct *new; 1924 int err = -ENOMEM; 1925 1926 if (is_vm_hugetlb_page(vma) && (addr & 1927 ~(huge_page_mask(hstate_vma(vma))))) 1928 return -EINVAL; 1929 1930 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 1931 if (!new) 1932 goto out_err; 1933 1934 /* most fields are the same, copy all, and then fixup */ 1935 *new = *vma; 1936 1937 INIT_LIST_HEAD(&new->anon_vma_chain); 1938 1939 if (new_below) 1940 new->vm_end = addr; 1941 else { 1942 new->vm_start = addr; 1943 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 1944 } 1945 1946 pol = mpol_dup(vma_policy(vma)); 1947 if (IS_ERR(pol)) { 1948 err = PTR_ERR(pol); 1949 goto out_free_vma; 1950 } 1951 vma_set_policy(new, pol); 1952 1953 if (anon_vma_clone(new, vma)) 1954 goto out_free_mpol; 1955 1956 if (new->vm_file) { 1957 get_file(new->vm_file); 1958 if (vma->vm_flags & VM_EXECUTABLE) 1959 added_exe_file_vma(mm); 1960 } 1961 1962 if (new->vm_ops && new->vm_ops->open) 1963 new->vm_ops->open(new); 1964 1965 if (new_below) 1966 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 1967 ((addr - new->vm_start) >> PAGE_SHIFT), new); 1968 else 1969 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 1970 1971 /* Success. */ 1972 if (!err) 1973 return 0; 1974 1975 /* Clean everything up if vma_adjust failed. */ 1976 if (new->vm_ops && new->vm_ops->close) 1977 new->vm_ops->close(new); 1978 if (new->vm_file) { 1979 if (vma->vm_flags & VM_EXECUTABLE) 1980 removed_exe_file_vma(mm); 1981 fput(new->vm_file); 1982 } 1983 unlink_anon_vmas(new); 1984 out_free_mpol: 1985 mpol_put(pol); 1986 out_free_vma: 1987 kmem_cache_free(vm_area_cachep, new); 1988 out_err: 1989 return err; 1990} 1991 1992/* 1993 * Split a vma into two pieces at address 'addr', a new vma is allocated 1994 * either for the first part or the tail. 1995 */ 1996int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 1997 unsigned long addr, int new_below) 1998{ 1999 if (mm->map_count >= sysctl_max_map_count) 2000 return -ENOMEM; 2001 2002 return __split_vma(mm, vma, addr, new_below); 2003} 2004 2005/* Munmap is split into 2 main parts -- this part which finds 2006 * what needs doing, and the areas themselves, which do the 2007 * work. This now handles partial unmappings. 2008 * Jeremy Fitzhardinge <jeremy@goop.org> 2009 */ 2010int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 2011{ 2012 unsigned long end; 2013 struct vm_area_struct *vma, *prev, *last; 2014 2015 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start) 2016 return -EINVAL; 2017 2018 if ((len = PAGE_ALIGN(len)) == 0) 2019 return -EINVAL; 2020 2021 /* Find the first overlapping VMA */ 2022 vma = find_vma_prev(mm, start, &prev); 2023 if (!vma) 2024 return 0; 2025 /* we have start < vma->vm_end */ 2026 2027 /* if it doesn't overlap, we have nothing.. */ 2028 end = start + len; 2029 if (vma->vm_start >= end) 2030 return 0; 2031 2032 /* 2033 * If we need to split any vma, do it now to save pain later. 2034 * 2035 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2036 * unmapped vm_area_struct will remain in use: so lower split_vma 2037 * places tmp vma above, and higher split_vma places tmp vma below. 2038 */ 2039 if (start > vma->vm_start) { 2040 int error; 2041 2042 /* 2043 * Make sure that map_count on return from munmap() will 2044 * not exceed its limit; but let map_count go just above 2045 * its limit temporarily, to help free resources as expected. 2046 */ 2047 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2048 return -ENOMEM; 2049 2050 error = __split_vma(mm, vma, start, 0); 2051 if (error) 2052 return error; 2053 prev = vma; 2054 } 2055 2056 /* Does it split the last one? */ 2057 last = find_vma(mm, end); 2058 if (last && end > last->vm_start) { 2059 int error = __split_vma(mm, last, end, 1); 2060 if (error) 2061 return error; 2062 } 2063 vma = prev? prev->vm_next: mm->mmap; 2064 2065 /* 2066 * unlock any mlock()ed ranges before detaching vmas 2067 */ 2068 if (mm->locked_vm) { 2069 struct vm_area_struct *tmp = vma; 2070 while (tmp && tmp->vm_start < end) { 2071 if (tmp->vm_flags & VM_LOCKED) { 2072 mm->locked_vm -= vma_pages(tmp); 2073 munlock_vma_pages_all(tmp); 2074 } 2075 tmp = tmp->vm_next; 2076 } 2077 } 2078 2079 /* 2080 * Remove the vma's, and unmap the actual pages 2081 */ 2082 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2083 unmap_region(mm, vma, prev, start, end); 2084 2085 /* Fix up all other VM information */ 2086 remove_vma_list(mm, vma); 2087 2088 return 0; 2089} 2090 2091EXPORT_SYMBOL(do_munmap); 2092 2093SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2094{ 2095 int ret; 2096 struct mm_struct *mm = current->mm; 2097 2098 profile_munmap(addr); 2099 2100 down_write(&mm->mmap_sem); 2101 ret = do_munmap(mm, addr, len); 2102 up_write(&mm->mmap_sem); 2103 return ret; 2104} 2105 2106static inline void verify_mm_writelocked(struct mm_struct *mm) 2107{ 2108#ifdef CONFIG_DEBUG_VM 2109 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2110 WARN_ON(1); 2111 up_read(&mm->mmap_sem); 2112 } 2113#endif 2114} 2115 2116/* 2117 * this is really a simplified "do_mmap". it only handles 2118 * anonymous maps. eventually we may be able to do some 2119 * brk-specific accounting here. 2120 */ 2121unsigned long do_brk(unsigned long addr, unsigned long len) 2122{ 2123 struct mm_struct * mm = current->mm; 2124 struct vm_area_struct * vma, * prev; 2125 unsigned long flags; 2126 struct rb_node ** rb_link, * rb_parent; 2127 pgoff_t pgoff = addr >> PAGE_SHIFT; 2128 int error; 2129 2130 len = PAGE_ALIGN(len); 2131 if (!len) 2132 return addr; 2133 2134 error = security_file_mmap(NULL, 0, 0, 0, addr, 1); 2135 if (error) 2136 return error; 2137 2138 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2139 2140 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2141 if (error & ~PAGE_MASK) 2142 return error; 2143 2144 /* 2145 * mlock MCL_FUTURE? 2146 */ 2147 if (mm->def_flags & VM_LOCKED) { 2148 unsigned long locked, lock_limit; 2149 locked = len >> PAGE_SHIFT; 2150 locked += mm->locked_vm; 2151 lock_limit = rlimit(RLIMIT_MEMLOCK); 2152 lock_limit >>= PAGE_SHIFT; 2153 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 2154 return -EAGAIN; 2155 } 2156 2157 /* 2158 * mm->mmap_sem is required to protect against another thread 2159 * changing the mappings in case we sleep. 2160 */ 2161 verify_mm_writelocked(mm); 2162 2163 /* 2164 * Clear old maps. this also does some error checking for us 2165 */ 2166 munmap_back: 2167 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 2168 if (vma && vma->vm_start < addr + len) { 2169 if (do_munmap(mm, addr, len)) 2170 return -ENOMEM; 2171 goto munmap_back; 2172 } 2173 2174 /* Check against address space limits *after* clearing old maps... */ 2175 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 2176 return -ENOMEM; 2177 2178 if (mm->map_count > sysctl_max_map_count) 2179 return -ENOMEM; 2180 2181 if (security_vm_enough_memory(len >> PAGE_SHIFT)) 2182 return -ENOMEM; 2183 2184 /* Can we just expand an old private anonymous mapping? */ 2185 vma = vma_merge(mm, prev, addr, addr + len, flags, 2186 NULL, NULL, pgoff, NULL); 2187 if (vma) 2188 goto out; 2189 2190 /* 2191 * create a vma struct for an anonymous mapping 2192 */ 2193 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2194 if (!vma) { 2195 vm_unacct_memory(len >> PAGE_SHIFT); 2196 return -ENOMEM; 2197 } 2198 2199 INIT_LIST_HEAD(&vma->anon_vma_chain); 2200 vma->vm_mm = mm; 2201 vma->vm_start = addr; 2202 vma->vm_end = addr + len; 2203 vma->vm_pgoff = pgoff; 2204 vma->vm_flags = flags; 2205 vma->vm_page_prot = vm_get_page_prot(flags); 2206 vma_link(mm, vma, prev, rb_link, rb_parent); 2207out: 2208 perf_event_mmap(vma); 2209 mm->total_vm += len >> PAGE_SHIFT; 2210 if (flags & VM_LOCKED) { 2211 if (!mlock_vma_pages_range(vma, addr, addr + len)) 2212 mm->locked_vm += (len >> PAGE_SHIFT); 2213 } 2214 return addr; 2215} 2216 2217EXPORT_SYMBOL(do_brk); 2218 2219/* Release all mmaps. */ 2220void exit_mmap(struct mm_struct *mm) 2221{ 2222 struct mmu_gather *tlb; 2223 struct vm_area_struct *vma; 2224 unsigned long nr_accounted = 0; 2225 unsigned long end; 2226 2227 /* mm's last user has gone, and its about to be pulled down */ 2228 mmu_notifier_release(mm); 2229 2230 if (mm->locked_vm) { 2231 vma = mm->mmap; 2232 while (vma) { 2233 if (vma->vm_flags & VM_LOCKED) 2234 munlock_vma_pages_all(vma); 2235 vma = vma->vm_next; 2236 } 2237 } 2238 2239 arch_exit_mmap(mm); 2240 2241 vma = mm->mmap; 2242 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2243 return; 2244 2245 lru_add_drain(); 2246 flush_cache_mm(mm); 2247 tlb = tlb_gather_mmu(mm, 1); 2248 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2249 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2250 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL); 2251 vm_unacct_memory(nr_accounted); 2252 2253 free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0); 2254 tlb_finish_mmu(tlb, 0, end); 2255 2256 /* 2257 * Walk the list again, actually closing and freeing it, 2258 * with preemption enabled, without holding any MM locks. 2259 */ 2260 while (vma) 2261 vma = remove_vma(vma); 2262 2263 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT); 2264} 2265 2266/* Insert vm structure into process list sorted by address 2267 * and into the inode's i_mmap tree. If vm_file is non-NULL 2268 * then i_mmap_lock is taken here. 2269 */ 2270int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma) 2271{ 2272 struct vm_area_struct * __vma, * prev; 2273 struct rb_node ** rb_link, * rb_parent; 2274 2275 /* 2276 * The vm_pgoff of a purely anonymous vma should be irrelevant 2277 * until its first write fault, when page's anon_vma and index 2278 * are set. But now set the vm_pgoff it will almost certainly 2279 * end up with (unless mremap moves it elsewhere before that 2280 * first wfault), so /proc/pid/maps tells a consistent story. 2281 * 2282 * By setting it to reflect the virtual start address of the 2283 * vma, merges and splits can happen in a seamless way, just 2284 * using the existing file pgoff checks and manipulations. 2285 * Similarly in do_mmap_pgoff and in do_brk. 2286 */ 2287 if (!vma->vm_file) { 2288 BUG_ON(vma->anon_vma); 2289 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2290 } 2291 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent); 2292 if (__vma && __vma->vm_start < vma->vm_end) 2293 return -ENOMEM; 2294 if ((vma->vm_flags & VM_ACCOUNT) && 2295 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2296 return -ENOMEM; 2297 vma_link(mm, vma, prev, rb_link, rb_parent); 2298 return 0; 2299} 2300 2301/* 2302 * Copy the vma structure to a new location in the same mm, 2303 * prior to moving page table entries, to effect an mremap move. 2304 */ 2305struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2306 unsigned long addr, unsigned long len, pgoff_t pgoff) 2307{ 2308 struct vm_area_struct *vma = *vmap; 2309 unsigned long vma_start = vma->vm_start; 2310 struct mm_struct *mm = vma->vm_mm; 2311 struct vm_area_struct *new_vma, *prev; 2312 struct rb_node **rb_link, *rb_parent; 2313 struct mempolicy *pol; 2314 2315 /* 2316 * If anonymous vma has not yet been faulted, update new pgoff 2317 * to match new location, to increase its chance of merging. 2318 */ 2319 if (!vma->vm_file && !vma->anon_vma) 2320 pgoff = addr >> PAGE_SHIFT; 2321 2322 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 2323 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2324 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma)); 2325 if (new_vma) { 2326 /* 2327 * Source vma may have been merged into new_vma 2328 */ 2329 if (vma_start >= new_vma->vm_start && 2330 vma_start < new_vma->vm_end) 2331 *vmap = new_vma; 2332 } else { 2333 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2334 if (new_vma) { 2335 *new_vma = *vma; 2336 pol = mpol_dup(vma_policy(vma)); 2337 if (IS_ERR(pol)) 2338 goto out_free_vma; 2339 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 2340 if (anon_vma_clone(new_vma, vma)) 2341 goto out_free_mempol; 2342 vma_set_policy(new_vma, pol); 2343 new_vma->vm_start = addr; 2344 new_vma->vm_end = addr + len; 2345 new_vma->vm_pgoff = pgoff; 2346 if (new_vma->vm_file) { 2347 get_file(new_vma->vm_file); 2348 if (vma->vm_flags & VM_EXECUTABLE) 2349 added_exe_file_vma(mm); 2350 } 2351 if (new_vma->vm_ops && new_vma->vm_ops->open) 2352 new_vma->vm_ops->open(new_vma); 2353 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2354 } 2355 } 2356 return new_vma; 2357 2358 out_free_mempol: 2359 mpol_put(pol); 2360 out_free_vma: 2361 kmem_cache_free(vm_area_cachep, new_vma); 2362 return NULL; 2363} 2364 2365/* 2366 * Return true if the calling process may expand its vm space by the passed 2367 * number of pages 2368 */ 2369int may_expand_vm(struct mm_struct *mm, unsigned long npages) 2370{ 2371 unsigned long cur = mm->total_vm; /* pages */ 2372 unsigned long lim; 2373 2374 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT; 2375 2376 if (cur + npages > lim) 2377 return 0; 2378 return 1; 2379} 2380 2381 2382static int special_mapping_fault(struct vm_area_struct *vma, 2383 struct vm_fault *vmf) 2384{ 2385 pgoff_t pgoff; 2386 struct page **pages; 2387 2388 /* 2389 * special mappings have no vm_file, and in that case, the mm 2390 * uses vm_pgoff internally. So we have to subtract it from here. 2391 * We are allowed to do this because we are the mm; do not copy 2392 * this code into drivers! 2393 */ 2394 pgoff = vmf->pgoff - vma->vm_pgoff; 2395 2396 for (pages = vma->vm_private_data; pgoff && *pages; ++pages) 2397 pgoff--; 2398 2399 if (*pages) { 2400 struct page *page = *pages; 2401 get_page(page); 2402 vmf->page = page; 2403 return 0; 2404 } 2405 2406 return VM_FAULT_SIGBUS; 2407} 2408 2409/* 2410 * Having a close hook prevents vma merging regardless of flags. 2411 */ 2412static void special_mapping_close(struct vm_area_struct *vma) 2413{ 2414} 2415 2416static const struct vm_operations_struct special_mapping_vmops = { 2417 .close = special_mapping_close, 2418 .fault = special_mapping_fault, 2419}; 2420 2421/* 2422 * Called with mm->mmap_sem held for writing. 2423 * Insert a new vma covering the given region, with the given flags. 2424 * Its pages are supplied by the given array of struct page *. 2425 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 2426 * The region past the last page supplied will always produce SIGBUS. 2427 * The array pointer and the pages it points to are assumed to stay alive 2428 * for as long as this mapping might exist. 2429 */ 2430int install_special_mapping(struct mm_struct *mm, 2431 unsigned long addr, unsigned long len, 2432 unsigned long vm_flags, struct page **pages) 2433{ 2434 int ret; 2435 struct vm_area_struct *vma; 2436 2437 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2438 if (unlikely(vma == NULL)) 2439 return -ENOMEM; 2440 2441 INIT_LIST_HEAD(&vma->anon_vma_chain); 2442 vma->vm_mm = mm; 2443 vma->vm_start = addr; 2444 vma->vm_end = addr + len; 2445 2446 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND; 2447 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 2448 2449 vma->vm_ops = &special_mapping_vmops; 2450 vma->vm_private_data = pages; 2451 2452 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1); 2453 if (ret) 2454 goto out; 2455 2456 ret = insert_vm_struct(mm, vma); 2457 if (ret) 2458 goto out; 2459 2460 mm->total_vm += len >> PAGE_SHIFT; 2461 2462 perf_event_mmap(vma); 2463 2464 return 0; 2465 2466out: 2467 kmem_cache_free(vm_area_cachep, vma); 2468 return ret; 2469} 2470 2471static DEFINE_MUTEX(mm_all_locks_mutex); 2472 2473static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 2474{ 2475 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) { 2476 /* 2477 * The LSB of head.next can't change from under us 2478 * because we hold the mm_all_locks_mutex. 2479 */ 2480 spin_lock_nest_lock(&anon_vma->root->lock, &mm->mmap_sem); 2481 /* 2482 * We can safely modify head.next after taking the 2483 * anon_vma->root->lock. If some other vma in this mm shares 2484 * the same anon_vma we won't take it again. 2485 * 2486 * No need of atomic instructions here, head.next 2487 * can't change from under us thanks to the 2488 * anon_vma->root->lock. 2489 */ 2490 if (__test_and_set_bit(0, (unsigned long *) 2491 &anon_vma->root->head.next)) 2492 BUG(); 2493 } 2494} 2495 2496static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 2497{ 2498 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 2499 /* 2500 * AS_MM_ALL_LOCKS can't change from under us because 2501 * we hold the mm_all_locks_mutex. 2502 * 2503 * Operations on ->flags have to be atomic because 2504 * even if AS_MM_ALL_LOCKS is stable thanks to the 2505 * mm_all_locks_mutex, there may be other cpus 2506 * changing other bitflags in parallel to us. 2507 */ 2508 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 2509 BUG(); 2510 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem); 2511 } 2512} 2513 2514/* 2515 * This operation locks against the VM for all pte/vma/mm related 2516 * operations that could ever happen on a certain mm. This includes 2517 * vmtruncate, try_to_unmap, and all page faults. 2518 * 2519 * The caller must take the mmap_sem in write mode before calling 2520 * mm_take_all_locks(). The caller isn't allowed to release the 2521 * mmap_sem until mm_drop_all_locks() returns. 2522 * 2523 * mmap_sem in write mode is required in order to block all operations 2524 * that could modify pagetables and free pages without need of 2525 * altering the vma layout (for example populate_range() with 2526 * nonlinear vmas). It's also needed in write mode to avoid new 2527 * anon_vmas to be associated with existing vmas. 2528 * 2529 * A single task can't take more than one mm_take_all_locks() in a row 2530 * or it would deadlock. 2531 * 2532 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in 2533 * mapping->flags avoid to take the same lock twice, if more than one 2534 * vma in this mm is backed by the same anon_vma or address_space. 2535 * 2536 * We can take all the locks in random order because the VM code 2537 * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never 2538 * takes more than one of them in a row. Secondly we're protected 2539 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex. 2540 * 2541 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 2542 * that may have to take thousand of locks. 2543 * 2544 * mm_take_all_locks() can fail if it's interrupted by signals. 2545 */ 2546int mm_take_all_locks(struct mm_struct *mm) 2547{ 2548 struct vm_area_struct *vma; 2549 struct anon_vma_chain *avc; 2550 int ret = -EINTR; 2551 2552 BUG_ON(down_read_trylock(&mm->mmap_sem)); 2553 2554 mutex_lock(&mm_all_locks_mutex); 2555 2556 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2557 if (signal_pending(current)) 2558 goto out_unlock; 2559 if (vma->vm_file && vma->vm_file->f_mapping) 2560 vm_lock_mapping(mm, vma->vm_file->f_mapping); 2561 } 2562 2563 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2564 if (signal_pending(current)) 2565 goto out_unlock; 2566 if (vma->anon_vma) 2567 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 2568 vm_lock_anon_vma(mm, avc->anon_vma); 2569 } 2570 2571 ret = 0; 2572 2573out_unlock: 2574 if (ret) 2575 mm_drop_all_locks(mm); 2576 2577 return ret; 2578} 2579 2580static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 2581{ 2582 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) { 2583 /* 2584 * The LSB of head.next can't change to 0 from under 2585 * us because we hold the mm_all_locks_mutex. 2586 * 2587 * We must however clear the bitflag before unlocking 2588 * the vma so the users using the anon_vma->head will 2589 * never see our bitflag. 2590 * 2591 * No need of atomic instructions here, head.next 2592 * can't change from under us until we release the 2593 * anon_vma->root->lock. 2594 */ 2595 if (!__test_and_clear_bit(0, (unsigned long *) 2596 &anon_vma->root->head.next)) 2597 BUG(); 2598 anon_vma_unlock(anon_vma); 2599 } 2600} 2601 2602static void vm_unlock_mapping(struct address_space *mapping) 2603{ 2604 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 2605 /* 2606 * AS_MM_ALL_LOCKS can't change to 0 from under us 2607 * because we hold the mm_all_locks_mutex. 2608 */ 2609 spin_unlock(&mapping->i_mmap_lock); 2610 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 2611 &mapping->flags)) 2612 BUG(); 2613 } 2614} 2615 2616/* 2617 * The mmap_sem cannot be released by the caller until 2618 * mm_drop_all_locks() returns. 2619 */ 2620void mm_drop_all_locks(struct mm_struct *mm) 2621{ 2622 struct vm_area_struct *vma; 2623 struct anon_vma_chain *avc; 2624 2625 BUG_ON(down_read_trylock(&mm->mmap_sem)); 2626 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 2627 2628 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2629 if (vma->anon_vma) 2630 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 2631 vm_unlock_anon_vma(avc->anon_vma); 2632 if (vma->vm_file && vma->vm_file->f_mapping) 2633 vm_unlock_mapping(vma->vm_file->f_mapping); 2634 } 2635 2636 mutex_unlock(&mm_all_locks_mutex); 2637} 2638 2639/* 2640 * initialise the VMA slab 2641 */ 2642void __init mmap_init(void) 2643{ 2644 int ret; 2645 2646 ret = percpu_counter_init(&vm_committed_as, 0); 2647 VM_BUG_ON(ret); 2648} 2649