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/export.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 #include <linux/audit.h> 32 #include <linux/khugepaged.h> 33 34 #include <asm/uaccess.h> 35 #include <asm/cacheflush.h> 36 #include <asm/tlb.h> 37 #include <asm/mmu_context.h> 38 39 #include "internal.h" 40 41 #ifndef arch_mmap_check 42 #define arch_mmap_check(addr, len, flags) (0) 43 #endif 44 45 #ifndef arch_rebalance_pgtables 46 #define arch_rebalance_pgtables(addr, len) (addr) 47 #endif 48 49 static void unmap_region(struct mm_struct *mm, 50 struct vm_area_struct *vma, struct vm_area_struct *prev, 51 unsigned long start, unsigned long end); 52 53 /* 54 * WARNING: the debugging will use recursive algorithms so never enable this 55 * unless you know what you are doing. 56 */ 57 #undef DEBUG_MM_RB 58 59 /* description of effects of mapping type and prot in current implementation. 60 * this is due to the limited x86 page protection hardware. The expected 61 * behavior is in parens: 62 * 63 * map_type prot 64 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 65 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 66 * w: (no) no w: (no) no w: (yes) yes w: (no) no 67 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 68 * 69 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 70 * w: (no) no w: (no) no w: (copy) copy w: (no) no 71 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 72 * 73 */ 74 pgprot_t protection_map[16] = { 75 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 76 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 77 }; 78 79 pgprot_t vm_get_page_prot(unsigned long vm_flags) 80 { 81 return __pgprot(pgprot_val(protection_map[vm_flags & 82 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) | 83 pgprot_val(arch_vm_get_page_prot(vm_flags))); 84 } 85 EXPORT_SYMBOL(vm_get_page_prot); 86 87 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */ 88 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */ 89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT; 90 /* 91 * Make sure vm_committed_as in one cacheline and not cacheline shared with 92 * other variables. It can be updated by several CPUs frequently. 93 */ 94 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp; 95 96 /* 97 * Check that a process has enough memory to allocate a new virtual 98 * mapping. 0 means there is enough memory for the allocation to 99 * succeed and -ENOMEM implies there is not. 100 * 101 * We currently support three overcommit policies, which are set via the 102 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting 103 * 104 * Strict overcommit modes added 2002 Feb 26 by Alan Cox. 105 * Additional code 2002 Jul 20 by Robert Love. 106 * 107 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise. 108 * 109 * Note this is a helper function intended to be used by LSMs which 110 * wish to use this logic. 111 */ 112 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin) 113 { 114 unsigned long free, allowed; 115 116 vm_acct_memory(pages); 117 118 /* 119 * Sometimes we want to use more memory than we have 120 */ 121 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS) 122 return 0; 123 124 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) { 125 free = global_page_state(NR_FREE_PAGES); 126 free += global_page_state(NR_FILE_PAGES); 127 128 /* 129 * shmem pages shouldn't be counted as free in this 130 * case, they can't be purged, only swapped out, and 131 * that won't affect the overall amount of available 132 * memory in the system. 133 */ 134 free -= global_page_state(NR_SHMEM); 135 136 free += nr_swap_pages; 137 138 /* 139 * Any slabs which are created with the 140 * SLAB_RECLAIM_ACCOUNT flag claim to have contents 141 * which are reclaimable, under pressure. The dentry 142 * cache and most inode caches should fall into this 143 */ 144 free += global_page_state(NR_SLAB_RECLAIMABLE); 145 146 /* 147 * Leave reserved pages. The pages are not for anonymous pages. 148 */ 149 if (free <= totalreserve_pages) 150 goto error; 151 else 152 free -= totalreserve_pages; 153 154 /* 155 * Leave the last 3% for root 156 */ 157 if (!cap_sys_admin) 158 free -= free / 32; 159 160 if (free > pages) 161 return 0; 162 163 goto error; 164 } 165 166 allowed = (totalram_pages - hugetlb_total_pages()) 167 * sysctl_overcommit_ratio / 100; 168 /* 169 * Leave the last 3% for root 170 */ 171 if (!cap_sys_admin) 172 allowed -= allowed / 32; 173 allowed += total_swap_pages; 174 175 /* Don't let a single process grow too big: 176 leave 3% of the size of this process for other processes */ 177 if (mm) 178 allowed -= mm->total_vm / 32; 179 180 if (percpu_counter_read_positive(&vm_committed_as) < allowed) 181 return 0; 182 error: 183 vm_unacct_memory(pages); 184 185 return -ENOMEM; 186 } 187 188 /* 189 * Requires inode->i_mapping->i_mmap_mutex 190 */ 191 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 192 struct file *file, struct address_space *mapping) 193 { 194 if (vma->vm_flags & VM_DENYWRITE) 195 atomic_inc(&file->f_path.dentry->d_inode->i_writecount); 196 if (vma->vm_flags & VM_SHARED) 197 mapping->i_mmap_writable--; 198 199 flush_dcache_mmap_lock(mapping); 200 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 201 list_del_init(&vma->shared.vm_set.list); 202 else 203 vma_prio_tree_remove(vma, &mapping->i_mmap); 204 flush_dcache_mmap_unlock(mapping); 205 } 206 207 /* 208 * Unlink a file-based vm structure from its prio_tree, to hide 209 * vma from rmap and vmtruncate before freeing its page tables. 210 */ 211 void unlink_file_vma(struct vm_area_struct *vma) 212 { 213 struct file *file = vma->vm_file; 214 215 if (file) { 216 struct address_space *mapping = file->f_mapping; 217 mutex_lock(&mapping->i_mmap_mutex); 218 __remove_shared_vm_struct(vma, file, mapping); 219 mutex_unlock(&mapping->i_mmap_mutex); 220 } 221 } 222 223 /* 224 * Close a vm structure and free it, returning the next. 225 */ 226 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 227 { 228 struct vm_area_struct *next = vma->vm_next; 229 230 might_sleep(); 231 if (vma->vm_ops && vma->vm_ops->close) 232 vma->vm_ops->close(vma); 233 if (vma->vm_file) { 234 fput(vma->vm_file); 235 if (vma->vm_flags & VM_EXECUTABLE) 236 removed_exe_file_vma(vma->vm_mm); 237 } 238 mpol_put(vma_policy(vma)); 239 kmem_cache_free(vm_area_cachep, vma); 240 return next; 241 } 242 243 SYSCALL_DEFINE1(brk, unsigned long, brk) 244 { 245 unsigned long rlim, retval; 246 unsigned long newbrk, oldbrk; 247 struct mm_struct *mm = current->mm; 248 unsigned long min_brk; 249 250 down_write(&mm->mmap_sem); 251 252 #ifdef CONFIG_COMPAT_BRK 253 /* 254 * CONFIG_COMPAT_BRK can still be overridden by setting 255 * randomize_va_space to 2, which will still cause mm->start_brk 256 * to be arbitrarily shifted 257 */ 258 if (current->brk_randomized) 259 min_brk = mm->start_brk; 260 else 261 min_brk = mm->end_data; 262 #else 263 min_brk = mm->start_brk; 264 #endif 265 if (brk < min_brk) 266 goto out; 267 268 /* 269 * Check against rlimit here. If this check is done later after the test 270 * of oldbrk with newbrk then it can escape the test and let the data 271 * segment grow beyond its set limit the in case where the limit is 272 * not page aligned -Ram Gupta 273 */ 274 rlim = rlimit(RLIMIT_DATA); 275 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) + 276 (mm->end_data - mm->start_data) > rlim) 277 goto out; 278 279 newbrk = PAGE_ALIGN(brk); 280 oldbrk = PAGE_ALIGN(mm->brk); 281 if (oldbrk == newbrk) 282 goto set_brk; 283 284 /* Always allow shrinking brk. */ 285 if (brk <= mm->brk) { 286 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 287 goto set_brk; 288 goto out; 289 } 290 291 /* Check against existing mmap mappings. */ 292 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 293 goto out; 294 295 /* Ok, looks good - let it rip. */ 296 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 297 goto out; 298 set_brk: 299 mm->brk = brk; 300 out: 301 retval = mm->brk; 302 up_write(&mm->mmap_sem); 303 return retval; 304 } 305 306 #ifdef DEBUG_MM_RB 307 static int browse_rb(struct rb_root *root) 308 { 309 int i = 0, j; 310 struct rb_node *nd, *pn = NULL; 311 unsigned long prev = 0, pend = 0; 312 313 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 314 struct vm_area_struct *vma; 315 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 316 if (vma->vm_start < prev) 317 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1; 318 if (vma->vm_start < pend) 319 printk("vm_start %lx pend %lx\n", vma->vm_start, pend); 320 if (vma->vm_start > vma->vm_end) 321 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start); 322 i++; 323 pn = nd; 324 prev = vma->vm_start; 325 pend = vma->vm_end; 326 } 327 j = 0; 328 for (nd = pn; nd; nd = rb_prev(nd)) { 329 j++; 330 } 331 if (i != j) 332 printk("backwards %d, forwards %d\n", j, i), i = 0; 333 return i; 334 } 335 336 void validate_mm(struct mm_struct *mm) 337 { 338 int bug = 0; 339 int i = 0; 340 struct vm_area_struct *tmp = mm->mmap; 341 while (tmp) { 342 tmp = tmp->vm_next; 343 i++; 344 } 345 if (i != mm->map_count) 346 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1; 347 i = browse_rb(&mm->mm_rb); 348 if (i != mm->map_count) 349 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1; 350 BUG_ON(bug); 351 } 352 #else 353 #define validate_mm(mm) do { } while (0) 354 #endif 355 356 static struct vm_area_struct * 357 find_vma_prepare(struct mm_struct *mm, unsigned long addr, 358 struct vm_area_struct **pprev, struct rb_node ***rb_link, 359 struct rb_node ** rb_parent) 360 { 361 struct vm_area_struct * vma; 362 struct rb_node ** __rb_link, * __rb_parent, * rb_prev; 363 364 __rb_link = &mm->mm_rb.rb_node; 365 rb_prev = __rb_parent = NULL; 366 vma = NULL; 367 368 while (*__rb_link) { 369 struct vm_area_struct *vma_tmp; 370 371 __rb_parent = *__rb_link; 372 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 373 374 if (vma_tmp->vm_end > addr) { 375 vma = vma_tmp; 376 if (vma_tmp->vm_start <= addr) 377 break; 378 __rb_link = &__rb_parent->rb_left; 379 } else { 380 rb_prev = __rb_parent; 381 __rb_link = &__rb_parent->rb_right; 382 } 383 } 384 385 *pprev = NULL; 386 if (rb_prev) 387 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 388 *rb_link = __rb_link; 389 *rb_parent = __rb_parent; 390 return vma; 391 } 392 393 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 394 struct rb_node **rb_link, struct rb_node *rb_parent) 395 { 396 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 397 rb_insert_color(&vma->vm_rb, &mm->mm_rb); 398 } 399 400 static void __vma_link_file(struct vm_area_struct *vma) 401 { 402 struct file *file; 403 404 file = vma->vm_file; 405 if (file) { 406 struct address_space *mapping = file->f_mapping; 407 408 if (vma->vm_flags & VM_DENYWRITE) 409 atomic_dec(&file->f_path.dentry->d_inode->i_writecount); 410 if (vma->vm_flags & VM_SHARED) 411 mapping->i_mmap_writable++; 412 413 flush_dcache_mmap_lock(mapping); 414 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 415 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear); 416 else 417 vma_prio_tree_insert(vma, &mapping->i_mmap); 418 flush_dcache_mmap_unlock(mapping); 419 } 420 } 421 422 static void 423 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 424 struct vm_area_struct *prev, struct rb_node **rb_link, 425 struct rb_node *rb_parent) 426 { 427 __vma_link_list(mm, vma, prev, rb_parent); 428 __vma_link_rb(mm, vma, rb_link, rb_parent); 429 } 430 431 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 432 struct vm_area_struct *prev, struct rb_node **rb_link, 433 struct rb_node *rb_parent) 434 { 435 struct address_space *mapping = NULL; 436 437 if (vma->vm_file) 438 mapping = vma->vm_file->f_mapping; 439 440 if (mapping) 441 mutex_lock(&mapping->i_mmap_mutex); 442 443 __vma_link(mm, vma, prev, rb_link, rb_parent); 444 __vma_link_file(vma); 445 446 if (mapping) 447 mutex_unlock(&mapping->i_mmap_mutex); 448 449 mm->map_count++; 450 validate_mm(mm); 451 } 452 453 /* 454 * Helper for vma_adjust in the split_vma insert case: 455 * insert vm structure into list and rbtree and anon_vma, 456 * but it has already been inserted into prio_tree earlier. 457 */ 458 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 459 { 460 struct vm_area_struct *__vma, *prev; 461 struct rb_node **rb_link, *rb_parent; 462 463 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent); 464 BUG_ON(__vma && __vma->vm_start < vma->vm_end); 465 __vma_link(mm, vma, prev, rb_link, rb_parent); 466 mm->map_count++; 467 } 468 469 static inline void 470 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 471 struct vm_area_struct *prev) 472 { 473 struct vm_area_struct *next = vma->vm_next; 474 475 prev->vm_next = next; 476 if (next) 477 next->vm_prev = prev; 478 rb_erase(&vma->vm_rb, &mm->mm_rb); 479 if (mm->mmap_cache == vma) 480 mm->mmap_cache = prev; 481 } 482 483 /* 484 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 485 * is already present in an i_mmap tree without adjusting the tree. 486 * The following helper function should be used when such adjustments 487 * are necessary. The "insert" vma (if any) is to be inserted 488 * before we drop the necessary locks. 489 */ 490 int vma_adjust(struct vm_area_struct *vma, unsigned long start, 491 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 492 { 493 struct mm_struct *mm = vma->vm_mm; 494 struct vm_area_struct *next = vma->vm_next; 495 struct vm_area_struct *importer = NULL; 496 struct address_space *mapping = NULL; 497 struct prio_tree_root *root = NULL; 498 struct anon_vma *anon_vma = NULL; 499 struct file *file = vma->vm_file; 500 long adjust_next = 0; 501 int remove_next = 0; 502 503 if (next && !insert) { 504 struct vm_area_struct *exporter = NULL; 505 506 if (end >= next->vm_end) { 507 /* 508 * vma expands, overlapping all the next, and 509 * perhaps the one after too (mprotect case 6). 510 */ 511 again: remove_next = 1 + (end > next->vm_end); 512 end = next->vm_end; 513 exporter = next; 514 importer = vma; 515 } else if (end > next->vm_start) { 516 /* 517 * vma expands, overlapping part of the next: 518 * mprotect case 5 shifting the boundary up. 519 */ 520 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 521 exporter = next; 522 importer = vma; 523 } else if (end < vma->vm_end) { 524 /* 525 * vma shrinks, and !insert tells it's not 526 * split_vma inserting another: so it must be 527 * mprotect case 4 shifting the boundary down. 528 */ 529 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT); 530 exporter = vma; 531 importer = next; 532 } 533 534 /* 535 * Easily overlooked: when mprotect shifts the boundary, 536 * make sure the expanding vma has anon_vma set if the 537 * shrinking vma had, to cover any anon pages imported. 538 */ 539 if (exporter && exporter->anon_vma && !importer->anon_vma) { 540 if (anon_vma_clone(importer, exporter)) 541 return -ENOMEM; 542 importer->anon_vma = exporter->anon_vma; 543 } 544 } 545 546 if (file) { 547 mapping = file->f_mapping; 548 if (!(vma->vm_flags & VM_NONLINEAR)) 549 root = &mapping->i_mmap; 550 mutex_lock(&mapping->i_mmap_mutex); 551 if (insert) { 552 /* 553 * Put into prio_tree now, so instantiated pages 554 * are visible to arm/parisc __flush_dcache_page 555 * throughout; but we cannot insert into address 556 * space until vma start or end is updated. 557 */ 558 __vma_link_file(insert); 559 } 560 } 561 562 vma_adjust_trans_huge(vma, start, end, adjust_next); 563 564 /* 565 * When changing only vma->vm_end, we don't really need anon_vma 566 * lock. This is a fairly rare case by itself, but the anon_vma 567 * lock may be shared between many sibling processes. Skipping 568 * the lock for brk adjustments makes a difference sometimes. 569 */ 570 if (vma->anon_vma && (importer || start != vma->vm_start)) { 571 anon_vma = vma->anon_vma; 572 anon_vma_lock(anon_vma); 573 } 574 575 if (root) { 576 flush_dcache_mmap_lock(mapping); 577 vma_prio_tree_remove(vma, root); 578 if (adjust_next) 579 vma_prio_tree_remove(next, root); 580 } 581 582 vma->vm_start = start; 583 vma->vm_end = end; 584 vma->vm_pgoff = pgoff; 585 if (adjust_next) { 586 next->vm_start += adjust_next << PAGE_SHIFT; 587 next->vm_pgoff += adjust_next; 588 } 589 590 if (root) { 591 if (adjust_next) 592 vma_prio_tree_insert(next, root); 593 vma_prio_tree_insert(vma, root); 594 flush_dcache_mmap_unlock(mapping); 595 } 596 597 if (remove_next) { 598 /* 599 * vma_merge has merged next into vma, and needs 600 * us to remove next before dropping the locks. 601 */ 602 __vma_unlink(mm, next, vma); 603 if (file) 604 __remove_shared_vm_struct(next, file, mapping); 605 } else if (insert) { 606 /* 607 * split_vma has split insert from vma, and needs 608 * us to insert it before dropping the locks 609 * (it may either follow vma or precede it). 610 */ 611 __insert_vm_struct(mm, insert); 612 } 613 614 if (anon_vma) 615 anon_vma_unlock(anon_vma); 616 if (mapping) 617 mutex_unlock(&mapping->i_mmap_mutex); 618 619 if (remove_next) { 620 if (file) { 621 fput(file); 622 if (next->vm_flags & VM_EXECUTABLE) 623 removed_exe_file_vma(mm); 624 } 625 if (next->anon_vma) 626 anon_vma_merge(vma, next); 627 mm->map_count--; 628 mpol_put(vma_policy(next)); 629 kmem_cache_free(vm_area_cachep, next); 630 /* 631 * In mprotect's case 6 (see comments on vma_merge), 632 * we must remove another next too. It would clutter 633 * up the code too much to do both in one go. 634 */ 635 if (remove_next == 2) { 636 next = vma->vm_next; 637 goto again; 638 } 639 } 640 641 validate_mm(mm); 642 643 return 0; 644 } 645 646 /* 647 * If the vma has a ->close operation then the driver probably needs to release 648 * per-vma resources, so we don't attempt to merge those. 649 */ 650 static inline int is_mergeable_vma(struct vm_area_struct *vma, 651 struct file *file, unsigned long vm_flags) 652 { 653 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */ 654 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR) 655 return 0; 656 if (vma->vm_file != file) 657 return 0; 658 if (vma->vm_ops && vma->vm_ops->close) 659 return 0; 660 return 1; 661 } 662 663 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 664 struct anon_vma *anon_vma2, 665 struct vm_area_struct *vma) 666 { 667 /* 668 * The list_is_singular() test is to avoid merging VMA cloned from 669 * parents. This can improve scalability caused by anon_vma lock. 670 */ 671 if ((!anon_vma1 || !anon_vma2) && (!vma || 672 list_is_singular(&vma->anon_vma_chain))) 673 return 1; 674 return anon_vma1 == anon_vma2; 675 } 676 677 /* 678 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 679 * in front of (at a lower virtual address and file offset than) the vma. 680 * 681 * We cannot merge two vmas if they have differently assigned (non-NULL) 682 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 683 * 684 * We don't check here for the merged mmap wrapping around the end of pagecache 685 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 686 * wrap, nor mmaps which cover the final page at index -1UL. 687 */ 688 static int 689 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 690 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 691 { 692 if (is_mergeable_vma(vma, file, vm_flags) && 693 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 694 if (vma->vm_pgoff == vm_pgoff) 695 return 1; 696 } 697 return 0; 698 } 699 700 /* 701 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 702 * beyond (at a higher virtual address and file offset than) the vma. 703 * 704 * We cannot merge two vmas if they have differently assigned (non-NULL) 705 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 706 */ 707 static int 708 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 709 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 710 { 711 if (is_mergeable_vma(vma, file, vm_flags) && 712 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 713 pgoff_t vm_pglen; 714 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; 715 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 716 return 1; 717 } 718 return 0; 719 } 720 721 /* 722 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 723 * whether that can be merged with its predecessor or its successor. 724 * Or both (it neatly fills a hole). 725 * 726 * In most cases - when called for mmap, brk or mremap - [addr,end) is 727 * certain not to be mapped by the time vma_merge is called; but when 728 * called for mprotect, it is certain to be already mapped (either at 729 * an offset within prev, or at the start of next), and the flags of 730 * this area are about to be changed to vm_flags - and the no-change 731 * case has already been eliminated. 732 * 733 * The following mprotect cases have to be considered, where AAAA is 734 * the area passed down from mprotect_fixup, never extending beyond one 735 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 736 * 737 * AAAA AAAA AAAA AAAA 738 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 739 * cannot merge might become might become might become 740 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 741 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 742 * mremap move: PPPPNNNNNNNN 8 743 * AAAA 744 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 745 * might become case 1 below case 2 below case 3 below 746 * 747 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 748 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 749 */ 750 struct vm_area_struct *vma_merge(struct mm_struct *mm, 751 struct vm_area_struct *prev, unsigned long addr, 752 unsigned long end, unsigned long vm_flags, 753 struct anon_vma *anon_vma, struct file *file, 754 pgoff_t pgoff, struct mempolicy *policy) 755 { 756 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 757 struct vm_area_struct *area, *next; 758 int err; 759 760 /* 761 * We later require that vma->vm_flags == vm_flags, 762 * so this tests vma->vm_flags & VM_SPECIAL, too. 763 */ 764 if (vm_flags & VM_SPECIAL) 765 return NULL; 766 767 if (prev) 768 next = prev->vm_next; 769 else 770 next = mm->mmap; 771 area = next; 772 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 773 next = next->vm_next; 774 775 /* 776 * Can it merge with the predecessor? 777 */ 778 if (prev && prev->vm_end == addr && 779 mpol_equal(vma_policy(prev), policy) && 780 can_vma_merge_after(prev, vm_flags, 781 anon_vma, file, pgoff)) { 782 /* 783 * OK, it can. Can we now merge in the successor as well? 784 */ 785 if (next && end == next->vm_start && 786 mpol_equal(policy, vma_policy(next)) && 787 can_vma_merge_before(next, vm_flags, 788 anon_vma, file, pgoff+pglen) && 789 is_mergeable_anon_vma(prev->anon_vma, 790 next->anon_vma, NULL)) { 791 /* cases 1, 6 */ 792 err = vma_adjust(prev, prev->vm_start, 793 next->vm_end, prev->vm_pgoff, NULL); 794 } else /* cases 2, 5, 7 */ 795 err = vma_adjust(prev, prev->vm_start, 796 end, prev->vm_pgoff, NULL); 797 if (err) 798 return NULL; 799 khugepaged_enter_vma_merge(prev); 800 return prev; 801 } 802 803 /* 804 * Can this new request be merged in front of next? 805 */ 806 if (next && end == next->vm_start && 807 mpol_equal(policy, vma_policy(next)) && 808 can_vma_merge_before(next, vm_flags, 809 anon_vma, file, pgoff+pglen)) { 810 if (prev && addr < prev->vm_end) /* case 4 */ 811 err = vma_adjust(prev, prev->vm_start, 812 addr, prev->vm_pgoff, NULL); 813 else /* cases 3, 8 */ 814 err = vma_adjust(area, addr, next->vm_end, 815 next->vm_pgoff - pglen, NULL); 816 if (err) 817 return NULL; 818 khugepaged_enter_vma_merge(area); 819 return area; 820 } 821 822 return NULL; 823 } 824 825 /* 826 * Rough compatbility check to quickly see if it's even worth looking 827 * at sharing an anon_vma. 828 * 829 * They need to have the same vm_file, and the flags can only differ 830 * in things that mprotect may change. 831 * 832 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 833 * we can merge the two vma's. For example, we refuse to merge a vma if 834 * there is a vm_ops->close() function, because that indicates that the 835 * driver is doing some kind of reference counting. But that doesn't 836 * really matter for the anon_vma sharing case. 837 */ 838 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 839 { 840 return a->vm_end == b->vm_start && 841 mpol_equal(vma_policy(a), vma_policy(b)) && 842 a->vm_file == b->vm_file && 843 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) && 844 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 845 } 846 847 /* 848 * Do some basic sanity checking to see if we can re-use the anon_vma 849 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 850 * the same as 'old', the other will be the new one that is trying 851 * to share the anon_vma. 852 * 853 * NOTE! This runs with mm_sem held for reading, so it is possible that 854 * the anon_vma of 'old' is concurrently in the process of being set up 855 * by another page fault trying to merge _that_. But that's ok: if it 856 * is being set up, that automatically means that it will be a singleton 857 * acceptable for merging, so we can do all of this optimistically. But 858 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer. 859 * 860 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 861 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 862 * is to return an anon_vma that is "complex" due to having gone through 863 * a fork). 864 * 865 * We also make sure that the two vma's are compatible (adjacent, 866 * and with the same memory policies). That's all stable, even with just 867 * a read lock on the mm_sem. 868 */ 869 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 870 { 871 if (anon_vma_compatible(a, b)) { 872 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma); 873 874 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 875 return anon_vma; 876 } 877 return NULL; 878 } 879 880 /* 881 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 882 * neighbouring vmas for a suitable anon_vma, before it goes off 883 * to allocate a new anon_vma. It checks because a repetitive 884 * sequence of mprotects and faults may otherwise lead to distinct 885 * anon_vmas being allocated, preventing vma merge in subsequent 886 * mprotect. 887 */ 888 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 889 { 890 struct anon_vma *anon_vma; 891 struct vm_area_struct *near; 892 893 near = vma->vm_next; 894 if (!near) 895 goto try_prev; 896 897 anon_vma = reusable_anon_vma(near, vma, near); 898 if (anon_vma) 899 return anon_vma; 900 try_prev: 901 near = vma->vm_prev; 902 if (!near) 903 goto none; 904 905 anon_vma = reusable_anon_vma(near, near, vma); 906 if (anon_vma) 907 return anon_vma; 908 none: 909 /* 910 * There's no absolute need to look only at touching neighbours: 911 * we could search further afield for "compatible" anon_vmas. 912 * But it would probably just be a waste of time searching, 913 * or lead to too many vmas hanging off the same anon_vma. 914 * We're trying to allow mprotect remerging later on, 915 * not trying to minimize memory used for anon_vmas. 916 */ 917 return NULL; 918 } 919 920 #ifdef CONFIG_PROC_FS 921 void vm_stat_account(struct mm_struct *mm, unsigned long flags, 922 struct file *file, long pages) 923 { 924 const unsigned long stack_flags 925 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN); 926 927 if (file) { 928 mm->shared_vm += pages; 929 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC) 930 mm->exec_vm += pages; 931 } else if (flags & stack_flags) 932 mm->stack_vm += pages; 933 if (flags & (VM_RESERVED|VM_IO)) 934 mm->reserved_vm += pages; 935 } 936 #endif /* CONFIG_PROC_FS */ 937 938 /* 939 * The caller must hold down_write(¤t->mm->mmap_sem). 940 */ 941 942 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr, 943 unsigned long len, unsigned long prot, 944 unsigned long flags, unsigned long pgoff) 945 { 946 struct mm_struct * mm = current->mm; 947 struct inode *inode; 948 vm_flags_t vm_flags; 949 int error; 950 unsigned long reqprot = prot; 951 952 /* 953 * Does the application expect PROT_READ to imply PROT_EXEC? 954 * 955 * (the exception is when the underlying filesystem is noexec 956 * mounted, in which case we dont add PROT_EXEC.) 957 */ 958 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 959 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC))) 960 prot |= PROT_EXEC; 961 962 if (!len) 963 return -EINVAL; 964 965 if (!(flags & MAP_FIXED)) 966 addr = round_hint_to_min(addr); 967 968 /* Careful about overflows.. */ 969 len = PAGE_ALIGN(len); 970 if (!len) 971 return -ENOMEM; 972 973 /* offset overflow? */ 974 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 975 return -EOVERFLOW; 976 977 /* Too many mappings? */ 978 if (mm->map_count > sysctl_max_map_count) 979 return -ENOMEM; 980 981 /* Obtain the address to map to. we verify (or select) it and ensure 982 * that it represents a valid section of the address space. 983 */ 984 addr = get_unmapped_area(file, addr, len, pgoff, flags); 985 if (addr & ~PAGE_MASK) 986 return addr; 987 988 /* Do simple checking here so the lower-level routines won't have 989 * to. we assume access permissions have been handled by the open 990 * of the memory object, so we don't do any here. 991 */ 992 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 993 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 994 995 if (flags & MAP_LOCKED) 996 if (!can_do_mlock()) 997 return -EPERM; 998 999 /* mlock MCL_FUTURE? */ 1000 if (vm_flags & VM_LOCKED) { 1001 unsigned long locked, lock_limit; 1002 locked = len >> PAGE_SHIFT; 1003 locked += mm->locked_vm; 1004 lock_limit = rlimit(RLIMIT_MEMLOCK); 1005 lock_limit >>= PAGE_SHIFT; 1006 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1007 return -EAGAIN; 1008 } 1009 1010 inode = file ? file->f_path.dentry->d_inode : NULL; 1011 1012 if (file) { 1013 switch (flags & MAP_TYPE) { 1014 case MAP_SHARED: 1015 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 1016 return -EACCES; 1017 1018 /* 1019 * Make sure we don't allow writing to an append-only 1020 * file.. 1021 */ 1022 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1023 return -EACCES; 1024 1025 /* 1026 * Make sure there are no mandatory locks on the file. 1027 */ 1028 if (locks_verify_locked(inode)) 1029 return -EAGAIN; 1030 1031 vm_flags |= VM_SHARED | VM_MAYSHARE; 1032 if (!(file->f_mode & FMODE_WRITE)) 1033 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1034 1035 /* fall through */ 1036 case MAP_PRIVATE: 1037 if (!(file->f_mode & FMODE_READ)) 1038 return -EACCES; 1039 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) { 1040 if (vm_flags & VM_EXEC) 1041 return -EPERM; 1042 vm_flags &= ~VM_MAYEXEC; 1043 } 1044 1045 if (!file->f_op || !file->f_op->mmap) 1046 return -ENODEV; 1047 break; 1048 1049 default: 1050 return -EINVAL; 1051 } 1052 } else { 1053 switch (flags & MAP_TYPE) { 1054 case MAP_SHARED: 1055 /* 1056 * Ignore pgoff. 1057 */ 1058 pgoff = 0; 1059 vm_flags |= VM_SHARED | VM_MAYSHARE; 1060 break; 1061 case MAP_PRIVATE: 1062 /* 1063 * Set pgoff according to addr for anon_vma. 1064 */ 1065 pgoff = addr >> PAGE_SHIFT; 1066 break; 1067 default: 1068 return -EINVAL; 1069 } 1070 } 1071 1072 error = security_file_mmap(file, reqprot, prot, flags, addr, 0); 1073 if (error) 1074 return error; 1075 1076 return mmap_region(file, addr, len, flags, vm_flags, pgoff); 1077 } 1078 EXPORT_SYMBOL(do_mmap_pgoff); 1079 1080 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1081 unsigned long, prot, unsigned long, flags, 1082 unsigned long, fd, unsigned long, pgoff) 1083 { 1084 struct file *file = NULL; 1085 unsigned long retval = -EBADF; 1086 1087 if (!(flags & MAP_ANONYMOUS)) { 1088 audit_mmap_fd(fd, flags); 1089 if (unlikely(flags & MAP_HUGETLB)) 1090 return -EINVAL; 1091 file = fget(fd); 1092 if (!file) 1093 goto out; 1094 } else if (flags & MAP_HUGETLB) { 1095 struct user_struct *user = NULL; 1096 /* 1097 * VM_NORESERVE is used because the reservations will be 1098 * taken when vm_ops->mmap() is called 1099 * A dummy user value is used because we are not locking 1100 * memory so no accounting is necessary 1101 */ 1102 len = ALIGN(len, huge_page_size(&default_hstate)); 1103 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE, 1104 &user, HUGETLB_ANONHUGE_INODE); 1105 if (IS_ERR(file)) 1106 return PTR_ERR(file); 1107 } 1108 1109 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1110 1111 down_write(¤t->mm->mmap_sem); 1112 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1113 up_write(¤t->mm->mmap_sem); 1114 1115 if (file) 1116 fput(file); 1117 out: 1118 return retval; 1119 } 1120 1121 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1122 struct mmap_arg_struct { 1123 unsigned long addr; 1124 unsigned long len; 1125 unsigned long prot; 1126 unsigned long flags; 1127 unsigned long fd; 1128 unsigned long offset; 1129 }; 1130 1131 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1132 { 1133 struct mmap_arg_struct a; 1134 1135 if (copy_from_user(&a, arg, sizeof(a))) 1136 return -EFAULT; 1137 if (a.offset & ~PAGE_MASK) 1138 return -EINVAL; 1139 1140 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1141 a.offset >> PAGE_SHIFT); 1142 } 1143 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1144 1145 /* 1146 * Some shared mappigns will want the pages marked read-only 1147 * to track write events. If so, we'll downgrade vm_page_prot 1148 * to the private version (using protection_map[] without the 1149 * VM_SHARED bit). 1150 */ 1151 int vma_wants_writenotify(struct vm_area_struct *vma) 1152 { 1153 vm_flags_t vm_flags = vma->vm_flags; 1154 1155 /* If it was private or non-writable, the write bit is already clear */ 1156 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1157 return 0; 1158 1159 /* The backer wishes to know when pages are first written to? */ 1160 if (vma->vm_ops && vma->vm_ops->page_mkwrite) 1161 return 1; 1162 1163 /* The open routine did something to the protections already? */ 1164 if (pgprot_val(vma->vm_page_prot) != 1165 pgprot_val(vm_get_page_prot(vm_flags))) 1166 return 0; 1167 1168 /* Specialty mapping? */ 1169 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE)) 1170 return 0; 1171 1172 /* Can the mapping track the dirty pages? */ 1173 return vma->vm_file && vma->vm_file->f_mapping && 1174 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1175 } 1176 1177 /* 1178 * We account for memory if it's a private writeable mapping, 1179 * not hugepages and VM_NORESERVE wasn't set. 1180 */ 1181 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1182 { 1183 /* 1184 * hugetlb has its own accounting separate from the core VM 1185 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1186 */ 1187 if (file && is_file_hugepages(file)) 1188 return 0; 1189 1190 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1191 } 1192 1193 unsigned long mmap_region(struct file *file, unsigned long addr, 1194 unsigned long len, unsigned long flags, 1195 vm_flags_t vm_flags, unsigned long pgoff) 1196 { 1197 struct mm_struct *mm = current->mm; 1198 struct vm_area_struct *vma, *prev; 1199 int correct_wcount = 0; 1200 int error; 1201 struct rb_node **rb_link, *rb_parent; 1202 unsigned long charged = 0; 1203 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL; 1204 1205 /* Clear old maps */ 1206 error = -ENOMEM; 1207 munmap_back: 1208 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 1209 if (vma && vma->vm_start < addr + len) { 1210 if (do_munmap(mm, addr, len)) 1211 return -ENOMEM; 1212 goto munmap_back; 1213 } 1214 1215 /* Check against address space limit. */ 1216 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 1217 return -ENOMEM; 1218 1219 /* 1220 * Set 'VM_NORESERVE' if we should not account for the 1221 * memory use of this mapping. 1222 */ 1223 if ((flags & MAP_NORESERVE)) { 1224 /* We honor MAP_NORESERVE if allowed to overcommit */ 1225 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1226 vm_flags |= VM_NORESERVE; 1227 1228 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1229 if (file && is_file_hugepages(file)) 1230 vm_flags |= VM_NORESERVE; 1231 } 1232 1233 /* 1234 * Private writable mapping: check memory availability 1235 */ 1236 if (accountable_mapping(file, vm_flags)) { 1237 charged = len >> PAGE_SHIFT; 1238 if (security_vm_enough_memory(charged)) 1239 return -ENOMEM; 1240 vm_flags |= VM_ACCOUNT; 1241 } 1242 1243 /* 1244 * Can we just expand an old mapping? 1245 */ 1246 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL); 1247 if (vma) 1248 goto out; 1249 1250 /* 1251 * Determine the object being mapped and call the appropriate 1252 * specific mapper. the address has already been validated, but 1253 * not unmapped, but the maps are removed from the list. 1254 */ 1255 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1256 if (!vma) { 1257 error = -ENOMEM; 1258 goto unacct_error; 1259 } 1260 1261 vma->vm_mm = mm; 1262 vma->vm_start = addr; 1263 vma->vm_end = addr + len; 1264 vma->vm_flags = vm_flags; 1265 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1266 vma->vm_pgoff = pgoff; 1267 INIT_LIST_HEAD(&vma->anon_vma_chain); 1268 1269 if (file) { 1270 error = -EINVAL; 1271 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1272 goto free_vma; 1273 if (vm_flags & VM_DENYWRITE) { 1274 error = deny_write_access(file); 1275 if (error) 1276 goto free_vma; 1277 correct_wcount = 1; 1278 } 1279 vma->vm_file = file; 1280 get_file(file); 1281 error = file->f_op->mmap(file, vma); 1282 if (error) 1283 goto unmap_and_free_vma; 1284 if (vm_flags & VM_EXECUTABLE) 1285 added_exe_file_vma(mm); 1286 1287 /* Can addr have changed?? 1288 * 1289 * Answer: Yes, several device drivers can do it in their 1290 * f_op->mmap method. -DaveM 1291 */ 1292 addr = vma->vm_start; 1293 pgoff = vma->vm_pgoff; 1294 vm_flags = vma->vm_flags; 1295 } else if (vm_flags & VM_SHARED) { 1296 error = shmem_zero_setup(vma); 1297 if (error) 1298 goto free_vma; 1299 } 1300 1301 if (vma_wants_writenotify(vma)) { 1302 pgprot_t pprot = vma->vm_page_prot; 1303 1304 /* Can vma->vm_page_prot have changed?? 1305 * 1306 * Answer: Yes, drivers may have changed it in their 1307 * f_op->mmap method. 1308 * 1309 * Ensures that vmas marked as uncached stay that way. 1310 */ 1311 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED); 1312 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot))) 1313 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 1314 } 1315 1316 vma_link(mm, vma, prev, rb_link, rb_parent); 1317 file = vma->vm_file; 1318 1319 /* Once vma denies write, undo our temporary denial count */ 1320 if (correct_wcount) 1321 atomic_inc(&inode->i_writecount); 1322 out: 1323 perf_event_mmap(vma); 1324 1325 mm->total_vm += len >> PAGE_SHIFT; 1326 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT); 1327 if (vm_flags & VM_LOCKED) { 1328 if (!mlock_vma_pages_range(vma, addr, addr + len)) 1329 mm->locked_vm += (len >> PAGE_SHIFT); 1330 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK)) 1331 make_pages_present(addr, addr + len); 1332 return addr; 1333 1334 unmap_and_free_vma: 1335 if (correct_wcount) 1336 atomic_inc(&inode->i_writecount); 1337 vma->vm_file = NULL; 1338 fput(file); 1339 1340 /* Undo any partial mapping done by a device driver. */ 1341 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1342 charged = 0; 1343 free_vma: 1344 kmem_cache_free(vm_area_cachep, vma); 1345 unacct_error: 1346 if (charged) 1347 vm_unacct_memory(charged); 1348 return error; 1349 } 1350 1351 /* Get an address range which is currently unmapped. 1352 * For shmat() with addr=0. 1353 * 1354 * Ugly calling convention alert: 1355 * Return value with the low bits set means error value, 1356 * ie 1357 * if (ret & ~PAGE_MASK) 1358 * error = ret; 1359 * 1360 * This function "knows" that -ENOMEM has the bits set. 1361 */ 1362 #ifndef HAVE_ARCH_UNMAPPED_AREA 1363 unsigned long 1364 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1365 unsigned long len, unsigned long pgoff, unsigned long flags) 1366 { 1367 struct mm_struct *mm = current->mm; 1368 struct vm_area_struct *vma; 1369 unsigned long start_addr; 1370 1371 if (len > TASK_SIZE) 1372 return -ENOMEM; 1373 1374 if (flags & MAP_FIXED) 1375 return addr; 1376 1377 if (addr) { 1378 addr = PAGE_ALIGN(addr); 1379 vma = find_vma(mm, addr); 1380 if (TASK_SIZE - len >= addr && 1381 (!vma || addr + len <= vma->vm_start)) 1382 return addr; 1383 } 1384 if (len > mm->cached_hole_size) { 1385 start_addr = addr = mm->free_area_cache; 1386 } else { 1387 start_addr = addr = TASK_UNMAPPED_BASE; 1388 mm->cached_hole_size = 0; 1389 } 1390 1391 full_search: 1392 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1393 /* At this point: (!vma || addr < vma->vm_end). */ 1394 if (TASK_SIZE - len < addr) { 1395 /* 1396 * Start a new search - just in case we missed 1397 * some holes. 1398 */ 1399 if (start_addr != TASK_UNMAPPED_BASE) { 1400 addr = TASK_UNMAPPED_BASE; 1401 start_addr = addr; 1402 mm->cached_hole_size = 0; 1403 goto full_search; 1404 } 1405 return -ENOMEM; 1406 } 1407 if (!vma || addr + len <= vma->vm_start) { 1408 /* 1409 * Remember the place where we stopped the search: 1410 */ 1411 mm->free_area_cache = addr + len; 1412 return addr; 1413 } 1414 if (addr + mm->cached_hole_size < vma->vm_start) 1415 mm->cached_hole_size = vma->vm_start - addr; 1416 addr = vma->vm_end; 1417 } 1418 } 1419 #endif 1420 1421 void arch_unmap_area(struct mm_struct *mm, unsigned long addr) 1422 { 1423 /* 1424 * Is this a new hole at the lowest possible address? 1425 */ 1426 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) { 1427 mm->free_area_cache = addr; 1428 mm->cached_hole_size = ~0UL; 1429 } 1430 } 1431 1432 /* 1433 * This mmap-allocator allocates new areas top-down from below the 1434 * stack's low limit (the base): 1435 */ 1436 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1437 unsigned long 1438 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1439 const unsigned long len, const unsigned long pgoff, 1440 const unsigned long flags) 1441 { 1442 struct vm_area_struct *vma; 1443 struct mm_struct *mm = current->mm; 1444 unsigned long addr = addr0; 1445 1446 /* requested length too big for entire address space */ 1447 if (len > TASK_SIZE) 1448 return -ENOMEM; 1449 1450 if (flags & MAP_FIXED) 1451 return addr; 1452 1453 /* requesting a specific address */ 1454 if (addr) { 1455 addr = PAGE_ALIGN(addr); 1456 vma = find_vma(mm, addr); 1457 if (TASK_SIZE - len >= addr && 1458 (!vma || addr + len <= vma->vm_start)) 1459 return addr; 1460 } 1461 1462 /* check if free_area_cache is useful for us */ 1463 if (len <= mm->cached_hole_size) { 1464 mm->cached_hole_size = 0; 1465 mm->free_area_cache = mm->mmap_base; 1466 } 1467 1468 /* either no address requested or can't fit in requested address hole */ 1469 addr = mm->free_area_cache; 1470 1471 /* make sure it can fit in the remaining address space */ 1472 if (addr > len) { 1473 vma = find_vma(mm, addr-len); 1474 if (!vma || addr <= vma->vm_start) 1475 /* remember the address as a hint for next time */ 1476 return (mm->free_area_cache = addr-len); 1477 } 1478 1479 if (mm->mmap_base < len) 1480 goto bottomup; 1481 1482 addr = mm->mmap_base-len; 1483 1484 do { 1485 /* 1486 * Lookup failure means no vma is above this address, 1487 * else if new region fits below vma->vm_start, 1488 * return with success: 1489 */ 1490 vma = find_vma(mm, addr); 1491 if (!vma || addr+len <= vma->vm_start) 1492 /* remember the address as a hint for next time */ 1493 return (mm->free_area_cache = addr); 1494 1495 /* remember the largest hole we saw so far */ 1496 if (addr + mm->cached_hole_size < vma->vm_start) 1497 mm->cached_hole_size = vma->vm_start - addr; 1498 1499 /* try just below the current vma->vm_start */ 1500 addr = vma->vm_start-len; 1501 } while (len < vma->vm_start); 1502 1503 bottomup: 1504 /* 1505 * A failed mmap() very likely causes application failure, 1506 * so fall back to the bottom-up function here. This scenario 1507 * can happen with large stack limits and large mmap() 1508 * allocations. 1509 */ 1510 mm->cached_hole_size = ~0UL; 1511 mm->free_area_cache = TASK_UNMAPPED_BASE; 1512 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags); 1513 /* 1514 * Restore the topdown base: 1515 */ 1516 mm->free_area_cache = mm->mmap_base; 1517 mm->cached_hole_size = ~0UL; 1518 1519 return addr; 1520 } 1521 #endif 1522 1523 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr) 1524 { 1525 /* 1526 * Is this a new hole at the highest possible address? 1527 */ 1528 if (addr > mm->free_area_cache) 1529 mm->free_area_cache = addr; 1530 1531 /* dont allow allocations above current base */ 1532 if (mm->free_area_cache > mm->mmap_base) 1533 mm->free_area_cache = mm->mmap_base; 1534 } 1535 1536 unsigned long 1537 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1538 unsigned long pgoff, unsigned long flags) 1539 { 1540 unsigned long (*get_area)(struct file *, unsigned long, 1541 unsigned long, unsigned long, unsigned long); 1542 1543 unsigned long error = arch_mmap_check(addr, len, flags); 1544 if (error) 1545 return error; 1546 1547 /* Careful about overflows.. */ 1548 if (len > TASK_SIZE) 1549 return -ENOMEM; 1550 1551 get_area = current->mm->get_unmapped_area; 1552 if (file && file->f_op && file->f_op->get_unmapped_area) 1553 get_area = file->f_op->get_unmapped_area; 1554 addr = get_area(file, addr, len, pgoff, flags); 1555 if (IS_ERR_VALUE(addr)) 1556 return addr; 1557 1558 if (addr > TASK_SIZE - len) 1559 return -ENOMEM; 1560 if (addr & ~PAGE_MASK) 1561 return -EINVAL; 1562 1563 return arch_rebalance_pgtables(addr, len); 1564 } 1565 1566 EXPORT_SYMBOL(get_unmapped_area); 1567 1568 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 1569 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 1570 { 1571 struct vm_area_struct *vma = NULL; 1572 1573 if (mm) { 1574 /* Check the cache first. */ 1575 /* (Cache hit rate is typically around 35%.) */ 1576 vma = mm->mmap_cache; 1577 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) { 1578 struct rb_node * rb_node; 1579 1580 rb_node = mm->mm_rb.rb_node; 1581 vma = NULL; 1582 1583 while (rb_node) { 1584 struct vm_area_struct * vma_tmp; 1585 1586 vma_tmp = rb_entry(rb_node, 1587 struct vm_area_struct, vm_rb); 1588 1589 if (vma_tmp->vm_end > addr) { 1590 vma = vma_tmp; 1591 if (vma_tmp->vm_start <= addr) 1592 break; 1593 rb_node = rb_node->rb_left; 1594 } else 1595 rb_node = rb_node->rb_right; 1596 } 1597 if (vma) 1598 mm->mmap_cache = vma; 1599 } 1600 } 1601 return vma; 1602 } 1603 1604 EXPORT_SYMBOL(find_vma); 1605 1606 /* 1607 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 1608 * Note: pprev is set to NULL when return value is NULL. 1609 */ 1610 struct vm_area_struct * 1611 find_vma_prev(struct mm_struct *mm, unsigned long addr, 1612 struct vm_area_struct **pprev) 1613 { 1614 struct vm_area_struct *vma; 1615 1616 vma = find_vma(mm, addr); 1617 *pprev = vma ? vma->vm_prev : NULL; 1618 return vma; 1619 } 1620 1621 /* 1622 * Verify that the stack growth is acceptable and 1623 * update accounting. This is shared with both the 1624 * grow-up and grow-down cases. 1625 */ 1626 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 1627 { 1628 struct mm_struct *mm = vma->vm_mm; 1629 struct rlimit *rlim = current->signal->rlim; 1630 unsigned long new_start; 1631 1632 /* address space limit tests */ 1633 if (!may_expand_vm(mm, grow)) 1634 return -ENOMEM; 1635 1636 /* Stack limit test */ 1637 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 1638 return -ENOMEM; 1639 1640 /* mlock limit tests */ 1641 if (vma->vm_flags & VM_LOCKED) { 1642 unsigned long locked; 1643 unsigned long limit; 1644 locked = mm->locked_vm + grow; 1645 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 1646 limit >>= PAGE_SHIFT; 1647 if (locked > limit && !capable(CAP_IPC_LOCK)) 1648 return -ENOMEM; 1649 } 1650 1651 /* Check to ensure the stack will not grow into a hugetlb-only region */ 1652 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 1653 vma->vm_end - size; 1654 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 1655 return -EFAULT; 1656 1657 /* 1658 * Overcommit.. This must be the final test, as it will 1659 * update security statistics. 1660 */ 1661 if (security_vm_enough_memory_mm(mm, grow)) 1662 return -ENOMEM; 1663 1664 /* Ok, everything looks good - let it rip */ 1665 mm->total_vm += grow; 1666 if (vma->vm_flags & VM_LOCKED) 1667 mm->locked_vm += grow; 1668 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow); 1669 return 0; 1670 } 1671 1672 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 1673 /* 1674 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 1675 * vma is the last one with address > vma->vm_end. Have to extend vma. 1676 */ 1677 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 1678 { 1679 int error; 1680 1681 if (!(vma->vm_flags & VM_GROWSUP)) 1682 return -EFAULT; 1683 1684 /* 1685 * We must make sure the anon_vma is allocated 1686 * so that the anon_vma locking is not a noop. 1687 */ 1688 if (unlikely(anon_vma_prepare(vma))) 1689 return -ENOMEM; 1690 vma_lock_anon_vma(vma); 1691 1692 /* 1693 * vma->vm_start/vm_end cannot change under us because the caller 1694 * is required to hold the mmap_sem in read mode. We need the 1695 * anon_vma lock to serialize against concurrent expand_stacks. 1696 * Also guard against wrapping around to address 0. 1697 */ 1698 if (address < PAGE_ALIGN(address+4)) 1699 address = PAGE_ALIGN(address+4); 1700 else { 1701 vma_unlock_anon_vma(vma); 1702 return -ENOMEM; 1703 } 1704 error = 0; 1705 1706 /* Somebody else might have raced and expanded it already */ 1707 if (address > vma->vm_end) { 1708 unsigned long size, grow; 1709 1710 size = address - vma->vm_start; 1711 grow = (address - vma->vm_end) >> PAGE_SHIFT; 1712 1713 error = -ENOMEM; 1714 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 1715 error = acct_stack_growth(vma, size, grow); 1716 if (!error) { 1717 vma->vm_end = address; 1718 perf_event_mmap(vma); 1719 } 1720 } 1721 } 1722 vma_unlock_anon_vma(vma); 1723 khugepaged_enter_vma_merge(vma); 1724 return error; 1725 } 1726 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 1727 1728 /* 1729 * vma is the first one with address < vma->vm_start. Have to extend vma. 1730 */ 1731 int expand_downwards(struct vm_area_struct *vma, 1732 unsigned long address) 1733 { 1734 int error; 1735 1736 /* 1737 * We must make sure the anon_vma is allocated 1738 * so that the anon_vma locking is not a noop. 1739 */ 1740 if (unlikely(anon_vma_prepare(vma))) 1741 return -ENOMEM; 1742 1743 address &= PAGE_MASK; 1744 error = security_file_mmap(NULL, 0, 0, 0, address, 1); 1745 if (error) 1746 return error; 1747 1748 vma_lock_anon_vma(vma); 1749 1750 /* 1751 * vma->vm_start/vm_end cannot change under us because the caller 1752 * is required to hold the mmap_sem in read mode. We need the 1753 * anon_vma lock to serialize against concurrent expand_stacks. 1754 */ 1755 1756 /* Somebody else might have raced and expanded it already */ 1757 if (address < vma->vm_start) { 1758 unsigned long size, grow; 1759 1760 size = vma->vm_end - address; 1761 grow = (vma->vm_start - address) >> PAGE_SHIFT; 1762 1763 error = -ENOMEM; 1764 if (grow <= vma->vm_pgoff) { 1765 error = acct_stack_growth(vma, size, grow); 1766 if (!error) { 1767 vma->vm_start = address; 1768 vma->vm_pgoff -= grow; 1769 perf_event_mmap(vma); 1770 } 1771 } 1772 } 1773 vma_unlock_anon_vma(vma); 1774 khugepaged_enter_vma_merge(vma); 1775 return error; 1776 } 1777 1778 #ifdef CONFIG_STACK_GROWSUP 1779 int expand_stack(struct vm_area_struct *vma, unsigned long address) 1780 { 1781 return expand_upwards(vma, address); 1782 } 1783 1784 struct vm_area_struct * 1785 find_extend_vma(struct mm_struct *mm, unsigned long addr) 1786 { 1787 struct vm_area_struct *vma, *prev; 1788 1789 addr &= PAGE_MASK; 1790 vma = find_vma_prev(mm, addr, &prev); 1791 if (vma && (vma->vm_start <= addr)) 1792 return vma; 1793 if (!prev || expand_stack(prev, addr)) 1794 return NULL; 1795 if (prev->vm_flags & VM_LOCKED) { 1796 mlock_vma_pages_range(prev, addr, prev->vm_end); 1797 } 1798 return prev; 1799 } 1800 #else 1801 int expand_stack(struct vm_area_struct *vma, unsigned long address) 1802 { 1803 return expand_downwards(vma, address); 1804 } 1805 1806 struct vm_area_struct * 1807 find_extend_vma(struct mm_struct * mm, unsigned long addr) 1808 { 1809 struct vm_area_struct * vma; 1810 unsigned long start; 1811 1812 addr &= PAGE_MASK; 1813 vma = find_vma(mm,addr); 1814 if (!vma) 1815 return NULL; 1816 if (vma->vm_start <= addr) 1817 return vma; 1818 if (!(vma->vm_flags & VM_GROWSDOWN)) 1819 return NULL; 1820 start = vma->vm_start; 1821 if (expand_stack(vma, addr)) 1822 return NULL; 1823 if (vma->vm_flags & VM_LOCKED) { 1824 mlock_vma_pages_range(vma, addr, start); 1825 } 1826 return vma; 1827 } 1828 #endif 1829 1830 /* 1831 * Ok - we have the memory areas we should free on the vma list, 1832 * so release them, and do the vma updates. 1833 * 1834 * Called with the mm semaphore held. 1835 */ 1836 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 1837 { 1838 /* Update high watermark before we lower total_vm */ 1839 update_hiwater_vm(mm); 1840 do { 1841 long nrpages = vma_pages(vma); 1842 1843 mm->total_vm -= nrpages; 1844 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages); 1845 vma = remove_vma(vma); 1846 } while (vma); 1847 validate_mm(mm); 1848 } 1849 1850 /* 1851 * Get rid of page table information in the indicated region. 1852 * 1853 * Called with the mm semaphore held. 1854 */ 1855 static void unmap_region(struct mm_struct *mm, 1856 struct vm_area_struct *vma, struct vm_area_struct *prev, 1857 unsigned long start, unsigned long end) 1858 { 1859 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap; 1860 struct mmu_gather tlb; 1861 unsigned long nr_accounted = 0; 1862 1863 lru_add_drain(); 1864 tlb_gather_mmu(&tlb, mm, 0); 1865 update_hiwater_rss(mm); 1866 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL); 1867 vm_unacct_memory(nr_accounted); 1868 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 1869 next ? next->vm_start : 0); 1870 tlb_finish_mmu(&tlb, start, end); 1871 } 1872 1873 /* 1874 * Create a list of vma's touched by the unmap, removing them from the mm's 1875 * vma list as we go.. 1876 */ 1877 static void 1878 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 1879 struct vm_area_struct *prev, unsigned long end) 1880 { 1881 struct vm_area_struct **insertion_point; 1882 struct vm_area_struct *tail_vma = NULL; 1883 unsigned long addr; 1884 1885 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 1886 vma->vm_prev = NULL; 1887 do { 1888 rb_erase(&vma->vm_rb, &mm->mm_rb); 1889 mm->map_count--; 1890 tail_vma = vma; 1891 vma = vma->vm_next; 1892 } while (vma && vma->vm_start < end); 1893 *insertion_point = vma; 1894 if (vma) 1895 vma->vm_prev = prev; 1896 tail_vma->vm_next = NULL; 1897 if (mm->unmap_area == arch_unmap_area) 1898 addr = prev ? prev->vm_end : mm->mmap_base; 1899 else 1900 addr = vma ? vma->vm_start : mm->mmap_base; 1901 mm->unmap_area(mm, addr); 1902 mm->mmap_cache = NULL; /* Kill the cache. */ 1903 } 1904 1905 /* 1906 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the 1907 * munmap path where it doesn't make sense to fail. 1908 */ 1909 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma, 1910 unsigned long addr, int new_below) 1911 { 1912 struct mempolicy *pol; 1913 struct vm_area_struct *new; 1914 int err = -ENOMEM; 1915 1916 if (is_vm_hugetlb_page(vma) && (addr & 1917 ~(huge_page_mask(hstate_vma(vma))))) 1918 return -EINVAL; 1919 1920 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 1921 if (!new) 1922 goto out_err; 1923 1924 /* most fields are the same, copy all, and then fixup */ 1925 *new = *vma; 1926 1927 INIT_LIST_HEAD(&new->anon_vma_chain); 1928 1929 if (new_below) 1930 new->vm_end = addr; 1931 else { 1932 new->vm_start = addr; 1933 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 1934 } 1935 1936 pol = mpol_dup(vma_policy(vma)); 1937 if (IS_ERR(pol)) { 1938 err = PTR_ERR(pol); 1939 goto out_free_vma; 1940 } 1941 vma_set_policy(new, pol); 1942 1943 if (anon_vma_clone(new, vma)) 1944 goto out_free_mpol; 1945 1946 if (new->vm_file) { 1947 get_file(new->vm_file); 1948 if (vma->vm_flags & VM_EXECUTABLE) 1949 added_exe_file_vma(mm); 1950 } 1951 1952 if (new->vm_ops && new->vm_ops->open) 1953 new->vm_ops->open(new); 1954 1955 if (new_below) 1956 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 1957 ((addr - new->vm_start) >> PAGE_SHIFT), new); 1958 else 1959 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 1960 1961 /* Success. */ 1962 if (!err) 1963 return 0; 1964 1965 /* Clean everything up if vma_adjust failed. */ 1966 if (new->vm_ops && new->vm_ops->close) 1967 new->vm_ops->close(new); 1968 if (new->vm_file) { 1969 if (vma->vm_flags & VM_EXECUTABLE) 1970 removed_exe_file_vma(mm); 1971 fput(new->vm_file); 1972 } 1973 unlink_anon_vmas(new); 1974 out_free_mpol: 1975 mpol_put(pol); 1976 out_free_vma: 1977 kmem_cache_free(vm_area_cachep, new); 1978 out_err: 1979 return err; 1980 } 1981 1982 /* 1983 * Split a vma into two pieces at address 'addr', a new vma is allocated 1984 * either for the first part or the tail. 1985 */ 1986 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 1987 unsigned long addr, int new_below) 1988 { 1989 if (mm->map_count >= sysctl_max_map_count) 1990 return -ENOMEM; 1991 1992 return __split_vma(mm, vma, addr, new_below); 1993 } 1994 1995 /* Munmap is split into 2 main parts -- this part which finds 1996 * what needs doing, and the areas themselves, which do the 1997 * work. This now handles partial unmappings. 1998 * Jeremy Fitzhardinge <jeremy@goop.org> 1999 */ 2000 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 2001 { 2002 unsigned long end; 2003 struct vm_area_struct *vma, *prev, *last; 2004 2005 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start) 2006 return -EINVAL; 2007 2008 if ((len = PAGE_ALIGN(len)) == 0) 2009 return -EINVAL; 2010 2011 /* Find the first overlapping VMA */ 2012 vma = find_vma(mm, start); 2013 if (!vma) 2014 return 0; 2015 prev = vma->vm_prev; 2016 /* we have start < vma->vm_end */ 2017 2018 /* if it doesn't overlap, we have nothing.. */ 2019 end = start + len; 2020 if (vma->vm_start >= end) 2021 return 0; 2022 2023 /* 2024 * If we need to split any vma, do it now to save pain later. 2025 * 2026 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2027 * unmapped vm_area_struct will remain in use: so lower split_vma 2028 * places tmp vma above, and higher split_vma places tmp vma below. 2029 */ 2030 if (start > vma->vm_start) { 2031 int error; 2032 2033 /* 2034 * Make sure that map_count on return from munmap() will 2035 * not exceed its limit; but let map_count go just above 2036 * its limit temporarily, to help free resources as expected. 2037 */ 2038 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2039 return -ENOMEM; 2040 2041 error = __split_vma(mm, vma, start, 0); 2042 if (error) 2043 return error; 2044 prev = vma; 2045 } 2046 2047 /* Does it split the last one? */ 2048 last = find_vma(mm, end); 2049 if (last && end > last->vm_start) { 2050 int error = __split_vma(mm, last, end, 1); 2051 if (error) 2052 return error; 2053 } 2054 vma = prev? prev->vm_next: mm->mmap; 2055 2056 /* 2057 * unlock any mlock()ed ranges before detaching vmas 2058 */ 2059 if (mm->locked_vm) { 2060 struct vm_area_struct *tmp = vma; 2061 while (tmp && tmp->vm_start < end) { 2062 if (tmp->vm_flags & VM_LOCKED) { 2063 mm->locked_vm -= vma_pages(tmp); 2064 munlock_vma_pages_all(tmp); 2065 } 2066 tmp = tmp->vm_next; 2067 } 2068 } 2069 2070 /* 2071 * Remove the vma's, and unmap the actual pages 2072 */ 2073 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2074 unmap_region(mm, vma, prev, start, end); 2075 2076 /* Fix up all other VM information */ 2077 remove_vma_list(mm, vma); 2078 2079 return 0; 2080 } 2081 2082 EXPORT_SYMBOL(do_munmap); 2083 2084 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2085 { 2086 int ret; 2087 struct mm_struct *mm = current->mm; 2088 2089 profile_munmap(addr); 2090 2091 down_write(&mm->mmap_sem); 2092 ret = do_munmap(mm, addr, len); 2093 up_write(&mm->mmap_sem); 2094 return ret; 2095 } 2096 2097 static inline void verify_mm_writelocked(struct mm_struct *mm) 2098 { 2099 #ifdef CONFIG_DEBUG_VM 2100 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2101 WARN_ON(1); 2102 up_read(&mm->mmap_sem); 2103 } 2104 #endif 2105 } 2106 2107 /* 2108 * this is really a simplified "do_mmap". it only handles 2109 * anonymous maps. eventually we may be able to do some 2110 * brk-specific accounting here. 2111 */ 2112 unsigned long do_brk(unsigned long addr, unsigned long len) 2113 { 2114 struct mm_struct * mm = current->mm; 2115 struct vm_area_struct * vma, * prev; 2116 unsigned long flags; 2117 struct rb_node ** rb_link, * rb_parent; 2118 pgoff_t pgoff = addr >> PAGE_SHIFT; 2119 int error; 2120 2121 len = PAGE_ALIGN(len); 2122 if (!len) 2123 return addr; 2124 2125 error = security_file_mmap(NULL, 0, 0, 0, addr, 1); 2126 if (error) 2127 return error; 2128 2129 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2130 2131 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2132 if (error & ~PAGE_MASK) 2133 return error; 2134 2135 /* 2136 * mlock MCL_FUTURE? 2137 */ 2138 if (mm->def_flags & VM_LOCKED) { 2139 unsigned long locked, lock_limit; 2140 locked = len >> PAGE_SHIFT; 2141 locked += mm->locked_vm; 2142 lock_limit = rlimit(RLIMIT_MEMLOCK); 2143 lock_limit >>= PAGE_SHIFT; 2144 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 2145 return -EAGAIN; 2146 } 2147 2148 /* 2149 * mm->mmap_sem is required to protect against another thread 2150 * changing the mappings in case we sleep. 2151 */ 2152 verify_mm_writelocked(mm); 2153 2154 /* 2155 * Clear old maps. this also does some error checking for us 2156 */ 2157 munmap_back: 2158 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 2159 if (vma && vma->vm_start < addr + len) { 2160 if (do_munmap(mm, addr, len)) 2161 return -ENOMEM; 2162 goto munmap_back; 2163 } 2164 2165 /* Check against address space limits *after* clearing old maps... */ 2166 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 2167 return -ENOMEM; 2168 2169 if (mm->map_count > sysctl_max_map_count) 2170 return -ENOMEM; 2171 2172 if (security_vm_enough_memory(len >> PAGE_SHIFT)) 2173 return -ENOMEM; 2174 2175 /* Can we just expand an old private anonymous mapping? */ 2176 vma = vma_merge(mm, prev, addr, addr + len, flags, 2177 NULL, NULL, pgoff, NULL); 2178 if (vma) 2179 goto out; 2180 2181 /* 2182 * create a vma struct for an anonymous mapping 2183 */ 2184 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2185 if (!vma) { 2186 vm_unacct_memory(len >> PAGE_SHIFT); 2187 return -ENOMEM; 2188 } 2189 2190 INIT_LIST_HEAD(&vma->anon_vma_chain); 2191 vma->vm_mm = mm; 2192 vma->vm_start = addr; 2193 vma->vm_end = addr + len; 2194 vma->vm_pgoff = pgoff; 2195 vma->vm_flags = flags; 2196 vma->vm_page_prot = vm_get_page_prot(flags); 2197 vma_link(mm, vma, prev, rb_link, rb_parent); 2198 out: 2199 perf_event_mmap(vma); 2200 mm->total_vm += len >> PAGE_SHIFT; 2201 if (flags & VM_LOCKED) { 2202 if (!mlock_vma_pages_range(vma, addr, addr + len)) 2203 mm->locked_vm += (len >> PAGE_SHIFT); 2204 } 2205 return addr; 2206 } 2207 2208 EXPORT_SYMBOL(do_brk); 2209 2210 /* Release all mmaps. */ 2211 void exit_mmap(struct mm_struct *mm) 2212 { 2213 struct mmu_gather tlb; 2214 struct vm_area_struct *vma; 2215 unsigned long nr_accounted = 0; 2216 unsigned long end; 2217 2218 /* mm's last user has gone, and its about to be pulled down */ 2219 mmu_notifier_release(mm); 2220 2221 if (mm->locked_vm) { 2222 vma = mm->mmap; 2223 while (vma) { 2224 if (vma->vm_flags & VM_LOCKED) 2225 munlock_vma_pages_all(vma); 2226 vma = vma->vm_next; 2227 } 2228 } 2229 2230 arch_exit_mmap(mm); 2231 2232 vma = mm->mmap; 2233 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2234 return; 2235 2236 lru_add_drain(); 2237 flush_cache_mm(mm); 2238 tlb_gather_mmu(&tlb, mm, 1); 2239 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2240 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2241 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL); 2242 vm_unacct_memory(nr_accounted); 2243 2244 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0); 2245 tlb_finish_mmu(&tlb, 0, end); 2246 2247 /* 2248 * Walk the list again, actually closing and freeing it, 2249 * with preemption enabled, without holding any MM locks. 2250 */ 2251 while (vma) 2252 vma = remove_vma(vma); 2253 2254 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT); 2255 } 2256 2257 /* Insert vm structure into process list sorted by address 2258 * and into the inode's i_mmap tree. If vm_file is non-NULL 2259 * then i_mmap_mutex is taken here. 2260 */ 2261 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma) 2262 { 2263 struct vm_area_struct * __vma, * prev; 2264 struct rb_node ** rb_link, * rb_parent; 2265 2266 /* 2267 * The vm_pgoff of a purely anonymous vma should be irrelevant 2268 * until its first write fault, when page's anon_vma and index 2269 * are set. But now set the vm_pgoff it will almost certainly 2270 * end up with (unless mremap moves it elsewhere before that 2271 * first wfault), so /proc/pid/maps tells a consistent story. 2272 * 2273 * By setting it to reflect the virtual start address of the 2274 * vma, merges and splits can happen in a seamless way, just 2275 * using the existing file pgoff checks and manipulations. 2276 * Similarly in do_mmap_pgoff and in do_brk. 2277 */ 2278 if (!vma->vm_file) { 2279 BUG_ON(vma->anon_vma); 2280 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2281 } 2282 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent); 2283 if (__vma && __vma->vm_start < vma->vm_end) 2284 return -ENOMEM; 2285 if ((vma->vm_flags & VM_ACCOUNT) && 2286 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2287 return -ENOMEM; 2288 vma_link(mm, vma, prev, rb_link, rb_parent); 2289 return 0; 2290 } 2291 2292 /* 2293 * Copy the vma structure to a new location in the same mm, 2294 * prior to moving page table entries, to effect an mremap move. 2295 */ 2296 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2297 unsigned long addr, unsigned long len, pgoff_t pgoff) 2298 { 2299 struct vm_area_struct *vma = *vmap; 2300 unsigned long vma_start = vma->vm_start; 2301 struct mm_struct *mm = vma->vm_mm; 2302 struct vm_area_struct *new_vma, *prev; 2303 struct rb_node **rb_link, *rb_parent; 2304 struct mempolicy *pol; 2305 bool faulted_in_anon_vma = true; 2306 2307 /* 2308 * If anonymous vma has not yet been faulted, update new pgoff 2309 * to match new location, to increase its chance of merging. 2310 */ 2311 if (unlikely(!vma->vm_file && !vma->anon_vma)) { 2312 pgoff = addr >> PAGE_SHIFT; 2313 faulted_in_anon_vma = false; 2314 } 2315 2316 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 2317 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2318 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma)); 2319 if (new_vma) { 2320 /* 2321 * Source vma may have been merged into new_vma 2322 */ 2323 if (unlikely(vma_start >= new_vma->vm_start && 2324 vma_start < new_vma->vm_end)) { 2325 /* 2326 * The only way we can get a vma_merge with 2327 * self during an mremap is if the vma hasn't 2328 * been faulted in yet and we were allowed to 2329 * reset the dst vma->vm_pgoff to the 2330 * destination address of the mremap to allow 2331 * the merge to happen. mremap must change the 2332 * vm_pgoff linearity between src and dst vmas 2333 * (in turn preventing a vma_merge) to be 2334 * safe. It is only safe to keep the vm_pgoff 2335 * linear if there are no pages mapped yet. 2336 */ 2337 VM_BUG_ON(faulted_in_anon_vma); 2338 *vmap = new_vma; 2339 } else 2340 anon_vma_moveto_tail(new_vma); 2341 } else { 2342 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2343 if (new_vma) { 2344 *new_vma = *vma; 2345 pol = mpol_dup(vma_policy(vma)); 2346 if (IS_ERR(pol)) 2347 goto out_free_vma; 2348 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 2349 if (anon_vma_clone(new_vma, vma)) 2350 goto out_free_mempol; 2351 vma_set_policy(new_vma, pol); 2352 new_vma->vm_start = addr; 2353 new_vma->vm_end = addr + len; 2354 new_vma->vm_pgoff = pgoff; 2355 if (new_vma->vm_file) { 2356 get_file(new_vma->vm_file); 2357 if (vma->vm_flags & VM_EXECUTABLE) 2358 added_exe_file_vma(mm); 2359 } 2360 if (new_vma->vm_ops && new_vma->vm_ops->open) 2361 new_vma->vm_ops->open(new_vma); 2362 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2363 } 2364 } 2365 return new_vma; 2366 2367 out_free_mempol: 2368 mpol_put(pol); 2369 out_free_vma: 2370 kmem_cache_free(vm_area_cachep, new_vma); 2371 return NULL; 2372 } 2373 2374 /* 2375 * Return true if the calling process may expand its vm space by the passed 2376 * number of pages 2377 */ 2378 int may_expand_vm(struct mm_struct *mm, unsigned long npages) 2379 { 2380 unsigned long cur = mm->total_vm; /* pages */ 2381 unsigned long lim; 2382 2383 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT; 2384 2385 if (cur + npages > lim) 2386 return 0; 2387 return 1; 2388 } 2389 2390 2391 static int special_mapping_fault(struct vm_area_struct *vma, 2392 struct vm_fault *vmf) 2393 { 2394 pgoff_t pgoff; 2395 struct page **pages; 2396 2397 /* 2398 * special mappings have no vm_file, and in that case, the mm 2399 * uses vm_pgoff internally. So we have to subtract it from here. 2400 * We are allowed to do this because we are the mm; do not copy 2401 * this code into drivers! 2402 */ 2403 pgoff = vmf->pgoff - vma->vm_pgoff; 2404 2405 for (pages = vma->vm_private_data; pgoff && *pages; ++pages) 2406 pgoff--; 2407 2408 if (*pages) { 2409 struct page *page = *pages; 2410 get_page(page); 2411 vmf->page = page; 2412 return 0; 2413 } 2414 2415 return VM_FAULT_SIGBUS; 2416 } 2417 2418 /* 2419 * Having a close hook prevents vma merging regardless of flags. 2420 */ 2421 static void special_mapping_close(struct vm_area_struct *vma) 2422 { 2423 } 2424 2425 static const struct vm_operations_struct special_mapping_vmops = { 2426 .close = special_mapping_close, 2427 .fault = special_mapping_fault, 2428 }; 2429 2430 /* 2431 * Called with mm->mmap_sem held for writing. 2432 * Insert a new vma covering the given region, with the given flags. 2433 * Its pages are supplied by the given array of struct page *. 2434 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 2435 * The region past the last page supplied will always produce SIGBUS. 2436 * The array pointer and the pages it points to are assumed to stay alive 2437 * for as long as this mapping might exist. 2438 */ 2439 int install_special_mapping(struct mm_struct *mm, 2440 unsigned long addr, unsigned long len, 2441 unsigned long vm_flags, struct page **pages) 2442 { 2443 int ret; 2444 struct vm_area_struct *vma; 2445 2446 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2447 if (unlikely(vma == NULL)) 2448 return -ENOMEM; 2449 2450 INIT_LIST_HEAD(&vma->anon_vma_chain); 2451 vma->vm_mm = mm; 2452 vma->vm_start = addr; 2453 vma->vm_end = addr + len; 2454 2455 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND; 2456 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 2457 2458 vma->vm_ops = &special_mapping_vmops; 2459 vma->vm_private_data = pages; 2460 2461 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1); 2462 if (ret) 2463 goto out; 2464 2465 ret = insert_vm_struct(mm, vma); 2466 if (ret) 2467 goto out; 2468 2469 mm->total_vm += len >> PAGE_SHIFT; 2470 2471 perf_event_mmap(vma); 2472 2473 return 0; 2474 2475 out: 2476 kmem_cache_free(vm_area_cachep, vma); 2477 return ret; 2478 } 2479 2480 static DEFINE_MUTEX(mm_all_locks_mutex); 2481 2482 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 2483 { 2484 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) { 2485 /* 2486 * The LSB of head.next can't change from under us 2487 * because we hold the mm_all_locks_mutex. 2488 */ 2489 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem); 2490 /* 2491 * We can safely modify head.next after taking the 2492 * anon_vma->root->mutex. If some other vma in this mm shares 2493 * the same anon_vma we won't take it again. 2494 * 2495 * No need of atomic instructions here, head.next 2496 * can't change from under us thanks to the 2497 * anon_vma->root->mutex. 2498 */ 2499 if (__test_and_set_bit(0, (unsigned long *) 2500 &anon_vma->root->head.next)) 2501 BUG(); 2502 } 2503 } 2504 2505 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 2506 { 2507 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 2508 /* 2509 * AS_MM_ALL_LOCKS can't change from under us because 2510 * we hold the mm_all_locks_mutex. 2511 * 2512 * Operations on ->flags have to be atomic because 2513 * even if AS_MM_ALL_LOCKS is stable thanks to the 2514 * mm_all_locks_mutex, there may be other cpus 2515 * changing other bitflags in parallel to us. 2516 */ 2517 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 2518 BUG(); 2519 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem); 2520 } 2521 } 2522 2523 /* 2524 * This operation locks against the VM for all pte/vma/mm related 2525 * operations that could ever happen on a certain mm. This includes 2526 * vmtruncate, try_to_unmap, and all page faults. 2527 * 2528 * The caller must take the mmap_sem in write mode before calling 2529 * mm_take_all_locks(). The caller isn't allowed to release the 2530 * mmap_sem until mm_drop_all_locks() returns. 2531 * 2532 * mmap_sem in write mode is required in order to block all operations 2533 * that could modify pagetables and free pages without need of 2534 * altering the vma layout (for example populate_range() with 2535 * nonlinear vmas). It's also needed in write mode to avoid new 2536 * anon_vmas to be associated with existing vmas. 2537 * 2538 * A single task can't take more than one mm_take_all_locks() in a row 2539 * or it would deadlock. 2540 * 2541 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in 2542 * mapping->flags avoid to take the same lock twice, if more than one 2543 * vma in this mm is backed by the same anon_vma or address_space. 2544 * 2545 * We can take all the locks in random order because the VM code 2546 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never 2547 * takes more than one of them in a row. Secondly we're protected 2548 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex. 2549 * 2550 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 2551 * that may have to take thousand of locks. 2552 * 2553 * mm_take_all_locks() can fail if it's interrupted by signals. 2554 */ 2555 int mm_take_all_locks(struct mm_struct *mm) 2556 { 2557 struct vm_area_struct *vma; 2558 struct anon_vma_chain *avc; 2559 2560 BUG_ON(down_read_trylock(&mm->mmap_sem)); 2561 2562 mutex_lock(&mm_all_locks_mutex); 2563 2564 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2565 if (signal_pending(current)) 2566 goto out_unlock; 2567 if (vma->vm_file && vma->vm_file->f_mapping) 2568 vm_lock_mapping(mm, vma->vm_file->f_mapping); 2569 } 2570 2571 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2572 if (signal_pending(current)) 2573 goto out_unlock; 2574 if (vma->anon_vma) 2575 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 2576 vm_lock_anon_vma(mm, avc->anon_vma); 2577 } 2578 2579 return 0; 2580 2581 out_unlock: 2582 mm_drop_all_locks(mm); 2583 return -EINTR; 2584 } 2585 2586 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 2587 { 2588 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) { 2589 /* 2590 * The LSB of head.next can't change to 0 from under 2591 * us because we hold the mm_all_locks_mutex. 2592 * 2593 * We must however clear the bitflag before unlocking 2594 * the vma so the users using the anon_vma->head will 2595 * never see our bitflag. 2596 * 2597 * No need of atomic instructions here, head.next 2598 * can't change from under us until we release the 2599 * anon_vma->root->mutex. 2600 */ 2601 if (!__test_and_clear_bit(0, (unsigned long *) 2602 &anon_vma->root->head.next)) 2603 BUG(); 2604 anon_vma_unlock(anon_vma); 2605 } 2606 } 2607 2608 static void vm_unlock_mapping(struct address_space *mapping) 2609 { 2610 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 2611 /* 2612 * AS_MM_ALL_LOCKS can't change to 0 from under us 2613 * because we hold the mm_all_locks_mutex. 2614 */ 2615 mutex_unlock(&mapping->i_mmap_mutex); 2616 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 2617 &mapping->flags)) 2618 BUG(); 2619 } 2620 } 2621 2622 /* 2623 * The mmap_sem cannot be released by the caller until 2624 * mm_drop_all_locks() returns. 2625 */ 2626 void mm_drop_all_locks(struct mm_struct *mm) 2627 { 2628 struct vm_area_struct *vma; 2629 struct anon_vma_chain *avc; 2630 2631 BUG_ON(down_read_trylock(&mm->mmap_sem)); 2632 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 2633 2634 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2635 if (vma->anon_vma) 2636 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 2637 vm_unlock_anon_vma(avc->anon_vma); 2638 if (vma->vm_file && vma->vm_file->f_mapping) 2639 vm_unlock_mapping(vma->vm_file->f_mapping); 2640 } 2641 2642 mutex_unlock(&mm_all_locks_mutex); 2643 } 2644 2645 /* 2646 * initialise the VMA slab 2647 */ 2648 void __init mmap_init(void) 2649 { 2650 int ret; 2651 2652 ret = percpu_counter_init(&vm_committed_as, 0); 2653 VM_BUG_ON(ret); 2654 } 2655