mmap.c revision 8314c4f24a0a5c9b1f7544e9fa83a1d5367ddaa7
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/ima.h> 24#include <linux/hugetlb.h> 25#include <linux/profile.h> 26#include <linux/module.h> 27#include <linux/mount.h> 28#include <linux/mempolicy.h> 29#include <linux/rmap.h> 30#include <linux/ksm.h> 31#include <linux/mmu_notifier.h> 32#include <linux/perf_event.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 49static 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 */ 74pgprot_t protection_map[16] = { 75 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 76 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 77}; 78 79pgprot_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} 85EXPORT_SYMBOL(vm_get_page_prot); 86 87int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */ 88int sysctl_overcommit_ratio = 50; /* default is 50% */ 89int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT; 90struct percpu_counter vm_committed_as; 91 92/* 93 * Check that a process has enough memory to allocate a new virtual 94 * mapping. 0 means there is enough memory for the allocation to 95 * succeed and -ENOMEM implies there is not. 96 * 97 * We currently support three overcommit policies, which are set via the 98 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting 99 * 100 * Strict overcommit modes added 2002 Feb 26 by Alan Cox. 101 * Additional code 2002 Jul 20 by Robert Love. 102 * 103 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise. 104 * 105 * Note this is a helper function intended to be used by LSMs which 106 * wish to use this logic. 107 */ 108int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin) 109{ 110 unsigned long free, allowed; 111 112 vm_acct_memory(pages); 113 114 /* 115 * Sometimes we want to use more memory than we have 116 */ 117 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS) 118 return 0; 119 120 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) { 121 unsigned long n; 122 123 free = global_page_state(NR_FILE_PAGES); 124 free += nr_swap_pages; 125 126 /* 127 * Any slabs which are created with the 128 * SLAB_RECLAIM_ACCOUNT flag claim to have contents 129 * which are reclaimable, under pressure. The dentry 130 * cache and most inode caches should fall into this 131 */ 132 free += global_page_state(NR_SLAB_RECLAIMABLE); 133 134 /* 135 * Leave the last 3% for root 136 */ 137 if (!cap_sys_admin) 138 free -= free / 32; 139 140 if (free > pages) 141 return 0; 142 143 /* 144 * nr_free_pages() is very expensive on large systems, 145 * only call if we're about to fail. 146 */ 147 n = nr_free_pages(); 148 149 /* 150 * Leave reserved pages. The pages are not for anonymous pages. 151 */ 152 if (n <= totalreserve_pages) 153 goto error; 154 else 155 n -= totalreserve_pages; 156 157 /* 158 * Leave the last 3% for root 159 */ 160 if (!cap_sys_admin) 161 n -= n / 32; 162 free += n; 163 164 if (free > pages) 165 return 0; 166 167 goto error; 168 } 169 170 allowed = (totalram_pages - hugetlb_total_pages()) 171 * sysctl_overcommit_ratio / 100; 172 /* 173 * Leave the last 3% for root 174 */ 175 if (!cap_sys_admin) 176 allowed -= allowed / 32; 177 allowed += total_swap_pages; 178 179 /* Don't let a single process grow too big: 180 leave 3% of the size of this process for other processes */ 181 if (mm) 182 allowed -= mm->total_vm / 32; 183 184 if (percpu_counter_read_positive(&vm_committed_as) < allowed) 185 return 0; 186error: 187 vm_unacct_memory(pages); 188 189 return -ENOMEM; 190} 191 192/* 193 * Requires inode->i_mapping->i_mmap_lock 194 */ 195static void __remove_shared_vm_struct(struct vm_area_struct *vma, 196 struct file *file, struct address_space *mapping) 197{ 198 if (vma->vm_flags & VM_DENYWRITE) 199 atomic_inc(&file->f_path.dentry->d_inode->i_writecount); 200 if (vma->vm_flags & VM_SHARED) 201 mapping->i_mmap_writable--; 202 203 flush_dcache_mmap_lock(mapping); 204 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 205 list_del_init(&vma->shared.vm_set.list); 206 else 207 vma_prio_tree_remove(vma, &mapping->i_mmap); 208 flush_dcache_mmap_unlock(mapping); 209} 210 211/* 212 * Unlink a file-based vm structure from its prio_tree, to hide 213 * vma from rmap and vmtruncate before freeing its page tables. 214 */ 215void unlink_file_vma(struct vm_area_struct *vma) 216{ 217 struct file *file = vma->vm_file; 218 219 if (file) { 220 struct address_space *mapping = file->f_mapping; 221 spin_lock(&mapping->i_mmap_lock); 222 __remove_shared_vm_struct(vma, file, mapping); 223 spin_unlock(&mapping->i_mmap_lock); 224 } 225} 226 227/* 228 * Close a vm structure and free it, returning the next. 229 */ 230static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 231{ 232 struct vm_area_struct *next = vma->vm_next; 233 234 might_sleep(); 235 if (vma->vm_ops && vma->vm_ops->close) 236 vma->vm_ops->close(vma); 237 if (vma->vm_file) { 238 fput(vma->vm_file); 239 if (vma->vm_flags & VM_EXECUTABLE) 240 removed_exe_file_vma(vma->vm_mm); 241 } 242 mpol_put(vma_policy(vma)); 243 kmem_cache_free(vm_area_cachep, vma); 244 return next; 245} 246 247SYSCALL_DEFINE1(brk, unsigned long, brk) 248{ 249 unsigned long rlim, retval; 250 unsigned long newbrk, oldbrk; 251 struct mm_struct *mm = current->mm; 252 unsigned long min_brk; 253 254 down_write(&mm->mmap_sem); 255 256#ifdef CONFIG_COMPAT_BRK 257 min_brk = mm->end_code; 258#else 259 min_brk = mm->start_brk; 260#endif 261 if (brk < min_brk) 262 goto out; 263 264 /* 265 * Check against rlimit here. If this check is done later after the test 266 * of oldbrk with newbrk then it can escape the test and let the data 267 * segment grow beyond its set limit the in case where the limit is 268 * not page aligned -Ram Gupta 269 */ 270 rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur; 271 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) + 272 (mm->end_data - mm->start_data) > rlim) 273 goto out; 274 275 newbrk = PAGE_ALIGN(brk); 276 oldbrk = PAGE_ALIGN(mm->brk); 277 if (oldbrk == newbrk) 278 goto set_brk; 279 280 /* Always allow shrinking brk. */ 281 if (brk <= mm->brk) { 282 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 283 goto set_brk; 284 goto out; 285 } 286 287 /* Check against existing mmap mappings. */ 288 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 289 goto out; 290 291 /* Ok, looks good - let it rip. */ 292 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 293 goto out; 294set_brk: 295 mm->brk = brk; 296out: 297 retval = mm->brk; 298 up_write(&mm->mmap_sem); 299 return retval; 300} 301 302#ifdef DEBUG_MM_RB 303static int browse_rb(struct rb_root *root) 304{ 305 int i = 0, j; 306 struct rb_node *nd, *pn = NULL; 307 unsigned long prev = 0, pend = 0; 308 309 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 310 struct vm_area_struct *vma; 311 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 312 if (vma->vm_start < prev) 313 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1; 314 if (vma->vm_start < pend) 315 printk("vm_start %lx pend %lx\n", vma->vm_start, pend); 316 if (vma->vm_start > vma->vm_end) 317 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start); 318 i++; 319 pn = nd; 320 prev = vma->vm_start; 321 pend = vma->vm_end; 322 } 323 j = 0; 324 for (nd = pn; nd; nd = rb_prev(nd)) { 325 j++; 326 } 327 if (i != j) 328 printk("backwards %d, forwards %d\n", j, i), i = 0; 329 return i; 330} 331 332void validate_mm(struct mm_struct *mm) 333{ 334 int bug = 0; 335 int i = 0; 336 struct vm_area_struct *tmp = mm->mmap; 337 while (tmp) { 338 tmp = tmp->vm_next; 339 i++; 340 } 341 if (i != mm->map_count) 342 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1; 343 i = browse_rb(&mm->mm_rb); 344 if (i != mm->map_count) 345 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1; 346 BUG_ON(bug); 347} 348#else 349#define validate_mm(mm) do { } while (0) 350#endif 351 352static struct vm_area_struct * 353find_vma_prepare(struct mm_struct *mm, unsigned long addr, 354 struct vm_area_struct **pprev, struct rb_node ***rb_link, 355 struct rb_node ** rb_parent) 356{ 357 struct vm_area_struct * vma; 358 struct rb_node ** __rb_link, * __rb_parent, * rb_prev; 359 360 __rb_link = &mm->mm_rb.rb_node; 361 rb_prev = __rb_parent = NULL; 362 vma = NULL; 363 364 while (*__rb_link) { 365 struct vm_area_struct *vma_tmp; 366 367 __rb_parent = *__rb_link; 368 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 369 370 if (vma_tmp->vm_end > addr) { 371 vma = vma_tmp; 372 if (vma_tmp->vm_start <= addr) 373 break; 374 __rb_link = &__rb_parent->rb_left; 375 } else { 376 rb_prev = __rb_parent; 377 __rb_link = &__rb_parent->rb_right; 378 } 379 } 380 381 *pprev = NULL; 382 if (rb_prev) 383 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 384 *rb_link = __rb_link; 385 *rb_parent = __rb_parent; 386 return vma; 387} 388 389static inline void 390__vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma, 391 struct vm_area_struct *prev, struct rb_node *rb_parent) 392{ 393 if (prev) { 394 vma->vm_next = prev->vm_next; 395 prev->vm_next = vma; 396 } else { 397 mm->mmap = vma; 398 if (rb_parent) 399 vma->vm_next = rb_entry(rb_parent, 400 struct vm_area_struct, vm_rb); 401 else 402 vma->vm_next = NULL; 403 } 404} 405 406void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 407 struct rb_node **rb_link, struct rb_node *rb_parent) 408{ 409 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 410 rb_insert_color(&vma->vm_rb, &mm->mm_rb); 411} 412 413static void __vma_link_file(struct vm_area_struct *vma) 414{ 415 struct file *file; 416 417 file = vma->vm_file; 418 if (file) { 419 struct address_space *mapping = file->f_mapping; 420 421 if (vma->vm_flags & VM_DENYWRITE) 422 atomic_dec(&file->f_path.dentry->d_inode->i_writecount); 423 if (vma->vm_flags & VM_SHARED) 424 mapping->i_mmap_writable++; 425 426 flush_dcache_mmap_lock(mapping); 427 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 428 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear); 429 else 430 vma_prio_tree_insert(vma, &mapping->i_mmap); 431 flush_dcache_mmap_unlock(mapping); 432 } 433} 434 435static void 436__vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 437 struct vm_area_struct *prev, struct rb_node **rb_link, 438 struct rb_node *rb_parent) 439{ 440 __vma_link_list(mm, vma, prev, rb_parent); 441 __vma_link_rb(mm, vma, rb_link, rb_parent); 442 __anon_vma_link(vma); 443} 444 445static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 446 struct vm_area_struct *prev, struct rb_node **rb_link, 447 struct rb_node *rb_parent) 448{ 449 struct address_space *mapping = NULL; 450 451 if (vma->vm_file) 452 mapping = vma->vm_file->f_mapping; 453 454 if (mapping) { 455 spin_lock(&mapping->i_mmap_lock); 456 vma->vm_truncate_count = mapping->truncate_count; 457 } 458 anon_vma_lock(vma); 459 460 __vma_link(mm, vma, prev, rb_link, rb_parent); 461 __vma_link_file(vma); 462 463 anon_vma_unlock(vma); 464 if (mapping) 465 spin_unlock(&mapping->i_mmap_lock); 466 467 mm->map_count++; 468 validate_mm(mm); 469} 470 471/* 472 * Helper for vma_adjust in the split_vma insert case: 473 * insert vm structure into list and rbtree and anon_vma, 474 * but it has already been inserted into prio_tree earlier. 475 */ 476static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 477{ 478 struct vm_area_struct *__vma, *prev; 479 struct rb_node **rb_link, *rb_parent; 480 481 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent); 482 BUG_ON(__vma && __vma->vm_start < vma->vm_end); 483 __vma_link(mm, vma, prev, rb_link, rb_parent); 484 mm->map_count++; 485} 486 487static inline void 488__vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 489 struct vm_area_struct *prev) 490{ 491 prev->vm_next = vma->vm_next; 492 rb_erase(&vma->vm_rb, &mm->mm_rb); 493 if (mm->mmap_cache == vma) 494 mm->mmap_cache = prev; 495} 496 497/* 498 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 499 * is already present in an i_mmap tree without adjusting the tree. 500 * The following helper function should be used when such adjustments 501 * are necessary. The "insert" vma (if any) is to be inserted 502 * before we drop the necessary locks. 503 */ 504void vma_adjust(struct vm_area_struct *vma, unsigned long start, 505 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 506{ 507 struct mm_struct *mm = vma->vm_mm; 508 struct vm_area_struct *next = vma->vm_next; 509 struct vm_area_struct *importer = NULL; 510 struct address_space *mapping = NULL; 511 struct prio_tree_root *root = NULL; 512 struct file *file = vma->vm_file; 513 struct anon_vma *anon_vma = NULL; 514 long adjust_next = 0; 515 int remove_next = 0; 516 517 if (next && !insert) { 518 if (end >= next->vm_end) { 519 /* 520 * vma expands, overlapping all the next, and 521 * perhaps the one after too (mprotect case 6). 522 */ 523again: remove_next = 1 + (end > next->vm_end); 524 end = next->vm_end; 525 anon_vma = next->anon_vma; 526 importer = vma; 527 } else if (end > next->vm_start) { 528 /* 529 * vma expands, overlapping part of the next: 530 * mprotect case 5 shifting the boundary up. 531 */ 532 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 533 anon_vma = next->anon_vma; 534 importer = vma; 535 } else if (end < vma->vm_end) { 536 /* 537 * vma shrinks, and !insert tells it's not 538 * split_vma inserting another: so it must be 539 * mprotect case 4 shifting the boundary down. 540 */ 541 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT); 542 anon_vma = next->anon_vma; 543 importer = next; 544 } 545 } 546 547 if (file) { 548 mapping = file->f_mapping; 549 if (!(vma->vm_flags & VM_NONLINEAR)) 550 root = &mapping->i_mmap; 551 spin_lock(&mapping->i_mmap_lock); 552 if (importer && 553 vma->vm_truncate_count != next->vm_truncate_count) { 554 /* 555 * unmap_mapping_range might be in progress: 556 * ensure that the expanding vma is rescanned. 557 */ 558 importer->vm_truncate_count = 0; 559 } 560 if (insert) { 561 insert->vm_truncate_count = vma->vm_truncate_count; 562 /* 563 * Put into prio_tree now, so instantiated pages 564 * are visible to arm/parisc __flush_dcache_page 565 * throughout; but we cannot insert into address 566 * space until vma start or end is updated. 567 */ 568 __vma_link_file(insert); 569 } 570 } 571 572 /* 573 * When changing only vma->vm_end, we don't really need 574 * anon_vma lock: but is that case worth optimizing out? 575 */ 576 if (vma->anon_vma) 577 anon_vma = vma->anon_vma; 578 if (anon_vma) { 579 spin_lock(&anon_vma->lock); 580 /* 581 * Easily overlooked: when mprotect shifts the boundary, 582 * make sure the expanding vma has anon_vma set if the 583 * shrinking vma had, to cover any anon pages imported. 584 */ 585 if (importer && !importer->anon_vma) { 586 importer->anon_vma = anon_vma; 587 __anon_vma_link(importer); 588 } 589 } 590 591 if (root) { 592 flush_dcache_mmap_lock(mapping); 593 vma_prio_tree_remove(vma, root); 594 if (adjust_next) 595 vma_prio_tree_remove(next, root); 596 } 597 598 vma->vm_start = start; 599 vma->vm_end = end; 600 vma->vm_pgoff = pgoff; 601 if (adjust_next) { 602 next->vm_start += adjust_next << PAGE_SHIFT; 603 next->vm_pgoff += adjust_next; 604 } 605 606 if (root) { 607 if (adjust_next) 608 vma_prio_tree_insert(next, root); 609 vma_prio_tree_insert(vma, root); 610 flush_dcache_mmap_unlock(mapping); 611 } 612 613 if (remove_next) { 614 /* 615 * vma_merge has merged next into vma, and needs 616 * us to remove next before dropping the locks. 617 */ 618 __vma_unlink(mm, next, vma); 619 if (file) 620 __remove_shared_vm_struct(next, file, mapping); 621 if (next->anon_vma) 622 __anon_vma_merge(vma, next); 623 } else if (insert) { 624 /* 625 * split_vma has split insert from vma, and needs 626 * us to insert it before dropping the locks 627 * (it may either follow vma or precede it). 628 */ 629 __insert_vm_struct(mm, insert); 630 } 631 632 if (anon_vma) 633 spin_unlock(&anon_vma->lock); 634 if (mapping) 635 spin_unlock(&mapping->i_mmap_lock); 636 637 if (remove_next) { 638 if (file) { 639 fput(file); 640 if (next->vm_flags & VM_EXECUTABLE) 641 removed_exe_file_vma(mm); 642 } 643 mm->map_count--; 644 mpol_put(vma_policy(next)); 645 kmem_cache_free(vm_area_cachep, next); 646 /* 647 * In mprotect's case 6 (see comments on vma_merge), 648 * we must remove another next too. It would clutter 649 * up the code too much to do both in one go. 650 */ 651 if (remove_next == 2) { 652 next = vma->vm_next; 653 goto again; 654 } 655 } 656 657 validate_mm(mm); 658} 659 660/* 661 * If the vma has a ->close operation then the driver probably needs to release 662 * per-vma resources, so we don't attempt to merge those. 663 */ 664static inline int is_mergeable_vma(struct vm_area_struct *vma, 665 struct file *file, unsigned long vm_flags) 666{ 667 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */ 668 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR) 669 return 0; 670 if (vma->vm_file != file) 671 return 0; 672 if (vma->vm_ops && vma->vm_ops->close) 673 return 0; 674 return 1; 675} 676 677static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 678 struct anon_vma *anon_vma2) 679{ 680 return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2); 681} 682 683/* 684 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 685 * in front of (at a lower virtual address and file offset than) the vma. 686 * 687 * We cannot merge two vmas if they have differently assigned (non-NULL) 688 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 689 * 690 * We don't check here for the merged mmap wrapping around the end of pagecache 691 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 692 * wrap, nor mmaps which cover the final page at index -1UL. 693 */ 694static int 695can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 696 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 697{ 698 if (is_mergeable_vma(vma, file, vm_flags) && 699 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) { 700 if (vma->vm_pgoff == vm_pgoff) 701 return 1; 702 } 703 return 0; 704} 705 706/* 707 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 708 * beyond (at a higher virtual address and file offset than) the vma. 709 * 710 * We cannot merge two vmas if they have differently assigned (non-NULL) 711 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 712 */ 713static int 714can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 715 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 716{ 717 if (is_mergeable_vma(vma, file, vm_flags) && 718 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) { 719 pgoff_t vm_pglen; 720 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; 721 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 722 return 1; 723 } 724 return 0; 725} 726 727/* 728 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 729 * whether that can be merged with its predecessor or its successor. 730 * Or both (it neatly fills a hole). 731 * 732 * In most cases - when called for mmap, brk or mremap - [addr,end) is 733 * certain not to be mapped by the time vma_merge is called; but when 734 * called for mprotect, it is certain to be already mapped (either at 735 * an offset within prev, or at the start of next), and the flags of 736 * this area are about to be changed to vm_flags - and the no-change 737 * case has already been eliminated. 738 * 739 * The following mprotect cases have to be considered, where AAAA is 740 * the area passed down from mprotect_fixup, never extending beyond one 741 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 742 * 743 * AAAA AAAA AAAA AAAA 744 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 745 * cannot merge might become might become might become 746 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 747 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 748 * mremap move: PPPPNNNNNNNN 8 749 * AAAA 750 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 751 * might become case 1 below case 2 below case 3 below 752 * 753 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 754 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 755 */ 756struct vm_area_struct *vma_merge(struct mm_struct *mm, 757 struct vm_area_struct *prev, unsigned long addr, 758 unsigned long end, unsigned long vm_flags, 759 struct anon_vma *anon_vma, struct file *file, 760 pgoff_t pgoff, struct mempolicy *policy) 761{ 762 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 763 struct vm_area_struct *area, *next; 764 765 /* 766 * We later require that vma->vm_flags == vm_flags, 767 * so this tests vma->vm_flags & VM_SPECIAL, too. 768 */ 769 if (vm_flags & VM_SPECIAL) 770 return NULL; 771 772 if (prev) 773 next = prev->vm_next; 774 else 775 next = mm->mmap; 776 area = next; 777 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 778 next = next->vm_next; 779 780 /* 781 * Can it merge with the predecessor? 782 */ 783 if (prev && prev->vm_end == addr && 784 mpol_equal(vma_policy(prev), policy) && 785 can_vma_merge_after(prev, vm_flags, 786 anon_vma, file, pgoff)) { 787 /* 788 * OK, it can. Can we now merge in the successor as well? 789 */ 790 if (next && end == next->vm_start && 791 mpol_equal(policy, vma_policy(next)) && 792 can_vma_merge_before(next, vm_flags, 793 anon_vma, file, pgoff+pglen) && 794 is_mergeable_anon_vma(prev->anon_vma, 795 next->anon_vma)) { 796 /* cases 1, 6 */ 797 vma_adjust(prev, prev->vm_start, 798 next->vm_end, prev->vm_pgoff, NULL); 799 } else /* cases 2, 5, 7 */ 800 vma_adjust(prev, prev->vm_start, 801 end, prev->vm_pgoff, NULL); 802 return prev; 803 } 804 805 /* 806 * Can this new request be merged in front of next? 807 */ 808 if (next && end == next->vm_start && 809 mpol_equal(policy, vma_policy(next)) && 810 can_vma_merge_before(next, vm_flags, 811 anon_vma, file, pgoff+pglen)) { 812 if (prev && addr < prev->vm_end) /* case 4 */ 813 vma_adjust(prev, prev->vm_start, 814 addr, prev->vm_pgoff, NULL); 815 else /* cases 3, 8 */ 816 vma_adjust(area, addr, next->vm_end, 817 next->vm_pgoff - pglen, NULL); 818 return area; 819 } 820 821 return NULL; 822} 823 824/* 825 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 826 * neighbouring vmas for a suitable anon_vma, before it goes off 827 * to allocate a new anon_vma. It checks because a repetitive 828 * sequence of mprotects and faults may otherwise lead to distinct 829 * anon_vmas being allocated, preventing vma merge in subsequent 830 * mprotect. 831 */ 832struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 833{ 834 struct vm_area_struct *near; 835 unsigned long vm_flags; 836 837 near = vma->vm_next; 838 if (!near) 839 goto try_prev; 840 841 /* 842 * Since only mprotect tries to remerge vmas, match flags 843 * which might be mprotected into each other later on. 844 * Neither mlock nor madvise tries to remerge at present, 845 * so leave their flags as obstructing a merge. 846 */ 847 vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC); 848 vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC); 849 850 if (near->anon_vma && vma->vm_end == near->vm_start && 851 mpol_equal(vma_policy(vma), vma_policy(near)) && 852 can_vma_merge_before(near, vm_flags, 853 NULL, vma->vm_file, vma->vm_pgoff + 854 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT))) 855 return near->anon_vma; 856try_prev: 857 /* 858 * It is potentially slow to have to call find_vma_prev here. 859 * But it's only on the first write fault on the vma, not 860 * every time, and we could devise a way to avoid it later 861 * (e.g. stash info in next's anon_vma_node when assigning 862 * an anon_vma, or when trying vma_merge). Another time. 863 */ 864 BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma); 865 if (!near) 866 goto none; 867 868 vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC); 869 vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC); 870 871 if (near->anon_vma && near->vm_end == vma->vm_start && 872 mpol_equal(vma_policy(near), vma_policy(vma)) && 873 can_vma_merge_after(near, vm_flags, 874 NULL, vma->vm_file, vma->vm_pgoff)) 875 return near->anon_vma; 876none: 877 /* 878 * There's no absolute need to look only at touching neighbours: 879 * we could search further afield for "compatible" anon_vmas. 880 * But it would probably just be a waste of time searching, 881 * or lead to too many vmas hanging off the same anon_vma. 882 * We're trying to allow mprotect remerging later on, 883 * not trying to minimize memory used for anon_vmas. 884 */ 885 return NULL; 886} 887 888#ifdef CONFIG_PROC_FS 889void vm_stat_account(struct mm_struct *mm, unsigned long flags, 890 struct file *file, long pages) 891{ 892 const unsigned long stack_flags 893 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN); 894 895 if (file) { 896 mm->shared_vm += pages; 897 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC) 898 mm->exec_vm += pages; 899 } else if (flags & stack_flags) 900 mm->stack_vm += pages; 901 if (flags & (VM_RESERVED|VM_IO)) 902 mm->reserved_vm += pages; 903} 904#endif /* CONFIG_PROC_FS */ 905 906/* 907 * The caller must hold down_write(¤t->mm->mmap_sem). 908 */ 909 910unsigned long do_mmap_pgoff(struct file *file, unsigned long addr, 911 unsigned long len, unsigned long prot, 912 unsigned long flags, unsigned long pgoff) 913{ 914 struct mm_struct * mm = current->mm; 915 struct inode *inode; 916 unsigned int vm_flags; 917 int error; 918 unsigned long reqprot = prot; 919 920 /* 921 * Does the application expect PROT_READ to imply PROT_EXEC? 922 * 923 * (the exception is when the underlying filesystem is noexec 924 * mounted, in which case we dont add PROT_EXEC.) 925 */ 926 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 927 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC))) 928 prot |= PROT_EXEC; 929 930 if (!len) 931 return -EINVAL; 932 933 if (!(flags & MAP_FIXED)) 934 addr = round_hint_to_min(addr); 935 936 error = arch_mmap_check(addr, len, flags); 937 if (error) 938 return error; 939 940 /* Careful about overflows.. */ 941 len = PAGE_ALIGN(len); 942 if (!len || len > TASK_SIZE) 943 return -ENOMEM; 944 945 /* offset overflow? */ 946 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 947 return -EOVERFLOW; 948 949 /* Too many mappings? */ 950 if (mm->map_count > sysctl_max_map_count) 951 return -ENOMEM; 952 953 /* Obtain the address to map to. we verify (or select) it and ensure 954 * that it represents a valid section of the address space. 955 */ 956 addr = get_unmapped_area(file, addr, len, pgoff, flags); 957 if (addr & ~PAGE_MASK) 958 return addr; 959 960 /* Do simple checking here so the lower-level routines won't have 961 * to. we assume access permissions have been handled by the open 962 * of the memory object, so we don't do any here. 963 */ 964 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 965 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 966 967 if (flags & MAP_LOCKED) { 968 if (!can_do_mlock()) 969 return -EPERM; 970 vm_flags |= VM_LOCKED; 971 } 972 973 /* mlock MCL_FUTURE? */ 974 if (vm_flags & VM_LOCKED) { 975 unsigned long locked, lock_limit; 976 locked = len >> PAGE_SHIFT; 977 locked += mm->locked_vm; 978 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur; 979 lock_limit >>= PAGE_SHIFT; 980 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 981 return -EAGAIN; 982 } 983 984 inode = file ? file->f_path.dentry->d_inode : NULL; 985 986 if (file) { 987 switch (flags & MAP_TYPE) { 988 case MAP_SHARED: 989 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 990 return -EACCES; 991 992 /* 993 * Make sure we don't allow writing to an append-only 994 * file.. 995 */ 996 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 997 return -EACCES; 998 999 /* 1000 * Make sure there are no mandatory locks on the file. 1001 */ 1002 if (locks_verify_locked(inode)) 1003 return -EAGAIN; 1004 1005 vm_flags |= VM_SHARED | VM_MAYSHARE; 1006 if (!(file->f_mode & FMODE_WRITE)) 1007 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1008 1009 /* fall through */ 1010 case MAP_PRIVATE: 1011 if (!(file->f_mode & FMODE_READ)) 1012 return -EACCES; 1013 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) { 1014 if (vm_flags & VM_EXEC) 1015 return -EPERM; 1016 vm_flags &= ~VM_MAYEXEC; 1017 } 1018 1019 if (!file->f_op || !file->f_op->mmap) 1020 return -ENODEV; 1021 break; 1022 1023 default: 1024 return -EINVAL; 1025 } 1026 } else { 1027 switch (flags & MAP_TYPE) { 1028 case MAP_SHARED: 1029 /* 1030 * Ignore pgoff. 1031 */ 1032 pgoff = 0; 1033 vm_flags |= VM_SHARED | VM_MAYSHARE; 1034 break; 1035 case MAP_PRIVATE: 1036 /* 1037 * Set pgoff according to addr for anon_vma. 1038 */ 1039 pgoff = addr >> PAGE_SHIFT; 1040 break; 1041 default: 1042 return -EINVAL; 1043 } 1044 } 1045 1046 error = security_file_mmap(file, reqprot, prot, flags, addr, 0); 1047 if (error) 1048 return error; 1049 error = ima_file_mmap(file, prot); 1050 if (error) 1051 return error; 1052 1053 return mmap_region(file, addr, len, flags, vm_flags, pgoff); 1054} 1055EXPORT_SYMBOL(do_mmap_pgoff); 1056 1057/* 1058 * Some shared mappigns will want the pages marked read-only 1059 * to track write events. If so, we'll downgrade vm_page_prot 1060 * to the private version (using protection_map[] without the 1061 * VM_SHARED bit). 1062 */ 1063int vma_wants_writenotify(struct vm_area_struct *vma) 1064{ 1065 unsigned int vm_flags = vma->vm_flags; 1066 1067 /* If it was private or non-writable, the write bit is already clear */ 1068 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1069 return 0; 1070 1071 /* The backer wishes to know when pages are first written to? */ 1072 if (vma->vm_ops && vma->vm_ops->page_mkwrite) 1073 return 1; 1074 1075 /* The open routine did something to the protections already? */ 1076 if (pgprot_val(vma->vm_page_prot) != 1077 pgprot_val(vm_get_page_prot(vm_flags))) 1078 return 0; 1079 1080 /* Specialty mapping? */ 1081 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE)) 1082 return 0; 1083 1084 /* Can the mapping track the dirty pages? */ 1085 return vma->vm_file && vma->vm_file->f_mapping && 1086 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1087} 1088 1089/* 1090 * We account for memory if it's a private writeable mapping, 1091 * not hugepages and VM_NORESERVE wasn't set. 1092 */ 1093static inline int accountable_mapping(struct file *file, unsigned int vm_flags) 1094{ 1095 /* 1096 * hugetlb has its own accounting separate from the core VM 1097 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1098 */ 1099 if (file && is_file_hugepages(file)) 1100 return 0; 1101 1102 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1103} 1104 1105unsigned long mmap_region(struct file *file, unsigned long addr, 1106 unsigned long len, unsigned long flags, 1107 unsigned int vm_flags, unsigned long pgoff) 1108{ 1109 struct mm_struct *mm = current->mm; 1110 struct vm_area_struct *vma, *prev; 1111 int correct_wcount = 0; 1112 int error; 1113 struct rb_node **rb_link, *rb_parent; 1114 unsigned long charged = 0; 1115 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL; 1116 1117 /* Clear old maps */ 1118 error = -ENOMEM; 1119munmap_back: 1120 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 1121 if (vma && vma->vm_start < addr + len) { 1122 if (do_munmap(mm, addr, len)) 1123 return -ENOMEM; 1124 goto munmap_back; 1125 } 1126 1127 /* Check against address space limit. */ 1128 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 1129 return -ENOMEM; 1130 1131 /* 1132 * Set 'VM_NORESERVE' if we should not account for the 1133 * memory use of this mapping. 1134 */ 1135 if ((flags & MAP_NORESERVE)) { 1136 /* We honor MAP_NORESERVE if allowed to overcommit */ 1137 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1138 vm_flags |= VM_NORESERVE; 1139 1140 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1141 if (file && is_file_hugepages(file)) 1142 vm_flags |= VM_NORESERVE; 1143 } 1144 1145 /* 1146 * Private writable mapping: check memory availability 1147 */ 1148 if (accountable_mapping(file, vm_flags)) { 1149 charged = len >> PAGE_SHIFT; 1150 if (security_vm_enough_memory(charged)) 1151 return -ENOMEM; 1152 vm_flags |= VM_ACCOUNT; 1153 } 1154 1155 /* 1156 * Can we just expand an old mapping? 1157 */ 1158 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL); 1159 if (vma) 1160 goto out; 1161 1162 /* 1163 * Determine the object being mapped and call the appropriate 1164 * specific mapper. the address has already been validated, but 1165 * not unmapped, but the maps are removed from the list. 1166 */ 1167 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1168 if (!vma) { 1169 error = -ENOMEM; 1170 goto unacct_error; 1171 } 1172 1173 vma->vm_mm = mm; 1174 vma->vm_start = addr; 1175 vma->vm_end = addr + len; 1176 vma->vm_flags = vm_flags; 1177 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1178 vma->vm_pgoff = pgoff; 1179 1180 if (file) { 1181 error = -EINVAL; 1182 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1183 goto free_vma; 1184 if (vm_flags & VM_DENYWRITE) { 1185 error = deny_write_access(file); 1186 if (error) 1187 goto free_vma; 1188 correct_wcount = 1; 1189 } 1190 vma->vm_file = file; 1191 get_file(file); 1192 error = file->f_op->mmap(file, vma); 1193 if (error) 1194 goto unmap_and_free_vma; 1195 if (vm_flags & VM_EXECUTABLE) 1196 added_exe_file_vma(mm); 1197 } else if (vm_flags & VM_SHARED) { 1198 error = shmem_zero_setup(vma); 1199 if (error) 1200 goto free_vma; 1201 } 1202 1203 /* Can addr have changed?? 1204 * 1205 * Answer: Yes, several device drivers can do it in their 1206 * f_op->mmap method. -DaveM 1207 */ 1208 addr = vma->vm_start; 1209 pgoff = vma->vm_pgoff; 1210 vm_flags = vma->vm_flags; 1211 1212 if (vma_wants_writenotify(vma)) 1213 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED); 1214 1215 vma_link(mm, vma, prev, rb_link, rb_parent); 1216 file = vma->vm_file; 1217 1218 /* Once vma denies write, undo our temporary denial count */ 1219 if (correct_wcount) 1220 atomic_inc(&inode->i_writecount); 1221out: 1222 perf_event_mmap(vma); 1223 1224 mm->total_vm += len >> PAGE_SHIFT; 1225 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT); 1226 if (vm_flags & VM_LOCKED) { 1227 /* 1228 * makes pages present; downgrades, drops, reacquires mmap_sem 1229 */ 1230 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len); 1231 if (nr_pages < 0) 1232 return nr_pages; /* vma gone! */ 1233 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages; 1234 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK)) 1235 make_pages_present(addr, addr + len); 1236 return addr; 1237 1238unmap_and_free_vma: 1239 if (correct_wcount) 1240 atomic_inc(&inode->i_writecount); 1241 vma->vm_file = NULL; 1242 fput(file); 1243 1244 /* Undo any partial mapping done by a device driver. */ 1245 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1246 charged = 0; 1247free_vma: 1248 kmem_cache_free(vm_area_cachep, vma); 1249unacct_error: 1250 if (charged) 1251 vm_unacct_memory(charged); 1252 return error; 1253} 1254 1255/* Get an address range which is currently unmapped. 1256 * For shmat() with addr=0. 1257 * 1258 * Ugly calling convention alert: 1259 * Return value with the low bits set means error value, 1260 * ie 1261 * if (ret & ~PAGE_MASK) 1262 * error = ret; 1263 * 1264 * This function "knows" that -ENOMEM has the bits set. 1265 */ 1266#ifndef HAVE_ARCH_UNMAPPED_AREA 1267unsigned long 1268arch_get_unmapped_area(struct file *filp, unsigned long addr, 1269 unsigned long len, unsigned long pgoff, unsigned long flags) 1270{ 1271 struct mm_struct *mm = current->mm; 1272 struct vm_area_struct *vma; 1273 unsigned long start_addr; 1274 1275 if (len > TASK_SIZE) 1276 return -ENOMEM; 1277 1278 if (flags & MAP_FIXED) 1279 return addr; 1280 1281 if (addr) { 1282 addr = PAGE_ALIGN(addr); 1283 vma = find_vma(mm, addr); 1284 if (TASK_SIZE - len >= addr && 1285 (!vma || addr + len <= vma->vm_start)) 1286 return addr; 1287 } 1288 if (len > mm->cached_hole_size) { 1289 start_addr = addr = mm->free_area_cache; 1290 } else { 1291 start_addr = addr = TASK_UNMAPPED_BASE; 1292 mm->cached_hole_size = 0; 1293 } 1294 1295full_search: 1296 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1297 /* At this point: (!vma || addr < vma->vm_end). */ 1298 if (TASK_SIZE - len < addr) { 1299 /* 1300 * Start a new search - just in case we missed 1301 * some holes. 1302 */ 1303 if (start_addr != TASK_UNMAPPED_BASE) { 1304 addr = TASK_UNMAPPED_BASE; 1305 start_addr = addr; 1306 mm->cached_hole_size = 0; 1307 goto full_search; 1308 } 1309 return -ENOMEM; 1310 } 1311 if (!vma || addr + len <= vma->vm_start) { 1312 /* 1313 * Remember the place where we stopped the search: 1314 */ 1315 mm->free_area_cache = addr + len; 1316 return addr; 1317 } 1318 if (addr + mm->cached_hole_size < vma->vm_start) 1319 mm->cached_hole_size = vma->vm_start - addr; 1320 addr = vma->vm_end; 1321 } 1322} 1323#endif 1324 1325void arch_unmap_area(struct mm_struct *mm, unsigned long addr) 1326{ 1327 /* 1328 * Is this a new hole at the lowest possible address? 1329 */ 1330 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) { 1331 mm->free_area_cache = addr; 1332 mm->cached_hole_size = ~0UL; 1333 } 1334} 1335 1336/* 1337 * This mmap-allocator allocates new areas top-down from below the 1338 * stack's low limit (the base): 1339 */ 1340#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1341unsigned long 1342arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1343 const unsigned long len, const unsigned long pgoff, 1344 const unsigned long flags) 1345{ 1346 struct vm_area_struct *vma; 1347 struct mm_struct *mm = current->mm; 1348 unsigned long addr = addr0; 1349 1350 /* requested length too big for entire address space */ 1351 if (len > TASK_SIZE) 1352 return -ENOMEM; 1353 1354 if (flags & MAP_FIXED) 1355 return addr; 1356 1357 /* requesting a specific address */ 1358 if (addr) { 1359 addr = PAGE_ALIGN(addr); 1360 vma = find_vma(mm, addr); 1361 if (TASK_SIZE - len >= addr && 1362 (!vma || addr + len <= vma->vm_start)) 1363 return addr; 1364 } 1365 1366 /* check if free_area_cache is useful for us */ 1367 if (len <= mm->cached_hole_size) { 1368 mm->cached_hole_size = 0; 1369 mm->free_area_cache = mm->mmap_base; 1370 } 1371 1372 /* either no address requested or can't fit in requested address hole */ 1373 addr = mm->free_area_cache; 1374 1375 /* make sure it can fit in the remaining address space */ 1376 if (addr > len) { 1377 vma = find_vma(mm, addr-len); 1378 if (!vma || addr <= vma->vm_start) 1379 /* remember the address as a hint for next time */ 1380 return (mm->free_area_cache = addr-len); 1381 } 1382 1383 if (mm->mmap_base < len) 1384 goto bottomup; 1385 1386 addr = mm->mmap_base-len; 1387 1388 do { 1389 /* 1390 * Lookup failure means no vma is above this address, 1391 * else if new region fits below vma->vm_start, 1392 * return with success: 1393 */ 1394 vma = find_vma(mm, addr); 1395 if (!vma || addr+len <= vma->vm_start) 1396 /* remember the address as a hint for next time */ 1397 return (mm->free_area_cache = addr); 1398 1399 /* remember the largest hole we saw so far */ 1400 if (addr + mm->cached_hole_size < vma->vm_start) 1401 mm->cached_hole_size = vma->vm_start - addr; 1402 1403 /* try just below the current vma->vm_start */ 1404 addr = vma->vm_start-len; 1405 } while (len < vma->vm_start); 1406 1407bottomup: 1408 /* 1409 * A failed mmap() very likely causes application failure, 1410 * so fall back to the bottom-up function here. This scenario 1411 * can happen with large stack limits and large mmap() 1412 * allocations. 1413 */ 1414 mm->cached_hole_size = ~0UL; 1415 mm->free_area_cache = TASK_UNMAPPED_BASE; 1416 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags); 1417 /* 1418 * Restore the topdown base: 1419 */ 1420 mm->free_area_cache = mm->mmap_base; 1421 mm->cached_hole_size = ~0UL; 1422 1423 return addr; 1424} 1425#endif 1426 1427void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr) 1428{ 1429 /* 1430 * Is this a new hole at the highest possible address? 1431 */ 1432 if (addr > mm->free_area_cache) 1433 mm->free_area_cache = addr; 1434 1435 /* dont allow allocations above current base */ 1436 if (mm->free_area_cache > mm->mmap_base) 1437 mm->free_area_cache = mm->mmap_base; 1438} 1439 1440unsigned long 1441get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1442 unsigned long pgoff, unsigned long flags) 1443{ 1444 unsigned long (*get_area)(struct file *, unsigned long, 1445 unsigned long, unsigned long, unsigned long); 1446 1447 get_area = current->mm->get_unmapped_area; 1448 if (file && file->f_op && file->f_op->get_unmapped_area) 1449 get_area = file->f_op->get_unmapped_area; 1450 addr = get_area(file, addr, len, pgoff, flags); 1451 if (IS_ERR_VALUE(addr)) 1452 return addr; 1453 1454 if (addr > TASK_SIZE - len) 1455 return -ENOMEM; 1456 if (addr & ~PAGE_MASK) 1457 return -EINVAL; 1458 1459 return arch_rebalance_pgtables(addr, len); 1460} 1461 1462EXPORT_SYMBOL(get_unmapped_area); 1463 1464/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 1465struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 1466{ 1467 struct vm_area_struct *vma = NULL; 1468 1469 if (mm) { 1470 /* Check the cache first. */ 1471 /* (Cache hit rate is typically around 35%.) */ 1472 vma = mm->mmap_cache; 1473 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) { 1474 struct rb_node * rb_node; 1475 1476 rb_node = mm->mm_rb.rb_node; 1477 vma = NULL; 1478 1479 while (rb_node) { 1480 struct vm_area_struct * vma_tmp; 1481 1482 vma_tmp = rb_entry(rb_node, 1483 struct vm_area_struct, vm_rb); 1484 1485 if (vma_tmp->vm_end > addr) { 1486 vma = vma_tmp; 1487 if (vma_tmp->vm_start <= addr) 1488 break; 1489 rb_node = rb_node->rb_left; 1490 } else 1491 rb_node = rb_node->rb_right; 1492 } 1493 if (vma) 1494 mm->mmap_cache = vma; 1495 } 1496 } 1497 return vma; 1498} 1499 1500EXPORT_SYMBOL(find_vma); 1501 1502/* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */ 1503struct vm_area_struct * 1504find_vma_prev(struct mm_struct *mm, unsigned long addr, 1505 struct vm_area_struct **pprev) 1506{ 1507 struct vm_area_struct *vma = NULL, *prev = NULL; 1508 struct rb_node *rb_node; 1509 if (!mm) 1510 goto out; 1511 1512 /* Guard against addr being lower than the first VMA */ 1513 vma = mm->mmap; 1514 1515 /* Go through the RB tree quickly. */ 1516 rb_node = mm->mm_rb.rb_node; 1517 1518 while (rb_node) { 1519 struct vm_area_struct *vma_tmp; 1520 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 1521 1522 if (addr < vma_tmp->vm_end) { 1523 rb_node = rb_node->rb_left; 1524 } else { 1525 prev = vma_tmp; 1526 if (!prev->vm_next || (addr < prev->vm_next->vm_end)) 1527 break; 1528 rb_node = rb_node->rb_right; 1529 } 1530 } 1531 1532out: 1533 *pprev = prev; 1534 return prev ? prev->vm_next : vma; 1535} 1536 1537/* 1538 * Verify that the stack growth is acceptable and 1539 * update accounting. This is shared with both the 1540 * grow-up and grow-down cases. 1541 */ 1542static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 1543{ 1544 struct mm_struct *mm = vma->vm_mm; 1545 struct rlimit *rlim = current->signal->rlim; 1546 unsigned long new_start; 1547 1548 /* address space limit tests */ 1549 if (!may_expand_vm(mm, grow)) 1550 return -ENOMEM; 1551 1552 /* Stack limit test */ 1553 if (size > rlim[RLIMIT_STACK].rlim_cur) 1554 return -ENOMEM; 1555 1556 /* mlock limit tests */ 1557 if (vma->vm_flags & VM_LOCKED) { 1558 unsigned long locked; 1559 unsigned long limit; 1560 locked = mm->locked_vm + grow; 1561 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT; 1562 if (locked > limit && !capable(CAP_IPC_LOCK)) 1563 return -ENOMEM; 1564 } 1565 1566 /* Check to ensure the stack will not grow into a hugetlb-only region */ 1567 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 1568 vma->vm_end - size; 1569 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 1570 return -EFAULT; 1571 1572 /* 1573 * Overcommit.. This must be the final test, as it will 1574 * update security statistics. 1575 */ 1576 if (security_vm_enough_memory_mm(mm, grow)) 1577 return -ENOMEM; 1578 1579 /* Ok, everything looks good - let it rip */ 1580 mm->total_vm += grow; 1581 if (vma->vm_flags & VM_LOCKED) 1582 mm->locked_vm += grow; 1583 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow); 1584 return 0; 1585} 1586 1587#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 1588/* 1589 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 1590 * vma is the last one with address > vma->vm_end. Have to extend vma. 1591 */ 1592#ifndef CONFIG_IA64 1593static 1594#endif 1595int expand_upwards(struct vm_area_struct *vma, unsigned long address) 1596{ 1597 int error; 1598 1599 if (!(vma->vm_flags & VM_GROWSUP)) 1600 return -EFAULT; 1601 1602 /* 1603 * We must make sure the anon_vma is allocated 1604 * so that the anon_vma locking is not a noop. 1605 */ 1606 if (unlikely(anon_vma_prepare(vma))) 1607 return -ENOMEM; 1608 anon_vma_lock(vma); 1609 1610 /* 1611 * vma->vm_start/vm_end cannot change under us because the caller 1612 * is required to hold the mmap_sem in read mode. We need the 1613 * anon_vma lock to serialize against concurrent expand_stacks. 1614 * Also guard against wrapping around to address 0. 1615 */ 1616 if (address < PAGE_ALIGN(address+4)) 1617 address = PAGE_ALIGN(address+4); 1618 else { 1619 anon_vma_unlock(vma); 1620 return -ENOMEM; 1621 } 1622 error = 0; 1623 1624 /* Somebody else might have raced and expanded it already */ 1625 if (address > vma->vm_end) { 1626 unsigned long size, grow; 1627 1628 size = address - vma->vm_start; 1629 grow = (address - vma->vm_end) >> PAGE_SHIFT; 1630 1631 error = acct_stack_growth(vma, size, grow); 1632 if (!error) 1633 vma->vm_end = address; 1634 } 1635 anon_vma_unlock(vma); 1636 return error; 1637} 1638#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 1639 1640/* 1641 * vma is the first one with address < vma->vm_start. Have to extend vma. 1642 */ 1643static int expand_downwards(struct vm_area_struct *vma, 1644 unsigned long address) 1645{ 1646 int error; 1647 1648 /* 1649 * We must make sure the anon_vma is allocated 1650 * so that the anon_vma locking is not a noop. 1651 */ 1652 if (unlikely(anon_vma_prepare(vma))) 1653 return -ENOMEM; 1654 1655 address &= PAGE_MASK; 1656 error = security_file_mmap(NULL, 0, 0, 0, address, 1); 1657 if (error) 1658 return error; 1659 1660 anon_vma_lock(vma); 1661 1662 /* 1663 * vma->vm_start/vm_end cannot change under us because the caller 1664 * is required to hold the mmap_sem in read mode. We need the 1665 * anon_vma lock to serialize against concurrent expand_stacks. 1666 */ 1667 1668 /* Somebody else might have raced and expanded it already */ 1669 if (address < vma->vm_start) { 1670 unsigned long size, grow; 1671 1672 size = vma->vm_end - address; 1673 grow = (vma->vm_start - address) >> PAGE_SHIFT; 1674 1675 error = acct_stack_growth(vma, size, grow); 1676 if (!error) { 1677 vma->vm_start = address; 1678 vma->vm_pgoff -= grow; 1679 } 1680 } 1681 anon_vma_unlock(vma); 1682 return error; 1683} 1684 1685int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address) 1686{ 1687 return expand_downwards(vma, address); 1688} 1689 1690#ifdef CONFIG_STACK_GROWSUP 1691int expand_stack(struct vm_area_struct *vma, unsigned long address) 1692{ 1693 return expand_upwards(vma, address); 1694} 1695 1696struct vm_area_struct * 1697find_extend_vma(struct mm_struct *mm, unsigned long addr) 1698{ 1699 struct vm_area_struct *vma, *prev; 1700 1701 addr &= PAGE_MASK; 1702 vma = find_vma_prev(mm, addr, &prev); 1703 if (vma && (vma->vm_start <= addr)) 1704 return vma; 1705 if (!prev || expand_stack(prev, addr)) 1706 return NULL; 1707 if (prev->vm_flags & VM_LOCKED) { 1708 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0) 1709 return NULL; /* vma gone! */ 1710 } 1711 return prev; 1712} 1713#else 1714int expand_stack(struct vm_area_struct *vma, unsigned long address) 1715{ 1716 return expand_downwards(vma, address); 1717} 1718 1719struct vm_area_struct * 1720find_extend_vma(struct mm_struct * mm, unsigned long addr) 1721{ 1722 struct vm_area_struct * vma; 1723 unsigned long start; 1724 1725 addr &= PAGE_MASK; 1726 vma = find_vma(mm,addr); 1727 if (!vma) 1728 return NULL; 1729 if (vma->vm_start <= addr) 1730 return vma; 1731 if (!(vma->vm_flags & VM_GROWSDOWN)) 1732 return NULL; 1733 start = vma->vm_start; 1734 if (expand_stack(vma, addr)) 1735 return NULL; 1736 if (vma->vm_flags & VM_LOCKED) { 1737 if (mlock_vma_pages_range(vma, addr, start) < 0) 1738 return NULL; /* vma gone! */ 1739 } 1740 return vma; 1741} 1742#endif 1743 1744/* 1745 * Ok - we have the memory areas we should free on the vma list, 1746 * so release them, and do the vma updates. 1747 * 1748 * Called with the mm semaphore held. 1749 */ 1750static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 1751{ 1752 /* Update high watermark before we lower total_vm */ 1753 update_hiwater_vm(mm); 1754 do { 1755 long nrpages = vma_pages(vma); 1756 1757 mm->total_vm -= nrpages; 1758 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages); 1759 vma = remove_vma(vma); 1760 } while (vma); 1761 validate_mm(mm); 1762} 1763 1764/* 1765 * Get rid of page table information in the indicated region. 1766 * 1767 * Called with the mm semaphore held. 1768 */ 1769static void unmap_region(struct mm_struct *mm, 1770 struct vm_area_struct *vma, struct vm_area_struct *prev, 1771 unsigned long start, unsigned long end) 1772{ 1773 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap; 1774 struct mmu_gather *tlb; 1775 unsigned long nr_accounted = 0; 1776 1777 lru_add_drain(); 1778 tlb = tlb_gather_mmu(mm, 0); 1779 update_hiwater_rss(mm); 1780 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL); 1781 vm_unacct_memory(nr_accounted); 1782 free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS, 1783 next? next->vm_start: 0); 1784 tlb_finish_mmu(tlb, start, end); 1785} 1786 1787/* 1788 * Create a list of vma's touched by the unmap, removing them from the mm's 1789 * vma list as we go.. 1790 */ 1791static void 1792detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 1793 struct vm_area_struct *prev, unsigned long end) 1794{ 1795 struct vm_area_struct **insertion_point; 1796 struct vm_area_struct *tail_vma = NULL; 1797 unsigned long addr; 1798 1799 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 1800 do { 1801 rb_erase(&vma->vm_rb, &mm->mm_rb); 1802 mm->map_count--; 1803 tail_vma = vma; 1804 vma = vma->vm_next; 1805 } while (vma && vma->vm_start < end); 1806 *insertion_point = vma; 1807 tail_vma->vm_next = NULL; 1808 if (mm->unmap_area == arch_unmap_area) 1809 addr = prev ? prev->vm_end : mm->mmap_base; 1810 else 1811 addr = vma ? vma->vm_start : mm->mmap_base; 1812 mm->unmap_area(mm, addr); 1813 mm->mmap_cache = NULL; /* Kill the cache. */ 1814} 1815 1816/* 1817 * Split a vma into two pieces at address 'addr', a new vma is allocated 1818 * either for the first part or the tail. 1819 */ 1820int split_vma(struct mm_struct * mm, struct vm_area_struct * vma, 1821 unsigned long addr, int new_below) 1822{ 1823 struct mempolicy *pol; 1824 struct vm_area_struct *new; 1825 1826 if (is_vm_hugetlb_page(vma) && (addr & 1827 ~(huge_page_mask(hstate_vma(vma))))) 1828 return -EINVAL; 1829 1830 if (mm->map_count >= sysctl_max_map_count) 1831 return -ENOMEM; 1832 1833 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 1834 if (!new) 1835 return -ENOMEM; 1836 1837 /* most fields are the same, copy all, and then fixup */ 1838 *new = *vma; 1839 1840 if (new_below) 1841 new->vm_end = addr; 1842 else { 1843 new->vm_start = addr; 1844 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 1845 } 1846 1847 pol = mpol_dup(vma_policy(vma)); 1848 if (IS_ERR(pol)) { 1849 kmem_cache_free(vm_area_cachep, new); 1850 return PTR_ERR(pol); 1851 } 1852 vma_set_policy(new, pol); 1853 1854 if (new->vm_file) { 1855 get_file(new->vm_file); 1856 if (vma->vm_flags & VM_EXECUTABLE) 1857 added_exe_file_vma(mm); 1858 } 1859 1860 if (new->vm_ops && new->vm_ops->open) 1861 new->vm_ops->open(new); 1862 1863 if (new_below) 1864 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 1865 ((addr - new->vm_start) >> PAGE_SHIFT), new); 1866 else 1867 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 1868 1869 return 0; 1870} 1871 1872/* Munmap is split into 2 main parts -- this part which finds 1873 * what needs doing, and the areas themselves, which do the 1874 * work. This now handles partial unmappings. 1875 * Jeremy Fitzhardinge <jeremy@goop.org> 1876 */ 1877int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 1878{ 1879 unsigned long end; 1880 struct vm_area_struct *vma, *prev, *last; 1881 1882 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start) 1883 return -EINVAL; 1884 1885 if ((len = PAGE_ALIGN(len)) == 0) 1886 return -EINVAL; 1887 1888 /* Find the first overlapping VMA */ 1889 vma = find_vma_prev(mm, start, &prev); 1890 if (!vma) 1891 return 0; 1892 /* we have start < vma->vm_end */ 1893 1894 /* if it doesn't overlap, we have nothing.. */ 1895 end = start + len; 1896 if (vma->vm_start >= end) 1897 return 0; 1898 1899 /* 1900 * If we need to split any vma, do it now to save pain later. 1901 * 1902 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 1903 * unmapped vm_area_struct will remain in use: so lower split_vma 1904 * places tmp vma above, and higher split_vma places tmp vma below. 1905 */ 1906 if (start > vma->vm_start) { 1907 int error = split_vma(mm, vma, start, 0); 1908 if (error) 1909 return error; 1910 prev = vma; 1911 } 1912 1913 /* Does it split the last one? */ 1914 last = find_vma(mm, end); 1915 if (last && end > last->vm_start) { 1916 int error = split_vma(mm, last, end, 1); 1917 if (error) 1918 return error; 1919 } 1920 vma = prev? prev->vm_next: mm->mmap; 1921 1922 /* 1923 * unlock any mlock()ed ranges before detaching vmas 1924 */ 1925 if (mm->locked_vm) { 1926 struct vm_area_struct *tmp = vma; 1927 while (tmp && tmp->vm_start < end) { 1928 if (tmp->vm_flags & VM_LOCKED) { 1929 mm->locked_vm -= vma_pages(tmp); 1930 munlock_vma_pages_all(tmp); 1931 } 1932 tmp = tmp->vm_next; 1933 } 1934 } 1935 1936 /* 1937 * Remove the vma's, and unmap the actual pages 1938 */ 1939 detach_vmas_to_be_unmapped(mm, vma, prev, end); 1940 unmap_region(mm, vma, prev, start, end); 1941 1942 /* Fix up all other VM information */ 1943 remove_vma_list(mm, vma); 1944 1945 return 0; 1946} 1947 1948EXPORT_SYMBOL(do_munmap); 1949 1950SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 1951{ 1952 int ret; 1953 struct mm_struct *mm = current->mm; 1954 1955 profile_munmap(addr); 1956 1957 down_write(&mm->mmap_sem); 1958 ret = do_munmap(mm, addr, len); 1959 up_write(&mm->mmap_sem); 1960 return ret; 1961} 1962 1963static inline void verify_mm_writelocked(struct mm_struct *mm) 1964{ 1965#ifdef CONFIG_DEBUG_VM 1966 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 1967 WARN_ON(1); 1968 up_read(&mm->mmap_sem); 1969 } 1970#endif 1971} 1972 1973/* 1974 * this is really a simplified "do_mmap". it only handles 1975 * anonymous maps. eventually we may be able to do some 1976 * brk-specific accounting here. 1977 */ 1978unsigned long do_brk(unsigned long addr, unsigned long len) 1979{ 1980 struct mm_struct * mm = current->mm; 1981 struct vm_area_struct * vma, * prev; 1982 unsigned long flags; 1983 struct rb_node ** rb_link, * rb_parent; 1984 pgoff_t pgoff = addr >> PAGE_SHIFT; 1985 int error; 1986 1987 len = PAGE_ALIGN(len); 1988 if (!len) 1989 return addr; 1990 1991 if ((addr + len) > TASK_SIZE || (addr + len) < addr) 1992 return -EINVAL; 1993 1994 if (is_hugepage_only_range(mm, addr, len)) 1995 return -EINVAL; 1996 1997 error = security_file_mmap(NULL, 0, 0, 0, addr, 1); 1998 if (error) 1999 return error; 2000 2001 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2002 2003 error = arch_mmap_check(addr, len, flags); 2004 if (error) 2005 return error; 2006 2007 /* 2008 * mlock MCL_FUTURE? 2009 */ 2010 if (mm->def_flags & VM_LOCKED) { 2011 unsigned long locked, lock_limit; 2012 locked = len >> PAGE_SHIFT; 2013 locked += mm->locked_vm; 2014 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur; 2015 lock_limit >>= PAGE_SHIFT; 2016 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 2017 return -EAGAIN; 2018 } 2019 2020 /* 2021 * mm->mmap_sem is required to protect against another thread 2022 * changing the mappings in case we sleep. 2023 */ 2024 verify_mm_writelocked(mm); 2025 2026 /* 2027 * Clear old maps. this also does some error checking for us 2028 */ 2029 munmap_back: 2030 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 2031 if (vma && vma->vm_start < addr + len) { 2032 if (do_munmap(mm, addr, len)) 2033 return -ENOMEM; 2034 goto munmap_back; 2035 } 2036 2037 /* Check against address space limits *after* clearing old maps... */ 2038 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 2039 return -ENOMEM; 2040 2041 if (mm->map_count > sysctl_max_map_count) 2042 return -ENOMEM; 2043 2044 if (security_vm_enough_memory(len >> PAGE_SHIFT)) 2045 return -ENOMEM; 2046 2047 /* Can we just expand an old private anonymous mapping? */ 2048 vma = vma_merge(mm, prev, addr, addr + len, flags, 2049 NULL, NULL, pgoff, NULL); 2050 if (vma) 2051 goto out; 2052 2053 /* 2054 * create a vma struct for an anonymous mapping 2055 */ 2056 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2057 if (!vma) { 2058 vm_unacct_memory(len >> PAGE_SHIFT); 2059 return -ENOMEM; 2060 } 2061 2062 vma->vm_mm = mm; 2063 vma->vm_start = addr; 2064 vma->vm_end = addr + len; 2065 vma->vm_pgoff = pgoff; 2066 vma->vm_flags = flags; 2067 vma->vm_page_prot = vm_get_page_prot(flags); 2068 vma_link(mm, vma, prev, rb_link, rb_parent); 2069out: 2070 mm->total_vm += len >> PAGE_SHIFT; 2071 if (flags & VM_LOCKED) { 2072 if (!mlock_vma_pages_range(vma, addr, addr + len)) 2073 mm->locked_vm += (len >> PAGE_SHIFT); 2074 } 2075 return addr; 2076} 2077 2078EXPORT_SYMBOL(do_brk); 2079 2080/* Release all mmaps. */ 2081void exit_mmap(struct mm_struct *mm) 2082{ 2083 struct mmu_gather *tlb; 2084 struct vm_area_struct *vma; 2085 unsigned long nr_accounted = 0; 2086 unsigned long end; 2087 2088 /* mm's last user has gone, and its about to be pulled down */ 2089 mmu_notifier_release(mm); 2090 2091 if (mm->locked_vm) { 2092 vma = mm->mmap; 2093 while (vma) { 2094 if (vma->vm_flags & VM_LOCKED) 2095 munlock_vma_pages_all(vma); 2096 vma = vma->vm_next; 2097 } 2098 } 2099 2100 arch_exit_mmap(mm); 2101 2102 vma = mm->mmap; 2103 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2104 return; 2105 2106 lru_add_drain(); 2107 flush_cache_mm(mm); 2108 tlb = tlb_gather_mmu(mm, 1); 2109 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2110 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2111 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL); 2112 vm_unacct_memory(nr_accounted); 2113 2114 free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0); 2115 tlb_finish_mmu(tlb, 0, end); 2116 2117 /* 2118 * Walk the list again, actually closing and freeing it, 2119 * with preemption enabled, without holding any MM locks. 2120 */ 2121 while (vma) 2122 vma = remove_vma(vma); 2123 2124 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT); 2125} 2126 2127/* Insert vm structure into process list sorted by address 2128 * and into the inode's i_mmap tree. If vm_file is non-NULL 2129 * then i_mmap_lock is taken here. 2130 */ 2131int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma) 2132{ 2133 struct vm_area_struct * __vma, * prev; 2134 struct rb_node ** rb_link, * rb_parent; 2135 2136 /* 2137 * The vm_pgoff of a purely anonymous vma should be irrelevant 2138 * until its first write fault, when page's anon_vma and index 2139 * are set. But now set the vm_pgoff it will almost certainly 2140 * end up with (unless mremap moves it elsewhere before that 2141 * first wfault), so /proc/pid/maps tells a consistent story. 2142 * 2143 * By setting it to reflect the virtual start address of the 2144 * vma, merges and splits can happen in a seamless way, just 2145 * using the existing file pgoff checks and manipulations. 2146 * Similarly in do_mmap_pgoff and in do_brk. 2147 */ 2148 if (!vma->vm_file) { 2149 BUG_ON(vma->anon_vma); 2150 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2151 } 2152 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent); 2153 if (__vma && __vma->vm_start < vma->vm_end) 2154 return -ENOMEM; 2155 if ((vma->vm_flags & VM_ACCOUNT) && 2156 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2157 return -ENOMEM; 2158 vma_link(mm, vma, prev, rb_link, rb_parent); 2159 return 0; 2160} 2161 2162/* 2163 * Copy the vma structure to a new location in the same mm, 2164 * prior to moving page table entries, to effect an mremap move. 2165 */ 2166struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2167 unsigned long addr, unsigned long len, pgoff_t pgoff) 2168{ 2169 struct vm_area_struct *vma = *vmap; 2170 unsigned long vma_start = vma->vm_start; 2171 struct mm_struct *mm = vma->vm_mm; 2172 struct vm_area_struct *new_vma, *prev; 2173 struct rb_node **rb_link, *rb_parent; 2174 struct mempolicy *pol; 2175 2176 /* 2177 * If anonymous vma has not yet been faulted, update new pgoff 2178 * to match new location, to increase its chance of merging. 2179 */ 2180 if (!vma->vm_file && !vma->anon_vma) 2181 pgoff = addr >> PAGE_SHIFT; 2182 2183 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent); 2184 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2185 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma)); 2186 if (new_vma) { 2187 /* 2188 * Source vma may have been merged into new_vma 2189 */ 2190 if (vma_start >= new_vma->vm_start && 2191 vma_start < new_vma->vm_end) 2192 *vmap = new_vma; 2193 } else { 2194 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2195 if (new_vma) { 2196 *new_vma = *vma; 2197 pol = mpol_dup(vma_policy(vma)); 2198 if (IS_ERR(pol)) { 2199 kmem_cache_free(vm_area_cachep, new_vma); 2200 return NULL; 2201 } 2202 vma_set_policy(new_vma, pol); 2203 new_vma->vm_start = addr; 2204 new_vma->vm_end = addr + len; 2205 new_vma->vm_pgoff = pgoff; 2206 if (new_vma->vm_file) { 2207 get_file(new_vma->vm_file); 2208 if (vma->vm_flags & VM_EXECUTABLE) 2209 added_exe_file_vma(mm); 2210 } 2211 if (new_vma->vm_ops && new_vma->vm_ops->open) 2212 new_vma->vm_ops->open(new_vma); 2213 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2214 } 2215 } 2216 return new_vma; 2217} 2218 2219/* 2220 * Return true if the calling process may expand its vm space by the passed 2221 * number of pages 2222 */ 2223int may_expand_vm(struct mm_struct *mm, unsigned long npages) 2224{ 2225 unsigned long cur = mm->total_vm; /* pages */ 2226 unsigned long lim; 2227 2228 lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT; 2229 2230 if (cur + npages > lim) 2231 return 0; 2232 return 1; 2233} 2234 2235 2236static int special_mapping_fault(struct vm_area_struct *vma, 2237 struct vm_fault *vmf) 2238{ 2239 pgoff_t pgoff; 2240 struct page **pages; 2241 2242 /* 2243 * special mappings have no vm_file, and in that case, the mm 2244 * uses vm_pgoff internally. So we have to subtract it from here. 2245 * We are allowed to do this because we are the mm; do not copy 2246 * this code into drivers! 2247 */ 2248 pgoff = vmf->pgoff - vma->vm_pgoff; 2249 2250 for (pages = vma->vm_private_data; pgoff && *pages; ++pages) 2251 pgoff--; 2252 2253 if (*pages) { 2254 struct page *page = *pages; 2255 get_page(page); 2256 vmf->page = page; 2257 return 0; 2258 } 2259 2260 return VM_FAULT_SIGBUS; 2261} 2262 2263/* 2264 * Having a close hook prevents vma merging regardless of flags. 2265 */ 2266static void special_mapping_close(struct vm_area_struct *vma) 2267{ 2268} 2269 2270static struct vm_operations_struct special_mapping_vmops = { 2271 .close = special_mapping_close, 2272 .fault = special_mapping_fault, 2273}; 2274 2275/* 2276 * Called with mm->mmap_sem held for writing. 2277 * Insert a new vma covering the given region, with the given flags. 2278 * Its pages are supplied by the given array of struct page *. 2279 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 2280 * The region past the last page supplied will always produce SIGBUS. 2281 * The array pointer and the pages it points to are assumed to stay alive 2282 * for as long as this mapping might exist. 2283 */ 2284int install_special_mapping(struct mm_struct *mm, 2285 unsigned long addr, unsigned long len, 2286 unsigned long vm_flags, struct page **pages) 2287{ 2288 struct vm_area_struct *vma; 2289 2290 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2291 if (unlikely(vma == NULL)) 2292 return -ENOMEM; 2293 2294 vma->vm_mm = mm; 2295 vma->vm_start = addr; 2296 vma->vm_end = addr + len; 2297 2298 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND; 2299 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 2300 2301 vma->vm_ops = &special_mapping_vmops; 2302 vma->vm_private_data = pages; 2303 2304 if (unlikely(insert_vm_struct(mm, vma))) { 2305 kmem_cache_free(vm_area_cachep, vma); 2306 return -ENOMEM; 2307 } 2308 2309 mm->total_vm += len >> PAGE_SHIFT; 2310 2311 perf_event_mmap(vma); 2312 2313 return 0; 2314} 2315 2316static DEFINE_MUTEX(mm_all_locks_mutex); 2317 2318static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 2319{ 2320 if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) { 2321 /* 2322 * The LSB of head.next can't change from under us 2323 * because we hold the mm_all_locks_mutex. 2324 */ 2325 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem); 2326 /* 2327 * We can safely modify head.next after taking the 2328 * anon_vma->lock. If some other vma in this mm shares 2329 * the same anon_vma we won't take it again. 2330 * 2331 * No need of atomic instructions here, head.next 2332 * can't change from under us thanks to the 2333 * anon_vma->lock. 2334 */ 2335 if (__test_and_set_bit(0, (unsigned long *) 2336 &anon_vma->head.next)) 2337 BUG(); 2338 } 2339} 2340 2341static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 2342{ 2343 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 2344 /* 2345 * AS_MM_ALL_LOCKS can't change from under us because 2346 * we hold the mm_all_locks_mutex. 2347 * 2348 * Operations on ->flags have to be atomic because 2349 * even if AS_MM_ALL_LOCKS is stable thanks to the 2350 * mm_all_locks_mutex, there may be other cpus 2351 * changing other bitflags in parallel to us. 2352 */ 2353 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 2354 BUG(); 2355 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem); 2356 } 2357} 2358 2359/* 2360 * This operation locks against the VM for all pte/vma/mm related 2361 * operations that could ever happen on a certain mm. This includes 2362 * vmtruncate, try_to_unmap, and all page faults. 2363 * 2364 * The caller must take the mmap_sem in write mode before calling 2365 * mm_take_all_locks(). The caller isn't allowed to release the 2366 * mmap_sem until mm_drop_all_locks() returns. 2367 * 2368 * mmap_sem in write mode is required in order to block all operations 2369 * that could modify pagetables and free pages without need of 2370 * altering the vma layout (for example populate_range() with 2371 * nonlinear vmas). It's also needed in write mode to avoid new 2372 * anon_vmas to be associated with existing vmas. 2373 * 2374 * A single task can't take more than one mm_take_all_locks() in a row 2375 * or it would deadlock. 2376 * 2377 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in 2378 * mapping->flags avoid to take the same lock twice, if more than one 2379 * vma in this mm is backed by the same anon_vma or address_space. 2380 * 2381 * We can take all the locks in random order because the VM code 2382 * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never 2383 * takes more than one of them in a row. Secondly we're protected 2384 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex. 2385 * 2386 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 2387 * that may have to take thousand of locks. 2388 * 2389 * mm_take_all_locks() can fail if it's interrupted by signals. 2390 */ 2391int mm_take_all_locks(struct mm_struct *mm) 2392{ 2393 struct vm_area_struct *vma; 2394 int ret = -EINTR; 2395 2396 BUG_ON(down_read_trylock(&mm->mmap_sem)); 2397 2398 mutex_lock(&mm_all_locks_mutex); 2399 2400 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2401 if (signal_pending(current)) 2402 goto out_unlock; 2403 if (vma->vm_file && vma->vm_file->f_mapping) 2404 vm_lock_mapping(mm, vma->vm_file->f_mapping); 2405 } 2406 2407 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2408 if (signal_pending(current)) 2409 goto out_unlock; 2410 if (vma->anon_vma) 2411 vm_lock_anon_vma(mm, vma->anon_vma); 2412 } 2413 2414 ret = 0; 2415 2416out_unlock: 2417 if (ret) 2418 mm_drop_all_locks(mm); 2419 2420 return ret; 2421} 2422 2423static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 2424{ 2425 if (test_bit(0, (unsigned long *) &anon_vma->head.next)) { 2426 /* 2427 * The LSB of head.next can't change to 0 from under 2428 * us because we hold the mm_all_locks_mutex. 2429 * 2430 * We must however clear the bitflag before unlocking 2431 * the vma so the users using the anon_vma->head will 2432 * never see our bitflag. 2433 * 2434 * No need of atomic instructions here, head.next 2435 * can't change from under us until we release the 2436 * anon_vma->lock. 2437 */ 2438 if (!__test_and_clear_bit(0, (unsigned long *) 2439 &anon_vma->head.next)) 2440 BUG(); 2441 spin_unlock(&anon_vma->lock); 2442 } 2443} 2444 2445static void vm_unlock_mapping(struct address_space *mapping) 2446{ 2447 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 2448 /* 2449 * AS_MM_ALL_LOCKS can't change to 0 from under us 2450 * because we hold the mm_all_locks_mutex. 2451 */ 2452 spin_unlock(&mapping->i_mmap_lock); 2453 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 2454 &mapping->flags)) 2455 BUG(); 2456 } 2457} 2458 2459/* 2460 * The mmap_sem cannot be released by the caller until 2461 * mm_drop_all_locks() returns. 2462 */ 2463void mm_drop_all_locks(struct mm_struct *mm) 2464{ 2465 struct vm_area_struct *vma; 2466 2467 BUG_ON(down_read_trylock(&mm->mmap_sem)); 2468 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 2469 2470 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2471 if (vma->anon_vma) 2472 vm_unlock_anon_vma(vma->anon_vma); 2473 if (vma->vm_file && vma->vm_file->f_mapping) 2474 vm_unlock_mapping(vma->vm_file->f_mapping); 2475 } 2476 2477 mutex_unlock(&mm_all_locks_mutex); 2478} 2479 2480/* 2481 * initialise the VMA slab 2482 */ 2483void __init mmap_init(void) 2484{ 2485 int ret; 2486 2487 ret = percpu_counter_init(&vm_committed_as, 0); 2488 VM_BUG_ON(ret); 2489} 2490