1 /* 2 * mm/mmap.c 3 * 4 * Written by obz. 5 * 6 * Address space accounting code <alan@lxorguk.ukuu.org.uk> 7 */ 8 9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 10 11 #include <linux/kernel.h> 12 #include <linux/slab.h> 13 #include <linux/backing-dev.h> 14 #include <linux/mm.h> 15 #include <linux/vmacache.h> 16 #include <linux/shm.h> 17 #include <linux/mman.h> 18 #include <linux/pagemap.h> 19 #include <linux/swap.h> 20 #include <linux/syscalls.h> 21 #include <linux/capability.h> 22 #include <linux/init.h> 23 #include <linux/file.h> 24 #include <linux/fs.h> 25 #include <linux/personality.h> 26 #include <linux/security.h> 27 #include <linux/hugetlb.h> 28 #include <linux/profile.h> 29 #include <linux/export.h> 30 #include <linux/mount.h> 31 #include <linux/mempolicy.h> 32 #include <linux/rmap.h> 33 #include <linux/mmu_notifier.h> 34 #include <linux/perf_event.h> 35 #include <linux/audit.h> 36 #include <linux/khugepaged.h> 37 #include <linux/uprobes.h> 38 #include <linux/rbtree_augmented.h> 39 #include <linux/sched/sysctl.h> 40 #include <linux/notifier.h> 41 #include <linux/memory.h> 42 #include <linux/printk.h> 43 44 #include <asm/uaccess.h> 45 #include <asm/cacheflush.h> 46 #include <asm/tlb.h> 47 #include <asm/mmu_context.h> 48 49 #include "internal.h" 50 51 #ifndef arch_mmap_check 52 #define arch_mmap_check(addr, len, flags) (0) 53 #endif 54 55 #ifndef arch_rebalance_pgtables 56 #define arch_rebalance_pgtables(addr, len) (addr) 57 #endif 58 59 static void unmap_region(struct mm_struct *mm, 60 struct vm_area_struct *vma, struct vm_area_struct *prev, 61 unsigned long start, unsigned long end); 62 63 /* description of effects of mapping type and prot in current implementation. 64 * this is due to the limited x86 page protection hardware. The expected 65 * behavior is in parens: 66 * 67 * map_type prot 68 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 69 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 70 * w: (no) no w: (no) no w: (yes) yes w: (no) no 71 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 72 * 73 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 74 * w: (no) no w: (no) no w: (copy) copy w: (no) no 75 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 76 * 77 */ 78 pgprot_t protection_map[16] = { 79 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 80 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 81 }; 82 83 pgprot_t vm_get_page_prot(unsigned long vm_flags) 84 { 85 return __pgprot(pgprot_val(protection_map[vm_flags & 86 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) | 87 pgprot_val(arch_vm_get_page_prot(vm_flags))); 88 } 89 EXPORT_SYMBOL(vm_get_page_prot); 90 91 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */ 92 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */ 93 unsigned long sysctl_overcommit_kbytes __read_mostly; 94 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT; 95 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */ 96 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */ 97 /* 98 * Make sure vm_committed_as in one cacheline and not cacheline shared with 99 * other variables. It can be updated by several CPUs frequently. 100 */ 101 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp; 102 103 /* 104 * The global memory commitment made in the system can be a metric 105 * that can be used to drive ballooning decisions when Linux is hosted 106 * as a guest. On Hyper-V, the host implements a policy engine for dynamically 107 * balancing memory across competing virtual machines that are hosted. 108 * Several metrics drive this policy engine including the guest reported 109 * memory commitment. 110 */ 111 unsigned long vm_memory_committed(void) 112 { 113 return percpu_counter_read_positive(&vm_committed_as); 114 } 115 EXPORT_SYMBOL_GPL(vm_memory_committed); 116 117 /* 118 * Check that a process has enough memory to allocate a new virtual 119 * mapping. 0 means there is enough memory for the allocation to 120 * succeed and -ENOMEM implies there is not. 121 * 122 * We currently support three overcommit policies, which are set via the 123 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting 124 * 125 * Strict overcommit modes added 2002 Feb 26 by Alan Cox. 126 * Additional code 2002 Jul 20 by Robert Love. 127 * 128 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise. 129 * 130 * Note this is a helper function intended to be used by LSMs which 131 * wish to use this logic. 132 */ 133 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin) 134 { 135 unsigned long free, allowed, reserve; 136 137 vm_acct_memory(pages); 138 139 /* 140 * Sometimes we want to use more memory than we have 141 */ 142 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS) 143 return 0; 144 145 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) { 146 free = global_page_state(NR_FREE_PAGES); 147 free += global_page_state(NR_FILE_PAGES); 148 149 /* 150 * shmem pages shouldn't be counted as free in this 151 * case, they can't be purged, only swapped out, and 152 * that won't affect the overall amount of available 153 * memory in the system. 154 */ 155 free -= global_page_state(NR_SHMEM); 156 157 free += get_nr_swap_pages(); 158 159 /* 160 * Any slabs which are created with the 161 * SLAB_RECLAIM_ACCOUNT flag claim to have contents 162 * which are reclaimable, under pressure. The dentry 163 * cache and most inode caches should fall into this 164 */ 165 free += global_page_state(NR_SLAB_RECLAIMABLE); 166 167 /* 168 * Leave reserved pages. The pages are not for anonymous pages. 169 */ 170 if (free <= totalreserve_pages) 171 goto error; 172 else 173 free -= totalreserve_pages; 174 175 /* 176 * Reserve some for root 177 */ 178 if (!cap_sys_admin) 179 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10); 180 181 if (free > pages) 182 return 0; 183 184 goto error; 185 } 186 187 allowed = vm_commit_limit(); 188 /* 189 * Reserve some for root 190 */ 191 if (!cap_sys_admin) 192 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10); 193 194 /* 195 * Don't let a single process grow so big a user can't recover 196 */ 197 if (mm) { 198 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10); 199 allowed -= min(mm->total_vm / 32, reserve); 200 } 201 202 if (percpu_counter_read_positive(&vm_committed_as) < allowed) 203 return 0; 204 error: 205 vm_unacct_memory(pages); 206 207 return -ENOMEM; 208 } 209 210 /* 211 * Requires inode->i_mapping->i_mmap_mutex 212 */ 213 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 214 struct file *file, struct address_space *mapping) 215 { 216 if (vma->vm_flags & VM_DENYWRITE) 217 atomic_inc(&file_inode(file)->i_writecount); 218 if (vma->vm_flags & VM_SHARED) 219 mapping->i_mmap_writable--; 220 221 flush_dcache_mmap_lock(mapping); 222 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 223 list_del_init(&vma->shared.nonlinear); 224 else 225 vma_interval_tree_remove(vma, &mapping->i_mmap); 226 flush_dcache_mmap_unlock(mapping); 227 } 228 229 /* 230 * Unlink a file-based vm structure from its interval tree, to hide 231 * vma from rmap and vmtruncate before freeing its page tables. 232 */ 233 void unlink_file_vma(struct vm_area_struct *vma) 234 { 235 struct file *file = vma->vm_file; 236 237 if (file) { 238 struct address_space *mapping = file->f_mapping; 239 mutex_lock(&mapping->i_mmap_mutex); 240 __remove_shared_vm_struct(vma, file, mapping); 241 mutex_unlock(&mapping->i_mmap_mutex); 242 } 243 } 244 245 /* 246 * Close a vm structure and free it, returning the next. 247 */ 248 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 249 { 250 struct vm_area_struct *next = vma->vm_next; 251 252 might_sleep(); 253 if (vma->vm_ops && vma->vm_ops->close) 254 vma->vm_ops->close(vma); 255 if (vma->vm_file) 256 fput(vma->vm_file); 257 mpol_put(vma_policy(vma)); 258 kmem_cache_free(vm_area_cachep, vma); 259 return next; 260 } 261 262 static unsigned long do_brk(unsigned long addr, unsigned long len); 263 264 SYSCALL_DEFINE1(brk, unsigned long, brk) 265 { 266 unsigned long rlim, retval; 267 unsigned long newbrk, oldbrk; 268 struct mm_struct *mm = current->mm; 269 unsigned long min_brk; 270 bool populate; 271 272 down_write(&mm->mmap_sem); 273 274 #ifdef CONFIG_COMPAT_BRK 275 /* 276 * CONFIG_COMPAT_BRK can still be overridden by setting 277 * randomize_va_space to 2, which will still cause mm->start_brk 278 * to be arbitrarily shifted 279 */ 280 if (current->brk_randomized) 281 min_brk = mm->start_brk; 282 else 283 min_brk = mm->end_data; 284 #else 285 min_brk = mm->start_brk; 286 #endif 287 if (brk < min_brk) 288 goto out; 289 290 /* 291 * Check against rlimit here. If this check is done later after the test 292 * of oldbrk with newbrk then it can escape the test and let the data 293 * segment grow beyond its set limit the in case where the limit is 294 * not page aligned -Ram Gupta 295 */ 296 rlim = rlimit(RLIMIT_DATA); 297 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) + 298 (mm->end_data - mm->start_data) > rlim) 299 goto out; 300 301 newbrk = PAGE_ALIGN(brk); 302 oldbrk = PAGE_ALIGN(mm->brk); 303 if (oldbrk == newbrk) 304 goto set_brk; 305 306 /* Always allow shrinking brk. */ 307 if (brk <= mm->brk) { 308 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 309 goto set_brk; 310 goto out; 311 } 312 313 /* Check against existing mmap mappings. */ 314 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 315 goto out; 316 317 /* Ok, looks good - let it rip. */ 318 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 319 goto out; 320 321 set_brk: 322 mm->brk = brk; 323 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 324 up_write(&mm->mmap_sem); 325 if (populate) 326 mm_populate(oldbrk, newbrk - oldbrk); 327 return brk; 328 329 out: 330 retval = mm->brk; 331 up_write(&mm->mmap_sem); 332 return retval; 333 } 334 335 static long vma_compute_subtree_gap(struct vm_area_struct *vma) 336 { 337 unsigned long max, subtree_gap; 338 max = vma->vm_start; 339 if (vma->vm_prev) 340 max -= vma->vm_prev->vm_end; 341 if (vma->vm_rb.rb_left) { 342 subtree_gap = rb_entry(vma->vm_rb.rb_left, 343 struct vm_area_struct, vm_rb)->rb_subtree_gap; 344 if (subtree_gap > max) 345 max = subtree_gap; 346 } 347 if (vma->vm_rb.rb_right) { 348 subtree_gap = rb_entry(vma->vm_rb.rb_right, 349 struct vm_area_struct, vm_rb)->rb_subtree_gap; 350 if (subtree_gap > max) 351 max = subtree_gap; 352 } 353 return max; 354 } 355 356 #ifdef CONFIG_DEBUG_VM_RB 357 static int browse_rb(struct rb_root *root) 358 { 359 int i = 0, j, bug = 0; 360 struct rb_node *nd, *pn = NULL; 361 unsigned long prev = 0, pend = 0; 362 363 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 364 struct vm_area_struct *vma; 365 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 366 if (vma->vm_start < prev) { 367 pr_info("vm_start %lx prev %lx\n", vma->vm_start, prev); 368 bug = 1; 369 } 370 if (vma->vm_start < pend) { 371 pr_info("vm_start %lx pend %lx\n", vma->vm_start, pend); 372 bug = 1; 373 } 374 if (vma->vm_start > vma->vm_end) { 375 pr_info("vm_end %lx < vm_start %lx\n", 376 vma->vm_end, vma->vm_start); 377 bug = 1; 378 } 379 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) { 380 pr_info("free gap %lx, correct %lx\n", 381 vma->rb_subtree_gap, 382 vma_compute_subtree_gap(vma)); 383 bug = 1; 384 } 385 i++; 386 pn = nd; 387 prev = vma->vm_start; 388 pend = vma->vm_end; 389 } 390 j = 0; 391 for (nd = pn; nd; nd = rb_prev(nd)) 392 j++; 393 if (i != j) { 394 pr_info("backwards %d, forwards %d\n", j, i); 395 bug = 1; 396 } 397 return bug ? -1 : i; 398 } 399 400 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore) 401 { 402 struct rb_node *nd; 403 404 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 405 struct vm_area_struct *vma; 406 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 407 BUG_ON(vma != ignore && 408 vma->rb_subtree_gap != vma_compute_subtree_gap(vma)); 409 } 410 } 411 412 static void validate_mm(struct mm_struct *mm) 413 { 414 int bug = 0; 415 int i = 0; 416 unsigned long highest_address = 0; 417 struct vm_area_struct *vma = mm->mmap; 418 while (vma) { 419 struct anon_vma_chain *avc; 420 vma_lock_anon_vma(vma); 421 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 422 anon_vma_interval_tree_verify(avc); 423 vma_unlock_anon_vma(vma); 424 highest_address = vma->vm_end; 425 vma = vma->vm_next; 426 i++; 427 } 428 if (i != mm->map_count) { 429 pr_info("map_count %d vm_next %d\n", mm->map_count, i); 430 bug = 1; 431 } 432 if (highest_address != mm->highest_vm_end) { 433 pr_info("mm->highest_vm_end %lx, found %lx\n", 434 mm->highest_vm_end, highest_address); 435 bug = 1; 436 } 437 i = browse_rb(&mm->mm_rb); 438 if (i != mm->map_count) { 439 pr_info("map_count %d rb %d\n", mm->map_count, i); 440 bug = 1; 441 } 442 BUG_ON(bug); 443 } 444 #else 445 #define validate_mm_rb(root, ignore) do { } while (0) 446 #define validate_mm(mm) do { } while (0) 447 #endif 448 449 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb, 450 unsigned long, rb_subtree_gap, vma_compute_subtree_gap) 451 452 /* 453 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or 454 * vma->vm_prev->vm_end values changed, without modifying the vma's position 455 * in the rbtree. 456 */ 457 static void vma_gap_update(struct vm_area_struct *vma) 458 { 459 /* 460 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback 461 * function that does exacltly what we want. 462 */ 463 vma_gap_callbacks_propagate(&vma->vm_rb, NULL); 464 } 465 466 static inline void vma_rb_insert(struct vm_area_struct *vma, 467 struct rb_root *root) 468 { 469 /* All rb_subtree_gap values must be consistent prior to insertion */ 470 validate_mm_rb(root, NULL); 471 472 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 473 } 474 475 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root) 476 { 477 /* 478 * All rb_subtree_gap values must be consistent prior to erase, 479 * with the possible exception of the vma being erased. 480 */ 481 validate_mm_rb(root, vma); 482 483 /* 484 * Note rb_erase_augmented is a fairly large inline function, 485 * so make sure we instantiate it only once with our desired 486 * augmented rbtree callbacks. 487 */ 488 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 489 } 490 491 /* 492 * vma has some anon_vma assigned, and is already inserted on that 493 * anon_vma's interval trees. 494 * 495 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 496 * vma must be removed from the anon_vma's interval trees using 497 * anon_vma_interval_tree_pre_update_vma(). 498 * 499 * After the update, the vma will be reinserted using 500 * anon_vma_interval_tree_post_update_vma(). 501 * 502 * The entire update must be protected by exclusive mmap_sem and by 503 * the root anon_vma's mutex. 504 */ 505 static inline void 506 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 507 { 508 struct anon_vma_chain *avc; 509 510 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 511 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 512 } 513 514 static inline void 515 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 516 { 517 struct anon_vma_chain *avc; 518 519 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 520 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 521 } 522 523 static int find_vma_links(struct mm_struct *mm, unsigned long addr, 524 unsigned long end, struct vm_area_struct **pprev, 525 struct rb_node ***rb_link, struct rb_node **rb_parent) 526 { 527 struct rb_node **__rb_link, *__rb_parent, *rb_prev; 528 529 __rb_link = &mm->mm_rb.rb_node; 530 rb_prev = __rb_parent = NULL; 531 532 while (*__rb_link) { 533 struct vm_area_struct *vma_tmp; 534 535 __rb_parent = *__rb_link; 536 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 537 538 if (vma_tmp->vm_end > addr) { 539 /* Fail if an existing vma overlaps the area */ 540 if (vma_tmp->vm_start < end) 541 return -ENOMEM; 542 __rb_link = &__rb_parent->rb_left; 543 } else { 544 rb_prev = __rb_parent; 545 __rb_link = &__rb_parent->rb_right; 546 } 547 } 548 549 *pprev = NULL; 550 if (rb_prev) 551 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 552 *rb_link = __rb_link; 553 *rb_parent = __rb_parent; 554 return 0; 555 } 556 557 static unsigned long count_vma_pages_range(struct mm_struct *mm, 558 unsigned long addr, unsigned long end) 559 { 560 unsigned long nr_pages = 0; 561 struct vm_area_struct *vma; 562 563 /* Find first overlaping mapping */ 564 vma = find_vma_intersection(mm, addr, end); 565 if (!vma) 566 return 0; 567 568 nr_pages = (min(end, vma->vm_end) - 569 max(addr, vma->vm_start)) >> PAGE_SHIFT; 570 571 /* Iterate over the rest of the overlaps */ 572 for (vma = vma->vm_next; vma; vma = vma->vm_next) { 573 unsigned long overlap_len; 574 575 if (vma->vm_start > end) 576 break; 577 578 overlap_len = min(end, vma->vm_end) - vma->vm_start; 579 nr_pages += overlap_len >> PAGE_SHIFT; 580 } 581 582 return nr_pages; 583 } 584 585 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 586 struct rb_node **rb_link, struct rb_node *rb_parent) 587 { 588 /* Update tracking information for the gap following the new vma. */ 589 if (vma->vm_next) 590 vma_gap_update(vma->vm_next); 591 else 592 mm->highest_vm_end = vma->vm_end; 593 594 /* 595 * vma->vm_prev wasn't known when we followed the rbtree to find the 596 * correct insertion point for that vma. As a result, we could not 597 * update the vma vm_rb parents rb_subtree_gap values on the way down. 598 * So, we first insert the vma with a zero rb_subtree_gap value 599 * (to be consistent with what we did on the way down), and then 600 * immediately update the gap to the correct value. Finally we 601 * rebalance the rbtree after all augmented values have been set. 602 */ 603 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 604 vma->rb_subtree_gap = 0; 605 vma_gap_update(vma); 606 vma_rb_insert(vma, &mm->mm_rb); 607 } 608 609 static void __vma_link_file(struct vm_area_struct *vma) 610 { 611 struct file *file; 612 613 file = vma->vm_file; 614 if (file) { 615 struct address_space *mapping = file->f_mapping; 616 617 if (vma->vm_flags & VM_DENYWRITE) 618 atomic_dec(&file_inode(file)->i_writecount); 619 if (vma->vm_flags & VM_SHARED) 620 mapping->i_mmap_writable++; 621 622 flush_dcache_mmap_lock(mapping); 623 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 624 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear); 625 else 626 vma_interval_tree_insert(vma, &mapping->i_mmap); 627 flush_dcache_mmap_unlock(mapping); 628 } 629 } 630 631 static void 632 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 633 struct vm_area_struct *prev, struct rb_node **rb_link, 634 struct rb_node *rb_parent) 635 { 636 __vma_link_list(mm, vma, prev, rb_parent); 637 __vma_link_rb(mm, vma, rb_link, rb_parent); 638 } 639 640 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 641 struct vm_area_struct *prev, struct rb_node **rb_link, 642 struct rb_node *rb_parent) 643 { 644 struct address_space *mapping = NULL; 645 646 if (vma->vm_file) { 647 mapping = vma->vm_file->f_mapping; 648 mutex_lock(&mapping->i_mmap_mutex); 649 } 650 651 __vma_link(mm, vma, prev, rb_link, rb_parent); 652 __vma_link_file(vma); 653 654 if (mapping) 655 mutex_unlock(&mapping->i_mmap_mutex); 656 657 mm->map_count++; 658 validate_mm(mm); 659 } 660 661 /* 662 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the 663 * mm's list and rbtree. It has already been inserted into the interval tree. 664 */ 665 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 666 { 667 struct vm_area_struct *prev; 668 struct rb_node **rb_link, *rb_parent; 669 670 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 671 &prev, &rb_link, &rb_parent)) 672 BUG(); 673 __vma_link(mm, vma, prev, rb_link, rb_parent); 674 mm->map_count++; 675 } 676 677 static inline void 678 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 679 struct vm_area_struct *prev) 680 { 681 struct vm_area_struct *next; 682 683 vma_rb_erase(vma, &mm->mm_rb); 684 prev->vm_next = next = vma->vm_next; 685 if (next) 686 next->vm_prev = prev; 687 688 /* Kill the cache */ 689 vmacache_invalidate(mm); 690 } 691 692 /* 693 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 694 * is already present in an i_mmap tree without adjusting the tree. 695 * The following helper function should be used when such adjustments 696 * are necessary. The "insert" vma (if any) is to be inserted 697 * before we drop the necessary locks. 698 */ 699 int vma_adjust(struct vm_area_struct *vma, unsigned long start, 700 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 701 { 702 struct mm_struct *mm = vma->vm_mm; 703 struct vm_area_struct *next = vma->vm_next; 704 struct vm_area_struct *importer = NULL; 705 struct address_space *mapping = NULL; 706 struct rb_root *root = NULL; 707 struct anon_vma *anon_vma = NULL; 708 struct file *file = vma->vm_file; 709 bool start_changed = false, end_changed = false; 710 long adjust_next = 0; 711 int remove_next = 0; 712 713 if (next && !insert) { 714 struct vm_area_struct *exporter = NULL; 715 716 if (end >= next->vm_end) { 717 /* 718 * vma expands, overlapping all the next, and 719 * perhaps the one after too (mprotect case 6). 720 */ 721 again: remove_next = 1 + (end > next->vm_end); 722 end = next->vm_end; 723 exporter = next; 724 importer = vma; 725 } else if (end > next->vm_start) { 726 /* 727 * vma expands, overlapping part of the next: 728 * mprotect case 5 shifting the boundary up. 729 */ 730 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 731 exporter = next; 732 importer = vma; 733 } else if (end < vma->vm_end) { 734 /* 735 * vma shrinks, and !insert tells it's not 736 * split_vma inserting another: so it must be 737 * mprotect case 4 shifting the boundary down. 738 */ 739 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT); 740 exporter = vma; 741 importer = next; 742 } 743 744 /* 745 * Easily overlooked: when mprotect shifts the boundary, 746 * make sure the expanding vma has anon_vma set if the 747 * shrinking vma had, to cover any anon pages imported. 748 */ 749 if (exporter && exporter->anon_vma && !importer->anon_vma) { 750 if (anon_vma_clone(importer, exporter)) 751 return -ENOMEM; 752 importer->anon_vma = exporter->anon_vma; 753 } 754 } 755 756 if (file) { 757 mapping = file->f_mapping; 758 if (!(vma->vm_flags & VM_NONLINEAR)) { 759 root = &mapping->i_mmap; 760 uprobe_munmap(vma, vma->vm_start, vma->vm_end); 761 762 if (adjust_next) 763 uprobe_munmap(next, next->vm_start, 764 next->vm_end); 765 } 766 767 mutex_lock(&mapping->i_mmap_mutex); 768 if (insert) { 769 /* 770 * Put into interval tree now, so instantiated pages 771 * are visible to arm/parisc __flush_dcache_page 772 * throughout; but we cannot insert into address 773 * space until vma start or end is updated. 774 */ 775 __vma_link_file(insert); 776 } 777 } 778 779 vma_adjust_trans_huge(vma, start, end, adjust_next); 780 781 anon_vma = vma->anon_vma; 782 if (!anon_vma && adjust_next) 783 anon_vma = next->anon_vma; 784 if (anon_vma) { 785 VM_BUG_ON(adjust_next && next->anon_vma && 786 anon_vma != next->anon_vma); 787 anon_vma_lock_write(anon_vma); 788 anon_vma_interval_tree_pre_update_vma(vma); 789 if (adjust_next) 790 anon_vma_interval_tree_pre_update_vma(next); 791 } 792 793 if (root) { 794 flush_dcache_mmap_lock(mapping); 795 vma_interval_tree_remove(vma, root); 796 if (adjust_next) 797 vma_interval_tree_remove(next, root); 798 } 799 800 if (start != vma->vm_start) { 801 vma->vm_start = start; 802 start_changed = true; 803 } 804 if (end != vma->vm_end) { 805 vma->vm_end = end; 806 end_changed = true; 807 } 808 vma->vm_pgoff = pgoff; 809 if (adjust_next) { 810 next->vm_start += adjust_next << PAGE_SHIFT; 811 next->vm_pgoff += adjust_next; 812 } 813 814 if (root) { 815 if (adjust_next) 816 vma_interval_tree_insert(next, root); 817 vma_interval_tree_insert(vma, root); 818 flush_dcache_mmap_unlock(mapping); 819 } 820 821 if (remove_next) { 822 /* 823 * vma_merge has merged next into vma, and needs 824 * us to remove next before dropping the locks. 825 */ 826 __vma_unlink(mm, next, vma); 827 if (file) 828 __remove_shared_vm_struct(next, file, mapping); 829 } else if (insert) { 830 /* 831 * split_vma has split insert from vma, and needs 832 * us to insert it before dropping the locks 833 * (it may either follow vma or precede it). 834 */ 835 __insert_vm_struct(mm, insert); 836 } else { 837 if (start_changed) 838 vma_gap_update(vma); 839 if (end_changed) { 840 if (!next) 841 mm->highest_vm_end = end; 842 else if (!adjust_next) 843 vma_gap_update(next); 844 } 845 } 846 847 if (anon_vma) { 848 anon_vma_interval_tree_post_update_vma(vma); 849 if (adjust_next) 850 anon_vma_interval_tree_post_update_vma(next); 851 anon_vma_unlock_write(anon_vma); 852 } 853 if (mapping) 854 mutex_unlock(&mapping->i_mmap_mutex); 855 856 if (root) { 857 uprobe_mmap(vma); 858 859 if (adjust_next) 860 uprobe_mmap(next); 861 } 862 863 if (remove_next) { 864 if (file) { 865 uprobe_munmap(next, next->vm_start, next->vm_end); 866 fput(file); 867 } 868 if (next->anon_vma) 869 anon_vma_merge(vma, next); 870 mm->map_count--; 871 mpol_put(vma_policy(next)); 872 kmem_cache_free(vm_area_cachep, next); 873 /* 874 * In mprotect's case 6 (see comments on vma_merge), 875 * we must remove another next too. It would clutter 876 * up the code too much to do both in one go. 877 */ 878 next = vma->vm_next; 879 if (remove_next == 2) 880 goto again; 881 else if (next) 882 vma_gap_update(next); 883 else 884 mm->highest_vm_end = end; 885 } 886 if (insert && file) 887 uprobe_mmap(insert); 888 889 validate_mm(mm); 890 891 return 0; 892 } 893 894 /* 895 * If the vma has a ->close operation then the driver probably needs to release 896 * per-vma resources, so we don't attempt to merge those. 897 */ 898 static inline int is_mergeable_vma(struct vm_area_struct *vma, 899 struct file *file, unsigned long vm_flags) 900 { 901 /* 902 * VM_SOFTDIRTY should not prevent from VMA merging, if we 903 * match the flags but dirty bit -- the caller should mark 904 * merged VMA as dirty. If dirty bit won't be excluded from 905 * comparison, we increase pressue on the memory system forcing 906 * the kernel to generate new VMAs when old one could be 907 * extended instead. 908 */ 909 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 910 return 0; 911 if (vma->vm_file != file) 912 return 0; 913 if (vma->vm_ops && vma->vm_ops->close) 914 return 0; 915 return 1; 916 } 917 918 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 919 struct anon_vma *anon_vma2, 920 struct vm_area_struct *vma) 921 { 922 /* 923 * The list_is_singular() test is to avoid merging VMA cloned from 924 * parents. This can improve scalability caused by anon_vma lock. 925 */ 926 if ((!anon_vma1 || !anon_vma2) && (!vma || 927 list_is_singular(&vma->anon_vma_chain))) 928 return 1; 929 return anon_vma1 == anon_vma2; 930 } 931 932 /* 933 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 934 * in front of (at a lower virtual address and file offset than) the vma. 935 * 936 * We cannot merge two vmas if they have differently assigned (non-NULL) 937 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 938 * 939 * We don't check here for the merged mmap wrapping around the end of pagecache 940 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 941 * wrap, nor mmaps which cover the final page at index -1UL. 942 */ 943 static int 944 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 945 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 946 { 947 if (is_mergeable_vma(vma, file, vm_flags) && 948 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 949 if (vma->vm_pgoff == vm_pgoff) 950 return 1; 951 } 952 return 0; 953 } 954 955 /* 956 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 957 * beyond (at a higher virtual address and file offset than) the vma. 958 * 959 * We cannot merge two vmas if they have differently assigned (non-NULL) 960 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 961 */ 962 static int 963 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 964 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 965 { 966 if (is_mergeable_vma(vma, file, vm_flags) && 967 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 968 pgoff_t vm_pglen; 969 vm_pglen = vma_pages(vma); 970 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 971 return 1; 972 } 973 return 0; 974 } 975 976 /* 977 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 978 * whether that can be merged with its predecessor or its successor. 979 * Or both (it neatly fills a hole). 980 * 981 * In most cases - when called for mmap, brk or mremap - [addr,end) is 982 * certain not to be mapped by the time vma_merge is called; but when 983 * called for mprotect, it is certain to be already mapped (either at 984 * an offset within prev, or at the start of next), and the flags of 985 * this area are about to be changed to vm_flags - and the no-change 986 * case has already been eliminated. 987 * 988 * The following mprotect cases have to be considered, where AAAA is 989 * the area passed down from mprotect_fixup, never extending beyond one 990 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 991 * 992 * AAAA AAAA AAAA AAAA 993 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 994 * cannot merge might become might become might become 995 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 996 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 997 * mremap move: PPPPNNNNNNNN 8 998 * AAAA 999 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 1000 * might become case 1 below case 2 below case 3 below 1001 * 1002 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 1003 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 1004 */ 1005 struct vm_area_struct *vma_merge(struct mm_struct *mm, 1006 struct vm_area_struct *prev, unsigned long addr, 1007 unsigned long end, unsigned long vm_flags, 1008 struct anon_vma *anon_vma, struct file *file, 1009 pgoff_t pgoff, struct mempolicy *policy) 1010 { 1011 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 1012 struct vm_area_struct *area, *next; 1013 int err; 1014 1015 /* 1016 * We later require that vma->vm_flags == vm_flags, 1017 * so this tests vma->vm_flags & VM_SPECIAL, too. 1018 */ 1019 if (vm_flags & VM_SPECIAL) 1020 return NULL; 1021 1022 if (prev) 1023 next = prev->vm_next; 1024 else 1025 next = mm->mmap; 1026 area = next; 1027 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 1028 next = next->vm_next; 1029 1030 /* 1031 * Can it merge with the predecessor? 1032 */ 1033 if (prev && prev->vm_end == addr && 1034 mpol_equal(vma_policy(prev), policy) && 1035 can_vma_merge_after(prev, vm_flags, 1036 anon_vma, file, pgoff)) { 1037 /* 1038 * OK, it can. Can we now merge in the successor as well? 1039 */ 1040 if (next && end == next->vm_start && 1041 mpol_equal(policy, vma_policy(next)) && 1042 can_vma_merge_before(next, vm_flags, 1043 anon_vma, file, pgoff+pglen) && 1044 is_mergeable_anon_vma(prev->anon_vma, 1045 next->anon_vma, NULL)) { 1046 /* cases 1, 6 */ 1047 err = vma_adjust(prev, prev->vm_start, 1048 next->vm_end, prev->vm_pgoff, NULL); 1049 } else /* cases 2, 5, 7 */ 1050 err = vma_adjust(prev, prev->vm_start, 1051 end, prev->vm_pgoff, NULL); 1052 if (err) 1053 return NULL; 1054 khugepaged_enter_vma_merge(prev); 1055 return prev; 1056 } 1057 1058 /* 1059 * Can this new request be merged in front of next? 1060 */ 1061 if (next && end == next->vm_start && 1062 mpol_equal(policy, vma_policy(next)) && 1063 can_vma_merge_before(next, vm_flags, 1064 anon_vma, file, pgoff+pglen)) { 1065 if (prev && addr < prev->vm_end) /* case 4 */ 1066 err = vma_adjust(prev, prev->vm_start, 1067 addr, prev->vm_pgoff, NULL); 1068 else /* cases 3, 8 */ 1069 err = vma_adjust(area, addr, next->vm_end, 1070 next->vm_pgoff - pglen, NULL); 1071 if (err) 1072 return NULL; 1073 khugepaged_enter_vma_merge(area); 1074 return area; 1075 } 1076 1077 return NULL; 1078 } 1079 1080 /* 1081 * Rough compatbility check to quickly see if it's even worth looking 1082 * at sharing an anon_vma. 1083 * 1084 * They need to have the same vm_file, and the flags can only differ 1085 * in things that mprotect may change. 1086 * 1087 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1088 * we can merge the two vma's. For example, we refuse to merge a vma if 1089 * there is a vm_ops->close() function, because that indicates that the 1090 * driver is doing some kind of reference counting. But that doesn't 1091 * really matter for the anon_vma sharing case. 1092 */ 1093 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1094 { 1095 return a->vm_end == b->vm_start && 1096 mpol_equal(vma_policy(a), vma_policy(b)) && 1097 a->vm_file == b->vm_file && 1098 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) && 1099 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1100 } 1101 1102 /* 1103 * Do some basic sanity checking to see if we can re-use the anon_vma 1104 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1105 * the same as 'old', the other will be the new one that is trying 1106 * to share the anon_vma. 1107 * 1108 * NOTE! This runs with mm_sem held for reading, so it is possible that 1109 * the anon_vma of 'old' is concurrently in the process of being set up 1110 * by another page fault trying to merge _that_. But that's ok: if it 1111 * is being set up, that automatically means that it will be a singleton 1112 * acceptable for merging, so we can do all of this optimistically. But 1113 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer. 1114 * 1115 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1116 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1117 * is to return an anon_vma that is "complex" due to having gone through 1118 * a fork). 1119 * 1120 * We also make sure that the two vma's are compatible (adjacent, 1121 * and with the same memory policies). That's all stable, even with just 1122 * a read lock on the mm_sem. 1123 */ 1124 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1125 { 1126 if (anon_vma_compatible(a, b)) { 1127 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma); 1128 1129 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1130 return anon_vma; 1131 } 1132 return NULL; 1133 } 1134 1135 /* 1136 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1137 * neighbouring vmas for a suitable anon_vma, before it goes off 1138 * to allocate a new anon_vma. It checks because a repetitive 1139 * sequence of mprotects and faults may otherwise lead to distinct 1140 * anon_vmas being allocated, preventing vma merge in subsequent 1141 * mprotect. 1142 */ 1143 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1144 { 1145 struct anon_vma *anon_vma; 1146 struct vm_area_struct *near; 1147 1148 near = vma->vm_next; 1149 if (!near) 1150 goto try_prev; 1151 1152 anon_vma = reusable_anon_vma(near, vma, near); 1153 if (anon_vma) 1154 return anon_vma; 1155 try_prev: 1156 near = vma->vm_prev; 1157 if (!near) 1158 goto none; 1159 1160 anon_vma = reusable_anon_vma(near, near, vma); 1161 if (anon_vma) 1162 return anon_vma; 1163 none: 1164 /* 1165 * There's no absolute need to look only at touching neighbours: 1166 * we could search further afield for "compatible" anon_vmas. 1167 * But it would probably just be a waste of time searching, 1168 * or lead to too many vmas hanging off the same anon_vma. 1169 * We're trying to allow mprotect remerging later on, 1170 * not trying to minimize memory used for anon_vmas. 1171 */ 1172 return NULL; 1173 } 1174 1175 #ifdef CONFIG_PROC_FS 1176 void vm_stat_account(struct mm_struct *mm, unsigned long flags, 1177 struct file *file, long pages) 1178 { 1179 const unsigned long stack_flags 1180 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN); 1181 1182 mm->total_vm += pages; 1183 1184 if (file) { 1185 mm->shared_vm += pages; 1186 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC) 1187 mm->exec_vm += pages; 1188 } else if (flags & stack_flags) 1189 mm->stack_vm += pages; 1190 } 1191 #endif /* CONFIG_PROC_FS */ 1192 1193 /* 1194 * If a hint addr is less than mmap_min_addr change hint to be as 1195 * low as possible but still greater than mmap_min_addr 1196 */ 1197 static inline unsigned long round_hint_to_min(unsigned long hint) 1198 { 1199 hint &= PAGE_MASK; 1200 if (((void *)hint != NULL) && 1201 (hint < mmap_min_addr)) 1202 return PAGE_ALIGN(mmap_min_addr); 1203 return hint; 1204 } 1205 1206 static inline int mlock_future_check(struct mm_struct *mm, 1207 unsigned long flags, 1208 unsigned long len) 1209 { 1210 unsigned long locked, lock_limit; 1211 1212 /* mlock MCL_FUTURE? */ 1213 if (flags & VM_LOCKED) { 1214 locked = len >> PAGE_SHIFT; 1215 locked += mm->locked_vm; 1216 lock_limit = rlimit(RLIMIT_MEMLOCK); 1217 lock_limit >>= PAGE_SHIFT; 1218 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1219 return -EAGAIN; 1220 } 1221 return 0; 1222 } 1223 1224 /* 1225 * The caller must hold down_write(¤t->mm->mmap_sem). 1226 */ 1227 1228 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr, 1229 unsigned long len, unsigned long prot, 1230 unsigned long flags, unsigned long pgoff, 1231 unsigned long *populate) 1232 { 1233 struct mm_struct * mm = current->mm; 1234 vm_flags_t vm_flags; 1235 1236 *populate = 0; 1237 1238 /* 1239 * Does the application expect PROT_READ to imply PROT_EXEC? 1240 * 1241 * (the exception is when the underlying filesystem is noexec 1242 * mounted, in which case we dont add PROT_EXEC.) 1243 */ 1244 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1245 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC))) 1246 prot |= PROT_EXEC; 1247 1248 if (!len) 1249 return -EINVAL; 1250 1251 if (!(flags & MAP_FIXED)) 1252 addr = round_hint_to_min(addr); 1253 1254 /* Careful about overflows.. */ 1255 len = PAGE_ALIGN(len); 1256 if (!len) 1257 return -ENOMEM; 1258 1259 /* offset overflow? */ 1260 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1261 return -EOVERFLOW; 1262 1263 /* Too many mappings? */ 1264 if (mm->map_count > sysctl_max_map_count) 1265 return -ENOMEM; 1266 1267 /* Obtain the address to map to. we verify (or select) it and ensure 1268 * that it represents a valid section of the address space. 1269 */ 1270 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1271 if (addr & ~PAGE_MASK) 1272 return addr; 1273 1274 /* Do simple checking here so the lower-level routines won't have 1275 * to. we assume access permissions have been handled by the open 1276 * of the memory object, so we don't do any here. 1277 */ 1278 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 1279 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1280 1281 if (flags & MAP_LOCKED) 1282 if (!can_do_mlock()) 1283 return -EPERM; 1284 1285 if (mlock_future_check(mm, vm_flags, len)) 1286 return -EAGAIN; 1287 1288 if (file) { 1289 struct inode *inode = file_inode(file); 1290 1291 switch (flags & MAP_TYPE) { 1292 case MAP_SHARED: 1293 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 1294 return -EACCES; 1295 1296 /* 1297 * Make sure we don't allow writing to an append-only 1298 * file.. 1299 */ 1300 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1301 return -EACCES; 1302 1303 /* 1304 * Make sure there are no mandatory locks on the file. 1305 */ 1306 if (locks_verify_locked(file)) 1307 return -EAGAIN; 1308 1309 vm_flags |= VM_SHARED | VM_MAYSHARE; 1310 if (!(file->f_mode & FMODE_WRITE)) 1311 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1312 1313 /* fall through */ 1314 case MAP_PRIVATE: 1315 if (!(file->f_mode & FMODE_READ)) 1316 return -EACCES; 1317 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) { 1318 if (vm_flags & VM_EXEC) 1319 return -EPERM; 1320 vm_flags &= ~VM_MAYEXEC; 1321 } 1322 1323 if (!file->f_op->mmap) 1324 return -ENODEV; 1325 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1326 return -EINVAL; 1327 break; 1328 1329 default: 1330 return -EINVAL; 1331 } 1332 } else { 1333 switch (flags & MAP_TYPE) { 1334 case MAP_SHARED: 1335 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1336 return -EINVAL; 1337 /* 1338 * Ignore pgoff. 1339 */ 1340 pgoff = 0; 1341 vm_flags |= VM_SHARED | VM_MAYSHARE; 1342 break; 1343 case MAP_PRIVATE: 1344 /* 1345 * Set pgoff according to addr for anon_vma. 1346 */ 1347 pgoff = addr >> PAGE_SHIFT; 1348 break; 1349 default: 1350 return -EINVAL; 1351 } 1352 } 1353 1354 /* 1355 * Set 'VM_NORESERVE' if we should not account for the 1356 * memory use of this mapping. 1357 */ 1358 if (flags & MAP_NORESERVE) { 1359 /* We honor MAP_NORESERVE if allowed to overcommit */ 1360 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1361 vm_flags |= VM_NORESERVE; 1362 1363 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1364 if (file && is_file_hugepages(file)) 1365 vm_flags |= VM_NORESERVE; 1366 } 1367 1368 addr = mmap_region(file, addr, len, vm_flags, pgoff); 1369 if (!IS_ERR_VALUE(addr) && 1370 ((vm_flags & VM_LOCKED) || 1371 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1372 *populate = len; 1373 return addr; 1374 } 1375 1376 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1377 unsigned long, prot, unsigned long, flags, 1378 unsigned long, fd, unsigned long, pgoff) 1379 { 1380 struct file *file = NULL; 1381 unsigned long retval = -EBADF; 1382 1383 if (!(flags & MAP_ANONYMOUS)) { 1384 audit_mmap_fd(fd, flags); 1385 file = fget(fd); 1386 if (!file) 1387 goto out; 1388 if (is_file_hugepages(file)) 1389 len = ALIGN(len, huge_page_size(hstate_file(file))); 1390 retval = -EINVAL; 1391 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file))) 1392 goto out_fput; 1393 } else if (flags & MAP_HUGETLB) { 1394 struct user_struct *user = NULL; 1395 struct hstate *hs; 1396 1397 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK); 1398 if (!hs) 1399 return -EINVAL; 1400 1401 len = ALIGN(len, huge_page_size(hs)); 1402 /* 1403 * VM_NORESERVE is used because the reservations will be 1404 * taken when vm_ops->mmap() is called 1405 * A dummy user value is used because we are not locking 1406 * memory so no accounting is necessary 1407 */ 1408 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1409 VM_NORESERVE, 1410 &user, HUGETLB_ANONHUGE_INODE, 1411 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1412 if (IS_ERR(file)) 1413 return PTR_ERR(file); 1414 } 1415 1416 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1417 1418 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1419 out_fput: 1420 if (file) 1421 fput(file); 1422 out: 1423 return retval; 1424 } 1425 1426 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1427 struct mmap_arg_struct { 1428 unsigned long addr; 1429 unsigned long len; 1430 unsigned long prot; 1431 unsigned long flags; 1432 unsigned long fd; 1433 unsigned long offset; 1434 }; 1435 1436 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1437 { 1438 struct mmap_arg_struct a; 1439 1440 if (copy_from_user(&a, arg, sizeof(a))) 1441 return -EFAULT; 1442 if (a.offset & ~PAGE_MASK) 1443 return -EINVAL; 1444 1445 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1446 a.offset >> PAGE_SHIFT); 1447 } 1448 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1449 1450 /* 1451 * Some shared mappigns will want the pages marked read-only 1452 * to track write events. If so, we'll downgrade vm_page_prot 1453 * to the private version (using protection_map[] without the 1454 * VM_SHARED bit). 1455 */ 1456 int vma_wants_writenotify(struct vm_area_struct *vma) 1457 { 1458 vm_flags_t vm_flags = vma->vm_flags; 1459 1460 /* If it was private or non-writable, the write bit is already clear */ 1461 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1462 return 0; 1463 1464 /* The backer wishes to know when pages are first written to? */ 1465 if (vma->vm_ops && vma->vm_ops->page_mkwrite) 1466 return 1; 1467 1468 /* The open routine did something to the protections already? */ 1469 if (pgprot_val(vma->vm_page_prot) != 1470 pgprot_val(vm_get_page_prot(vm_flags))) 1471 return 0; 1472 1473 /* Specialty mapping? */ 1474 if (vm_flags & VM_PFNMAP) 1475 return 0; 1476 1477 /* Can the mapping track the dirty pages? */ 1478 return vma->vm_file && vma->vm_file->f_mapping && 1479 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1480 } 1481 1482 /* 1483 * We account for memory if it's a private writeable mapping, 1484 * not hugepages and VM_NORESERVE wasn't set. 1485 */ 1486 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1487 { 1488 /* 1489 * hugetlb has its own accounting separate from the core VM 1490 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1491 */ 1492 if (file && is_file_hugepages(file)) 1493 return 0; 1494 1495 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1496 } 1497 1498 unsigned long mmap_region(struct file *file, unsigned long addr, 1499 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff) 1500 { 1501 struct mm_struct *mm = current->mm; 1502 struct vm_area_struct *vma, *prev; 1503 int error; 1504 struct rb_node **rb_link, *rb_parent; 1505 unsigned long charged = 0; 1506 1507 /* Check against address space limit. */ 1508 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) { 1509 unsigned long nr_pages; 1510 1511 /* 1512 * MAP_FIXED may remove pages of mappings that intersects with 1513 * requested mapping. Account for the pages it would unmap. 1514 */ 1515 if (!(vm_flags & MAP_FIXED)) 1516 return -ENOMEM; 1517 1518 nr_pages = count_vma_pages_range(mm, addr, addr + len); 1519 1520 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages)) 1521 return -ENOMEM; 1522 } 1523 1524 /* Clear old maps */ 1525 error = -ENOMEM; 1526 munmap_back: 1527 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) { 1528 if (do_munmap(mm, addr, len)) 1529 return -ENOMEM; 1530 goto munmap_back; 1531 } 1532 1533 /* 1534 * Private writable mapping: check memory availability 1535 */ 1536 if (accountable_mapping(file, vm_flags)) { 1537 charged = len >> PAGE_SHIFT; 1538 if (security_vm_enough_memory_mm(mm, charged)) 1539 return -ENOMEM; 1540 vm_flags |= VM_ACCOUNT; 1541 } 1542 1543 /* 1544 * Can we just expand an old mapping? 1545 */ 1546 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL); 1547 if (vma) 1548 goto out; 1549 1550 /* 1551 * Determine the object being mapped and call the appropriate 1552 * specific mapper. the address has already been validated, but 1553 * not unmapped, but the maps are removed from the list. 1554 */ 1555 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1556 if (!vma) { 1557 error = -ENOMEM; 1558 goto unacct_error; 1559 } 1560 1561 vma->vm_mm = mm; 1562 vma->vm_start = addr; 1563 vma->vm_end = addr + len; 1564 vma->vm_flags = vm_flags; 1565 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1566 vma->vm_pgoff = pgoff; 1567 INIT_LIST_HEAD(&vma->anon_vma_chain); 1568 1569 if (file) { 1570 if (vm_flags & VM_DENYWRITE) { 1571 error = deny_write_access(file); 1572 if (error) 1573 goto free_vma; 1574 } 1575 vma->vm_file = get_file(file); 1576 error = file->f_op->mmap(file, vma); 1577 if (error) 1578 goto unmap_and_free_vma; 1579 1580 /* Can addr have changed?? 1581 * 1582 * Answer: Yes, several device drivers can do it in their 1583 * f_op->mmap method. -DaveM 1584 * Bug: If addr is changed, prev, rb_link, rb_parent should 1585 * be updated for vma_link() 1586 */ 1587 WARN_ON_ONCE(addr != vma->vm_start); 1588 1589 addr = vma->vm_start; 1590 vm_flags = vma->vm_flags; 1591 } else if (vm_flags & VM_SHARED) { 1592 error = shmem_zero_setup(vma); 1593 if (error) 1594 goto free_vma; 1595 } 1596 1597 if (vma_wants_writenotify(vma)) { 1598 pgprot_t pprot = vma->vm_page_prot; 1599 1600 /* Can vma->vm_page_prot have changed?? 1601 * 1602 * Answer: Yes, drivers may have changed it in their 1603 * f_op->mmap method. 1604 * 1605 * Ensures that vmas marked as uncached stay that way. 1606 */ 1607 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED); 1608 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot))) 1609 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 1610 } 1611 1612 vma_link(mm, vma, prev, rb_link, rb_parent); 1613 /* Once vma denies write, undo our temporary denial count */ 1614 if (vm_flags & VM_DENYWRITE) 1615 allow_write_access(file); 1616 file = vma->vm_file; 1617 out: 1618 perf_event_mmap(vma); 1619 1620 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT); 1621 if (vm_flags & VM_LOCKED) { 1622 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) || 1623 vma == get_gate_vma(current->mm))) 1624 mm->locked_vm += (len >> PAGE_SHIFT); 1625 else 1626 vma->vm_flags &= ~VM_LOCKED; 1627 } 1628 1629 if (file) 1630 uprobe_mmap(vma); 1631 1632 /* 1633 * New (or expanded) vma always get soft dirty status. 1634 * Otherwise user-space soft-dirty page tracker won't 1635 * be able to distinguish situation when vma area unmapped, 1636 * then new mapped in-place (which must be aimed as 1637 * a completely new data area). 1638 */ 1639 vma->vm_flags |= VM_SOFTDIRTY; 1640 1641 return addr; 1642 1643 unmap_and_free_vma: 1644 if (vm_flags & VM_DENYWRITE) 1645 allow_write_access(file); 1646 vma->vm_file = NULL; 1647 fput(file); 1648 1649 /* Undo any partial mapping done by a device driver. */ 1650 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1651 charged = 0; 1652 free_vma: 1653 kmem_cache_free(vm_area_cachep, vma); 1654 unacct_error: 1655 if (charged) 1656 vm_unacct_memory(charged); 1657 return error; 1658 } 1659 1660 unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1661 { 1662 /* 1663 * We implement the search by looking for an rbtree node that 1664 * immediately follows a suitable gap. That is, 1665 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length; 1666 * - gap_end = vma->vm_start >= info->low_limit + length; 1667 * - gap_end - gap_start >= length 1668 */ 1669 1670 struct mm_struct *mm = current->mm; 1671 struct vm_area_struct *vma; 1672 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1673 1674 /* Adjust search length to account for worst case alignment overhead */ 1675 length = info->length + info->align_mask; 1676 if (length < info->length) 1677 return -ENOMEM; 1678 1679 /* Adjust search limits by the desired length */ 1680 if (info->high_limit < length) 1681 return -ENOMEM; 1682 high_limit = info->high_limit - length; 1683 1684 if (info->low_limit > high_limit) 1685 return -ENOMEM; 1686 low_limit = info->low_limit + length; 1687 1688 /* Check if rbtree root looks promising */ 1689 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1690 goto check_highest; 1691 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1692 if (vma->rb_subtree_gap < length) 1693 goto check_highest; 1694 1695 while (true) { 1696 /* Visit left subtree if it looks promising */ 1697 gap_end = vma->vm_start; 1698 if (gap_end >= low_limit && vma->vm_rb.rb_left) { 1699 struct vm_area_struct *left = 1700 rb_entry(vma->vm_rb.rb_left, 1701 struct vm_area_struct, vm_rb); 1702 if (left->rb_subtree_gap >= length) { 1703 vma = left; 1704 continue; 1705 } 1706 } 1707 1708 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1709 check_current: 1710 /* Check if current node has a suitable gap */ 1711 if (gap_start > high_limit) 1712 return -ENOMEM; 1713 if (gap_end >= low_limit && gap_end - gap_start >= length) 1714 goto found; 1715 1716 /* Visit right subtree if it looks promising */ 1717 if (vma->vm_rb.rb_right) { 1718 struct vm_area_struct *right = 1719 rb_entry(vma->vm_rb.rb_right, 1720 struct vm_area_struct, vm_rb); 1721 if (right->rb_subtree_gap >= length) { 1722 vma = right; 1723 continue; 1724 } 1725 } 1726 1727 /* Go back up the rbtree to find next candidate node */ 1728 while (true) { 1729 struct rb_node *prev = &vma->vm_rb; 1730 if (!rb_parent(prev)) 1731 goto check_highest; 1732 vma = rb_entry(rb_parent(prev), 1733 struct vm_area_struct, vm_rb); 1734 if (prev == vma->vm_rb.rb_left) { 1735 gap_start = vma->vm_prev->vm_end; 1736 gap_end = vma->vm_start; 1737 goto check_current; 1738 } 1739 } 1740 } 1741 1742 check_highest: 1743 /* Check highest gap, which does not precede any rbtree node */ 1744 gap_start = mm->highest_vm_end; 1745 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */ 1746 if (gap_start > high_limit) 1747 return -ENOMEM; 1748 1749 found: 1750 /* We found a suitable gap. Clip it with the original low_limit. */ 1751 if (gap_start < info->low_limit) 1752 gap_start = info->low_limit; 1753 1754 /* Adjust gap address to the desired alignment */ 1755 gap_start += (info->align_offset - gap_start) & info->align_mask; 1756 1757 VM_BUG_ON(gap_start + info->length > info->high_limit); 1758 VM_BUG_ON(gap_start + info->length > gap_end); 1759 return gap_start; 1760 } 1761 1762 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1763 { 1764 struct mm_struct *mm = current->mm; 1765 struct vm_area_struct *vma; 1766 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1767 1768 /* Adjust search length to account for worst case alignment overhead */ 1769 length = info->length + info->align_mask; 1770 if (length < info->length) 1771 return -ENOMEM; 1772 1773 /* 1774 * Adjust search limits by the desired length. 1775 * See implementation comment at top of unmapped_area(). 1776 */ 1777 gap_end = info->high_limit; 1778 if (gap_end < length) 1779 return -ENOMEM; 1780 high_limit = gap_end - length; 1781 1782 if (info->low_limit > high_limit) 1783 return -ENOMEM; 1784 low_limit = info->low_limit + length; 1785 1786 /* Check highest gap, which does not precede any rbtree node */ 1787 gap_start = mm->highest_vm_end; 1788 if (gap_start <= high_limit) 1789 goto found_highest; 1790 1791 /* Check if rbtree root looks promising */ 1792 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1793 return -ENOMEM; 1794 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1795 if (vma->rb_subtree_gap < length) 1796 return -ENOMEM; 1797 1798 while (true) { 1799 /* Visit right subtree if it looks promising */ 1800 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1801 if (gap_start <= high_limit && vma->vm_rb.rb_right) { 1802 struct vm_area_struct *right = 1803 rb_entry(vma->vm_rb.rb_right, 1804 struct vm_area_struct, vm_rb); 1805 if (right->rb_subtree_gap >= length) { 1806 vma = right; 1807 continue; 1808 } 1809 } 1810 1811 check_current: 1812 /* Check if current node has a suitable gap */ 1813 gap_end = vma->vm_start; 1814 if (gap_end < low_limit) 1815 return -ENOMEM; 1816 if (gap_start <= high_limit && gap_end - gap_start >= length) 1817 goto found; 1818 1819 /* Visit left subtree if it looks promising */ 1820 if (vma->vm_rb.rb_left) { 1821 struct vm_area_struct *left = 1822 rb_entry(vma->vm_rb.rb_left, 1823 struct vm_area_struct, vm_rb); 1824 if (left->rb_subtree_gap >= length) { 1825 vma = left; 1826 continue; 1827 } 1828 } 1829 1830 /* Go back up the rbtree to find next candidate node */ 1831 while (true) { 1832 struct rb_node *prev = &vma->vm_rb; 1833 if (!rb_parent(prev)) 1834 return -ENOMEM; 1835 vma = rb_entry(rb_parent(prev), 1836 struct vm_area_struct, vm_rb); 1837 if (prev == vma->vm_rb.rb_right) { 1838 gap_start = vma->vm_prev ? 1839 vma->vm_prev->vm_end : 0; 1840 goto check_current; 1841 } 1842 } 1843 } 1844 1845 found: 1846 /* We found a suitable gap. Clip it with the original high_limit. */ 1847 if (gap_end > info->high_limit) 1848 gap_end = info->high_limit; 1849 1850 found_highest: 1851 /* Compute highest gap address at the desired alignment */ 1852 gap_end -= info->length; 1853 gap_end -= (gap_end - info->align_offset) & info->align_mask; 1854 1855 VM_BUG_ON(gap_end < info->low_limit); 1856 VM_BUG_ON(gap_end < gap_start); 1857 return gap_end; 1858 } 1859 1860 /* Get an address range which is currently unmapped. 1861 * For shmat() with addr=0. 1862 * 1863 * Ugly calling convention alert: 1864 * Return value with the low bits set means error value, 1865 * ie 1866 * if (ret & ~PAGE_MASK) 1867 * error = ret; 1868 * 1869 * This function "knows" that -ENOMEM has the bits set. 1870 */ 1871 #ifndef HAVE_ARCH_UNMAPPED_AREA 1872 unsigned long 1873 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1874 unsigned long len, unsigned long pgoff, unsigned long flags) 1875 { 1876 struct mm_struct *mm = current->mm; 1877 struct vm_area_struct *vma; 1878 struct vm_unmapped_area_info info; 1879 1880 if (len > TASK_SIZE - mmap_min_addr) 1881 return -ENOMEM; 1882 1883 if (flags & MAP_FIXED) 1884 return addr; 1885 1886 if (addr) { 1887 addr = PAGE_ALIGN(addr); 1888 vma = find_vma(mm, addr); 1889 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1890 (!vma || addr + len <= vma->vm_start)) 1891 return addr; 1892 } 1893 1894 info.flags = 0; 1895 info.length = len; 1896 info.low_limit = mm->mmap_base; 1897 info.high_limit = TASK_SIZE; 1898 info.align_mask = 0; 1899 return vm_unmapped_area(&info); 1900 } 1901 #endif 1902 1903 /* 1904 * This mmap-allocator allocates new areas top-down from below the 1905 * stack's low limit (the base): 1906 */ 1907 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1908 unsigned long 1909 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1910 const unsigned long len, const unsigned long pgoff, 1911 const unsigned long flags) 1912 { 1913 struct vm_area_struct *vma; 1914 struct mm_struct *mm = current->mm; 1915 unsigned long addr = addr0; 1916 struct vm_unmapped_area_info info; 1917 1918 /* requested length too big for entire address space */ 1919 if (len > TASK_SIZE - mmap_min_addr) 1920 return -ENOMEM; 1921 1922 if (flags & MAP_FIXED) 1923 return addr; 1924 1925 /* requesting a specific address */ 1926 if (addr) { 1927 addr = PAGE_ALIGN(addr); 1928 vma = find_vma(mm, addr); 1929 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1930 (!vma || addr + len <= vma->vm_start)) 1931 return addr; 1932 } 1933 1934 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 1935 info.length = len; 1936 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 1937 info.high_limit = mm->mmap_base; 1938 info.align_mask = 0; 1939 addr = vm_unmapped_area(&info); 1940 1941 /* 1942 * A failed mmap() very likely causes application failure, 1943 * so fall back to the bottom-up function here. This scenario 1944 * can happen with large stack limits and large mmap() 1945 * allocations. 1946 */ 1947 if (addr & ~PAGE_MASK) { 1948 VM_BUG_ON(addr != -ENOMEM); 1949 info.flags = 0; 1950 info.low_limit = TASK_UNMAPPED_BASE; 1951 info.high_limit = TASK_SIZE; 1952 addr = vm_unmapped_area(&info); 1953 } 1954 1955 return addr; 1956 } 1957 #endif 1958 1959 unsigned long 1960 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1961 unsigned long pgoff, unsigned long flags) 1962 { 1963 unsigned long (*get_area)(struct file *, unsigned long, 1964 unsigned long, unsigned long, unsigned long); 1965 1966 unsigned long error = arch_mmap_check(addr, len, flags); 1967 if (error) 1968 return error; 1969 1970 /* Careful about overflows.. */ 1971 if (len > TASK_SIZE) 1972 return -ENOMEM; 1973 1974 get_area = current->mm->get_unmapped_area; 1975 if (file && file->f_op->get_unmapped_area) 1976 get_area = file->f_op->get_unmapped_area; 1977 addr = get_area(file, addr, len, pgoff, flags); 1978 if (IS_ERR_VALUE(addr)) 1979 return addr; 1980 1981 if (addr > TASK_SIZE - len) 1982 return -ENOMEM; 1983 if (addr & ~PAGE_MASK) 1984 return -EINVAL; 1985 1986 addr = arch_rebalance_pgtables(addr, len); 1987 error = security_mmap_addr(addr); 1988 return error ? error : addr; 1989 } 1990 1991 EXPORT_SYMBOL(get_unmapped_area); 1992 1993 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 1994 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 1995 { 1996 struct rb_node *rb_node; 1997 struct vm_area_struct *vma; 1998 1999 /* Check the cache first. */ 2000 vma = vmacache_find(mm, addr); 2001 if (likely(vma)) 2002 return vma; 2003 2004 rb_node = mm->mm_rb.rb_node; 2005 vma = NULL; 2006 2007 while (rb_node) { 2008 struct vm_area_struct *tmp; 2009 2010 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2011 2012 if (tmp->vm_end > addr) { 2013 vma = tmp; 2014 if (tmp->vm_start <= addr) 2015 break; 2016 rb_node = rb_node->rb_left; 2017 } else 2018 rb_node = rb_node->rb_right; 2019 } 2020 2021 if (vma) 2022 vmacache_update(addr, vma); 2023 return vma; 2024 } 2025 2026 EXPORT_SYMBOL(find_vma); 2027 2028 /* 2029 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 2030 */ 2031 struct vm_area_struct * 2032 find_vma_prev(struct mm_struct *mm, unsigned long addr, 2033 struct vm_area_struct **pprev) 2034 { 2035 struct vm_area_struct *vma; 2036 2037 vma = find_vma(mm, addr); 2038 if (vma) { 2039 *pprev = vma->vm_prev; 2040 } else { 2041 struct rb_node *rb_node = mm->mm_rb.rb_node; 2042 *pprev = NULL; 2043 while (rb_node) { 2044 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2045 rb_node = rb_node->rb_right; 2046 } 2047 } 2048 return vma; 2049 } 2050 2051 /* 2052 * Verify that the stack growth is acceptable and 2053 * update accounting. This is shared with both the 2054 * grow-up and grow-down cases. 2055 */ 2056 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 2057 { 2058 struct mm_struct *mm = vma->vm_mm; 2059 struct rlimit *rlim = current->signal->rlim; 2060 unsigned long new_start; 2061 2062 /* address space limit tests */ 2063 if (!may_expand_vm(mm, grow)) 2064 return -ENOMEM; 2065 2066 /* Stack limit test */ 2067 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 2068 return -ENOMEM; 2069 2070 /* mlock limit tests */ 2071 if (vma->vm_flags & VM_LOCKED) { 2072 unsigned long locked; 2073 unsigned long limit; 2074 locked = mm->locked_vm + grow; 2075 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 2076 limit >>= PAGE_SHIFT; 2077 if (locked > limit && !capable(CAP_IPC_LOCK)) 2078 return -ENOMEM; 2079 } 2080 2081 /* Check to ensure the stack will not grow into a hugetlb-only region */ 2082 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 2083 vma->vm_end - size; 2084 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 2085 return -EFAULT; 2086 2087 /* 2088 * Overcommit.. This must be the final test, as it will 2089 * update security statistics. 2090 */ 2091 if (security_vm_enough_memory_mm(mm, grow)) 2092 return -ENOMEM; 2093 2094 /* Ok, everything looks good - let it rip */ 2095 if (vma->vm_flags & VM_LOCKED) 2096 mm->locked_vm += grow; 2097 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow); 2098 return 0; 2099 } 2100 2101 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 2102 /* 2103 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 2104 * vma is the last one with address > vma->vm_end. Have to extend vma. 2105 */ 2106 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 2107 { 2108 int error; 2109 2110 if (!(vma->vm_flags & VM_GROWSUP)) 2111 return -EFAULT; 2112 2113 /* 2114 * We must make sure the anon_vma is allocated 2115 * so that the anon_vma locking is not a noop. 2116 */ 2117 if (unlikely(anon_vma_prepare(vma))) 2118 return -ENOMEM; 2119 vma_lock_anon_vma(vma); 2120 2121 /* 2122 * vma->vm_start/vm_end cannot change under us because the caller 2123 * is required to hold the mmap_sem in read mode. We need the 2124 * anon_vma lock to serialize against concurrent expand_stacks. 2125 * Also guard against wrapping around to address 0. 2126 */ 2127 if (address < PAGE_ALIGN(address+4)) 2128 address = PAGE_ALIGN(address+4); 2129 else { 2130 vma_unlock_anon_vma(vma); 2131 return -ENOMEM; 2132 } 2133 error = 0; 2134 2135 /* Somebody else might have raced and expanded it already */ 2136 if (address > vma->vm_end) { 2137 unsigned long size, grow; 2138 2139 size = address - vma->vm_start; 2140 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2141 2142 error = -ENOMEM; 2143 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2144 error = acct_stack_growth(vma, size, grow); 2145 if (!error) { 2146 /* 2147 * vma_gap_update() doesn't support concurrent 2148 * updates, but we only hold a shared mmap_sem 2149 * lock here, so we need to protect against 2150 * concurrent vma expansions. 2151 * vma_lock_anon_vma() doesn't help here, as 2152 * we don't guarantee that all growable vmas 2153 * in a mm share the same root anon vma. 2154 * So, we reuse mm->page_table_lock to guard 2155 * against concurrent vma expansions. 2156 */ 2157 spin_lock(&vma->vm_mm->page_table_lock); 2158 anon_vma_interval_tree_pre_update_vma(vma); 2159 vma->vm_end = address; 2160 anon_vma_interval_tree_post_update_vma(vma); 2161 if (vma->vm_next) 2162 vma_gap_update(vma->vm_next); 2163 else 2164 vma->vm_mm->highest_vm_end = address; 2165 spin_unlock(&vma->vm_mm->page_table_lock); 2166 2167 perf_event_mmap(vma); 2168 } 2169 } 2170 } 2171 vma_unlock_anon_vma(vma); 2172 khugepaged_enter_vma_merge(vma); 2173 validate_mm(vma->vm_mm); 2174 return error; 2175 } 2176 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2177 2178 /* 2179 * vma is the first one with address < vma->vm_start. Have to extend vma. 2180 */ 2181 int expand_downwards(struct vm_area_struct *vma, 2182 unsigned long address) 2183 { 2184 int error; 2185 2186 /* 2187 * We must make sure the anon_vma is allocated 2188 * so that the anon_vma locking is not a noop. 2189 */ 2190 if (unlikely(anon_vma_prepare(vma))) 2191 return -ENOMEM; 2192 2193 address &= PAGE_MASK; 2194 error = security_mmap_addr(address); 2195 if (error) 2196 return error; 2197 2198 vma_lock_anon_vma(vma); 2199 2200 /* 2201 * vma->vm_start/vm_end cannot change under us because the caller 2202 * is required to hold the mmap_sem in read mode. We need the 2203 * anon_vma lock to serialize against concurrent expand_stacks. 2204 */ 2205 2206 /* Somebody else might have raced and expanded it already */ 2207 if (address < vma->vm_start) { 2208 unsigned long size, grow; 2209 2210 size = vma->vm_end - address; 2211 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2212 2213 error = -ENOMEM; 2214 if (grow <= vma->vm_pgoff) { 2215 error = acct_stack_growth(vma, size, grow); 2216 if (!error) { 2217 /* 2218 * vma_gap_update() doesn't support concurrent 2219 * updates, but we only hold a shared mmap_sem 2220 * lock here, so we need to protect against 2221 * concurrent vma expansions. 2222 * vma_lock_anon_vma() doesn't help here, as 2223 * we don't guarantee that all growable vmas 2224 * in a mm share the same root anon vma. 2225 * So, we reuse mm->page_table_lock to guard 2226 * against concurrent vma expansions. 2227 */ 2228 spin_lock(&vma->vm_mm->page_table_lock); 2229 anon_vma_interval_tree_pre_update_vma(vma); 2230 vma->vm_start = address; 2231 vma->vm_pgoff -= grow; 2232 anon_vma_interval_tree_post_update_vma(vma); 2233 vma_gap_update(vma); 2234 spin_unlock(&vma->vm_mm->page_table_lock); 2235 2236 perf_event_mmap(vma); 2237 } 2238 } 2239 } 2240 vma_unlock_anon_vma(vma); 2241 khugepaged_enter_vma_merge(vma); 2242 validate_mm(vma->vm_mm); 2243 return error; 2244 } 2245 2246 /* 2247 * Note how expand_stack() refuses to expand the stack all the way to 2248 * abut the next virtual mapping, *unless* that mapping itself is also 2249 * a stack mapping. We want to leave room for a guard page, after all 2250 * (the guard page itself is not added here, that is done by the 2251 * actual page faulting logic) 2252 * 2253 * This matches the behavior of the guard page logic (see mm/memory.c: 2254 * check_stack_guard_page()), which only allows the guard page to be 2255 * removed under these circumstances. 2256 */ 2257 #ifdef CONFIG_STACK_GROWSUP 2258 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2259 { 2260 struct vm_area_struct *next; 2261 2262 address &= PAGE_MASK; 2263 next = vma->vm_next; 2264 if (next && next->vm_start == address + PAGE_SIZE) { 2265 if (!(next->vm_flags & VM_GROWSUP)) 2266 return -ENOMEM; 2267 } 2268 return expand_upwards(vma, address); 2269 } 2270 2271 struct vm_area_struct * 2272 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2273 { 2274 struct vm_area_struct *vma, *prev; 2275 2276 addr &= PAGE_MASK; 2277 vma = find_vma_prev(mm, addr, &prev); 2278 if (vma && (vma->vm_start <= addr)) 2279 return vma; 2280 if (!prev || expand_stack(prev, addr)) 2281 return NULL; 2282 if (prev->vm_flags & VM_LOCKED) 2283 __mlock_vma_pages_range(prev, addr, prev->vm_end, NULL); 2284 return prev; 2285 } 2286 #else 2287 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2288 { 2289 struct vm_area_struct *prev; 2290 2291 address &= PAGE_MASK; 2292 prev = vma->vm_prev; 2293 if (prev && prev->vm_end == address) { 2294 if (!(prev->vm_flags & VM_GROWSDOWN)) 2295 return -ENOMEM; 2296 } 2297 return expand_downwards(vma, address); 2298 } 2299 2300 struct vm_area_struct * 2301 find_extend_vma(struct mm_struct * mm, unsigned long addr) 2302 { 2303 struct vm_area_struct * vma; 2304 unsigned long start; 2305 2306 addr &= PAGE_MASK; 2307 vma = find_vma(mm,addr); 2308 if (!vma) 2309 return NULL; 2310 if (vma->vm_start <= addr) 2311 return vma; 2312 if (!(vma->vm_flags & VM_GROWSDOWN)) 2313 return NULL; 2314 start = vma->vm_start; 2315 if (expand_stack(vma, addr)) 2316 return NULL; 2317 if (vma->vm_flags & VM_LOCKED) 2318 __mlock_vma_pages_range(vma, addr, start, NULL); 2319 return vma; 2320 } 2321 #endif 2322 2323 /* 2324 * Ok - we have the memory areas we should free on the vma list, 2325 * so release them, and do the vma updates. 2326 * 2327 * Called with the mm semaphore held. 2328 */ 2329 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 2330 { 2331 unsigned long nr_accounted = 0; 2332 2333 /* Update high watermark before we lower total_vm */ 2334 update_hiwater_vm(mm); 2335 do { 2336 long nrpages = vma_pages(vma); 2337 2338 if (vma->vm_flags & VM_ACCOUNT) 2339 nr_accounted += nrpages; 2340 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages); 2341 vma = remove_vma(vma); 2342 } while (vma); 2343 vm_unacct_memory(nr_accounted); 2344 validate_mm(mm); 2345 } 2346 2347 /* 2348 * Get rid of page table information in the indicated region. 2349 * 2350 * Called with the mm semaphore held. 2351 */ 2352 static void unmap_region(struct mm_struct *mm, 2353 struct vm_area_struct *vma, struct vm_area_struct *prev, 2354 unsigned long start, unsigned long end) 2355 { 2356 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap; 2357 struct mmu_gather tlb; 2358 2359 lru_add_drain(); 2360 tlb_gather_mmu(&tlb, mm, start, end); 2361 update_hiwater_rss(mm); 2362 unmap_vmas(&tlb, vma, start, end); 2363 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2364 next ? next->vm_start : USER_PGTABLES_CEILING); 2365 tlb_finish_mmu(&tlb, start, end); 2366 } 2367 2368 /* 2369 * Create a list of vma's touched by the unmap, removing them from the mm's 2370 * vma list as we go.. 2371 */ 2372 static void 2373 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 2374 struct vm_area_struct *prev, unsigned long end) 2375 { 2376 struct vm_area_struct **insertion_point; 2377 struct vm_area_struct *tail_vma = NULL; 2378 2379 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 2380 vma->vm_prev = NULL; 2381 do { 2382 vma_rb_erase(vma, &mm->mm_rb); 2383 mm->map_count--; 2384 tail_vma = vma; 2385 vma = vma->vm_next; 2386 } while (vma && vma->vm_start < end); 2387 *insertion_point = vma; 2388 if (vma) { 2389 vma->vm_prev = prev; 2390 vma_gap_update(vma); 2391 } else 2392 mm->highest_vm_end = prev ? prev->vm_end : 0; 2393 tail_vma->vm_next = NULL; 2394 2395 /* Kill the cache */ 2396 vmacache_invalidate(mm); 2397 } 2398 2399 /* 2400 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the 2401 * munmap path where it doesn't make sense to fail. 2402 */ 2403 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma, 2404 unsigned long addr, int new_below) 2405 { 2406 struct vm_area_struct *new; 2407 int err = -ENOMEM; 2408 2409 if (is_vm_hugetlb_page(vma) && (addr & 2410 ~(huge_page_mask(hstate_vma(vma))))) 2411 return -EINVAL; 2412 2413 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2414 if (!new) 2415 goto out_err; 2416 2417 /* most fields are the same, copy all, and then fixup */ 2418 *new = *vma; 2419 2420 INIT_LIST_HEAD(&new->anon_vma_chain); 2421 2422 if (new_below) 2423 new->vm_end = addr; 2424 else { 2425 new->vm_start = addr; 2426 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2427 } 2428 2429 err = vma_dup_policy(vma, new); 2430 if (err) 2431 goto out_free_vma; 2432 2433 if (anon_vma_clone(new, vma)) 2434 goto out_free_mpol; 2435 2436 if (new->vm_file) 2437 get_file(new->vm_file); 2438 2439 if (new->vm_ops && new->vm_ops->open) 2440 new->vm_ops->open(new); 2441 2442 if (new_below) 2443 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 2444 ((addr - new->vm_start) >> PAGE_SHIFT), new); 2445 else 2446 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 2447 2448 /* Success. */ 2449 if (!err) 2450 return 0; 2451 2452 /* Clean everything up if vma_adjust failed. */ 2453 if (new->vm_ops && new->vm_ops->close) 2454 new->vm_ops->close(new); 2455 if (new->vm_file) 2456 fput(new->vm_file); 2457 unlink_anon_vmas(new); 2458 out_free_mpol: 2459 mpol_put(vma_policy(new)); 2460 out_free_vma: 2461 kmem_cache_free(vm_area_cachep, new); 2462 out_err: 2463 return err; 2464 } 2465 2466 /* 2467 * Split a vma into two pieces at address 'addr', a new vma is allocated 2468 * either for the first part or the tail. 2469 */ 2470 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2471 unsigned long addr, int new_below) 2472 { 2473 if (mm->map_count >= sysctl_max_map_count) 2474 return -ENOMEM; 2475 2476 return __split_vma(mm, vma, addr, new_below); 2477 } 2478 2479 /* Munmap is split into 2 main parts -- this part which finds 2480 * what needs doing, and the areas themselves, which do the 2481 * work. This now handles partial unmappings. 2482 * Jeremy Fitzhardinge <jeremy@goop.org> 2483 */ 2484 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 2485 { 2486 unsigned long end; 2487 struct vm_area_struct *vma, *prev, *last; 2488 2489 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start) 2490 return -EINVAL; 2491 2492 if ((len = PAGE_ALIGN(len)) == 0) 2493 return -EINVAL; 2494 2495 /* Find the first overlapping VMA */ 2496 vma = find_vma(mm, start); 2497 if (!vma) 2498 return 0; 2499 prev = vma->vm_prev; 2500 /* we have start < vma->vm_end */ 2501 2502 /* if it doesn't overlap, we have nothing.. */ 2503 end = start + len; 2504 if (vma->vm_start >= end) 2505 return 0; 2506 2507 /* 2508 * If we need to split any vma, do it now to save pain later. 2509 * 2510 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2511 * unmapped vm_area_struct will remain in use: so lower split_vma 2512 * places tmp vma above, and higher split_vma places tmp vma below. 2513 */ 2514 if (start > vma->vm_start) { 2515 int error; 2516 2517 /* 2518 * Make sure that map_count on return from munmap() will 2519 * not exceed its limit; but let map_count go just above 2520 * its limit temporarily, to help free resources as expected. 2521 */ 2522 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2523 return -ENOMEM; 2524 2525 error = __split_vma(mm, vma, start, 0); 2526 if (error) 2527 return error; 2528 prev = vma; 2529 } 2530 2531 /* Does it split the last one? */ 2532 last = find_vma(mm, end); 2533 if (last && end > last->vm_start) { 2534 int error = __split_vma(mm, last, end, 1); 2535 if (error) 2536 return error; 2537 } 2538 vma = prev? prev->vm_next: mm->mmap; 2539 2540 /* 2541 * unlock any mlock()ed ranges before detaching vmas 2542 */ 2543 if (mm->locked_vm) { 2544 struct vm_area_struct *tmp = vma; 2545 while (tmp && tmp->vm_start < end) { 2546 if (tmp->vm_flags & VM_LOCKED) { 2547 mm->locked_vm -= vma_pages(tmp); 2548 munlock_vma_pages_all(tmp); 2549 } 2550 tmp = tmp->vm_next; 2551 } 2552 } 2553 2554 /* 2555 * Remove the vma's, and unmap the actual pages 2556 */ 2557 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2558 unmap_region(mm, vma, prev, start, end); 2559 2560 /* Fix up all other VM information */ 2561 remove_vma_list(mm, vma); 2562 2563 return 0; 2564 } 2565 2566 int vm_munmap(unsigned long start, size_t len) 2567 { 2568 int ret; 2569 struct mm_struct *mm = current->mm; 2570 2571 down_write(&mm->mmap_sem); 2572 ret = do_munmap(mm, start, len); 2573 up_write(&mm->mmap_sem); 2574 return ret; 2575 } 2576 EXPORT_SYMBOL(vm_munmap); 2577 2578 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2579 { 2580 profile_munmap(addr); 2581 return vm_munmap(addr, len); 2582 } 2583 2584 static inline void verify_mm_writelocked(struct mm_struct *mm) 2585 { 2586 #ifdef CONFIG_DEBUG_VM 2587 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2588 WARN_ON(1); 2589 up_read(&mm->mmap_sem); 2590 } 2591 #endif 2592 } 2593 2594 /* 2595 * this is really a simplified "do_mmap". it only handles 2596 * anonymous maps. eventually we may be able to do some 2597 * brk-specific accounting here. 2598 */ 2599 static unsigned long do_brk(unsigned long addr, unsigned long len) 2600 { 2601 struct mm_struct * mm = current->mm; 2602 struct vm_area_struct * vma, * prev; 2603 unsigned long flags; 2604 struct rb_node ** rb_link, * rb_parent; 2605 pgoff_t pgoff = addr >> PAGE_SHIFT; 2606 int error; 2607 2608 len = PAGE_ALIGN(len); 2609 if (!len) 2610 return addr; 2611 2612 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2613 2614 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2615 if (error & ~PAGE_MASK) 2616 return error; 2617 2618 error = mlock_future_check(mm, mm->def_flags, len); 2619 if (error) 2620 return error; 2621 2622 /* 2623 * mm->mmap_sem is required to protect against another thread 2624 * changing the mappings in case we sleep. 2625 */ 2626 verify_mm_writelocked(mm); 2627 2628 /* 2629 * Clear old maps. this also does some error checking for us 2630 */ 2631 munmap_back: 2632 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) { 2633 if (do_munmap(mm, addr, len)) 2634 return -ENOMEM; 2635 goto munmap_back; 2636 } 2637 2638 /* Check against address space limits *after* clearing old maps... */ 2639 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 2640 return -ENOMEM; 2641 2642 if (mm->map_count > sysctl_max_map_count) 2643 return -ENOMEM; 2644 2645 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 2646 return -ENOMEM; 2647 2648 /* Can we just expand an old private anonymous mapping? */ 2649 vma = vma_merge(mm, prev, addr, addr + len, flags, 2650 NULL, NULL, pgoff, NULL); 2651 if (vma) 2652 goto out; 2653 2654 /* 2655 * create a vma struct for an anonymous mapping 2656 */ 2657 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2658 if (!vma) { 2659 vm_unacct_memory(len >> PAGE_SHIFT); 2660 return -ENOMEM; 2661 } 2662 2663 INIT_LIST_HEAD(&vma->anon_vma_chain); 2664 vma->vm_mm = mm; 2665 vma->vm_start = addr; 2666 vma->vm_end = addr + len; 2667 vma->vm_pgoff = pgoff; 2668 vma->vm_flags = flags; 2669 vma->vm_page_prot = vm_get_page_prot(flags); 2670 vma_link(mm, vma, prev, rb_link, rb_parent); 2671 out: 2672 perf_event_mmap(vma); 2673 mm->total_vm += len >> PAGE_SHIFT; 2674 if (flags & VM_LOCKED) 2675 mm->locked_vm += (len >> PAGE_SHIFT); 2676 vma->vm_flags |= VM_SOFTDIRTY; 2677 return addr; 2678 } 2679 2680 unsigned long vm_brk(unsigned long addr, unsigned long len) 2681 { 2682 struct mm_struct *mm = current->mm; 2683 unsigned long ret; 2684 bool populate; 2685 2686 down_write(&mm->mmap_sem); 2687 ret = do_brk(addr, len); 2688 populate = ((mm->def_flags & VM_LOCKED) != 0); 2689 up_write(&mm->mmap_sem); 2690 if (populate) 2691 mm_populate(addr, len); 2692 return ret; 2693 } 2694 EXPORT_SYMBOL(vm_brk); 2695 2696 /* Release all mmaps. */ 2697 void exit_mmap(struct mm_struct *mm) 2698 { 2699 struct mmu_gather tlb; 2700 struct vm_area_struct *vma; 2701 unsigned long nr_accounted = 0; 2702 2703 /* mm's last user has gone, and its about to be pulled down */ 2704 mmu_notifier_release(mm); 2705 2706 if (mm->locked_vm) { 2707 vma = mm->mmap; 2708 while (vma) { 2709 if (vma->vm_flags & VM_LOCKED) 2710 munlock_vma_pages_all(vma); 2711 vma = vma->vm_next; 2712 } 2713 } 2714 2715 arch_exit_mmap(mm); 2716 2717 vma = mm->mmap; 2718 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2719 return; 2720 2721 lru_add_drain(); 2722 flush_cache_mm(mm); 2723 tlb_gather_mmu(&tlb, mm, 0, -1); 2724 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2725 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2726 unmap_vmas(&tlb, vma, 0, -1); 2727 2728 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING); 2729 tlb_finish_mmu(&tlb, 0, -1); 2730 2731 /* 2732 * Walk the list again, actually closing and freeing it, 2733 * with preemption enabled, without holding any MM locks. 2734 */ 2735 while (vma) { 2736 if (vma->vm_flags & VM_ACCOUNT) 2737 nr_accounted += vma_pages(vma); 2738 vma = remove_vma(vma); 2739 } 2740 vm_unacct_memory(nr_accounted); 2741 2742 WARN_ON(atomic_long_read(&mm->nr_ptes) > 2743 (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT); 2744 } 2745 2746 /* Insert vm structure into process list sorted by address 2747 * and into the inode's i_mmap tree. If vm_file is non-NULL 2748 * then i_mmap_mutex is taken here. 2749 */ 2750 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 2751 { 2752 struct vm_area_struct *prev; 2753 struct rb_node **rb_link, *rb_parent; 2754 2755 /* 2756 * The vm_pgoff of a purely anonymous vma should be irrelevant 2757 * until its first write fault, when page's anon_vma and index 2758 * are set. But now set the vm_pgoff it will almost certainly 2759 * end up with (unless mremap moves it elsewhere before that 2760 * first wfault), so /proc/pid/maps tells a consistent story. 2761 * 2762 * By setting it to reflect the virtual start address of the 2763 * vma, merges and splits can happen in a seamless way, just 2764 * using the existing file pgoff checks and manipulations. 2765 * Similarly in do_mmap_pgoff and in do_brk. 2766 */ 2767 if (!vma->vm_file) { 2768 BUG_ON(vma->anon_vma); 2769 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2770 } 2771 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 2772 &prev, &rb_link, &rb_parent)) 2773 return -ENOMEM; 2774 if ((vma->vm_flags & VM_ACCOUNT) && 2775 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2776 return -ENOMEM; 2777 2778 vma_link(mm, vma, prev, rb_link, rb_parent); 2779 return 0; 2780 } 2781 2782 /* 2783 * Copy the vma structure to a new location in the same mm, 2784 * prior to moving page table entries, to effect an mremap move. 2785 */ 2786 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2787 unsigned long addr, unsigned long len, pgoff_t pgoff, 2788 bool *need_rmap_locks) 2789 { 2790 struct vm_area_struct *vma = *vmap; 2791 unsigned long vma_start = vma->vm_start; 2792 struct mm_struct *mm = vma->vm_mm; 2793 struct vm_area_struct *new_vma, *prev; 2794 struct rb_node **rb_link, *rb_parent; 2795 bool faulted_in_anon_vma = true; 2796 2797 /* 2798 * If anonymous vma has not yet been faulted, update new pgoff 2799 * to match new location, to increase its chance of merging. 2800 */ 2801 if (unlikely(!vma->vm_file && !vma->anon_vma)) { 2802 pgoff = addr >> PAGE_SHIFT; 2803 faulted_in_anon_vma = false; 2804 } 2805 2806 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) 2807 return NULL; /* should never get here */ 2808 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2809 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma)); 2810 if (new_vma) { 2811 /* 2812 * Source vma may have been merged into new_vma 2813 */ 2814 if (unlikely(vma_start >= new_vma->vm_start && 2815 vma_start < new_vma->vm_end)) { 2816 /* 2817 * The only way we can get a vma_merge with 2818 * self during an mremap is if the vma hasn't 2819 * been faulted in yet and we were allowed to 2820 * reset the dst vma->vm_pgoff to the 2821 * destination address of the mremap to allow 2822 * the merge to happen. mremap must change the 2823 * vm_pgoff linearity between src and dst vmas 2824 * (in turn preventing a vma_merge) to be 2825 * safe. It is only safe to keep the vm_pgoff 2826 * linear if there are no pages mapped yet. 2827 */ 2828 VM_BUG_ON(faulted_in_anon_vma); 2829 *vmap = vma = new_vma; 2830 } 2831 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 2832 } else { 2833 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2834 if (new_vma) { 2835 *new_vma = *vma; 2836 new_vma->vm_start = addr; 2837 new_vma->vm_end = addr + len; 2838 new_vma->vm_pgoff = pgoff; 2839 if (vma_dup_policy(vma, new_vma)) 2840 goto out_free_vma; 2841 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 2842 if (anon_vma_clone(new_vma, vma)) 2843 goto out_free_mempol; 2844 if (new_vma->vm_file) 2845 get_file(new_vma->vm_file); 2846 if (new_vma->vm_ops && new_vma->vm_ops->open) 2847 new_vma->vm_ops->open(new_vma); 2848 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2849 *need_rmap_locks = false; 2850 } 2851 } 2852 return new_vma; 2853 2854 out_free_mempol: 2855 mpol_put(vma_policy(new_vma)); 2856 out_free_vma: 2857 kmem_cache_free(vm_area_cachep, new_vma); 2858 return NULL; 2859 } 2860 2861 /* 2862 * Return true if the calling process may expand its vm space by the passed 2863 * number of pages 2864 */ 2865 int may_expand_vm(struct mm_struct *mm, unsigned long npages) 2866 { 2867 unsigned long cur = mm->total_vm; /* pages */ 2868 unsigned long lim; 2869 2870 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT; 2871 2872 if (cur + npages > lim) 2873 return 0; 2874 return 1; 2875 } 2876 2877 static int special_mapping_fault(struct vm_area_struct *vma, 2878 struct vm_fault *vmf); 2879 2880 /* 2881 * Having a close hook prevents vma merging regardless of flags. 2882 */ 2883 static void special_mapping_close(struct vm_area_struct *vma) 2884 { 2885 } 2886 2887 static const char *special_mapping_name(struct vm_area_struct *vma) 2888 { 2889 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 2890 } 2891 2892 static const struct vm_operations_struct special_mapping_vmops = { 2893 .close = special_mapping_close, 2894 .fault = special_mapping_fault, 2895 .name = special_mapping_name, 2896 }; 2897 2898 static const struct vm_operations_struct legacy_special_mapping_vmops = { 2899 .close = special_mapping_close, 2900 .fault = special_mapping_fault, 2901 }; 2902 2903 static int special_mapping_fault(struct vm_area_struct *vma, 2904 struct vm_fault *vmf) 2905 { 2906 pgoff_t pgoff; 2907 struct page **pages; 2908 2909 /* 2910 * special mappings have no vm_file, and in that case, the mm 2911 * uses vm_pgoff internally. So we have to subtract it from here. 2912 * We are allowed to do this because we are the mm; do not copy 2913 * this code into drivers! 2914 */ 2915 pgoff = vmf->pgoff - vma->vm_pgoff; 2916 2917 if (vma->vm_ops == &legacy_special_mapping_vmops) 2918 pages = vma->vm_private_data; 2919 else 2920 pages = ((struct vm_special_mapping *)vma->vm_private_data)-> 2921 pages; 2922 2923 for (; pgoff && *pages; ++pages) 2924 pgoff--; 2925 2926 if (*pages) { 2927 struct page *page = *pages; 2928 get_page(page); 2929 vmf->page = page; 2930 return 0; 2931 } 2932 2933 return VM_FAULT_SIGBUS; 2934 } 2935 2936 static struct vm_area_struct *__install_special_mapping( 2937 struct mm_struct *mm, 2938 unsigned long addr, unsigned long len, 2939 unsigned long vm_flags, const struct vm_operations_struct *ops, 2940 void *priv) 2941 { 2942 int ret; 2943 struct vm_area_struct *vma; 2944 2945 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2946 if (unlikely(vma == NULL)) 2947 return ERR_PTR(-ENOMEM); 2948 2949 INIT_LIST_HEAD(&vma->anon_vma_chain); 2950 vma->vm_mm = mm; 2951 vma->vm_start = addr; 2952 vma->vm_end = addr + len; 2953 2954 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY; 2955 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 2956 2957 vma->vm_ops = ops; 2958 vma->vm_private_data = priv; 2959 2960 ret = insert_vm_struct(mm, vma); 2961 if (ret) 2962 goto out; 2963 2964 mm->total_vm += len >> PAGE_SHIFT; 2965 2966 perf_event_mmap(vma); 2967 2968 return vma; 2969 2970 out: 2971 kmem_cache_free(vm_area_cachep, vma); 2972 return ERR_PTR(ret); 2973 } 2974 2975 /* 2976 * Called with mm->mmap_sem held for writing. 2977 * Insert a new vma covering the given region, with the given flags. 2978 * Its pages are supplied by the given array of struct page *. 2979 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 2980 * The region past the last page supplied will always produce SIGBUS. 2981 * The array pointer and the pages it points to are assumed to stay alive 2982 * for as long as this mapping might exist. 2983 */ 2984 struct vm_area_struct *_install_special_mapping( 2985 struct mm_struct *mm, 2986 unsigned long addr, unsigned long len, 2987 unsigned long vm_flags, const struct vm_special_mapping *spec) 2988 { 2989 return __install_special_mapping(mm, addr, len, vm_flags, 2990 &special_mapping_vmops, (void *)spec); 2991 } 2992 2993 int install_special_mapping(struct mm_struct *mm, 2994 unsigned long addr, unsigned long len, 2995 unsigned long vm_flags, struct page **pages) 2996 { 2997 struct vm_area_struct *vma = __install_special_mapping( 2998 mm, addr, len, vm_flags, &legacy_special_mapping_vmops, 2999 (void *)pages); 3000 3001 return PTR_ERR_OR_ZERO(vma); 3002 } 3003 3004 static DEFINE_MUTEX(mm_all_locks_mutex); 3005 3006 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3007 { 3008 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3009 /* 3010 * The LSB of head.next can't change from under us 3011 * because we hold the mm_all_locks_mutex. 3012 */ 3013 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem); 3014 /* 3015 * We can safely modify head.next after taking the 3016 * anon_vma->root->rwsem. If some other vma in this mm shares 3017 * the same anon_vma we won't take it again. 3018 * 3019 * No need of atomic instructions here, head.next 3020 * can't change from under us thanks to the 3021 * anon_vma->root->rwsem. 3022 */ 3023 if (__test_and_set_bit(0, (unsigned long *) 3024 &anon_vma->root->rb_root.rb_node)) 3025 BUG(); 3026 } 3027 } 3028 3029 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3030 { 3031 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3032 /* 3033 * AS_MM_ALL_LOCKS can't change from under us because 3034 * we hold the mm_all_locks_mutex. 3035 * 3036 * Operations on ->flags have to be atomic because 3037 * even if AS_MM_ALL_LOCKS is stable thanks to the 3038 * mm_all_locks_mutex, there may be other cpus 3039 * changing other bitflags in parallel to us. 3040 */ 3041 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3042 BUG(); 3043 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem); 3044 } 3045 } 3046 3047 /* 3048 * This operation locks against the VM for all pte/vma/mm related 3049 * operations that could ever happen on a certain mm. This includes 3050 * vmtruncate, try_to_unmap, and all page faults. 3051 * 3052 * The caller must take the mmap_sem in write mode before calling 3053 * mm_take_all_locks(). The caller isn't allowed to release the 3054 * mmap_sem until mm_drop_all_locks() returns. 3055 * 3056 * mmap_sem in write mode is required in order to block all operations 3057 * that could modify pagetables and free pages without need of 3058 * altering the vma layout (for example populate_range() with 3059 * nonlinear vmas). It's also needed in write mode to avoid new 3060 * anon_vmas to be associated with existing vmas. 3061 * 3062 * A single task can't take more than one mm_take_all_locks() in a row 3063 * or it would deadlock. 3064 * 3065 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3066 * mapping->flags avoid to take the same lock twice, if more than one 3067 * vma in this mm is backed by the same anon_vma or address_space. 3068 * 3069 * We can take all the locks in random order because the VM code 3070 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never 3071 * takes more than one of them in a row. Secondly we're protected 3072 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex. 3073 * 3074 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3075 * that may have to take thousand of locks. 3076 * 3077 * mm_take_all_locks() can fail if it's interrupted by signals. 3078 */ 3079 int mm_take_all_locks(struct mm_struct *mm) 3080 { 3081 struct vm_area_struct *vma; 3082 struct anon_vma_chain *avc; 3083 3084 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3085 3086 mutex_lock(&mm_all_locks_mutex); 3087 3088 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3089 if (signal_pending(current)) 3090 goto out_unlock; 3091 if (vma->vm_file && vma->vm_file->f_mapping) 3092 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3093 } 3094 3095 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3096 if (signal_pending(current)) 3097 goto out_unlock; 3098 if (vma->anon_vma) 3099 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3100 vm_lock_anon_vma(mm, avc->anon_vma); 3101 } 3102 3103 return 0; 3104 3105 out_unlock: 3106 mm_drop_all_locks(mm); 3107 return -EINTR; 3108 } 3109 3110 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3111 { 3112 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3113 /* 3114 * The LSB of head.next can't change to 0 from under 3115 * us because we hold the mm_all_locks_mutex. 3116 * 3117 * We must however clear the bitflag before unlocking 3118 * the vma so the users using the anon_vma->rb_root will 3119 * never see our bitflag. 3120 * 3121 * No need of atomic instructions here, head.next 3122 * can't change from under us until we release the 3123 * anon_vma->root->rwsem. 3124 */ 3125 if (!__test_and_clear_bit(0, (unsigned long *) 3126 &anon_vma->root->rb_root.rb_node)) 3127 BUG(); 3128 anon_vma_unlock_write(anon_vma); 3129 } 3130 } 3131 3132 static void vm_unlock_mapping(struct address_space *mapping) 3133 { 3134 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3135 /* 3136 * AS_MM_ALL_LOCKS can't change to 0 from under us 3137 * because we hold the mm_all_locks_mutex. 3138 */ 3139 mutex_unlock(&mapping->i_mmap_mutex); 3140 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3141 &mapping->flags)) 3142 BUG(); 3143 } 3144 } 3145 3146 /* 3147 * The mmap_sem cannot be released by the caller until 3148 * mm_drop_all_locks() returns. 3149 */ 3150 void mm_drop_all_locks(struct mm_struct *mm) 3151 { 3152 struct vm_area_struct *vma; 3153 struct anon_vma_chain *avc; 3154 3155 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3156 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3157 3158 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3159 if (vma->anon_vma) 3160 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3161 vm_unlock_anon_vma(avc->anon_vma); 3162 if (vma->vm_file && vma->vm_file->f_mapping) 3163 vm_unlock_mapping(vma->vm_file->f_mapping); 3164 } 3165 3166 mutex_unlock(&mm_all_locks_mutex); 3167 } 3168 3169 /* 3170 * initialise the VMA slab 3171 */ 3172 void __init mmap_init(void) 3173 { 3174 int ret; 3175 3176 ret = percpu_counter_init(&vm_committed_as, 0); 3177 VM_BUG_ON(ret); 3178 } 3179 3180 /* 3181 * Initialise sysctl_user_reserve_kbytes. 3182 * 3183 * This is intended to prevent a user from starting a single memory hogging 3184 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3185 * mode. 3186 * 3187 * The default value is min(3% of free memory, 128MB) 3188 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3189 */ 3190 static int init_user_reserve(void) 3191 { 3192 unsigned long free_kbytes; 3193 3194 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3195 3196 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3197 return 0; 3198 } 3199 subsys_initcall(init_user_reserve); 3200 3201 /* 3202 * Initialise sysctl_admin_reserve_kbytes. 3203 * 3204 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3205 * to log in and kill a memory hogging process. 3206 * 3207 * Systems with more than 256MB will reserve 8MB, enough to recover 3208 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3209 * only reserve 3% of free pages by default. 3210 */ 3211 static int init_admin_reserve(void) 3212 { 3213 unsigned long free_kbytes; 3214 3215 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3216 3217 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3218 return 0; 3219 } 3220 subsys_initcall(init_admin_reserve); 3221 3222 /* 3223 * Reinititalise user and admin reserves if memory is added or removed. 3224 * 3225 * The default user reserve max is 128MB, and the default max for the 3226 * admin reserve is 8MB. These are usually, but not always, enough to 3227 * enable recovery from a memory hogging process using login/sshd, a shell, 3228 * and tools like top. It may make sense to increase or even disable the 3229 * reserve depending on the existence of swap or variations in the recovery 3230 * tools. So, the admin may have changed them. 3231 * 3232 * If memory is added and the reserves have been eliminated or increased above 3233 * the default max, then we'll trust the admin. 3234 * 3235 * If memory is removed and there isn't enough free memory, then we 3236 * need to reset the reserves. 3237 * 3238 * Otherwise keep the reserve set by the admin. 3239 */ 3240 static int reserve_mem_notifier(struct notifier_block *nb, 3241 unsigned long action, void *data) 3242 { 3243 unsigned long tmp, free_kbytes; 3244 3245 switch (action) { 3246 case MEM_ONLINE: 3247 /* Default max is 128MB. Leave alone if modified by operator. */ 3248 tmp = sysctl_user_reserve_kbytes; 3249 if (0 < tmp && tmp < (1UL << 17)) 3250 init_user_reserve(); 3251 3252 /* Default max is 8MB. Leave alone if modified by operator. */ 3253 tmp = sysctl_admin_reserve_kbytes; 3254 if (0 < tmp && tmp < (1UL << 13)) 3255 init_admin_reserve(); 3256 3257 break; 3258 case MEM_OFFLINE: 3259 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3260 3261 if (sysctl_user_reserve_kbytes > free_kbytes) { 3262 init_user_reserve(); 3263 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3264 sysctl_user_reserve_kbytes); 3265 } 3266 3267 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3268 init_admin_reserve(); 3269 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3270 sysctl_admin_reserve_kbytes); 3271 } 3272 break; 3273 default: 3274 break; 3275 } 3276 return NOTIFY_OK; 3277 } 3278 3279 static struct notifier_block reserve_mem_nb = { 3280 .notifier_call = reserve_mem_notifier, 3281 }; 3282 3283 static int __meminit init_reserve_notifier(void) 3284 { 3285 if (register_hotmemory_notifier(&reserve_mem_nb)) 3286 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3287 3288 return 0; 3289 } 3290 subsys_initcall(init_reserve_notifier); 3291