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/mmdebug.h> 35 #include <linux/perf_event.h> 36 #include <linux/audit.h> 37 #include <linux/khugepaged.h> 38 #include <linux/uprobes.h> 39 #include <linux/rbtree_augmented.h> 40 #include <linux/sched/sysctl.h> 41 #include <linux/notifier.h> 42 #include <linux/memory.h> 43 #include <linux/printk.h> 44 45 #include <asm/uaccess.h> 46 #include <asm/cacheflush.h> 47 #include <asm/tlb.h> 48 #include <asm/mmu_context.h> 49 50 #include "internal.h" 51 52 #ifndef arch_mmap_check 53 #define arch_mmap_check(addr, len, flags) (0) 54 #endif 55 56 #ifndef arch_rebalance_pgtables 57 #define arch_rebalance_pgtables(addr, len) (addr) 58 #endif 59 60 static void unmap_region(struct mm_struct *mm, 61 struct vm_area_struct *vma, struct vm_area_struct *prev, 62 unsigned long start, unsigned long end); 63 64 /* description of effects of mapping type and prot in current implementation. 65 * this is due to the limited x86 page protection hardware. The expected 66 * behavior is in parens: 67 * 68 * map_type prot 69 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 70 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 71 * w: (no) no w: (no) no w: (yes) yes w: (no) no 72 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 73 * 74 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 75 * w: (no) no w: (no) no w: (copy) copy w: (no) no 76 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 77 * 78 */ 79 pgprot_t protection_map[16] = { 80 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 81 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 82 }; 83 84 pgprot_t vm_get_page_prot(unsigned long vm_flags) 85 { 86 return __pgprot(pgprot_val(protection_map[vm_flags & 87 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) | 88 pgprot_val(arch_vm_get_page_prot(vm_flags))); 89 } 90 EXPORT_SYMBOL(vm_get_page_prot); 91 92 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags) 93 { 94 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags)); 95 } 96 97 /* Update vma->vm_page_prot to reflect vma->vm_flags. */ 98 void vma_set_page_prot(struct vm_area_struct *vma) 99 { 100 unsigned long vm_flags = vma->vm_flags; 101 102 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags); 103 if (vma_wants_writenotify(vma)) { 104 vm_flags &= ~VM_SHARED; 105 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, 106 vm_flags); 107 } 108 } 109 110 111 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */ 112 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */ 113 unsigned long sysctl_overcommit_kbytes __read_mostly; 114 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT; 115 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */ 116 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */ 117 /* 118 * Make sure vm_committed_as in one cacheline and not cacheline shared with 119 * other variables. It can be updated by several CPUs frequently. 120 */ 121 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp; 122 123 /* 124 * The global memory commitment made in the system can be a metric 125 * that can be used to drive ballooning decisions when Linux is hosted 126 * as a guest. On Hyper-V, the host implements a policy engine for dynamically 127 * balancing memory across competing virtual machines that are hosted. 128 * Several metrics drive this policy engine including the guest reported 129 * memory commitment. 130 */ 131 unsigned long vm_memory_committed(void) 132 { 133 return percpu_counter_read_positive(&vm_committed_as); 134 } 135 EXPORT_SYMBOL_GPL(vm_memory_committed); 136 137 /* 138 * Check that a process has enough memory to allocate a new virtual 139 * mapping. 0 means there is enough memory for the allocation to 140 * succeed and -ENOMEM implies there is not. 141 * 142 * We currently support three overcommit policies, which are set via the 143 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting 144 * 145 * Strict overcommit modes added 2002 Feb 26 by Alan Cox. 146 * Additional code 2002 Jul 20 by Robert Love. 147 * 148 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise. 149 * 150 * Note this is a helper function intended to be used by LSMs which 151 * wish to use this logic. 152 */ 153 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin) 154 { 155 unsigned long free, allowed, reserve; 156 157 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) < 158 -(s64)vm_committed_as_batch * num_online_cpus(), 159 "memory commitment underflow"); 160 161 vm_acct_memory(pages); 162 163 /* 164 * Sometimes we want to use more memory than we have 165 */ 166 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS) 167 return 0; 168 169 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) { 170 free = global_page_state(NR_FREE_PAGES); 171 free += global_page_state(NR_FILE_PAGES); 172 173 /* 174 * shmem pages shouldn't be counted as free in this 175 * case, they can't be purged, only swapped out, and 176 * that won't affect the overall amount of available 177 * memory in the system. 178 */ 179 free -= global_page_state(NR_SHMEM); 180 181 free += get_nr_swap_pages(); 182 183 /* 184 * Any slabs which are created with the 185 * SLAB_RECLAIM_ACCOUNT flag claim to have contents 186 * which are reclaimable, under pressure. The dentry 187 * cache and most inode caches should fall into this 188 */ 189 free += global_page_state(NR_SLAB_RECLAIMABLE); 190 191 /* 192 * Leave reserved pages. The pages are not for anonymous pages. 193 */ 194 if (free <= totalreserve_pages) 195 goto error; 196 else 197 free -= totalreserve_pages; 198 199 /* 200 * Reserve some for root 201 */ 202 if (!cap_sys_admin) 203 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10); 204 205 if (free > pages) 206 return 0; 207 208 goto error; 209 } 210 211 allowed = vm_commit_limit(); 212 /* 213 * Reserve some for root 214 */ 215 if (!cap_sys_admin) 216 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10); 217 218 /* 219 * Don't let a single process grow so big a user can't recover 220 */ 221 if (mm) { 222 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10); 223 allowed -= min(mm->total_vm / 32, reserve); 224 } 225 226 if (percpu_counter_read_positive(&vm_committed_as) < allowed) 227 return 0; 228 error: 229 vm_unacct_memory(pages); 230 231 return -ENOMEM; 232 } 233 234 /* 235 * Requires inode->i_mapping->i_mmap_rwsem 236 */ 237 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 238 struct file *file, struct address_space *mapping) 239 { 240 if (vma->vm_flags & VM_DENYWRITE) 241 atomic_inc(&file_inode(file)->i_writecount); 242 if (vma->vm_flags & VM_SHARED) 243 mapping_unmap_writable(mapping); 244 245 flush_dcache_mmap_lock(mapping); 246 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 247 list_del_init(&vma->shared.nonlinear); 248 else 249 vma_interval_tree_remove(vma, &mapping->i_mmap); 250 flush_dcache_mmap_unlock(mapping); 251 } 252 253 /* 254 * Unlink a file-based vm structure from its interval tree, to hide 255 * vma from rmap and vmtruncate before freeing its page tables. 256 */ 257 void unlink_file_vma(struct vm_area_struct *vma) 258 { 259 struct file *file = vma->vm_file; 260 261 if (file) { 262 struct address_space *mapping = file->f_mapping; 263 i_mmap_lock_write(mapping); 264 __remove_shared_vm_struct(vma, file, mapping); 265 i_mmap_unlock_write(mapping); 266 } 267 } 268 269 /* 270 * Close a vm structure and free it, returning the next. 271 */ 272 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 273 { 274 struct vm_area_struct *next = vma->vm_next; 275 276 might_sleep(); 277 if (vma->vm_ops && vma->vm_ops->close) 278 vma->vm_ops->close(vma); 279 if (vma->vm_file) 280 fput(vma->vm_file); 281 mpol_put(vma_policy(vma)); 282 kmem_cache_free(vm_area_cachep, vma); 283 return next; 284 } 285 286 static unsigned long do_brk(unsigned long addr, unsigned long len); 287 288 SYSCALL_DEFINE1(brk, unsigned long, brk) 289 { 290 unsigned long retval; 291 unsigned long newbrk, oldbrk; 292 struct mm_struct *mm = current->mm; 293 unsigned long min_brk; 294 bool populate; 295 296 down_write(&mm->mmap_sem); 297 298 #ifdef CONFIG_COMPAT_BRK 299 /* 300 * CONFIG_COMPAT_BRK can still be overridden by setting 301 * randomize_va_space to 2, which will still cause mm->start_brk 302 * to be arbitrarily shifted 303 */ 304 if (current->brk_randomized) 305 min_brk = mm->start_brk; 306 else 307 min_brk = mm->end_data; 308 #else 309 min_brk = mm->start_brk; 310 #endif 311 if (brk < min_brk) 312 goto out; 313 314 /* 315 * Check against rlimit here. If this check is done later after the test 316 * of oldbrk with newbrk then it can escape the test and let the data 317 * segment grow beyond its set limit the in case where the limit is 318 * not page aligned -Ram Gupta 319 */ 320 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 321 mm->end_data, mm->start_data)) 322 goto out; 323 324 newbrk = PAGE_ALIGN(brk); 325 oldbrk = PAGE_ALIGN(mm->brk); 326 if (oldbrk == newbrk) 327 goto set_brk; 328 329 /* Always allow shrinking brk. */ 330 if (brk <= mm->brk) { 331 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 332 goto set_brk; 333 goto out; 334 } 335 336 /* Check against existing mmap mappings. */ 337 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 338 goto out; 339 340 /* Ok, looks good - let it rip. */ 341 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 342 goto out; 343 344 set_brk: 345 mm->brk = brk; 346 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 347 up_write(&mm->mmap_sem); 348 if (populate) 349 mm_populate(oldbrk, newbrk - oldbrk); 350 return brk; 351 352 out: 353 retval = mm->brk; 354 up_write(&mm->mmap_sem); 355 return retval; 356 } 357 358 static long vma_compute_subtree_gap(struct vm_area_struct *vma) 359 { 360 unsigned long max, subtree_gap; 361 max = vma->vm_start; 362 if (vma->vm_prev) 363 max -= vma->vm_prev->vm_end; 364 if (vma->vm_rb.rb_left) { 365 subtree_gap = rb_entry(vma->vm_rb.rb_left, 366 struct vm_area_struct, vm_rb)->rb_subtree_gap; 367 if (subtree_gap > max) 368 max = subtree_gap; 369 } 370 if (vma->vm_rb.rb_right) { 371 subtree_gap = rb_entry(vma->vm_rb.rb_right, 372 struct vm_area_struct, vm_rb)->rb_subtree_gap; 373 if (subtree_gap > max) 374 max = subtree_gap; 375 } 376 return max; 377 } 378 379 #ifdef CONFIG_DEBUG_VM_RB 380 static int browse_rb(struct rb_root *root) 381 { 382 int i = 0, j, bug = 0; 383 struct rb_node *nd, *pn = NULL; 384 unsigned long prev = 0, pend = 0; 385 386 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 387 struct vm_area_struct *vma; 388 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 389 if (vma->vm_start < prev) { 390 pr_emerg("vm_start %lx < prev %lx\n", 391 vma->vm_start, prev); 392 bug = 1; 393 } 394 if (vma->vm_start < pend) { 395 pr_emerg("vm_start %lx < pend %lx\n", 396 vma->vm_start, pend); 397 bug = 1; 398 } 399 if (vma->vm_start > vma->vm_end) { 400 pr_emerg("vm_start %lx > vm_end %lx\n", 401 vma->vm_start, vma->vm_end); 402 bug = 1; 403 } 404 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) { 405 pr_emerg("free gap %lx, correct %lx\n", 406 vma->rb_subtree_gap, 407 vma_compute_subtree_gap(vma)); 408 bug = 1; 409 } 410 i++; 411 pn = nd; 412 prev = vma->vm_start; 413 pend = vma->vm_end; 414 } 415 j = 0; 416 for (nd = pn; nd; nd = rb_prev(nd)) 417 j++; 418 if (i != j) { 419 pr_emerg("backwards %d, forwards %d\n", j, i); 420 bug = 1; 421 } 422 return bug ? -1 : i; 423 } 424 425 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore) 426 { 427 struct rb_node *nd; 428 429 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 430 struct vm_area_struct *vma; 431 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 432 VM_BUG_ON_VMA(vma != ignore && 433 vma->rb_subtree_gap != vma_compute_subtree_gap(vma), 434 vma); 435 } 436 } 437 438 static void validate_mm(struct mm_struct *mm) 439 { 440 int bug = 0; 441 int i = 0; 442 unsigned long highest_address = 0; 443 struct vm_area_struct *vma = mm->mmap; 444 445 while (vma) { 446 struct anon_vma_chain *avc; 447 448 vma_lock_anon_vma(vma); 449 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 450 anon_vma_interval_tree_verify(avc); 451 vma_unlock_anon_vma(vma); 452 highest_address = vma->vm_end; 453 vma = vma->vm_next; 454 i++; 455 } 456 if (i != mm->map_count) { 457 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i); 458 bug = 1; 459 } 460 if (highest_address != mm->highest_vm_end) { 461 pr_emerg("mm->highest_vm_end %lx, found %lx\n", 462 mm->highest_vm_end, highest_address); 463 bug = 1; 464 } 465 i = browse_rb(&mm->mm_rb); 466 if (i != mm->map_count) { 467 if (i != -1) 468 pr_emerg("map_count %d rb %d\n", mm->map_count, i); 469 bug = 1; 470 } 471 VM_BUG_ON_MM(bug, mm); 472 } 473 #else 474 #define validate_mm_rb(root, ignore) do { } while (0) 475 #define validate_mm(mm) do { } while (0) 476 #endif 477 478 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb, 479 unsigned long, rb_subtree_gap, vma_compute_subtree_gap) 480 481 /* 482 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or 483 * vma->vm_prev->vm_end values changed, without modifying the vma's position 484 * in the rbtree. 485 */ 486 static void vma_gap_update(struct vm_area_struct *vma) 487 { 488 /* 489 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback 490 * function that does exacltly what we want. 491 */ 492 vma_gap_callbacks_propagate(&vma->vm_rb, NULL); 493 } 494 495 static inline void vma_rb_insert(struct vm_area_struct *vma, 496 struct rb_root *root) 497 { 498 /* All rb_subtree_gap values must be consistent prior to insertion */ 499 validate_mm_rb(root, NULL); 500 501 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 502 } 503 504 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root) 505 { 506 /* 507 * All rb_subtree_gap values must be consistent prior to erase, 508 * with the possible exception of the vma being erased. 509 */ 510 validate_mm_rb(root, vma); 511 512 /* 513 * Note rb_erase_augmented is a fairly large inline function, 514 * so make sure we instantiate it only once with our desired 515 * augmented rbtree callbacks. 516 */ 517 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 518 } 519 520 /* 521 * vma has some anon_vma assigned, and is already inserted on that 522 * anon_vma's interval trees. 523 * 524 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 525 * vma must be removed from the anon_vma's interval trees using 526 * anon_vma_interval_tree_pre_update_vma(). 527 * 528 * After the update, the vma will be reinserted using 529 * anon_vma_interval_tree_post_update_vma(). 530 * 531 * The entire update must be protected by exclusive mmap_sem and by 532 * the root anon_vma's mutex. 533 */ 534 static inline void 535 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 536 { 537 struct anon_vma_chain *avc; 538 539 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 540 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 541 } 542 543 static inline void 544 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 545 { 546 struct anon_vma_chain *avc; 547 548 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 549 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 550 } 551 552 static int find_vma_links(struct mm_struct *mm, unsigned long addr, 553 unsigned long end, struct vm_area_struct **pprev, 554 struct rb_node ***rb_link, struct rb_node **rb_parent) 555 { 556 struct rb_node **__rb_link, *__rb_parent, *rb_prev; 557 558 __rb_link = &mm->mm_rb.rb_node; 559 rb_prev = __rb_parent = NULL; 560 561 while (*__rb_link) { 562 struct vm_area_struct *vma_tmp; 563 564 __rb_parent = *__rb_link; 565 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 566 567 if (vma_tmp->vm_end > addr) { 568 /* Fail if an existing vma overlaps the area */ 569 if (vma_tmp->vm_start < end) 570 return -ENOMEM; 571 __rb_link = &__rb_parent->rb_left; 572 } else { 573 rb_prev = __rb_parent; 574 __rb_link = &__rb_parent->rb_right; 575 } 576 } 577 578 *pprev = NULL; 579 if (rb_prev) 580 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 581 *rb_link = __rb_link; 582 *rb_parent = __rb_parent; 583 return 0; 584 } 585 586 static unsigned long count_vma_pages_range(struct mm_struct *mm, 587 unsigned long addr, unsigned long end) 588 { 589 unsigned long nr_pages = 0; 590 struct vm_area_struct *vma; 591 592 /* Find first overlaping mapping */ 593 vma = find_vma_intersection(mm, addr, end); 594 if (!vma) 595 return 0; 596 597 nr_pages = (min(end, vma->vm_end) - 598 max(addr, vma->vm_start)) >> PAGE_SHIFT; 599 600 /* Iterate over the rest of the overlaps */ 601 for (vma = vma->vm_next; vma; vma = vma->vm_next) { 602 unsigned long overlap_len; 603 604 if (vma->vm_start > end) 605 break; 606 607 overlap_len = min(end, vma->vm_end) - vma->vm_start; 608 nr_pages += overlap_len >> PAGE_SHIFT; 609 } 610 611 return nr_pages; 612 } 613 614 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 615 struct rb_node **rb_link, struct rb_node *rb_parent) 616 { 617 /* Update tracking information for the gap following the new vma. */ 618 if (vma->vm_next) 619 vma_gap_update(vma->vm_next); 620 else 621 mm->highest_vm_end = vma->vm_end; 622 623 /* 624 * vma->vm_prev wasn't known when we followed the rbtree to find the 625 * correct insertion point for that vma. As a result, we could not 626 * update the vma vm_rb parents rb_subtree_gap values on the way down. 627 * So, we first insert the vma with a zero rb_subtree_gap value 628 * (to be consistent with what we did on the way down), and then 629 * immediately update the gap to the correct value. Finally we 630 * rebalance the rbtree after all augmented values have been set. 631 */ 632 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 633 vma->rb_subtree_gap = 0; 634 vma_gap_update(vma); 635 vma_rb_insert(vma, &mm->mm_rb); 636 } 637 638 static void __vma_link_file(struct vm_area_struct *vma) 639 { 640 struct file *file; 641 642 file = vma->vm_file; 643 if (file) { 644 struct address_space *mapping = file->f_mapping; 645 646 if (vma->vm_flags & VM_DENYWRITE) 647 atomic_dec(&file_inode(file)->i_writecount); 648 if (vma->vm_flags & VM_SHARED) 649 atomic_inc(&mapping->i_mmap_writable); 650 651 flush_dcache_mmap_lock(mapping); 652 if (unlikely(vma->vm_flags & VM_NONLINEAR)) 653 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear); 654 else 655 vma_interval_tree_insert(vma, &mapping->i_mmap); 656 flush_dcache_mmap_unlock(mapping); 657 } 658 } 659 660 static void 661 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 662 struct vm_area_struct *prev, struct rb_node **rb_link, 663 struct rb_node *rb_parent) 664 { 665 __vma_link_list(mm, vma, prev, rb_parent); 666 __vma_link_rb(mm, vma, rb_link, rb_parent); 667 } 668 669 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 670 struct vm_area_struct *prev, struct rb_node **rb_link, 671 struct rb_node *rb_parent) 672 { 673 struct address_space *mapping = NULL; 674 675 if (vma->vm_file) { 676 mapping = vma->vm_file->f_mapping; 677 i_mmap_lock_write(mapping); 678 } 679 680 __vma_link(mm, vma, prev, rb_link, rb_parent); 681 __vma_link_file(vma); 682 683 if (mapping) 684 i_mmap_unlock_write(mapping); 685 686 mm->map_count++; 687 validate_mm(mm); 688 } 689 690 /* 691 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the 692 * mm's list and rbtree. It has already been inserted into the interval tree. 693 */ 694 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 695 { 696 struct vm_area_struct *prev; 697 struct rb_node **rb_link, *rb_parent; 698 699 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 700 &prev, &rb_link, &rb_parent)) 701 BUG(); 702 __vma_link(mm, vma, prev, rb_link, rb_parent); 703 mm->map_count++; 704 } 705 706 static inline void 707 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 708 struct vm_area_struct *prev) 709 { 710 struct vm_area_struct *next; 711 712 vma_rb_erase(vma, &mm->mm_rb); 713 prev->vm_next = next = vma->vm_next; 714 if (next) 715 next->vm_prev = prev; 716 717 /* Kill the cache */ 718 vmacache_invalidate(mm); 719 } 720 721 /* 722 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 723 * is already present in an i_mmap tree without adjusting the tree. 724 * The following helper function should be used when such adjustments 725 * are necessary. The "insert" vma (if any) is to be inserted 726 * before we drop the necessary locks. 727 */ 728 int vma_adjust(struct vm_area_struct *vma, unsigned long start, 729 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 730 { 731 struct mm_struct *mm = vma->vm_mm; 732 struct vm_area_struct *next = vma->vm_next; 733 struct vm_area_struct *importer = NULL; 734 struct address_space *mapping = NULL; 735 struct rb_root *root = NULL; 736 struct anon_vma *anon_vma = NULL; 737 struct file *file = vma->vm_file; 738 bool start_changed = false, end_changed = false; 739 long adjust_next = 0; 740 int remove_next = 0; 741 742 if (next && !insert) { 743 struct vm_area_struct *exporter = NULL; 744 745 if (end >= next->vm_end) { 746 /* 747 * vma expands, overlapping all the next, and 748 * perhaps the one after too (mprotect case 6). 749 */ 750 again: remove_next = 1 + (end > next->vm_end); 751 end = next->vm_end; 752 exporter = next; 753 importer = vma; 754 } else if (end > next->vm_start) { 755 /* 756 * vma expands, overlapping part of the next: 757 * mprotect case 5 shifting the boundary up. 758 */ 759 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 760 exporter = next; 761 importer = vma; 762 } else if (end < vma->vm_end) { 763 /* 764 * vma shrinks, and !insert tells it's not 765 * split_vma inserting another: so it must be 766 * mprotect case 4 shifting the boundary down. 767 */ 768 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT); 769 exporter = vma; 770 importer = next; 771 } 772 773 /* 774 * Easily overlooked: when mprotect shifts the boundary, 775 * make sure the expanding vma has anon_vma set if the 776 * shrinking vma had, to cover any anon pages imported. 777 */ 778 if (exporter && exporter->anon_vma && !importer->anon_vma) { 779 int error; 780 781 importer->anon_vma = exporter->anon_vma; 782 error = anon_vma_clone(importer, exporter); 783 if (error) { 784 importer->anon_vma = NULL; 785 return error; 786 } 787 } 788 } 789 790 if (file) { 791 mapping = file->f_mapping; 792 if (!(vma->vm_flags & VM_NONLINEAR)) { 793 root = &mapping->i_mmap; 794 uprobe_munmap(vma, vma->vm_start, vma->vm_end); 795 796 if (adjust_next) 797 uprobe_munmap(next, next->vm_start, 798 next->vm_end); 799 } 800 801 i_mmap_lock_write(mapping); 802 if (insert) { 803 /* 804 * Put into interval tree now, so instantiated pages 805 * are visible to arm/parisc __flush_dcache_page 806 * throughout; but we cannot insert into address 807 * space until vma start or end is updated. 808 */ 809 __vma_link_file(insert); 810 } 811 } 812 813 vma_adjust_trans_huge(vma, start, end, adjust_next); 814 815 anon_vma = vma->anon_vma; 816 if (!anon_vma && adjust_next) 817 anon_vma = next->anon_vma; 818 if (anon_vma) { 819 VM_BUG_ON_VMA(adjust_next && next->anon_vma && 820 anon_vma != next->anon_vma, next); 821 anon_vma_lock_write(anon_vma); 822 anon_vma_interval_tree_pre_update_vma(vma); 823 if (adjust_next) 824 anon_vma_interval_tree_pre_update_vma(next); 825 } 826 827 if (root) { 828 flush_dcache_mmap_lock(mapping); 829 vma_interval_tree_remove(vma, root); 830 if (adjust_next) 831 vma_interval_tree_remove(next, root); 832 } 833 834 if (start != vma->vm_start) { 835 vma->vm_start = start; 836 start_changed = true; 837 } 838 if (end != vma->vm_end) { 839 vma->vm_end = end; 840 end_changed = true; 841 } 842 vma->vm_pgoff = pgoff; 843 if (adjust_next) { 844 next->vm_start += adjust_next << PAGE_SHIFT; 845 next->vm_pgoff += adjust_next; 846 } 847 848 if (root) { 849 if (adjust_next) 850 vma_interval_tree_insert(next, root); 851 vma_interval_tree_insert(vma, root); 852 flush_dcache_mmap_unlock(mapping); 853 } 854 855 if (remove_next) { 856 /* 857 * vma_merge has merged next into vma, and needs 858 * us to remove next before dropping the locks. 859 */ 860 __vma_unlink(mm, next, vma); 861 if (file) 862 __remove_shared_vm_struct(next, file, mapping); 863 } else if (insert) { 864 /* 865 * split_vma has split insert from vma, and needs 866 * us to insert it before dropping the locks 867 * (it may either follow vma or precede it). 868 */ 869 __insert_vm_struct(mm, insert); 870 } else { 871 if (start_changed) 872 vma_gap_update(vma); 873 if (end_changed) { 874 if (!next) 875 mm->highest_vm_end = end; 876 else if (!adjust_next) 877 vma_gap_update(next); 878 } 879 } 880 881 if (anon_vma) { 882 anon_vma_interval_tree_post_update_vma(vma); 883 if (adjust_next) 884 anon_vma_interval_tree_post_update_vma(next); 885 anon_vma_unlock_write(anon_vma); 886 } 887 if (mapping) 888 i_mmap_unlock_write(mapping); 889 890 if (root) { 891 uprobe_mmap(vma); 892 893 if (adjust_next) 894 uprobe_mmap(next); 895 } 896 897 if (remove_next) { 898 if (file) { 899 uprobe_munmap(next, next->vm_start, next->vm_end); 900 fput(file); 901 } 902 if (next->anon_vma) 903 anon_vma_merge(vma, next); 904 mm->map_count--; 905 mpol_put(vma_policy(next)); 906 kmem_cache_free(vm_area_cachep, next); 907 /* 908 * In mprotect's case 6 (see comments on vma_merge), 909 * we must remove another next too. It would clutter 910 * up the code too much to do both in one go. 911 */ 912 next = vma->vm_next; 913 if (remove_next == 2) 914 goto again; 915 else if (next) 916 vma_gap_update(next); 917 else 918 mm->highest_vm_end = end; 919 } 920 if (insert && file) 921 uprobe_mmap(insert); 922 923 validate_mm(mm); 924 925 return 0; 926 } 927 928 /* 929 * If the vma has a ->close operation then the driver probably needs to release 930 * per-vma resources, so we don't attempt to merge those. 931 */ 932 static inline int is_mergeable_vma(struct vm_area_struct *vma, 933 struct file *file, unsigned long vm_flags) 934 { 935 /* 936 * VM_SOFTDIRTY should not prevent from VMA merging, if we 937 * match the flags but dirty bit -- the caller should mark 938 * merged VMA as dirty. If dirty bit won't be excluded from 939 * comparison, we increase pressue on the memory system forcing 940 * the kernel to generate new VMAs when old one could be 941 * extended instead. 942 */ 943 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 944 return 0; 945 if (vma->vm_file != file) 946 return 0; 947 if (vma->vm_ops && vma->vm_ops->close) 948 return 0; 949 return 1; 950 } 951 952 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 953 struct anon_vma *anon_vma2, 954 struct vm_area_struct *vma) 955 { 956 /* 957 * The list_is_singular() test is to avoid merging VMA cloned from 958 * parents. This can improve scalability caused by anon_vma lock. 959 */ 960 if ((!anon_vma1 || !anon_vma2) && (!vma || 961 list_is_singular(&vma->anon_vma_chain))) 962 return 1; 963 return anon_vma1 == anon_vma2; 964 } 965 966 /* 967 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 968 * in front of (at a lower virtual address and file offset than) the vma. 969 * 970 * We cannot merge two vmas if they have differently assigned (non-NULL) 971 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 972 * 973 * We don't check here for the merged mmap wrapping around the end of pagecache 974 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 975 * wrap, nor mmaps which cover the final page at index -1UL. 976 */ 977 static int 978 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 979 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 980 { 981 if (is_mergeable_vma(vma, file, vm_flags) && 982 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 983 if (vma->vm_pgoff == vm_pgoff) 984 return 1; 985 } 986 return 0; 987 } 988 989 /* 990 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 991 * beyond (at a higher virtual address and file offset than) the vma. 992 * 993 * We cannot merge two vmas if they have differently assigned (non-NULL) 994 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 995 */ 996 static int 997 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 998 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 999 { 1000 if (is_mergeable_vma(vma, file, vm_flags) && 1001 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1002 pgoff_t vm_pglen; 1003 vm_pglen = vma_pages(vma); 1004 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 1005 return 1; 1006 } 1007 return 0; 1008 } 1009 1010 /* 1011 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 1012 * whether that can be merged with its predecessor or its successor. 1013 * Or both (it neatly fills a hole). 1014 * 1015 * In most cases - when called for mmap, brk or mremap - [addr,end) is 1016 * certain not to be mapped by the time vma_merge is called; but when 1017 * called for mprotect, it is certain to be already mapped (either at 1018 * an offset within prev, or at the start of next), and the flags of 1019 * this area are about to be changed to vm_flags - and the no-change 1020 * case has already been eliminated. 1021 * 1022 * The following mprotect cases have to be considered, where AAAA is 1023 * the area passed down from mprotect_fixup, never extending beyond one 1024 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 1025 * 1026 * AAAA AAAA AAAA AAAA 1027 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 1028 * cannot merge might become might become might become 1029 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 1030 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 1031 * mremap move: PPPPNNNNNNNN 8 1032 * AAAA 1033 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 1034 * might become case 1 below case 2 below case 3 below 1035 * 1036 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 1037 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 1038 */ 1039 struct vm_area_struct *vma_merge(struct mm_struct *mm, 1040 struct vm_area_struct *prev, unsigned long addr, 1041 unsigned long end, unsigned long vm_flags, 1042 struct anon_vma *anon_vma, struct file *file, 1043 pgoff_t pgoff, struct mempolicy *policy) 1044 { 1045 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 1046 struct vm_area_struct *area, *next; 1047 int err; 1048 1049 /* 1050 * We later require that vma->vm_flags == vm_flags, 1051 * so this tests vma->vm_flags & VM_SPECIAL, too. 1052 */ 1053 if (vm_flags & VM_SPECIAL) 1054 return NULL; 1055 1056 if (prev) 1057 next = prev->vm_next; 1058 else 1059 next = mm->mmap; 1060 area = next; 1061 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 1062 next = next->vm_next; 1063 1064 /* 1065 * Can it merge with the predecessor? 1066 */ 1067 if (prev && prev->vm_end == addr && 1068 mpol_equal(vma_policy(prev), policy) && 1069 can_vma_merge_after(prev, vm_flags, 1070 anon_vma, file, pgoff)) { 1071 /* 1072 * OK, it can. Can we now merge in the successor as well? 1073 */ 1074 if (next && end == next->vm_start && 1075 mpol_equal(policy, vma_policy(next)) && 1076 can_vma_merge_before(next, vm_flags, 1077 anon_vma, file, pgoff+pglen) && 1078 is_mergeable_anon_vma(prev->anon_vma, 1079 next->anon_vma, NULL)) { 1080 /* cases 1, 6 */ 1081 err = vma_adjust(prev, prev->vm_start, 1082 next->vm_end, prev->vm_pgoff, NULL); 1083 } else /* cases 2, 5, 7 */ 1084 err = vma_adjust(prev, prev->vm_start, 1085 end, prev->vm_pgoff, NULL); 1086 if (err) 1087 return NULL; 1088 khugepaged_enter_vma_merge(prev, vm_flags); 1089 return prev; 1090 } 1091 1092 /* 1093 * Can this new request be merged in front of next? 1094 */ 1095 if (next && end == next->vm_start && 1096 mpol_equal(policy, vma_policy(next)) && 1097 can_vma_merge_before(next, vm_flags, 1098 anon_vma, file, pgoff+pglen)) { 1099 if (prev && addr < prev->vm_end) /* case 4 */ 1100 err = vma_adjust(prev, prev->vm_start, 1101 addr, prev->vm_pgoff, NULL); 1102 else /* cases 3, 8 */ 1103 err = vma_adjust(area, addr, next->vm_end, 1104 next->vm_pgoff - pglen, NULL); 1105 if (err) 1106 return NULL; 1107 khugepaged_enter_vma_merge(area, vm_flags); 1108 return area; 1109 } 1110 1111 return NULL; 1112 } 1113 1114 /* 1115 * Rough compatbility check to quickly see if it's even worth looking 1116 * at sharing an anon_vma. 1117 * 1118 * They need to have the same vm_file, and the flags can only differ 1119 * in things that mprotect may change. 1120 * 1121 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1122 * we can merge the two vma's. For example, we refuse to merge a vma if 1123 * there is a vm_ops->close() function, because that indicates that the 1124 * driver is doing some kind of reference counting. But that doesn't 1125 * really matter for the anon_vma sharing case. 1126 */ 1127 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1128 { 1129 return a->vm_end == b->vm_start && 1130 mpol_equal(vma_policy(a), vma_policy(b)) && 1131 a->vm_file == b->vm_file && 1132 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) && 1133 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1134 } 1135 1136 /* 1137 * Do some basic sanity checking to see if we can re-use the anon_vma 1138 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1139 * the same as 'old', the other will be the new one that is trying 1140 * to share the anon_vma. 1141 * 1142 * NOTE! This runs with mm_sem held for reading, so it is possible that 1143 * the anon_vma of 'old' is concurrently in the process of being set up 1144 * by another page fault trying to merge _that_. But that's ok: if it 1145 * is being set up, that automatically means that it will be a singleton 1146 * acceptable for merging, so we can do all of this optimistically. But 1147 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer. 1148 * 1149 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1150 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1151 * is to return an anon_vma that is "complex" due to having gone through 1152 * a fork). 1153 * 1154 * We also make sure that the two vma's are compatible (adjacent, 1155 * and with the same memory policies). That's all stable, even with just 1156 * a read lock on the mm_sem. 1157 */ 1158 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1159 { 1160 if (anon_vma_compatible(a, b)) { 1161 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma); 1162 1163 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1164 return anon_vma; 1165 } 1166 return NULL; 1167 } 1168 1169 /* 1170 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1171 * neighbouring vmas for a suitable anon_vma, before it goes off 1172 * to allocate a new anon_vma. It checks because a repetitive 1173 * sequence of mprotects and faults may otherwise lead to distinct 1174 * anon_vmas being allocated, preventing vma merge in subsequent 1175 * mprotect. 1176 */ 1177 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1178 { 1179 struct anon_vma *anon_vma; 1180 struct vm_area_struct *near; 1181 1182 near = vma->vm_next; 1183 if (!near) 1184 goto try_prev; 1185 1186 anon_vma = reusable_anon_vma(near, vma, near); 1187 if (anon_vma) 1188 return anon_vma; 1189 try_prev: 1190 near = vma->vm_prev; 1191 if (!near) 1192 goto none; 1193 1194 anon_vma = reusable_anon_vma(near, near, vma); 1195 if (anon_vma) 1196 return anon_vma; 1197 none: 1198 /* 1199 * There's no absolute need to look only at touching neighbours: 1200 * we could search further afield for "compatible" anon_vmas. 1201 * But it would probably just be a waste of time searching, 1202 * or lead to too many vmas hanging off the same anon_vma. 1203 * We're trying to allow mprotect remerging later on, 1204 * not trying to minimize memory used for anon_vmas. 1205 */ 1206 return NULL; 1207 } 1208 1209 #ifdef CONFIG_PROC_FS 1210 void vm_stat_account(struct mm_struct *mm, unsigned long flags, 1211 struct file *file, long pages) 1212 { 1213 const unsigned long stack_flags 1214 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN); 1215 1216 mm->total_vm += pages; 1217 1218 if (file) { 1219 mm->shared_vm += pages; 1220 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC) 1221 mm->exec_vm += pages; 1222 } else if (flags & stack_flags) 1223 mm->stack_vm += pages; 1224 } 1225 #endif /* CONFIG_PROC_FS */ 1226 1227 /* 1228 * If a hint addr is less than mmap_min_addr change hint to be as 1229 * low as possible but still greater than mmap_min_addr 1230 */ 1231 static inline unsigned long round_hint_to_min(unsigned long hint) 1232 { 1233 hint &= PAGE_MASK; 1234 if (((void *)hint != NULL) && 1235 (hint < mmap_min_addr)) 1236 return PAGE_ALIGN(mmap_min_addr); 1237 return hint; 1238 } 1239 1240 static inline int mlock_future_check(struct mm_struct *mm, 1241 unsigned long flags, 1242 unsigned long len) 1243 { 1244 unsigned long locked, lock_limit; 1245 1246 /* mlock MCL_FUTURE? */ 1247 if (flags & VM_LOCKED) { 1248 locked = len >> PAGE_SHIFT; 1249 locked += mm->locked_vm; 1250 lock_limit = rlimit(RLIMIT_MEMLOCK); 1251 lock_limit >>= PAGE_SHIFT; 1252 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1253 return -EAGAIN; 1254 } 1255 return 0; 1256 } 1257 1258 /* 1259 * The caller must hold down_write(¤t->mm->mmap_sem). 1260 */ 1261 1262 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr, 1263 unsigned long len, unsigned long prot, 1264 unsigned long flags, unsigned long pgoff, 1265 unsigned long *populate) 1266 { 1267 struct mm_struct *mm = current->mm; 1268 vm_flags_t vm_flags; 1269 1270 *populate = 0; 1271 1272 /* 1273 * Does the application expect PROT_READ to imply PROT_EXEC? 1274 * 1275 * (the exception is when the underlying filesystem is noexec 1276 * mounted, in which case we dont add PROT_EXEC.) 1277 */ 1278 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1279 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC))) 1280 prot |= PROT_EXEC; 1281 1282 if (!len) 1283 return -EINVAL; 1284 1285 if (!(flags & MAP_FIXED)) 1286 addr = round_hint_to_min(addr); 1287 1288 /* Careful about overflows.. */ 1289 len = PAGE_ALIGN(len); 1290 if (!len) 1291 return -ENOMEM; 1292 1293 /* offset overflow? */ 1294 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1295 return -EOVERFLOW; 1296 1297 /* Too many mappings? */ 1298 if (mm->map_count > sysctl_max_map_count) 1299 return -ENOMEM; 1300 1301 /* Obtain the address to map to. we verify (or select) it and ensure 1302 * that it represents a valid section of the address space. 1303 */ 1304 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1305 if (addr & ~PAGE_MASK) 1306 return addr; 1307 1308 /* Do simple checking here so the lower-level routines won't have 1309 * to. we assume access permissions have been handled by the open 1310 * of the memory object, so we don't do any here. 1311 */ 1312 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 1313 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1314 1315 if (flags & MAP_LOCKED) 1316 if (!can_do_mlock()) 1317 return -EPERM; 1318 1319 if (mlock_future_check(mm, vm_flags, len)) 1320 return -EAGAIN; 1321 1322 if (file) { 1323 struct inode *inode = file_inode(file); 1324 1325 switch (flags & MAP_TYPE) { 1326 case MAP_SHARED: 1327 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 1328 return -EACCES; 1329 1330 /* 1331 * Make sure we don't allow writing to an append-only 1332 * file.. 1333 */ 1334 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1335 return -EACCES; 1336 1337 /* 1338 * Make sure there are no mandatory locks on the file. 1339 */ 1340 if (locks_verify_locked(file)) 1341 return -EAGAIN; 1342 1343 vm_flags |= VM_SHARED | VM_MAYSHARE; 1344 if (!(file->f_mode & FMODE_WRITE)) 1345 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1346 1347 /* fall through */ 1348 case MAP_PRIVATE: 1349 if (!(file->f_mode & FMODE_READ)) 1350 return -EACCES; 1351 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) { 1352 if (vm_flags & VM_EXEC) 1353 return -EPERM; 1354 vm_flags &= ~VM_MAYEXEC; 1355 } 1356 1357 if (!file->f_op->mmap) 1358 return -ENODEV; 1359 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1360 return -EINVAL; 1361 break; 1362 1363 default: 1364 return -EINVAL; 1365 } 1366 } else { 1367 switch (flags & MAP_TYPE) { 1368 case MAP_SHARED: 1369 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1370 return -EINVAL; 1371 /* 1372 * Ignore pgoff. 1373 */ 1374 pgoff = 0; 1375 vm_flags |= VM_SHARED | VM_MAYSHARE; 1376 break; 1377 case MAP_PRIVATE: 1378 /* 1379 * Set pgoff according to addr for anon_vma. 1380 */ 1381 pgoff = addr >> PAGE_SHIFT; 1382 break; 1383 default: 1384 return -EINVAL; 1385 } 1386 } 1387 1388 /* 1389 * Set 'VM_NORESERVE' if we should not account for the 1390 * memory use of this mapping. 1391 */ 1392 if (flags & MAP_NORESERVE) { 1393 /* We honor MAP_NORESERVE if allowed to overcommit */ 1394 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1395 vm_flags |= VM_NORESERVE; 1396 1397 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1398 if (file && is_file_hugepages(file)) 1399 vm_flags |= VM_NORESERVE; 1400 } 1401 1402 addr = mmap_region(file, addr, len, vm_flags, pgoff); 1403 if (!IS_ERR_VALUE(addr) && 1404 ((vm_flags & VM_LOCKED) || 1405 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1406 *populate = len; 1407 return addr; 1408 } 1409 1410 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1411 unsigned long, prot, unsigned long, flags, 1412 unsigned long, fd, unsigned long, pgoff) 1413 { 1414 struct file *file = NULL; 1415 unsigned long retval = -EBADF; 1416 1417 if (!(flags & MAP_ANONYMOUS)) { 1418 audit_mmap_fd(fd, flags); 1419 file = fget(fd); 1420 if (!file) 1421 goto out; 1422 if (is_file_hugepages(file)) 1423 len = ALIGN(len, huge_page_size(hstate_file(file))); 1424 retval = -EINVAL; 1425 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file))) 1426 goto out_fput; 1427 } else if (flags & MAP_HUGETLB) { 1428 struct user_struct *user = NULL; 1429 struct hstate *hs; 1430 1431 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK); 1432 if (!hs) 1433 return -EINVAL; 1434 1435 len = ALIGN(len, huge_page_size(hs)); 1436 /* 1437 * VM_NORESERVE is used because the reservations will be 1438 * taken when vm_ops->mmap() is called 1439 * A dummy user value is used because we are not locking 1440 * memory so no accounting is necessary 1441 */ 1442 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1443 VM_NORESERVE, 1444 &user, HUGETLB_ANONHUGE_INODE, 1445 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1446 if (IS_ERR(file)) 1447 return PTR_ERR(file); 1448 } 1449 1450 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1451 1452 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1453 out_fput: 1454 if (file) 1455 fput(file); 1456 out: 1457 return retval; 1458 } 1459 1460 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1461 struct mmap_arg_struct { 1462 unsigned long addr; 1463 unsigned long len; 1464 unsigned long prot; 1465 unsigned long flags; 1466 unsigned long fd; 1467 unsigned long offset; 1468 }; 1469 1470 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1471 { 1472 struct mmap_arg_struct a; 1473 1474 if (copy_from_user(&a, arg, sizeof(a))) 1475 return -EFAULT; 1476 if (a.offset & ~PAGE_MASK) 1477 return -EINVAL; 1478 1479 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1480 a.offset >> PAGE_SHIFT); 1481 } 1482 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1483 1484 /* 1485 * Some shared mappigns will want the pages marked read-only 1486 * to track write events. If so, we'll downgrade vm_page_prot 1487 * to the private version (using protection_map[] without the 1488 * VM_SHARED bit). 1489 */ 1490 int vma_wants_writenotify(struct vm_area_struct *vma) 1491 { 1492 vm_flags_t vm_flags = vma->vm_flags; 1493 1494 /* If it was private or non-writable, the write bit is already clear */ 1495 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1496 return 0; 1497 1498 /* The backer wishes to know when pages are first written to? */ 1499 if (vma->vm_ops && vma->vm_ops->page_mkwrite) 1500 return 1; 1501 1502 /* The open routine did something to the protections that pgprot_modify 1503 * won't preserve? */ 1504 if (pgprot_val(vma->vm_page_prot) != 1505 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags))) 1506 return 0; 1507 1508 /* Do we need to track softdirty? */ 1509 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY)) 1510 return 1; 1511 1512 /* Specialty mapping? */ 1513 if (vm_flags & VM_PFNMAP) 1514 return 0; 1515 1516 /* Can the mapping track the dirty pages? */ 1517 return vma->vm_file && vma->vm_file->f_mapping && 1518 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1519 } 1520 1521 /* 1522 * We account for memory if it's a private writeable mapping, 1523 * not hugepages and VM_NORESERVE wasn't set. 1524 */ 1525 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1526 { 1527 /* 1528 * hugetlb has its own accounting separate from the core VM 1529 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1530 */ 1531 if (file && is_file_hugepages(file)) 1532 return 0; 1533 1534 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1535 } 1536 1537 unsigned long mmap_region(struct file *file, unsigned long addr, 1538 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff) 1539 { 1540 struct mm_struct *mm = current->mm; 1541 struct vm_area_struct *vma, *prev; 1542 int error; 1543 struct rb_node **rb_link, *rb_parent; 1544 unsigned long charged = 0; 1545 1546 /* Check against address space limit. */ 1547 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) { 1548 unsigned long nr_pages; 1549 1550 /* 1551 * MAP_FIXED may remove pages of mappings that intersects with 1552 * requested mapping. Account for the pages it would unmap. 1553 */ 1554 if (!(vm_flags & MAP_FIXED)) 1555 return -ENOMEM; 1556 1557 nr_pages = count_vma_pages_range(mm, addr, addr + len); 1558 1559 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages)) 1560 return -ENOMEM; 1561 } 1562 1563 /* Clear old maps */ 1564 error = -ENOMEM; 1565 munmap_back: 1566 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) { 1567 if (do_munmap(mm, addr, len)) 1568 return -ENOMEM; 1569 goto munmap_back; 1570 } 1571 1572 /* 1573 * Private writable mapping: check memory availability 1574 */ 1575 if (accountable_mapping(file, vm_flags)) { 1576 charged = len >> PAGE_SHIFT; 1577 if (security_vm_enough_memory_mm(mm, charged)) 1578 return -ENOMEM; 1579 vm_flags |= VM_ACCOUNT; 1580 } 1581 1582 /* 1583 * Can we just expand an old mapping? 1584 */ 1585 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL); 1586 if (vma) 1587 goto out; 1588 1589 /* 1590 * Determine the object being mapped and call the appropriate 1591 * specific mapper. the address has already been validated, but 1592 * not unmapped, but the maps are removed from the list. 1593 */ 1594 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1595 if (!vma) { 1596 error = -ENOMEM; 1597 goto unacct_error; 1598 } 1599 1600 vma->vm_mm = mm; 1601 vma->vm_start = addr; 1602 vma->vm_end = addr + len; 1603 vma->vm_flags = vm_flags; 1604 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1605 vma->vm_pgoff = pgoff; 1606 INIT_LIST_HEAD(&vma->anon_vma_chain); 1607 1608 if (file) { 1609 if (vm_flags & VM_DENYWRITE) { 1610 error = deny_write_access(file); 1611 if (error) 1612 goto free_vma; 1613 } 1614 if (vm_flags & VM_SHARED) { 1615 error = mapping_map_writable(file->f_mapping); 1616 if (error) 1617 goto allow_write_and_free_vma; 1618 } 1619 1620 /* ->mmap() can change vma->vm_file, but must guarantee that 1621 * vma_link() below can deny write-access if VM_DENYWRITE is set 1622 * and map writably if VM_SHARED is set. This usually means the 1623 * new file must not have been exposed to user-space, yet. 1624 */ 1625 vma->vm_file = get_file(file); 1626 error = file->f_op->mmap(file, vma); 1627 if (error) 1628 goto unmap_and_free_vma; 1629 1630 /* Can addr have changed?? 1631 * 1632 * Answer: Yes, several device drivers can do it in their 1633 * f_op->mmap method. -DaveM 1634 * Bug: If addr is changed, prev, rb_link, rb_parent should 1635 * be updated for vma_link() 1636 */ 1637 WARN_ON_ONCE(addr != vma->vm_start); 1638 1639 addr = vma->vm_start; 1640 vm_flags = vma->vm_flags; 1641 } else if (vm_flags & VM_SHARED) { 1642 error = shmem_zero_setup(vma); 1643 if (error) 1644 goto free_vma; 1645 } 1646 1647 vma_link(mm, vma, prev, rb_link, rb_parent); 1648 /* Once vma denies write, undo our temporary denial count */ 1649 if (file) { 1650 if (vm_flags & VM_SHARED) 1651 mapping_unmap_writable(file->f_mapping); 1652 if (vm_flags & VM_DENYWRITE) 1653 allow_write_access(file); 1654 } 1655 file = vma->vm_file; 1656 out: 1657 perf_event_mmap(vma); 1658 1659 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT); 1660 if (vm_flags & VM_LOCKED) { 1661 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) || 1662 vma == get_gate_vma(current->mm))) 1663 mm->locked_vm += (len >> PAGE_SHIFT); 1664 else 1665 vma->vm_flags &= ~VM_LOCKED; 1666 } 1667 1668 if (file) 1669 uprobe_mmap(vma); 1670 1671 /* 1672 * New (or expanded) vma always get soft dirty status. 1673 * Otherwise user-space soft-dirty page tracker won't 1674 * be able to distinguish situation when vma area unmapped, 1675 * then new mapped in-place (which must be aimed as 1676 * a completely new data area). 1677 */ 1678 vma->vm_flags |= VM_SOFTDIRTY; 1679 1680 vma_set_page_prot(vma); 1681 1682 return addr; 1683 1684 unmap_and_free_vma: 1685 vma->vm_file = NULL; 1686 fput(file); 1687 1688 /* Undo any partial mapping done by a device driver. */ 1689 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1690 charged = 0; 1691 if (vm_flags & VM_SHARED) 1692 mapping_unmap_writable(file->f_mapping); 1693 allow_write_and_free_vma: 1694 if (vm_flags & VM_DENYWRITE) 1695 allow_write_access(file); 1696 free_vma: 1697 kmem_cache_free(vm_area_cachep, vma); 1698 unacct_error: 1699 if (charged) 1700 vm_unacct_memory(charged); 1701 return error; 1702 } 1703 1704 unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1705 { 1706 /* 1707 * We implement the search by looking for an rbtree node that 1708 * immediately follows a suitable gap. That is, 1709 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length; 1710 * - gap_end = vma->vm_start >= info->low_limit + length; 1711 * - gap_end - gap_start >= length 1712 */ 1713 1714 struct mm_struct *mm = current->mm; 1715 struct vm_area_struct *vma; 1716 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1717 1718 /* Adjust search length to account for worst case alignment overhead */ 1719 length = info->length + info->align_mask; 1720 if (length < info->length) 1721 return -ENOMEM; 1722 1723 /* Adjust search limits by the desired length */ 1724 if (info->high_limit < length) 1725 return -ENOMEM; 1726 high_limit = info->high_limit - length; 1727 1728 if (info->low_limit > high_limit) 1729 return -ENOMEM; 1730 low_limit = info->low_limit + length; 1731 1732 /* Check if rbtree root looks promising */ 1733 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1734 goto check_highest; 1735 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1736 if (vma->rb_subtree_gap < length) 1737 goto check_highest; 1738 1739 while (true) { 1740 /* Visit left subtree if it looks promising */ 1741 gap_end = vma->vm_start; 1742 if (gap_end >= low_limit && vma->vm_rb.rb_left) { 1743 struct vm_area_struct *left = 1744 rb_entry(vma->vm_rb.rb_left, 1745 struct vm_area_struct, vm_rb); 1746 if (left->rb_subtree_gap >= length) { 1747 vma = left; 1748 continue; 1749 } 1750 } 1751 1752 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1753 check_current: 1754 /* Check if current node has a suitable gap */ 1755 if (gap_start > high_limit) 1756 return -ENOMEM; 1757 if (gap_end >= low_limit && gap_end - gap_start >= length) 1758 goto found; 1759 1760 /* Visit right subtree if it looks promising */ 1761 if (vma->vm_rb.rb_right) { 1762 struct vm_area_struct *right = 1763 rb_entry(vma->vm_rb.rb_right, 1764 struct vm_area_struct, vm_rb); 1765 if (right->rb_subtree_gap >= length) { 1766 vma = right; 1767 continue; 1768 } 1769 } 1770 1771 /* Go back up the rbtree to find next candidate node */ 1772 while (true) { 1773 struct rb_node *prev = &vma->vm_rb; 1774 if (!rb_parent(prev)) 1775 goto check_highest; 1776 vma = rb_entry(rb_parent(prev), 1777 struct vm_area_struct, vm_rb); 1778 if (prev == vma->vm_rb.rb_left) { 1779 gap_start = vma->vm_prev->vm_end; 1780 gap_end = vma->vm_start; 1781 goto check_current; 1782 } 1783 } 1784 } 1785 1786 check_highest: 1787 /* Check highest gap, which does not precede any rbtree node */ 1788 gap_start = mm->highest_vm_end; 1789 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */ 1790 if (gap_start > high_limit) 1791 return -ENOMEM; 1792 1793 found: 1794 /* We found a suitable gap. Clip it with the original low_limit. */ 1795 if (gap_start < info->low_limit) 1796 gap_start = info->low_limit; 1797 1798 /* Adjust gap address to the desired alignment */ 1799 gap_start += (info->align_offset - gap_start) & info->align_mask; 1800 1801 VM_BUG_ON(gap_start + info->length > info->high_limit); 1802 VM_BUG_ON(gap_start + info->length > gap_end); 1803 return gap_start; 1804 } 1805 1806 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1807 { 1808 struct mm_struct *mm = current->mm; 1809 struct vm_area_struct *vma; 1810 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1811 1812 /* Adjust search length to account for worst case alignment overhead */ 1813 length = info->length + info->align_mask; 1814 if (length < info->length) 1815 return -ENOMEM; 1816 1817 /* 1818 * Adjust search limits by the desired length. 1819 * See implementation comment at top of unmapped_area(). 1820 */ 1821 gap_end = info->high_limit; 1822 if (gap_end < length) 1823 return -ENOMEM; 1824 high_limit = gap_end - length; 1825 1826 if (info->low_limit > high_limit) 1827 return -ENOMEM; 1828 low_limit = info->low_limit + length; 1829 1830 /* Check highest gap, which does not precede any rbtree node */ 1831 gap_start = mm->highest_vm_end; 1832 if (gap_start <= high_limit) 1833 goto found_highest; 1834 1835 /* Check if rbtree root looks promising */ 1836 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1837 return -ENOMEM; 1838 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1839 if (vma->rb_subtree_gap < length) 1840 return -ENOMEM; 1841 1842 while (true) { 1843 /* Visit right subtree if it looks promising */ 1844 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1845 if (gap_start <= high_limit && vma->vm_rb.rb_right) { 1846 struct vm_area_struct *right = 1847 rb_entry(vma->vm_rb.rb_right, 1848 struct vm_area_struct, vm_rb); 1849 if (right->rb_subtree_gap >= length) { 1850 vma = right; 1851 continue; 1852 } 1853 } 1854 1855 check_current: 1856 /* Check if current node has a suitable gap */ 1857 gap_end = vma->vm_start; 1858 if (gap_end < low_limit) 1859 return -ENOMEM; 1860 if (gap_start <= high_limit && gap_end - gap_start >= length) 1861 goto found; 1862 1863 /* Visit left subtree if it looks promising */ 1864 if (vma->vm_rb.rb_left) { 1865 struct vm_area_struct *left = 1866 rb_entry(vma->vm_rb.rb_left, 1867 struct vm_area_struct, vm_rb); 1868 if (left->rb_subtree_gap >= length) { 1869 vma = left; 1870 continue; 1871 } 1872 } 1873 1874 /* Go back up the rbtree to find next candidate node */ 1875 while (true) { 1876 struct rb_node *prev = &vma->vm_rb; 1877 if (!rb_parent(prev)) 1878 return -ENOMEM; 1879 vma = rb_entry(rb_parent(prev), 1880 struct vm_area_struct, vm_rb); 1881 if (prev == vma->vm_rb.rb_right) { 1882 gap_start = vma->vm_prev ? 1883 vma->vm_prev->vm_end : 0; 1884 goto check_current; 1885 } 1886 } 1887 } 1888 1889 found: 1890 /* We found a suitable gap. Clip it with the original high_limit. */ 1891 if (gap_end > info->high_limit) 1892 gap_end = info->high_limit; 1893 1894 found_highest: 1895 /* Compute highest gap address at the desired alignment */ 1896 gap_end -= info->length; 1897 gap_end -= (gap_end - info->align_offset) & info->align_mask; 1898 1899 VM_BUG_ON(gap_end < info->low_limit); 1900 VM_BUG_ON(gap_end < gap_start); 1901 return gap_end; 1902 } 1903 1904 /* Get an address range which is currently unmapped. 1905 * For shmat() with addr=0. 1906 * 1907 * Ugly calling convention alert: 1908 * Return value with the low bits set means error value, 1909 * ie 1910 * if (ret & ~PAGE_MASK) 1911 * error = ret; 1912 * 1913 * This function "knows" that -ENOMEM has the bits set. 1914 */ 1915 #ifndef HAVE_ARCH_UNMAPPED_AREA 1916 unsigned long 1917 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1918 unsigned long len, unsigned long pgoff, unsigned long flags) 1919 { 1920 struct mm_struct *mm = current->mm; 1921 struct vm_area_struct *vma; 1922 struct vm_unmapped_area_info info; 1923 1924 if (len > TASK_SIZE - mmap_min_addr) 1925 return -ENOMEM; 1926 1927 if (flags & MAP_FIXED) 1928 return addr; 1929 1930 if (addr) { 1931 addr = PAGE_ALIGN(addr); 1932 vma = find_vma(mm, addr); 1933 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1934 (!vma || addr + len <= vma->vm_start)) 1935 return addr; 1936 } 1937 1938 info.flags = 0; 1939 info.length = len; 1940 info.low_limit = mm->mmap_base; 1941 info.high_limit = TASK_SIZE; 1942 info.align_mask = 0; 1943 return vm_unmapped_area(&info); 1944 } 1945 #endif 1946 1947 /* 1948 * This mmap-allocator allocates new areas top-down from below the 1949 * stack's low limit (the base): 1950 */ 1951 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1952 unsigned long 1953 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1954 const unsigned long len, const unsigned long pgoff, 1955 const unsigned long flags) 1956 { 1957 struct vm_area_struct *vma; 1958 struct mm_struct *mm = current->mm; 1959 unsigned long addr = addr0; 1960 struct vm_unmapped_area_info info; 1961 1962 /* requested length too big for entire address space */ 1963 if (len > TASK_SIZE - mmap_min_addr) 1964 return -ENOMEM; 1965 1966 if (flags & MAP_FIXED) 1967 return addr; 1968 1969 /* requesting a specific address */ 1970 if (addr) { 1971 addr = PAGE_ALIGN(addr); 1972 vma = find_vma(mm, addr); 1973 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1974 (!vma || addr + len <= vma->vm_start)) 1975 return addr; 1976 } 1977 1978 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 1979 info.length = len; 1980 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 1981 info.high_limit = mm->mmap_base; 1982 info.align_mask = 0; 1983 addr = vm_unmapped_area(&info); 1984 1985 /* 1986 * A failed mmap() very likely causes application failure, 1987 * so fall back to the bottom-up function here. This scenario 1988 * can happen with large stack limits and large mmap() 1989 * allocations. 1990 */ 1991 if (addr & ~PAGE_MASK) { 1992 VM_BUG_ON(addr != -ENOMEM); 1993 info.flags = 0; 1994 info.low_limit = TASK_UNMAPPED_BASE; 1995 info.high_limit = TASK_SIZE; 1996 addr = vm_unmapped_area(&info); 1997 } 1998 1999 return addr; 2000 } 2001 #endif 2002 2003 unsigned long 2004 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 2005 unsigned long pgoff, unsigned long flags) 2006 { 2007 unsigned long (*get_area)(struct file *, unsigned long, 2008 unsigned long, unsigned long, unsigned long); 2009 2010 unsigned long error = arch_mmap_check(addr, len, flags); 2011 if (error) 2012 return error; 2013 2014 /* Careful about overflows.. */ 2015 if (len > TASK_SIZE) 2016 return -ENOMEM; 2017 2018 get_area = current->mm->get_unmapped_area; 2019 if (file && file->f_op->get_unmapped_area) 2020 get_area = file->f_op->get_unmapped_area; 2021 addr = get_area(file, addr, len, pgoff, flags); 2022 if (IS_ERR_VALUE(addr)) 2023 return addr; 2024 2025 if (addr > TASK_SIZE - len) 2026 return -ENOMEM; 2027 if (addr & ~PAGE_MASK) 2028 return -EINVAL; 2029 2030 addr = arch_rebalance_pgtables(addr, len); 2031 error = security_mmap_addr(addr); 2032 return error ? error : addr; 2033 } 2034 2035 EXPORT_SYMBOL(get_unmapped_area); 2036 2037 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 2038 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 2039 { 2040 struct rb_node *rb_node; 2041 struct vm_area_struct *vma; 2042 2043 /* Check the cache first. */ 2044 vma = vmacache_find(mm, addr); 2045 if (likely(vma)) 2046 return vma; 2047 2048 rb_node = mm->mm_rb.rb_node; 2049 vma = NULL; 2050 2051 while (rb_node) { 2052 struct vm_area_struct *tmp; 2053 2054 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2055 2056 if (tmp->vm_end > addr) { 2057 vma = tmp; 2058 if (tmp->vm_start <= addr) 2059 break; 2060 rb_node = rb_node->rb_left; 2061 } else 2062 rb_node = rb_node->rb_right; 2063 } 2064 2065 if (vma) 2066 vmacache_update(addr, vma); 2067 return vma; 2068 } 2069 2070 EXPORT_SYMBOL(find_vma); 2071 2072 /* 2073 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 2074 */ 2075 struct vm_area_struct * 2076 find_vma_prev(struct mm_struct *mm, unsigned long addr, 2077 struct vm_area_struct **pprev) 2078 { 2079 struct vm_area_struct *vma; 2080 2081 vma = find_vma(mm, addr); 2082 if (vma) { 2083 *pprev = vma->vm_prev; 2084 } else { 2085 struct rb_node *rb_node = mm->mm_rb.rb_node; 2086 *pprev = NULL; 2087 while (rb_node) { 2088 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2089 rb_node = rb_node->rb_right; 2090 } 2091 } 2092 return vma; 2093 } 2094 2095 /* 2096 * Verify that the stack growth is acceptable and 2097 * update accounting. This is shared with both the 2098 * grow-up and grow-down cases. 2099 */ 2100 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 2101 { 2102 struct mm_struct *mm = vma->vm_mm; 2103 struct rlimit *rlim = current->signal->rlim; 2104 unsigned long new_start, actual_size; 2105 2106 /* address space limit tests */ 2107 if (!may_expand_vm(mm, grow)) 2108 return -ENOMEM; 2109 2110 /* Stack limit test */ 2111 actual_size = size; 2112 if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN))) 2113 actual_size -= PAGE_SIZE; 2114 if (actual_size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 2115 return -ENOMEM; 2116 2117 /* mlock limit tests */ 2118 if (vma->vm_flags & VM_LOCKED) { 2119 unsigned long locked; 2120 unsigned long limit; 2121 locked = mm->locked_vm + grow; 2122 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 2123 limit >>= PAGE_SHIFT; 2124 if (locked > limit && !capable(CAP_IPC_LOCK)) 2125 return -ENOMEM; 2126 } 2127 2128 /* Check to ensure the stack will not grow into a hugetlb-only region */ 2129 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 2130 vma->vm_end - size; 2131 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 2132 return -EFAULT; 2133 2134 /* 2135 * Overcommit.. This must be the final test, as it will 2136 * update security statistics. 2137 */ 2138 if (security_vm_enough_memory_mm(mm, grow)) 2139 return -ENOMEM; 2140 2141 /* Ok, everything looks good - let it rip */ 2142 if (vma->vm_flags & VM_LOCKED) 2143 mm->locked_vm += grow; 2144 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow); 2145 return 0; 2146 } 2147 2148 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 2149 /* 2150 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 2151 * vma is the last one with address > vma->vm_end. Have to extend vma. 2152 */ 2153 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 2154 { 2155 int error; 2156 2157 if (!(vma->vm_flags & VM_GROWSUP)) 2158 return -EFAULT; 2159 2160 /* 2161 * We must make sure the anon_vma is allocated 2162 * so that the anon_vma locking is not a noop. 2163 */ 2164 if (unlikely(anon_vma_prepare(vma))) 2165 return -ENOMEM; 2166 vma_lock_anon_vma(vma); 2167 2168 /* 2169 * vma->vm_start/vm_end cannot change under us because the caller 2170 * is required to hold the mmap_sem in read mode. We need the 2171 * anon_vma lock to serialize against concurrent expand_stacks. 2172 * Also guard against wrapping around to address 0. 2173 */ 2174 if (address < PAGE_ALIGN(address+4)) 2175 address = PAGE_ALIGN(address+4); 2176 else { 2177 vma_unlock_anon_vma(vma); 2178 return -ENOMEM; 2179 } 2180 error = 0; 2181 2182 /* Somebody else might have raced and expanded it already */ 2183 if (address > vma->vm_end) { 2184 unsigned long size, grow; 2185 2186 size = address - vma->vm_start; 2187 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2188 2189 error = -ENOMEM; 2190 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2191 error = acct_stack_growth(vma, size, grow); 2192 if (!error) { 2193 /* 2194 * vma_gap_update() doesn't support concurrent 2195 * updates, but we only hold a shared mmap_sem 2196 * lock here, so we need to protect against 2197 * concurrent vma expansions. 2198 * vma_lock_anon_vma() doesn't help here, as 2199 * we don't guarantee that all growable vmas 2200 * in a mm share the same root anon vma. 2201 * So, we reuse mm->page_table_lock to guard 2202 * against concurrent vma expansions. 2203 */ 2204 spin_lock(&vma->vm_mm->page_table_lock); 2205 anon_vma_interval_tree_pre_update_vma(vma); 2206 vma->vm_end = address; 2207 anon_vma_interval_tree_post_update_vma(vma); 2208 if (vma->vm_next) 2209 vma_gap_update(vma->vm_next); 2210 else 2211 vma->vm_mm->highest_vm_end = address; 2212 spin_unlock(&vma->vm_mm->page_table_lock); 2213 2214 perf_event_mmap(vma); 2215 } 2216 } 2217 } 2218 vma_unlock_anon_vma(vma); 2219 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2220 validate_mm(vma->vm_mm); 2221 return error; 2222 } 2223 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2224 2225 /* 2226 * vma is the first one with address < vma->vm_start. Have to extend vma. 2227 */ 2228 int expand_downwards(struct vm_area_struct *vma, 2229 unsigned long address) 2230 { 2231 int error; 2232 2233 /* 2234 * We must make sure the anon_vma is allocated 2235 * so that the anon_vma locking is not a noop. 2236 */ 2237 if (unlikely(anon_vma_prepare(vma))) 2238 return -ENOMEM; 2239 2240 address &= PAGE_MASK; 2241 error = security_mmap_addr(address); 2242 if (error) 2243 return error; 2244 2245 vma_lock_anon_vma(vma); 2246 2247 /* 2248 * vma->vm_start/vm_end cannot change under us because the caller 2249 * is required to hold the mmap_sem in read mode. We need the 2250 * anon_vma lock to serialize against concurrent expand_stacks. 2251 */ 2252 2253 /* Somebody else might have raced and expanded it already */ 2254 if (address < vma->vm_start) { 2255 unsigned long size, grow; 2256 2257 size = vma->vm_end - address; 2258 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2259 2260 error = -ENOMEM; 2261 if (grow <= vma->vm_pgoff) { 2262 error = acct_stack_growth(vma, size, grow); 2263 if (!error) { 2264 /* 2265 * vma_gap_update() doesn't support concurrent 2266 * updates, but we only hold a shared mmap_sem 2267 * lock here, so we need to protect against 2268 * concurrent vma expansions. 2269 * vma_lock_anon_vma() doesn't help here, as 2270 * we don't guarantee that all growable vmas 2271 * in a mm share the same root anon vma. 2272 * So, we reuse mm->page_table_lock to guard 2273 * against concurrent vma expansions. 2274 */ 2275 spin_lock(&vma->vm_mm->page_table_lock); 2276 anon_vma_interval_tree_pre_update_vma(vma); 2277 vma->vm_start = address; 2278 vma->vm_pgoff -= grow; 2279 anon_vma_interval_tree_post_update_vma(vma); 2280 vma_gap_update(vma); 2281 spin_unlock(&vma->vm_mm->page_table_lock); 2282 2283 perf_event_mmap(vma); 2284 } 2285 } 2286 } 2287 vma_unlock_anon_vma(vma); 2288 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2289 validate_mm(vma->vm_mm); 2290 return error; 2291 } 2292 2293 /* 2294 * Note how expand_stack() refuses to expand the stack all the way to 2295 * abut the next virtual mapping, *unless* that mapping itself is also 2296 * a stack mapping. We want to leave room for a guard page, after all 2297 * (the guard page itself is not added here, that is done by the 2298 * actual page faulting logic) 2299 * 2300 * This matches the behavior of the guard page logic (see mm/memory.c: 2301 * check_stack_guard_page()), which only allows the guard page to be 2302 * removed under these circumstances. 2303 */ 2304 #ifdef CONFIG_STACK_GROWSUP 2305 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2306 { 2307 struct vm_area_struct *next; 2308 2309 address &= PAGE_MASK; 2310 next = vma->vm_next; 2311 if (next && next->vm_start == address + PAGE_SIZE) { 2312 if (!(next->vm_flags & VM_GROWSUP)) 2313 return -ENOMEM; 2314 } 2315 return expand_upwards(vma, address); 2316 } 2317 2318 struct vm_area_struct * 2319 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2320 { 2321 struct vm_area_struct *vma, *prev; 2322 2323 addr &= PAGE_MASK; 2324 vma = find_vma_prev(mm, addr, &prev); 2325 if (vma && (vma->vm_start <= addr)) 2326 return vma; 2327 if (!prev || expand_stack(prev, addr)) 2328 return NULL; 2329 if (prev->vm_flags & VM_LOCKED) 2330 __mlock_vma_pages_range(prev, addr, prev->vm_end, NULL); 2331 return prev; 2332 } 2333 #else 2334 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2335 { 2336 struct vm_area_struct *prev; 2337 2338 address &= PAGE_MASK; 2339 prev = vma->vm_prev; 2340 if (prev && prev->vm_end == address) { 2341 if (!(prev->vm_flags & VM_GROWSDOWN)) 2342 return -ENOMEM; 2343 } 2344 return expand_downwards(vma, address); 2345 } 2346 2347 struct vm_area_struct * 2348 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2349 { 2350 struct vm_area_struct *vma; 2351 unsigned long start; 2352 2353 addr &= PAGE_MASK; 2354 vma = find_vma(mm, addr); 2355 if (!vma) 2356 return NULL; 2357 if (vma->vm_start <= addr) 2358 return vma; 2359 if (!(vma->vm_flags & VM_GROWSDOWN)) 2360 return NULL; 2361 start = vma->vm_start; 2362 if (expand_stack(vma, addr)) 2363 return NULL; 2364 if (vma->vm_flags & VM_LOCKED) 2365 __mlock_vma_pages_range(vma, addr, start, NULL); 2366 return vma; 2367 } 2368 #endif 2369 2370 EXPORT_SYMBOL_GPL(find_extend_vma); 2371 2372 /* 2373 * Ok - we have the memory areas we should free on the vma list, 2374 * so release them, and do the vma updates. 2375 * 2376 * Called with the mm semaphore held. 2377 */ 2378 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 2379 { 2380 unsigned long nr_accounted = 0; 2381 2382 /* Update high watermark before we lower total_vm */ 2383 update_hiwater_vm(mm); 2384 do { 2385 long nrpages = vma_pages(vma); 2386 2387 if (vma->vm_flags & VM_ACCOUNT) 2388 nr_accounted += nrpages; 2389 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages); 2390 vma = remove_vma(vma); 2391 } while (vma); 2392 vm_unacct_memory(nr_accounted); 2393 validate_mm(mm); 2394 } 2395 2396 /* 2397 * Get rid of page table information in the indicated region. 2398 * 2399 * Called with the mm semaphore held. 2400 */ 2401 static void unmap_region(struct mm_struct *mm, 2402 struct vm_area_struct *vma, struct vm_area_struct *prev, 2403 unsigned long start, unsigned long end) 2404 { 2405 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap; 2406 struct mmu_gather tlb; 2407 2408 lru_add_drain(); 2409 tlb_gather_mmu(&tlb, mm, start, end); 2410 update_hiwater_rss(mm); 2411 unmap_vmas(&tlb, vma, start, end); 2412 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2413 next ? next->vm_start : USER_PGTABLES_CEILING); 2414 tlb_finish_mmu(&tlb, start, end); 2415 } 2416 2417 /* 2418 * Create a list of vma's touched by the unmap, removing them from the mm's 2419 * vma list as we go.. 2420 */ 2421 static void 2422 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 2423 struct vm_area_struct *prev, unsigned long end) 2424 { 2425 struct vm_area_struct **insertion_point; 2426 struct vm_area_struct *tail_vma = NULL; 2427 2428 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 2429 vma->vm_prev = NULL; 2430 do { 2431 vma_rb_erase(vma, &mm->mm_rb); 2432 mm->map_count--; 2433 tail_vma = vma; 2434 vma = vma->vm_next; 2435 } while (vma && vma->vm_start < end); 2436 *insertion_point = vma; 2437 if (vma) { 2438 vma->vm_prev = prev; 2439 vma_gap_update(vma); 2440 } else 2441 mm->highest_vm_end = prev ? prev->vm_end : 0; 2442 tail_vma->vm_next = NULL; 2443 2444 /* Kill the cache */ 2445 vmacache_invalidate(mm); 2446 } 2447 2448 /* 2449 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the 2450 * munmap path where it doesn't make sense to fail. 2451 */ 2452 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2453 unsigned long addr, int new_below) 2454 { 2455 struct vm_area_struct *new; 2456 int err = -ENOMEM; 2457 2458 if (is_vm_hugetlb_page(vma) && (addr & 2459 ~(huge_page_mask(hstate_vma(vma))))) 2460 return -EINVAL; 2461 2462 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2463 if (!new) 2464 goto out_err; 2465 2466 /* most fields are the same, copy all, and then fixup */ 2467 *new = *vma; 2468 2469 INIT_LIST_HEAD(&new->anon_vma_chain); 2470 2471 if (new_below) 2472 new->vm_end = addr; 2473 else { 2474 new->vm_start = addr; 2475 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2476 } 2477 2478 err = vma_dup_policy(vma, new); 2479 if (err) 2480 goto out_free_vma; 2481 2482 err = anon_vma_clone(new, vma); 2483 if (err) 2484 goto out_free_mpol; 2485 2486 if (new->vm_file) 2487 get_file(new->vm_file); 2488 2489 if (new->vm_ops && new->vm_ops->open) 2490 new->vm_ops->open(new); 2491 2492 if (new_below) 2493 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 2494 ((addr - new->vm_start) >> PAGE_SHIFT), new); 2495 else 2496 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 2497 2498 /* Success. */ 2499 if (!err) 2500 return 0; 2501 2502 /* Clean everything up if vma_adjust failed. */ 2503 if (new->vm_ops && new->vm_ops->close) 2504 new->vm_ops->close(new); 2505 if (new->vm_file) 2506 fput(new->vm_file); 2507 unlink_anon_vmas(new); 2508 out_free_mpol: 2509 mpol_put(vma_policy(new)); 2510 out_free_vma: 2511 kmem_cache_free(vm_area_cachep, new); 2512 out_err: 2513 return err; 2514 } 2515 2516 /* 2517 * Split a vma into two pieces at address 'addr', a new vma is allocated 2518 * either for the first part or the tail. 2519 */ 2520 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2521 unsigned long addr, int new_below) 2522 { 2523 if (mm->map_count >= sysctl_max_map_count) 2524 return -ENOMEM; 2525 2526 return __split_vma(mm, vma, addr, new_below); 2527 } 2528 2529 /* Munmap is split into 2 main parts -- this part which finds 2530 * what needs doing, and the areas themselves, which do the 2531 * work. This now handles partial unmappings. 2532 * Jeremy Fitzhardinge <jeremy@goop.org> 2533 */ 2534 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 2535 { 2536 unsigned long end; 2537 struct vm_area_struct *vma, *prev, *last; 2538 2539 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start) 2540 return -EINVAL; 2541 2542 len = PAGE_ALIGN(len); 2543 if (len == 0) 2544 return -EINVAL; 2545 2546 /* Find the first overlapping VMA */ 2547 vma = find_vma(mm, start); 2548 if (!vma) 2549 return 0; 2550 prev = vma->vm_prev; 2551 /* we have start < vma->vm_end */ 2552 2553 /* if it doesn't overlap, we have nothing.. */ 2554 end = start + len; 2555 if (vma->vm_start >= end) 2556 return 0; 2557 2558 /* 2559 * If we need to split any vma, do it now to save pain later. 2560 * 2561 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2562 * unmapped vm_area_struct will remain in use: so lower split_vma 2563 * places tmp vma above, and higher split_vma places tmp vma below. 2564 */ 2565 if (start > vma->vm_start) { 2566 int error; 2567 2568 /* 2569 * Make sure that map_count on return from munmap() will 2570 * not exceed its limit; but let map_count go just above 2571 * its limit temporarily, to help free resources as expected. 2572 */ 2573 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2574 return -ENOMEM; 2575 2576 error = __split_vma(mm, vma, start, 0); 2577 if (error) 2578 return error; 2579 prev = vma; 2580 } 2581 2582 /* Does it split the last one? */ 2583 last = find_vma(mm, end); 2584 if (last && end > last->vm_start) { 2585 int error = __split_vma(mm, last, end, 1); 2586 if (error) 2587 return error; 2588 } 2589 vma = prev ? prev->vm_next : mm->mmap; 2590 2591 /* 2592 * unlock any mlock()ed ranges before detaching vmas 2593 */ 2594 if (mm->locked_vm) { 2595 struct vm_area_struct *tmp = vma; 2596 while (tmp && tmp->vm_start < end) { 2597 if (tmp->vm_flags & VM_LOCKED) { 2598 mm->locked_vm -= vma_pages(tmp); 2599 munlock_vma_pages_all(tmp); 2600 } 2601 tmp = tmp->vm_next; 2602 } 2603 } 2604 2605 /* 2606 * Remove the vma's, and unmap the actual pages 2607 */ 2608 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2609 unmap_region(mm, vma, prev, start, end); 2610 2611 arch_unmap(mm, vma, start, end); 2612 2613 /* Fix up all other VM information */ 2614 remove_vma_list(mm, vma); 2615 2616 return 0; 2617 } 2618 2619 int vm_munmap(unsigned long start, size_t len) 2620 { 2621 int ret; 2622 struct mm_struct *mm = current->mm; 2623 2624 down_write(&mm->mmap_sem); 2625 ret = do_munmap(mm, start, len); 2626 up_write(&mm->mmap_sem); 2627 return ret; 2628 } 2629 EXPORT_SYMBOL(vm_munmap); 2630 2631 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2632 { 2633 profile_munmap(addr); 2634 return vm_munmap(addr, len); 2635 } 2636 2637 static inline void verify_mm_writelocked(struct mm_struct *mm) 2638 { 2639 #ifdef CONFIG_DEBUG_VM 2640 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2641 WARN_ON(1); 2642 up_read(&mm->mmap_sem); 2643 } 2644 #endif 2645 } 2646 2647 /* 2648 * this is really a simplified "do_mmap". it only handles 2649 * anonymous maps. eventually we may be able to do some 2650 * brk-specific accounting here. 2651 */ 2652 static unsigned long do_brk(unsigned long addr, unsigned long len) 2653 { 2654 struct mm_struct *mm = current->mm; 2655 struct vm_area_struct *vma, *prev; 2656 unsigned long flags; 2657 struct rb_node **rb_link, *rb_parent; 2658 pgoff_t pgoff = addr >> PAGE_SHIFT; 2659 int error; 2660 2661 len = PAGE_ALIGN(len); 2662 if (!len) 2663 return addr; 2664 2665 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2666 2667 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2668 if (error & ~PAGE_MASK) 2669 return error; 2670 2671 error = mlock_future_check(mm, mm->def_flags, len); 2672 if (error) 2673 return error; 2674 2675 /* 2676 * mm->mmap_sem is required to protect against another thread 2677 * changing the mappings in case we sleep. 2678 */ 2679 verify_mm_writelocked(mm); 2680 2681 /* 2682 * Clear old maps. this also does some error checking for us 2683 */ 2684 munmap_back: 2685 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) { 2686 if (do_munmap(mm, addr, len)) 2687 return -ENOMEM; 2688 goto munmap_back; 2689 } 2690 2691 /* Check against address space limits *after* clearing old maps... */ 2692 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 2693 return -ENOMEM; 2694 2695 if (mm->map_count > sysctl_max_map_count) 2696 return -ENOMEM; 2697 2698 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 2699 return -ENOMEM; 2700 2701 /* Can we just expand an old private anonymous mapping? */ 2702 vma = vma_merge(mm, prev, addr, addr + len, flags, 2703 NULL, NULL, pgoff, NULL); 2704 if (vma) 2705 goto out; 2706 2707 /* 2708 * create a vma struct for an anonymous mapping 2709 */ 2710 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2711 if (!vma) { 2712 vm_unacct_memory(len >> PAGE_SHIFT); 2713 return -ENOMEM; 2714 } 2715 2716 INIT_LIST_HEAD(&vma->anon_vma_chain); 2717 vma->vm_mm = mm; 2718 vma->vm_start = addr; 2719 vma->vm_end = addr + len; 2720 vma->vm_pgoff = pgoff; 2721 vma->vm_flags = flags; 2722 vma->vm_page_prot = vm_get_page_prot(flags); 2723 vma_link(mm, vma, prev, rb_link, rb_parent); 2724 out: 2725 perf_event_mmap(vma); 2726 mm->total_vm += len >> PAGE_SHIFT; 2727 if (flags & VM_LOCKED) 2728 mm->locked_vm += (len >> PAGE_SHIFT); 2729 vma->vm_flags |= VM_SOFTDIRTY; 2730 return addr; 2731 } 2732 2733 unsigned long vm_brk(unsigned long addr, unsigned long len) 2734 { 2735 struct mm_struct *mm = current->mm; 2736 unsigned long ret; 2737 bool populate; 2738 2739 down_write(&mm->mmap_sem); 2740 ret = do_brk(addr, len); 2741 populate = ((mm->def_flags & VM_LOCKED) != 0); 2742 up_write(&mm->mmap_sem); 2743 if (populate) 2744 mm_populate(addr, len); 2745 return ret; 2746 } 2747 EXPORT_SYMBOL(vm_brk); 2748 2749 /* Release all mmaps. */ 2750 void exit_mmap(struct mm_struct *mm) 2751 { 2752 struct mmu_gather tlb; 2753 struct vm_area_struct *vma; 2754 unsigned long nr_accounted = 0; 2755 2756 /* mm's last user has gone, and its about to be pulled down */ 2757 mmu_notifier_release(mm); 2758 2759 if (mm->locked_vm) { 2760 vma = mm->mmap; 2761 while (vma) { 2762 if (vma->vm_flags & VM_LOCKED) 2763 munlock_vma_pages_all(vma); 2764 vma = vma->vm_next; 2765 } 2766 } 2767 2768 arch_exit_mmap(mm); 2769 2770 vma = mm->mmap; 2771 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2772 return; 2773 2774 lru_add_drain(); 2775 flush_cache_mm(mm); 2776 tlb_gather_mmu(&tlb, mm, 0, -1); 2777 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2778 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2779 unmap_vmas(&tlb, vma, 0, -1); 2780 2781 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING); 2782 tlb_finish_mmu(&tlb, 0, -1); 2783 2784 /* 2785 * Walk the list again, actually closing and freeing it, 2786 * with preemption enabled, without holding any MM locks. 2787 */ 2788 while (vma) { 2789 if (vma->vm_flags & VM_ACCOUNT) 2790 nr_accounted += vma_pages(vma); 2791 vma = remove_vma(vma); 2792 } 2793 vm_unacct_memory(nr_accounted); 2794 2795 WARN_ON(atomic_long_read(&mm->nr_ptes) > 2796 (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT); 2797 } 2798 2799 /* Insert vm structure into process list sorted by address 2800 * and into the inode's i_mmap tree. If vm_file is non-NULL 2801 * then i_mmap_rwsem is taken here. 2802 */ 2803 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 2804 { 2805 struct vm_area_struct *prev; 2806 struct rb_node **rb_link, *rb_parent; 2807 2808 /* 2809 * The vm_pgoff of a purely anonymous vma should be irrelevant 2810 * until its first write fault, when page's anon_vma and index 2811 * are set. But now set the vm_pgoff it will almost certainly 2812 * end up with (unless mremap moves it elsewhere before that 2813 * first wfault), so /proc/pid/maps tells a consistent story. 2814 * 2815 * By setting it to reflect the virtual start address of the 2816 * vma, merges and splits can happen in a seamless way, just 2817 * using the existing file pgoff checks and manipulations. 2818 * Similarly in do_mmap_pgoff and in do_brk. 2819 */ 2820 if (!vma->vm_file) { 2821 BUG_ON(vma->anon_vma); 2822 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2823 } 2824 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 2825 &prev, &rb_link, &rb_parent)) 2826 return -ENOMEM; 2827 if ((vma->vm_flags & VM_ACCOUNT) && 2828 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2829 return -ENOMEM; 2830 2831 vma_link(mm, vma, prev, rb_link, rb_parent); 2832 return 0; 2833 } 2834 2835 /* 2836 * Copy the vma structure to a new location in the same mm, 2837 * prior to moving page table entries, to effect an mremap move. 2838 */ 2839 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2840 unsigned long addr, unsigned long len, pgoff_t pgoff, 2841 bool *need_rmap_locks) 2842 { 2843 struct vm_area_struct *vma = *vmap; 2844 unsigned long vma_start = vma->vm_start; 2845 struct mm_struct *mm = vma->vm_mm; 2846 struct vm_area_struct *new_vma, *prev; 2847 struct rb_node **rb_link, *rb_parent; 2848 bool faulted_in_anon_vma = true; 2849 2850 /* 2851 * If anonymous vma has not yet been faulted, update new pgoff 2852 * to match new location, to increase its chance of merging. 2853 */ 2854 if (unlikely(!vma->vm_file && !vma->anon_vma)) { 2855 pgoff = addr >> PAGE_SHIFT; 2856 faulted_in_anon_vma = false; 2857 } 2858 2859 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) 2860 return NULL; /* should never get here */ 2861 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2862 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma)); 2863 if (new_vma) { 2864 /* 2865 * Source vma may have been merged into new_vma 2866 */ 2867 if (unlikely(vma_start >= new_vma->vm_start && 2868 vma_start < new_vma->vm_end)) { 2869 /* 2870 * The only way we can get a vma_merge with 2871 * self during an mremap is if the vma hasn't 2872 * been faulted in yet and we were allowed to 2873 * reset the dst vma->vm_pgoff to the 2874 * destination address of the mremap to allow 2875 * the merge to happen. mremap must change the 2876 * vm_pgoff linearity between src and dst vmas 2877 * (in turn preventing a vma_merge) to be 2878 * safe. It is only safe to keep the vm_pgoff 2879 * linear if there are no pages mapped yet. 2880 */ 2881 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 2882 *vmap = vma = new_vma; 2883 } 2884 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 2885 } else { 2886 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2887 if (new_vma) { 2888 *new_vma = *vma; 2889 new_vma->vm_start = addr; 2890 new_vma->vm_end = addr + len; 2891 new_vma->vm_pgoff = pgoff; 2892 if (vma_dup_policy(vma, new_vma)) 2893 goto out_free_vma; 2894 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 2895 if (anon_vma_clone(new_vma, vma)) 2896 goto out_free_mempol; 2897 if (new_vma->vm_file) 2898 get_file(new_vma->vm_file); 2899 if (new_vma->vm_ops && new_vma->vm_ops->open) 2900 new_vma->vm_ops->open(new_vma); 2901 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2902 *need_rmap_locks = false; 2903 } 2904 } 2905 return new_vma; 2906 2907 out_free_mempol: 2908 mpol_put(vma_policy(new_vma)); 2909 out_free_vma: 2910 kmem_cache_free(vm_area_cachep, new_vma); 2911 return NULL; 2912 } 2913 2914 /* 2915 * Return true if the calling process may expand its vm space by the passed 2916 * number of pages 2917 */ 2918 int may_expand_vm(struct mm_struct *mm, unsigned long npages) 2919 { 2920 unsigned long cur = mm->total_vm; /* pages */ 2921 unsigned long lim; 2922 2923 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT; 2924 2925 if (cur + npages > lim) 2926 return 0; 2927 return 1; 2928 } 2929 2930 static int special_mapping_fault(struct vm_area_struct *vma, 2931 struct vm_fault *vmf); 2932 2933 /* 2934 * Having a close hook prevents vma merging regardless of flags. 2935 */ 2936 static void special_mapping_close(struct vm_area_struct *vma) 2937 { 2938 } 2939 2940 static const char *special_mapping_name(struct vm_area_struct *vma) 2941 { 2942 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 2943 } 2944 2945 static const struct vm_operations_struct special_mapping_vmops = { 2946 .close = special_mapping_close, 2947 .fault = special_mapping_fault, 2948 .name = special_mapping_name, 2949 }; 2950 2951 static const struct vm_operations_struct legacy_special_mapping_vmops = { 2952 .close = special_mapping_close, 2953 .fault = special_mapping_fault, 2954 }; 2955 2956 static int special_mapping_fault(struct vm_area_struct *vma, 2957 struct vm_fault *vmf) 2958 { 2959 pgoff_t pgoff; 2960 struct page **pages; 2961 2962 /* 2963 * special mappings have no vm_file, and in that case, the mm 2964 * uses vm_pgoff internally. So we have to subtract it from here. 2965 * We are allowed to do this because we are the mm; do not copy 2966 * this code into drivers! 2967 */ 2968 pgoff = vmf->pgoff - vma->vm_pgoff; 2969 2970 if (vma->vm_ops == &legacy_special_mapping_vmops) 2971 pages = vma->vm_private_data; 2972 else 2973 pages = ((struct vm_special_mapping *)vma->vm_private_data)-> 2974 pages; 2975 2976 for (; pgoff && *pages; ++pages) 2977 pgoff--; 2978 2979 if (*pages) { 2980 struct page *page = *pages; 2981 get_page(page); 2982 vmf->page = page; 2983 return 0; 2984 } 2985 2986 return VM_FAULT_SIGBUS; 2987 } 2988 2989 static struct vm_area_struct *__install_special_mapping( 2990 struct mm_struct *mm, 2991 unsigned long addr, unsigned long len, 2992 unsigned long vm_flags, const struct vm_operations_struct *ops, 2993 void *priv) 2994 { 2995 int ret; 2996 struct vm_area_struct *vma; 2997 2998 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2999 if (unlikely(vma == NULL)) 3000 return ERR_PTR(-ENOMEM); 3001 3002 INIT_LIST_HEAD(&vma->anon_vma_chain); 3003 vma->vm_mm = mm; 3004 vma->vm_start = addr; 3005 vma->vm_end = addr + len; 3006 3007 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY; 3008 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3009 3010 vma->vm_ops = ops; 3011 vma->vm_private_data = priv; 3012 3013 ret = insert_vm_struct(mm, vma); 3014 if (ret) 3015 goto out; 3016 3017 mm->total_vm += len >> PAGE_SHIFT; 3018 3019 perf_event_mmap(vma); 3020 3021 return vma; 3022 3023 out: 3024 kmem_cache_free(vm_area_cachep, vma); 3025 return ERR_PTR(ret); 3026 } 3027 3028 /* 3029 * Called with mm->mmap_sem held for writing. 3030 * Insert a new vma covering the given region, with the given flags. 3031 * Its pages are supplied by the given array of struct page *. 3032 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3033 * The region past the last page supplied will always produce SIGBUS. 3034 * The array pointer and the pages it points to are assumed to stay alive 3035 * for as long as this mapping might exist. 3036 */ 3037 struct vm_area_struct *_install_special_mapping( 3038 struct mm_struct *mm, 3039 unsigned long addr, unsigned long len, 3040 unsigned long vm_flags, const struct vm_special_mapping *spec) 3041 { 3042 return __install_special_mapping(mm, addr, len, vm_flags, 3043 &special_mapping_vmops, (void *)spec); 3044 } 3045 3046 int install_special_mapping(struct mm_struct *mm, 3047 unsigned long addr, unsigned long len, 3048 unsigned long vm_flags, struct page **pages) 3049 { 3050 struct vm_area_struct *vma = __install_special_mapping( 3051 mm, addr, len, vm_flags, &legacy_special_mapping_vmops, 3052 (void *)pages); 3053 3054 return PTR_ERR_OR_ZERO(vma); 3055 } 3056 3057 static DEFINE_MUTEX(mm_all_locks_mutex); 3058 3059 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3060 { 3061 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3062 /* 3063 * The LSB of head.next can't change from under us 3064 * because we hold the mm_all_locks_mutex. 3065 */ 3066 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem); 3067 /* 3068 * We can safely modify head.next after taking the 3069 * anon_vma->root->rwsem. If some other vma in this mm shares 3070 * the same anon_vma we won't take it again. 3071 * 3072 * No need of atomic instructions here, head.next 3073 * can't change from under us thanks to the 3074 * anon_vma->root->rwsem. 3075 */ 3076 if (__test_and_set_bit(0, (unsigned long *) 3077 &anon_vma->root->rb_root.rb_node)) 3078 BUG(); 3079 } 3080 } 3081 3082 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3083 { 3084 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3085 /* 3086 * AS_MM_ALL_LOCKS can't change from under us because 3087 * we hold the mm_all_locks_mutex. 3088 * 3089 * Operations on ->flags have to be atomic because 3090 * even if AS_MM_ALL_LOCKS is stable thanks to the 3091 * mm_all_locks_mutex, there may be other cpus 3092 * changing other bitflags in parallel to us. 3093 */ 3094 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3095 BUG(); 3096 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem); 3097 } 3098 } 3099 3100 /* 3101 * This operation locks against the VM for all pte/vma/mm related 3102 * operations that could ever happen on a certain mm. This includes 3103 * vmtruncate, try_to_unmap, and all page faults. 3104 * 3105 * The caller must take the mmap_sem in write mode before calling 3106 * mm_take_all_locks(). The caller isn't allowed to release the 3107 * mmap_sem until mm_drop_all_locks() returns. 3108 * 3109 * mmap_sem in write mode is required in order to block all operations 3110 * that could modify pagetables and free pages without need of 3111 * altering the vma layout (for example populate_range() with 3112 * nonlinear vmas). It's also needed in write mode to avoid new 3113 * anon_vmas to be associated with existing vmas. 3114 * 3115 * A single task can't take more than one mm_take_all_locks() in a row 3116 * or it would deadlock. 3117 * 3118 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3119 * mapping->flags avoid to take the same lock twice, if more than one 3120 * vma in this mm is backed by the same anon_vma or address_space. 3121 * 3122 * We can take all the locks in random order because the VM code 3123 * taking i_mmap_rwsem or anon_vma->rwsem outside the mmap_sem never 3124 * takes more than one of them in a row. Secondly we're protected 3125 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex. 3126 * 3127 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3128 * that may have to take thousand of locks. 3129 * 3130 * mm_take_all_locks() can fail if it's interrupted by signals. 3131 */ 3132 int mm_take_all_locks(struct mm_struct *mm) 3133 { 3134 struct vm_area_struct *vma; 3135 struct anon_vma_chain *avc; 3136 3137 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3138 3139 mutex_lock(&mm_all_locks_mutex); 3140 3141 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3142 if (signal_pending(current)) 3143 goto out_unlock; 3144 if (vma->vm_file && vma->vm_file->f_mapping) 3145 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3146 } 3147 3148 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3149 if (signal_pending(current)) 3150 goto out_unlock; 3151 if (vma->anon_vma) 3152 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3153 vm_lock_anon_vma(mm, avc->anon_vma); 3154 } 3155 3156 return 0; 3157 3158 out_unlock: 3159 mm_drop_all_locks(mm); 3160 return -EINTR; 3161 } 3162 3163 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3164 { 3165 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3166 /* 3167 * The LSB of head.next can't change to 0 from under 3168 * us because we hold the mm_all_locks_mutex. 3169 * 3170 * We must however clear the bitflag before unlocking 3171 * the vma so the users using the anon_vma->rb_root will 3172 * never see our bitflag. 3173 * 3174 * No need of atomic instructions here, head.next 3175 * can't change from under us until we release the 3176 * anon_vma->root->rwsem. 3177 */ 3178 if (!__test_and_clear_bit(0, (unsigned long *) 3179 &anon_vma->root->rb_root.rb_node)) 3180 BUG(); 3181 anon_vma_unlock_write(anon_vma); 3182 } 3183 } 3184 3185 static void vm_unlock_mapping(struct address_space *mapping) 3186 { 3187 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3188 /* 3189 * AS_MM_ALL_LOCKS can't change to 0 from under us 3190 * because we hold the mm_all_locks_mutex. 3191 */ 3192 i_mmap_unlock_write(mapping); 3193 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3194 &mapping->flags)) 3195 BUG(); 3196 } 3197 } 3198 3199 /* 3200 * The mmap_sem cannot be released by the caller until 3201 * mm_drop_all_locks() returns. 3202 */ 3203 void mm_drop_all_locks(struct mm_struct *mm) 3204 { 3205 struct vm_area_struct *vma; 3206 struct anon_vma_chain *avc; 3207 3208 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3209 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3210 3211 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3212 if (vma->anon_vma) 3213 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3214 vm_unlock_anon_vma(avc->anon_vma); 3215 if (vma->vm_file && vma->vm_file->f_mapping) 3216 vm_unlock_mapping(vma->vm_file->f_mapping); 3217 } 3218 3219 mutex_unlock(&mm_all_locks_mutex); 3220 } 3221 3222 /* 3223 * initialise the VMA slab 3224 */ 3225 void __init mmap_init(void) 3226 { 3227 int ret; 3228 3229 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3230 VM_BUG_ON(ret); 3231 } 3232 3233 /* 3234 * Initialise sysctl_user_reserve_kbytes. 3235 * 3236 * This is intended to prevent a user from starting a single memory hogging 3237 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3238 * mode. 3239 * 3240 * The default value is min(3% of free memory, 128MB) 3241 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3242 */ 3243 static int init_user_reserve(void) 3244 { 3245 unsigned long free_kbytes; 3246 3247 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3248 3249 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3250 return 0; 3251 } 3252 subsys_initcall(init_user_reserve); 3253 3254 /* 3255 * Initialise sysctl_admin_reserve_kbytes. 3256 * 3257 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3258 * to log in and kill a memory hogging process. 3259 * 3260 * Systems with more than 256MB will reserve 8MB, enough to recover 3261 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3262 * only reserve 3% of free pages by default. 3263 */ 3264 static int init_admin_reserve(void) 3265 { 3266 unsigned long free_kbytes; 3267 3268 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3269 3270 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3271 return 0; 3272 } 3273 subsys_initcall(init_admin_reserve); 3274 3275 /* 3276 * Reinititalise user and admin reserves if memory is added or removed. 3277 * 3278 * The default user reserve max is 128MB, and the default max for the 3279 * admin reserve is 8MB. These are usually, but not always, enough to 3280 * enable recovery from a memory hogging process using login/sshd, a shell, 3281 * and tools like top. It may make sense to increase or even disable the 3282 * reserve depending on the existence of swap or variations in the recovery 3283 * tools. So, the admin may have changed them. 3284 * 3285 * If memory is added and the reserves have been eliminated or increased above 3286 * the default max, then we'll trust the admin. 3287 * 3288 * If memory is removed and there isn't enough free memory, then we 3289 * need to reset the reserves. 3290 * 3291 * Otherwise keep the reserve set by the admin. 3292 */ 3293 static int reserve_mem_notifier(struct notifier_block *nb, 3294 unsigned long action, void *data) 3295 { 3296 unsigned long tmp, free_kbytes; 3297 3298 switch (action) { 3299 case MEM_ONLINE: 3300 /* Default max is 128MB. Leave alone if modified by operator. */ 3301 tmp = sysctl_user_reserve_kbytes; 3302 if (0 < tmp && tmp < (1UL << 17)) 3303 init_user_reserve(); 3304 3305 /* Default max is 8MB. Leave alone if modified by operator. */ 3306 tmp = sysctl_admin_reserve_kbytes; 3307 if (0 < tmp && tmp < (1UL << 13)) 3308 init_admin_reserve(); 3309 3310 break; 3311 case MEM_OFFLINE: 3312 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3313 3314 if (sysctl_user_reserve_kbytes > free_kbytes) { 3315 init_user_reserve(); 3316 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3317 sysctl_user_reserve_kbytes); 3318 } 3319 3320 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3321 init_admin_reserve(); 3322 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3323 sysctl_admin_reserve_kbytes); 3324 } 3325 break; 3326 default: 3327 break; 3328 } 3329 return NOTIFY_OK; 3330 } 3331 3332 static struct notifier_block reserve_mem_nb = { 3333 .notifier_call = reserve_mem_notifier, 3334 }; 3335 3336 static int __meminit init_reserve_notifier(void) 3337 { 3338 if (register_hotmemory_notifier(&reserve_mem_nb)) 3339 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3340 3341 return 0; 3342 } 3343 subsys_initcall(init_reserve_notifier); 3344