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 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_t(long, 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 vma_interval_tree_remove(vma, &mapping->i_mmap); 247 flush_dcache_mmap_unlock(mapping); 248 } 249 250 /* 251 * Unlink a file-based vm structure from its interval tree, to hide 252 * vma from rmap and vmtruncate before freeing its page tables. 253 */ 254 void unlink_file_vma(struct vm_area_struct *vma) 255 { 256 struct file *file = vma->vm_file; 257 258 if (file) { 259 struct address_space *mapping = file->f_mapping; 260 i_mmap_lock_write(mapping); 261 __remove_shared_vm_struct(vma, file, mapping); 262 i_mmap_unlock_write(mapping); 263 } 264 } 265 266 /* 267 * Close a vm structure and free it, returning the next. 268 */ 269 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 270 { 271 struct vm_area_struct *next = vma->vm_next; 272 273 might_sleep(); 274 if (vma->vm_ops && vma->vm_ops->close) 275 vma->vm_ops->close(vma); 276 if (vma->vm_file) 277 fput(vma->vm_file); 278 mpol_put(vma_policy(vma)); 279 kmem_cache_free(vm_area_cachep, vma); 280 return next; 281 } 282 283 static unsigned long do_brk(unsigned long addr, unsigned long len); 284 285 SYSCALL_DEFINE1(brk, unsigned long, brk) 286 { 287 unsigned long retval; 288 unsigned long newbrk, oldbrk; 289 struct mm_struct *mm = current->mm; 290 unsigned long min_brk; 291 bool populate; 292 293 down_write(&mm->mmap_sem); 294 295 #ifdef CONFIG_COMPAT_BRK 296 /* 297 * CONFIG_COMPAT_BRK can still be overridden by setting 298 * randomize_va_space to 2, which will still cause mm->start_brk 299 * to be arbitrarily shifted 300 */ 301 if (current->brk_randomized) 302 min_brk = mm->start_brk; 303 else 304 min_brk = mm->end_data; 305 #else 306 min_brk = mm->start_brk; 307 #endif 308 if (brk < min_brk) 309 goto out; 310 311 /* 312 * Check against rlimit here. If this check is done later after the test 313 * of oldbrk with newbrk then it can escape the test and let the data 314 * segment grow beyond its set limit the in case where the limit is 315 * not page aligned -Ram Gupta 316 */ 317 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 318 mm->end_data, mm->start_data)) 319 goto out; 320 321 newbrk = PAGE_ALIGN(brk); 322 oldbrk = PAGE_ALIGN(mm->brk); 323 if (oldbrk == newbrk) 324 goto set_brk; 325 326 /* Always allow shrinking brk. */ 327 if (brk <= mm->brk) { 328 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 329 goto set_brk; 330 goto out; 331 } 332 333 /* Check against existing mmap mappings. */ 334 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 335 goto out; 336 337 /* Ok, looks good - let it rip. */ 338 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 339 goto out; 340 341 set_brk: 342 mm->brk = brk; 343 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 344 up_write(&mm->mmap_sem); 345 if (populate) 346 mm_populate(oldbrk, newbrk - oldbrk); 347 return brk; 348 349 out: 350 retval = mm->brk; 351 up_write(&mm->mmap_sem); 352 return retval; 353 } 354 355 static long vma_compute_subtree_gap(struct vm_area_struct *vma) 356 { 357 unsigned long max, subtree_gap; 358 max = vma->vm_start; 359 if (vma->vm_prev) 360 max -= vma->vm_prev->vm_end; 361 if (vma->vm_rb.rb_left) { 362 subtree_gap = rb_entry(vma->vm_rb.rb_left, 363 struct vm_area_struct, vm_rb)->rb_subtree_gap; 364 if (subtree_gap > max) 365 max = subtree_gap; 366 } 367 if (vma->vm_rb.rb_right) { 368 subtree_gap = rb_entry(vma->vm_rb.rb_right, 369 struct vm_area_struct, vm_rb)->rb_subtree_gap; 370 if (subtree_gap > max) 371 max = subtree_gap; 372 } 373 return max; 374 } 375 376 #ifdef CONFIG_DEBUG_VM_RB 377 static int browse_rb(struct rb_root *root) 378 { 379 int i = 0, j, bug = 0; 380 struct rb_node *nd, *pn = NULL; 381 unsigned long prev = 0, pend = 0; 382 383 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 384 struct vm_area_struct *vma; 385 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 386 if (vma->vm_start < prev) { 387 pr_emerg("vm_start %lx < prev %lx\n", 388 vma->vm_start, prev); 389 bug = 1; 390 } 391 if (vma->vm_start < pend) { 392 pr_emerg("vm_start %lx < pend %lx\n", 393 vma->vm_start, pend); 394 bug = 1; 395 } 396 if (vma->vm_start > vma->vm_end) { 397 pr_emerg("vm_start %lx > vm_end %lx\n", 398 vma->vm_start, vma->vm_end); 399 bug = 1; 400 } 401 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) { 402 pr_emerg("free gap %lx, correct %lx\n", 403 vma->rb_subtree_gap, 404 vma_compute_subtree_gap(vma)); 405 bug = 1; 406 } 407 i++; 408 pn = nd; 409 prev = vma->vm_start; 410 pend = vma->vm_end; 411 } 412 j = 0; 413 for (nd = pn; nd; nd = rb_prev(nd)) 414 j++; 415 if (i != j) { 416 pr_emerg("backwards %d, forwards %d\n", j, i); 417 bug = 1; 418 } 419 return bug ? -1 : i; 420 } 421 422 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore) 423 { 424 struct rb_node *nd; 425 426 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 427 struct vm_area_struct *vma; 428 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 429 VM_BUG_ON_VMA(vma != ignore && 430 vma->rb_subtree_gap != vma_compute_subtree_gap(vma), 431 vma); 432 } 433 } 434 435 static void validate_mm(struct mm_struct *mm) 436 { 437 int bug = 0; 438 int i = 0; 439 unsigned long highest_address = 0; 440 struct vm_area_struct *vma = mm->mmap; 441 442 while (vma) { 443 struct anon_vma_chain *avc; 444 445 vma_lock_anon_vma(vma); 446 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 447 anon_vma_interval_tree_verify(avc); 448 vma_unlock_anon_vma(vma); 449 highest_address = vma->vm_end; 450 vma = vma->vm_next; 451 i++; 452 } 453 if (i != mm->map_count) { 454 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i); 455 bug = 1; 456 } 457 if (highest_address != mm->highest_vm_end) { 458 pr_emerg("mm->highest_vm_end %lx, found %lx\n", 459 mm->highest_vm_end, highest_address); 460 bug = 1; 461 } 462 i = browse_rb(&mm->mm_rb); 463 if (i != mm->map_count) { 464 if (i != -1) 465 pr_emerg("map_count %d rb %d\n", mm->map_count, i); 466 bug = 1; 467 } 468 VM_BUG_ON_MM(bug, mm); 469 } 470 #else 471 #define validate_mm_rb(root, ignore) do { } while (0) 472 #define validate_mm(mm) do { } while (0) 473 #endif 474 475 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb, 476 unsigned long, rb_subtree_gap, vma_compute_subtree_gap) 477 478 /* 479 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or 480 * vma->vm_prev->vm_end values changed, without modifying the vma's position 481 * in the rbtree. 482 */ 483 static void vma_gap_update(struct vm_area_struct *vma) 484 { 485 /* 486 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback 487 * function that does exacltly what we want. 488 */ 489 vma_gap_callbacks_propagate(&vma->vm_rb, NULL); 490 } 491 492 static inline void vma_rb_insert(struct vm_area_struct *vma, 493 struct rb_root *root) 494 { 495 /* All rb_subtree_gap values must be consistent prior to insertion */ 496 validate_mm_rb(root, NULL); 497 498 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 499 } 500 501 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root) 502 { 503 /* 504 * All rb_subtree_gap values must be consistent prior to erase, 505 * with the possible exception of the vma being erased. 506 */ 507 validate_mm_rb(root, vma); 508 509 /* 510 * Note rb_erase_augmented is a fairly large inline function, 511 * so make sure we instantiate it only once with our desired 512 * augmented rbtree callbacks. 513 */ 514 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 515 } 516 517 /* 518 * vma has some anon_vma assigned, and is already inserted on that 519 * anon_vma's interval trees. 520 * 521 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 522 * vma must be removed from the anon_vma's interval trees using 523 * anon_vma_interval_tree_pre_update_vma(). 524 * 525 * After the update, the vma will be reinserted using 526 * anon_vma_interval_tree_post_update_vma(). 527 * 528 * The entire update must be protected by exclusive mmap_sem and by 529 * the root anon_vma's mutex. 530 */ 531 static inline void 532 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 533 { 534 struct anon_vma_chain *avc; 535 536 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 537 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 538 } 539 540 static inline void 541 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 542 { 543 struct anon_vma_chain *avc; 544 545 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 546 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 547 } 548 549 static int find_vma_links(struct mm_struct *mm, unsigned long addr, 550 unsigned long end, struct vm_area_struct **pprev, 551 struct rb_node ***rb_link, struct rb_node **rb_parent) 552 { 553 struct rb_node **__rb_link, *__rb_parent, *rb_prev; 554 555 __rb_link = &mm->mm_rb.rb_node; 556 rb_prev = __rb_parent = NULL; 557 558 while (*__rb_link) { 559 struct vm_area_struct *vma_tmp; 560 561 __rb_parent = *__rb_link; 562 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 563 564 if (vma_tmp->vm_end > addr) { 565 /* Fail if an existing vma overlaps the area */ 566 if (vma_tmp->vm_start < end) 567 return -ENOMEM; 568 __rb_link = &__rb_parent->rb_left; 569 } else { 570 rb_prev = __rb_parent; 571 __rb_link = &__rb_parent->rb_right; 572 } 573 } 574 575 *pprev = NULL; 576 if (rb_prev) 577 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 578 *rb_link = __rb_link; 579 *rb_parent = __rb_parent; 580 return 0; 581 } 582 583 static unsigned long count_vma_pages_range(struct mm_struct *mm, 584 unsigned long addr, unsigned long end) 585 { 586 unsigned long nr_pages = 0; 587 struct vm_area_struct *vma; 588 589 /* Find first overlaping mapping */ 590 vma = find_vma_intersection(mm, addr, end); 591 if (!vma) 592 return 0; 593 594 nr_pages = (min(end, vma->vm_end) - 595 max(addr, vma->vm_start)) >> PAGE_SHIFT; 596 597 /* Iterate over the rest of the overlaps */ 598 for (vma = vma->vm_next; vma; vma = vma->vm_next) { 599 unsigned long overlap_len; 600 601 if (vma->vm_start > end) 602 break; 603 604 overlap_len = min(end, vma->vm_end) - vma->vm_start; 605 nr_pages += overlap_len >> PAGE_SHIFT; 606 } 607 608 return nr_pages; 609 } 610 611 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 612 struct rb_node **rb_link, struct rb_node *rb_parent) 613 { 614 /* Update tracking information for the gap following the new vma. */ 615 if (vma->vm_next) 616 vma_gap_update(vma->vm_next); 617 else 618 mm->highest_vm_end = vma->vm_end; 619 620 /* 621 * vma->vm_prev wasn't known when we followed the rbtree to find the 622 * correct insertion point for that vma. As a result, we could not 623 * update the vma vm_rb parents rb_subtree_gap values on the way down. 624 * So, we first insert the vma with a zero rb_subtree_gap value 625 * (to be consistent with what we did on the way down), and then 626 * immediately update the gap to the correct value. Finally we 627 * rebalance the rbtree after all augmented values have been set. 628 */ 629 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 630 vma->rb_subtree_gap = 0; 631 vma_gap_update(vma); 632 vma_rb_insert(vma, &mm->mm_rb); 633 } 634 635 static void __vma_link_file(struct vm_area_struct *vma) 636 { 637 struct file *file; 638 639 file = vma->vm_file; 640 if (file) { 641 struct address_space *mapping = file->f_mapping; 642 643 if (vma->vm_flags & VM_DENYWRITE) 644 atomic_dec(&file_inode(file)->i_writecount); 645 if (vma->vm_flags & VM_SHARED) 646 atomic_inc(&mapping->i_mmap_writable); 647 648 flush_dcache_mmap_lock(mapping); 649 vma_interval_tree_insert(vma, &mapping->i_mmap); 650 flush_dcache_mmap_unlock(mapping); 651 } 652 } 653 654 static void 655 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 656 struct vm_area_struct *prev, struct rb_node **rb_link, 657 struct rb_node *rb_parent) 658 { 659 __vma_link_list(mm, vma, prev, rb_parent); 660 __vma_link_rb(mm, vma, rb_link, rb_parent); 661 } 662 663 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 664 struct vm_area_struct *prev, struct rb_node **rb_link, 665 struct rb_node *rb_parent) 666 { 667 struct address_space *mapping = NULL; 668 669 if (vma->vm_file) { 670 mapping = vma->vm_file->f_mapping; 671 i_mmap_lock_write(mapping); 672 } 673 674 __vma_link(mm, vma, prev, rb_link, rb_parent); 675 __vma_link_file(vma); 676 677 if (mapping) 678 i_mmap_unlock_write(mapping); 679 680 mm->map_count++; 681 validate_mm(mm); 682 } 683 684 /* 685 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the 686 * mm's list and rbtree. It has already been inserted into the interval tree. 687 */ 688 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 689 { 690 struct vm_area_struct *prev; 691 struct rb_node **rb_link, *rb_parent; 692 693 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 694 &prev, &rb_link, &rb_parent)) 695 BUG(); 696 __vma_link(mm, vma, prev, rb_link, rb_parent); 697 mm->map_count++; 698 } 699 700 static inline void 701 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 702 struct vm_area_struct *prev) 703 { 704 struct vm_area_struct *next; 705 706 vma_rb_erase(vma, &mm->mm_rb); 707 prev->vm_next = next = vma->vm_next; 708 if (next) 709 next->vm_prev = prev; 710 711 /* Kill the cache */ 712 vmacache_invalidate(mm); 713 } 714 715 /* 716 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 717 * is already present in an i_mmap tree without adjusting the tree. 718 * The following helper function should be used when such adjustments 719 * are necessary. The "insert" vma (if any) is to be inserted 720 * before we drop the necessary locks. 721 */ 722 int vma_adjust(struct vm_area_struct *vma, unsigned long start, 723 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 724 { 725 struct mm_struct *mm = vma->vm_mm; 726 struct vm_area_struct *next = vma->vm_next; 727 struct vm_area_struct *importer = NULL; 728 struct address_space *mapping = NULL; 729 struct rb_root *root = NULL; 730 struct anon_vma *anon_vma = NULL; 731 struct file *file = vma->vm_file; 732 bool start_changed = false, end_changed = false; 733 long adjust_next = 0; 734 int remove_next = 0; 735 736 if (next && !insert) { 737 struct vm_area_struct *exporter = NULL; 738 739 if (end >= next->vm_end) { 740 /* 741 * vma expands, overlapping all the next, and 742 * perhaps the one after too (mprotect case 6). 743 */ 744 again: remove_next = 1 + (end > next->vm_end); 745 end = next->vm_end; 746 exporter = next; 747 importer = vma; 748 } else if (end > next->vm_start) { 749 /* 750 * vma expands, overlapping part of the next: 751 * mprotect case 5 shifting the boundary up. 752 */ 753 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 754 exporter = next; 755 importer = vma; 756 } else if (end < vma->vm_end) { 757 /* 758 * vma shrinks, and !insert tells it's not 759 * split_vma inserting another: so it must be 760 * mprotect case 4 shifting the boundary down. 761 */ 762 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT); 763 exporter = vma; 764 importer = next; 765 } 766 767 /* 768 * Easily overlooked: when mprotect shifts the boundary, 769 * make sure the expanding vma has anon_vma set if the 770 * shrinking vma had, to cover any anon pages imported. 771 */ 772 if (exporter && exporter->anon_vma && !importer->anon_vma) { 773 int error; 774 775 importer->anon_vma = exporter->anon_vma; 776 error = anon_vma_clone(importer, exporter); 777 if (error) { 778 importer->anon_vma = NULL; 779 return error; 780 } 781 } 782 } 783 784 if (file) { 785 mapping = file->f_mapping; 786 root = &mapping->i_mmap; 787 uprobe_munmap(vma, vma->vm_start, vma->vm_end); 788 789 if (adjust_next) 790 uprobe_munmap(next, next->vm_start, next->vm_end); 791 792 i_mmap_lock_write(mapping); 793 if (insert) { 794 /* 795 * Put into interval tree now, so instantiated pages 796 * are visible to arm/parisc __flush_dcache_page 797 * throughout; but we cannot insert into address 798 * space until vma start or end is updated. 799 */ 800 __vma_link_file(insert); 801 } 802 } 803 804 vma_adjust_trans_huge(vma, start, end, adjust_next); 805 806 anon_vma = vma->anon_vma; 807 if (!anon_vma && adjust_next) 808 anon_vma = next->anon_vma; 809 if (anon_vma) { 810 VM_BUG_ON_VMA(adjust_next && next->anon_vma && 811 anon_vma != next->anon_vma, next); 812 anon_vma_lock_write(anon_vma); 813 anon_vma_interval_tree_pre_update_vma(vma); 814 if (adjust_next) 815 anon_vma_interval_tree_pre_update_vma(next); 816 } 817 818 if (root) { 819 flush_dcache_mmap_lock(mapping); 820 vma_interval_tree_remove(vma, root); 821 if (adjust_next) 822 vma_interval_tree_remove(next, root); 823 } 824 825 if (start != vma->vm_start) { 826 vma->vm_start = start; 827 start_changed = true; 828 } 829 if (end != vma->vm_end) { 830 vma->vm_end = end; 831 end_changed = true; 832 } 833 vma->vm_pgoff = pgoff; 834 if (adjust_next) { 835 next->vm_start += adjust_next << PAGE_SHIFT; 836 next->vm_pgoff += adjust_next; 837 } 838 839 if (root) { 840 if (adjust_next) 841 vma_interval_tree_insert(next, root); 842 vma_interval_tree_insert(vma, root); 843 flush_dcache_mmap_unlock(mapping); 844 } 845 846 if (remove_next) { 847 /* 848 * vma_merge has merged next into vma, and needs 849 * us to remove next before dropping the locks. 850 */ 851 __vma_unlink(mm, next, vma); 852 if (file) 853 __remove_shared_vm_struct(next, file, mapping); 854 } else if (insert) { 855 /* 856 * split_vma has split insert from vma, and needs 857 * us to insert it before dropping the locks 858 * (it may either follow vma or precede it). 859 */ 860 __insert_vm_struct(mm, insert); 861 } else { 862 if (start_changed) 863 vma_gap_update(vma); 864 if (end_changed) { 865 if (!next) 866 mm->highest_vm_end = end; 867 else if (!adjust_next) 868 vma_gap_update(next); 869 } 870 } 871 872 if (anon_vma) { 873 anon_vma_interval_tree_post_update_vma(vma); 874 if (adjust_next) 875 anon_vma_interval_tree_post_update_vma(next); 876 anon_vma_unlock_write(anon_vma); 877 } 878 if (mapping) 879 i_mmap_unlock_write(mapping); 880 881 if (root) { 882 uprobe_mmap(vma); 883 884 if (adjust_next) 885 uprobe_mmap(next); 886 } 887 888 if (remove_next) { 889 if (file) { 890 uprobe_munmap(next, next->vm_start, next->vm_end); 891 fput(file); 892 } 893 if (next->anon_vma) 894 anon_vma_merge(vma, next); 895 mm->map_count--; 896 mpol_put(vma_policy(next)); 897 kmem_cache_free(vm_area_cachep, next); 898 /* 899 * In mprotect's case 6 (see comments on vma_merge), 900 * we must remove another next too. It would clutter 901 * up the code too much to do both in one go. 902 */ 903 next = vma->vm_next; 904 if (remove_next == 2) 905 goto again; 906 else if (next) 907 vma_gap_update(next); 908 else 909 mm->highest_vm_end = end; 910 } 911 if (insert && file) 912 uprobe_mmap(insert); 913 914 validate_mm(mm); 915 916 return 0; 917 } 918 919 /* 920 * If the vma has a ->close operation then the driver probably needs to release 921 * per-vma resources, so we don't attempt to merge those. 922 */ 923 static inline int is_mergeable_vma(struct vm_area_struct *vma, 924 struct file *file, unsigned long vm_flags) 925 { 926 /* 927 * VM_SOFTDIRTY should not prevent from VMA merging, if we 928 * match the flags but dirty bit -- the caller should mark 929 * merged VMA as dirty. If dirty bit won't be excluded from 930 * comparison, we increase pressue on the memory system forcing 931 * the kernel to generate new VMAs when old one could be 932 * extended instead. 933 */ 934 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 935 return 0; 936 if (vma->vm_file != file) 937 return 0; 938 if (vma->vm_ops && vma->vm_ops->close) 939 return 0; 940 return 1; 941 } 942 943 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 944 struct anon_vma *anon_vma2, 945 struct vm_area_struct *vma) 946 { 947 /* 948 * The list_is_singular() test is to avoid merging VMA cloned from 949 * parents. This can improve scalability caused by anon_vma lock. 950 */ 951 if ((!anon_vma1 || !anon_vma2) && (!vma || 952 list_is_singular(&vma->anon_vma_chain))) 953 return 1; 954 return anon_vma1 == anon_vma2; 955 } 956 957 /* 958 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 959 * in front of (at a lower virtual address and file offset than) the vma. 960 * 961 * We cannot merge two vmas if they have differently assigned (non-NULL) 962 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 963 * 964 * We don't check here for the merged mmap wrapping around the end of pagecache 965 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 966 * wrap, nor mmaps which cover the final page at index -1UL. 967 */ 968 static int 969 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 970 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 971 { 972 if (is_mergeable_vma(vma, file, vm_flags) && 973 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 974 if (vma->vm_pgoff == vm_pgoff) 975 return 1; 976 } 977 return 0; 978 } 979 980 /* 981 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 982 * beyond (at a higher virtual address and file offset than) the vma. 983 * 984 * We cannot merge two vmas if they have differently assigned (non-NULL) 985 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 986 */ 987 static int 988 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 989 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff) 990 { 991 if (is_mergeable_vma(vma, file, vm_flags) && 992 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 993 pgoff_t vm_pglen; 994 vm_pglen = vma_pages(vma); 995 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 996 return 1; 997 } 998 return 0; 999 } 1000 1001 /* 1002 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 1003 * whether that can be merged with its predecessor or its successor. 1004 * Or both (it neatly fills a hole). 1005 * 1006 * In most cases - when called for mmap, brk or mremap - [addr,end) is 1007 * certain not to be mapped by the time vma_merge is called; but when 1008 * called for mprotect, it is certain to be already mapped (either at 1009 * an offset within prev, or at the start of next), and the flags of 1010 * this area are about to be changed to vm_flags - and the no-change 1011 * case has already been eliminated. 1012 * 1013 * The following mprotect cases have to be considered, where AAAA is 1014 * the area passed down from mprotect_fixup, never extending beyond one 1015 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 1016 * 1017 * AAAA AAAA AAAA AAAA 1018 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 1019 * cannot merge might become might become might become 1020 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 1021 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 1022 * mremap move: PPPPNNNNNNNN 8 1023 * AAAA 1024 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 1025 * might become case 1 below case 2 below case 3 below 1026 * 1027 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 1028 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 1029 */ 1030 struct vm_area_struct *vma_merge(struct mm_struct *mm, 1031 struct vm_area_struct *prev, unsigned long addr, 1032 unsigned long end, unsigned long vm_flags, 1033 struct anon_vma *anon_vma, struct file *file, 1034 pgoff_t pgoff, struct mempolicy *policy) 1035 { 1036 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 1037 struct vm_area_struct *area, *next; 1038 int err; 1039 1040 /* 1041 * We later require that vma->vm_flags == vm_flags, 1042 * so this tests vma->vm_flags & VM_SPECIAL, too. 1043 */ 1044 if (vm_flags & VM_SPECIAL) 1045 return NULL; 1046 1047 if (prev) 1048 next = prev->vm_next; 1049 else 1050 next = mm->mmap; 1051 area = next; 1052 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 1053 next = next->vm_next; 1054 1055 /* 1056 * Can it merge with the predecessor? 1057 */ 1058 if (prev && prev->vm_end == addr && 1059 mpol_equal(vma_policy(prev), policy) && 1060 can_vma_merge_after(prev, vm_flags, 1061 anon_vma, file, pgoff)) { 1062 /* 1063 * OK, it can. Can we now merge in the successor as well? 1064 */ 1065 if (next && end == next->vm_start && 1066 mpol_equal(policy, vma_policy(next)) && 1067 can_vma_merge_before(next, vm_flags, 1068 anon_vma, file, pgoff+pglen) && 1069 is_mergeable_anon_vma(prev->anon_vma, 1070 next->anon_vma, NULL)) { 1071 /* cases 1, 6 */ 1072 err = vma_adjust(prev, prev->vm_start, 1073 next->vm_end, prev->vm_pgoff, NULL); 1074 } else /* cases 2, 5, 7 */ 1075 err = vma_adjust(prev, prev->vm_start, 1076 end, prev->vm_pgoff, NULL); 1077 if (err) 1078 return NULL; 1079 khugepaged_enter_vma_merge(prev, vm_flags); 1080 return prev; 1081 } 1082 1083 /* 1084 * Can this new request be merged in front of next? 1085 */ 1086 if (next && end == next->vm_start && 1087 mpol_equal(policy, vma_policy(next)) && 1088 can_vma_merge_before(next, vm_flags, 1089 anon_vma, file, pgoff+pglen)) { 1090 if (prev && addr < prev->vm_end) /* case 4 */ 1091 err = vma_adjust(prev, prev->vm_start, 1092 addr, prev->vm_pgoff, NULL); 1093 else /* cases 3, 8 */ 1094 err = vma_adjust(area, addr, next->vm_end, 1095 next->vm_pgoff - pglen, NULL); 1096 if (err) 1097 return NULL; 1098 khugepaged_enter_vma_merge(area, vm_flags); 1099 return area; 1100 } 1101 1102 return NULL; 1103 } 1104 1105 /* 1106 * Rough compatbility check to quickly see if it's even worth looking 1107 * at sharing an anon_vma. 1108 * 1109 * They need to have the same vm_file, and the flags can only differ 1110 * in things that mprotect may change. 1111 * 1112 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1113 * we can merge the two vma's. For example, we refuse to merge a vma if 1114 * there is a vm_ops->close() function, because that indicates that the 1115 * driver is doing some kind of reference counting. But that doesn't 1116 * really matter for the anon_vma sharing case. 1117 */ 1118 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1119 { 1120 return a->vm_end == b->vm_start && 1121 mpol_equal(vma_policy(a), vma_policy(b)) && 1122 a->vm_file == b->vm_file && 1123 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) && 1124 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1125 } 1126 1127 /* 1128 * Do some basic sanity checking to see if we can re-use the anon_vma 1129 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1130 * the same as 'old', the other will be the new one that is trying 1131 * to share the anon_vma. 1132 * 1133 * NOTE! This runs with mm_sem held for reading, so it is possible that 1134 * the anon_vma of 'old' is concurrently in the process of being set up 1135 * by another page fault trying to merge _that_. But that's ok: if it 1136 * is being set up, that automatically means that it will be a singleton 1137 * acceptable for merging, so we can do all of this optimistically. But 1138 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer. 1139 * 1140 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1141 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1142 * is to return an anon_vma that is "complex" due to having gone through 1143 * a fork). 1144 * 1145 * We also make sure that the two vma's are compatible (adjacent, 1146 * and with the same memory policies). That's all stable, even with just 1147 * a read lock on the mm_sem. 1148 */ 1149 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1150 { 1151 if (anon_vma_compatible(a, b)) { 1152 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma); 1153 1154 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1155 return anon_vma; 1156 } 1157 return NULL; 1158 } 1159 1160 /* 1161 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1162 * neighbouring vmas for a suitable anon_vma, before it goes off 1163 * to allocate a new anon_vma. It checks because a repetitive 1164 * sequence of mprotects and faults may otherwise lead to distinct 1165 * anon_vmas being allocated, preventing vma merge in subsequent 1166 * mprotect. 1167 */ 1168 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1169 { 1170 struct anon_vma *anon_vma; 1171 struct vm_area_struct *near; 1172 1173 near = vma->vm_next; 1174 if (!near) 1175 goto try_prev; 1176 1177 anon_vma = reusable_anon_vma(near, vma, near); 1178 if (anon_vma) 1179 return anon_vma; 1180 try_prev: 1181 near = vma->vm_prev; 1182 if (!near) 1183 goto none; 1184 1185 anon_vma = reusable_anon_vma(near, near, vma); 1186 if (anon_vma) 1187 return anon_vma; 1188 none: 1189 /* 1190 * There's no absolute need to look only at touching neighbours: 1191 * we could search further afield for "compatible" anon_vmas. 1192 * But it would probably just be a waste of time searching, 1193 * or lead to too many vmas hanging off the same anon_vma. 1194 * We're trying to allow mprotect remerging later on, 1195 * not trying to minimize memory used for anon_vmas. 1196 */ 1197 return NULL; 1198 } 1199 1200 #ifdef CONFIG_PROC_FS 1201 void vm_stat_account(struct mm_struct *mm, unsigned long flags, 1202 struct file *file, long pages) 1203 { 1204 const unsigned long stack_flags 1205 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN); 1206 1207 mm->total_vm += pages; 1208 1209 if (file) { 1210 mm->shared_vm += pages; 1211 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC) 1212 mm->exec_vm += pages; 1213 } else if (flags & stack_flags) 1214 mm->stack_vm += pages; 1215 } 1216 #endif /* CONFIG_PROC_FS */ 1217 1218 /* 1219 * If a hint addr is less than mmap_min_addr change hint to be as 1220 * low as possible but still greater than mmap_min_addr 1221 */ 1222 static inline unsigned long round_hint_to_min(unsigned long hint) 1223 { 1224 hint &= PAGE_MASK; 1225 if (((void *)hint != NULL) && 1226 (hint < mmap_min_addr)) 1227 return PAGE_ALIGN(mmap_min_addr); 1228 return hint; 1229 } 1230 1231 static inline int mlock_future_check(struct mm_struct *mm, 1232 unsigned long flags, 1233 unsigned long len) 1234 { 1235 unsigned long locked, lock_limit; 1236 1237 /* mlock MCL_FUTURE? */ 1238 if (flags & VM_LOCKED) { 1239 locked = len >> PAGE_SHIFT; 1240 locked += mm->locked_vm; 1241 lock_limit = rlimit(RLIMIT_MEMLOCK); 1242 lock_limit >>= PAGE_SHIFT; 1243 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1244 return -EAGAIN; 1245 } 1246 return 0; 1247 } 1248 1249 /* 1250 * The caller must hold down_write(¤t->mm->mmap_sem). 1251 */ 1252 1253 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr, 1254 unsigned long len, unsigned long prot, 1255 unsigned long flags, unsigned long pgoff, 1256 unsigned long *populate) 1257 { 1258 struct mm_struct *mm = current->mm; 1259 vm_flags_t vm_flags; 1260 1261 *populate = 0; 1262 1263 /* 1264 * Does the application expect PROT_READ to imply PROT_EXEC? 1265 * 1266 * (the exception is when the underlying filesystem is noexec 1267 * mounted, in which case we dont add PROT_EXEC.) 1268 */ 1269 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1270 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC))) 1271 prot |= PROT_EXEC; 1272 1273 if (!len) 1274 return -EINVAL; 1275 1276 if (!(flags & MAP_FIXED)) 1277 addr = round_hint_to_min(addr); 1278 1279 /* Careful about overflows.. */ 1280 len = PAGE_ALIGN(len); 1281 if (!len) 1282 return -ENOMEM; 1283 1284 /* offset overflow? */ 1285 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1286 return -EOVERFLOW; 1287 1288 /* Too many mappings? */ 1289 if (mm->map_count > sysctl_max_map_count) 1290 return -ENOMEM; 1291 1292 /* Obtain the address to map to. we verify (or select) it and ensure 1293 * that it represents a valid section of the address space. 1294 */ 1295 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1296 if (addr & ~PAGE_MASK) 1297 return addr; 1298 1299 /* Do simple checking here so the lower-level routines won't have 1300 * to. we assume access permissions have been handled by the open 1301 * of the memory object, so we don't do any here. 1302 */ 1303 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 1304 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1305 1306 if (flags & MAP_LOCKED) 1307 if (!can_do_mlock()) 1308 return -EPERM; 1309 1310 if (mlock_future_check(mm, vm_flags, len)) 1311 return -EAGAIN; 1312 1313 if (file) { 1314 struct inode *inode = file_inode(file); 1315 1316 switch (flags & MAP_TYPE) { 1317 case MAP_SHARED: 1318 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 1319 return -EACCES; 1320 1321 /* 1322 * Make sure we don't allow writing to an append-only 1323 * file.. 1324 */ 1325 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1326 return -EACCES; 1327 1328 /* 1329 * Make sure there are no mandatory locks on the file. 1330 */ 1331 if (locks_verify_locked(file)) 1332 return -EAGAIN; 1333 1334 vm_flags |= VM_SHARED | VM_MAYSHARE; 1335 if (!(file->f_mode & FMODE_WRITE)) 1336 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1337 1338 /* fall through */ 1339 case MAP_PRIVATE: 1340 if (!(file->f_mode & FMODE_READ)) 1341 return -EACCES; 1342 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) { 1343 if (vm_flags & VM_EXEC) 1344 return -EPERM; 1345 vm_flags &= ~VM_MAYEXEC; 1346 } 1347 1348 if (!file->f_op->mmap) 1349 return -ENODEV; 1350 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1351 return -EINVAL; 1352 break; 1353 1354 default: 1355 return -EINVAL; 1356 } 1357 } else { 1358 switch (flags & MAP_TYPE) { 1359 case MAP_SHARED: 1360 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1361 return -EINVAL; 1362 /* 1363 * Ignore pgoff. 1364 */ 1365 pgoff = 0; 1366 vm_flags |= VM_SHARED | VM_MAYSHARE; 1367 break; 1368 case MAP_PRIVATE: 1369 /* 1370 * Set pgoff according to addr for anon_vma. 1371 */ 1372 pgoff = addr >> PAGE_SHIFT; 1373 break; 1374 default: 1375 return -EINVAL; 1376 } 1377 } 1378 1379 /* 1380 * Set 'VM_NORESERVE' if we should not account for the 1381 * memory use of this mapping. 1382 */ 1383 if (flags & MAP_NORESERVE) { 1384 /* We honor MAP_NORESERVE if allowed to overcommit */ 1385 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1386 vm_flags |= VM_NORESERVE; 1387 1388 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1389 if (file && is_file_hugepages(file)) 1390 vm_flags |= VM_NORESERVE; 1391 } 1392 1393 addr = mmap_region(file, addr, len, vm_flags, pgoff); 1394 if (!IS_ERR_VALUE(addr) && 1395 ((vm_flags & VM_LOCKED) || 1396 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1397 *populate = len; 1398 return addr; 1399 } 1400 1401 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1402 unsigned long, prot, unsigned long, flags, 1403 unsigned long, fd, unsigned long, pgoff) 1404 { 1405 struct file *file = NULL; 1406 unsigned long retval = -EBADF; 1407 1408 if (!(flags & MAP_ANONYMOUS)) { 1409 audit_mmap_fd(fd, flags); 1410 file = fget(fd); 1411 if (!file) 1412 goto out; 1413 if (is_file_hugepages(file)) 1414 len = ALIGN(len, huge_page_size(hstate_file(file))); 1415 retval = -EINVAL; 1416 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file))) 1417 goto out_fput; 1418 } else if (flags & MAP_HUGETLB) { 1419 struct user_struct *user = NULL; 1420 struct hstate *hs; 1421 1422 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK); 1423 if (!hs) 1424 return -EINVAL; 1425 1426 len = ALIGN(len, huge_page_size(hs)); 1427 /* 1428 * VM_NORESERVE is used because the reservations will be 1429 * taken when vm_ops->mmap() is called 1430 * A dummy user value is used because we are not locking 1431 * memory so no accounting is necessary 1432 */ 1433 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1434 VM_NORESERVE, 1435 &user, HUGETLB_ANONHUGE_INODE, 1436 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1437 if (IS_ERR(file)) 1438 return PTR_ERR(file); 1439 } 1440 1441 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1442 1443 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1444 out_fput: 1445 if (file) 1446 fput(file); 1447 out: 1448 return retval; 1449 } 1450 1451 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1452 struct mmap_arg_struct { 1453 unsigned long addr; 1454 unsigned long len; 1455 unsigned long prot; 1456 unsigned long flags; 1457 unsigned long fd; 1458 unsigned long offset; 1459 }; 1460 1461 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1462 { 1463 struct mmap_arg_struct a; 1464 1465 if (copy_from_user(&a, arg, sizeof(a))) 1466 return -EFAULT; 1467 if (a.offset & ~PAGE_MASK) 1468 return -EINVAL; 1469 1470 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1471 a.offset >> PAGE_SHIFT); 1472 } 1473 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1474 1475 /* 1476 * Some shared mappigns will want the pages marked read-only 1477 * to track write events. If so, we'll downgrade vm_page_prot 1478 * to the private version (using protection_map[] without the 1479 * VM_SHARED bit). 1480 */ 1481 int vma_wants_writenotify(struct vm_area_struct *vma) 1482 { 1483 vm_flags_t vm_flags = vma->vm_flags; 1484 1485 /* If it was private or non-writable, the write bit is already clear */ 1486 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1487 return 0; 1488 1489 /* The backer wishes to know when pages are first written to? */ 1490 if (vma->vm_ops && vma->vm_ops->page_mkwrite) 1491 return 1; 1492 1493 /* The open routine did something to the protections that pgprot_modify 1494 * won't preserve? */ 1495 if (pgprot_val(vma->vm_page_prot) != 1496 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags))) 1497 return 0; 1498 1499 /* Do we need to track softdirty? */ 1500 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY)) 1501 return 1; 1502 1503 /* Specialty mapping? */ 1504 if (vm_flags & VM_PFNMAP) 1505 return 0; 1506 1507 /* Can the mapping track the dirty pages? */ 1508 return vma->vm_file && vma->vm_file->f_mapping && 1509 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1510 } 1511 1512 /* 1513 * We account for memory if it's a private writeable mapping, 1514 * not hugepages and VM_NORESERVE wasn't set. 1515 */ 1516 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1517 { 1518 /* 1519 * hugetlb has its own accounting separate from the core VM 1520 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1521 */ 1522 if (file && is_file_hugepages(file)) 1523 return 0; 1524 1525 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1526 } 1527 1528 unsigned long mmap_region(struct file *file, unsigned long addr, 1529 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff) 1530 { 1531 struct mm_struct *mm = current->mm; 1532 struct vm_area_struct *vma, *prev; 1533 int error; 1534 struct rb_node **rb_link, *rb_parent; 1535 unsigned long charged = 0; 1536 1537 /* Check against address space limit. */ 1538 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) { 1539 unsigned long nr_pages; 1540 1541 /* 1542 * MAP_FIXED may remove pages of mappings that intersects with 1543 * requested mapping. Account for the pages it would unmap. 1544 */ 1545 if (!(vm_flags & MAP_FIXED)) 1546 return -ENOMEM; 1547 1548 nr_pages = count_vma_pages_range(mm, addr, addr + len); 1549 1550 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages)) 1551 return -ENOMEM; 1552 } 1553 1554 /* Clear old maps */ 1555 error = -ENOMEM; 1556 munmap_back: 1557 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) { 1558 if (do_munmap(mm, addr, len)) 1559 return -ENOMEM; 1560 goto munmap_back; 1561 } 1562 1563 /* 1564 * Private writable mapping: check memory availability 1565 */ 1566 if (accountable_mapping(file, vm_flags)) { 1567 charged = len >> PAGE_SHIFT; 1568 if (security_vm_enough_memory_mm(mm, charged)) 1569 return -ENOMEM; 1570 vm_flags |= VM_ACCOUNT; 1571 } 1572 1573 /* 1574 * Can we just expand an old mapping? 1575 */ 1576 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL); 1577 if (vma) 1578 goto out; 1579 1580 /* 1581 * Determine the object being mapped and call the appropriate 1582 * specific mapper. the address has already been validated, but 1583 * not unmapped, but the maps are removed from the list. 1584 */ 1585 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1586 if (!vma) { 1587 error = -ENOMEM; 1588 goto unacct_error; 1589 } 1590 1591 vma->vm_mm = mm; 1592 vma->vm_start = addr; 1593 vma->vm_end = addr + len; 1594 vma->vm_flags = vm_flags; 1595 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1596 vma->vm_pgoff = pgoff; 1597 INIT_LIST_HEAD(&vma->anon_vma_chain); 1598 1599 if (file) { 1600 if (vm_flags & VM_DENYWRITE) { 1601 error = deny_write_access(file); 1602 if (error) 1603 goto free_vma; 1604 } 1605 if (vm_flags & VM_SHARED) { 1606 error = mapping_map_writable(file->f_mapping); 1607 if (error) 1608 goto allow_write_and_free_vma; 1609 } 1610 1611 /* ->mmap() can change vma->vm_file, but must guarantee that 1612 * vma_link() below can deny write-access if VM_DENYWRITE is set 1613 * and map writably if VM_SHARED is set. This usually means the 1614 * new file must not have been exposed to user-space, yet. 1615 */ 1616 vma->vm_file = get_file(file); 1617 error = file->f_op->mmap(file, vma); 1618 if (error) 1619 goto unmap_and_free_vma; 1620 1621 /* Can addr have changed?? 1622 * 1623 * Answer: Yes, several device drivers can do it in their 1624 * f_op->mmap method. -DaveM 1625 * Bug: If addr is changed, prev, rb_link, rb_parent should 1626 * be updated for vma_link() 1627 */ 1628 WARN_ON_ONCE(addr != vma->vm_start); 1629 1630 addr = vma->vm_start; 1631 vm_flags = vma->vm_flags; 1632 } else if (vm_flags & VM_SHARED) { 1633 error = shmem_zero_setup(vma); 1634 if (error) 1635 goto free_vma; 1636 } 1637 1638 vma_link(mm, vma, prev, rb_link, rb_parent); 1639 /* Once vma denies write, undo our temporary denial count */ 1640 if (file) { 1641 if (vm_flags & VM_SHARED) 1642 mapping_unmap_writable(file->f_mapping); 1643 if (vm_flags & VM_DENYWRITE) 1644 allow_write_access(file); 1645 } 1646 file = vma->vm_file; 1647 out: 1648 perf_event_mmap(vma); 1649 1650 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT); 1651 if (vm_flags & VM_LOCKED) { 1652 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) || 1653 vma == get_gate_vma(current->mm))) 1654 mm->locked_vm += (len >> PAGE_SHIFT); 1655 else 1656 vma->vm_flags &= ~VM_LOCKED; 1657 } 1658 1659 if (file) 1660 uprobe_mmap(vma); 1661 1662 /* 1663 * New (or expanded) vma always get soft dirty status. 1664 * Otherwise user-space soft-dirty page tracker won't 1665 * be able to distinguish situation when vma area unmapped, 1666 * then new mapped in-place (which must be aimed as 1667 * a completely new data area). 1668 */ 1669 vma->vm_flags |= VM_SOFTDIRTY; 1670 1671 vma_set_page_prot(vma); 1672 1673 return addr; 1674 1675 unmap_and_free_vma: 1676 vma->vm_file = NULL; 1677 fput(file); 1678 1679 /* Undo any partial mapping done by a device driver. */ 1680 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1681 charged = 0; 1682 if (vm_flags & VM_SHARED) 1683 mapping_unmap_writable(file->f_mapping); 1684 allow_write_and_free_vma: 1685 if (vm_flags & VM_DENYWRITE) 1686 allow_write_access(file); 1687 free_vma: 1688 kmem_cache_free(vm_area_cachep, vma); 1689 unacct_error: 1690 if (charged) 1691 vm_unacct_memory(charged); 1692 return error; 1693 } 1694 1695 unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1696 { 1697 /* 1698 * We implement the search by looking for an rbtree node that 1699 * immediately follows a suitable gap. That is, 1700 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length; 1701 * - gap_end = vma->vm_start >= info->low_limit + length; 1702 * - gap_end - gap_start >= length 1703 */ 1704 1705 struct mm_struct *mm = current->mm; 1706 struct vm_area_struct *vma; 1707 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1708 1709 /* Adjust search length to account for worst case alignment overhead */ 1710 length = info->length + info->align_mask; 1711 if (length < info->length) 1712 return -ENOMEM; 1713 1714 /* Adjust search limits by the desired length */ 1715 if (info->high_limit < length) 1716 return -ENOMEM; 1717 high_limit = info->high_limit - length; 1718 1719 if (info->low_limit > high_limit) 1720 return -ENOMEM; 1721 low_limit = info->low_limit + length; 1722 1723 /* Check if rbtree root looks promising */ 1724 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1725 goto check_highest; 1726 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1727 if (vma->rb_subtree_gap < length) 1728 goto check_highest; 1729 1730 while (true) { 1731 /* Visit left subtree if it looks promising */ 1732 gap_end = vma->vm_start; 1733 if (gap_end >= low_limit && vma->vm_rb.rb_left) { 1734 struct vm_area_struct *left = 1735 rb_entry(vma->vm_rb.rb_left, 1736 struct vm_area_struct, vm_rb); 1737 if (left->rb_subtree_gap >= length) { 1738 vma = left; 1739 continue; 1740 } 1741 } 1742 1743 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1744 check_current: 1745 /* Check if current node has a suitable gap */ 1746 if (gap_start > high_limit) 1747 return -ENOMEM; 1748 if (gap_end >= low_limit && gap_end - gap_start >= length) 1749 goto found; 1750 1751 /* Visit right subtree if it looks promising */ 1752 if (vma->vm_rb.rb_right) { 1753 struct vm_area_struct *right = 1754 rb_entry(vma->vm_rb.rb_right, 1755 struct vm_area_struct, vm_rb); 1756 if (right->rb_subtree_gap >= length) { 1757 vma = right; 1758 continue; 1759 } 1760 } 1761 1762 /* Go back up the rbtree to find next candidate node */ 1763 while (true) { 1764 struct rb_node *prev = &vma->vm_rb; 1765 if (!rb_parent(prev)) 1766 goto check_highest; 1767 vma = rb_entry(rb_parent(prev), 1768 struct vm_area_struct, vm_rb); 1769 if (prev == vma->vm_rb.rb_left) { 1770 gap_start = vma->vm_prev->vm_end; 1771 gap_end = vma->vm_start; 1772 goto check_current; 1773 } 1774 } 1775 } 1776 1777 check_highest: 1778 /* Check highest gap, which does not precede any rbtree node */ 1779 gap_start = mm->highest_vm_end; 1780 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */ 1781 if (gap_start > high_limit) 1782 return -ENOMEM; 1783 1784 found: 1785 /* We found a suitable gap. Clip it with the original low_limit. */ 1786 if (gap_start < info->low_limit) 1787 gap_start = info->low_limit; 1788 1789 /* Adjust gap address to the desired alignment */ 1790 gap_start += (info->align_offset - gap_start) & info->align_mask; 1791 1792 VM_BUG_ON(gap_start + info->length > info->high_limit); 1793 VM_BUG_ON(gap_start + info->length > gap_end); 1794 return gap_start; 1795 } 1796 1797 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1798 { 1799 struct mm_struct *mm = current->mm; 1800 struct vm_area_struct *vma; 1801 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1802 1803 /* Adjust search length to account for worst case alignment overhead */ 1804 length = info->length + info->align_mask; 1805 if (length < info->length) 1806 return -ENOMEM; 1807 1808 /* 1809 * Adjust search limits by the desired length. 1810 * See implementation comment at top of unmapped_area(). 1811 */ 1812 gap_end = info->high_limit; 1813 if (gap_end < length) 1814 return -ENOMEM; 1815 high_limit = gap_end - length; 1816 1817 if (info->low_limit > high_limit) 1818 return -ENOMEM; 1819 low_limit = info->low_limit + length; 1820 1821 /* Check highest gap, which does not precede any rbtree node */ 1822 gap_start = mm->highest_vm_end; 1823 if (gap_start <= high_limit) 1824 goto found_highest; 1825 1826 /* Check if rbtree root looks promising */ 1827 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1828 return -ENOMEM; 1829 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1830 if (vma->rb_subtree_gap < length) 1831 return -ENOMEM; 1832 1833 while (true) { 1834 /* Visit right subtree if it looks promising */ 1835 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1836 if (gap_start <= high_limit && vma->vm_rb.rb_right) { 1837 struct vm_area_struct *right = 1838 rb_entry(vma->vm_rb.rb_right, 1839 struct vm_area_struct, vm_rb); 1840 if (right->rb_subtree_gap >= length) { 1841 vma = right; 1842 continue; 1843 } 1844 } 1845 1846 check_current: 1847 /* Check if current node has a suitable gap */ 1848 gap_end = vma->vm_start; 1849 if (gap_end < low_limit) 1850 return -ENOMEM; 1851 if (gap_start <= high_limit && gap_end - gap_start >= length) 1852 goto found; 1853 1854 /* Visit left subtree if it looks promising */ 1855 if (vma->vm_rb.rb_left) { 1856 struct vm_area_struct *left = 1857 rb_entry(vma->vm_rb.rb_left, 1858 struct vm_area_struct, vm_rb); 1859 if (left->rb_subtree_gap >= length) { 1860 vma = left; 1861 continue; 1862 } 1863 } 1864 1865 /* Go back up the rbtree to find next candidate node */ 1866 while (true) { 1867 struct rb_node *prev = &vma->vm_rb; 1868 if (!rb_parent(prev)) 1869 return -ENOMEM; 1870 vma = rb_entry(rb_parent(prev), 1871 struct vm_area_struct, vm_rb); 1872 if (prev == vma->vm_rb.rb_right) { 1873 gap_start = vma->vm_prev ? 1874 vma->vm_prev->vm_end : 0; 1875 goto check_current; 1876 } 1877 } 1878 } 1879 1880 found: 1881 /* We found a suitable gap. Clip it with the original high_limit. */ 1882 if (gap_end > info->high_limit) 1883 gap_end = info->high_limit; 1884 1885 found_highest: 1886 /* Compute highest gap address at the desired alignment */ 1887 gap_end -= info->length; 1888 gap_end -= (gap_end - info->align_offset) & info->align_mask; 1889 1890 VM_BUG_ON(gap_end < info->low_limit); 1891 VM_BUG_ON(gap_end < gap_start); 1892 return gap_end; 1893 } 1894 1895 /* Get an address range which is currently unmapped. 1896 * For shmat() with addr=0. 1897 * 1898 * Ugly calling convention alert: 1899 * Return value with the low bits set means error value, 1900 * ie 1901 * if (ret & ~PAGE_MASK) 1902 * error = ret; 1903 * 1904 * This function "knows" that -ENOMEM has the bits set. 1905 */ 1906 #ifndef HAVE_ARCH_UNMAPPED_AREA 1907 unsigned long 1908 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1909 unsigned long len, unsigned long pgoff, unsigned long flags) 1910 { 1911 struct mm_struct *mm = current->mm; 1912 struct vm_area_struct *vma; 1913 struct vm_unmapped_area_info info; 1914 1915 if (len > TASK_SIZE - mmap_min_addr) 1916 return -ENOMEM; 1917 1918 if (flags & MAP_FIXED) 1919 return addr; 1920 1921 if (addr) { 1922 addr = PAGE_ALIGN(addr); 1923 vma = find_vma(mm, addr); 1924 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1925 (!vma || addr + len <= vma->vm_start)) 1926 return addr; 1927 } 1928 1929 info.flags = 0; 1930 info.length = len; 1931 info.low_limit = mm->mmap_base; 1932 info.high_limit = TASK_SIZE; 1933 info.align_mask = 0; 1934 return vm_unmapped_area(&info); 1935 } 1936 #endif 1937 1938 /* 1939 * This mmap-allocator allocates new areas top-down from below the 1940 * stack's low limit (the base): 1941 */ 1942 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1943 unsigned long 1944 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1945 const unsigned long len, const unsigned long pgoff, 1946 const unsigned long flags) 1947 { 1948 struct vm_area_struct *vma; 1949 struct mm_struct *mm = current->mm; 1950 unsigned long addr = addr0; 1951 struct vm_unmapped_area_info info; 1952 1953 /* requested length too big for entire address space */ 1954 if (len > TASK_SIZE - mmap_min_addr) 1955 return -ENOMEM; 1956 1957 if (flags & MAP_FIXED) 1958 return addr; 1959 1960 /* requesting a specific address */ 1961 if (addr) { 1962 addr = PAGE_ALIGN(addr); 1963 vma = find_vma(mm, addr); 1964 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1965 (!vma || addr + len <= vma->vm_start)) 1966 return addr; 1967 } 1968 1969 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 1970 info.length = len; 1971 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 1972 info.high_limit = mm->mmap_base; 1973 info.align_mask = 0; 1974 addr = vm_unmapped_area(&info); 1975 1976 /* 1977 * A failed mmap() very likely causes application failure, 1978 * so fall back to the bottom-up function here. This scenario 1979 * can happen with large stack limits and large mmap() 1980 * allocations. 1981 */ 1982 if (addr & ~PAGE_MASK) { 1983 VM_BUG_ON(addr != -ENOMEM); 1984 info.flags = 0; 1985 info.low_limit = TASK_UNMAPPED_BASE; 1986 info.high_limit = TASK_SIZE; 1987 addr = vm_unmapped_area(&info); 1988 } 1989 1990 return addr; 1991 } 1992 #endif 1993 1994 unsigned long 1995 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1996 unsigned long pgoff, unsigned long flags) 1997 { 1998 unsigned long (*get_area)(struct file *, unsigned long, 1999 unsigned long, unsigned long, unsigned long); 2000 2001 unsigned long error = arch_mmap_check(addr, len, flags); 2002 if (error) 2003 return error; 2004 2005 /* Careful about overflows.. */ 2006 if (len > TASK_SIZE) 2007 return -ENOMEM; 2008 2009 get_area = current->mm->get_unmapped_area; 2010 if (file && file->f_op->get_unmapped_area) 2011 get_area = file->f_op->get_unmapped_area; 2012 addr = get_area(file, addr, len, pgoff, flags); 2013 if (IS_ERR_VALUE(addr)) 2014 return addr; 2015 2016 if (addr > TASK_SIZE - len) 2017 return -ENOMEM; 2018 if (addr & ~PAGE_MASK) 2019 return -EINVAL; 2020 2021 addr = arch_rebalance_pgtables(addr, len); 2022 error = security_mmap_addr(addr); 2023 return error ? error : addr; 2024 } 2025 2026 EXPORT_SYMBOL(get_unmapped_area); 2027 2028 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 2029 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 2030 { 2031 struct rb_node *rb_node; 2032 struct vm_area_struct *vma; 2033 2034 /* Check the cache first. */ 2035 vma = vmacache_find(mm, addr); 2036 if (likely(vma)) 2037 return vma; 2038 2039 rb_node = mm->mm_rb.rb_node; 2040 vma = NULL; 2041 2042 while (rb_node) { 2043 struct vm_area_struct *tmp; 2044 2045 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2046 2047 if (tmp->vm_end > addr) { 2048 vma = tmp; 2049 if (tmp->vm_start <= addr) 2050 break; 2051 rb_node = rb_node->rb_left; 2052 } else 2053 rb_node = rb_node->rb_right; 2054 } 2055 2056 if (vma) 2057 vmacache_update(addr, vma); 2058 return vma; 2059 } 2060 2061 EXPORT_SYMBOL(find_vma); 2062 2063 /* 2064 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 2065 */ 2066 struct vm_area_struct * 2067 find_vma_prev(struct mm_struct *mm, unsigned long addr, 2068 struct vm_area_struct **pprev) 2069 { 2070 struct vm_area_struct *vma; 2071 2072 vma = find_vma(mm, addr); 2073 if (vma) { 2074 *pprev = vma->vm_prev; 2075 } else { 2076 struct rb_node *rb_node = mm->mm_rb.rb_node; 2077 *pprev = NULL; 2078 while (rb_node) { 2079 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2080 rb_node = rb_node->rb_right; 2081 } 2082 } 2083 return vma; 2084 } 2085 2086 /* 2087 * Verify that the stack growth is acceptable and 2088 * update accounting. This is shared with both the 2089 * grow-up and grow-down cases. 2090 */ 2091 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 2092 { 2093 struct mm_struct *mm = vma->vm_mm; 2094 struct rlimit *rlim = current->signal->rlim; 2095 unsigned long new_start, actual_size; 2096 2097 /* address space limit tests */ 2098 if (!may_expand_vm(mm, grow)) 2099 return -ENOMEM; 2100 2101 /* Stack limit test */ 2102 actual_size = size; 2103 if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN))) 2104 actual_size -= PAGE_SIZE; 2105 if (actual_size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 2106 return -ENOMEM; 2107 2108 /* mlock limit tests */ 2109 if (vma->vm_flags & VM_LOCKED) { 2110 unsigned long locked; 2111 unsigned long limit; 2112 locked = mm->locked_vm + grow; 2113 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 2114 limit >>= PAGE_SHIFT; 2115 if (locked > limit && !capable(CAP_IPC_LOCK)) 2116 return -ENOMEM; 2117 } 2118 2119 /* Check to ensure the stack will not grow into a hugetlb-only region */ 2120 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 2121 vma->vm_end - size; 2122 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 2123 return -EFAULT; 2124 2125 /* 2126 * Overcommit.. This must be the final test, as it will 2127 * update security statistics. 2128 */ 2129 if (security_vm_enough_memory_mm(mm, grow)) 2130 return -ENOMEM; 2131 2132 /* Ok, everything looks good - let it rip */ 2133 if (vma->vm_flags & VM_LOCKED) 2134 mm->locked_vm += grow; 2135 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow); 2136 return 0; 2137 } 2138 2139 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 2140 /* 2141 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 2142 * vma is the last one with address > vma->vm_end. Have to extend vma. 2143 */ 2144 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 2145 { 2146 int error; 2147 2148 if (!(vma->vm_flags & VM_GROWSUP)) 2149 return -EFAULT; 2150 2151 /* 2152 * We must make sure the anon_vma is allocated 2153 * so that the anon_vma locking is not a noop. 2154 */ 2155 if (unlikely(anon_vma_prepare(vma))) 2156 return -ENOMEM; 2157 vma_lock_anon_vma(vma); 2158 2159 /* 2160 * vma->vm_start/vm_end cannot change under us because the caller 2161 * is required to hold the mmap_sem in read mode. We need the 2162 * anon_vma lock to serialize against concurrent expand_stacks. 2163 * Also guard against wrapping around to address 0. 2164 */ 2165 if (address < PAGE_ALIGN(address+4)) 2166 address = PAGE_ALIGN(address+4); 2167 else { 2168 vma_unlock_anon_vma(vma); 2169 return -ENOMEM; 2170 } 2171 error = 0; 2172 2173 /* Somebody else might have raced and expanded it already */ 2174 if (address > vma->vm_end) { 2175 unsigned long size, grow; 2176 2177 size = address - vma->vm_start; 2178 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2179 2180 error = -ENOMEM; 2181 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2182 error = acct_stack_growth(vma, size, grow); 2183 if (!error) { 2184 /* 2185 * vma_gap_update() doesn't support concurrent 2186 * updates, but we only hold a shared mmap_sem 2187 * lock here, so we need to protect against 2188 * concurrent vma expansions. 2189 * vma_lock_anon_vma() doesn't help here, as 2190 * we don't guarantee that all growable vmas 2191 * in a mm share the same root anon vma. 2192 * So, we reuse mm->page_table_lock to guard 2193 * against concurrent vma expansions. 2194 */ 2195 spin_lock(&vma->vm_mm->page_table_lock); 2196 anon_vma_interval_tree_pre_update_vma(vma); 2197 vma->vm_end = address; 2198 anon_vma_interval_tree_post_update_vma(vma); 2199 if (vma->vm_next) 2200 vma_gap_update(vma->vm_next); 2201 else 2202 vma->vm_mm->highest_vm_end = address; 2203 spin_unlock(&vma->vm_mm->page_table_lock); 2204 2205 perf_event_mmap(vma); 2206 } 2207 } 2208 } 2209 vma_unlock_anon_vma(vma); 2210 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2211 validate_mm(vma->vm_mm); 2212 return error; 2213 } 2214 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2215 2216 /* 2217 * vma is the first one with address < vma->vm_start. Have to extend vma. 2218 */ 2219 int expand_downwards(struct vm_area_struct *vma, 2220 unsigned long address) 2221 { 2222 int error; 2223 2224 /* 2225 * We must make sure the anon_vma is allocated 2226 * so that the anon_vma locking is not a noop. 2227 */ 2228 if (unlikely(anon_vma_prepare(vma))) 2229 return -ENOMEM; 2230 2231 address &= PAGE_MASK; 2232 error = security_mmap_addr(address); 2233 if (error) 2234 return error; 2235 2236 vma_lock_anon_vma(vma); 2237 2238 /* 2239 * vma->vm_start/vm_end cannot change under us because the caller 2240 * is required to hold the mmap_sem in read mode. We need the 2241 * anon_vma lock to serialize against concurrent expand_stacks. 2242 */ 2243 2244 /* Somebody else might have raced and expanded it already */ 2245 if (address < vma->vm_start) { 2246 unsigned long size, grow; 2247 2248 size = vma->vm_end - address; 2249 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2250 2251 error = -ENOMEM; 2252 if (grow <= vma->vm_pgoff) { 2253 error = acct_stack_growth(vma, size, grow); 2254 if (!error) { 2255 /* 2256 * vma_gap_update() doesn't support concurrent 2257 * updates, but we only hold a shared mmap_sem 2258 * lock here, so we need to protect against 2259 * concurrent vma expansions. 2260 * vma_lock_anon_vma() doesn't help here, as 2261 * we don't guarantee that all growable vmas 2262 * in a mm share the same root anon vma. 2263 * So, we reuse mm->page_table_lock to guard 2264 * against concurrent vma expansions. 2265 */ 2266 spin_lock(&vma->vm_mm->page_table_lock); 2267 anon_vma_interval_tree_pre_update_vma(vma); 2268 vma->vm_start = address; 2269 vma->vm_pgoff -= grow; 2270 anon_vma_interval_tree_post_update_vma(vma); 2271 vma_gap_update(vma); 2272 spin_unlock(&vma->vm_mm->page_table_lock); 2273 2274 perf_event_mmap(vma); 2275 } 2276 } 2277 } 2278 vma_unlock_anon_vma(vma); 2279 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2280 validate_mm(vma->vm_mm); 2281 return error; 2282 } 2283 2284 /* 2285 * Note how expand_stack() refuses to expand the stack all the way to 2286 * abut the next virtual mapping, *unless* that mapping itself is also 2287 * a stack mapping. We want to leave room for a guard page, after all 2288 * (the guard page itself is not added here, that is done by the 2289 * actual page faulting logic) 2290 * 2291 * This matches the behavior of the guard page logic (see mm/memory.c: 2292 * check_stack_guard_page()), which only allows the guard page to be 2293 * removed under these circumstances. 2294 */ 2295 #ifdef CONFIG_STACK_GROWSUP 2296 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2297 { 2298 struct vm_area_struct *next; 2299 2300 address &= PAGE_MASK; 2301 next = vma->vm_next; 2302 if (next && next->vm_start == address + PAGE_SIZE) { 2303 if (!(next->vm_flags & VM_GROWSUP)) 2304 return -ENOMEM; 2305 } 2306 return expand_upwards(vma, address); 2307 } 2308 2309 struct vm_area_struct * 2310 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2311 { 2312 struct vm_area_struct *vma, *prev; 2313 2314 addr &= PAGE_MASK; 2315 vma = find_vma_prev(mm, addr, &prev); 2316 if (vma && (vma->vm_start <= addr)) 2317 return vma; 2318 if (!prev || expand_stack(prev, addr)) 2319 return NULL; 2320 if (prev->vm_flags & VM_LOCKED) 2321 __mlock_vma_pages_range(prev, addr, prev->vm_end, NULL); 2322 return prev; 2323 } 2324 #else 2325 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2326 { 2327 struct vm_area_struct *prev; 2328 2329 address &= PAGE_MASK; 2330 prev = vma->vm_prev; 2331 if (prev && prev->vm_end == address) { 2332 if (!(prev->vm_flags & VM_GROWSDOWN)) 2333 return -ENOMEM; 2334 } 2335 return expand_downwards(vma, address); 2336 } 2337 2338 struct vm_area_struct * 2339 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2340 { 2341 struct vm_area_struct *vma; 2342 unsigned long start; 2343 2344 addr &= PAGE_MASK; 2345 vma = find_vma(mm, addr); 2346 if (!vma) 2347 return NULL; 2348 if (vma->vm_start <= addr) 2349 return vma; 2350 if (!(vma->vm_flags & VM_GROWSDOWN)) 2351 return NULL; 2352 start = vma->vm_start; 2353 if (expand_stack(vma, addr)) 2354 return NULL; 2355 if (vma->vm_flags & VM_LOCKED) 2356 __mlock_vma_pages_range(vma, addr, start, NULL); 2357 return vma; 2358 } 2359 #endif 2360 2361 EXPORT_SYMBOL_GPL(find_extend_vma); 2362 2363 /* 2364 * Ok - we have the memory areas we should free on the vma list, 2365 * so release them, and do the vma updates. 2366 * 2367 * Called with the mm semaphore held. 2368 */ 2369 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 2370 { 2371 unsigned long nr_accounted = 0; 2372 2373 /* Update high watermark before we lower total_vm */ 2374 update_hiwater_vm(mm); 2375 do { 2376 long nrpages = vma_pages(vma); 2377 2378 if (vma->vm_flags & VM_ACCOUNT) 2379 nr_accounted += nrpages; 2380 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages); 2381 vma = remove_vma(vma); 2382 } while (vma); 2383 vm_unacct_memory(nr_accounted); 2384 validate_mm(mm); 2385 } 2386 2387 /* 2388 * Get rid of page table information in the indicated region. 2389 * 2390 * Called with the mm semaphore held. 2391 */ 2392 static void unmap_region(struct mm_struct *mm, 2393 struct vm_area_struct *vma, struct vm_area_struct *prev, 2394 unsigned long start, unsigned long end) 2395 { 2396 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap; 2397 struct mmu_gather tlb; 2398 2399 lru_add_drain(); 2400 tlb_gather_mmu(&tlb, mm, start, end); 2401 update_hiwater_rss(mm); 2402 unmap_vmas(&tlb, vma, start, end); 2403 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2404 next ? next->vm_start : USER_PGTABLES_CEILING); 2405 tlb_finish_mmu(&tlb, start, end); 2406 } 2407 2408 /* 2409 * Create a list of vma's touched by the unmap, removing them from the mm's 2410 * vma list as we go.. 2411 */ 2412 static void 2413 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 2414 struct vm_area_struct *prev, unsigned long end) 2415 { 2416 struct vm_area_struct **insertion_point; 2417 struct vm_area_struct *tail_vma = NULL; 2418 2419 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 2420 vma->vm_prev = NULL; 2421 do { 2422 vma_rb_erase(vma, &mm->mm_rb); 2423 mm->map_count--; 2424 tail_vma = vma; 2425 vma = vma->vm_next; 2426 } while (vma && vma->vm_start < end); 2427 *insertion_point = vma; 2428 if (vma) { 2429 vma->vm_prev = prev; 2430 vma_gap_update(vma); 2431 } else 2432 mm->highest_vm_end = prev ? prev->vm_end : 0; 2433 tail_vma->vm_next = NULL; 2434 2435 /* Kill the cache */ 2436 vmacache_invalidate(mm); 2437 } 2438 2439 /* 2440 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the 2441 * munmap path where it doesn't make sense to fail. 2442 */ 2443 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2444 unsigned long addr, int new_below) 2445 { 2446 struct vm_area_struct *new; 2447 int err = -ENOMEM; 2448 2449 if (is_vm_hugetlb_page(vma) && (addr & 2450 ~(huge_page_mask(hstate_vma(vma))))) 2451 return -EINVAL; 2452 2453 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2454 if (!new) 2455 goto out_err; 2456 2457 /* most fields are the same, copy all, and then fixup */ 2458 *new = *vma; 2459 2460 INIT_LIST_HEAD(&new->anon_vma_chain); 2461 2462 if (new_below) 2463 new->vm_end = addr; 2464 else { 2465 new->vm_start = addr; 2466 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2467 } 2468 2469 err = vma_dup_policy(vma, new); 2470 if (err) 2471 goto out_free_vma; 2472 2473 err = anon_vma_clone(new, vma); 2474 if (err) 2475 goto out_free_mpol; 2476 2477 if (new->vm_file) 2478 get_file(new->vm_file); 2479 2480 if (new->vm_ops && new->vm_ops->open) 2481 new->vm_ops->open(new); 2482 2483 if (new_below) 2484 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 2485 ((addr - new->vm_start) >> PAGE_SHIFT), new); 2486 else 2487 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 2488 2489 /* Success. */ 2490 if (!err) 2491 return 0; 2492 2493 /* Clean everything up if vma_adjust failed. */ 2494 if (new->vm_ops && new->vm_ops->close) 2495 new->vm_ops->close(new); 2496 if (new->vm_file) 2497 fput(new->vm_file); 2498 unlink_anon_vmas(new); 2499 out_free_mpol: 2500 mpol_put(vma_policy(new)); 2501 out_free_vma: 2502 kmem_cache_free(vm_area_cachep, new); 2503 out_err: 2504 return err; 2505 } 2506 2507 /* 2508 * Split a vma into two pieces at address 'addr', a new vma is allocated 2509 * either for the first part or the tail. 2510 */ 2511 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2512 unsigned long addr, int new_below) 2513 { 2514 if (mm->map_count >= sysctl_max_map_count) 2515 return -ENOMEM; 2516 2517 return __split_vma(mm, vma, addr, new_below); 2518 } 2519 2520 /* Munmap is split into 2 main parts -- this part which finds 2521 * what needs doing, and the areas themselves, which do the 2522 * work. This now handles partial unmappings. 2523 * Jeremy Fitzhardinge <jeremy@goop.org> 2524 */ 2525 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 2526 { 2527 unsigned long end; 2528 struct vm_area_struct *vma, *prev, *last; 2529 2530 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start) 2531 return -EINVAL; 2532 2533 len = PAGE_ALIGN(len); 2534 if (len == 0) 2535 return -EINVAL; 2536 2537 /* Find the first overlapping VMA */ 2538 vma = find_vma(mm, start); 2539 if (!vma) 2540 return 0; 2541 prev = vma->vm_prev; 2542 /* we have start < vma->vm_end */ 2543 2544 /* if it doesn't overlap, we have nothing.. */ 2545 end = start + len; 2546 if (vma->vm_start >= end) 2547 return 0; 2548 2549 /* 2550 * If we need to split any vma, do it now to save pain later. 2551 * 2552 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2553 * unmapped vm_area_struct will remain in use: so lower split_vma 2554 * places tmp vma above, and higher split_vma places tmp vma below. 2555 */ 2556 if (start > vma->vm_start) { 2557 int error; 2558 2559 /* 2560 * Make sure that map_count on return from munmap() will 2561 * not exceed its limit; but let map_count go just above 2562 * its limit temporarily, to help free resources as expected. 2563 */ 2564 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2565 return -ENOMEM; 2566 2567 error = __split_vma(mm, vma, start, 0); 2568 if (error) 2569 return error; 2570 prev = vma; 2571 } 2572 2573 /* Does it split the last one? */ 2574 last = find_vma(mm, end); 2575 if (last && end > last->vm_start) { 2576 int error = __split_vma(mm, last, end, 1); 2577 if (error) 2578 return error; 2579 } 2580 vma = prev ? prev->vm_next : mm->mmap; 2581 2582 /* 2583 * unlock any mlock()ed ranges before detaching vmas 2584 */ 2585 if (mm->locked_vm) { 2586 struct vm_area_struct *tmp = vma; 2587 while (tmp && tmp->vm_start < end) { 2588 if (tmp->vm_flags & VM_LOCKED) { 2589 mm->locked_vm -= vma_pages(tmp); 2590 munlock_vma_pages_all(tmp); 2591 } 2592 tmp = tmp->vm_next; 2593 } 2594 } 2595 2596 /* 2597 * Remove the vma's, and unmap the actual pages 2598 */ 2599 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2600 unmap_region(mm, vma, prev, start, end); 2601 2602 arch_unmap(mm, vma, start, end); 2603 2604 /* Fix up all other VM information */ 2605 remove_vma_list(mm, vma); 2606 2607 return 0; 2608 } 2609 2610 int vm_munmap(unsigned long start, size_t len) 2611 { 2612 int ret; 2613 struct mm_struct *mm = current->mm; 2614 2615 down_write(&mm->mmap_sem); 2616 ret = do_munmap(mm, start, len); 2617 up_write(&mm->mmap_sem); 2618 return ret; 2619 } 2620 EXPORT_SYMBOL(vm_munmap); 2621 2622 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2623 { 2624 profile_munmap(addr); 2625 return vm_munmap(addr, len); 2626 } 2627 2628 2629 /* 2630 * Emulation of deprecated remap_file_pages() syscall. 2631 */ 2632 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size, 2633 unsigned long, prot, unsigned long, pgoff, unsigned long, flags) 2634 { 2635 2636 struct mm_struct *mm = current->mm; 2637 struct vm_area_struct *vma; 2638 unsigned long populate = 0; 2639 unsigned long ret = -EINVAL; 2640 struct file *file; 2641 2642 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. " 2643 "See Documentation/vm/remap_file_pages.txt.\n", 2644 current->comm, current->pid); 2645 2646 if (prot) 2647 return ret; 2648 start = start & PAGE_MASK; 2649 size = size & PAGE_MASK; 2650 2651 if (start + size <= start) 2652 return ret; 2653 2654 /* Does pgoff wrap? */ 2655 if (pgoff + (size >> PAGE_SHIFT) < pgoff) 2656 return ret; 2657 2658 down_write(&mm->mmap_sem); 2659 vma = find_vma(mm, start); 2660 2661 if (!vma || !(vma->vm_flags & VM_SHARED)) 2662 goto out; 2663 2664 if (start < vma->vm_start || start + size > vma->vm_end) 2665 goto out; 2666 2667 if (pgoff == linear_page_index(vma, start)) { 2668 ret = 0; 2669 goto out; 2670 } 2671 2672 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0; 2673 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0; 2674 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0; 2675 2676 flags &= MAP_NONBLOCK; 2677 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE; 2678 if (vma->vm_flags & VM_LOCKED) { 2679 flags |= MAP_LOCKED; 2680 /* drop PG_Mlocked flag for over-mapped range */ 2681 munlock_vma_pages_range(vma, start, start + size); 2682 } 2683 2684 file = get_file(vma->vm_file); 2685 ret = do_mmap_pgoff(vma->vm_file, start, size, 2686 prot, flags, pgoff, &populate); 2687 fput(file); 2688 out: 2689 up_write(&mm->mmap_sem); 2690 if (populate) 2691 mm_populate(ret, populate); 2692 if (!IS_ERR_VALUE(ret)) 2693 ret = 0; 2694 return ret; 2695 } 2696 2697 static inline void verify_mm_writelocked(struct mm_struct *mm) 2698 { 2699 #ifdef CONFIG_DEBUG_VM 2700 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2701 WARN_ON(1); 2702 up_read(&mm->mmap_sem); 2703 } 2704 #endif 2705 } 2706 2707 /* 2708 * this is really a simplified "do_mmap". it only handles 2709 * anonymous maps. eventually we may be able to do some 2710 * brk-specific accounting here. 2711 */ 2712 static unsigned long do_brk(unsigned long addr, unsigned long len) 2713 { 2714 struct mm_struct *mm = current->mm; 2715 struct vm_area_struct *vma, *prev; 2716 unsigned long flags; 2717 struct rb_node **rb_link, *rb_parent; 2718 pgoff_t pgoff = addr >> PAGE_SHIFT; 2719 int error; 2720 2721 len = PAGE_ALIGN(len); 2722 if (!len) 2723 return addr; 2724 2725 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2726 2727 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2728 if (error & ~PAGE_MASK) 2729 return error; 2730 2731 error = mlock_future_check(mm, mm->def_flags, len); 2732 if (error) 2733 return error; 2734 2735 /* 2736 * mm->mmap_sem is required to protect against another thread 2737 * changing the mappings in case we sleep. 2738 */ 2739 verify_mm_writelocked(mm); 2740 2741 /* 2742 * Clear old maps. this also does some error checking for us 2743 */ 2744 munmap_back: 2745 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) { 2746 if (do_munmap(mm, addr, len)) 2747 return -ENOMEM; 2748 goto munmap_back; 2749 } 2750 2751 /* Check against address space limits *after* clearing old maps... */ 2752 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) 2753 return -ENOMEM; 2754 2755 if (mm->map_count > sysctl_max_map_count) 2756 return -ENOMEM; 2757 2758 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 2759 return -ENOMEM; 2760 2761 /* Can we just expand an old private anonymous mapping? */ 2762 vma = vma_merge(mm, prev, addr, addr + len, flags, 2763 NULL, NULL, pgoff, NULL); 2764 if (vma) 2765 goto out; 2766 2767 /* 2768 * create a vma struct for an anonymous mapping 2769 */ 2770 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2771 if (!vma) { 2772 vm_unacct_memory(len >> PAGE_SHIFT); 2773 return -ENOMEM; 2774 } 2775 2776 INIT_LIST_HEAD(&vma->anon_vma_chain); 2777 vma->vm_mm = mm; 2778 vma->vm_start = addr; 2779 vma->vm_end = addr + len; 2780 vma->vm_pgoff = pgoff; 2781 vma->vm_flags = flags; 2782 vma->vm_page_prot = vm_get_page_prot(flags); 2783 vma_link(mm, vma, prev, rb_link, rb_parent); 2784 out: 2785 perf_event_mmap(vma); 2786 mm->total_vm += len >> PAGE_SHIFT; 2787 if (flags & VM_LOCKED) 2788 mm->locked_vm += (len >> PAGE_SHIFT); 2789 vma->vm_flags |= VM_SOFTDIRTY; 2790 return addr; 2791 } 2792 2793 unsigned long vm_brk(unsigned long addr, unsigned long len) 2794 { 2795 struct mm_struct *mm = current->mm; 2796 unsigned long ret; 2797 bool populate; 2798 2799 down_write(&mm->mmap_sem); 2800 ret = do_brk(addr, len); 2801 populate = ((mm->def_flags & VM_LOCKED) != 0); 2802 up_write(&mm->mmap_sem); 2803 if (populate) 2804 mm_populate(addr, len); 2805 return ret; 2806 } 2807 EXPORT_SYMBOL(vm_brk); 2808 2809 /* Release all mmaps. */ 2810 void exit_mmap(struct mm_struct *mm) 2811 { 2812 struct mmu_gather tlb; 2813 struct vm_area_struct *vma; 2814 unsigned long nr_accounted = 0; 2815 2816 /* mm's last user has gone, and its about to be pulled down */ 2817 mmu_notifier_release(mm); 2818 2819 if (mm->locked_vm) { 2820 vma = mm->mmap; 2821 while (vma) { 2822 if (vma->vm_flags & VM_LOCKED) 2823 munlock_vma_pages_all(vma); 2824 vma = vma->vm_next; 2825 } 2826 } 2827 2828 arch_exit_mmap(mm); 2829 2830 vma = mm->mmap; 2831 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2832 return; 2833 2834 lru_add_drain(); 2835 flush_cache_mm(mm); 2836 tlb_gather_mmu(&tlb, mm, 0, -1); 2837 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2838 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2839 unmap_vmas(&tlb, vma, 0, -1); 2840 2841 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING); 2842 tlb_finish_mmu(&tlb, 0, -1); 2843 2844 /* 2845 * Walk the list again, actually closing and freeing it, 2846 * with preemption enabled, without holding any MM locks. 2847 */ 2848 while (vma) { 2849 if (vma->vm_flags & VM_ACCOUNT) 2850 nr_accounted += vma_pages(vma); 2851 vma = remove_vma(vma); 2852 } 2853 vm_unacct_memory(nr_accounted); 2854 } 2855 2856 /* Insert vm structure into process list sorted by address 2857 * and into the inode's i_mmap tree. If vm_file is non-NULL 2858 * then i_mmap_rwsem is taken here. 2859 */ 2860 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 2861 { 2862 struct vm_area_struct *prev; 2863 struct rb_node **rb_link, *rb_parent; 2864 2865 /* 2866 * The vm_pgoff of a purely anonymous vma should be irrelevant 2867 * until its first write fault, when page's anon_vma and index 2868 * are set. But now set the vm_pgoff it will almost certainly 2869 * end up with (unless mremap moves it elsewhere before that 2870 * first wfault), so /proc/pid/maps tells a consistent story. 2871 * 2872 * By setting it to reflect the virtual start address of the 2873 * vma, merges and splits can happen in a seamless way, just 2874 * using the existing file pgoff checks and manipulations. 2875 * Similarly in do_mmap_pgoff and in do_brk. 2876 */ 2877 if (!vma->vm_file) { 2878 BUG_ON(vma->anon_vma); 2879 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2880 } 2881 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 2882 &prev, &rb_link, &rb_parent)) 2883 return -ENOMEM; 2884 if ((vma->vm_flags & VM_ACCOUNT) && 2885 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2886 return -ENOMEM; 2887 2888 vma_link(mm, vma, prev, rb_link, rb_parent); 2889 return 0; 2890 } 2891 2892 /* 2893 * Copy the vma structure to a new location in the same mm, 2894 * prior to moving page table entries, to effect an mremap move. 2895 */ 2896 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2897 unsigned long addr, unsigned long len, pgoff_t pgoff, 2898 bool *need_rmap_locks) 2899 { 2900 struct vm_area_struct *vma = *vmap; 2901 unsigned long vma_start = vma->vm_start; 2902 struct mm_struct *mm = vma->vm_mm; 2903 struct vm_area_struct *new_vma, *prev; 2904 struct rb_node **rb_link, *rb_parent; 2905 bool faulted_in_anon_vma = true; 2906 2907 /* 2908 * If anonymous vma has not yet been faulted, update new pgoff 2909 * to match new location, to increase its chance of merging. 2910 */ 2911 if (unlikely(!vma->vm_file && !vma->anon_vma)) { 2912 pgoff = addr >> PAGE_SHIFT; 2913 faulted_in_anon_vma = false; 2914 } 2915 2916 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) 2917 return NULL; /* should never get here */ 2918 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2919 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma)); 2920 if (new_vma) { 2921 /* 2922 * Source vma may have been merged into new_vma 2923 */ 2924 if (unlikely(vma_start >= new_vma->vm_start && 2925 vma_start < new_vma->vm_end)) { 2926 /* 2927 * The only way we can get a vma_merge with 2928 * self during an mremap is if the vma hasn't 2929 * been faulted in yet and we were allowed to 2930 * reset the dst vma->vm_pgoff to the 2931 * destination address of the mremap to allow 2932 * the merge to happen. mremap must change the 2933 * vm_pgoff linearity between src and dst vmas 2934 * (in turn preventing a vma_merge) to be 2935 * safe. It is only safe to keep the vm_pgoff 2936 * linear if there are no pages mapped yet. 2937 */ 2938 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 2939 *vmap = vma = new_vma; 2940 } 2941 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 2942 } else { 2943 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2944 if (new_vma) { 2945 *new_vma = *vma; 2946 new_vma->vm_start = addr; 2947 new_vma->vm_end = addr + len; 2948 new_vma->vm_pgoff = pgoff; 2949 if (vma_dup_policy(vma, new_vma)) 2950 goto out_free_vma; 2951 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 2952 if (anon_vma_clone(new_vma, vma)) 2953 goto out_free_mempol; 2954 if (new_vma->vm_file) 2955 get_file(new_vma->vm_file); 2956 if (new_vma->vm_ops && new_vma->vm_ops->open) 2957 new_vma->vm_ops->open(new_vma); 2958 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2959 *need_rmap_locks = false; 2960 } 2961 } 2962 return new_vma; 2963 2964 out_free_mempol: 2965 mpol_put(vma_policy(new_vma)); 2966 out_free_vma: 2967 kmem_cache_free(vm_area_cachep, new_vma); 2968 return NULL; 2969 } 2970 2971 /* 2972 * Return true if the calling process may expand its vm space by the passed 2973 * number of pages 2974 */ 2975 int may_expand_vm(struct mm_struct *mm, unsigned long npages) 2976 { 2977 unsigned long cur = mm->total_vm; /* pages */ 2978 unsigned long lim; 2979 2980 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT; 2981 2982 if (cur + npages > lim) 2983 return 0; 2984 return 1; 2985 } 2986 2987 static int special_mapping_fault(struct vm_area_struct *vma, 2988 struct vm_fault *vmf); 2989 2990 /* 2991 * Having a close hook prevents vma merging regardless of flags. 2992 */ 2993 static void special_mapping_close(struct vm_area_struct *vma) 2994 { 2995 } 2996 2997 static const char *special_mapping_name(struct vm_area_struct *vma) 2998 { 2999 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 3000 } 3001 3002 static const struct vm_operations_struct special_mapping_vmops = { 3003 .close = special_mapping_close, 3004 .fault = special_mapping_fault, 3005 .name = special_mapping_name, 3006 }; 3007 3008 static const struct vm_operations_struct legacy_special_mapping_vmops = { 3009 .close = special_mapping_close, 3010 .fault = special_mapping_fault, 3011 }; 3012 3013 static int special_mapping_fault(struct vm_area_struct *vma, 3014 struct vm_fault *vmf) 3015 { 3016 pgoff_t pgoff; 3017 struct page **pages; 3018 3019 /* 3020 * special mappings have no vm_file, and in that case, the mm 3021 * uses vm_pgoff internally. So we have to subtract it from here. 3022 * We are allowed to do this because we are the mm; do not copy 3023 * this code into drivers! 3024 */ 3025 pgoff = vmf->pgoff - vma->vm_pgoff; 3026 3027 if (vma->vm_ops == &legacy_special_mapping_vmops) 3028 pages = vma->vm_private_data; 3029 else 3030 pages = ((struct vm_special_mapping *)vma->vm_private_data)-> 3031 pages; 3032 3033 for (; pgoff && *pages; ++pages) 3034 pgoff--; 3035 3036 if (*pages) { 3037 struct page *page = *pages; 3038 get_page(page); 3039 vmf->page = page; 3040 return 0; 3041 } 3042 3043 return VM_FAULT_SIGBUS; 3044 } 3045 3046 static struct vm_area_struct *__install_special_mapping( 3047 struct mm_struct *mm, 3048 unsigned long addr, unsigned long len, 3049 unsigned long vm_flags, const struct vm_operations_struct *ops, 3050 void *priv) 3051 { 3052 int ret; 3053 struct vm_area_struct *vma; 3054 3055 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 3056 if (unlikely(vma == NULL)) 3057 return ERR_PTR(-ENOMEM); 3058 3059 INIT_LIST_HEAD(&vma->anon_vma_chain); 3060 vma->vm_mm = mm; 3061 vma->vm_start = addr; 3062 vma->vm_end = addr + len; 3063 3064 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY; 3065 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3066 3067 vma->vm_ops = ops; 3068 vma->vm_private_data = priv; 3069 3070 ret = insert_vm_struct(mm, vma); 3071 if (ret) 3072 goto out; 3073 3074 mm->total_vm += len >> PAGE_SHIFT; 3075 3076 perf_event_mmap(vma); 3077 3078 return vma; 3079 3080 out: 3081 kmem_cache_free(vm_area_cachep, vma); 3082 return ERR_PTR(ret); 3083 } 3084 3085 /* 3086 * Called with mm->mmap_sem held for writing. 3087 * Insert a new vma covering the given region, with the given flags. 3088 * Its pages are supplied by the given array of struct page *. 3089 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3090 * The region past the last page supplied will always produce SIGBUS. 3091 * The array pointer and the pages it points to are assumed to stay alive 3092 * for as long as this mapping might exist. 3093 */ 3094 struct vm_area_struct *_install_special_mapping( 3095 struct mm_struct *mm, 3096 unsigned long addr, unsigned long len, 3097 unsigned long vm_flags, const struct vm_special_mapping *spec) 3098 { 3099 return __install_special_mapping(mm, addr, len, vm_flags, 3100 &special_mapping_vmops, (void *)spec); 3101 } 3102 3103 int install_special_mapping(struct mm_struct *mm, 3104 unsigned long addr, unsigned long len, 3105 unsigned long vm_flags, struct page **pages) 3106 { 3107 struct vm_area_struct *vma = __install_special_mapping( 3108 mm, addr, len, vm_flags, &legacy_special_mapping_vmops, 3109 (void *)pages); 3110 3111 return PTR_ERR_OR_ZERO(vma); 3112 } 3113 3114 static DEFINE_MUTEX(mm_all_locks_mutex); 3115 3116 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3117 { 3118 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3119 /* 3120 * The LSB of head.next can't change from under us 3121 * because we hold the mm_all_locks_mutex. 3122 */ 3123 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem); 3124 /* 3125 * We can safely modify head.next after taking the 3126 * anon_vma->root->rwsem. If some other vma in this mm shares 3127 * the same anon_vma we won't take it again. 3128 * 3129 * No need of atomic instructions here, head.next 3130 * can't change from under us thanks to the 3131 * anon_vma->root->rwsem. 3132 */ 3133 if (__test_and_set_bit(0, (unsigned long *) 3134 &anon_vma->root->rb_root.rb_node)) 3135 BUG(); 3136 } 3137 } 3138 3139 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3140 { 3141 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3142 /* 3143 * AS_MM_ALL_LOCKS can't change from under us because 3144 * we hold the mm_all_locks_mutex. 3145 * 3146 * Operations on ->flags have to be atomic because 3147 * even if AS_MM_ALL_LOCKS is stable thanks to the 3148 * mm_all_locks_mutex, there may be other cpus 3149 * changing other bitflags in parallel to us. 3150 */ 3151 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3152 BUG(); 3153 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem); 3154 } 3155 } 3156 3157 /* 3158 * This operation locks against the VM for all pte/vma/mm related 3159 * operations that could ever happen on a certain mm. This includes 3160 * vmtruncate, try_to_unmap, and all page faults. 3161 * 3162 * The caller must take the mmap_sem in write mode before calling 3163 * mm_take_all_locks(). The caller isn't allowed to release the 3164 * mmap_sem until mm_drop_all_locks() returns. 3165 * 3166 * mmap_sem in write mode is required in order to block all operations 3167 * that could modify pagetables and free pages without need of 3168 * altering the vma layout. It's also needed in write mode to avoid new 3169 * anon_vmas to be associated with existing vmas. 3170 * 3171 * A single task can't take more than one mm_take_all_locks() in a row 3172 * or it would deadlock. 3173 * 3174 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3175 * mapping->flags avoid to take the same lock twice, if more than one 3176 * vma in this mm is backed by the same anon_vma or address_space. 3177 * 3178 * We can take all the locks in random order because the VM code 3179 * taking i_mmap_rwsem or anon_vma->rwsem outside the mmap_sem never 3180 * takes more than one of them in a row. Secondly we're protected 3181 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex. 3182 * 3183 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3184 * that may have to take thousand of locks. 3185 * 3186 * mm_take_all_locks() can fail if it's interrupted by signals. 3187 */ 3188 int mm_take_all_locks(struct mm_struct *mm) 3189 { 3190 struct vm_area_struct *vma; 3191 struct anon_vma_chain *avc; 3192 3193 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3194 3195 mutex_lock(&mm_all_locks_mutex); 3196 3197 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3198 if (signal_pending(current)) 3199 goto out_unlock; 3200 if (vma->vm_file && vma->vm_file->f_mapping) 3201 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3202 } 3203 3204 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3205 if (signal_pending(current)) 3206 goto out_unlock; 3207 if (vma->anon_vma) 3208 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3209 vm_lock_anon_vma(mm, avc->anon_vma); 3210 } 3211 3212 return 0; 3213 3214 out_unlock: 3215 mm_drop_all_locks(mm); 3216 return -EINTR; 3217 } 3218 3219 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3220 { 3221 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3222 /* 3223 * The LSB of head.next can't change to 0 from under 3224 * us because we hold the mm_all_locks_mutex. 3225 * 3226 * We must however clear the bitflag before unlocking 3227 * the vma so the users using the anon_vma->rb_root will 3228 * never see our bitflag. 3229 * 3230 * No need of atomic instructions here, head.next 3231 * can't change from under us until we release the 3232 * anon_vma->root->rwsem. 3233 */ 3234 if (!__test_and_clear_bit(0, (unsigned long *) 3235 &anon_vma->root->rb_root.rb_node)) 3236 BUG(); 3237 anon_vma_unlock_write(anon_vma); 3238 } 3239 } 3240 3241 static void vm_unlock_mapping(struct address_space *mapping) 3242 { 3243 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3244 /* 3245 * AS_MM_ALL_LOCKS can't change to 0 from under us 3246 * because we hold the mm_all_locks_mutex. 3247 */ 3248 i_mmap_unlock_write(mapping); 3249 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3250 &mapping->flags)) 3251 BUG(); 3252 } 3253 } 3254 3255 /* 3256 * The mmap_sem cannot be released by the caller until 3257 * mm_drop_all_locks() returns. 3258 */ 3259 void mm_drop_all_locks(struct mm_struct *mm) 3260 { 3261 struct vm_area_struct *vma; 3262 struct anon_vma_chain *avc; 3263 3264 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3265 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3266 3267 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3268 if (vma->anon_vma) 3269 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3270 vm_unlock_anon_vma(avc->anon_vma); 3271 if (vma->vm_file && vma->vm_file->f_mapping) 3272 vm_unlock_mapping(vma->vm_file->f_mapping); 3273 } 3274 3275 mutex_unlock(&mm_all_locks_mutex); 3276 } 3277 3278 /* 3279 * initialise the VMA slab 3280 */ 3281 void __init mmap_init(void) 3282 { 3283 int ret; 3284 3285 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3286 VM_BUG_ON(ret); 3287 } 3288 3289 /* 3290 * Initialise sysctl_user_reserve_kbytes. 3291 * 3292 * This is intended to prevent a user from starting a single memory hogging 3293 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3294 * mode. 3295 * 3296 * The default value is min(3% of free memory, 128MB) 3297 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3298 */ 3299 static int init_user_reserve(void) 3300 { 3301 unsigned long free_kbytes; 3302 3303 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3304 3305 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3306 return 0; 3307 } 3308 subsys_initcall(init_user_reserve); 3309 3310 /* 3311 * Initialise sysctl_admin_reserve_kbytes. 3312 * 3313 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3314 * to log in and kill a memory hogging process. 3315 * 3316 * Systems with more than 256MB will reserve 8MB, enough to recover 3317 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3318 * only reserve 3% of free pages by default. 3319 */ 3320 static int init_admin_reserve(void) 3321 { 3322 unsigned long free_kbytes; 3323 3324 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3325 3326 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3327 return 0; 3328 } 3329 subsys_initcall(init_admin_reserve); 3330 3331 /* 3332 * Reinititalise user and admin reserves if memory is added or removed. 3333 * 3334 * The default user reserve max is 128MB, and the default max for the 3335 * admin reserve is 8MB. These are usually, but not always, enough to 3336 * enable recovery from a memory hogging process using login/sshd, a shell, 3337 * and tools like top. It may make sense to increase or even disable the 3338 * reserve depending on the existence of swap or variations in the recovery 3339 * tools. So, the admin may have changed them. 3340 * 3341 * If memory is added and the reserves have been eliminated or increased above 3342 * the default max, then we'll trust the admin. 3343 * 3344 * If memory is removed and there isn't enough free memory, then we 3345 * need to reset the reserves. 3346 * 3347 * Otherwise keep the reserve set by the admin. 3348 */ 3349 static int reserve_mem_notifier(struct notifier_block *nb, 3350 unsigned long action, void *data) 3351 { 3352 unsigned long tmp, free_kbytes; 3353 3354 switch (action) { 3355 case MEM_ONLINE: 3356 /* Default max is 128MB. Leave alone if modified by operator. */ 3357 tmp = sysctl_user_reserve_kbytes; 3358 if (0 < tmp && tmp < (1UL << 17)) 3359 init_user_reserve(); 3360 3361 /* Default max is 8MB. Leave alone if modified by operator. */ 3362 tmp = sysctl_admin_reserve_kbytes; 3363 if (0 < tmp && tmp < (1UL << 13)) 3364 init_admin_reserve(); 3365 3366 break; 3367 case MEM_OFFLINE: 3368 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3369 3370 if (sysctl_user_reserve_kbytes > free_kbytes) { 3371 init_user_reserve(); 3372 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3373 sysctl_user_reserve_kbytes); 3374 } 3375 3376 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3377 init_admin_reserve(); 3378 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3379 sysctl_admin_reserve_kbytes); 3380 } 3381 break; 3382 default: 3383 break; 3384 } 3385 return NOTIFY_OK; 3386 } 3387 3388 static struct notifier_block reserve_mem_nb = { 3389 .notifier_call = reserve_mem_notifier, 3390 }; 3391 3392 static int __meminit init_reserve_notifier(void) 3393 { 3394 if (register_hotmemory_notifier(&reserve_mem_nb)) 3395 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3396 3397 return 0; 3398 } 3399 subsys_initcall(init_reserve_notifier); 3400