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 #include <linux/userfaultfd_k.h> 45 #include <linux/moduleparam.h> 46 47 #include <asm/uaccess.h> 48 #include <asm/cacheflush.h> 49 #include <asm/tlb.h> 50 #include <asm/mmu_context.h> 51 52 #include "internal.h" 53 54 #ifndef arch_mmap_check 55 #define arch_mmap_check(addr, len, flags) (0) 56 #endif 57 58 #ifndef arch_rebalance_pgtables 59 #define arch_rebalance_pgtables(addr, len) (addr) 60 #endif 61 62 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS 63 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN; 64 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX; 65 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS; 66 #endif 67 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS 68 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN; 69 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX; 70 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS; 71 #endif 72 73 static bool ignore_rlimit_data = true; 74 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644); 75 76 static void unmap_region(struct mm_struct *mm, 77 struct vm_area_struct *vma, struct vm_area_struct *prev, 78 unsigned long start, unsigned long end); 79 80 /* description of effects of mapping type and prot in current implementation. 81 * this is due to the limited x86 page protection hardware. The expected 82 * behavior is in parens: 83 * 84 * map_type prot 85 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 86 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 87 * w: (no) no w: (no) no w: (yes) yes w: (no) no 88 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 89 * 90 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 91 * w: (no) no w: (no) no w: (copy) copy w: (no) no 92 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 93 * 94 */ 95 pgprot_t protection_map[16] = { 96 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 97 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 98 }; 99 100 pgprot_t vm_get_page_prot(unsigned long vm_flags) 101 { 102 return __pgprot(pgprot_val(protection_map[vm_flags & 103 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) | 104 pgprot_val(arch_vm_get_page_prot(vm_flags))); 105 } 106 EXPORT_SYMBOL(vm_get_page_prot); 107 108 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags) 109 { 110 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags)); 111 } 112 113 /* Update vma->vm_page_prot to reflect vma->vm_flags. */ 114 void vma_set_page_prot(struct vm_area_struct *vma) 115 { 116 unsigned long vm_flags = vma->vm_flags; 117 118 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags); 119 if (vma_wants_writenotify(vma)) { 120 vm_flags &= ~VM_SHARED; 121 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, 122 vm_flags); 123 } 124 } 125 126 127 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */ 128 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */ 129 unsigned long sysctl_overcommit_kbytes __read_mostly; 130 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT; 131 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */ 132 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */ 133 /* 134 * Make sure vm_committed_as in one cacheline and not cacheline shared with 135 * other variables. It can be updated by several CPUs frequently. 136 */ 137 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp; 138 139 /* 140 * The global memory commitment made in the system can be a metric 141 * that can be used to drive ballooning decisions when Linux is hosted 142 * as a guest. On Hyper-V, the host implements a policy engine for dynamically 143 * balancing memory across competing virtual machines that are hosted. 144 * Several metrics drive this policy engine including the guest reported 145 * memory commitment. 146 */ 147 unsigned long vm_memory_committed(void) 148 { 149 return percpu_counter_read_positive(&vm_committed_as); 150 } 151 EXPORT_SYMBOL_GPL(vm_memory_committed); 152 153 /* 154 * Check that a process has enough memory to allocate a new virtual 155 * mapping. 0 means there is enough memory for the allocation to 156 * succeed and -ENOMEM implies there is not. 157 * 158 * We currently support three overcommit policies, which are set via the 159 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting 160 * 161 * Strict overcommit modes added 2002 Feb 26 by Alan Cox. 162 * Additional code 2002 Jul 20 by Robert Love. 163 * 164 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise. 165 * 166 * Note this is a helper function intended to be used by LSMs which 167 * wish to use this logic. 168 */ 169 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin) 170 { 171 long free, allowed, reserve; 172 173 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) < 174 -(s64)vm_committed_as_batch * num_online_cpus(), 175 "memory commitment underflow"); 176 177 vm_acct_memory(pages); 178 179 /* 180 * Sometimes we want to use more memory than we have 181 */ 182 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS) 183 return 0; 184 185 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) { 186 free = global_page_state(NR_FREE_PAGES); 187 free += global_page_state(NR_FILE_PAGES); 188 189 /* 190 * shmem pages shouldn't be counted as free in this 191 * case, they can't be purged, only swapped out, and 192 * that won't affect the overall amount of available 193 * memory in the system. 194 */ 195 free -= global_page_state(NR_SHMEM); 196 197 free += get_nr_swap_pages(); 198 199 /* 200 * Any slabs which are created with the 201 * SLAB_RECLAIM_ACCOUNT flag claim to have contents 202 * which are reclaimable, under pressure. The dentry 203 * cache and most inode caches should fall into this 204 */ 205 free += global_page_state(NR_SLAB_RECLAIMABLE); 206 207 /* 208 * Leave reserved pages. The pages are not for anonymous pages. 209 */ 210 if (free <= totalreserve_pages) 211 goto error; 212 else 213 free -= totalreserve_pages; 214 215 /* 216 * Reserve some for root 217 */ 218 if (!cap_sys_admin) 219 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10); 220 221 if (free > pages) 222 return 0; 223 224 goto error; 225 } 226 227 allowed = vm_commit_limit(); 228 /* 229 * Reserve some for root 230 */ 231 if (!cap_sys_admin) 232 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10); 233 234 /* 235 * Don't let a single process grow so big a user can't recover 236 */ 237 if (mm) { 238 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10); 239 allowed -= min_t(long, mm->total_vm / 32, reserve); 240 } 241 242 if (percpu_counter_read_positive(&vm_committed_as) < allowed) 243 return 0; 244 error: 245 vm_unacct_memory(pages); 246 247 return -ENOMEM; 248 } 249 250 /* 251 * Requires inode->i_mapping->i_mmap_rwsem 252 */ 253 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 254 struct file *file, struct address_space *mapping) 255 { 256 if (vma->vm_flags & VM_DENYWRITE) 257 atomic_inc(&file_inode(file)->i_writecount); 258 if (vma->vm_flags & VM_SHARED) 259 mapping_unmap_writable(mapping); 260 261 flush_dcache_mmap_lock(mapping); 262 vma_interval_tree_remove(vma, &mapping->i_mmap); 263 flush_dcache_mmap_unlock(mapping); 264 } 265 266 /* 267 * Unlink a file-based vm structure from its interval tree, to hide 268 * vma from rmap and vmtruncate before freeing its page tables. 269 */ 270 void unlink_file_vma(struct vm_area_struct *vma) 271 { 272 struct file *file = vma->vm_file; 273 274 if (file) { 275 struct address_space *mapping = file->f_mapping; 276 i_mmap_lock_write(mapping); 277 __remove_shared_vm_struct(vma, file, mapping); 278 i_mmap_unlock_write(mapping); 279 } 280 } 281 282 /* 283 * Close a vm structure and free it, returning the next. 284 */ 285 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 286 { 287 struct vm_area_struct *next = vma->vm_next; 288 289 might_sleep(); 290 if (vma->vm_ops && vma->vm_ops->close) 291 vma->vm_ops->close(vma); 292 if (vma->vm_file) 293 fput(vma->vm_file); 294 mpol_put(vma_policy(vma)); 295 kmem_cache_free(vm_area_cachep, vma); 296 return next; 297 } 298 299 static unsigned long do_brk(unsigned long addr, unsigned long len); 300 301 SYSCALL_DEFINE1(brk, unsigned long, brk) 302 { 303 unsigned long retval; 304 unsigned long newbrk, oldbrk; 305 struct mm_struct *mm = current->mm; 306 unsigned long min_brk; 307 bool populate; 308 309 down_write(&mm->mmap_sem); 310 311 #ifdef CONFIG_COMPAT_BRK 312 /* 313 * CONFIG_COMPAT_BRK can still be overridden by setting 314 * randomize_va_space to 2, which will still cause mm->start_brk 315 * to be arbitrarily shifted 316 */ 317 if (current->brk_randomized) 318 min_brk = mm->start_brk; 319 else 320 min_brk = mm->end_data; 321 #else 322 min_brk = mm->start_brk; 323 #endif 324 if (brk < min_brk) 325 goto out; 326 327 /* 328 * Check against rlimit here. If this check is done later after the test 329 * of oldbrk with newbrk then it can escape the test and let the data 330 * segment grow beyond its set limit the in case where the limit is 331 * not page aligned -Ram Gupta 332 */ 333 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 334 mm->end_data, mm->start_data)) 335 goto out; 336 337 newbrk = PAGE_ALIGN(brk); 338 oldbrk = PAGE_ALIGN(mm->brk); 339 if (oldbrk == newbrk) 340 goto set_brk; 341 342 /* Always allow shrinking brk. */ 343 if (brk <= mm->brk) { 344 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 345 goto set_brk; 346 goto out; 347 } 348 349 /* Check against existing mmap mappings. */ 350 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 351 goto out; 352 353 /* Ok, looks good - let it rip. */ 354 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 355 goto out; 356 357 set_brk: 358 mm->brk = brk; 359 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 360 up_write(&mm->mmap_sem); 361 if (populate) 362 mm_populate(oldbrk, newbrk - oldbrk); 363 return brk; 364 365 out: 366 retval = mm->brk; 367 up_write(&mm->mmap_sem); 368 return retval; 369 } 370 371 static long vma_compute_subtree_gap(struct vm_area_struct *vma) 372 { 373 unsigned long max, subtree_gap; 374 max = vma->vm_start; 375 if (vma->vm_prev) 376 max -= vma->vm_prev->vm_end; 377 if (vma->vm_rb.rb_left) { 378 subtree_gap = rb_entry(vma->vm_rb.rb_left, 379 struct vm_area_struct, vm_rb)->rb_subtree_gap; 380 if (subtree_gap > max) 381 max = subtree_gap; 382 } 383 if (vma->vm_rb.rb_right) { 384 subtree_gap = rb_entry(vma->vm_rb.rb_right, 385 struct vm_area_struct, vm_rb)->rb_subtree_gap; 386 if (subtree_gap > max) 387 max = subtree_gap; 388 } 389 return max; 390 } 391 392 #ifdef CONFIG_DEBUG_VM_RB 393 static int browse_rb(struct mm_struct *mm) 394 { 395 struct rb_root *root = &mm->mm_rb; 396 int i = 0, j, bug = 0; 397 struct rb_node *nd, *pn = NULL; 398 unsigned long prev = 0, pend = 0; 399 400 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 401 struct vm_area_struct *vma; 402 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 403 if (vma->vm_start < prev) { 404 pr_emerg("vm_start %lx < prev %lx\n", 405 vma->vm_start, prev); 406 bug = 1; 407 } 408 if (vma->vm_start < pend) { 409 pr_emerg("vm_start %lx < pend %lx\n", 410 vma->vm_start, pend); 411 bug = 1; 412 } 413 if (vma->vm_start > vma->vm_end) { 414 pr_emerg("vm_start %lx > vm_end %lx\n", 415 vma->vm_start, vma->vm_end); 416 bug = 1; 417 } 418 spin_lock(&mm->page_table_lock); 419 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) { 420 pr_emerg("free gap %lx, correct %lx\n", 421 vma->rb_subtree_gap, 422 vma_compute_subtree_gap(vma)); 423 bug = 1; 424 } 425 spin_unlock(&mm->page_table_lock); 426 i++; 427 pn = nd; 428 prev = vma->vm_start; 429 pend = vma->vm_end; 430 } 431 j = 0; 432 for (nd = pn; nd; nd = rb_prev(nd)) 433 j++; 434 if (i != j) { 435 pr_emerg("backwards %d, forwards %d\n", j, i); 436 bug = 1; 437 } 438 return bug ? -1 : i; 439 } 440 441 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore) 442 { 443 struct rb_node *nd; 444 445 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 446 struct vm_area_struct *vma; 447 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 448 VM_BUG_ON_VMA(vma != ignore && 449 vma->rb_subtree_gap != vma_compute_subtree_gap(vma), 450 vma); 451 } 452 } 453 454 static void validate_mm(struct mm_struct *mm) 455 { 456 int bug = 0; 457 int i = 0; 458 unsigned long highest_address = 0; 459 struct vm_area_struct *vma = mm->mmap; 460 461 while (vma) { 462 struct anon_vma *anon_vma = vma->anon_vma; 463 struct anon_vma_chain *avc; 464 465 if (anon_vma) { 466 anon_vma_lock_read(anon_vma); 467 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 468 anon_vma_interval_tree_verify(avc); 469 anon_vma_unlock_read(anon_vma); 470 } 471 472 highest_address = vma->vm_end; 473 vma = vma->vm_next; 474 i++; 475 } 476 if (i != mm->map_count) { 477 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i); 478 bug = 1; 479 } 480 if (highest_address != mm->highest_vm_end) { 481 pr_emerg("mm->highest_vm_end %lx, found %lx\n", 482 mm->highest_vm_end, highest_address); 483 bug = 1; 484 } 485 i = browse_rb(mm); 486 if (i != mm->map_count) { 487 if (i != -1) 488 pr_emerg("map_count %d rb %d\n", mm->map_count, i); 489 bug = 1; 490 } 491 VM_BUG_ON_MM(bug, mm); 492 } 493 #else 494 #define validate_mm_rb(root, ignore) do { } while (0) 495 #define validate_mm(mm) do { } while (0) 496 #endif 497 498 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb, 499 unsigned long, rb_subtree_gap, vma_compute_subtree_gap) 500 501 /* 502 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or 503 * vma->vm_prev->vm_end values changed, without modifying the vma's position 504 * in the rbtree. 505 */ 506 static void vma_gap_update(struct vm_area_struct *vma) 507 { 508 /* 509 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback 510 * function that does exacltly what we want. 511 */ 512 vma_gap_callbacks_propagate(&vma->vm_rb, NULL); 513 } 514 515 static inline void vma_rb_insert(struct vm_area_struct *vma, 516 struct rb_root *root) 517 { 518 /* All rb_subtree_gap values must be consistent prior to insertion */ 519 validate_mm_rb(root, NULL); 520 521 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 522 } 523 524 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root) 525 { 526 /* 527 * All rb_subtree_gap values must be consistent prior to erase, 528 * with the possible exception of the vma being erased. 529 */ 530 validate_mm_rb(root, vma); 531 532 /* 533 * Note rb_erase_augmented is a fairly large inline function, 534 * so make sure we instantiate it only once with our desired 535 * augmented rbtree callbacks. 536 */ 537 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 538 } 539 540 /* 541 * vma has some anon_vma assigned, and is already inserted on that 542 * anon_vma's interval trees. 543 * 544 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 545 * vma must be removed from the anon_vma's interval trees using 546 * anon_vma_interval_tree_pre_update_vma(). 547 * 548 * After the update, the vma will be reinserted using 549 * anon_vma_interval_tree_post_update_vma(). 550 * 551 * The entire update must be protected by exclusive mmap_sem and by 552 * the root anon_vma's mutex. 553 */ 554 static inline void 555 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 556 { 557 struct anon_vma_chain *avc; 558 559 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 560 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 561 } 562 563 static inline void 564 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 565 { 566 struct anon_vma_chain *avc; 567 568 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 569 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 570 } 571 572 static int find_vma_links(struct mm_struct *mm, unsigned long addr, 573 unsigned long end, struct vm_area_struct **pprev, 574 struct rb_node ***rb_link, struct rb_node **rb_parent) 575 { 576 struct rb_node **__rb_link, *__rb_parent, *rb_prev; 577 578 __rb_link = &mm->mm_rb.rb_node; 579 rb_prev = __rb_parent = NULL; 580 581 while (*__rb_link) { 582 struct vm_area_struct *vma_tmp; 583 584 __rb_parent = *__rb_link; 585 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 586 587 if (vma_tmp->vm_end > addr) { 588 /* Fail if an existing vma overlaps the area */ 589 if (vma_tmp->vm_start < end) 590 return -ENOMEM; 591 __rb_link = &__rb_parent->rb_left; 592 } else { 593 rb_prev = __rb_parent; 594 __rb_link = &__rb_parent->rb_right; 595 } 596 } 597 598 *pprev = NULL; 599 if (rb_prev) 600 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 601 *rb_link = __rb_link; 602 *rb_parent = __rb_parent; 603 return 0; 604 } 605 606 static unsigned long count_vma_pages_range(struct mm_struct *mm, 607 unsigned long addr, unsigned long end) 608 { 609 unsigned long nr_pages = 0; 610 struct vm_area_struct *vma; 611 612 /* Find first overlaping mapping */ 613 vma = find_vma_intersection(mm, addr, end); 614 if (!vma) 615 return 0; 616 617 nr_pages = (min(end, vma->vm_end) - 618 max(addr, vma->vm_start)) >> PAGE_SHIFT; 619 620 /* Iterate over the rest of the overlaps */ 621 for (vma = vma->vm_next; vma; vma = vma->vm_next) { 622 unsigned long overlap_len; 623 624 if (vma->vm_start > end) 625 break; 626 627 overlap_len = min(end, vma->vm_end) - vma->vm_start; 628 nr_pages += overlap_len >> PAGE_SHIFT; 629 } 630 631 return nr_pages; 632 } 633 634 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 635 struct rb_node **rb_link, struct rb_node *rb_parent) 636 { 637 /* Update tracking information for the gap following the new vma. */ 638 if (vma->vm_next) 639 vma_gap_update(vma->vm_next); 640 else 641 mm->highest_vm_end = vma->vm_end; 642 643 /* 644 * vma->vm_prev wasn't known when we followed the rbtree to find the 645 * correct insertion point for that vma. As a result, we could not 646 * update the vma vm_rb parents rb_subtree_gap values on the way down. 647 * So, we first insert the vma with a zero rb_subtree_gap value 648 * (to be consistent with what we did on the way down), and then 649 * immediately update the gap to the correct value. Finally we 650 * rebalance the rbtree after all augmented values have been set. 651 */ 652 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 653 vma->rb_subtree_gap = 0; 654 vma_gap_update(vma); 655 vma_rb_insert(vma, &mm->mm_rb); 656 } 657 658 static void __vma_link_file(struct vm_area_struct *vma) 659 { 660 struct file *file; 661 662 file = vma->vm_file; 663 if (file) { 664 struct address_space *mapping = file->f_mapping; 665 666 if (vma->vm_flags & VM_DENYWRITE) 667 atomic_dec(&file_inode(file)->i_writecount); 668 if (vma->vm_flags & VM_SHARED) 669 atomic_inc(&mapping->i_mmap_writable); 670 671 flush_dcache_mmap_lock(mapping); 672 vma_interval_tree_insert(vma, &mapping->i_mmap); 673 flush_dcache_mmap_unlock(mapping); 674 } 675 } 676 677 static void 678 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 679 struct vm_area_struct *prev, struct rb_node **rb_link, 680 struct rb_node *rb_parent) 681 { 682 __vma_link_list(mm, vma, prev, rb_parent); 683 __vma_link_rb(mm, vma, rb_link, rb_parent); 684 } 685 686 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 687 struct vm_area_struct *prev, struct rb_node **rb_link, 688 struct rb_node *rb_parent) 689 { 690 struct address_space *mapping = NULL; 691 692 if (vma->vm_file) { 693 mapping = vma->vm_file->f_mapping; 694 i_mmap_lock_write(mapping); 695 } 696 697 __vma_link(mm, vma, prev, rb_link, rb_parent); 698 __vma_link_file(vma); 699 700 if (mapping) 701 i_mmap_unlock_write(mapping); 702 703 mm->map_count++; 704 validate_mm(mm); 705 } 706 707 /* 708 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the 709 * mm's list and rbtree. It has already been inserted into the interval tree. 710 */ 711 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 712 { 713 struct vm_area_struct *prev; 714 struct rb_node **rb_link, *rb_parent; 715 716 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 717 &prev, &rb_link, &rb_parent)) 718 BUG(); 719 __vma_link(mm, vma, prev, rb_link, rb_parent); 720 mm->map_count++; 721 } 722 723 static inline void 724 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 725 struct vm_area_struct *prev) 726 { 727 struct vm_area_struct *next; 728 729 vma_rb_erase(vma, &mm->mm_rb); 730 prev->vm_next = next = vma->vm_next; 731 if (next) 732 next->vm_prev = prev; 733 734 /* Kill the cache */ 735 vmacache_invalidate(mm); 736 } 737 738 /* 739 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 740 * is already present in an i_mmap tree without adjusting the tree. 741 * The following helper function should be used when such adjustments 742 * are necessary. The "insert" vma (if any) is to be inserted 743 * before we drop the necessary locks. 744 */ 745 int vma_adjust(struct vm_area_struct *vma, unsigned long start, 746 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 747 { 748 struct mm_struct *mm = vma->vm_mm; 749 struct vm_area_struct *next = vma->vm_next; 750 struct vm_area_struct *importer = NULL; 751 struct address_space *mapping = NULL; 752 struct rb_root *root = NULL; 753 struct anon_vma *anon_vma = NULL; 754 struct file *file = vma->vm_file; 755 bool start_changed = false, end_changed = false; 756 long adjust_next = 0; 757 int remove_next = 0; 758 759 if (next && !insert) { 760 struct vm_area_struct *exporter = NULL; 761 762 if (end >= next->vm_end) { 763 /* 764 * vma expands, overlapping all the next, and 765 * perhaps the one after too (mprotect case 6). 766 */ 767 again: remove_next = 1 + (end > next->vm_end); 768 end = next->vm_end; 769 exporter = next; 770 importer = vma; 771 } else if (end > next->vm_start) { 772 /* 773 * vma expands, overlapping part of the next: 774 * mprotect case 5 shifting the boundary up. 775 */ 776 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 777 exporter = next; 778 importer = vma; 779 } else if (end < vma->vm_end) { 780 /* 781 * vma shrinks, and !insert tells it's not 782 * split_vma inserting another: so it must be 783 * mprotect case 4 shifting the boundary down. 784 */ 785 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT); 786 exporter = vma; 787 importer = next; 788 } 789 790 /* 791 * Easily overlooked: when mprotect shifts the boundary, 792 * make sure the expanding vma has anon_vma set if the 793 * shrinking vma had, to cover any anon pages imported. 794 */ 795 if (exporter && exporter->anon_vma && !importer->anon_vma) { 796 int error; 797 798 importer->anon_vma = exporter->anon_vma; 799 error = anon_vma_clone(importer, exporter); 800 if (error) 801 return error; 802 } 803 } 804 805 if (file) { 806 mapping = file->f_mapping; 807 root = &mapping->i_mmap; 808 uprobe_munmap(vma, vma->vm_start, vma->vm_end); 809 810 if (adjust_next) 811 uprobe_munmap(next, next->vm_start, next->vm_end); 812 813 i_mmap_lock_write(mapping); 814 if (insert) { 815 /* 816 * Put into interval tree now, so instantiated pages 817 * are visible to arm/parisc __flush_dcache_page 818 * throughout; but we cannot insert into address 819 * space until vma start or end is updated. 820 */ 821 __vma_link_file(insert); 822 } 823 } 824 825 vma_adjust_trans_huge(vma, start, end, adjust_next); 826 827 anon_vma = vma->anon_vma; 828 if (!anon_vma && adjust_next) 829 anon_vma = next->anon_vma; 830 if (anon_vma) { 831 VM_BUG_ON_VMA(adjust_next && next->anon_vma && 832 anon_vma != next->anon_vma, next); 833 anon_vma_lock_write(anon_vma); 834 anon_vma_interval_tree_pre_update_vma(vma); 835 if (adjust_next) 836 anon_vma_interval_tree_pre_update_vma(next); 837 } 838 839 if (root) { 840 flush_dcache_mmap_lock(mapping); 841 vma_interval_tree_remove(vma, root); 842 if (adjust_next) 843 vma_interval_tree_remove(next, root); 844 } 845 846 if (start != vma->vm_start) { 847 vma->vm_start = start; 848 start_changed = true; 849 } 850 if (end != vma->vm_end) { 851 vma->vm_end = end; 852 end_changed = true; 853 } 854 vma->vm_pgoff = pgoff; 855 if (adjust_next) { 856 next->vm_start += adjust_next << PAGE_SHIFT; 857 next->vm_pgoff += adjust_next; 858 } 859 860 if (root) { 861 if (adjust_next) 862 vma_interval_tree_insert(next, root); 863 vma_interval_tree_insert(vma, root); 864 flush_dcache_mmap_unlock(mapping); 865 } 866 867 if (remove_next) { 868 /* 869 * vma_merge has merged next into vma, and needs 870 * us to remove next before dropping the locks. 871 */ 872 __vma_unlink(mm, next, vma); 873 if (file) 874 __remove_shared_vm_struct(next, file, mapping); 875 } else if (insert) { 876 /* 877 * split_vma has split insert from vma, and needs 878 * us to insert it before dropping the locks 879 * (it may either follow vma or precede it). 880 */ 881 __insert_vm_struct(mm, insert); 882 } else { 883 if (start_changed) 884 vma_gap_update(vma); 885 if (end_changed) { 886 if (!next) 887 mm->highest_vm_end = end; 888 else if (!adjust_next) 889 vma_gap_update(next); 890 } 891 } 892 893 if (anon_vma) { 894 anon_vma_interval_tree_post_update_vma(vma); 895 if (adjust_next) 896 anon_vma_interval_tree_post_update_vma(next); 897 anon_vma_unlock_write(anon_vma); 898 } 899 if (mapping) 900 i_mmap_unlock_write(mapping); 901 902 if (root) { 903 uprobe_mmap(vma); 904 905 if (adjust_next) 906 uprobe_mmap(next); 907 } 908 909 if (remove_next) { 910 if (file) { 911 uprobe_munmap(next, next->vm_start, next->vm_end); 912 fput(file); 913 } 914 if (next->anon_vma) 915 anon_vma_merge(vma, next); 916 mm->map_count--; 917 mpol_put(vma_policy(next)); 918 kmem_cache_free(vm_area_cachep, next); 919 /* 920 * In mprotect's case 6 (see comments on vma_merge), 921 * we must remove another next too. It would clutter 922 * up the code too much to do both in one go. 923 */ 924 next = vma->vm_next; 925 if (remove_next == 2) 926 goto again; 927 else if (next) 928 vma_gap_update(next); 929 else 930 mm->highest_vm_end = end; 931 } 932 if (insert && file) 933 uprobe_mmap(insert); 934 935 validate_mm(mm); 936 937 return 0; 938 } 939 940 /* 941 * If the vma has a ->close operation then the driver probably needs to release 942 * per-vma resources, so we don't attempt to merge those. 943 */ 944 static inline int is_mergeable_vma(struct vm_area_struct *vma, 945 struct file *file, unsigned long vm_flags, 946 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 947 { 948 /* 949 * VM_SOFTDIRTY should not prevent from VMA merging, if we 950 * match the flags but dirty bit -- the caller should mark 951 * merged VMA as dirty. If dirty bit won't be excluded from 952 * comparison, we increase pressue on the memory system forcing 953 * the kernel to generate new VMAs when old one could be 954 * extended instead. 955 */ 956 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 957 return 0; 958 if (vma->vm_file != file) 959 return 0; 960 if (vma->vm_ops && vma->vm_ops->close) 961 return 0; 962 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx)) 963 return 0; 964 return 1; 965 } 966 967 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 968 struct anon_vma *anon_vma2, 969 struct vm_area_struct *vma) 970 { 971 /* 972 * The list_is_singular() test is to avoid merging VMA cloned from 973 * parents. This can improve scalability caused by anon_vma lock. 974 */ 975 if ((!anon_vma1 || !anon_vma2) && (!vma || 976 list_is_singular(&vma->anon_vma_chain))) 977 return 1; 978 return anon_vma1 == anon_vma2; 979 } 980 981 /* 982 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 983 * in front of (at a lower virtual address and file offset than) the vma. 984 * 985 * We cannot merge two vmas if they have differently assigned (non-NULL) 986 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 987 * 988 * We don't check here for the merged mmap wrapping around the end of pagecache 989 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 990 * wrap, nor mmaps which cover the final page at index -1UL. 991 */ 992 static int 993 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 994 struct anon_vma *anon_vma, struct file *file, 995 pgoff_t vm_pgoff, 996 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 997 { 998 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) && 999 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1000 if (vma->vm_pgoff == vm_pgoff) 1001 return 1; 1002 } 1003 return 0; 1004 } 1005 1006 /* 1007 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 1008 * beyond (at a higher virtual address and file offset than) the vma. 1009 * 1010 * We cannot merge two vmas if they have differently assigned (non-NULL) 1011 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 1012 */ 1013 static int 1014 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 1015 struct anon_vma *anon_vma, struct file *file, 1016 pgoff_t vm_pgoff, 1017 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1018 { 1019 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) && 1020 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1021 pgoff_t vm_pglen; 1022 vm_pglen = vma_pages(vma); 1023 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 1024 return 1; 1025 } 1026 return 0; 1027 } 1028 1029 /* 1030 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 1031 * whether that can be merged with its predecessor or its successor. 1032 * Or both (it neatly fills a hole). 1033 * 1034 * In most cases - when called for mmap, brk or mremap - [addr,end) is 1035 * certain not to be mapped by the time vma_merge is called; but when 1036 * called for mprotect, it is certain to be already mapped (either at 1037 * an offset within prev, or at the start of next), and the flags of 1038 * this area are about to be changed to vm_flags - and the no-change 1039 * case has already been eliminated. 1040 * 1041 * The following mprotect cases have to be considered, where AAAA is 1042 * the area passed down from mprotect_fixup, never extending beyond one 1043 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 1044 * 1045 * AAAA AAAA AAAA AAAA 1046 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 1047 * cannot merge might become might become might become 1048 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 1049 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 1050 * mremap move: PPPPNNNNNNNN 8 1051 * AAAA 1052 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 1053 * might become case 1 below case 2 below case 3 below 1054 * 1055 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 1056 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 1057 */ 1058 struct vm_area_struct *vma_merge(struct mm_struct *mm, 1059 struct vm_area_struct *prev, unsigned long addr, 1060 unsigned long end, unsigned long vm_flags, 1061 struct anon_vma *anon_vma, struct file *file, 1062 pgoff_t pgoff, struct mempolicy *policy, 1063 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1064 { 1065 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 1066 struct vm_area_struct *area, *next; 1067 int err; 1068 1069 /* 1070 * We later require that vma->vm_flags == vm_flags, 1071 * so this tests vma->vm_flags & VM_SPECIAL, too. 1072 */ 1073 if (vm_flags & VM_SPECIAL) 1074 return NULL; 1075 1076 if (prev) 1077 next = prev->vm_next; 1078 else 1079 next = mm->mmap; 1080 area = next; 1081 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 1082 next = next->vm_next; 1083 1084 /* 1085 * Can it merge with the predecessor? 1086 */ 1087 if (prev && prev->vm_end == addr && 1088 mpol_equal(vma_policy(prev), policy) && 1089 can_vma_merge_after(prev, vm_flags, 1090 anon_vma, file, pgoff, 1091 vm_userfaultfd_ctx)) { 1092 /* 1093 * OK, it can. Can we now merge in the successor as well? 1094 */ 1095 if (next && end == next->vm_start && 1096 mpol_equal(policy, vma_policy(next)) && 1097 can_vma_merge_before(next, vm_flags, 1098 anon_vma, file, 1099 pgoff+pglen, 1100 vm_userfaultfd_ctx) && 1101 is_mergeable_anon_vma(prev->anon_vma, 1102 next->anon_vma, NULL)) { 1103 /* cases 1, 6 */ 1104 err = vma_adjust(prev, prev->vm_start, 1105 next->vm_end, prev->vm_pgoff, NULL); 1106 } else /* cases 2, 5, 7 */ 1107 err = vma_adjust(prev, prev->vm_start, 1108 end, prev->vm_pgoff, NULL); 1109 if (err) 1110 return NULL; 1111 khugepaged_enter_vma_merge(prev, vm_flags); 1112 return prev; 1113 } 1114 1115 /* 1116 * Can this new request be merged in front of next? 1117 */ 1118 if (next && end == next->vm_start && 1119 mpol_equal(policy, vma_policy(next)) && 1120 can_vma_merge_before(next, vm_flags, 1121 anon_vma, file, pgoff+pglen, 1122 vm_userfaultfd_ctx)) { 1123 if (prev && addr < prev->vm_end) /* case 4 */ 1124 err = vma_adjust(prev, prev->vm_start, 1125 addr, prev->vm_pgoff, NULL); 1126 else /* cases 3, 8 */ 1127 err = vma_adjust(area, addr, next->vm_end, 1128 next->vm_pgoff - pglen, NULL); 1129 if (err) 1130 return NULL; 1131 khugepaged_enter_vma_merge(area, vm_flags); 1132 return area; 1133 } 1134 1135 return NULL; 1136 } 1137 1138 /* 1139 * Rough compatbility check to quickly see if it's even worth looking 1140 * at sharing an anon_vma. 1141 * 1142 * They need to have the same vm_file, and the flags can only differ 1143 * in things that mprotect may change. 1144 * 1145 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1146 * we can merge the two vma's. For example, we refuse to merge a vma if 1147 * there is a vm_ops->close() function, because that indicates that the 1148 * driver is doing some kind of reference counting. But that doesn't 1149 * really matter for the anon_vma sharing case. 1150 */ 1151 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1152 { 1153 return a->vm_end == b->vm_start && 1154 mpol_equal(vma_policy(a), vma_policy(b)) && 1155 a->vm_file == b->vm_file && 1156 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) && 1157 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1158 } 1159 1160 /* 1161 * Do some basic sanity checking to see if we can re-use the anon_vma 1162 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1163 * the same as 'old', the other will be the new one that is trying 1164 * to share the anon_vma. 1165 * 1166 * NOTE! This runs with mm_sem held for reading, so it is possible that 1167 * the anon_vma of 'old' is concurrently in the process of being set up 1168 * by another page fault trying to merge _that_. But that's ok: if it 1169 * is being set up, that automatically means that it will be a singleton 1170 * acceptable for merging, so we can do all of this optimistically. But 1171 * we do that READ_ONCE() to make sure that we never re-load the pointer. 1172 * 1173 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1174 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1175 * is to return an anon_vma that is "complex" due to having gone through 1176 * a fork). 1177 * 1178 * We also make sure that the two vma's are compatible (adjacent, 1179 * and with the same memory policies). That's all stable, even with just 1180 * a read lock on the mm_sem. 1181 */ 1182 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1183 { 1184 if (anon_vma_compatible(a, b)) { 1185 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma); 1186 1187 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1188 return anon_vma; 1189 } 1190 return NULL; 1191 } 1192 1193 /* 1194 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1195 * neighbouring vmas for a suitable anon_vma, before it goes off 1196 * to allocate a new anon_vma. It checks because a repetitive 1197 * sequence of mprotects and faults may otherwise lead to distinct 1198 * anon_vmas being allocated, preventing vma merge in subsequent 1199 * mprotect. 1200 */ 1201 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1202 { 1203 struct anon_vma *anon_vma; 1204 struct vm_area_struct *near; 1205 1206 near = vma->vm_next; 1207 if (!near) 1208 goto try_prev; 1209 1210 anon_vma = reusable_anon_vma(near, vma, near); 1211 if (anon_vma) 1212 return anon_vma; 1213 try_prev: 1214 near = vma->vm_prev; 1215 if (!near) 1216 goto none; 1217 1218 anon_vma = reusable_anon_vma(near, near, vma); 1219 if (anon_vma) 1220 return anon_vma; 1221 none: 1222 /* 1223 * There's no absolute need to look only at touching neighbours: 1224 * we could search further afield for "compatible" anon_vmas. 1225 * But it would probably just be a waste of time searching, 1226 * or lead to too many vmas hanging off the same anon_vma. 1227 * We're trying to allow mprotect remerging later on, 1228 * not trying to minimize memory used for anon_vmas. 1229 */ 1230 return NULL; 1231 } 1232 1233 /* 1234 * If a hint addr is less than mmap_min_addr change hint to be as 1235 * low as possible but still greater than mmap_min_addr 1236 */ 1237 static inline unsigned long round_hint_to_min(unsigned long hint) 1238 { 1239 hint &= PAGE_MASK; 1240 if (((void *)hint != NULL) && 1241 (hint < mmap_min_addr)) 1242 return PAGE_ALIGN(mmap_min_addr); 1243 return hint; 1244 } 1245 1246 static inline int mlock_future_check(struct mm_struct *mm, 1247 unsigned long flags, 1248 unsigned long len) 1249 { 1250 unsigned long locked, lock_limit; 1251 1252 /* mlock MCL_FUTURE? */ 1253 if (flags & VM_LOCKED) { 1254 locked = len >> PAGE_SHIFT; 1255 locked += mm->locked_vm; 1256 lock_limit = rlimit(RLIMIT_MEMLOCK); 1257 lock_limit >>= PAGE_SHIFT; 1258 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1259 return -EAGAIN; 1260 } 1261 return 0; 1262 } 1263 1264 /* 1265 * The caller must hold down_write(¤t->mm->mmap_sem). 1266 */ 1267 unsigned long do_mmap(struct file *file, unsigned long addr, 1268 unsigned long len, unsigned long prot, 1269 unsigned long flags, vm_flags_t vm_flags, 1270 unsigned long pgoff, unsigned long *populate) 1271 { 1272 struct mm_struct *mm = current->mm; 1273 1274 *populate = 0; 1275 1276 if (!len) 1277 return -EINVAL; 1278 1279 /* 1280 * Does the application expect PROT_READ to imply PROT_EXEC? 1281 * 1282 * (the exception is when the underlying filesystem is noexec 1283 * mounted, in which case we dont add PROT_EXEC.) 1284 */ 1285 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1286 if (!(file && path_noexec(&file->f_path))) 1287 prot |= PROT_EXEC; 1288 1289 if (!(flags & MAP_FIXED)) 1290 addr = round_hint_to_min(addr); 1291 1292 /* Careful about overflows.. */ 1293 len = PAGE_ALIGN(len); 1294 if (!len) 1295 return -ENOMEM; 1296 1297 /* offset overflow? */ 1298 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1299 return -EOVERFLOW; 1300 1301 /* Too many mappings? */ 1302 if (mm->map_count > sysctl_max_map_count) 1303 return -ENOMEM; 1304 1305 /* Obtain the address to map to. we verify (or select) it and ensure 1306 * that it represents a valid section of the address space. 1307 */ 1308 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1309 if (offset_in_page(addr)) 1310 return addr; 1311 1312 /* Do simple checking here so the lower-level routines won't have 1313 * to. we assume access permissions have been handled by the open 1314 * of the memory object, so we don't do any here. 1315 */ 1316 vm_flags |= calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 1317 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1318 1319 if (flags & MAP_LOCKED) 1320 if (!can_do_mlock()) 1321 return -EPERM; 1322 1323 if (mlock_future_check(mm, vm_flags, len)) 1324 return -EAGAIN; 1325 1326 if (file) { 1327 struct inode *inode = file_inode(file); 1328 1329 switch (flags & MAP_TYPE) { 1330 case MAP_SHARED: 1331 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 1332 return -EACCES; 1333 1334 /* 1335 * Make sure we don't allow writing to an append-only 1336 * file.. 1337 */ 1338 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1339 return -EACCES; 1340 1341 /* 1342 * Make sure there are no mandatory locks on the file. 1343 */ 1344 if (locks_verify_locked(file)) 1345 return -EAGAIN; 1346 1347 vm_flags |= VM_SHARED | VM_MAYSHARE; 1348 if (!(file->f_mode & FMODE_WRITE)) 1349 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1350 1351 /* fall through */ 1352 case MAP_PRIVATE: 1353 if (!(file->f_mode & FMODE_READ)) 1354 return -EACCES; 1355 if (path_noexec(&file->f_path)) { 1356 if (vm_flags & VM_EXEC) 1357 return -EPERM; 1358 vm_flags &= ~VM_MAYEXEC; 1359 } 1360 1361 if (!file->f_op->mmap) 1362 return -ENODEV; 1363 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1364 return -EINVAL; 1365 break; 1366 1367 default: 1368 return -EINVAL; 1369 } 1370 } else { 1371 switch (flags & MAP_TYPE) { 1372 case MAP_SHARED: 1373 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1374 return -EINVAL; 1375 /* 1376 * Ignore pgoff. 1377 */ 1378 pgoff = 0; 1379 vm_flags |= VM_SHARED | VM_MAYSHARE; 1380 break; 1381 case MAP_PRIVATE: 1382 /* 1383 * Set pgoff according to addr for anon_vma. 1384 */ 1385 pgoff = addr >> PAGE_SHIFT; 1386 break; 1387 default: 1388 return -EINVAL; 1389 } 1390 } 1391 1392 /* 1393 * Set 'VM_NORESERVE' if we should not account for the 1394 * memory use of this mapping. 1395 */ 1396 if (flags & MAP_NORESERVE) { 1397 /* We honor MAP_NORESERVE if allowed to overcommit */ 1398 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1399 vm_flags |= VM_NORESERVE; 1400 1401 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1402 if (file && is_file_hugepages(file)) 1403 vm_flags |= VM_NORESERVE; 1404 } 1405 1406 addr = mmap_region(file, addr, len, vm_flags, pgoff); 1407 if (!IS_ERR_VALUE(addr) && 1408 ((vm_flags & VM_LOCKED) || 1409 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1410 *populate = len; 1411 return addr; 1412 } 1413 1414 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1415 unsigned long, prot, unsigned long, flags, 1416 unsigned long, fd, unsigned long, pgoff) 1417 { 1418 struct file *file = NULL; 1419 unsigned long retval; 1420 1421 if (!(flags & MAP_ANONYMOUS)) { 1422 audit_mmap_fd(fd, flags); 1423 file = fget(fd); 1424 if (!file) 1425 return -EBADF; 1426 if (is_file_hugepages(file)) 1427 len = ALIGN(len, huge_page_size(hstate_file(file))); 1428 retval = -EINVAL; 1429 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file))) 1430 goto out_fput; 1431 } else if (flags & MAP_HUGETLB) { 1432 struct user_struct *user = NULL; 1433 struct hstate *hs; 1434 1435 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK); 1436 if (!hs) 1437 return -EINVAL; 1438 1439 len = ALIGN(len, huge_page_size(hs)); 1440 /* 1441 * VM_NORESERVE is used because the reservations will be 1442 * taken when vm_ops->mmap() is called 1443 * A dummy user value is used because we are not locking 1444 * memory so no accounting is necessary 1445 */ 1446 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1447 VM_NORESERVE, 1448 &user, HUGETLB_ANONHUGE_INODE, 1449 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1450 if (IS_ERR(file)) 1451 return PTR_ERR(file); 1452 } 1453 1454 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1455 1456 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1457 out_fput: 1458 if (file) 1459 fput(file); 1460 return retval; 1461 } 1462 1463 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1464 struct mmap_arg_struct { 1465 unsigned long addr; 1466 unsigned long len; 1467 unsigned long prot; 1468 unsigned long flags; 1469 unsigned long fd; 1470 unsigned long offset; 1471 }; 1472 1473 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1474 { 1475 struct mmap_arg_struct a; 1476 1477 if (copy_from_user(&a, arg, sizeof(a))) 1478 return -EFAULT; 1479 if (offset_in_page(a.offset)) 1480 return -EINVAL; 1481 1482 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1483 a.offset >> PAGE_SHIFT); 1484 } 1485 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1486 1487 /* 1488 * Some shared mappigns will want the pages marked read-only 1489 * to track write events. If so, we'll downgrade vm_page_prot 1490 * to the private version (using protection_map[] without the 1491 * VM_SHARED bit). 1492 */ 1493 int vma_wants_writenotify(struct vm_area_struct *vma) 1494 { 1495 vm_flags_t vm_flags = vma->vm_flags; 1496 const struct vm_operations_struct *vm_ops = vma->vm_ops; 1497 1498 /* If it was private or non-writable, the write bit is already clear */ 1499 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1500 return 0; 1501 1502 /* The backer wishes to know when pages are first written to? */ 1503 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite)) 1504 return 1; 1505 1506 /* The open routine did something to the protections that pgprot_modify 1507 * won't preserve? */ 1508 if (pgprot_val(vma->vm_page_prot) != 1509 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags))) 1510 return 0; 1511 1512 /* Do we need to track softdirty? */ 1513 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY)) 1514 return 1; 1515 1516 /* Specialty mapping? */ 1517 if (vm_flags & VM_PFNMAP) 1518 return 0; 1519 1520 /* Can the mapping track the dirty pages? */ 1521 return vma->vm_file && vma->vm_file->f_mapping && 1522 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1523 } 1524 1525 /* 1526 * We account for memory if it's a private writeable mapping, 1527 * not hugepages and VM_NORESERVE wasn't set. 1528 */ 1529 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1530 { 1531 /* 1532 * hugetlb has its own accounting separate from the core VM 1533 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1534 */ 1535 if (file && is_file_hugepages(file)) 1536 return 0; 1537 1538 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1539 } 1540 1541 unsigned long mmap_region(struct file *file, unsigned long addr, 1542 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff) 1543 { 1544 struct mm_struct *mm = current->mm; 1545 struct vm_area_struct *vma, *prev; 1546 int error; 1547 struct rb_node **rb_link, *rb_parent; 1548 unsigned long charged = 0; 1549 1550 /* Check against address space limit. */ 1551 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) { 1552 unsigned long nr_pages; 1553 1554 /* 1555 * MAP_FIXED may remove pages of mappings that intersects with 1556 * requested mapping. Account for the pages it would unmap. 1557 */ 1558 nr_pages = count_vma_pages_range(mm, addr, addr + len); 1559 1560 if (!may_expand_vm(mm, vm_flags, 1561 (len >> PAGE_SHIFT) - nr_pages)) 1562 return -ENOMEM; 1563 } 1564 1565 /* Clear old maps */ 1566 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link, 1567 &rb_parent)) { 1568 if (do_munmap(mm, addr, len)) 1569 return -ENOMEM; 1570 } 1571 1572 /* 1573 * Private writable mapping: check memory availability 1574 */ 1575 if (accountable_mapping(file, vm_flags)) { 1576 charged = len >> PAGE_SHIFT; 1577 if (security_vm_enough_memory_mm(mm, charged)) 1578 return -ENOMEM; 1579 vm_flags |= VM_ACCOUNT; 1580 } 1581 1582 /* 1583 * Can we just expand an old mapping? 1584 */ 1585 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, 1586 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX); 1587 if (vma) 1588 goto out; 1589 1590 /* 1591 * Determine the object being mapped and call the appropriate 1592 * specific mapper. the address has already been validated, but 1593 * not unmapped, but the maps are removed from the list. 1594 */ 1595 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1596 if (!vma) { 1597 error = -ENOMEM; 1598 goto unacct_error; 1599 } 1600 1601 vma->vm_mm = mm; 1602 vma->vm_start = addr; 1603 vma->vm_end = addr + len; 1604 vma->vm_flags = vm_flags; 1605 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1606 vma->vm_pgoff = pgoff; 1607 INIT_LIST_HEAD(&vma->anon_vma_chain); 1608 1609 if (file) { 1610 if (vm_flags & VM_DENYWRITE) { 1611 error = deny_write_access(file); 1612 if (error) 1613 goto free_vma; 1614 } 1615 if (vm_flags & VM_SHARED) { 1616 error = mapping_map_writable(file->f_mapping); 1617 if (error) 1618 goto allow_write_and_free_vma; 1619 } 1620 1621 /* ->mmap() can change vma->vm_file, but must guarantee that 1622 * vma_link() below can deny write-access if VM_DENYWRITE is set 1623 * and map writably if VM_SHARED is set. This usually means the 1624 * new file must not have been exposed to user-space, yet. 1625 */ 1626 vma->vm_file = get_file(file); 1627 error = file->f_op->mmap(file, vma); 1628 if (error) 1629 goto unmap_and_free_vma; 1630 1631 /* Can addr have changed?? 1632 * 1633 * Answer: Yes, several device drivers can do it in their 1634 * f_op->mmap method. -DaveM 1635 * Bug: If addr is changed, prev, rb_link, rb_parent should 1636 * be updated for vma_link() 1637 */ 1638 WARN_ON_ONCE(addr != vma->vm_start); 1639 1640 addr = vma->vm_start; 1641 vm_flags = vma->vm_flags; 1642 } else if (vm_flags & VM_SHARED) { 1643 error = shmem_zero_setup(vma); 1644 if (error) 1645 goto free_vma; 1646 } 1647 1648 vma_link(mm, vma, prev, rb_link, rb_parent); 1649 /* Once vma denies write, undo our temporary denial count */ 1650 if (file) { 1651 if (vm_flags & VM_SHARED) 1652 mapping_unmap_writable(file->f_mapping); 1653 if (vm_flags & VM_DENYWRITE) 1654 allow_write_access(file); 1655 } 1656 file = vma->vm_file; 1657 out: 1658 perf_event_mmap(vma); 1659 1660 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT); 1661 if (vm_flags & VM_LOCKED) { 1662 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) || 1663 vma == get_gate_vma(current->mm))) 1664 mm->locked_vm += (len >> PAGE_SHIFT); 1665 else 1666 vma->vm_flags &= VM_LOCKED_CLEAR_MASK; 1667 } 1668 1669 if (file) 1670 uprobe_mmap(vma); 1671 1672 /* 1673 * New (or expanded) vma always get soft dirty status. 1674 * Otherwise user-space soft-dirty page tracker won't 1675 * be able to distinguish situation when vma area unmapped, 1676 * then new mapped in-place (which must be aimed as 1677 * a completely new data area). 1678 */ 1679 vma->vm_flags |= VM_SOFTDIRTY; 1680 1681 vma_set_page_prot(vma); 1682 1683 return addr; 1684 1685 unmap_and_free_vma: 1686 vma->vm_file = NULL; 1687 fput(file); 1688 1689 /* Undo any partial mapping done by a device driver. */ 1690 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1691 charged = 0; 1692 if (vm_flags & VM_SHARED) 1693 mapping_unmap_writable(file->f_mapping); 1694 allow_write_and_free_vma: 1695 if (vm_flags & VM_DENYWRITE) 1696 allow_write_access(file); 1697 free_vma: 1698 kmem_cache_free(vm_area_cachep, vma); 1699 unacct_error: 1700 if (charged) 1701 vm_unacct_memory(charged); 1702 return error; 1703 } 1704 1705 unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1706 { 1707 /* 1708 * We implement the search by looking for an rbtree node that 1709 * immediately follows a suitable gap. That is, 1710 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length; 1711 * - gap_end = vma->vm_start >= info->low_limit + length; 1712 * - gap_end - gap_start >= length 1713 */ 1714 1715 struct mm_struct *mm = current->mm; 1716 struct vm_area_struct *vma; 1717 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1718 1719 /* Adjust search length to account for worst case alignment overhead */ 1720 length = info->length + info->align_mask; 1721 if (length < info->length) 1722 return -ENOMEM; 1723 1724 /* Adjust search limits by the desired length */ 1725 if (info->high_limit < length) 1726 return -ENOMEM; 1727 high_limit = info->high_limit - length; 1728 1729 if (info->low_limit > high_limit) 1730 return -ENOMEM; 1731 low_limit = info->low_limit + length; 1732 1733 /* Check if rbtree root looks promising */ 1734 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1735 goto check_highest; 1736 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1737 if (vma->rb_subtree_gap < length) 1738 goto check_highest; 1739 1740 while (true) { 1741 /* Visit left subtree if it looks promising */ 1742 gap_end = vma->vm_start; 1743 if (gap_end >= low_limit && vma->vm_rb.rb_left) { 1744 struct vm_area_struct *left = 1745 rb_entry(vma->vm_rb.rb_left, 1746 struct vm_area_struct, vm_rb); 1747 if (left->rb_subtree_gap >= length) { 1748 vma = left; 1749 continue; 1750 } 1751 } 1752 1753 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1754 check_current: 1755 /* Check if current node has a suitable gap */ 1756 if (gap_start > high_limit) 1757 return -ENOMEM; 1758 if (gap_end >= low_limit && gap_end - gap_start >= length) 1759 goto found; 1760 1761 /* Visit right subtree if it looks promising */ 1762 if (vma->vm_rb.rb_right) { 1763 struct vm_area_struct *right = 1764 rb_entry(vma->vm_rb.rb_right, 1765 struct vm_area_struct, vm_rb); 1766 if (right->rb_subtree_gap >= length) { 1767 vma = right; 1768 continue; 1769 } 1770 } 1771 1772 /* Go back up the rbtree to find next candidate node */ 1773 while (true) { 1774 struct rb_node *prev = &vma->vm_rb; 1775 if (!rb_parent(prev)) 1776 goto check_highest; 1777 vma = rb_entry(rb_parent(prev), 1778 struct vm_area_struct, vm_rb); 1779 if (prev == vma->vm_rb.rb_left) { 1780 gap_start = vma->vm_prev->vm_end; 1781 gap_end = vma->vm_start; 1782 goto check_current; 1783 } 1784 } 1785 } 1786 1787 check_highest: 1788 /* Check highest gap, which does not precede any rbtree node */ 1789 gap_start = mm->highest_vm_end; 1790 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */ 1791 if (gap_start > high_limit) 1792 return -ENOMEM; 1793 1794 found: 1795 /* We found a suitable gap. Clip it with the original low_limit. */ 1796 if (gap_start < info->low_limit) 1797 gap_start = info->low_limit; 1798 1799 /* Adjust gap address to the desired alignment */ 1800 gap_start += (info->align_offset - gap_start) & info->align_mask; 1801 1802 VM_BUG_ON(gap_start + info->length > info->high_limit); 1803 VM_BUG_ON(gap_start + info->length > gap_end); 1804 return gap_start; 1805 } 1806 1807 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1808 { 1809 struct mm_struct *mm = current->mm; 1810 struct vm_area_struct *vma; 1811 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1812 1813 /* Adjust search length to account for worst case alignment overhead */ 1814 length = info->length + info->align_mask; 1815 if (length < info->length) 1816 return -ENOMEM; 1817 1818 /* 1819 * Adjust search limits by the desired length. 1820 * See implementation comment at top of unmapped_area(). 1821 */ 1822 gap_end = info->high_limit; 1823 if (gap_end < length) 1824 return -ENOMEM; 1825 high_limit = gap_end - length; 1826 1827 if (info->low_limit > high_limit) 1828 return -ENOMEM; 1829 low_limit = info->low_limit + length; 1830 1831 /* Check highest gap, which does not precede any rbtree node */ 1832 gap_start = mm->highest_vm_end; 1833 if (gap_start <= high_limit) 1834 goto found_highest; 1835 1836 /* Check if rbtree root looks promising */ 1837 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1838 return -ENOMEM; 1839 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1840 if (vma->rb_subtree_gap < length) 1841 return -ENOMEM; 1842 1843 while (true) { 1844 /* Visit right subtree if it looks promising */ 1845 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1846 if (gap_start <= high_limit && vma->vm_rb.rb_right) { 1847 struct vm_area_struct *right = 1848 rb_entry(vma->vm_rb.rb_right, 1849 struct vm_area_struct, vm_rb); 1850 if (right->rb_subtree_gap >= length) { 1851 vma = right; 1852 continue; 1853 } 1854 } 1855 1856 check_current: 1857 /* Check if current node has a suitable gap */ 1858 gap_end = vma->vm_start; 1859 if (gap_end < low_limit) 1860 return -ENOMEM; 1861 if (gap_start <= high_limit && gap_end - gap_start >= length) 1862 goto found; 1863 1864 /* Visit left subtree if it looks promising */ 1865 if (vma->vm_rb.rb_left) { 1866 struct vm_area_struct *left = 1867 rb_entry(vma->vm_rb.rb_left, 1868 struct vm_area_struct, vm_rb); 1869 if (left->rb_subtree_gap >= length) { 1870 vma = left; 1871 continue; 1872 } 1873 } 1874 1875 /* Go back up the rbtree to find next candidate node */ 1876 while (true) { 1877 struct rb_node *prev = &vma->vm_rb; 1878 if (!rb_parent(prev)) 1879 return -ENOMEM; 1880 vma = rb_entry(rb_parent(prev), 1881 struct vm_area_struct, vm_rb); 1882 if (prev == vma->vm_rb.rb_right) { 1883 gap_start = vma->vm_prev ? 1884 vma->vm_prev->vm_end : 0; 1885 goto check_current; 1886 } 1887 } 1888 } 1889 1890 found: 1891 /* We found a suitable gap. Clip it with the original high_limit. */ 1892 if (gap_end > info->high_limit) 1893 gap_end = info->high_limit; 1894 1895 found_highest: 1896 /* Compute highest gap address at the desired alignment */ 1897 gap_end -= info->length; 1898 gap_end -= (gap_end - info->align_offset) & info->align_mask; 1899 1900 VM_BUG_ON(gap_end < info->low_limit); 1901 VM_BUG_ON(gap_end < gap_start); 1902 return gap_end; 1903 } 1904 1905 /* Get an address range which is currently unmapped. 1906 * For shmat() with addr=0. 1907 * 1908 * Ugly calling convention alert: 1909 * Return value with the low bits set means error value, 1910 * ie 1911 * if (ret & ~PAGE_MASK) 1912 * error = ret; 1913 * 1914 * This function "knows" that -ENOMEM has the bits set. 1915 */ 1916 #ifndef HAVE_ARCH_UNMAPPED_AREA 1917 unsigned long 1918 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1919 unsigned long len, unsigned long pgoff, unsigned long flags) 1920 { 1921 struct mm_struct *mm = current->mm; 1922 struct vm_area_struct *vma; 1923 struct vm_unmapped_area_info info; 1924 1925 if (len > TASK_SIZE - mmap_min_addr) 1926 return -ENOMEM; 1927 1928 if (flags & MAP_FIXED) 1929 return addr; 1930 1931 if (addr) { 1932 addr = PAGE_ALIGN(addr); 1933 vma = find_vma(mm, addr); 1934 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1935 (!vma || addr + len <= vma->vm_start)) 1936 return addr; 1937 } 1938 1939 info.flags = 0; 1940 info.length = len; 1941 info.low_limit = mm->mmap_base; 1942 info.high_limit = TASK_SIZE; 1943 info.align_mask = 0; 1944 return vm_unmapped_area(&info); 1945 } 1946 #endif 1947 1948 /* 1949 * This mmap-allocator allocates new areas top-down from below the 1950 * stack's low limit (the base): 1951 */ 1952 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1953 unsigned long 1954 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1955 const unsigned long len, const unsigned long pgoff, 1956 const unsigned long flags) 1957 { 1958 struct vm_area_struct *vma; 1959 struct mm_struct *mm = current->mm; 1960 unsigned long addr = addr0; 1961 struct vm_unmapped_area_info info; 1962 1963 /* requested length too big for entire address space */ 1964 if (len > TASK_SIZE - mmap_min_addr) 1965 return -ENOMEM; 1966 1967 if (flags & MAP_FIXED) 1968 return addr; 1969 1970 /* requesting a specific address */ 1971 if (addr) { 1972 addr = PAGE_ALIGN(addr); 1973 vma = find_vma(mm, addr); 1974 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1975 (!vma || addr + len <= vma->vm_start)) 1976 return addr; 1977 } 1978 1979 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 1980 info.length = len; 1981 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 1982 info.high_limit = mm->mmap_base; 1983 info.align_mask = 0; 1984 addr = vm_unmapped_area(&info); 1985 1986 /* 1987 * A failed mmap() very likely causes application failure, 1988 * so fall back to the bottom-up function here. This scenario 1989 * can happen with large stack limits and large mmap() 1990 * allocations. 1991 */ 1992 if (offset_in_page(addr)) { 1993 VM_BUG_ON(addr != -ENOMEM); 1994 info.flags = 0; 1995 info.low_limit = TASK_UNMAPPED_BASE; 1996 info.high_limit = TASK_SIZE; 1997 addr = vm_unmapped_area(&info); 1998 } 1999 2000 return addr; 2001 } 2002 #endif 2003 2004 unsigned long 2005 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 2006 unsigned long pgoff, unsigned long flags) 2007 { 2008 unsigned long (*get_area)(struct file *, unsigned long, 2009 unsigned long, unsigned long, unsigned long); 2010 2011 unsigned long error = arch_mmap_check(addr, len, flags); 2012 if (error) 2013 return error; 2014 2015 /* Careful about overflows.. */ 2016 if (len > TASK_SIZE) 2017 return -ENOMEM; 2018 2019 get_area = current->mm->get_unmapped_area; 2020 if (file && file->f_op->get_unmapped_area) 2021 get_area = file->f_op->get_unmapped_area; 2022 addr = get_area(file, addr, len, pgoff, flags); 2023 if (IS_ERR_VALUE(addr)) 2024 return addr; 2025 2026 if (addr > TASK_SIZE - len) 2027 return -ENOMEM; 2028 if (offset_in_page(addr)) 2029 return -EINVAL; 2030 2031 addr = arch_rebalance_pgtables(addr, len); 2032 error = security_mmap_addr(addr); 2033 return error ? error : addr; 2034 } 2035 2036 EXPORT_SYMBOL(get_unmapped_area); 2037 2038 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 2039 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 2040 { 2041 struct rb_node *rb_node; 2042 struct vm_area_struct *vma; 2043 2044 /* Check the cache first. */ 2045 vma = vmacache_find(mm, addr); 2046 if (likely(vma)) 2047 return vma; 2048 2049 rb_node = mm->mm_rb.rb_node; 2050 2051 while (rb_node) { 2052 struct vm_area_struct *tmp; 2053 2054 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2055 2056 if (tmp->vm_end > addr) { 2057 vma = tmp; 2058 if (tmp->vm_start <= addr) 2059 break; 2060 rb_node = rb_node->rb_left; 2061 } else 2062 rb_node = rb_node->rb_right; 2063 } 2064 2065 if (vma) 2066 vmacache_update(addr, vma); 2067 return vma; 2068 } 2069 2070 EXPORT_SYMBOL(find_vma); 2071 2072 /* 2073 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 2074 */ 2075 struct vm_area_struct * 2076 find_vma_prev(struct mm_struct *mm, unsigned long addr, 2077 struct vm_area_struct **pprev) 2078 { 2079 struct vm_area_struct *vma; 2080 2081 vma = find_vma(mm, addr); 2082 if (vma) { 2083 *pprev = vma->vm_prev; 2084 } else { 2085 struct rb_node *rb_node = mm->mm_rb.rb_node; 2086 *pprev = NULL; 2087 while (rb_node) { 2088 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2089 rb_node = rb_node->rb_right; 2090 } 2091 } 2092 return vma; 2093 } 2094 2095 /* 2096 * Verify that the stack growth is acceptable and 2097 * update accounting. This is shared with both the 2098 * grow-up and grow-down cases. 2099 */ 2100 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 2101 { 2102 struct mm_struct *mm = vma->vm_mm; 2103 struct rlimit *rlim = current->signal->rlim; 2104 unsigned long new_start, actual_size; 2105 2106 /* address space limit tests */ 2107 if (!may_expand_vm(mm, vma->vm_flags, grow)) 2108 return -ENOMEM; 2109 2110 /* Stack limit test */ 2111 actual_size = size; 2112 if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN))) 2113 actual_size -= PAGE_SIZE; 2114 if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 2115 return -ENOMEM; 2116 2117 /* mlock limit tests */ 2118 if (vma->vm_flags & VM_LOCKED) { 2119 unsigned long locked; 2120 unsigned long limit; 2121 locked = mm->locked_vm + grow; 2122 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 2123 limit >>= PAGE_SHIFT; 2124 if (locked > limit && !capable(CAP_IPC_LOCK)) 2125 return -ENOMEM; 2126 } 2127 2128 /* Check to ensure the stack will not grow into a hugetlb-only region */ 2129 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 2130 vma->vm_end - size; 2131 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 2132 return -EFAULT; 2133 2134 /* 2135 * Overcommit.. This must be the final test, as it will 2136 * update security statistics. 2137 */ 2138 if (security_vm_enough_memory_mm(mm, grow)) 2139 return -ENOMEM; 2140 2141 return 0; 2142 } 2143 2144 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 2145 /* 2146 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 2147 * vma is the last one with address > vma->vm_end. Have to extend vma. 2148 */ 2149 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 2150 { 2151 struct mm_struct *mm = vma->vm_mm; 2152 int error = 0; 2153 2154 if (!(vma->vm_flags & VM_GROWSUP)) 2155 return -EFAULT; 2156 2157 /* Guard against wrapping around to address 0. */ 2158 if (address < PAGE_ALIGN(address+4)) 2159 address = PAGE_ALIGN(address+4); 2160 else 2161 return -ENOMEM; 2162 2163 /* We must make sure the anon_vma is allocated. */ 2164 if (unlikely(anon_vma_prepare(vma))) 2165 return -ENOMEM; 2166 2167 /* 2168 * vma->vm_start/vm_end cannot change under us because the caller 2169 * is required to hold the mmap_sem in read mode. We need the 2170 * anon_vma lock to serialize against concurrent expand_stacks. 2171 */ 2172 anon_vma_lock_write(vma->anon_vma); 2173 2174 /* Somebody else might have raced and expanded it already */ 2175 if (address > vma->vm_end) { 2176 unsigned long size, grow; 2177 2178 size = address - vma->vm_start; 2179 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2180 2181 error = -ENOMEM; 2182 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2183 error = acct_stack_growth(vma, size, grow); 2184 if (!error) { 2185 /* 2186 * vma_gap_update() doesn't support concurrent 2187 * updates, but we only hold a shared mmap_sem 2188 * lock here, so we need to protect against 2189 * concurrent vma expansions. 2190 * anon_vma_lock_write() doesn't help here, as 2191 * we don't guarantee that all growable vmas 2192 * in a mm share the same root anon vma. 2193 * So, we reuse mm->page_table_lock to guard 2194 * against concurrent vma expansions. 2195 */ 2196 spin_lock(&mm->page_table_lock); 2197 if (vma->vm_flags & VM_LOCKED) 2198 mm->locked_vm += grow; 2199 vm_stat_account(mm, vma->vm_flags, grow); 2200 anon_vma_interval_tree_pre_update_vma(vma); 2201 vma->vm_end = address; 2202 anon_vma_interval_tree_post_update_vma(vma); 2203 if (vma->vm_next) 2204 vma_gap_update(vma->vm_next); 2205 else 2206 mm->highest_vm_end = address; 2207 spin_unlock(&mm->page_table_lock); 2208 2209 perf_event_mmap(vma); 2210 } 2211 } 2212 } 2213 anon_vma_unlock_write(vma->anon_vma); 2214 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2215 validate_mm(mm); 2216 return error; 2217 } 2218 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2219 2220 /* 2221 * vma is the first one with address < vma->vm_start. Have to extend vma. 2222 */ 2223 int expand_downwards(struct vm_area_struct *vma, 2224 unsigned long address) 2225 { 2226 struct mm_struct *mm = vma->vm_mm; 2227 int error; 2228 2229 address &= PAGE_MASK; 2230 error = security_mmap_addr(address); 2231 if (error) 2232 return error; 2233 2234 /* We must make sure the anon_vma is allocated. */ 2235 if (unlikely(anon_vma_prepare(vma))) 2236 return -ENOMEM; 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 anon_vma_lock_write(vma->anon_vma); 2244 2245 /* Somebody else might have raced and expanded it already */ 2246 if (address < vma->vm_start) { 2247 unsigned long size, grow; 2248 2249 size = vma->vm_end - address; 2250 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2251 2252 error = -ENOMEM; 2253 if (grow <= vma->vm_pgoff) { 2254 error = acct_stack_growth(vma, size, grow); 2255 if (!error) { 2256 /* 2257 * vma_gap_update() doesn't support concurrent 2258 * updates, but we only hold a shared mmap_sem 2259 * lock here, so we need to protect against 2260 * concurrent vma expansions. 2261 * anon_vma_lock_write() doesn't help here, as 2262 * we don't guarantee that all growable vmas 2263 * in a mm share the same root anon vma. 2264 * So, we reuse mm->page_table_lock to guard 2265 * against concurrent vma expansions. 2266 */ 2267 spin_lock(&mm->page_table_lock); 2268 if (vma->vm_flags & VM_LOCKED) 2269 mm->locked_vm += grow; 2270 vm_stat_account(mm, vma->vm_flags, grow); 2271 anon_vma_interval_tree_pre_update_vma(vma); 2272 vma->vm_start = address; 2273 vma->vm_pgoff -= grow; 2274 anon_vma_interval_tree_post_update_vma(vma); 2275 vma_gap_update(vma); 2276 spin_unlock(&mm->page_table_lock); 2277 2278 perf_event_mmap(vma); 2279 } 2280 } 2281 } 2282 anon_vma_unlock_write(vma->anon_vma); 2283 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2284 validate_mm(mm); 2285 return error; 2286 } 2287 2288 /* 2289 * Note how expand_stack() refuses to expand the stack all the way to 2290 * abut the next virtual mapping, *unless* that mapping itself is also 2291 * a stack mapping. We want to leave room for a guard page, after all 2292 * (the guard page itself is not added here, that is done by the 2293 * actual page faulting logic) 2294 * 2295 * This matches the behavior of the guard page logic (see mm/memory.c: 2296 * check_stack_guard_page()), which only allows the guard page to be 2297 * removed under these circumstances. 2298 */ 2299 #ifdef CONFIG_STACK_GROWSUP 2300 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2301 { 2302 struct vm_area_struct *next; 2303 2304 address &= PAGE_MASK; 2305 next = vma->vm_next; 2306 if (next && next->vm_start == address + PAGE_SIZE) { 2307 if (!(next->vm_flags & VM_GROWSUP)) 2308 return -ENOMEM; 2309 } 2310 return expand_upwards(vma, address); 2311 } 2312 2313 struct vm_area_struct * 2314 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2315 { 2316 struct vm_area_struct *vma, *prev; 2317 2318 addr &= PAGE_MASK; 2319 vma = find_vma_prev(mm, addr, &prev); 2320 if (vma && (vma->vm_start <= addr)) 2321 return vma; 2322 if (!prev || expand_stack(prev, addr)) 2323 return NULL; 2324 if (prev->vm_flags & VM_LOCKED) 2325 populate_vma_page_range(prev, addr, prev->vm_end, NULL); 2326 return prev; 2327 } 2328 #else 2329 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2330 { 2331 struct vm_area_struct *prev; 2332 2333 address &= PAGE_MASK; 2334 prev = vma->vm_prev; 2335 if (prev && prev->vm_end == address) { 2336 if (!(prev->vm_flags & VM_GROWSDOWN)) 2337 return -ENOMEM; 2338 } 2339 return expand_downwards(vma, address); 2340 } 2341 2342 struct vm_area_struct * 2343 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2344 { 2345 struct vm_area_struct *vma; 2346 unsigned long start; 2347 2348 addr &= PAGE_MASK; 2349 vma = find_vma(mm, addr); 2350 if (!vma) 2351 return NULL; 2352 if (vma->vm_start <= addr) 2353 return vma; 2354 if (!(vma->vm_flags & VM_GROWSDOWN)) 2355 return NULL; 2356 start = vma->vm_start; 2357 if (expand_stack(vma, addr)) 2358 return NULL; 2359 if (vma->vm_flags & VM_LOCKED) 2360 populate_vma_page_range(vma, addr, start, NULL); 2361 return vma; 2362 } 2363 #endif 2364 2365 EXPORT_SYMBOL_GPL(find_extend_vma); 2366 2367 /* 2368 * Ok - we have the memory areas we should free on the vma list, 2369 * so release them, and do the vma updates. 2370 * 2371 * Called with the mm semaphore held. 2372 */ 2373 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 2374 { 2375 unsigned long nr_accounted = 0; 2376 2377 /* Update high watermark before we lower total_vm */ 2378 update_hiwater_vm(mm); 2379 do { 2380 long nrpages = vma_pages(vma); 2381 2382 if (vma->vm_flags & VM_ACCOUNT) 2383 nr_accounted += nrpages; 2384 vm_stat_account(mm, vma->vm_flags, -nrpages); 2385 vma = remove_vma(vma); 2386 } while (vma); 2387 vm_unacct_memory(nr_accounted); 2388 validate_mm(mm); 2389 } 2390 2391 /* 2392 * Get rid of page table information in the indicated region. 2393 * 2394 * Called with the mm semaphore held. 2395 */ 2396 static void unmap_region(struct mm_struct *mm, 2397 struct vm_area_struct *vma, struct vm_area_struct *prev, 2398 unsigned long start, unsigned long end) 2399 { 2400 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap; 2401 struct mmu_gather tlb; 2402 2403 lru_add_drain(); 2404 tlb_gather_mmu(&tlb, mm, start, end); 2405 update_hiwater_rss(mm); 2406 unmap_vmas(&tlb, vma, start, end); 2407 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2408 next ? next->vm_start : USER_PGTABLES_CEILING); 2409 tlb_finish_mmu(&tlb, start, end); 2410 } 2411 2412 /* 2413 * Create a list of vma's touched by the unmap, removing them from the mm's 2414 * vma list as we go.. 2415 */ 2416 static void 2417 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 2418 struct vm_area_struct *prev, unsigned long end) 2419 { 2420 struct vm_area_struct **insertion_point; 2421 struct vm_area_struct *tail_vma = NULL; 2422 2423 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 2424 vma->vm_prev = NULL; 2425 do { 2426 vma_rb_erase(vma, &mm->mm_rb); 2427 mm->map_count--; 2428 tail_vma = vma; 2429 vma = vma->vm_next; 2430 } while (vma && vma->vm_start < end); 2431 *insertion_point = vma; 2432 if (vma) { 2433 vma->vm_prev = prev; 2434 vma_gap_update(vma); 2435 } else 2436 mm->highest_vm_end = prev ? prev->vm_end : 0; 2437 tail_vma->vm_next = NULL; 2438 2439 /* Kill the cache */ 2440 vmacache_invalidate(mm); 2441 } 2442 2443 /* 2444 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the 2445 * munmap path where it doesn't make sense to fail. 2446 */ 2447 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2448 unsigned long addr, int new_below) 2449 { 2450 struct vm_area_struct *new; 2451 int err; 2452 2453 if (is_vm_hugetlb_page(vma) && (addr & 2454 ~(huge_page_mask(hstate_vma(vma))))) 2455 return -EINVAL; 2456 2457 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2458 if (!new) 2459 return -ENOMEM; 2460 2461 /* most fields are the same, copy all, and then fixup */ 2462 *new = *vma; 2463 2464 INIT_LIST_HEAD(&new->anon_vma_chain); 2465 2466 if (new_below) 2467 new->vm_end = addr; 2468 else { 2469 new->vm_start = addr; 2470 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2471 } 2472 2473 err = vma_dup_policy(vma, new); 2474 if (err) 2475 goto out_free_vma; 2476 2477 err = anon_vma_clone(new, vma); 2478 if (err) 2479 goto out_free_mpol; 2480 2481 if (new->vm_file) 2482 get_file(new->vm_file); 2483 2484 if (new->vm_ops && new->vm_ops->open) 2485 new->vm_ops->open(new); 2486 2487 if (new_below) 2488 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 2489 ((addr - new->vm_start) >> PAGE_SHIFT), new); 2490 else 2491 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 2492 2493 /* Success. */ 2494 if (!err) 2495 return 0; 2496 2497 /* Clean everything up if vma_adjust failed. */ 2498 if (new->vm_ops && new->vm_ops->close) 2499 new->vm_ops->close(new); 2500 if (new->vm_file) 2501 fput(new->vm_file); 2502 unlink_anon_vmas(new); 2503 out_free_mpol: 2504 mpol_put(vma_policy(new)); 2505 out_free_vma: 2506 kmem_cache_free(vm_area_cachep, new); 2507 return err; 2508 } 2509 2510 /* 2511 * Split a vma into two pieces at address 'addr', a new vma is allocated 2512 * either for the first part or the tail. 2513 */ 2514 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2515 unsigned long addr, int new_below) 2516 { 2517 if (mm->map_count >= sysctl_max_map_count) 2518 return -ENOMEM; 2519 2520 return __split_vma(mm, vma, addr, new_below); 2521 } 2522 2523 /* Munmap is split into 2 main parts -- this part which finds 2524 * what needs doing, and the areas themselves, which do the 2525 * work. This now handles partial unmappings. 2526 * Jeremy Fitzhardinge <jeremy@goop.org> 2527 */ 2528 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 2529 { 2530 unsigned long end; 2531 struct vm_area_struct *vma, *prev, *last; 2532 2533 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start) 2534 return -EINVAL; 2535 2536 len = PAGE_ALIGN(len); 2537 if (len == 0) 2538 return -EINVAL; 2539 2540 /* Find the first overlapping VMA */ 2541 vma = find_vma(mm, start); 2542 if (!vma) 2543 return 0; 2544 prev = vma->vm_prev; 2545 /* we have start < vma->vm_end */ 2546 2547 /* if it doesn't overlap, we have nothing.. */ 2548 end = start + len; 2549 if (vma->vm_start >= end) 2550 return 0; 2551 2552 /* 2553 * If we need to split any vma, do it now to save pain later. 2554 * 2555 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2556 * unmapped vm_area_struct will remain in use: so lower split_vma 2557 * places tmp vma above, and higher split_vma places tmp vma below. 2558 */ 2559 if (start > vma->vm_start) { 2560 int error; 2561 2562 /* 2563 * Make sure that map_count on return from munmap() will 2564 * not exceed its limit; but let map_count go just above 2565 * its limit temporarily, to help free resources as expected. 2566 */ 2567 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2568 return -ENOMEM; 2569 2570 error = __split_vma(mm, vma, start, 0); 2571 if (error) 2572 return error; 2573 prev = vma; 2574 } 2575 2576 /* Does it split the last one? */ 2577 last = find_vma(mm, end); 2578 if (last && end > last->vm_start) { 2579 int error = __split_vma(mm, last, end, 1); 2580 if (error) 2581 return error; 2582 } 2583 vma = prev ? prev->vm_next : mm->mmap; 2584 2585 /* 2586 * unlock any mlock()ed ranges before detaching vmas 2587 */ 2588 if (mm->locked_vm) { 2589 struct vm_area_struct *tmp = vma; 2590 while (tmp && tmp->vm_start < end) { 2591 if (tmp->vm_flags & VM_LOCKED) { 2592 mm->locked_vm -= vma_pages(tmp); 2593 munlock_vma_pages_all(tmp); 2594 } 2595 tmp = tmp->vm_next; 2596 } 2597 } 2598 2599 /* 2600 * Remove the vma's, and unmap the actual pages 2601 */ 2602 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2603 unmap_region(mm, vma, prev, start, end); 2604 2605 arch_unmap(mm, vma, start, end); 2606 2607 /* Fix up all other VM information */ 2608 remove_vma_list(mm, vma); 2609 2610 return 0; 2611 } 2612 2613 int vm_munmap(unsigned long start, size_t len) 2614 { 2615 int ret; 2616 struct mm_struct *mm = current->mm; 2617 2618 down_write(&mm->mmap_sem); 2619 ret = do_munmap(mm, start, len); 2620 up_write(&mm->mmap_sem); 2621 return ret; 2622 } 2623 EXPORT_SYMBOL(vm_munmap); 2624 2625 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2626 { 2627 profile_munmap(addr); 2628 return vm_munmap(addr, len); 2629 } 2630 2631 2632 /* 2633 * Emulation of deprecated remap_file_pages() syscall. 2634 */ 2635 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size, 2636 unsigned long, prot, unsigned long, pgoff, unsigned long, flags) 2637 { 2638 2639 struct mm_struct *mm = current->mm; 2640 struct vm_area_struct *vma; 2641 unsigned long populate = 0; 2642 unsigned long ret = -EINVAL; 2643 struct file *file; 2644 2645 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. " 2646 "See Documentation/vm/remap_file_pages.txt.\n", 2647 current->comm, current->pid); 2648 2649 if (prot) 2650 return ret; 2651 start = start & PAGE_MASK; 2652 size = size & PAGE_MASK; 2653 2654 if (start + size <= start) 2655 return ret; 2656 2657 /* Does pgoff wrap? */ 2658 if (pgoff + (size >> PAGE_SHIFT) < pgoff) 2659 return ret; 2660 2661 down_write(&mm->mmap_sem); 2662 vma = find_vma(mm, start); 2663 2664 if (!vma || !(vma->vm_flags & VM_SHARED)) 2665 goto out; 2666 2667 if (start < vma->vm_start || start + size > vma->vm_end) 2668 goto out; 2669 2670 if (pgoff == linear_page_index(vma, start)) { 2671 ret = 0; 2672 goto out; 2673 } 2674 2675 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0; 2676 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0; 2677 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0; 2678 2679 flags &= MAP_NONBLOCK; 2680 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE; 2681 if (vma->vm_flags & VM_LOCKED) { 2682 flags |= MAP_LOCKED; 2683 /* drop PG_Mlocked flag for over-mapped range */ 2684 munlock_vma_pages_range(vma, start, start + size); 2685 } 2686 2687 file = get_file(vma->vm_file); 2688 ret = do_mmap_pgoff(vma->vm_file, start, size, 2689 prot, flags, pgoff, &populate); 2690 fput(file); 2691 out: 2692 up_write(&mm->mmap_sem); 2693 if (populate) 2694 mm_populate(ret, populate); 2695 if (!IS_ERR_VALUE(ret)) 2696 ret = 0; 2697 return ret; 2698 } 2699 2700 static inline void verify_mm_writelocked(struct mm_struct *mm) 2701 { 2702 #ifdef CONFIG_DEBUG_VM 2703 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2704 WARN_ON(1); 2705 up_read(&mm->mmap_sem); 2706 } 2707 #endif 2708 } 2709 2710 /* 2711 * this is really a simplified "do_mmap". it only handles 2712 * anonymous maps. eventually we may be able to do some 2713 * brk-specific accounting here. 2714 */ 2715 static unsigned long do_brk(unsigned long addr, unsigned long len) 2716 { 2717 struct mm_struct *mm = current->mm; 2718 struct vm_area_struct *vma, *prev; 2719 unsigned long flags; 2720 struct rb_node **rb_link, *rb_parent; 2721 pgoff_t pgoff = addr >> PAGE_SHIFT; 2722 int error; 2723 2724 len = PAGE_ALIGN(len); 2725 if (!len) 2726 return addr; 2727 2728 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2729 2730 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2731 if (offset_in_page(error)) 2732 return error; 2733 2734 error = mlock_future_check(mm, mm->def_flags, len); 2735 if (error) 2736 return error; 2737 2738 /* 2739 * mm->mmap_sem is required to protect against another thread 2740 * changing the mappings in case we sleep. 2741 */ 2742 verify_mm_writelocked(mm); 2743 2744 /* 2745 * Clear old maps. this also does some error checking for us 2746 */ 2747 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link, 2748 &rb_parent)) { 2749 if (do_munmap(mm, addr, len)) 2750 return -ENOMEM; 2751 } 2752 2753 /* Check against address space limits *after* clearing old maps... */ 2754 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT)) 2755 return -ENOMEM; 2756 2757 if (mm->map_count > sysctl_max_map_count) 2758 return -ENOMEM; 2759 2760 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 2761 return -ENOMEM; 2762 2763 /* Can we just expand an old private anonymous mapping? */ 2764 vma = vma_merge(mm, prev, addr, addr + len, flags, 2765 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX); 2766 if (vma) 2767 goto out; 2768 2769 /* 2770 * create a vma struct for an anonymous mapping 2771 */ 2772 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2773 if (!vma) { 2774 vm_unacct_memory(len >> PAGE_SHIFT); 2775 return -ENOMEM; 2776 } 2777 2778 INIT_LIST_HEAD(&vma->anon_vma_chain); 2779 vma->vm_mm = mm; 2780 vma->vm_start = addr; 2781 vma->vm_end = addr + len; 2782 vma->vm_pgoff = pgoff; 2783 vma->vm_flags = flags; 2784 vma->vm_page_prot = vm_get_page_prot(flags); 2785 vma_link(mm, vma, prev, rb_link, rb_parent); 2786 out: 2787 perf_event_mmap(vma); 2788 mm->total_vm += len >> PAGE_SHIFT; 2789 mm->data_vm += len >> PAGE_SHIFT; 2790 if (flags & VM_LOCKED) 2791 mm->locked_vm += (len >> PAGE_SHIFT); 2792 vma->vm_flags |= VM_SOFTDIRTY; 2793 return addr; 2794 } 2795 2796 unsigned long vm_brk(unsigned long addr, unsigned long len) 2797 { 2798 struct mm_struct *mm = current->mm; 2799 unsigned long ret; 2800 bool populate; 2801 2802 down_write(&mm->mmap_sem); 2803 ret = do_brk(addr, len); 2804 populate = ((mm->def_flags & VM_LOCKED) != 0); 2805 up_write(&mm->mmap_sem); 2806 if (populate) 2807 mm_populate(addr, len); 2808 return ret; 2809 } 2810 EXPORT_SYMBOL(vm_brk); 2811 2812 /* Release all mmaps. */ 2813 void exit_mmap(struct mm_struct *mm) 2814 { 2815 struct mmu_gather tlb; 2816 struct vm_area_struct *vma; 2817 unsigned long nr_accounted = 0; 2818 2819 /* mm's last user has gone, and its about to be pulled down */ 2820 mmu_notifier_release(mm); 2821 2822 if (mm->locked_vm) { 2823 vma = mm->mmap; 2824 while (vma) { 2825 if (vma->vm_flags & VM_LOCKED) 2826 munlock_vma_pages_all(vma); 2827 vma = vma->vm_next; 2828 } 2829 } 2830 2831 arch_exit_mmap(mm); 2832 2833 vma = mm->mmap; 2834 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2835 return; 2836 2837 lru_add_drain(); 2838 flush_cache_mm(mm); 2839 tlb_gather_mmu(&tlb, mm, 0, -1); 2840 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2841 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2842 unmap_vmas(&tlb, vma, 0, -1); 2843 2844 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING); 2845 tlb_finish_mmu(&tlb, 0, -1); 2846 2847 /* 2848 * Walk the list again, actually closing and freeing it, 2849 * with preemption enabled, without holding any MM locks. 2850 */ 2851 while (vma) { 2852 if (vma->vm_flags & VM_ACCOUNT) 2853 nr_accounted += vma_pages(vma); 2854 vma = remove_vma(vma); 2855 } 2856 vm_unacct_memory(nr_accounted); 2857 } 2858 2859 /* Insert vm structure into process list sorted by address 2860 * and into the inode's i_mmap tree. If vm_file is non-NULL 2861 * then i_mmap_rwsem is taken here. 2862 */ 2863 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 2864 { 2865 struct vm_area_struct *prev; 2866 struct rb_node **rb_link, *rb_parent; 2867 2868 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 2869 &prev, &rb_link, &rb_parent)) 2870 return -ENOMEM; 2871 if ((vma->vm_flags & VM_ACCOUNT) && 2872 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2873 return -ENOMEM; 2874 2875 /* 2876 * The vm_pgoff of a purely anonymous vma should be irrelevant 2877 * until its first write fault, when page's anon_vma and index 2878 * are set. But now set the vm_pgoff it will almost certainly 2879 * end up with (unless mremap moves it elsewhere before that 2880 * first wfault), so /proc/pid/maps tells a consistent story. 2881 * 2882 * By setting it to reflect the virtual start address of the 2883 * vma, merges and splits can happen in a seamless way, just 2884 * using the existing file pgoff checks and manipulations. 2885 * Similarly in do_mmap_pgoff and in do_brk. 2886 */ 2887 if (vma_is_anonymous(vma)) { 2888 BUG_ON(vma->anon_vma); 2889 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2890 } 2891 2892 vma_link(mm, vma, prev, rb_link, rb_parent); 2893 return 0; 2894 } 2895 2896 /* 2897 * Copy the vma structure to a new location in the same mm, 2898 * prior to moving page table entries, to effect an mremap move. 2899 */ 2900 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2901 unsigned long addr, unsigned long len, pgoff_t pgoff, 2902 bool *need_rmap_locks) 2903 { 2904 struct vm_area_struct *vma = *vmap; 2905 unsigned long vma_start = vma->vm_start; 2906 struct mm_struct *mm = vma->vm_mm; 2907 struct vm_area_struct *new_vma, *prev; 2908 struct rb_node **rb_link, *rb_parent; 2909 bool faulted_in_anon_vma = true; 2910 2911 /* 2912 * If anonymous vma has not yet been faulted, update new pgoff 2913 * to match new location, to increase its chance of merging. 2914 */ 2915 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) { 2916 pgoff = addr >> PAGE_SHIFT; 2917 faulted_in_anon_vma = false; 2918 } 2919 2920 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) 2921 return NULL; /* should never get here */ 2922 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2923 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 2924 vma->vm_userfaultfd_ctx); 2925 if (new_vma) { 2926 /* 2927 * Source vma may have been merged into new_vma 2928 */ 2929 if (unlikely(vma_start >= new_vma->vm_start && 2930 vma_start < new_vma->vm_end)) { 2931 /* 2932 * The only way we can get a vma_merge with 2933 * self during an mremap is if the vma hasn't 2934 * been faulted in yet and we were allowed to 2935 * reset the dst vma->vm_pgoff to the 2936 * destination address of the mremap to allow 2937 * the merge to happen. mremap must change the 2938 * vm_pgoff linearity between src and dst vmas 2939 * (in turn preventing a vma_merge) to be 2940 * safe. It is only safe to keep the vm_pgoff 2941 * linear if there are no pages mapped yet. 2942 */ 2943 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 2944 *vmap = vma = new_vma; 2945 } 2946 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 2947 } else { 2948 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2949 if (!new_vma) 2950 goto out; 2951 *new_vma = *vma; 2952 new_vma->vm_start = addr; 2953 new_vma->vm_end = addr + len; 2954 new_vma->vm_pgoff = pgoff; 2955 if (vma_dup_policy(vma, new_vma)) 2956 goto out_free_vma; 2957 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 2958 if (anon_vma_clone(new_vma, vma)) 2959 goto out_free_mempol; 2960 if (new_vma->vm_file) 2961 get_file(new_vma->vm_file); 2962 if (new_vma->vm_ops && new_vma->vm_ops->open) 2963 new_vma->vm_ops->open(new_vma); 2964 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2965 *need_rmap_locks = false; 2966 } 2967 return new_vma; 2968 2969 out_free_mempol: 2970 mpol_put(vma_policy(new_vma)); 2971 out_free_vma: 2972 kmem_cache_free(vm_area_cachep, new_vma); 2973 out: 2974 return NULL; 2975 } 2976 2977 /* 2978 * Return true if the calling process may expand its vm space by the passed 2979 * number of pages 2980 */ 2981 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages) 2982 { 2983 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT) 2984 return false; 2985 2986 if (is_data_mapping(flags) && 2987 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) { 2988 if (ignore_rlimit_data) 2989 pr_warn_once("%s (%d): VmData %lu exceed data ulimit " 2990 "%lu. Will be forbidden soon.\n", 2991 current->comm, current->pid, 2992 (mm->data_vm + npages) << PAGE_SHIFT, 2993 rlimit(RLIMIT_DATA)); 2994 else 2995 return false; 2996 } 2997 2998 return true; 2999 } 3000 3001 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages) 3002 { 3003 mm->total_vm += npages; 3004 3005 if (is_exec_mapping(flags)) 3006 mm->exec_vm += npages; 3007 else if (is_stack_mapping(flags)) 3008 mm->stack_vm += npages; 3009 else if (is_data_mapping(flags)) 3010 mm->data_vm += npages; 3011 } 3012 3013 static int special_mapping_fault(struct vm_area_struct *vma, 3014 struct vm_fault *vmf); 3015 3016 /* 3017 * Having a close hook prevents vma merging regardless of flags. 3018 */ 3019 static void special_mapping_close(struct vm_area_struct *vma) 3020 { 3021 } 3022 3023 static const char *special_mapping_name(struct vm_area_struct *vma) 3024 { 3025 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 3026 } 3027 3028 static const struct vm_operations_struct special_mapping_vmops = { 3029 .close = special_mapping_close, 3030 .fault = special_mapping_fault, 3031 .name = special_mapping_name, 3032 }; 3033 3034 static const struct vm_operations_struct legacy_special_mapping_vmops = { 3035 .close = special_mapping_close, 3036 .fault = special_mapping_fault, 3037 }; 3038 3039 static int special_mapping_fault(struct vm_area_struct *vma, 3040 struct vm_fault *vmf) 3041 { 3042 pgoff_t pgoff; 3043 struct page **pages; 3044 3045 if (vma->vm_ops == &legacy_special_mapping_vmops) 3046 pages = vma->vm_private_data; 3047 else 3048 pages = ((struct vm_special_mapping *)vma->vm_private_data)-> 3049 pages; 3050 3051 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages) 3052 pgoff--; 3053 3054 if (*pages) { 3055 struct page *page = *pages; 3056 get_page(page); 3057 vmf->page = page; 3058 return 0; 3059 } 3060 3061 return VM_FAULT_SIGBUS; 3062 } 3063 3064 static struct vm_area_struct *__install_special_mapping( 3065 struct mm_struct *mm, 3066 unsigned long addr, unsigned long len, 3067 unsigned long vm_flags, void *priv, 3068 const struct vm_operations_struct *ops) 3069 { 3070 int ret; 3071 struct vm_area_struct *vma; 3072 3073 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 3074 if (unlikely(vma == NULL)) 3075 return ERR_PTR(-ENOMEM); 3076 3077 INIT_LIST_HEAD(&vma->anon_vma_chain); 3078 vma->vm_mm = mm; 3079 vma->vm_start = addr; 3080 vma->vm_end = addr + len; 3081 3082 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY; 3083 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3084 3085 vma->vm_ops = ops; 3086 vma->vm_private_data = priv; 3087 3088 ret = insert_vm_struct(mm, vma); 3089 if (ret) 3090 goto out; 3091 3092 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT); 3093 3094 perf_event_mmap(vma); 3095 3096 return vma; 3097 3098 out: 3099 kmem_cache_free(vm_area_cachep, vma); 3100 return ERR_PTR(ret); 3101 } 3102 3103 /* 3104 * Called with mm->mmap_sem held for writing. 3105 * Insert a new vma covering the given region, with the given flags. 3106 * Its pages are supplied by the given array of struct page *. 3107 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3108 * The region past the last page supplied will always produce SIGBUS. 3109 * The array pointer and the pages it points to are assumed to stay alive 3110 * for as long as this mapping might exist. 3111 */ 3112 struct vm_area_struct *_install_special_mapping( 3113 struct mm_struct *mm, 3114 unsigned long addr, unsigned long len, 3115 unsigned long vm_flags, const struct vm_special_mapping *spec) 3116 { 3117 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec, 3118 &special_mapping_vmops); 3119 } 3120 3121 int install_special_mapping(struct mm_struct *mm, 3122 unsigned long addr, unsigned long len, 3123 unsigned long vm_flags, struct page **pages) 3124 { 3125 struct vm_area_struct *vma = __install_special_mapping( 3126 mm, addr, len, vm_flags, (void *)pages, 3127 &legacy_special_mapping_vmops); 3128 3129 return PTR_ERR_OR_ZERO(vma); 3130 } 3131 3132 static DEFINE_MUTEX(mm_all_locks_mutex); 3133 3134 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3135 { 3136 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3137 /* 3138 * The LSB of head.next can't change from under us 3139 * because we hold the mm_all_locks_mutex. 3140 */ 3141 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem); 3142 /* 3143 * We can safely modify head.next after taking the 3144 * anon_vma->root->rwsem. If some other vma in this mm shares 3145 * the same anon_vma we won't take it again. 3146 * 3147 * No need of atomic instructions here, head.next 3148 * can't change from under us thanks to the 3149 * anon_vma->root->rwsem. 3150 */ 3151 if (__test_and_set_bit(0, (unsigned long *) 3152 &anon_vma->root->rb_root.rb_node)) 3153 BUG(); 3154 } 3155 } 3156 3157 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3158 { 3159 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3160 /* 3161 * AS_MM_ALL_LOCKS can't change from under us because 3162 * we hold the mm_all_locks_mutex. 3163 * 3164 * Operations on ->flags have to be atomic because 3165 * even if AS_MM_ALL_LOCKS is stable thanks to the 3166 * mm_all_locks_mutex, there may be other cpus 3167 * changing other bitflags in parallel to us. 3168 */ 3169 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3170 BUG(); 3171 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem); 3172 } 3173 } 3174 3175 /* 3176 * This operation locks against the VM for all pte/vma/mm related 3177 * operations that could ever happen on a certain mm. This includes 3178 * vmtruncate, try_to_unmap, and all page faults. 3179 * 3180 * The caller must take the mmap_sem in write mode before calling 3181 * mm_take_all_locks(). The caller isn't allowed to release the 3182 * mmap_sem until mm_drop_all_locks() returns. 3183 * 3184 * mmap_sem in write mode is required in order to block all operations 3185 * that could modify pagetables and free pages without need of 3186 * altering the vma layout. It's also needed in write mode to avoid new 3187 * anon_vmas to be associated with existing vmas. 3188 * 3189 * A single task can't take more than one mm_take_all_locks() in a row 3190 * or it would deadlock. 3191 * 3192 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3193 * mapping->flags avoid to take the same lock twice, if more than one 3194 * vma in this mm is backed by the same anon_vma or address_space. 3195 * 3196 * We take locks in following order, accordingly to comment at beginning 3197 * of mm/rmap.c: 3198 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for 3199 * hugetlb mapping); 3200 * - all i_mmap_rwsem locks; 3201 * - all anon_vma->rwseml 3202 * 3203 * We can take all locks within these types randomly because the VM code 3204 * doesn't nest them and we protected from parallel mm_take_all_locks() by 3205 * mm_all_locks_mutex. 3206 * 3207 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3208 * that may have to take thousand of locks. 3209 * 3210 * mm_take_all_locks() can fail if it's interrupted by signals. 3211 */ 3212 int mm_take_all_locks(struct mm_struct *mm) 3213 { 3214 struct vm_area_struct *vma; 3215 struct anon_vma_chain *avc; 3216 3217 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3218 3219 mutex_lock(&mm_all_locks_mutex); 3220 3221 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3222 if (signal_pending(current)) 3223 goto out_unlock; 3224 if (vma->vm_file && vma->vm_file->f_mapping && 3225 is_vm_hugetlb_page(vma)) 3226 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3227 } 3228 3229 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3230 if (signal_pending(current)) 3231 goto out_unlock; 3232 if (vma->vm_file && vma->vm_file->f_mapping && 3233 !is_vm_hugetlb_page(vma)) 3234 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3235 } 3236 3237 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3238 if (signal_pending(current)) 3239 goto out_unlock; 3240 if (vma->anon_vma) 3241 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3242 vm_lock_anon_vma(mm, avc->anon_vma); 3243 } 3244 3245 return 0; 3246 3247 out_unlock: 3248 mm_drop_all_locks(mm); 3249 return -EINTR; 3250 } 3251 3252 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3253 { 3254 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3255 /* 3256 * The LSB of head.next can't change to 0 from under 3257 * us because we hold the mm_all_locks_mutex. 3258 * 3259 * We must however clear the bitflag before unlocking 3260 * the vma so the users using the anon_vma->rb_root will 3261 * never see our bitflag. 3262 * 3263 * No need of atomic instructions here, head.next 3264 * can't change from under us until we release the 3265 * anon_vma->root->rwsem. 3266 */ 3267 if (!__test_and_clear_bit(0, (unsigned long *) 3268 &anon_vma->root->rb_root.rb_node)) 3269 BUG(); 3270 anon_vma_unlock_write(anon_vma); 3271 } 3272 } 3273 3274 static void vm_unlock_mapping(struct address_space *mapping) 3275 { 3276 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3277 /* 3278 * AS_MM_ALL_LOCKS can't change to 0 from under us 3279 * because we hold the mm_all_locks_mutex. 3280 */ 3281 i_mmap_unlock_write(mapping); 3282 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3283 &mapping->flags)) 3284 BUG(); 3285 } 3286 } 3287 3288 /* 3289 * The mmap_sem cannot be released by the caller until 3290 * mm_drop_all_locks() returns. 3291 */ 3292 void mm_drop_all_locks(struct mm_struct *mm) 3293 { 3294 struct vm_area_struct *vma; 3295 struct anon_vma_chain *avc; 3296 3297 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3298 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3299 3300 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3301 if (vma->anon_vma) 3302 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3303 vm_unlock_anon_vma(avc->anon_vma); 3304 if (vma->vm_file && vma->vm_file->f_mapping) 3305 vm_unlock_mapping(vma->vm_file->f_mapping); 3306 } 3307 3308 mutex_unlock(&mm_all_locks_mutex); 3309 } 3310 3311 /* 3312 * initialise the VMA slab 3313 */ 3314 void __init mmap_init(void) 3315 { 3316 int ret; 3317 3318 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3319 VM_BUG_ON(ret); 3320 } 3321 3322 /* 3323 * Initialise sysctl_user_reserve_kbytes. 3324 * 3325 * This is intended to prevent a user from starting a single memory hogging 3326 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3327 * mode. 3328 * 3329 * The default value is min(3% of free memory, 128MB) 3330 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3331 */ 3332 static int init_user_reserve(void) 3333 { 3334 unsigned long free_kbytes; 3335 3336 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3337 3338 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3339 return 0; 3340 } 3341 subsys_initcall(init_user_reserve); 3342 3343 /* 3344 * Initialise sysctl_admin_reserve_kbytes. 3345 * 3346 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3347 * to log in and kill a memory hogging process. 3348 * 3349 * Systems with more than 256MB will reserve 8MB, enough to recover 3350 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3351 * only reserve 3% of free pages by default. 3352 */ 3353 static int init_admin_reserve(void) 3354 { 3355 unsigned long free_kbytes; 3356 3357 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3358 3359 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3360 return 0; 3361 } 3362 subsys_initcall(init_admin_reserve); 3363 3364 /* 3365 * Reinititalise user and admin reserves if memory is added or removed. 3366 * 3367 * The default user reserve max is 128MB, and the default max for the 3368 * admin reserve is 8MB. These are usually, but not always, enough to 3369 * enable recovery from a memory hogging process using login/sshd, a shell, 3370 * and tools like top. It may make sense to increase or even disable the 3371 * reserve depending on the existence of swap or variations in the recovery 3372 * tools. So, the admin may have changed them. 3373 * 3374 * If memory is added and the reserves have been eliminated or increased above 3375 * the default max, then we'll trust the admin. 3376 * 3377 * If memory is removed and there isn't enough free memory, then we 3378 * need to reset the reserves. 3379 * 3380 * Otherwise keep the reserve set by the admin. 3381 */ 3382 static int reserve_mem_notifier(struct notifier_block *nb, 3383 unsigned long action, void *data) 3384 { 3385 unsigned long tmp, free_kbytes; 3386 3387 switch (action) { 3388 case MEM_ONLINE: 3389 /* Default max is 128MB. Leave alone if modified by operator. */ 3390 tmp = sysctl_user_reserve_kbytes; 3391 if (0 < tmp && tmp < (1UL << 17)) 3392 init_user_reserve(); 3393 3394 /* Default max is 8MB. Leave alone if modified by operator. */ 3395 tmp = sysctl_admin_reserve_kbytes; 3396 if (0 < tmp && tmp < (1UL << 13)) 3397 init_admin_reserve(); 3398 3399 break; 3400 case MEM_OFFLINE: 3401 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3402 3403 if (sysctl_user_reserve_kbytes > free_kbytes) { 3404 init_user_reserve(); 3405 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3406 sysctl_user_reserve_kbytes); 3407 } 3408 3409 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3410 init_admin_reserve(); 3411 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3412 sysctl_admin_reserve_kbytes); 3413 } 3414 break; 3415 default: 3416 break; 3417 } 3418 return NOTIFY_OK; 3419 } 3420 3421 static struct notifier_block reserve_mem_nb = { 3422 .notifier_call = reserve_mem_notifier, 3423 }; 3424 3425 static int __meminit init_reserve_notifier(void) 3426 { 3427 if (register_hotmemory_notifier(&reserve_mem_nb)) 3428 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3429 3430 return 0; 3431 } 3432 subsys_initcall(init_reserve_notifier); 3433