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 46 #include <asm/uaccess.h> 47 #include <asm/cacheflush.h> 48 #include <asm/tlb.h> 49 #include <asm/mmu_context.h> 50 51 #include "internal.h" 52 53 #ifndef arch_mmap_check 54 #define arch_mmap_check(addr, len, flags) (0) 55 #endif 56 57 #ifndef arch_rebalance_pgtables 58 #define arch_rebalance_pgtables(addr, len) (addr) 59 #endif 60 61 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS 62 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN; 63 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX; 64 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS; 65 #endif 66 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS 67 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN; 68 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX; 69 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS; 70 #endif 71 72 73 static void unmap_region(struct mm_struct *mm, 74 struct vm_area_struct *vma, struct vm_area_struct *prev, 75 unsigned long start, unsigned long end); 76 77 /* description of effects of mapping type and prot in current implementation. 78 * this is due to the limited x86 page protection hardware. The expected 79 * behavior is in parens: 80 * 81 * map_type prot 82 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 83 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 84 * w: (no) no w: (no) no w: (yes) yes w: (no) no 85 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 86 * 87 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 88 * w: (no) no w: (no) no w: (copy) copy w: (no) no 89 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 90 * 91 */ 92 pgprot_t protection_map[16] = { 93 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 94 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 95 }; 96 97 pgprot_t vm_get_page_prot(unsigned long vm_flags) 98 { 99 return __pgprot(pgprot_val(protection_map[vm_flags & 100 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) | 101 pgprot_val(arch_vm_get_page_prot(vm_flags))); 102 } 103 EXPORT_SYMBOL(vm_get_page_prot); 104 105 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags) 106 { 107 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags)); 108 } 109 110 /* Update vma->vm_page_prot to reflect vma->vm_flags. */ 111 void vma_set_page_prot(struct vm_area_struct *vma) 112 { 113 unsigned long vm_flags = vma->vm_flags; 114 115 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags); 116 if (vma_wants_writenotify(vma)) { 117 vm_flags &= ~VM_SHARED; 118 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, 119 vm_flags); 120 } 121 } 122 123 124 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */ 125 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */ 126 unsigned long sysctl_overcommit_kbytes __read_mostly; 127 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT; 128 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */ 129 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */ 130 /* 131 * Make sure vm_committed_as in one cacheline and not cacheline shared with 132 * other variables. It can be updated by several CPUs frequently. 133 */ 134 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp; 135 136 /* 137 * The global memory commitment made in the system can be a metric 138 * that can be used to drive ballooning decisions when Linux is hosted 139 * as a guest. On Hyper-V, the host implements a policy engine for dynamically 140 * balancing memory across competing virtual machines that are hosted. 141 * Several metrics drive this policy engine including the guest reported 142 * memory commitment. 143 */ 144 unsigned long vm_memory_committed(void) 145 { 146 return percpu_counter_read_positive(&vm_committed_as); 147 } 148 EXPORT_SYMBOL_GPL(vm_memory_committed); 149 150 /* 151 * Check that a process has enough memory to allocate a new virtual 152 * mapping. 0 means there is enough memory for the allocation to 153 * succeed and -ENOMEM implies there is not. 154 * 155 * We currently support three overcommit policies, which are set via the 156 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting 157 * 158 * Strict overcommit modes added 2002 Feb 26 by Alan Cox. 159 * Additional code 2002 Jul 20 by Robert Love. 160 * 161 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise. 162 * 163 * Note this is a helper function intended to be used by LSMs which 164 * wish to use this logic. 165 */ 166 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin) 167 { 168 long free, allowed, reserve; 169 170 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) < 171 -(s64)vm_committed_as_batch * num_online_cpus(), 172 "memory commitment underflow"); 173 174 vm_acct_memory(pages); 175 176 /* 177 * Sometimes we want to use more memory than we have 178 */ 179 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS) 180 return 0; 181 182 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) { 183 free = global_page_state(NR_FREE_PAGES); 184 free += global_page_state(NR_FILE_PAGES); 185 186 /* 187 * shmem pages shouldn't be counted as free in this 188 * case, they can't be purged, only swapped out, and 189 * that won't affect the overall amount of available 190 * memory in the system. 191 */ 192 free -= global_page_state(NR_SHMEM); 193 194 free += get_nr_swap_pages(); 195 196 /* 197 * Any slabs which are created with the 198 * SLAB_RECLAIM_ACCOUNT flag claim to have contents 199 * which are reclaimable, under pressure. The dentry 200 * cache and most inode caches should fall into this 201 */ 202 free += global_page_state(NR_SLAB_RECLAIMABLE); 203 204 /* 205 * Leave reserved pages. The pages are not for anonymous pages. 206 */ 207 if (free <= totalreserve_pages) 208 goto error; 209 else 210 free -= totalreserve_pages; 211 212 /* 213 * Reserve some for root 214 */ 215 if (!cap_sys_admin) 216 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10); 217 218 if (free > pages) 219 return 0; 220 221 goto error; 222 } 223 224 allowed = vm_commit_limit(); 225 /* 226 * Reserve some for root 227 */ 228 if (!cap_sys_admin) 229 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10); 230 231 /* 232 * Don't let a single process grow so big a user can't recover 233 */ 234 if (mm) { 235 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10); 236 allowed -= min_t(long, mm->total_vm / 32, reserve); 237 } 238 239 if (percpu_counter_read_positive(&vm_committed_as) < allowed) 240 return 0; 241 error: 242 vm_unacct_memory(pages); 243 244 return -ENOMEM; 245 } 246 247 /* 248 * Requires inode->i_mapping->i_mmap_rwsem 249 */ 250 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 251 struct file *file, struct address_space *mapping) 252 { 253 if (vma->vm_flags & VM_DENYWRITE) 254 atomic_inc(&file_inode(file)->i_writecount); 255 if (vma->vm_flags & VM_SHARED) 256 mapping_unmap_writable(mapping); 257 258 flush_dcache_mmap_lock(mapping); 259 vma_interval_tree_remove(vma, &mapping->i_mmap); 260 flush_dcache_mmap_unlock(mapping); 261 } 262 263 /* 264 * Unlink a file-based vm structure from its interval tree, to hide 265 * vma from rmap and vmtruncate before freeing its page tables. 266 */ 267 void unlink_file_vma(struct vm_area_struct *vma) 268 { 269 struct file *file = vma->vm_file; 270 271 if (file) { 272 struct address_space *mapping = file->f_mapping; 273 i_mmap_lock_write(mapping); 274 __remove_shared_vm_struct(vma, file, mapping); 275 i_mmap_unlock_write(mapping); 276 } 277 } 278 279 /* 280 * Close a vm structure and free it, returning the next. 281 */ 282 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 283 { 284 struct vm_area_struct *next = vma->vm_next; 285 286 might_sleep(); 287 if (vma->vm_ops && vma->vm_ops->close) 288 vma->vm_ops->close(vma); 289 if (vma->vm_file) 290 fput(vma->vm_file); 291 mpol_put(vma_policy(vma)); 292 kmem_cache_free(vm_area_cachep, vma); 293 return next; 294 } 295 296 static unsigned long do_brk(unsigned long addr, unsigned long len); 297 298 SYSCALL_DEFINE1(brk, unsigned long, brk) 299 { 300 unsigned long retval; 301 unsigned long newbrk, oldbrk; 302 struct mm_struct *mm = current->mm; 303 unsigned long min_brk; 304 bool populate; 305 306 down_write(&mm->mmap_sem); 307 308 #ifdef CONFIG_COMPAT_BRK 309 /* 310 * CONFIG_COMPAT_BRK can still be overridden by setting 311 * randomize_va_space to 2, which will still cause mm->start_brk 312 * to be arbitrarily shifted 313 */ 314 if (current->brk_randomized) 315 min_brk = mm->start_brk; 316 else 317 min_brk = mm->end_data; 318 #else 319 min_brk = mm->start_brk; 320 #endif 321 if (brk < min_brk) 322 goto out; 323 324 /* 325 * Check against rlimit here. If this check is done later after the test 326 * of oldbrk with newbrk then it can escape the test and let the data 327 * segment grow beyond its set limit the in case where the limit is 328 * not page aligned -Ram Gupta 329 */ 330 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 331 mm->end_data, mm->start_data)) 332 goto out; 333 334 newbrk = PAGE_ALIGN(brk); 335 oldbrk = PAGE_ALIGN(mm->brk); 336 if (oldbrk == newbrk) 337 goto set_brk; 338 339 /* Always allow shrinking brk. */ 340 if (brk <= mm->brk) { 341 if (!do_munmap(mm, newbrk, oldbrk-newbrk)) 342 goto set_brk; 343 goto out; 344 } 345 346 /* Check against existing mmap mappings. */ 347 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE)) 348 goto out; 349 350 /* Ok, looks good - let it rip. */ 351 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk) 352 goto out; 353 354 set_brk: 355 mm->brk = brk; 356 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 357 up_write(&mm->mmap_sem); 358 if (populate) 359 mm_populate(oldbrk, newbrk - oldbrk); 360 return brk; 361 362 out: 363 retval = mm->brk; 364 up_write(&mm->mmap_sem); 365 return retval; 366 } 367 368 static long vma_compute_subtree_gap(struct vm_area_struct *vma) 369 { 370 unsigned long max, subtree_gap; 371 max = vma->vm_start; 372 if (vma->vm_prev) 373 max -= vma->vm_prev->vm_end; 374 if (vma->vm_rb.rb_left) { 375 subtree_gap = rb_entry(vma->vm_rb.rb_left, 376 struct vm_area_struct, vm_rb)->rb_subtree_gap; 377 if (subtree_gap > max) 378 max = subtree_gap; 379 } 380 if (vma->vm_rb.rb_right) { 381 subtree_gap = rb_entry(vma->vm_rb.rb_right, 382 struct vm_area_struct, vm_rb)->rb_subtree_gap; 383 if (subtree_gap > max) 384 max = subtree_gap; 385 } 386 return max; 387 } 388 389 #ifdef CONFIG_DEBUG_VM_RB 390 static int browse_rb(struct rb_root *root) 391 { 392 int i = 0, j, bug = 0; 393 struct rb_node *nd, *pn = NULL; 394 unsigned long prev = 0, pend = 0; 395 396 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 397 struct vm_area_struct *vma; 398 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 399 if (vma->vm_start < prev) { 400 pr_emerg("vm_start %lx < prev %lx\n", 401 vma->vm_start, prev); 402 bug = 1; 403 } 404 if (vma->vm_start < pend) { 405 pr_emerg("vm_start %lx < pend %lx\n", 406 vma->vm_start, pend); 407 bug = 1; 408 } 409 if (vma->vm_start > vma->vm_end) { 410 pr_emerg("vm_start %lx > vm_end %lx\n", 411 vma->vm_start, vma->vm_end); 412 bug = 1; 413 } 414 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) { 415 pr_emerg("free gap %lx, correct %lx\n", 416 vma->rb_subtree_gap, 417 vma_compute_subtree_gap(vma)); 418 bug = 1; 419 } 420 i++; 421 pn = nd; 422 prev = vma->vm_start; 423 pend = vma->vm_end; 424 } 425 j = 0; 426 for (nd = pn; nd; nd = rb_prev(nd)) 427 j++; 428 if (i != j) { 429 pr_emerg("backwards %d, forwards %d\n", j, i); 430 bug = 1; 431 } 432 return bug ? -1 : i; 433 } 434 435 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore) 436 { 437 struct rb_node *nd; 438 439 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 440 struct vm_area_struct *vma; 441 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 442 VM_BUG_ON_VMA(vma != ignore && 443 vma->rb_subtree_gap != vma_compute_subtree_gap(vma), 444 vma); 445 } 446 } 447 448 static void validate_mm(struct mm_struct *mm) 449 { 450 int bug = 0; 451 int i = 0; 452 unsigned long highest_address = 0; 453 struct vm_area_struct *vma = mm->mmap; 454 455 while (vma) { 456 struct anon_vma_chain *avc; 457 458 vma_lock_anon_vma(vma); 459 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 460 anon_vma_interval_tree_verify(avc); 461 vma_unlock_anon_vma(vma); 462 highest_address = vma->vm_end; 463 vma = vma->vm_next; 464 i++; 465 } 466 if (i != mm->map_count) { 467 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i); 468 bug = 1; 469 } 470 if (highest_address != mm->highest_vm_end) { 471 pr_emerg("mm->highest_vm_end %lx, found %lx\n", 472 mm->highest_vm_end, highest_address); 473 bug = 1; 474 } 475 i = browse_rb(&mm->mm_rb); 476 if (i != mm->map_count) { 477 if (i != -1) 478 pr_emerg("map_count %d rb %d\n", mm->map_count, i); 479 bug = 1; 480 } 481 VM_BUG_ON_MM(bug, mm); 482 } 483 #else 484 #define validate_mm_rb(root, ignore) do { } while (0) 485 #define validate_mm(mm) do { } while (0) 486 #endif 487 488 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb, 489 unsigned long, rb_subtree_gap, vma_compute_subtree_gap) 490 491 /* 492 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or 493 * vma->vm_prev->vm_end values changed, without modifying the vma's position 494 * in the rbtree. 495 */ 496 static void vma_gap_update(struct vm_area_struct *vma) 497 { 498 /* 499 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback 500 * function that does exacltly what we want. 501 */ 502 vma_gap_callbacks_propagate(&vma->vm_rb, NULL); 503 } 504 505 static inline void vma_rb_insert(struct vm_area_struct *vma, 506 struct rb_root *root) 507 { 508 /* All rb_subtree_gap values must be consistent prior to insertion */ 509 validate_mm_rb(root, NULL); 510 511 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 512 } 513 514 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root) 515 { 516 /* 517 * All rb_subtree_gap values must be consistent prior to erase, 518 * with the possible exception of the vma being erased. 519 */ 520 validate_mm_rb(root, vma); 521 522 /* 523 * Note rb_erase_augmented is a fairly large inline function, 524 * so make sure we instantiate it only once with our desired 525 * augmented rbtree callbacks. 526 */ 527 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 528 } 529 530 /* 531 * vma has some anon_vma assigned, and is already inserted on that 532 * anon_vma's interval trees. 533 * 534 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 535 * vma must be removed from the anon_vma's interval trees using 536 * anon_vma_interval_tree_pre_update_vma(). 537 * 538 * After the update, the vma will be reinserted using 539 * anon_vma_interval_tree_post_update_vma(). 540 * 541 * The entire update must be protected by exclusive mmap_sem and by 542 * the root anon_vma's mutex. 543 */ 544 static inline void 545 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 546 { 547 struct anon_vma_chain *avc; 548 549 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 550 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 551 } 552 553 static inline void 554 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 555 { 556 struct anon_vma_chain *avc; 557 558 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 559 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 560 } 561 562 static int find_vma_links(struct mm_struct *mm, unsigned long addr, 563 unsigned long end, struct vm_area_struct **pprev, 564 struct rb_node ***rb_link, struct rb_node **rb_parent) 565 { 566 struct rb_node **__rb_link, *__rb_parent, *rb_prev; 567 568 __rb_link = &mm->mm_rb.rb_node; 569 rb_prev = __rb_parent = NULL; 570 571 while (*__rb_link) { 572 struct vm_area_struct *vma_tmp; 573 574 __rb_parent = *__rb_link; 575 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 576 577 if (vma_tmp->vm_end > addr) { 578 /* Fail if an existing vma overlaps the area */ 579 if (vma_tmp->vm_start < end) 580 return -ENOMEM; 581 __rb_link = &__rb_parent->rb_left; 582 } else { 583 rb_prev = __rb_parent; 584 __rb_link = &__rb_parent->rb_right; 585 } 586 } 587 588 *pprev = NULL; 589 if (rb_prev) 590 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 591 *rb_link = __rb_link; 592 *rb_parent = __rb_parent; 593 return 0; 594 } 595 596 static unsigned long count_vma_pages_range(struct mm_struct *mm, 597 unsigned long addr, unsigned long end) 598 { 599 unsigned long nr_pages = 0; 600 struct vm_area_struct *vma; 601 602 /* Find first overlaping mapping */ 603 vma = find_vma_intersection(mm, addr, end); 604 if (!vma) 605 return 0; 606 607 nr_pages = (min(end, vma->vm_end) - 608 max(addr, vma->vm_start)) >> PAGE_SHIFT; 609 610 /* Iterate over the rest of the overlaps */ 611 for (vma = vma->vm_next; vma; vma = vma->vm_next) { 612 unsigned long overlap_len; 613 614 if (vma->vm_start > end) 615 break; 616 617 overlap_len = min(end, vma->vm_end) - vma->vm_start; 618 nr_pages += overlap_len >> PAGE_SHIFT; 619 } 620 621 return nr_pages; 622 } 623 624 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 625 struct rb_node **rb_link, struct rb_node *rb_parent) 626 { 627 /* Update tracking information for the gap following the new vma. */ 628 if (vma->vm_next) 629 vma_gap_update(vma->vm_next); 630 else 631 mm->highest_vm_end = vma->vm_end; 632 633 /* 634 * vma->vm_prev wasn't known when we followed the rbtree to find the 635 * correct insertion point for that vma. As a result, we could not 636 * update the vma vm_rb parents rb_subtree_gap values on the way down. 637 * So, we first insert the vma with a zero rb_subtree_gap value 638 * (to be consistent with what we did on the way down), and then 639 * immediately update the gap to the correct value. Finally we 640 * rebalance the rbtree after all augmented values have been set. 641 */ 642 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 643 vma->rb_subtree_gap = 0; 644 vma_gap_update(vma); 645 vma_rb_insert(vma, &mm->mm_rb); 646 } 647 648 static void __vma_link_file(struct vm_area_struct *vma) 649 { 650 struct file *file; 651 652 file = vma->vm_file; 653 if (file) { 654 struct address_space *mapping = file->f_mapping; 655 656 if (vma->vm_flags & VM_DENYWRITE) 657 atomic_dec(&file_inode(file)->i_writecount); 658 if (vma->vm_flags & VM_SHARED) 659 atomic_inc(&mapping->i_mmap_writable); 660 661 flush_dcache_mmap_lock(mapping); 662 vma_interval_tree_insert(vma, &mapping->i_mmap); 663 flush_dcache_mmap_unlock(mapping); 664 } 665 } 666 667 static void 668 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 669 struct vm_area_struct *prev, struct rb_node **rb_link, 670 struct rb_node *rb_parent) 671 { 672 __vma_link_list(mm, vma, prev, rb_parent); 673 __vma_link_rb(mm, vma, rb_link, rb_parent); 674 } 675 676 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 677 struct vm_area_struct *prev, struct rb_node **rb_link, 678 struct rb_node *rb_parent) 679 { 680 struct address_space *mapping = NULL; 681 682 if (vma->vm_file) { 683 mapping = vma->vm_file->f_mapping; 684 i_mmap_lock_write(mapping); 685 } 686 687 __vma_link(mm, vma, prev, rb_link, rb_parent); 688 __vma_link_file(vma); 689 690 if (mapping) 691 i_mmap_unlock_write(mapping); 692 693 mm->map_count++; 694 validate_mm(mm); 695 } 696 697 /* 698 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the 699 * mm's list and rbtree. It has already been inserted into the interval tree. 700 */ 701 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 702 { 703 struct vm_area_struct *prev; 704 struct rb_node **rb_link, *rb_parent; 705 706 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 707 &prev, &rb_link, &rb_parent)) 708 BUG(); 709 __vma_link(mm, vma, prev, rb_link, rb_parent); 710 mm->map_count++; 711 } 712 713 static inline void 714 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma, 715 struct vm_area_struct *prev) 716 { 717 struct vm_area_struct *next; 718 719 vma_rb_erase(vma, &mm->mm_rb); 720 prev->vm_next = next = vma->vm_next; 721 if (next) 722 next->vm_prev = prev; 723 724 /* Kill the cache */ 725 vmacache_invalidate(mm); 726 } 727 728 /* 729 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 730 * is already present in an i_mmap tree without adjusting the tree. 731 * The following helper function should be used when such adjustments 732 * are necessary. The "insert" vma (if any) is to be inserted 733 * before we drop the necessary locks. 734 */ 735 int vma_adjust(struct vm_area_struct *vma, unsigned long start, 736 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert) 737 { 738 struct mm_struct *mm = vma->vm_mm; 739 struct vm_area_struct *next = vma->vm_next; 740 struct vm_area_struct *importer = NULL; 741 struct address_space *mapping = NULL; 742 struct rb_root *root = NULL; 743 struct anon_vma *anon_vma = NULL; 744 struct file *file = vma->vm_file; 745 bool start_changed = false, end_changed = false; 746 long adjust_next = 0; 747 int remove_next = 0; 748 749 if (next && !insert) { 750 struct vm_area_struct *exporter = NULL; 751 752 if (end >= next->vm_end) { 753 /* 754 * vma expands, overlapping all the next, and 755 * perhaps the one after too (mprotect case 6). 756 */ 757 again: remove_next = 1 + (end > next->vm_end); 758 end = next->vm_end; 759 exporter = next; 760 importer = vma; 761 } else if (end > next->vm_start) { 762 /* 763 * vma expands, overlapping part of the next: 764 * mprotect case 5 shifting the boundary up. 765 */ 766 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 767 exporter = next; 768 importer = vma; 769 } else if (end < vma->vm_end) { 770 /* 771 * vma shrinks, and !insert tells it's not 772 * split_vma inserting another: so it must be 773 * mprotect case 4 shifting the boundary down. 774 */ 775 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT); 776 exporter = vma; 777 importer = next; 778 } 779 780 /* 781 * Easily overlooked: when mprotect shifts the boundary, 782 * make sure the expanding vma has anon_vma set if the 783 * shrinking vma had, to cover any anon pages imported. 784 */ 785 if (exporter && exporter->anon_vma && !importer->anon_vma) { 786 int error; 787 788 importer->anon_vma = exporter->anon_vma; 789 error = anon_vma_clone(importer, exporter); 790 if (error) 791 return error; 792 } 793 } 794 795 if (file) { 796 mapping = file->f_mapping; 797 root = &mapping->i_mmap; 798 uprobe_munmap(vma, vma->vm_start, vma->vm_end); 799 800 if (adjust_next) 801 uprobe_munmap(next, next->vm_start, next->vm_end); 802 803 i_mmap_lock_write(mapping); 804 if (insert) { 805 /* 806 * Put into interval tree now, so instantiated pages 807 * are visible to arm/parisc __flush_dcache_page 808 * throughout; but we cannot insert into address 809 * space until vma start or end is updated. 810 */ 811 __vma_link_file(insert); 812 } 813 } 814 815 vma_adjust_trans_huge(vma, start, end, adjust_next); 816 817 anon_vma = vma->anon_vma; 818 if (!anon_vma && adjust_next) 819 anon_vma = next->anon_vma; 820 if (anon_vma) { 821 VM_BUG_ON_VMA(adjust_next && next->anon_vma && 822 anon_vma != next->anon_vma, next); 823 anon_vma_lock_write(anon_vma); 824 anon_vma_interval_tree_pre_update_vma(vma); 825 if (adjust_next) 826 anon_vma_interval_tree_pre_update_vma(next); 827 } 828 829 if (root) { 830 flush_dcache_mmap_lock(mapping); 831 vma_interval_tree_remove(vma, root); 832 if (adjust_next) 833 vma_interval_tree_remove(next, root); 834 } 835 836 if (start != vma->vm_start) { 837 vma->vm_start = start; 838 start_changed = true; 839 } 840 if (end != vma->vm_end) { 841 vma->vm_end = end; 842 end_changed = true; 843 } 844 vma->vm_pgoff = pgoff; 845 if (adjust_next) { 846 next->vm_start += adjust_next << PAGE_SHIFT; 847 next->vm_pgoff += adjust_next; 848 } 849 850 if (root) { 851 if (adjust_next) 852 vma_interval_tree_insert(next, root); 853 vma_interval_tree_insert(vma, root); 854 flush_dcache_mmap_unlock(mapping); 855 } 856 857 if (remove_next) { 858 /* 859 * vma_merge has merged next into vma, and needs 860 * us to remove next before dropping the locks. 861 */ 862 __vma_unlink(mm, next, vma); 863 if (file) 864 __remove_shared_vm_struct(next, file, mapping); 865 } else if (insert) { 866 /* 867 * split_vma has split insert from vma, and needs 868 * us to insert it before dropping the locks 869 * (it may either follow vma or precede it). 870 */ 871 __insert_vm_struct(mm, insert); 872 } else { 873 if (start_changed) 874 vma_gap_update(vma); 875 if (end_changed) { 876 if (!next) 877 mm->highest_vm_end = end; 878 else if (!adjust_next) 879 vma_gap_update(next); 880 } 881 } 882 883 if (anon_vma) { 884 anon_vma_interval_tree_post_update_vma(vma); 885 if (adjust_next) 886 anon_vma_interval_tree_post_update_vma(next); 887 anon_vma_unlock_write(anon_vma); 888 } 889 if (mapping) 890 i_mmap_unlock_write(mapping); 891 892 if (root) { 893 uprobe_mmap(vma); 894 895 if (adjust_next) 896 uprobe_mmap(next); 897 } 898 899 if (remove_next) { 900 if (file) { 901 uprobe_munmap(next, next->vm_start, next->vm_end); 902 fput(file); 903 } 904 if (next->anon_vma) 905 anon_vma_merge(vma, next); 906 mm->map_count--; 907 mpol_put(vma_policy(next)); 908 kmem_cache_free(vm_area_cachep, next); 909 /* 910 * In mprotect's case 6 (see comments on vma_merge), 911 * we must remove another next too. It would clutter 912 * up the code too much to do both in one go. 913 */ 914 next = vma->vm_next; 915 if (remove_next == 2) 916 goto again; 917 else if (next) 918 vma_gap_update(next); 919 else 920 mm->highest_vm_end = end; 921 } 922 if (insert && file) 923 uprobe_mmap(insert); 924 925 validate_mm(mm); 926 927 return 0; 928 } 929 930 /* 931 * If the vma has a ->close operation then the driver probably needs to release 932 * per-vma resources, so we don't attempt to merge those. 933 */ 934 static inline int is_mergeable_vma(struct vm_area_struct *vma, 935 struct file *file, unsigned long vm_flags, 936 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 937 { 938 /* 939 * VM_SOFTDIRTY should not prevent from VMA merging, if we 940 * match the flags but dirty bit -- the caller should mark 941 * merged VMA as dirty. If dirty bit won't be excluded from 942 * comparison, we increase pressue on the memory system forcing 943 * the kernel to generate new VMAs when old one could be 944 * extended instead. 945 */ 946 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 947 return 0; 948 if (vma->vm_file != file) 949 return 0; 950 if (vma->vm_ops && vma->vm_ops->close) 951 return 0; 952 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx)) 953 return 0; 954 return 1; 955 } 956 957 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 958 struct anon_vma *anon_vma2, 959 struct vm_area_struct *vma) 960 { 961 /* 962 * The list_is_singular() test is to avoid merging VMA cloned from 963 * parents. This can improve scalability caused by anon_vma lock. 964 */ 965 if ((!anon_vma1 || !anon_vma2) && (!vma || 966 list_is_singular(&vma->anon_vma_chain))) 967 return 1; 968 return anon_vma1 == anon_vma2; 969 } 970 971 /* 972 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 973 * in front of (at a lower virtual address and file offset than) the vma. 974 * 975 * We cannot merge two vmas if they have differently assigned (non-NULL) 976 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 977 * 978 * We don't check here for the merged mmap wrapping around the end of pagecache 979 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 980 * wrap, nor mmaps which cover the final page at index -1UL. 981 */ 982 static int 983 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 984 struct anon_vma *anon_vma, struct file *file, 985 pgoff_t vm_pgoff, 986 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 987 { 988 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) && 989 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 990 if (vma->vm_pgoff == vm_pgoff) 991 return 1; 992 } 993 return 0; 994 } 995 996 /* 997 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 998 * beyond (at a higher virtual address and file offset than) the vma. 999 * 1000 * We cannot merge two vmas if they have differently assigned (non-NULL) 1001 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 1002 */ 1003 static int 1004 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 1005 struct anon_vma *anon_vma, struct file *file, 1006 pgoff_t vm_pgoff, 1007 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1008 { 1009 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) && 1010 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1011 pgoff_t vm_pglen; 1012 vm_pglen = vma_pages(vma); 1013 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 1014 return 1; 1015 } 1016 return 0; 1017 } 1018 1019 /* 1020 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 1021 * whether that can be merged with its predecessor or its successor. 1022 * Or both (it neatly fills a hole). 1023 * 1024 * In most cases - when called for mmap, brk or mremap - [addr,end) is 1025 * certain not to be mapped by the time vma_merge is called; but when 1026 * called for mprotect, it is certain to be already mapped (either at 1027 * an offset within prev, or at the start of next), and the flags of 1028 * this area are about to be changed to vm_flags - and the no-change 1029 * case has already been eliminated. 1030 * 1031 * The following mprotect cases have to be considered, where AAAA is 1032 * the area passed down from mprotect_fixup, never extending beyond one 1033 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 1034 * 1035 * AAAA AAAA AAAA AAAA 1036 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 1037 * cannot merge might become might become might become 1038 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 1039 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 1040 * mremap move: PPPPNNNNNNNN 8 1041 * AAAA 1042 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 1043 * might become case 1 below case 2 below case 3 below 1044 * 1045 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX: 1046 * mprotect_fixup updates vm_flags & vm_page_prot on successful return. 1047 */ 1048 struct vm_area_struct *vma_merge(struct mm_struct *mm, 1049 struct vm_area_struct *prev, unsigned long addr, 1050 unsigned long end, unsigned long vm_flags, 1051 struct anon_vma *anon_vma, struct file *file, 1052 pgoff_t pgoff, struct mempolicy *policy, 1053 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1054 { 1055 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 1056 struct vm_area_struct *area, *next; 1057 int err; 1058 1059 /* 1060 * We later require that vma->vm_flags == vm_flags, 1061 * so this tests vma->vm_flags & VM_SPECIAL, too. 1062 */ 1063 if (vm_flags & VM_SPECIAL) 1064 return NULL; 1065 1066 if (prev) 1067 next = prev->vm_next; 1068 else 1069 next = mm->mmap; 1070 area = next; 1071 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 1072 next = next->vm_next; 1073 1074 /* 1075 * Can it merge with the predecessor? 1076 */ 1077 if (prev && prev->vm_end == addr && 1078 mpol_equal(vma_policy(prev), policy) && 1079 can_vma_merge_after(prev, vm_flags, 1080 anon_vma, file, pgoff, 1081 vm_userfaultfd_ctx)) { 1082 /* 1083 * OK, it can. Can we now merge in the successor as well? 1084 */ 1085 if (next && end == next->vm_start && 1086 mpol_equal(policy, vma_policy(next)) && 1087 can_vma_merge_before(next, vm_flags, 1088 anon_vma, file, 1089 pgoff+pglen, 1090 vm_userfaultfd_ctx) && 1091 is_mergeable_anon_vma(prev->anon_vma, 1092 next->anon_vma, NULL)) { 1093 /* cases 1, 6 */ 1094 err = vma_adjust(prev, prev->vm_start, 1095 next->vm_end, prev->vm_pgoff, NULL); 1096 } else /* cases 2, 5, 7 */ 1097 err = vma_adjust(prev, prev->vm_start, 1098 end, prev->vm_pgoff, NULL); 1099 if (err) 1100 return NULL; 1101 khugepaged_enter_vma_merge(prev, vm_flags); 1102 return prev; 1103 } 1104 1105 /* 1106 * Can this new request be merged in front of next? 1107 */ 1108 if (next && end == next->vm_start && 1109 mpol_equal(policy, vma_policy(next)) && 1110 can_vma_merge_before(next, vm_flags, 1111 anon_vma, file, pgoff+pglen, 1112 vm_userfaultfd_ctx)) { 1113 if (prev && addr < prev->vm_end) /* case 4 */ 1114 err = vma_adjust(prev, prev->vm_start, 1115 addr, prev->vm_pgoff, NULL); 1116 else /* cases 3, 8 */ 1117 err = vma_adjust(area, addr, next->vm_end, 1118 next->vm_pgoff - pglen, NULL); 1119 if (err) 1120 return NULL; 1121 khugepaged_enter_vma_merge(area, vm_flags); 1122 return area; 1123 } 1124 1125 return NULL; 1126 } 1127 1128 /* 1129 * Rough compatbility check to quickly see if it's even worth looking 1130 * at sharing an anon_vma. 1131 * 1132 * They need to have the same vm_file, and the flags can only differ 1133 * in things that mprotect may change. 1134 * 1135 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1136 * we can merge the two vma's. For example, we refuse to merge a vma if 1137 * there is a vm_ops->close() function, because that indicates that the 1138 * driver is doing some kind of reference counting. But that doesn't 1139 * really matter for the anon_vma sharing case. 1140 */ 1141 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1142 { 1143 return a->vm_end == b->vm_start && 1144 mpol_equal(vma_policy(a), vma_policy(b)) && 1145 a->vm_file == b->vm_file && 1146 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) && 1147 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1148 } 1149 1150 /* 1151 * Do some basic sanity checking to see if we can re-use the anon_vma 1152 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1153 * the same as 'old', the other will be the new one that is trying 1154 * to share the anon_vma. 1155 * 1156 * NOTE! This runs with mm_sem held for reading, so it is possible that 1157 * the anon_vma of 'old' is concurrently in the process of being set up 1158 * by another page fault trying to merge _that_. But that's ok: if it 1159 * is being set up, that automatically means that it will be a singleton 1160 * acceptable for merging, so we can do all of this optimistically. But 1161 * we do that READ_ONCE() to make sure that we never re-load the pointer. 1162 * 1163 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1164 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1165 * is to return an anon_vma that is "complex" due to having gone through 1166 * a fork). 1167 * 1168 * We also make sure that the two vma's are compatible (adjacent, 1169 * and with the same memory policies). That's all stable, even with just 1170 * a read lock on the mm_sem. 1171 */ 1172 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1173 { 1174 if (anon_vma_compatible(a, b)) { 1175 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma); 1176 1177 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1178 return anon_vma; 1179 } 1180 return NULL; 1181 } 1182 1183 /* 1184 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1185 * neighbouring vmas for a suitable anon_vma, before it goes off 1186 * to allocate a new anon_vma. It checks because a repetitive 1187 * sequence of mprotects and faults may otherwise lead to distinct 1188 * anon_vmas being allocated, preventing vma merge in subsequent 1189 * mprotect. 1190 */ 1191 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1192 { 1193 struct anon_vma *anon_vma; 1194 struct vm_area_struct *near; 1195 1196 near = vma->vm_next; 1197 if (!near) 1198 goto try_prev; 1199 1200 anon_vma = reusable_anon_vma(near, vma, near); 1201 if (anon_vma) 1202 return anon_vma; 1203 try_prev: 1204 near = vma->vm_prev; 1205 if (!near) 1206 goto none; 1207 1208 anon_vma = reusable_anon_vma(near, near, vma); 1209 if (anon_vma) 1210 return anon_vma; 1211 none: 1212 /* 1213 * There's no absolute need to look only at touching neighbours: 1214 * we could search further afield for "compatible" anon_vmas. 1215 * But it would probably just be a waste of time searching, 1216 * or lead to too many vmas hanging off the same anon_vma. 1217 * We're trying to allow mprotect remerging later on, 1218 * not trying to minimize memory used for anon_vmas. 1219 */ 1220 return NULL; 1221 } 1222 1223 /* 1224 * If a hint addr is less than mmap_min_addr change hint to be as 1225 * low as possible but still greater than mmap_min_addr 1226 */ 1227 static inline unsigned long round_hint_to_min(unsigned long hint) 1228 { 1229 hint &= PAGE_MASK; 1230 if (((void *)hint != NULL) && 1231 (hint < mmap_min_addr)) 1232 return PAGE_ALIGN(mmap_min_addr); 1233 return hint; 1234 } 1235 1236 static inline int mlock_future_check(struct mm_struct *mm, 1237 unsigned long flags, 1238 unsigned long len) 1239 { 1240 unsigned long locked, lock_limit; 1241 1242 /* mlock MCL_FUTURE? */ 1243 if (flags & VM_LOCKED) { 1244 locked = len >> PAGE_SHIFT; 1245 locked += mm->locked_vm; 1246 lock_limit = rlimit(RLIMIT_MEMLOCK); 1247 lock_limit >>= PAGE_SHIFT; 1248 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1249 return -EAGAIN; 1250 } 1251 return 0; 1252 } 1253 1254 /* 1255 * The caller must hold down_write(¤t->mm->mmap_sem). 1256 */ 1257 unsigned long do_mmap(struct file *file, unsigned long addr, 1258 unsigned long len, unsigned long prot, 1259 unsigned long flags, vm_flags_t vm_flags, 1260 unsigned long pgoff, unsigned long *populate) 1261 { 1262 struct mm_struct *mm = current->mm; 1263 1264 *populate = 0; 1265 1266 if (!len) 1267 return -EINVAL; 1268 1269 /* 1270 * Does the application expect PROT_READ to imply PROT_EXEC? 1271 * 1272 * (the exception is when the underlying filesystem is noexec 1273 * mounted, in which case we dont add PROT_EXEC.) 1274 */ 1275 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1276 if (!(file && path_noexec(&file->f_path))) 1277 prot |= PROT_EXEC; 1278 1279 if (!(flags & MAP_FIXED)) 1280 addr = round_hint_to_min(addr); 1281 1282 /* Careful about overflows.. */ 1283 len = PAGE_ALIGN(len); 1284 if (!len) 1285 return -ENOMEM; 1286 1287 /* offset overflow? */ 1288 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1289 return -EOVERFLOW; 1290 1291 /* Too many mappings? */ 1292 if (mm->map_count > sysctl_max_map_count) 1293 return -ENOMEM; 1294 1295 /* Obtain the address to map to. we verify (or select) it and ensure 1296 * that it represents a valid section of the address space. 1297 */ 1298 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1299 if (offset_in_page(addr)) 1300 return addr; 1301 1302 /* Do simple checking here so the lower-level routines won't have 1303 * to. we assume access permissions have been handled by the open 1304 * of the memory object, so we don't do any here. 1305 */ 1306 vm_flags |= calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) | 1307 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1308 1309 if (flags & MAP_LOCKED) 1310 if (!can_do_mlock()) 1311 return -EPERM; 1312 1313 if (mlock_future_check(mm, vm_flags, len)) 1314 return -EAGAIN; 1315 1316 if (file) { 1317 struct inode *inode = file_inode(file); 1318 1319 switch (flags & MAP_TYPE) { 1320 case MAP_SHARED: 1321 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 1322 return -EACCES; 1323 1324 /* 1325 * Make sure we don't allow writing to an append-only 1326 * file.. 1327 */ 1328 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1329 return -EACCES; 1330 1331 /* 1332 * Make sure there are no mandatory locks on the file. 1333 */ 1334 if (locks_verify_locked(file)) 1335 return -EAGAIN; 1336 1337 vm_flags |= VM_SHARED | VM_MAYSHARE; 1338 if (!(file->f_mode & FMODE_WRITE)) 1339 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1340 1341 /* fall through */ 1342 case MAP_PRIVATE: 1343 if (!(file->f_mode & FMODE_READ)) 1344 return -EACCES; 1345 if (path_noexec(&file->f_path)) { 1346 if (vm_flags & VM_EXEC) 1347 return -EPERM; 1348 vm_flags &= ~VM_MAYEXEC; 1349 } 1350 1351 if (!file->f_op->mmap) 1352 return -ENODEV; 1353 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1354 return -EINVAL; 1355 break; 1356 1357 default: 1358 return -EINVAL; 1359 } 1360 } else { 1361 switch (flags & MAP_TYPE) { 1362 case MAP_SHARED: 1363 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1364 return -EINVAL; 1365 /* 1366 * Ignore pgoff. 1367 */ 1368 pgoff = 0; 1369 vm_flags |= VM_SHARED | VM_MAYSHARE; 1370 break; 1371 case MAP_PRIVATE: 1372 /* 1373 * Set pgoff according to addr for anon_vma. 1374 */ 1375 pgoff = addr >> PAGE_SHIFT; 1376 break; 1377 default: 1378 return -EINVAL; 1379 } 1380 } 1381 1382 /* 1383 * Set 'VM_NORESERVE' if we should not account for the 1384 * memory use of this mapping. 1385 */ 1386 if (flags & MAP_NORESERVE) { 1387 /* We honor MAP_NORESERVE if allowed to overcommit */ 1388 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1389 vm_flags |= VM_NORESERVE; 1390 1391 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1392 if (file && is_file_hugepages(file)) 1393 vm_flags |= VM_NORESERVE; 1394 } 1395 1396 addr = mmap_region(file, addr, len, vm_flags, pgoff); 1397 if (!IS_ERR_VALUE(addr) && 1398 ((vm_flags & VM_LOCKED) || 1399 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1400 *populate = len; 1401 return addr; 1402 } 1403 1404 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1405 unsigned long, prot, unsigned long, flags, 1406 unsigned long, fd, unsigned long, pgoff) 1407 { 1408 struct file *file = NULL; 1409 unsigned long retval; 1410 1411 if (!(flags & MAP_ANONYMOUS)) { 1412 audit_mmap_fd(fd, flags); 1413 file = fget(fd); 1414 if (!file) 1415 return -EBADF; 1416 if (is_file_hugepages(file)) 1417 len = ALIGN(len, huge_page_size(hstate_file(file))); 1418 retval = -EINVAL; 1419 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file))) 1420 goto out_fput; 1421 } else if (flags & MAP_HUGETLB) { 1422 struct user_struct *user = NULL; 1423 struct hstate *hs; 1424 1425 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK); 1426 if (!hs) 1427 return -EINVAL; 1428 1429 len = ALIGN(len, huge_page_size(hs)); 1430 /* 1431 * VM_NORESERVE is used because the reservations will be 1432 * taken when vm_ops->mmap() is called 1433 * A dummy user value is used because we are not locking 1434 * memory so no accounting is necessary 1435 */ 1436 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1437 VM_NORESERVE, 1438 &user, HUGETLB_ANONHUGE_INODE, 1439 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1440 if (IS_ERR(file)) 1441 return PTR_ERR(file); 1442 } 1443 1444 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1445 1446 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1447 out_fput: 1448 if (file) 1449 fput(file); 1450 return retval; 1451 } 1452 1453 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1454 struct mmap_arg_struct { 1455 unsigned long addr; 1456 unsigned long len; 1457 unsigned long prot; 1458 unsigned long flags; 1459 unsigned long fd; 1460 unsigned long offset; 1461 }; 1462 1463 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1464 { 1465 struct mmap_arg_struct a; 1466 1467 if (copy_from_user(&a, arg, sizeof(a))) 1468 return -EFAULT; 1469 if (offset_in_page(a.offset)) 1470 return -EINVAL; 1471 1472 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1473 a.offset >> PAGE_SHIFT); 1474 } 1475 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1476 1477 /* 1478 * Some shared mappigns will want the pages marked read-only 1479 * to track write events. If so, we'll downgrade vm_page_prot 1480 * to the private version (using protection_map[] without the 1481 * VM_SHARED bit). 1482 */ 1483 int vma_wants_writenotify(struct vm_area_struct *vma) 1484 { 1485 vm_flags_t vm_flags = vma->vm_flags; 1486 const struct vm_operations_struct *vm_ops = vma->vm_ops; 1487 1488 /* If it was private or non-writable, the write bit is already clear */ 1489 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1490 return 0; 1491 1492 /* The backer wishes to know when pages are first written to? */ 1493 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite)) 1494 return 1; 1495 1496 /* The open routine did something to the protections that pgprot_modify 1497 * won't preserve? */ 1498 if (pgprot_val(vma->vm_page_prot) != 1499 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags))) 1500 return 0; 1501 1502 /* Do we need to track softdirty? */ 1503 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY)) 1504 return 1; 1505 1506 /* Specialty mapping? */ 1507 if (vm_flags & VM_PFNMAP) 1508 return 0; 1509 1510 /* Can the mapping track the dirty pages? */ 1511 return vma->vm_file && vma->vm_file->f_mapping && 1512 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1513 } 1514 1515 /* 1516 * We account for memory if it's a private writeable mapping, 1517 * not hugepages and VM_NORESERVE wasn't set. 1518 */ 1519 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1520 { 1521 /* 1522 * hugetlb has its own accounting separate from the core VM 1523 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1524 */ 1525 if (file && is_file_hugepages(file)) 1526 return 0; 1527 1528 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1529 } 1530 1531 unsigned long mmap_region(struct file *file, unsigned long addr, 1532 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff) 1533 { 1534 struct mm_struct *mm = current->mm; 1535 struct vm_area_struct *vma, *prev; 1536 int error; 1537 struct rb_node **rb_link, *rb_parent; 1538 unsigned long charged = 0; 1539 1540 /* Check against address space limit. */ 1541 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) { 1542 unsigned long nr_pages; 1543 1544 /* 1545 * MAP_FIXED may remove pages of mappings that intersects with 1546 * requested mapping. Account for the pages it would unmap. 1547 */ 1548 nr_pages = count_vma_pages_range(mm, addr, addr + len); 1549 1550 if (!may_expand_vm(mm, vm_flags, 1551 (len >> PAGE_SHIFT) - nr_pages)) 1552 return -ENOMEM; 1553 } 1554 1555 /* Clear old maps */ 1556 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link, 1557 &rb_parent)) { 1558 if (do_munmap(mm, addr, len)) 1559 return -ENOMEM; 1560 } 1561 1562 /* 1563 * Private writable mapping: check memory availability 1564 */ 1565 if (accountable_mapping(file, vm_flags)) { 1566 charged = len >> PAGE_SHIFT; 1567 if (security_vm_enough_memory_mm(mm, charged)) 1568 return -ENOMEM; 1569 vm_flags |= VM_ACCOUNT; 1570 } 1571 1572 /* 1573 * Can we just expand an old mapping? 1574 */ 1575 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, 1576 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX); 1577 if (vma) 1578 goto out; 1579 1580 /* 1581 * Determine the object being mapped and call the appropriate 1582 * specific mapper. the address has already been validated, but 1583 * not unmapped, but the maps are removed from the list. 1584 */ 1585 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1586 if (!vma) { 1587 error = -ENOMEM; 1588 goto unacct_error; 1589 } 1590 1591 vma->vm_mm = mm; 1592 vma->vm_start = addr; 1593 vma->vm_end = addr + len; 1594 vma->vm_flags = vm_flags; 1595 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1596 vma->vm_pgoff = pgoff; 1597 INIT_LIST_HEAD(&vma->anon_vma_chain); 1598 1599 if (file) { 1600 if (vm_flags & VM_DENYWRITE) { 1601 error = deny_write_access(file); 1602 if (error) 1603 goto free_vma; 1604 } 1605 if (vm_flags & VM_SHARED) { 1606 error = mapping_map_writable(file->f_mapping); 1607 if (error) 1608 goto allow_write_and_free_vma; 1609 } 1610 1611 /* ->mmap() can change vma->vm_file, but must guarantee that 1612 * vma_link() below can deny write-access if VM_DENYWRITE is set 1613 * and map writably if VM_SHARED is set. This usually means the 1614 * new file must not have been exposed to user-space, yet. 1615 */ 1616 vma->vm_file = get_file(file); 1617 error = file->f_op->mmap(file, vma); 1618 if (error) 1619 goto unmap_and_free_vma; 1620 1621 /* Can addr have changed?? 1622 * 1623 * Answer: Yes, several device drivers can do it in their 1624 * f_op->mmap method. -DaveM 1625 * Bug: If addr is changed, prev, rb_link, rb_parent should 1626 * be updated for vma_link() 1627 */ 1628 WARN_ON_ONCE(addr != vma->vm_start); 1629 1630 addr = vma->vm_start; 1631 vm_flags = vma->vm_flags; 1632 } else if (vm_flags & VM_SHARED) { 1633 error = shmem_zero_setup(vma); 1634 if (error) 1635 goto free_vma; 1636 } 1637 1638 vma_link(mm, vma, prev, rb_link, rb_parent); 1639 /* Once vma denies write, undo our temporary denial count */ 1640 if (file) { 1641 if (vm_flags & VM_SHARED) 1642 mapping_unmap_writable(file->f_mapping); 1643 if (vm_flags & VM_DENYWRITE) 1644 allow_write_access(file); 1645 } 1646 file = vma->vm_file; 1647 out: 1648 perf_event_mmap(vma); 1649 1650 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT); 1651 if (vm_flags & VM_LOCKED) { 1652 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) || 1653 vma == get_gate_vma(current->mm))) 1654 mm->locked_vm += (len >> PAGE_SHIFT); 1655 else 1656 vma->vm_flags &= VM_LOCKED_CLEAR_MASK; 1657 } 1658 1659 if (file) 1660 uprobe_mmap(vma); 1661 1662 /* 1663 * New (or expanded) vma always get soft dirty status. 1664 * Otherwise user-space soft-dirty page tracker won't 1665 * be able to distinguish situation when vma area unmapped, 1666 * then new mapped in-place (which must be aimed as 1667 * a completely new data area). 1668 */ 1669 vma->vm_flags |= VM_SOFTDIRTY; 1670 1671 vma_set_page_prot(vma); 1672 1673 return addr; 1674 1675 unmap_and_free_vma: 1676 vma->vm_file = NULL; 1677 fput(file); 1678 1679 /* Undo any partial mapping done by a device driver. */ 1680 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1681 charged = 0; 1682 if (vm_flags & VM_SHARED) 1683 mapping_unmap_writable(file->f_mapping); 1684 allow_write_and_free_vma: 1685 if (vm_flags & VM_DENYWRITE) 1686 allow_write_access(file); 1687 free_vma: 1688 kmem_cache_free(vm_area_cachep, vma); 1689 unacct_error: 1690 if (charged) 1691 vm_unacct_memory(charged); 1692 return error; 1693 } 1694 1695 unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1696 { 1697 /* 1698 * We implement the search by looking for an rbtree node that 1699 * immediately follows a suitable gap. That is, 1700 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length; 1701 * - gap_end = vma->vm_start >= info->low_limit + length; 1702 * - gap_end - gap_start >= length 1703 */ 1704 1705 struct mm_struct *mm = current->mm; 1706 struct vm_area_struct *vma; 1707 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1708 1709 /* Adjust search length to account for worst case alignment overhead */ 1710 length = info->length + info->align_mask; 1711 if (length < info->length) 1712 return -ENOMEM; 1713 1714 /* Adjust search limits by the desired length */ 1715 if (info->high_limit < length) 1716 return -ENOMEM; 1717 high_limit = info->high_limit - length; 1718 1719 if (info->low_limit > high_limit) 1720 return -ENOMEM; 1721 low_limit = info->low_limit + length; 1722 1723 /* Check if rbtree root looks promising */ 1724 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1725 goto check_highest; 1726 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1727 if (vma->rb_subtree_gap < length) 1728 goto check_highest; 1729 1730 while (true) { 1731 /* Visit left subtree if it looks promising */ 1732 gap_end = vma->vm_start; 1733 if (gap_end >= low_limit && vma->vm_rb.rb_left) { 1734 struct vm_area_struct *left = 1735 rb_entry(vma->vm_rb.rb_left, 1736 struct vm_area_struct, vm_rb); 1737 if (left->rb_subtree_gap >= length) { 1738 vma = left; 1739 continue; 1740 } 1741 } 1742 1743 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1744 check_current: 1745 /* Check if current node has a suitable gap */ 1746 if (gap_start > high_limit) 1747 return -ENOMEM; 1748 if (gap_end >= low_limit && gap_end - gap_start >= length) 1749 goto found; 1750 1751 /* Visit right subtree if it looks promising */ 1752 if (vma->vm_rb.rb_right) { 1753 struct vm_area_struct *right = 1754 rb_entry(vma->vm_rb.rb_right, 1755 struct vm_area_struct, vm_rb); 1756 if (right->rb_subtree_gap >= length) { 1757 vma = right; 1758 continue; 1759 } 1760 } 1761 1762 /* Go back up the rbtree to find next candidate node */ 1763 while (true) { 1764 struct rb_node *prev = &vma->vm_rb; 1765 if (!rb_parent(prev)) 1766 goto check_highest; 1767 vma = rb_entry(rb_parent(prev), 1768 struct vm_area_struct, vm_rb); 1769 if (prev == vma->vm_rb.rb_left) { 1770 gap_start = vma->vm_prev->vm_end; 1771 gap_end = vma->vm_start; 1772 goto check_current; 1773 } 1774 } 1775 } 1776 1777 check_highest: 1778 /* Check highest gap, which does not precede any rbtree node */ 1779 gap_start = mm->highest_vm_end; 1780 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */ 1781 if (gap_start > high_limit) 1782 return -ENOMEM; 1783 1784 found: 1785 /* We found a suitable gap. Clip it with the original low_limit. */ 1786 if (gap_start < info->low_limit) 1787 gap_start = info->low_limit; 1788 1789 /* Adjust gap address to the desired alignment */ 1790 gap_start += (info->align_offset - gap_start) & info->align_mask; 1791 1792 VM_BUG_ON(gap_start + info->length > info->high_limit); 1793 VM_BUG_ON(gap_start + info->length > gap_end); 1794 return gap_start; 1795 } 1796 1797 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1798 { 1799 struct mm_struct *mm = current->mm; 1800 struct vm_area_struct *vma; 1801 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1802 1803 /* Adjust search length to account for worst case alignment overhead */ 1804 length = info->length + info->align_mask; 1805 if (length < info->length) 1806 return -ENOMEM; 1807 1808 /* 1809 * Adjust search limits by the desired length. 1810 * See implementation comment at top of unmapped_area(). 1811 */ 1812 gap_end = info->high_limit; 1813 if (gap_end < length) 1814 return -ENOMEM; 1815 high_limit = gap_end - length; 1816 1817 if (info->low_limit > high_limit) 1818 return -ENOMEM; 1819 low_limit = info->low_limit + length; 1820 1821 /* Check highest gap, which does not precede any rbtree node */ 1822 gap_start = mm->highest_vm_end; 1823 if (gap_start <= high_limit) 1824 goto found_highest; 1825 1826 /* Check if rbtree root looks promising */ 1827 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1828 return -ENOMEM; 1829 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1830 if (vma->rb_subtree_gap < length) 1831 return -ENOMEM; 1832 1833 while (true) { 1834 /* Visit right subtree if it looks promising */ 1835 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0; 1836 if (gap_start <= high_limit && vma->vm_rb.rb_right) { 1837 struct vm_area_struct *right = 1838 rb_entry(vma->vm_rb.rb_right, 1839 struct vm_area_struct, vm_rb); 1840 if (right->rb_subtree_gap >= length) { 1841 vma = right; 1842 continue; 1843 } 1844 } 1845 1846 check_current: 1847 /* Check if current node has a suitable gap */ 1848 gap_end = vma->vm_start; 1849 if (gap_end < low_limit) 1850 return -ENOMEM; 1851 if (gap_start <= high_limit && gap_end - gap_start >= length) 1852 goto found; 1853 1854 /* Visit left subtree if it looks promising */ 1855 if (vma->vm_rb.rb_left) { 1856 struct vm_area_struct *left = 1857 rb_entry(vma->vm_rb.rb_left, 1858 struct vm_area_struct, vm_rb); 1859 if (left->rb_subtree_gap >= length) { 1860 vma = left; 1861 continue; 1862 } 1863 } 1864 1865 /* Go back up the rbtree to find next candidate node */ 1866 while (true) { 1867 struct rb_node *prev = &vma->vm_rb; 1868 if (!rb_parent(prev)) 1869 return -ENOMEM; 1870 vma = rb_entry(rb_parent(prev), 1871 struct vm_area_struct, vm_rb); 1872 if (prev == vma->vm_rb.rb_right) { 1873 gap_start = vma->vm_prev ? 1874 vma->vm_prev->vm_end : 0; 1875 goto check_current; 1876 } 1877 } 1878 } 1879 1880 found: 1881 /* We found a suitable gap. Clip it with the original high_limit. */ 1882 if (gap_end > info->high_limit) 1883 gap_end = info->high_limit; 1884 1885 found_highest: 1886 /* Compute highest gap address at the desired alignment */ 1887 gap_end -= info->length; 1888 gap_end -= (gap_end - info->align_offset) & info->align_mask; 1889 1890 VM_BUG_ON(gap_end < info->low_limit); 1891 VM_BUG_ON(gap_end < gap_start); 1892 return gap_end; 1893 } 1894 1895 /* Get an address range which is currently unmapped. 1896 * For shmat() with addr=0. 1897 * 1898 * Ugly calling convention alert: 1899 * Return value with the low bits set means error value, 1900 * ie 1901 * if (ret & ~PAGE_MASK) 1902 * error = ret; 1903 * 1904 * This function "knows" that -ENOMEM has the bits set. 1905 */ 1906 #ifndef HAVE_ARCH_UNMAPPED_AREA 1907 unsigned long 1908 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1909 unsigned long len, unsigned long pgoff, unsigned long flags) 1910 { 1911 struct mm_struct *mm = current->mm; 1912 struct vm_area_struct *vma; 1913 struct vm_unmapped_area_info info; 1914 1915 if (len > TASK_SIZE - mmap_min_addr) 1916 return -ENOMEM; 1917 1918 if (flags & MAP_FIXED) 1919 return addr; 1920 1921 if (addr) { 1922 addr = PAGE_ALIGN(addr); 1923 vma = find_vma(mm, addr); 1924 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1925 (!vma || addr + len <= vma->vm_start)) 1926 return addr; 1927 } 1928 1929 info.flags = 0; 1930 info.length = len; 1931 info.low_limit = mm->mmap_base; 1932 info.high_limit = TASK_SIZE; 1933 info.align_mask = 0; 1934 return vm_unmapped_area(&info); 1935 } 1936 #endif 1937 1938 /* 1939 * This mmap-allocator allocates new areas top-down from below the 1940 * stack's low limit (the base): 1941 */ 1942 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1943 unsigned long 1944 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 1945 const unsigned long len, const unsigned long pgoff, 1946 const unsigned long flags) 1947 { 1948 struct vm_area_struct *vma; 1949 struct mm_struct *mm = current->mm; 1950 unsigned long addr = addr0; 1951 struct vm_unmapped_area_info info; 1952 1953 /* requested length too big for entire address space */ 1954 if (len > TASK_SIZE - mmap_min_addr) 1955 return -ENOMEM; 1956 1957 if (flags & MAP_FIXED) 1958 return addr; 1959 1960 /* requesting a specific address */ 1961 if (addr) { 1962 addr = PAGE_ALIGN(addr); 1963 vma = find_vma(mm, addr); 1964 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1965 (!vma || addr + len <= vma->vm_start)) 1966 return addr; 1967 } 1968 1969 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 1970 info.length = len; 1971 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 1972 info.high_limit = mm->mmap_base; 1973 info.align_mask = 0; 1974 addr = vm_unmapped_area(&info); 1975 1976 /* 1977 * A failed mmap() very likely causes application failure, 1978 * so fall back to the bottom-up function here. This scenario 1979 * can happen with large stack limits and large mmap() 1980 * allocations. 1981 */ 1982 if (offset_in_page(addr)) { 1983 VM_BUG_ON(addr != -ENOMEM); 1984 info.flags = 0; 1985 info.low_limit = TASK_UNMAPPED_BASE; 1986 info.high_limit = TASK_SIZE; 1987 addr = vm_unmapped_area(&info); 1988 } 1989 1990 return addr; 1991 } 1992 #endif 1993 1994 unsigned long 1995 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1996 unsigned long pgoff, unsigned long flags) 1997 { 1998 unsigned long (*get_area)(struct file *, unsigned long, 1999 unsigned long, unsigned long, unsigned long); 2000 2001 unsigned long error = arch_mmap_check(addr, len, flags); 2002 if (error) 2003 return error; 2004 2005 /* Careful about overflows.. */ 2006 if (len > TASK_SIZE) 2007 return -ENOMEM; 2008 2009 get_area = current->mm->get_unmapped_area; 2010 if (file && file->f_op->get_unmapped_area) 2011 get_area = file->f_op->get_unmapped_area; 2012 addr = get_area(file, addr, len, pgoff, flags); 2013 if (IS_ERR_VALUE(addr)) 2014 return addr; 2015 2016 if (addr > TASK_SIZE - len) 2017 return -ENOMEM; 2018 if (offset_in_page(addr)) 2019 return -EINVAL; 2020 2021 addr = arch_rebalance_pgtables(addr, len); 2022 error = security_mmap_addr(addr); 2023 return error ? error : addr; 2024 } 2025 2026 EXPORT_SYMBOL(get_unmapped_area); 2027 2028 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 2029 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 2030 { 2031 struct rb_node *rb_node; 2032 struct vm_area_struct *vma; 2033 2034 /* Check the cache first. */ 2035 vma = vmacache_find(mm, addr); 2036 if (likely(vma)) 2037 return vma; 2038 2039 rb_node = mm->mm_rb.rb_node; 2040 2041 while (rb_node) { 2042 struct vm_area_struct *tmp; 2043 2044 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2045 2046 if (tmp->vm_end > addr) { 2047 vma = tmp; 2048 if (tmp->vm_start <= addr) 2049 break; 2050 rb_node = rb_node->rb_left; 2051 } else 2052 rb_node = rb_node->rb_right; 2053 } 2054 2055 if (vma) 2056 vmacache_update(addr, vma); 2057 return vma; 2058 } 2059 2060 EXPORT_SYMBOL(find_vma); 2061 2062 /* 2063 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 2064 */ 2065 struct vm_area_struct * 2066 find_vma_prev(struct mm_struct *mm, unsigned long addr, 2067 struct vm_area_struct **pprev) 2068 { 2069 struct vm_area_struct *vma; 2070 2071 vma = find_vma(mm, addr); 2072 if (vma) { 2073 *pprev = vma->vm_prev; 2074 } else { 2075 struct rb_node *rb_node = mm->mm_rb.rb_node; 2076 *pprev = NULL; 2077 while (rb_node) { 2078 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2079 rb_node = rb_node->rb_right; 2080 } 2081 } 2082 return vma; 2083 } 2084 2085 /* 2086 * Verify that the stack growth is acceptable and 2087 * update accounting. This is shared with both the 2088 * grow-up and grow-down cases. 2089 */ 2090 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow) 2091 { 2092 struct mm_struct *mm = vma->vm_mm; 2093 struct rlimit *rlim = current->signal->rlim; 2094 unsigned long new_start, actual_size; 2095 2096 /* address space limit tests */ 2097 if (!may_expand_vm(mm, vma->vm_flags, grow)) 2098 return -ENOMEM; 2099 2100 /* Stack limit test */ 2101 actual_size = size; 2102 if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN))) 2103 actual_size -= PAGE_SIZE; 2104 if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 2105 return -ENOMEM; 2106 2107 /* mlock limit tests */ 2108 if (vma->vm_flags & VM_LOCKED) { 2109 unsigned long locked; 2110 unsigned long limit; 2111 locked = mm->locked_vm + grow; 2112 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 2113 limit >>= PAGE_SHIFT; 2114 if (locked > limit && !capable(CAP_IPC_LOCK)) 2115 return -ENOMEM; 2116 } 2117 2118 /* Check to ensure the stack will not grow into a hugetlb-only region */ 2119 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 2120 vma->vm_end - size; 2121 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 2122 return -EFAULT; 2123 2124 /* 2125 * Overcommit.. This must be the final test, as it will 2126 * update security statistics. 2127 */ 2128 if (security_vm_enough_memory_mm(mm, grow)) 2129 return -ENOMEM; 2130 2131 return 0; 2132 } 2133 2134 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 2135 /* 2136 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 2137 * vma is the last one with address > vma->vm_end. Have to extend vma. 2138 */ 2139 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 2140 { 2141 struct mm_struct *mm = vma->vm_mm; 2142 int error; 2143 2144 if (!(vma->vm_flags & VM_GROWSUP)) 2145 return -EFAULT; 2146 2147 /* 2148 * We must make sure the anon_vma is allocated 2149 * so that the anon_vma locking is not a noop. 2150 */ 2151 if (unlikely(anon_vma_prepare(vma))) 2152 return -ENOMEM; 2153 vma_lock_anon_vma(vma); 2154 2155 /* 2156 * vma->vm_start/vm_end cannot change under us because the caller 2157 * is required to hold the mmap_sem in read mode. We need the 2158 * anon_vma lock to serialize against concurrent expand_stacks. 2159 * Also guard against wrapping around to address 0. 2160 */ 2161 if (address < PAGE_ALIGN(address+4)) 2162 address = PAGE_ALIGN(address+4); 2163 else { 2164 vma_unlock_anon_vma(vma); 2165 return -ENOMEM; 2166 } 2167 error = 0; 2168 2169 /* Somebody else might have raced and expanded it already */ 2170 if (address > vma->vm_end) { 2171 unsigned long size, grow; 2172 2173 size = address - vma->vm_start; 2174 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2175 2176 error = -ENOMEM; 2177 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2178 error = acct_stack_growth(vma, size, grow); 2179 if (!error) { 2180 /* 2181 * vma_gap_update() doesn't support concurrent 2182 * updates, but we only hold a shared mmap_sem 2183 * lock here, so we need to protect against 2184 * concurrent vma expansions. 2185 * vma_lock_anon_vma() doesn't help here, as 2186 * we don't guarantee that all growable vmas 2187 * in a mm share the same root anon vma. 2188 * So, we reuse mm->page_table_lock to guard 2189 * against concurrent vma expansions. 2190 */ 2191 spin_lock(&mm->page_table_lock); 2192 if (vma->vm_flags & VM_LOCKED) 2193 mm->locked_vm += grow; 2194 vm_stat_account(mm, vma->vm_flags, grow); 2195 anon_vma_interval_tree_pre_update_vma(vma); 2196 vma->vm_end = address; 2197 anon_vma_interval_tree_post_update_vma(vma); 2198 if (vma->vm_next) 2199 vma_gap_update(vma->vm_next); 2200 else 2201 mm->highest_vm_end = address; 2202 spin_unlock(&mm->page_table_lock); 2203 2204 perf_event_mmap(vma); 2205 } 2206 } 2207 } 2208 vma_unlock_anon_vma(vma); 2209 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2210 validate_mm(mm); 2211 return error; 2212 } 2213 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2214 2215 /* 2216 * vma is the first one with address < vma->vm_start. Have to extend vma. 2217 */ 2218 int expand_downwards(struct vm_area_struct *vma, 2219 unsigned long address) 2220 { 2221 struct mm_struct *mm = vma->vm_mm; 2222 int error; 2223 2224 /* 2225 * We must make sure the anon_vma is allocated 2226 * so that the anon_vma locking is not a noop. 2227 */ 2228 if (unlikely(anon_vma_prepare(vma))) 2229 return -ENOMEM; 2230 2231 address &= PAGE_MASK; 2232 error = security_mmap_addr(address); 2233 if (error) 2234 return error; 2235 2236 vma_lock_anon_vma(vma); 2237 2238 /* 2239 * vma->vm_start/vm_end cannot change under us because the caller 2240 * is required to hold the mmap_sem in read mode. We need the 2241 * anon_vma lock to serialize against concurrent expand_stacks. 2242 */ 2243 2244 /* Somebody else might have raced and expanded it already */ 2245 if (address < vma->vm_start) { 2246 unsigned long size, grow; 2247 2248 size = vma->vm_end - address; 2249 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2250 2251 error = -ENOMEM; 2252 if (grow <= vma->vm_pgoff) { 2253 error = acct_stack_growth(vma, size, grow); 2254 if (!error) { 2255 /* 2256 * vma_gap_update() doesn't support concurrent 2257 * updates, but we only hold a shared mmap_sem 2258 * lock here, so we need to protect against 2259 * concurrent vma expansions. 2260 * vma_lock_anon_vma() doesn't help here, as 2261 * we don't guarantee that all growable vmas 2262 * in a mm share the same root anon vma. 2263 * So, we reuse mm->page_table_lock to guard 2264 * against concurrent vma expansions. 2265 */ 2266 spin_lock(&mm->page_table_lock); 2267 if (vma->vm_flags & VM_LOCKED) 2268 mm->locked_vm += grow; 2269 vm_stat_account(mm, vma->vm_flags, grow); 2270 anon_vma_interval_tree_pre_update_vma(vma); 2271 vma->vm_start = address; 2272 vma->vm_pgoff -= grow; 2273 anon_vma_interval_tree_post_update_vma(vma); 2274 vma_gap_update(vma); 2275 spin_unlock(&mm->page_table_lock); 2276 2277 perf_event_mmap(vma); 2278 } 2279 } 2280 } 2281 vma_unlock_anon_vma(vma); 2282 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2283 validate_mm(mm); 2284 return error; 2285 } 2286 2287 /* 2288 * Note how expand_stack() refuses to expand the stack all the way to 2289 * abut the next virtual mapping, *unless* that mapping itself is also 2290 * a stack mapping. We want to leave room for a guard page, after all 2291 * (the guard page itself is not added here, that is done by the 2292 * actual page faulting logic) 2293 * 2294 * This matches the behavior of the guard page logic (see mm/memory.c: 2295 * check_stack_guard_page()), which only allows the guard page to be 2296 * removed under these circumstances. 2297 */ 2298 #ifdef CONFIG_STACK_GROWSUP 2299 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2300 { 2301 struct vm_area_struct *next; 2302 2303 address &= PAGE_MASK; 2304 next = vma->vm_next; 2305 if (next && next->vm_start == address + PAGE_SIZE) { 2306 if (!(next->vm_flags & VM_GROWSUP)) 2307 return -ENOMEM; 2308 } 2309 return expand_upwards(vma, address); 2310 } 2311 2312 struct vm_area_struct * 2313 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2314 { 2315 struct vm_area_struct *vma, *prev; 2316 2317 addr &= PAGE_MASK; 2318 vma = find_vma_prev(mm, addr, &prev); 2319 if (vma && (vma->vm_start <= addr)) 2320 return vma; 2321 if (!prev || expand_stack(prev, addr)) 2322 return NULL; 2323 if (prev->vm_flags & VM_LOCKED) 2324 populate_vma_page_range(prev, addr, prev->vm_end, NULL); 2325 return prev; 2326 } 2327 #else 2328 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2329 { 2330 struct vm_area_struct *prev; 2331 2332 address &= PAGE_MASK; 2333 prev = vma->vm_prev; 2334 if (prev && prev->vm_end == address) { 2335 if (!(prev->vm_flags & VM_GROWSDOWN)) 2336 return -ENOMEM; 2337 } 2338 return expand_downwards(vma, address); 2339 } 2340 2341 struct vm_area_struct * 2342 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2343 { 2344 struct vm_area_struct *vma; 2345 unsigned long start; 2346 2347 addr &= PAGE_MASK; 2348 vma = find_vma(mm, addr); 2349 if (!vma) 2350 return NULL; 2351 if (vma->vm_start <= addr) 2352 return vma; 2353 if (!(vma->vm_flags & VM_GROWSDOWN)) 2354 return NULL; 2355 start = vma->vm_start; 2356 if (expand_stack(vma, addr)) 2357 return NULL; 2358 if (vma->vm_flags & VM_LOCKED) 2359 populate_vma_page_range(vma, addr, start, NULL); 2360 return vma; 2361 } 2362 #endif 2363 2364 EXPORT_SYMBOL_GPL(find_extend_vma); 2365 2366 /* 2367 * Ok - we have the memory areas we should free on the vma list, 2368 * so release them, and do the vma updates. 2369 * 2370 * Called with the mm semaphore held. 2371 */ 2372 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 2373 { 2374 unsigned long nr_accounted = 0; 2375 2376 /* Update high watermark before we lower total_vm */ 2377 update_hiwater_vm(mm); 2378 do { 2379 long nrpages = vma_pages(vma); 2380 2381 if (vma->vm_flags & VM_ACCOUNT) 2382 nr_accounted += nrpages; 2383 vm_stat_account(mm, vma->vm_flags, -nrpages); 2384 vma = remove_vma(vma); 2385 } while (vma); 2386 vm_unacct_memory(nr_accounted); 2387 validate_mm(mm); 2388 } 2389 2390 /* 2391 * Get rid of page table information in the indicated region. 2392 * 2393 * Called with the mm semaphore held. 2394 */ 2395 static void unmap_region(struct mm_struct *mm, 2396 struct vm_area_struct *vma, struct vm_area_struct *prev, 2397 unsigned long start, unsigned long end) 2398 { 2399 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap; 2400 struct mmu_gather tlb; 2401 2402 lru_add_drain(); 2403 tlb_gather_mmu(&tlb, mm, start, end); 2404 update_hiwater_rss(mm); 2405 unmap_vmas(&tlb, vma, start, end); 2406 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2407 next ? next->vm_start : USER_PGTABLES_CEILING); 2408 tlb_finish_mmu(&tlb, start, end); 2409 } 2410 2411 /* 2412 * Create a list of vma's touched by the unmap, removing them from the mm's 2413 * vma list as we go.. 2414 */ 2415 static void 2416 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 2417 struct vm_area_struct *prev, unsigned long end) 2418 { 2419 struct vm_area_struct **insertion_point; 2420 struct vm_area_struct *tail_vma = NULL; 2421 2422 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 2423 vma->vm_prev = NULL; 2424 do { 2425 vma_rb_erase(vma, &mm->mm_rb); 2426 mm->map_count--; 2427 tail_vma = vma; 2428 vma = vma->vm_next; 2429 } while (vma && vma->vm_start < end); 2430 *insertion_point = vma; 2431 if (vma) { 2432 vma->vm_prev = prev; 2433 vma_gap_update(vma); 2434 } else 2435 mm->highest_vm_end = prev ? prev->vm_end : 0; 2436 tail_vma->vm_next = NULL; 2437 2438 /* Kill the cache */ 2439 vmacache_invalidate(mm); 2440 } 2441 2442 /* 2443 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the 2444 * munmap path where it doesn't make sense to fail. 2445 */ 2446 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2447 unsigned long addr, int new_below) 2448 { 2449 struct vm_area_struct *new; 2450 int err; 2451 2452 if (is_vm_hugetlb_page(vma) && (addr & 2453 ~(huge_page_mask(hstate_vma(vma))))) 2454 return -EINVAL; 2455 2456 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2457 if (!new) 2458 return -ENOMEM; 2459 2460 /* most fields are the same, copy all, and then fixup */ 2461 *new = *vma; 2462 2463 INIT_LIST_HEAD(&new->anon_vma_chain); 2464 2465 if (new_below) 2466 new->vm_end = addr; 2467 else { 2468 new->vm_start = addr; 2469 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2470 } 2471 2472 err = vma_dup_policy(vma, new); 2473 if (err) 2474 goto out_free_vma; 2475 2476 err = anon_vma_clone(new, vma); 2477 if (err) 2478 goto out_free_mpol; 2479 2480 if (new->vm_file) 2481 get_file(new->vm_file); 2482 2483 if (new->vm_ops && new->vm_ops->open) 2484 new->vm_ops->open(new); 2485 2486 if (new_below) 2487 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 2488 ((addr - new->vm_start) >> PAGE_SHIFT), new); 2489 else 2490 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 2491 2492 /* Success. */ 2493 if (!err) 2494 return 0; 2495 2496 /* Clean everything up if vma_adjust failed. */ 2497 if (new->vm_ops && new->vm_ops->close) 2498 new->vm_ops->close(new); 2499 if (new->vm_file) 2500 fput(new->vm_file); 2501 unlink_anon_vmas(new); 2502 out_free_mpol: 2503 mpol_put(vma_policy(new)); 2504 out_free_vma: 2505 kmem_cache_free(vm_area_cachep, new); 2506 return err; 2507 } 2508 2509 /* 2510 * Split a vma into two pieces at address 'addr', a new vma is allocated 2511 * either for the first part or the tail. 2512 */ 2513 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2514 unsigned long addr, int new_below) 2515 { 2516 if (mm->map_count >= sysctl_max_map_count) 2517 return -ENOMEM; 2518 2519 return __split_vma(mm, vma, addr, new_below); 2520 } 2521 2522 /* Munmap is split into 2 main parts -- this part which finds 2523 * what needs doing, and the areas themselves, which do the 2524 * work. This now handles partial unmappings. 2525 * Jeremy Fitzhardinge <jeremy@goop.org> 2526 */ 2527 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len) 2528 { 2529 unsigned long end; 2530 struct vm_area_struct *vma, *prev, *last; 2531 2532 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start) 2533 return -EINVAL; 2534 2535 len = PAGE_ALIGN(len); 2536 if (len == 0) 2537 return -EINVAL; 2538 2539 /* Find the first overlapping VMA */ 2540 vma = find_vma(mm, start); 2541 if (!vma) 2542 return 0; 2543 prev = vma->vm_prev; 2544 /* we have start < vma->vm_end */ 2545 2546 /* if it doesn't overlap, we have nothing.. */ 2547 end = start + len; 2548 if (vma->vm_start >= end) 2549 return 0; 2550 2551 /* 2552 * If we need to split any vma, do it now to save pain later. 2553 * 2554 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2555 * unmapped vm_area_struct will remain in use: so lower split_vma 2556 * places tmp vma above, and higher split_vma places tmp vma below. 2557 */ 2558 if (start > vma->vm_start) { 2559 int error; 2560 2561 /* 2562 * Make sure that map_count on return from munmap() will 2563 * not exceed its limit; but let map_count go just above 2564 * its limit temporarily, to help free resources as expected. 2565 */ 2566 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2567 return -ENOMEM; 2568 2569 error = __split_vma(mm, vma, start, 0); 2570 if (error) 2571 return error; 2572 prev = vma; 2573 } 2574 2575 /* Does it split the last one? */ 2576 last = find_vma(mm, end); 2577 if (last && end > last->vm_start) { 2578 int error = __split_vma(mm, last, end, 1); 2579 if (error) 2580 return error; 2581 } 2582 vma = prev ? prev->vm_next : mm->mmap; 2583 2584 /* 2585 * unlock any mlock()ed ranges before detaching vmas 2586 */ 2587 if (mm->locked_vm) { 2588 struct vm_area_struct *tmp = vma; 2589 while (tmp && tmp->vm_start < end) { 2590 if (tmp->vm_flags & VM_LOCKED) { 2591 mm->locked_vm -= vma_pages(tmp); 2592 munlock_vma_pages_all(tmp); 2593 } 2594 tmp = tmp->vm_next; 2595 } 2596 } 2597 2598 /* 2599 * Remove the vma's, and unmap the actual pages 2600 */ 2601 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2602 unmap_region(mm, vma, prev, start, end); 2603 2604 arch_unmap(mm, vma, start, end); 2605 2606 /* Fix up all other VM information */ 2607 remove_vma_list(mm, vma); 2608 2609 return 0; 2610 } 2611 2612 int vm_munmap(unsigned long start, size_t len) 2613 { 2614 int ret; 2615 struct mm_struct *mm = current->mm; 2616 2617 down_write(&mm->mmap_sem); 2618 ret = do_munmap(mm, start, len); 2619 up_write(&mm->mmap_sem); 2620 return ret; 2621 } 2622 EXPORT_SYMBOL(vm_munmap); 2623 2624 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2625 { 2626 profile_munmap(addr); 2627 return vm_munmap(addr, len); 2628 } 2629 2630 2631 /* 2632 * Emulation of deprecated remap_file_pages() syscall. 2633 */ 2634 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size, 2635 unsigned long, prot, unsigned long, pgoff, unsigned long, flags) 2636 { 2637 2638 struct mm_struct *mm = current->mm; 2639 struct vm_area_struct *vma; 2640 unsigned long populate = 0; 2641 unsigned long ret = -EINVAL; 2642 struct file *file; 2643 2644 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. " 2645 "See Documentation/vm/remap_file_pages.txt.\n", 2646 current->comm, current->pid); 2647 2648 if (prot) 2649 return ret; 2650 start = start & PAGE_MASK; 2651 size = size & PAGE_MASK; 2652 2653 if (start + size <= start) 2654 return ret; 2655 2656 /* Does pgoff wrap? */ 2657 if (pgoff + (size >> PAGE_SHIFT) < pgoff) 2658 return ret; 2659 2660 down_write(&mm->mmap_sem); 2661 vma = find_vma(mm, start); 2662 2663 if (!vma || !(vma->vm_flags & VM_SHARED)) 2664 goto out; 2665 2666 if (start < vma->vm_start || start + size > vma->vm_end) 2667 goto out; 2668 2669 if (pgoff == linear_page_index(vma, start)) { 2670 ret = 0; 2671 goto out; 2672 } 2673 2674 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0; 2675 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0; 2676 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0; 2677 2678 flags &= MAP_NONBLOCK; 2679 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE; 2680 if (vma->vm_flags & VM_LOCKED) { 2681 flags |= MAP_LOCKED; 2682 /* drop PG_Mlocked flag for over-mapped range */ 2683 munlock_vma_pages_range(vma, start, start + size); 2684 } 2685 2686 file = get_file(vma->vm_file); 2687 ret = do_mmap_pgoff(vma->vm_file, start, size, 2688 prot, flags, pgoff, &populate); 2689 fput(file); 2690 out: 2691 up_write(&mm->mmap_sem); 2692 if (populate) 2693 mm_populate(ret, populate); 2694 if (!IS_ERR_VALUE(ret)) 2695 ret = 0; 2696 return ret; 2697 } 2698 2699 static inline void verify_mm_writelocked(struct mm_struct *mm) 2700 { 2701 #ifdef CONFIG_DEBUG_VM 2702 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2703 WARN_ON(1); 2704 up_read(&mm->mmap_sem); 2705 } 2706 #endif 2707 } 2708 2709 /* 2710 * this is really a simplified "do_mmap". it only handles 2711 * anonymous maps. eventually we may be able to do some 2712 * brk-specific accounting here. 2713 */ 2714 static unsigned long do_brk(unsigned long addr, unsigned long len) 2715 { 2716 struct mm_struct *mm = current->mm; 2717 struct vm_area_struct *vma, *prev; 2718 unsigned long flags; 2719 struct rb_node **rb_link, *rb_parent; 2720 pgoff_t pgoff = addr >> PAGE_SHIFT; 2721 int error; 2722 2723 len = PAGE_ALIGN(len); 2724 if (!len) 2725 return addr; 2726 2727 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2728 2729 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2730 if (offset_in_page(error)) 2731 return error; 2732 2733 error = mlock_future_check(mm, mm->def_flags, len); 2734 if (error) 2735 return error; 2736 2737 /* 2738 * mm->mmap_sem is required to protect against another thread 2739 * changing the mappings in case we sleep. 2740 */ 2741 verify_mm_writelocked(mm); 2742 2743 /* 2744 * Clear old maps. this also does some error checking for us 2745 */ 2746 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link, 2747 &rb_parent)) { 2748 if (do_munmap(mm, addr, len)) 2749 return -ENOMEM; 2750 } 2751 2752 /* Check against address space limits *after* clearing old maps... */ 2753 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT)) 2754 return -ENOMEM; 2755 2756 if (mm->map_count > sysctl_max_map_count) 2757 return -ENOMEM; 2758 2759 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 2760 return -ENOMEM; 2761 2762 /* Can we just expand an old private anonymous mapping? */ 2763 vma = vma_merge(mm, prev, addr, addr + len, flags, 2764 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX); 2765 if (vma) 2766 goto out; 2767 2768 /* 2769 * create a vma struct for an anonymous mapping 2770 */ 2771 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2772 if (!vma) { 2773 vm_unacct_memory(len >> PAGE_SHIFT); 2774 return -ENOMEM; 2775 } 2776 2777 INIT_LIST_HEAD(&vma->anon_vma_chain); 2778 vma->vm_mm = mm; 2779 vma->vm_start = addr; 2780 vma->vm_end = addr + len; 2781 vma->vm_pgoff = pgoff; 2782 vma->vm_flags = flags; 2783 vma->vm_page_prot = vm_get_page_prot(flags); 2784 vma_link(mm, vma, prev, rb_link, rb_parent); 2785 out: 2786 perf_event_mmap(vma); 2787 mm->total_vm += len >> PAGE_SHIFT; 2788 mm->data_vm += len >> PAGE_SHIFT; 2789 if (flags & VM_LOCKED) 2790 mm->locked_vm += (len >> PAGE_SHIFT); 2791 vma->vm_flags |= VM_SOFTDIRTY; 2792 return addr; 2793 } 2794 2795 unsigned long vm_brk(unsigned long addr, unsigned long len) 2796 { 2797 struct mm_struct *mm = current->mm; 2798 unsigned long ret; 2799 bool populate; 2800 2801 down_write(&mm->mmap_sem); 2802 ret = do_brk(addr, len); 2803 populate = ((mm->def_flags & VM_LOCKED) != 0); 2804 up_write(&mm->mmap_sem); 2805 if (populate) 2806 mm_populate(addr, len); 2807 return ret; 2808 } 2809 EXPORT_SYMBOL(vm_brk); 2810 2811 /* Release all mmaps. */ 2812 void exit_mmap(struct mm_struct *mm) 2813 { 2814 struct mmu_gather tlb; 2815 struct vm_area_struct *vma; 2816 unsigned long nr_accounted = 0; 2817 2818 /* mm's last user has gone, and its about to be pulled down */ 2819 mmu_notifier_release(mm); 2820 2821 if (mm->locked_vm) { 2822 vma = mm->mmap; 2823 while (vma) { 2824 if (vma->vm_flags & VM_LOCKED) 2825 munlock_vma_pages_all(vma); 2826 vma = vma->vm_next; 2827 } 2828 } 2829 2830 arch_exit_mmap(mm); 2831 2832 vma = mm->mmap; 2833 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2834 return; 2835 2836 lru_add_drain(); 2837 flush_cache_mm(mm); 2838 tlb_gather_mmu(&tlb, mm, 0, -1); 2839 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2840 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2841 unmap_vmas(&tlb, vma, 0, -1); 2842 2843 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING); 2844 tlb_finish_mmu(&tlb, 0, -1); 2845 2846 /* 2847 * Walk the list again, actually closing and freeing it, 2848 * with preemption enabled, without holding any MM locks. 2849 */ 2850 while (vma) { 2851 if (vma->vm_flags & VM_ACCOUNT) 2852 nr_accounted += vma_pages(vma); 2853 vma = remove_vma(vma); 2854 } 2855 vm_unacct_memory(nr_accounted); 2856 } 2857 2858 /* Insert vm structure into process list sorted by address 2859 * and into the inode's i_mmap tree. If vm_file is non-NULL 2860 * then i_mmap_rwsem is taken here. 2861 */ 2862 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 2863 { 2864 struct vm_area_struct *prev; 2865 struct rb_node **rb_link, *rb_parent; 2866 2867 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 2868 &prev, &rb_link, &rb_parent)) 2869 return -ENOMEM; 2870 if ((vma->vm_flags & VM_ACCOUNT) && 2871 security_vm_enough_memory_mm(mm, vma_pages(vma))) 2872 return -ENOMEM; 2873 2874 /* 2875 * The vm_pgoff of a purely anonymous vma should be irrelevant 2876 * until its first write fault, when page's anon_vma and index 2877 * are set. But now set the vm_pgoff it will almost certainly 2878 * end up with (unless mremap moves it elsewhere before that 2879 * first wfault), so /proc/pid/maps tells a consistent story. 2880 * 2881 * By setting it to reflect the virtual start address of the 2882 * vma, merges and splits can happen in a seamless way, just 2883 * using the existing file pgoff checks and manipulations. 2884 * Similarly in do_mmap_pgoff and in do_brk. 2885 */ 2886 if (vma_is_anonymous(vma)) { 2887 BUG_ON(vma->anon_vma); 2888 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 2889 } 2890 2891 vma_link(mm, vma, prev, rb_link, rb_parent); 2892 return 0; 2893 } 2894 2895 /* 2896 * Copy the vma structure to a new location in the same mm, 2897 * prior to moving page table entries, to effect an mremap move. 2898 */ 2899 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 2900 unsigned long addr, unsigned long len, pgoff_t pgoff, 2901 bool *need_rmap_locks) 2902 { 2903 struct vm_area_struct *vma = *vmap; 2904 unsigned long vma_start = vma->vm_start; 2905 struct mm_struct *mm = vma->vm_mm; 2906 struct vm_area_struct *new_vma, *prev; 2907 struct rb_node **rb_link, *rb_parent; 2908 bool faulted_in_anon_vma = true; 2909 2910 /* 2911 * If anonymous vma has not yet been faulted, update new pgoff 2912 * to match new location, to increase its chance of merging. 2913 */ 2914 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) { 2915 pgoff = addr >> PAGE_SHIFT; 2916 faulted_in_anon_vma = false; 2917 } 2918 2919 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) 2920 return NULL; /* should never get here */ 2921 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 2922 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 2923 vma->vm_userfaultfd_ctx); 2924 if (new_vma) { 2925 /* 2926 * Source vma may have been merged into new_vma 2927 */ 2928 if (unlikely(vma_start >= new_vma->vm_start && 2929 vma_start < new_vma->vm_end)) { 2930 /* 2931 * The only way we can get a vma_merge with 2932 * self during an mremap is if the vma hasn't 2933 * been faulted in yet and we were allowed to 2934 * reset the dst vma->vm_pgoff to the 2935 * destination address of the mremap to allow 2936 * the merge to happen. mremap must change the 2937 * vm_pgoff linearity between src and dst vmas 2938 * (in turn preventing a vma_merge) to be 2939 * safe. It is only safe to keep the vm_pgoff 2940 * linear if there are no pages mapped yet. 2941 */ 2942 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 2943 *vmap = vma = new_vma; 2944 } 2945 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 2946 } else { 2947 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2948 if (!new_vma) 2949 goto out; 2950 *new_vma = *vma; 2951 new_vma->vm_start = addr; 2952 new_vma->vm_end = addr + len; 2953 new_vma->vm_pgoff = pgoff; 2954 if (vma_dup_policy(vma, new_vma)) 2955 goto out_free_vma; 2956 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 2957 if (anon_vma_clone(new_vma, vma)) 2958 goto out_free_mempol; 2959 if (new_vma->vm_file) 2960 get_file(new_vma->vm_file); 2961 if (new_vma->vm_ops && new_vma->vm_ops->open) 2962 new_vma->vm_ops->open(new_vma); 2963 vma_link(mm, new_vma, prev, rb_link, rb_parent); 2964 *need_rmap_locks = false; 2965 } 2966 return new_vma; 2967 2968 out_free_mempol: 2969 mpol_put(vma_policy(new_vma)); 2970 out_free_vma: 2971 kmem_cache_free(vm_area_cachep, new_vma); 2972 out: 2973 return NULL; 2974 } 2975 2976 /* 2977 * Return true if the calling process may expand its vm space by the passed 2978 * number of pages 2979 */ 2980 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages) 2981 { 2982 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT) 2983 return false; 2984 2985 if ((flags & (VM_WRITE | VM_SHARED | (VM_STACK_FLAGS & 2986 (VM_GROWSUP | VM_GROWSDOWN)))) == VM_WRITE) 2987 return mm->data_vm + npages <= rlimit(RLIMIT_DATA); 2988 2989 return true; 2990 } 2991 2992 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages) 2993 { 2994 mm->total_vm += npages; 2995 2996 if ((flags & (VM_EXEC | VM_WRITE)) == VM_EXEC) 2997 mm->exec_vm += npages; 2998 else if (flags & (VM_STACK_FLAGS & (VM_GROWSUP | VM_GROWSDOWN))) 2999 mm->stack_vm += npages; 3000 else if ((flags & (VM_WRITE | VM_SHARED)) == VM_WRITE) 3001 mm->data_vm += npages; 3002 } 3003 3004 static int special_mapping_fault(struct vm_area_struct *vma, 3005 struct vm_fault *vmf); 3006 3007 /* 3008 * Having a close hook prevents vma merging regardless of flags. 3009 */ 3010 static void special_mapping_close(struct vm_area_struct *vma) 3011 { 3012 } 3013 3014 static const char *special_mapping_name(struct vm_area_struct *vma) 3015 { 3016 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 3017 } 3018 3019 static const struct vm_operations_struct special_mapping_vmops = { 3020 .close = special_mapping_close, 3021 .fault = special_mapping_fault, 3022 .name = special_mapping_name, 3023 }; 3024 3025 static const struct vm_operations_struct legacy_special_mapping_vmops = { 3026 .close = special_mapping_close, 3027 .fault = special_mapping_fault, 3028 }; 3029 3030 static int special_mapping_fault(struct vm_area_struct *vma, 3031 struct vm_fault *vmf) 3032 { 3033 pgoff_t pgoff; 3034 struct page **pages; 3035 3036 if (vma->vm_ops == &legacy_special_mapping_vmops) 3037 pages = vma->vm_private_data; 3038 else 3039 pages = ((struct vm_special_mapping *)vma->vm_private_data)-> 3040 pages; 3041 3042 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages) 3043 pgoff--; 3044 3045 if (*pages) { 3046 struct page *page = *pages; 3047 get_page(page); 3048 vmf->page = page; 3049 return 0; 3050 } 3051 3052 return VM_FAULT_SIGBUS; 3053 } 3054 3055 static struct vm_area_struct *__install_special_mapping( 3056 struct mm_struct *mm, 3057 unsigned long addr, unsigned long len, 3058 unsigned long vm_flags, void *priv, 3059 const struct vm_operations_struct *ops) 3060 { 3061 int ret; 3062 struct vm_area_struct *vma; 3063 3064 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 3065 if (unlikely(vma == NULL)) 3066 return ERR_PTR(-ENOMEM); 3067 3068 INIT_LIST_HEAD(&vma->anon_vma_chain); 3069 vma->vm_mm = mm; 3070 vma->vm_start = addr; 3071 vma->vm_end = addr + len; 3072 3073 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY; 3074 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3075 3076 vma->vm_ops = ops; 3077 vma->vm_private_data = priv; 3078 3079 ret = insert_vm_struct(mm, vma); 3080 if (ret) 3081 goto out; 3082 3083 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT); 3084 3085 perf_event_mmap(vma); 3086 3087 return vma; 3088 3089 out: 3090 kmem_cache_free(vm_area_cachep, vma); 3091 return ERR_PTR(ret); 3092 } 3093 3094 /* 3095 * Called with mm->mmap_sem held for writing. 3096 * Insert a new vma covering the given region, with the given flags. 3097 * Its pages are supplied by the given array of struct page *. 3098 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3099 * The region past the last page supplied will always produce SIGBUS. 3100 * The array pointer and the pages it points to are assumed to stay alive 3101 * for as long as this mapping might exist. 3102 */ 3103 struct vm_area_struct *_install_special_mapping( 3104 struct mm_struct *mm, 3105 unsigned long addr, unsigned long len, 3106 unsigned long vm_flags, const struct vm_special_mapping *spec) 3107 { 3108 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec, 3109 &special_mapping_vmops); 3110 } 3111 3112 int install_special_mapping(struct mm_struct *mm, 3113 unsigned long addr, unsigned long len, 3114 unsigned long vm_flags, struct page **pages) 3115 { 3116 struct vm_area_struct *vma = __install_special_mapping( 3117 mm, addr, len, vm_flags, (void *)pages, 3118 &legacy_special_mapping_vmops); 3119 3120 return PTR_ERR_OR_ZERO(vma); 3121 } 3122 3123 static DEFINE_MUTEX(mm_all_locks_mutex); 3124 3125 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3126 { 3127 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3128 /* 3129 * The LSB of head.next can't change from under us 3130 * because we hold the mm_all_locks_mutex. 3131 */ 3132 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem); 3133 /* 3134 * We can safely modify head.next after taking the 3135 * anon_vma->root->rwsem. If some other vma in this mm shares 3136 * the same anon_vma we won't take it again. 3137 * 3138 * No need of atomic instructions here, head.next 3139 * can't change from under us thanks to the 3140 * anon_vma->root->rwsem. 3141 */ 3142 if (__test_and_set_bit(0, (unsigned long *) 3143 &anon_vma->root->rb_root.rb_node)) 3144 BUG(); 3145 } 3146 } 3147 3148 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3149 { 3150 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3151 /* 3152 * AS_MM_ALL_LOCKS can't change from under us because 3153 * we hold the mm_all_locks_mutex. 3154 * 3155 * Operations on ->flags have to be atomic because 3156 * even if AS_MM_ALL_LOCKS is stable thanks to the 3157 * mm_all_locks_mutex, there may be other cpus 3158 * changing other bitflags in parallel to us. 3159 */ 3160 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3161 BUG(); 3162 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem); 3163 } 3164 } 3165 3166 /* 3167 * This operation locks against the VM for all pte/vma/mm related 3168 * operations that could ever happen on a certain mm. This includes 3169 * vmtruncate, try_to_unmap, and all page faults. 3170 * 3171 * The caller must take the mmap_sem in write mode before calling 3172 * mm_take_all_locks(). The caller isn't allowed to release the 3173 * mmap_sem until mm_drop_all_locks() returns. 3174 * 3175 * mmap_sem in write mode is required in order to block all operations 3176 * that could modify pagetables and free pages without need of 3177 * altering the vma layout. It's also needed in write mode to avoid new 3178 * anon_vmas to be associated with existing vmas. 3179 * 3180 * A single task can't take more than one mm_take_all_locks() in a row 3181 * or it would deadlock. 3182 * 3183 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3184 * mapping->flags avoid to take the same lock twice, if more than one 3185 * vma in this mm is backed by the same anon_vma or address_space. 3186 * 3187 * We take locks in following order, accordingly to comment at beginning 3188 * of mm/rmap.c: 3189 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for 3190 * hugetlb mapping); 3191 * - all i_mmap_rwsem locks; 3192 * - all anon_vma->rwseml 3193 * 3194 * We can take all locks within these types randomly because the VM code 3195 * doesn't nest them and we protected from parallel mm_take_all_locks() by 3196 * mm_all_locks_mutex. 3197 * 3198 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3199 * that may have to take thousand of locks. 3200 * 3201 * mm_take_all_locks() can fail if it's interrupted by signals. 3202 */ 3203 int mm_take_all_locks(struct mm_struct *mm) 3204 { 3205 struct vm_area_struct *vma; 3206 struct anon_vma_chain *avc; 3207 3208 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3209 3210 mutex_lock(&mm_all_locks_mutex); 3211 3212 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3213 if (signal_pending(current)) 3214 goto out_unlock; 3215 if (vma->vm_file && vma->vm_file->f_mapping && 3216 is_vm_hugetlb_page(vma)) 3217 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3218 } 3219 3220 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3221 if (signal_pending(current)) 3222 goto out_unlock; 3223 if (vma->vm_file && vma->vm_file->f_mapping && 3224 !is_vm_hugetlb_page(vma)) 3225 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3226 } 3227 3228 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3229 if (signal_pending(current)) 3230 goto out_unlock; 3231 if (vma->anon_vma) 3232 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3233 vm_lock_anon_vma(mm, avc->anon_vma); 3234 } 3235 3236 return 0; 3237 3238 out_unlock: 3239 mm_drop_all_locks(mm); 3240 return -EINTR; 3241 } 3242 3243 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3244 { 3245 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3246 /* 3247 * The LSB of head.next can't change to 0 from under 3248 * us because we hold the mm_all_locks_mutex. 3249 * 3250 * We must however clear the bitflag before unlocking 3251 * the vma so the users using the anon_vma->rb_root will 3252 * never see our bitflag. 3253 * 3254 * No need of atomic instructions here, head.next 3255 * can't change from under us until we release the 3256 * anon_vma->root->rwsem. 3257 */ 3258 if (!__test_and_clear_bit(0, (unsigned long *) 3259 &anon_vma->root->rb_root.rb_node)) 3260 BUG(); 3261 anon_vma_unlock_write(anon_vma); 3262 } 3263 } 3264 3265 static void vm_unlock_mapping(struct address_space *mapping) 3266 { 3267 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3268 /* 3269 * AS_MM_ALL_LOCKS can't change to 0 from under us 3270 * because we hold the mm_all_locks_mutex. 3271 */ 3272 i_mmap_unlock_write(mapping); 3273 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3274 &mapping->flags)) 3275 BUG(); 3276 } 3277 } 3278 3279 /* 3280 * The mmap_sem cannot be released by the caller until 3281 * mm_drop_all_locks() returns. 3282 */ 3283 void mm_drop_all_locks(struct mm_struct *mm) 3284 { 3285 struct vm_area_struct *vma; 3286 struct anon_vma_chain *avc; 3287 3288 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3289 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3290 3291 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3292 if (vma->anon_vma) 3293 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3294 vm_unlock_anon_vma(avc->anon_vma); 3295 if (vma->vm_file && vma->vm_file->f_mapping) 3296 vm_unlock_mapping(vma->vm_file->f_mapping); 3297 } 3298 3299 mutex_unlock(&mm_all_locks_mutex); 3300 } 3301 3302 /* 3303 * initialise the VMA slab 3304 */ 3305 void __init mmap_init(void) 3306 { 3307 int ret; 3308 3309 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3310 VM_BUG_ON(ret); 3311 } 3312 3313 /* 3314 * Initialise sysctl_user_reserve_kbytes. 3315 * 3316 * This is intended to prevent a user from starting a single memory hogging 3317 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3318 * mode. 3319 * 3320 * The default value is min(3% of free memory, 128MB) 3321 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3322 */ 3323 static int init_user_reserve(void) 3324 { 3325 unsigned long free_kbytes; 3326 3327 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3328 3329 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3330 return 0; 3331 } 3332 subsys_initcall(init_user_reserve); 3333 3334 /* 3335 * Initialise sysctl_admin_reserve_kbytes. 3336 * 3337 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3338 * to log in and kill a memory hogging process. 3339 * 3340 * Systems with more than 256MB will reserve 8MB, enough to recover 3341 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3342 * only reserve 3% of free pages by default. 3343 */ 3344 static int init_admin_reserve(void) 3345 { 3346 unsigned long free_kbytes; 3347 3348 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3349 3350 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3351 return 0; 3352 } 3353 subsys_initcall(init_admin_reserve); 3354 3355 /* 3356 * Reinititalise user and admin reserves if memory is added or removed. 3357 * 3358 * The default user reserve max is 128MB, and the default max for the 3359 * admin reserve is 8MB. These are usually, but not always, enough to 3360 * enable recovery from a memory hogging process using login/sshd, a shell, 3361 * and tools like top. It may make sense to increase or even disable the 3362 * reserve depending on the existence of swap or variations in the recovery 3363 * tools. So, the admin may have changed them. 3364 * 3365 * If memory is added and the reserves have been eliminated or increased above 3366 * the default max, then we'll trust the admin. 3367 * 3368 * If memory is removed and there isn't enough free memory, then we 3369 * need to reset the reserves. 3370 * 3371 * Otherwise keep the reserve set by the admin. 3372 */ 3373 static int reserve_mem_notifier(struct notifier_block *nb, 3374 unsigned long action, void *data) 3375 { 3376 unsigned long tmp, free_kbytes; 3377 3378 switch (action) { 3379 case MEM_ONLINE: 3380 /* Default max is 128MB. Leave alone if modified by operator. */ 3381 tmp = sysctl_user_reserve_kbytes; 3382 if (0 < tmp && tmp < (1UL << 17)) 3383 init_user_reserve(); 3384 3385 /* Default max is 8MB. Leave alone if modified by operator. */ 3386 tmp = sysctl_admin_reserve_kbytes; 3387 if (0 < tmp && tmp < (1UL << 13)) 3388 init_admin_reserve(); 3389 3390 break; 3391 case MEM_OFFLINE: 3392 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3393 3394 if (sysctl_user_reserve_kbytes > free_kbytes) { 3395 init_user_reserve(); 3396 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3397 sysctl_user_reserve_kbytes); 3398 } 3399 3400 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3401 init_admin_reserve(); 3402 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3403 sysctl_admin_reserve_kbytes); 3404 } 3405 break; 3406 default: 3407 break; 3408 } 3409 return NOTIFY_OK; 3410 } 3411 3412 static struct notifier_block reserve_mem_nb = { 3413 .notifier_call = reserve_mem_notifier, 3414 }; 3415 3416 static int __meminit init_reserve_notifier(void) 3417 { 3418 if (register_hotmemory_notifier(&reserve_mem_nb)) 3419 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3420 3421 return 0; 3422 } 3423 subsys_initcall(init_reserve_notifier); 3424