1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * mm/mmap.c 4 * 5 * Written by obz. 6 * 7 * Address space accounting code <alan@lxorguk.ukuu.org.uk> 8 */ 9 10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 11 12 #include <linux/kernel.h> 13 #include <linux/slab.h> 14 #include <linux/backing-dev.h> 15 #include <linux/mm.h> 16 #include <linux/mm_inline.h> 17 #include <linux/vmacache.h> 18 #include <linux/shm.h> 19 #include <linux/mman.h> 20 #include <linux/pagemap.h> 21 #include <linux/swap.h> 22 #include <linux/syscalls.h> 23 #include <linux/capability.h> 24 #include <linux/init.h> 25 #include <linux/file.h> 26 #include <linux/fs.h> 27 #include <linux/personality.h> 28 #include <linux/security.h> 29 #include <linux/hugetlb.h> 30 #include <linux/shmem_fs.h> 31 #include <linux/profile.h> 32 #include <linux/export.h> 33 #include <linux/mount.h> 34 #include <linux/mempolicy.h> 35 #include <linux/rmap.h> 36 #include <linux/mmu_notifier.h> 37 #include <linux/mmdebug.h> 38 #include <linux/perf_event.h> 39 #include <linux/audit.h> 40 #include <linux/khugepaged.h> 41 #include <linux/uprobes.h> 42 #include <linux/rbtree_augmented.h> 43 #include <linux/notifier.h> 44 #include <linux/memory.h> 45 #include <linux/printk.h> 46 #include <linux/userfaultfd_k.h> 47 #include <linux/moduleparam.h> 48 #include <linux/pkeys.h> 49 #include <linux/oom.h> 50 #include <linux/sched/mm.h> 51 52 #include <linux/uaccess.h> 53 #include <asm/cacheflush.h> 54 #include <asm/tlb.h> 55 #include <asm/mmu_context.h> 56 57 #define CREATE_TRACE_POINTS 58 #include <trace/events/mmap.h> 59 60 #include "internal.h" 61 62 #ifndef arch_mmap_check 63 #define arch_mmap_check(addr, len, flags) (0) 64 #endif 65 66 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS 67 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN; 68 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX; 69 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS; 70 #endif 71 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS 72 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN; 73 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX; 74 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS; 75 #endif 76 77 static bool ignore_rlimit_data; 78 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644); 79 80 static void unmap_region(struct mm_struct *mm, 81 struct vm_area_struct *vma, struct vm_area_struct *prev, 82 unsigned long start, unsigned long end); 83 84 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags) 85 { 86 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags)); 87 } 88 89 /* Update vma->vm_page_prot to reflect vma->vm_flags. */ 90 void vma_set_page_prot(struct vm_area_struct *vma) 91 { 92 unsigned long vm_flags = vma->vm_flags; 93 pgprot_t vm_page_prot; 94 95 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags); 96 if (vma_wants_writenotify(vma, vm_page_prot)) { 97 vm_flags &= ~VM_SHARED; 98 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags); 99 } 100 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */ 101 WRITE_ONCE(vma->vm_page_prot, vm_page_prot); 102 } 103 104 /* 105 * Requires inode->i_mapping->i_mmap_rwsem 106 */ 107 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 108 struct file *file, struct address_space *mapping) 109 { 110 if (vma->vm_flags & VM_SHARED) 111 mapping_unmap_writable(mapping); 112 113 flush_dcache_mmap_lock(mapping); 114 vma_interval_tree_remove(vma, &mapping->i_mmap); 115 flush_dcache_mmap_unlock(mapping); 116 } 117 118 /* 119 * Unlink a file-based vm structure from its interval tree, to hide 120 * vma from rmap and vmtruncate before freeing its page tables. 121 */ 122 void unlink_file_vma(struct vm_area_struct *vma) 123 { 124 struct file *file = vma->vm_file; 125 126 if (file) { 127 struct address_space *mapping = file->f_mapping; 128 i_mmap_lock_write(mapping); 129 __remove_shared_vm_struct(vma, file, mapping); 130 i_mmap_unlock_write(mapping); 131 } 132 } 133 134 /* 135 * Close a vm structure and free it, returning the next. 136 */ 137 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 138 { 139 struct vm_area_struct *next = vma->vm_next; 140 141 might_sleep(); 142 if (vma->vm_ops && vma->vm_ops->close) 143 vma->vm_ops->close(vma); 144 if (vma->vm_file) 145 fput(vma->vm_file); 146 mpol_put(vma_policy(vma)); 147 vm_area_free(vma); 148 return next; 149 } 150 151 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, 152 struct list_head *uf); 153 SYSCALL_DEFINE1(brk, unsigned long, brk) 154 { 155 unsigned long newbrk, oldbrk, origbrk; 156 struct mm_struct *mm = current->mm; 157 struct vm_area_struct *next; 158 unsigned long min_brk; 159 bool populate; 160 bool downgraded = false; 161 LIST_HEAD(uf); 162 163 if (mmap_write_lock_killable(mm)) 164 return -EINTR; 165 166 origbrk = mm->brk; 167 168 #ifdef CONFIG_COMPAT_BRK 169 /* 170 * CONFIG_COMPAT_BRK can still be overridden by setting 171 * randomize_va_space to 2, which will still cause mm->start_brk 172 * to be arbitrarily shifted 173 */ 174 if (current->brk_randomized) 175 min_brk = mm->start_brk; 176 else 177 min_brk = mm->end_data; 178 #else 179 min_brk = mm->start_brk; 180 #endif 181 if (brk < min_brk) 182 goto out; 183 184 /* 185 * Check against rlimit here. If this check is done later after the test 186 * of oldbrk with newbrk then it can escape the test and let the data 187 * segment grow beyond its set limit the in case where the limit is 188 * not page aligned -Ram Gupta 189 */ 190 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 191 mm->end_data, mm->start_data)) 192 goto out; 193 194 newbrk = PAGE_ALIGN(brk); 195 oldbrk = PAGE_ALIGN(mm->brk); 196 if (oldbrk == newbrk) { 197 mm->brk = brk; 198 goto success; 199 } 200 201 /* 202 * Always allow shrinking brk. 203 * __do_munmap() may downgrade mmap_lock to read. 204 */ 205 if (brk <= mm->brk) { 206 int ret; 207 208 /* 209 * mm->brk must to be protected by write mmap_lock so update it 210 * before downgrading mmap_lock. When __do_munmap() fails, 211 * mm->brk will be restored from origbrk. 212 */ 213 mm->brk = brk; 214 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true); 215 if (ret < 0) { 216 mm->brk = origbrk; 217 goto out; 218 } else if (ret == 1) { 219 downgraded = true; 220 } 221 goto success; 222 } 223 224 /* Check against existing mmap mappings. */ 225 next = find_vma(mm, oldbrk); 226 if (next && newbrk + PAGE_SIZE > vm_start_gap(next)) 227 goto out; 228 229 /* Ok, looks good - let it rip. */ 230 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0) 231 goto out; 232 mm->brk = brk; 233 234 success: 235 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 236 if (downgraded) 237 mmap_read_unlock(mm); 238 else 239 mmap_write_unlock(mm); 240 userfaultfd_unmap_complete(mm, &uf); 241 if (populate) 242 mm_populate(oldbrk, newbrk - oldbrk); 243 return brk; 244 245 out: 246 mmap_write_unlock(mm); 247 return origbrk; 248 } 249 250 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma) 251 { 252 unsigned long gap, prev_end; 253 254 /* 255 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we 256 * allow two stack_guard_gaps between them here, and when choosing 257 * an unmapped area; whereas when expanding we only require one. 258 * That's a little inconsistent, but keeps the code here simpler. 259 */ 260 gap = vm_start_gap(vma); 261 if (vma->vm_prev) { 262 prev_end = vm_end_gap(vma->vm_prev); 263 if (gap > prev_end) 264 gap -= prev_end; 265 else 266 gap = 0; 267 } 268 return gap; 269 } 270 271 #ifdef CONFIG_DEBUG_VM_RB 272 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma) 273 { 274 unsigned long max = vma_compute_gap(vma), subtree_gap; 275 if (vma->vm_rb.rb_left) { 276 subtree_gap = rb_entry(vma->vm_rb.rb_left, 277 struct vm_area_struct, vm_rb)->rb_subtree_gap; 278 if (subtree_gap > max) 279 max = subtree_gap; 280 } 281 if (vma->vm_rb.rb_right) { 282 subtree_gap = rb_entry(vma->vm_rb.rb_right, 283 struct vm_area_struct, vm_rb)->rb_subtree_gap; 284 if (subtree_gap > max) 285 max = subtree_gap; 286 } 287 return max; 288 } 289 290 static int browse_rb(struct mm_struct *mm) 291 { 292 struct rb_root *root = &mm->mm_rb; 293 int i = 0, j, bug = 0; 294 struct rb_node *nd, *pn = NULL; 295 unsigned long prev = 0, pend = 0; 296 297 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 298 struct vm_area_struct *vma; 299 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 300 if (vma->vm_start < prev) { 301 pr_emerg("vm_start %lx < prev %lx\n", 302 vma->vm_start, prev); 303 bug = 1; 304 } 305 if (vma->vm_start < pend) { 306 pr_emerg("vm_start %lx < pend %lx\n", 307 vma->vm_start, pend); 308 bug = 1; 309 } 310 if (vma->vm_start > vma->vm_end) { 311 pr_emerg("vm_start %lx > vm_end %lx\n", 312 vma->vm_start, vma->vm_end); 313 bug = 1; 314 } 315 spin_lock(&mm->page_table_lock); 316 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) { 317 pr_emerg("free gap %lx, correct %lx\n", 318 vma->rb_subtree_gap, 319 vma_compute_subtree_gap(vma)); 320 bug = 1; 321 } 322 spin_unlock(&mm->page_table_lock); 323 i++; 324 pn = nd; 325 prev = vma->vm_start; 326 pend = vma->vm_end; 327 } 328 j = 0; 329 for (nd = pn; nd; nd = rb_prev(nd)) 330 j++; 331 if (i != j) { 332 pr_emerg("backwards %d, forwards %d\n", j, i); 333 bug = 1; 334 } 335 return bug ? -1 : i; 336 } 337 338 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore) 339 { 340 struct rb_node *nd; 341 342 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 343 struct vm_area_struct *vma; 344 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 345 VM_BUG_ON_VMA(vma != ignore && 346 vma->rb_subtree_gap != vma_compute_subtree_gap(vma), 347 vma); 348 } 349 } 350 351 static void validate_mm(struct mm_struct *mm) 352 { 353 int bug = 0; 354 int i = 0; 355 unsigned long highest_address = 0; 356 struct vm_area_struct *vma = mm->mmap; 357 358 while (vma) { 359 struct anon_vma *anon_vma = vma->anon_vma; 360 struct anon_vma_chain *avc; 361 362 if (anon_vma) { 363 anon_vma_lock_read(anon_vma); 364 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 365 anon_vma_interval_tree_verify(avc); 366 anon_vma_unlock_read(anon_vma); 367 } 368 369 highest_address = vm_end_gap(vma); 370 vma = vma->vm_next; 371 i++; 372 } 373 if (i != mm->map_count) { 374 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i); 375 bug = 1; 376 } 377 if (highest_address != mm->highest_vm_end) { 378 pr_emerg("mm->highest_vm_end %lx, found %lx\n", 379 mm->highest_vm_end, highest_address); 380 bug = 1; 381 } 382 i = browse_rb(mm); 383 if (i != mm->map_count) { 384 if (i != -1) 385 pr_emerg("map_count %d rb %d\n", mm->map_count, i); 386 bug = 1; 387 } 388 VM_BUG_ON_MM(bug, mm); 389 } 390 #else 391 #define validate_mm_rb(root, ignore) do { } while (0) 392 #define validate_mm(mm) do { } while (0) 393 #endif 394 395 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks, 396 struct vm_area_struct, vm_rb, 397 unsigned long, rb_subtree_gap, vma_compute_gap) 398 399 /* 400 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or 401 * vma->vm_prev->vm_end values changed, without modifying the vma's position 402 * in the rbtree. 403 */ 404 static void vma_gap_update(struct vm_area_struct *vma) 405 { 406 /* 407 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created 408 * a callback function that does exactly what we want. 409 */ 410 vma_gap_callbacks_propagate(&vma->vm_rb, NULL); 411 } 412 413 static inline void vma_rb_insert(struct vm_area_struct *vma, 414 struct rb_root *root) 415 { 416 /* All rb_subtree_gap values must be consistent prior to insertion */ 417 validate_mm_rb(root, NULL); 418 419 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 420 } 421 422 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root) 423 { 424 /* 425 * Note rb_erase_augmented is a fairly large inline function, 426 * so make sure we instantiate it only once with our desired 427 * augmented rbtree callbacks. 428 */ 429 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 430 } 431 432 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma, 433 struct rb_root *root, 434 struct vm_area_struct *ignore) 435 { 436 /* 437 * All rb_subtree_gap values must be consistent prior to erase, 438 * with the possible exception of 439 * 440 * a. the "next" vma being erased if next->vm_start was reduced in 441 * __vma_adjust() -> __vma_unlink() 442 * b. the vma being erased in detach_vmas_to_be_unmapped() -> 443 * vma_rb_erase() 444 */ 445 validate_mm_rb(root, ignore); 446 447 __vma_rb_erase(vma, root); 448 } 449 450 static __always_inline void vma_rb_erase(struct vm_area_struct *vma, 451 struct rb_root *root) 452 { 453 vma_rb_erase_ignore(vma, root, vma); 454 } 455 456 /* 457 * vma has some anon_vma assigned, and is already inserted on that 458 * anon_vma's interval trees. 459 * 460 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 461 * vma must be removed from the anon_vma's interval trees using 462 * anon_vma_interval_tree_pre_update_vma(). 463 * 464 * After the update, the vma will be reinserted using 465 * anon_vma_interval_tree_post_update_vma(). 466 * 467 * The entire update must be protected by exclusive mmap_lock and by 468 * the root anon_vma's mutex. 469 */ 470 static inline void 471 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 472 { 473 struct anon_vma_chain *avc; 474 475 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 476 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 477 } 478 479 static inline void 480 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 481 { 482 struct anon_vma_chain *avc; 483 484 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 485 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 486 } 487 488 static int find_vma_links(struct mm_struct *mm, unsigned long addr, 489 unsigned long end, struct vm_area_struct **pprev, 490 struct rb_node ***rb_link, struct rb_node **rb_parent) 491 { 492 struct rb_node **__rb_link, *__rb_parent, *rb_prev; 493 494 mmap_assert_locked(mm); 495 __rb_link = &mm->mm_rb.rb_node; 496 rb_prev = __rb_parent = NULL; 497 498 while (*__rb_link) { 499 struct vm_area_struct *vma_tmp; 500 501 __rb_parent = *__rb_link; 502 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 503 504 if (vma_tmp->vm_end > addr) { 505 /* Fail if an existing vma overlaps the area */ 506 if (vma_tmp->vm_start < end) 507 return -ENOMEM; 508 __rb_link = &__rb_parent->rb_left; 509 } else { 510 rb_prev = __rb_parent; 511 __rb_link = &__rb_parent->rb_right; 512 } 513 } 514 515 *pprev = NULL; 516 if (rb_prev) 517 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 518 *rb_link = __rb_link; 519 *rb_parent = __rb_parent; 520 return 0; 521 } 522 523 /* 524 * vma_next() - Get the next VMA. 525 * @mm: The mm_struct. 526 * @vma: The current vma. 527 * 528 * If @vma is NULL, return the first vma in the mm. 529 * 530 * Returns: The next VMA after @vma. 531 */ 532 static inline struct vm_area_struct *vma_next(struct mm_struct *mm, 533 struct vm_area_struct *vma) 534 { 535 if (!vma) 536 return mm->mmap; 537 538 return vma->vm_next; 539 } 540 541 /* 542 * munmap_vma_range() - munmap VMAs that overlap a range. 543 * @mm: The mm struct 544 * @start: The start of the range. 545 * @len: The length of the range. 546 * @pprev: pointer to the pointer that will be set to previous vm_area_struct 547 * @rb_link: the rb_node 548 * @rb_parent: the parent rb_node 549 * 550 * Find all the vm_area_struct that overlap from @start to 551 * @end and munmap them. Set @pprev to the previous vm_area_struct. 552 * 553 * Returns: -ENOMEM on munmap failure or 0 on success. 554 */ 555 static inline int 556 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len, 557 struct vm_area_struct **pprev, struct rb_node ***link, 558 struct rb_node **parent, struct list_head *uf) 559 { 560 561 while (find_vma_links(mm, start, start + len, pprev, link, parent)) 562 if (do_munmap(mm, start, len, uf)) 563 return -ENOMEM; 564 565 return 0; 566 } 567 static unsigned long count_vma_pages_range(struct mm_struct *mm, 568 unsigned long addr, unsigned long end) 569 { 570 unsigned long nr_pages = 0; 571 struct vm_area_struct *vma; 572 573 /* Find first overlapping mapping */ 574 vma = find_vma_intersection(mm, addr, end); 575 if (!vma) 576 return 0; 577 578 nr_pages = (min(end, vma->vm_end) - 579 max(addr, vma->vm_start)) >> PAGE_SHIFT; 580 581 /* Iterate over the rest of the overlaps */ 582 for (vma = vma->vm_next; vma; vma = vma->vm_next) { 583 unsigned long overlap_len; 584 585 if (vma->vm_start > end) 586 break; 587 588 overlap_len = min(end, vma->vm_end) - vma->vm_start; 589 nr_pages += overlap_len >> PAGE_SHIFT; 590 } 591 592 return nr_pages; 593 } 594 595 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 596 struct rb_node **rb_link, struct rb_node *rb_parent) 597 { 598 /* Update tracking information for the gap following the new vma. */ 599 if (vma->vm_next) 600 vma_gap_update(vma->vm_next); 601 else 602 mm->highest_vm_end = vm_end_gap(vma); 603 604 /* 605 * vma->vm_prev wasn't known when we followed the rbtree to find the 606 * correct insertion point for that vma. As a result, we could not 607 * update the vma vm_rb parents rb_subtree_gap values on the way down. 608 * So, we first insert the vma with a zero rb_subtree_gap value 609 * (to be consistent with what we did on the way down), and then 610 * immediately update the gap to the correct value. Finally we 611 * rebalance the rbtree after all augmented values have been set. 612 */ 613 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 614 vma->rb_subtree_gap = 0; 615 vma_gap_update(vma); 616 vma_rb_insert(vma, &mm->mm_rb); 617 } 618 619 static void __vma_link_file(struct vm_area_struct *vma) 620 { 621 struct file *file; 622 623 file = vma->vm_file; 624 if (file) { 625 struct address_space *mapping = file->f_mapping; 626 627 if (vma->vm_flags & VM_SHARED) 628 mapping_allow_writable(mapping); 629 630 flush_dcache_mmap_lock(mapping); 631 vma_interval_tree_insert(vma, &mapping->i_mmap); 632 flush_dcache_mmap_unlock(mapping); 633 } 634 } 635 636 static void 637 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 638 struct vm_area_struct *prev, struct rb_node **rb_link, 639 struct rb_node *rb_parent) 640 { 641 __vma_link_list(mm, vma, prev); 642 __vma_link_rb(mm, vma, rb_link, rb_parent); 643 } 644 645 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 646 struct vm_area_struct *prev, struct rb_node **rb_link, 647 struct rb_node *rb_parent) 648 { 649 struct address_space *mapping = NULL; 650 651 if (vma->vm_file) { 652 mapping = vma->vm_file->f_mapping; 653 i_mmap_lock_write(mapping); 654 } 655 656 __vma_link(mm, vma, prev, rb_link, rb_parent); 657 __vma_link_file(vma); 658 659 if (mapping) 660 i_mmap_unlock_write(mapping); 661 662 mm->map_count++; 663 validate_mm(mm); 664 } 665 666 /* 667 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the 668 * mm's list and rbtree. It has already been inserted into the interval tree. 669 */ 670 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 671 { 672 struct vm_area_struct *prev; 673 struct rb_node **rb_link, *rb_parent; 674 675 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 676 &prev, &rb_link, &rb_parent)) 677 BUG(); 678 __vma_link(mm, vma, prev, rb_link, rb_parent); 679 mm->map_count++; 680 } 681 682 static __always_inline void __vma_unlink(struct mm_struct *mm, 683 struct vm_area_struct *vma, 684 struct vm_area_struct *ignore) 685 { 686 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore); 687 __vma_unlink_list(mm, vma); 688 /* Kill the cache */ 689 vmacache_invalidate(mm); 690 } 691 692 /* 693 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 694 * is already present in an i_mmap tree without adjusting the tree. 695 * The following helper function should be used when such adjustments 696 * are necessary. The "insert" vma (if any) is to be inserted 697 * before we drop the necessary locks. 698 */ 699 int __vma_adjust(struct vm_area_struct *vma, unsigned long start, 700 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert, 701 struct vm_area_struct *expand) 702 { 703 struct mm_struct *mm = vma->vm_mm; 704 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma; 705 struct address_space *mapping = NULL; 706 struct rb_root_cached *root = NULL; 707 struct anon_vma *anon_vma = NULL; 708 struct file *file = vma->vm_file; 709 bool start_changed = false, end_changed = false; 710 long adjust_next = 0; 711 int remove_next = 0; 712 713 if (next && !insert) { 714 struct vm_area_struct *exporter = NULL, *importer = NULL; 715 716 if (end >= next->vm_end) { 717 /* 718 * vma expands, overlapping all the next, and 719 * perhaps the one after too (mprotect case 6). 720 * The only other cases that gets here are 721 * case 1, case 7 and case 8. 722 */ 723 if (next == expand) { 724 /* 725 * The only case where we don't expand "vma" 726 * and we expand "next" instead is case 8. 727 */ 728 VM_WARN_ON(end != next->vm_end); 729 /* 730 * remove_next == 3 means we're 731 * removing "vma" and that to do so we 732 * swapped "vma" and "next". 733 */ 734 remove_next = 3; 735 VM_WARN_ON(file != next->vm_file); 736 swap(vma, next); 737 } else { 738 VM_WARN_ON(expand != vma); 739 /* 740 * case 1, 6, 7, remove_next == 2 is case 6, 741 * remove_next == 1 is case 1 or 7. 742 */ 743 remove_next = 1 + (end > next->vm_end); 744 VM_WARN_ON(remove_next == 2 && 745 end != next->vm_next->vm_end); 746 /* trim end to next, for case 6 first pass */ 747 end = next->vm_end; 748 } 749 750 exporter = next; 751 importer = vma; 752 753 /* 754 * If next doesn't have anon_vma, import from vma after 755 * next, if the vma overlaps with it. 756 */ 757 if (remove_next == 2 && !next->anon_vma) 758 exporter = next->vm_next; 759 760 } else if (end > next->vm_start) { 761 /* 762 * vma expands, overlapping part of the next: 763 * mprotect case 5 shifting the boundary up. 764 */ 765 adjust_next = (end - next->vm_start); 766 exporter = next; 767 importer = vma; 768 VM_WARN_ON(expand != importer); 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); 776 exporter = vma; 777 importer = next; 778 VM_WARN_ON(expand != importer); 779 } 780 781 /* 782 * Easily overlooked: when mprotect shifts the boundary, 783 * make sure the expanding vma has anon_vma set if the 784 * shrinking vma had, to cover any anon pages imported. 785 */ 786 if (exporter && exporter->anon_vma && !importer->anon_vma) { 787 int error; 788 789 importer->anon_vma = exporter->anon_vma; 790 error = anon_vma_clone(importer, exporter); 791 if (error) 792 return error; 793 } 794 } 795 again: 796 vma_adjust_trans_huge(orig_vma, start, end, adjust_next); 797 798 if (file) { 799 mapping = file->f_mapping; 800 root = &mapping->i_mmap; 801 uprobe_munmap(vma, vma->vm_start, vma->vm_end); 802 803 if (adjust_next) 804 uprobe_munmap(next, next->vm_start, next->vm_end); 805 806 i_mmap_lock_write(mapping); 807 if (insert) { 808 /* 809 * Put into interval tree now, so instantiated pages 810 * are visible to arm/parisc __flush_dcache_page 811 * throughout; but we cannot insert into address 812 * space until vma start or end is updated. 813 */ 814 __vma_link_file(insert); 815 } 816 } 817 818 anon_vma = vma->anon_vma; 819 if (!anon_vma && adjust_next) 820 anon_vma = next->anon_vma; 821 if (anon_vma) { 822 VM_WARN_ON(adjust_next && next->anon_vma && 823 anon_vma != next->anon_vma); 824 anon_vma_lock_write(anon_vma); 825 anon_vma_interval_tree_pre_update_vma(vma); 826 if (adjust_next) 827 anon_vma_interval_tree_pre_update_vma(next); 828 } 829 830 if (file) { 831 flush_dcache_mmap_lock(mapping); 832 vma_interval_tree_remove(vma, root); 833 if (adjust_next) 834 vma_interval_tree_remove(next, root); 835 } 836 837 if (start != vma->vm_start) { 838 vma->vm_start = start; 839 start_changed = true; 840 } 841 if (end != vma->vm_end) { 842 vma->vm_end = end; 843 end_changed = true; 844 } 845 vma->vm_pgoff = pgoff; 846 if (adjust_next) { 847 next->vm_start += adjust_next; 848 next->vm_pgoff += adjust_next >> PAGE_SHIFT; 849 } 850 851 if (file) { 852 if (adjust_next) 853 vma_interval_tree_insert(next, root); 854 vma_interval_tree_insert(vma, root); 855 flush_dcache_mmap_unlock(mapping); 856 } 857 858 if (remove_next) { 859 /* 860 * vma_merge has merged next into vma, and needs 861 * us to remove next before dropping the locks. 862 */ 863 if (remove_next != 3) 864 __vma_unlink(mm, next, next); 865 else 866 /* 867 * vma is not before next if they've been 868 * swapped. 869 * 870 * pre-swap() next->vm_start was reduced so 871 * tell validate_mm_rb to ignore pre-swap() 872 * "next" (which is stored in post-swap() 873 * "vma"). 874 */ 875 __vma_unlink(mm, next, vma); 876 if (file) 877 __remove_shared_vm_struct(next, file, mapping); 878 } else if (insert) { 879 /* 880 * split_vma has split insert from vma, and needs 881 * us to insert it before dropping the locks 882 * (it may either follow vma or precede it). 883 */ 884 __insert_vm_struct(mm, insert); 885 } else { 886 if (start_changed) 887 vma_gap_update(vma); 888 if (end_changed) { 889 if (!next) 890 mm->highest_vm_end = vm_end_gap(vma); 891 else if (!adjust_next) 892 vma_gap_update(next); 893 } 894 } 895 896 if (anon_vma) { 897 anon_vma_interval_tree_post_update_vma(vma); 898 if (adjust_next) 899 anon_vma_interval_tree_post_update_vma(next); 900 anon_vma_unlock_write(anon_vma); 901 } 902 903 if (file) { 904 i_mmap_unlock_write(mapping); 905 uprobe_mmap(vma); 906 907 if (adjust_next) 908 uprobe_mmap(next); 909 } 910 911 if (remove_next) { 912 if (file) { 913 uprobe_munmap(next, next->vm_start, next->vm_end); 914 fput(file); 915 } 916 if (next->anon_vma) 917 anon_vma_merge(vma, next); 918 mm->map_count--; 919 mpol_put(vma_policy(next)); 920 vm_area_free(next); 921 /* 922 * In mprotect's case 6 (see comments on vma_merge), 923 * we must remove another next too. It would clutter 924 * up the code too much to do both in one go. 925 */ 926 if (remove_next != 3) { 927 /* 928 * If "next" was removed and vma->vm_end was 929 * expanded (up) over it, in turn 930 * "next->vm_prev->vm_end" changed and the 931 * "vma->vm_next" gap must be updated. 932 */ 933 next = vma->vm_next; 934 } else { 935 /* 936 * For the scope of the comment "next" and 937 * "vma" considered pre-swap(): if "vma" was 938 * removed, next->vm_start was expanded (down) 939 * over it and the "next" gap must be updated. 940 * Because of the swap() the post-swap() "vma" 941 * actually points to pre-swap() "next" 942 * (post-swap() "next" as opposed is now a 943 * dangling pointer). 944 */ 945 next = vma; 946 } 947 if (remove_next == 2) { 948 remove_next = 1; 949 end = next->vm_end; 950 goto again; 951 } 952 else if (next) 953 vma_gap_update(next); 954 else { 955 /* 956 * If remove_next == 2 we obviously can't 957 * reach this path. 958 * 959 * If remove_next == 3 we can't reach this 960 * path because pre-swap() next is always not 961 * NULL. pre-swap() "next" is not being 962 * removed and its next->vm_end is not altered 963 * (and furthermore "end" already matches 964 * next->vm_end in remove_next == 3). 965 * 966 * We reach this only in the remove_next == 1 967 * case if the "next" vma that was removed was 968 * the highest vma of the mm. However in such 969 * case next->vm_end == "end" and the extended 970 * "vma" has vma->vm_end == next->vm_end so 971 * mm->highest_vm_end doesn't need any update 972 * in remove_next == 1 case. 973 */ 974 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma)); 975 } 976 } 977 if (insert && file) 978 uprobe_mmap(insert); 979 980 validate_mm(mm); 981 982 return 0; 983 } 984 985 /* 986 * If the vma has a ->close operation then the driver probably needs to release 987 * per-vma resources, so we don't attempt to merge those. 988 */ 989 static inline int is_mergeable_vma(struct vm_area_struct *vma, 990 struct file *file, unsigned long vm_flags, 991 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 992 struct anon_vma_name *anon_name) 993 { 994 /* 995 * VM_SOFTDIRTY should not prevent from VMA merging, if we 996 * match the flags but dirty bit -- the caller should mark 997 * merged VMA as dirty. If dirty bit won't be excluded from 998 * comparison, we increase pressure on the memory system forcing 999 * the kernel to generate new VMAs when old one could be 1000 * extended instead. 1001 */ 1002 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 1003 return 0; 1004 if (vma->vm_file != file) 1005 return 0; 1006 if (vma->vm_ops && vma->vm_ops->close) 1007 return 0; 1008 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx)) 1009 return 0; 1010 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name)) 1011 return 0; 1012 return 1; 1013 } 1014 1015 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 1016 struct anon_vma *anon_vma2, 1017 struct vm_area_struct *vma) 1018 { 1019 /* 1020 * The list_is_singular() test is to avoid merging VMA cloned from 1021 * parents. This can improve scalability caused by anon_vma lock. 1022 */ 1023 if ((!anon_vma1 || !anon_vma2) && (!vma || 1024 list_is_singular(&vma->anon_vma_chain))) 1025 return 1; 1026 return anon_vma1 == anon_vma2; 1027 } 1028 1029 /* 1030 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 1031 * in front of (at a lower virtual address and file offset than) the vma. 1032 * 1033 * We cannot merge two vmas if they have differently assigned (non-NULL) 1034 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 1035 * 1036 * We don't check here for the merged mmap wrapping around the end of pagecache 1037 * indices (16TB on ia32) because do_mmap() does not permit mmap's which 1038 * wrap, nor mmaps which cover the final page at index -1UL. 1039 */ 1040 static int 1041 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 1042 struct anon_vma *anon_vma, struct file *file, 1043 pgoff_t vm_pgoff, 1044 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 1045 struct anon_vma_name *anon_name) 1046 { 1047 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) && 1048 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1049 if (vma->vm_pgoff == vm_pgoff) 1050 return 1; 1051 } 1052 return 0; 1053 } 1054 1055 /* 1056 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 1057 * beyond (at a higher virtual address and file offset than) the vma. 1058 * 1059 * We cannot merge two vmas if they have differently assigned (non-NULL) 1060 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 1061 */ 1062 static int 1063 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 1064 struct anon_vma *anon_vma, struct file *file, 1065 pgoff_t vm_pgoff, 1066 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 1067 struct anon_vma_name *anon_name) 1068 { 1069 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) && 1070 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1071 pgoff_t vm_pglen; 1072 vm_pglen = vma_pages(vma); 1073 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 1074 return 1; 1075 } 1076 return 0; 1077 } 1078 1079 /* 1080 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name), 1081 * figure out whether that can be merged with its predecessor or its 1082 * successor. Or both (it neatly fills a hole). 1083 * 1084 * In most cases - when called for mmap, brk or mremap - [addr,end) is 1085 * certain not to be mapped by the time vma_merge is called; but when 1086 * called for mprotect, it is certain to be already mapped (either at 1087 * an offset within prev, or at the start of next), and the flags of 1088 * this area are about to be changed to vm_flags - and the no-change 1089 * case has already been eliminated. 1090 * 1091 * The following mprotect cases have to be considered, where AAAA is 1092 * the area passed down from mprotect_fixup, never extending beyond one 1093 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 1094 * 1095 * AAAA AAAA AAAA 1096 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN 1097 * cannot merge might become might become 1098 * PPNNNNNNNNNN PPPPPPPPPPNN 1099 * mmap, brk or case 4 below case 5 below 1100 * mremap move: 1101 * AAAA AAAA 1102 * PPPP NNNN PPPPNNNNXXXX 1103 * might become might become 1104 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or 1105 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or 1106 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8 1107 * 1108 * It is important for case 8 that the vma NNNN overlapping the 1109 * region AAAA is never going to extended over XXXX. Instead XXXX must 1110 * be extended in region AAAA and NNNN must be removed. This way in 1111 * all cases where vma_merge succeeds, the moment vma_adjust drops the 1112 * rmap_locks, the properties of the merged vma will be already 1113 * correct for the whole merged range. Some of those properties like 1114 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must 1115 * be correct for the whole merged range immediately after the 1116 * rmap_locks are released. Otherwise if XXXX would be removed and 1117 * NNNN would be extended over the XXXX range, remove_migration_ptes 1118 * or other rmap walkers (if working on addresses beyond the "end" 1119 * parameter) may establish ptes with the wrong permissions of NNNN 1120 * instead of the right permissions of XXXX. 1121 */ 1122 struct vm_area_struct *vma_merge(struct mm_struct *mm, 1123 struct vm_area_struct *prev, unsigned long addr, 1124 unsigned long end, unsigned long vm_flags, 1125 struct anon_vma *anon_vma, struct file *file, 1126 pgoff_t pgoff, struct mempolicy *policy, 1127 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 1128 struct anon_vma_name *anon_name) 1129 { 1130 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 1131 struct vm_area_struct *area, *next; 1132 int err; 1133 1134 /* 1135 * We later require that vma->vm_flags == vm_flags, 1136 * so this tests vma->vm_flags & VM_SPECIAL, too. 1137 */ 1138 if (vm_flags & VM_SPECIAL) 1139 return NULL; 1140 1141 next = vma_next(mm, prev); 1142 area = next; 1143 if (area && area->vm_end == end) /* cases 6, 7, 8 */ 1144 next = next->vm_next; 1145 1146 /* verify some invariant that must be enforced by the caller */ 1147 VM_WARN_ON(prev && addr <= prev->vm_start); 1148 VM_WARN_ON(area && end > area->vm_end); 1149 VM_WARN_ON(addr >= end); 1150 1151 /* 1152 * Can it merge with the predecessor? 1153 */ 1154 if (prev && prev->vm_end == addr && 1155 mpol_equal(vma_policy(prev), policy) && 1156 can_vma_merge_after(prev, vm_flags, 1157 anon_vma, file, pgoff, 1158 vm_userfaultfd_ctx, anon_name)) { 1159 /* 1160 * OK, it can. Can we now merge in the successor as well? 1161 */ 1162 if (next && end == next->vm_start && 1163 mpol_equal(policy, vma_policy(next)) && 1164 can_vma_merge_before(next, vm_flags, 1165 anon_vma, file, 1166 pgoff+pglen, 1167 vm_userfaultfd_ctx, anon_name) && 1168 is_mergeable_anon_vma(prev->anon_vma, 1169 next->anon_vma, NULL)) { 1170 /* cases 1, 6 */ 1171 err = __vma_adjust(prev, prev->vm_start, 1172 next->vm_end, prev->vm_pgoff, NULL, 1173 prev); 1174 } else /* cases 2, 5, 7 */ 1175 err = __vma_adjust(prev, prev->vm_start, 1176 end, prev->vm_pgoff, NULL, prev); 1177 if (err) 1178 return NULL; 1179 khugepaged_enter_vma(prev, vm_flags); 1180 return prev; 1181 } 1182 1183 /* 1184 * Can this new request be merged in front of next? 1185 */ 1186 if (next && end == next->vm_start && 1187 mpol_equal(policy, vma_policy(next)) && 1188 can_vma_merge_before(next, vm_flags, 1189 anon_vma, file, pgoff+pglen, 1190 vm_userfaultfd_ctx, anon_name)) { 1191 if (prev && addr < prev->vm_end) /* case 4 */ 1192 err = __vma_adjust(prev, prev->vm_start, 1193 addr, prev->vm_pgoff, NULL, next); 1194 else { /* cases 3, 8 */ 1195 err = __vma_adjust(area, addr, next->vm_end, 1196 next->vm_pgoff - pglen, NULL, next); 1197 /* 1198 * In case 3 area is already equal to next and 1199 * this is a noop, but in case 8 "area" has 1200 * been removed and next was expanded over it. 1201 */ 1202 area = next; 1203 } 1204 if (err) 1205 return NULL; 1206 khugepaged_enter_vma(area, vm_flags); 1207 return area; 1208 } 1209 1210 return NULL; 1211 } 1212 1213 /* 1214 * Rough compatibility check to quickly see if it's even worth looking 1215 * at sharing an anon_vma. 1216 * 1217 * They need to have the same vm_file, and the flags can only differ 1218 * in things that mprotect may change. 1219 * 1220 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1221 * we can merge the two vma's. For example, we refuse to merge a vma if 1222 * there is a vm_ops->close() function, because that indicates that the 1223 * driver is doing some kind of reference counting. But that doesn't 1224 * really matter for the anon_vma sharing case. 1225 */ 1226 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1227 { 1228 return a->vm_end == b->vm_start && 1229 mpol_equal(vma_policy(a), vma_policy(b)) && 1230 a->vm_file == b->vm_file && 1231 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) && 1232 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1233 } 1234 1235 /* 1236 * Do some basic sanity checking to see if we can re-use the anon_vma 1237 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1238 * the same as 'old', the other will be the new one that is trying 1239 * to share the anon_vma. 1240 * 1241 * NOTE! This runs with mmap_lock held for reading, so it is possible that 1242 * the anon_vma of 'old' is concurrently in the process of being set up 1243 * by another page fault trying to merge _that_. But that's ok: if it 1244 * is being set up, that automatically means that it will be a singleton 1245 * acceptable for merging, so we can do all of this optimistically. But 1246 * we do that READ_ONCE() to make sure that we never re-load the pointer. 1247 * 1248 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1249 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1250 * is to return an anon_vma that is "complex" due to having gone through 1251 * a fork). 1252 * 1253 * We also make sure that the two vma's are compatible (adjacent, 1254 * and with the same memory policies). That's all stable, even with just 1255 * a read lock on the mmap_lock. 1256 */ 1257 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1258 { 1259 if (anon_vma_compatible(a, b)) { 1260 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma); 1261 1262 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1263 return anon_vma; 1264 } 1265 return NULL; 1266 } 1267 1268 /* 1269 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1270 * neighbouring vmas for a suitable anon_vma, before it goes off 1271 * to allocate a new anon_vma. It checks because a repetitive 1272 * sequence of mprotects and faults may otherwise lead to distinct 1273 * anon_vmas being allocated, preventing vma merge in subsequent 1274 * mprotect. 1275 */ 1276 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1277 { 1278 struct anon_vma *anon_vma = NULL; 1279 1280 /* Try next first. */ 1281 if (vma->vm_next) { 1282 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next); 1283 if (anon_vma) 1284 return anon_vma; 1285 } 1286 1287 /* Try prev next. */ 1288 if (vma->vm_prev) 1289 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma); 1290 1291 /* 1292 * We might reach here with anon_vma == NULL if we can't find 1293 * any reusable anon_vma. 1294 * There's no absolute need to look only at touching neighbours: 1295 * we could search further afield for "compatible" anon_vmas. 1296 * But it would probably just be a waste of time searching, 1297 * or lead to too many vmas hanging off the same anon_vma. 1298 * We're trying to allow mprotect remerging later on, 1299 * not trying to minimize memory used for anon_vmas. 1300 */ 1301 return anon_vma; 1302 } 1303 1304 /* 1305 * If a hint addr is less than mmap_min_addr change hint to be as 1306 * low as possible but still greater than mmap_min_addr 1307 */ 1308 static inline unsigned long round_hint_to_min(unsigned long hint) 1309 { 1310 hint &= PAGE_MASK; 1311 if (((void *)hint != NULL) && 1312 (hint < mmap_min_addr)) 1313 return PAGE_ALIGN(mmap_min_addr); 1314 return hint; 1315 } 1316 1317 int mlock_future_check(struct mm_struct *mm, unsigned long flags, 1318 unsigned long len) 1319 { 1320 unsigned long locked, lock_limit; 1321 1322 /* mlock MCL_FUTURE? */ 1323 if (flags & VM_LOCKED) { 1324 locked = len >> PAGE_SHIFT; 1325 locked += mm->locked_vm; 1326 lock_limit = rlimit(RLIMIT_MEMLOCK); 1327 lock_limit >>= PAGE_SHIFT; 1328 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1329 return -EAGAIN; 1330 } 1331 return 0; 1332 } 1333 1334 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode) 1335 { 1336 if (S_ISREG(inode->i_mode)) 1337 return MAX_LFS_FILESIZE; 1338 1339 if (S_ISBLK(inode->i_mode)) 1340 return MAX_LFS_FILESIZE; 1341 1342 if (S_ISSOCK(inode->i_mode)) 1343 return MAX_LFS_FILESIZE; 1344 1345 /* Special "we do even unsigned file positions" case */ 1346 if (file->f_mode & FMODE_UNSIGNED_OFFSET) 1347 return 0; 1348 1349 /* Yes, random drivers might want more. But I'm tired of buggy drivers */ 1350 return ULONG_MAX; 1351 } 1352 1353 static inline bool file_mmap_ok(struct file *file, struct inode *inode, 1354 unsigned long pgoff, unsigned long len) 1355 { 1356 u64 maxsize = file_mmap_size_max(file, inode); 1357 1358 if (maxsize && len > maxsize) 1359 return false; 1360 maxsize -= len; 1361 if (pgoff > maxsize >> PAGE_SHIFT) 1362 return false; 1363 return true; 1364 } 1365 1366 /* 1367 * The caller must write-lock current->mm->mmap_lock. 1368 */ 1369 unsigned long do_mmap(struct file *file, unsigned long addr, 1370 unsigned long len, unsigned long prot, 1371 unsigned long flags, unsigned long pgoff, 1372 unsigned long *populate, struct list_head *uf) 1373 { 1374 struct mm_struct *mm = current->mm; 1375 vm_flags_t vm_flags; 1376 int pkey = 0; 1377 1378 *populate = 0; 1379 1380 if (!len) 1381 return -EINVAL; 1382 1383 /* 1384 * Does the application expect PROT_READ to imply PROT_EXEC? 1385 * 1386 * (the exception is when the underlying filesystem is noexec 1387 * mounted, in which case we dont add PROT_EXEC.) 1388 */ 1389 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1390 if (!(file && path_noexec(&file->f_path))) 1391 prot |= PROT_EXEC; 1392 1393 /* force arch specific MAP_FIXED handling in get_unmapped_area */ 1394 if (flags & MAP_FIXED_NOREPLACE) 1395 flags |= MAP_FIXED; 1396 1397 if (!(flags & MAP_FIXED)) 1398 addr = round_hint_to_min(addr); 1399 1400 /* Careful about overflows.. */ 1401 len = PAGE_ALIGN(len); 1402 if (!len) 1403 return -ENOMEM; 1404 1405 /* offset overflow? */ 1406 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1407 return -EOVERFLOW; 1408 1409 /* Too many mappings? */ 1410 if (mm->map_count > sysctl_max_map_count) 1411 return -ENOMEM; 1412 1413 /* Obtain the address to map to. we verify (or select) it and ensure 1414 * that it represents a valid section of the address space. 1415 */ 1416 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1417 if (IS_ERR_VALUE(addr)) 1418 return addr; 1419 1420 if (flags & MAP_FIXED_NOREPLACE) { 1421 if (find_vma_intersection(mm, addr, addr + len)) 1422 return -EEXIST; 1423 } 1424 1425 if (prot == PROT_EXEC) { 1426 pkey = execute_only_pkey(mm); 1427 if (pkey < 0) 1428 pkey = 0; 1429 } 1430 1431 /* Do simple checking here so the lower-level routines won't have 1432 * to. we assume access permissions have been handled by the open 1433 * of the memory object, so we don't do any here. 1434 */ 1435 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) | 1436 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1437 1438 if (flags & MAP_LOCKED) 1439 if (!can_do_mlock()) 1440 return -EPERM; 1441 1442 if (mlock_future_check(mm, vm_flags, len)) 1443 return -EAGAIN; 1444 1445 if (file) { 1446 struct inode *inode = file_inode(file); 1447 unsigned long flags_mask; 1448 1449 if (!file_mmap_ok(file, inode, pgoff, len)) 1450 return -EOVERFLOW; 1451 1452 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags; 1453 1454 switch (flags & MAP_TYPE) { 1455 case MAP_SHARED: 1456 /* 1457 * Force use of MAP_SHARED_VALIDATE with non-legacy 1458 * flags. E.g. MAP_SYNC is dangerous to use with 1459 * MAP_SHARED as you don't know which consistency model 1460 * you will get. We silently ignore unsupported flags 1461 * with MAP_SHARED to preserve backward compatibility. 1462 */ 1463 flags &= LEGACY_MAP_MASK; 1464 fallthrough; 1465 case MAP_SHARED_VALIDATE: 1466 if (flags & ~flags_mask) 1467 return -EOPNOTSUPP; 1468 if (prot & PROT_WRITE) { 1469 if (!(file->f_mode & FMODE_WRITE)) 1470 return -EACCES; 1471 if (IS_SWAPFILE(file->f_mapping->host)) 1472 return -ETXTBSY; 1473 } 1474 1475 /* 1476 * Make sure we don't allow writing to an append-only 1477 * file.. 1478 */ 1479 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1480 return -EACCES; 1481 1482 vm_flags |= VM_SHARED | VM_MAYSHARE; 1483 if (!(file->f_mode & FMODE_WRITE)) 1484 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1485 fallthrough; 1486 case MAP_PRIVATE: 1487 if (!(file->f_mode & FMODE_READ)) 1488 return -EACCES; 1489 if (path_noexec(&file->f_path)) { 1490 if (vm_flags & VM_EXEC) 1491 return -EPERM; 1492 vm_flags &= ~VM_MAYEXEC; 1493 } 1494 1495 if (!file->f_op->mmap) 1496 return -ENODEV; 1497 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1498 return -EINVAL; 1499 break; 1500 1501 default: 1502 return -EINVAL; 1503 } 1504 } else { 1505 switch (flags & MAP_TYPE) { 1506 case MAP_SHARED: 1507 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1508 return -EINVAL; 1509 /* 1510 * Ignore pgoff. 1511 */ 1512 pgoff = 0; 1513 vm_flags |= VM_SHARED | VM_MAYSHARE; 1514 break; 1515 case MAP_PRIVATE: 1516 /* 1517 * Set pgoff according to addr for anon_vma. 1518 */ 1519 pgoff = addr >> PAGE_SHIFT; 1520 break; 1521 default: 1522 return -EINVAL; 1523 } 1524 } 1525 1526 /* 1527 * Set 'VM_NORESERVE' if we should not account for the 1528 * memory use of this mapping. 1529 */ 1530 if (flags & MAP_NORESERVE) { 1531 /* We honor MAP_NORESERVE if allowed to overcommit */ 1532 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1533 vm_flags |= VM_NORESERVE; 1534 1535 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1536 if (file && is_file_hugepages(file)) 1537 vm_flags |= VM_NORESERVE; 1538 } 1539 1540 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf); 1541 if (!IS_ERR_VALUE(addr) && 1542 ((vm_flags & VM_LOCKED) || 1543 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1544 *populate = len; 1545 return addr; 1546 } 1547 1548 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len, 1549 unsigned long prot, unsigned long flags, 1550 unsigned long fd, unsigned long pgoff) 1551 { 1552 struct file *file = NULL; 1553 unsigned long retval; 1554 1555 if (!(flags & MAP_ANONYMOUS)) { 1556 audit_mmap_fd(fd, flags); 1557 file = fget(fd); 1558 if (!file) 1559 return -EBADF; 1560 if (is_file_hugepages(file)) { 1561 len = ALIGN(len, huge_page_size(hstate_file(file))); 1562 } else if (unlikely(flags & MAP_HUGETLB)) { 1563 retval = -EINVAL; 1564 goto out_fput; 1565 } 1566 } else if (flags & MAP_HUGETLB) { 1567 struct hstate *hs; 1568 1569 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1570 if (!hs) 1571 return -EINVAL; 1572 1573 len = ALIGN(len, huge_page_size(hs)); 1574 /* 1575 * VM_NORESERVE is used because the reservations will be 1576 * taken when vm_ops->mmap() is called 1577 */ 1578 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1579 VM_NORESERVE, 1580 HUGETLB_ANONHUGE_INODE, 1581 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1582 if (IS_ERR(file)) 1583 return PTR_ERR(file); 1584 } 1585 1586 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1587 out_fput: 1588 if (file) 1589 fput(file); 1590 return retval; 1591 } 1592 1593 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1594 unsigned long, prot, unsigned long, flags, 1595 unsigned long, fd, unsigned long, pgoff) 1596 { 1597 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff); 1598 } 1599 1600 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1601 struct mmap_arg_struct { 1602 unsigned long addr; 1603 unsigned long len; 1604 unsigned long prot; 1605 unsigned long flags; 1606 unsigned long fd; 1607 unsigned long offset; 1608 }; 1609 1610 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1611 { 1612 struct mmap_arg_struct a; 1613 1614 if (copy_from_user(&a, arg, sizeof(a))) 1615 return -EFAULT; 1616 if (offset_in_page(a.offset)) 1617 return -EINVAL; 1618 1619 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1620 a.offset >> PAGE_SHIFT); 1621 } 1622 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1623 1624 /* 1625 * Some shared mappings will want the pages marked read-only 1626 * to track write events. If so, we'll downgrade vm_page_prot 1627 * to the private version (using protection_map[] without the 1628 * VM_SHARED bit). 1629 */ 1630 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot) 1631 { 1632 vm_flags_t vm_flags = vma->vm_flags; 1633 const struct vm_operations_struct *vm_ops = vma->vm_ops; 1634 1635 /* If it was private or non-writable, the write bit is already clear */ 1636 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1637 return 0; 1638 1639 /* The backer wishes to know when pages are first written to? */ 1640 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite)) 1641 return 1; 1642 1643 /* The open routine did something to the protections that pgprot_modify 1644 * won't preserve? */ 1645 if (pgprot_val(vm_page_prot) != 1646 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags))) 1647 return 0; 1648 1649 /* 1650 * Do we need to track softdirty? hugetlb does not support softdirty 1651 * tracking yet. 1652 */ 1653 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma)) 1654 return 1; 1655 1656 /* Specialty mapping? */ 1657 if (vm_flags & VM_PFNMAP) 1658 return 0; 1659 1660 /* Can the mapping track the dirty pages? */ 1661 return vma->vm_file && vma->vm_file->f_mapping && 1662 mapping_can_writeback(vma->vm_file->f_mapping); 1663 } 1664 1665 /* 1666 * We account for memory if it's a private writeable mapping, 1667 * not hugepages and VM_NORESERVE wasn't set. 1668 */ 1669 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1670 { 1671 /* 1672 * hugetlb has its own accounting separate from the core VM 1673 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1674 */ 1675 if (file && is_file_hugepages(file)) 1676 return 0; 1677 1678 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1679 } 1680 1681 unsigned long mmap_region(struct file *file, unsigned long addr, 1682 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff, 1683 struct list_head *uf) 1684 { 1685 struct mm_struct *mm = current->mm; 1686 struct vm_area_struct *vma, *prev, *merge; 1687 int error; 1688 struct rb_node **rb_link, *rb_parent; 1689 unsigned long charged = 0; 1690 1691 /* Check against address space limit. */ 1692 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) { 1693 unsigned long nr_pages; 1694 1695 /* 1696 * MAP_FIXED may remove pages of mappings that intersects with 1697 * requested mapping. Account for the pages it would unmap. 1698 */ 1699 nr_pages = count_vma_pages_range(mm, addr, addr + len); 1700 1701 if (!may_expand_vm(mm, vm_flags, 1702 (len >> PAGE_SHIFT) - nr_pages)) 1703 return -ENOMEM; 1704 } 1705 1706 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */ 1707 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf)) 1708 return -ENOMEM; 1709 /* 1710 * Private writable mapping: check memory availability 1711 */ 1712 if (accountable_mapping(file, vm_flags)) { 1713 charged = len >> PAGE_SHIFT; 1714 if (security_vm_enough_memory_mm(mm, charged)) 1715 return -ENOMEM; 1716 vm_flags |= VM_ACCOUNT; 1717 } 1718 1719 /* 1720 * Can we just expand an old mapping? 1721 */ 1722 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, 1723 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX, NULL); 1724 if (vma) 1725 goto out; 1726 1727 /* 1728 * Determine the object being mapped and call the appropriate 1729 * specific mapper. the address has already been validated, but 1730 * not unmapped, but the maps are removed from the list. 1731 */ 1732 vma = vm_area_alloc(mm); 1733 if (!vma) { 1734 error = -ENOMEM; 1735 goto unacct_error; 1736 } 1737 1738 vma->vm_start = addr; 1739 vma->vm_end = addr + len; 1740 vma->vm_flags = vm_flags; 1741 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1742 vma->vm_pgoff = pgoff; 1743 1744 if (file) { 1745 if (vm_flags & VM_SHARED) { 1746 error = mapping_map_writable(file->f_mapping); 1747 if (error) 1748 goto free_vma; 1749 } 1750 1751 vma->vm_file = get_file(file); 1752 error = call_mmap(file, vma); 1753 if (error) 1754 goto unmap_and_free_vma; 1755 1756 /* Can addr have changed?? 1757 * 1758 * Answer: Yes, several device drivers can do it in their 1759 * f_op->mmap method. -DaveM 1760 * Bug: If addr is changed, prev, rb_link, rb_parent should 1761 * be updated for vma_link() 1762 */ 1763 WARN_ON_ONCE(addr != vma->vm_start); 1764 1765 addr = vma->vm_start; 1766 1767 /* If vm_flags changed after call_mmap(), we should try merge vma again 1768 * as we may succeed this time. 1769 */ 1770 if (unlikely(vm_flags != vma->vm_flags && prev)) { 1771 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags, 1772 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL); 1773 if (merge) { 1774 /* ->mmap() can change vma->vm_file and fput the original file. So 1775 * fput the vma->vm_file here or we would add an extra fput for file 1776 * and cause general protection fault ultimately. 1777 */ 1778 fput(vma->vm_file); 1779 vm_area_free(vma); 1780 vma = merge; 1781 /* Update vm_flags to pick up the change. */ 1782 vm_flags = vma->vm_flags; 1783 goto unmap_writable; 1784 } 1785 } 1786 1787 vm_flags = vma->vm_flags; 1788 } else if (vm_flags & VM_SHARED) { 1789 error = shmem_zero_setup(vma); 1790 if (error) 1791 goto free_vma; 1792 } else { 1793 vma_set_anonymous(vma); 1794 } 1795 1796 /* Allow architectures to sanity-check the vm_flags */ 1797 if (!arch_validate_flags(vma->vm_flags)) { 1798 error = -EINVAL; 1799 if (file) 1800 goto unmap_and_free_vma; 1801 else 1802 goto free_vma; 1803 } 1804 1805 vma_link(mm, vma, prev, rb_link, rb_parent); 1806 1807 /* 1808 * vma_merge() calls khugepaged_enter_vma() either, the below 1809 * call covers the non-merge case. 1810 */ 1811 khugepaged_enter_vma(vma, vma->vm_flags); 1812 1813 /* Once vma denies write, undo our temporary denial count */ 1814 unmap_writable: 1815 if (file && vm_flags & VM_SHARED) 1816 mapping_unmap_writable(file->f_mapping); 1817 file = vma->vm_file; 1818 out: 1819 perf_event_mmap(vma); 1820 1821 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT); 1822 if (vm_flags & VM_LOCKED) { 1823 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) || 1824 is_vm_hugetlb_page(vma) || 1825 vma == get_gate_vma(current->mm)) 1826 vma->vm_flags &= VM_LOCKED_CLEAR_MASK; 1827 else 1828 mm->locked_vm += (len >> PAGE_SHIFT); 1829 } 1830 1831 if (file) 1832 uprobe_mmap(vma); 1833 1834 /* 1835 * New (or expanded) vma always get soft dirty status. 1836 * Otherwise user-space soft-dirty page tracker won't 1837 * be able to distinguish situation when vma area unmapped, 1838 * then new mapped in-place (which must be aimed as 1839 * a completely new data area). 1840 */ 1841 vma->vm_flags |= VM_SOFTDIRTY; 1842 1843 vma_set_page_prot(vma); 1844 1845 return addr; 1846 1847 unmap_and_free_vma: 1848 fput(vma->vm_file); 1849 vma->vm_file = NULL; 1850 1851 /* Undo any partial mapping done by a device driver. */ 1852 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1853 if (vm_flags & VM_SHARED) 1854 mapping_unmap_writable(file->f_mapping); 1855 free_vma: 1856 vm_area_free(vma); 1857 unacct_error: 1858 if (charged) 1859 vm_unacct_memory(charged); 1860 return error; 1861 } 1862 1863 static unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1864 { 1865 /* 1866 * We implement the search by looking for an rbtree node that 1867 * immediately follows a suitable gap. That is, 1868 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length; 1869 * - gap_end = vma->vm_start >= info->low_limit + length; 1870 * - gap_end - gap_start >= length 1871 */ 1872 1873 struct mm_struct *mm = current->mm; 1874 struct vm_area_struct *vma; 1875 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1876 1877 /* Adjust search length to account for worst case alignment overhead */ 1878 length = info->length + info->align_mask; 1879 if (length < info->length) 1880 return -ENOMEM; 1881 1882 /* Adjust search limits by the desired length */ 1883 if (info->high_limit < length) 1884 return -ENOMEM; 1885 high_limit = info->high_limit - length; 1886 1887 if (info->low_limit > high_limit) 1888 return -ENOMEM; 1889 low_limit = info->low_limit + length; 1890 1891 /* Check if rbtree root looks promising */ 1892 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1893 goto check_highest; 1894 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1895 if (vma->rb_subtree_gap < length) 1896 goto check_highest; 1897 1898 while (true) { 1899 /* Visit left subtree if it looks promising */ 1900 gap_end = vm_start_gap(vma); 1901 if (gap_end >= low_limit && vma->vm_rb.rb_left) { 1902 struct vm_area_struct *left = 1903 rb_entry(vma->vm_rb.rb_left, 1904 struct vm_area_struct, vm_rb); 1905 if (left->rb_subtree_gap >= length) { 1906 vma = left; 1907 continue; 1908 } 1909 } 1910 1911 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0; 1912 check_current: 1913 /* Check if current node has a suitable gap */ 1914 if (gap_start > high_limit) 1915 return -ENOMEM; 1916 if (gap_end >= low_limit && 1917 gap_end > gap_start && gap_end - gap_start >= length) 1918 goto found; 1919 1920 /* Visit right subtree if it looks promising */ 1921 if (vma->vm_rb.rb_right) { 1922 struct vm_area_struct *right = 1923 rb_entry(vma->vm_rb.rb_right, 1924 struct vm_area_struct, vm_rb); 1925 if (right->rb_subtree_gap >= length) { 1926 vma = right; 1927 continue; 1928 } 1929 } 1930 1931 /* Go back up the rbtree to find next candidate node */ 1932 while (true) { 1933 struct rb_node *prev = &vma->vm_rb; 1934 if (!rb_parent(prev)) 1935 goto check_highest; 1936 vma = rb_entry(rb_parent(prev), 1937 struct vm_area_struct, vm_rb); 1938 if (prev == vma->vm_rb.rb_left) { 1939 gap_start = vm_end_gap(vma->vm_prev); 1940 gap_end = vm_start_gap(vma); 1941 goto check_current; 1942 } 1943 } 1944 } 1945 1946 check_highest: 1947 /* Check highest gap, which does not precede any rbtree node */ 1948 gap_start = mm->highest_vm_end; 1949 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */ 1950 if (gap_start > high_limit) 1951 return -ENOMEM; 1952 1953 found: 1954 /* We found a suitable gap. Clip it with the original low_limit. */ 1955 if (gap_start < info->low_limit) 1956 gap_start = info->low_limit; 1957 1958 /* Adjust gap address to the desired alignment */ 1959 gap_start += (info->align_offset - gap_start) & info->align_mask; 1960 1961 VM_BUG_ON(gap_start + info->length > info->high_limit); 1962 VM_BUG_ON(gap_start + info->length > gap_end); 1963 return gap_start; 1964 } 1965 1966 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1967 { 1968 struct mm_struct *mm = current->mm; 1969 struct vm_area_struct *vma; 1970 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1971 1972 /* Adjust search length to account for worst case alignment overhead */ 1973 length = info->length + info->align_mask; 1974 if (length < info->length) 1975 return -ENOMEM; 1976 1977 /* 1978 * Adjust search limits by the desired length. 1979 * See implementation comment at top of unmapped_area(). 1980 */ 1981 gap_end = info->high_limit; 1982 if (gap_end < length) 1983 return -ENOMEM; 1984 high_limit = gap_end - length; 1985 1986 if (info->low_limit > high_limit) 1987 return -ENOMEM; 1988 low_limit = info->low_limit + length; 1989 1990 /* Check highest gap, which does not precede any rbtree node */ 1991 gap_start = mm->highest_vm_end; 1992 if (gap_start <= high_limit) 1993 goto found_highest; 1994 1995 /* Check if rbtree root looks promising */ 1996 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1997 return -ENOMEM; 1998 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1999 if (vma->rb_subtree_gap < length) 2000 return -ENOMEM; 2001 2002 while (true) { 2003 /* Visit right subtree if it looks promising */ 2004 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0; 2005 if (gap_start <= high_limit && vma->vm_rb.rb_right) { 2006 struct vm_area_struct *right = 2007 rb_entry(vma->vm_rb.rb_right, 2008 struct vm_area_struct, vm_rb); 2009 if (right->rb_subtree_gap >= length) { 2010 vma = right; 2011 continue; 2012 } 2013 } 2014 2015 check_current: 2016 /* Check if current node has a suitable gap */ 2017 gap_end = vm_start_gap(vma); 2018 if (gap_end < low_limit) 2019 return -ENOMEM; 2020 if (gap_start <= high_limit && 2021 gap_end > gap_start && gap_end - gap_start >= length) 2022 goto found; 2023 2024 /* Visit left subtree if it looks promising */ 2025 if (vma->vm_rb.rb_left) { 2026 struct vm_area_struct *left = 2027 rb_entry(vma->vm_rb.rb_left, 2028 struct vm_area_struct, vm_rb); 2029 if (left->rb_subtree_gap >= length) { 2030 vma = left; 2031 continue; 2032 } 2033 } 2034 2035 /* Go back up the rbtree to find next candidate node */ 2036 while (true) { 2037 struct rb_node *prev = &vma->vm_rb; 2038 if (!rb_parent(prev)) 2039 return -ENOMEM; 2040 vma = rb_entry(rb_parent(prev), 2041 struct vm_area_struct, vm_rb); 2042 if (prev == vma->vm_rb.rb_right) { 2043 gap_start = vma->vm_prev ? 2044 vm_end_gap(vma->vm_prev) : 0; 2045 goto check_current; 2046 } 2047 } 2048 } 2049 2050 found: 2051 /* We found a suitable gap. Clip it with the original high_limit. */ 2052 if (gap_end > info->high_limit) 2053 gap_end = info->high_limit; 2054 2055 found_highest: 2056 /* Compute highest gap address at the desired alignment */ 2057 gap_end -= info->length; 2058 gap_end -= (gap_end - info->align_offset) & info->align_mask; 2059 2060 VM_BUG_ON(gap_end < info->low_limit); 2061 VM_BUG_ON(gap_end < gap_start); 2062 return gap_end; 2063 } 2064 2065 /* 2066 * Search for an unmapped address range. 2067 * 2068 * We are looking for a range that: 2069 * - does not intersect with any VMA; 2070 * - is contained within the [low_limit, high_limit) interval; 2071 * - is at least the desired size. 2072 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask) 2073 */ 2074 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info) 2075 { 2076 unsigned long addr; 2077 2078 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN) 2079 addr = unmapped_area_topdown(info); 2080 else 2081 addr = unmapped_area(info); 2082 2083 trace_vm_unmapped_area(addr, info); 2084 return addr; 2085 } 2086 2087 /* Get an address range which is currently unmapped. 2088 * For shmat() with addr=0. 2089 * 2090 * Ugly calling convention alert: 2091 * Return value with the low bits set means error value, 2092 * ie 2093 * if (ret & ~PAGE_MASK) 2094 * error = ret; 2095 * 2096 * This function "knows" that -ENOMEM has the bits set. 2097 */ 2098 unsigned long 2099 generic_get_unmapped_area(struct file *filp, unsigned long addr, 2100 unsigned long len, unsigned long pgoff, 2101 unsigned long flags) 2102 { 2103 struct mm_struct *mm = current->mm; 2104 struct vm_area_struct *vma, *prev; 2105 struct vm_unmapped_area_info info; 2106 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); 2107 2108 if (len > mmap_end - mmap_min_addr) 2109 return -ENOMEM; 2110 2111 if (flags & MAP_FIXED) 2112 return addr; 2113 2114 if (addr) { 2115 addr = PAGE_ALIGN(addr); 2116 vma = find_vma_prev(mm, addr, &prev); 2117 if (mmap_end - len >= addr && addr >= mmap_min_addr && 2118 (!vma || addr + len <= vm_start_gap(vma)) && 2119 (!prev || addr >= vm_end_gap(prev))) 2120 return addr; 2121 } 2122 2123 info.flags = 0; 2124 info.length = len; 2125 info.low_limit = mm->mmap_base; 2126 info.high_limit = mmap_end; 2127 info.align_mask = 0; 2128 info.align_offset = 0; 2129 return vm_unmapped_area(&info); 2130 } 2131 2132 #ifndef HAVE_ARCH_UNMAPPED_AREA 2133 unsigned long 2134 arch_get_unmapped_area(struct file *filp, unsigned long addr, 2135 unsigned long len, unsigned long pgoff, 2136 unsigned long flags) 2137 { 2138 return generic_get_unmapped_area(filp, addr, len, pgoff, flags); 2139 } 2140 #endif 2141 2142 /* 2143 * This mmap-allocator allocates new areas top-down from below the 2144 * stack's low limit (the base): 2145 */ 2146 unsigned long 2147 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr, 2148 unsigned long len, unsigned long pgoff, 2149 unsigned long flags) 2150 { 2151 struct vm_area_struct *vma, *prev; 2152 struct mm_struct *mm = current->mm; 2153 struct vm_unmapped_area_info info; 2154 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); 2155 2156 /* requested length too big for entire address space */ 2157 if (len > mmap_end - mmap_min_addr) 2158 return -ENOMEM; 2159 2160 if (flags & MAP_FIXED) 2161 return addr; 2162 2163 /* requesting a specific address */ 2164 if (addr) { 2165 addr = PAGE_ALIGN(addr); 2166 vma = find_vma_prev(mm, addr, &prev); 2167 if (mmap_end - len >= addr && addr >= mmap_min_addr && 2168 (!vma || addr + len <= vm_start_gap(vma)) && 2169 (!prev || addr >= vm_end_gap(prev))) 2170 return addr; 2171 } 2172 2173 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 2174 info.length = len; 2175 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 2176 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base); 2177 info.align_mask = 0; 2178 info.align_offset = 0; 2179 addr = vm_unmapped_area(&info); 2180 2181 /* 2182 * A failed mmap() very likely causes application failure, 2183 * so fall back to the bottom-up function here. This scenario 2184 * can happen with large stack limits and large mmap() 2185 * allocations. 2186 */ 2187 if (offset_in_page(addr)) { 2188 VM_BUG_ON(addr != -ENOMEM); 2189 info.flags = 0; 2190 info.low_limit = TASK_UNMAPPED_BASE; 2191 info.high_limit = mmap_end; 2192 addr = vm_unmapped_area(&info); 2193 } 2194 2195 return addr; 2196 } 2197 2198 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 2199 unsigned long 2200 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr, 2201 unsigned long len, unsigned long pgoff, 2202 unsigned long flags) 2203 { 2204 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags); 2205 } 2206 #endif 2207 2208 unsigned long 2209 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 2210 unsigned long pgoff, unsigned long flags) 2211 { 2212 unsigned long (*get_area)(struct file *, unsigned long, 2213 unsigned long, unsigned long, unsigned long); 2214 2215 unsigned long error = arch_mmap_check(addr, len, flags); 2216 if (error) 2217 return error; 2218 2219 /* Careful about overflows.. */ 2220 if (len > TASK_SIZE) 2221 return -ENOMEM; 2222 2223 get_area = current->mm->get_unmapped_area; 2224 if (file) { 2225 if (file->f_op->get_unmapped_area) 2226 get_area = file->f_op->get_unmapped_area; 2227 } else if (flags & MAP_SHARED) { 2228 /* 2229 * mmap_region() will call shmem_zero_setup() to create a file, 2230 * so use shmem's get_unmapped_area in case it can be huge. 2231 * do_mmap() will clear pgoff, so match alignment. 2232 */ 2233 pgoff = 0; 2234 get_area = shmem_get_unmapped_area; 2235 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { 2236 /* Ensures that larger anonymous mappings are THP aligned. */ 2237 get_area = thp_get_unmapped_area; 2238 } 2239 2240 addr = get_area(file, addr, len, pgoff, flags); 2241 if (IS_ERR_VALUE(addr)) 2242 return addr; 2243 2244 if (addr > TASK_SIZE - len) 2245 return -ENOMEM; 2246 if (offset_in_page(addr)) 2247 return -EINVAL; 2248 2249 error = security_mmap_addr(addr); 2250 return error ? error : addr; 2251 } 2252 2253 EXPORT_SYMBOL(get_unmapped_area); 2254 2255 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 2256 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 2257 { 2258 struct rb_node *rb_node; 2259 struct vm_area_struct *vma; 2260 2261 mmap_assert_locked(mm); 2262 /* Check the cache first. */ 2263 vma = vmacache_find(mm, addr); 2264 if (likely(vma)) 2265 return vma; 2266 2267 rb_node = mm->mm_rb.rb_node; 2268 2269 while (rb_node) { 2270 struct vm_area_struct *tmp; 2271 2272 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2273 2274 if (tmp->vm_end > addr) { 2275 vma = tmp; 2276 if (tmp->vm_start <= addr) 2277 break; 2278 rb_node = rb_node->rb_left; 2279 } else 2280 rb_node = rb_node->rb_right; 2281 } 2282 2283 if (vma) 2284 vmacache_update(addr, vma); 2285 return vma; 2286 } 2287 2288 EXPORT_SYMBOL(find_vma); 2289 2290 /* 2291 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 2292 */ 2293 struct vm_area_struct * 2294 find_vma_prev(struct mm_struct *mm, unsigned long addr, 2295 struct vm_area_struct **pprev) 2296 { 2297 struct vm_area_struct *vma; 2298 2299 vma = find_vma(mm, addr); 2300 if (vma) { 2301 *pprev = vma->vm_prev; 2302 } else { 2303 struct rb_node *rb_node = rb_last(&mm->mm_rb); 2304 2305 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL; 2306 } 2307 return vma; 2308 } 2309 2310 /* 2311 * Verify that the stack growth is acceptable and 2312 * update accounting. This is shared with both the 2313 * grow-up and grow-down cases. 2314 */ 2315 static int acct_stack_growth(struct vm_area_struct *vma, 2316 unsigned long size, unsigned long grow) 2317 { 2318 struct mm_struct *mm = vma->vm_mm; 2319 unsigned long new_start; 2320 2321 /* address space limit tests */ 2322 if (!may_expand_vm(mm, vma->vm_flags, grow)) 2323 return -ENOMEM; 2324 2325 /* Stack limit test */ 2326 if (size > rlimit(RLIMIT_STACK)) 2327 return -ENOMEM; 2328 2329 /* mlock limit tests */ 2330 if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT)) 2331 return -ENOMEM; 2332 2333 /* Check to ensure the stack will not grow into a hugetlb-only region */ 2334 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 2335 vma->vm_end - size; 2336 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 2337 return -EFAULT; 2338 2339 /* 2340 * Overcommit.. This must be the final test, as it will 2341 * update security statistics. 2342 */ 2343 if (security_vm_enough_memory_mm(mm, grow)) 2344 return -ENOMEM; 2345 2346 return 0; 2347 } 2348 2349 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 2350 /* 2351 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 2352 * vma is the last one with address > vma->vm_end. Have to extend vma. 2353 */ 2354 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 2355 { 2356 struct mm_struct *mm = vma->vm_mm; 2357 struct vm_area_struct *next; 2358 unsigned long gap_addr; 2359 int error = 0; 2360 2361 if (!(vma->vm_flags & VM_GROWSUP)) 2362 return -EFAULT; 2363 2364 /* Guard against exceeding limits of the address space. */ 2365 address &= PAGE_MASK; 2366 if (address >= (TASK_SIZE & PAGE_MASK)) 2367 return -ENOMEM; 2368 address += PAGE_SIZE; 2369 2370 /* Enforce stack_guard_gap */ 2371 gap_addr = address + stack_guard_gap; 2372 2373 /* Guard against overflow */ 2374 if (gap_addr < address || gap_addr > TASK_SIZE) 2375 gap_addr = TASK_SIZE; 2376 2377 next = vma->vm_next; 2378 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) { 2379 if (!(next->vm_flags & VM_GROWSUP)) 2380 return -ENOMEM; 2381 /* Check that both stack segments have the same anon_vma? */ 2382 } 2383 2384 /* We must make sure the anon_vma is allocated. */ 2385 if (unlikely(anon_vma_prepare(vma))) 2386 return -ENOMEM; 2387 2388 /* 2389 * vma->vm_start/vm_end cannot change under us because the caller 2390 * is required to hold the mmap_lock in read mode. We need the 2391 * anon_vma lock to serialize against concurrent expand_stacks. 2392 */ 2393 anon_vma_lock_write(vma->anon_vma); 2394 2395 /* Somebody else might have raced and expanded it already */ 2396 if (address > vma->vm_end) { 2397 unsigned long size, grow; 2398 2399 size = address - vma->vm_start; 2400 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2401 2402 error = -ENOMEM; 2403 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2404 error = acct_stack_growth(vma, size, grow); 2405 if (!error) { 2406 /* 2407 * vma_gap_update() doesn't support concurrent 2408 * updates, but we only hold a shared mmap_lock 2409 * lock here, so we need to protect against 2410 * concurrent vma expansions. 2411 * anon_vma_lock_write() doesn't help here, as 2412 * we don't guarantee that all growable vmas 2413 * in a mm share the same root anon vma. 2414 * So, we reuse mm->page_table_lock to guard 2415 * against concurrent vma expansions. 2416 */ 2417 spin_lock(&mm->page_table_lock); 2418 if (vma->vm_flags & VM_LOCKED) 2419 mm->locked_vm += grow; 2420 vm_stat_account(mm, vma->vm_flags, grow); 2421 anon_vma_interval_tree_pre_update_vma(vma); 2422 vma->vm_end = address; 2423 anon_vma_interval_tree_post_update_vma(vma); 2424 if (vma->vm_next) 2425 vma_gap_update(vma->vm_next); 2426 else 2427 mm->highest_vm_end = vm_end_gap(vma); 2428 spin_unlock(&mm->page_table_lock); 2429 2430 perf_event_mmap(vma); 2431 } 2432 } 2433 } 2434 anon_vma_unlock_write(vma->anon_vma); 2435 khugepaged_enter_vma(vma, vma->vm_flags); 2436 validate_mm(mm); 2437 return error; 2438 } 2439 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2440 2441 /* 2442 * vma is the first one with address < vma->vm_start. Have to extend vma. 2443 */ 2444 int expand_downwards(struct vm_area_struct *vma, 2445 unsigned long address) 2446 { 2447 struct mm_struct *mm = vma->vm_mm; 2448 struct vm_area_struct *prev; 2449 int error = 0; 2450 2451 address &= PAGE_MASK; 2452 if (address < mmap_min_addr) 2453 return -EPERM; 2454 2455 /* Enforce stack_guard_gap */ 2456 prev = vma->vm_prev; 2457 /* Check that both stack segments have the same anon_vma? */ 2458 if (prev && !(prev->vm_flags & VM_GROWSDOWN) && 2459 vma_is_accessible(prev)) { 2460 if (address - prev->vm_end < stack_guard_gap) 2461 return -ENOMEM; 2462 } 2463 2464 /* We must make sure the anon_vma is allocated. */ 2465 if (unlikely(anon_vma_prepare(vma))) 2466 return -ENOMEM; 2467 2468 /* 2469 * vma->vm_start/vm_end cannot change under us because the caller 2470 * is required to hold the mmap_lock in read mode. We need the 2471 * anon_vma lock to serialize against concurrent expand_stacks. 2472 */ 2473 anon_vma_lock_write(vma->anon_vma); 2474 2475 /* Somebody else might have raced and expanded it already */ 2476 if (address < vma->vm_start) { 2477 unsigned long size, grow; 2478 2479 size = vma->vm_end - address; 2480 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2481 2482 error = -ENOMEM; 2483 if (grow <= vma->vm_pgoff) { 2484 error = acct_stack_growth(vma, size, grow); 2485 if (!error) { 2486 /* 2487 * vma_gap_update() doesn't support concurrent 2488 * updates, but we only hold a shared mmap_lock 2489 * lock here, so we need to protect against 2490 * concurrent vma expansions. 2491 * anon_vma_lock_write() doesn't help here, as 2492 * we don't guarantee that all growable vmas 2493 * in a mm share the same root anon vma. 2494 * So, we reuse mm->page_table_lock to guard 2495 * against concurrent vma expansions. 2496 */ 2497 spin_lock(&mm->page_table_lock); 2498 if (vma->vm_flags & VM_LOCKED) 2499 mm->locked_vm += grow; 2500 vm_stat_account(mm, vma->vm_flags, grow); 2501 anon_vma_interval_tree_pre_update_vma(vma); 2502 vma->vm_start = address; 2503 vma->vm_pgoff -= grow; 2504 anon_vma_interval_tree_post_update_vma(vma); 2505 vma_gap_update(vma); 2506 spin_unlock(&mm->page_table_lock); 2507 2508 perf_event_mmap(vma); 2509 } 2510 } 2511 } 2512 anon_vma_unlock_write(vma->anon_vma); 2513 khugepaged_enter_vma(vma, vma->vm_flags); 2514 validate_mm(mm); 2515 return error; 2516 } 2517 2518 /* enforced gap between the expanding stack and other mappings. */ 2519 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT; 2520 2521 static int __init cmdline_parse_stack_guard_gap(char *p) 2522 { 2523 unsigned long val; 2524 char *endptr; 2525 2526 val = simple_strtoul(p, &endptr, 10); 2527 if (!*endptr) 2528 stack_guard_gap = val << PAGE_SHIFT; 2529 2530 return 1; 2531 } 2532 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap); 2533 2534 #ifdef CONFIG_STACK_GROWSUP 2535 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2536 { 2537 return expand_upwards(vma, address); 2538 } 2539 2540 struct vm_area_struct * 2541 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2542 { 2543 struct vm_area_struct *vma, *prev; 2544 2545 addr &= PAGE_MASK; 2546 vma = find_vma_prev(mm, addr, &prev); 2547 if (vma && (vma->vm_start <= addr)) 2548 return vma; 2549 if (!prev || expand_stack(prev, addr)) 2550 return NULL; 2551 if (prev->vm_flags & VM_LOCKED) 2552 populate_vma_page_range(prev, addr, prev->vm_end, NULL); 2553 return prev; 2554 } 2555 #else 2556 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2557 { 2558 return expand_downwards(vma, address); 2559 } 2560 2561 struct vm_area_struct * 2562 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2563 { 2564 struct vm_area_struct *vma; 2565 unsigned long start; 2566 2567 addr &= PAGE_MASK; 2568 vma = find_vma(mm, addr); 2569 if (!vma) 2570 return NULL; 2571 if (vma->vm_start <= addr) 2572 return vma; 2573 if (!(vma->vm_flags & VM_GROWSDOWN)) 2574 return NULL; 2575 start = vma->vm_start; 2576 if (expand_stack(vma, addr)) 2577 return NULL; 2578 if (vma->vm_flags & VM_LOCKED) 2579 populate_vma_page_range(vma, addr, start, NULL); 2580 return vma; 2581 } 2582 #endif 2583 2584 EXPORT_SYMBOL_GPL(find_extend_vma); 2585 2586 /* 2587 * Ok - we have the memory areas we should free on the vma list, 2588 * so release them, and do the vma updates. 2589 * 2590 * Called with the mm semaphore held. 2591 */ 2592 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 2593 { 2594 unsigned long nr_accounted = 0; 2595 2596 /* Update high watermark before we lower total_vm */ 2597 update_hiwater_vm(mm); 2598 do { 2599 long nrpages = vma_pages(vma); 2600 2601 if (vma->vm_flags & VM_ACCOUNT) 2602 nr_accounted += nrpages; 2603 vm_stat_account(mm, vma->vm_flags, -nrpages); 2604 vma = remove_vma(vma); 2605 } while (vma); 2606 vm_unacct_memory(nr_accounted); 2607 validate_mm(mm); 2608 } 2609 2610 /* 2611 * Get rid of page table information in the indicated region. 2612 * 2613 * Called with the mm semaphore held. 2614 */ 2615 static void unmap_region(struct mm_struct *mm, 2616 struct vm_area_struct *vma, struct vm_area_struct *prev, 2617 unsigned long start, unsigned long end) 2618 { 2619 struct vm_area_struct *next = vma_next(mm, prev); 2620 struct mmu_gather tlb; 2621 2622 lru_add_drain(); 2623 tlb_gather_mmu(&tlb, mm); 2624 update_hiwater_rss(mm); 2625 unmap_vmas(&tlb, vma, start, end); 2626 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2627 next ? next->vm_start : USER_PGTABLES_CEILING); 2628 tlb_finish_mmu(&tlb); 2629 } 2630 2631 /* 2632 * Create a list of vma's touched by the unmap, removing them from the mm's 2633 * vma list as we go.. 2634 */ 2635 static bool 2636 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 2637 struct vm_area_struct *prev, unsigned long end) 2638 { 2639 struct vm_area_struct **insertion_point; 2640 struct vm_area_struct *tail_vma = NULL; 2641 2642 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 2643 vma->vm_prev = NULL; 2644 do { 2645 vma_rb_erase(vma, &mm->mm_rb); 2646 if (vma->vm_flags & VM_LOCKED) 2647 mm->locked_vm -= vma_pages(vma); 2648 mm->map_count--; 2649 tail_vma = vma; 2650 vma = vma->vm_next; 2651 } while (vma && vma->vm_start < end); 2652 *insertion_point = vma; 2653 if (vma) { 2654 vma->vm_prev = prev; 2655 vma_gap_update(vma); 2656 } else 2657 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0; 2658 tail_vma->vm_next = NULL; 2659 2660 /* Kill the cache */ 2661 vmacache_invalidate(mm); 2662 2663 /* 2664 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or 2665 * VM_GROWSUP VMA. Such VMAs can change their size under 2666 * down_read(mmap_lock) and collide with the VMA we are about to unmap. 2667 */ 2668 if (vma && (vma->vm_flags & VM_GROWSDOWN)) 2669 return false; 2670 if (prev && (prev->vm_flags & VM_GROWSUP)) 2671 return false; 2672 return true; 2673 } 2674 2675 /* 2676 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it 2677 * has already been checked or doesn't make sense to fail. 2678 */ 2679 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2680 unsigned long addr, int new_below) 2681 { 2682 struct vm_area_struct *new; 2683 int err; 2684 2685 if (vma->vm_ops && vma->vm_ops->may_split) { 2686 err = vma->vm_ops->may_split(vma, addr); 2687 if (err) 2688 return err; 2689 } 2690 2691 new = vm_area_dup(vma); 2692 if (!new) 2693 return -ENOMEM; 2694 2695 if (new_below) 2696 new->vm_end = addr; 2697 else { 2698 new->vm_start = addr; 2699 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2700 } 2701 2702 err = vma_dup_policy(vma, new); 2703 if (err) 2704 goto out_free_vma; 2705 2706 err = anon_vma_clone(new, vma); 2707 if (err) 2708 goto out_free_mpol; 2709 2710 if (new->vm_file) 2711 get_file(new->vm_file); 2712 2713 if (new->vm_ops && new->vm_ops->open) 2714 new->vm_ops->open(new); 2715 2716 if (new_below) 2717 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 2718 ((addr - new->vm_start) >> PAGE_SHIFT), new); 2719 else 2720 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 2721 2722 /* Success. */ 2723 if (!err) 2724 return 0; 2725 2726 /* Clean everything up if vma_adjust failed. */ 2727 if (new->vm_ops && new->vm_ops->close) 2728 new->vm_ops->close(new); 2729 if (new->vm_file) 2730 fput(new->vm_file); 2731 unlink_anon_vmas(new); 2732 out_free_mpol: 2733 mpol_put(vma_policy(new)); 2734 out_free_vma: 2735 vm_area_free(new); 2736 return err; 2737 } 2738 2739 /* 2740 * Split a vma into two pieces at address 'addr', a new vma is allocated 2741 * either for the first part or the tail. 2742 */ 2743 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2744 unsigned long addr, int new_below) 2745 { 2746 if (mm->map_count >= sysctl_max_map_count) 2747 return -ENOMEM; 2748 2749 return __split_vma(mm, vma, addr, new_below); 2750 } 2751 2752 /* Munmap is split into 2 main parts -- this part which finds 2753 * what needs doing, and the areas themselves, which do the 2754 * work. This now handles partial unmappings. 2755 * Jeremy Fitzhardinge <jeremy@goop.org> 2756 */ 2757 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len, 2758 struct list_head *uf, bool downgrade) 2759 { 2760 unsigned long end; 2761 struct vm_area_struct *vma, *prev, *last; 2762 2763 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start) 2764 return -EINVAL; 2765 2766 len = PAGE_ALIGN(len); 2767 end = start + len; 2768 if (len == 0) 2769 return -EINVAL; 2770 2771 /* 2772 * arch_unmap() might do unmaps itself. It must be called 2773 * and finish any rbtree manipulation before this code 2774 * runs and also starts to manipulate the rbtree. 2775 */ 2776 arch_unmap(mm, start, end); 2777 2778 /* Find the first overlapping VMA where start < vma->vm_end */ 2779 vma = find_vma_intersection(mm, start, end); 2780 if (!vma) 2781 return 0; 2782 prev = vma->vm_prev; 2783 2784 /* 2785 * If we need to split any vma, do it now to save pain later. 2786 * 2787 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2788 * unmapped vm_area_struct will remain in use: so lower split_vma 2789 * places tmp vma above, and higher split_vma places tmp vma below. 2790 */ 2791 if (start > vma->vm_start) { 2792 int error; 2793 2794 /* 2795 * Make sure that map_count on return from munmap() will 2796 * not exceed its limit; but let map_count go just above 2797 * its limit temporarily, to help free resources as expected. 2798 */ 2799 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2800 return -ENOMEM; 2801 2802 error = __split_vma(mm, vma, start, 0); 2803 if (error) 2804 return error; 2805 prev = vma; 2806 } 2807 2808 /* Does it split the last one? */ 2809 last = find_vma(mm, end); 2810 if (last && end > last->vm_start) { 2811 int error = __split_vma(mm, last, end, 1); 2812 if (error) 2813 return error; 2814 } 2815 vma = vma_next(mm, prev); 2816 2817 if (unlikely(uf)) { 2818 /* 2819 * If userfaultfd_unmap_prep returns an error the vmas 2820 * will remain split, but userland will get a 2821 * highly unexpected error anyway. This is no 2822 * different than the case where the first of the two 2823 * __split_vma fails, but we don't undo the first 2824 * split, despite we could. This is unlikely enough 2825 * failure that it's not worth optimizing it for. 2826 */ 2827 int error = userfaultfd_unmap_prep(vma, start, end, uf); 2828 if (error) 2829 return error; 2830 } 2831 2832 /* Detach vmas from rbtree */ 2833 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end)) 2834 downgrade = false; 2835 2836 if (downgrade) 2837 mmap_write_downgrade(mm); 2838 2839 unmap_region(mm, vma, prev, start, end); 2840 2841 /* Fix up all other VM information */ 2842 remove_vma_list(mm, vma); 2843 2844 return downgrade ? 1 : 0; 2845 } 2846 2847 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len, 2848 struct list_head *uf) 2849 { 2850 return __do_munmap(mm, start, len, uf, false); 2851 } 2852 2853 static int __vm_munmap(unsigned long start, size_t len, bool downgrade) 2854 { 2855 int ret; 2856 struct mm_struct *mm = current->mm; 2857 LIST_HEAD(uf); 2858 2859 if (mmap_write_lock_killable(mm)) 2860 return -EINTR; 2861 2862 ret = __do_munmap(mm, start, len, &uf, downgrade); 2863 /* 2864 * Returning 1 indicates mmap_lock is downgraded. 2865 * But 1 is not legal return value of vm_munmap() and munmap(), reset 2866 * it to 0 before return. 2867 */ 2868 if (ret == 1) { 2869 mmap_read_unlock(mm); 2870 ret = 0; 2871 } else 2872 mmap_write_unlock(mm); 2873 2874 userfaultfd_unmap_complete(mm, &uf); 2875 return ret; 2876 } 2877 2878 int vm_munmap(unsigned long start, size_t len) 2879 { 2880 return __vm_munmap(start, len, false); 2881 } 2882 EXPORT_SYMBOL(vm_munmap); 2883 2884 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2885 { 2886 addr = untagged_addr(addr); 2887 return __vm_munmap(addr, len, true); 2888 } 2889 2890 2891 /* 2892 * Emulation of deprecated remap_file_pages() syscall. 2893 */ 2894 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size, 2895 unsigned long, prot, unsigned long, pgoff, unsigned long, flags) 2896 { 2897 2898 struct mm_struct *mm = current->mm; 2899 struct vm_area_struct *vma; 2900 unsigned long populate = 0; 2901 unsigned long ret = -EINVAL; 2902 struct file *file; 2903 2904 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n", 2905 current->comm, current->pid); 2906 2907 if (prot) 2908 return ret; 2909 start = start & PAGE_MASK; 2910 size = size & PAGE_MASK; 2911 2912 if (start + size <= start) 2913 return ret; 2914 2915 /* Does pgoff wrap? */ 2916 if (pgoff + (size >> PAGE_SHIFT) < pgoff) 2917 return ret; 2918 2919 if (mmap_write_lock_killable(mm)) 2920 return -EINTR; 2921 2922 vma = vma_lookup(mm, start); 2923 2924 if (!vma || !(vma->vm_flags & VM_SHARED)) 2925 goto out; 2926 2927 if (start + size > vma->vm_end) { 2928 struct vm_area_struct *next; 2929 2930 for (next = vma->vm_next; next; next = next->vm_next) { 2931 /* hole between vmas ? */ 2932 if (next->vm_start != next->vm_prev->vm_end) 2933 goto out; 2934 2935 if (next->vm_file != vma->vm_file) 2936 goto out; 2937 2938 if (next->vm_flags != vma->vm_flags) 2939 goto out; 2940 2941 if (start + size <= next->vm_end) 2942 break; 2943 } 2944 2945 if (!next) 2946 goto out; 2947 } 2948 2949 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0; 2950 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0; 2951 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0; 2952 2953 flags &= MAP_NONBLOCK; 2954 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE; 2955 if (vma->vm_flags & VM_LOCKED) 2956 flags |= MAP_LOCKED; 2957 2958 file = get_file(vma->vm_file); 2959 ret = do_mmap(vma->vm_file, start, size, 2960 prot, flags, pgoff, &populate, NULL); 2961 fput(file); 2962 out: 2963 mmap_write_unlock(mm); 2964 if (populate) 2965 mm_populate(ret, populate); 2966 if (!IS_ERR_VALUE(ret)) 2967 ret = 0; 2968 return ret; 2969 } 2970 2971 /* 2972 * this is really a simplified "do_mmap". it only handles 2973 * anonymous maps. eventually we may be able to do some 2974 * brk-specific accounting here. 2975 */ 2976 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf) 2977 { 2978 struct mm_struct *mm = current->mm; 2979 struct vm_area_struct *vma, *prev; 2980 struct rb_node **rb_link, *rb_parent; 2981 pgoff_t pgoff = addr >> PAGE_SHIFT; 2982 int error; 2983 unsigned long mapped_addr; 2984 2985 /* Until we need other flags, refuse anything except VM_EXEC. */ 2986 if ((flags & (~VM_EXEC)) != 0) 2987 return -EINVAL; 2988 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2989 2990 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2991 if (IS_ERR_VALUE(mapped_addr)) 2992 return mapped_addr; 2993 2994 error = mlock_future_check(mm, mm->def_flags, len); 2995 if (error) 2996 return error; 2997 2998 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */ 2999 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf)) 3000 return -ENOMEM; 3001 3002 /* Check against address space limits *after* clearing old maps... */ 3003 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT)) 3004 return -ENOMEM; 3005 3006 if (mm->map_count > sysctl_max_map_count) 3007 return -ENOMEM; 3008 3009 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 3010 return -ENOMEM; 3011 3012 /* Can we just expand an old private anonymous mapping? */ 3013 vma = vma_merge(mm, prev, addr, addr + len, flags, 3014 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX, NULL); 3015 if (vma) 3016 goto out; 3017 3018 /* 3019 * create a vma struct for an anonymous mapping 3020 */ 3021 vma = vm_area_alloc(mm); 3022 if (!vma) { 3023 vm_unacct_memory(len >> PAGE_SHIFT); 3024 return -ENOMEM; 3025 } 3026 3027 vma_set_anonymous(vma); 3028 vma->vm_start = addr; 3029 vma->vm_end = addr + len; 3030 vma->vm_pgoff = pgoff; 3031 vma->vm_flags = flags; 3032 vma->vm_page_prot = vm_get_page_prot(flags); 3033 vma_link(mm, vma, prev, rb_link, rb_parent); 3034 out: 3035 perf_event_mmap(vma); 3036 mm->total_vm += len >> PAGE_SHIFT; 3037 mm->data_vm += len >> PAGE_SHIFT; 3038 if (flags & VM_LOCKED) 3039 mm->locked_vm += (len >> PAGE_SHIFT); 3040 vma->vm_flags |= VM_SOFTDIRTY; 3041 return 0; 3042 } 3043 3044 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags) 3045 { 3046 struct mm_struct *mm = current->mm; 3047 unsigned long len; 3048 int ret; 3049 bool populate; 3050 LIST_HEAD(uf); 3051 3052 len = PAGE_ALIGN(request); 3053 if (len < request) 3054 return -ENOMEM; 3055 if (!len) 3056 return 0; 3057 3058 if (mmap_write_lock_killable(mm)) 3059 return -EINTR; 3060 3061 ret = do_brk_flags(addr, len, flags, &uf); 3062 populate = ((mm->def_flags & VM_LOCKED) != 0); 3063 mmap_write_unlock(mm); 3064 userfaultfd_unmap_complete(mm, &uf); 3065 if (populate && !ret) 3066 mm_populate(addr, len); 3067 return ret; 3068 } 3069 EXPORT_SYMBOL(vm_brk_flags); 3070 3071 int vm_brk(unsigned long addr, unsigned long len) 3072 { 3073 return vm_brk_flags(addr, len, 0); 3074 } 3075 EXPORT_SYMBOL(vm_brk); 3076 3077 /* Release all mmaps. */ 3078 void exit_mmap(struct mm_struct *mm) 3079 { 3080 struct mmu_gather tlb; 3081 struct vm_area_struct *vma; 3082 unsigned long nr_accounted = 0; 3083 3084 /* mm's last user has gone, and its about to be pulled down */ 3085 mmu_notifier_release(mm); 3086 3087 if (unlikely(mm_is_oom_victim(mm))) { 3088 /* 3089 * Manually reap the mm to free as much memory as possible. 3090 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard 3091 * this mm from further consideration. Taking mm->mmap_lock for 3092 * write after setting MMF_OOM_SKIP will guarantee that the oom 3093 * reaper will not run on this mm again after mmap_lock is 3094 * dropped. 3095 * 3096 * Nothing can be holding mm->mmap_lock here and the above call 3097 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in 3098 * __oom_reap_task_mm() will not block. 3099 */ 3100 (void)__oom_reap_task_mm(mm); 3101 set_bit(MMF_OOM_SKIP, &mm->flags); 3102 } 3103 3104 mmap_write_lock(mm); 3105 arch_exit_mmap(mm); 3106 3107 vma = mm->mmap; 3108 if (!vma) { 3109 /* Can happen if dup_mmap() received an OOM */ 3110 mmap_write_unlock(mm); 3111 return; 3112 } 3113 3114 lru_add_drain(); 3115 flush_cache_mm(mm); 3116 tlb_gather_mmu_fullmm(&tlb, mm); 3117 /* update_hiwater_rss(mm) here? but nobody should be looking */ 3118 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 3119 unmap_vmas(&tlb, vma, 0, -1); 3120 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING); 3121 tlb_finish_mmu(&tlb); 3122 3123 /* Walk the list again, actually closing and freeing it. */ 3124 while (vma) { 3125 if (vma->vm_flags & VM_ACCOUNT) 3126 nr_accounted += vma_pages(vma); 3127 vma = remove_vma(vma); 3128 cond_resched(); 3129 } 3130 mm->mmap = NULL; 3131 mmap_write_unlock(mm); 3132 vm_unacct_memory(nr_accounted); 3133 } 3134 3135 /* Insert vm structure into process list sorted by address 3136 * and into the inode's i_mmap tree. If vm_file is non-NULL 3137 * then i_mmap_rwsem is taken here. 3138 */ 3139 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 3140 { 3141 struct vm_area_struct *prev; 3142 struct rb_node **rb_link, *rb_parent; 3143 3144 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 3145 &prev, &rb_link, &rb_parent)) 3146 return -ENOMEM; 3147 if ((vma->vm_flags & VM_ACCOUNT) && 3148 security_vm_enough_memory_mm(mm, vma_pages(vma))) 3149 return -ENOMEM; 3150 3151 /* 3152 * The vm_pgoff of a purely anonymous vma should be irrelevant 3153 * until its first write fault, when page's anon_vma and index 3154 * are set. But now set the vm_pgoff it will almost certainly 3155 * end up with (unless mremap moves it elsewhere before that 3156 * first wfault), so /proc/pid/maps tells a consistent story. 3157 * 3158 * By setting it to reflect the virtual start address of the 3159 * vma, merges and splits can happen in a seamless way, just 3160 * using the existing file pgoff checks and manipulations. 3161 * Similarly in do_mmap and in do_brk_flags. 3162 */ 3163 if (vma_is_anonymous(vma)) { 3164 BUG_ON(vma->anon_vma); 3165 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 3166 } 3167 3168 vma_link(mm, vma, prev, rb_link, rb_parent); 3169 return 0; 3170 } 3171 3172 /* 3173 * Copy the vma structure to a new location in the same mm, 3174 * prior to moving page table entries, to effect an mremap move. 3175 */ 3176 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 3177 unsigned long addr, unsigned long len, pgoff_t pgoff, 3178 bool *need_rmap_locks) 3179 { 3180 struct vm_area_struct *vma = *vmap; 3181 unsigned long vma_start = vma->vm_start; 3182 struct mm_struct *mm = vma->vm_mm; 3183 struct vm_area_struct *new_vma, *prev; 3184 struct rb_node **rb_link, *rb_parent; 3185 bool faulted_in_anon_vma = true; 3186 3187 /* 3188 * If anonymous vma has not yet been faulted, update new pgoff 3189 * to match new location, to increase its chance of merging. 3190 */ 3191 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) { 3192 pgoff = addr >> PAGE_SHIFT; 3193 faulted_in_anon_vma = false; 3194 } 3195 3196 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) 3197 return NULL; /* should never get here */ 3198 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 3199 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 3200 vma->vm_userfaultfd_ctx, anon_vma_name(vma)); 3201 if (new_vma) { 3202 /* 3203 * Source vma may have been merged into new_vma 3204 */ 3205 if (unlikely(vma_start >= new_vma->vm_start && 3206 vma_start < new_vma->vm_end)) { 3207 /* 3208 * The only way we can get a vma_merge with 3209 * self during an mremap is if the vma hasn't 3210 * been faulted in yet and we were allowed to 3211 * reset the dst vma->vm_pgoff to the 3212 * destination address of the mremap to allow 3213 * the merge to happen. mremap must change the 3214 * vm_pgoff linearity between src and dst vmas 3215 * (in turn preventing a vma_merge) to be 3216 * safe. It is only safe to keep the vm_pgoff 3217 * linear if there are no pages mapped yet. 3218 */ 3219 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 3220 *vmap = vma = new_vma; 3221 } 3222 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 3223 } else { 3224 new_vma = vm_area_dup(vma); 3225 if (!new_vma) 3226 goto out; 3227 new_vma->vm_start = addr; 3228 new_vma->vm_end = addr + len; 3229 new_vma->vm_pgoff = pgoff; 3230 if (vma_dup_policy(vma, new_vma)) 3231 goto out_free_vma; 3232 if (anon_vma_clone(new_vma, vma)) 3233 goto out_free_mempol; 3234 if (new_vma->vm_file) 3235 get_file(new_vma->vm_file); 3236 if (new_vma->vm_ops && new_vma->vm_ops->open) 3237 new_vma->vm_ops->open(new_vma); 3238 vma_link(mm, new_vma, prev, rb_link, rb_parent); 3239 *need_rmap_locks = false; 3240 } 3241 return new_vma; 3242 3243 out_free_mempol: 3244 mpol_put(vma_policy(new_vma)); 3245 out_free_vma: 3246 vm_area_free(new_vma); 3247 out: 3248 return NULL; 3249 } 3250 3251 /* 3252 * Return true if the calling process may expand its vm space by the passed 3253 * number of pages 3254 */ 3255 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages) 3256 { 3257 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT) 3258 return false; 3259 3260 if (is_data_mapping(flags) && 3261 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) { 3262 /* Workaround for Valgrind */ 3263 if (rlimit(RLIMIT_DATA) == 0 && 3264 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT) 3265 return true; 3266 3267 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n", 3268 current->comm, current->pid, 3269 (mm->data_vm + npages) << PAGE_SHIFT, 3270 rlimit(RLIMIT_DATA), 3271 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data"); 3272 3273 if (!ignore_rlimit_data) 3274 return false; 3275 } 3276 3277 return true; 3278 } 3279 3280 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages) 3281 { 3282 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages); 3283 3284 if (is_exec_mapping(flags)) 3285 mm->exec_vm += npages; 3286 else if (is_stack_mapping(flags)) 3287 mm->stack_vm += npages; 3288 else if (is_data_mapping(flags)) 3289 mm->data_vm += npages; 3290 } 3291 3292 static vm_fault_t special_mapping_fault(struct vm_fault *vmf); 3293 3294 /* 3295 * Having a close hook prevents vma merging regardless of flags. 3296 */ 3297 static void special_mapping_close(struct vm_area_struct *vma) 3298 { 3299 } 3300 3301 static const char *special_mapping_name(struct vm_area_struct *vma) 3302 { 3303 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 3304 } 3305 3306 static int special_mapping_mremap(struct vm_area_struct *new_vma) 3307 { 3308 struct vm_special_mapping *sm = new_vma->vm_private_data; 3309 3310 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm)) 3311 return -EFAULT; 3312 3313 if (sm->mremap) 3314 return sm->mremap(sm, new_vma); 3315 3316 return 0; 3317 } 3318 3319 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr) 3320 { 3321 /* 3322 * Forbid splitting special mappings - kernel has expectations over 3323 * the number of pages in mapping. Together with VM_DONTEXPAND 3324 * the size of vma should stay the same over the special mapping's 3325 * lifetime. 3326 */ 3327 return -EINVAL; 3328 } 3329 3330 static const struct vm_operations_struct special_mapping_vmops = { 3331 .close = special_mapping_close, 3332 .fault = special_mapping_fault, 3333 .mremap = special_mapping_mremap, 3334 .name = special_mapping_name, 3335 /* vDSO code relies that VVAR can't be accessed remotely */ 3336 .access = NULL, 3337 .may_split = special_mapping_split, 3338 }; 3339 3340 static const struct vm_operations_struct legacy_special_mapping_vmops = { 3341 .close = special_mapping_close, 3342 .fault = special_mapping_fault, 3343 }; 3344 3345 static vm_fault_t special_mapping_fault(struct vm_fault *vmf) 3346 { 3347 struct vm_area_struct *vma = vmf->vma; 3348 pgoff_t pgoff; 3349 struct page **pages; 3350 3351 if (vma->vm_ops == &legacy_special_mapping_vmops) { 3352 pages = vma->vm_private_data; 3353 } else { 3354 struct vm_special_mapping *sm = vma->vm_private_data; 3355 3356 if (sm->fault) 3357 return sm->fault(sm, vmf->vma, vmf); 3358 3359 pages = sm->pages; 3360 } 3361 3362 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages) 3363 pgoff--; 3364 3365 if (*pages) { 3366 struct page *page = *pages; 3367 get_page(page); 3368 vmf->page = page; 3369 return 0; 3370 } 3371 3372 return VM_FAULT_SIGBUS; 3373 } 3374 3375 static struct vm_area_struct *__install_special_mapping( 3376 struct mm_struct *mm, 3377 unsigned long addr, unsigned long len, 3378 unsigned long vm_flags, void *priv, 3379 const struct vm_operations_struct *ops) 3380 { 3381 int ret; 3382 struct vm_area_struct *vma; 3383 3384 vma = vm_area_alloc(mm); 3385 if (unlikely(vma == NULL)) 3386 return ERR_PTR(-ENOMEM); 3387 3388 vma->vm_start = addr; 3389 vma->vm_end = addr + len; 3390 3391 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY; 3392 vma->vm_flags &= VM_LOCKED_CLEAR_MASK; 3393 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3394 3395 vma->vm_ops = ops; 3396 vma->vm_private_data = priv; 3397 3398 ret = insert_vm_struct(mm, vma); 3399 if (ret) 3400 goto out; 3401 3402 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT); 3403 3404 perf_event_mmap(vma); 3405 3406 return vma; 3407 3408 out: 3409 vm_area_free(vma); 3410 return ERR_PTR(ret); 3411 } 3412 3413 bool vma_is_special_mapping(const struct vm_area_struct *vma, 3414 const struct vm_special_mapping *sm) 3415 { 3416 return vma->vm_private_data == sm && 3417 (vma->vm_ops == &special_mapping_vmops || 3418 vma->vm_ops == &legacy_special_mapping_vmops); 3419 } 3420 3421 /* 3422 * Called with mm->mmap_lock held for writing. 3423 * Insert a new vma covering the given region, with the given flags. 3424 * Its pages are supplied by the given array of struct page *. 3425 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3426 * The region past the last page supplied will always produce SIGBUS. 3427 * The array pointer and the pages it points to are assumed to stay alive 3428 * for as long as this mapping might exist. 3429 */ 3430 struct vm_area_struct *_install_special_mapping( 3431 struct mm_struct *mm, 3432 unsigned long addr, unsigned long len, 3433 unsigned long vm_flags, const struct vm_special_mapping *spec) 3434 { 3435 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec, 3436 &special_mapping_vmops); 3437 } 3438 3439 int install_special_mapping(struct mm_struct *mm, 3440 unsigned long addr, unsigned long len, 3441 unsigned long vm_flags, struct page **pages) 3442 { 3443 struct vm_area_struct *vma = __install_special_mapping( 3444 mm, addr, len, vm_flags, (void *)pages, 3445 &legacy_special_mapping_vmops); 3446 3447 return PTR_ERR_OR_ZERO(vma); 3448 } 3449 3450 static DEFINE_MUTEX(mm_all_locks_mutex); 3451 3452 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3453 { 3454 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) { 3455 /* 3456 * The LSB of head.next can't change from under us 3457 * because we hold the mm_all_locks_mutex. 3458 */ 3459 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock); 3460 /* 3461 * We can safely modify head.next after taking the 3462 * anon_vma->root->rwsem. If some other vma in this mm shares 3463 * the same anon_vma we won't take it again. 3464 * 3465 * No need of atomic instructions here, head.next 3466 * can't change from under us thanks to the 3467 * anon_vma->root->rwsem. 3468 */ 3469 if (__test_and_set_bit(0, (unsigned long *) 3470 &anon_vma->root->rb_root.rb_root.rb_node)) 3471 BUG(); 3472 } 3473 } 3474 3475 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3476 { 3477 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3478 /* 3479 * AS_MM_ALL_LOCKS can't change from under us because 3480 * we hold the mm_all_locks_mutex. 3481 * 3482 * Operations on ->flags have to be atomic because 3483 * even if AS_MM_ALL_LOCKS is stable thanks to the 3484 * mm_all_locks_mutex, there may be other cpus 3485 * changing other bitflags in parallel to us. 3486 */ 3487 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3488 BUG(); 3489 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock); 3490 } 3491 } 3492 3493 /* 3494 * This operation locks against the VM for all pte/vma/mm related 3495 * operations that could ever happen on a certain mm. This includes 3496 * vmtruncate, try_to_unmap, and all page faults. 3497 * 3498 * The caller must take the mmap_lock in write mode before calling 3499 * mm_take_all_locks(). The caller isn't allowed to release the 3500 * mmap_lock until mm_drop_all_locks() returns. 3501 * 3502 * mmap_lock in write mode is required in order to block all operations 3503 * that could modify pagetables and free pages without need of 3504 * altering the vma layout. It's also needed in write mode to avoid new 3505 * anon_vmas to be associated with existing vmas. 3506 * 3507 * A single task can't take more than one mm_take_all_locks() in a row 3508 * or it would deadlock. 3509 * 3510 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3511 * mapping->flags avoid to take the same lock twice, if more than one 3512 * vma in this mm is backed by the same anon_vma or address_space. 3513 * 3514 * We take locks in following order, accordingly to comment at beginning 3515 * of mm/rmap.c: 3516 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for 3517 * hugetlb mapping); 3518 * - all i_mmap_rwsem locks; 3519 * - all anon_vma->rwseml 3520 * 3521 * We can take all locks within these types randomly because the VM code 3522 * doesn't nest them and we protected from parallel mm_take_all_locks() by 3523 * mm_all_locks_mutex. 3524 * 3525 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3526 * that may have to take thousand of locks. 3527 * 3528 * mm_take_all_locks() can fail if it's interrupted by signals. 3529 */ 3530 int mm_take_all_locks(struct mm_struct *mm) 3531 { 3532 struct vm_area_struct *vma; 3533 struct anon_vma_chain *avc; 3534 3535 mmap_assert_write_locked(mm); 3536 3537 mutex_lock(&mm_all_locks_mutex); 3538 3539 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3540 if (signal_pending(current)) 3541 goto out_unlock; 3542 if (vma->vm_file && vma->vm_file->f_mapping && 3543 is_vm_hugetlb_page(vma)) 3544 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3545 } 3546 3547 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3548 if (signal_pending(current)) 3549 goto out_unlock; 3550 if (vma->vm_file && vma->vm_file->f_mapping && 3551 !is_vm_hugetlb_page(vma)) 3552 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3553 } 3554 3555 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3556 if (signal_pending(current)) 3557 goto out_unlock; 3558 if (vma->anon_vma) 3559 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3560 vm_lock_anon_vma(mm, avc->anon_vma); 3561 } 3562 3563 return 0; 3564 3565 out_unlock: 3566 mm_drop_all_locks(mm); 3567 return -EINTR; 3568 } 3569 3570 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3571 { 3572 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) { 3573 /* 3574 * The LSB of head.next can't change to 0 from under 3575 * us because we hold the mm_all_locks_mutex. 3576 * 3577 * We must however clear the bitflag before unlocking 3578 * the vma so the users using the anon_vma->rb_root will 3579 * never see our bitflag. 3580 * 3581 * No need of atomic instructions here, head.next 3582 * can't change from under us until we release the 3583 * anon_vma->root->rwsem. 3584 */ 3585 if (!__test_and_clear_bit(0, (unsigned long *) 3586 &anon_vma->root->rb_root.rb_root.rb_node)) 3587 BUG(); 3588 anon_vma_unlock_write(anon_vma); 3589 } 3590 } 3591 3592 static void vm_unlock_mapping(struct address_space *mapping) 3593 { 3594 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3595 /* 3596 * AS_MM_ALL_LOCKS can't change to 0 from under us 3597 * because we hold the mm_all_locks_mutex. 3598 */ 3599 i_mmap_unlock_write(mapping); 3600 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3601 &mapping->flags)) 3602 BUG(); 3603 } 3604 } 3605 3606 /* 3607 * The mmap_lock cannot be released by the caller until 3608 * mm_drop_all_locks() returns. 3609 */ 3610 void mm_drop_all_locks(struct mm_struct *mm) 3611 { 3612 struct vm_area_struct *vma; 3613 struct anon_vma_chain *avc; 3614 3615 mmap_assert_write_locked(mm); 3616 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3617 3618 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3619 if (vma->anon_vma) 3620 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3621 vm_unlock_anon_vma(avc->anon_vma); 3622 if (vma->vm_file && vma->vm_file->f_mapping) 3623 vm_unlock_mapping(vma->vm_file->f_mapping); 3624 } 3625 3626 mutex_unlock(&mm_all_locks_mutex); 3627 } 3628 3629 /* 3630 * initialise the percpu counter for VM 3631 */ 3632 void __init mmap_init(void) 3633 { 3634 int ret; 3635 3636 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3637 VM_BUG_ON(ret); 3638 } 3639 3640 /* 3641 * Initialise sysctl_user_reserve_kbytes. 3642 * 3643 * This is intended to prevent a user from starting a single memory hogging 3644 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3645 * mode. 3646 * 3647 * The default value is min(3% of free memory, 128MB) 3648 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3649 */ 3650 static int init_user_reserve(void) 3651 { 3652 unsigned long free_kbytes; 3653 3654 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3655 3656 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3657 return 0; 3658 } 3659 subsys_initcall(init_user_reserve); 3660 3661 /* 3662 * Initialise sysctl_admin_reserve_kbytes. 3663 * 3664 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3665 * to log in and kill a memory hogging process. 3666 * 3667 * Systems with more than 256MB will reserve 8MB, enough to recover 3668 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3669 * only reserve 3% of free pages by default. 3670 */ 3671 static int init_admin_reserve(void) 3672 { 3673 unsigned long free_kbytes; 3674 3675 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3676 3677 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3678 return 0; 3679 } 3680 subsys_initcall(init_admin_reserve); 3681 3682 /* 3683 * Reinititalise user and admin reserves if memory is added or removed. 3684 * 3685 * The default user reserve max is 128MB, and the default max for the 3686 * admin reserve is 8MB. These are usually, but not always, enough to 3687 * enable recovery from a memory hogging process using login/sshd, a shell, 3688 * and tools like top. It may make sense to increase or even disable the 3689 * reserve depending on the existence of swap or variations in the recovery 3690 * tools. So, the admin may have changed them. 3691 * 3692 * If memory is added and the reserves have been eliminated or increased above 3693 * the default max, then we'll trust the admin. 3694 * 3695 * If memory is removed and there isn't enough free memory, then we 3696 * need to reset the reserves. 3697 * 3698 * Otherwise keep the reserve set by the admin. 3699 */ 3700 static int reserve_mem_notifier(struct notifier_block *nb, 3701 unsigned long action, void *data) 3702 { 3703 unsigned long tmp, free_kbytes; 3704 3705 switch (action) { 3706 case MEM_ONLINE: 3707 /* Default max is 128MB. Leave alone if modified by operator. */ 3708 tmp = sysctl_user_reserve_kbytes; 3709 if (0 < tmp && tmp < (1UL << 17)) 3710 init_user_reserve(); 3711 3712 /* Default max is 8MB. Leave alone if modified by operator. */ 3713 tmp = sysctl_admin_reserve_kbytes; 3714 if (0 < tmp && tmp < (1UL << 13)) 3715 init_admin_reserve(); 3716 3717 break; 3718 case MEM_OFFLINE: 3719 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3720 3721 if (sysctl_user_reserve_kbytes > free_kbytes) { 3722 init_user_reserve(); 3723 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3724 sysctl_user_reserve_kbytes); 3725 } 3726 3727 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3728 init_admin_reserve(); 3729 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3730 sysctl_admin_reserve_kbytes); 3731 } 3732 break; 3733 default: 3734 break; 3735 } 3736 return NOTIFY_OK; 3737 } 3738 3739 static struct notifier_block reserve_mem_nb = { 3740 .notifier_call = reserve_mem_notifier, 3741 }; 3742 3743 static int __meminit init_reserve_notifier(void) 3744 { 3745 if (register_hotmemory_notifier(&reserve_mem_nb)) 3746 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3747 3748 return 0; 3749 } 3750 subsys_initcall(init_reserve_notifier); 3751