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/vmacache.h> 17 #include <linux/shm.h> 18 #include <linux/mman.h> 19 #include <linux/pagemap.h> 20 #include <linux/swap.h> 21 #include <linux/syscalls.h> 22 #include <linux/capability.h> 23 #include <linux/init.h> 24 #include <linux/file.h> 25 #include <linux/fs.h> 26 #include <linux/personality.h> 27 #include <linux/security.h> 28 #include <linux/hugetlb.h> 29 #include <linux/shmem_fs.h> 30 #include <linux/profile.h> 31 #include <linux/export.h> 32 #include <linux/mount.h> 33 #include <linux/mempolicy.h> 34 #include <linux/rmap.h> 35 #include <linux/mmu_notifier.h> 36 #include <linux/mmdebug.h> 37 #include <linux/perf_event.h> 38 #include <linux/audit.h> 39 #include <linux/khugepaged.h> 40 #include <linux/uprobes.h> 41 #include <linux/rbtree_augmented.h> 42 #include <linux/notifier.h> 43 #include <linux/memory.h> 44 #include <linux/printk.h> 45 #include <linux/userfaultfd_k.h> 46 #include <linux/moduleparam.h> 47 #include <linux/pkeys.h> 48 #include <linux/oom.h> 49 #include <linux/sched/mm.h> 50 51 #include <linux/uaccess.h> 52 #include <asm/cacheflush.h> 53 #include <asm/tlb.h> 54 #include <asm/mmu_context.h> 55 56 #define CREATE_TRACE_POINTS 57 #include <trace/events/mmap.h> 58 59 #include "internal.h" 60 61 #ifndef arch_mmap_check 62 #define arch_mmap_check(addr, len, flags) (0) 63 #endif 64 65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS 66 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN; 67 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX; 68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS; 69 #endif 70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS 71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN; 72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX; 73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS; 74 #endif 75 76 static bool ignore_rlimit_data; 77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644); 78 79 static void unmap_region(struct mm_struct *mm, 80 struct vm_area_struct *vma, struct vm_area_struct *prev, 81 unsigned long start, unsigned long end); 82 83 /* description of effects of mapping type and prot in current implementation. 84 * this is due to the limited x86 page protection hardware. The expected 85 * behavior is in parens: 86 * 87 * map_type prot 88 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 89 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 90 * w: (no) no w: (no) no w: (yes) yes w: (no) no 91 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 92 * 93 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 94 * w: (no) no w: (no) no w: (copy) copy w: (no) no 95 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 96 */ 97 pgprot_t protection_map[16] __ro_after_init = { 98 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 99 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 100 }; 101 102 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT 103 static inline pgprot_t arch_filter_pgprot(pgprot_t prot) 104 { 105 return prot; 106 } 107 #endif 108 109 pgprot_t vm_get_page_prot(unsigned long vm_flags) 110 { 111 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags & 112 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) | 113 pgprot_val(arch_vm_get_page_prot(vm_flags))); 114 115 return arch_filter_pgprot(ret); 116 } 117 EXPORT_SYMBOL(vm_get_page_prot); 118 119 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags) 120 { 121 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags)); 122 } 123 124 /* Update vma->vm_page_prot to reflect vma->vm_flags. */ 125 void vma_set_page_prot(struct vm_area_struct *vma) 126 { 127 unsigned long vm_flags = vma->vm_flags; 128 pgprot_t vm_page_prot; 129 130 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags); 131 if (vma_wants_writenotify(vma, vm_page_prot)) { 132 vm_flags &= ~VM_SHARED; 133 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags); 134 } 135 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */ 136 WRITE_ONCE(vma->vm_page_prot, vm_page_prot); 137 } 138 139 /* 140 * Requires inode->i_mapping->i_mmap_rwsem 141 */ 142 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 143 struct file *file, struct address_space *mapping) 144 { 145 if (vma->vm_flags & VM_DENYWRITE) 146 allow_write_access(file); 147 if (vma->vm_flags & VM_SHARED) 148 mapping_unmap_writable(mapping); 149 150 flush_dcache_mmap_lock(mapping); 151 vma_interval_tree_remove(vma, &mapping->i_mmap); 152 flush_dcache_mmap_unlock(mapping); 153 } 154 155 /* 156 * Unlink a file-based vm structure from its interval tree, to hide 157 * vma from rmap and vmtruncate before freeing its page tables. 158 */ 159 void unlink_file_vma(struct vm_area_struct *vma) 160 { 161 struct file *file = vma->vm_file; 162 163 if (file) { 164 struct address_space *mapping = file->f_mapping; 165 i_mmap_lock_write(mapping); 166 __remove_shared_vm_struct(vma, file, mapping); 167 i_mmap_unlock_write(mapping); 168 } 169 } 170 171 /* 172 * Close a vm structure and free it, returning the next. 173 */ 174 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 175 { 176 struct vm_area_struct *next = vma->vm_next; 177 178 might_sleep(); 179 if (vma->vm_ops && vma->vm_ops->close) 180 vma->vm_ops->close(vma); 181 if (vma->vm_file) 182 fput(vma->vm_file); 183 mpol_put(vma_policy(vma)); 184 vm_area_free(vma); 185 return next; 186 } 187 188 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, 189 struct list_head *uf); 190 SYSCALL_DEFINE1(brk, unsigned long, brk) 191 { 192 unsigned long retval; 193 unsigned long newbrk, oldbrk, origbrk; 194 struct mm_struct *mm = current->mm; 195 struct vm_area_struct *next; 196 unsigned long min_brk; 197 bool populate; 198 bool downgraded = false; 199 LIST_HEAD(uf); 200 201 if (mmap_write_lock_killable(mm)) 202 return -EINTR; 203 204 origbrk = mm->brk; 205 206 #ifdef CONFIG_COMPAT_BRK 207 /* 208 * CONFIG_COMPAT_BRK can still be overridden by setting 209 * randomize_va_space to 2, which will still cause mm->start_brk 210 * to be arbitrarily shifted 211 */ 212 if (current->brk_randomized) 213 min_brk = mm->start_brk; 214 else 215 min_brk = mm->end_data; 216 #else 217 min_brk = mm->start_brk; 218 #endif 219 if (brk < min_brk) 220 goto out; 221 222 /* 223 * Check against rlimit here. If this check is done later after the test 224 * of oldbrk with newbrk then it can escape the test and let the data 225 * segment grow beyond its set limit the in case where the limit is 226 * not page aligned -Ram Gupta 227 */ 228 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 229 mm->end_data, mm->start_data)) 230 goto out; 231 232 newbrk = PAGE_ALIGN(brk); 233 oldbrk = PAGE_ALIGN(mm->brk); 234 if (oldbrk == newbrk) { 235 mm->brk = brk; 236 goto success; 237 } 238 239 /* 240 * Always allow shrinking brk. 241 * __do_munmap() may downgrade mmap_lock to read. 242 */ 243 if (brk <= mm->brk) { 244 int ret; 245 246 /* 247 * mm->brk must to be protected by write mmap_lock so update it 248 * before downgrading mmap_lock. When __do_munmap() fails, 249 * mm->brk will be restored from origbrk. 250 */ 251 mm->brk = brk; 252 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true); 253 if (ret < 0) { 254 mm->brk = origbrk; 255 goto out; 256 } else if (ret == 1) { 257 downgraded = true; 258 } 259 goto success; 260 } 261 262 /* Check against existing mmap mappings. */ 263 next = find_vma(mm, oldbrk); 264 if (next && newbrk + PAGE_SIZE > vm_start_gap(next)) 265 goto out; 266 267 /* Ok, looks good - let it rip. */ 268 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0) 269 goto out; 270 mm->brk = brk; 271 272 success: 273 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 274 if (downgraded) 275 mmap_read_unlock(mm); 276 else 277 mmap_write_unlock(mm); 278 userfaultfd_unmap_complete(mm, &uf); 279 if (populate) 280 mm_populate(oldbrk, newbrk - oldbrk); 281 return brk; 282 283 out: 284 retval = origbrk; 285 mmap_write_unlock(mm); 286 return retval; 287 } 288 289 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma) 290 { 291 unsigned long gap, prev_end; 292 293 /* 294 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we 295 * allow two stack_guard_gaps between them here, and when choosing 296 * an unmapped area; whereas when expanding we only require one. 297 * That's a little inconsistent, but keeps the code here simpler. 298 */ 299 gap = vm_start_gap(vma); 300 if (vma->vm_prev) { 301 prev_end = vm_end_gap(vma->vm_prev); 302 if (gap > prev_end) 303 gap -= prev_end; 304 else 305 gap = 0; 306 } 307 return gap; 308 } 309 310 #ifdef CONFIG_DEBUG_VM_RB 311 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma) 312 { 313 unsigned long max = vma_compute_gap(vma), subtree_gap; 314 if (vma->vm_rb.rb_left) { 315 subtree_gap = rb_entry(vma->vm_rb.rb_left, 316 struct vm_area_struct, vm_rb)->rb_subtree_gap; 317 if (subtree_gap > max) 318 max = subtree_gap; 319 } 320 if (vma->vm_rb.rb_right) { 321 subtree_gap = rb_entry(vma->vm_rb.rb_right, 322 struct vm_area_struct, vm_rb)->rb_subtree_gap; 323 if (subtree_gap > max) 324 max = subtree_gap; 325 } 326 return max; 327 } 328 329 static int browse_rb(struct mm_struct *mm) 330 { 331 struct rb_root *root = &mm->mm_rb; 332 int i = 0, j, bug = 0; 333 struct rb_node *nd, *pn = NULL; 334 unsigned long prev = 0, pend = 0; 335 336 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 337 struct vm_area_struct *vma; 338 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 339 if (vma->vm_start < prev) { 340 pr_emerg("vm_start %lx < prev %lx\n", 341 vma->vm_start, prev); 342 bug = 1; 343 } 344 if (vma->vm_start < pend) { 345 pr_emerg("vm_start %lx < pend %lx\n", 346 vma->vm_start, pend); 347 bug = 1; 348 } 349 if (vma->vm_start > vma->vm_end) { 350 pr_emerg("vm_start %lx > vm_end %lx\n", 351 vma->vm_start, vma->vm_end); 352 bug = 1; 353 } 354 spin_lock(&mm->page_table_lock); 355 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) { 356 pr_emerg("free gap %lx, correct %lx\n", 357 vma->rb_subtree_gap, 358 vma_compute_subtree_gap(vma)); 359 bug = 1; 360 } 361 spin_unlock(&mm->page_table_lock); 362 i++; 363 pn = nd; 364 prev = vma->vm_start; 365 pend = vma->vm_end; 366 } 367 j = 0; 368 for (nd = pn; nd; nd = rb_prev(nd)) 369 j++; 370 if (i != j) { 371 pr_emerg("backwards %d, forwards %d\n", j, i); 372 bug = 1; 373 } 374 return bug ? -1 : i; 375 } 376 377 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore) 378 { 379 struct rb_node *nd; 380 381 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 382 struct vm_area_struct *vma; 383 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 384 VM_BUG_ON_VMA(vma != ignore && 385 vma->rb_subtree_gap != vma_compute_subtree_gap(vma), 386 vma); 387 } 388 } 389 390 static void validate_mm(struct mm_struct *mm) 391 { 392 int bug = 0; 393 int i = 0; 394 unsigned long highest_address = 0; 395 struct vm_area_struct *vma = mm->mmap; 396 397 while (vma) { 398 struct anon_vma *anon_vma = vma->anon_vma; 399 struct anon_vma_chain *avc; 400 401 if (anon_vma) { 402 anon_vma_lock_read(anon_vma); 403 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 404 anon_vma_interval_tree_verify(avc); 405 anon_vma_unlock_read(anon_vma); 406 } 407 408 highest_address = vm_end_gap(vma); 409 vma = vma->vm_next; 410 i++; 411 } 412 if (i != mm->map_count) { 413 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i); 414 bug = 1; 415 } 416 if (highest_address != mm->highest_vm_end) { 417 pr_emerg("mm->highest_vm_end %lx, found %lx\n", 418 mm->highest_vm_end, highest_address); 419 bug = 1; 420 } 421 i = browse_rb(mm); 422 if (i != mm->map_count) { 423 if (i != -1) 424 pr_emerg("map_count %d rb %d\n", mm->map_count, i); 425 bug = 1; 426 } 427 VM_BUG_ON_MM(bug, mm); 428 } 429 #else 430 #define validate_mm_rb(root, ignore) do { } while (0) 431 #define validate_mm(mm) do { } while (0) 432 #endif 433 434 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks, 435 struct vm_area_struct, vm_rb, 436 unsigned long, rb_subtree_gap, vma_compute_gap) 437 438 /* 439 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or 440 * vma->vm_prev->vm_end values changed, without modifying the vma's position 441 * in the rbtree. 442 */ 443 static void vma_gap_update(struct vm_area_struct *vma) 444 { 445 /* 446 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created 447 * a callback function that does exactly what we want. 448 */ 449 vma_gap_callbacks_propagate(&vma->vm_rb, NULL); 450 } 451 452 static inline void vma_rb_insert(struct vm_area_struct *vma, 453 struct rb_root *root) 454 { 455 /* All rb_subtree_gap values must be consistent prior to insertion */ 456 validate_mm_rb(root, NULL); 457 458 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 459 } 460 461 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root) 462 { 463 /* 464 * Note rb_erase_augmented is a fairly large inline function, 465 * so make sure we instantiate it only once with our desired 466 * augmented rbtree callbacks. 467 */ 468 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 469 } 470 471 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma, 472 struct rb_root *root, 473 struct vm_area_struct *ignore) 474 { 475 /* 476 * All rb_subtree_gap values must be consistent prior to erase, 477 * with the possible exception of 478 * 479 * a. the "next" vma being erased if next->vm_start was reduced in 480 * __vma_adjust() -> __vma_unlink() 481 * b. the vma being erased in detach_vmas_to_be_unmapped() -> 482 * vma_rb_erase() 483 */ 484 validate_mm_rb(root, ignore); 485 486 __vma_rb_erase(vma, root); 487 } 488 489 static __always_inline void vma_rb_erase(struct vm_area_struct *vma, 490 struct rb_root *root) 491 { 492 vma_rb_erase_ignore(vma, root, vma); 493 } 494 495 /* 496 * vma has some anon_vma assigned, and is already inserted on that 497 * anon_vma's interval trees. 498 * 499 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 500 * vma must be removed from the anon_vma's interval trees using 501 * anon_vma_interval_tree_pre_update_vma(). 502 * 503 * After the update, the vma will be reinserted using 504 * anon_vma_interval_tree_post_update_vma(). 505 * 506 * The entire update must be protected by exclusive mmap_lock and by 507 * the root anon_vma's mutex. 508 */ 509 static inline void 510 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 511 { 512 struct anon_vma_chain *avc; 513 514 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 515 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 516 } 517 518 static inline void 519 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 520 { 521 struct anon_vma_chain *avc; 522 523 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 524 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 525 } 526 527 static int find_vma_links(struct mm_struct *mm, unsigned long addr, 528 unsigned long end, struct vm_area_struct **pprev, 529 struct rb_node ***rb_link, struct rb_node **rb_parent) 530 { 531 struct rb_node **__rb_link, *__rb_parent, *rb_prev; 532 533 __rb_link = &mm->mm_rb.rb_node; 534 rb_prev = __rb_parent = NULL; 535 536 while (*__rb_link) { 537 struct vm_area_struct *vma_tmp; 538 539 __rb_parent = *__rb_link; 540 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 541 542 if (vma_tmp->vm_end > addr) { 543 /* Fail if an existing vma overlaps the area */ 544 if (vma_tmp->vm_start < end) 545 return -ENOMEM; 546 __rb_link = &__rb_parent->rb_left; 547 } else { 548 rb_prev = __rb_parent; 549 __rb_link = &__rb_parent->rb_right; 550 } 551 } 552 553 *pprev = NULL; 554 if (rb_prev) 555 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 556 *rb_link = __rb_link; 557 *rb_parent = __rb_parent; 558 return 0; 559 } 560 561 /* 562 * vma_next() - Get the next VMA. 563 * @mm: The mm_struct. 564 * @vma: The current vma. 565 * 566 * If @vma is NULL, return the first vma in the mm. 567 * 568 * Returns: The next VMA after @vma. 569 */ 570 static inline struct vm_area_struct *vma_next(struct mm_struct *mm, 571 struct vm_area_struct *vma) 572 { 573 if (!vma) 574 return mm->mmap; 575 576 return vma->vm_next; 577 } 578 579 /* 580 * munmap_vma_range() - munmap VMAs that overlap a range. 581 * @mm: The mm struct 582 * @start: The start of the range. 583 * @len: The length of the range. 584 * @pprev: pointer to the pointer that will be set to previous vm_area_struct 585 * @rb_link: the rb_node 586 * @rb_parent: the parent rb_node 587 * 588 * Find all the vm_area_struct that overlap from @start to 589 * @end and munmap them. Set @pprev to the previous vm_area_struct. 590 * 591 * Returns: -ENOMEM on munmap failure or 0 on success. 592 */ 593 static inline int 594 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len, 595 struct vm_area_struct **pprev, struct rb_node ***link, 596 struct rb_node **parent, struct list_head *uf) 597 { 598 599 while (find_vma_links(mm, start, start + len, pprev, link, parent)) 600 if (do_munmap(mm, start, len, uf)) 601 return -ENOMEM; 602 603 return 0; 604 } 605 static unsigned long count_vma_pages_range(struct mm_struct *mm, 606 unsigned long addr, unsigned long end) 607 { 608 unsigned long nr_pages = 0; 609 struct vm_area_struct *vma; 610 611 /* Find first overlaping mapping */ 612 vma = find_vma_intersection(mm, addr, end); 613 if (!vma) 614 return 0; 615 616 nr_pages = (min(end, vma->vm_end) - 617 max(addr, vma->vm_start)) >> PAGE_SHIFT; 618 619 /* Iterate over the rest of the overlaps */ 620 for (vma = vma->vm_next; vma; vma = vma->vm_next) { 621 unsigned long overlap_len; 622 623 if (vma->vm_start > end) 624 break; 625 626 overlap_len = min(end, vma->vm_end) - vma->vm_start; 627 nr_pages += overlap_len >> PAGE_SHIFT; 628 } 629 630 return nr_pages; 631 } 632 633 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 634 struct rb_node **rb_link, struct rb_node *rb_parent) 635 { 636 /* Update tracking information for the gap following the new vma. */ 637 if (vma->vm_next) 638 vma_gap_update(vma->vm_next); 639 else 640 mm->highest_vm_end = vm_end_gap(vma); 641 642 /* 643 * vma->vm_prev wasn't known when we followed the rbtree to find the 644 * correct insertion point for that vma. As a result, we could not 645 * update the vma vm_rb parents rb_subtree_gap values on the way down. 646 * So, we first insert the vma with a zero rb_subtree_gap value 647 * (to be consistent with what we did on the way down), and then 648 * immediately update the gap to the correct value. Finally we 649 * rebalance the rbtree after all augmented values have been set. 650 */ 651 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 652 vma->rb_subtree_gap = 0; 653 vma_gap_update(vma); 654 vma_rb_insert(vma, &mm->mm_rb); 655 } 656 657 static void __vma_link_file(struct vm_area_struct *vma) 658 { 659 struct file *file; 660 661 file = vma->vm_file; 662 if (file) { 663 struct address_space *mapping = file->f_mapping; 664 665 if (vma->vm_flags & VM_DENYWRITE) 666 put_write_access(file_inode(file)); 667 if (vma->vm_flags & VM_SHARED) 668 mapping_allow_writable(mapping); 669 670 flush_dcache_mmap_lock(mapping); 671 vma_interval_tree_insert(vma, &mapping->i_mmap); 672 flush_dcache_mmap_unlock(mapping); 673 } 674 } 675 676 static void 677 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 678 struct vm_area_struct *prev, struct rb_node **rb_link, 679 struct rb_node *rb_parent) 680 { 681 __vma_link_list(mm, vma, prev); 682 __vma_link_rb(mm, vma, rb_link, rb_parent); 683 } 684 685 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 686 struct vm_area_struct *prev, struct rb_node **rb_link, 687 struct rb_node *rb_parent) 688 { 689 struct address_space *mapping = NULL; 690 691 if (vma->vm_file) { 692 mapping = vma->vm_file->f_mapping; 693 i_mmap_lock_write(mapping); 694 } 695 696 __vma_link(mm, vma, prev, rb_link, rb_parent); 697 __vma_link_file(vma); 698 699 if (mapping) 700 i_mmap_unlock_write(mapping); 701 702 mm->map_count++; 703 validate_mm(mm); 704 } 705 706 /* 707 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the 708 * mm's list and rbtree. It has already been inserted into the interval tree. 709 */ 710 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 711 { 712 struct vm_area_struct *prev; 713 struct rb_node **rb_link, *rb_parent; 714 715 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 716 &prev, &rb_link, &rb_parent)) 717 BUG(); 718 __vma_link(mm, vma, prev, rb_link, rb_parent); 719 mm->map_count++; 720 } 721 722 static __always_inline void __vma_unlink(struct mm_struct *mm, 723 struct vm_area_struct *vma, 724 struct vm_area_struct *ignore) 725 { 726 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore); 727 __vma_unlink_list(mm, vma); 728 /* Kill the cache */ 729 vmacache_invalidate(mm); 730 } 731 732 /* 733 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 734 * is already present in an i_mmap tree without adjusting the tree. 735 * The following helper function should be used when such adjustments 736 * are necessary. The "insert" vma (if any) is to be inserted 737 * before we drop the necessary locks. 738 */ 739 int __vma_adjust(struct vm_area_struct *vma, unsigned long start, 740 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert, 741 struct vm_area_struct *expand) 742 { 743 struct mm_struct *mm = vma->vm_mm; 744 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma; 745 struct address_space *mapping = NULL; 746 struct rb_root_cached *root = NULL; 747 struct anon_vma *anon_vma = NULL; 748 struct file *file = vma->vm_file; 749 bool start_changed = false, end_changed = false; 750 long adjust_next = 0; 751 int remove_next = 0; 752 753 if (next && !insert) { 754 struct vm_area_struct *exporter = NULL, *importer = NULL; 755 756 if (end >= next->vm_end) { 757 /* 758 * vma expands, overlapping all the next, and 759 * perhaps the one after too (mprotect case 6). 760 * The only other cases that gets here are 761 * case 1, case 7 and case 8. 762 */ 763 if (next == expand) { 764 /* 765 * The only case where we don't expand "vma" 766 * and we expand "next" instead is case 8. 767 */ 768 VM_WARN_ON(end != next->vm_end); 769 /* 770 * remove_next == 3 means we're 771 * removing "vma" and that to do so we 772 * swapped "vma" and "next". 773 */ 774 remove_next = 3; 775 VM_WARN_ON(file != next->vm_file); 776 swap(vma, next); 777 } else { 778 VM_WARN_ON(expand != vma); 779 /* 780 * case 1, 6, 7, remove_next == 2 is case 6, 781 * remove_next == 1 is case 1 or 7. 782 */ 783 remove_next = 1 + (end > next->vm_end); 784 VM_WARN_ON(remove_next == 2 && 785 end != next->vm_next->vm_end); 786 /* trim end to next, for case 6 first pass */ 787 end = next->vm_end; 788 } 789 790 exporter = next; 791 importer = vma; 792 793 /* 794 * If next doesn't have anon_vma, import from vma after 795 * next, if the vma overlaps with it. 796 */ 797 if (remove_next == 2 && !next->anon_vma) 798 exporter = next->vm_next; 799 800 } else if (end > next->vm_start) { 801 /* 802 * vma expands, overlapping part of the next: 803 * mprotect case 5 shifting the boundary up. 804 */ 805 adjust_next = (end - next->vm_start); 806 exporter = next; 807 importer = vma; 808 VM_WARN_ON(expand != importer); 809 } else if (end < vma->vm_end) { 810 /* 811 * vma shrinks, and !insert tells it's not 812 * split_vma inserting another: so it must be 813 * mprotect case 4 shifting the boundary down. 814 */ 815 adjust_next = -(vma->vm_end - end); 816 exporter = vma; 817 importer = next; 818 VM_WARN_ON(expand != importer); 819 } 820 821 /* 822 * Easily overlooked: when mprotect shifts the boundary, 823 * make sure the expanding vma has anon_vma set if the 824 * shrinking vma had, to cover any anon pages imported. 825 */ 826 if (exporter && exporter->anon_vma && !importer->anon_vma) { 827 int error; 828 829 importer->anon_vma = exporter->anon_vma; 830 error = anon_vma_clone(importer, exporter); 831 if (error) 832 return error; 833 } 834 } 835 again: 836 vma_adjust_trans_huge(orig_vma, start, end, adjust_next); 837 838 if (file) { 839 mapping = file->f_mapping; 840 root = &mapping->i_mmap; 841 uprobe_munmap(vma, vma->vm_start, vma->vm_end); 842 843 if (adjust_next) 844 uprobe_munmap(next, next->vm_start, next->vm_end); 845 846 i_mmap_lock_write(mapping); 847 if (insert) { 848 /* 849 * Put into interval tree now, so instantiated pages 850 * are visible to arm/parisc __flush_dcache_page 851 * throughout; but we cannot insert into address 852 * space until vma start or end is updated. 853 */ 854 __vma_link_file(insert); 855 } 856 } 857 858 anon_vma = vma->anon_vma; 859 if (!anon_vma && adjust_next) 860 anon_vma = next->anon_vma; 861 if (anon_vma) { 862 VM_WARN_ON(adjust_next && next->anon_vma && 863 anon_vma != next->anon_vma); 864 anon_vma_lock_write(anon_vma); 865 anon_vma_interval_tree_pre_update_vma(vma); 866 if (adjust_next) 867 anon_vma_interval_tree_pre_update_vma(next); 868 } 869 870 if (file) { 871 flush_dcache_mmap_lock(mapping); 872 vma_interval_tree_remove(vma, root); 873 if (adjust_next) 874 vma_interval_tree_remove(next, root); 875 } 876 877 if (start != vma->vm_start) { 878 vma->vm_start = start; 879 start_changed = true; 880 } 881 if (end != vma->vm_end) { 882 vma->vm_end = end; 883 end_changed = true; 884 } 885 vma->vm_pgoff = pgoff; 886 if (adjust_next) { 887 next->vm_start += adjust_next; 888 next->vm_pgoff += adjust_next >> PAGE_SHIFT; 889 } 890 891 if (file) { 892 if (adjust_next) 893 vma_interval_tree_insert(next, root); 894 vma_interval_tree_insert(vma, root); 895 flush_dcache_mmap_unlock(mapping); 896 } 897 898 if (remove_next) { 899 /* 900 * vma_merge has merged next into vma, and needs 901 * us to remove next before dropping the locks. 902 */ 903 if (remove_next != 3) 904 __vma_unlink(mm, next, next); 905 else 906 /* 907 * vma is not before next if they've been 908 * swapped. 909 * 910 * pre-swap() next->vm_start was reduced so 911 * tell validate_mm_rb to ignore pre-swap() 912 * "next" (which is stored in post-swap() 913 * "vma"). 914 */ 915 __vma_unlink(mm, next, vma); 916 if (file) 917 __remove_shared_vm_struct(next, file, mapping); 918 } else if (insert) { 919 /* 920 * split_vma has split insert from vma, and needs 921 * us to insert it before dropping the locks 922 * (it may either follow vma or precede it). 923 */ 924 __insert_vm_struct(mm, insert); 925 } else { 926 if (start_changed) 927 vma_gap_update(vma); 928 if (end_changed) { 929 if (!next) 930 mm->highest_vm_end = vm_end_gap(vma); 931 else if (!adjust_next) 932 vma_gap_update(next); 933 } 934 } 935 936 if (anon_vma) { 937 anon_vma_interval_tree_post_update_vma(vma); 938 if (adjust_next) 939 anon_vma_interval_tree_post_update_vma(next); 940 anon_vma_unlock_write(anon_vma); 941 } 942 943 if (file) { 944 i_mmap_unlock_write(mapping); 945 uprobe_mmap(vma); 946 947 if (adjust_next) 948 uprobe_mmap(next); 949 } 950 951 if (remove_next) { 952 if (file) { 953 uprobe_munmap(next, next->vm_start, next->vm_end); 954 fput(file); 955 } 956 if (next->anon_vma) 957 anon_vma_merge(vma, next); 958 mm->map_count--; 959 mpol_put(vma_policy(next)); 960 vm_area_free(next); 961 /* 962 * In mprotect's case 6 (see comments on vma_merge), 963 * we must remove another next too. It would clutter 964 * up the code too much to do both in one go. 965 */ 966 if (remove_next != 3) { 967 /* 968 * If "next" was removed and vma->vm_end was 969 * expanded (up) over it, in turn 970 * "next->vm_prev->vm_end" changed and the 971 * "vma->vm_next" gap must be updated. 972 */ 973 next = vma->vm_next; 974 } else { 975 /* 976 * For the scope of the comment "next" and 977 * "vma" considered pre-swap(): if "vma" was 978 * removed, next->vm_start was expanded (down) 979 * over it and the "next" gap must be updated. 980 * Because of the swap() the post-swap() "vma" 981 * actually points to pre-swap() "next" 982 * (post-swap() "next" as opposed is now a 983 * dangling pointer). 984 */ 985 next = vma; 986 } 987 if (remove_next == 2) { 988 remove_next = 1; 989 end = next->vm_end; 990 goto again; 991 } 992 else if (next) 993 vma_gap_update(next); 994 else { 995 /* 996 * If remove_next == 2 we obviously can't 997 * reach this path. 998 * 999 * If remove_next == 3 we can't reach this 1000 * path because pre-swap() next is always not 1001 * NULL. pre-swap() "next" is not being 1002 * removed and its next->vm_end is not altered 1003 * (and furthermore "end" already matches 1004 * next->vm_end in remove_next == 3). 1005 * 1006 * We reach this only in the remove_next == 1 1007 * case if the "next" vma that was removed was 1008 * the highest vma of the mm. However in such 1009 * case next->vm_end == "end" and the extended 1010 * "vma" has vma->vm_end == next->vm_end so 1011 * mm->highest_vm_end doesn't need any update 1012 * in remove_next == 1 case. 1013 */ 1014 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma)); 1015 } 1016 } 1017 if (insert && file) 1018 uprobe_mmap(insert); 1019 1020 validate_mm(mm); 1021 1022 return 0; 1023 } 1024 1025 /* 1026 * If the vma has a ->close operation then the driver probably needs to release 1027 * per-vma resources, so we don't attempt to merge those. 1028 */ 1029 static inline int is_mergeable_vma(struct vm_area_struct *vma, 1030 struct file *file, unsigned long vm_flags, 1031 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1032 { 1033 /* 1034 * VM_SOFTDIRTY should not prevent from VMA merging, if we 1035 * match the flags but dirty bit -- the caller should mark 1036 * merged VMA as dirty. If dirty bit won't be excluded from 1037 * comparison, we increase pressure on the memory system forcing 1038 * the kernel to generate new VMAs when old one could be 1039 * extended instead. 1040 */ 1041 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 1042 return 0; 1043 if (vma->vm_file != file) 1044 return 0; 1045 if (vma->vm_ops && vma->vm_ops->close) 1046 return 0; 1047 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx)) 1048 return 0; 1049 return 1; 1050 } 1051 1052 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 1053 struct anon_vma *anon_vma2, 1054 struct vm_area_struct *vma) 1055 { 1056 /* 1057 * The list_is_singular() test is to avoid merging VMA cloned from 1058 * parents. This can improve scalability caused by anon_vma lock. 1059 */ 1060 if ((!anon_vma1 || !anon_vma2) && (!vma || 1061 list_is_singular(&vma->anon_vma_chain))) 1062 return 1; 1063 return anon_vma1 == anon_vma2; 1064 } 1065 1066 /* 1067 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 1068 * in front of (at a lower virtual address and file offset than) the vma. 1069 * 1070 * We cannot merge two vmas if they have differently assigned (non-NULL) 1071 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 1072 * 1073 * We don't check here for the merged mmap wrapping around the end of pagecache 1074 * indices (16TB on ia32) because do_mmap() does not permit mmap's which 1075 * wrap, nor mmaps which cover the final page at index -1UL. 1076 */ 1077 static int 1078 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 1079 struct anon_vma *anon_vma, struct file *file, 1080 pgoff_t vm_pgoff, 1081 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1082 { 1083 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) && 1084 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1085 if (vma->vm_pgoff == vm_pgoff) 1086 return 1; 1087 } 1088 return 0; 1089 } 1090 1091 /* 1092 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 1093 * beyond (at a higher virtual address and file offset than) the vma. 1094 * 1095 * We cannot merge two vmas if they have differently assigned (non-NULL) 1096 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 1097 */ 1098 static int 1099 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 1100 struct anon_vma *anon_vma, struct file *file, 1101 pgoff_t vm_pgoff, 1102 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1103 { 1104 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) && 1105 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1106 pgoff_t vm_pglen; 1107 vm_pglen = vma_pages(vma); 1108 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 1109 return 1; 1110 } 1111 return 0; 1112 } 1113 1114 /* 1115 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 1116 * whether that can be merged with its predecessor or its successor. 1117 * Or both (it neatly fills a hole). 1118 * 1119 * In most cases - when called for mmap, brk or mremap - [addr,end) is 1120 * certain not to be mapped by the time vma_merge is called; but when 1121 * called for mprotect, it is certain to be already mapped (either at 1122 * an offset within prev, or at the start of next), and the flags of 1123 * this area are about to be changed to vm_flags - and the no-change 1124 * case has already been eliminated. 1125 * 1126 * The following mprotect cases have to be considered, where AAAA is 1127 * the area passed down from mprotect_fixup, never extending beyond one 1128 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 1129 * 1130 * AAAA AAAA AAAA 1131 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN 1132 * cannot merge might become might become 1133 * PPNNNNNNNNNN PPPPPPPPPPNN 1134 * mmap, brk or case 4 below case 5 below 1135 * mremap move: 1136 * AAAA AAAA 1137 * PPPP NNNN PPPPNNNNXXXX 1138 * might become might become 1139 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or 1140 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or 1141 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8 1142 * 1143 * It is important for case 8 that the vma NNNN overlapping the 1144 * region AAAA is never going to extended over XXXX. Instead XXXX must 1145 * be extended in region AAAA and NNNN must be removed. This way in 1146 * all cases where vma_merge succeeds, the moment vma_adjust drops the 1147 * rmap_locks, the properties of the merged vma will be already 1148 * correct for the whole merged range. Some of those properties like 1149 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must 1150 * be correct for the whole merged range immediately after the 1151 * rmap_locks are released. Otherwise if XXXX would be removed and 1152 * NNNN would be extended over the XXXX range, remove_migration_ptes 1153 * or other rmap walkers (if working on addresses beyond the "end" 1154 * parameter) may establish ptes with the wrong permissions of NNNN 1155 * instead of the right permissions of XXXX. 1156 */ 1157 struct vm_area_struct *vma_merge(struct mm_struct *mm, 1158 struct vm_area_struct *prev, unsigned long addr, 1159 unsigned long end, unsigned long vm_flags, 1160 struct anon_vma *anon_vma, struct file *file, 1161 pgoff_t pgoff, struct mempolicy *policy, 1162 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1163 { 1164 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 1165 struct vm_area_struct *area, *next; 1166 int err; 1167 1168 /* 1169 * We later require that vma->vm_flags == vm_flags, 1170 * so this tests vma->vm_flags & VM_SPECIAL, too. 1171 */ 1172 if (vm_flags & VM_SPECIAL) 1173 return NULL; 1174 1175 next = vma_next(mm, prev); 1176 area = next; 1177 if (area && area->vm_end == end) /* cases 6, 7, 8 */ 1178 next = next->vm_next; 1179 1180 /* verify some invariant that must be enforced by the caller */ 1181 VM_WARN_ON(prev && addr <= prev->vm_start); 1182 VM_WARN_ON(area && end > area->vm_end); 1183 VM_WARN_ON(addr >= end); 1184 1185 /* 1186 * Can it merge with the predecessor? 1187 */ 1188 if (prev && prev->vm_end == addr && 1189 mpol_equal(vma_policy(prev), policy) && 1190 can_vma_merge_after(prev, vm_flags, 1191 anon_vma, file, pgoff, 1192 vm_userfaultfd_ctx)) { 1193 /* 1194 * OK, it can. Can we now merge in the successor as well? 1195 */ 1196 if (next && end == next->vm_start && 1197 mpol_equal(policy, vma_policy(next)) && 1198 can_vma_merge_before(next, vm_flags, 1199 anon_vma, file, 1200 pgoff+pglen, 1201 vm_userfaultfd_ctx) && 1202 is_mergeable_anon_vma(prev->anon_vma, 1203 next->anon_vma, NULL)) { 1204 /* cases 1, 6 */ 1205 err = __vma_adjust(prev, prev->vm_start, 1206 next->vm_end, prev->vm_pgoff, NULL, 1207 prev); 1208 } else /* cases 2, 5, 7 */ 1209 err = __vma_adjust(prev, prev->vm_start, 1210 end, prev->vm_pgoff, NULL, prev); 1211 if (err) 1212 return NULL; 1213 khugepaged_enter_vma_merge(prev, vm_flags); 1214 return prev; 1215 } 1216 1217 /* 1218 * Can this new request be merged in front of next? 1219 */ 1220 if (next && end == next->vm_start && 1221 mpol_equal(policy, vma_policy(next)) && 1222 can_vma_merge_before(next, vm_flags, 1223 anon_vma, file, pgoff+pglen, 1224 vm_userfaultfd_ctx)) { 1225 if (prev && addr < prev->vm_end) /* case 4 */ 1226 err = __vma_adjust(prev, prev->vm_start, 1227 addr, prev->vm_pgoff, NULL, next); 1228 else { /* cases 3, 8 */ 1229 err = __vma_adjust(area, addr, next->vm_end, 1230 next->vm_pgoff - pglen, NULL, next); 1231 /* 1232 * In case 3 area is already equal to next and 1233 * this is a noop, but in case 8 "area" has 1234 * been removed and next was expanded over it. 1235 */ 1236 area = next; 1237 } 1238 if (err) 1239 return NULL; 1240 khugepaged_enter_vma_merge(area, vm_flags); 1241 return area; 1242 } 1243 1244 return NULL; 1245 } 1246 1247 /* 1248 * Rough compatibility check to quickly see if it's even worth looking 1249 * at sharing an anon_vma. 1250 * 1251 * They need to have the same vm_file, and the flags can only differ 1252 * in things that mprotect may change. 1253 * 1254 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1255 * we can merge the two vma's. For example, we refuse to merge a vma if 1256 * there is a vm_ops->close() function, because that indicates that the 1257 * driver is doing some kind of reference counting. But that doesn't 1258 * really matter for the anon_vma sharing case. 1259 */ 1260 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1261 { 1262 return a->vm_end == b->vm_start && 1263 mpol_equal(vma_policy(a), vma_policy(b)) && 1264 a->vm_file == b->vm_file && 1265 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) && 1266 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1267 } 1268 1269 /* 1270 * Do some basic sanity checking to see if we can re-use the anon_vma 1271 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1272 * the same as 'old', the other will be the new one that is trying 1273 * to share the anon_vma. 1274 * 1275 * NOTE! This runs with mm_sem held for reading, so it is possible that 1276 * the anon_vma of 'old' is concurrently in the process of being set up 1277 * by another page fault trying to merge _that_. But that's ok: if it 1278 * is being set up, that automatically means that it will be a singleton 1279 * acceptable for merging, so we can do all of this optimistically. But 1280 * we do that READ_ONCE() to make sure that we never re-load the pointer. 1281 * 1282 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1283 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1284 * is to return an anon_vma that is "complex" due to having gone through 1285 * a fork). 1286 * 1287 * We also make sure that the two vma's are compatible (adjacent, 1288 * and with the same memory policies). That's all stable, even with just 1289 * a read lock on the mm_sem. 1290 */ 1291 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1292 { 1293 if (anon_vma_compatible(a, b)) { 1294 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma); 1295 1296 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1297 return anon_vma; 1298 } 1299 return NULL; 1300 } 1301 1302 /* 1303 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1304 * neighbouring vmas for a suitable anon_vma, before it goes off 1305 * to allocate a new anon_vma. It checks because a repetitive 1306 * sequence of mprotects and faults may otherwise lead to distinct 1307 * anon_vmas being allocated, preventing vma merge in subsequent 1308 * mprotect. 1309 */ 1310 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1311 { 1312 struct anon_vma *anon_vma = NULL; 1313 1314 /* Try next first. */ 1315 if (vma->vm_next) { 1316 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next); 1317 if (anon_vma) 1318 return anon_vma; 1319 } 1320 1321 /* Try prev next. */ 1322 if (vma->vm_prev) 1323 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma); 1324 1325 /* 1326 * We might reach here with anon_vma == NULL if we can't find 1327 * any reusable anon_vma. 1328 * There's no absolute need to look only at touching neighbours: 1329 * we could search further afield for "compatible" anon_vmas. 1330 * But it would probably just be a waste of time searching, 1331 * or lead to too many vmas hanging off the same anon_vma. 1332 * We're trying to allow mprotect remerging later on, 1333 * not trying to minimize memory used for anon_vmas. 1334 */ 1335 return anon_vma; 1336 } 1337 1338 /* 1339 * If a hint addr is less than mmap_min_addr change hint to be as 1340 * low as possible but still greater than mmap_min_addr 1341 */ 1342 static inline unsigned long round_hint_to_min(unsigned long hint) 1343 { 1344 hint &= PAGE_MASK; 1345 if (((void *)hint != NULL) && 1346 (hint < mmap_min_addr)) 1347 return PAGE_ALIGN(mmap_min_addr); 1348 return hint; 1349 } 1350 1351 static inline int mlock_future_check(struct mm_struct *mm, 1352 unsigned long flags, 1353 unsigned long len) 1354 { 1355 unsigned long locked, lock_limit; 1356 1357 /* mlock MCL_FUTURE? */ 1358 if (flags & VM_LOCKED) { 1359 locked = len >> PAGE_SHIFT; 1360 locked += mm->locked_vm; 1361 lock_limit = rlimit(RLIMIT_MEMLOCK); 1362 lock_limit >>= PAGE_SHIFT; 1363 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1364 return -EAGAIN; 1365 } 1366 return 0; 1367 } 1368 1369 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode) 1370 { 1371 if (S_ISREG(inode->i_mode)) 1372 return MAX_LFS_FILESIZE; 1373 1374 if (S_ISBLK(inode->i_mode)) 1375 return MAX_LFS_FILESIZE; 1376 1377 if (S_ISSOCK(inode->i_mode)) 1378 return MAX_LFS_FILESIZE; 1379 1380 /* Special "we do even unsigned file positions" case */ 1381 if (file->f_mode & FMODE_UNSIGNED_OFFSET) 1382 return 0; 1383 1384 /* Yes, random drivers might want more. But I'm tired of buggy drivers */ 1385 return ULONG_MAX; 1386 } 1387 1388 static inline bool file_mmap_ok(struct file *file, struct inode *inode, 1389 unsigned long pgoff, unsigned long len) 1390 { 1391 u64 maxsize = file_mmap_size_max(file, inode); 1392 1393 if (maxsize && len > maxsize) 1394 return false; 1395 maxsize -= len; 1396 if (pgoff > maxsize >> PAGE_SHIFT) 1397 return false; 1398 return true; 1399 } 1400 1401 /* 1402 * The caller must write-lock current->mm->mmap_lock. 1403 */ 1404 unsigned long do_mmap(struct file *file, unsigned long addr, 1405 unsigned long len, unsigned long prot, 1406 unsigned long flags, unsigned long pgoff, 1407 unsigned long *populate, struct list_head *uf) 1408 { 1409 struct mm_struct *mm = current->mm; 1410 vm_flags_t vm_flags; 1411 int pkey = 0; 1412 1413 *populate = 0; 1414 1415 if (!len) 1416 return -EINVAL; 1417 1418 /* 1419 * Does the application expect PROT_READ to imply PROT_EXEC? 1420 * 1421 * (the exception is when the underlying filesystem is noexec 1422 * mounted, in which case we dont add PROT_EXEC.) 1423 */ 1424 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1425 if (!(file && path_noexec(&file->f_path))) 1426 prot |= PROT_EXEC; 1427 1428 /* force arch specific MAP_FIXED handling in get_unmapped_area */ 1429 if (flags & MAP_FIXED_NOREPLACE) 1430 flags |= MAP_FIXED; 1431 1432 if (!(flags & MAP_FIXED)) 1433 addr = round_hint_to_min(addr); 1434 1435 /* Careful about overflows.. */ 1436 len = PAGE_ALIGN(len); 1437 if (!len) 1438 return -ENOMEM; 1439 1440 /* offset overflow? */ 1441 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1442 return -EOVERFLOW; 1443 1444 /* Too many mappings? */ 1445 if (mm->map_count > sysctl_max_map_count) 1446 return -ENOMEM; 1447 1448 /* Obtain the address to map to. we verify (or select) it and ensure 1449 * that it represents a valid section of the address space. 1450 */ 1451 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1452 if (IS_ERR_VALUE(addr)) 1453 return addr; 1454 1455 if (flags & MAP_FIXED_NOREPLACE) { 1456 struct vm_area_struct *vma = find_vma(mm, addr); 1457 1458 if (vma && vma->vm_start < addr + len) 1459 return -EEXIST; 1460 } 1461 1462 if (prot == PROT_EXEC) { 1463 pkey = execute_only_pkey(mm); 1464 if (pkey < 0) 1465 pkey = 0; 1466 } 1467 1468 /* Do simple checking here so the lower-level routines won't have 1469 * to. we assume access permissions have been handled by the open 1470 * of the memory object, so we don't do any here. 1471 */ 1472 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) | 1473 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1474 1475 if (flags & MAP_LOCKED) 1476 if (!can_do_mlock()) 1477 return -EPERM; 1478 1479 if (mlock_future_check(mm, vm_flags, len)) 1480 return -EAGAIN; 1481 1482 if (file) { 1483 struct inode *inode = file_inode(file); 1484 unsigned long flags_mask; 1485 1486 if (!file_mmap_ok(file, inode, pgoff, len)) 1487 return -EOVERFLOW; 1488 1489 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags; 1490 1491 switch (flags & MAP_TYPE) { 1492 case MAP_SHARED: 1493 /* 1494 * Force use of MAP_SHARED_VALIDATE with non-legacy 1495 * flags. E.g. MAP_SYNC is dangerous to use with 1496 * MAP_SHARED as you don't know which consistency model 1497 * you will get. We silently ignore unsupported flags 1498 * with MAP_SHARED to preserve backward compatibility. 1499 */ 1500 flags &= LEGACY_MAP_MASK; 1501 fallthrough; 1502 case MAP_SHARED_VALIDATE: 1503 if (flags & ~flags_mask) 1504 return -EOPNOTSUPP; 1505 if (prot & PROT_WRITE) { 1506 if (!(file->f_mode & FMODE_WRITE)) 1507 return -EACCES; 1508 if (IS_SWAPFILE(file->f_mapping->host)) 1509 return -ETXTBSY; 1510 } 1511 1512 /* 1513 * Make sure we don't allow writing to an append-only 1514 * file.. 1515 */ 1516 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1517 return -EACCES; 1518 1519 /* 1520 * Make sure there are no mandatory locks on the file. 1521 */ 1522 if (locks_verify_locked(file)) 1523 return -EAGAIN; 1524 1525 vm_flags |= VM_SHARED | VM_MAYSHARE; 1526 if (!(file->f_mode & FMODE_WRITE)) 1527 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1528 fallthrough; 1529 case MAP_PRIVATE: 1530 if (!(file->f_mode & FMODE_READ)) 1531 return -EACCES; 1532 if (path_noexec(&file->f_path)) { 1533 if (vm_flags & VM_EXEC) 1534 return -EPERM; 1535 vm_flags &= ~VM_MAYEXEC; 1536 } 1537 1538 if (!file->f_op->mmap) 1539 return -ENODEV; 1540 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1541 return -EINVAL; 1542 break; 1543 1544 default: 1545 return -EINVAL; 1546 } 1547 } else { 1548 switch (flags & MAP_TYPE) { 1549 case MAP_SHARED: 1550 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1551 return -EINVAL; 1552 /* 1553 * Ignore pgoff. 1554 */ 1555 pgoff = 0; 1556 vm_flags |= VM_SHARED | VM_MAYSHARE; 1557 break; 1558 case MAP_PRIVATE: 1559 /* 1560 * Set pgoff according to addr for anon_vma. 1561 */ 1562 pgoff = addr >> PAGE_SHIFT; 1563 break; 1564 default: 1565 return -EINVAL; 1566 } 1567 } 1568 1569 /* 1570 * Set 'VM_NORESERVE' if we should not account for the 1571 * memory use of this mapping. 1572 */ 1573 if (flags & MAP_NORESERVE) { 1574 /* We honor MAP_NORESERVE if allowed to overcommit */ 1575 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1576 vm_flags |= VM_NORESERVE; 1577 1578 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1579 if (file && is_file_hugepages(file)) 1580 vm_flags |= VM_NORESERVE; 1581 } 1582 1583 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf); 1584 if (!IS_ERR_VALUE(addr) && 1585 ((vm_flags & VM_LOCKED) || 1586 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1587 *populate = len; 1588 return addr; 1589 } 1590 1591 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len, 1592 unsigned long prot, unsigned long flags, 1593 unsigned long fd, unsigned long pgoff) 1594 { 1595 struct file *file = NULL; 1596 unsigned long retval; 1597 1598 if (!(flags & MAP_ANONYMOUS)) { 1599 audit_mmap_fd(fd, flags); 1600 file = fget(fd); 1601 if (!file) 1602 return -EBADF; 1603 if (is_file_hugepages(file)) { 1604 len = ALIGN(len, huge_page_size(hstate_file(file))); 1605 } else if (unlikely(flags & MAP_HUGETLB)) { 1606 retval = -EINVAL; 1607 goto out_fput; 1608 } 1609 } else if (flags & MAP_HUGETLB) { 1610 struct user_struct *user = NULL; 1611 struct hstate *hs; 1612 1613 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1614 if (!hs) 1615 return -EINVAL; 1616 1617 len = ALIGN(len, huge_page_size(hs)); 1618 /* 1619 * VM_NORESERVE is used because the reservations will be 1620 * taken when vm_ops->mmap() is called 1621 * A dummy user value is used because we are not locking 1622 * memory so no accounting is necessary 1623 */ 1624 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1625 VM_NORESERVE, 1626 &user, HUGETLB_ANONHUGE_INODE, 1627 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1628 if (IS_ERR(file)) 1629 return PTR_ERR(file); 1630 } 1631 1632 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1633 1634 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1635 out_fput: 1636 if (file) 1637 fput(file); 1638 return retval; 1639 } 1640 1641 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1642 unsigned long, prot, unsigned long, flags, 1643 unsigned long, fd, unsigned long, pgoff) 1644 { 1645 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff); 1646 } 1647 1648 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1649 struct mmap_arg_struct { 1650 unsigned long addr; 1651 unsigned long len; 1652 unsigned long prot; 1653 unsigned long flags; 1654 unsigned long fd; 1655 unsigned long offset; 1656 }; 1657 1658 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1659 { 1660 struct mmap_arg_struct a; 1661 1662 if (copy_from_user(&a, arg, sizeof(a))) 1663 return -EFAULT; 1664 if (offset_in_page(a.offset)) 1665 return -EINVAL; 1666 1667 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1668 a.offset >> PAGE_SHIFT); 1669 } 1670 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1671 1672 /* 1673 * Some shared mappings will want the pages marked read-only 1674 * to track write events. If so, we'll downgrade vm_page_prot 1675 * to the private version (using protection_map[] without the 1676 * VM_SHARED bit). 1677 */ 1678 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot) 1679 { 1680 vm_flags_t vm_flags = vma->vm_flags; 1681 const struct vm_operations_struct *vm_ops = vma->vm_ops; 1682 1683 /* If it was private or non-writable, the write bit is already clear */ 1684 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1685 return 0; 1686 1687 /* The backer wishes to know when pages are first written to? */ 1688 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite)) 1689 return 1; 1690 1691 /* The open routine did something to the protections that pgprot_modify 1692 * won't preserve? */ 1693 if (pgprot_val(vm_page_prot) != 1694 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags))) 1695 return 0; 1696 1697 /* Do we need to track softdirty? */ 1698 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY)) 1699 return 1; 1700 1701 /* Specialty mapping? */ 1702 if (vm_flags & VM_PFNMAP) 1703 return 0; 1704 1705 /* Can the mapping track the dirty pages? */ 1706 return vma->vm_file && vma->vm_file->f_mapping && 1707 mapping_can_writeback(vma->vm_file->f_mapping); 1708 } 1709 1710 /* 1711 * We account for memory if it's a private writeable mapping, 1712 * not hugepages and VM_NORESERVE wasn't set. 1713 */ 1714 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1715 { 1716 /* 1717 * hugetlb has its own accounting separate from the core VM 1718 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1719 */ 1720 if (file && is_file_hugepages(file)) 1721 return 0; 1722 1723 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1724 } 1725 1726 unsigned long mmap_region(struct file *file, unsigned long addr, 1727 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff, 1728 struct list_head *uf) 1729 { 1730 struct mm_struct *mm = current->mm; 1731 struct vm_area_struct *vma, *prev, *merge; 1732 int error; 1733 struct rb_node **rb_link, *rb_parent; 1734 unsigned long charged = 0; 1735 1736 /* Check against address space limit. */ 1737 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) { 1738 unsigned long nr_pages; 1739 1740 /* 1741 * MAP_FIXED may remove pages of mappings that intersects with 1742 * requested mapping. Account for the pages it would unmap. 1743 */ 1744 nr_pages = count_vma_pages_range(mm, addr, addr + len); 1745 1746 if (!may_expand_vm(mm, vm_flags, 1747 (len >> PAGE_SHIFT) - nr_pages)) 1748 return -ENOMEM; 1749 } 1750 1751 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */ 1752 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf)) 1753 return -ENOMEM; 1754 /* 1755 * Private writable mapping: check memory availability 1756 */ 1757 if (accountable_mapping(file, vm_flags)) { 1758 charged = len >> PAGE_SHIFT; 1759 if (security_vm_enough_memory_mm(mm, charged)) 1760 return -ENOMEM; 1761 vm_flags |= VM_ACCOUNT; 1762 } 1763 1764 /* 1765 * Can we just expand an old mapping? 1766 */ 1767 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, 1768 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX); 1769 if (vma) 1770 goto out; 1771 1772 /* 1773 * Determine the object being mapped and call the appropriate 1774 * specific mapper. the address has already been validated, but 1775 * not unmapped, but the maps are removed from the list. 1776 */ 1777 vma = vm_area_alloc(mm); 1778 if (!vma) { 1779 error = -ENOMEM; 1780 goto unacct_error; 1781 } 1782 1783 vma->vm_start = addr; 1784 vma->vm_end = addr + len; 1785 vma->vm_flags = vm_flags; 1786 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1787 vma->vm_pgoff = pgoff; 1788 1789 if (file) { 1790 if (vm_flags & VM_DENYWRITE) { 1791 error = deny_write_access(file); 1792 if (error) 1793 goto free_vma; 1794 } 1795 if (vm_flags & VM_SHARED) { 1796 error = mapping_map_writable(file->f_mapping); 1797 if (error) 1798 goto allow_write_and_free_vma; 1799 } 1800 1801 /* ->mmap() can change vma->vm_file, but must guarantee that 1802 * vma_link() below can deny write-access if VM_DENYWRITE is set 1803 * and map writably if VM_SHARED is set. This usually means the 1804 * new file must not have been exposed to user-space, yet. 1805 */ 1806 vma->vm_file = get_file(file); 1807 error = call_mmap(file, vma); 1808 if (error) 1809 goto unmap_and_free_vma; 1810 1811 /* Can addr have changed?? 1812 * 1813 * Answer: Yes, several device drivers can do it in their 1814 * f_op->mmap method. -DaveM 1815 * Bug: If addr is changed, prev, rb_link, rb_parent should 1816 * be updated for vma_link() 1817 */ 1818 WARN_ON_ONCE(addr != vma->vm_start); 1819 1820 addr = vma->vm_start; 1821 1822 /* If vm_flags changed after call_mmap(), we should try merge vma again 1823 * as we may succeed this time. 1824 */ 1825 if (unlikely(vm_flags != vma->vm_flags && prev)) { 1826 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags, 1827 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX); 1828 if (merge) { 1829 /* ->mmap() can change vma->vm_file and fput the original file. So 1830 * fput the vma->vm_file here or we would add an extra fput for file 1831 * and cause general protection fault ultimately. 1832 */ 1833 fput(vma->vm_file); 1834 vm_area_free(vma); 1835 vma = merge; 1836 /* Update vm_flags to pick up the change. */ 1837 vm_flags = vma->vm_flags; 1838 goto unmap_writable; 1839 } 1840 } 1841 1842 vm_flags = vma->vm_flags; 1843 } else if (vm_flags & VM_SHARED) { 1844 error = shmem_zero_setup(vma); 1845 if (error) 1846 goto free_vma; 1847 } else { 1848 vma_set_anonymous(vma); 1849 } 1850 1851 /* Allow architectures to sanity-check the vm_flags */ 1852 if (!arch_validate_flags(vma->vm_flags)) { 1853 error = -EINVAL; 1854 if (file) 1855 goto unmap_and_free_vma; 1856 else 1857 goto free_vma; 1858 } 1859 1860 vma_link(mm, vma, prev, rb_link, rb_parent); 1861 /* Once vma denies write, undo our temporary denial count */ 1862 if (file) { 1863 unmap_writable: 1864 if (vm_flags & VM_SHARED) 1865 mapping_unmap_writable(file->f_mapping); 1866 if (vm_flags & VM_DENYWRITE) 1867 allow_write_access(file); 1868 } 1869 file = vma->vm_file; 1870 out: 1871 perf_event_mmap(vma); 1872 1873 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT); 1874 if (vm_flags & VM_LOCKED) { 1875 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) || 1876 is_vm_hugetlb_page(vma) || 1877 vma == get_gate_vma(current->mm)) 1878 vma->vm_flags &= VM_LOCKED_CLEAR_MASK; 1879 else 1880 mm->locked_vm += (len >> PAGE_SHIFT); 1881 } 1882 1883 if (file) 1884 uprobe_mmap(vma); 1885 1886 /* 1887 * New (or expanded) vma always get soft dirty status. 1888 * Otherwise user-space soft-dirty page tracker won't 1889 * be able to distinguish situation when vma area unmapped, 1890 * then new mapped in-place (which must be aimed as 1891 * a completely new data area). 1892 */ 1893 vma->vm_flags |= VM_SOFTDIRTY; 1894 1895 vma_set_page_prot(vma); 1896 1897 return addr; 1898 1899 unmap_and_free_vma: 1900 fput(vma->vm_file); 1901 vma->vm_file = NULL; 1902 1903 /* Undo any partial mapping done by a device driver. */ 1904 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1905 charged = 0; 1906 if (vm_flags & VM_SHARED) 1907 mapping_unmap_writable(file->f_mapping); 1908 allow_write_and_free_vma: 1909 if (vm_flags & VM_DENYWRITE) 1910 allow_write_access(file); 1911 free_vma: 1912 vm_area_free(vma); 1913 unacct_error: 1914 if (charged) 1915 vm_unacct_memory(charged); 1916 return error; 1917 } 1918 1919 static unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1920 { 1921 /* 1922 * We implement the search by looking for an rbtree node that 1923 * immediately follows a suitable gap. That is, 1924 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length; 1925 * - gap_end = vma->vm_start >= info->low_limit + length; 1926 * - gap_end - gap_start >= length 1927 */ 1928 1929 struct mm_struct *mm = current->mm; 1930 struct vm_area_struct *vma; 1931 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1932 1933 /* Adjust search length to account for worst case alignment overhead */ 1934 length = info->length + info->align_mask; 1935 if (length < info->length) 1936 return -ENOMEM; 1937 1938 /* Adjust search limits by the desired length */ 1939 if (info->high_limit < length) 1940 return -ENOMEM; 1941 high_limit = info->high_limit - length; 1942 1943 if (info->low_limit > high_limit) 1944 return -ENOMEM; 1945 low_limit = info->low_limit + length; 1946 1947 /* Check if rbtree root looks promising */ 1948 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1949 goto check_highest; 1950 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1951 if (vma->rb_subtree_gap < length) 1952 goto check_highest; 1953 1954 while (true) { 1955 /* Visit left subtree if it looks promising */ 1956 gap_end = vm_start_gap(vma); 1957 if (gap_end >= low_limit && vma->vm_rb.rb_left) { 1958 struct vm_area_struct *left = 1959 rb_entry(vma->vm_rb.rb_left, 1960 struct vm_area_struct, vm_rb); 1961 if (left->rb_subtree_gap >= length) { 1962 vma = left; 1963 continue; 1964 } 1965 } 1966 1967 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0; 1968 check_current: 1969 /* Check if current node has a suitable gap */ 1970 if (gap_start > high_limit) 1971 return -ENOMEM; 1972 if (gap_end >= low_limit && 1973 gap_end > gap_start && gap_end - gap_start >= length) 1974 goto found; 1975 1976 /* Visit right subtree if it looks promising */ 1977 if (vma->vm_rb.rb_right) { 1978 struct vm_area_struct *right = 1979 rb_entry(vma->vm_rb.rb_right, 1980 struct vm_area_struct, vm_rb); 1981 if (right->rb_subtree_gap >= length) { 1982 vma = right; 1983 continue; 1984 } 1985 } 1986 1987 /* Go back up the rbtree to find next candidate node */ 1988 while (true) { 1989 struct rb_node *prev = &vma->vm_rb; 1990 if (!rb_parent(prev)) 1991 goto check_highest; 1992 vma = rb_entry(rb_parent(prev), 1993 struct vm_area_struct, vm_rb); 1994 if (prev == vma->vm_rb.rb_left) { 1995 gap_start = vm_end_gap(vma->vm_prev); 1996 gap_end = vm_start_gap(vma); 1997 goto check_current; 1998 } 1999 } 2000 } 2001 2002 check_highest: 2003 /* Check highest gap, which does not precede any rbtree node */ 2004 gap_start = mm->highest_vm_end; 2005 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */ 2006 if (gap_start > high_limit) 2007 return -ENOMEM; 2008 2009 found: 2010 /* We found a suitable gap. Clip it with the original low_limit. */ 2011 if (gap_start < info->low_limit) 2012 gap_start = info->low_limit; 2013 2014 /* Adjust gap address to the desired alignment */ 2015 gap_start += (info->align_offset - gap_start) & info->align_mask; 2016 2017 VM_BUG_ON(gap_start + info->length > info->high_limit); 2018 VM_BUG_ON(gap_start + info->length > gap_end); 2019 return gap_start; 2020 } 2021 2022 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 2023 { 2024 struct mm_struct *mm = current->mm; 2025 struct vm_area_struct *vma; 2026 unsigned long length, low_limit, high_limit, gap_start, gap_end; 2027 2028 /* Adjust search length to account for worst case alignment overhead */ 2029 length = info->length + info->align_mask; 2030 if (length < info->length) 2031 return -ENOMEM; 2032 2033 /* 2034 * Adjust search limits by the desired length. 2035 * See implementation comment at top of unmapped_area(). 2036 */ 2037 gap_end = info->high_limit; 2038 if (gap_end < length) 2039 return -ENOMEM; 2040 high_limit = gap_end - length; 2041 2042 if (info->low_limit > high_limit) 2043 return -ENOMEM; 2044 low_limit = info->low_limit + length; 2045 2046 /* Check highest gap, which does not precede any rbtree node */ 2047 gap_start = mm->highest_vm_end; 2048 if (gap_start <= high_limit) 2049 goto found_highest; 2050 2051 /* Check if rbtree root looks promising */ 2052 if (RB_EMPTY_ROOT(&mm->mm_rb)) 2053 return -ENOMEM; 2054 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 2055 if (vma->rb_subtree_gap < length) 2056 return -ENOMEM; 2057 2058 while (true) { 2059 /* Visit right subtree if it looks promising */ 2060 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0; 2061 if (gap_start <= high_limit && vma->vm_rb.rb_right) { 2062 struct vm_area_struct *right = 2063 rb_entry(vma->vm_rb.rb_right, 2064 struct vm_area_struct, vm_rb); 2065 if (right->rb_subtree_gap >= length) { 2066 vma = right; 2067 continue; 2068 } 2069 } 2070 2071 check_current: 2072 /* Check if current node has a suitable gap */ 2073 gap_end = vm_start_gap(vma); 2074 if (gap_end < low_limit) 2075 return -ENOMEM; 2076 if (gap_start <= high_limit && 2077 gap_end > gap_start && gap_end - gap_start >= length) 2078 goto found; 2079 2080 /* Visit left subtree if it looks promising */ 2081 if (vma->vm_rb.rb_left) { 2082 struct vm_area_struct *left = 2083 rb_entry(vma->vm_rb.rb_left, 2084 struct vm_area_struct, vm_rb); 2085 if (left->rb_subtree_gap >= length) { 2086 vma = left; 2087 continue; 2088 } 2089 } 2090 2091 /* Go back up the rbtree to find next candidate node */ 2092 while (true) { 2093 struct rb_node *prev = &vma->vm_rb; 2094 if (!rb_parent(prev)) 2095 return -ENOMEM; 2096 vma = rb_entry(rb_parent(prev), 2097 struct vm_area_struct, vm_rb); 2098 if (prev == vma->vm_rb.rb_right) { 2099 gap_start = vma->vm_prev ? 2100 vm_end_gap(vma->vm_prev) : 0; 2101 goto check_current; 2102 } 2103 } 2104 } 2105 2106 found: 2107 /* We found a suitable gap. Clip it with the original high_limit. */ 2108 if (gap_end > info->high_limit) 2109 gap_end = info->high_limit; 2110 2111 found_highest: 2112 /* Compute highest gap address at the desired alignment */ 2113 gap_end -= info->length; 2114 gap_end -= (gap_end - info->align_offset) & info->align_mask; 2115 2116 VM_BUG_ON(gap_end < info->low_limit); 2117 VM_BUG_ON(gap_end < gap_start); 2118 return gap_end; 2119 } 2120 2121 /* 2122 * Search for an unmapped address range. 2123 * 2124 * We are looking for a range that: 2125 * - does not intersect with any VMA; 2126 * - is contained within the [low_limit, high_limit) interval; 2127 * - is at least the desired size. 2128 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask) 2129 */ 2130 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info) 2131 { 2132 unsigned long addr; 2133 2134 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN) 2135 addr = unmapped_area_topdown(info); 2136 else 2137 addr = unmapped_area(info); 2138 2139 trace_vm_unmapped_area(addr, info); 2140 return addr; 2141 } 2142 2143 #ifndef arch_get_mmap_end 2144 #define arch_get_mmap_end(addr) (TASK_SIZE) 2145 #endif 2146 2147 #ifndef arch_get_mmap_base 2148 #define arch_get_mmap_base(addr, base) (base) 2149 #endif 2150 2151 /* Get an address range which is currently unmapped. 2152 * For shmat() with addr=0. 2153 * 2154 * Ugly calling convention alert: 2155 * Return value with the low bits set means error value, 2156 * ie 2157 * if (ret & ~PAGE_MASK) 2158 * error = ret; 2159 * 2160 * This function "knows" that -ENOMEM has the bits set. 2161 */ 2162 #ifndef HAVE_ARCH_UNMAPPED_AREA 2163 unsigned long 2164 arch_get_unmapped_area(struct file *filp, unsigned long addr, 2165 unsigned long len, unsigned long pgoff, unsigned long flags) 2166 { 2167 struct mm_struct *mm = current->mm; 2168 struct vm_area_struct *vma, *prev; 2169 struct vm_unmapped_area_info info; 2170 const unsigned long mmap_end = arch_get_mmap_end(addr); 2171 2172 if (len > mmap_end - mmap_min_addr) 2173 return -ENOMEM; 2174 2175 if (flags & MAP_FIXED) 2176 return addr; 2177 2178 if (addr) { 2179 addr = PAGE_ALIGN(addr); 2180 vma = find_vma_prev(mm, addr, &prev); 2181 if (mmap_end - len >= addr && addr >= mmap_min_addr && 2182 (!vma || addr + len <= vm_start_gap(vma)) && 2183 (!prev || addr >= vm_end_gap(prev))) 2184 return addr; 2185 } 2186 2187 info.flags = 0; 2188 info.length = len; 2189 info.low_limit = mm->mmap_base; 2190 info.high_limit = mmap_end; 2191 info.align_mask = 0; 2192 info.align_offset = 0; 2193 return vm_unmapped_area(&info); 2194 } 2195 #endif 2196 2197 /* 2198 * This mmap-allocator allocates new areas top-down from below the 2199 * stack's low limit (the base): 2200 */ 2201 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 2202 unsigned long 2203 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr, 2204 unsigned long len, unsigned long pgoff, 2205 unsigned long flags) 2206 { 2207 struct vm_area_struct *vma, *prev; 2208 struct mm_struct *mm = current->mm; 2209 struct vm_unmapped_area_info info; 2210 const unsigned long mmap_end = arch_get_mmap_end(addr); 2211 2212 /* requested length too big for entire address space */ 2213 if (len > mmap_end - mmap_min_addr) 2214 return -ENOMEM; 2215 2216 if (flags & MAP_FIXED) 2217 return addr; 2218 2219 /* requesting a specific address */ 2220 if (addr) { 2221 addr = PAGE_ALIGN(addr); 2222 vma = find_vma_prev(mm, addr, &prev); 2223 if (mmap_end - len >= addr && addr >= mmap_min_addr && 2224 (!vma || addr + len <= vm_start_gap(vma)) && 2225 (!prev || addr >= vm_end_gap(prev))) 2226 return addr; 2227 } 2228 2229 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 2230 info.length = len; 2231 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 2232 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base); 2233 info.align_mask = 0; 2234 info.align_offset = 0; 2235 addr = vm_unmapped_area(&info); 2236 2237 /* 2238 * A failed mmap() very likely causes application failure, 2239 * so fall back to the bottom-up function here. This scenario 2240 * can happen with large stack limits and large mmap() 2241 * allocations. 2242 */ 2243 if (offset_in_page(addr)) { 2244 VM_BUG_ON(addr != -ENOMEM); 2245 info.flags = 0; 2246 info.low_limit = TASK_UNMAPPED_BASE; 2247 info.high_limit = mmap_end; 2248 addr = vm_unmapped_area(&info); 2249 } 2250 2251 return addr; 2252 } 2253 #endif 2254 2255 unsigned long 2256 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 2257 unsigned long pgoff, unsigned long flags) 2258 { 2259 unsigned long (*get_area)(struct file *, unsigned long, 2260 unsigned long, unsigned long, unsigned long); 2261 2262 unsigned long error = arch_mmap_check(addr, len, flags); 2263 if (error) 2264 return error; 2265 2266 /* Careful about overflows.. */ 2267 if (len > TASK_SIZE) 2268 return -ENOMEM; 2269 2270 get_area = current->mm->get_unmapped_area; 2271 if (file) { 2272 if (file->f_op->get_unmapped_area) 2273 get_area = file->f_op->get_unmapped_area; 2274 } else if (flags & MAP_SHARED) { 2275 /* 2276 * mmap_region() will call shmem_zero_setup() to create a file, 2277 * so use shmem's get_unmapped_area in case it can be huge. 2278 * do_mmap() will clear pgoff, so match alignment. 2279 */ 2280 pgoff = 0; 2281 get_area = shmem_get_unmapped_area; 2282 } 2283 2284 addr = get_area(file, addr, len, pgoff, flags); 2285 if (IS_ERR_VALUE(addr)) 2286 return addr; 2287 2288 if (addr > TASK_SIZE - len) 2289 return -ENOMEM; 2290 if (offset_in_page(addr)) 2291 return -EINVAL; 2292 2293 error = security_mmap_addr(addr); 2294 return error ? error : addr; 2295 } 2296 2297 EXPORT_SYMBOL(get_unmapped_area); 2298 2299 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 2300 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 2301 { 2302 struct rb_node *rb_node; 2303 struct vm_area_struct *vma; 2304 2305 /* Check the cache first. */ 2306 vma = vmacache_find(mm, addr); 2307 if (likely(vma)) 2308 return vma; 2309 2310 rb_node = mm->mm_rb.rb_node; 2311 2312 while (rb_node) { 2313 struct vm_area_struct *tmp; 2314 2315 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2316 2317 if (tmp->vm_end > addr) { 2318 vma = tmp; 2319 if (tmp->vm_start <= addr) 2320 break; 2321 rb_node = rb_node->rb_left; 2322 } else 2323 rb_node = rb_node->rb_right; 2324 } 2325 2326 if (vma) 2327 vmacache_update(addr, vma); 2328 return vma; 2329 } 2330 2331 EXPORT_SYMBOL(find_vma); 2332 2333 /* 2334 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 2335 */ 2336 struct vm_area_struct * 2337 find_vma_prev(struct mm_struct *mm, unsigned long addr, 2338 struct vm_area_struct **pprev) 2339 { 2340 struct vm_area_struct *vma; 2341 2342 vma = find_vma(mm, addr); 2343 if (vma) { 2344 *pprev = vma->vm_prev; 2345 } else { 2346 struct rb_node *rb_node = rb_last(&mm->mm_rb); 2347 2348 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL; 2349 } 2350 return vma; 2351 } 2352 2353 /* 2354 * Verify that the stack growth is acceptable and 2355 * update accounting. This is shared with both the 2356 * grow-up and grow-down cases. 2357 */ 2358 static int acct_stack_growth(struct vm_area_struct *vma, 2359 unsigned long size, unsigned long grow) 2360 { 2361 struct mm_struct *mm = vma->vm_mm; 2362 unsigned long new_start; 2363 2364 /* address space limit tests */ 2365 if (!may_expand_vm(mm, vma->vm_flags, grow)) 2366 return -ENOMEM; 2367 2368 /* Stack limit test */ 2369 if (size > rlimit(RLIMIT_STACK)) 2370 return -ENOMEM; 2371 2372 /* mlock limit tests */ 2373 if (vma->vm_flags & VM_LOCKED) { 2374 unsigned long locked; 2375 unsigned long limit; 2376 locked = mm->locked_vm + grow; 2377 limit = rlimit(RLIMIT_MEMLOCK); 2378 limit >>= PAGE_SHIFT; 2379 if (locked > limit && !capable(CAP_IPC_LOCK)) 2380 return -ENOMEM; 2381 } 2382 2383 /* Check to ensure the stack will not grow into a hugetlb-only region */ 2384 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 2385 vma->vm_end - size; 2386 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 2387 return -EFAULT; 2388 2389 /* 2390 * Overcommit.. This must be the final test, as it will 2391 * update security statistics. 2392 */ 2393 if (security_vm_enough_memory_mm(mm, grow)) 2394 return -ENOMEM; 2395 2396 return 0; 2397 } 2398 2399 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 2400 /* 2401 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 2402 * vma is the last one with address > vma->vm_end. Have to extend vma. 2403 */ 2404 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 2405 { 2406 struct mm_struct *mm = vma->vm_mm; 2407 struct vm_area_struct *next; 2408 unsigned long gap_addr; 2409 int error = 0; 2410 2411 if (!(vma->vm_flags & VM_GROWSUP)) 2412 return -EFAULT; 2413 2414 /* Guard against exceeding limits of the address space. */ 2415 address &= PAGE_MASK; 2416 if (address >= (TASK_SIZE & PAGE_MASK)) 2417 return -ENOMEM; 2418 address += PAGE_SIZE; 2419 2420 /* Enforce stack_guard_gap */ 2421 gap_addr = address + stack_guard_gap; 2422 2423 /* Guard against overflow */ 2424 if (gap_addr < address || gap_addr > TASK_SIZE) 2425 gap_addr = TASK_SIZE; 2426 2427 next = vma->vm_next; 2428 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) { 2429 if (!(next->vm_flags & VM_GROWSUP)) 2430 return -ENOMEM; 2431 /* Check that both stack segments have the same anon_vma? */ 2432 } 2433 2434 /* We must make sure the anon_vma is allocated. */ 2435 if (unlikely(anon_vma_prepare(vma))) 2436 return -ENOMEM; 2437 2438 /* 2439 * vma->vm_start/vm_end cannot change under us because the caller 2440 * is required to hold the mmap_lock in read mode. We need the 2441 * anon_vma lock to serialize against concurrent expand_stacks. 2442 */ 2443 anon_vma_lock_write(vma->anon_vma); 2444 2445 /* Somebody else might have raced and expanded it already */ 2446 if (address > vma->vm_end) { 2447 unsigned long size, grow; 2448 2449 size = address - vma->vm_start; 2450 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2451 2452 error = -ENOMEM; 2453 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2454 error = acct_stack_growth(vma, size, grow); 2455 if (!error) { 2456 /* 2457 * vma_gap_update() doesn't support concurrent 2458 * updates, but we only hold a shared mmap_lock 2459 * lock here, so we need to protect against 2460 * concurrent vma expansions. 2461 * anon_vma_lock_write() doesn't help here, as 2462 * we don't guarantee that all growable vmas 2463 * in a mm share the same root anon vma. 2464 * So, we reuse mm->page_table_lock to guard 2465 * against concurrent vma expansions. 2466 */ 2467 spin_lock(&mm->page_table_lock); 2468 if (vma->vm_flags & VM_LOCKED) 2469 mm->locked_vm += grow; 2470 vm_stat_account(mm, vma->vm_flags, grow); 2471 anon_vma_interval_tree_pre_update_vma(vma); 2472 vma->vm_end = address; 2473 anon_vma_interval_tree_post_update_vma(vma); 2474 if (vma->vm_next) 2475 vma_gap_update(vma->vm_next); 2476 else 2477 mm->highest_vm_end = vm_end_gap(vma); 2478 spin_unlock(&mm->page_table_lock); 2479 2480 perf_event_mmap(vma); 2481 } 2482 } 2483 } 2484 anon_vma_unlock_write(vma->anon_vma); 2485 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2486 validate_mm(mm); 2487 return error; 2488 } 2489 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2490 2491 /* 2492 * vma is the first one with address < vma->vm_start. Have to extend vma. 2493 */ 2494 int expand_downwards(struct vm_area_struct *vma, 2495 unsigned long address) 2496 { 2497 struct mm_struct *mm = vma->vm_mm; 2498 struct vm_area_struct *prev; 2499 int error = 0; 2500 2501 address &= PAGE_MASK; 2502 if (address < mmap_min_addr) 2503 return -EPERM; 2504 2505 /* Enforce stack_guard_gap */ 2506 prev = vma->vm_prev; 2507 /* Check that both stack segments have the same anon_vma? */ 2508 if (prev && !(prev->vm_flags & VM_GROWSDOWN) && 2509 vma_is_accessible(prev)) { 2510 if (address - prev->vm_end < stack_guard_gap) 2511 return -ENOMEM; 2512 } 2513 2514 /* We must make sure the anon_vma is allocated. */ 2515 if (unlikely(anon_vma_prepare(vma))) 2516 return -ENOMEM; 2517 2518 /* 2519 * vma->vm_start/vm_end cannot change under us because the caller 2520 * is required to hold the mmap_lock in read mode. We need the 2521 * anon_vma lock to serialize against concurrent expand_stacks. 2522 */ 2523 anon_vma_lock_write(vma->anon_vma); 2524 2525 /* Somebody else might have raced and expanded it already */ 2526 if (address < vma->vm_start) { 2527 unsigned long size, grow; 2528 2529 size = vma->vm_end - address; 2530 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2531 2532 error = -ENOMEM; 2533 if (grow <= vma->vm_pgoff) { 2534 error = acct_stack_growth(vma, size, grow); 2535 if (!error) { 2536 /* 2537 * vma_gap_update() doesn't support concurrent 2538 * updates, but we only hold a shared mmap_lock 2539 * lock here, so we need to protect against 2540 * concurrent vma expansions. 2541 * anon_vma_lock_write() doesn't help here, as 2542 * we don't guarantee that all growable vmas 2543 * in a mm share the same root anon vma. 2544 * So, we reuse mm->page_table_lock to guard 2545 * against concurrent vma expansions. 2546 */ 2547 spin_lock(&mm->page_table_lock); 2548 if (vma->vm_flags & VM_LOCKED) 2549 mm->locked_vm += grow; 2550 vm_stat_account(mm, vma->vm_flags, grow); 2551 anon_vma_interval_tree_pre_update_vma(vma); 2552 vma->vm_start = address; 2553 vma->vm_pgoff -= grow; 2554 anon_vma_interval_tree_post_update_vma(vma); 2555 vma_gap_update(vma); 2556 spin_unlock(&mm->page_table_lock); 2557 2558 perf_event_mmap(vma); 2559 } 2560 } 2561 } 2562 anon_vma_unlock_write(vma->anon_vma); 2563 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2564 validate_mm(mm); 2565 return error; 2566 } 2567 2568 /* enforced gap between the expanding stack and other mappings. */ 2569 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT; 2570 2571 static int __init cmdline_parse_stack_guard_gap(char *p) 2572 { 2573 unsigned long val; 2574 char *endptr; 2575 2576 val = simple_strtoul(p, &endptr, 10); 2577 if (!*endptr) 2578 stack_guard_gap = val << PAGE_SHIFT; 2579 2580 return 0; 2581 } 2582 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap); 2583 2584 #ifdef CONFIG_STACK_GROWSUP 2585 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2586 { 2587 return expand_upwards(vma, address); 2588 } 2589 2590 struct vm_area_struct * 2591 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2592 { 2593 struct vm_area_struct *vma, *prev; 2594 2595 addr &= PAGE_MASK; 2596 vma = find_vma_prev(mm, addr, &prev); 2597 if (vma && (vma->vm_start <= addr)) 2598 return vma; 2599 /* don't alter vm_end if the coredump is running */ 2600 if (!prev || expand_stack(prev, addr)) 2601 return NULL; 2602 if (prev->vm_flags & VM_LOCKED) 2603 populate_vma_page_range(prev, addr, prev->vm_end, NULL); 2604 return prev; 2605 } 2606 #else 2607 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2608 { 2609 return expand_downwards(vma, address); 2610 } 2611 2612 struct vm_area_struct * 2613 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2614 { 2615 struct vm_area_struct *vma; 2616 unsigned long start; 2617 2618 addr &= PAGE_MASK; 2619 vma = find_vma(mm, addr); 2620 if (!vma) 2621 return NULL; 2622 if (vma->vm_start <= addr) 2623 return vma; 2624 if (!(vma->vm_flags & VM_GROWSDOWN)) 2625 return NULL; 2626 start = vma->vm_start; 2627 if (expand_stack(vma, addr)) 2628 return NULL; 2629 if (vma->vm_flags & VM_LOCKED) 2630 populate_vma_page_range(vma, addr, start, NULL); 2631 return vma; 2632 } 2633 #endif 2634 2635 EXPORT_SYMBOL_GPL(find_extend_vma); 2636 2637 /* 2638 * Ok - we have the memory areas we should free on the vma list, 2639 * so release them, and do the vma updates. 2640 * 2641 * Called with the mm semaphore held. 2642 */ 2643 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 2644 { 2645 unsigned long nr_accounted = 0; 2646 2647 /* Update high watermark before we lower total_vm */ 2648 update_hiwater_vm(mm); 2649 do { 2650 long nrpages = vma_pages(vma); 2651 2652 if (vma->vm_flags & VM_ACCOUNT) 2653 nr_accounted += nrpages; 2654 vm_stat_account(mm, vma->vm_flags, -nrpages); 2655 vma = remove_vma(vma); 2656 } while (vma); 2657 vm_unacct_memory(nr_accounted); 2658 validate_mm(mm); 2659 } 2660 2661 /* 2662 * Get rid of page table information in the indicated region. 2663 * 2664 * Called with the mm semaphore held. 2665 */ 2666 static void unmap_region(struct mm_struct *mm, 2667 struct vm_area_struct *vma, struct vm_area_struct *prev, 2668 unsigned long start, unsigned long end) 2669 { 2670 struct vm_area_struct *next = vma_next(mm, prev); 2671 struct mmu_gather tlb; 2672 2673 lru_add_drain(); 2674 tlb_gather_mmu(&tlb, mm, start, end); 2675 update_hiwater_rss(mm); 2676 unmap_vmas(&tlb, vma, start, end); 2677 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2678 next ? next->vm_start : USER_PGTABLES_CEILING); 2679 tlb_finish_mmu(&tlb, start, end); 2680 } 2681 2682 /* 2683 * Create a list of vma's touched by the unmap, removing them from the mm's 2684 * vma list as we go.. 2685 */ 2686 static bool 2687 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 2688 struct vm_area_struct *prev, unsigned long end) 2689 { 2690 struct vm_area_struct **insertion_point; 2691 struct vm_area_struct *tail_vma = NULL; 2692 2693 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 2694 vma->vm_prev = NULL; 2695 do { 2696 vma_rb_erase(vma, &mm->mm_rb); 2697 mm->map_count--; 2698 tail_vma = vma; 2699 vma = vma->vm_next; 2700 } while (vma && vma->vm_start < end); 2701 *insertion_point = vma; 2702 if (vma) { 2703 vma->vm_prev = prev; 2704 vma_gap_update(vma); 2705 } else 2706 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0; 2707 tail_vma->vm_next = NULL; 2708 2709 /* Kill the cache */ 2710 vmacache_invalidate(mm); 2711 2712 /* 2713 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or 2714 * VM_GROWSUP VMA. Such VMAs can change their size under 2715 * down_read(mmap_lock) and collide with the VMA we are about to unmap. 2716 */ 2717 if (vma && (vma->vm_flags & VM_GROWSDOWN)) 2718 return false; 2719 if (prev && (prev->vm_flags & VM_GROWSUP)) 2720 return false; 2721 return true; 2722 } 2723 2724 /* 2725 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it 2726 * has already been checked or doesn't make sense to fail. 2727 */ 2728 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2729 unsigned long addr, int new_below) 2730 { 2731 struct vm_area_struct *new; 2732 int err; 2733 2734 if (vma->vm_ops && vma->vm_ops->may_split) { 2735 err = vma->vm_ops->may_split(vma, addr); 2736 if (err) 2737 return err; 2738 } 2739 2740 new = vm_area_dup(vma); 2741 if (!new) 2742 return -ENOMEM; 2743 2744 if (new_below) 2745 new->vm_end = addr; 2746 else { 2747 new->vm_start = addr; 2748 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2749 } 2750 2751 err = vma_dup_policy(vma, new); 2752 if (err) 2753 goto out_free_vma; 2754 2755 err = anon_vma_clone(new, vma); 2756 if (err) 2757 goto out_free_mpol; 2758 2759 if (new->vm_file) 2760 get_file(new->vm_file); 2761 2762 if (new->vm_ops && new->vm_ops->open) 2763 new->vm_ops->open(new); 2764 2765 if (new_below) 2766 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 2767 ((addr - new->vm_start) >> PAGE_SHIFT), new); 2768 else 2769 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 2770 2771 /* Success. */ 2772 if (!err) 2773 return 0; 2774 2775 /* Clean everything up if vma_adjust failed. */ 2776 if (new->vm_ops && new->vm_ops->close) 2777 new->vm_ops->close(new); 2778 if (new->vm_file) 2779 fput(new->vm_file); 2780 unlink_anon_vmas(new); 2781 out_free_mpol: 2782 mpol_put(vma_policy(new)); 2783 out_free_vma: 2784 vm_area_free(new); 2785 return err; 2786 } 2787 2788 /* 2789 * Split a vma into two pieces at address 'addr', a new vma is allocated 2790 * either for the first part or the tail. 2791 */ 2792 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2793 unsigned long addr, int new_below) 2794 { 2795 if (mm->map_count >= sysctl_max_map_count) 2796 return -ENOMEM; 2797 2798 return __split_vma(mm, vma, addr, new_below); 2799 } 2800 2801 /* Munmap is split into 2 main parts -- this part which finds 2802 * what needs doing, and the areas themselves, which do the 2803 * work. This now handles partial unmappings. 2804 * Jeremy Fitzhardinge <jeremy@goop.org> 2805 */ 2806 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len, 2807 struct list_head *uf, bool downgrade) 2808 { 2809 unsigned long end; 2810 struct vm_area_struct *vma, *prev, *last; 2811 2812 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start) 2813 return -EINVAL; 2814 2815 len = PAGE_ALIGN(len); 2816 end = start + len; 2817 if (len == 0) 2818 return -EINVAL; 2819 2820 /* 2821 * arch_unmap() might do unmaps itself. It must be called 2822 * and finish any rbtree manipulation before this code 2823 * runs and also starts to manipulate the rbtree. 2824 */ 2825 arch_unmap(mm, start, end); 2826 2827 /* Find the first overlapping VMA */ 2828 vma = find_vma(mm, start); 2829 if (!vma) 2830 return 0; 2831 prev = vma->vm_prev; 2832 /* we have start < vma->vm_end */ 2833 2834 /* if it doesn't overlap, we have nothing.. */ 2835 if (vma->vm_start >= end) 2836 return 0; 2837 2838 /* 2839 * If we need to split any vma, do it now to save pain later. 2840 * 2841 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2842 * unmapped vm_area_struct will remain in use: so lower split_vma 2843 * places tmp vma above, and higher split_vma places tmp vma below. 2844 */ 2845 if (start > vma->vm_start) { 2846 int error; 2847 2848 /* 2849 * Make sure that map_count on return from munmap() will 2850 * not exceed its limit; but let map_count go just above 2851 * its limit temporarily, to help free resources as expected. 2852 */ 2853 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2854 return -ENOMEM; 2855 2856 error = __split_vma(mm, vma, start, 0); 2857 if (error) 2858 return error; 2859 prev = vma; 2860 } 2861 2862 /* Does it split the last one? */ 2863 last = find_vma(mm, end); 2864 if (last && end > last->vm_start) { 2865 int error = __split_vma(mm, last, end, 1); 2866 if (error) 2867 return error; 2868 } 2869 vma = vma_next(mm, prev); 2870 2871 if (unlikely(uf)) { 2872 /* 2873 * If userfaultfd_unmap_prep returns an error the vmas 2874 * will remain splitted, but userland will get a 2875 * highly unexpected error anyway. This is no 2876 * different than the case where the first of the two 2877 * __split_vma fails, but we don't undo the first 2878 * split, despite we could. This is unlikely enough 2879 * failure that it's not worth optimizing it for. 2880 */ 2881 int error = userfaultfd_unmap_prep(vma, start, end, uf); 2882 if (error) 2883 return error; 2884 } 2885 2886 /* 2887 * unlock any mlock()ed ranges before detaching vmas 2888 */ 2889 if (mm->locked_vm) { 2890 struct vm_area_struct *tmp = vma; 2891 while (tmp && tmp->vm_start < end) { 2892 if (tmp->vm_flags & VM_LOCKED) { 2893 mm->locked_vm -= vma_pages(tmp); 2894 munlock_vma_pages_all(tmp); 2895 } 2896 2897 tmp = tmp->vm_next; 2898 } 2899 } 2900 2901 /* Detach vmas from rbtree */ 2902 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end)) 2903 downgrade = false; 2904 2905 if (downgrade) 2906 mmap_write_downgrade(mm); 2907 2908 unmap_region(mm, vma, prev, start, end); 2909 2910 /* Fix up all other VM information */ 2911 remove_vma_list(mm, vma); 2912 2913 return downgrade ? 1 : 0; 2914 } 2915 2916 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len, 2917 struct list_head *uf) 2918 { 2919 return __do_munmap(mm, start, len, uf, false); 2920 } 2921 2922 static int __vm_munmap(unsigned long start, size_t len, bool downgrade) 2923 { 2924 int ret; 2925 struct mm_struct *mm = current->mm; 2926 LIST_HEAD(uf); 2927 2928 if (mmap_write_lock_killable(mm)) 2929 return -EINTR; 2930 2931 ret = __do_munmap(mm, start, len, &uf, downgrade); 2932 /* 2933 * Returning 1 indicates mmap_lock is downgraded. 2934 * But 1 is not legal return value of vm_munmap() and munmap(), reset 2935 * it to 0 before return. 2936 */ 2937 if (ret == 1) { 2938 mmap_read_unlock(mm); 2939 ret = 0; 2940 } else 2941 mmap_write_unlock(mm); 2942 2943 userfaultfd_unmap_complete(mm, &uf); 2944 return ret; 2945 } 2946 2947 int vm_munmap(unsigned long start, size_t len) 2948 { 2949 return __vm_munmap(start, len, false); 2950 } 2951 EXPORT_SYMBOL(vm_munmap); 2952 2953 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2954 { 2955 addr = untagged_addr(addr); 2956 profile_munmap(addr); 2957 return __vm_munmap(addr, len, true); 2958 } 2959 2960 2961 /* 2962 * Emulation of deprecated remap_file_pages() syscall. 2963 */ 2964 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size, 2965 unsigned long, prot, unsigned long, pgoff, unsigned long, flags) 2966 { 2967 2968 struct mm_struct *mm = current->mm; 2969 struct vm_area_struct *vma; 2970 unsigned long populate = 0; 2971 unsigned long ret = -EINVAL; 2972 struct file *file; 2973 2974 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n", 2975 current->comm, current->pid); 2976 2977 if (prot) 2978 return ret; 2979 start = start & PAGE_MASK; 2980 size = size & PAGE_MASK; 2981 2982 if (start + size <= start) 2983 return ret; 2984 2985 /* Does pgoff wrap? */ 2986 if (pgoff + (size >> PAGE_SHIFT) < pgoff) 2987 return ret; 2988 2989 if (mmap_write_lock_killable(mm)) 2990 return -EINTR; 2991 2992 vma = find_vma(mm, start); 2993 2994 if (!vma || !(vma->vm_flags & VM_SHARED)) 2995 goto out; 2996 2997 if (start < vma->vm_start) 2998 goto out; 2999 3000 if (start + size > vma->vm_end) { 3001 struct vm_area_struct *next; 3002 3003 for (next = vma->vm_next; next; next = next->vm_next) { 3004 /* hole between vmas ? */ 3005 if (next->vm_start != next->vm_prev->vm_end) 3006 goto out; 3007 3008 if (next->vm_file != vma->vm_file) 3009 goto out; 3010 3011 if (next->vm_flags != vma->vm_flags) 3012 goto out; 3013 3014 if (start + size <= next->vm_end) 3015 break; 3016 } 3017 3018 if (!next) 3019 goto out; 3020 } 3021 3022 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0; 3023 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0; 3024 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0; 3025 3026 flags &= MAP_NONBLOCK; 3027 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE; 3028 if (vma->vm_flags & VM_LOCKED) { 3029 struct vm_area_struct *tmp; 3030 flags |= MAP_LOCKED; 3031 3032 /* drop PG_Mlocked flag for over-mapped range */ 3033 for (tmp = vma; tmp->vm_start >= start + size; 3034 tmp = tmp->vm_next) { 3035 /* 3036 * Split pmd and munlock page on the border 3037 * of the range. 3038 */ 3039 vma_adjust_trans_huge(tmp, start, start + size, 0); 3040 3041 munlock_vma_pages_range(tmp, 3042 max(tmp->vm_start, start), 3043 min(tmp->vm_end, start + size)); 3044 } 3045 } 3046 3047 file = get_file(vma->vm_file); 3048 ret = do_mmap(vma->vm_file, start, size, 3049 prot, flags, pgoff, &populate, NULL); 3050 fput(file); 3051 out: 3052 mmap_write_unlock(mm); 3053 if (populate) 3054 mm_populate(ret, populate); 3055 if (!IS_ERR_VALUE(ret)) 3056 ret = 0; 3057 return ret; 3058 } 3059 3060 /* 3061 * this is really a simplified "do_mmap". it only handles 3062 * anonymous maps. eventually we may be able to do some 3063 * brk-specific accounting here. 3064 */ 3065 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf) 3066 { 3067 struct mm_struct *mm = current->mm; 3068 struct vm_area_struct *vma, *prev; 3069 struct rb_node **rb_link, *rb_parent; 3070 pgoff_t pgoff = addr >> PAGE_SHIFT; 3071 int error; 3072 unsigned long mapped_addr; 3073 3074 /* Until we need other flags, refuse anything except VM_EXEC. */ 3075 if ((flags & (~VM_EXEC)) != 0) 3076 return -EINVAL; 3077 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 3078 3079 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 3080 if (IS_ERR_VALUE(mapped_addr)) 3081 return mapped_addr; 3082 3083 error = mlock_future_check(mm, mm->def_flags, len); 3084 if (error) 3085 return error; 3086 3087 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */ 3088 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf)) 3089 return -ENOMEM; 3090 3091 /* Check against address space limits *after* clearing old maps... */ 3092 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT)) 3093 return -ENOMEM; 3094 3095 if (mm->map_count > sysctl_max_map_count) 3096 return -ENOMEM; 3097 3098 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 3099 return -ENOMEM; 3100 3101 /* Can we just expand an old private anonymous mapping? */ 3102 vma = vma_merge(mm, prev, addr, addr + len, flags, 3103 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX); 3104 if (vma) 3105 goto out; 3106 3107 /* 3108 * create a vma struct for an anonymous mapping 3109 */ 3110 vma = vm_area_alloc(mm); 3111 if (!vma) { 3112 vm_unacct_memory(len >> PAGE_SHIFT); 3113 return -ENOMEM; 3114 } 3115 3116 vma_set_anonymous(vma); 3117 vma->vm_start = addr; 3118 vma->vm_end = addr + len; 3119 vma->vm_pgoff = pgoff; 3120 vma->vm_flags = flags; 3121 vma->vm_page_prot = vm_get_page_prot(flags); 3122 vma_link(mm, vma, prev, rb_link, rb_parent); 3123 out: 3124 perf_event_mmap(vma); 3125 mm->total_vm += len >> PAGE_SHIFT; 3126 mm->data_vm += len >> PAGE_SHIFT; 3127 if (flags & VM_LOCKED) 3128 mm->locked_vm += (len >> PAGE_SHIFT); 3129 vma->vm_flags |= VM_SOFTDIRTY; 3130 return 0; 3131 } 3132 3133 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags) 3134 { 3135 struct mm_struct *mm = current->mm; 3136 unsigned long len; 3137 int ret; 3138 bool populate; 3139 LIST_HEAD(uf); 3140 3141 len = PAGE_ALIGN(request); 3142 if (len < request) 3143 return -ENOMEM; 3144 if (!len) 3145 return 0; 3146 3147 if (mmap_write_lock_killable(mm)) 3148 return -EINTR; 3149 3150 ret = do_brk_flags(addr, len, flags, &uf); 3151 populate = ((mm->def_flags & VM_LOCKED) != 0); 3152 mmap_write_unlock(mm); 3153 userfaultfd_unmap_complete(mm, &uf); 3154 if (populate && !ret) 3155 mm_populate(addr, len); 3156 return ret; 3157 } 3158 EXPORT_SYMBOL(vm_brk_flags); 3159 3160 int vm_brk(unsigned long addr, unsigned long len) 3161 { 3162 return vm_brk_flags(addr, len, 0); 3163 } 3164 EXPORT_SYMBOL(vm_brk); 3165 3166 /* Release all mmaps. */ 3167 void exit_mmap(struct mm_struct *mm) 3168 { 3169 struct mmu_gather tlb; 3170 struct vm_area_struct *vma; 3171 unsigned long nr_accounted = 0; 3172 3173 /* mm's last user has gone, and its about to be pulled down */ 3174 mmu_notifier_release(mm); 3175 3176 if (unlikely(mm_is_oom_victim(mm))) { 3177 /* 3178 * Manually reap the mm to free as much memory as possible. 3179 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard 3180 * this mm from further consideration. Taking mm->mmap_lock for 3181 * write after setting MMF_OOM_SKIP will guarantee that the oom 3182 * reaper will not run on this mm again after mmap_lock is 3183 * dropped. 3184 * 3185 * Nothing can be holding mm->mmap_lock here and the above call 3186 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in 3187 * __oom_reap_task_mm() will not block. 3188 * 3189 * This needs to be done before calling munlock_vma_pages_all(), 3190 * which clears VM_LOCKED, otherwise the oom reaper cannot 3191 * reliably test it. 3192 */ 3193 (void)__oom_reap_task_mm(mm); 3194 3195 set_bit(MMF_OOM_SKIP, &mm->flags); 3196 mmap_write_lock(mm); 3197 mmap_write_unlock(mm); 3198 } 3199 3200 if (mm->locked_vm) { 3201 vma = mm->mmap; 3202 while (vma) { 3203 if (vma->vm_flags & VM_LOCKED) 3204 munlock_vma_pages_all(vma); 3205 vma = vma->vm_next; 3206 } 3207 } 3208 3209 arch_exit_mmap(mm); 3210 3211 vma = mm->mmap; 3212 if (!vma) /* Can happen if dup_mmap() received an OOM */ 3213 return; 3214 3215 lru_add_drain(); 3216 flush_cache_mm(mm); 3217 tlb_gather_mmu(&tlb, mm, 0, -1); 3218 /* update_hiwater_rss(mm) here? but nobody should be looking */ 3219 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 3220 unmap_vmas(&tlb, vma, 0, -1); 3221 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING); 3222 tlb_finish_mmu(&tlb, 0, -1); 3223 3224 /* 3225 * Walk the list again, actually closing and freeing it, 3226 * with preemption enabled, without holding any MM locks. 3227 */ 3228 while (vma) { 3229 if (vma->vm_flags & VM_ACCOUNT) 3230 nr_accounted += vma_pages(vma); 3231 vma = remove_vma(vma); 3232 cond_resched(); 3233 } 3234 vm_unacct_memory(nr_accounted); 3235 } 3236 3237 /* Insert vm structure into process list sorted by address 3238 * and into the inode's i_mmap tree. If vm_file is non-NULL 3239 * then i_mmap_rwsem is taken here. 3240 */ 3241 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 3242 { 3243 struct vm_area_struct *prev; 3244 struct rb_node **rb_link, *rb_parent; 3245 3246 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 3247 &prev, &rb_link, &rb_parent)) 3248 return -ENOMEM; 3249 if ((vma->vm_flags & VM_ACCOUNT) && 3250 security_vm_enough_memory_mm(mm, vma_pages(vma))) 3251 return -ENOMEM; 3252 3253 /* 3254 * The vm_pgoff of a purely anonymous vma should be irrelevant 3255 * until its first write fault, when page's anon_vma and index 3256 * are set. But now set the vm_pgoff it will almost certainly 3257 * end up with (unless mremap moves it elsewhere before that 3258 * first wfault), so /proc/pid/maps tells a consistent story. 3259 * 3260 * By setting it to reflect the virtual start address of the 3261 * vma, merges and splits can happen in a seamless way, just 3262 * using the existing file pgoff checks and manipulations. 3263 * Similarly in do_mmap and in do_brk_flags. 3264 */ 3265 if (vma_is_anonymous(vma)) { 3266 BUG_ON(vma->anon_vma); 3267 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 3268 } 3269 3270 vma_link(mm, vma, prev, rb_link, rb_parent); 3271 return 0; 3272 } 3273 3274 /* 3275 * Copy the vma structure to a new location in the same mm, 3276 * prior to moving page table entries, to effect an mremap move. 3277 */ 3278 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 3279 unsigned long addr, unsigned long len, pgoff_t pgoff, 3280 bool *need_rmap_locks) 3281 { 3282 struct vm_area_struct *vma = *vmap; 3283 unsigned long vma_start = vma->vm_start; 3284 struct mm_struct *mm = vma->vm_mm; 3285 struct vm_area_struct *new_vma, *prev; 3286 struct rb_node **rb_link, *rb_parent; 3287 bool faulted_in_anon_vma = true; 3288 3289 /* 3290 * If anonymous vma has not yet been faulted, update new pgoff 3291 * to match new location, to increase its chance of merging. 3292 */ 3293 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) { 3294 pgoff = addr >> PAGE_SHIFT; 3295 faulted_in_anon_vma = false; 3296 } 3297 3298 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) 3299 return NULL; /* should never get here */ 3300 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 3301 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 3302 vma->vm_userfaultfd_ctx); 3303 if (new_vma) { 3304 /* 3305 * Source vma may have been merged into new_vma 3306 */ 3307 if (unlikely(vma_start >= new_vma->vm_start && 3308 vma_start < new_vma->vm_end)) { 3309 /* 3310 * The only way we can get a vma_merge with 3311 * self during an mremap is if the vma hasn't 3312 * been faulted in yet and we were allowed to 3313 * reset the dst vma->vm_pgoff to the 3314 * destination address of the mremap to allow 3315 * the merge to happen. mremap must change the 3316 * vm_pgoff linearity between src and dst vmas 3317 * (in turn preventing a vma_merge) to be 3318 * safe. It is only safe to keep the vm_pgoff 3319 * linear if there are no pages mapped yet. 3320 */ 3321 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 3322 *vmap = vma = new_vma; 3323 } 3324 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 3325 } else { 3326 new_vma = vm_area_dup(vma); 3327 if (!new_vma) 3328 goto out; 3329 new_vma->vm_start = addr; 3330 new_vma->vm_end = addr + len; 3331 new_vma->vm_pgoff = pgoff; 3332 if (vma_dup_policy(vma, new_vma)) 3333 goto out_free_vma; 3334 if (anon_vma_clone(new_vma, vma)) 3335 goto out_free_mempol; 3336 if (new_vma->vm_file) 3337 get_file(new_vma->vm_file); 3338 if (new_vma->vm_ops && new_vma->vm_ops->open) 3339 new_vma->vm_ops->open(new_vma); 3340 vma_link(mm, new_vma, prev, rb_link, rb_parent); 3341 *need_rmap_locks = false; 3342 } 3343 return new_vma; 3344 3345 out_free_mempol: 3346 mpol_put(vma_policy(new_vma)); 3347 out_free_vma: 3348 vm_area_free(new_vma); 3349 out: 3350 return NULL; 3351 } 3352 3353 /* 3354 * Return true if the calling process may expand its vm space by the passed 3355 * number of pages 3356 */ 3357 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages) 3358 { 3359 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT) 3360 return false; 3361 3362 if (is_data_mapping(flags) && 3363 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) { 3364 /* Workaround for Valgrind */ 3365 if (rlimit(RLIMIT_DATA) == 0 && 3366 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT) 3367 return true; 3368 3369 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n", 3370 current->comm, current->pid, 3371 (mm->data_vm + npages) << PAGE_SHIFT, 3372 rlimit(RLIMIT_DATA), 3373 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data"); 3374 3375 if (!ignore_rlimit_data) 3376 return false; 3377 } 3378 3379 return true; 3380 } 3381 3382 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages) 3383 { 3384 mm->total_vm += npages; 3385 3386 if (is_exec_mapping(flags)) 3387 mm->exec_vm += npages; 3388 else if (is_stack_mapping(flags)) 3389 mm->stack_vm += npages; 3390 else if (is_data_mapping(flags)) 3391 mm->data_vm += npages; 3392 } 3393 3394 static vm_fault_t special_mapping_fault(struct vm_fault *vmf); 3395 3396 /* 3397 * Having a close hook prevents vma merging regardless of flags. 3398 */ 3399 static void special_mapping_close(struct vm_area_struct *vma) 3400 { 3401 } 3402 3403 static const char *special_mapping_name(struct vm_area_struct *vma) 3404 { 3405 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 3406 } 3407 3408 static int special_mapping_mremap(struct vm_area_struct *new_vma, 3409 unsigned long flags) 3410 { 3411 struct vm_special_mapping *sm = new_vma->vm_private_data; 3412 3413 if (flags & MREMAP_DONTUNMAP) 3414 return -EINVAL; 3415 3416 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm)) 3417 return -EFAULT; 3418 3419 if (sm->mremap) 3420 return sm->mremap(sm, new_vma); 3421 3422 return 0; 3423 } 3424 3425 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr) 3426 { 3427 /* 3428 * Forbid splitting special mappings - kernel has expectations over 3429 * the number of pages in mapping. Together with VM_DONTEXPAND 3430 * the size of vma should stay the same over the special mapping's 3431 * lifetime. 3432 */ 3433 return -EINVAL; 3434 } 3435 3436 static const struct vm_operations_struct special_mapping_vmops = { 3437 .close = special_mapping_close, 3438 .fault = special_mapping_fault, 3439 .mremap = special_mapping_mremap, 3440 .name = special_mapping_name, 3441 /* vDSO code relies that VVAR can't be accessed remotely */ 3442 .access = NULL, 3443 .may_split = special_mapping_split, 3444 }; 3445 3446 static const struct vm_operations_struct legacy_special_mapping_vmops = { 3447 .close = special_mapping_close, 3448 .fault = special_mapping_fault, 3449 }; 3450 3451 static vm_fault_t special_mapping_fault(struct vm_fault *vmf) 3452 { 3453 struct vm_area_struct *vma = vmf->vma; 3454 pgoff_t pgoff; 3455 struct page **pages; 3456 3457 if (vma->vm_ops == &legacy_special_mapping_vmops) { 3458 pages = vma->vm_private_data; 3459 } else { 3460 struct vm_special_mapping *sm = vma->vm_private_data; 3461 3462 if (sm->fault) 3463 return sm->fault(sm, vmf->vma, vmf); 3464 3465 pages = sm->pages; 3466 } 3467 3468 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages) 3469 pgoff--; 3470 3471 if (*pages) { 3472 struct page *page = *pages; 3473 get_page(page); 3474 vmf->page = page; 3475 return 0; 3476 } 3477 3478 return VM_FAULT_SIGBUS; 3479 } 3480 3481 static struct vm_area_struct *__install_special_mapping( 3482 struct mm_struct *mm, 3483 unsigned long addr, unsigned long len, 3484 unsigned long vm_flags, void *priv, 3485 const struct vm_operations_struct *ops) 3486 { 3487 int ret; 3488 struct vm_area_struct *vma; 3489 3490 vma = vm_area_alloc(mm); 3491 if (unlikely(vma == NULL)) 3492 return ERR_PTR(-ENOMEM); 3493 3494 vma->vm_start = addr; 3495 vma->vm_end = addr + len; 3496 3497 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY; 3498 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3499 3500 vma->vm_ops = ops; 3501 vma->vm_private_data = priv; 3502 3503 ret = insert_vm_struct(mm, vma); 3504 if (ret) 3505 goto out; 3506 3507 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT); 3508 3509 perf_event_mmap(vma); 3510 3511 return vma; 3512 3513 out: 3514 vm_area_free(vma); 3515 return ERR_PTR(ret); 3516 } 3517 3518 bool vma_is_special_mapping(const struct vm_area_struct *vma, 3519 const struct vm_special_mapping *sm) 3520 { 3521 return vma->vm_private_data == sm && 3522 (vma->vm_ops == &special_mapping_vmops || 3523 vma->vm_ops == &legacy_special_mapping_vmops); 3524 } 3525 3526 /* 3527 * Called with mm->mmap_lock held for writing. 3528 * Insert a new vma covering the given region, with the given flags. 3529 * Its pages are supplied by the given array of struct page *. 3530 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3531 * The region past the last page supplied will always produce SIGBUS. 3532 * The array pointer and the pages it points to are assumed to stay alive 3533 * for as long as this mapping might exist. 3534 */ 3535 struct vm_area_struct *_install_special_mapping( 3536 struct mm_struct *mm, 3537 unsigned long addr, unsigned long len, 3538 unsigned long vm_flags, const struct vm_special_mapping *spec) 3539 { 3540 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec, 3541 &special_mapping_vmops); 3542 } 3543 3544 int install_special_mapping(struct mm_struct *mm, 3545 unsigned long addr, unsigned long len, 3546 unsigned long vm_flags, struct page **pages) 3547 { 3548 struct vm_area_struct *vma = __install_special_mapping( 3549 mm, addr, len, vm_flags, (void *)pages, 3550 &legacy_special_mapping_vmops); 3551 3552 return PTR_ERR_OR_ZERO(vma); 3553 } 3554 3555 static DEFINE_MUTEX(mm_all_locks_mutex); 3556 3557 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3558 { 3559 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) { 3560 /* 3561 * The LSB of head.next can't change from under us 3562 * because we hold the mm_all_locks_mutex. 3563 */ 3564 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock); 3565 /* 3566 * We can safely modify head.next after taking the 3567 * anon_vma->root->rwsem. If some other vma in this mm shares 3568 * the same anon_vma we won't take it again. 3569 * 3570 * No need of atomic instructions here, head.next 3571 * can't change from under us thanks to the 3572 * anon_vma->root->rwsem. 3573 */ 3574 if (__test_and_set_bit(0, (unsigned long *) 3575 &anon_vma->root->rb_root.rb_root.rb_node)) 3576 BUG(); 3577 } 3578 } 3579 3580 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3581 { 3582 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3583 /* 3584 * AS_MM_ALL_LOCKS can't change from under us because 3585 * we hold the mm_all_locks_mutex. 3586 * 3587 * Operations on ->flags have to be atomic because 3588 * even if AS_MM_ALL_LOCKS is stable thanks to the 3589 * mm_all_locks_mutex, there may be other cpus 3590 * changing other bitflags in parallel to us. 3591 */ 3592 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3593 BUG(); 3594 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock); 3595 } 3596 } 3597 3598 /* 3599 * This operation locks against the VM for all pte/vma/mm related 3600 * operations that could ever happen on a certain mm. This includes 3601 * vmtruncate, try_to_unmap, and all page faults. 3602 * 3603 * The caller must take the mmap_lock in write mode before calling 3604 * mm_take_all_locks(). The caller isn't allowed to release the 3605 * mmap_lock until mm_drop_all_locks() returns. 3606 * 3607 * mmap_lock in write mode is required in order to block all operations 3608 * that could modify pagetables and free pages without need of 3609 * altering the vma layout. It's also needed in write mode to avoid new 3610 * anon_vmas to be associated with existing vmas. 3611 * 3612 * A single task can't take more than one mm_take_all_locks() in a row 3613 * or it would deadlock. 3614 * 3615 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3616 * mapping->flags avoid to take the same lock twice, if more than one 3617 * vma in this mm is backed by the same anon_vma or address_space. 3618 * 3619 * We take locks in following order, accordingly to comment at beginning 3620 * of mm/rmap.c: 3621 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for 3622 * hugetlb mapping); 3623 * - all i_mmap_rwsem locks; 3624 * - all anon_vma->rwseml 3625 * 3626 * We can take all locks within these types randomly because the VM code 3627 * doesn't nest them and we protected from parallel mm_take_all_locks() by 3628 * mm_all_locks_mutex. 3629 * 3630 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3631 * that may have to take thousand of locks. 3632 * 3633 * mm_take_all_locks() can fail if it's interrupted by signals. 3634 */ 3635 int mm_take_all_locks(struct mm_struct *mm) 3636 { 3637 struct vm_area_struct *vma; 3638 struct anon_vma_chain *avc; 3639 3640 BUG_ON(mmap_read_trylock(mm)); 3641 3642 mutex_lock(&mm_all_locks_mutex); 3643 3644 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3645 if (signal_pending(current)) 3646 goto out_unlock; 3647 if (vma->vm_file && vma->vm_file->f_mapping && 3648 is_vm_hugetlb_page(vma)) 3649 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3650 } 3651 3652 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3653 if (signal_pending(current)) 3654 goto out_unlock; 3655 if (vma->vm_file && vma->vm_file->f_mapping && 3656 !is_vm_hugetlb_page(vma)) 3657 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3658 } 3659 3660 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3661 if (signal_pending(current)) 3662 goto out_unlock; 3663 if (vma->anon_vma) 3664 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3665 vm_lock_anon_vma(mm, avc->anon_vma); 3666 } 3667 3668 return 0; 3669 3670 out_unlock: 3671 mm_drop_all_locks(mm); 3672 return -EINTR; 3673 } 3674 3675 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3676 { 3677 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) { 3678 /* 3679 * The LSB of head.next can't change to 0 from under 3680 * us because we hold the mm_all_locks_mutex. 3681 * 3682 * We must however clear the bitflag before unlocking 3683 * the vma so the users using the anon_vma->rb_root will 3684 * never see our bitflag. 3685 * 3686 * No need of atomic instructions here, head.next 3687 * can't change from under us until we release the 3688 * anon_vma->root->rwsem. 3689 */ 3690 if (!__test_and_clear_bit(0, (unsigned long *) 3691 &anon_vma->root->rb_root.rb_root.rb_node)) 3692 BUG(); 3693 anon_vma_unlock_write(anon_vma); 3694 } 3695 } 3696 3697 static void vm_unlock_mapping(struct address_space *mapping) 3698 { 3699 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3700 /* 3701 * AS_MM_ALL_LOCKS can't change to 0 from under us 3702 * because we hold the mm_all_locks_mutex. 3703 */ 3704 i_mmap_unlock_write(mapping); 3705 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3706 &mapping->flags)) 3707 BUG(); 3708 } 3709 } 3710 3711 /* 3712 * The mmap_lock cannot be released by the caller until 3713 * mm_drop_all_locks() returns. 3714 */ 3715 void mm_drop_all_locks(struct mm_struct *mm) 3716 { 3717 struct vm_area_struct *vma; 3718 struct anon_vma_chain *avc; 3719 3720 BUG_ON(mmap_read_trylock(mm)); 3721 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3722 3723 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3724 if (vma->anon_vma) 3725 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3726 vm_unlock_anon_vma(avc->anon_vma); 3727 if (vma->vm_file && vma->vm_file->f_mapping) 3728 vm_unlock_mapping(vma->vm_file->f_mapping); 3729 } 3730 3731 mutex_unlock(&mm_all_locks_mutex); 3732 } 3733 3734 /* 3735 * initialise the percpu counter for VM 3736 */ 3737 void __init mmap_init(void) 3738 { 3739 int ret; 3740 3741 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3742 VM_BUG_ON(ret); 3743 } 3744 3745 /* 3746 * Initialise sysctl_user_reserve_kbytes. 3747 * 3748 * This is intended to prevent a user from starting a single memory hogging 3749 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3750 * mode. 3751 * 3752 * The default value is min(3% of free memory, 128MB) 3753 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3754 */ 3755 static int init_user_reserve(void) 3756 { 3757 unsigned long free_kbytes; 3758 3759 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3760 3761 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3762 return 0; 3763 } 3764 subsys_initcall(init_user_reserve); 3765 3766 /* 3767 * Initialise sysctl_admin_reserve_kbytes. 3768 * 3769 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3770 * to log in and kill a memory hogging process. 3771 * 3772 * Systems with more than 256MB will reserve 8MB, enough to recover 3773 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3774 * only reserve 3% of free pages by default. 3775 */ 3776 static int init_admin_reserve(void) 3777 { 3778 unsigned long free_kbytes; 3779 3780 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3781 3782 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3783 return 0; 3784 } 3785 subsys_initcall(init_admin_reserve); 3786 3787 /* 3788 * Reinititalise user and admin reserves if memory is added or removed. 3789 * 3790 * The default user reserve max is 128MB, and the default max for the 3791 * admin reserve is 8MB. These are usually, but not always, enough to 3792 * enable recovery from a memory hogging process using login/sshd, a shell, 3793 * and tools like top. It may make sense to increase or even disable the 3794 * reserve depending on the existence of swap or variations in the recovery 3795 * tools. So, the admin may have changed them. 3796 * 3797 * If memory is added and the reserves have been eliminated or increased above 3798 * the default max, then we'll trust the admin. 3799 * 3800 * If memory is removed and there isn't enough free memory, then we 3801 * need to reset the reserves. 3802 * 3803 * Otherwise keep the reserve set by the admin. 3804 */ 3805 static int reserve_mem_notifier(struct notifier_block *nb, 3806 unsigned long action, void *data) 3807 { 3808 unsigned long tmp, free_kbytes; 3809 3810 switch (action) { 3811 case MEM_ONLINE: 3812 /* Default max is 128MB. Leave alone if modified by operator. */ 3813 tmp = sysctl_user_reserve_kbytes; 3814 if (0 < tmp && tmp < (1UL << 17)) 3815 init_user_reserve(); 3816 3817 /* Default max is 8MB. Leave alone if modified by operator. */ 3818 tmp = sysctl_admin_reserve_kbytes; 3819 if (0 < tmp && tmp < (1UL << 13)) 3820 init_admin_reserve(); 3821 3822 break; 3823 case MEM_OFFLINE: 3824 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3825 3826 if (sysctl_user_reserve_kbytes > free_kbytes) { 3827 init_user_reserve(); 3828 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3829 sysctl_user_reserve_kbytes); 3830 } 3831 3832 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3833 init_admin_reserve(); 3834 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3835 sysctl_admin_reserve_kbytes); 3836 } 3837 break; 3838 default: 3839 break; 3840 } 3841 return NOTIFY_OK; 3842 } 3843 3844 static struct notifier_block reserve_mem_nb = { 3845 .notifier_call = reserve_mem_notifier, 3846 }; 3847 3848 static int __meminit init_reserve_notifier(void) 3849 { 3850 if (register_hotmemory_notifier(&reserve_mem_nb)) 3851 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3852 3853 return 0; 3854 } 3855 subsys_initcall(init_reserve_notifier); 3856