1 /* 2 * mm/mmap.c 3 * 4 * Written by obz. 5 * 6 * Address space accounting code <alan@lxorguk.ukuu.org.uk> 7 */ 8 9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 10 11 #include <linux/kernel.h> 12 #include <linux/slab.h> 13 #include <linux/backing-dev.h> 14 #include <linux/mm.h> 15 #include <linux/vmacache.h> 16 #include <linux/shm.h> 17 #include <linux/mman.h> 18 #include <linux/pagemap.h> 19 #include <linux/swap.h> 20 #include <linux/syscalls.h> 21 #include <linux/capability.h> 22 #include <linux/init.h> 23 #include <linux/file.h> 24 #include <linux/fs.h> 25 #include <linux/personality.h> 26 #include <linux/security.h> 27 #include <linux/hugetlb.h> 28 #include <linux/shmem_fs.h> 29 #include <linux/profile.h> 30 #include <linux/export.h> 31 #include <linux/mount.h> 32 #include <linux/mempolicy.h> 33 #include <linux/rmap.h> 34 #include <linux/mmu_notifier.h> 35 #include <linux/mmdebug.h> 36 #include <linux/perf_event.h> 37 #include <linux/audit.h> 38 #include <linux/khugepaged.h> 39 #include <linux/uprobes.h> 40 #include <linux/rbtree_augmented.h> 41 #include <linux/notifier.h> 42 #include <linux/memory.h> 43 #include <linux/printk.h> 44 #include <linux/userfaultfd_k.h> 45 #include <linux/moduleparam.h> 46 #include <linux/pkeys.h> 47 48 #include <linux/uaccess.h> 49 #include <asm/cacheflush.h> 50 #include <asm/tlb.h> 51 #include <asm/mmu_context.h> 52 53 #include "internal.h" 54 55 #ifndef arch_mmap_check 56 #define arch_mmap_check(addr, len, flags) (0) 57 #endif 58 59 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS 60 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN; 61 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX; 62 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS; 63 #endif 64 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS 65 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN; 66 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX; 67 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS; 68 #endif 69 70 static bool ignore_rlimit_data; 71 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644); 72 73 static void unmap_region(struct mm_struct *mm, 74 struct vm_area_struct *vma, struct vm_area_struct *prev, 75 unsigned long start, unsigned long end); 76 77 /* description of effects of mapping type and prot in current implementation. 78 * this is due to the limited x86 page protection hardware. The expected 79 * behavior is in parens: 80 * 81 * map_type prot 82 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC 83 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes 84 * w: (no) no w: (no) no w: (yes) yes w: (no) no 85 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 86 * 87 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes 88 * w: (no) no w: (no) no w: (copy) copy w: (no) no 89 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes 90 * 91 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and 92 * MAP_PRIVATE: 93 * r: (no) no 94 * w: (no) no 95 * x: (yes) yes 96 */ 97 pgprot_t protection_map[16] = { 98 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111, 99 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111 100 }; 101 102 pgprot_t vm_get_page_prot(unsigned long vm_flags) 103 { 104 return __pgprot(pgprot_val(protection_map[vm_flags & 105 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) | 106 pgprot_val(arch_vm_get_page_prot(vm_flags))); 107 } 108 EXPORT_SYMBOL(vm_get_page_prot); 109 110 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags) 111 { 112 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags)); 113 } 114 115 /* Update vma->vm_page_prot to reflect vma->vm_flags. */ 116 void vma_set_page_prot(struct vm_area_struct *vma) 117 { 118 unsigned long vm_flags = vma->vm_flags; 119 pgprot_t vm_page_prot; 120 121 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags); 122 if (vma_wants_writenotify(vma, vm_page_prot)) { 123 vm_flags &= ~VM_SHARED; 124 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags); 125 } 126 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */ 127 WRITE_ONCE(vma->vm_page_prot, vm_page_prot); 128 } 129 130 /* 131 * Requires inode->i_mapping->i_mmap_rwsem 132 */ 133 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 134 struct file *file, struct address_space *mapping) 135 { 136 if (vma->vm_flags & VM_DENYWRITE) 137 atomic_inc(&file_inode(file)->i_writecount); 138 if (vma->vm_flags & VM_SHARED) 139 mapping_unmap_writable(mapping); 140 141 flush_dcache_mmap_lock(mapping); 142 vma_interval_tree_remove(vma, &mapping->i_mmap); 143 flush_dcache_mmap_unlock(mapping); 144 } 145 146 /* 147 * Unlink a file-based vm structure from its interval tree, to hide 148 * vma from rmap and vmtruncate before freeing its page tables. 149 */ 150 void unlink_file_vma(struct vm_area_struct *vma) 151 { 152 struct file *file = vma->vm_file; 153 154 if (file) { 155 struct address_space *mapping = file->f_mapping; 156 i_mmap_lock_write(mapping); 157 __remove_shared_vm_struct(vma, file, mapping); 158 i_mmap_unlock_write(mapping); 159 } 160 } 161 162 /* 163 * Close a vm structure and free it, returning the next. 164 */ 165 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma) 166 { 167 struct vm_area_struct *next = vma->vm_next; 168 169 might_sleep(); 170 if (vma->vm_ops && vma->vm_ops->close) 171 vma->vm_ops->close(vma); 172 if (vma->vm_file) 173 fput(vma->vm_file); 174 mpol_put(vma_policy(vma)); 175 kmem_cache_free(vm_area_cachep, vma); 176 return next; 177 } 178 179 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf); 180 181 SYSCALL_DEFINE1(brk, unsigned long, brk) 182 { 183 unsigned long retval; 184 unsigned long newbrk, oldbrk; 185 struct mm_struct *mm = current->mm; 186 struct vm_area_struct *next; 187 unsigned long min_brk; 188 bool populate; 189 LIST_HEAD(uf); 190 191 if (down_write_killable(&mm->mmap_sem)) 192 return -EINTR; 193 194 #ifdef CONFIG_COMPAT_BRK 195 /* 196 * CONFIG_COMPAT_BRK can still be overridden by setting 197 * randomize_va_space to 2, which will still cause mm->start_brk 198 * to be arbitrarily shifted 199 */ 200 if (current->brk_randomized) 201 min_brk = mm->start_brk; 202 else 203 min_brk = mm->end_data; 204 #else 205 min_brk = mm->start_brk; 206 #endif 207 if (brk < min_brk) 208 goto out; 209 210 /* 211 * Check against rlimit here. If this check is done later after the test 212 * of oldbrk with newbrk then it can escape the test and let the data 213 * segment grow beyond its set limit the in case where the limit is 214 * not page aligned -Ram Gupta 215 */ 216 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 217 mm->end_data, mm->start_data)) 218 goto out; 219 220 newbrk = PAGE_ALIGN(brk); 221 oldbrk = PAGE_ALIGN(mm->brk); 222 if (oldbrk == newbrk) 223 goto set_brk; 224 225 /* Always allow shrinking brk. */ 226 if (brk <= mm->brk) { 227 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf)) 228 goto set_brk; 229 goto out; 230 } 231 232 /* Check against existing mmap mappings. */ 233 next = find_vma(mm, oldbrk); 234 if (next && newbrk + PAGE_SIZE > vm_start_gap(next)) 235 goto out; 236 237 /* Ok, looks good - let it rip. */ 238 if (do_brk(oldbrk, newbrk-oldbrk, &uf) < 0) 239 goto out; 240 241 set_brk: 242 mm->brk = brk; 243 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 244 up_write(&mm->mmap_sem); 245 userfaultfd_unmap_complete(mm, &uf); 246 if (populate) 247 mm_populate(oldbrk, newbrk - oldbrk); 248 return brk; 249 250 out: 251 retval = mm->brk; 252 up_write(&mm->mmap_sem); 253 return retval; 254 } 255 256 static long vma_compute_subtree_gap(struct vm_area_struct *vma) 257 { 258 unsigned long max, prev_end, subtree_gap; 259 260 /* 261 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we 262 * allow two stack_guard_gaps between them here, and when choosing 263 * an unmapped area; whereas when expanding we only require one. 264 * That's a little inconsistent, but keeps the code here simpler. 265 */ 266 max = vm_start_gap(vma); 267 if (vma->vm_prev) { 268 prev_end = vm_end_gap(vma->vm_prev); 269 if (max > prev_end) 270 max -= prev_end; 271 else 272 max = 0; 273 } 274 if (vma->vm_rb.rb_left) { 275 subtree_gap = rb_entry(vma->vm_rb.rb_left, 276 struct vm_area_struct, vm_rb)->rb_subtree_gap; 277 if (subtree_gap > max) 278 max = subtree_gap; 279 } 280 if (vma->vm_rb.rb_right) { 281 subtree_gap = rb_entry(vma->vm_rb.rb_right, 282 struct vm_area_struct, vm_rb)->rb_subtree_gap; 283 if (subtree_gap > max) 284 max = subtree_gap; 285 } 286 return max; 287 } 288 289 #ifdef CONFIG_DEBUG_VM_RB 290 static int browse_rb(struct mm_struct *mm) 291 { 292 struct rb_root *root = &mm->mm_rb; 293 int i = 0, j, bug = 0; 294 struct rb_node *nd, *pn = NULL; 295 unsigned long prev = 0, pend = 0; 296 297 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 298 struct vm_area_struct *vma; 299 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 300 if (vma->vm_start < prev) { 301 pr_emerg("vm_start %lx < prev %lx\n", 302 vma->vm_start, prev); 303 bug = 1; 304 } 305 if (vma->vm_start < pend) { 306 pr_emerg("vm_start %lx < pend %lx\n", 307 vma->vm_start, pend); 308 bug = 1; 309 } 310 if (vma->vm_start > vma->vm_end) { 311 pr_emerg("vm_start %lx > vm_end %lx\n", 312 vma->vm_start, vma->vm_end); 313 bug = 1; 314 } 315 spin_lock(&mm->page_table_lock); 316 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) { 317 pr_emerg("free gap %lx, correct %lx\n", 318 vma->rb_subtree_gap, 319 vma_compute_subtree_gap(vma)); 320 bug = 1; 321 } 322 spin_unlock(&mm->page_table_lock); 323 i++; 324 pn = nd; 325 prev = vma->vm_start; 326 pend = vma->vm_end; 327 } 328 j = 0; 329 for (nd = pn; nd; nd = rb_prev(nd)) 330 j++; 331 if (i != j) { 332 pr_emerg("backwards %d, forwards %d\n", j, i); 333 bug = 1; 334 } 335 return bug ? -1 : i; 336 } 337 338 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore) 339 { 340 struct rb_node *nd; 341 342 for (nd = rb_first(root); nd; nd = rb_next(nd)) { 343 struct vm_area_struct *vma; 344 vma = rb_entry(nd, struct vm_area_struct, vm_rb); 345 VM_BUG_ON_VMA(vma != ignore && 346 vma->rb_subtree_gap != vma_compute_subtree_gap(vma), 347 vma); 348 } 349 } 350 351 static void validate_mm(struct mm_struct *mm) 352 { 353 int bug = 0; 354 int i = 0; 355 unsigned long highest_address = 0; 356 struct vm_area_struct *vma = mm->mmap; 357 358 while (vma) { 359 struct anon_vma *anon_vma = vma->anon_vma; 360 struct anon_vma_chain *avc; 361 362 if (anon_vma) { 363 anon_vma_lock_read(anon_vma); 364 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 365 anon_vma_interval_tree_verify(avc); 366 anon_vma_unlock_read(anon_vma); 367 } 368 369 highest_address = vm_end_gap(vma); 370 vma = vma->vm_next; 371 i++; 372 } 373 if (i != mm->map_count) { 374 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i); 375 bug = 1; 376 } 377 if (highest_address != mm->highest_vm_end) { 378 pr_emerg("mm->highest_vm_end %lx, found %lx\n", 379 mm->highest_vm_end, highest_address); 380 bug = 1; 381 } 382 i = browse_rb(mm); 383 if (i != mm->map_count) { 384 if (i != -1) 385 pr_emerg("map_count %d rb %d\n", mm->map_count, i); 386 bug = 1; 387 } 388 VM_BUG_ON_MM(bug, mm); 389 } 390 #else 391 #define validate_mm_rb(root, ignore) do { } while (0) 392 #define validate_mm(mm) do { } while (0) 393 #endif 394 395 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb, 396 unsigned long, rb_subtree_gap, vma_compute_subtree_gap) 397 398 /* 399 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or 400 * vma->vm_prev->vm_end values changed, without modifying the vma's position 401 * in the rbtree. 402 */ 403 static void vma_gap_update(struct vm_area_struct *vma) 404 { 405 /* 406 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback 407 * function that does exacltly what we want. 408 */ 409 vma_gap_callbacks_propagate(&vma->vm_rb, NULL); 410 } 411 412 static inline void vma_rb_insert(struct vm_area_struct *vma, 413 struct rb_root *root) 414 { 415 /* All rb_subtree_gap values must be consistent prior to insertion */ 416 validate_mm_rb(root, NULL); 417 418 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 419 } 420 421 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root) 422 { 423 /* 424 * Note rb_erase_augmented is a fairly large inline function, 425 * so make sure we instantiate it only once with our desired 426 * augmented rbtree callbacks. 427 */ 428 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks); 429 } 430 431 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma, 432 struct rb_root *root, 433 struct vm_area_struct *ignore) 434 { 435 /* 436 * All rb_subtree_gap values must be consistent prior to erase, 437 * with the possible exception of the "next" vma being erased if 438 * next->vm_start was reduced. 439 */ 440 validate_mm_rb(root, ignore); 441 442 __vma_rb_erase(vma, root); 443 } 444 445 static __always_inline void vma_rb_erase(struct vm_area_struct *vma, 446 struct rb_root *root) 447 { 448 /* 449 * All rb_subtree_gap values must be consistent prior to erase, 450 * with the possible exception of the vma being erased. 451 */ 452 validate_mm_rb(root, vma); 453 454 __vma_rb_erase(vma, root); 455 } 456 457 /* 458 * vma has some anon_vma assigned, and is already inserted on that 459 * anon_vma's interval trees. 460 * 461 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 462 * vma must be removed from the anon_vma's interval trees using 463 * anon_vma_interval_tree_pre_update_vma(). 464 * 465 * After the update, the vma will be reinserted using 466 * anon_vma_interval_tree_post_update_vma(). 467 * 468 * The entire update must be protected by exclusive mmap_sem and by 469 * the root anon_vma's mutex. 470 */ 471 static inline void 472 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 473 { 474 struct anon_vma_chain *avc; 475 476 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 477 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 478 } 479 480 static inline void 481 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 482 { 483 struct anon_vma_chain *avc; 484 485 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 486 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 487 } 488 489 static int find_vma_links(struct mm_struct *mm, unsigned long addr, 490 unsigned long end, struct vm_area_struct **pprev, 491 struct rb_node ***rb_link, struct rb_node **rb_parent) 492 { 493 struct rb_node **__rb_link, *__rb_parent, *rb_prev; 494 495 __rb_link = &mm->mm_rb.rb_node; 496 rb_prev = __rb_parent = NULL; 497 498 while (*__rb_link) { 499 struct vm_area_struct *vma_tmp; 500 501 __rb_parent = *__rb_link; 502 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb); 503 504 if (vma_tmp->vm_end > addr) { 505 /* Fail if an existing vma overlaps the area */ 506 if (vma_tmp->vm_start < end) 507 return -ENOMEM; 508 __rb_link = &__rb_parent->rb_left; 509 } else { 510 rb_prev = __rb_parent; 511 __rb_link = &__rb_parent->rb_right; 512 } 513 } 514 515 *pprev = NULL; 516 if (rb_prev) 517 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb); 518 *rb_link = __rb_link; 519 *rb_parent = __rb_parent; 520 return 0; 521 } 522 523 static unsigned long count_vma_pages_range(struct mm_struct *mm, 524 unsigned long addr, unsigned long end) 525 { 526 unsigned long nr_pages = 0; 527 struct vm_area_struct *vma; 528 529 /* Find first overlaping mapping */ 530 vma = find_vma_intersection(mm, addr, end); 531 if (!vma) 532 return 0; 533 534 nr_pages = (min(end, vma->vm_end) - 535 max(addr, vma->vm_start)) >> PAGE_SHIFT; 536 537 /* Iterate over the rest of the overlaps */ 538 for (vma = vma->vm_next; vma; vma = vma->vm_next) { 539 unsigned long overlap_len; 540 541 if (vma->vm_start > end) 542 break; 543 544 overlap_len = min(end, vma->vm_end) - vma->vm_start; 545 nr_pages += overlap_len >> PAGE_SHIFT; 546 } 547 548 return nr_pages; 549 } 550 551 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma, 552 struct rb_node **rb_link, struct rb_node *rb_parent) 553 { 554 /* Update tracking information for the gap following the new vma. */ 555 if (vma->vm_next) 556 vma_gap_update(vma->vm_next); 557 else 558 mm->highest_vm_end = vm_end_gap(vma); 559 560 /* 561 * vma->vm_prev wasn't known when we followed the rbtree to find the 562 * correct insertion point for that vma. As a result, we could not 563 * update the vma vm_rb parents rb_subtree_gap values on the way down. 564 * So, we first insert the vma with a zero rb_subtree_gap value 565 * (to be consistent with what we did on the way down), and then 566 * immediately update the gap to the correct value. Finally we 567 * rebalance the rbtree after all augmented values have been set. 568 */ 569 rb_link_node(&vma->vm_rb, rb_parent, rb_link); 570 vma->rb_subtree_gap = 0; 571 vma_gap_update(vma); 572 vma_rb_insert(vma, &mm->mm_rb); 573 } 574 575 static void __vma_link_file(struct vm_area_struct *vma) 576 { 577 struct file *file; 578 579 file = vma->vm_file; 580 if (file) { 581 struct address_space *mapping = file->f_mapping; 582 583 if (vma->vm_flags & VM_DENYWRITE) 584 atomic_dec(&file_inode(file)->i_writecount); 585 if (vma->vm_flags & VM_SHARED) 586 atomic_inc(&mapping->i_mmap_writable); 587 588 flush_dcache_mmap_lock(mapping); 589 vma_interval_tree_insert(vma, &mapping->i_mmap); 590 flush_dcache_mmap_unlock(mapping); 591 } 592 } 593 594 static void 595 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 596 struct vm_area_struct *prev, struct rb_node **rb_link, 597 struct rb_node *rb_parent) 598 { 599 __vma_link_list(mm, vma, prev, rb_parent); 600 __vma_link_rb(mm, vma, rb_link, rb_parent); 601 } 602 603 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma, 604 struct vm_area_struct *prev, struct rb_node **rb_link, 605 struct rb_node *rb_parent) 606 { 607 struct address_space *mapping = NULL; 608 609 if (vma->vm_file) { 610 mapping = vma->vm_file->f_mapping; 611 i_mmap_lock_write(mapping); 612 } 613 614 __vma_link(mm, vma, prev, rb_link, rb_parent); 615 __vma_link_file(vma); 616 617 if (mapping) 618 i_mmap_unlock_write(mapping); 619 620 mm->map_count++; 621 validate_mm(mm); 622 } 623 624 /* 625 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the 626 * mm's list and rbtree. It has already been inserted into the interval tree. 627 */ 628 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 629 { 630 struct vm_area_struct *prev; 631 struct rb_node **rb_link, *rb_parent; 632 633 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 634 &prev, &rb_link, &rb_parent)) 635 BUG(); 636 __vma_link(mm, vma, prev, rb_link, rb_parent); 637 mm->map_count++; 638 } 639 640 static __always_inline void __vma_unlink_common(struct mm_struct *mm, 641 struct vm_area_struct *vma, 642 struct vm_area_struct *prev, 643 bool has_prev, 644 struct vm_area_struct *ignore) 645 { 646 struct vm_area_struct *next; 647 648 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore); 649 next = vma->vm_next; 650 if (has_prev) 651 prev->vm_next = next; 652 else { 653 prev = vma->vm_prev; 654 if (prev) 655 prev->vm_next = next; 656 else 657 mm->mmap = next; 658 } 659 if (next) 660 next->vm_prev = prev; 661 662 /* Kill the cache */ 663 vmacache_invalidate(mm); 664 } 665 666 static inline void __vma_unlink_prev(struct mm_struct *mm, 667 struct vm_area_struct *vma, 668 struct vm_area_struct *prev) 669 { 670 __vma_unlink_common(mm, vma, prev, true, vma); 671 } 672 673 /* 674 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that 675 * is already present in an i_mmap tree without adjusting the tree. 676 * The following helper function should be used when such adjustments 677 * are necessary. The "insert" vma (if any) is to be inserted 678 * before we drop the necessary locks. 679 */ 680 int __vma_adjust(struct vm_area_struct *vma, unsigned long start, 681 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert, 682 struct vm_area_struct *expand) 683 { 684 struct mm_struct *mm = vma->vm_mm; 685 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma; 686 struct address_space *mapping = NULL; 687 struct rb_root *root = NULL; 688 struct anon_vma *anon_vma = NULL; 689 struct file *file = vma->vm_file; 690 bool start_changed = false, end_changed = false; 691 long adjust_next = 0; 692 int remove_next = 0; 693 694 if (next && !insert) { 695 struct vm_area_struct *exporter = NULL, *importer = NULL; 696 697 if (end >= next->vm_end) { 698 /* 699 * vma expands, overlapping all the next, and 700 * perhaps the one after too (mprotect case 6). 701 * The only other cases that gets here are 702 * case 1, case 7 and case 8. 703 */ 704 if (next == expand) { 705 /* 706 * The only case where we don't expand "vma" 707 * and we expand "next" instead is case 8. 708 */ 709 VM_WARN_ON(end != next->vm_end); 710 /* 711 * remove_next == 3 means we're 712 * removing "vma" and that to do so we 713 * swapped "vma" and "next". 714 */ 715 remove_next = 3; 716 VM_WARN_ON(file != next->vm_file); 717 swap(vma, next); 718 } else { 719 VM_WARN_ON(expand != vma); 720 /* 721 * case 1, 6, 7, remove_next == 2 is case 6, 722 * remove_next == 1 is case 1 or 7. 723 */ 724 remove_next = 1 + (end > next->vm_end); 725 VM_WARN_ON(remove_next == 2 && 726 end != next->vm_next->vm_end); 727 VM_WARN_ON(remove_next == 1 && 728 end != next->vm_end); 729 /* trim end to next, for case 6 first pass */ 730 end = next->vm_end; 731 } 732 733 exporter = next; 734 importer = vma; 735 736 /* 737 * If next doesn't have anon_vma, import from vma after 738 * next, if the vma overlaps with it. 739 */ 740 if (remove_next == 2 && !next->anon_vma) 741 exporter = next->vm_next; 742 743 } else if (end > next->vm_start) { 744 /* 745 * vma expands, overlapping part of the next: 746 * mprotect case 5 shifting the boundary up. 747 */ 748 adjust_next = (end - next->vm_start) >> PAGE_SHIFT; 749 exporter = next; 750 importer = vma; 751 VM_WARN_ON(expand != importer); 752 } else if (end < vma->vm_end) { 753 /* 754 * vma shrinks, and !insert tells it's not 755 * split_vma inserting another: so it must be 756 * mprotect case 4 shifting the boundary down. 757 */ 758 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT); 759 exporter = vma; 760 importer = next; 761 VM_WARN_ON(expand != importer); 762 } 763 764 /* 765 * Easily overlooked: when mprotect shifts the boundary, 766 * make sure the expanding vma has anon_vma set if the 767 * shrinking vma had, to cover any anon pages imported. 768 */ 769 if (exporter && exporter->anon_vma && !importer->anon_vma) { 770 int error; 771 772 importer->anon_vma = exporter->anon_vma; 773 error = anon_vma_clone(importer, exporter); 774 if (error) 775 return error; 776 } 777 } 778 again: 779 vma_adjust_trans_huge(orig_vma, start, end, adjust_next); 780 781 if (file) { 782 mapping = file->f_mapping; 783 root = &mapping->i_mmap; 784 uprobe_munmap(vma, vma->vm_start, vma->vm_end); 785 786 if (adjust_next) 787 uprobe_munmap(next, next->vm_start, next->vm_end); 788 789 i_mmap_lock_write(mapping); 790 if (insert) { 791 /* 792 * Put into interval tree now, so instantiated pages 793 * are visible to arm/parisc __flush_dcache_page 794 * throughout; but we cannot insert into address 795 * space until vma start or end is updated. 796 */ 797 __vma_link_file(insert); 798 } 799 } 800 801 anon_vma = vma->anon_vma; 802 if (!anon_vma && adjust_next) 803 anon_vma = next->anon_vma; 804 if (anon_vma) { 805 VM_WARN_ON(adjust_next && next->anon_vma && 806 anon_vma != next->anon_vma); 807 anon_vma_lock_write(anon_vma); 808 anon_vma_interval_tree_pre_update_vma(vma); 809 if (adjust_next) 810 anon_vma_interval_tree_pre_update_vma(next); 811 } 812 813 if (root) { 814 flush_dcache_mmap_lock(mapping); 815 vma_interval_tree_remove(vma, root); 816 if (adjust_next) 817 vma_interval_tree_remove(next, root); 818 } 819 820 if (start != vma->vm_start) { 821 vma->vm_start = start; 822 start_changed = true; 823 } 824 if (end != vma->vm_end) { 825 vma->vm_end = end; 826 end_changed = true; 827 } 828 vma->vm_pgoff = pgoff; 829 if (adjust_next) { 830 next->vm_start += adjust_next << PAGE_SHIFT; 831 next->vm_pgoff += adjust_next; 832 } 833 834 if (root) { 835 if (adjust_next) 836 vma_interval_tree_insert(next, root); 837 vma_interval_tree_insert(vma, root); 838 flush_dcache_mmap_unlock(mapping); 839 } 840 841 if (remove_next) { 842 /* 843 * vma_merge has merged next into vma, and needs 844 * us to remove next before dropping the locks. 845 */ 846 if (remove_next != 3) 847 __vma_unlink_prev(mm, next, vma); 848 else 849 /* 850 * vma is not before next if they've been 851 * swapped. 852 * 853 * pre-swap() next->vm_start was reduced so 854 * tell validate_mm_rb to ignore pre-swap() 855 * "next" (which is stored in post-swap() 856 * "vma"). 857 */ 858 __vma_unlink_common(mm, next, NULL, false, vma); 859 if (file) 860 __remove_shared_vm_struct(next, file, mapping); 861 } else if (insert) { 862 /* 863 * split_vma has split insert from vma, and needs 864 * us to insert it before dropping the locks 865 * (it may either follow vma or precede it). 866 */ 867 __insert_vm_struct(mm, insert); 868 } else { 869 if (start_changed) 870 vma_gap_update(vma); 871 if (end_changed) { 872 if (!next) 873 mm->highest_vm_end = vm_end_gap(vma); 874 else if (!adjust_next) 875 vma_gap_update(next); 876 } 877 } 878 879 if (anon_vma) { 880 anon_vma_interval_tree_post_update_vma(vma); 881 if (adjust_next) 882 anon_vma_interval_tree_post_update_vma(next); 883 anon_vma_unlock_write(anon_vma); 884 } 885 if (mapping) 886 i_mmap_unlock_write(mapping); 887 888 if (root) { 889 uprobe_mmap(vma); 890 891 if (adjust_next) 892 uprobe_mmap(next); 893 } 894 895 if (remove_next) { 896 if (file) { 897 uprobe_munmap(next, next->vm_start, next->vm_end); 898 fput(file); 899 } 900 if (next->anon_vma) 901 anon_vma_merge(vma, next); 902 mm->map_count--; 903 mpol_put(vma_policy(next)); 904 kmem_cache_free(vm_area_cachep, next); 905 /* 906 * In mprotect's case 6 (see comments on vma_merge), 907 * we must remove another next too. It would clutter 908 * up the code too much to do both in one go. 909 */ 910 if (remove_next != 3) { 911 /* 912 * If "next" was removed and vma->vm_end was 913 * expanded (up) over it, in turn 914 * "next->vm_prev->vm_end" changed and the 915 * "vma->vm_next" gap must be updated. 916 */ 917 next = vma->vm_next; 918 } else { 919 /* 920 * For the scope of the comment "next" and 921 * "vma" considered pre-swap(): if "vma" was 922 * removed, next->vm_start was expanded (down) 923 * over it and the "next" gap must be updated. 924 * Because of the swap() the post-swap() "vma" 925 * actually points to pre-swap() "next" 926 * (post-swap() "next" as opposed is now a 927 * dangling pointer). 928 */ 929 next = vma; 930 } 931 if (remove_next == 2) { 932 remove_next = 1; 933 end = next->vm_end; 934 goto again; 935 } 936 else if (next) 937 vma_gap_update(next); 938 else { 939 /* 940 * If remove_next == 2 we obviously can't 941 * reach this path. 942 * 943 * If remove_next == 3 we can't reach this 944 * path because pre-swap() next is always not 945 * NULL. pre-swap() "next" is not being 946 * removed and its next->vm_end is not altered 947 * (and furthermore "end" already matches 948 * next->vm_end in remove_next == 3). 949 * 950 * We reach this only in the remove_next == 1 951 * case if the "next" vma that was removed was 952 * the highest vma of the mm. However in such 953 * case next->vm_end == "end" and the extended 954 * "vma" has vma->vm_end == next->vm_end so 955 * mm->highest_vm_end doesn't need any update 956 * in remove_next == 1 case. 957 */ 958 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma)); 959 } 960 } 961 if (insert && file) 962 uprobe_mmap(insert); 963 964 validate_mm(mm); 965 966 return 0; 967 } 968 969 /* 970 * If the vma has a ->close operation then the driver probably needs to release 971 * per-vma resources, so we don't attempt to merge those. 972 */ 973 static inline int is_mergeable_vma(struct vm_area_struct *vma, 974 struct file *file, unsigned long vm_flags, 975 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 976 { 977 /* 978 * VM_SOFTDIRTY should not prevent from VMA merging, if we 979 * match the flags but dirty bit -- the caller should mark 980 * merged VMA as dirty. If dirty bit won't be excluded from 981 * comparison, we increase pressue on the memory system forcing 982 * the kernel to generate new VMAs when old one could be 983 * extended instead. 984 */ 985 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 986 return 0; 987 if (vma->vm_file != file) 988 return 0; 989 if (vma->vm_ops && vma->vm_ops->close) 990 return 0; 991 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx)) 992 return 0; 993 return 1; 994 } 995 996 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 997 struct anon_vma *anon_vma2, 998 struct vm_area_struct *vma) 999 { 1000 /* 1001 * The list_is_singular() test is to avoid merging VMA cloned from 1002 * parents. This can improve scalability caused by anon_vma lock. 1003 */ 1004 if ((!anon_vma1 || !anon_vma2) && (!vma || 1005 list_is_singular(&vma->anon_vma_chain))) 1006 return 1; 1007 return anon_vma1 == anon_vma2; 1008 } 1009 1010 /* 1011 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 1012 * in front of (at a lower virtual address and file offset than) the vma. 1013 * 1014 * We cannot merge two vmas if they have differently assigned (non-NULL) 1015 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 1016 * 1017 * We don't check here for the merged mmap wrapping around the end of pagecache 1018 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which 1019 * wrap, nor mmaps which cover the final page at index -1UL. 1020 */ 1021 static int 1022 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 1023 struct anon_vma *anon_vma, struct file *file, 1024 pgoff_t vm_pgoff, 1025 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1026 { 1027 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) && 1028 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1029 if (vma->vm_pgoff == vm_pgoff) 1030 return 1; 1031 } 1032 return 0; 1033 } 1034 1035 /* 1036 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 1037 * beyond (at a higher virtual address and file offset than) the vma. 1038 * 1039 * We cannot merge two vmas if they have differently assigned (non-NULL) 1040 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 1041 */ 1042 static int 1043 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 1044 struct anon_vma *anon_vma, struct file *file, 1045 pgoff_t vm_pgoff, 1046 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1047 { 1048 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) && 1049 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 1050 pgoff_t vm_pglen; 1051 vm_pglen = vma_pages(vma); 1052 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 1053 return 1; 1054 } 1055 return 0; 1056 } 1057 1058 /* 1059 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out 1060 * whether that can be merged with its predecessor or its successor. 1061 * Or both (it neatly fills a hole). 1062 * 1063 * In most cases - when called for mmap, brk or mremap - [addr,end) is 1064 * certain not to be mapped by the time vma_merge is called; but when 1065 * called for mprotect, it is certain to be already mapped (either at 1066 * an offset within prev, or at the start of next), and the flags of 1067 * this area are about to be changed to vm_flags - and the no-change 1068 * case has already been eliminated. 1069 * 1070 * The following mprotect cases have to be considered, where AAAA is 1071 * the area passed down from mprotect_fixup, never extending beyond one 1072 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 1073 * 1074 * AAAA AAAA AAAA AAAA 1075 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX 1076 * cannot merge might become might become might become 1077 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or 1078 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or 1079 * mremap move: PPPPXXXXXXXX 8 1080 * AAAA 1081 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN 1082 * might become case 1 below case 2 below case 3 below 1083 * 1084 * It is important for case 8 that the the vma NNNN overlapping the 1085 * region AAAA is never going to extended over XXXX. Instead XXXX must 1086 * be extended in region AAAA and NNNN must be removed. This way in 1087 * all cases where vma_merge succeeds, the moment vma_adjust drops the 1088 * rmap_locks, the properties of the merged vma will be already 1089 * correct for the whole merged range. Some of those properties like 1090 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must 1091 * be correct for the whole merged range immediately after the 1092 * rmap_locks are released. Otherwise if XXXX would be removed and 1093 * NNNN would be extended over the XXXX range, remove_migration_ptes 1094 * or other rmap walkers (if working on addresses beyond the "end" 1095 * parameter) may establish ptes with the wrong permissions of NNNN 1096 * instead of the right permissions of XXXX. 1097 */ 1098 struct vm_area_struct *vma_merge(struct mm_struct *mm, 1099 struct vm_area_struct *prev, unsigned long addr, 1100 unsigned long end, unsigned long vm_flags, 1101 struct anon_vma *anon_vma, struct file *file, 1102 pgoff_t pgoff, struct mempolicy *policy, 1103 struct vm_userfaultfd_ctx vm_userfaultfd_ctx) 1104 { 1105 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 1106 struct vm_area_struct *area, *next; 1107 int err; 1108 1109 /* 1110 * We later require that vma->vm_flags == vm_flags, 1111 * so this tests vma->vm_flags & VM_SPECIAL, too. 1112 */ 1113 if (vm_flags & VM_SPECIAL) 1114 return NULL; 1115 1116 if (prev) 1117 next = prev->vm_next; 1118 else 1119 next = mm->mmap; 1120 area = next; 1121 if (area && area->vm_end == end) /* cases 6, 7, 8 */ 1122 next = next->vm_next; 1123 1124 /* verify some invariant that must be enforced by the caller */ 1125 VM_WARN_ON(prev && addr <= prev->vm_start); 1126 VM_WARN_ON(area && end > area->vm_end); 1127 VM_WARN_ON(addr >= end); 1128 1129 /* 1130 * Can it merge with the predecessor? 1131 */ 1132 if (prev && prev->vm_end == addr && 1133 mpol_equal(vma_policy(prev), policy) && 1134 can_vma_merge_after(prev, vm_flags, 1135 anon_vma, file, pgoff, 1136 vm_userfaultfd_ctx)) { 1137 /* 1138 * OK, it can. Can we now merge in the successor as well? 1139 */ 1140 if (next && end == next->vm_start && 1141 mpol_equal(policy, vma_policy(next)) && 1142 can_vma_merge_before(next, vm_flags, 1143 anon_vma, file, 1144 pgoff+pglen, 1145 vm_userfaultfd_ctx) && 1146 is_mergeable_anon_vma(prev->anon_vma, 1147 next->anon_vma, NULL)) { 1148 /* cases 1, 6 */ 1149 err = __vma_adjust(prev, prev->vm_start, 1150 next->vm_end, prev->vm_pgoff, NULL, 1151 prev); 1152 } else /* cases 2, 5, 7 */ 1153 err = __vma_adjust(prev, prev->vm_start, 1154 end, prev->vm_pgoff, NULL, prev); 1155 if (err) 1156 return NULL; 1157 khugepaged_enter_vma_merge(prev, vm_flags); 1158 return prev; 1159 } 1160 1161 /* 1162 * Can this new request be merged in front of next? 1163 */ 1164 if (next && end == next->vm_start && 1165 mpol_equal(policy, vma_policy(next)) && 1166 can_vma_merge_before(next, vm_flags, 1167 anon_vma, file, pgoff+pglen, 1168 vm_userfaultfd_ctx)) { 1169 if (prev && addr < prev->vm_end) /* case 4 */ 1170 err = __vma_adjust(prev, prev->vm_start, 1171 addr, prev->vm_pgoff, NULL, next); 1172 else { /* cases 3, 8 */ 1173 err = __vma_adjust(area, addr, next->vm_end, 1174 next->vm_pgoff - pglen, NULL, next); 1175 /* 1176 * In case 3 area is already equal to next and 1177 * this is a noop, but in case 8 "area" has 1178 * been removed and next was expanded over it. 1179 */ 1180 area = next; 1181 } 1182 if (err) 1183 return NULL; 1184 khugepaged_enter_vma_merge(area, vm_flags); 1185 return area; 1186 } 1187 1188 return NULL; 1189 } 1190 1191 /* 1192 * Rough compatbility check to quickly see if it's even worth looking 1193 * at sharing an anon_vma. 1194 * 1195 * They need to have the same vm_file, and the flags can only differ 1196 * in things that mprotect may change. 1197 * 1198 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1199 * we can merge the two vma's. For example, we refuse to merge a vma if 1200 * there is a vm_ops->close() function, because that indicates that the 1201 * driver is doing some kind of reference counting. But that doesn't 1202 * really matter for the anon_vma sharing case. 1203 */ 1204 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1205 { 1206 return a->vm_end == b->vm_start && 1207 mpol_equal(vma_policy(a), vma_policy(b)) && 1208 a->vm_file == b->vm_file && 1209 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) && 1210 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1211 } 1212 1213 /* 1214 * Do some basic sanity checking to see if we can re-use the anon_vma 1215 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1216 * the same as 'old', the other will be the new one that is trying 1217 * to share the anon_vma. 1218 * 1219 * NOTE! This runs with mm_sem held for reading, so it is possible that 1220 * the anon_vma of 'old' is concurrently in the process of being set up 1221 * by another page fault trying to merge _that_. But that's ok: if it 1222 * is being set up, that automatically means that it will be a singleton 1223 * acceptable for merging, so we can do all of this optimistically. But 1224 * we do that READ_ONCE() to make sure that we never re-load the pointer. 1225 * 1226 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1227 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1228 * is to return an anon_vma that is "complex" due to having gone through 1229 * a fork). 1230 * 1231 * We also make sure that the two vma's are compatible (adjacent, 1232 * and with the same memory policies). That's all stable, even with just 1233 * a read lock on the mm_sem. 1234 */ 1235 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1236 { 1237 if (anon_vma_compatible(a, b)) { 1238 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma); 1239 1240 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1241 return anon_vma; 1242 } 1243 return NULL; 1244 } 1245 1246 /* 1247 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1248 * neighbouring vmas for a suitable anon_vma, before it goes off 1249 * to allocate a new anon_vma. It checks because a repetitive 1250 * sequence of mprotects and faults may otherwise lead to distinct 1251 * anon_vmas being allocated, preventing vma merge in subsequent 1252 * mprotect. 1253 */ 1254 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1255 { 1256 struct anon_vma *anon_vma; 1257 struct vm_area_struct *near; 1258 1259 near = vma->vm_next; 1260 if (!near) 1261 goto try_prev; 1262 1263 anon_vma = reusable_anon_vma(near, vma, near); 1264 if (anon_vma) 1265 return anon_vma; 1266 try_prev: 1267 near = vma->vm_prev; 1268 if (!near) 1269 goto none; 1270 1271 anon_vma = reusable_anon_vma(near, near, vma); 1272 if (anon_vma) 1273 return anon_vma; 1274 none: 1275 /* 1276 * There's no absolute need to look only at touching neighbours: 1277 * we could search further afield for "compatible" anon_vmas. 1278 * But it would probably just be a waste of time searching, 1279 * or lead to too many vmas hanging off the same anon_vma. 1280 * We're trying to allow mprotect remerging later on, 1281 * not trying to minimize memory used for anon_vmas. 1282 */ 1283 return NULL; 1284 } 1285 1286 /* 1287 * If a hint addr is less than mmap_min_addr change hint to be as 1288 * low as possible but still greater than mmap_min_addr 1289 */ 1290 static inline unsigned long round_hint_to_min(unsigned long hint) 1291 { 1292 hint &= PAGE_MASK; 1293 if (((void *)hint != NULL) && 1294 (hint < mmap_min_addr)) 1295 return PAGE_ALIGN(mmap_min_addr); 1296 return hint; 1297 } 1298 1299 static inline int mlock_future_check(struct mm_struct *mm, 1300 unsigned long flags, 1301 unsigned long len) 1302 { 1303 unsigned long locked, lock_limit; 1304 1305 /* mlock MCL_FUTURE? */ 1306 if (flags & VM_LOCKED) { 1307 locked = len >> PAGE_SHIFT; 1308 locked += mm->locked_vm; 1309 lock_limit = rlimit(RLIMIT_MEMLOCK); 1310 lock_limit >>= PAGE_SHIFT; 1311 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1312 return -EAGAIN; 1313 } 1314 return 0; 1315 } 1316 1317 /* 1318 * The caller must hold down_write(¤t->mm->mmap_sem). 1319 */ 1320 unsigned long do_mmap(struct file *file, unsigned long addr, 1321 unsigned long len, unsigned long prot, 1322 unsigned long flags, vm_flags_t vm_flags, 1323 unsigned long pgoff, unsigned long *populate, 1324 struct list_head *uf) 1325 { 1326 struct mm_struct *mm = current->mm; 1327 int pkey = 0; 1328 1329 *populate = 0; 1330 1331 if (!len) 1332 return -EINVAL; 1333 1334 /* 1335 * Does the application expect PROT_READ to imply PROT_EXEC? 1336 * 1337 * (the exception is when the underlying filesystem is noexec 1338 * mounted, in which case we dont add PROT_EXEC.) 1339 */ 1340 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1341 if (!(file && path_noexec(&file->f_path))) 1342 prot |= PROT_EXEC; 1343 1344 if (!(flags & MAP_FIXED)) 1345 addr = round_hint_to_min(addr); 1346 1347 /* Careful about overflows.. */ 1348 len = PAGE_ALIGN(len); 1349 if (!len) 1350 return -ENOMEM; 1351 1352 /* offset overflow? */ 1353 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1354 return -EOVERFLOW; 1355 1356 /* Too many mappings? */ 1357 if (mm->map_count > sysctl_max_map_count) 1358 return -ENOMEM; 1359 1360 /* Obtain the address to map to. we verify (or select) it and ensure 1361 * that it represents a valid section of the address space. 1362 */ 1363 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1364 if (offset_in_page(addr)) 1365 return addr; 1366 1367 if (prot == PROT_EXEC) { 1368 pkey = execute_only_pkey(mm); 1369 if (pkey < 0) 1370 pkey = 0; 1371 } 1372 1373 /* Do simple checking here so the lower-level routines won't have 1374 * to. we assume access permissions have been handled by the open 1375 * of the memory object, so we don't do any here. 1376 */ 1377 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) | 1378 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1379 1380 if (flags & MAP_LOCKED) 1381 if (!can_do_mlock()) 1382 return -EPERM; 1383 1384 if (mlock_future_check(mm, vm_flags, len)) 1385 return -EAGAIN; 1386 1387 if (file) { 1388 struct inode *inode = file_inode(file); 1389 1390 switch (flags & MAP_TYPE) { 1391 case MAP_SHARED: 1392 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE)) 1393 return -EACCES; 1394 1395 /* 1396 * Make sure we don't allow writing to an append-only 1397 * file.. 1398 */ 1399 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1400 return -EACCES; 1401 1402 /* 1403 * Make sure there are no mandatory locks on the file. 1404 */ 1405 if (locks_verify_locked(file)) 1406 return -EAGAIN; 1407 1408 vm_flags |= VM_SHARED | VM_MAYSHARE; 1409 if (!(file->f_mode & FMODE_WRITE)) 1410 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1411 1412 /* fall through */ 1413 case MAP_PRIVATE: 1414 if (!(file->f_mode & FMODE_READ)) 1415 return -EACCES; 1416 if (path_noexec(&file->f_path)) { 1417 if (vm_flags & VM_EXEC) 1418 return -EPERM; 1419 vm_flags &= ~VM_MAYEXEC; 1420 } 1421 1422 if (!file->f_op->mmap) 1423 return -ENODEV; 1424 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1425 return -EINVAL; 1426 break; 1427 1428 default: 1429 return -EINVAL; 1430 } 1431 } else { 1432 switch (flags & MAP_TYPE) { 1433 case MAP_SHARED: 1434 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1435 return -EINVAL; 1436 /* 1437 * Ignore pgoff. 1438 */ 1439 pgoff = 0; 1440 vm_flags |= VM_SHARED | VM_MAYSHARE; 1441 break; 1442 case MAP_PRIVATE: 1443 /* 1444 * Set pgoff according to addr for anon_vma. 1445 */ 1446 pgoff = addr >> PAGE_SHIFT; 1447 break; 1448 default: 1449 return -EINVAL; 1450 } 1451 } 1452 1453 /* 1454 * Set 'VM_NORESERVE' if we should not account for the 1455 * memory use of this mapping. 1456 */ 1457 if (flags & MAP_NORESERVE) { 1458 /* We honor MAP_NORESERVE if allowed to overcommit */ 1459 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1460 vm_flags |= VM_NORESERVE; 1461 1462 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1463 if (file && is_file_hugepages(file)) 1464 vm_flags |= VM_NORESERVE; 1465 } 1466 1467 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf); 1468 if (!IS_ERR_VALUE(addr) && 1469 ((vm_flags & VM_LOCKED) || 1470 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1471 *populate = len; 1472 return addr; 1473 } 1474 1475 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1476 unsigned long, prot, unsigned long, flags, 1477 unsigned long, fd, unsigned long, pgoff) 1478 { 1479 struct file *file = NULL; 1480 unsigned long retval; 1481 1482 if (!(flags & MAP_ANONYMOUS)) { 1483 audit_mmap_fd(fd, flags); 1484 file = fget(fd); 1485 if (!file) 1486 return -EBADF; 1487 if (is_file_hugepages(file)) 1488 len = ALIGN(len, huge_page_size(hstate_file(file))); 1489 retval = -EINVAL; 1490 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file))) 1491 goto out_fput; 1492 } else if (flags & MAP_HUGETLB) { 1493 struct user_struct *user = NULL; 1494 struct hstate *hs; 1495 1496 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1497 if (!hs) 1498 return -EINVAL; 1499 1500 len = ALIGN(len, huge_page_size(hs)); 1501 /* 1502 * VM_NORESERVE is used because the reservations will be 1503 * taken when vm_ops->mmap() is called 1504 * A dummy user value is used because we are not locking 1505 * memory so no accounting is necessary 1506 */ 1507 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1508 VM_NORESERVE, 1509 &user, HUGETLB_ANONHUGE_INODE, 1510 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1511 if (IS_ERR(file)) 1512 return PTR_ERR(file); 1513 } 1514 1515 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE); 1516 1517 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1518 out_fput: 1519 if (file) 1520 fput(file); 1521 return retval; 1522 } 1523 1524 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1525 struct mmap_arg_struct { 1526 unsigned long addr; 1527 unsigned long len; 1528 unsigned long prot; 1529 unsigned long flags; 1530 unsigned long fd; 1531 unsigned long offset; 1532 }; 1533 1534 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1535 { 1536 struct mmap_arg_struct a; 1537 1538 if (copy_from_user(&a, arg, sizeof(a))) 1539 return -EFAULT; 1540 if (offset_in_page(a.offset)) 1541 return -EINVAL; 1542 1543 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1544 a.offset >> PAGE_SHIFT); 1545 } 1546 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1547 1548 /* 1549 * Some shared mappigns will want the pages marked read-only 1550 * to track write events. If so, we'll downgrade vm_page_prot 1551 * to the private version (using protection_map[] without the 1552 * VM_SHARED bit). 1553 */ 1554 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot) 1555 { 1556 vm_flags_t vm_flags = vma->vm_flags; 1557 const struct vm_operations_struct *vm_ops = vma->vm_ops; 1558 1559 /* If it was private or non-writable, the write bit is already clear */ 1560 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1561 return 0; 1562 1563 /* The backer wishes to know when pages are first written to? */ 1564 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite)) 1565 return 1; 1566 1567 /* The open routine did something to the protections that pgprot_modify 1568 * won't preserve? */ 1569 if (pgprot_val(vm_page_prot) != 1570 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags))) 1571 return 0; 1572 1573 /* Do we need to track softdirty? */ 1574 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY)) 1575 return 1; 1576 1577 /* Specialty mapping? */ 1578 if (vm_flags & VM_PFNMAP) 1579 return 0; 1580 1581 /* Can the mapping track the dirty pages? */ 1582 return vma->vm_file && vma->vm_file->f_mapping && 1583 mapping_cap_account_dirty(vma->vm_file->f_mapping); 1584 } 1585 1586 /* 1587 * We account for memory if it's a private writeable mapping, 1588 * not hugepages and VM_NORESERVE wasn't set. 1589 */ 1590 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1591 { 1592 /* 1593 * hugetlb has its own accounting separate from the core VM 1594 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1595 */ 1596 if (file && is_file_hugepages(file)) 1597 return 0; 1598 1599 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1600 } 1601 1602 unsigned long mmap_region(struct file *file, unsigned long addr, 1603 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff, 1604 struct list_head *uf) 1605 { 1606 struct mm_struct *mm = current->mm; 1607 struct vm_area_struct *vma, *prev; 1608 int error; 1609 struct rb_node **rb_link, *rb_parent; 1610 unsigned long charged = 0; 1611 1612 /* Check against address space limit. */ 1613 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) { 1614 unsigned long nr_pages; 1615 1616 /* 1617 * MAP_FIXED may remove pages of mappings that intersects with 1618 * requested mapping. Account for the pages it would unmap. 1619 */ 1620 nr_pages = count_vma_pages_range(mm, addr, addr + len); 1621 1622 if (!may_expand_vm(mm, vm_flags, 1623 (len >> PAGE_SHIFT) - nr_pages)) 1624 return -ENOMEM; 1625 } 1626 1627 /* Clear old maps */ 1628 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link, 1629 &rb_parent)) { 1630 if (do_munmap(mm, addr, len, uf)) 1631 return -ENOMEM; 1632 } 1633 1634 /* 1635 * Private writable mapping: check memory availability 1636 */ 1637 if (accountable_mapping(file, vm_flags)) { 1638 charged = len >> PAGE_SHIFT; 1639 if (security_vm_enough_memory_mm(mm, charged)) 1640 return -ENOMEM; 1641 vm_flags |= VM_ACCOUNT; 1642 } 1643 1644 /* 1645 * Can we just expand an old mapping? 1646 */ 1647 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, 1648 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX); 1649 if (vma) 1650 goto out; 1651 1652 /* 1653 * Determine the object being mapped and call the appropriate 1654 * specific mapper. the address has already been validated, but 1655 * not unmapped, but the maps are removed from the list. 1656 */ 1657 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 1658 if (!vma) { 1659 error = -ENOMEM; 1660 goto unacct_error; 1661 } 1662 1663 vma->vm_mm = mm; 1664 vma->vm_start = addr; 1665 vma->vm_end = addr + len; 1666 vma->vm_flags = vm_flags; 1667 vma->vm_page_prot = vm_get_page_prot(vm_flags); 1668 vma->vm_pgoff = pgoff; 1669 INIT_LIST_HEAD(&vma->anon_vma_chain); 1670 1671 if (file) { 1672 if (vm_flags & VM_DENYWRITE) { 1673 error = deny_write_access(file); 1674 if (error) 1675 goto free_vma; 1676 } 1677 if (vm_flags & VM_SHARED) { 1678 error = mapping_map_writable(file->f_mapping); 1679 if (error) 1680 goto allow_write_and_free_vma; 1681 } 1682 1683 /* ->mmap() can change vma->vm_file, but must guarantee that 1684 * vma_link() below can deny write-access if VM_DENYWRITE is set 1685 * and map writably if VM_SHARED is set. This usually means the 1686 * new file must not have been exposed to user-space, yet. 1687 */ 1688 vma->vm_file = get_file(file); 1689 error = call_mmap(file, vma); 1690 if (error) 1691 goto unmap_and_free_vma; 1692 1693 /* Can addr have changed?? 1694 * 1695 * Answer: Yes, several device drivers can do it in their 1696 * f_op->mmap method. -DaveM 1697 * Bug: If addr is changed, prev, rb_link, rb_parent should 1698 * be updated for vma_link() 1699 */ 1700 WARN_ON_ONCE(addr != vma->vm_start); 1701 1702 addr = vma->vm_start; 1703 vm_flags = vma->vm_flags; 1704 } else if (vm_flags & VM_SHARED) { 1705 error = shmem_zero_setup(vma); 1706 if (error) 1707 goto free_vma; 1708 } 1709 1710 vma_link(mm, vma, prev, rb_link, rb_parent); 1711 /* Once vma denies write, undo our temporary denial count */ 1712 if (file) { 1713 if (vm_flags & VM_SHARED) 1714 mapping_unmap_writable(file->f_mapping); 1715 if (vm_flags & VM_DENYWRITE) 1716 allow_write_access(file); 1717 } 1718 file = vma->vm_file; 1719 out: 1720 perf_event_mmap(vma); 1721 1722 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT); 1723 if (vm_flags & VM_LOCKED) { 1724 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) || 1725 vma == get_gate_vma(current->mm))) 1726 mm->locked_vm += (len >> PAGE_SHIFT); 1727 else 1728 vma->vm_flags &= VM_LOCKED_CLEAR_MASK; 1729 } 1730 1731 if (file) 1732 uprobe_mmap(vma); 1733 1734 /* 1735 * New (or expanded) vma always get soft dirty status. 1736 * Otherwise user-space soft-dirty page tracker won't 1737 * be able to distinguish situation when vma area unmapped, 1738 * then new mapped in-place (which must be aimed as 1739 * a completely new data area). 1740 */ 1741 vma->vm_flags |= VM_SOFTDIRTY; 1742 1743 vma_set_page_prot(vma); 1744 1745 return addr; 1746 1747 unmap_and_free_vma: 1748 vma->vm_file = NULL; 1749 fput(file); 1750 1751 /* Undo any partial mapping done by a device driver. */ 1752 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end); 1753 charged = 0; 1754 if (vm_flags & VM_SHARED) 1755 mapping_unmap_writable(file->f_mapping); 1756 allow_write_and_free_vma: 1757 if (vm_flags & VM_DENYWRITE) 1758 allow_write_access(file); 1759 free_vma: 1760 kmem_cache_free(vm_area_cachep, vma); 1761 unacct_error: 1762 if (charged) 1763 vm_unacct_memory(charged); 1764 return error; 1765 } 1766 1767 unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1768 { 1769 /* 1770 * We implement the search by looking for an rbtree node that 1771 * immediately follows a suitable gap. That is, 1772 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length; 1773 * - gap_end = vma->vm_start >= info->low_limit + length; 1774 * - gap_end - gap_start >= length 1775 */ 1776 1777 struct mm_struct *mm = current->mm; 1778 struct vm_area_struct *vma; 1779 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1780 1781 /* Adjust search length to account for worst case alignment overhead */ 1782 length = info->length + info->align_mask; 1783 if (length < info->length) 1784 return -ENOMEM; 1785 1786 /* Adjust search limits by the desired length */ 1787 if (info->high_limit < length) 1788 return -ENOMEM; 1789 high_limit = info->high_limit - length; 1790 1791 if (info->low_limit > high_limit) 1792 return -ENOMEM; 1793 low_limit = info->low_limit + length; 1794 1795 /* Check if rbtree root looks promising */ 1796 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1797 goto check_highest; 1798 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1799 if (vma->rb_subtree_gap < length) 1800 goto check_highest; 1801 1802 while (true) { 1803 /* Visit left subtree if it looks promising */ 1804 gap_end = vm_start_gap(vma); 1805 if (gap_end >= low_limit && vma->vm_rb.rb_left) { 1806 struct vm_area_struct *left = 1807 rb_entry(vma->vm_rb.rb_left, 1808 struct vm_area_struct, vm_rb); 1809 if (left->rb_subtree_gap >= length) { 1810 vma = left; 1811 continue; 1812 } 1813 } 1814 1815 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0; 1816 check_current: 1817 /* Check if current node has a suitable gap */ 1818 if (gap_start > high_limit) 1819 return -ENOMEM; 1820 if (gap_end >= low_limit && 1821 gap_end > gap_start && gap_end - gap_start >= length) 1822 goto found; 1823 1824 /* Visit right subtree if it looks promising */ 1825 if (vma->vm_rb.rb_right) { 1826 struct vm_area_struct *right = 1827 rb_entry(vma->vm_rb.rb_right, 1828 struct vm_area_struct, vm_rb); 1829 if (right->rb_subtree_gap >= length) { 1830 vma = right; 1831 continue; 1832 } 1833 } 1834 1835 /* Go back up the rbtree to find next candidate node */ 1836 while (true) { 1837 struct rb_node *prev = &vma->vm_rb; 1838 if (!rb_parent(prev)) 1839 goto check_highest; 1840 vma = rb_entry(rb_parent(prev), 1841 struct vm_area_struct, vm_rb); 1842 if (prev == vma->vm_rb.rb_left) { 1843 gap_start = vm_end_gap(vma->vm_prev); 1844 gap_end = vm_start_gap(vma); 1845 goto check_current; 1846 } 1847 } 1848 } 1849 1850 check_highest: 1851 /* Check highest gap, which does not precede any rbtree node */ 1852 gap_start = mm->highest_vm_end; 1853 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */ 1854 if (gap_start > high_limit) 1855 return -ENOMEM; 1856 1857 found: 1858 /* We found a suitable gap. Clip it with the original low_limit. */ 1859 if (gap_start < info->low_limit) 1860 gap_start = info->low_limit; 1861 1862 /* Adjust gap address to the desired alignment */ 1863 gap_start += (info->align_offset - gap_start) & info->align_mask; 1864 1865 VM_BUG_ON(gap_start + info->length > info->high_limit); 1866 VM_BUG_ON(gap_start + info->length > gap_end); 1867 return gap_start; 1868 } 1869 1870 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1871 { 1872 struct mm_struct *mm = current->mm; 1873 struct vm_area_struct *vma; 1874 unsigned long length, low_limit, high_limit, gap_start, gap_end; 1875 1876 /* Adjust search length to account for worst case alignment overhead */ 1877 length = info->length + info->align_mask; 1878 if (length < info->length) 1879 return -ENOMEM; 1880 1881 /* 1882 * Adjust search limits by the desired length. 1883 * See implementation comment at top of unmapped_area(). 1884 */ 1885 gap_end = info->high_limit; 1886 if (gap_end < length) 1887 return -ENOMEM; 1888 high_limit = gap_end - length; 1889 1890 if (info->low_limit > high_limit) 1891 return -ENOMEM; 1892 low_limit = info->low_limit + length; 1893 1894 /* Check highest gap, which does not precede any rbtree node */ 1895 gap_start = mm->highest_vm_end; 1896 if (gap_start <= high_limit) 1897 goto found_highest; 1898 1899 /* Check if rbtree root looks promising */ 1900 if (RB_EMPTY_ROOT(&mm->mm_rb)) 1901 return -ENOMEM; 1902 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb); 1903 if (vma->rb_subtree_gap < length) 1904 return -ENOMEM; 1905 1906 while (true) { 1907 /* Visit right subtree if it looks promising */ 1908 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0; 1909 if (gap_start <= high_limit && vma->vm_rb.rb_right) { 1910 struct vm_area_struct *right = 1911 rb_entry(vma->vm_rb.rb_right, 1912 struct vm_area_struct, vm_rb); 1913 if (right->rb_subtree_gap >= length) { 1914 vma = right; 1915 continue; 1916 } 1917 } 1918 1919 check_current: 1920 /* Check if current node has a suitable gap */ 1921 gap_end = vm_start_gap(vma); 1922 if (gap_end < low_limit) 1923 return -ENOMEM; 1924 if (gap_start <= high_limit && 1925 gap_end > gap_start && gap_end - gap_start >= length) 1926 goto found; 1927 1928 /* Visit left subtree if it looks promising */ 1929 if (vma->vm_rb.rb_left) { 1930 struct vm_area_struct *left = 1931 rb_entry(vma->vm_rb.rb_left, 1932 struct vm_area_struct, vm_rb); 1933 if (left->rb_subtree_gap >= length) { 1934 vma = left; 1935 continue; 1936 } 1937 } 1938 1939 /* Go back up the rbtree to find next candidate node */ 1940 while (true) { 1941 struct rb_node *prev = &vma->vm_rb; 1942 if (!rb_parent(prev)) 1943 return -ENOMEM; 1944 vma = rb_entry(rb_parent(prev), 1945 struct vm_area_struct, vm_rb); 1946 if (prev == vma->vm_rb.rb_right) { 1947 gap_start = vma->vm_prev ? 1948 vm_end_gap(vma->vm_prev) : 0; 1949 goto check_current; 1950 } 1951 } 1952 } 1953 1954 found: 1955 /* We found a suitable gap. Clip it with the original high_limit. */ 1956 if (gap_end > info->high_limit) 1957 gap_end = info->high_limit; 1958 1959 found_highest: 1960 /* Compute highest gap address at the desired alignment */ 1961 gap_end -= info->length; 1962 gap_end -= (gap_end - info->align_offset) & info->align_mask; 1963 1964 VM_BUG_ON(gap_end < info->low_limit); 1965 VM_BUG_ON(gap_end < gap_start); 1966 return gap_end; 1967 } 1968 1969 /* Get an address range which is currently unmapped. 1970 * For shmat() with addr=0. 1971 * 1972 * Ugly calling convention alert: 1973 * Return value with the low bits set means error value, 1974 * ie 1975 * if (ret & ~PAGE_MASK) 1976 * error = ret; 1977 * 1978 * This function "knows" that -ENOMEM has the bits set. 1979 */ 1980 #ifndef HAVE_ARCH_UNMAPPED_AREA 1981 unsigned long 1982 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1983 unsigned long len, unsigned long pgoff, unsigned long flags) 1984 { 1985 struct mm_struct *mm = current->mm; 1986 struct vm_area_struct *vma, *prev; 1987 struct vm_unmapped_area_info info; 1988 1989 if (len > TASK_SIZE - mmap_min_addr) 1990 return -ENOMEM; 1991 1992 if (flags & MAP_FIXED) 1993 return addr; 1994 1995 if (addr) { 1996 addr = PAGE_ALIGN(addr); 1997 vma = find_vma_prev(mm, addr, &prev); 1998 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 1999 (!vma || addr + len <= vm_start_gap(vma)) && 2000 (!prev || addr >= vm_end_gap(prev))) 2001 return addr; 2002 } 2003 2004 info.flags = 0; 2005 info.length = len; 2006 info.low_limit = mm->mmap_base; 2007 info.high_limit = TASK_SIZE; 2008 info.align_mask = 0; 2009 return vm_unmapped_area(&info); 2010 } 2011 #endif 2012 2013 /* 2014 * This mmap-allocator allocates new areas top-down from below the 2015 * stack's low limit (the base): 2016 */ 2017 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 2018 unsigned long 2019 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 2020 const unsigned long len, const unsigned long pgoff, 2021 const unsigned long flags) 2022 { 2023 struct vm_area_struct *vma, *prev; 2024 struct mm_struct *mm = current->mm; 2025 unsigned long addr = addr0; 2026 struct vm_unmapped_area_info info; 2027 2028 /* requested length too big for entire address space */ 2029 if (len > TASK_SIZE - mmap_min_addr) 2030 return -ENOMEM; 2031 2032 if (flags & MAP_FIXED) 2033 return addr; 2034 2035 /* requesting a specific address */ 2036 if (addr) { 2037 addr = PAGE_ALIGN(addr); 2038 vma = find_vma_prev(mm, addr, &prev); 2039 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr && 2040 (!vma || addr + len <= vm_start_gap(vma)) && 2041 (!prev || addr >= vm_end_gap(prev))) 2042 return addr; 2043 } 2044 2045 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 2046 info.length = len; 2047 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 2048 info.high_limit = mm->mmap_base; 2049 info.align_mask = 0; 2050 addr = vm_unmapped_area(&info); 2051 2052 /* 2053 * A failed mmap() very likely causes application failure, 2054 * so fall back to the bottom-up function here. This scenario 2055 * can happen with large stack limits and large mmap() 2056 * allocations. 2057 */ 2058 if (offset_in_page(addr)) { 2059 VM_BUG_ON(addr != -ENOMEM); 2060 info.flags = 0; 2061 info.low_limit = TASK_UNMAPPED_BASE; 2062 info.high_limit = TASK_SIZE; 2063 addr = vm_unmapped_area(&info); 2064 } 2065 2066 return addr; 2067 } 2068 #endif 2069 2070 unsigned long 2071 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 2072 unsigned long pgoff, unsigned long flags) 2073 { 2074 unsigned long (*get_area)(struct file *, unsigned long, 2075 unsigned long, unsigned long, unsigned long); 2076 2077 unsigned long error = arch_mmap_check(addr, len, flags); 2078 if (error) 2079 return error; 2080 2081 /* Careful about overflows.. */ 2082 if (len > TASK_SIZE) 2083 return -ENOMEM; 2084 2085 get_area = current->mm->get_unmapped_area; 2086 if (file) { 2087 if (file->f_op->get_unmapped_area) 2088 get_area = file->f_op->get_unmapped_area; 2089 } else if (flags & MAP_SHARED) { 2090 /* 2091 * mmap_region() will call shmem_zero_setup() to create a file, 2092 * so use shmem's get_unmapped_area in case it can be huge. 2093 * do_mmap_pgoff() will clear pgoff, so match alignment. 2094 */ 2095 pgoff = 0; 2096 get_area = shmem_get_unmapped_area; 2097 } 2098 2099 addr = get_area(file, addr, len, pgoff, flags); 2100 if (IS_ERR_VALUE(addr)) 2101 return addr; 2102 2103 if (addr > TASK_SIZE - len) 2104 return -ENOMEM; 2105 if (offset_in_page(addr)) 2106 return -EINVAL; 2107 2108 error = security_mmap_addr(addr); 2109 return error ? error : addr; 2110 } 2111 2112 EXPORT_SYMBOL(get_unmapped_area); 2113 2114 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */ 2115 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 2116 { 2117 struct rb_node *rb_node; 2118 struct vm_area_struct *vma; 2119 2120 /* Check the cache first. */ 2121 vma = vmacache_find(mm, addr); 2122 if (likely(vma)) 2123 return vma; 2124 2125 rb_node = mm->mm_rb.rb_node; 2126 2127 while (rb_node) { 2128 struct vm_area_struct *tmp; 2129 2130 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2131 2132 if (tmp->vm_end > addr) { 2133 vma = tmp; 2134 if (tmp->vm_start <= addr) 2135 break; 2136 rb_node = rb_node->rb_left; 2137 } else 2138 rb_node = rb_node->rb_right; 2139 } 2140 2141 if (vma) 2142 vmacache_update(addr, vma); 2143 return vma; 2144 } 2145 2146 EXPORT_SYMBOL(find_vma); 2147 2148 /* 2149 * Same as find_vma, but also return a pointer to the previous VMA in *pprev. 2150 */ 2151 struct vm_area_struct * 2152 find_vma_prev(struct mm_struct *mm, unsigned long addr, 2153 struct vm_area_struct **pprev) 2154 { 2155 struct vm_area_struct *vma; 2156 2157 vma = find_vma(mm, addr); 2158 if (vma) { 2159 *pprev = vma->vm_prev; 2160 } else { 2161 struct rb_node *rb_node = mm->mm_rb.rb_node; 2162 *pprev = NULL; 2163 while (rb_node) { 2164 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb); 2165 rb_node = rb_node->rb_right; 2166 } 2167 } 2168 return vma; 2169 } 2170 2171 /* 2172 * Verify that the stack growth is acceptable and 2173 * update accounting. This is shared with both the 2174 * grow-up and grow-down cases. 2175 */ 2176 static int acct_stack_growth(struct vm_area_struct *vma, 2177 unsigned long size, unsigned long grow) 2178 { 2179 struct mm_struct *mm = vma->vm_mm; 2180 struct rlimit *rlim = current->signal->rlim; 2181 unsigned long new_start; 2182 2183 /* address space limit tests */ 2184 if (!may_expand_vm(mm, vma->vm_flags, grow)) 2185 return -ENOMEM; 2186 2187 /* Stack limit test */ 2188 if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur)) 2189 return -ENOMEM; 2190 2191 /* mlock limit tests */ 2192 if (vma->vm_flags & VM_LOCKED) { 2193 unsigned long locked; 2194 unsigned long limit; 2195 locked = mm->locked_vm + grow; 2196 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur); 2197 limit >>= PAGE_SHIFT; 2198 if (locked > limit && !capable(CAP_IPC_LOCK)) 2199 return -ENOMEM; 2200 } 2201 2202 /* Check to ensure the stack will not grow into a hugetlb-only region */ 2203 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 2204 vma->vm_end - size; 2205 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 2206 return -EFAULT; 2207 2208 /* 2209 * Overcommit.. This must be the final test, as it will 2210 * update security statistics. 2211 */ 2212 if (security_vm_enough_memory_mm(mm, grow)) 2213 return -ENOMEM; 2214 2215 return 0; 2216 } 2217 2218 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 2219 /* 2220 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 2221 * vma is the last one with address > vma->vm_end. Have to extend vma. 2222 */ 2223 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 2224 { 2225 struct mm_struct *mm = vma->vm_mm; 2226 struct vm_area_struct *next; 2227 unsigned long gap_addr; 2228 int error = 0; 2229 2230 if (!(vma->vm_flags & VM_GROWSUP)) 2231 return -EFAULT; 2232 2233 /* Guard against exceeding limits of the address space. */ 2234 address &= PAGE_MASK; 2235 if (address >= TASK_SIZE) 2236 return -ENOMEM; 2237 address += PAGE_SIZE; 2238 2239 /* Enforce stack_guard_gap */ 2240 gap_addr = address + stack_guard_gap; 2241 2242 /* Guard against overflow */ 2243 if (gap_addr < address || gap_addr > TASK_SIZE) 2244 gap_addr = TASK_SIZE; 2245 2246 next = vma->vm_next; 2247 if (next && next->vm_start < gap_addr) { 2248 if (!(next->vm_flags & VM_GROWSUP)) 2249 return -ENOMEM; 2250 /* Check that both stack segments have the same anon_vma? */ 2251 } 2252 2253 /* We must make sure the anon_vma is allocated. */ 2254 if (unlikely(anon_vma_prepare(vma))) 2255 return -ENOMEM; 2256 2257 /* 2258 * vma->vm_start/vm_end cannot change under us because the caller 2259 * is required to hold the mmap_sem in read mode. We need the 2260 * anon_vma lock to serialize against concurrent expand_stacks. 2261 */ 2262 anon_vma_lock_write(vma->anon_vma); 2263 2264 /* Somebody else might have raced and expanded it already */ 2265 if (address > vma->vm_end) { 2266 unsigned long size, grow; 2267 2268 size = address - vma->vm_start; 2269 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2270 2271 error = -ENOMEM; 2272 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2273 error = acct_stack_growth(vma, size, grow); 2274 if (!error) { 2275 /* 2276 * vma_gap_update() doesn't support concurrent 2277 * updates, but we only hold a shared mmap_sem 2278 * lock here, so we need to protect against 2279 * concurrent vma expansions. 2280 * anon_vma_lock_write() doesn't help here, as 2281 * we don't guarantee that all growable vmas 2282 * in a mm share the same root anon vma. 2283 * So, we reuse mm->page_table_lock to guard 2284 * against concurrent vma expansions. 2285 */ 2286 spin_lock(&mm->page_table_lock); 2287 if (vma->vm_flags & VM_LOCKED) 2288 mm->locked_vm += grow; 2289 vm_stat_account(mm, vma->vm_flags, grow); 2290 anon_vma_interval_tree_pre_update_vma(vma); 2291 vma->vm_end = address; 2292 anon_vma_interval_tree_post_update_vma(vma); 2293 if (vma->vm_next) 2294 vma_gap_update(vma->vm_next); 2295 else 2296 mm->highest_vm_end = vm_end_gap(vma); 2297 spin_unlock(&mm->page_table_lock); 2298 2299 perf_event_mmap(vma); 2300 } 2301 } 2302 } 2303 anon_vma_unlock_write(vma->anon_vma); 2304 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2305 validate_mm(mm); 2306 return error; 2307 } 2308 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2309 2310 /* 2311 * vma is the first one with address < vma->vm_start. Have to extend vma. 2312 */ 2313 int expand_downwards(struct vm_area_struct *vma, 2314 unsigned long address) 2315 { 2316 struct mm_struct *mm = vma->vm_mm; 2317 struct vm_area_struct *prev; 2318 unsigned long gap_addr; 2319 int error; 2320 2321 address &= PAGE_MASK; 2322 error = security_mmap_addr(address); 2323 if (error) 2324 return error; 2325 2326 /* Enforce stack_guard_gap */ 2327 gap_addr = address - stack_guard_gap; 2328 if (gap_addr > address) 2329 return -ENOMEM; 2330 prev = vma->vm_prev; 2331 if (prev && prev->vm_end > gap_addr) { 2332 if (!(prev->vm_flags & VM_GROWSDOWN)) 2333 return -ENOMEM; 2334 /* Check that both stack segments have the same anon_vma? */ 2335 } 2336 2337 /* We must make sure the anon_vma is allocated. */ 2338 if (unlikely(anon_vma_prepare(vma))) 2339 return -ENOMEM; 2340 2341 /* 2342 * vma->vm_start/vm_end cannot change under us because the caller 2343 * is required to hold the mmap_sem in read mode. We need the 2344 * anon_vma lock to serialize against concurrent expand_stacks. 2345 */ 2346 anon_vma_lock_write(vma->anon_vma); 2347 2348 /* Somebody else might have raced and expanded it already */ 2349 if (address < vma->vm_start) { 2350 unsigned long size, grow; 2351 2352 size = vma->vm_end - address; 2353 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2354 2355 error = -ENOMEM; 2356 if (grow <= vma->vm_pgoff) { 2357 error = acct_stack_growth(vma, size, grow); 2358 if (!error) { 2359 /* 2360 * vma_gap_update() doesn't support concurrent 2361 * updates, but we only hold a shared mmap_sem 2362 * lock here, so we need to protect against 2363 * concurrent vma expansions. 2364 * anon_vma_lock_write() doesn't help here, as 2365 * we don't guarantee that all growable vmas 2366 * in a mm share the same root anon vma. 2367 * So, we reuse mm->page_table_lock to guard 2368 * against concurrent vma expansions. 2369 */ 2370 spin_lock(&mm->page_table_lock); 2371 if (vma->vm_flags & VM_LOCKED) 2372 mm->locked_vm += grow; 2373 vm_stat_account(mm, vma->vm_flags, grow); 2374 anon_vma_interval_tree_pre_update_vma(vma); 2375 vma->vm_start = address; 2376 vma->vm_pgoff -= grow; 2377 anon_vma_interval_tree_post_update_vma(vma); 2378 vma_gap_update(vma); 2379 spin_unlock(&mm->page_table_lock); 2380 2381 perf_event_mmap(vma); 2382 } 2383 } 2384 } 2385 anon_vma_unlock_write(vma->anon_vma); 2386 khugepaged_enter_vma_merge(vma, vma->vm_flags); 2387 validate_mm(mm); 2388 return error; 2389 } 2390 2391 /* enforced gap between the expanding stack and other mappings. */ 2392 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT; 2393 2394 static int __init cmdline_parse_stack_guard_gap(char *p) 2395 { 2396 unsigned long val; 2397 char *endptr; 2398 2399 val = simple_strtoul(p, &endptr, 10); 2400 if (!*endptr) 2401 stack_guard_gap = val << PAGE_SHIFT; 2402 2403 return 0; 2404 } 2405 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap); 2406 2407 #ifdef CONFIG_STACK_GROWSUP 2408 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2409 { 2410 return expand_upwards(vma, address); 2411 } 2412 2413 struct vm_area_struct * 2414 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2415 { 2416 struct vm_area_struct *vma, *prev; 2417 2418 addr &= PAGE_MASK; 2419 vma = find_vma_prev(mm, addr, &prev); 2420 if (vma && (vma->vm_start <= addr)) 2421 return vma; 2422 if (!prev || expand_stack(prev, addr)) 2423 return NULL; 2424 if (prev->vm_flags & VM_LOCKED) 2425 populate_vma_page_range(prev, addr, prev->vm_end, NULL); 2426 return prev; 2427 } 2428 #else 2429 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2430 { 2431 return expand_downwards(vma, address); 2432 } 2433 2434 struct vm_area_struct * 2435 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2436 { 2437 struct vm_area_struct *vma; 2438 unsigned long start; 2439 2440 addr &= PAGE_MASK; 2441 vma = find_vma(mm, addr); 2442 if (!vma) 2443 return NULL; 2444 if (vma->vm_start <= addr) 2445 return vma; 2446 if (!(vma->vm_flags & VM_GROWSDOWN)) 2447 return NULL; 2448 start = vma->vm_start; 2449 if (expand_stack(vma, addr)) 2450 return NULL; 2451 if (vma->vm_flags & VM_LOCKED) 2452 populate_vma_page_range(vma, addr, start, NULL); 2453 return vma; 2454 } 2455 #endif 2456 2457 EXPORT_SYMBOL_GPL(find_extend_vma); 2458 2459 /* 2460 * Ok - we have the memory areas we should free on the vma list, 2461 * so release them, and do the vma updates. 2462 * 2463 * Called with the mm semaphore held. 2464 */ 2465 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma) 2466 { 2467 unsigned long nr_accounted = 0; 2468 2469 /* Update high watermark before we lower total_vm */ 2470 update_hiwater_vm(mm); 2471 do { 2472 long nrpages = vma_pages(vma); 2473 2474 if (vma->vm_flags & VM_ACCOUNT) 2475 nr_accounted += nrpages; 2476 vm_stat_account(mm, vma->vm_flags, -nrpages); 2477 vma = remove_vma(vma); 2478 } while (vma); 2479 vm_unacct_memory(nr_accounted); 2480 validate_mm(mm); 2481 } 2482 2483 /* 2484 * Get rid of page table information in the indicated region. 2485 * 2486 * Called with the mm semaphore held. 2487 */ 2488 static void unmap_region(struct mm_struct *mm, 2489 struct vm_area_struct *vma, struct vm_area_struct *prev, 2490 unsigned long start, unsigned long end) 2491 { 2492 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap; 2493 struct mmu_gather tlb; 2494 2495 lru_add_drain(); 2496 tlb_gather_mmu(&tlb, mm, start, end); 2497 update_hiwater_rss(mm); 2498 unmap_vmas(&tlb, vma, start, end); 2499 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2500 next ? next->vm_start : USER_PGTABLES_CEILING); 2501 tlb_finish_mmu(&tlb, start, end); 2502 } 2503 2504 /* 2505 * Create a list of vma's touched by the unmap, removing them from the mm's 2506 * vma list as we go.. 2507 */ 2508 static void 2509 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma, 2510 struct vm_area_struct *prev, unsigned long end) 2511 { 2512 struct vm_area_struct **insertion_point; 2513 struct vm_area_struct *tail_vma = NULL; 2514 2515 insertion_point = (prev ? &prev->vm_next : &mm->mmap); 2516 vma->vm_prev = NULL; 2517 do { 2518 vma_rb_erase(vma, &mm->mm_rb); 2519 mm->map_count--; 2520 tail_vma = vma; 2521 vma = vma->vm_next; 2522 } while (vma && vma->vm_start < end); 2523 *insertion_point = vma; 2524 if (vma) { 2525 vma->vm_prev = prev; 2526 vma_gap_update(vma); 2527 } else 2528 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0; 2529 tail_vma->vm_next = NULL; 2530 2531 /* Kill the cache */ 2532 vmacache_invalidate(mm); 2533 } 2534 2535 /* 2536 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it 2537 * has already been checked or doesn't make sense to fail. 2538 */ 2539 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2540 unsigned long addr, int new_below) 2541 { 2542 struct vm_area_struct *new; 2543 int err; 2544 2545 if (is_vm_hugetlb_page(vma) && (addr & 2546 ~(huge_page_mask(hstate_vma(vma))))) 2547 return -EINVAL; 2548 2549 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 2550 if (!new) 2551 return -ENOMEM; 2552 2553 /* most fields are the same, copy all, and then fixup */ 2554 *new = *vma; 2555 2556 INIT_LIST_HEAD(&new->anon_vma_chain); 2557 2558 if (new_below) 2559 new->vm_end = addr; 2560 else { 2561 new->vm_start = addr; 2562 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2563 } 2564 2565 err = vma_dup_policy(vma, new); 2566 if (err) 2567 goto out_free_vma; 2568 2569 err = anon_vma_clone(new, vma); 2570 if (err) 2571 goto out_free_mpol; 2572 2573 if (new->vm_file) 2574 get_file(new->vm_file); 2575 2576 if (new->vm_ops && new->vm_ops->open) 2577 new->vm_ops->open(new); 2578 2579 if (new_below) 2580 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff + 2581 ((addr - new->vm_start) >> PAGE_SHIFT), new); 2582 else 2583 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new); 2584 2585 /* Success. */ 2586 if (!err) 2587 return 0; 2588 2589 /* Clean everything up if vma_adjust failed. */ 2590 if (new->vm_ops && new->vm_ops->close) 2591 new->vm_ops->close(new); 2592 if (new->vm_file) 2593 fput(new->vm_file); 2594 unlink_anon_vmas(new); 2595 out_free_mpol: 2596 mpol_put(vma_policy(new)); 2597 out_free_vma: 2598 kmem_cache_free(vm_area_cachep, new); 2599 return err; 2600 } 2601 2602 /* 2603 * Split a vma into two pieces at address 'addr', a new vma is allocated 2604 * either for the first part or the tail. 2605 */ 2606 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma, 2607 unsigned long addr, int new_below) 2608 { 2609 if (mm->map_count >= sysctl_max_map_count) 2610 return -ENOMEM; 2611 2612 return __split_vma(mm, vma, addr, new_below); 2613 } 2614 2615 /* Munmap is split into 2 main parts -- this part which finds 2616 * what needs doing, and the areas themselves, which do the 2617 * work. This now handles partial unmappings. 2618 * Jeremy Fitzhardinge <jeremy@goop.org> 2619 */ 2620 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len, 2621 struct list_head *uf) 2622 { 2623 unsigned long end; 2624 struct vm_area_struct *vma, *prev, *last; 2625 2626 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start) 2627 return -EINVAL; 2628 2629 len = PAGE_ALIGN(len); 2630 if (len == 0) 2631 return -EINVAL; 2632 2633 /* Find the first overlapping VMA */ 2634 vma = find_vma(mm, start); 2635 if (!vma) 2636 return 0; 2637 prev = vma->vm_prev; 2638 /* we have start < vma->vm_end */ 2639 2640 /* if it doesn't overlap, we have nothing.. */ 2641 end = start + len; 2642 if (vma->vm_start >= end) 2643 return 0; 2644 2645 if (uf) { 2646 int error = userfaultfd_unmap_prep(vma, start, end, uf); 2647 2648 if (error) 2649 return error; 2650 } 2651 2652 /* 2653 * If we need to split any vma, do it now to save pain later. 2654 * 2655 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2656 * unmapped vm_area_struct will remain in use: so lower split_vma 2657 * places tmp vma above, and higher split_vma places tmp vma below. 2658 */ 2659 if (start > vma->vm_start) { 2660 int error; 2661 2662 /* 2663 * Make sure that map_count on return from munmap() will 2664 * not exceed its limit; but let map_count go just above 2665 * its limit temporarily, to help free resources as expected. 2666 */ 2667 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2668 return -ENOMEM; 2669 2670 error = __split_vma(mm, vma, start, 0); 2671 if (error) 2672 return error; 2673 prev = vma; 2674 } 2675 2676 /* Does it split the last one? */ 2677 last = find_vma(mm, end); 2678 if (last && end > last->vm_start) { 2679 int error = __split_vma(mm, last, end, 1); 2680 if (error) 2681 return error; 2682 } 2683 vma = prev ? prev->vm_next : mm->mmap; 2684 2685 /* 2686 * unlock any mlock()ed ranges before detaching vmas 2687 */ 2688 if (mm->locked_vm) { 2689 struct vm_area_struct *tmp = vma; 2690 while (tmp && tmp->vm_start < end) { 2691 if (tmp->vm_flags & VM_LOCKED) { 2692 mm->locked_vm -= vma_pages(tmp); 2693 munlock_vma_pages_all(tmp); 2694 } 2695 tmp = tmp->vm_next; 2696 } 2697 } 2698 2699 /* 2700 * Remove the vma's, and unmap the actual pages 2701 */ 2702 detach_vmas_to_be_unmapped(mm, vma, prev, end); 2703 unmap_region(mm, vma, prev, start, end); 2704 2705 arch_unmap(mm, vma, start, end); 2706 2707 /* Fix up all other VM information */ 2708 remove_vma_list(mm, vma); 2709 2710 return 0; 2711 } 2712 2713 int vm_munmap(unsigned long start, size_t len) 2714 { 2715 int ret; 2716 struct mm_struct *mm = current->mm; 2717 LIST_HEAD(uf); 2718 2719 if (down_write_killable(&mm->mmap_sem)) 2720 return -EINTR; 2721 2722 ret = do_munmap(mm, start, len, &uf); 2723 up_write(&mm->mmap_sem); 2724 userfaultfd_unmap_complete(mm, &uf); 2725 return ret; 2726 } 2727 EXPORT_SYMBOL(vm_munmap); 2728 2729 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2730 { 2731 profile_munmap(addr); 2732 return vm_munmap(addr, len); 2733 } 2734 2735 2736 /* 2737 * Emulation of deprecated remap_file_pages() syscall. 2738 */ 2739 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size, 2740 unsigned long, prot, unsigned long, pgoff, unsigned long, flags) 2741 { 2742 2743 struct mm_struct *mm = current->mm; 2744 struct vm_area_struct *vma; 2745 unsigned long populate = 0; 2746 unsigned long ret = -EINVAL; 2747 struct file *file; 2748 2749 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n", 2750 current->comm, current->pid); 2751 2752 if (prot) 2753 return ret; 2754 start = start & PAGE_MASK; 2755 size = size & PAGE_MASK; 2756 2757 if (start + size <= start) 2758 return ret; 2759 2760 /* Does pgoff wrap? */ 2761 if (pgoff + (size >> PAGE_SHIFT) < pgoff) 2762 return ret; 2763 2764 if (down_write_killable(&mm->mmap_sem)) 2765 return -EINTR; 2766 2767 vma = find_vma(mm, start); 2768 2769 if (!vma || !(vma->vm_flags & VM_SHARED)) 2770 goto out; 2771 2772 if (start < vma->vm_start) 2773 goto out; 2774 2775 if (start + size > vma->vm_end) { 2776 struct vm_area_struct *next; 2777 2778 for (next = vma->vm_next; next; next = next->vm_next) { 2779 /* hole between vmas ? */ 2780 if (next->vm_start != next->vm_prev->vm_end) 2781 goto out; 2782 2783 if (next->vm_file != vma->vm_file) 2784 goto out; 2785 2786 if (next->vm_flags != vma->vm_flags) 2787 goto out; 2788 2789 if (start + size <= next->vm_end) 2790 break; 2791 } 2792 2793 if (!next) 2794 goto out; 2795 } 2796 2797 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0; 2798 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0; 2799 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0; 2800 2801 flags &= MAP_NONBLOCK; 2802 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE; 2803 if (vma->vm_flags & VM_LOCKED) { 2804 struct vm_area_struct *tmp; 2805 flags |= MAP_LOCKED; 2806 2807 /* drop PG_Mlocked flag for over-mapped range */ 2808 for (tmp = vma; tmp->vm_start >= start + size; 2809 tmp = tmp->vm_next) { 2810 /* 2811 * Split pmd and munlock page on the border 2812 * of the range. 2813 */ 2814 vma_adjust_trans_huge(tmp, start, start + size, 0); 2815 2816 munlock_vma_pages_range(tmp, 2817 max(tmp->vm_start, start), 2818 min(tmp->vm_end, start + size)); 2819 } 2820 } 2821 2822 file = get_file(vma->vm_file); 2823 ret = do_mmap_pgoff(vma->vm_file, start, size, 2824 prot, flags, pgoff, &populate, NULL); 2825 fput(file); 2826 out: 2827 up_write(&mm->mmap_sem); 2828 if (populate) 2829 mm_populate(ret, populate); 2830 if (!IS_ERR_VALUE(ret)) 2831 ret = 0; 2832 return ret; 2833 } 2834 2835 static inline void verify_mm_writelocked(struct mm_struct *mm) 2836 { 2837 #ifdef CONFIG_DEBUG_VM 2838 if (unlikely(down_read_trylock(&mm->mmap_sem))) { 2839 WARN_ON(1); 2840 up_read(&mm->mmap_sem); 2841 } 2842 #endif 2843 } 2844 2845 /* 2846 * this is really a simplified "do_mmap". it only handles 2847 * anonymous maps. eventually we may be able to do some 2848 * brk-specific accounting here. 2849 */ 2850 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, struct list_head *uf) 2851 { 2852 struct mm_struct *mm = current->mm; 2853 struct vm_area_struct *vma, *prev; 2854 unsigned long len; 2855 struct rb_node **rb_link, *rb_parent; 2856 pgoff_t pgoff = addr >> PAGE_SHIFT; 2857 int error; 2858 2859 len = PAGE_ALIGN(request); 2860 if (len < request) 2861 return -ENOMEM; 2862 if (!len) 2863 return 0; 2864 2865 /* Until we need other flags, refuse anything except VM_EXEC. */ 2866 if ((flags & (~VM_EXEC)) != 0) 2867 return -EINVAL; 2868 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2869 2870 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 2871 if (offset_in_page(error)) 2872 return error; 2873 2874 error = mlock_future_check(mm, mm->def_flags, len); 2875 if (error) 2876 return error; 2877 2878 /* 2879 * mm->mmap_sem is required to protect against another thread 2880 * changing the mappings in case we sleep. 2881 */ 2882 verify_mm_writelocked(mm); 2883 2884 /* 2885 * Clear old maps. this also does some error checking for us 2886 */ 2887 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link, 2888 &rb_parent)) { 2889 if (do_munmap(mm, addr, len, uf)) 2890 return -ENOMEM; 2891 } 2892 2893 /* Check against address space limits *after* clearing old maps... */ 2894 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT)) 2895 return -ENOMEM; 2896 2897 if (mm->map_count > sysctl_max_map_count) 2898 return -ENOMEM; 2899 2900 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 2901 return -ENOMEM; 2902 2903 /* Can we just expand an old private anonymous mapping? */ 2904 vma = vma_merge(mm, prev, addr, addr + len, flags, 2905 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX); 2906 if (vma) 2907 goto out; 2908 2909 /* 2910 * create a vma struct for an anonymous mapping 2911 */ 2912 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 2913 if (!vma) { 2914 vm_unacct_memory(len >> PAGE_SHIFT); 2915 return -ENOMEM; 2916 } 2917 2918 INIT_LIST_HEAD(&vma->anon_vma_chain); 2919 vma->vm_mm = mm; 2920 vma->vm_start = addr; 2921 vma->vm_end = addr + len; 2922 vma->vm_pgoff = pgoff; 2923 vma->vm_flags = flags; 2924 vma->vm_page_prot = vm_get_page_prot(flags); 2925 vma_link(mm, vma, prev, rb_link, rb_parent); 2926 out: 2927 perf_event_mmap(vma); 2928 mm->total_vm += len >> PAGE_SHIFT; 2929 mm->data_vm += len >> PAGE_SHIFT; 2930 if (flags & VM_LOCKED) 2931 mm->locked_vm += (len >> PAGE_SHIFT); 2932 vma->vm_flags |= VM_SOFTDIRTY; 2933 return 0; 2934 } 2935 2936 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf) 2937 { 2938 return do_brk_flags(addr, len, 0, uf); 2939 } 2940 2941 int vm_brk_flags(unsigned long addr, unsigned long len, unsigned long flags) 2942 { 2943 struct mm_struct *mm = current->mm; 2944 int ret; 2945 bool populate; 2946 LIST_HEAD(uf); 2947 2948 if (down_write_killable(&mm->mmap_sem)) 2949 return -EINTR; 2950 2951 ret = do_brk_flags(addr, len, flags, &uf); 2952 populate = ((mm->def_flags & VM_LOCKED) != 0); 2953 up_write(&mm->mmap_sem); 2954 userfaultfd_unmap_complete(mm, &uf); 2955 if (populate && !ret) 2956 mm_populate(addr, len); 2957 return ret; 2958 } 2959 EXPORT_SYMBOL(vm_brk_flags); 2960 2961 int vm_brk(unsigned long addr, unsigned long len) 2962 { 2963 return vm_brk_flags(addr, len, 0); 2964 } 2965 EXPORT_SYMBOL(vm_brk); 2966 2967 /* Release all mmaps. */ 2968 void exit_mmap(struct mm_struct *mm) 2969 { 2970 struct mmu_gather tlb; 2971 struct vm_area_struct *vma; 2972 unsigned long nr_accounted = 0; 2973 2974 /* mm's last user has gone, and its about to be pulled down */ 2975 mmu_notifier_release(mm); 2976 2977 if (mm->locked_vm) { 2978 vma = mm->mmap; 2979 while (vma) { 2980 if (vma->vm_flags & VM_LOCKED) 2981 munlock_vma_pages_all(vma); 2982 vma = vma->vm_next; 2983 } 2984 } 2985 2986 arch_exit_mmap(mm); 2987 2988 vma = mm->mmap; 2989 if (!vma) /* Can happen if dup_mmap() received an OOM */ 2990 return; 2991 2992 lru_add_drain(); 2993 flush_cache_mm(mm); 2994 tlb_gather_mmu(&tlb, mm, 0, -1); 2995 /* update_hiwater_rss(mm) here? but nobody should be looking */ 2996 /* Use -1 here to ensure all VMAs in the mm are unmapped */ 2997 unmap_vmas(&tlb, vma, 0, -1); 2998 2999 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING); 3000 tlb_finish_mmu(&tlb, 0, -1); 3001 3002 /* 3003 * Walk the list again, actually closing and freeing it, 3004 * with preemption enabled, without holding any MM locks. 3005 */ 3006 while (vma) { 3007 if (vma->vm_flags & VM_ACCOUNT) 3008 nr_accounted += vma_pages(vma); 3009 vma = remove_vma(vma); 3010 } 3011 vm_unacct_memory(nr_accounted); 3012 } 3013 3014 /* Insert vm structure into process list sorted by address 3015 * and into the inode's i_mmap tree. If vm_file is non-NULL 3016 * then i_mmap_rwsem is taken here. 3017 */ 3018 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 3019 { 3020 struct vm_area_struct *prev; 3021 struct rb_node **rb_link, *rb_parent; 3022 3023 if (find_vma_links(mm, vma->vm_start, vma->vm_end, 3024 &prev, &rb_link, &rb_parent)) 3025 return -ENOMEM; 3026 if ((vma->vm_flags & VM_ACCOUNT) && 3027 security_vm_enough_memory_mm(mm, vma_pages(vma))) 3028 return -ENOMEM; 3029 3030 /* 3031 * The vm_pgoff of a purely anonymous vma should be irrelevant 3032 * until its first write fault, when page's anon_vma and index 3033 * are set. But now set the vm_pgoff it will almost certainly 3034 * end up with (unless mremap moves it elsewhere before that 3035 * first wfault), so /proc/pid/maps tells a consistent story. 3036 * 3037 * By setting it to reflect the virtual start address of the 3038 * vma, merges and splits can happen in a seamless way, just 3039 * using the existing file pgoff checks and manipulations. 3040 * Similarly in do_mmap_pgoff and in do_brk. 3041 */ 3042 if (vma_is_anonymous(vma)) { 3043 BUG_ON(vma->anon_vma); 3044 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 3045 } 3046 3047 vma_link(mm, vma, prev, rb_link, rb_parent); 3048 return 0; 3049 } 3050 3051 /* 3052 * Copy the vma structure to a new location in the same mm, 3053 * prior to moving page table entries, to effect an mremap move. 3054 */ 3055 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 3056 unsigned long addr, unsigned long len, pgoff_t pgoff, 3057 bool *need_rmap_locks) 3058 { 3059 struct vm_area_struct *vma = *vmap; 3060 unsigned long vma_start = vma->vm_start; 3061 struct mm_struct *mm = vma->vm_mm; 3062 struct vm_area_struct *new_vma, *prev; 3063 struct rb_node **rb_link, *rb_parent; 3064 bool faulted_in_anon_vma = true; 3065 3066 /* 3067 * If anonymous vma has not yet been faulted, update new pgoff 3068 * to match new location, to increase its chance of merging. 3069 */ 3070 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) { 3071 pgoff = addr >> PAGE_SHIFT; 3072 faulted_in_anon_vma = false; 3073 } 3074 3075 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) 3076 return NULL; /* should never get here */ 3077 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags, 3078 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 3079 vma->vm_userfaultfd_ctx); 3080 if (new_vma) { 3081 /* 3082 * Source vma may have been merged into new_vma 3083 */ 3084 if (unlikely(vma_start >= new_vma->vm_start && 3085 vma_start < new_vma->vm_end)) { 3086 /* 3087 * The only way we can get a vma_merge with 3088 * self during an mremap is if the vma hasn't 3089 * been faulted in yet and we were allowed to 3090 * reset the dst vma->vm_pgoff to the 3091 * destination address of the mremap to allow 3092 * the merge to happen. mremap must change the 3093 * vm_pgoff linearity between src and dst vmas 3094 * (in turn preventing a vma_merge) to be 3095 * safe. It is only safe to keep the vm_pgoff 3096 * linear if there are no pages mapped yet. 3097 */ 3098 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 3099 *vmap = vma = new_vma; 3100 } 3101 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 3102 } else { 3103 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL); 3104 if (!new_vma) 3105 goto out; 3106 *new_vma = *vma; 3107 new_vma->vm_start = addr; 3108 new_vma->vm_end = addr + len; 3109 new_vma->vm_pgoff = pgoff; 3110 if (vma_dup_policy(vma, new_vma)) 3111 goto out_free_vma; 3112 INIT_LIST_HEAD(&new_vma->anon_vma_chain); 3113 if (anon_vma_clone(new_vma, vma)) 3114 goto out_free_mempol; 3115 if (new_vma->vm_file) 3116 get_file(new_vma->vm_file); 3117 if (new_vma->vm_ops && new_vma->vm_ops->open) 3118 new_vma->vm_ops->open(new_vma); 3119 vma_link(mm, new_vma, prev, rb_link, rb_parent); 3120 *need_rmap_locks = false; 3121 } 3122 return new_vma; 3123 3124 out_free_mempol: 3125 mpol_put(vma_policy(new_vma)); 3126 out_free_vma: 3127 kmem_cache_free(vm_area_cachep, new_vma); 3128 out: 3129 return NULL; 3130 } 3131 3132 /* 3133 * Return true if the calling process may expand its vm space by the passed 3134 * number of pages 3135 */ 3136 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages) 3137 { 3138 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT) 3139 return false; 3140 3141 if (is_data_mapping(flags) && 3142 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) { 3143 /* Workaround for Valgrind */ 3144 if (rlimit(RLIMIT_DATA) == 0 && 3145 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT) 3146 return true; 3147 if (!ignore_rlimit_data) { 3148 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n", 3149 current->comm, current->pid, 3150 (mm->data_vm + npages) << PAGE_SHIFT, 3151 rlimit(RLIMIT_DATA)); 3152 return false; 3153 } 3154 } 3155 3156 return true; 3157 } 3158 3159 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages) 3160 { 3161 mm->total_vm += npages; 3162 3163 if (is_exec_mapping(flags)) 3164 mm->exec_vm += npages; 3165 else if (is_stack_mapping(flags)) 3166 mm->stack_vm += npages; 3167 else if (is_data_mapping(flags)) 3168 mm->data_vm += npages; 3169 } 3170 3171 static int special_mapping_fault(struct vm_fault *vmf); 3172 3173 /* 3174 * Having a close hook prevents vma merging regardless of flags. 3175 */ 3176 static void special_mapping_close(struct vm_area_struct *vma) 3177 { 3178 } 3179 3180 static const char *special_mapping_name(struct vm_area_struct *vma) 3181 { 3182 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 3183 } 3184 3185 static int special_mapping_mremap(struct vm_area_struct *new_vma) 3186 { 3187 struct vm_special_mapping *sm = new_vma->vm_private_data; 3188 3189 if (sm->mremap) 3190 return sm->mremap(sm, new_vma); 3191 return 0; 3192 } 3193 3194 static const struct vm_operations_struct special_mapping_vmops = { 3195 .close = special_mapping_close, 3196 .fault = special_mapping_fault, 3197 .mremap = special_mapping_mremap, 3198 .name = special_mapping_name, 3199 }; 3200 3201 static const struct vm_operations_struct legacy_special_mapping_vmops = { 3202 .close = special_mapping_close, 3203 .fault = special_mapping_fault, 3204 }; 3205 3206 static int special_mapping_fault(struct vm_fault *vmf) 3207 { 3208 struct vm_area_struct *vma = vmf->vma; 3209 pgoff_t pgoff; 3210 struct page **pages; 3211 3212 if (vma->vm_ops == &legacy_special_mapping_vmops) { 3213 pages = vma->vm_private_data; 3214 } else { 3215 struct vm_special_mapping *sm = vma->vm_private_data; 3216 3217 if (sm->fault) 3218 return sm->fault(sm, vmf->vma, vmf); 3219 3220 pages = sm->pages; 3221 } 3222 3223 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages) 3224 pgoff--; 3225 3226 if (*pages) { 3227 struct page *page = *pages; 3228 get_page(page); 3229 vmf->page = page; 3230 return 0; 3231 } 3232 3233 return VM_FAULT_SIGBUS; 3234 } 3235 3236 static struct vm_area_struct *__install_special_mapping( 3237 struct mm_struct *mm, 3238 unsigned long addr, unsigned long len, 3239 unsigned long vm_flags, void *priv, 3240 const struct vm_operations_struct *ops) 3241 { 3242 int ret; 3243 struct vm_area_struct *vma; 3244 3245 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL); 3246 if (unlikely(vma == NULL)) 3247 return ERR_PTR(-ENOMEM); 3248 3249 INIT_LIST_HEAD(&vma->anon_vma_chain); 3250 vma->vm_mm = mm; 3251 vma->vm_start = addr; 3252 vma->vm_end = addr + len; 3253 3254 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY; 3255 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3256 3257 vma->vm_ops = ops; 3258 vma->vm_private_data = priv; 3259 3260 ret = insert_vm_struct(mm, vma); 3261 if (ret) 3262 goto out; 3263 3264 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT); 3265 3266 perf_event_mmap(vma); 3267 3268 return vma; 3269 3270 out: 3271 kmem_cache_free(vm_area_cachep, vma); 3272 return ERR_PTR(ret); 3273 } 3274 3275 bool vma_is_special_mapping(const struct vm_area_struct *vma, 3276 const struct vm_special_mapping *sm) 3277 { 3278 return vma->vm_private_data == sm && 3279 (vma->vm_ops == &special_mapping_vmops || 3280 vma->vm_ops == &legacy_special_mapping_vmops); 3281 } 3282 3283 /* 3284 * Called with mm->mmap_sem held for writing. 3285 * Insert a new vma covering the given region, with the given flags. 3286 * Its pages are supplied by the given array of struct page *. 3287 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3288 * The region past the last page supplied will always produce SIGBUS. 3289 * The array pointer and the pages it points to are assumed to stay alive 3290 * for as long as this mapping might exist. 3291 */ 3292 struct vm_area_struct *_install_special_mapping( 3293 struct mm_struct *mm, 3294 unsigned long addr, unsigned long len, 3295 unsigned long vm_flags, const struct vm_special_mapping *spec) 3296 { 3297 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec, 3298 &special_mapping_vmops); 3299 } 3300 3301 int install_special_mapping(struct mm_struct *mm, 3302 unsigned long addr, unsigned long len, 3303 unsigned long vm_flags, struct page **pages) 3304 { 3305 struct vm_area_struct *vma = __install_special_mapping( 3306 mm, addr, len, vm_flags, (void *)pages, 3307 &legacy_special_mapping_vmops); 3308 3309 return PTR_ERR_OR_ZERO(vma); 3310 } 3311 3312 static DEFINE_MUTEX(mm_all_locks_mutex); 3313 3314 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3315 { 3316 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3317 /* 3318 * The LSB of head.next can't change from under us 3319 * because we hold the mm_all_locks_mutex. 3320 */ 3321 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem); 3322 /* 3323 * We can safely modify head.next after taking the 3324 * anon_vma->root->rwsem. If some other vma in this mm shares 3325 * the same anon_vma we won't take it again. 3326 * 3327 * No need of atomic instructions here, head.next 3328 * can't change from under us thanks to the 3329 * anon_vma->root->rwsem. 3330 */ 3331 if (__test_and_set_bit(0, (unsigned long *) 3332 &anon_vma->root->rb_root.rb_node)) 3333 BUG(); 3334 } 3335 } 3336 3337 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3338 { 3339 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3340 /* 3341 * AS_MM_ALL_LOCKS can't change from under us because 3342 * we hold the mm_all_locks_mutex. 3343 * 3344 * Operations on ->flags have to be atomic because 3345 * even if AS_MM_ALL_LOCKS is stable thanks to the 3346 * mm_all_locks_mutex, there may be other cpus 3347 * changing other bitflags in parallel to us. 3348 */ 3349 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3350 BUG(); 3351 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem); 3352 } 3353 } 3354 3355 /* 3356 * This operation locks against the VM for all pte/vma/mm related 3357 * operations that could ever happen on a certain mm. This includes 3358 * vmtruncate, try_to_unmap, and all page faults. 3359 * 3360 * The caller must take the mmap_sem in write mode before calling 3361 * mm_take_all_locks(). The caller isn't allowed to release the 3362 * mmap_sem until mm_drop_all_locks() returns. 3363 * 3364 * mmap_sem in write mode is required in order to block all operations 3365 * that could modify pagetables and free pages without need of 3366 * altering the vma layout. It's also needed in write mode to avoid new 3367 * anon_vmas to be associated with existing vmas. 3368 * 3369 * A single task can't take more than one mm_take_all_locks() in a row 3370 * or it would deadlock. 3371 * 3372 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3373 * mapping->flags avoid to take the same lock twice, if more than one 3374 * vma in this mm is backed by the same anon_vma or address_space. 3375 * 3376 * We take locks in following order, accordingly to comment at beginning 3377 * of mm/rmap.c: 3378 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for 3379 * hugetlb mapping); 3380 * - all i_mmap_rwsem locks; 3381 * - all anon_vma->rwseml 3382 * 3383 * We can take all locks within these types randomly because the VM code 3384 * doesn't nest them and we protected from parallel mm_take_all_locks() by 3385 * mm_all_locks_mutex. 3386 * 3387 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3388 * that may have to take thousand of locks. 3389 * 3390 * mm_take_all_locks() can fail if it's interrupted by signals. 3391 */ 3392 int mm_take_all_locks(struct mm_struct *mm) 3393 { 3394 struct vm_area_struct *vma; 3395 struct anon_vma_chain *avc; 3396 3397 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3398 3399 mutex_lock(&mm_all_locks_mutex); 3400 3401 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3402 if (signal_pending(current)) 3403 goto out_unlock; 3404 if (vma->vm_file && vma->vm_file->f_mapping && 3405 is_vm_hugetlb_page(vma)) 3406 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3407 } 3408 3409 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3410 if (signal_pending(current)) 3411 goto out_unlock; 3412 if (vma->vm_file && vma->vm_file->f_mapping && 3413 !is_vm_hugetlb_page(vma)) 3414 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3415 } 3416 3417 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3418 if (signal_pending(current)) 3419 goto out_unlock; 3420 if (vma->anon_vma) 3421 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3422 vm_lock_anon_vma(mm, avc->anon_vma); 3423 } 3424 3425 return 0; 3426 3427 out_unlock: 3428 mm_drop_all_locks(mm); 3429 return -EINTR; 3430 } 3431 3432 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3433 { 3434 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) { 3435 /* 3436 * The LSB of head.next can't change to 0 from under 3437 * us because we hold the mm_all_locks_mutex. 3438 * 3439 * We must however clear the bitflag before unlocking 3440 * the vma so the users using the anon_vma->rb_root will 3441 * never see our bitflag. 3442 * 3443 * No need of atomic instructions here, head.next 3444 * can't change from under us until we release the 3445 * anon_vma->root->rwsem. 3446 */ 3447 if (!__test_and_clear_bit(0, (unsigned long *) 3448 &anon_vma->root->rb_root.rb_node)) 3449 BUG(); 3450 anon_vma_unlock_write(anon_vma); 3451 } 3452 } 3453 3454 static void vm_unlock_mapping(struct address_space *mapping) 3455 { 3456 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3457 /* 3458 * AS_MM_ALL_LOCKS can't change to 0 from under us 3459 * because we hold the mm_all_locks_mutex. 3460 */ 3461 i_mmap_unlock_write(mapping); 3462 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3463 &mapping->flags)) 3464 BUG(); 3465 } 3466 } 3467 3468 /* 3469 * The mmap_sem cannot be released by the caller until 3470 * mm_drop_all_locks() returns. 3471 */ 3472 void mm_drop_all_locks(struct mm_struct *mm) 3473 { 3474 struct vm_area_struct *vma; 3475 struct anon_vma_chain *avc; 3476 3477 BUG_ON(down_read_trylock(&mm->mmap_sem)); 3478 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3479 3480 for (vma = mm->mmap; vma; vma = vma->vm_next) { 3481 if (vma->anon_vma) 3482 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3483 vm_unlock_anon_vma(avc->anon_vma); 3484 if (vma->vm_file && vma->vm_file->f_mapping) 3485 vm_unlock_mapping(vma->vm_file->f_mapping); 3486 } 3487 3488 mutex_unlock(&mm_all_locks_mutex); 3489 } 3490 3491 /* 3492 * initialise the percpu counter for VM 3493 */ 3494 void __init mmap_init(void) 3495 { 3496 int ret; 3497 3498 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3499 VM_BUG_ON(ret); 3500 } 3501 3502 /* 3503 * Initialise sysctl_user_reserve_kbytes. 3504 * 3505 * This is intended to prevent a user from starting a single memory hogging 3506 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3507 * mode. 3508 * 3509 * The default value is min(3% of free memory, 128MB) 3510 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3511 */ 3512 static int init_user_reserve(void) 3513 { 3514 unsigned long free_kbytes; 3515 3516 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3517 3518 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3519 return 0; 3520 } 3521 subsys_initcall(init_user_reserve); 3522 3523 /* 3524 * Initialise sysctl_admin_reserve_kbytes. 3525 * 3526 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3527 * to log in and kill a memory hogging process. 3528 * 3529 * Systems with more than 256MB will reserve 8MB, enough to recover 3530 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3531 * only reserve 3% of free pages by default. 3532 */ 3533 static int init_admin_reserve(void) 3534 { 3535 unsigned long free_kbytes; 3536 3537 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3538 3539 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3540 return 0; 3541 } 3542 subsys_initcall(init_admin_reserve); 3543 3544 /* 3545 * Reinititalise user and admin reserves if memory is added or removed. 3546 * 3547 * The default user reserve max is 128MB, and the default max for the 3548 * admin reserve is 8MB. These are usually, but not always, enough to 3549 * enable recovery from a memory hogging process using login/sshd, a shell, 3550 * and tools like top. It may make sense to increase or even disable the 3551 * reserve depending on the existence of swap or variations in the recovery 3552 * tools. So, the admin may have changed them. 3553 * 3554 * If memory is added and the reserves have been eliminated or increased above 3555 * the default max, then we'll trust the admin. 3556 * 3557 * If memory is removed and there isn't enough free memory, then we 3558 * need to reset the reserves. 3559 * 3560 * Otherwise keep the reserve set by the admin. 3561 */ 3562 static int reserve_mem_notifier(struct notifier_block *nb, 3563 unsigned long action, void *data) 3564 { 3565 unsigned long tmp, free_kbytes; 3566 3567 switch (action) { 3568 case MEM_ONLINE: 3569 /* Default max is 128MB. Leave alone if modified by operator. */ 3570 tmp = sysctl_user_reserve_kbytes; 3571 if (0 < tmp && tmp < (1UL << 17)) 3572 init_user_reserve(); 3573 3574 /* Default max is 8MB. Leave alone if modified by operator. */ 3575 tmp = sysctl_admin_reserve_kbytes; 3576 if (0 < tmp && tmp < (1UL << 13)) 3577 init_admin_reserve(); 3578 3579 break; 3580 case MEM_OFFLINE: 3581 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3582 3583 if (sysctl_user_reserve_kbytes > free_kbytes) { 3584 init_user_reserve(); 3585 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3586 sysctl_user_reserve_kbytes); 3587 } 3588 3589 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3590 init_admin_reserve(); 3591 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3592 sysctl_admin_reserve_kbytes); 3593 } 3594 break; 3595 default: 3596 break; 3597 } 3598 return NOTIFY_OK; 3599 } 3600 3601 static struct notifier_block reserve_mem_nb = { 3602 .notifier_call = reserve_mem_notifier, 3603 }; 3604 3605 static int __meminit init_reserve_notifier(void) 3606 { 3607 if (register_hotmemory_notifier(&reserve_mem_nb)) 3608 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3609 3610 return 0; 3611 } 3612 subsys_initcall(init_reserve_notifier); 3613