1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * mm/mmap.c 4 * 5 * Written by obz. 6 * 7 * Address space accounting code <alan@lxorguk.ukuu.org.uk> 8 */ 9 10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 11 12 #include <linux/kernel.h> 13 #include <linux/slab.h> 14 #include <linux/backing-dev.h> 15 #include <linux/mm.h> 16 #include <linux/mm_inline.h> 17 #include <linux/shm.h> 18 #include <linux/mman.h> 19 #include <linux/pagemap.h> 20 #include <linux/swap.h> 21 #include <linux/syscalls.h> 22 #include <linux/capability.h> 23 #include <linux/init.h> 24 #include <linux/file.h> 25 #include <linux/fs.h> 26 #include <linux/personality.h> 27 #include <linux/security.h> 28 #include <linux/hugetlb.h> 29 #include <linux/shmem_fs.h> 30 #include <linux/profile.h> 31 #include <linux/export.h> 32 #include <linux/mount.h> 33 #include <linux/mempolicy.h> 34 #include <linux/rmap.h> 35 #include <linux/mmu_notifier.h> 36 #include <linux/mmdebug.h> 37 #include <linux/perf_event.h> 38 #include <linux/audit.h> 39 #include <linux/khugepaged.h> 40 #include <linux/uprobes.h> 41 #include <linux/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 #include <linux/oom.h> 48 #include <linux/sched/mm.h> 49 50 #include <linux/uaccess.h> 51 #include <asm/cacheflush.h> 52 #include <asm/tlb.h> 53 #include <asm/mmu_context.h> 54 55 #define CREATE_TRACE_POINTS 56 #include <trace/events/mmap.h> 57 58 #include "internal.h" 59 60 #ifndef arch_mmap_check 61 #define arch_mmap_check(addr, len, flags) (0) 62 #endif 63 64 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS 65 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN; 66 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX; 67 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS; 68 #endif 69 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS 70 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN; 71 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX; 72 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS; 73 #endif 74 75 static bool ignore_rlimit_data; 76 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644); 77 78 static void unmap_region(struct mm_struct *mm, struct maple_tree *mt, 79 struct vm_area_struct *vma, struct vm_area_struct *prev, 80 struct vm_area_struct *next, unsigned long start, 81 unsigned long end, bool mm_wr_locked); 82 83 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags) 84 { 85 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags)); 86 } 87 88 /* Update vma->vm_page_prot to reflect vma->vm_flags. */ 89 void vma_set_page_prot(struct vm_area_struct *vma) 90 { 91 unsigned long vm_flags = vma->vm_flags; 92 pgprot_t vm_page_prot; 93 94 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags); 95 if (vma_wants_writenotify(vma, vm_page_prot)) { 96 vm_flags &= ~VM_SHARED; 97 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags); 98 } 99 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */ 100 WRITE_ONCE(vma->vm_page_prot, vm_page_prot); 101 } 102 103 /* 104 * Requires inode->i_mapping->i_mmap_rwsem 105 */ 106 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 107 struct file *file, struct address_space *mapping) 108 { 109 if (vma->vm_flags & VM_SHARED) 110 mapping_unmap_writable(mapping); 111 112 flush_dcache_mmap_lock(mapping); 113 vma_interval_tree_remove(vma, &mapping->i_mmap); 114 flush_dcache_mmap_unlock(mapping); 115 } 116 117 /* 118 * Unlink a file-based vm structure from its interval tree, to hide 119 * vma from rmap and vmtruncate before freeing its page tables. 120 */ 121 void unlink_file_vma(struct vm_area_struct *vma) 122 { 123 struct file *file = vma->vm_file; 124 125 if (file) { 126 struct address_space *mapping = file->f_mapping; 127 i_mmap_lock_write(mapping); 128 __remove_shared_vm_struct(vma, file, mapping); 129 i_mmap_unlock_write(mapping); 130 } 131 } 132 133 /* 134 * Close a vm structure and free it. 135 */ 136 static void remove_vma(struct vm_area_struct *vma) 137 { 138 might_sleep(); 139 if (vma->vm_ops && vma->vm_ops->close) 140 vma->vm_ops->close(vma); 141 if (vma->vm_file) 142 fput(vma->vm_file); 143 mpol_put(vma_policy(vma)); 144 vm_area_free(vma); 145 } 146 147 static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi, 148 unsigned long min) 149 { 150 return mas_prev(&vmi->mas, min); 151 } 152 153 static inline int vma_iter_clear_gfp(struct vma_iterator *vmi, 154 unsigned long start, unsigned long end, gfp_t gfp) 155 { 156 vmi->mas.index = start; 157 vmi->mas.last = end - 1; 158 mas_store_gfp(&vmi->mas, NULL, gfp); 159 if (unlikely(mas_is_err(&vmi->mas))) 160 return -ENOMEM; 161 162 return 0; 163 } 164 165 /* 166 * check_brk_limits() - Use platform specific check of range & verify mlock 167 * limits. 168 * @addr: The address to check 169 * @len: The size of increase. 170 * 171 * Return: 0 on success. 172 */ 173 static int check_brk_limits(unsigned long addr, unsigned long len) 174 { 175 unsigned long mapped_addr; 176 177 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 178 if (IS_ERR_VALUE(mapped_addr)) 179 return mapped_addr; 180 181 return mlock_future_check(current->mm, current->mm->def_flags, len); 182 } 183 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma, 184 unsigned long addr, unsigned long request, unsigned long flags); 185 SYSCALL_DEFINE1(brk, unsigned long, brk) 186 { 187 unsigned long newbrk, oldbrk, origbrk; 188 struct mm_struct *mm = current->mm; 189 struct vm_area_struct *brkvma, *next = NULL; 190 unsigned long min_brk; 191 bool populate; 192 bool downgraded = false; 193 LIST_HEAD(uf); 194 struct vma_iterator vmi; 195 196 if (mmap_write_lock_killable(mm)) 197 return -EINTR; 198 199 origbrk = mm->brk; 200 201 #ifdef CONFIG_COMPAT_BRK 202 /* 203 * CONFIG_COMPAT_BRK can still be overridden by setting 204 * randomize_va_space to 2, which will still cause mm->start_brk 205 * to be arbitrarily shifted 206 */ 207 if (current->brk_randomized) 208 min_brk = mm->start_brk; 209 else 210 min_brk = mm->end_data; 211 #else 212 min_brk = mm->start_brk; 213 #endif 214 if (brk < min_brk) 215 goto out; 216 217 /* 218 * Check against rlimit here. If this check is done later after the test 219 * of oldbrk with newbrk then it can escape the test and let the data 220 * segment grow beyond its set limit the in case where the limit is 221 * not page aligned -Ram Gupta 222 */ 223 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 224 mm->end_data, mm->start_data)) 225 goto out; 226 227 newbrk = PAGE_ALIGN(brk); 228 oldbrk = PAGE_ALIGN(mm->brk); 229 if (oldbrk == newbrk) { 230 mm->brk = brk; 231 goto success; 232 } 233 234 /* 235 * Always allow shrinking brk. 236 * do_vma_munmap() may downgrade mmap_lock to read. 237 */ 238 if (brk <= mm->brk) { 239 int ret; 240 241 /* Search one past newbrk */ 242 vma_iter_init(&vmi, mm, newbrk); 243 brkvma = vma_find(&vmi, oldbrk); 244 if (!brkvma || brkvma->vm_start >= oldbrk) 245 goto out; /* mapping intersects with an existing non-brk vma. */ 246 /* 247 * mm->brk must be protected by write mmap_lock. 248 * do_vma_munmap() may downgrade the lock, so update it 249 * before calling do_vma_munmap(). 250 */ 251 mm->brk = brk; 252 ret = do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true); 253 if (ret == 1) { 254 downgraded = true; 255 goto success; 256 } else if (!ret) 257 goto success; 258 259 mm->brk = origbrk; 260 goto out; 261 } 262 263 if (check_brk_limits(oldbrk, newbrk - oldbrk)) 264 goto out; 265 266 /* 267 * Only check if the next VMA is within the stack_guard_gap of the 268 * expansion area 269 */ 270 vma_iter_init(&vmi, mm, oldbrk); 271 next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap); 272 if (next && newbrk + PAGE_SIZE > vm_start_gap(next)) 273 goto out; 274 275 brkvma = vma_prev_limit(&vmi, mm->start_brk); 276 /* Ok, looks good - let it rip. */ 277 if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0) 278 goto out; 279 280 mm->brk = brk; 281 282 success: 283 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 284 if (downgraded) 285 mmap_read_unlock(mm); 286 else 287 mmap_write_unlock(mm); 288 userfaultfd_unmap_complete(mm, &uf); 289 if (populate) 290 mm_populate(oldbrk, newbrk - oldbrk); 291 return brk; 292 293 out: 294 mmap_write_unlock(mm); 295 return origbrk; 296 } 297 298 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE) 299 extern void mt_validate(struct maple_tree *mt); 300 extern void mt_dump(const struct maple_tree *mt); 301 302 /* Validate the maple tree */ 303 static void validate_mm_mt(struct mm_struct *mm) 304 { 305 struct maple_tree *mt = &mm->mm_mt; 306 struct vm_area_struct *vma_mt; 307 308 MA_STATE(mas, mt, 0, 0); 309 310 mt_validate(&mm->mm_mt); 311 mas_for_each(&mas, vma_mt, ULONG_MAX) { 312 if ((vma_mt->vm_start != mas.index) || 313 (vma_mt->vm_end - 1 != mas.last)) { 314 pr_emerg("issue in %s\n", current->comm); 315 dump_stack(); 316 dump_vma(vma_mt); 317 pr_emerg("mt piv: %p %lu - %lu\n", vma_mt, 318 mas.index, mas.last); 319 pr_emerg("mt vma: %p %lu - %lu\n", vma_mt, 320 vma_mt->vm_start, vma_mt->vm_end); 321 322 mt_dump(mas.tree); 323 if (vma_mt->vm_end != mas.last + 1) { 324 pr_err("vma: %p vma_mt %lu-%lu\tmt %lu-%lu\n", 325 mm, vma_mt->vm_start, vma_mt->vm_end, 326 mas.index, mas.last); 327 mt_dump(mas.tree); 328 } 329 VM_BUG_ON_MM(vma_mt->vm_end != mas.last + 1, mm); 330 if (vma_mt->vm_start != mas.index) { 331 pr_err("vma: %p vma_mt %p %lu - %lu doesn't match\n", 332 mm, vma_mt, vma_mt->vm_start, vma_mt->vm_end); 333 mt_dump(mas.tree); 334 } 335 VM_BUG_ON_MM(vma_mt->vm_start != mas.index, mm); 336 } 337 } 338 } 339 340 static void validate_mm(struct mm_struct *mm) 341 { 342 int bug = 0; 343 int i = 0; 344 struct vm_area_struct *vma; 345 MA_STATE(mas, &mm->mm_mt, 0, 0); 346 347 validate_mm_mt(mm); 348 349 mas_for_each(&mas, vma, ULONG_MAX) { 350 #ifdef CONFIG_DEBUG_VM_RB 351 struct anon_vma *anon_vma = vma->anon_vma; 352 struct anon_vma_chain *avc; 353 354 if (anon_vma) { 355 anon_vma_lock_read(anon_vma); 356 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 357 anon_vma_interval_tree_verify(avc); 358 anon_vma_unlock_read(anon_vma); 359 } 360 #endif 361 i++; 362 } 363 if (i != mm->map_count) { 364 pr_emerg("map_count %d mas_for_each %d\n", mm->map_count, i); 365 bug = 1; 366 } 367 VM_BUG_ON_MM(bug, mm); 368 } 369 370 #else /* !CONFIG_DEBUG_VM_MAPLE_TREE */ 371 #define validate_mm_mt(root) do { } while (0) 372 #define validate_mm(mm) do { } while (0) 373 #endif /* CONFIG_DEBUG_VM_MAPLE_TREE */ 374 375 /* 376 * vma has some anon_vma assigned, and is already inserted on that 377 * anon_vma's interval trees. 378 * 379 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 380 * vma must be removed from the anon_vma's interval trees using 381 * anon_vma_interval_tree_pre_update_vma(). 382 * 383 * After the update, the vma will be reinserted using 384 * anon_vma_interval_tree_post_update_vma(). 385 * 386 * The entire update must be protected by exclusive mmap_lock and by 387 * the root anon_vma's mutex. 388 */ 389 static inline void 390 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 391 { 392 struct anon_vma_chain *avc; 393 394 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 395 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 396 } 397 398 static inline void 399 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 400 { 401 struct anon_vma_chain *avc; 402 403 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 404 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 405 } 406 407 static unsigned long count_vma_pages_range(struct mm_struct *mm, 408 unsigned long addr, unsigned long end) 409 { 410 VMA_ITERATOR(vmi, mm, addr); 411 struct vm_area_struct *vma; 412 unsigned long nr_pages = 0; 413 414 for_each_vma_range(vmi, vma, end) { 415 unsigned long vm_start = max(addr, vma->vm_start); 416 unsigned long vm_end = min(end, vma->vm_end); 417 418 nr_pages += PHYS_PFN(vm_end - vm_start); 419 } 420 421 return nr_pages; 422 } 423 424 static void __vma_link_file(struct vm_area_struct *vma, 425 struct address_space *mapping) 426 { 427 if (vma->vm_flags & VM_SHARED) 428 mapping_allow_writable(mapping); 429 430 flush_dcache_mmap_lock(mapping); 431 vma_interval_tree_insert(vma, &mapping->i_mmap); 432 flush_dcache_mmap_unlock(mapping); 433 } 434 435 static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma) 436 { 437 VMA_ITERATOR(vmi, mm, 0); 438 struct address_space *mapping = NULL; 439 440 if (vma_iter_prealloc(&vmi)) 441 return -ENOMEM; 442 443 if (vma->vm_file) { 444 mapping = vma->vm_file->f_mapping; 445 i_mmap_lock_write(mapping); 446 } 447 448 vma_iter_store(&vmi, vma); 449 450 if (mapping) { 451 __vma_link_file(vma, mapping); 452 i_mmap_unlock_write(mapping); 453 } 454 455 mm->map_count++; 456 validate_mm(mm); 457 return 0; 458 } 459 460 /* 461 * init_multi_vma_prep() - Initializer for struct vma_prepare 462 * @vp: The vma_prepare struct 463 * @vma: The vma that will be altered once locked 464 * @next: The next vma if it is to be adjusted 465 * @remove: The first vma to be removed 466 * @remove2: The second vma to be removed 467 */ 468 static inline void init_multi_vma_prep(struct vma_prepare *vp, 469 struct vm_area_struct *vma, struct vm_area_struct *next, 470 struct vm_area_struct *remove, struct vm_area_struct *remove2) 471 { 472 memset(vp, 0, sizeof(struct vma_prepare)); 473 vp->vma = vma; 474 vp->anon_vma = vma->anon_vma; 475 vp->remove = remove; 476 vp->remove2 = remove2; 477 vp->adj_next = next; 478 if (!vp->anon_vma && next) 479 vp->anon_vma = next->anon_vma; 480 481 vp->file = vma->vm_file; 482 if (vp->file) 483 vp->mapping = vma->vm_file->f_mapping; 484 485 } 486 487 /* 488 * init_vma_prep() - Initializer wrapper for vma_prepare struct 489 * @vp: The vma_prepare struct 490 * @vma: The vma that will be altered once locked 491 */ 492 static inline void init_vma_prep(struct vma_prepare *vp, 493 struct vm_area_struct *vma) 494 { 495 init_multi_vma_prep(vp, vma, NULL, NULL, NULL); 496 } 497 498 499 /* 500 * vma_prepare() - Helper function for handling locking VMAs prior to altering 501 * @vp: The initialized vma_prepare struct 502 */ 503 static inline void vma_prepare(struct vma_prepare *vp) 504 { 505 if (vp->file) { 506 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end); 507 508 if (vp->adj_next) 509 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start, 510 vp->adj_next->vm_end); 511 512 i_mmap_lock_write(vp->mapping); 513 if (vp->insert && vp->insert->vm_file) { 514 /* 515 * Put into interval tree now, so instantiated pages 516 * are visible to arm/parisc __flush_dcache_page 517 * throughout; but we cannot insert into address 518 * space until vma start or end is updated. 519 */ 520 __vma_link_file(vp->insert, 521 vp->insert->vm_file->f_mapping); 522 } 523 } 524 525 if (vp->anon_vma) { 526 anon_vma_lock_write(vp->anon_vma); 527 anon_vma_interval_tree_pre_update_vma(vp->vma); 528 if (vp->adj_next) 529 anon_vma_interval_tree_pre_update_vma(vp->adj_next); 530 } 531 532 if (vp->file) { 533 flush_dcache_mmap_lock(vp->mapping); 534 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap); 535 if (vp->adj_next) 536 vma_interval_tree_remove(vp->adj_next, 537 &vp->mapping->i_mmap); 538 } 539 540 } 541 542 /* 543 * vma_complete- Helper function for handling the unlocking after altering VMAs, 544 * or for inserting a VMA. 545 * 546 * @vp: The vma_prepare struct 547 * @vmi: The vma iterator 548 * @mm: The mm_struct 549 */ 550 static inline void vma_complete(struct vma_prepare *vp, 551 struct vma_iterator *vmi, struct mm_struct *mm) 552 { 553 if (vp->file) { 554 if (vp->adj_next) 555 vma_interval_tree_insert(vp->adj_next, 556 &vp->mapping->i_mmap); 557 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap); 558 flush_dcache_mmap_unlock(vp->mapping); 559 } 560 561 if (vp->remove && vp->file) { 562 __remove_shared_vm_struct(vp->remove, vp->file, vp->mapping); 563 if (vp->remove2) 564 __remove_shared_vm_struct(vp->remove2, vp->file, 565 vp->mapping); 566 } else if (vp->insert) { 567 /* 568 * split_vma has split insert from vma, and needs 569 * us to insert it before dropping the locks 570 * (it may either follow vma or precede it). 571 */ 572 vma_iter_store(vmi, vp->insert); 573 mm->map_count++; 574 } 575 576 if (vp->anon_vma) { 577 anon_vma_interval_tree_post_update_vma(vp->vma); 578 if (vp->adj_next) 579 anon_vma_interval_tree_post_update_vma(vp->adj_next); 580 anon_vma_unlock_write(vp->anon_vma); 581 } 582 583 if (vp->file) { 584 i_mmap_unlock_write(vp->mapping); 585 uprobe_mmap(vp->vma); 586 587 if (vp->adj_next) 588 uprobe_mmap(vp->adj_next); 589 } 590 591 if (vp->remove) { 592 again: 593 if (vp->file) { 594 uprobe_munmap(vp->remove, vp->remove->vm_start, 595 vp->remove->vm_end); 596 fput(vp->file); 597 } 598 if (vp->remove->anon_vma) 599 anon_vma_merge(vp->vma, vp->remove); 600 mm->map_count--; 601 mpol_put(vma_policy(vp->remove)); 602 if (!vp->remove2) 603 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end); 604 vm_area_free(vp->remove); 605 606 /* 607 * In mprotect's case 6 (see comments on vma_merge), 608 * we must remove the one after next as well. 609 */ 610 if (vp->remove2) { 611 vp->remove = vp->remove2; 612 vp->remove2 = NULL; 613 goto again; 614 } 615 } 616 if (vp->insert && vp->file) 617 uprobe_mmap(vp->insert); 618 } 619 620 /* 621 * dup_anon_vma() - Helper function to duplicate anon_vma 622 * @dst: The destination VMA 623 * @src: The source VMA 624 * 625 * Returns: 0 on success. 626 */ 627 static inline int dup_anon_vma(struct vm_area_struct *dst, 628 struct vm_area_struct *src) 629 { 630 /* 631 * Easily overlooked: when mprotect shifts the boundary, make sure the 632 * expanding vma has anon_vma set if the shrinking vma had, to cover any 633 * anon pages imported. 634 */ 635 if (src->anon_vma && !dst->anon_vma) { 636 dst->anon_vma = src->anon_vma; 637 return anon_vma_clone(dst, src); 638 } 639 640 return 0; 641 } 642 643 /* 644 * vma_expand - Expand an existing VMA 645 * 646 * @vmi: The vma iterator 647 * @vma: The vma to expand 648 * @start: The start of the vma 649 * @end: The exclusive end of the vma 650 * @pgoff: The page offset of vma 651 * @next: The current of next vma. 652 * 653 * Expand @vma to @start and @end. Can expand off the start and end. Will 654 * expand over @next if it's different from @vma and @end == @next->vm_end. 655 * Checking if the @vma can expand and merge with @next needs to be handled by 656 * the caller. 657 * 658 * Returns: 0 on success 659 */ 660 int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma, 661 unsigned long start, unsigned long end, pgoff_t pgoff, 662 struct vm_area_struct *next) 663 { 664 bool remove_next = false; 665 struct vma_prepare vp; 666 667 if (next && (vma != next) && (end == next->vm_end)) { 668 int ret; 669 670 remove_next = true; 671 ret = dup_anon_vma(vma, next); 672 if (ret) 673 return ret; 674 } 675 676 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL); 677 /* Not merging but overwriting any part of next is not handled. */ 678 VM_WARN_ON(next && !vp.remove && 679 next != vma && end > next->vm_start); 680 /* Only handles expanding */ 681 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end); 682 683 if (vma_iter_prealloc(vmi)) 684 goto nomem; 685 686 vma_adjust_trans_huge(vma, start, end, 0); 687 /* VMA iterator points to previous, so set to start if necessary */ 688 if (vma_iter_addr(vmi) != start) 689 vma_iter_set(vmi, start); 690 691 vma_prepare(&vp); 692 vma->vm_start = start; 693 vma->vm_end = end; 694 vma->vm_pgoff = pgoff; 695 /* Note: mas must be pointing to the expanding VMA */ 696 vma_iter_store(vmi, vma); 697 698 vma_complete(&vp, vmi, vma->vm_mm); 699 validate_mm(vma->vm_mm); 700 return 0; 701 702 nomem: 703 return -ENOMEM; 704 } 705 706 /* 707 * vma_shrink() - Reduce an existing VMAs memory area 708 * @vmi: The vma iterator 709 * @vma: The VMA to modify 710 * @start: The new start 711 * @end: The new end 712 * 713 * Returns: 0 on success, -ENOMEM otherwise 714 */ 715 int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma, 716 unsigned long start, unsigned long end, pgoff_t pgoff) 717 { 718 struct vma_prepare vp; 719 720 WARN_ON((vma->vm_start != start) && (vma->vm_end != end)); 721 722 if (vma_iter_prealloc(vmi)) 723 return -ENOMEM; 724 725 init_vma_prep(&vp, vma); 726 vma_adjust_trans_huge(vma, start, end, 0); 727 vma_prepare(&vp); 728 729 if (vma->vm_start < start) 730 vma_iter_clear(vmi, vma->vm_start, start); 731 732 if (vma->vm_end > end) 733 vma_iter_clear(vmi, end, vma->vm_end); 734 735 vma->vm_start = start; 736 vma->vm_end = end; 737 vma->vm_pgoff = pgoff; 738 vma_complete(&vp, vmi, vma->vm_mm); 739 validate_mm(vma->vm_mm); 740 return 0; 741 } 742 743 /* 744 * If the vma has a ->close operation then the driver probably needs to release 745 * per-vma resources, so we don't attempt to merge those. 746 */ 747 static inline int is_mergeable_vma(struct vm_area_struct *vma, 748 struct file *file, unsigned long vm_flags, 749 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 750 struct anon_vma_name *anon_name) 751 { 752 /* 753 * VM_SOFTDIRTY should not prevent from VMA merging, if we 754 * match the flags but dirty bit -- the caller should mark 755 * merged VMA as dirty. If dirty bit won't be excluded from 756 * comparison, we increase pressure on the memory system forcing 757 * the kernel to generate new VMAs when old one could be 758 * extended instead. 759 */ 760 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 761 return 0; 762 if (vma->vm_file != file) 763 return 0; 764 if (vma->vm_ops && vma->vm_ops->close) 765 return 0; 766 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx)) 767 return 0; 768 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name)) 769 return 0; 770 return 1; 771 } 772 773 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1, 774 struct anon_vma *anon_vma2, 775 struct vm_area_struct *vma) 776 { 777 /* 778 * The list_is_singular() test is to avoid merging VMA cloned from 779 * parents. This can improve scalability caused by anon_vma lock. 780 */ 781 if ((!anon_vma1 || !anon_vma2) && (!vma || 782 list_is_singular(&vma->anon_vma_chain))) 783 return 1; 784 return anon_vma1 == anon_vma2; 785 } 786 787 /* 788 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 789 * in front of (at a lower virtual address and file offset than) the vma. 790 * 791 * We cannot merge two vmas if they have differently assigned (non-NULL) 792 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 793 * 794 * We don't check here for the merged mmap wrapping around the end of pagecache 795 * indices (16TB on ia32) because do_mmap() does not permit mmap's which 796 * wrap, nor mmaps which cover the final page at index -1UL. 797 */ 798 static int 799 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 800 struct anon_vma *anon_vma, struct file *file, 801 pgoff_t vm_pgoff, 802 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 803 struct anon_vma_name *anon_name) 804 { 805 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) && 806 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 807 if (vma->vm_pgoff == vm_pgoff) 808 return 1; 809 } 810 return 0; 811 } 812 813 /* 814 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 815 * beyond (at a higher virtual address and file offset than) the vma. 816 * 817 * We cannot merge two vmas if they have differently assigned (non-NULL) 818 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 819 */ 820 static int 821 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 822 struct anon_vma *anon_vma, struct file *file, 823 pgoff_t vm_pgoff, 824 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 825 struct anon_vma_name *anon_name) 826 { 827 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) && 828 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 829 pgoff_t vm_pglen; 830 vm_pglen = vma_pages(vma); 831 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 832 return 1; 833 } 834 return 0; 835 } 836 837 /* 838 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name), 839 * figure out whether that can be merged with its predecessor or its 840 * successor. Or both (it neatly fills a hole). 841 * 842 * In most cases - when called for mmap, brk or mremap - [addr,end) is 843 * certain not to be mapped by the time vma_merge is called; but when 844 * called for mprotect, it is certain to be already mapped (either at 845 * an offset within prev, or at the start of next), and the flags of 846 * this area are about to be changed to vm_flags - and the no-change 847 * case has already been eliminated. 848 * 849 * The following mprotect cases have to be considered, where AAAA is 850 * the area passed down from mprotect_fixup, never extending beyond one 851 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after: 852 * 853 * AAAA AAAA AAAA 854 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN 855 * cannot merge might become might become 856 * PPNNNNNNNNNN PPPPPPPPPPNN 857 * mmap, brk or case 4 below case 5 below 858 * mremap move: 859 * AAAA AAAA 860 * PPPP NNNN PPPPNNNNXXXX 861 * might become might become 862 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or 863 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or 864 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8 865 * 866 * It is important for case 8 that the vma NNNN overlapping the 867 * region AAAA is never going to extended over XXXX. Instead XXXX must 868 * be extended in region AAAA and NNNN must be removed. This way in 869 * all cases where vma_merge succeeds, the moment vma_merge drops the 870 * rmap_locks, the properties of the merged vma will be already 871 * correct for the whole merged range. Some of those properties like 872 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must 873 * be correct for the whole merged range immediately after the 874 * rmap_locks are released. Otherwise if XXXX would be removed and 875 * NNNN would be extended over the XXXX range, remove_migration_ptes 876 * or other rmap walkers (if working on addresses beyond the "end" 877 * parameter) may establish ptes with the wrong permissions of NNNN 878 * instead of the right permissions of XXXX. 879 * 880 * In the code below: 881 * PPPP is represented by *prev 882 * NNNN is represented by *mid (and possibly equal to *next) 883 * XXXX is represented by *next or not represented at all. 884 * AAAA is not represented - it will be merged or the function will return NULL 885 */ 886 struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm, 887 struct vm_area_struct *prev, unsigned long addr, 888 unsigned long end, unsigned long vm_flags, 889 struct anon_vma *anon_vma, struct file *file, 890 pgoff_t pgoff, struct mempolicy *policy, 891 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 892 struct anon_vma_name *anon_name) 893 { 894 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 895 pgoff_t vma_pgoff; 896 struct vm_area_struct *mid, *next, *res = NULL; 897 struct vm_area_struct *vma, *adjust, *remove, *remove2; 898 int err = -1; 899 bool merge_prev = false; 900 bool merge_next = false; 901 bool vma_expanded = false; 902 struct vma_prepare vp; 903 unsigned long vma_end = end; 904 long adj_next = 0; 905 unsigned long vma_start = addr; 906 907 validate_mm(mm); 908 /* 909 * We later require that vma->vm_flags == vm_flags, 910 * so this tests vma->vm_flags & VM_SPECIAL, too. 911 */ 912 if (vm_flags & VM_SPECIAL) 913 return NULL; 914 915 next = find_vma(mm, prev ? prev->vm_end : 0); 916 mid = next; 917 if (next && next->vm_end == end) /* cases 6, 7, 8 */ 918 next = find_vma(mm, next->vm_end); 919 920 /* verify some invariant that must be enforced by the caller */ 921 VM_WARN_ON(prev && addr <= prev->vm_start); 922 VM_WARN_ON(mid && end > mid->vm_end); 923 VM_WARN_ON(addr >= end); 924 925 if (prev) { 926 res = prev; 927 vma = prev; 928 vma_start = prev->vm_start; 929 vma_pgoff = prev->vm_pgoff; 930 /* Can we merge the predecessor? */ 931 if (prev->vm_end == addr && mpol_equal(vma_policy(prev), policy) 932 && can_vma_merge_after(prev, vm_flags, anon_vma, file, 933 pgoff, vm_userfaultfd_ctx, anon_name)) { 934 merge_prev = true; 935 vma_prev(vmi); 936 } 937 } 938 /* Can we merge the successor? */ 939 if (next && end == next->vm_start && 940 mpol_equal(policy, vma_policy(next)) && 941 can_vma_merge_before(next, vm_flags, 942 anon_vma, file, pgoff+pglen, 943 vm_userfaultfd_ctx, anon_name)) { 944 merge_next = true; 945 } 946 947 remove = remove2 = adjust = NULL; 948 /* Can we merge both the predecessor and the successor? */ 949 if (merge_prev && merge_next && 950 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) { 951 remove = mid; /* case 1 */ 952 vma_end = next->vm_end; 953 err = dup_anon_vma(res, remove); 954 if (mid != next) { /* case 6 */ 955 remove2 = next; 956 if (!remove->anon_vma) 957 err = dup_anon_vma(res, remove2); 958 } 959 } else if (merge_prev) { 960 err = 0; /* case 2 */ 961 if (mid && end > mid->vm_start) { 962 err = dup_anon_vma(res, mid); 963 if (end == mid->vm_end) { /* case 7 */ 964 remove = mid; 965 } else { /* case 5 */ 966 adjust = mid; 967 adj_next = (end - mid->vm_start); 968 } 969 } 970 } else if (merge_next) { 971 res = next; 972 if (prev && addr < prev->vm_end) { /* case 4 */ 973 vma_end = addr; 974 adjust = mid; 975 adj_next = -(vma->vm_end - addr); 976 err = dup_anon_vma(adjust, prev); 977 } else { 978 vma = next; /* case 3 */ 979 vma_start = addr; 980 vma_end = next->vm_end; 981 vma_pgoff = mid->vm_pgoff; 982 err = 0; 983 if (mid != next) { /* case 8 */ 984 remove = mid; 985 err = dup_anon_vma(res, remove); 986 } 987 } 988 } 989 990 /* Cannot merge or error in anon_vma clone */ 991 if (err) 992 return NULL; 993 994 if (vma_iter_prealloc(vmi)) 995 return NULL; 996 997 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_next); 998 init_multi_vma_prep(&vp, vma, adjust, remove, remove2); 999 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma && 1000 vp.anon_vma != adjust->anon_vma); 1001 1002 vma_prepare(&vp); 1003 if (vma_start < vma->vm_start || vma_end > vma->vm_end) 1004 vma_expanded = true; 1005 1006 vma->vm_start = vma_start; 1007 vma->vm_end = vma_end; 1008 vma->vm_pgoff = vma_pgoff; 1009 1010 if (vma_expanded) 1011 vma_iter_store(vmi, vma); 1012 1013 if (adj_next) { 1014 adjust->vm_start += adj_next; 1015 adjust->vm_pgoff += adj_next >> PAGE_SHIFT; 1016 if (adj_next < 0) { 1017 WARN_ON(vma_expanded); 1018 vma_iter_store(vmi, next); 1019 } 1020 } 1021 1022 vma_complete(&vp, vmi, mm); 1023 vma_iter_free(vmi); 1024 validate_mm(mm); 1025 khugepaged_enter_vma(res, vm_flags); 1026 1027 return res; 1028 } 1029 1030 /* 1031 * Rough compatibility check to quickly see if it's even worth looking 1032 * at sharing an anon_vma. 1033 * 1034 * They need to have the same vm_file, and the flags can only differ 1035 * in things that mprotect may change. 1036 * 1037 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1038 * we can merge the two vma's. For example, we refuse to merge a vma if 1039 * there is a vm_ops->close() function, because that indicates that the 1040 * driver is doing some kind of reference counting. But that doesn't 1041 * really matter for the anon_vma sharing case. 1042 */ 1043 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1044 { 1045 return a->vm_end == b->vm_start && 1046 mpol_equal(vma_policy(a), vma_policy(b)) && 1047 a->vm_file == b->vm_file && 1048 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) && 1049 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1050 } 1051 1052 /* 1053 * Do some basic sanity checking to see if we can re-use the anon_vma 1054 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1055 * the same as 'old', the other will be the new one that is trying 1056 * to share the anon_vma. 1057 * 1058 * NOTE! This runs with mmap_lock held for reading, so it is possible that 1059 * the anon_vma of 'old' is concurrently in the process of being set up 1060 * by another page fault trying to merge _that_. But that's ok: if it 1061 * is being set up, that automatically means that it will be a singleton 1062 * acceptable for merging, so we can do all of this optimistically. But 1063 * we do that READ_ONCE() to make sure that we never re-load the pointer. 1064 * 1065 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1066 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1067 * is to return an anon_vma that is "complex" due to having gone through 1068 * a fork). 1069 * 1070 * We also make sure that the two vma's are compatible (adjacent, 1071 * and with the same memory policies). That's all stable, even with just 1072 * a read lock on the mmap_lock. 1073 */ 1074 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1075 { 1076 if (anon_vma_compatible(a, b)) { 1077 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma); 1078 1079 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1080 return anon_vma; 1081 } 1082 return NULL; 1083 } 1084 1085 /* 1086 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1087 * neighbouring vmas for a suitable anon_vma, before it goes off 1088 * to allocate a new anon_vma. It checks because a repetitive 1089 * sequence of mprotects and faults may otherwise lead to distinct 1090 * anon_vmas being allocated, preventing vma merge in subsequent 1091 * mprotect. 1092 */ 1093 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1094 { 1095 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end); 1096 struct anon_vma *anon_vma = NULL; 1097 struct vm_area_struct *prev, *next; 1098 1099 /* Try next first. */ 1100 next = mas_walk(&mas); 1101 if (next) { 1102 anon_vma = reusable_anon_vma(next, vma, next); 1103 if (anon_vma) 1104 return anon_vma; 1105 } 1106 1107 prev = mas_prev(&mas, 0); 1108 VM_BUG_ON_VMA(prev != vma, vma); 1109 prev = mas_prev(&mas, 0); 1110 /* Try prev next. */ 1111 if (prev) 1112 anon_vma = reusable_anon_vma(prev, prev, vma); 1113 1114 /* 1115 * We might reach here with anon_vma == NULL if we can't find 1116 * any reusable anon_vma. 1117 * There's no absolute need to look only at touching neighbours: 1118 * we could search further afield for "compatible" anon_vmas. 1119 * But it would probably just be a waste of time searching, 1120 * or lead to too many vmas hanging off the same anon_vma. 1121 * We're trying to allow mprotect remerging later on, 1122 * not trying to minimize memory used for anon_vmas. 1123 */ 1124 return anon_vma; 1125 } 1126 1127 /* 1128 * If a hint addr is less than mmap_min_addr change hint to be as 1129 * low as possible but still greater than mmap_min_addr 1130 */ 1131 static inline unsigned long round_hint_to_min(unsigned long hint) 1132 { 1133 hint &= PAGE_MASK; 1134 if (((void *)hint != NULL) && 1135 (hint < mmap_min_addr)) 1136 return PAGE_ALIGN(mmap_min_addr); 1137 return hint; 1138 } 1139 1140 int mlock_future_check(struct mm_struct *mm, unsigned long flags, 1141 unsigned long len) 1142 { 1143 unsigned long locked, lock_limit; 1144 1145 /* mlock MCL_FUTURE? */ 1146 if (flags & VM_LOCKED) { 1147 locked = len >> PAGE_SHIFT; 1148 locked += mm->locked_vm; 1149 lock_limit = rlimit(RLIMIT_MEMLOCK); 1150 lock_limit >>= PAGE_SHIFT; 1151 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) 1152 return -EAGAIN; 1153 } 1154 return 0; 1155 } 1156 1157 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode) 1158 { 1159 if (S_ISREG(inode->i_mode)) 1160 return MAX_LFS_FILESIZE; 1161 1162 if (S_ISBLK(inode->i_mode)) 1163 return MAX_LFS_FILESIZE; 1164 1165 if (S_ISSOCK(inode->i_mode)) 1166 return MAX_LFS_FILESIZE; 1167 1168 /* Special "we do even unsigned file positions" case */ 1169 if (file->f_mode & FMODE_UNSIGNED_OFFSET) 1170 return 0; 1171 1172 /* Yes, random drivers might want more. But I'm tired of buggy drivers */ 1173 return ULONG_MAX; 1174 } 1175 1176 static inline bool file_mmap_ok(struct file *file, struct inode *inode, 1177 unsigned long pgoff, unsigned long len) 1178 { 1179 u64 maxsize = file_mmap_size_max(file, inode); 1180 1181 if (maxsize && len > maxsize) 1182 return false; 1183 maxsize -= len; 1184 if (pgoff > maxsize >> PAGE_SHIFT) 1185 return false; 1186 return true; 1187 } 1188 1189 /* 1190 * The caller must write-lock current->mm->mmap_lock. 1191 */ 1192 unsigned long do_mmap(struct file *file, unsigned long addr, 1193 unsigned long len, unsigned long prot, 1194 unsigned long flags, unsigned long pgoff, 1195 unsigned long *populate, struct list_head *uf) 1196 { 1197 struct mm_struct *mm = current->mm; 1198 vm_flags_t vm_flags; 1199 int pkey = 0; 1200 1201 validate_mm(mm); 1202 *populate = 0; 1203 1204 if (!len) 1205 return -EINVAL; 1206 1207 /* 1208 * Does the application expect PROT_READ to imply PROT_EXEC? 1209 * 1210 * (the exception is when the underlying filesystem is noexec 1211 * mounted, in which case we dont add PROT_EXEC.) 1212 */ 1213 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1214 if (!(file && path_noexec(&file->f_path))) 1215 prot |= PROT_EXEC; 1216 1217 /* force arch specific MAP_FIXED handling in get_unmapped_area */ 1218 if (flags & MAP_FIXED_NOREPLACE) 1219 flags |= MAP_FIXED; 1220 1221 if (!(flags & MAP_FIXED)) 1222 addr = round_hint_to_min(addr); 1223 1224 /* Careful about overflows.. */ 1225 len = PAGE_ALIGN(len); 1226 if (!len) 1227 return -ENOMEM; 1228 1229 /* offset overflow? */ 1230 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1231 return -EOVERFLOW; 1232 1233 /* Too many mappings? */ 1234 if (mm->map_count > sysctl_max_map_count) 1235 return -ENOMEM; 1236 1237 /* Obtain the address to map to. we verify (or select) it and ensure 1238 * that it represents a valid section of the address space. 1239 */ 1240 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1241 if (IS_ERR_VALUE(addr)) 1242 return addr; 1243 1244 if (flags & MAP_FIXED_NOREPLACE) { 1245 if (find_vma_intersection(mm, addr, addr + len)) 1246 return -EEXIST; 1247 } 1248 1249 if (prot == PROT_EXEC) { 1250 pkey = execute_only_pkey(mm); 1251 if (pkey < 0) 1252 pkey = 0; 1253 } 1254 1255 /* Do simple checking here so the lower-level routines won't have 1256 * to. we assume access permissions have been handled by the open 1257 * of the memory object, so we don't do any here. 1258 */ 1259 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) | 1260 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1261 1262 if (flags & MAP_LOCKED) 1263 if (!can_do_mlock()) 1264 return -EPERM; 1265 1266 if (mlock_future_check(mm, vm_flags, len)) 1267 return -EAGAIN; 1268 1269 if (file) { 1270 struct inode *inode = file_inode(file); 1271 unsigned long flags_mask; 1272 1273 if (!file_mmap_ok(file, inode, pgoff, len)) 1274 return -EOVERFLOW; 1275 1276 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags; 1277 1278 switch (flags & MAP_TYPE) { 1279 case MAP_SHARED: 1280 /* 1281 * Force use of MAP_SHARED_VALIDATE with non-legacy 1282 * flags. E.g. MAP_SYNC is dangerous to use with 1283 * MAP_SHARED as you don't know which consistency model 1284 * you will get. We silently ignore unsupported flags 1285 * with MAP_SHARED to preserve backward compatibility. 1286 */ 1287 flags &= LEGACY_MAP_MASK; 1288 fallthrough; 1289 case MAP_SHARED_VALIDATE: 1290 if (flags & ~flags_mask) 1291 return -EOPNOTSUPP; 1292 if (prot & PROT_WRITE) { 1293 if (!(file->f_mode & FMODE_WRITE)) 1294 return -EACCES; 1295 if (IS_SWAPFILE(file->f_mapping->host)) 1296 return -ETXTBSY; 1297 } 1298 1299 /* 1300 * Make sure we don't allow writing to an append-only 1301 * file.. 1302 */ 1303 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1304 return -EACCES; 1305 1306 vm_flags |= VM_SHARED | VM_MAYSHARE; 1307 if (!(file->f_mode & FMODE_WRITE)) 1308 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1309 fallthrough; 1310 case MAP_PRIVATE: 1311 if (!(file->f_mode & FMODE_READ)) 1312 return -EACCES; 1313 if (path_noexec(&file->f_path)) { 1314 if (vm_flags & VM_EXEC) 1315 return -EPERM; 1316 vm_flags &= ~VM_MAYEXEC; 1317 } 1318 1319 if (!file->f_op->mmap) 1320 return -ENODEV; 1321 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1322 return -EINVAL; 1323 break; 1324 1325 default: 1326 return -EINVAL; 1327 } 1328 } else { 1329 switch (flags & MAP_TYPE) { 1330 case MAP_SHARED: 1331 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1332 return -EINVAL; 1333 /* 1334 * Ignore pgoff. 1335 */ 1336 pgoff = 0; 1337 vm_flags |= VM_SHARED | VM_MAYSHARE; 1338 break; 1339 case MAP_PRIVATE: 1340 /* 1341 * Set pgoff according to addr for anon_vma. 1342 */ 1343 pgoff = addr >> PAGE_SHIFT; 1344 break; 1345 default: 1346 return -EINVAL; 1347 } 1348 } 1349 1350 /* 1351 * Set 'VM_NORESERVE' if we should not account for the 1352 * memory use of this mapping. 1353 */ 1354 if (flags & MAP_NORESERVE) { 1355 /* We honor MAP_NORESERVE if allowed to overcommit */ 1356 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1357 vm_flags |= VM_NORESERVE; 1358 1359 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1360 if (file && is_file_hugepages(file)) 1361 vm_flags |= VM_NORESERVE; 1362 } 1363 1364 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf); 1365 if (!IS_ERR_VALUE(addr) && 1366 ((vm_flags & VM_LOCKED) || 1367 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1368 *populate = len; 1369 return addr; 1370 } 1371 1372 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len, 1373 unsigned long prot, unsigned long flags, 1374 unsigned long fd, unsigned long pgoff) 1375 { 1376 struct file *file = NULL; 1377 unsigned long retval; 1378 1379 if (!(flags & MAP_ANONYMOUS)) { 1380 audit_mmap_fd(fd, flags); 1381 file = fget(fd); 1382 if (!file) 1383 return -EBADF; 1384 if (is_file_hugepages(file)) { 1385 len = ALIGN(len, huge_page_size(hstate_file(file))); 1386 } else if (unlikely(flags & MAP_HUGETLB)) { 1387 retval = -EINVAL; 1388 goto out_fput; 1389 } 1390 } else if (flags & MAP_HUGETLB) { 1391 struct hstate *hs; 1392 1393 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1394 if (!hs) 1395 return -EINVAL; 1396 1397 len = ALIGN(len, huge_page_size(hs)); 1398 /* 1399 * VM_NORESERVE is used because the reservations will be 1400 * taken when vm_ops->mmap() is called 1401 */ 1402 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1403 VM_NORESERVE, 1404 HUGETLB_ANONHUGE_INODE, 1405 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1406 if (IS_ERR(file)) 1407 return PTR_ERR(file); 1408 } 1409 1410 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1411 out_fput: 1412 if (file) 1413 fput(file); 1414 return retval; 1415 } 1416 1417 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1418 unsigned long, prot, unsigned long, flags, 1419 unsigned long, fd, unsigned long, pgoff) 1420 { 1421 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff); 1422 } 1423 1424 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1425 struct mmap_arg_struct { 1426 unsigned long addr; 1427 unsigned long len; 1428 unsigned long prot; 1429 unsigned long flags; 1430 unsigned long fd; 1431 unsigned long offset; 1432 }; 1433 1434 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1435 { 1436 struct mmap_arg_struct a; 1437 1438 if (copy_from_user(&a, arg, sizeof(a))) 1439 return -EFAULT; 1440 if (offset_in_page(a.offset)) 1441 return -EINVAL; 1442 1443 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1444 a.offset >> PAGE_SHIFT); 1445 } 1446 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1447 1448 /* 1449 * Some shared mappings will want the pages marked read-only 1450 * to track write events. If so, we'll downgrade vm_page_prot 1451 * to the private version (using protection_map[] without the 1452 * VM_SHARED bit). 1453 */ 1454 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot) 1455 { 1456 vm_flags_t vm_flags = vma->vm_flags; 1457 const struct vm_operations_struct *vm_ops = vma->vm_ops; 1458 1459 /* If it was private or non-writable, the write bit is already clear */ 1460 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED))) 1461 return 0; 1462 1463 /* The backer wishes to know when pages are first written to? */ 1464 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite)) 1465 return 1; 1466 1467 /* The open routine did something to the protections that pgprot_modify 1468 * won't preserve? */ 1469 if (pgprot_val(vm_page_prot) != 1470 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags))) 1471 return 0; 1472 1473 /* 1474 * Do we need to track softdirty? hugetlb does not support softdirty 1475 * tracking yet. 1476 */ 1477 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma)) 1478 return 1; 1479 1480 /* Do we need write faults for uffd-wp tracking? */ 1481 if (userfaultfd_wp(vma)) 1482 return 1; 1483 1484 /* Specialty mapping? */ 1485 if (vm_flags & VM_PFNMAP) 1486 return 0; 1487 1488 /* Can the mapping track the dirty pages? */ 1489 return vma->vm_file && vma->vm_file->f_mapping && 1490 mapping_can_writeback(vma->vm_file->f_mapping); 1491 } 1492 1493 /* 1494 * We account for memory if it's a private writeable mapping, 1495 * not hugepages and VM_NORESERVE wasn't set. 1496 */ 1497 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1498 { 1499 /* 1500 * hugetlb has its own accounting separate from the core VM 1501 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1502 */ 1503 if (file && is_file_hugepages(file)) 1504 return 0; 1505 1506 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1507 } 1508 1509 /** 1510 * unmapped_area() - Find an area between the low_limit and the high_limit with 1511 * the correct alignment and offset, all from @info. Note: current->mm is used 1512 * for the search. 1513 * 1514 * @info: The unmapped area information including the range [low_limit - 1515 * high_limit), the alignment offset and mask. 1516 * 1517 * Return: A memory address or -ENOMEM. 1518 */ 1519 static unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1520 { 1521 unsigned long length, gap; 1522 1523 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0); 1524 1525 /* Adjust search length to account for worst case alignment overhead */ 1526 length = info->length + info->align_mask; 1527 if (length < info->length) 1528 return -ENOMEM; 1529 1530 if (mas_empty_area(&mas, info->low_limit, info->high_limit - 1, 1531 length)) 1532 return -ENOMEM; 1533 1534 gap = mas.index; 1535 gap += (info->align_offset - gap) & info->align_mask; 1536 return gap; 1537 } 1538 1539 /** 1540 * unmapped_area_topdown() - Find an area between the low_limit and the 1541 * high_limit with the correct alignment and offset at the highest available 1542 * address, all from @info. Note: current->mm is used for the search. 1543 * 1544 * @info: The unmapped area information including the range [low_limit - 1545 * high_limit), the alignment offset and mask. 1546 * 1547 * Return: A memory address or -ENOMEM. 1548 */ 1549 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1550 { 1551 unsigned long length, gap; 1552 1553 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0); 1554 /* Adjust search length to account for worst case alignment overhead */ 1555 length = info->length + info->align_mask; 1556 if (length < info->length) 1557 return -ENOMEM; 1558 1559 if (mas_empty_area_rev(&mas, info->low_limit, info->high_limit - 1, 1560 length)) 1561 return -ENOMEM; 1562 1563 gap = mas.last + 1 - info->length; 1564 gap -= (gap - info->align_offset) & info->align_mask; 1565 return gap; 1566 } 1567 1568 /* 1569 * Search for an unmapped address range. 1570 * 1571 * We are looking for a range that: 1572 * - does not intersect with any VMA; 1573 * - is contained within the [low_limit, high_limit) interval; 1574 * - is at least the desired size. 1575 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask) 1576 */ 1577 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info) 1578 { 1579 unsigned long addr; 1580 1581 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN) 1582 addr = unmapped_area_topdown(info); 1583 else 1584 addr = unmapped_area(info); 1585 1586 trace_vm_unmapped_area(addr, info); 1587 return addr; 1588 } 1589 1590 /* Get an address range which is currently unmapped. 1591 * For shmat() with addr=0. 1592 * 1593 * Ugly calling convention alert: 1594 * Return value with the low bits set means error value, 1595 * ie 1596 * if (ret & ~PAGE_MASK) 1597 * error = ret; 1598 * 1599 * This function "knows" that -ENOMEM has the bits set. 1600 */ 1601 unsigned long 1602 generic_get_unmapped_area(struct file *filp, unsigned long addr, 1603 unsigned long len, unsigned long pgoff, 1604 unsigned long flags) 1605 { 1606 struct mm_struct *mm = current->mm; 1607 struct vm_area_struct *vma, *prev; 1608 struct vm_unmapped_area_info info; 1609 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); 1610 1611 if (len > mmap_end - mmap_min_addr) 1612 return -ENOMEM; 1613 1614 if (flags & MAP_FIXED) 1615 return addr; 1616 1617 if (addr) { 1618 addr = PAGE_ALIGN(addr); 1619 vma = find_vma_prev(mm, addr, &prev); 1620 if (mmap_end - len >= addr && addr >= mmap_min_addr && 1621 (!vma || addr + len <= vm_start_gap(vma)) && 1622 (!prev || addr >= vm_end_gap(prev))) 1623 return addr; 1624 } 1625 1626 info.flags = 0; 1627 info.length = len; 1628 info.low_limit = mm->mmap_base; 1629 info.high_limit = mmap_end; 1630 info.align_mask = 0; 1631 info.align_offset = 0; 1632 return vm_unmapped_area(&info); 1633 } 1634 1635 #ifndef HAVE_ARCH_UNMAPPED_AREA 1636 unsigned long 1637 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1638 unsigned long len, unsigned long pgoff, 1639 unsigned long flags) 1640 { 1641 return generic_get_unmapped_area(filp, addr, len, pgoff, flags); 1642 } 1643 #endif 1644 1645 /* 1646 * This mmap-allocator allocates new areas top-down from below the 1647 * stack's low limit (the base): 1648 */ 1649 unsigned long 1650 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr, 1651 unsigned long len, unsigned long pgoff, 1652 unsigned long flags) 1653 { 1654 struct vm_area_struct *vma, *prev; 1655 struct mm_struct *mm = current->mm; 1656 struct vm_unmapped_area_info info; 1657 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); 1658 1659 /* requested length too big for entire address space */ 1660 if (len > mmap_end - mmap_min_addr) 1661 return -ENOMEM; 1662 1663 if (flags & MAP_FIXED) 1664 return addr; 1665 1666 /* requesting a specific address */ 1667 if (addr) { 1668 addr = PAGE_ALIGN(addr); 1669 vma = find_vma_prev(mm, addr, &prev); 1670 if (mmap_end - len >= addr && addr >= mmap_min_addr && 1671 (!vma || addr + len <= vm_start_gap(vma)) && 1672 (!prev || addr >= vm_end_gap(prev))) 1673 return addr; 1674 } 1675 1676 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 1677 info.length = len; 1678 info.low_limit = max(PAGE_SIZE, mmap_min_addr); 1679 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base); 1680 info.align_mask = 0; 1681 info.align_offset = 0; 1682 addr = vm_unmapped_area(&info); 1683 1684 /* 1685 * A failed mmap() very likely causes application failure, 1686 * so fall back to the bottom-up function here. This scenario 1687 * can happen with large stack limits and large mmap() 1688 * allocations. 1689 */ 1690 if (offset_in_page(addr)) { 1691 VM_BUG_ON(addr != -ENOMEM); 1692 info.flags = 0; 1693 info.low_limit = TASK_UNMAPPED_BASE; 1694 info.high_limit = mmap_end; 1695 addr = vm_unmapped_area(&info); 1696 } 1697 1698 return addr; 1699 } 1700 1701 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1702 unsigned long 1703 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr, 1704 unsigned long len, unsigned long pgoff, 1705 unsigned long flags) 1706 { 1707 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags); 1708 } 1709 #endif 1710 1711 unsigned long 1712 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1713 unsigned long pgoff, unsigned long flags) 1714 { 1715 unsigned long (*get_area)(struct file *, unsigned long, 1716 unsigned long, unsigned long, unsigned long); 1717 1718 unsigned long error = arch_mmap_check(addr, len, flags); 1719 if (error) 1720 return error; 1721 1722 /* Careful about overflows.. */ 1723 if (len > TASK_SIZE) 1724 return -ENOMEM; 1725 1726 get_area = current->mm->get_unmapped_area; 1727 if (file) { 1728 if (file->f_op->get_unmapped_area) 1729 get_area = file->f_op->get_unmapped_area; 1730 } else if (flags & MAP_SHARED) { 1731 /* 1732 * mmap_region() will call shmem_zero_setup() to create a file, 1733 * so use shmem's get_unmapped_area in case it can be huge. 1734 * do_mmap() will clear pgoff, so match alignment. 1735 */ 1736 pgoff = 0; 1737 get_area = shmem_get_unmapped_area; 1738 } 1739 1740 addr = get_area(file, addr, len, pgoff, flags); 1741 if (IS_ERR_VALUE(addr)) 1742 return addr; 1743 1744 if (addr > TASK_SIZE - len) 1745 return -ENOMEM; 1746 if (offset_in_page(addr)) 1747 return -EINVAL; 1748 1749 error = security_mmap_addr(addr); 1750 return error ? error : addr; 1751 } 1752 1753 EXPORT_SYMBOL(get_unmapped_area); 1754 1755 /** 1756 * find_vma_intersection() - Look up the first VMA which intersects the interval 1757 * @mm: The process address space. 1758 * @start_addr: The inclusive start user address. 1759 * @end_addr: The exclusive end user address. 1760 * 1761 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes 1762 * start_addr < end_addr. 1763 */ 1764 struct vm_area_struct *find_vma_intersection(struct mm_struct *mm, 1765 unsigned long start_addr, 1766 unsigned long end_addr) 1767 { 1768 unsigned long index = start_addr; 1769 1770 mmap_assert_locked(mm); 1771 return mt_find(&mm->mm_mt, &index, end_addr - 1); 1772 } 1773 EXPORT_SYMBOL(find_vma_intersection); 1774 1775 /** 1776 * find_vma() - Find the VMA for a given address, or the next VMA. 1777 * @mm: The mm_struct to check 1778 * @addr: The address 1779 * 1780 * Returns: The VMA associated with addr, or the next VMA. 1781 * May return %NULL in the case of no VMA at addr or above. 1782 */ 1783 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 1784 { 1785 unsigned long index = addr; 1786 1787 mmap_assert_locked(mm); 1788 return mt_find(&mm->mm_mt, &index, ULONG_MAX); 1789 } 1790 EXPORT_SYMBOL(find_vma); 1791 1792 /** 1793 * find_vma_prev() - Find the VMA for a given address, or the next vma and 1794 * set %pprev to the previous VMA, if any. 1795 * @mm: The mm_struct to check 1796 * @addr: The address 1797 * @pprev: The pointer to set to the previous VMA 1798 * 1799 * Note that RCU lock is missing here since the external mmap_lock() is used 1800 * instead. 1801 * 1802 * Returns: The VMA associated with @addr, or the next vma. 1803 * May return %NULL in the case of no vma at addr or above. 1804 */ 1805 struct vm_area_struct * 1806 find_vma_prev(struct mm_struct *mm, unsigned long addr, 1807 struct vm_area_struct **pprev) 1808 { 1809 struct vm_area_struct *vma; 1810 MA_STATE(mas, &mm->mm_mt, addr, addr); 1811 1812 vma = mas_walk(&mas); 1813 *pprev = mas_prev(&mas, 0); 1814 if (!vma) 1815 vma = mas_next(&mas, ULONG_MAX); 1816 return vma; 1817 } 1818 1819 /* 1820 * Verify that the stack growth is acceptable and 1821 * update accounting. This is shared with both the 1822 * grow-up and grow-down cases. 1823 */ 1824 static int acct_stack_growth(struct vm_area_struct *vma, 1825 unsigned long size, unsigned long grow) 1826 { 1827 struct mm_struct *mm = vma->vm_mm; 1828 unsigned long new_start; 1829 1830 /* address space limit tests */ 1831 if (!may_expand_vm(mm, vma->vm_flags, grow)) 1832 return -ENOMEM; 1833 1834 /* Stack limit test */ 1835 if (size > rlimit(RLIMIT_STACK)) 1836 return -ENOMEM; 1837 1838 /* mlock limit tests */ 1839 if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT)) 1840 return -ENOMEM; 1841 1842 /* Check to ensure the stack will not grow into a hugetlb-only region */ 1843 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 1844 vma->vm_end - size; 1845 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 1846 return -EFAULT; 1847 1848 /* 1849 * Overcommit.. This must be the final test, as it will 1850 * update security statistics. 1851 */ 1852 if (security_vm_enough_memory_mm(mm, grow)) 1853 return -ENOMEM; 1854 1855 return 0; 1856 } 1857 1858 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 1859 /* 1860 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 1861 * vma is the last one with address > vma->vm_end. Have to extend vma. 1862 */ 1863 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 1864 { 1865 struct mm_struct *mm = vma->vm_mm; 1866 struct vm_area_struct *next; 1867 unsigned long gap_addr; 1868 int error = 0; 1869 MA_STATE(mas, &mm->mm_mt, 0, 0); 1870 1871 if (!(vma->vm_flags & VM_GROWSUP)) 1872 return -EFAULT; 1873 1874 /* Guard against exceeding limits of the address space. */ 1875 address &= PAGE_MASK; 1876 if (address >= (TASK_SIZE & PAGE_MASK)) 1877 return -ENOMEM; 1878 address += PAGE_SIZE; 1879 1880 /* Enforce stack_guard_gap */ 1881 gap_addr = address + stack_guard_gap; 1882 1883 /* Guard against overflow */ 1884 if (gap_addr < address || gap_addr > TASK_SIZE) 1885 gap_addr = TASK_SIZE; 1886 1887 next = find_vma_intersection(mm, vma->vm_end, gap_addr); 1888 if (next && vma_is_accessible(next)) { 1889 if (!(next->vm_flags & VM_GROWSUP)) 1890 return -ENOMEM; 1891 /* Check that both stack segments have the same anon_vma? */ 1892 } 1893 1894 if (mas_preallocate(&mas, GFP_KERNEL)) 1895 return -ENOMEM; 1896 1897 /* We must make sure the anon_vma is allocated. */ 1898 if (unlikely(anon_vma_prepare(vma))) { 1899 mas_destroy(&mas); 1900 return -ENOMEM; 1901 } 1902 1903 /* 1904 * vma->vm_start/vm_end cannot change under us because the caller 1905 * is required to hold the mmap_lock in read mode. We need the 1906 * anon_vma lock to serialize against concurrent expand_stacks. 1907 */ 1908 anon_vma_lock_write(vma->anon_vma); 1909 1910 /* Somebody else might have raced and expanded it already */ 1911 if (address > vma->vm_end) { 1912 unsigned long size, grow; 1913 1914 size = address - vma->vm_start; 1915 grow = (address - vma->vm_end) >> PAGE_SHIFT; 1916 1917 error = -ENOMEM; 1918 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 1919 error = acct_stack_growth(vma, size, grow); 1920 if (!error) { 1921 /* 1922 * We only hold a shared mmap_lock lock here, so 1923 * we need to protect against concurrent vma 1924 * expansions. anon_vma_lock_write() doesn't 1925 * help here, as we don't guarantee that all 1926 * growable vmas in a mm share the same root 1927 * anon vma. So, we reuse mm->page_table_lock 1928 * to guard against concurrent vma expansions. 1929 */ 1930 spin_lock(&mm->page_table_lock); 1931 if (vma->vm_flags & VM_LOCKED) 1932 mm->locked_vm += grow; 1933 vm_stat_account(mm, vma->vm_flags, grow); 1934 anon_vma_interval_tree_pre_update_vma(vma); 1935 vma->vm_end = address; 1936 /* Overwrite old entry in mtree. */ 1937 mas_set_range(&mas, vma->vm_start, address - 1); 1938 mas_store_prealloc(&mas, vma); 1939 anon_vma_interval_tree_post_update_vma(vma); 1940 spin_unlock(&mm->page_table_lock); 1941 1942 perf_event_mmap(vma); 1943 } 1944 } 1945 } 1946 anon_vma_unlock_write(vma->anon_vma); 1947 khugepaged_enter_vma(vma, vma->vm_flags); 1948 mas_destroy(&mas); 1949 return error; 1950 } 1951 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 1952 1953 /* 1954 * vma is the first one with address < vma->vm_start. Have to extend vma. 1955 */ 1956 int expand_downwards(struct vm_area_struct *vma, unsigned long address) 1957 { 1958 struct mm_struct *mm = vma->vm_mm; 1959 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start); 1960 struct vm_area_struct *prev; 1961 int error = 0; 1962 1963 address &= PAGE_MASK; 1964 if (address < mmap_min_addr) 1965 return -EPERM; 1966 1967 /* Enforce stack_guard_gap */ 1968 prev = mas_prev(&mas, 0); 1969 /* Check that both stack segments have the same anon_vma? */ 1970 if (prev && !(prev->vm_flags & VM_GROWSDOWN) && 1971 vma_is_accessible(prev)) { 1972 if (address - prev->vm_end < stack_guard_gap) 1973 return -ENOMEM; 1974 } 1975 1976 if (mas_preallocate(&mas, GFP_KERNEL)) 1977 return -ENOMEM; 1978 1979 /* We must make sure the anon_vma is allocated. */ 1980 if (unlikely(anon_vma_prepare(vma))) { 1981 mas_destroy(&mas); 1982 return -ENOMEM; 1983 } 1984 1985 /* 1986 * vma->vm_start/vm_end cannot change under us because the caller 1987 * is required to hold the mmap_lock in read mode. We need the 1988 * anon_vma lock to serialize against concurrent expand_stacks. 1989 */ 1990 anon_vma_lock_write(vma->anon_vma); 1991 1992 /* Somebody else might have raced and expanded it already */ 1993 if (address < vma->vm_start) { 1994 unsigned long size, grow; 1995 1996 size = vma->vm_end - address; 1997 grow = (vma->vm_start - address) >> PAGE_SHIFT; 1998 1999 error = -ENOMEM; 2000 if (grow <= vma->vm_pgoff) { 2001 error = acct_stack_growth(vma, size, grow); 2002 if (!error) { 2003 /* 2004 * We only hold a shared mmap_lock lock here, so 2005 * we need to protect against concurrent vma 2006 * expansions. anon_vma_lock_write() doesn't 2007 * help here, as we don't guarantee that all 2008 * growable vmas in a mm share the same root 2009 * anon vma. So, we reuse mm->page_table_lock 2010 * to guard against concurrent vma expansions. 2011 */ 2012 spin_lock(&mm->page_table_lock); 2013 if (vma->vm_flags & VM_LOCKED) 2014 mm->locked_vm += grow; 2015 vm_stat_account(mm, vma->vm_flags, grow); 2016 anon_vma_interval_tree_pre_update_vma(vma); 2017 vma->vm_start = address; 2018 vma->vm_pgoff -= grow; 2019 /* Overwrite old entry in mtree. */ 2020 mas_set_range(&mas, address, vma->vm_end - 1); 2021 mas_store_prealloc(&mas, vma); 2022 anon_vma_interval_tree_post_update_vma(vma); 2023 spin_unlock(&mm->page_table_lock); 2024 2025 perf_event_mmap(vma); 2026 } 2027 } 2028 } 2029 anon_vma_unlock_write(vma->anon_vma); 2030 khugepaged_enter_vma(vma, vma->vm_flags); 2031 mas_destroy(&mas); 2032 return error; 2033 } 2034 2035 /* enforced gap between the expanding stack and other mappings. */ 2036 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT; 2037 2038 static int __init cmdline_parse_stack_guard_gap(char *p) 2039 { 2040 unsigned long val; 2041 char *endptr; 2042 2043 val = simple_strtoul(p, &endptr, 10); 2044 if (!*endptr) 2045 stack_guard_gap = val << PAGE_SHIFT; 2046 2047 return 1; 2048 } 2049 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap); 2050 2051 #ifdef CONFIG_STACK_GROWSUP 2052 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2053 { 2054 return expand_upwards(vma, address); 2055 } 2056 2057 struct vm_area_struct * 2058 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2059 { 2060 struct vm_area_struct *vma, *prev; 2061 2062 addr &= PAGE_MASK; 2063 vma = find_vma_prev(mm, addr, &prev); 2064 if (vma && (vma->vm_start <= addr)) 2065 return vma; 2066 if (!prev || expand_stack(prev, addr)) 2067 return NULL; 2068 if (prev->vm_flags & VM_LOCKED) 2069 populate_vma_page_range(prev, addr, prev->vm_end, NULL); 2070 return prev; 2071 } 2072 #else 2073 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2074 { 2075 return expand_downwards(vma, address); 2076 } 2077 2078 struct vm_area_struct * 2079 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2080 { 2081 struct vm_area_struct *vma; 2082 unsigned long start; 2083 2084 addr &= PAGE_MASK; 2085 vma = find_vma(mm, addr); 2086 if (!vma) 2087 return NULL; 2088 if (vma->vm_start <= addr) 2089 return vma; 2090 if (!(vma->vm_flags & VM_GROWSDOWN)) 2091 return NULL; 2092 start = vma->vm_start; 2093 if (expand_stack(vma, addr)) 2094 return NULL; 2095 if (vma->vm_flags & VM_LOCKED) 2096 populate_vma_page_range(vma, addr, start, NULL); 2097 return vma; 2098 } 2099 #endif 2100 2101 EXPORT_SYMBOL_GPL(find_extend_vma); 2102 2103 /* 2104 * Ok - we have the memory areas we should free on a maple tree so release them, 2105 * and do the vma updates. 2106 * 2107 * Called with the mm semaphore held. 2108 */ 2109 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas) 2110 { 2111 unsigned long nr_accounted = 0; 2112 struct vm_area_struct *vma; 2113 2114 /* Update high watermark before we lower total_vm */ 2115 update_hiwater_vm(mm); 2116 mas_for_each(mas, vma, ULONG_MAX) { 2117 long nrpages = vma_pages(vma); 2118 2119 if (vma->vm_flags & VM_ACCOUNT) 2120 nr_accounted += nrpages; 2121 vm_stat_account(mm, vma->vm_flags, -nrpages); 2122 remove_vma(vma); 2123 } 2124 vm_unacct_memory(nr_accounted); 2125 validate_mm(mm); 2126 } 2127 2128 /* 2129 * Get rid of page table information in the indicated region. 2130 * 2131 * Called with the mm semaphore held. 2132 */ 2133 static void unmap_region(struct mm_struct *mm, struct maple_tree *mt, 2134 struct vm_area_struct *vma, struct vm_area_struct *prev, 2135 struct vm_area_struct *next, 2136 unsigned long start, unsigned long end, bool mm_wr_locked) 2137 { 2138 struct mmu_gather tlb; 2139 2140 lru_add_drain(); 2141 tlb_gather_mmu(&tlb, mm); 2142 update_hiwater_rss(mm); 2143 unmap_vmas(&tlb, mt, vma, start, end, mm_wr_locked); 2144 free_pgtables(&tlb, mt, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2145 next ? next->vm_start : USER_PGTABLES_CEILING); 2146 tlb_finish_mmu(&tlb); 2147 } 2148 2149 /* 2150 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it 2151 * has already been checked or doesn't make sense to fail. 2152 * VMA Iterator will point to the end VMA. 2153 */ 2154 int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma, 2155 unsigned long addr, int new_below) 2156 { 2157 struct vma_prepare vp; 2158 struct vm_area_struct *new; 2159 int err; 2160 2161 validate_mm_mt(vma->vm_mm); 2162 2163 WARN_ON(vma->vm_start >= addr); 2164 WARN_ON(vma->vm_end <= addr); 2165 2166 if (vma->vm_ops && vma->vm_ops->may_split) { 2167 err = vma->vm_ops->may_split(vma, addr); 2168 if (err) 2169 return err; 2170 } 2171 2172 new = vm_area_dup(vma); 2173 if (!new) 2174 return -ENOMEM; 2175 2176 err = -ENOMEM; 2177 if (vma_iter_prealloc(vmi)) 2178 goto out_free_vma; 2179 2180 if (new_below) { 2181 new->vm_end = addr; 2182 } else { 2183 new->vm_start = addr; 2184 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2185 } 2186 2187 err = vma_dup_policy(vma, new); 2188 if (err) 2189 goto out_free_vmi; 2190 2191 err = anon_vma_clone(new, vma); 2192 if (err) 2193 goto out_free_mpol; 2194 2195 if (new->vm_file) 2196 get_file(new->vm_file); 2197 2198 if (new->vm_ops && new->vm_ops->open) 2199 new->vm_ops->open(new); 2200 2201 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0); 2202 init_vma_prep(&vp, vma); 2203 vp.insert = new; 2204 vma_prepare(&vp); 2205 2206 if (new_below) { 2207 vma->vm_start = addr; 2208 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT; 2209 } else { 2210 vma->vm_end = addr; 2211 } 2212 2213 /* vma_complete stores the new vma */ 2214 vma_complete(&vp, vmi, vma->vm_mm); 2215 2216 /* Success. */ 2217 if (new_below) 2218 vma_next(vmi); 2219 validate_mm_mt(vma->vm_mm); 2220 return 0; 2221 2222 out_free_mpol: 2223 mpol_put(vma_policy(new)); 2224 out_free_vmi: 2225 vma_iter_free(vmi); 2226 out_free_vma: 2227 vm_area_free(new); 2228 validate_mm_mt(vma->vm_mm); 2229 return err; 2230 } 2231 2232 /* 2233 * Split a vma into two pieces at address 'addr', a new vma is allocated 2234 * either for the first part or the tail. 2235 */ 2236 int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma, 2237 unsigned long addr, int new_below) 2238 { 2239 if (vma->vm_mm->map_count >= sysctl_max_map_count) 2240 return -ENOMEM; 2241 2242 return __split_vma(vmi, vma, addr, new_below); 2243 } 2244 2245 static inline int munmap_sidetree(struct vm_area_struct *vma, 2246 struct ma_state *mas_detach) 2247 { 2248 mas_set_range(mas_detach, vma->vm_start, vma->vm_end - 1); 2249 if (mas_store_gfp(mas_detach, vma, GFP_KERNEL)) 2250 return -ENOMEM; 2251 2252 if (vma->vm_flags & VM_LOCKED) 2253 vma->vm_mm->locked_vm -= vma_pages(vma); 2254 2255 return 0; 2256 } 2257 2258 /* 2259 * do_vmi_align_munmap() - munmap the aligned region from @start to @end. 2260 * @vmi: The vma iterator 2261 * @vma: The starting vm_area_struct 2262 * @mm: The mm_struct 2263 * @start: The aligned start address to munmap. 2264 * @end: The aligned end address to munmap. 2265 * @uf: The userfaultfd list_head 2266 * @downgrade: Set to true to attempt a write downgrade of the mmap_lock 2267 * 2268 * If @downgrade is true, check return code for potential release of the lock. 2269 */ 2270 static int 2271 do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma, 2272 struct mm_struct *mm, unsigned long start, 2273 unsigned long end, struct list_head *uf, bool downgrade) 2274 { 2275 struct vm_area_struct *prev, *next = NULL; 2276 struct maple_tree mt_detach; 2277 int count = 0; 2278 int error = -ENOMEM; 2279 MA_STATE(mas_detach, &mt_detach, 0, 0); 2280 mt_init_flags(&mt_detach, MT_FLAGS_LOCK_EXTERN); 2281 mt_set_external_lock(&mt_detach, &mm->mmap_lock); 2282 2283 /* 2284 * If we need to split any vma, do it now to save pain later. 2285 * 2286 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2287 * unmapped vm_area_struct will remain in use: so lower split_vma 2288 * places tmp vma above, and higher split_vma places tmp vma below. 2289 */ 2290 2291 /* Does it split the first one? */ 2292 if (start > vma->vm_start) { 2293 2294 /* 2295 * Make sure that map_count on return from munmap() will 2296 * not exceed its limit; but let map_count go just above 2297 * its limit temporarily, to help free resources as expected. 2298 */ 2299 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2300 goto map_count_exceeded; 2301 2302 error = __split_vma(vmi, vma, start, 0); 2303 if (error) 2304 goto start_split_failed; 2305 2306 vma = vma_iter_load(vmi); 2307 } 2308 2309 prev = vma_prev(vmi); 2310 if (unlikely((!prev))) 2311 vma_iter_set(vmi, start); 2312 2313 /* 2314 * Detach a range of VMAs from the mm. Using next as a temp variable as 2315 * it is always overwritten. 2316 */ 2317 for_each_vma_range(*vmi, next, end) { 2318 /* Does it split the end? */ 2319 if (next->vm_end > end) { 2320 error = __split_vma(vmi, next, end, 0); 2321 if (error) 2322 goto end_split_failed; 2323 } 2324 error = munmap_sidetree(next, &mas_detach); 2325 if (error) 2326 goto munmap_sidetree_failed; 2327 2328 count++; 2329 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE 2330 BUG_ON(next->vm_start < start); 2331 BUG_ON(next->vm_start > end); 2332 #endif 2333 } 2334 2335 next = vma_next(vmi); 2336 if (unlikely(uf)) { 2337 /* 2338 * If userfaultfd_unmap_prep returns an error the vmas 2339 * will remain split, but userland will get a 2340 * highly unexpected error anyway. This is no 2341 * different than the case where the first of the two 2342 * __split_vma fails, but we don't undo the first 2343 * split, despite we could. This is unlikely enough 2344 * failure that it's not worth optimizing it for. 2345 */ 2346 error = userfaultfd_unmap_prep(mm, start, end, uf); 2347 2348 if (error) 2349 goto userfaultfd_error; 2350 } 2351 2352 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE) 2353 /* Make sure no VMAs are about to be lost. */ 2354 { 2355 MA_STATE(test, &mt_detach, start, end - 1); 2356 struct vm_area_struct *vma_mas, *vma_test; 2357 int test_count = 0; 2358 2359 vma_iter_set(vmi, start); 2360 rcu_read_lock(); 2361 vma_test = mas_find(&test, end - 1); 2362 for_each_vma_range(*vmi, vma_mas, end) { 2363 BUG_ON(vma_mas != vma_test); 2364 test_count++; 2365 vma_test = mas_next(&test, end - 1); 2366 } 2367 rcu_read_unlock(); 2368 BUG_ON(count != test_count); 2369 } 2370 #endif 2371 /* Point of no return */ 2372 vma_iter_set(vmi, start); 2373 if (vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL)) 2374 return -ENOMEM; 2375 2376 mm->map_count -= count; 2377 /* 2378 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or 2379 * VM_GROWSUP VMA. Such VMAs can change their size under 2380 * down_read(mmap_lock) and collide with the VMA we are about to unmap. 2381 */ 2382 if (downgrade) { 2383 if (next && (next->vm_flags & VM_GROWSDOWN)) 2384 downgrade = false; 2385 else if (prev && (prev->vm_flags & VM_GROWSUP)) 2386 downgrade = false; 2387 else 2388 mmap_write_downgrade(mm); 2389 } 2390 2391 /* 2392 * We can free page tables without write-locking mmap_lock because VMAs 2393 * were isolated before we downgraded mmap_lock. 2394 */ 2395 unmap_region(mm, &mt_detach, vma, prev, next, start, end, !downgrade); 2396 /* Statistics and freeing VMAs */ 2397 mas_set(&mas_detach, start); 2398 remove_mt(mm, &mas_detach); 2399 __mt_destroy(&mt_detach); 2400 2401 2402 validate_mm(mm); 2403 return downgrade ? 1 : 0; 2404 2405 userfaultfd_error: 2406 munmap_sidetree_failed: 2407 end_split_failed: 2408 __mt_destroy(&mt_detach); 2409 start_split_failed: 2410 map_count_exceeded: 2411 return error; 2412 } 2413 2414 /* 2415 * do_vmi_munmap() - munmap a given range. 2416 * @vmi: The vma iterator 2417 * @mm: The mm_struct 2418 * @start: The start address to munmap 2419 * @len: The length of the range to munmap 2420 * @uf: The userfaultfd list_head 2421 * @downgrade: set to true if the user wants to attempt to write_downgrade the 2422 * mmap_lock 2423 * 2424 * This function takes a @mas that is either pointing to the previous VMA or set 2425 * to MA_START and sets it up to remove the mapping(s). The @len will be 2426 * aligned and any arch_unmap work will be preformed. 2427 * 2428 * Returns: -EINVAL on failure, 1 on success and unlock, 0 otherwise. 2429 */ 2430 int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm, 2431 unsigned long start, size_t len, struct list_head *uf, 2432 bool downgrade) 2433 { 2434 unsigned long end; 2435 struct vm_area_struct *vma; 2436 2437 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start) 2438 return -EINVAL; 2439 2440 end = start + PAGE_ALIGN(len); 2441 if (end == start) 2442 return -EINVAL; 2443 2444 /* arch_unmap() might do unmaps itself. */ 2445 arch_unmap(mm, start, end); 2446 2447 /* Find the first overlapping VMA */ 2448 vma = vma_find(vmi, end); 2449 if (!vma) 2450 return 0; 2451 2452 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, downgrade); 2453 } 2454 2455 /* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls. 2456 * @mm: The mm_struct 2457 * @start: The start address to munmap 2458 * @len: The length to be munmapped. 2459 * @uf: The userfaultfd list_head 2460 */ 2461 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len, 2462 struct list_head *uf) 2463 { 2464 VMA_ITERATOR(vmi, mm, start); 2465 2466 return do_vmi_munmap(&vmi, mm, start, len, uf, false); 2467 } 2468 2469 unsigned long mmap_region(struct file *file, unsigned long addr, 2470 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff, 2471 struct list_head *uf) 2472 { 2473 struct mm_struct *mm = current->mm; 2474 struct vm_area_struct *vma = NULL; 2475 struct vm_area_struct *next, *prev, *merge; 2476 pgoff_t pglen = len >> PAGE_SHIFT; 2477 unsigned long charged = 0; 2478 unsigned long end = addr + len; 2479 unsigned long merge_start = addr, merge_end = end; 2480 pgoff_t vm_pgoff; 2481 int error; 2482 VMA_ITERATOR(vmi, mm, addr); 2483 2484 /* Check against address space limit. */ 2485 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) { 2486 unsigned long nr_pages; 2487 2488 /* 2489 * MAP_FIXED may remove pages of mappings that intersects with 2490 * requested mapping. Account for the pages it would unmap. 2491 */ 2492 nr_pages = count_vma_pages_range(mm, addr, end); 2493 2494 if (!may_expand_vm(mm, vm_flags, 2495 (len >> PAGE_SHIFT) - nr_pages)) 2496 return -ENOMEM; 2497 } 2498 2499 /* Unmap any existing mapping in the area */ 2500 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false)) 2501 return -ENOMEM; 2502 2503 /* 2504 * Private writable mapping: check memory availability 2505 */ 2506 if (accountable_mapping(file, vm_flags)) { 2507 charged = len >> PAGE_SHIFT; 2508 if (security_vm_enough_memory_mm(mm, charged)) 2509 return -ENOMEM; 2510 vm_flags |= VM_ACCOUNT; 2511 } 2512 2513 next = vma_next(&vmi); 2514 prev = vma_prev(&vmi); 2515 if (vm_flags & VM_SPECIAL) 2516 goto cannot_expand; 2517 2518 /* Attempt to expand an old mapping */ 2519 /* Check next */ 2520 if (next && next->vm_start == end && !vma_policy(next) && 2521 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen, 2522 NULL_VM_UFFD_CTX, NULL)) { 2523 merge_end = next->vm_end; 2524 vma = next; 2525 vm_pgoff = next->vm_pgoff - pglen; 2526 } 2527 2528 /* Check prev */ 2529 if (prev && prev->vm_end == addr && !vma_policy(prev) && 2530 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file, 2531 pgoff, vma->vm_userfaultfd_ctx, NULL) : 2532 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff, 2533 NULL_VM_UFFD_CTX, NULL))) { 2534 merge_start = prev->vm_start; 2535 vma = prev; 2536 vm_pgoff = prev->vm_pgoff; 2537 } 2538 2539 2540 /* Actually expand, if possible */ 2541 if (vma && 2542 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) { 2543 khugepaged_enter_vma(vma, vm_flags); 2544 goto expanded; 2545 } 2546 2547 cannot_expand: 2548 /* 2549 * Determine the object being mapped and call the appropriate 2550 * specific mapper. the address has already been validated, but 2551 * not unmapped, but the maps are removed from the list. 2552 */ 2553 vma = vm_area_alloc(mm); 2554 if (!vma) { 2555 error = -ENOMEM; 2556 goto unacct_error; 2557 } 2558 2559 vma_iter_set(&vmi, addr); 2560 vma->vm_start = addr; 2561 vma->vm_end = end; 2562 vm_flags_init(vma, vm_flags); 2563 vma->vm_page_prot = vm_get_page_prot(vm_flags); 2564 vma->vm_pgoff = pgoff; 2565 2566 if (file) { 2567 if (vm_flags & VM_SHARED) { 2568 error = mapping_map_writable(file->f_mapping); 2569 if (error) 2570 goto free_vma; 2571 } 2572 2573 vma->vm_file = get_file(file); 2574 error = call_mmap(file, vma); 2575 if (error) 2576 goto unmap_and_free_vma; 2577 2578 /* 2579 * Expansion is handled above, merging is handled below. 2580 * Drivers should not alter the address of the VMA. 2581 */ 2582 error = -EINVAL; 2583 if (WARN_ON((addr != vma->vm_start))) 2584 goto close_and_free_vma; 2585 2586 vma_iter_set(&vmi, addr); 2587 /* 2588 * If vm_flags changed after call_mmap(), we should try merge 2589 * vma again as we may succeed this time. 2590 */ 2591 if (unlikely(vm_flags != vma->vm_flags && prev)) { 2592 merge = vma_merge(&vmi, mm, prev, vma->vm_start, 2593 vma->vm_end, vma->vm_flags, NULL, 2594 vma->vm_file, vma->vm_pgoff, NULL, 2595 NULL_VM_UFFD_CTX, NULL); 2596 if (merge) { 2597 /* 2598 * ->mmap() can change vma->vm_file and fput 2599 * the original file. So fput the vma->vm_file 2600 * here or we would add an extra fput for file 2601 * and cause general protection fault 2602 * ultimately. 2603 */ 2604 fput(vma->vm_file); 2605 vm_area_free(vma); 2606 vma = merge; 2607 /* Update vm_flags to pick up the change. */ 2608 vm_flags = vma->vm_flags; 2609 goto unmap_writable; 2610 } 2611 } 2612 2613 vm_flags = vma->vm_flags; 2614 } else if (vm_flags & VM_SHARED) { 2615 error = shmem_zero_setup(vma); 2616 if (error) 2617 goto free_vma; 2618 } else { 2619 vma_set_anonymous(vma); 2620 } 2621 2622 if (map_deny_write_exec(vma, vma->vm_flags)) { 2623 error = -EACCES; 2624 if (file) 2625 goto close_and_free_vma; 2626 else if (vma->vm_file) 2627 goto unmap_and_free_vma; 2628 else 2629 goto free_vma; 2630 } 2631 2632 /* Allow architectures to sanity-check the vm_flags */ 2633 error = -EINVAL; 2634 if (!arch_validate_flags(vma->vm_flags)) 2635 goto close_and_free_vma; 2636 2637 error = -ENOMEM; 2638 if (vma_iter_prealloc(&vmi)) 2639 goto close_and_free_vma; 2640 2641 if (vma->vm_file) 2642 i_mmap_lock_write(vma->vm_file->f_mapping); 2643 2644 vma_iter_store(&vmi, vma); 2645 mm->map_count++; 2646 if (vma->vm_file) { 2647 if (vma->vm_flags & VM_SHARED) 2648 mapping_allow_writable(vma->vm_file->f_mapping); 2649 2650 flush_dcache_mmap_lock(vma->vm_file->f_mapping); 2651 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap); 2652 flush_dcache_mmap_unlock(vma->vm_file->f_mapping); 2653 i_mmap_unlock_write(vma->vm_file->f_mapping); 2654 } 2655 2656 /* 2657 * vma_merge() calls khugepaged_enter_vma() either, the below 2658 * call covers the non-merge case. 2659 */ 2660 khugepaged_enter_vma(vma, vma->vm_flags); 2661 2662 /* Once vma denies write, undo our temporary denial count */ 2663 unmap_writable: 2664 if (file && vm_flags & VM_SHARED) 2665 mapping_unmap_writable(file->f_mapping); 2666 file = vma->vm_file; 2667 expanded: 2668 perf_event_mmap(vma); 2669 2670 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT); 2671 if (vm_flags & VM_LOCKED) { 2672 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) || 2673 is_vm_hugetlb_page(vma) || 2674 vma == get_gate_vma(current->mm)) 2675 vm_flags_clear(vma, VM_LOCKED_MASK); 2676 else 2677 mm->locked_vm += (len >> PAGE_SHIFT); 2678 } 2679 2680 if (file) 2681 uprobe_mmap(vma); 2682 2683 /* 2684 * New (or expanded) vma always get soft dirty status. 2685 * Otherwise user-space soft-dirty page tracker won't 2686 * be able to distinguish situation when vma area unmapped, 2687 * then new mapped in-place (which must be aimed as 2688 * a completely new data area). 2689 */ 2690 vm_flags_set(vma, VM_SOFTDIRTY); 2691 2692 vma_set_page_prot(vma); 2693 2694 validate_mm(mm); 2695 return addr; 2696 2697 close_and_free_vma: 2698 if (file && vma->vm_ops && vma->vm_ops->close) 2699 vma->vm_ops->close(vma); 2700 2701 if (file || vma->vm_file) { 2702 unmap_and_free_vma: 2703 fput(vma->vm_file); 2704 vma->vm_file = NULL; 2705 2706 /* Undo any partial mapping done by a device driver. */ 2707 unmap_region(mm, &mm->mm_mt, vma, prev, next, vma->vm_start, 2708 vma->vm_end, true); 2709 } 2710 if (file && (vm_flags & VM_SHARED)) 2711 mapping_unmap_writable(file->f_mapping); 2712 free_vma: 2713 vm_area_free(vma); 2714 unacct_error: 2715 if (charged) 2716 vm_unacct_memory(charged); 2717 validate_mm(mm); 2718 return error; 2719 } 2720 2721 static int __vm_munmap(unsigned long start, size_t len, bool downgrade) 2722 { 2723 int ret; 2724 struct mm_struct *mm = current->mm; 2725 LIST_HEAD(uf); 2726 VMA_ITERATOR(vmi, mm, start); 2727 2728 if (mmap_write_lock_killable(mm)) 2729 return -EINTR; 2730 2731 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, downgrade); 2732 /* 2733 * Returning 1 indicates mmap_lock is downgraded. 2734 * But 1 is not legal return value of vm_munmap() and munmap(), reset 2735 * it to 0 before return. 2736 */ 2737 if (ret == 1) { 2738 mmap_read_unlock(mm); 2739 ret = 0; 2740 } else 2741 mmap_write_unlock(mm); 2742 2743 userfaultfd_unmap_complete(mm, &uf); 2744 return ret; 2745 } 2746 2747 int vm_munmap(unsigned long start, size_t len) 2748 { 2749 return __vm_munmap(start, len, false); 2750 } 2751 EXPORT_SYMBOL(vm_munmap); 2752 2753 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2754 { 2755 addr = untagged_addr(addr); 2756 return __vm_munmap(addr, len, true); 2757 } 2758 2759 2760 /* 2761 * Emulation of deprecated remap_file_pages() syscall. 2762 */ 2763 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size, 2764 unsigned long, prot, unsigned long, pgoff, unsigned long, flags) 2765 { 2766 2767 struct mm_struct *mm = current->mm; 2768 struct vm_area_struct *vma; 2769 unsigned long populate = 0; 2770 unsigned long ret = -EINVAL; 2771 struct file *file; 2772 2773 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n", 2774 current->comm, current->pid); 2775 2776 if (prot) 2777 return ret; 2778 start = start & PAGE_MASK; 2779 size = size & PAGE_MASK; 2780 2781 if (start + size <= start) 2782 return ret; 2783 2784 /* Does pgoff wrap? */ 2785 if (pgoff + (size >> PAGE_SHIFT) < pgoff) 2786 return ret; 2787 2788 if (mmap_write_lock_killable(mm)) 2789 return -EINTR; 2790 2791 vma = vma_lookup(mm, start); 2792 2793 if (!vma || !(vma->vm_flags & VM_SHARED)) 2794 goto out; 2795 2796 if (start + size > vma->vm_end) { 2797 VMA_ITERATOR(vmi, mm, vma->vm_end); 2798 struct vm_area_struct *next, *prev = vma; 2799 2800 for_each_vma_range(vmi, next, start + size) { 2801 /* hole between vmas ? */ 2802 if (next->vm_start != prev->vm_end) 2803 goto out; 2804 2805 if (next->vm_file != vma->vm_file) 2806 goto out; 2807 2808 if (next->vm_flags != vma->vm_flags) 2809 goto out; 2810 2811 if (start + size <= next->vm_end) 2812 break; 2813 2814 prev = next; 2815 } 2816 2817 if (!next) 2818 goto out; 2819 } 2820 2821 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0; 2822 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0; 2823 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0; 2824 2825 flags &= MAP_NONBLOCK; 2826 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE; 2827 if (vma->vm_flags & VM_LOCKED) 2828 flags |= MAP_LOCKED; 2829 2830 file = get_file(vma->vm_file); 2831 ret = do_mmap(vma->vm_file, start, size, 2832 prot, flags, pgoff, &populate, NULL); 2833 fput(file); 2834 out: 2835 mmap_write_unlock(mm); 2836 if (populate) 2837 mm_populate(ret, populate); 2838 if (!IS_ERR_VALUE(ret)) 2839 ret = 0; 2840 return ret; 2841 } 2842 2843 /* 2844 * do_vma_munmap() - Unmap a full or partial vma. 2845 * @vmi: The vma iterator pointing at the vma 2846 * @vma: The first vma to be munmapped 2847 * @start: the start of the address to unmap 2848 * @end: The end of the address to unmap 2849 * @uf: The userfaultfd list_head 2850 * @downgrade: Attempt to downgrade or not 2851 * 2852 * Returns: 0 on success and not downgraded, 1 on success and downgraded. 2853 * unmaps a VMA mapping when the vma iterator is already in position. 2854 * Does not handle alignment. 2855 */ 2856 int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma, 2857 unsigned long start, unsigned long end, 2858 struct list_head *uf, bool downgrade) 2859 { 2860 struct mm_struct *mm = vma->vm_mm; 2861 int ret; 2862 2863 arch_unmap(mm, start, end); 2864 ret = do_vmi_align_munmap(vmi, vma, mm, start, end, uf, downgrade); 2865 validate_mm_mt(mm); 2866 return ret; 2867 } 2868 2869 /* 2870 * do_brk_flags() - Increase the brk vma if the flags match. 2871 * @vmi: The vma iterator 2872 * @addr: The start address 2873 * @len: The length of the increase 2874 * @vma: The vma, 2875 * @flags: The VMA Flags 2876 * 2877 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags 2878 * do not match then create a new anonymous VMA. Eventually we may be able to 2879 * do some brk-specific accounting here. 2880 */ 2881 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma, 2882 unsigned long addr, unsigned long len, unsigned long flags) 2883 { 2884 struct mm_struct *mm = current->mm; 2885 struct vma_prepare vp; 2886 2887 validate_mm_mt(mm); 2888 /* 2889 * Check against address space limits by the changed size 2890 * Note: This happens *after* clearing old mappings in some code paths. 2891 */ 2892 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2893 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT)) 2894 return -ENOMEM; 2895 2896 if (mm->map_count > sysctl_max_map_count) 2897 return -ENOMEM; 2898 2899 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 2900 return -ENOMEM; 2901 2902 /* 2903 * Expand the existing vma if possible; Note that singular lists do not 2904 * occur after forking, so the expand will only happen on new VMAs. 2905 */ 2906 if (vma && vma->vm_end == addr && !vma_policy(vma) && 2907 can_vma_merge_after(vma, flags, NULL, NULL, 2908 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) { 2909 if (vma_iter_prealloc(vmi)) 2910 goto unacct_fail; 2911 2912 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0); 2913 init_vma_prep(&vp, vma); 2914 vma_prepare(&vp); 2915 vma->vm_end = addr + len; 2916 vm_flags_set(vma, VM_SOFTDIRTY); 2917 vma_iter_store(vmi, vma); 2918 2919 vma_complete(&vp, vmi, mm); 2920 khugepaged_enter_vma(vma, flags); 2921 goto out; 2922 } 2923 2924 /* create a vma struct for an anonymous mapping */ 2925 vma = vm_area_alloc(mm); 2926 if (!vma) 2927 goto unacct_fail; 2928 2929 vma_set_anonymous(vma); 2930 vma->vm_start = addr; 2931 vma->vm_end = addr + len; 2932 vma->vm_pgoff = addr >> PAGE_SHIFT; 2933 vm_flags_init(vma, flags); 2934 vma->vm_page_prot = vm_get_page_prot(flags); 2935 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL)) 2936 goto mas_store_fail; 2937 2938 mm->map_count++; 2939 out: 2940 perf_event_mmap(vma); 2941 mm->total_vm += len >> PAGE_SHIFT; 2942 mm->data_vm += len >> PAGE_SHIFT; 2943 if (flags & VM_LOCKED) 2944 mm->locked_vm += (len >> PAGE_SHIFT); 2945 vm_flags_set(vma, VM_SOFTDIRTY); 2946 validate_mm(mm); 2947 return 0; 2948 2949 mas_store_fail: 2950 vm_area_free(vma); 2951 unacct_fail: 2952 vm_unacct_memory(len >> PAGE_SHIFT); 2953 return -ENOMEM; 2954 } 2955 2956 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags) 2957 { 2958 struct mm_struct *mm = current->mm; 2959 struct vm_area_struct *vma = NULL; 2960 unsigned long len; 2961 int ret; 2962 bool populate; 2963 LIST_HEAD(uf); 2964 VMA_ITERATOR(vmi, mm, addr); 2965 2966 len = PAGE_ALIGN(request); 2967 if (len < request) 2968 return -ENOMEM; 2969 if (!len) 2970 return 0; 2971 2972 if (mmap_write_lock_killable(mm)) 2973 return -EINTR; 2974 2975 /* Until we need other flags, refuse anything except VM_EXEC. */ 2976 if ((flags & (~VM_EXEC)) != 0) 2977 return -EINVAL; 2978 2979 ret = check_brk_limits(addr, len); 2980 if (ret) 2981 goto limits_failed; 2982 2983 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0); 2984 if (ret) 2985 goto munmap_failed; 2986 2987 vma = vma_prev(&vmi); 2988 ret = do_brk_flags(&vmi, vma, addr, len, flags); 2989 populate = ((mm->def_flags & VM_LOCKED) != 0); 2990 mmap_write_unlock(mm); 2991 userfaultfd_unmap_complete(mm, &uf); 2992 if (populate && !ret) 2993 mm_populate(addr, len); 2994 return ret; 2995 2996 munmap_failed: 2997 limits_failed: 2998 mmap_write_unlock(mm); 2999 return ret; 3000 } 3001 EXPORT_SYMBOL(vm_brk_flags); 3002 3003 int vm_brk(unsigned long addr, unsigned long len) 3004 { 3005 return vm_brk_flags(addr, len, 0); 3006 } 3007 EXPORT_SYMBOL(vm_brk); 3008 3009 /* Release all mmaps. */ 3010 void exit_mmap(struct mm_struct *mm) 3011 { 3012 struct mmu_gather tlb; 3013 struct vm_area_struct *vma; 3014 unsigned long nr_accounted = 0; 3015 MA_STATE(mas, &mm->mm_mt, 0, 0); 3016 int count = 0; 3017 3018 /* mm's last user has gone, and its about to be pulled down */ 3019 mmu_notifier_release(mm); 3020 3021 mmap_read_lock(mm); 3022 arch_exit_mmap(mm); 3023 3024 vma = mas_find(&mas, ULONG_MAX); 3025 if (!vma) { 3026 /* Can happen if dup_mmap() received an OOM */ 3027 mmap_read_unlock(mm); 3028 return; 3029 } 3030 3031 lru_add_drain(); 3032 flush_cache_mm(mm); 3033 tlb_gather_mmu_fullmm(&tlb, mm); 3034 /* update_hiwater_rss(mm) here? but nobody should be looking */ 3035 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */ 3036 unmap_vmas(&tlb, &mm->mm_mt, vma, 0, ULONG_MAX, false); 3037 mmap_read_unlock(mm); 3038 3039 /* 3040 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper 3041 * because the memory has been already freed. 3042 */ 3043 set_bit(MMF_OOM_SKIP, &mm->flags); 3044 mmap_write_lock(mm); 3045 free_pgtables(&tlb, &mm->mm_mt, vma, FIRST_USER_ADDRESS, 3046 USER_PGTABLES_CEILING); 3047 tlb_finish_mmu(&tlb); 3048 3049 /* 3050 * Walk the list again, actually closing and freeing it, with preemption 3051 * enabled, without holding any MM locks besides the unreachable 3052 * mmap_write_lock. 3053 */ 3054 do { 3055 if (vma->vm_flags & VM_ACCOUNT) 3056 nr_accounted += vma_pages(vma); 3057 remove_vma(vma); 3058 count++; 3059 cond_resched(); 3060 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL); 3061 3062 BUG_ON(count != mm->map_count); 3063 3064 trace_exit_mmap(mm); 3065 __mt_destroy(&mm->mm_mt); 3066 mmap_write_unlock(mm); 3067 vm_unacct_memory(nr_accounted); 3068 } 3069 3070 /* Insert vm structure into process list sorted by address 3071 * and into the inode's i_mmap tree. If vm_file is non-NULL 3072 * then i_mmap_rwsem is taken here. 3073 */ 3074 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 3075 { 3076 unsigned long charged = vma_pages(vma); 3077 3078 3079 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end)) 3080 return -ENOMEM; 3081 3082 if ((vma->vm_flags & VM_ACCOUNT) && 3083 security_vm_enough_memory_mm(mm, charged)) 3084 return -ENOMEM; 3085 3086 /* 3087 * The vm_pgoff of a purely anonymous vma should be irrelevant 3088 * until its first write fault, when page's anon_vma and index 3089 * are set. But now set the vm_pgoff it will almost certainly 3090 * end up with (unless mremap moves it elsewhere before that 3091 * first wfault), so /proc/pid/maps tells a consistent story. 3092 * 3093 * By setting it to reflect the virtual start address of the 3094 * vma, merges and splits can happen in a seamless way, just 3095 * using the existing file pgoff checks and manipulations. 3096 * Similarly in do_mmap and in do_brk_flags. 3097 */ 3098 if (vma_is_anonymous(vma)) { 3099 BUG_ON(vma->anon_vma); 3100 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 3101 } 3102 3103 if (vma_link(mm, vma)) { 3104 vm_unacct_memory(charged); 3105 return -ENOMEM; 3106 } 3107 3108 return 0; 3109 } 3110 3111 /* 3112 * Copy the vma structure to a new location in the same mm, 3113 * prior to moving page table entries, to effect an mremap move. 3114 */ 3115 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 3116 unsigned long addr, unsigned long len, pgoff_t pgoff, 3117 bool *need_rmap_locks) 3118 { 3119 struct vm_area_struct *vma = *vmap; 3120 unsigned long vma_start = vma->vm_start; 3121 struct mm_struct *mm = vma->vm_mm; 3122 struct vm_area_struct *new_vma, *prev; 3123 bool faulted_in_anon_vma = true; 3124 VMA_ITERATOR(vmi, mm, addr); 3125 3126 validate_mm_mt(mm); 3127 /* 3128 * If anonymous vma has not yet been faulted, update new pgoff 3129 * to match new location, to increase its chance of merging. 3130 */ 3131 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) { 3132 pgoff = addr >> PAGE_SHIFT; 3133 faulted_in_anon_vma = false; 3134 } 3135 3136 new_vma = find_vma_prev(mm, addr, &prev); 3137 if (new_vma && new_vma->vm_start < addr + len) 3138 return NULL; /* should never get here */ 3139 3140 new_vma = vma_merge(&vmi, mm, prev, addr, addr + len, vma->vm_flags, 3141 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 3142 vma->vm_userfaultfd_ctx, anon_vma_name(vma)); 3143 if (new_vma) { 3144 /* 3145 * Source vma may have been merged into new_vma 3146 */ 3147 if (unlikely(vma_start >= new_vma->vm_start && 3148 vma_start < new_vma->vm_end)) { 3149 /* 3150 * The only way we can get a vma_merge with 3151 * self during an mremap is if the vma hasn't 3152 * been faulted in yet and we were allowed to 3153 * reset the dst vma->vm_pgoff to the 3154 * destination address of the mremap to allow 3155 * the merge to happen. mremap must change the 3156 * vm_pgoff linearity between src and dst vmas 3157 * (in turn preventing a vma_merge) to be 3158 * safe. It is only safe to keep the vm_pgoff 3159 * linear if there are no pages mapped yet. 3160 */ 3161 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 3162 *vmap = vma = new_vma; 3163 } 3164 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 3165 } else { 3166 new_vma = vm_area_dup(vma); 3167 if (!new_vma) 3168 goto out; 3169 new_vma->vm_start = addr; 3170 new_vma->vm_end = addr + len; 3171 new_vma->vm_pgoff = pgoff; 3172 if (vma_dup_policy(vma, new_vma)) 3173 goto out_free_vma; 3174 if (anon_vma_clone(new_vma, vma)) 3175 goto out_free_mempol; 3176 if (new_vma->vm_file) 3177 get_file(new_vma->vm_file); 3178 if (new_vma->vm_ops && new_vma->vm_ops->open) 3179 new_vma->vm_ops->open(new_vma); 3180 if (vma_link(mm, new_vma)) 3181 goto out_vma_link; 3182 *need_rmap_locks = false; 3183 } 3184 validate_mm_mt(mm); 3185 return new_vma; 3186 3187 out_vma_link: 3188 if (new_vma->vm_ops && new_vma->vm_ops->close) 3189 new_vma->vm_ops->close(new_vma); 3190 3191 if (new_vma->vm_file) 3192 fput(new_vma->vm_file); 3193 3194 unlink_anon_vmas(new_vma); 3195 out_free_mempol: 3196 mpol_put(vma_policy(new_vma)); 3197 out_free_vma: 3198 vm_area_free(new_vma); 3199 out: 3200 validate_mm_mt(mm); 3201 return NULL; 3202 } 3203 3204 /* 3205 * Return true if the calling process may expand its vm space by the passed 3206 * number of pages 3207 */ 3208 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages) 3209 { 3210 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT) 3211 return false; 3212 3213 if (is_data_mapping(flags) && 3214 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) { 3215 /* Workaround for Valgrind */ 3216 if (rlimit(RLIMIT_DATA) == 0 && 3217 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT) 3218 return true; 3219 3220 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n", 3221 current->comm, current->pid, 3222 (mm->data_vm + npages) << PAGE_SHIFT, 3223 rlimit(RLIMIT_DATA), 3224 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data"); 3225 3226 if (!ignore_rlimit_data) 3227 return false; 3228 } 3229 3230 return true; 3231 } 3232 3233 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages) 3234 { 3235 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages); 3236 3237 if (is_exec_mapping(flags)) 3238 mm->exec_vm += npages; 3239 else if (is_stack_mapping(flags)) 3240 mm->stack_vm += npages; 3241 else if (is_data_mapping(flags)) 3242 mm->data_vm += npages; 3243 } 3244 3245 static vm_fault_t special_mapping_fault(struct vm_fault *vmf); 3246 3247 /* 3248 * Having a close hook prevents vma merging regardless of flags. 3249 */ 3250 static void special_mapping_close(struct vm_area_struct *vma) 3251 { 3252 } 3253 3254 static const char *special_mapping_name(struct vm_area_struct *vma) 3255 { 3256 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 3257 } 3258 3259 static int special_mapping_mremap(struct vm_area_struct *new_vma) 3260 { 3261 struct vm_special_mapping *sm = new_vma->vm_private_data; 3262 3263 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm)) 3264 return -EFAULT; 3265 3266 if (sm->mremap) 3267 return sm->mremap(sm, new_vma); 3268 3269 return 0; 3270 } 3271 3272 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr) 3273 { 3274 /* 3275 * Forbid splitting special mappings - kernel has expectations over 3276 * the number of pages in mapping. Together with VM_DONTEXPAND 3277 * the size of vma should stay the same over the special mapping's 3278 * lifetime. 3279 */ 3280 return -EINVAL; 3281 } 3282 3283 static const struct vm_operations_struct special_mapping_vmops = { 3284 .close = special_mapping_close, 3285 .fault = special_mapping_fault, 3286 .mremap = special_mapping_mremap, 3287 .name = special_mapping_name, 3288 /* vDSO code relies that VVAR can't be accessed remotely */ 3289 .access = NULL, 3290 .may_split = special_mapping_split, 3291 }; 3292 3293 static const struct vm_operations_struct legacy_special_mapping_vmops = { 3294 .close = special_mapping_close, 3295 .fault = special_mapping_fault, 3296 }; 3297 3298 static vm_fault_t special_mapping_fault(struct vm_fault *vmf) 3299 { 3300 struct vm_area_struct *vma = vmf->vma; 3301 pgoff_t pgoff; 3302 struct page **pages; 3303 3304 if (vma->vm_ops == &legacy_special_mapping_vmops) { 3305 pages = vma->vm_private_data; 3306 } else { 3307 struct vm_special_mapping *sm = vma->vm_private_data; 3308 3309 if (sm->fault) 3310 return sm->fault(sm, vmf->vma, vmf); 3311 3312 pages = sm->pages; 3313 } 3314 3315 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages) 3316 pgoff--; 3317 3318 if (*pages) { 3319 struct page *page = *pages; 3320 get_page(page); 3321 vmf->page = page; 3322 return 0; 3323 } 3324 3325 return VM_FAULT_SIGBUS; 3326 } 3327 3328 static struct vm_area_struct *__install_special_mapping( 3329 struct mm_struct *mm, 3330 unsigned long addr, unsigned long len, 3331 unsigned long vm_flags, void *priv, 3332 const struct vm_operations_struct *ops) 3333 { 3334 int ret; 3335 struct vm_area_struct *vma; 3336 3337 validate_mm_mt(mm); 3338 vma = vm_area_alloc(mm); 3339 if (unlikely(vma == NULL)) 3340 return ERR_PTR(-ENOMEM); 3341 3342 vma->vm_start = addr; 3343 vma->vm_end = addr + len; 3344 3345 vm_flags_init(vma, (vm_flags | mm->def_flags | 3346 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK); 3347 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3348 3349 vma->vm_ops = ops; 3350 vma->vm_private_data = priv; 3351 3352 ret = insert_vm_struct(mm, vma); 3353 if (ret) 3354 goto out; 3355 3356 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT); 3357 3358 perf_event_mmap(vma); 3359 3360 validate_mm_mt(mm); 3361 return vma; 3362 3363 out: 3364 vm_area_free(vma); 3365 validate_mm_mt(mm); 3366 return ERR_PTR(ret); 3367 } 3368 3369 bool vma_is_special_mapping(const struct vm_area_struct *vma, 3370 const struct vm_special_mapping *sm) 3371 { 3372 return vma->vm_private_data == sm && 3373 (vma->vm_ops == &special_mapping_vmops || 3374 vma->vm_ops == &legacy_special_mapping_vmops); 3375 } 3376 3377 /* 3378 * Called with mm->mmap_lock held for writing. 3379 * Insert a new vma covering the given region, with the given flags. 3380 * Its pages are supplied by the given array of struct page *. 3381 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3382 * The region past the last page supplied will always produce SIGBUS. 3383 * The array pointer and the pages it points to are assumed to stay alive 3384 * for as long as this mapping might exist. 3385 */ 3386 struct vm_area_struct *_install_special_mapping( 3387 struct mm_struct *mm, 3388 unsigned long addr, unsigned long len, 3389 unsigned long vm_flags, const struct vm_special_mapping *spec) 3390 { 3391 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec, 3392 &special_mapping_vmops); 3393 } 3394 3395 int install_special_mapping(struct mm_struct *mm, 3396 unsigned long addr, unsigned long len, 3397 unsigned long vm_flags, struct page **pages) 3398 { 3399 struct vm_area_struct *vma = __install_special_mapping( 3400 mm, addr, len, vm_flags, (void *)pages, 3401 &legacy_special_mapping_vmops); 3402 3403 return PTR_ERR_OR_ZERO(vma); 3404 } 3405 3406 static DEFINE_MUTEX(mm_all_locks_mutex); 3407 3408 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3409 { 3410 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) { 3411 /* 3412 * The LSB of head.next can't change from under us 3413 * because we hold the mm_all_locks_mutex. 3414 */ 3415 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock); 3416 /* 3417 * We can safely modify head.next after taking the 3418 * anon_vma->root->rwsem. If some other vma in this mm shares 3419 * the same anon_vma we won't take it again. 3420 * 3421 * No need of atomic instructions here, head.next 3422 * can't change from under us thanks to the 3423 * anon_vma->root->rwsem. 3424 */ 3425 if (__test_and_set_bit(0, (unsigned long *) 3426 &anon_vma->root->rb_root.rb_root.rb_node)) 3427 BUG(); 3428 } 3429 } 3430 3431 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3432 { 3433 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3434 /* 3435 * AS_MM_ALL_LOCKS can't change from under us because 3436 * we hold the mm_all_locks_mutex. 3437 * 3438 * Operations on ->flags have to be atomic because 3439 * even if AS_MM_ALL_LOCKS is stable thanks to the 3440 * mm_all_locks_mutex, there may be other cpus 3441 * changing other bitflags in parallel to us. 3442 */ 3443 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3444 BUG(); 3445 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock); 3446 } 3447 } 3448 3449 /* 3450 * This operation locks against the VM for all pte/vma/mm related 3451 * operations that could ever happen on a certain mm. This includes 3452 * vmtruncate, try_to_unmap, and all page faults. 3453 * 3454 * The caller must take the mmap_lock in write mode before calling 3455 * mm_take_all_locks(). The caller isn't allowed to release the 3456 * mmap_lock until mm_drop_all_locks() returns. 3457 * 3458 * mmap_lock in write mode is required in order to block all operations 3459 * that could modify pagetables and free pages without need of 3460 * altering the vma layout. It's also needed in write mode to avoid new 3461 * anon_vmas to be associated with existing vmas. 3462 * 3463 * A single task can't take more than one mm_take_all_locks() in a row 3464 * or it would deadlock. 3465 * 3466 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3467 * mapping->flags avoid to take the same lock twice, if more than one 3468 * vma in this mm is backed by the same anon_vma or address_space. 3469 * 3470 * We take locks in following order, accordingly to comment at beginning 3471 * of mm/rmap.c: 3472 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for 3473 * hugetlb mapping); 3474 * - all i_mmap_rwsem locks; 3475 * - all anon_vma->rwseml 3476 * 3477 * We can take all locks within these types randomly because the VM code 3478 * doesn't nest them and we protected from parallel mm_take_all_locks() by 3479 * mm_all_locks_mutex. 3480 * 3481 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3482 * that may have to take thousand of locks. 3483 * 3484 * mm_take_all_locks() can fail if it's interrupted by signals. 3485 */ 3486 int mm_take_all_locks(struct mm_struct *mm) 3487 { 3488 struct vm_area_struct *vma; 3489 struct anon_vma_chain *avc; 3490 MA_STATE(mas, &mm->mm_mt, 0, 0); 3491 3492 mmap_assert_write_locked(mm); 3493 3494 mutex_lock(&mm_all_locks_mutex); 3495 3496 mas_for_each(&mas, vma, ULONG_MAX) { 3497 if (signal_pending(current)) 3498 goto out_unlock; 3499 if (vma->vm_file && vma->vm_file->f_mapping && 3500 is_vm_hugetlb_page(vma)) 3501 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3502 } 3503 3504 mas_set(&mas, 0); 3505 mas_for_each(&mas, vma, ULONG_MAX) { 3506 if (signal_pending(current)) 3507 goto out_unlock; 3508 if (vma->vm_file && vma->vm_file->f_mapping && 3509 !is_vm_hugetlb_page(vma)) 3510 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3511 } 3512 3513 mas_set(&mas, 0); 3514 mas_for_each(&mas, vma, ULONG_MAX) { 3515 if (signal_pending(current)) 3516 goto out_unlock; 3517 if (vma->anon_vma) 3518 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3519 vm_lock_anon_vma(mm, avc->anon_vma); 3520 } 3521 3522 return 0; 3523 3524 out_unlock: 3525 mm_drop_all_locks(mm); 3526 return -EINTR; 3527 } 3528 3529 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3530 { 3531 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) { 3532 /* 3533 * The LSB of head.next can't change to 0 from under 3534 * us because we hold the mm_all_locks_mutex. 3535 * 3536 * We must however clear the bitflag before unlocking 3537 * the vma so the users using the anon_vma->rb_root will 3538 * never see our bitflag. 3539 * 3540 * No need of atomic instructions here, head.next 3541 * can't change from under us until we release the 3542 * anon_vma->root->rwsem. 3543 */ 3544 if (!__test_and_clear_bit(0, (unsigned long *) 3545 &anon_vma->root->rb_root.rb_root.rb_node)) 3546 BUG(); 3547 anon_vma_unlock_write(anon_vma); 3548 } 3549 } 3550 3551 static void vm_unlock_mapping(struct address_space *mapping) 3552 { 3553 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3554 /* 3555 * AS_MM_ALL_LOCKS can't change to 0 from under us 3556 * because we hold the mm_all_locks_mutex. 3557 */ 3558 i_mmap_unlock_write(mapping); 3559 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3560 &mapping->flags)) 3561 BUG(); 3562 } 3563 } 3564 3565 /* 3566 * The mmap_lock cannot be released by the caller until 3567 * mm_drop_all_locks() returns. 3568 */ 3569 void mm_drop_all_locks(struct mm_struct *mm) 3570 { 3571 struct vm_area_struct *vma; 3572 struct anon_vma_chain *avc; 3573 MA_STATE(mas, &mm->mm_mt, 0, 0); 3574 3575 mmap_assert_write_locked(mm); 3576 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3577 3578 mas_for_each(&mas, vma, ULONG_MAX) { 3579 if (vma->anon_vma) 3580 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3581 vm_unlock_anon_vma(avc->anon_vma); 3582 if (vma->vm_file && vma->vm_file->f_mapping) 3583 vm_unlock_mapping(vma->vm_file->f_mapping); 3584 } 3585 3586 mutex_unlock(&mm_all_locks_mutex); 3587 } 3588 3589 /* 3590 * initialise the percpu counter for VM 3591 */ 3592 void __init mmap_init(void) 3593 { 3594 int ret; 3595 3596 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3597 VM_BUG_ON(ret); 3598 } 3599 3600 /* 3601 * Initialise sysctl_user_reserve_kbytes. 3602 * 3603 * This is intended to prevent a user from starting a single memory hogging 3604 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3605 * mode. 3606 * 3607 * The default value is min(3% of free memory, 128MB) 3608 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3609 */ 3610 static int init_user_reserve(void) 3611 { 3612 unsigned long free_kbytes; 3613 3614 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3615 3616 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3617 return 0; 3618 } 3619 subsys_initcall(init_user_reserve); 3620 3621 /* 3622 * Initialise sysctl_admin_reserve_kbytes. 3623 * 3624 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3625 * to log in and kill a memory hogging process. 3626 * 3627 * Systems with more than 256MB will reserve 8MB, enough to recover 3628 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3629 * only reserve 3% of free pages by default. 3630 */ 3631 static int init_admin_reserve(void) 3632 { 3633 unsigned long free_kbytes; 3634 3635 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3636 3637 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3638 return 0; 3639 } 3640 subsys_initcall(init_admin_reserve); 3641 3642 /* 3643 * Reinititalise user and admin reserves if memory is added or removed. 3644 * 3645 * The default user reserve max is 128MB, and the default max for the 3646 * admin reserve is 8MB. These are usually, but not always, enough to 3647 * enable recovery from a memory hogging process using login/sshd, a shell, 3648 * and tools like top. It may make sense to increase or even disable the 3649 * reserve depending on the existence of swap or variations in the recovery 3650 * tools. So, the admin may have changed them. 3651 * 3652 * If memory is added and the reserves have been eliminated or increased above 3653 * the default max, then we'll trust the admin. 3654 * 3655 * If memory is removed and there isn't enough free memory, then we 3656 * need to reset the reserves. 3657 * 3658 * Otherwise keep the reserve set by the admin. 3659 */ 3660 static int reserve_mem_notifier(struct notifier_block *nb, 3661 unsigned long action, void *data) 3662 { 3663 unsigned long tmp, free_kbytes; 3664 3665 switch (action) { 3666 case MEM_ONLINE: 3667 /* Default max is 128MB. Leave alone if modified by operator. */ 3668 tmp = sysctl_user_reserve_kbytes; 3669 if (0 < tmp && tmp < (1UL << 17)) 3670 init_user_reserve(); 3671 3672 /* Default max is 8MB. Leave alone if modified by operator. */ 3673 tmp = sysctl_admin_reserve_kbytes; 3674 if (0 < tmp && tmp < (1UL << 13)) 3675 init_admin_reserve(); 3676 3677 break; 3678 case MEM_OFFLINE: 3679 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3680 3681 if (sysctl_user_reserve_kbytes > free_kbytes) { 3682 init_user_reserve(); 3683 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3684 sysctl_user_reserve_kbytes); 3685 } 3686 3687 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3688 init_admin_reserve(); 3689 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3690 sysctl_admin_reserve_kbytes); 3691 } 3692 break; 3693 default: 3694 break; 3695 } 3696 return NOTIFY_OK; 3697 } 3698 3699 static int __meminit init_reserve_notifier(void) 3700 { 3701 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI)) 3702 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3703 3704 return 0; 3705 } 3706 subsys_initcall(init_reserve_notifier); 3707