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 #include <linux/ksm.h> 50 51 #include <linux/uaccess.h> 52 #include <asm/cacheflush.h> 53 #include <asm/tlb.h> 54 #include <asm/mmu_context.h> 55 56 #define CREATE_TRACE_POINTS 57 #include <trace/events/mmap.h> 58 59 #include "internal.h" 60 61 #ifndef arch_mmap_check 62 #define arch_mmap_check(addr, len, flags) (0) 63 #endif 64 65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS 66 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN; 67 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX; 68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS; 69 #endif 70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS 71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN; 72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX; 73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS; 74 #endif 75 76 static bool ignore_rlimit_data; 77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644); 78 79 static void unmap_region(struct mm_struct *mm, struct maple_tree *mt, 80 struct vm_area_struct *vma, struct vm_area_struct *prev, 81 struct vm_area_struct *next, unsigned long start, 82 unsigned long end, bool mm_wr_locked); 83 84 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags) 85 { 86 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags)); 87 } 88 89 /* Update vma->vm_page_prot to reflect vma->vm_flags. */ 90 void vma_set_page_prot(struct vm_area_struct *vma) 91 { 92 unsigned long vm_flags = vma->vm_flags; 93 pgprot_t vm_page_prot; 94 95 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags); 96 if (vma_wants_writenotify(vma, vm_page_prot)) { 97 vm_flags &= ~VM_SHARED; 98 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags); 99 } 100 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */ 101 WRITE_ONCE(vma->vm_page_prot, vm_page_prot); 102 } 103 104 /* 105 * Requires inode->i_mapping->i_mmap_rwsem 106 */ 107 static void __remove_shared_vm_struct(struct vm_area_struct *vma, 108 struct file *file, struct address_space *mapping) 109 { 110 if (vma->vm_flags & VM_SHARED) 111 mapping_unmap_writable(mapping); 112 113 flush_dcache_mmap_lock(mapping); 114 vma_interval_tree_remove(vma, &mapping->i_mmap); 115 flush_dcache_mmap_unlock(mapping); 116 } 117 118 /* 119 * Unlink a file-based vm structure from its interval tree, to hide 120 * vma from rmap and vmtruncate before freeing its page tables. 121 */ 122 void unlink_file_vma(struct vm_area_struct *vma) 123 { 124 struct file *file = vma->vm_file; 125 126 if (file) { 127 struct address_space *mapping = file->f_mapping; 128 i_mmap_lock_write(mapping); 129 __remove_shared_vm_struct(vma, file, mapping); 130 i_mmap_unlock_write(mapping); 131 } 132 } 133 134 /* 135 * Close a vm structure and free it. 136 */ 137 static void remove_vma(struct vm_area_struct *vma, bool unreachable) 138 { 139 might_sleep(); 140 if (vma->vm_ops && vma->vm_ops->close) 141 vma->vm_ops->close(vma); 142 if (vma->vm_file) 143 fput(vma->vm_file); 144 mpol_put(vma_policy(vma)); 145 if (unreachable) 146 __vm_area_free(vma); 147 else 148 vm_area_free(vma); 149 } 150 151 static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi, 152 unsigned long min) 153 { 154 return mas_prev(&vmi->mas, min); 155 } 156 157 static inline int vma_iter_clear_gfp(struct vma_iterator *vmi, 158 unsigned long start, unsigned long end, gfp_t gfp) 159 { 160 vmi->mas.index = start; 161 vmi->mas.last = end - 1; 162 mas_store_gfp(&vmi->mas, NULL, gfp); 163 if (unlikely(mas_is_err(&vmi->mas))) 164 return -ENOMEM; 165 166 return 0; 167 } 168 169 /* 170 * check_brk_limits() - Use platform specific check of range & verify mlock 171 * limits. 172 * @addr: The address to check 173 * @len: The size of increase. 174 * 175 * Return: 0 on success. 176 */ 177 static int check_brk_limits(unsigned long addr, unsigned long len) 178 { 179 unsigned long mapped_addr; 180 181 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED); 182 if (IS_ERR_VALUE(mapped_addr)) 183 return mapped_addr; 184 185 return mlock_future_ok(current->mm, current->mm->def_flags, len) 186 ? 0 : -EAGAIN; 187 } 188 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma, 189 unsigned long addr, unsigned long request, unsigned long flags); 190 SYSCALL_DEFINE1(brk, unsigned long, brk) 191 { 192 unsigned long newbrk, oldbrk, origbrk; 193 struct mm_struct *mm = current->mm; 194 struct vm_area_struct *brkvma, *next = NULL; 195 unsigned long min_brk; 196 bool populate; 197 bool downgraded = false; 198 LIST_HEAD(uf); 199 struct vma_iterator vmi; 200 201 if (mmap_write_lock_killable(mm)) 202 return -EINTR; 203 204 origbrk = mm->brk; 205 206 #ifdef CONFIG_COMPAT_BRK 207 /* 208 * CONFIG_COMPAT_BRK can still be overridden by setting 209 * randomize_va_space to 2, which will still cause mm->start_brk 210 * to be arbitrarily shifted 211 */ 212 if (current->brk_randomized) 213 min_brk = mm->start_brk; 214 else 215 min_brk = mm->end_data; 216 #else 217 min_brk = mm->start_brk; 218 #endif 219 if (brk < min_brk) 220 goto out; 221 222 /* 223 * Check against rlimit here. If this check is done later after the test 224 * of oldbrk with newbrk then it can escape the test and let the data 225 * segment grow beyond its set limit the in case where the limit is 226 * not page aligned -Ram Gupta 227 */ 228 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk, 229 mm->end_data, mm->start_data)) 230 goto out; 231 232 newbrk = PAGE_ALIGN(brk); 233 oldbrk = PAGE_ALIGN(mm->brk); 234 if (oldbrk == newbrk) { 235 mm->brk = brk; 236 goto success; 237 } 238 239 /* 240 * Always allow shrinking brk. 241 * do_vma_munmap() may downgrade mmap_lock to read. 242 */ 243 if (brk <= mm->brk) { 244 int ret; 245 246 /* Search one past newbrk */ 247 vma_iter_init(&vmi, mm, newbrk); 248 brkvma = vma_find(&vmi, oldbrk); 249 if (!brkvma || brkvma->vm_start >= oldbrk) 250 goto out; /* mapping intersects with an existing non-brk vma. */ 251 /* 252 * mm->brk must be protected by write mmap_lock. 253 * do_vma_munmap() may downgrade the lock, so update it 254 * before calling do_vma_munmap(). 255 */ 256 mm->brk = brk; 257 ret = do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true); 258 if (ret == 1) { 259 downgraded = true; 260 goto success; 261 } else if (!ret) 262 goto success; 263 264 mm->brk = origbrk; 265 goto out; 266 } 267 268 if (check_brk_limits(oldbrk, newbrk - oldbrk)) 269 goto out; 270 271 /* 272 * Only check if the next VMA is within the stack_guard_gap of the 273 * expansion area 274 */ 275 vma_iter_init(&vmi, mm, oldbrk); 276 next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap); 277 if (next && newbrk + PAGE_SIZE > vm_start_gap(next)) 278 goto out; 279 280 brkvma = vma_prev_limit(&vmi, mm->start_brk); 281 /* Ok, looks good - let it rip. */ 282 if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0) 283 goto out; 284 285 mm->brk = brk; 286 287 success: 288 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0; 289 if (downgraded) 290 mmap_read_unlock(mm); 291 else 292 mmap_write_unlock(mm); 293 userfaultfd_unmap_complete(mm, &uf); 294 if (populate) 295 mm_populate(oldbrk, newbrk - oldbrk); 296 return brk; 297 298 out: 299 mmap_write_unlock(mm); 300 return origbrk; 301 } 302 303 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE) 304 static void validate_mm(struct mm_struct *mm) 305 { 306 int bug = 0; 307 int i = 0; 308 struct vm_area_struct *vma; 309 VMA_ITERATOR(vmi, mm, 0); 310 311 mt_validate(&mm->mm_mt); 312 for_each_vma(vmi, vma) { 313 #ifdef CONFIG_DEBUG_VM_RB 314 struct anon_vma *anon_vma = vma->anon_vma; 315 struct anon_vma_chain *avc; 316 #endif 317 unsigned long vmi_start, vmi_end; 318 bool warn = 0; 319 320 vmi_start = vma_iter_addr(&vmi); 321 vmi_end = vma_iter_end(&vmi); 322 if (VM_WARN_ON_ONCE_MM(vma->vm_end != vmi_end, mm)) 323 warn = 1; 324 325 if (VM_WARN_ON_ONCE_MM(vma->vm_start != vmi_start, mm)) 326 warn = 1; 327 328 if (warn) { 329 pr_emerg("issue in %s\n", current->comm); 330 dump_stack(); 331 dump_vma(vma); 332 pr_emerg("tree range: %px start %lx end %lx\n", vma, 333 vmi_start, vmi_end - 1); 334 vma_iter_dump_tree(&vmi); 335 } 336 337 #ifdef CONFIG_DEBUG_VM_RB 338 if (anon_vma) { 339 anon_vma_lock_read(anon_vma); 340 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 341 anon_vma_interval_tree_verify(avc); 342 anon_vma_unlock_read(anon_vma); 343 } 344 #endif 345 i++; 346 } 347 if (i != mm->map_count) { 348 pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i); 349 bug = 1; 350 } 351 VM_BUG_ON_MM(bug, mm); 352 } 353 354 #else /* !CONFIG_DEBUG_VM_MAPLE_TREE */ 355 #define validate_mm(mm) do { } while (0) 356 #endif /* CONFIG_DEBUG_VM_MAPLE_TREE */ 357 358 /* 359 * vma has some anon_vma assigned, and is already inserted on that 360 * anon_vma's interval trees. 361 * 362 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the 363 * vma must be removed from the anon_vma's interval trees using 364 * anon_vma_interval_tree_pre_update_vma(). 365 * 366 * After the update, the vma will be reinserted using 367 * anon_vma_interval_tree_post_update_vma(). 368 * 369 * The entire update must be protected by exclusive mmap_lock and by 370 * the root anon_vma's mutex. 371 */ 372 static inline void 373 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma) 374 { 375 struct anon_vma_chain *avc; 376 377 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 378 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root); 379 } 380 381 static inline void 382 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma) 383 { 384 struct anon_vma_chain *avc; 385 386 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 387 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root); 388 } 389 390 static unsigned long count_vma_pages_range(struct mm_struct *mm, 391 unsigned long addr, unsigned long end) 392 { 393 VMA_ITERATOR(vmi, mm, addr); 394 struct vm_area_struct *vma; 395 unsigned long nr_pages = 0; 396 397 for_each_vma_range(vmi, vma, end) { 398 unsigned long vm_start = max(addr, vma->vm_start); 399 unsigned long vm_end = min(end, vma->vm_end); 400 401 nr_pages += PHYS_PFN(vm_end - vm_start); 402 } 403 404 return nr_pages; 405 } 406 407 static void __vma_link_file(struct vm_area_struct *vma, 408 struct address_space *mapping) 409 { 410 if (vma->vm_flags & VM_SHARED) 411 mapping_allow_writable(mapping); 412 413 flush_dcache_mmap_lock(mapping); 414 vma_interval_tree_insert(vma, &mapping->i_mmap); 415 flush_dcache_mmap_unlock(mapping); 416 } 417 418 static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma) 419 { 420 VMA_ITERATOR(vmi, mm, 0); 421 struct address_space *mapping = NULL; 422 423 if (vma_iter_prealloc(&vmi)) 424 return -ENOMEM; 425 426 if (vma->vm_file) { 427 mapping = vma->vm_file->f_mapping; 428 i_mmap_lock_write(mapping); 429 } 430 431 vma_iter_store(&vmi, vma); 432 433 if (mapping) { 434 __vma_link_file(vma, mapping); 435 i_mmap_unlock_write(mapping); 436 } 437 438 mm->map_count++; 439 validate_mm(mm); 440 return 0; 441 } 442 443 /* 444 * init_multi_vma_prep() - Initializer for struct vma_prepare 445 * @vp: The vma_prepare struct 446 * @vma: The vma that will be altered once locked 447 * @next: The next vma if it is to be adjusted 448 * @remove: The first vma to be removed 449 * @remove2: The second vma to be removed 450 */ 451 static inline void init_multi_vma_prep(struct vma_prepare *vp, 452 struct vm_area_struct *vma, struct vm_area_struct *next, 453 struct vm_area_struct *remove, struct vm_area_struct *remove2) 454 { 455 memset(vp, 0, sizeof(struct vma_prepare)); 456 vp->vma = vma; 457 vp->anon_vma = vma->anon_vma; 458 vp->remove = remove; 459 vp->remove2 = remove2; 460 vp->adj_next = next; 461 if (!vp->anon_vma && next) 462 vp->anon_vma = next->anon_vma; 463 464 vp->file = vma->vm_file; 465 if (vp->file) 466 vp->mapping = vma->vm_file->f_mapping; 467 468 } 469 470 /* 471 * init_vma_prep() - Initializer wrapper for vma_prepare struct 472 * @vp: The vma_prepare struct 473 * @vma: The vma that will be altered once locked 474 */ 475 static inline void init_vma_prep(struct vma_prepare *vp, 476 struct vm_area_struct *vma) 477 { 478 init_multi_vma_prep(vp, vma, NULL, NULL, NULL); 479 } 480 481 482 /* 483 * vma_prepare() - Helper function for handling locking VMAs prior to altering 484 * @vp: The initialized vma_prepare struct 485 */ 486 static inline void vma_prepare(struct vma_prepare *vp) 487 { 488 vma_start_write(vp->vma); 489 if (vp->adj_next) 490 vma_start_write(vp->adj_next); 491 /* vp->insert is always a newly created VMA, no need for locking */ 492 if (vp->remove) 493 vma_start_write(vp->remove); 494 if (vp->remove2) 495 vma_start_write(vp->remove2); 496 497 if (vp->file) { 498 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end); 499 500 if (vp->adj_next) 501 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start, 502 vp->adj_next->vm_end); 503 504 i_mmap_lock_write(vp->mapping); 505 if (vp->insert && vp->insert->vm_file) { 506 /* 507 * Put into interval tree now, so instantiated pages 508 * are visible to arm/parisc __flush_dcache_page 509 * throughout; but we cannot insert into address 510 * space until vma start or end is updated. 511 */ 512 __vma_link_file(vp->insert, 513 vp->insert->vm_file->f_mapping); 514 } 515 } 516 517 if (vp->anon_vma) { 518 anon_vma_lock_write(vp->anon_vma); 519 anon_vma_interval_tree_pre_update_vma(vp->vma); 520 if (vp->adj_next) 521 anon_vma_interval_tree_pre_update_vma(vp->adj_next); 522 } 523 524 if (vp->file) { 525 flush_dcache_mmap_lock(vp->mapping); 526 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap); 527 if (vp->adj_next) 528 vma_interval_tree_remove(vp->adj_next, 529 &vp->mapping->i_mmap); 530 } 531 532 } 533 534 /* 535 * vma_complete- Helper function for handling the unlocking after altering VMAs, 536 * or for inserting a VMA. 537 * 538 * @vp: The vma_prepare struct 539 * @vmi: The vma iterator 540 * @mm: The mm_struct 541 */ 542 static inline void vma_complete(struct vma_prepare *vp, 543 struct vma_iterator *vmi, struct mm_struct *mm) 544 { 545 if (vp->file) { 546 if (vp->adj_next) 547 vma_interval_tree_insert(vp->adj_next, 548 &vp->mapping->i_mmap); 549 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap); 550 flush_dcache_mmap_unlock(vp->mapping); 551 } 552 553 if (vp->remove && vp->file) { 554 __remove_shared_vm_struct(vp->remove, vp->file, vp->mapping); 555 if (vp->remove2) 556 __remove_shared_vm_struct(vp->remove2, vp->file, 557 vp->mapping); 558 } else if (vp->insert) { 559 /* 560 * split_vma has split insert from vma, and needs 561 * us to insert it before dropping the locks 562 * (it may either follow vma or precede it). 563 */ 564 vma_iter_store(vmi, vp->insert); 565 mm->map_count++; 566 } 567 568 if (vp->anon_vma) { 569 anon_vma_interval_tree_post_update_vma(vp->vma); 570 if (vp->adj_next) 571 anon_vma_interval_tree_post_update_vma(vp->adj_next); 572 anon_vma_unlock_write(vp->anon_vma); 573 } 574 575 if (vp->file) { 576 i_mmap_unlock_write(vp->mapping); 577 uprobe_mmap(vp->vma); 578 579 if (vp->adj_next) 580 uprobe_mmap(vp->adj_next); 581 } 582 583 if (vp->remove) { 584 again: 585 vma_mark_detached(vp->remove, true); 586 if (vp->file) { 587 uprobe_munmap(vp->remove, vp->remove->vm_start, 588 vp->remove->vm_end); 589 fput(vp->file); 590 } 591 if (vp->remove->anon_vma) 592 anon_vma_merge(vp->vma, vp->remove); 593 mm->map_count--; 594 mpol_put(vma_policy(vp->remove)); 595 if (!vp->remove2) 596 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end); 597 vm_area_free(vp->remove); 598 599 /* 600 * In mprotect's case 6 (see comments on vma_merge), 601 * we are removing both mid and next vmas 602 */ 603 if (vp->remove2) { 604 vp->remove = vp->remove2; 605 vp->remove2 = NULL; 606 goto again; 607 } 608 } 609 if (vp->insert && vp->file) 610 uprobe_mmap(vp->insert); 611 } 612 613 /* 614 * dup_anon_vma() - Helper function to duplicate anon_vma 615 * @dst: The destination VMA 616 * @src: The source VMA 617 * 618 * Returns: 0 on success. 619 */ 620 static inline int dup_anon_vma(struct vm_area_struct *dst, 621 struct vm_area_struct *src) 622 { 623 /* 624 * Easily overlooked: when mprotect shifts the boundary, make sure the 625 * expanding vma has anon_vma set if the shrinking vma had, to cover any 626 * anon pages imported. 627 */ 628 if (src->anon_vma && !dst->anon_vma) { 629 dst->anon_vma = src->anon_vma; 630 return anon_vma_clone(dst, src); 631 } 632 633 return 0; 634 } 635 636 /* 637 * vma_expand - Expand an existing VMA 638 * 639 * @vmi: The vma iterator 640 * @vma: The vma to expand 641 * @start: The start of the vma 642 * @end: The exclusive end of the vma 643 * @pgoff: The page offset of vma 644 * @next: The current of next vma. 645 * 646 * Expand @vma to @start and @end. Can expand off the start and end. Will 647 * expand over @next if it's different from @vma and @end == @next->vm_end. 648 * Checking if the @vma can expand and merge with @next needs to be handled by 649 * the caller. 650 * 651 * Returns: 0 on success 652 */ 653 int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma, 654 unsigned long start, unsigned long end, pgoff_t pgoff, 655 struct vm_area_struct *next) 656 { 657 bool remove_next = false; 658 struct vma_prepare vp; 659 660 if (next && (vma != next) && (end == next->vm_end)) { 661 int ret; 662 663 remove_next = true; 664 ret = dup_anon_vma(vma, next); 665 if (ret) 666 return ret; 667 } 668 669 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL); 670 /* Not merging but overwriting any part of next is not handled. */ 671 VM_WARN_ON(next && !vp.remove && 672 next != vma && end > next->vm_start); 673 /* Only handles expanding */ 674 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end); 675 676 if (vma_iter_prealloc(vmi)) 677 goto nomem; 678 679 vma_prepare(&vp); 680 vma_adjust_trans_huge(vma, start, end, 0); 681 /* VMA iterator points to previous, so set to start if necessary */ 682 if (vma_iter_addr(vmi) != start) 683 vma_iter_set(vmi, start); 684 685 vma->vm_start = start; 686 vma->vm_end = end; 687 vma->vm_pgoff = pgoff; 688 /* Note: mas must be pointing to the expanding VMA */ 689 vma_iter_store(vmi, vma); 690 691 vma_complete(&vp, vmi, vma->vm_mm); 692 validate_mm(vma->vm_mm); 693 return 0; 694 695 nomem: 696 return -ENOMEM; 697 } 698 699 /* 700 * vma_shrink() - Reduce an existing VMAs memory area 701 * @vmi: The vma iterator 702 * @vma: The VMA to modify 703 * @start: The new start 704 * @end: The new end 705 * 706 * Returns: 0 on success, -ENOMEM otherwise 707 */ 708 int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma, 709 unsigned long start, unsigned long end, pgoff_t pgoff) 710 { 711 struct vma_prepare vp; 712 713 WARN_ON((vma->vm_start != start) && (vma->vm_end != end)); 714 715 if (vma_iter_prealloc(vmi)) 716 return -ENOMEM; 717 718 init_vma_prep(&vp, vma); 719 vma_prepare(&vp); 720 vma_adjust_trans_huge(vma, start, end, 0); 721 722 if (vma->vm_start < start) 723 vma_iter_clear(vmi, vma->vm_start, start); 724 725 if (vma->vm_end > end) 726 vma_iter_clear(vmi, end, vma->vm_end); 727 728 vma->vm_start = start; 729 vma->vm_end = end; 730 vma->vm_pgoff = pgoff; 731 vma_complete(&vp, vmi, vma->vm_mm); 732 validate_mm(vma->vm_mm); 733 return 0; 734 } 735 736 /* 737 * If the vma has a ->close operation then the driver probably needs to release 738 * per-vma resources, so we don't attempt to merge those if the caller indicates 739 * the current vma may be removed as part of the merge. 740 */ 741 static inline bool is_mergeable_vma(struct vm_area_struct *vma, 742 struct file *file, unsigned long vm_flags, 743 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 744 struct anon_vma_name *anon_name, bool may_remove_vma) 745 { 746 /* 747 * VM_SOFTDIRTY should not prevent from VMA merging, if we 748 * match the flags but dirty bit -- the caller should mark 749 * merged VMA as dirty. If dirty bit won't be excluded from 750 * comparison, we increase pressure on the memory system forcing 751 * the kernel to generate new VMAs when old one could be 752 * extended instead. 753 */ 754 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY) 755 return false; 756 if (vma->vm_file != file) 757 return false; 758 if (may_remove_vma && vma->vm_ops && vma->vm_ops->close) 759 return false; 760 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx)) 761 return false; 762 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name)) 763 return false; 764 return true; 765 } 766 767 static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1, 768 struct anon_vma *anon_vma2, struct vm_area_struct *vma) 769 { 770 /* 771 * The list_is_singular() test is to avoid merging VMA cloned from 772 * parents. This can improve scalability caused by anon_vma lock. 773 */ 774 if ((!anon_vma1 || !anon_vma2) && (!vma || 775 list_is_singular(&vma->anon_vma_chain))) 776 return true; 777 return anon_vma1 == anon_vma2; 778 } 779 780 /* 781 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 782 * in front of (at a lower virtual address and file offset than) the vma. 783 * 784 * We cannot merge two vmas if they have differently assigned (non-NULL) 785 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 786 * 787 * We don't check here for the merged mmap wrapping around the end of pagecache 788 * indices (16TB on ia32) because do_mmap() does not permit mmap's which 789 * wrap, nor mmaps which cover the final page at index -1UL. 790 * 791 * We assume the vma may be removed as part of the merge. 792 */ 793 static bool 794 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags, 795 struct anon_vma *anon_vma, struct file *file, 796 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 797 struct anon_vma_name *anon_name) 798 { 799 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) && 800 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 801 if (vma->vm_pgoff == vm_pgoff) 802 return true; 803 } 804 return false; 805 } 806 807 /* 808 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff) 809 * beyond (at a higher virtual address and file offset than) the vma. 810 * 811 * We cannot merge two vmas if they have differently assigned (non-NULL) 812 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible. 813 * 814 * We assume that vma is not removed as part of the merge. 815 */ 816 static bool 817 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags, 818 struct anon_vma *anon_vma, struct file *file, 819 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 820 struct anon_vma_name *anon_name) 821 { 822 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) && 823 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) { 824 pgoff_t vm_pglen; 825 vm_pglen = vma_pages(vma); 826 if (vma->vm_pgoff + vm_pglen == vm_pgoff) 827 return true; 828 } 829 return false; 830 } 831 832 /* 833 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name), 834 * figure out whether that can be merged with its predecessor or its 835 * successor. Or both (it neatly fills a hole). 836 * 837 * In most cases - when called for mmap, brk or mremap - [addr,end) is 838 * certain not to be mapped by the time vma_merge is called; but when 839 * called for mprotect, it is certain to be already mapped (either at 840 * an offset within prev, or at the start of next), and the flags of 841 * this area are about to be changed to vm_flags - and the no-change 842 * case has already been eliminated. 843 * 844 * The following mprotect cases have to be considered, where **** is 845 * the area passed down from mprotect_fixup, never extending beyond one 846 * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts 847 * at the same address as **** and is of the same or larger span, and 848 * NNNN the next vma after ****: 849 * 850 * **** **** **** 851 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPCCCCCC 852 * cannot merge might become might become 853 * PPNNNNNNNNNN PPPPPPPPPPCC 854 * mmap, brk or case 4 below case 5 below 855 * mremap move: 856 * **** **** 857 * PPPP NNNN PPPPCCCCNNNN 858 * might become might become 859 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or 860 * PPPPPPPPNNNN 2 or PPPPPPPPNNNN 7 or 861 * PPPPNNNNNNNN 3 PPPPNNNNNNNN 8 862 * 863 * It is important for case 8 that the vma CCCC overlapping the 864 * region **** is never going to extended over NNNN. Instead NNNN must 865 * be extended in region **** and CCCC must be removed. This way in 866 * all cases where vma_merge succeeds, the moment vma_merge drops the 867 * rmap_locks, the properties of the merged vma will be already 868 * correct for the whole merged range. Some of those properties like 869 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must 870 * be correct for the whole merged range immediately after the 871 * rmap_locks are released. Otherwise if NNNN would be removed and 872 * CCCC would be extended over the NNNN range, remove_migration_ptes 873 * or other rmap walkers (if working on addresses beyond the "end" 874 * parameter) may establish ptes with the wrong permissions of CCCC 875 * instead of the right permissions of NNNN. 876 * 877 * In the code below: 878 * PPPP is represented by *prev 879 * CCCC is represented by *curr or not represented at all (NULL) 880 * NNNN is represented by *next or not represented at all (NULL) 881 * **** is not represented - it will be merged and the vma containing the 882 * area is returned, or the function will return NULL 883 */ 884 struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm, 885 struct vm_area_struct *prev, unsigned long addr, 886 unsigned long end, unsigned long vm_flags, 887 struct anon_vma *anon_vma, struct file *file, 888 pgoff_t pgoff, struct mempolicy *policy, 889 struct vm_userfaultfd_ctx vm_userfaultfd_ctx, 890 struct anon_vma_name *anon_name) 891 { 892 struct vm_area_struct *curr, *next, *res; 893 struct vm_area_struct *vma, *adjust, *remove, *remove2; 894 struct vma_prepare vp; 895 pgoff_t vma_pgoff; 896 int err = 0; 897 bool merge_prev = false; 898 bool merge_next = false; 899 bool vma_expanded = false; 900 unsigned long vma_start = addr; 901 unsigned long vma_end = end; 902 pgoff_t pglen = (end - addr) >> PAGE_SHIFT; 903 long adj_start = 0; 904 905 validate_mm(mm); 906 /* 907 * We later require that vma->vm_flags == vm_flags, 908 * so this tests vma->vm_flags & VM_SPECIAL, too. 909 */ 910 if (vm_flags & VM_SPECIAL) 911 return NULL; 912 913 /* Does the input range span an existing VMA? (cases 5 - 8) */ 914 curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end); 915 916 if (!curr || /* cases 1 - 4 */ 917 end == curr->vm_end) /* cases 6 - 8, adjacent VMA */ 918 next = vma_lookup(mm, end); 919 else 920 next = NULL; /* case 5 */ 921 922 if (prev) { 923 vma_start = prev->vm_start; 924 vma_pgoff = prev->vm_pgoff; 925 926 /* Can we merge the predecessor? */ 927 if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy) 928 && can_vma_merge_after(prev, vm_flags, anon_vma, file, 929 pgoff, vm_userfaultfd_ctx, anon_name)) { 930 merge_prev = true; 931 vma_prev(vmi); 932 } 933 } 934 935 /* Can we merge the successor? */ 936 if (next && mpol_equal(policy, vma_policy(next)) && 937 can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen, 938 vm_userfaultfd_ctx, anon_name)) { 939 merge_next = true; 940 } 941 942 /* Verify some invariant that must be enforced by the caller. */ 943 VM_WARN_ON(prev && addr <= prev->vm_start); 944 VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end)); 945 VM_WARN_ON(addr >= end); 946 947 if (!merge_prev && !merge_next) 948 return NULL; /* Not mergeable. */ 949 950 res = vma = prev; 951 remove = remove2 = adjust = NULL; 952 953 /* Can we merge both the predecessor and the successor? */ 954 if (merge_prev && merge_next && 955 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) { 956 remove = next; /* case 1 */ 957 vma_end = next->vm_end; 958 err = dup_anon_vma(prev, next); 959 if (curr) { /* case 6 */ 960 remove = curr; 961 remove2 = next; 962 if (!next->anon_vma) 963 err = dup_anon_vma(prev, curr); 964 } 965 } else if (merge_prev) { /* case 2 */ 966 if (curr) { 967 err = dup_anon_vma(prev, curr); 968 if (end == curr->vm_end) { /* case 7 */ 969 remove = curr; 970 } else { /* case 5 */ 971 adjust = curr; 972 adj_start = (end - curr->vm_start); 973 } 974 } 975 } else { /* merge_next */ 976 res = next; 977 if (prev && addr < prev->vm_end) { /* case 4 */ 978 vma_end = addr; 979 adjust = next; 980 adj_start = -(prev->vm_end - addr); 981 err = dup_anon_vma(next, prev); 982 } else { 983 /* 984 * Note that cases 3 and 8 are the ONLY ones where prev 985 * is permitted to be (but is not necessarily) NULL. 986 */ 987 vma = next; /* case 3 */ 988 vma_start = addr; 989 vma_end = next->vm_end; 990 vma_pgoff = next->vm_pgoff - pglen; 991 if (curr) { /* case 8 */ 992 vma_pgoff = curr->vm_pgoff; 993 remove = curr; 994 err = dup_anon_vma(next, curr); 995 } 996 } 997 } 998 999 /* Error in anon_vma clone. */ 1000 if (err) 1001 return NULL; 1002 1003 if (vma_iter_prealloc(vmi)) 1004 return NULL; 1005 1006 init_multi_vma_prep(&vp, vma, adjust, remove, remove2); 1007 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma && 1008 vp.anon_vma != adjust->anon_vma); 1009 1010 vma_prepare(&vp); 1011 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start); 1012 if (vma_start < vma->vm_start || vma_end > vma->vm_end) 1013 vma_expanded = true; 1014 1015 vma->vm_start = vma_start; 1016 vma->vm_end = vma_end; 1017 vma->vm_pgoff = vma_pgoff; 1018 1019 if (vma_expanded) 1020 vma_iter_store(vmi, vma); 1021 1022 if (adj_start) { 1023 adjust->vm_start += adj_start; 1024 adjust->vm_pgoff += adj_start >> PAGE_SHIFT; 1025 if (adj_start < 0) { 1026 WARN_ON(vma_expanded); 1027 vma_iter_store(vmi, next); 1028 } 1029 } 1030 1031 vma_complete(&vp, vmi, mm); 1032 vma_iter_free(vmi); 1033 validate_mm(mm); 1034 khugepaged_enter_vma(res, vm_flags); 1035 1036 return res; 1037 } 1038 1039 /* 1040 * Rough compatibility check to quickly see if it's even worth looking 1041 * at sharing an anon_vma. 1042 * 1043 * They need to have the same vm_file, and the flags can only differ 1044 * in things that mprotect may change. 1045 * 1046 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that 1047 * we can merge the two vma's. For example, we refuse to merge a vma if 1048 * there is a vm_ops->close() function, because that indicates that the 1049 * driver is doing some kind of reference counting. But that doesn't 1050 * really matter for the anon_vma sharing case. 1051 */ 1052 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b) 1053 { 1054 return a->vm_end == b->vm_start && 1055 mpol_equal(vma_policy(a), vma_policy(b)) && 1056 a->vm_file == b->vm_file && 1057 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) && 1058 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT); 1059 } 1060 1061 /* 1062 * Do some basic sanity checking to see if we can re-use the anon_vma 1063 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be 1064 * the same as 'old', the other will be the new one that is trying 1065 * to share the anon_vma. 1066 * 1067 * NOTE! This runs with mmap_lock held for reading, so it is possible that 1068 * the anon_vma of 'old' is concurrently in the process of being set up 1069 * by another page fault trying to merge _that_. But that's ok: if it 1070 * is being set up, that automatically means that it will be a singleton 1071 * acceptable for merging, so we can do all of this optimistically. But 1072 * we do that READ_ONCE() to make sure that we never re-load the pointer. 1073 * 1074 * IOW: that the "list_is_singular()" test on the anon_vma_chain only 1075 * matters for the 'stable anon_vma' case (ie the thing we want to avoid 1076 * is to return an anon_vma that is "complex" due to having gone through 1077 * a fork). 1078 * 1079 * We also make sure that the two vma's are compatible (adjacent, 1080 * and with the same memory policies). That's all stable, even with just 1081 * a read lock on the mmap_lock. 1082 */ 1083 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b) 1084 { 1085 if (anon_vma_compatible(a, b)) { 1086 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma); 1087 1088 if (anon_vma && list_is_singular(&old->anon_vma_chain)) 1089 return anon_vma; 1090 } 1091 return NULL; 1092 } 1093 1094 /* 1095 * find_mergeable_anon_vma is used by anon_vma_prepare, to check 1096 * neighbouring vmas for a suitable anon_vma, before it goes off 1097 * to allocate a new anon_vma. It checks because a repetitive 1098 * sequence of mprotects and faults may otherwise lead to distinct 1099 * anon_vmas being allocated, preventing vma merge in subsequent 1100 * mprotect. 1101 */ 1102 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma) 1103 { 1104 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end); 1105 struct anon_vma *anon_vma = NULL; 1106 struct vm_area_struct *prev, *next; 1107 1108 /* Try next first. */ 1109 next = mas_walk(&mas); 1110 if (next) { 1111 anon_vma = reusable_anon_vma(next, vma, next); 1112 if (anon_vma) 1113 return anon_vma; 1114 } 1115 1116 prev = mas_prev(&mas, 0); 1117 VM_BUG_ON_VMA(prev != vma, vma); 1118 prev = mas_prev(&mas, 0); 1119 /* Try prev next. */ 1120 if (prev) 1121 anon_vma = reusable_anon_vma(prev, prev, vma); 1122 1123 /* 1124 * We might reach here with anon_vma == NULL if we can't find 1125 * any reusable anon_vma. 1126 * There's no absolute need to look only at touching neighbours: 1127 * we could search further afield for "compatible" anon_vmas. 1128 * But it would probably just be a waste of time searching, 1129 * or lead to too many vmas hanging off the same anon_vma. 1130 * We're trying to allow mprotect remerging later on, 1131 * not trying to minimize memory used for anon_vmas. 1132 */ 1133 return anon_vma; 1134 } 1135 1136 /* 1137 * If a hint addr is less than mmap_min_addr change hint to be as 1138 * low as possible but still greater than mmap_min_addr 1139 */ 1140 static inline unsigned long round_hint_to_min(unsigned long hint) 1141 { 1142 hint &= PAGE_MASK; 1143 if (((void *)hint != NULL) && 1144 (hint < mmap_min_addr)) 1145 return PAGE_ALIGN(mmap_min_addr); 1146 return hint; 1147 } 1148 1149 bool mlock_future_ok(struct mm_struct *mm, unsigned long flags, 1150 unsigned long bytes) 1151 { 1152 unsigned long locked_pages, limit_pages; 1153 1154 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK)) 1155 return true; 1156 1157 locked_pages = bytes >> PAGE_SHIFT; 1158 locked_pages += mm->locked_vm; 1159 1160 limit_pages = rlimit(RLIMIT_MEMLOCK); 1161 limit_pages >>= PAGE_SHIFT; 1162 1163 return locked_pages <= limit_pages; 1164 } 1165 1166 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode) 1167 { 1168 if (S_ISREG(inode->i_mode)) 1169 return MAX_LFS_FILESIZE; 1170 1171 if (S_ISBLK(inode->i_mode)) 1172 return MAX_LFS_FILESIZE; 1173 1174 if (S_ISSOCK(inode->i_mode)) 1175 return MAX_LFS_FILESIZE; 1176 1177 /* Special "we do even unsigned file positions" case */ 1178 if (file->f_mode & FMODE_UNSIGNED_OFFSET) 1179 return 0; 1180 1181 /* Yes, random drivers might want more. But I'm tired of buggy drivers */ 1182 return ULONG_MAX; 1183 } 1184 1185 static inline bool file_mmap_ok(struct file *file, struct inode *inode, 1186 unsigned long pgoff, unsigned long len) 1187 { 1188 u64 maxsize = file_mmap_size_max(file, inode); 1189 1190 if (maxsize && len > maxsize) 1191 return false; 1192 maxsize -= len; 1193 if (pgoff > maxsize >> PAGE_SHIFT) 1194 return false; 1195 return true; 1196 } 1197 1198 /* 1199 * The caller must write-lock current->mm->mmap_lock. 1200 */ 1201 unsigned long do_mmap(struct file *file, unsigned long addr, 1202 unsigned long len, unsigned long prot, 1203 unsigned long flags, unsigned long pgoff, 1204 unsigned long *populate, struct list_head *uf) 1205 { 1206 struct mm_struct *mm = current->mm; 1207 vm_flags_t vm_flags; 1208 int pkey = 0; 1209 1210 validate_mm(mm); 1211 *populate = 0; 1212 1213 if (!len) 1214 return -EINVAL; 1215 1216 /* 1217 * Does the application expect PROT_READ to imply PROT_EXEC? 1218 * 1219 * (the exception is when the underlying filesystem is noexec 1220 * mounted, in which case we dont add PROT_EXEC.) 1221 */ 1222 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 1223 if (!(file && path_noexec(&file->f_path))) 1224 prot |= PROT_EXEC; 1225 1226 /* force arch specific MAP_FIXED handling in get_unmapped_area */ 1227 if (flags & MAP_FIXED_NOREPLACE) 1228 flags |= MAP_FIXED; 1229 1230 if (!(flags & MAP_FIXED)) 1231 addr = round_hint_to_min(addr); 1232 1233 /* Careful about overflows.. */ 1234 len = PAGE_ALIGN(len); 1235 if (!len) 1236 return -ENOMEM; 1237 1238 /* offset overflow? */ 1239 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff) 1240 return -EOVERFLOW; 1241 1242 /* Too many mappings? */ 1243 if (mm->map_count > sysctl_max_map_count) 1244 return -ENOMEM; 1245 1246 /* Obtain the address to map to. we verify (or select) it and ensure 1247 * that it represents a valid section of the address space. 1248 */ 1249 addr = get_unmapped_area(file, addr, len, pgoff, flags); 1250 if (IS_ERR_VALUE(addr)) 1251 return addr; 1252 1253 if (flags & MAP_FIXED_NOREPLACE) { 1254 if (find_vma_intersection(mm, addr, addr + len)) 1255 return -EEXIST; 1256 } 1257 1258 if (prot == PROT_EXEC) { 1259 pkey = execute_only_pkey(mm); 1260 if (pkey < 0) 1261 pkey = 0; 1262 } 1263 1264 /* Do simple checking here so the lower-level routines won't have 1265 * to. we assume access permissions have been handled by the open 1266 * of the memory object, so we don't do any here. 1267 */ 1268 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) | 1269 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC; 1270 1271 if (flags & MAP_LOCKED) 1272 if (!can_do_mlock()) 1273 return -EPERM; 1274 1275 if (!mlock_future_ok(mm, vm_flags, len)) 1276 return -EAGAIN; 1277 1278 if (file) { 1279 struct inode *inode = file_inode(file); 1280 unsigned long flags_mask; 1281 1282 if (!file_mmap_ok(file, inode, pgoff, len)) 1283 return -EOVERFLOW; 1284 1285 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags; 1286 1287 switch (flags & MAP_TYPE) { 1288 case MAP_SHARED: 1289 /* 1290 * Force use of MAP_SHARED_VALIDATE with non-legacy 1291 * flags. E.g. MAP_SYNC is dangerous to use with 1292 * MAP_SHARED as you don't know which consistency model 1293 * you will get. We silently ignore unsupported flags 1294 * with MAP_SHARED to preserve backward compatibility. 1295 */ 1296 flags &= LEGACY_MAP_MASK; 1297 fallthrough; 1298 case MAP_SHARED_VALIDATE: 1299 if (flags & ~flags_mask) 1300 return -EOPNOTSUPP; 1301 if (prot & PROT_WRITE) { 1302 if (!(file->f_mode & FMODE_WRITE)) 1303 return -EACCES; 1304 if (IS_SWAPFILE(file->f_mapping->host)) 1305 return -ETXTBSY; 1306 } 1307 1308 /* 1309 * Make sure we don't allow writing to an append-only 1310 * file.. 1311 */ 1312 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE)) 1313 return -EACCES; 1314 1315 vm_flags |= VM_SHARED | VM_MAYSHARE; 1316 if (!(file->f_mode & FMODE_WRITE)) 1317 vm_flags &= ~(VM_MAYWRITE | VM_SHARED); 1318 fallthrough; 1319 case MAP_PRIVATE: 1320 if (!(file->f_mode & FMODE_READ)) 1321 return -EACCES; 1322 if (path_noexec(&file->f_path)) { 1323 if (vm_flags & VM_EXEC) 1324 return -EPERM; 1325 vm_flags &= ~VM_MAYEXEC; 1326 } 1327 1328 if (!file->f_op->mmap) 1329 return -ENODEV; 1330 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1331 return -EINVAL; 1332 break; 1333 1334 default: 1335 return -EINVAL; 1336 } 1337 } else { 1338 switch (flags & MAP_TYPE) { 1339 case MAP_SHARED: 1340 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP)) 1341 return -EINVAL; 1342 /* 1343 * Ignore pgoff. 1344 */ 1345 pgoff = 0; 1346 vm_flags |= VM_SHARED | VM_MAYSHARE; 1347 break; 1348 case MAP_PRIVATE: 1349 /* 1350 * Set pgoff according to addr for anon_vma. 1351 */ 1352 pgoff = addr >> PAGE_SHIFT; 1353 break; 1354 default: 1355 return -EINVAL; 1356 } 1357 } 1358 1359 /* 1360 * Set 'VM_NORESERVE' if we should not account for the 1361 * memory use of this mapping. 1362 */ 1363 if (flags & MAP_NORESERVE) { 1364 /* We honor MAP_NORESERVE if allowed to overcommit */ 1365 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER) 1366 vm_flags |= VM_NORESERVE; 1367 1368 /* hugetlb applies strict overcommit unless MAP_NORESERVE */ 1369 if (file && is_file_hugepages(file)) 1370 vm_flags |= VM_NORESERVE; 1371 } 1372 1373 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf); 1374 if (!IS_ERR_VALUE(addr) && 1375 ((vm_flags & VM_LOCKED) || 1376 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE)) 1377 *populate = len; 1378 return addr; 1379 } 1380 1381 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len, 1382 unsigned long prot, unsigned long flags, 1383 unsigned long fd, unsigned long pgoff) 1384 { 1385 struct file *file = NULL; 1386 unsigned long retval; 1387 1388 if (!(flags & MAP_ANONYMOUS)) { 1389 audit_mmap_fd(fd, flags); 1390 file = fget(fd); 1391 if (!file) 1392 return -EBADF; 1393 if (is_file_hugepages(file)) { 1394 len = ALIGN(len, huge_page_size(hstate_file(file))); 1395 } else if (unlikely(flags & MAP_HUGETLB)) { 1396 retval = -EINVAL; 1397 goto out_fput; 1398 } 1399 } else if (flags & MAP_HUGETLB) { 1400 struct hstate *hs; 1401 1402 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1403 if (!hs) 1404 return -EINVAL; 1405 1406 len = ALIGN(len, huge_page_size(hs)); 1407 /* 1408 * VM_NORESERVE is used because the reservations will be 1409 * taken when vm_ops->mmap() is called 1410 */ 1411 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, 1412 VM_NORESERVE, 1413 HUGETLB_ANONHUGE_INODE, 1414 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK); 1415 if (IS_ERR(file)) 1416 return PTR_ERR(file); 1417 } 1418 1419 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff); 1420 out_fput: 1421 if (file) 1422 fput(file); 1423 return retval; 1424 } 1425 1426 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len, 1427 unsigned long, prot, unsigned long, flags, 1428 unsigned long, fd, unsigned long, pgoff) 1429 { 1430 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff); 1431 } 1432 1433 #ifdef __ARCH_WANT_SYS_OLD_MMAP 1434 struct mmap_arg_struct { 1435 unsigned long addr; 1436 unsigned long len; 1437 unsigned long prot; 1438 unsigned long flags; 1439 unsigned long fd; 1440 unsigned long offset; 1441 }; 1442 1443 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg) 1444 { 1445 struct mmap_arg_struct a; 1446 1447 if (copy_from_user(&a, arg, sizeof(a))) 1448 return -EFAULT; 1449 if (offset_in_page(a.offset)) 1450 return -EINVAL; 1451 1452 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd, 1453 a.offset >> PAGE_SHIFT); 1454 } 1455 #endif /* __ARCH_WANT_SYS_OLD_MMAP */ 1456 1457 static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops) 1458 { 1459 return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite); 1460 } 1461 1462 static bool vma_is_shared_writable(struct vm_area_struct *vma) 1463 { 1464 return (vma->vm_flags & (VM_WRITE | VM_SHARED)) == 1465 (VM_WRITE | VM_SHARED); 1466 } 1467 1468 static bool vma_fs_can_writeback(struct vm_area_struct *vma) 1469 { 1470 /* No managed pages to writeback. */ 1471 if (vma->vm_flags & VM_PFNMAP) 1472 return false; 1473 1474 return vma->vm_file && vma->vm_file->f_mapping && 1475 mapping_can_writeback(vma->vm_file->f_mapping); 1476 } 1477 1478 /* 1479 * Does this VMA require the underlying folios to have their dirty state 1480 * tracked? 1481 */ 1482 bool vma_needs_dirty_tracking(struct vm_area_struct *vma) 1483 { 1484 /* Only shared, writable VMAs require dirty tracking. */ 1485 if (!vma_is_shared_writable(vma)) 1486 return false; 1487 1488 /* Does the filesystem need to be notified? */ 1489 if (vm_ops_needs_writenotify(vma->vm_ops)) 1490 return true; 1491 1492 /* 1493 * Even if the filesystem doesn't indicate a need for writenotify, if it 1494 * can writeback, dirty tracking is still required. 1495 */ 1496 return vma_fs_can_writeback(vma); 1497 } 1498 1499 /* 1500 * Some shared mappings will want the pages marked read-only 1501 * to track write events. If so, we'll downgrade vm_page_prot 1502 * to the private version (using protection_map[] without the 1503 * VM_SHARED bit). 1504 */ 1505 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot) 1506 { 1507 /* If it was private or non-writable, the write bit is already clear */ 1508 if (!vma_is_shared_writable(vma)) 1509 return 0; 1510 1511 /* The backer wishes to know when pages are first written to? */ 1512 if (vm_ops_needs_writenotify(vma->vm_ops)) 1513 return 1; 1514 1515 /* The open routine did something to the protections that pgprot_modify 1516 * won't preserve? */ 1517 if (pgprot_val(vm_page_prot) != 1518 pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags))) 1519 return 0; 1520 1521 /* 1522 * Do we need to track softdirty? hugetlb does not support softdirty 1523 * tracking yet. 1524 */ 1525 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma)) 1526 return 1; 1527 1528 /* Do we need write faults for uffd-wp tracking? */ 1529 if (userfaultfd_wp(vma)) 1530 return 1; 1531 1532 /* Can the mapping track the dirty pages? */ 1533 return vma_fs_can_writeback(vma); 1534 } 1535 1536 /* 1537 * We account for memory if it's a private writeable mapping, 1538 * not hugepages and VM_NORESERVE wasn't set. 1539 */ 1540 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags) 1541 { 1542 /* 1543 * hugetlb has its own accounting separate from the core VM 1544 * VM_HUGETLB may not be set yet so we cannot check for that flag. 1545 */ 1546 if (file && is_file_hugepages(file)) 1547 return 0; 1548 1549 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE; 1550 } 1551 1552 /** 1553 * unmapped_area() - Find an area between the low_limit and the high_limit with 1554 * the correct alignment and offset, all from @info. Note: current->mm is used 1555 * for the search. 1556 * 1557 * @info: The unmapped area information including the range [low_limit - 1558 * high_limit), the alignment offset and mask. 1559 * 1560 * Return: A memory address or -ENOMEM. 1561 */ 1562 static unsigned long unmapped_area(struct vm_unmapped_area_info *info) 1563 { 1564 unsigned long length, gap; 1565 unsigned long low_limit, high_limit; 1566 struct vm_area_struct *tmp; 1567 1568 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0); 1569 1570 /* Adjust search length to account for worst case alignment overhead */ 1571 length = info->length + info->align_mask; 1572 if (length < info->length) 1573 return -ENOMEM; 1574 1575 low_limit = info->low_limit; 1576 if (low_limit < mmap_min_addr) 1577 low_limit = mmap_min_addr; 1578 high_limit = info->high_limit; 1579 retry: 1580 if (mas_empty_area(&mas, low_limit, high_limit - 1, length)) 1581 return -ENOMEM; 1582 1583 gap = mas.index; 1584 gap += (info->align_offset - gap) & info->align_mask; 1585 tmp = mas_next(&mas, ULONG_MAX); 1586 if (tmp && (tmp->vm_flags & VM_GROWSDOWN)) { /* Avoid prev check if possible */ 1587 if (vm_start_gap(tmp) < gap + length - 1) { 1588 low_limit = tmp->vm_end; 1589 mas_reset(&mas); 1590 goto retry; 1591 } 1592 } else { 1593 tmp = mas_prev(&mas, 0); 1594 if (tmp && vm_end_gap(tmp) > gap) { 1595 low_limit = vm_end_gap(tmp); 1596 mas_reset(&mas); 1597 goto retry; 1598 } 1599 } 1600 1601 return gap; 1602 } 1603 1604 /** 1605 * unmapped_area_topdown() - Find an area between the low_limit and the 1606 * high_limit with the correct alignment and offset at the highest available 1607 * address, all from @info. Note: current->mm is used for the search. 1608 * 1609 * @info: The unmapped area information including the range [low_limit - 1610 * high_limit), the alignment offset and mask. 1611 * 1612 * Return: A memory address or -ENOMEM. 1613 */ 1614 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info) 1615 { 1616 unsigned long length, gap, gap_end; 1617 unsigned long low_limit, high_limit; 1618 struct vm_area_struct *tmp; 1619 1620 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0); 1621 /* Adjust search length to account for worst case alignment overhead */ 1622 length = info->length + info->align_mask; 1623 if (length < info->length) 1624 return -ENOMEM; 1625 1626 low_limit = info->low_limit; 1627 if (low_limit < mmap_min_addr) 1628 low_limit = mmap_min_addr; 1629 high_limit = info->high_limit; 1630 retry: 1631 if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length)) 1632 return -ENOMEM; 1633 1634 gap = mas.last + 1 - info->length; 1635 gap -= (gap - info->align_offset) & info->align_mask; 1636 gap_end = mas.last; 1637 tmp = mas_next(&mas, ULONG_MAX); 1638 if (tmp && (tmp->vm_flags & VM_GROWSDOWN)) { /* Avoid prev check if possible */ 1639 if (vm_start_gap(tmp) <= gap_end) { 1640 high_limit = vm_start_gap(tmp); 1641 mas_reset(&mas); 1642 goto retry; 1643 } 1644 } else { 1645 tmp = mas_prev(&mas, 0); 1646 if (tmp && vm_end_gap(tmp) > gap) { 1647 high_limit = tmp->vm_start; 1648 mas_reset(&mas); 1649 goto retry; 1650 } 1651 } 1652 1653 return gap; 1654 } 1655 1656 /* 1657 * Search for an unmapped address range. 1658 * 1659 * We are looking for a range that: 1660 * - does not intersect with any VMA; 1661 * - is contained within the [low_limit, high_limit) interval; 1662 * - is at least the desired size. 1663 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask) 1664 */ 1665 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info) 1666 { 1667 unsigned long addr; 1668 1669 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN) 1670 addr = unmapped_area_topdown(info); 1671 else 1672 addr = unmapped_area(info); 1673 1674 trace_vm_unmapped_area(addr, info); 1675 return addr; 1676 } 1677 1678 /* Get an address range which is currently unmapped. 1679 * For shmat() with addr=0. 1680 * 1681 * Ugly calling convention alert: 1682 * Return value with the low bits set means error value, 1683 * ie 1684 * if (ret & ~PAGE_MASK) 1685 * error = ret; 1686 * 1687 * This function "knows" that -ENOMEM has the bits set. 1688 */ 1689 unsigned long 1690 generic_get_unmapped_area(struct file *filp, unsigned long addr, 1691 unsigned long len, unsigned long pgoff, 1692 unsigned long flags) 1693 { 1694 struct mm_struct *mm = current->mm; 1695 struct vm_area_struct *vma, *prev; 1696 struct vm_unmapped_area_info info; 1697 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); 1698 1699 if (len > mmap_end - mmap_min_addr) 1700 return -ENOMEM; 1701 1702 if (flags & MAP_FIXED) 1703 return addr; 1704 1705 if (addr) { 1706 addr = PAGE_ALIGN(addr); 1707 vma = find_vma_prev(mm, addr, &prev); 1708 if (mmap_end - len >= addr && addr >= mmap_min_addr && 1709 (!vma || addr + len <= vm_start_gap(vma)) && 1710 (!prev || addr >= vm_end_gap(prev))) 1711 return addr; 1712 } 1713 1714 info.flags = 0; 1715 info.length = len; 1716 info.low_limit = mm->mmap_base; 1717 info.high_limit = mmap_end; 1718 info.align_mask = 0; 1719 info.align_offset = 0; 1720 return vm_unmapped_area(&info); 1721 } 1722 1723 #ifndef HAVE_ARCH_UNMAPPED_AREA 1724 unsigned long 1725 arch_get_unmapped_area(struct file *filp, unsigned long addr, 1726 unsigned long len, unsigned long pgoff, 1727 unsigned long flags) 1728 { 1729 return generic_get_unmapped_area(filp, addr, len, pgoff, flags); 1730 } 1731 #endif 1732 1733 /* 1734 * This mmap-allocator allocates new areas top-down from below the 1735 * stack's low limit (the base): 1736 */ 1737 unsigned long 1738 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr, 1739 unsigned long len, unsigned long pgoff, 1740 unsigned long flags) 1741 { 1742 struct vm_area_struct *vma, *prev; 1743 struct mm_struct *mm = current->mm; 1744 struct vm_unmapped_area_info info; 1745 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); 1746 1747 /* requested length too big for entire address space */ 1748 if (len > mmap_end - mmap_min_addr) 1749 return -ENOMEM; 1750 1751 if (flags & MAP_FIXED) 1752 return addr; 1753 1754 /* requesting a specific address */ 1755 if (addr) { 1756 addr = PAGE_ALIGN(addr); 1757 vma = find_vma_prev(mm, addr, &prev); 1758 if (mmap_end - len >= addr && addr >= mmap_min_addr && 1759 (!vma || addr + len <= vm_start_gap(vma)) && 1760 (!prev || addr >= vm_end_gap(prev))) 1761 return addr; 1762 } 1763 1764 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 1765 info.length = len; 1766 info.low_limit = PAGE_SIZE; 1767 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base); 1768 info.align_mask = 0; 1769 info.align_offset = 0; 1770 addr = vm_unmapped_area(&info); 1771 1772 /* 1773 * A failed mmap() very likely causes application failure, 1774 * so fall back to the bottom-up function here. This scenario 1775 * can happen with large stack limits and large mmap() 1776 * allocations. 1777 */ 1778 if (offset_in_page(addr)) { 1779 VM_BUG_ON(addr != -ENOMEM); 1780 info.flags = 0; 1781 info.low_limit = TASK_UNMAPPED_BASE; 1782 info.high_limit = mmap_end; 1783 addr = vm_unmapped_area(&info); 1784 } 1785 1786 return addr; 1787 } 1788 1789 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN 1790 unsigned long 1791 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr, 1792 unsigned long len, unsigned long pgoff, 1793 unsigned long flags) 1794 { 1795 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags); 1796 } 1797 #endif 1798 1799 unsigned long 1800 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, 1801 unsigned long pgoff, unsigned long flags) 1802 { 1803 unsigned long (*get_area)(struct file *, unsigned long, 1804 unsigned long, unsigned long, unsigned long); 1805 1806 unsigned long error = arch_mmap_check(addr, len, flags); 1807 if (error) 1808 return error; 1809 1810 /* Careful about overflows.. */ 1811 if (len > TASK_SIZE) 1812 return -ENOMEM; 1813 1814 get_area = current->mm->get_unmapped_area; 1815 if (file) { 1816 if (file->f_op->get_unmapped_area) 1817 get_area = file->f_op->get_unmapped_area; 1818 } else if (flags & MAP_SHARED) { 1819 /* 1820 * mmap_region() will call shmem_zero_setup() to create a file, 1821 * so use shmem's get_unmapped_area in case it can be huge. 1822 * do_mmap() will clear pgoff, so match alignment. 1823 */ 1824 pgoff = 0; 1825 get_area = shmem_get_unmapped_area; 1826 } 1827 1828 addr = get_area(file, addr, len, pgoff, flags); 1829 if (IS_ERR_VALUE(addr)) 1830 return addr; 1831 1832 if (addr > TASK_SIZE - len) 1833 return -ENOMEM; 1834 if (offset_in_page(addr)) 1835 return -EINVAL; 1836 1837 error = security_mmap_addr(addr); 1838 return error ? error : addr; 1839 } 1840 1841 EXPORT_SYMBOL(get_unmapped_area); 1842 1843 /** 1844 * find_vma_intersection() - Look up the first VMA which intersects the interval 1845 * @mm: The process address space. 1846 * @start_addr: The inclusive start user address. 1847 * @end_addr: The exclusive end user address. 1848 * 1849 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes 1850 * start_addr < end_addr. 1851 */ 1852 struct vm_area_struct *find_vma_intersection(struct mm_struct *mm, 1853 unsigned long start_addr, 1854 unsigned long end_addr) 1855 { 1856 unsigned long index = start_addr; 1857 1858 mmap_assert_locked(mm); 1859 return mt_find(&mm->mm_mt, &index, end_addr - 1); 1860 } 1861 EXPORT_SYMBOL(find_vma_intersection); 1862 1863 /** 1864 * find_vma() - Find the VMA for a given address, or the next VMA. 1865 * @mm: The mm_struct to check 1866 * @addr: The address 1867 * 1868 * Returns: The VMA associated with addr, or the next VMA. 1869 * May return %NULL in the case of no VMA at addr or above. 1870 */ 1871 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr) 1872 { 1873 unsigned long index = addr; 1874 1875 mmap_assert_locked(mm); 1876 return mt_find(&mm->mm_mt, &index, ULONG_MAX); 1877 } 1878 EXPORT_SYMBOL(find_vma); 1879 1880 /** 1881 * find_vma_prev() - Find the VMA for a given address, or the next vma and 1882 * set %pprev to the previous VMA, if any. 1883 * @mm: The mm_struct to check 1884 * @addr: The address 1885 * @pprev: The pointer to set to the previous VMA 1886 * 1887 * Note that RCU lock is missing here since the external mmap_lock() is used 1888 * instead. 1889 * 1890 * Returns: The VMA associated with @addr, or the next vma. 1891 * May return %NULL in the case of no vma at addr or above. 1892 */ 1893 struct vm_area_struct * 1894 find_vma_prev(struct mm_struct *mm, unsigned long addr, 1895 struct vm_area_struct **pprev) 1896 { 1897 struct vm_area_struct *vma; 1898 MA_STATE(mas, &mm->mm_mt, addr, addr); 1899 1900 vma = mas_walk(&mas); 1901 *pprev = mas_prev(&mas, 0); 1902 if (!vma) 1903 vma = mas_next(&mas, ULONG_MAX); 1904 return vma; 1905 } 1906 1907 /* 1908 * Verify that the stack growth is acceptable and 1909 * update accounting. This is shared with both the 1910 * grow-up and grow-down cases. 1911 */ 1912 static int acct_stack_growth(struct vm_area_struct *vma, 1913 unsigned long size, unsigned long grow) 1914 { 1915 struct mm_struct *mm = vma->vm_mm; 1916 unsigned long new_start; 1917 1918 /* address space limit tests */ 1919 if (!may_expand_vm(mm, vma->vm_flags, grow)) 1920 return -ENOMEM; 1921 1922 /* Stack limit test */ 1923 if (size > rlimit(RLIMIT_STACK)) 1924 return -ENOMEM; 1925 1926 /* mlock limit tests */ 1927 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT)) 1928 return -ENOMEM; 1929 1930 /* Check to ensure the stack will not grow into a hugetlb-only region */ 1931 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start : 1932 vma->vm_end - size; 1933 if (is_hugepage_only_range(vma->vm_mm, new_start, size)) 1934 return -EFAULT; 1935 1936 /* 1937 * Overcommit.. This must be the final test, as it will 1938 * update security statistics. 1939 */ 1940 if (security_vm_enough_memory_mm(mm, grow)) 1941 return -ENOMEM; 1942 1943 return 0; 1944 } 1945 1946 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64) 1947 /* 1948 * PA-RISC uses this for its stack; IA64 for its Register Backing Store. 1949 * vma is the last one with address > vma->vm_end. Have to extend vma. 1950 */ 1951 int expand_upwards(struct vm_area_struct *vma, unsigned long address) 1952 { 1953 struct mm_struct *mm = vma->vm_mm; 1954 struct vm_area_struct *next; 1955 unsigned long gap_addr; 1956 int error = 0; 1957 MA_STATE(mas, &mm->mm_mt, 0, 0); 1958 1959 if (!(vma->vm_flags & VM_GROWSUP)) 1960 return -EFAULT; 1961 1962 /* Guard against exceeding limits of the address space. */ 1963 address &= PAGE_MASK; 1964 if (address >= (TASK_SIZE & PAGE_MASK)) 1965 return -ENOMEM; 1966 address += PAGE_SIZE; 1967 1968 /* Enforce stack_guard_gap */ 1969 gap_addr = address + stack_guard_gap; 1970 1971 /* Guard against overflow */ 1972 if (gap_addr < address || gap_addr > TASK_SIZE) 1973 gap_addr = TASK_SIZE; 1974 1975 next = find_vma_intersection(mm, vma->vm_end, gap_addr); 1976 if (next && vma_is_accessible(next)) { 1977 if (!(next->vm_flags & VM_GROWSUP)) 1978 return -ENOMEM; 1979 /* Check that both stack segments have the same anon_vma? */ 1980 } 1981 1982 if (mas_preallocate(&mas, GFP_KERNEL)) 1983 return -ENOMEM; 1984 1985 /* We must make sure the anon_vma is allocated. */ 1986 if (unlikely(anon_vma_prepare(vma))) { 1987 mas_destroy(&mas); 1988 return -ENOMEM; 1989 } 1990 1991 /* 1992 * vma->vm_start/vm_end cannot change under us because the caller 1993 * is required to hold the mmap_lock in read mode. We need the 1994 * anon_vma lock to serialize against concurrent expand_stacks. 1995 */ 1996 anon_vma_lock_write(vma->anon_vma); 1997 1998 /* Somebody else might have raced and expanded it already */ 1999 if (address > vma->vm_end) { 2000 unsigned long size, grow; 2001 2002 size = address - vma->vm_start; 2003 grow = (address - vma->vm_end) >> PAGE_SHIFT; 2004 2005 error = -ENOMEM; 2006 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) { 2007 error = acct_stack_growth(vma, size, grow); 2008 if (!error) { 2009 /* 2010 * We only hold a shared mmap_lock lock here, so 2011 * we need to protect against concurrent vma 2012 * expansions. anon_vma_lock_write() doesn't 2013 * help here, as we don't guarantee that all 2014 * growable vmas in a mm share the same root 2015 * anon vma. So, we reuse mm->page_table_lock 2016 * to guard against concurrent vma expansions. 2017 */ 2018 spin_lock(&mm->page_table_lock); 2019 if (vma->vm_flags & VM_LOCKED) 2020 mm->locked_vm += grow; 2021 vm_stat_account(mm, vma->vm_flags, grow); 2022 anon_vma_interval_tree_pre_update_vma(vma); 2023 vma->vm_end = address; 2024 /* Overwrite old entry in mtree. */ 2025 mas_set_range(&mas, vma->vm_start, address - 1); 2026 mas_store_prealloc(&mas, vma); 2027 anon_vma_interval_tree_post_update_vma(vma); 2028 spin_unlock(&mm->page_table_lock); 2029 2030 perf_event_mmap(vma); 2031 } 2032 } 2033 } 2034 anon_vma_unlock_write(vma->anon_vma); 2035 khugepaged_enter_vma(vma, vma->vm_flags); 2036 mas_destroy(&mas); 2037 return error; 2038 } 2039 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */ 2040 2041 /* 2042 * vma is the first one with address < vma->vm_start. Have to extend vma. 2043 */ 2044 int expand_downwards(struct vm_area_struct *vma, unsigned long address) 2045 { 2046 struct mm_struct *mm = vma->vm_mm; 2047 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start); 2048 struct vm_area_struct *prev; 2049 int error = 0; 2050 2051 address &= PAGE_MASK; 2052 if (address < mmap_min_addr) 2053 return -EPERM; 2054 2055 /* Enforce stack_guard_gap */ 2056 prev = mas_prev(&mas, 0); 2057 /* Check that both stack segments have the same anon_vma? */ 2058 if (prev && !(prev->vm_flags & VM_GROWSDOWN) && 2059 vma_is_accessible(prev)) { 2060 if (address - prev->vm_end < stack_guard_gap) 2061 return -ENOMEM; 2062 } 2063 2064 if (mas_preallocate(&mas, GFP_KERNEL)) 2065 return -ENOMEM; 2066 2067 /* We must make sure the anon_vma is allocated. */ 2068 if (unlikely(anon_vma_prepare(vma))) { 2069 mas_destroy(&mas); 2070 return -ENOMEM; 2071 } 2072 2073 /* 2074 * vma->vm_start/vm_end cannot change under us because the caller 2075 * is required to hold the mmap_lock in read mode. We need the 2076 * anon_vma lock to serialize against concurrent expand_stacks. 2077 */ 2078 anon_vma_lock_write(vma->anon_vma); 2079 2080 /* Somebody else might have raced and expanded it already */ 2081 if (address < vma->vm_start) { 2082 unsigned long size, grow; 2083 2084 size = vma->vm_end - address; 2085 grow = (vma->vm_start - address) >> PAGE_SHIFT; 2086 2087 error = -ENOMEM; 2088 if (grow <= vma->vm_pgoff) { 2089 error = acct_stack_growth(vma, size, grow); 2090 if (!error) { 2091 /* 2092 * We only hold a shared mmap_lock lock here, so 2093 * we need to protect against concurrent vma 2094 * expansions. anon_vma_lock_write() doesn't 2095 * help here, as we don't guarantee that all 2096 * growable vmas in a mm share the same root 2097 * anon vma. So, we reuse mm->page_table_lock 2098 * to guard against concurrent vma expansions. 2099 */ 2100 spin_lock(&mm->page_table_lock); 2101 if (vma->vm_flags & VM_LOCKED) 2102 mm->locked_vm += grow; 2103 vm_stat_account(mm, vma->vm_flags, grow); 2104 anon_vma_interval_tree_pre_update_vma(vma); 2105 vma->vm_start = address; 2106 vma->vm_pgoff -= grow; 2107 /* Overwrite old entry in mtree. */ 2108 mas_set_range(&mas, address, vma->vm_end - 1); 2109 mas_store_prealloc(&mas, vma); 2110 anon_vma_interval_tree_post_update_vma(vma); 2111 spin_unlock(&mm->page_table_lock); 2112 2113 perf_event_mmap(vma); 2114 } 2115 } 2116 } 2117 anon_vma_unlock_write(vma->anon_vma); 2118 khugepaged_enter_vma(vma, vma->vm_flags); 2119 mas_destroy(&mas); 2120 return error; 2121 } 2122 2123 /* enforced gap between the expanding stack and other mappings. */ 2124 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT; 2125 2126 static int __init cmdline_parse_stack_guard_gap(char *p) 2127 { 2128 unsigned long val; 2129 char *endptr; 2130 2131 val = simple_strtoul(p, &endptr, 10); 2132 if (!*endptr) 2133 stack_guard_gap = val << PAGE_SHIFT; 2134 2135 return 1; 2136 } 2137 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap); 2138 2139 #ifdef CONFIG_STACK_GROWSUP 2140 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2141 { 2142 return expand_upwards(vma, address); 2143 } 2144 2145 struct vm_area_struct * 2146 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2147 { 2148 struct vm_area_struct *vma, *prev; 2149 2150 addr &= PAGE_MASK; 2151 vma = find_vma_prev(mm, addr, &prev); 2152 if (vma && (vma->vm_start <= addr)) 2153 return vma; 2154 if (!prev || expand_stack(prev, addr)) 2155 return NULL; 2156 if (prev->vm_flags & VM_LOCKED) 2157 populate_vma_page_range(prev, addr, prev->vm_end, NULL); 2158 return prev; 2159 } 2160 #else 2161 int expand_stack(struct vm_area_struct *vma, unsigned long address) 2162 { 2163 return expand_downwards(vma, address); 2164 } 2165 2166 struct vm_area_struct * 2167 find_extend_vma(struct mm_struct *mm, unsigned long addr) 2168 { 2169 struct vm_area_struct *vma; 2170 unsigned long start; 2171 2172 addr &= PAGE_MASK; 2173 vma = find_vma(mm, addr); 2174 if (!vma) 2175 return NULL; 2176 if (vma->vm_start <= addr) 2177 return vma; 2178 if (!(vma->vm_flags & VM_GROWSDOWN)) 2179 return NULL; 2180 start = vma->vm_start; 2181 if (expand_stack(vma, addr)) 2182 return NULL; 2183 if (vma->vm_flags & VM_LOCKED) 2184 populate_vma_page_range(vma, addr, start, NULL); 2185 return vma; 2186 } 2187 #endif 2188 2189 EXPORT_SYMBOL_GPL(find_extend_vma); 2190 2191 /* 2192 * Ok - we have the memory areas we should free on a maple tree so release them, 2193 * and do the vma updates. 2194 * 2195 * Called with the mm semaphore held. 2196 */ 2197 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas) 2198 { 2199 unsigned long nr_accounted = 0; 2200 struct vm_area_struct *vma; 2201 2202 /* Update high watermark before we lower total_vm */ 2203 update_hiwater_vm(mm); 2204 mas_for_each(mas, vma, ULONG_MAX) { 2205 long nrpages = vma_pages(vma); 2206 2207 if (vma->vm_flags & VM_ACCOUNT) 2208 nr_accounted += nrpages; 2209 vm_stat_account(mm, vma->vm_flags, -nrpages); 2210 remove_vma(vma, false); 2211 } 2212 vm_unacct_memory(nr_accounted); 2213 validate_mm(mm); 2214 } 2215 2216 /* 2217 * Get rid of page table information in the indicated region. 2218 * 2219 * Called with the mm semaphore held. 2220 */ 2221 static void unmap_region(struct mm_struct *mm, struct maple_tree *mt, 2222 struct vm_area_struct *vma, struct vm_area_struct *prev, 2223 struct vm_area_struct *next, 2224 unsigned long start, unsigned long end, bool mm_wr_locked) 2225 { 2226 struct mmu_gather tlb; 2227 2228 lru_add_drain(); 2229 tlb_gather_mmu(&tlb, mm); 2230 update_hiwater_rss(mm); 2231 unmap_vmas(&tlb, mt, vma, start, end, mm_wr_locked); 2232 free_pgtables(&tlb, mt, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS, 2233 next ? next->vm_start : USER_PGTABLES_CEILING, 2234 mm_wr_locked); 2235 tlb_finish_mmu(&tlb); 2236 } 2237 2238 /* 2239 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it 2240 * has already been checked or doesn't make sense to fail. 2241 * VMA Iterator will point to the end VMA. 2242 */ 2243 int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma, 2244 unsigned long addr, int new_below) 2245 { 2246 struct vma_prepare vp; 2247 struct vm_area_struct *new; 2248 int err; 2249 2250 validate_mm(vma->vm_mm); 2251 2252 WARN_ON(vma->vm_start >= addr); 2253 WARN_ON(vma->vm_end <= addr); 2254 2255 if (vma->vm_ops && vma->vm_ops->may_split) { 2256 err = vma->vm_ops->may_split(vma, addr); 2257 if (err) 2258 return err; 2259 } 2260 2261 new = vm_area_dup(vma); 2262 if (!new) 2263 return -ENOMEM; 2264 2265 err = -ENOMEM; 2266 if (vma_iter_prealloc(vmi)) 2267 goto out_free_vma; 2268 2269 if (new_below) { 2270 new->vm_end = addr; 2271 } else { 2272 new->vm_start = addr; 2273 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT); 2274 } 2275 2276 err = vma_dup_policy(vma, new); 2277 if (err) 2278 goto out_free_vmi; 2279 2280 err = anon_vma_clone(new, vma); 2281 if (err) 2282 goto out_free_mpol; 2283 2284 if (new->vm_file) 2285 get_file(new->vm_file); 2286 2287 if (new->vm_ops && new->vm_ops->open) 2288 new->vm_ops->open(new); 2289 2290 init_vma_prep(&vp, vma); 2291 vp.insert = new; 2292 vma_prepare(&vp); 2293 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0); 2294 2295 if (new_below) { 2296 vma->vm_start = addr; 2297 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT; 2298 } else { 2299 vma->vm_end = addr; 2300 } 2301 2302 /* vma_complete stores the new vma */ 2303 vma_complete(&vp, vmi, vma->vm_mm); 2304 2305 /* Success. */ 2306 if (new_below) 2307 vma_next(vmi); 2308 validate_mm(vma->vm_mm); 2309 return 0; 2310 2311 out_free_mpol: 2312 mpol_put(vma_policy(new)); 2313 out_free_vmi: 2314 vma_iter_free(vmi); 2315 out_free_vma: 2316 vm_area_free(new); 2317 validate_mm(vma->vm_mm); 2318 return err; 2319 } 2320 2321 /* 2322 * Split a vma into two pieces at address 'addr', a new vma is allocated 2323 * either for the first part or the tail. 2324 */ 2325 int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma, 2326 unsigned long addr, int new_below) 2327 { 2328 if (vma->vm_mm->map_count >= sysctl_max_map_count) 2329 return -ENOMEM; 2330 2331 return __split_vma(vmi, vma, addr, new_below); 2332 } 2333 2334 /* 2335 * do_vmi_align_munmap() - munmap the aligned region from @start to @end. 2336 * @vmi: The vma iterator 2337 * @vma: The starting vm_area_struct 2338 * @mm: The mm_struct 2339 * @start: The aligned start address to munmap. 2340 * @end: The aligned end address to munmap. 2341 * @uf: The userfaultfd list_head 2342 * @downgrade: Set to true to attempt a write downgrade of the mmap_lock 2343 * 2344 * If @downgrade is true, check return code for potential release of the lock. 2345 */ 2346 static int 2347 do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma, 2348 struct mm_struct *mm, unsigned long start, 2349 unsigned long end, struct list_head *uf, bool downgrade) 2350 { 2351 struct vm_area_struct *prev, *next = NULL; 2352 struct maple_tree mt_detach; 2353 int count = 0; 2354 int error = -ENOMEM; 2355 unsigned long locked_vm = 0; 2356 MA_STATE(mas_detach, &mt_detach, 0, 0); 2357 mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK); 2358 mt_set_external_lock(&mt_detach, &mm->mmap_lock); 2359 2360 /* 2361 * If we need to split any vma, do it now to save pain later. 2362 * 2363 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially 2364 * unmapped vm_area_struct will remain in use: so lower split_vma 2365 * places tmp vma above, and higher split_vma places tmp vma below. 2366 */ 2367 2368 /* Does it split the first one? */ 2369 if (start > vma->vm_start) { 2370 2371 /* 2372 * Make sure that map_count on return from munmap() will 2373 * not exceed its limit; but let map_count go just above 2374 * its limit temporarily, to help free resources as expected. 2375 */ 2376 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count) 2377 goto map_count_exceeded; 2378 2379 error = __split_vma(vmi, vma, start, 0); 2380 if (error) 2381 goto start_split_failed; 2382 2383 vma = vma_iter_load(vmi); 2384 } 2385 2386 prev = vma_prev(vmi); 2387 if (unlikely((!prev))) 2388 vma_iter_set(vmi, start); 2389 2390 /* 2391 * Detach a range of VMAs from the mm. Using next as a temp variable as 2392 * it is always overwritten. 2393 */ 2394 for_each_vma_range(*vmi, next, end) { 2395 /* Does it split the end? */ 2396 if (next->vm_end > end) { 2397 error = __split_vma(vmi, next, end, 0); 2398 if (error) 2399 goto end_split_failed; 2400 } 2401 vma_start_write(next); 2402 mas_set_range(&mas_detach, next->vm_start, next->vm_end - 1); 2403 if (mas_store_gfp(&mas_detach, next, GFP_KERNEL)) 2404 goto munmap_gather_failed; 2405 vma_mark_detached(next, true); 2406 if (next->vm_flags & VM_LOCKED) 2407 locked_vm += vma_pages(next); 2408 2409 count++; 2410 if (unlikely(uf)) { 2411 /* 2412 * If userfaultfd_unmap_prep returns an error the vmas 2413 * will remain split, but userland will get a 2414 * highly unexpected error anyway. This is no 2415 * different than the case where the first of the two 2416 * __split_vma fails, but we don't undo the first 2417 * split, despite we could. This is unlikely enough 2418 * failure that it's not worth optimizing it for. 2419 */ 2420 error = userfaultfd_unmap_prep(next, start, end, uf); 2421 2422 if (error) 2423 goto userfaultfd_error; 2424 } 2425 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE 2426 BUG_ON(next->vm_start < start); 2427 BUG_ON(next->vm_start > end); 2428 #endif 2429 } 2430 2431 if (vma_iter_end(vmi) > end) 2432 next = vma_iter_load(vmi); 2433 2434 if (!next) 2435 next = vma_next(vmi); 2436 2437 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE) 2438 /* Make sure no VMAs are about to be lost. */ 2439 { 2440 MA_STATE(test, &mt_detach, start, end - 1); 2441 struct vm_area_struct *vma_mas, *vma_test; 2442 int test_count = 0; 2443 2444 vma_iter_set(vmi, start); 2445 rcu_read_lock(); 2446 vma_test = mas_find(&test, end - 1); 2447 for_each_vma_range(*vmi, vma_mas, end) { 2448 BUG_ON(vma_mas != vma_test); 2449 test_count++; 2450 vma_test = mas_next(&test, end - 1); 2451 } 2452 rcu_read_unlock(); 2453 BUG_ON(count != test_count); 2454 } 2455 #endif 2456 /* Point of no return */ 2457 error = -ENOMEM; 2458 vma_iter_set(vmi, start); 2459 if (vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL)) 2460 goto clear_tree_failed; 2461 2462 mm->locked_vm -= locked_vm; 2463 mm->map_count -= count; 2464 /* 2465 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or 2466 * VM_GROWSUP VMA. Such VMAs can change their size under 2467 * down_read(mmap_lock) and collide with the VMA we are about to unmap. 2468 */ 2469 if (downgrade) { 2470 if (next && (next->vm_flags & VM_GROWSDOWN)) 2471 downgrade = false; 2472 else if (prev && (prev->vm_flags & VM_GROWSUP)) 2473 downgrade = false; 2474 else 2475 mmap_write_downgrade(mm); 2476 } 2477 2478 /* 2479 * We can free page tables without write-locking mmap_lock because VMAs 2480 * were isolated before we downgraded mmap_lock. 2481 */ 2482 unmap_region(mm, &mt_detach, vma, prev, next, start, end, !downgrade); 2483 /* Statistics and freeing VMAs */ 2484 mas_set(&mas_detach, start); 2485 remove_mt(mm, &mas_detach); 2486 __mt_destroy(&mt_detach); 2487 2488 2489 validate_mm(mm); 2490 return downgrade ? 1 : 0; 2491 2492 clear_tree_failed: 2493 userfaultfd_error: 2494 munmap_gather_failed: 2495 end_split_failed: 2496 mas_set(&mas_detach, 0); 2497 mas_for_each(&mas_detach, next, end) 2498 vma_mark_detached(next, false); 2499 2500 __mt_destroy(&mt_detach); 2501 start_split_failed: 2502 map_count_exceeded: 2503 return error; 2504 } 2505 2506 /* 2507 * do_vmi_munmap() - munmap a given range. 2508 * @vmi: The vma iterator 2509 * @mm: The mm_struct 2510 * @start: The start address to munmap 2511 * @len: The length of the range to munmap 2512 * @uf: The userfaultfd list_head 2513 * @downgrade: set to true if the user wants to attempt to write_downgrade the 2514 * mmap_lock 2515 * 2516 * This function takes a @mas that is either pointing to the previous VMA or set 2517 * to MA_START and sets it up to remove the mapping(s). The @len will be 2518 * aligned and any arch_unmap work will be preformed. 2519 * 2520 * Returns: -EINVAL on failure, 1 on success and unlock, 0 otherwise. 2521 */ 2522 int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm, 2523 unsigned long start, size_t len, struct list_head *uf, 2524 bool downgrade) 2525 { 2526 unsigned long end; 2527 struct vm_area_struct *vma; 2528 2529 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start) 2530 return -EINVAL; 2531 2532 end = start + PAGE_ALIGN(len); 2533 if (end == start) 2534 return -EINVAL; 2535 2536 /* arch_unmap() might do unmaps itself. */ 2537 arch_unmap(mm, start, end); 2538 2539 /* Find the first overlapping VMA */ 2540 vma = vma_find(vmi, end); 2541 if (!vma) 2542 return 0; 2543 2544 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, downgrade); 2545 } 2546 2547 /* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls. 2548 * @mm: The mm_struct 2549 * @start: The start address to munmap 2550 * @len: The length to be munmapped. 2551 * @uf: The userfaultfd list_head 2552 */ 2553 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len, 2554 struct list_head *uf) 2555 { 2556 VMA_ITERATOR(vmi, mm, start); 2557 2558 return do_vmi_munmap(&vmi, mm, start, len, uf, false); 2559 } 2560 2561 unsigned long mmap_region(struct file *file, unsigned long addr, 2562 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff, 2563 struct list_head *uf) 2564 { 2565 struct mm_struct *mm = current->mm; 2566 struct vm_area_struct *vma = NULL; 2567 struct vm_area_struct *next, *prev, *merge; 2568 pgoff_t pglen = len >> PAGE_SHIFT; 2569 unsigned long charged = 0; 2570 unsigned long end = addr + len; 2571 unsigned long merge_start = addr, merge_end = end; 2572 pgoff_t vm_pgoff; 2573 int error; 2574 VMA_ITERATOR(vmi, mm, addr); 2575 2576 /* Check against address space limit. */ 2577 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) { 2578 unsigned long nr_pages; 2579 2580 /* 2581 * MAP_FIXED may remove pages of mappings that intersects with 2582 * requested mapping. Account for the pages it would unmap. 2583 */ 2584 nr_pages = count_vma_pages_range(mm, addr, end); 2585 2586 if (!may_expand_vm(mm, vm_flags, 2587 (len >> PAGE_SHIFT) - nr_pages)) 2588 return -ENOMEM; 2589 } 2590 2591 /* Unmap any existing mapping in the area */ 2592 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false)) 2593 return -ENOMEM; 2594 2595 /* 2596 * Private writable mapping: check memory availability 2597 */ 2598 if (accountable_mapping(file, vm_flags)) { 2599 charged = len >> PAGE_SHIFT; 2600 if (security_vm_enough_memory_mm(mm, charged)) 2601 return -ENOMEM; 2602 vm_flags |= VM_ACCOUNT; 2603 } 2604 2605 next = vma_next(&vmi); 2606 prev = vma_prev(&vmi); 2607 if (vm_flags & VM_SPECIAL) 2608 goto cannot_expand; 2609 2610 /* Attempt to expand an old mapping */ 2611 /* Check next */ 2612 if (next && next->vm_start == end && !vma_policy(next) && 2613 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen, 2614 NULL_VM_UFFD_CTX, NULL)) { 2615 merge_end = next->vm_end; 2616 vma = next; 2617 vm_pgoff = next->vm_pgoff - pglen; 2618 } 2619 2620 /* Check prev */ 2621 if (prev && prev->vm_end == addr && !vma_policy(prev) && 2622 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file, 2623 pgoff, vma->vm_userfaultfd_ctx, NULL) : 2624 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff, 2625 NULL_VM_UFFD_CTX, NULL))) { 2626 merge_start = prev->vm_start; 2627 vma = prev; 2628 vm_pgoff = prev->vm_pgoff; 2629 } 2630 2631 2632 /* Actually expand, if possible */ 2633 if (vma && 2634 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) { 2635 khugepaged_enter_vma(vma, vm_flags); 2636 goto expanded; 2637 } 2638 2639 cannot_expand: 2640 if (prev) 2641 vma_iter_next_range(&vmi); 2642 2643 /* 2644 * Determine the object being mapped and call the appropriate 2645 * specific mapper. the address has already been validated, but 2646 * not unmapped, but the maps are removed from the list. 2647 */ 2648 vma = vm_area_alloc(mm); 2649 if (!vma) { 2650 error = -ENOMEM; 2651 goto unacct_error; 2652 } 2653 2654 vma_iter_set(&vmi, addr); 2655 vma->vm_start = addr; 2656 vma->vm_end = end; 2657 vm_flags_init(vma, vm_flags); 2658 vma->vm_page_prot = vm_get_page_prot(vm_flags); 2659 vma->vm_pgoff = pgoff; 2660 2661 if (file) { 2662 if (vm_flags & VM_SHARED) { 2663 error = mapping_map_writable(file->f_mapping); 2664 if (error) 2665 goto free_vma; 2666 } 2667 2668 vma->vm_file = get_file(file); 2669 error = call_mmap(file, vma); 2670 if (error) 2671 goto unmap_and_free_vma; 2672 2673 /* 2674 * Expansion is handled above, merging is handled below. 2675 * Drivers should not alter the address of the VMA. 2676 */ 2677 error = -EINVAL; 2678 if (WARN_ON((addr != vma->vm_start))) 2679 goto close_and_free_vma; 2680 2681 vma_iter_set(&vmi, addr); 2682 /* 2683 * If vm_flags changed after call_mmap(), we should try merge 2684 * vma again as we may succeed this time. 2685 */ 2686 if (unlikely(vm_flags != vma->vm_flags && prev)) { 2687 merge = vma_merge(&vmi, mm, prev, vma->vm_start, 2688 vma->vm_end, vma->vm_flags, NULL, 2689 vma->vm_file, vma->vm_pgoff, NULL, 2690 NULL_VM_UFFD_CTX, NULL); 2691 if (merge) { 2692 /* 2693 * ->mmap() can change vma->vm_file and fput 2694 * the original file. So fput the vma->vm_file 2695 * here or we would add an extra fput for file 2696 * and cause general protection fault 2697 * ultimately. 2698 */ 2699 fput(vma->vm_file); 2700 vm_area_free(vma); 2701 vma = merge; 2702 /* Update vm_flags to pick up the change. */ 2703 vm_flags = vma->vm_flags; 2704 goto unmap_writable; 2705 } 2706 } 2707 2708 vm_flags = vma->vm_flags; 2709 } else if (vm_flags & VM_SHARED) { 2710 error = shmem_zero_setup(vma); 2711 if (error) 2712 goto free_vma; 2713 } else { 2714 vma_set_anonymous(vma); 2715 } 2716 2717 if (map_deny_write_exec(vma, vma->vm_flags)) { 2718 error = -EACCES; 2719 goto close_and_free_vma; 2720 } 2721 2722 /* Allow architectures to sanity-check the vm_flags */ 2723 error = -EINVAL; 2724 if (!arch_validate_flags(vma->vm_flags)) 2725 goto close_and_free_vma; 2726 2727 error = -ENOMEM; 2728 if (vma_iter_prealloc(&vmi)) 2729 goto close_and_free_vma; 2730 2731 if (vma->vm_file) 2732 i_mmap_lock_write(vma->vm_file->f_mapping); 2733 2734 vma_iter_store(&vmi, vma); 2735 mm->map_count++; 2736 if (vma->vm_file) { 2737 if (vma->vm_flags & VM_SHARED) 2738 mapping_allow_writable(vma->vm_file->f_mapping); 2739 2740 flush_dcache_mmap_lock(vma->vm_file->f_mapping); 2741 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap); 2742 flush_dcache_mmap_unlock(vma->vm_file->f_mapping); 2743 i_mmap_unlock_write(vma->vm_file->f_mapping); 2744 } 2745 2746 /* 2747 * vma_merge() calls khugepaged_enter_vma() either, the below 2748 * call covers the non-merge case. 2749 */ 2750 khugepaged_enter_vma(vma, vma->vm_flags); 2751 2752 /* Once vma denies write, undo our temporary denial count */ 2753 unmap_writable: 2754 if (file && vm_flags & VM_SHARED) 2755 mapping_unmap_writable(file->f_mapping); 2756 file = vma->vm_file; 2757 ksm_add_vma(vma); 2758 expanded: 2759 perf_event_mmap(vma); 2760 2761 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT); 2762 if (vm_flags & VM_LOCKED) { 2763 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) || 2764 is_vm_hugetlb_page(vma) || 2765 vma == get_gate_vma(current->mm)) 2766 vm_flags_clear(vma, VM_LOCKED_MASK); 2767 else 2768 mm->locked_vm += (len >> PAGE_SHIFT); 2769 } 2770 2771 if (file) 2772 uprobe_mmap(vma); 2773 2774 /* 2775 * New (or expanded) vma always get soft dirty status. 2776 * Otherwise user-space soft-dirty page tracker won't 2777 * be able to distinguish situation when vma area unmapped, 2778 * then new mapped in-place (which must be aimed as 2779 * a completely new data area). 2780 */ 2781 vm_flags_set(vma, VM_SOFTDIRTY); 2782 2783 vma_set_page_prot(vma); 2784 2785 validate_mm(mm); 2786 return addr; 2787 2788 close_and_free_vma: 2789 if (file && vma->vm_ops && vma->vm_ops->close) 2790 vma->vm_ops->close(vma); 2791 2792 if (file || vma->vm_file) { 2793 unmap_and_free_vma: 2794 fput(vma->vm_file); 2795 vma->vm_file = NULL; 2796 2797 /* Undo any partial mapping done by a device driver. */ 2798 unmap_region(mm, &mm->mm_mt, vma, prev, next, vma->vm_start, 2799 vma->vm_end, true); 2800 } 2801 if (file && (vm_flags & VM_SHARED)) 2802 mapping_unmap_writable(file->f_mapping); 2803 free_vma: 2804 vm_area_free(vma); 2805 unacct_error: 2806 if (charged) 2807 vm_unacct_memory(charged); 2808 validate_mm(mm); 2809 return error; 2810 } 2811 2812 static int __vm_munmap(unsigned long start, size_t len, bool downgrade) 2813 { 2814 int ret; 2815 struct mm_struct *mm = current->mm; 2816 LIST_HEAD(uf); 2817 VMA_ITERATOR(vmi, mm, start); 2818 2819 if (mmap_write_lock_killable(mm)) 2820 return -EINTR; 2821 2822 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, downgrade); 2823 /* 2824 * Returning 1 indicates mmap_lock is downgraded. 2825 * But 1 is not legal return value of vm_munmap() and munmap(), reset 2826 * it to 0 before return. 2827 */ 2828 if (ret == 1) { 2829 mmap_read_unlock(mm); 2830 ret = 0; 2831 } else 2832 mmap_write_unlock(mm); 2833 2834 userfaultfd_unmap_complete(mm, &uf); 2835 return ret; 2836 } 2837 2838 int vm_munmap(unsigned long start, size_t len) 2839 { 2840 return __vm_munmap(start, len, false); 2841 } 2842 EXPORT_SYMBOL(vm_munmap); 2843 2844 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len) 2845 { 2846 addr = untagged_addr(addr); 2847 return __vm_munmap(addr, len, true); 2848 } 2849 2850 2851 /* 2852 * Emulation of deprecated remap_file_pages() syscall. 2853 */ 2854 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size, 2855 unsigned long, prot, unsigned long, pgoff, unsigned long, flags) 2856 { 2857 2858 struct mm_struct *mm = current->mm; 2859 struct vm_area_struct *vma; 2860 unsigned long populate = 0; 2861 unsigned long ret = -EINVAL; 2862 struct file *file; 2863 2864 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n", 2865 current->comm, current->pid); 2866 2867 if (prot) 2868 return ret; 2869 start = start & PAGE_MASK; 2870 size = size & PAGE_MASK; 2871 2872 if (start + size <= start) 2873 return ret; 2874 2875 /* Does pgoff wrap? */ 2876 if (pgoff + (size >> PAGE_SHIFT) < pgoff) 2877 return ret; 2878 2879 if (mmap_write_lock_killable(mm)) 2880 return -EINTR; 2881 2882 vma = vma_lookup(mm, start); 2883 2884 if (!vma || !(vma->vm_flags & VM_SHARED)) 2885 goto out; 2886 2887 if (start + size > vma->vm_end) { 2888 VMA_ITERATOR(vmi, mm, vma->vm_end); 2889 struct vm_area_struct *next, *prev = vma; 2890 2891 for_each_vma_range(vmi, next, start + size) { 2892 /* hole between vmas ? */ 2893 if (next->vm_start != prev->vm_end) 2894 goto out; 2895 2896 if (next->vm_file != vma->vm_file) 2897 goto out; 2898 2899 if (next->vm_flags != vma->vm_flags) 2900 goto out; 2901 2902 if (start + size <= next->vm_end) 2903 break; 2904 2905 prev = next; 2906 } 2907 2908 if (!next) 2909 goto out; 2910 } 2911 2912 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0; 2913 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0; 2914 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0; 2915 2916 flags &= MAP_NONBLOCK; 2917 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE; 2918 if (vma->vm_flags & VM_LOCKED) 2919 flags |= MAP_LOCKED; 2920 2921 file = get_file(vma->vm_file); 2922 ret = do_mmap(vma->vm_file, start, size, 2923 prot, flags, pgoff, &populate, NULL); 2924 fput(file); 2925 out: 2926 mmap_write_unlock(mm); 2927 if (populate) 2928 mm_populate(ret, populate); 2929 if (!IS_ERR_VALUE(ret)) 2930 ret = 0; 2931 return ret; 2932 } 2933 2934 /* 2935 * do_vma_munmap() - Unmap a full or partial vma. 2936 * @vmi: The vma iterator pointing at the vma 2937 * @vma: The first vma to be munmapped 2938 * @start: the start of the address to unmap 2939 * @end: The end of the address to unmap 2940 * @uf: The userfaultfd list_head 2941 * @downgrade: Attempt to downgrade or not 2942 * 2943 * Returns: 0 on success and not downgraded, 1 on success and downgraded. 2944 * unmaps a VMA mapping when the vma iterator is already in position. 2945 * Does not handle alignment. 2946 */ 2947 int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma, 2948 unsigned long start, unsigned long end, 2949 struct list_head *uf, bool downgrade) 2950 { 2951 struct mm_struct *mm = vma->vm_mm; 2952 int ret; 2953 2954 arch_unmap(mm, start, end); 2955 ret = do_vmi_align_munmap(vmi, vma, mm, start, end, uf, downgrade); 2956 validate_mm(mm); 2957 return ret; 2958 } 2959 2960 /* 2961 * do_brk_flags() - Increase the brk vma if the flags match. 2962 * @vmi: The vma iterator 2963 * @addr: The start address 2964 * @len: The length of the increase 2965 * @vma: The vma, 2966 * @flags: The VMA Flags 2967 * 2968 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags 2969 * do not match then create a new anonymous VMA. Eventually we may be able to 2970 * do some brk-specific accounting here. 2971 */ 2972 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma, 2973 unsigned long addr, unsigned long len, unsigned long flags) 2974 { 2975 struct mm_struct *mm = current->mm; 2976 struct vma_prepare vp; 2977 2978 validate_mm(mm); 2979 /* 2980 * Check against address space limits by the changed size 2981 * Note: This happens *after* clearing old mappings in some code paths. 2982 */ 2983 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags; 2984 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT)) 2985 return -ENOMEM; 2986 2987 if (mm->map_count > sysctl_max_map_count) 2988 return -ENOMEM; 2989 2990 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT)) 2991 return -ENOMEM; 2992 2993 /* 2994 * Expand the existing vma if possible; Note that singular lists do not 2995 * occur after forking, so the expand will only happen on new VMAs. 2996 */ 2997 if (vma && vma->vm_end == addr && !vma_policy(vma) && 2998 can_vma_merge_after(vma, flags, NULL, NULL, 2999 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) { 3000 if (vma_iter_prealloc(vmi)) 3001 goto unacct_fail; 3002 3003 init_vma_prep(&vp, vma); 3004 vma_prepare(&vp); 3005 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0); 3006 vma->vm_end = addr + len; 3007 vm_flags_set(vma, VM_SOFTDIRTY); 3008 vma_iter_store(vmi, vma); 3009 3010 vma_complete(&vp, vmi, mm); 3011 khugepaged_enter_vma(vma, flags); 3012 goto out; 3013 } 3014 3015 /* create a vma struct for an anonymous mapping */ 3016 vma = vm_area_alloc(mm); 3017 if (!vma) 3018 goto unacct_fail; 3019 3020 vma_set_anonymous(vma); 3021 vma->vm_start = addr; 3022 vma->vm_end = addr + len; 3023 vma->vm_pgoff = addr >> PAGE_SHIFT; 3024 vm_flags_init(vma, flags); 3025 vma->vm_page_prot = vm_get_page_prot(flags); 3026 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL)) 3027 goto mas_store_fail; 3028 3029 mm->map_count++; 3030 ksm_add_vma(vma); 3031 out: 3032 perf_event_mmap(vma); 3033 mm->total_vm += len >> PAGE_SHIFT; 3034 mm->data_vm += len >> PAGE_SHIFT; 3035 if (flags & VM_LOCKED) 3036 mm->locked_vm += (len >> PAGE_SHIFT); 3037 vm_flags_set(vma, VM_SOFTDIRTY); 3038 validate_mm(mm); 3039 return 0; 3040 3041 mas_store_fail: 3042 vm_area_free(vma); 3043 unacct_fail: 3044 vm_unacct_memory(len >> PAGE_SHIFT); 3045 return -ENOMEM; 3046 } 3047 3048 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags) 3049 { 3050 struct mm_struct *mm = current->mm; 3051 struct vm_area_struct *vma = NULL; 3052 unsigned long len; 3053 int ret; 3054 bool populate; 3055 LIST_HEAD(uf); 3056 VMA_ITERATOR(vmi, mm, addr); 3057 3058 len = PAGE_ALIGN(request); 3059 if (len < request) 3060 return -ENOMEM; 3061 if (!len) 3062 return 0; 3063 3064 if (mmap_write_lock_killable(mm)) 3065 return -EINTR; 3066 3067 /* Until we need other flags, refuse anything except VM_EXEC. */ 3068 if ((flags & (~VM_EXEC)) != 0) 3069 return -EINVAL; 3070 3071 ret = check_brk_limits(addr, len); 3072 if (ret) 3073 goto limits_failed; 3074 3075 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0); 3076 if (ret) 3077 goto munmap_failed; 3078 3079 vma = vma_prev(&vmi); 3080 ret = do_brk_flags(&vmi, vma, addr, len, flags); 3081 populate = ((mm->def_flags & VM_LOCKED) != 0); 3082 mmap_write_unlock(mm); 3083 userfaultfd_unmap_complete(mm, &uf); 3084 if (populate && !ret) 3085 mm_populate(addr, len); 3086 return ret; 3087 3088 munmap_failed: 3089 limits_failed: 3090 mmap_write_unlock(mm); 3091 return ret; 3092 } 3093 EXPORT_SYMBOL(vm_brk_flags); 3094 3095 int vm_brk(unsigned long addr, unsigned long len) 3096 { 3097 return vm_brk_flags(addr, len, 0); 3098 } 3099 EXPORT_SYMBOL(vm_brk); 3100 3101 /* Release all mmaps. */ 3102 void exit_mmap(struct mm_struct *mm) 3103 { 3104 struct mmu_gather tlb; 3105 struct vm_area_struct *vma; 3106 unsigned long nr_accounted = 0; 3107 MA_STATE(mas, &mm->mm_mt, 0, 0); 3108 int count = 0; 3109 3110 /* mm's last user has gone, and its about to be pulled down */ 3111 mmu_notifier_release(mm); 3112 3113 mmap_read_lock(mm); 3114 arch_exit_mmap(mm); 3115 3116 vma = mas_find(&mas, ULONG_MAX); 3117 if (!vma) { 3118 /* Can happen if dup_mmap() received an OOM */ 3119 mmap_read_unlock(mm); 3120 return; 3121 } 3122 3123 lru_add_drain(); 3124 flush_cache_mm(mm); 3125 tlb_gather_mmu_fullmm(&tlb, mm); 3126 /* update_hiwater_rss(mm) here? but nobody should be looking */ 3127 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */ 3128 unmap_vmas(&tlb, &mm->mm_mt, vma, 0, ULONG_MAX, false); 3129 mmap_read_unlock(mm); 3130 3131 /* 3132 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper 3133 * because the memory has been already freed. 3134 */ 3135 set_bit(MMF_OOM_SKIP, &mm->flags); 3136 mmap_write_lock(mm); 3137 mt_clear_in_rcu(&mm->mm_mt); 3138 free_pgtables(&tlb, &mm->mm_mt, vma, FIRST_USER_ADDRESS, 3139 USER_PGTABLES_CEILING, true); 3140 tlb_finish_mmu(&tlb); 3141 3142 /* 3143 * Walk the list again, actually closing and freeing it, with preemption 3144 * enabled, without holding any MM locks besides the unreachable 3145 * mmap_write_lock. 3146 */ 3147 do { 3148 if (vma->vm_flags & VM_ACCOUNT) 3149 nr_accounted += vma_pages(vma); 3150 remove_vma(vma, true); 3151 count++; 3152 cond_resched(); 3153 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL); 3154 3155 BUG_ON(count != mm->map_count); 3156 3157 trace_exit_mmap(mm); 3158 __mt_destroy(&mm->mm_mt); 3159 mmap_write_unlock(mm); 3160 vm_unacct_memory(nr_accounted); 3161 } 3162 3163 /* Insert vm structure into process list sorted by address 3164 * and into the inode's i_mmap tree. If vm_file is non-NULL 3165 * then i_mmap_rwsem is taken here. 3166 */ 3167 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma) 3168 { 3169 unsigned long charged = vma_pages(vma); 3170 3171 3172 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end)) 3173 return -ENOMEM; 3174 3175 if ((vma->vm_flags & VM_ACCOUNT) && 3176 security_vm_enough_memory_mm(mm, charged)) 3177 return -ENOMEM; 3178 3179 /* 3180 * The vm_pgoff of a purely anonymous vma should be irrelevant 3181 * until its first write fault, when page's anon_vma and index 3182 * are set. But now set the vm_pgoff it will almost certainly 3183 * end up with (unless mremap moves it elsewhere before that 3184 * first wfault), so /proc/pid/maps tells a consistent story. 3185 * 3186 * By setting it to reflect the virtual start address of the 3187 * vma, merges and splits can happen in a seamless way, just 3188 * using the existing file pgoff checks and manipulations. 3189 * Similarly in do_mmap and in do_brk_flags. 3190 */ 3191 if (vma_is_anonymous(vma)) { 3192 BUG_ON(vma->anon_vma); 3193 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; 3194 } 3195 3196 if (vma_link(mm, vma)) { 3197 vm_unacct_memory(charged); 3198 return -ENOMEM; 3199 } 3200 3201 return 0; 3202 } 3203 3204 /* 3205 * Copy the vma structure to a new location in the same mm, 3206 * prior to moving page table entries, to effect an mremap move. 3207 */ 3208 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap, 3209 unsigned long addr, unsigned long len, pgoff_t pgoff, 3210 bool *need_rmap_locks) 3211 { 3212 struct vm_area_struct *vma = *vmap; 3213 unsigned long vma_start = vma->vm_start; 3214 struct mm_struct *mm = vma->vm_mm; 3215 struct vm_area_struct *new_vma, *prev; 3216 bool faulted_in_anon_vma = true; 3217 VMA_ITERATOR(vmi, mm, addr); 3218 3219 validate_mm(mm); 3220 /* 3221 * If anonymous vma has not yet been faulted, update new pgoff 3222 * to match new location, to increase its chance of merging. 3223 */ 3224 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) { 3225 pgoff = addr >> PAGE_SHIFT; 3226 faulted_in_anon_vma = false; 3227 } 3228 3229 new_vma = find_vma_prev(mm, addr, &prev); 3230 if (new_vma && new_vma->vm_start < addr + len) 3231 return NULL; /* should never get here */ 3232 3233 new_vma = vma_merge(&vmi, mm, prev, addr, addr + len, vma->vm_flags, 3234 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 3235 vma->vm_userfaultfd_ctx, anon_vma_name(vma)); 3236 if (new_vma) { 3237 /* 3238 * Source vma may have been merged into new_vma 3239 */ 3240 if (unlikely(vma_start >= new_vma->vm_start && 3241 vma_start < new_vma->vm_end)) { 3242 /* 3243 * The only way we can get a vma_merge with 3244 * self during an mremap is if the vma hasn't 3245 * been faulted in yet and we were allowed to 3246 * reset the dst vma->vm_pgoff to the 3247 * destination address of the mremap to allow 3248 * the merge to happen. mremap must change the 3249 * vm_pgoff linearity between src and dst vmas 3250 * (in turn preventing a vma_merge) to be 3251 * safe. It is only safe to keep the vm_pgoff 3252 * linear if there are no pages mapped yet. 3253 */ 3254 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma); 3255 *vmap = vma = new_vma; 3256 } 3257 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff); 3258 } else { 3259 new_vma = vm_area_dup(vma); 3260 if (!new_vma) 3261 goto out; 3262 new_vma->vm_start = addr; 3263 new_vma->vm_end = addr + len; 3264 new_vma->vm_pgoff = pgoff; 3265 if (vma_dup_policy(vma, new_vma)) 3266 goto out_free_vma; 3267 if (anon_vma_clone(new_vma, vma)) 3268 goto out_free_mempol; 3269 if (new_vma->vm_file) 3270 get_file(new_vma->vm_file); 3271 if (new_vma->vm_ops && new_vma->vm_ops->open) 3272 new_vma->vm_ops->open(new_vma); 3273 vma_start_write(new_vma); 3274 if (vma_link(mm, new_vma)) 3275 goto out_vma_link; 3276 *need_rmap_locks = false; 3277 } 3278 validate_mm(mm); 3279 return new_vma; 3280 3281 out_vma_link: 3282 if (new_vma->vm_ops && new_vma->vm_ops->close) 3283 new_vma->vm_ops->close(new_vma); 3284 3285 if (new_vma->vm_file) 3286 fput(new_vma->vm_file); 3287 3288 unlink_anon_vmas(new_vma); 3289 out_free_mempol: 3290 mpol_put(vma_policy(new_vma)); 3291 out_free_vma: 3292 vm_area_free(new_vma); 3293 out: 3294 validate_mm(mm); 3295 return NULL; 3296 } 3297 3298 /* 3299 * Return true if the calling process may expand its vm space by the passed 3300 * number of pages 3301 */ 3302 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages) 3303 { 3304 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT) 3305 return false; 3306 3307 if (is_data_mapping(flags) && 3308 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) { 3309 /* Workaround for Valgrind */ 3310 if (rlimit(RLIMIT_DATA) == 0 && 3311 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT) 3312 return true; 3313 3314 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n", 3315 current->comm, current->pid, 3316 (mm->data_vm + npages) << PAGE_SHIFT, 3317 rlimit(RLIMIT_DATA), 3318 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data"); 3319 3320 if (!ignore_rlimit_data) 3321 return false; 3322 } 3323 3324 return true; 3325 } 3326 3327 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages) 3328 { 3329 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages); 3330 3331 if (is_exec_mapping(flags)) 3332 mm->exec_vm += npages; 3333 else if (is_stack_mapping(flags)) 3334 mm->stack_vm += npages; 3335 else if (is_data_mapping(flags)) 3336 mm->data_vm += npages; 3337 } 3338 3339 static vm_fault_t special_mapping_fault(struct vm_fault *vmf); 3340 3341 /* 3342 * Having a close hook prevents vma merging regardless of flags. 3343 */ 3344 static void special_mapping_close(struct vm_area_struct *vma) 3345 { 3346 } 3347 3348 static const char *special_mapping_name(struct vm_area_struct *vma) 3349 { 3350 return ((struct vm_special_mapping *)vma->vm_private_data)->name; 3351 } 3352 3353 static int special_mapping_mremap(struct vm_area_struct *new_vma) 3354 { 3355 struct vm_special_mapping *sm = new_vma->vm_private_data; 3356 3357 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm)) 3358 return -EFAULT; 3359 3360 if (sm->mremap) 3361 return sm->mremap(sm, new_vma); 3362 3363 return 0; 3364 } 3365 3366 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr) 3367 { 3368 /* 3369 * Forbid splitting special mappings - kernel has expectations over 3370 * the number of pages in mapping. Together with VM_DONTEXPAND 3371 * the size of vma should stay the same over the special mapping's 3372 * lifetime. 3373 */ 3374 return -EINVAL; 3375 } 3376 3377 static const struct vm_operations_struct special_mapping_vmops = { 3378 .close = special_mapping_close, 3379 .fault = special_mapping_fault, 3380 .mremap = special_mapping_mremap, 3381 .name = special_mapping_name, 3382 /* vDSO code relies that VVAR can't be accessed remotely */ 3383 .access = NULL, 3384 .may_split = special_mapping_split, 3385 }; 3386 3387 static const struct vm_operations_struct legacy_special_mapping_vmops = { 3388 .close = special_mapping_close, 3389 .fault = special_mapping_fault, 3390 }; 3391 3392 static vm_fault_t special_mapping_fault(struct vm_fault *vmf) 3393 { 3394 struct vm_area_struct *vma = vmf->vma; 3395 pgoff_t pgoff; 3396 struct page **pages; 3397 3398 if (vma->vm_ops == &legacy_special_mapping_vmops) { 3399 pages = vma->vm_private_data; 3400 } else { 3401 struct vm_special_mapping *sm = vma->vm_private_data; 3402 3403 if (sm->fault) 3404 return sm->fault(sm, vmf->vma, vmf); 3405 3406 pages = sm->pages; 3407 } 3408 3409 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages) 3410 pgoff--; 3411 3412 if (*pages) { 3413 struct page *page = *pages; 3414 get_page(page); 3415 vmf->page = page; 3416 return 0; 3417 } 3418 3419 return VM_FAULT_SIGBUS; 3420 } 3421 3422 static struct vm_area_struct *__install_special_mapping( 3423 struct mm_struct *mm, 3424 unsigned long addr, unsigned long len, 3425 unsigned long vm_flags, void *priv, 3426 const struct vm_operations_struct *ops) 3427 { 3428 int ret; 3429 struct vm_area_struct *vma; 3430 3431 validate_mm(mm); 3432 vma = vm_area_alloc(mm); 3433 if (unlikely(vma == NULL)) 3434 return ERR_PTR(-ENOMEM); 3435 3436 vma->vm_start = addr; 3437 vma->vm_end = addr + len; 3438 3439 vm_flags_init(vma, (vm_flags | mm->def_flags | 3440 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK); 3441 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 3442 3443 vma->vm_ops = ops; 3444 vma->vm_private_data = priv; 3445 3446 ret = insert_vm_struct(mm, vma); 3447 if (ret) 3448 goto out; 3449 3450 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT); 3451 3452 perf_event_mmap(vma); 3453 3454 validate_mm(mm); 3455 return vma; 3456 3457 out: 3458 vm_area_free(vma); 3459 validate_mm(mm); 3460 return ERR_PTR(ret); 3461 } 3462 3463 bool vma_is_special_mapping(const struct vm_area_struct *vma, 3464 const struct vm_special_mapping *sm) 3465 { 3466 return vma->vm_private_data == sm && 3467 (vma->vm_ops == &special_mapping_vmops || 3468 vma->vm_ops == &legacy_special_mapping_vmops); 3469 } 3470 3471 /* 3472 * Called with mm->mmap_lock held for writing. 3473 * Insert a new vma covering the given region, with the given flags. 3474 * Its pages are supplied by the given array of struct page *. 3475 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated. 3476 * The region past the last page supplied will always produce SIGBUS. 3477 * The array pointer and the pages it points to are assumed to stay alive 3478 * for as long as this mapping might exist. 3479 */ 3480 struct vm_area_struct *_install_special_mapping( 3481 struct mm_struct *mm, 3482 unsigned long addr, unsigned long len, 3483 unsigned long vm_flags, const struct vm_special_mapping *spec) 3484 { 3485 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec, 3486 &special_mapping_vmops); 3487 } 3488 3489 int install_special_mapping(struct mm_struct *mm, 3490 unsigned long addr, unsigned long len, 3491 unsigned long vm_flags, struct page **pages) 3492 { 3493 struct vm_area_struct *vma = __install_special_mapping( 3494 mm, addr, len, vm_flags, (void *)pages, 3495 &legacy_special_mapping_vmops); 3496 3497 return PTR_ERR_OR_ZERO(vma); 3498 } 3499 3500 static DEFINE_MUTEX(mm_all_locks_mutex); 3501 3502 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma) 3503 { 3504 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) { 3505 /* 3506 * The LSB of head.next can't change from under us 3507 * because we hold the mm_all_locks_mutex. 3508 */ 3509 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock); 3510 /* 3511 * We can safely modify head.next after taking the 3512 * anon_vma->root->rwsem. If some other vma in this mm shares 3513 * the same anon_vma we won't take it again. 3514 * 3515 * No need of atomic instructions here, head.next 3516 * can't change from under us thanks to the 3517 * anon_vma->root->rwsem. 3518 */ 3519 if (__test_and_set_bit(0, (unsigned long *) 3520 &anon_vma->root->rb_root.rb_root.rb_node)) 3521 BUG(); 3522 } 3523 } 3524 3525 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping) 3526 { 3527 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3528 /* 3529 * AS_MM_ALL_LOCKS can't change from under us because 3530 * we hold the mm_all_locks_mutex. 3531 * 3532 * Operations on ->flags have to be atomic because 3533 * even if AS_MM_ALL_LOCKS is stable thanks to the 3534 * mm_all_locks_mutex, there may be other cpus 3535 * changing other bitflags in parallel to us. 3536 */ 3537 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags)) 3538 BUG(); 3539 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock); 3540 } 3541 } 3542 3543 /* 3544 * This operation locks against the VM for all pte/vma/mm related 3545 * operations that could ever happen on a certain mm. This includes 3546 * vmtruncate, try_to_unmap, and all page faults. 3547 * 3548 * The caller must take the mmap_lock in write mode before calling 3549 * mm_take_all_locks(). The caller isn't allowed to release the 3550 * mmap_lock until mm_drop_all_locks() returns. 3551 * 3552 * mmap_lock in write mode is required in order to block all operations 3553 * that could modify pagetables and free pages without need of 3554 * altering the vma layout. It's also needed in write mode to avoid new 3555 * anon_vmas to be associated with existing vmas. 3556 * 3557 * A single task can't take more than one mm_take_all_locks() in a row 3558 * or it would deadlock. 3559 * 3560 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in 3561 * mapping->flags avoid to take the same lock twice, if more than one 3562 * vma in this mm is backed by the same anon_vma or address_space. 3563 * 3564 * We take locks in following order, accordingly to comment at beginning 3565 * of mm/rmap.c: 3566 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for 3567 * hugetlb mapping); 3568 * - all vmas marked locked 3569 * - all i_mmap_rwsem locks; 3570 * - all anon_vma->rwseml 3571 * 3572 * We can take all locks within these types randomly because the VM code 3573 * doesn't nest them and we protected from parallel mm_take_all_locks() by 3574 * mm_all_locks_mutex. 3575 * 3576 * mm_take_all_locks() and mm_drop_all_locks are expensive operations 3577 * that may have to take thousand of locks. 3578 * 3579 * mm_take_all_locks() can fail if it's interrupted by signals. 3580 */ 3581 int mm_take_all_locks(struct mm_struct *mm) 3582 { 3583 struct vm_area_struct *vma; 3584 struct anon_vma_chain *avc; 3585 MA_STATE(mas, &mm->mm_mt, 0, 0); 3586 3587 mmap_assert_write_locked(mm); 3588 3589 mutex_lock(&mm_all_locks_mutex); 3590 3591 mas_for_each(&mas, vma, ULONG_MAX) { 3592 if (signal_pending(current)) 3593 goto out_unlock; 3594 vma_start_write(vma); 3595 } 3596 3597 mas_set(&mas, 0); 3598 mas_for_each(&mas, vma, ULONG_MAX) { 3599 if (signal_pending(current)) 3600 goto out_unlock; 3601 if (vma->vm_file && vma->vm_file->f_mapping && 3602 is_vm_hugetlb_page(vma)) 3603 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3604 } 3605 3606 mas_set(&mas, 0); 3607 mas_for_each(&mas, vma, ULONG_MAX) { 3608 if (signal_pending(current)) 3609 goto out_unlock; 3610 if (vma->vm_file && vma->vm_file->f_mapping && 3611 !is_vm_hugetlb_page(vma)) 3612 vm_lock_mapping(mm, vma->vm_file->f_mapping); 3613 } 3614 3615 mas_set(&mas, 0); 3616 mas_for_each(&mas, vma, ULONG_MAX) { 3617 if (signal_pending(current)) 3618 goto out_unlock; 3619 if (vma->anon_vma) 3620 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3621 vm_lock_anon_vma(mm, avc->anon_vma); 3622 } 3623 3624 return 0; 3625 3626 out_unlock: 3627 mm_drop_all_locks(mm); 3628 return -EINTR; 3629 } 3630 3631 static void vm_unlock_anon_vma(struct anon_vma *anon_vma) 3632 { 3633 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) { 3634 /* 3635 * The LSB of head.next can't change to 0 from under 3636 * us because we hold the mm_all_locks_mutex. 3637 * 3638 * We must however clear the bitflag before unlocking 3639 * the vma so the users using the anon_vma->rb_root will 3640 * never see our bitflag. 3641 * 3642 * No need of atomic instructions here, head.next 3643 * can't change from under us until we release the 3644 * anon_vma->root->rwsem. 3645 */ 3646 if (!__test_and_clear_bit(0, (unsigned long *) 3647 &anon_vma->root->rb_root.rb_root.rb_node)) 3648 BUG(); 3649 anon_vma_unlock_write(anon_vma); 3650 } 3651 } 3652 3653 static void vm_unlock_mapping(struct address_space *mapping) 3654 { 3655 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) { 3656 /* 3657 * AS_MM_ALL_LOCKS can't change to 0 from under us 3658 * because we hold the mm_all_locks_mutex. 3659 */ 3660 i_mmap_unlock_write(mapping); 3661 if (!test_and_clear_bit(AS_MM_ALL_LOCKS, 3662 &mapping->flags)) 3663 BUG(); 3664 } 3665 } 3666 3667 /* 3668 * The mmap_lock cannot be released by the caller until 3669 * mm_drop_all_locks() returns. 3670 */ 3671 void mm_drop_all_locks(struct mm_struct *mm) 3672 { 3673 struct vm_area_struct *vma; 3674 struct anon_vma_chain *avc; 3675 MA_STATE(mas, &mm->mm_mt, 0, 0); 3676 3677 mmap_assert_write_locked(mm); 3678 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex)); 3679 3680 mas_for_each(&mas, vma, ULONG_MAX) { 3681 if (vma->anon_vma) 3682 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) 3683 vm_unlock_anon_vma(avc->anon_vma); 3684 if (vma->vm_file && vma->vm_file->f_mapping) 3685 vm_unlock_mapping(vma->vm_file->f_mapping); 3686 } 3687 vma_end_write_all(mm); 3688 3689 mutex_unlock(&mm_all_locks_mutex); 3690 } 3691 3692 /* 3693 * initialise the percpu counter for VM 3694 */ 3695 void __init mmap_init(void) 3696 { 3697 int ret; 3698 3699 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL); 3700 VM_BUG_ON(ret); 3701 } 3702 3703 /* 3704 * Initialise sysctl_user_reserve_kbytes. 3705 * 3706 * This is intended to prevent a user from starting a single memory hogging 3707 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER 3708 * mode. 3709 * 3710 * The default value is min(3% of free memory, 128MB) 3711 * 128MB is enough to recover with sshd/login, bash, and top/kill. 3712 */ 3713 static int init_user_reserve(void) 3714 { 3715 unsigned long free_kbytes; 3716 3717 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3718 3719 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17); 3720 return 0; 3721 } 3722 subsys_initcall(init_user_reserve); 3723 3724 /* 3725 * Initialise sysctl_admin_reserve_kbytes. 3726 * 3727 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin 3728 * to log in and kill a memory hogging process. 3729 * 3730 * Systems with more than 256MB will reserve 8MB, enough to recover 3731 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will 3732 * only reserve 3% of free pages by default. 3733 */ 3734 static int init_admin_reserve(void) 3735 { 3736 unsigned long free_kbytes; 3737 3738 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3739 3740 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13); 3741 return 0; 3742 } 3743 subsys_initcall(init_admin_reserve); 3744 3745 /* 3746 * Reinititalise user and admin reserves if memory is added or removed. 3747 * 3748 * The default user reserve max is 128MB, and the default max for the 3749 * admin reserve is 8MB. These are usually, but not always, enough to 3750 * enable recovery from a memory hogging process using login/sshd, a shell, 3751 * and tools like top. It may make sense to increase or even disable the 3752 * reserve depending on the existence of swap or variations in the recovery 3753 * tools. So, the admin may have changed them. 3754 * 3755 * If memory is added and the reserves have been eliminated or increased above 3756 * the default max, then we'll trust the admin. 3757 * 3758 * If memory is removed and there isn't enough free memory, then we 3759 * need to reset the reserves. 3760 * 3761 * Otherwise keep the reserve set by the admin. 3762 */ 3763 static int reserve_mem_notifier(struct notifier_block *nb, 3764 unsigned long action, void *data) 3765 { 3766 unsigned long tmp, free_kbytes; 3767 3768 switch (action) { 3769 case MEM_ONLINE: 3770 /* Default max is 128MB. Leave alone if modified by operator. */ 3771 tmp = sysctl_user_reserve_kbytes; 3772 if (0 < tmp && tmp < (1UL << 17)) 3773 init_user_reserve(); 3774 3775 /* Default max is 8MB. Leave alone if modified by operator. */ 3776 tmp = sysctl_admin_reserve_kbytes; 3777 if (0 < tmp && tmp < (1UL << 13)) 3778 init_admin_reserve(); 3779 3780 break; 3781 case MEM_OFFLINE: 3782 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10); 3783 3784 if (sysctl_user_reserve_kbytes > free_kbytes) { 3785 init_user_reserve(); 3786 pr_info("vm.user_reserve_kbytes reset to %lu\n", 3787 sysctl_user_reserve_kbytes); 3788 } 3789 3790 if (sysctl_admin_reserve_kbytes > free_kbytes) { 3791 init_admin_reserve(); 3792 pr_info("vm.admin_reserve_kbytes reset to %lu\n", 3793 sysctl_admin_reserve_kbytes); 3794 } 3795 break; 3796 default: 3797 break; 3798 } 3799 return NOTIFY_OK; 3800 } 3801 3802 static int __meminit init_reserve_notifier(void) 3803 { 3804 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI)) 3805 pr_err("Failed registering memory add/remove notifier for admin reserve\n"); 3806 3807 return 0; 3808 } 3809 subsys_initcall(init_reserve_notifier); 3810