| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * mm/mremap.c |
| 4 | * |
| 5 | * (C) Copyright 1996 Linus Torvalds |
| 6 | * |
| 7 | * Address space accounting code <alan@lxorguk.ukuu.org.uk> |
| 8 | * (C) Copyright 2002 Red Hat Inc, All Rights Reserved |
| 9 | */ |
| 10 | |
| 11 | #include <linux/mm.h> |
| 12 | #include <linux/hugetlb.h> |
| 13 | #include <linux/shm.h> |
| 14 | #include <linux/ksm.h> |
| 15 | #include <linux/mman.h> |
| 16 | #include <linux/swap.h> |
| 17 | #include <linux/capability.h> |
| 18 | #include <linux/fs.h> |
| 19 | #include <linux/swapops.h> |
| 20 | #include <linux/highmem.h> |
| 21 | #include <linux/security.h> |
| 22 | #include <linux/syscalls.h> |
| 23 | #include <linux/mmu_notifier.h> |
| 24 | #include <linux/uaccess.h> |
| 25 | #include <linux/userfaultfd_k.h> |
| 26 | |
| 27 | #include <asm/cacheflush.h> |
| 28 | #include <asm/tlbflush.h> |
| 29 | |
| 30 | #include "internal.h" |
| 31 | |
| 32 | static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr) |
| 33 | { |
| 34 | pgd_t *pgd; |
| 35 | p4d_t *p4d; |
| 36 | pud_t *pud; |
| 37 | |
| 38 | pgd = pgd_offset(mm, addr); |
| 39 | if (pgd_none_or_clear_bad(pgd)) |
| 40 | return NULL; |
| 41 | |
| 42 | p4d = p4d_offset(pgd, addr); |
| 43 | if (p4d_none_or_clear_bad(p4d)) |
| 44 | return NULL; |
| 45 | |
| 46 | pud = pud_offset(p4d, addr); |
| 47 | if (pud_none_or_clear_bad(pud)) |
| 48 | return NULL; |
| 49 | |
| 50 | return pud; |
| 51 | } |
| 52 | |
| 53 | static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr) |
| 54 | { |
| 55 | pud_t *pud; |
| 56 | pmd_t *pmd; |
| 57 | |
| 58 | pud = get_old_pud(mm, addr); |
| 59 | if (!pud) |
| 60 | return NULL; |
| 61 | |
| 62 | pmd = pmd_offset(pud, addr); |
| 63 | if (pmd_none(*pmd)) |
| 64 | return NULL; |
| 65 | |
| 66 | return pmd; |
| 67 | } |
| 68 | |
| 69 | static pud_t *alloc_new_pud(struct mm_struct *mm, struct vm_area_struct *vma, |
| 70 | unsigned long addr) |
| 71 | { |
| 72 | pgd_t *pgd; |
| 73 | p4d_t *p4d; |
| 74 | |
| 75 | pgd = pgd_offset(mm, addr); |
| 76 | p4d = p4d_alloc(mm, pgd, addr); |
| 77 | if (!p4d) |
| 78 | return NULL; |
| 79 | |
| 80 | return pud_alloc(mm, p4d, addr); |
| 81 | } |
| 82 | |
| 83 | static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma, |
| 84 | unsigned long addr) |
| 85 | { |
| 86 | pud_t *pud; |
| 87 | pmd_t *pmd; |
| 88 | |
| 89 | pud = alloc_new_pud(mm, vma, addr); |
| 90 | if (!pud) |
| 91 | return NULL; |
| 92 | |
| 93 | pmd = pmd_alloc(mm, pud, addr); |
| 94 | if (!pmd) |
| 95 | return NULL; |
| 96 | |
| 97 | VM_BUG_ON(pmd_trans_huge(*pmd)); |
| 98 | |
| 99 | return pmd; |
| 100 | } |
| 101 | |
| 102 | static void take_rmap_locks(struct vm_area_struct *vma) |
| 103 | { |
| 104 | if (vma->vm_file) |
| 105 | i_mmap_lock_write(vma->vm_file->f_mapping); |
| 106 | if (vma->anon_vma) |
| 107 | anon_vma_lock_write(vma->anon_vma); |
| 108 | } |
| 109 | |
| 110 | static void drop_rmap_locks(struct vm_area_struct *vma) |
| 111 | { |
| 112 | if (vma->anon_vma) |
| 113 | anon_vma_unlock_write(vma->anon_vma); |
| 114 | if (vma->vm_file) |
| 115 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
| 116 | } |
| 117 | |
| 118 | static pte_t move_soft_dirty_pte(pte_t pte) |
| 119 | { |
| 120 | /* |
| 121 | * Set soft dirty bit so we can notice |
| 122 | * in userspace the ptes were moved. |
| 123 | */ |
| 124 | #ifdef CONFIG_MEM_SOFT_DIRTY |
| 125 | if (pte_present(pte)) |
| 126 | pte = pte_mksoft_dirty(pte); |
| 127 | else if (is_swap_pte(pte)) |
| 128 | pte = pte_swp_mksoft_dirty(pte); |
| 129 | #endif |
| 130 | return pte; |
| 131 | } |
| 132 | |
| 133 | static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd, |
| 134 | unsigned long old_addr, unsigned long old_end, |
| 135 | struct vm_area_struct *new_vma, pmd_t *new_pmd, |
| 136 | unsigned long new_addr, bool need_rmap_locks) |
| 137 | { |
| 138 | struct mm_struct *mm = vma->vm_mm; |
| 139 | pte_t *old_pte, *new_pte, pte; |
| 140 | spinlock_t *old_ptl, *new_ptl; |
| 141 | bool force_flush = false; |
| 142 | unsigned long len = old_end - old_addr; |
| 143 | |
| 144 | /* |
| 145 | * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma |
| 146 | * locks to ensure that rmap will always observe either the old or the |
| 147 | * new ptes. This is the easiest way to avoid races with |
| 148 | * truncate_pagecache(), page migration, etc... |
| 149 | * |
| 150 | * When need_rmap_locks is false, we use other ways to avoid |
| 151 | * such races: |
| 152 | * |
| 153 | * - During exec() shift_arg_pages(), we use a specially tagged vma |
| 154 | * which rmap call sites look for using vma_is_temporary_stack(). |
| 155 | * |
| 156 | * - During mremap(), new_vma is often known to be placed after vma |
| 157 | * in rmap traversal order. This ensures rmap will always observe |
| 158 | * either the old pte, or the new pte, or both (the page table locks |
| 159 | * serialize access to individual ptes, but only rmap traversal |
| 160 | * order guarantees that we won't miss both the old and new ptes). |
| 161 | */ |
| 162 | if (need_rmap_locks) |
| 163 | take_rmap_locks(vma); |
| 164 | |
| 165 | /* |
| 166 | * We don't have to worry about the ordering of src and dst |
| 167 | * pte locks because exclusive mmap_lock prevents deadlock. |
| 168 | */ |
| 169 | old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl); |
| 170 | new_pte = pte_offset_map(new_pmd, new_addr); |
| 171 | new_ptl = pte_lockptr(mm, new_pmd); |
| 172 | if (new_ptl != old_ptl) |
| 173 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); |
| 174 | flush_tlb_batched_pending(vma->vm_mm); |
| 175 | arch_enter_lazy_mmu_mode(); |
| 176 | |
| 177 | for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE, |
| 178 | new_pte++, new_addr += PAGE_SIZE) { |
| 179 | if (pte_none(*old_pte)) |
| 180 | continue; |
| 181 | |
| 182 | pte = ptep_get_and_clear(mm, old_addr, old_pte); |
| 183 | /* |
| 184 | * If we are remapping a valid PTE, make sure |
| 185 | * to flush TLB before we drop the PTL for the |
| 186 | * PTE. |
| 187 | * |
| 188 | * NOTE! Both old and new PTL matter: the old one |
| 189 | * for racing with page_mkclean(), the new one to |
| 190 | * make sure the physical page stays valid until |
| 191 | * the TLB entry for the old mapping has been |
| 192 | * flushed. |
| 193 | */ |
| 194 | if (pte_present(pte)) |
| 195 | force_flush = true; |
| 196 | pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr); |
| 197 | pte = move_soft_dirty_pte(pte); |
| 198 | set_pte_at(mm, new_addr, new_pte, pte); |
| 199 | } |
| 200 | |
| 201 | arch_leave_lazy_mmu_mode(); |
| 202 | if (force_flush) |
| 203 | flush_tlb_range(vma, old_end - len, old_end); |
| 204 | if (new_ptl != old_ptl) |
| 205 | spin_unlock(new_ptl); |
| 206 | pte_unmap(new_pte - 1); |
| 207 | pte_unmap_unlock(old_pte - 1, old_ptl); |
| 208 | if (need_rmap_locks) |
| 209 | drop_rmap_locks(vma); |
| 210 | } |
| 211 | |
| 212 | #ifdef CONFIG_HAVE_MOVE_PMD |
| 213 | static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr, |
| 214 | unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) |
| 215 | { |
| 216 | spinlock_t *old_ptl, *new_ptl; |
| 217 | struct mm_struct *mm = vma->vm_mm; |
| 218 | pmd_t pmd; |
| 219 | |
| 220 | /* |
| 221 | * The destination pmd shouldn't be established, free_pgtables() |
| 222 | * should have released it. |
| 223 | * |
| 224 | * However, there's a case during execve() where we use mremap |
| 225 | * to move the initial stack, and in that case the target area |
| 226 | * may overlap the source area (always moving down). |
| 227 | * |
| 228 | * If everything is PMD-aligned, that works fine, as moving |
| 229 | * each pmd down will clear the source pmd. But if we first |
| 230 | * have a few 4kB-only pages that get moved down, and then |
| 231 | * hit the "now the rest is PMD-aligned, let's do everything |
| 232 | * one pmd at a time", we will still have the old (now empty |
| 233 | * of any 4kB pages, but still there) PMD in the page table |
| 234 | * tree. |
| 235 | * |
| 236 | * Warn on it once - because we really should try to figure |
| 237 | * out how to do this better - but then say "I won't move |
| 238 | * this pmd". |
| 239 | * |
| 240 | * One alternative might be to just unmap the target pmd at |
| 241 | * this point, and verify that it really is empty. We'll see. |
| 242 | */ |
| 243 | if (WARN_ON_ONCE(!pmd_none(*new_pmd))) |
| 244 | return false; |
| 245 | |
| 246 | /* |
| 247 | * We don't have to worry about the ordering of src and dst |
| 248 | * ptlocks because exclusive mmap_lock prevents deadlock. |
| 249 | */ |
| 250 | old_ptl = pmd_lock(vma->vm_mm, old_pmd); |
| 251 | new_ptl = pmd_lockptr(mm, new_pmd); |
| 252 | if (new_ptl != old_ptl) |
| 253 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); |
| 254 | |
| 255 | /* Clear the pmd */ |
| 256 | pmd = *old_pmd; |
| 257 | pmd_clear(old_pmd); |
| 258 | |
| 259 | VM_BUG_ON(!pmd_none(*new_pmd)); |
| 260 | |
| 261 | /* Set the new pmd */ |
| 262 | set_pmd_at(mm, new_addr, new_pmd, pmd); |
| 263 | flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); |
| 264 | if (new_ptl != old_ptl) |
| 265 | spin_unlock(new_ptl); |
| 266 | spin_unlock(old_ptl); |
| 267 | |
| 268 | return true; |
| 269 | } |
| 270 | #else |
| 271 | static inline bool move_normal_pmd(struct vm_area_struct *vma, |
| 272 | unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd, |
| 273 | pmd_t *new_pmd) |
| 274 | { |
| 275 | return false; |
| 276 | } |
| 277 | #endif |
| 278 | |
| 279 | #ifdef CONFIG_HAVE_MOVE_PUD |
| 280 | static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr, |
| 281 | unsigned long new_addr, pud_t *old_pud, pud_t *new_pud) |
| 282 | { |
| 283 | spinlock_t *old_ptl, *new_ptl; |
| 284 | struct mm_struct *mm = vma->vm_mm; |
| 285 | pud_t pud; |
| 286 | |
| 287 | /* |
| 288 | * The destination pud shouldn't be established, free_pgtables() |
| 289 | * should have released it. |
| 290 | */ |
| 291 | if (WARN_ON_ONCE(!pud_none(*new_pud))) |
| 292 | return false; |
| 293 | |
| 294 | /* |
| 295 | * We don't have to worry about the ordering of src and dst |
| 296 | * ptlocks because exclusive mmap_lock prevents deadlock. |
| 297 | */ |
| 298 | old_ptl = pud_lock(vma->vm_mm, old_pud); |
| 299 | new_ptl = pud_lockptr(mm, new_pud); |
| 300 | if (new_ptl != old_ptl) |
| 301 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); |
| 302 | |
| 303 | /* Clear the pud */ |
| 304 | pud = *old_pud; |
| 305 | pud_clear(old_pud); |
| 306 | |
| 307 | VM_BUG_ON(!pud_none(*new_pud)); |
| 308 | |
| 309 | /* Set the new pud */ |
| 310 | set_pud_at(mm, new_addr, new_pud, pud); |
| 311 | flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE); |
| 312 | if (new_ptl != old_ptl) |
| 313 | spin_unlock(new_ptl); |
| 314 | spin_unlock(old_ptl); |
| 315 | |
| 316 | return true; |
| 317 | } |
| 318 | #else |
| 319 | static inline bool move_normal_pud(struct vm_area_struct *vma, |
| 320 | unsigned long old_addr, unsigned long new_addr, pud_t *old_pud, |
| 321 | pud_t *new_pud) |
| 322 | { |
| 323 | return false; |
| 324 | } |
| 325 | #endif |
| 326 | |
| 327 | enum pgt_entry { |
| 328 | NORMAL_PMD, |
| 329 | HPAGE_PMD, |
| 330 | NORMAL_PUD, |
| 331 | }; |
| 332 | |
| 333 | /* |
| 334 | * Returns an extent of the corresponding size for the pgt_entry specified if |
| 335 | * valid. Else returns a smaller extent bounded by the end of the source and |
| 336 | * destination pgt_entry. |
| 337 | */ |
| 338 | static __always_inline unsigned long get_extent(enum pgt_entry entry, |
| 339 | unsigned long old_addr, unsigned long old_end, |
| 340 | unsigned long new_addr) |
| 341 | { |
| 342 | unsigned long next, extent, mask, size; |
| 343 | |
| 344 | switch (entry) { |
| 345 | case HPAGE_PMD: |
| 346 | case NORMAL_PMD: |
| 347 | mask = PMD_MASK; |
| 348 | size = PMD_SIZE; |
| 349 | break; |
| 350 | case NORMAL_PUD: |
| 351 | mask = PUD_MASK; |
| 352 | size = PUD_SIZE; |
| 353 | break; |
| 354 | default: |
| 355 | BUILD_BUG(); |
| 356 | break; |
| 357 | } |
| 358 | |
| 359 | next = (old_addr + size) & mask; |
| 360 | /* even if next overflowed, extent below will be ok */ |
| 361 | extent = next - old_addr; |
| 362 | if (extent > old_end - old_addr) |
| 363 | extent = old_end - old_addr; |
| 364 | next = (new_addr + size) & mask; |
| 365 | if (extent > next - new_addr) |
| 366 | extent = next - new_addr; |
| 367 | return extent; |
| 368 | } |
| 369 | |
| 370 | /* |
| 371 | * Attempts to speedup the move by moving entry at the level corresponding to |
| 372 | * pgt_entry. Returns true if the move was successful, else false. |
| 373 | */ |
| 374 | static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma, |
| 375 | unsigned long old_addr, unsigned long new_addr, |
| 376 | void *old_entry, void *new_entry, bool need_rmap_locks) |
| 377 | { |
| 378 | bool moved = false; |
| 379 | |
| 380 | /* See comment in move_ptes() */ |
| 381 | if (need_rmap_locks) |
| 382 | take_rmap_locks(vma); |
| 383 | |
| 384 | switch (entry) { |
| 385 | case NORMAL_PMD: |
| 386 | moved = move_normal_pmd(vma, old_addr, new_addr, old_entry, |
| 387 | new_entry); |
| 388 | break; |
| 389 | case NORMAL_PUD: |
| 390 | moved = move_normal_pud(vma, old_addr, new_addr, old_entry, |
| 391 | new_entry); |
| 392 | break; |
| 393 | case HPAGE_PMD: |
| 394 | moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && |
| 395 | move_huge_pmd(vma, old_addr, new_addr, old_entry, |
| 396 | new_entry); |
| 397 | break; |
| 398 | default: |
| 399 | WARN_ON_ONCE(1); |
| 400 | break; |
| 401 | } |
| 402 | |
| 403 | if (need_rmap_locks) |
| 404 | drop_rmap_locks(vma); |
| 405 | |
| 406 | return moved; |
| 407 | } |
| 408 | |
| 409 | unsigned long move_page_tables(struct vm_area_struct *vma, |
| 410 | unsigned long old_addr, struct vm_area_struct *new_vma, |
| 411 | unsigned long new_addr, unsigned long len, |
| 412 | bool need_rmap_locks) |
| 413 | { |
| 414 | unsigned long extent, old_end; |
| 415 | struct mmu_notifier_range range; |
| 416 | pmd_t *old_pmd, *new_pmd; |
| 417 | |
| 418 | old_end = old_addr + len; |
| 419 | flush_cache_range(vma, old_addr, old_end); |
| 420 | |
| 421 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm, |
| 422 | old_addr, old_end); |
| 423 | mmu_notifier_invalidate_range_start(&range); |
| 424 | |
| 425 | for (; old_addr < old_end; old_addr += extent, new_addr += extent) { |
| 426 | cond_resched(); |
| 427 | /* |
| 428 | * If extent is PUD-sized try to speed up the move by moving at the |
| 429 | * PUD level if possible. |
| 430 | */ |
| 431 | extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr); |
| 432 | if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) { |
| 433 | pud_t *old_pud, *new_pud; |
| 434 | |
| 435 | old_pud = get_old_pud(vma->vm_mm, old_addr); |
| 436 | if (!old_pud) |
| 437 | continue; |
| 438 | new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr); |
| 439 | if (!new_pud) |
| 440 | break; |
| 441 | if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr, |
| 442 | old_pud, new_pud, need_rmap_locks)) |
| 443 | continue; |
| 444 | } |
| 445 | |
| 446 | extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr); |
| 447 | old_pmd = get_old_pmd(vma->vm_mm, old_addr); |
| 448 | if (!old_pmd) |
| 449 | continue; |
| 450 | new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr); |
| 451 | if (!new_pmd) |
| 452 | break; |
| 453 | if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) || |
| 454 | pmd_devmap(*old_pmd)) { |
| 455 | if (extent == HPAGE_PMD_SIZE && |
| 456 | move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr, |
| 457 | old_pmd, new_pmd, need_rmap_locks)) |
| 458 | continue; |
| 459 | split_huge_pmd(vma, old_pmd, old_addr); |
| 460 | if (pmd_trans_unstable(old_pmd)) |
| 461 | continue; |
| 462 | } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) && |
| 463 | extent == PMD_SIZE) { |
| 464 | /* |
| 465 | * If the extent is PMD-sized, try to speed the move by |
| 466 | * moving at the PMD level if possible. |
| 467 | */ |
| 468 | if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr, |
| 469 | old_pmd, new_pmd, need_rmap_locks)) |
| 470 | continue; |
| 471 | } |
| 472 | |
| 473 | if (pte_alloc(new_vma->vm_mm, new_pmd)) |
| 474 | break; |
| 475 | move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma, |
| 476 | new_pmd, new_addr, need_rmap_locks); |
| 477 | } |
| 478 | |
| 479 | mmu_notifier_invalidate_range_end(&range); |
| 480 | |
| 481 | return len + old_addr - old_end; /* how much done */ |
| 482 | } |
| 483 | |
| 484 | static unsigned long move_vma(struct vm_area_struct *vma, |
| 485 | unsigned long old_addr, unsigned long old_len, |
| 486 | unsigned long new_len, unsigned long new_addr, |
| 487 | bool *locked, unsigned long flags, |
| 488 | struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap) |
| 489 | { |
| 490 | struct mm_struct *mm = vma->vm_mm; |
| 491 | struct vm_area_struct *new_vma; |
| 492 | unsigned long vm_flags = vma->vm_flags; |
| 493 | unsigned long new_pgoff; |
| 494 | unsigned long moved_len; |
| 495 | unsigned long excess = 0; |
| 496 | unsigned long hiwater_vm; |
| 497 | int split = 0; |
| 498 | int err = 0; |
| 499 | bool need_rmap_locks; |
| 500 | |
| 501 | /* |
| 502 | * We'd prefer to avoid failure later on in do_munmap: |
| 503 | * which may split one vma into three before unmapping. |
| 504 | */ |
| 505 | if (mm->map_count >= sysctl_max_map_count - 3) |
| 506 | return -ENOMEM; |
| 507 | |
| 508 | if (vma->vm_ops && vma->vm_ops->may_split) { |
| 509 | if (vma->vm_start != old_addr) |
| 510 | err = vma->vm_ops->may_split(vma, old_addr); |
| 511 | if (!err && vma->vm_end != old_addr + old_len) |
| 512 | err = vma->vm_ops->may_split(vma, old_addr + old_len); |
| 513 | if (err) |
| 514 | return err; |
| 515 | } |
| 516 | |
| 517 | /* |
| 518 | * Advise KSM to break any KSM pages in the area to be moved: |
| 519 | * it would be confusing if they were to turn up at the new |
| 520 | * location, where they happen to coincide with different KSM |
| 521 | * pages recently unmapped. But leave vma->vm_flags as it was, |
| 522 | * so KSM can come around to merge on vma and new_vma afterwards. |
| 523 | */ |
| 524 | err = ksm_madvise(vma, old_addr, old_addr + old_len, |
| 525 | MADV_UNMERGEABLE, &vm_flags); |
| 526 | if (err) |
| 527 | return err; |
| 528 | |
| 529 | if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT)) { |
| 530 | if (security_vm_enough_memory_mm(mm, new_len >> PAGE_SHIFT)) |
| 531 | return -ENOMEM; |
| 532 | } |
| 533 | |
| 534 | new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT); |
| 535 | new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff, |
| 536 | &need_rmap_locks); |
| 537 | if (!new_vma) { |
| 538 | if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT)) |
| 539 | vm_unacct_memory(new_len >> PAGE_SHIFT); |
| 540 | return -ENOMEM; |
| 541 | } |
| 542 | |
| 543 | moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len, |
| 544 | need_rmap_locks); |
| 545 | if (moved_len < old_len) { |
| 546 | err = -ENOMEM; |
| 547 | } else if (vma->vm_ops && vma->vm_ops->mremap) { |
| 548 | err = vma->vm_ops->mremap(new_vma); |
| 549 | } |
| 550 | |
| 551 | if (unlikely(err)) { |
| 552 | /* |
| 553 | * On error, move entries back from new area to old, |
| 554 | * which will succeed since page tables still there, |
| 555 | * and then proceed to unmap new area instead of old. |
| 556 | */ |
| 557 | move_page_tables(new_vma, new_addr, vma, old_addr, moved_len, |
| 558 | true); |
| 559 | vma = new_vma; |
| 560 | old_len = new_len; |
| 561 | old_addr = new_addr; |
| 562 | new_addr = err; |
| 563 | } else { |
| 564 | mremap_userfaultfd_prep(new_vma, uf); |
| 565 | } |
| 566 | |
| 567 | /* Conceal VM_ACCOUNT so old reservation is not undone */ |
| 568 | if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) { |
| 569 | vma->vm_flags &= ~VM_ACCOUNT; |
| 570 | excess = vma->vm_end - vma->vm_start - old_len; |
| 571 | if (old_addr > vma->vm_start && |
| 572 | old_addr + old_len < vma->vm_end) |
| 573 | split = 1; |
| 574 | } |
| 575 | |
| 576 | /* |
| 577 | * If we failed to move page tables we still do total_vm increment |
| 578 | * since do_munmap() will decrement it by old_len == new_len. |
| 579 | * |
| 580 | * Since total_vm is about to be raised artificially high for a |
| 581 | * moment, we need to restore high watermark afterwards: if stats |
| 582 | * are taken meanwhile, total_vm and hiwater_vm appear too high. |
| 583 | * If this were a serious issue, we'd add a flag to do_munmap(). |
| 584 | */ |
| 585 | hiwater_vm = mm->hiwater_vm; |
| 586 | vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT); |
| 587 | |
| 588 | /* Tell pfnmap has moved from this vma */ |
| 589 | if (unlikely(vma->vm_flags & VM_PFNMAP)) |
| 590 | untrack_pfn_moved(vma); |
| 591 | |
| 592 | if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) { |
| 593 | /* We always clear VM_LOCKED[ONFAULT] on the old vma */ |
| 594 | vma->vm_flags &= VM_LOCKED_CLEAR_MASK; |
| 595 | |
| 596 | /* |
| 597 | * anon_vma links of the old vma is no longer needed after its page |
| 598 | * table has been moved. |
| 599 | */ |
| 600 | if (new_vma != vma && vma->vm_start == old_addr && |
| 601 | vma->vm_end == (old_addr + old_len)) |
| 602 | unlink_anon_vmas(vma); |
| 603 | |
| 604 | /* Because we won't unmap we don't need to touch locked_vm */ |
| 605 | return new_addr; |
| 606 | } |
| 607 | |
| 608 | if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) { |
| 609 | /* OOM: unable to split vma, just get accounts right */ |
| 610 | if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) |
| 611 | vm_acct_memory(new_len >> PAGE_SHIFT); |
| 612 | excess = 0; |
| 613 | } |
| 614 | |
| 615 | if (vm_flags & VM_LOCKED) { |
| 616 | mm->locked_vm += new_len >> PAGE_SHIFT; |
| 617 | *locked = true; |
| 618 | } |
| 619 | |
| 620 | mm->hiwater_vm = hiwater_vm; |
| 621 | |
| 622 | /* Restore VM_ACCOUNT if one or two pieces of vma left */ |
| 623 | if (excess) { |
| 624 | vma->vm_flags |= VM_ACCOUNT; |
| 625 | if (split) |
| 626 | vma->vm_next->vm_flags |= VM_ACCOUNT; |
| 627 | } |
| 628 | |
| 629 | return new_addr; |
| 630 | } |
| 631 | |
| 632 | static struct vm_area_struct *vma_to_resize(unsigned long addr, |
| 633 | unsigned long old_len, unsigned long new_len, unsigned long flags, |
| 634 | unsigned long *p) |
| 635 | { |
| 636 | struct mm_struct *mm = current->mm; |
| 637 | struct vm_area_struct *vma; |
| 638 | unsigned long pgoff; |
| 639 | |
| 640 | vma = vma_lookup(mm, addr); |
| 641 | if (!vma) |
| 642 | return ERR_PTR(-EFAULT); |
| 643 | |
| 644 | /* |
| 645 | * !old_len is a special case where an attempt is made to 'duplicate' |
| 646 | * a mapping. This makes no sense for private mappings as it will |
| 647 | * instead create a fresh/new mapping unrelated to the original. This |
| 648 | * is contrary to the basic idea of mremap which creates new mappings |
| 649 | * based on the original. There are no known use cases for this |
| 650 | * behavior. As a result, fail such attempts. |
| 651 | */ |
| 652 | if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) { |
| 653 | pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid); |
| 654 | return ERR_PTR(-EINVAL); |
| 655 | } |
| 656 | |
| 657 | if ((flags & MREMAP_DONTUNMAP) && |
| 658 | (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))) |
| 659 | return ERR_PTR(-EINVAL); |
| 660 | |
| 661 | if (is_vm_hugetlb_page(vma)) |
| 662 | return ERR_PTR(-EINVAL); |
| 663 | |
| 664 | /* We can't remap across vm area boundaries */ |
| 665 | if (old_len > vma->vm_end - addr) |
| 666 | return ERR_PTR(-EFAULT); |
| 667 | |
| 668 | if (new_len == old_len) |
| 669 | return vma; |
| 670 | |
| 671 | /* Need to be careful about a growing mapping */ |
| 672 | pgoff = (addr - vma->vm_start) >> PAGE_SHIFT; |
| 673 | pgoff += vma->vm_pgoff; |
| 674 | if (pgoff + (new_len >> PAGE_SHIFT) < pgoff) |
| 675 | return ERR_PTR(-EINVAL); |
| 676 | |
| 677 | if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) |
| 678 | return ERR_PTR(-EFAULT); |
| 679 | |
| 680 | if (vma->vm_flags & VM_LOCKED) { |
| 681 | unsigned long locked, lock_limit; |
| 682 | locked = mm->locked_vm << PAGE_SHIFT; |
| 683 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
| 684 | locked += new_len - old_len; |
| 685 | if (locked > lock_limit && !capable(CAP_IPC_LOCK)) |
| 686 | return ERR_PTR(-EAGAIN); |
| 687 | } |
| 688 | |
| 689 | if (!may_expand_vm(mm, vma->vm_flags, |
| 690 | (new_len - old_len) >> PAGE_SHIFT)) |
| 691 | return ERR_PTR(-ENOMEM); |
| 692 | |
| 693 | if (vma->vm_flags & VM_ACCOUNT) { |
| 694 | unsigned long charged = (new_len - old_len) >> PAGE_SHIFT; |
| 695 | if (security_vm_enough_memory_mm(mm, charged)) |
| 696 | return ERR_PTR(-ENOMEM); |
| 697 | *p = charged; |
| 698 | } |
| 699 | |
| 700 | return vma; |
| 701 | } |
| 702 | |
| 703 | static unsigned long mremap_to(unsigned long addr, unsigned long old_len, |
| 704 | unsigned long new_addr, unsigned long new_len, bool *locked, |
| 705 | unsigned long flags, struct vm_userfaultfd_ctx *uf, |
| 706 | struct list_head *uf_unmap_early, |
| 707 | struct list_head *uf_unmap) |
| 708 | { |
| 709 | struct mm_struct *mm = current->mm; |
| 710 | struct vm_area_struct *vma; |
| 711 | unsigned long ret = -EINVAL; |
| 712 | unsigned long charged = 0; |
| 713 | unsigned long map_flags = 0; |
| 714 | |
| 715 | if (offset_in_page(new_addr)) |
| 716 | goto out; |
| 717 | |
| 718 | if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len) |
| 719 | goto out; |
| 720 | |
| 721 | /* Ensure the old/new locations do not overlap */ |
| 722 | if (addr + old_len > new_addr && new_addr + new_len > addr) |
| 723 | goto out; |
| 724 | |
| 725 | /* |
| 726 | * move_vma() need us to stay 4 maps below the threshold, otherwise |
| 727 | * it will bail out at the very beginning. |
| 728 | * That is a problem if we have already unmaped the regions here |
| 729 | * (new_addr, and old_addr), because userspace will not know the |
| 730 | * state of the vma's after it gets -ENOMEM. |
| 731 | * So, to avoid such scenario we can pre-compute if the whole |
| 732 | * operation has high chances to success map-wise. |
| 733 | * Worst-scenario case is when both vma's (new_addr and old_addr) get |
| 734 | * split in 3 before unmapping it. |
| 735 | * That means 2 more maps (1 for each) to the ones we already hold. |
| 736 | * Check whether current map count plus 2 still leads us to 4 maps below |
| 737 | * the threshold, otherwise return -ENOMEM here to be more safe. |
| 738 | */ |
| 739 | if ((mm->map_count + 2) >= sysctl_max_map_count - 3) |
| 740 | return -ENOMEM; |
| 741 | |
| 742 | if (flags & MREMAP_FIXED) { |
| 743 | ret = do_munmap(mm, new_addr, new_len, uf_unmap_early); |
| 744 | if (ret) |
| 745 | goto out; |
| 746 | } |
| 747 | |
| 748 | if (old_len >= new_len) { |
| 749 | ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap); |
| 750 | if (ret && old_len != new_len) |
| 751 | goto out; |
| 752 | old_len = new_len; |
| 753 | } |
| 754 | |
| 755 | vma = vma_to_resize(addr, old_len, new_len, flags, &charged); |
| 756 | if (IS_ERR(vma)) { |
| 757 | ret = PTR_ERR(vma); |
| 758 | goto out; |
| 759 | } |
| 760 | |
| 761 | /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */ |
| 762 | if (flags & MREMAP_DONTUNMAP && |
| 763 | !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) { |
| 764 | ret = -ENOMEM; |
| 765 | goto out; |
| 766 | } |
| 767 | |
| 768 | if (flags & MREMAP_FIXED) |
| 769 | map_flags |= MAP_FIXED; |
| 770 | |
| 771 | if (vma->vm_flags & VM_MAYSHARE) |
| 772 | map_flags |= MAP_SHARED; |
| 773 | |
| 774 | ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff + |
| 775 | ((addr - vma->vm_start) >> PAGE_SHIFT), |
| 776 | map_flags); |
| 777 | if (IS_ERR_VALUE(ret)) |
| 778 | goto out1; |
| 779 | |
| 780 | /* We got a new mapping */ |
| 781 | if (!(flags & MREMAP_FIXED)) |
| 782 | new_addr = ret; |
| 783 | |
| 784 | ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf, |
| 785 | uf_unmap); |
| 786 | |
| 787 | if (!(offset_in_page(ret))) |
| 788 | goto out; |
| 789 | |
| 790 | out1: |
| 791 | vm_unacct_memory(charged); |
| 792 | |
| 793 | out: |
| 794 | return ret; |
| 795 | } |
| 796 | |
| 797 | static int vma_expandable(struct vm_area_struct *vma, unsigned long delta) |
| 798 | { |
| 799 | unsigned long end = vma->vm_end + delta; |
| 800 | if (end < vma->vm_end) /* overflow */ |
| 801 | return 0; |
| 802 | if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */ |
| 803 | return 0; |
| 804 | if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start, |
| 805 | 0, MAP_FIXED) & ~PAGE_MASK) |
| 806 | return 0; |
| 807 | return 1; |
| 808 | } |
| 809 | |
| 810 | /* |
| 811 | * Expand (or shrink) an existing mapping, potentially moving it at the |
| 812 | * same time (controlled by the MREMAP_MAYMOVE flag and available VM space) |
| 813 | * |
| 814 | * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise |
| 815 | * This option implies MREMAP_MAYMOVE. |
| 816 | */ |
| 817 | SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len, |
| 818 | unsigned long, new_len, unsigned long, flags, |
| 819 | unsigned long, new_addr) |
| 820 | { |
| 821 | struct mm_struct *mm = current->mm; |
| 822 | struct vm_area_struct *vma; |
| 823 | unsigned long ret = -EINVAL; |
| 824 | unsigned long charged = 0; |
| 825 | bool locked = false; |
| 826 | bool downgraded = false; |
| 827 | struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX; |
| 828 | LIST_HEAD(uf_unmap_early); |
| 829 | LIST_HEAD(uf_unmap); |
| 830 | |
| 831 | /* |
| 832 | * There is a deliberate asymmetry here: we strip the pointer tag |
| 833 | * from the old address but leave the new address alone. This is |
| 834 | * for consistency with mmap(), where we prevent the creation of |
| 835 | * aliasing mappings in userspace by leaving the tag bits of the |
| 836 | * mapping address intact. A non-zero tag will cause the subsequent |
| 837 | * range checks to reject the address as invalid. |
| 838 | * |
| 839 | * See Documentation/arm64/tagged-address-abi.rst for more information. |
| 840 | */ |
| 841 | addr = untagged_addr(addr); |
| 842 | |
| 843 | if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP)) |
| 844 | return ret; |
| 845 | |
| 846 | if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE)) |
| 847 | return ret; |
| 848 | |
| 849 | /* |
| 850 | * MREMAP_DONTUNMAP is always a move and it does not allow resizing |
| 851 | * in the process. |
| 852 | */ |
| 853 | if (flags & MREMAP_DONTUNMAP && |
| 854 | (!(flags & MREMAP_MAYMOVE) || old_len != new_len)) |
| 855 | return ret; |
| 856 | |
| 857 | |
| 858 | if (offset_in_page(addr)) |
| 859 | return ret; |
| 860 | |
| 861 | old_len = PAGE_ALIGN(old_len); |
| 862 | new_len = PAGE_ALIGN(new_len); |
| 863 | |
| 864 | /* |
| 865 | * We allow a zero old-len as a special case |
| 866 | * for DOS-emu "duplicate shm area" thing. But |
| 867 | * a zero new-len is nonsensical. |
| 868 | */ |
| 869 | if (!new_len) |
| 870 | return ret; |
| 871 | |
| 872 | if (mmap_write_lock_killable(current->mm)) |
| 873 | return -EINTR; |
| 874 | |
| 875 | if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) { |
| 876 | ret = mremap_to(addr, old_len, new_addr, new_len, |
| 877 | &locked, flags, &uf, &uf_unmap_early, |
| 878 | &uf_unmap); |
| 879 | goto out; |
| 880 | } |
| 881 | |
| 882 | /* |
| 883 | * Always allow a shrinking remap: that just unmaps |
| 884 | * the unnecessary pages.. |
| 885 | * __do_munmap does all the needed commit accounting, and |
| 886 | * downgrades mmap_lock to read if so directed. |
| 887 | */ |
| 888 | if (old_len >= new_len) { |
| 889 | int retval; |
| 890 | |
| 891 | retval = __do_munmap(mm, addr+new_len, old_len - new_len, |
| 892 | &uf_unmap, true); |
| 893 | if (retval < 0 && old_len != new_len) { |
| 894 | ret = retval; |
| 895 | goto out; |
| 896 | /* Returning 1 indicates mmap_lock is downgraded to read. */ |
| 897 | } else if (retval == 1) |
| 898 | downgraded = true; |
| 899 | ret = addr; |
| 900 | goto out; |
| 901 | } |
| 902 | |
| 903 | /* |
| 904 | * Ok, we need to grow.. |
| 905 | */ |
| 906 | vma = vma_to_resize(addr, old_len, new_len, flags, &charged); |
| 907 | if (IS_ERR(vma)) { |
| 908 | ret = PTR_ERR(vma); |
| 909 | goto out; |
| 910 | } |
| 911 | |
| 912 | /* old_len exactly to the end of the area.. |
| 913 | */ |
| 914 | if (old_len == vma->vm_end - addr) { |
| 915 | /* can we just expand the current mapping? */ |
| 916 | if (vma_expandable(vma, new_len - old_len)) { |
| 917 | int pages = (new_len - old_len) >> PAGE_SHIFT; |
| 918 | |
| 919 | if (vma_adjust(vma, vma->vm_start, addr + new_len, |
| 920 | vma->vm_pgoff, NULL)) { |
| 921 | ret = -ENOMEM; |
| 922 | goto out; |
| 923 | } |
| 924 | |
| 925 | vm_stat_account(mm, vma->vm_flags, pages); |
| 926 | if (vma->vm_flags & VM_LOCKED) { |
| 927 | mm->locked_vm += pages; |
| 928 | locked = true; |
| 929 | new_addr = addr; |
| 930 | } |
| 931 | ret = addr; |
| 932 | goto out; |
| 933 | } |
| 934 | } |
| 935 | |
| 936 | /* |
| 937 | * We weren't able to just expand or shrink the area, |
| 938 | * we need to create a new one and move it.. |
| 939 | */ |
| 940 | ret = -ENOMEM; |
| 941 | if (flags & MREMAP_MAYMOVE) { |
| 942 | unsigned long map_flags = 0; |
| 943 | if (vma->vm_flags & VM_MAYSHARE) |
| 944 | map_flags |= MAP_SHARED; |
| 945 | |
| 946 | new_addr = get_unmapped_area(vma->vm_file, 0, new_len, |
| 947 | vma->vm_pgoff + |
| 948 | ((addr - vma->vm_start) >> PAGE_SHIFT), |
| 949 | map_flags); |
| 950 | if (IS_ERR_VALUE(new_addr)) { |
| 951 | ret = new_addr; |
| 952 | goto out; |
| 953 | } |
| 954 | |
| 955 | ret = move_vma(vma, addr, old_len, new_len, new_addr, |
| 956 | &locked, flags, &uf, &uf_unmap); |
| 957 | } |
| 958 | out: |
| 959 | if (offset_in_page(ret)) { |
| 960 | vm_unacct_memory(charged); |
| 961 | locked = false; |
| 962 | } |
| 963 | if (downgraded) |
| 964 | mmap_read_unlock(current->mm); |
| 965 | else |
| 966 | mmap_write_unlock(current->mm); |
| 967 | if (locked && new_len > old_len) |
| 968 | mm_populate(new_addr + old_len, new_len - old_len); |
| 969 | userfaultfd_unmap_complete(mm, &uf_unmap_early); |
| 970 | mremap_userfaultfd_complete(&uf, addr, ret, old_len); |
| 971 | userfaultfd_unmap_complete(mm, &uf_unmap); |
| 972 | return ret; |
| 973 | } |