| 1 | |
| 2 | |
| 3 | * mm/mremap.c
|
| 4 | *
|
| 5 | * (C) Copyright 1996 Linus Torvalds
|
| 6 | *
|
| 7 | * Address space accounting code <[email protected]>
|
| 8 | * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
|
| 9 |
|
| 10 |
|
| 11 | #include <linux/mm.h>
|
| 12 | #include <linux/mm_inline.h>
|
| 13 | #include <linux/hugetlb.h>
|
| 14 | #include <linux/shm.h>
|
| 15 | #include <linux/ksm.h>
|
| 16 | #include <linux/mman.h>
|
| 17 | #include <linux/swap.h>
|
| 18 | #include <linux/capability.h>
|
| 19 | #include <linux/fs.h>
|
| 20 | #include <linux/leafops.h>
|
| 21 | #include <linux/highmem.h>
|
| 22 | #include <linux/security.h>
|
| 23 | #include <linux/syscalls.h>
|
| 24 | #include <linux/mmu_notifier.h>
|
| 25 | #include <linux/uaccess.h>
|
| 26 | #include <linux/userfaultfd_k.h>
|
| 27 | #include <linux/mempolicy.h>
|
| 28 | #include <linux/pgalloc.h>
|
| 29 |
|
| 30 | #include <asm/cacheflush.h>
|
| 31 | #include <asm/tlb.h>
|
| 32 |
|
| 33 | #include "internal.h"
|
| 34 |
|
| 35 |
|
| 36 | enum mremap_type {
|
| 37 | MREMAP_INVALID,
|
| 38 | MREMAP_NO_RESIZE,
|
| 39 | MREMAP_SHRINK,
|
| 40 | MREMAP_EXPAND,
|
| 41 | };
|
| 42 |
|
| 43 | |
| 44 | * Describes a VMA mremap() operation and is threaded throughout it.
|
| 45 | *
|
| 46 | * Any of the fields may be mutated by the operation, however these values will
|
| 47 | * always accurately reflect the remap (for instance, we may adjust lengths and
|
| 48 | * delta to account for hugetlb alignment).
|
| 49 |
|
| 50 | struct vma_remap_struct {
|
| 51 |
|
| 52 | unsigned long addr;
|
| 53 | unsigned long old_len;
|
| 54 | unsigned long new_len;
|
| 55 | const unsigned long flags;
|
| 56 | unsigned long new_addr;
|
| 57 |
|
| 58 |
|
| 59 | struct vm_userfaultfd_ctx *uf;
|
| 60 | struct list_head *uf_unmap_early;
|
| 61 | struct list_head *uf_unmap;
|
| 62 |
|
| 63 |
|
| 64 | struct vm_area_struct *vma;
|
| 65 |
|
| 66 |
|
| 67 | unsigned long delta;
|
| 68 | bool populate_expand;
|
| 69 | enum mremap_type remap_type;
|
| 70 | bool mmap_locked;
|
| 71 | unsigned long charged;
|
| 72 | bool vmi_needs_invalidate;
|
| 73 | };
|
| 74 |
|
| 75 | static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr)
|
| 76 | {
|
| 77 | pgd_t *pgd;
|
| 78 | p4d_t *p4d;
|
| 79 | pud_t *pud;
|
| 80 |
|
| 81 | pgd = pgd_offset(mm, addr);
|
| 82 | if (pgd_none_or_clear_bad(pgd))
|
| 83 | return NULL;
|
| 84 |
|
| 85 | p4d = p4d_offset(pgd, addr);
|
| 86 | if (p4d_none_or_clear_bad(p4d))
|
| 87 | return NULL;
|
| 88 |
|
| 89 | pud = pud_offset(p4d, addr);
|
| 90 | if (pud_none_or_clear_bad(pud))
|
| 91 | return NULL;
|
| 92 |
|
| 93 | return pud;
|
| 94 | }
|
| 95 |
|
| 96 | static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
|
| 97 | {
|
| 98 | pud_t *pud;
|
| 99 | pmd_t *pmd;
|
| 100 |
|
| 101 | pud = get_old_pud(mm, addr);
|
| 102 | if (!pud)
|
| 103 | return NULL;
|
| 104 |
|
| 105 | pmd = pmd_offset(pud, addr);
|
| 106 | if (pmd_none(*pmd))
|
| 107 | return NULL;
|
| 108 |
|
| 109 | return pmd;
|
| 110 | }
|
| 111 |
|
| 112 | static pud_t *alloc_new_pud(struct mm_struct *mm, unsigned long addr)
|
| 113 | {
|
| 114 | pgd_t *pgd;
|
| 115 | p4d_t *p4d;
|
| 116 |
|
| 117 | pgd = pgd_offset(mm, addr);
|
| 118 | p4d = p4d_alloc(mm, pgd, addr);
|
| 119 | if (!p4d)
|
| 120 | return NULL;
|
| 121 |
|
| 122 | return pud_alloc(mm, p4d, addr);
|
| 123 | }
|
| 124 |
|
| 125 | static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
|
| 126 | {
|
| 127 | pud_t *pud;
|
| 128 | pmd_t *pmd;
|
| 129 |
|
| 130 | pud = alloc_new_pud(mm, addr);
|
| 131 | if (!pud)
|
| 132 | return NULL;
|
| 133 |
|
| 134 | pmd = pmd_alloc(mm, pud, addr);
|
| 135 | if (!pmd)
|
| 136 | return NULL;
|
| 137 |
|
| 138 | VM_BUG_ON(pmd_trans_huge(*pmd));
|
| 139 |
|
| 140 | return pmd;
|
| 141 | }
|
| 142 |
|
| 143 | static void take_rmap_locks(struct vm_area_struct *vma)
|
| 144 | {
|
| 145 | if (vma->vm_file)
|
| 146 | i_mmap_lock_write(vma->vm_file->f_mapping);
|
| 147 | if (vma->anon_vma)
|
| 148 | anon_vma_lock_write(vma->anon_vma);
|
| 149 | }
|
| 150 |
|
| 151 | static void drop_rmap_locks(struct vm_area_struct *vma)
|
| 152 | {
|
| 153 | if (vma->anon_vma)
|
| 154 | anon_vma_unlock_write(vma->anon_vma);
|
| 155 | if (vma->vm_file)
|
| 156 | i_mmap_unlock_write(vma->vm_file->f_mapping);
|
| 157 | }
|
| 158 |
|
| 159 | static pte_t move_soft_dirty_pte(pte_t pte)
|
| 160 | {
|
| 161 | if (pte_none(pte))
|
| 162 | return pte;
|
| 163 |
|
| 164 | |
| 165 | * Set soft dirty bit so we can notice
|
| 166 | * in userspace the ptes were moved.
|
| 167 |
|
| 168 | if (pgtable_supports_soft_dirty()) {
|
| 169 | if (pte_present(pte))
|
| 170 | pte = pte_mksoft_dirty(pte);
|
| 171 | else
|
| 172 | pte = pte_swp_mksoft_dirty(pte);
|
| 173 | }
|
| 174 |
|
| 175 | return pte;
|
| 176 | }
|
| 177 |
|
| 178 | static int mremap_folio_pte_batch(struct vm_area_struct *vma, unsigned long addr,
|
| 179 | pte_t *ptep, pte_t pte, int max_nr)
|
| 180 | {
|
| 181 | struct folio *folio;
|
| 182 |
|
| 183 | if (max_nr == 1)
|
| 184 | return 1;
|
| 185 |
|
| 186 |
|
| 187 | if (pte_batch_hint(ptep, pte) == 1)
|
| 188 | return 1;
|
| 189 |
|
| 190 | folio = vm_normal_folio(vma, addr, pte);
|
| 191 | if (!folio || !folio_test_large(folio))
|
| 192 | return 1;
|
| 193 |
|
| 194 | return folio_pte_batch_flags(folio, NULL, ptep, &pte, max_nr, FPB_RESPECT_WRITE);
|
| 195 | }
|
| 196 |
|
| 197 | static int move_ptes(struct pagetable_move_control *pmc,
|
| 198 | unsigned long extent, pmd_t *old_pmd, pmd_t *new_pmd)
|
| 199 | {
|
| 200 | struct vm_area_struct *vma = pmc->old;
|
| 201 | bool need_clear_uffd_wp = vma_has_uffd_without_event_remap(vma);
|
| 202 | struct mm_struct *mm = vma->vm_mm;
|
| 203 | pte_t *old_ptep, *new_ptep;
|
| 204 | pte_t old_pte, pte;
|
| 205 | pmd_t dummy_pmdval;
|
| 206 | spinlock_t *old_ptl, *new_ptl;
|
| 207 | bool force_flush = false;
|
| 208 | unsigned long old_addr = pmc->old_addr;
|
| 209 | unsigned long new_addr = pmc->new_addr;
|
| 210 | unsigned long old_end = old_addr + extent;
|
| 211 | unsigned long len = old_end - old_addr;
|
| 212 | int max_nr_ptes;
|
| 213 | int nr_ptes;
|
| 214 | int err = 0;
|
| 215 |
|
| 216 | |
| 217 | * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
|
| 218 | * locks to ensure that rmap will always observe either the old or the
|
| 219 | * new ptes. This is the easiest way to avoid races with
|
| 220 | * truncate_pagecache(), page migration, etc...
|
| 221 | *
|
| 222 | * When need_rmap_locks is false, we use other ways to avoid
|
| 223 | * such races:
|
| 224 | *
|
| 225 | * - During exec() shift_arg_pages(), we use a specially tagged vma
|
| 226 | * which rmap call sites look for using vma_is_temporary_stack().
|
| 227 | *
|
| 228 | * - During mremap(), new_vma is often known to be placed after vma
|
| 229 | * in rmap traversal order. This ensures rmap will always observe
|
| 230 | * either the old pte, or the new pte, or both (the page table locks
|
| 231 | * serialize access to individual ptes, but only rmap traversal
|
| 232 | * order guarantees that we won't miss both the old and new ptes).
|
| 233 |
|
| 234 | if (pmc->need_rmap_locks)
|
| 235 | take_rmap_locks(vma);
|
| 236 |
|
| 237 | |
| 238 | * We don't have to worry about the ordering of src and dst
|
| 239 | * pte locks because exclusive mmap_lock prevents deadlock.
|
| 240 |
|
| 241 | old_ptep = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
|
| 242 | if (!old_ptep) {
|
| 243 | err = -EAGAIN;
|
| 244 | goto out;
|
| 245 | }
|
| 246 | |
| 247 | * Now new_pte is none, so collapse_scan_file() path can not find
|
| 248 | * this by traversing file->f_mapping, so there is no concurrency with
|
| 249 | * retract_page_tables(). In addition, we already hold the exclusive
|
| 250 | * mmap_lock, so this new_pte page is stable, so there is no need to get
|
| 251 | * pmdval and do pmd_same() check.
|
| 252 |
|
| 253 | new_ptep = pte_offset_map_rw_nolock(mm, new_pmd, new_addr, &dummy_pmdval,
|
| 254 | &new_ptl);
|
| 255 | if (!new_ptep) {
|
| 256 | pte_unmap_unlock(old_ptep, old_ptl);
|
| 257 | err = -EAGAIN;
|
| 258 | goto out;
|
| 259 | }
|
| 260 | if (new_ptl != old_ptl)
|
| 261 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
|
| 262 | flush_tlb_batched_pending(vma->vm_mm);
|
| 263 | lazy_mmu_mode_enable();
|
| 264 |
|
| 265 | for (; old_addr < old_end; old_ptep += nr_ptes, old_addr += nr_ptes * PAGE_SIZE,
|
| 266 | new_ptep += nr_ptes, new_addr += nr_ptes * PAGE_SIZE) {
|
| 267 | VM_WARN_ON_ONCE(!pte_none(ptep_get(new_ptep)));
|
| 268 |
|
| 269 | nr_ptes = 1;
|
| 270 | max_nr_ptes = (old_end - old_addr) >> PAGE_SHIFT;
|
| 271 | old_pte = ptep_get(old_ptep);
|
| 272 | if (pte_none(old_pte))
|
| 273 | continue;
|
| 274 |
|
| 275 | |
| 276 | * If we are remapping a valid PTE, make sure
|
| 277 | * to flush TLB before we drop the PTL for the
|
| 278 | * PTE.
|
| 279 | *
|
| 280 | * NOTE! Both old and new PTL matter: the old one
|
| 281 | * for racing with folio_mkclean(), the new one to
|
| 282 | * make sure the physical page stays valid until
|
| 283 | * the TLB entry for the old mapping has been
|
| 284 | * flushed.
|
| 285 |
|
| 286 | if (pte_present(old_pte)) {
|
| 287 | nr_ptes = mremap_folio_pte_batch(vma, old_addr, old_ptep,
|
| 288 | old_pte, max_nr_ptes);
|
| 289 | force_flush = true;
|
| 290 | }
|
| 291 | pte = get_and_clear_ptes(mm, old_addr, old_ptep, nr_ptes);
|
| 292 | pte = move_pte(pte, old_addr, new_addr);
|
| 293 | pte = move_soft_dirty_pte(pte);
|
| 294 |
|
| 295 | if (need_clear_uffd_wp && pte_is_uffd_wp_marker(pte))
|
| 296 | pte_clear(mm, new_addr, new_ptep);
|
| 297 | else {
|
| 298 | if (need_clear_uffd_wp) {
|
| 299 | if (pte_present(pte)) {
|
| 300 | |
| 301 | * See __copy_present_ptes(): normalise
|
| 302 | * RWP PTEs so the destination starts
|
| 303 | * accessible instead of taking a
|
| 304 | * numa-hinting fault on first access.
|
| 305 |
|
| 306 | if (userfaultfd_rwp(vma) && pte_uffd(pte))
|
| 307 | pte = pte_modify(pte, vma->vm_page_prot);
|
| 308 | pte = pte_clear_uffd(pte);
|
| 309 | } else {
|
| 310 | pte = pte_swp_clear_uffd(pte);
|
| 311 | }
|
| 312 | }
|
| 313 | set_ptes(mm, new_addr, new_ptep, pte, nr_ptes);
|
| 314 | }
|
| 315 | }
|
| 316 |
|
| 317 | lazy_mmu_mode_disable();
|
| 318 | if (force_flush)
|
| 319 | flush_tlb_range(vma, old_end - len, old_end);
|
| 320 | if (new_ptl != old_ptl)
|
| 321 | spin_unlock(new_ptl);
|
| 322 | pte_unmap(new_ptep - 1);
|
| 323 | pte_unmap_unlock(old_ptep - 1, old_ptl);
|
| 324 | out:
|
| 325 | if (pmc->need_rmap_locks)
|
| 326 | drop_rmap_locks(vma);
|
| 327 | return err;
|
| 328 | }
|
| 329 |
|
| 330 | #ifndef arch_supports_page_table_move
|
| 331 | #define arch_supports_page_table_move arch_supports_page_table_move
|
| 332 | static inline bool arch_supports_page_table_move(void)
|
| 333 | {
|
| 334 | return IS_ENABLED(CONFIG_HAVE_MOVE_PMD) ||
|
| 335 | IS_ENABLED(CONFIG_HAVE_MOVE_PUD);
|
| 336 | }
|
| 337 | #endif
|
| 338 |
|
| 339 | static inline bool uffd_supports_page_table_move(struct pagetable_move_control *pmc)
|
| 340 | {
|
| 341 | |
| 342 | * If we are moving a VMA that has uffd-wp registered but with
|
| 343 | * remap events disabled (new VMA will not be registered with uffd), we
|
| 344 | * need to ensure that the uffd-wp state is cleared from all pgtables.
|
| 345 | * This means recursing into lower page tables in move_page_tables().
|
| 346 | *
|
| 347 | * We might get called with VMAs reversed when recovering from a
|
| 348 | * failed page table move. In that case, the
|
| 349 | * "old"-but-actually-"originally new" VMA during recovery will not have
|
| 350 | * a uffd context. Recursing into lower page tables during the original
|
| 351 | * move but not during the recovery move will cause trouble, because we
|
| 352 | * run into already-existing page tables. So check both VMAs.
|
| 353 |
|
| 354 | return !vma_has_uffd_without_event_remap(pmc->old) &&
|
| 355 | !vma_has_uffd_without_event_remap(pmc->new);
|
| 356 | }
|
| 357 |
|
| 358 | #ifdef CONFIG_HAVE_MOVE_PMD
|
| 359 | static bool move_normal_pmd(struct pagetable_move_control *pmc,
|
| 360 | pmd_t *old_pmd, pmd_t *new_pmd)
|
| 361 | {
|
| 362 | spinlock_t *old_ptl, *new_ptl;
|
| 363 | struct vm_area_struct *vma = pmc->old;
|
| 364 | struct mm_struct *mm = vma->vm_mm;
|
| 365 | bool res = false;
|
| 366 | pmd_t pmd;
|
| 367 |
|
| 368 | if (!arch_supports_page_table_move())
|
| 369 | return false;
|
| 370 | if (!uffd_supports_page_table_move(pmc))
|
| 371 | return false;
|
| 372 | |
| 373 | * The destination pmd shouldn't be established, free_pgtables()
|
| 374 | * should have released it.
|
| 375 | *
|
| 376 | * However, there's a case during execve() where we use mremap
|
| 377 | * to move the initial stack, and in that case the target area
|
| 378 | * may overlap the source area (always moving down).
|
| 379 | *
|
| 380 | * If everything is PMD-aligned, that works fine, as moving
|
| 381 | * each pmd down will clear the source pmd. But if we first
|
| 382 | * have a few 4kB-only pages that get moved down, and then
|
| 383 | * hit the "now the rest is PMD-aligned, let's do everything
|
| 384 | * one pmd at a time", we will still have the old (now empty
|
| 385 | * of any 4kB pages, but still there) PMD in the page table
|
| 386 | * tree.
|
| 387 | *
|
| 388 | * Warn on it once - because we really should try to figure
|
| 389 | * out how to do this better - but then say "I won't move
|
| 390 | * this pmd".
|
| 391 | *
|
| 392 | * One alternative might be to just unmap the target pmd at
|
| 393 | * this point, and verify that it really is empty. We'll see.
|
| 394 |
|
| 395 | if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
|
| 396 | return false;
|
| 397 |
|
| 398 | |
| 399 | * We don't have to worry about the ordering of src and dst
|
| 400 | * ptlocks because exclusive mmap_lock prevents deadlock.
|
| 401 |
|
| 402 | old_ptl = pmd_lock(mm, old_pmd);
|
| 403 | new_ptl = pmd_lockptr(mm, new_pmd);
|
| 404 | if (new_ptl != old_ptl)
|
| 405 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
|
| 406 |
|
| 407 | pmd = *old_pmd;
|
| 408 |
|
| 409 |
|
| 410 | if (unlikely(!pmd_present(pmd) || pmd_leaf(pmd)))
|
| 411 | goto out_unlock;
|
| 412 |
|
| 413 | pmd_clear(old_pmd);
|
| 414 | res = true;
|
| 415 |
|
| 416 | VM_BUG_ON(!pmd_none(*new_pmd));
|
| 417 |
|
| 418 | pmd_populate(mm, new_pmd, pmd_pgtable(pmd));
|
| 419 | flush_tlb_range(vma, pmc->old_addr, pmc->old_addr + PMD_SIZE);
|
| 420 | out_unlock:
|
| 421 | if (new_ptl != old_ptl)
|
| 422 | spin_unlock(new_ptl);
|
| 423 | spin_unlock(old_ptl);
|
| 424 |
|
| 425 | return res;
|
| 426 | }
|
| 427 | #else
|
| 428 | static inline bool move_normal_pmd(struct pagetable_move_control *pmc,
|
| 429 | pmd_t *old_pmd, pmd_t *new_pmd)
|
| 430 | {
|
| 431 | return false;
|
| 432 | }
|
| 433 | #endif
|
| 434 |
|
| 435 | #if CONFIG_PGTABLE_LEVELS > 2 && defined(CONFIG_HAVE_MOVE_PUD)
|
| 436 | static bool move_normal_pud(struct pagetable_move_control *pmc,
|
| 437 | pud_t *old_pud, pud_t *new_pud)
|
| 438 | {
|
| 439 | spinlock_t *old_ptl, *new_ptl;
|
| 440 | struct vm_area_struct *vma = pmc->old;
|
| 441 | struct mm_struct *mm = vma->vm_mm;
|
| 442 | pud_t pud;
|
| 443 |
|
| 444 | if (!arch_supports_page_table_move())
|
| 445 | return false;
|
| 446 | if (!uffd_supports_page_table_move(pmc))
|
| 447 | return false;
|
| 448 | |
| 449 | * The destination pud shouldn't be established, free_pgtables()
|
| 450 | * should have released it.
|
| 451 |
|
| 452 | if (WARN_ON_ONCE(!pud_none(*new_pud)))
|
| 453 | return false;
|
| 454 |
|
| 455 | |
| 456 | * We don't have to worry about the ordering of src and dst
|
| 457 | * ptlocks because exclusive mmap_lock prevents deadlock.
|
| 458 |
|
| 459 | old_ptl = pud_lock(mm, old_pud);
|
| 460 | new_ptl = pud_lockptr(mm, new_pud);
|
| 461 | if (new_ptl != old_ptl)
|
| 462 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
|
| 463 |
|
| 464 |
|
| 465 | pud = *old_pud;
|
| 466 | pud_clear(old_pud);
|
| 467 |
|
| 468 | VM_BUG_ON(!pud_none(*new_pud));
|
| 469 |
|
| 470 | pud_populate(mm, new_pud, pud_pgtable(pud));
|
| 471 | flush_tlb_range(vma, pmc->old_addr, pmc->old_addr + PUD_SIZE);
|
| 472 | if (new_ptl != old_ptl)
|
| 473 | spin_unlock(new_ptl);
|
| 474 | spin_unlock(old_ptl);
|
| 475 |
|
| 476 | return true;
|
| 477 | }
|
| 478 | #else
|
| 479 | static inline bool move_normal_pud(struct pagetable_move_control *pmc,
|
| 480 | pud_t *old_pud, pud_t *new_pud)
|
| 481 | {
|
| 482 | return false;
|
| 483 | }
|
| 484 | #endif
|
| 485 |
|
| 486 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
|
| 487 | static bool move_huge_pud(struct pagetable_move_control *pmc,
|
| 488 | pud_t *old_pud, pud_t *new_pud)
|
| 489 | {
|
| 490 | spinlock_t *old_ptl, *new_ptl;
|
| 491 | struct vm_area_struct *vma = pmc->old;
|
| 492 | struct mm_struct *mm = vma->vm_mm;
|
| 493 | pud_t pud;
|
| 494 |
|
| 495 | |
| 496 | * The destination pud shouldn't be established, free_pgtables()
|
| 497 | * should have released it.
|
| 498 |
|
| 499 | if (WARN_ON_ONCE(!pud_none(*new_pud)))
|
| 500 | return false;
|
| 501 |
|
| 502 | |
| 503 | * We don't have to worry about the ordering of src and dst
|
| 504 | * ptlocks because exclusive mmap_lock prevents deadlock.
|
| 505 |
|
| 506 | old_ptl = pud_lock(mm, old_pud);
|
| 507 | new_ptl = pud_lockptr(mm, new_pud);
|
| 508 | if (new_ptl != old_ptl)
|
| 509 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
|
| 510 |
|
| 511 |
|
| 512 | pud = *old_pud;
|
| 513 | pud_clear(old_pud);
|
| 514 |
|
| 515 | VM_BUG_ON(!pud_none(*new_pud));
|
| 516 |
|
| 517 |
|
| 518 |
|
| 519 | set_pud_at(mm, pmc->new_addr, new_pud, pud);
|
| 520 | flush_pud_tlb_range(vma, pmc->old_addr, pmc->old_addr + HPAGE_PUD_SIZE);
|
| 521 | if (new_ptl != old_ptl)
|
| 522 | spin_unlock(new_ptl);
|
| 523 | spin_unlock(old_ptl);
|
| 524 |
|
| 525 | return true;
|
| 526 | }
|
| 527 | #else
|
| 528 | static bool move_huge_pud(struct pagetable_move_control *pmc,
|
| 529 | pud_t *old_pud, pud_t *new_pud)
|
| 530 |
|
| 531 | {
|
| 532 | WARN_ON_ONCE(1);
|
| 533 | return false;
|
| 534 |
|
| 535 | }
|
| 536 | #endif
|
| 537 |
|
| 538 | enum pgt_entry {
|
| 539 | NORMAL_PMD,
|
| 540 | HPAGE_PMD,
|
| 541 | NORMAL_PUD,
|
| 542 | HPAGE_PUD,
|
| 543 | };
|
| 544 |
|
| 545 | |
| 546 | * Returns an extent of the corresponding size for the pgt_entry specified if
|
| 547 | * valid. Else returns a smaller extent bounded by the end of the source and
|
| 548 | * destination pgt_entry.
|
| 549 |
|
| 550 | static __always_inline unsigned long get_extent(enum pgt_entry entry,
|
| 551 | struct pagetable_move_control *pmc)
|
| 552 | {
|
| 553 | unsigned long next, extent, mask, size;
|
| 554 | unsigned long old_addr = pmc->old_addr;
|
| 555 | unsigned long old_end = pmc->old_end;
|
| 556 | unsigned long new_addr = pmc->new_addr;
|
| 557 |
|
| 558 | switch (entry) {
|
| 559 | case HPAGE_PMD:
|
| 560 | case NORMAL_PMD:
|
| 561 | mask = PMD_MASK;
|
| 562 | size = PMD_SIZE;
|
| 563 | break;
|
| 564 | case HPAGE_PUD:
|
| 565 | case NORMAL_PUD:
|
| 566 | mask = PUD_MASK;
|
| 567 | size = PUD_SIZE;
|
| 568 | break;
|
| 569 | default:
|
| 570 | BUILD_BUG();
|
| 571 | break;
|
| 572 | }
|
| 573 |
|
| 574 | next = (old_addr + size) & mask;
|
| 575 |
|
| 576 | extent = next - old_addr;
|
| 577 | if (extent > old_end - old_addr)
|
| 578 | extent = old_end - old_addr;
|
| 579 | next = (new_addr + size) & mask;
|
| 580 | if (extent > next - new_addr)
|
| 581 | extent = next - new_addr;
|
| 582 | return extent;
|
| 583 | }
|
| 584 |
|
| 585 | |
| 586 | * Should move_pgt_entry() acquire the rmap locks? This is either expressed in
|
| 587 | * the PMC, or overridden in the case of normal, larger page tables.
|
| 588 |
|
| 589 | static bool should_take_rmap_locks(struct pagetable_move_control *pmc,
|
| 590 | enum pgt_entry entry)
|
| 591 | {
|
| 592 | switch (entry) {
|
| 593 | case NORMAL_PMD:
|
| 594 | case NORMAL_PUD:
|
| 595 | return true;
|
| 596 | default:
|
| 597 | return pmc->need_rmap_locks;
|
| 598 | }
|
| 599 | }
|
| 600 |
|
| 601 | |
| 602 | * Attempts to speedup the move by moving entry at the level corresponding to
|
| 603 | * pgt_entry. Returns true if the move was successful, else false.
|
| 604 |
|
| 605 | static bool move_pgt_entry(struct pagetable_move_control *pmc,
|
| 606 | enum pgt_entry entry, void *old_entry, void *new_entry)
|
| 607 | {
|
| 608 | bool moved = false;
|
| 609 | bool need_rmap_locks = should_take_rmap_locks(pmc, entry);
|
| 610 |
|
| 611 |
|
| 612 | if (need_rmap_locks)
|
| 613 | take_rmap_locks(pmc->old);
|
| 614 |
|
| 615 | switch (entry) {
|
| 616 | case NORMAL_PMD:
|
| 617 | moved = move_normal_pmd(pmc, old_entry, new_entry);
|
| 618 | break;
|
| 619 | case NORMAL_PUD:
|
| 620 | moved = move_normal_pud(pmc, old_entry, new_entry);
|
| 621 | break;
|
| 622 | case HPAGE_PMD:
|
| 623 | moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
|
| 624 | move_huge_pmd(pmc->old, pmc->old_addr, pmc->new_addr, old_entry,
|
| 625 | new_entry);
|
| 626 | break;
|
| 627 | case HPAGE_PUD:
|
| 628 | moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
|
| 629 | move_huge_pud(pmc, old_entry, new_entry);
|
| 630 | break;
|
| 631 |
|
| 632 | default:
|
| 633 | WARN_ON_ONCE(1);
|
| 634 | break;
|
| 635 | }
|
| 636 |
|
| 637 | if (need_rmap_locks)
|
| 638 | drop_rmap_locks(pmc->old);
|
| 639 |
|
| 640 | return moved;
|
| 641 | }
|
| 642 |
|
| 643 | |
| 644 | * A helper to check if aligning down is OK. The aligned address should fall
|
| 645 | * on *no mapping*. For the stack moving down, that's a special move within
|
| 646 | * the VMA that is created to span the source and destination of the move,
|
| 647 | * so we make an exception for it.
|
| 648 |
|
| 649 | static bool can_align_down(struct pagetable_move_control *pmc,
|
| 650 | struct vm_area_struct *vma, unsigned long addr_to_align,
|
| 651 | unsigned long mask)
|
| 652 | {
|
| 653 | unsigned long addr_masked = addr_to_align & mask;
|
| 654 |
|
| 655 | |
| 656 | * If @addr_to_align of either source or destination is not the beginning
|
| 657 | * of the corresponding VMA, we can't align down or we will destroy part
|
| 658 | * of the current mapping.
|
| 659 |
|
| 660 | if (!pmc->for_stack && vma->vm_start != addr_to_align)
|
| 661 | return false;
|
| 662 |
|
| 663 |
|
| 664 | if (pmc->for_stack && addr_masked >= vma->vm_start)
|
| 665 | return true;
|
| 666 |
|
| 667 | |
| 668 | * Make sure the realignment doesn't cause the address to fall on an
|
| 669 | * existing mapping.
|
| 670 |
|
| 671 | return find_vma_intersection(vma->vm_mm, addr_masked, vma->vm_start) == NULL;
|
| 672 | }
|
| 673 |
|
| 674 | |
| 675 | * Determine if are in fact able to realign for efficiency to a higher page
|
| 676 | * table boundary.
|
| 677 |
|
| 678 | static bool can_realign_addr(struct pagetable_move_control *pmc,
|
| 679 | unsigned long pagetable_mask)
|
| 680 | {
|
| 681 | unsigned long align_mask = ~pagetable_mask;
|
| 682 | unsigned long old_align = pmc->old_addr & align_mask;
|
| 683 | unsigned long new_align = pmc->new_addr & align_mask;
|
| 684 | unsigned long pagetable_size = align_mask + 1;
|
| 685 | unsigned long old_align_next = pagetable_size - old_align;
|
| 686 |
|
| 687 | |
| 688 | * We don't want to have to go hunting for VMAs from the end of the old
|
| 689 | * VMA to the next page table boundary, also we want to make sure the
|
| 690 | * operation is worthwhile.
|
| 691 | *
|
| 692 | * So ensure that we only perform this realignment if the end of the
|
| 693 | * range being copied reaches or crosses the page table boundary.
|
| 694 | *
|
| 695 | * boundary boundary
|
| 696 | * .<- old_align -> .
|
| 697 | * . |----------------.-----------|
|
| 698 | * . | vma . |
|
| 699 | * . |----------------.-----------|
|
| 700 | * . <----------------.----------->
|
| 701 | * . len_in
|
| 702 | * <------------------------------->
|
| 703 | * . pagetable_size .
|
| 704 | * . <---------------->
|
| 705 | * . old_align_next .
|
| 706 |
|
| 707 | if (pmc->len_in < old_align_next)
|
| 708 | return false;
|
| 709 |
|
| 710 |
|
| 711 | if (old_align == 0)
|
| 712 | return false;
|
| 713 |
|
| 714 |
|
| 715 | if (old_align != new_align)
|
| 716 | return false;
|
| 717 |
|
| 718 |
|
| 719 | if (!can_align_down(pmc, pmc->old, pmc->old_addr, pagetable_mask) ||
|
| 720 | !can_align_down(pmc, pmc->new, pmc->new_addr, pagetable_mask))
|
| 721 | return false;
|
| 722 |
|
| 723 | return true;
|
| 724 | }
|
| 725 |
|
| 726 | |
| 727 | * Opportunistically realign to specified boundary for faster copy.
|
| 728 | *
|
| 729 | * Consider an mremap() of a VMA with page table boundaries as below, and no
|
| 730 | * preceding VMAs from the lower page table boundary to the start of the VMA,
|
| 731 | * with the end of the range reaching or crossing the page table boundary.
|
| 732 | *
|
| 733 | * boundary boundary
|
| 734 | * . |----------------.-----------|
|
| 735 | * . | vma . |
|
| 736 | * . |----------------.-----------|
|
| 737 | * . pmc->old_addr . pmc->old_end
|
| 738 | * . <---------------------------->
|
| 739 | * . move these page tables
|
| 740 | *
|
| 741 | * If we proceed with moving page tables in this scenario, we will have a lot of
|
| 742 | * work to do traversing old page tables and establishing new ones in the
|
| 743 | * destination across multiple lower level page tables.
|
| 744 | *
|
| 745 | * The idea here is simply to align pmc->old_addr, pmc->new_addr down to the
|
| 746 | * page table boundary, so we can simply copy a single page table entry for the
|
| 747 | * aligned portion of the VMA instead:
|
| 748 | *
|
| 749 | * boundary boundary
|
| 750 | * . |----------------.-----------|
|
| 751 | * . | vma . |
|
| 752 | * . |----------------.-----------|
|
| 753 | * pmc->old_addr . pmc->old_end
|
| 754 | * <------------------------------------------->
|
| 755 | * . move these page tables
|
| 756 |
|
| 757 | static void try_realign_addr(struct pagetable_move_control *pmc,
|
| 758 | unsigned long pagetable_mask)
|
| 759 | {
|
| 760 |
|
| 761 | if (!can_realign_addr(pmc, pagetable_mask))
|
| 762 | return;
|
| 763 |
|
| 764 | |
| 765 | * Simply align to page table boundaries. Note that we do NOT update the
|
| 766 | * pmc->old_end value, and since the move_page_tables() operation spans
|
| 767 | * from [old_addr, old_end) (offsetting new_addr as it is performed),
|
| 768 | * this simply changes the start of the copy, not the end.
|
| 769 |
|
| 770 | pmc->old_addr &= pagetable_mask;
|
| 771 | pmc->new_addr &= pagetable_mask;
|
| 772 | }
|
| 773 |
|
| 774 |
|
| 775 | static bool pmc_done(struct pagetable_move_control *pmc)
|
| 776 | {
|
| 777 | return pmc->old_addr >= pmc->old_end;
|
| 778 | }
|
| 779 |
|
| 780 |
|
| 781 | static void pmc_next(struct pagetable_move_control *pmc, unsigned long extent)
|
| 782 | {
|
| 783 | pmc->old_addr += extent;
|
| 784 | pmc->new_addr += extent;
|
| 785 | }
|
| 786 |
|
| 787 | |
| 788 | * Determine how many bytes in the specified input range have had their page
|
| 789 | * tables moved so far.
|
| 790 |
|
| 791 | static unsigned long pmc_progress(struct pagetable_move_control *pmc)
|
| 792 | {
|
| 793 | unsigned long orig_old_addr = pmc->old_end - pmc->len_in;
|
| 794 | unsigned long old_addr = pmc->old_addr;
|
| 795 |
|
| 796 | |
| 797 | * Prevent negative return values when {old,new}_addr was realigned but
|
| 798 | * we broke out of the loop in move_page_tables() for the first PMD
|
| 799 | * itself.
|
| 800 |
|
| 801 | return old_addr < orig_old_addr ? 0 : old_addr - orig_old_addr;
|
| 802 | }
|
| 803 |
|
| 804 | unsigned long move_page_tables(struct pagetable_move_control *pmc)
|
| 805 | {
|
| 806 | unsigned long extent;
|
| 807 | struct mmu_notifier_range range;
|
| 808 | pmd_t *old_pmd, *new_pmd;
|
| 809 | pud_t *old_pud, *new_pud;
|
| 810 | struct mm_struct *mm = pmc->old->vm_mm;
|
| 811 |
|
| 812 | if (!pmc->len_in)
|
| 813 | return 0;
|
| 814 |
|
| 815 | if (is_vm_hugetlb_page(pmc->old))
|
| 816 | return move_hugetlb_page_tables(pmc->old, pmc->new, pmc->old_addr,
|
| 817 | pmc->new_addr, pmc->len_in);
|
| 818 |
|
| 819 | |
| 820 | * If possible, realign addresses to PMD boundary for faster copy.
|
| 821 | * Only realign if the mremap copying hits a PMD boundary.
|
| 822 |
|
| 823 | try_realign_addr(pmc, PMD_MASK);
|
| 824 |
|
| 825 | flush_cache_range(pmc->old, pmc->old_addr, pmc->old_end);
|
| 826 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, mm,
|
| 827 | pmc->old_addr, pmc->old_end);
|
| 828 | mmu_notifier_invalidate_range_start(&range);
|
| 829 |
|
| 830 | for (; !pmc_done(pmc); pmc_next(pmc, extent)) {
|
| 831 | cond_resched();
|
| 832 | |
| 833 | * If extent is PUD-sized try to speed up the move by moving at the
|
| 834 | * PUD level if possible.
|
| 835 |
|
| 836 | extent = get_extent(NORMAL_PUD, pmc);
|
| 837 |
|
| 838 | old_pud = get_old_pud(mm, pmc->old_addr);
|
| 839 | if (!old_pud)
|
| 840 | continue;
|
| 841 | new_pud = alloc_new_pud(mm, pmc->new_addr);
|
| 842 | if (!new_pud)
|
| 843 | break;
|
| 844 | if (pud_trans_huge(*old_pud)) {
|
| 845 | if (extent == HPAGE_PUD_SIZE) {
|
| 846 | move_pgt_entry(pmc, HPAGE_PUD, old_pud, new_pud);
|
| 847 |
|
| 848 | continue;
|
| 849 | }
|
| 850 | } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
|
| 851 | if (move_pgt_entry(pmc, NORMAL_PUD, old_pud, new_pud))
|
| 852 | continue;
|
| 853 | }
|
| 854 |
|
| 855 | extent = get_extent(NORMAL_PMD, pmc);
|
| 856 | old_pmd = get_old_pmd(mm, pmc->old_addr);
|
| 857 | if (!old_pmd)
|
| 858 | continue;
|
| 859 | new_pmd = alloc_new_pmd(mm, pmc->new_addr);
|
| 860 | if (!new_pmd)
|
| 861 | break;
|
| 862 | again:
|
| 863 | if (pmd_is_huge(*old_pmd)) {
|
| 864 | if (extent == HPAGE_PMD_SIZE &&
|
| 865 | move_pgt_entry(pmc, HPAGE_PMD, old_pmd, new_pmd))
|
| 866 | continue;
|
| 867 | split_huge_pmd(pmc->old, old_pmd, pmc->old_addr);
|
| 868 | } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
|
| 869 | extent == PMD_SIZE) {
|
| 870 | |
| 871 | * If the extent is PMD-sized, try to speed the move by
|
| 872 | * moving at the PMD level if possible.
|
| 873 |
|
| 874 | if (move_pgt_entry(pmc, NORMAL_PMD, old_pmd, new_pmd))
|
| 875 | continue;
|
| 876 | }
|
| 877 | if (pmd_none(*old_pmd))
|
| 878 | continue;
|
| 879 | if (pte_alloc(pmc->new->vm_mm, new_pmd))
|
| 880 | break;
|
| 881 | if (move_ptes(pmc, extent, old_pmd, new_pmd) < 0)
|
| 882 | goto again;
|
| 883 | }
|
| 884 |
|
| 885 | mmu_notifier_invalidate_range_end(&range);
|
| 886 |
|
| 887 | return pmc_progress(pmc);
|
| 888 | }
|
| 889 |
|
| 890 |
|
| 891 | static void vrm_set_delta(struct vma_remap_struct *vrm)
|
| 892 | {
|
| 893 | vrm->delta = abs_diff(vrm->old_len, vrm->new_len);
|
| 894 | }
|
| 895 |
|
| 896 |
|
| 897 | static enum mremap_type vrm_remap_type(struct vma_remap_struct *vrm)
|
| 898 | {
|
| 899 | if (vrm->delta == 0)
|
| 900 | return MREMAP_NO_RESIZE;
|
| 901 |
|
| 902 | if (vrm->old_len > vrm->new_len)
|
| 903 | return MREMAP_SHRINK;
|
| 904 |
|
| 905 | return MREMAP_EXPAND;
|
| 906 | }
|
| 907 |
|
| 908 | |
| 909 | * When moving a VMA to vrm->new_adr, does this result in the new and old VMAs
|
| 910 | * overlapping?
|
| 911 |
|
| 912 | static bool vrm_overlaps(struct vma_remap_struct *vrm)
|
| 913 | {
|
| 914 | unsigned long start_old = vrm->addr;
|
| 915 | unsigned long start_new = vrm->new_addr;
|
| 916 | unsigned long end_old = vrm->addr + vrm->old_len;
|
| 917 | unsigned long end_new = vrm->new_addr + vrm->new_len;
|
| 918 |
|
| 919 | |
| 920 | * start_old end_old
|
| 921 | * |-----------|
|
| 922 | * | |
|
| 923 | * |-----------|
|
| 924 | * |-------------|
|
| 925 | * | |
|
| 926 | * |-------------|
|
| 927 | * start_new end_new
|
| 928 |
|
| 929 | if (end_old > start_new && end_new > start_old)
|
| 930 | return true;
|
| 931 |
|
| 932 | return false;
|
| 933 | }
|
| 934 |
|
| 935 | |
| 936 | * Will a new address definitely be assigned? This either if the user specifies
|
| 937 | * it via MREMAP_FIXED, or if MREMAP_DONTUNMAP is used, indicating we will
|
| 938 | * always determine a target address.
|
| 939 |
|
| 940 | static bool vrm_implies_new_addr(struct vma_remap_struct *vrm)
|
| 941 | {
|
| 942 | return vrm->flags & (MREMAP_FIXED | MREMAP_DONTUNMAP);
|
| 943 | }
|
| 944 |
|
| 945 | |
| 946 | * Find an unmapped area for the requested vrm->new_addr.
|
| 947 | *
|
| 948 | * If MREMAP_FIXED then this is equivalent to a MAP_FIXED mmap() call. If only
|
| 949 | * MREMAP_DONTUNMAP is set, then this is equivalent to providing a hint to
|
| 950 | * mmap(), otherwise this is equivalent to mmap() specifying a NULL address.
|
| 951 | *
|
| 952 | * Returns 0 on success (with vrm->new_addr updated), or an error code upon
|
| 953 | * failure.
|
| 954 |
|
| 955 | static unsigned long vrm_set_new_addr(struct vma_remap_struct *vrm)
|
| 956 | {
|
| 957 | struct vm_area_struct *vma = vrm->vma;
|
| 958 | unsigned long map_flags = 0;
|
| 959 |
|
| 960 | const pgoff_t pgoff = linear_page_index(vma, vrm->addr);
|
| 961 | unsigned long new_addr = vrm_implies_new_addr(vrm) ? vrm->new_addr : 0;
|
| 962 | unsigned long res;
|
| 963 |
|
| 964 | if (vrm->flags & MREMAP_FIXED)
|
| 965 | map_flags |= MAP_FIXED;
|
| 966 | if (vma_test(vma, VMA_MAYSHARE_BIT))
|
| 967 | map_flags |= MAP_SHARED;
|
| 968 |
|
| 969 | res = get_unmapped_area(vma->vm_file, new_addr, vrm->new_len, pgoff,
|
| 970 | map_flags);
|
| 971 | if (IS_ERR_VALUE(res))
|
| 972 | return res;
|
| 973 |
|
| 974 | vrm->new_addr = res;
|
| 975 | return 0;
|
| 976 | }
|
| 977 |
|
| 978 | |
| 979 | * Keep track of pages which have been added to the memory mapping. If the VMA
|
| 980 | * is accounted, also check to see if there is sufficient memory.
|
| 981 | *
|
| 982 | * Returns true on success, false if insufficient memory to charge.
|
| 983 |
|
| 984 | static bool vrm_calc_charge(struct vma_remap_struct *vrm)
|
| 985 | {
|
| 986 | unsigned long charged;
|
| 987 |
|
| 988 | if (!vma_test(vrm->vma, VMA_ACCOUNT_BIT))
|
| 989 | return true;
|
| 990 |
|
| 991 | |
| 992 | * If we don't unmap the old mapping, then we account the entirety of
|
| 993 | * the length of the new one. Otherwise it's just the delta in size.
|
| 994 |
|
| 995 | if (vrm->flags & MREMAP_DONTUNMAP)
|
| 996 | charged = vrm->new_len >> PAGE_SHIFT;
|
| 997 | else
|
| 998 | charged = vrm->delta >> PAGE_SHIFT;
|
| 999 |
|
| 1000 |
|
| 1001 |
|
| 1002 | if (security_vm_enough_memory_mm(current->mm, charged))
|
| 1003 | return false;
|
| 1004 |
|
| 1005 | vrm->charged = charged;
|
| 1006 | return true;
|
| 1007 | }
|
| 1008 |
|
| 1009 | |
| 1010 | * an error has occurred so we will not be using vrm->charged memory. Unaccount
|
| 1011 | * this memory if the VMA is accounted.
|
| 1012 |
|
| 1013 | static void vrm_uncharge(struct vma_remap_struct *vrm)
|
| 1014 | {
|
| 1015 | if (!vma_test(vrm->vma, VMA_ACCOUNT_BIT))
|
| 1016 | return;
|
| 1017 |
|
| 1018 | vm_unacct_memory(vrm->charged);
|
| 1019 | vrm->charged = 0;
|
| 1020 | }
|
| 1021 |
|
| 1022 | |
| 1023 | * Update mm exec_vm, stack_vm, data_vm, and locked_vm fields as needed to
|
| 1024 | * account for 'bytes' memory used, and if locked, indicate this in the VRM so
|
| 1025 | * we can handle this correctly later.
|
| 1026 |
|
| 1027 | static void vrm_stat_account(struct vma_remap_struct *vrm,
|
| 1028 | unsigned long bytes)
|
| 1029 | {
|
| 1030 | unsigned long pages = bytes >> PAGE_SHIFT;
|
| 1031 | struct mm_struct *mm = current->mm;
|
| 1032 | struct vm_area_struct *vma = vrm->vma;
|
| 1033 |
|
| 1034 | vm_stat_account(mm, vma->vm_flags, pages);
|
| 1035 | if (vma_test(vma, VMA_LOCKED_BIT))
|
| 1036 | mm->locked_vm += pages;
|
| 1037 | }
|
| 1038 |
|
| 1039 | static bool __check_map_count_against_split(struct mm_struct *mm,
|
| 1040 | bool pre_split,
|
| 1041 | bool before_unmaps)
|
| 1042 | {
|
| 1043 | const int sys_map_count = get_sysctl_max_map_count();
|
| 1044 | int map_count = mm->map_count;
|
| 1045 |
|
| 1046 | mmap_assert_write_locked(mm);
|
| 1047 |
|
| 1048 | |
| 1049 | * At the point of shrinking the VMA, if new_len < old_len, we unmap
|
| 1050 | * thusly in the worst case:
|
| 1051 | *
|
| 1052 | * old_addr+old_len old_addr+old_len
|
| 1053 | * |---------------.----.---------| |---------------| |---------|
|
| 1054 | * | . . | -> | +1 | -1 | +1 |
|
| 1055 | * |---------------.----.---------| |---------------| |---------|
|
| 1056 | * old_addr+new_len old_addr+new_len
|
| 1057 | *
|
| 1058 | * At the point of removing the portion of an existing VMA to make space
|
| 1059 | * for the moved VMA if MREMAP_FIXED, we unmap thusly in the worst case:
|
| 1060 | *
|
| 1061 | * new_addr new_addr+new_len new_addr new_addr+new_len
|
| 1062 | * |----.---------------.---------| |----| |---------|
|
| 1063 | * | . . | -> | +1 | -1 | +1 |
|
| 1064 | * |----.---------------.---------| |----| |---------|
|
| 1065 | *
|
| 1066 | * Therefore, before we consider the move anything, we have to account
|
| 1067 | * for 2 additional VMAs possibly being created upon these unmappings.
|
| 1068 |
|
| 1069 | if (before_unmaps)
|
| 1070 | map_count += 2;
|
| 1071 |
|
| 1072 | |
| 1073 | * At the point of MOVING the VMA:
|
| 1074 | *
|
| 1075 | * We start by copying a VMA, which creates an additional VMA if no
|
| 1076 | * merge occurs, then if not MREMAP_DONTUNMAP, we unmap the source VMA.
|
| 1077 | * In the worst case we might then observe:
|
| 1078 | *
|
| 1079 | * new_addr new_addr+new_len new_addr new_addr+new_len
|
| 1080 | * |----| |---------| |----|---------------|---------|
|
| 1081 | * | | | | -> | | +1 | |
|
| 1082 | * |----| |---------| |----|---------------|---------|
|
| 1083 | *
|
| 1084 | * old_addr old_addr+old_len old_addr old_addr+old_len
|
| 1085 | * |----.---------------.---------| |----| |---------|
|
| 1086 | * | . . | -> | +1 | -1 | +1 |
|
| 1087 | * |----.---------------.---------| |----| |---------|
|
| 1088 | *
|
| 1089 | * Therefore we must check to ensure we have headroom of 2 additional
|
| 1090 | * VMAs.
|
| 1091 |
|
| 1092 | map_count += 2;
|
| 1093 |
|
| 1094 |
|
| 1095 | if (pre_split)
|
| 1096 | map_count++;
|
| 1097 |
|
| 1098 | return map_count <= sys_map_count;
|
| 1099 | }
|
| 1100 |
|
| 1101 | static bool needs_pre_split(struct vma_remap_struct *vrm)
|
| 1102 | {
|
| 1103 | |
| 1104 | * An MREMAP_DONTUNMAP of a mlock()'d VMA needs to unlock the
|
| 1105 | * source VMA, so split in this case.
|
| 1106 |
|
| 1107 | return (vrm->flags & MREMAP_DONTUNMAP) &&
|
| 1108 | vma_test(vrm->vma, VMA_LOCKED_BIT);
|
| 1109 | }
|
| 1110 |
|
| 1111 |
|
| 1112 | static bool check_map_count_against_split(struct vma_remap_struct *vrm)
|
| 1113 | {
|
| 1114 | return __check_map_count_against_split(current->mm,
|
| 1115 | needs_pre_split(vrm), false);
|
| 1116 | }
|
| 1117 |
|
| 1118 |
|
| 1119 | static bool check_map_count_against_split_early(struct vma_remap_struct *vrm)
|
| 1120 | {
|
| 1121 | return __check_map_count_against_split(current->mm,
|
| 1122 | vrm->flags & MREMAP_DONTUNMAP, true);
|
| 1123 | }
|
| 1124 |
|
| 1125 | |
| 1126 | * Perform checks and preparation before attempting to write a VMA prior to it
|
| 1127 | * being moved.
|
| 1128 |
|
| 1129 | static unsigned long prep_move_vma(struct vma_remap_struct *vrm)
|
| 1130 | {
|
| 1131 | unsigned long err = 0;
|
| 1132 | struct vm_area_struct *vma = vrm->vma;
|
| 1133 | unsigned long old_addr = vrm->addr;
|
| 1134 | unsigned long old_len = vrm->old_len;
|
| 1135 | vm_flags_t dummy = vma->vm_flags;
|
| 1136 | const bool split_before = vma->vm_start != old_addr;
|
| 1137 | const bool split_after = vma->vm_end != old_addr + old_len;
|
| 1138 |
|
| 1139 |
|
| 1140 | if (!check_map_count_against_split(vrm))
|
| 1141 | return -ENOMEM;
|
| 1142 |
|
| 1143 | if (vma->vm_ops && vma->vm_ops->may_split) {
|
| 1144 | if (split_before)
|
| 1145 | err = vma->vm_ops->may_split(vma, old_addr);
|
| 1146 | if (!err && split_after)
|
| 1147 | err = vma->vm_ops->may_split(vma, old_addr + old_len);
|
| 1148 | if (err)
|
| 1149 | return err;
|
| 1150 | }
|
| 1151 |
|
| 1152 | |
| 1153 | * Advise KSM to break any KSM pages in the area to be moved:
|
| 1154 | * it would be confusing if they were to turn up at the new
|
| 1155 | * location, where they happen to coincide with different KSM
|
| 1156 | * pages recently unmapped. But leave vma->vm_flags as it was,
|
| 1157 | * so KSM can come around to merge on vma and new_vma afterwards.
|
| 1158 |
|
| 1159 | err = ksm_madvise(vma, old_addr, old_addr + old_len,
|
| 1160 | MADV_UNMERGEABLE, &dummy);
|
| 1161 | if (err)
|
| 1162 | return err;
|
| 1163 |
|
| 1164 | |
| 1165 | * To account mlock()'d pages correctly in the MREMAP_DONTUNMAP
|
| 1166 | * case perform any split ahead of time for an mlock()'d VMA.
|
| 1167 |
|
| 1168 | if (needs_pre_split(vrm)) {
|
| 1169 | VMA_ITERATOR(vmi, vma->vm_mm, old_addr);
|
| 1170 |
|
| 1171 | if (split_before)
|
| 1172 | err = split_vma(&vmi, vma, old_addr, 1);
|
| 1173 | if (!err && split_after)
|
| 1174 | err = split_vma(&vmi, vma, old_addr + old_len, 0);
|
| 1175 | vrm->vmi_needs_invalidate = true;
|
| 1176 | return err;
|
| 1177 | }
|
| 1178 |
|
| 1179 | return 0;
|
| 1180 | }
|
| 1181 |
|
| 1182 | |
| 1183 | * Unmap source VMA for VMA move, turning it from a copy to a move, being
|
| 1184 | * careful to ensure we do not underflow memory account while doing so if an
|
| 1185 | * accountable move.
|
| 1186 | *
|
| 1187 | * This is best effort, if we fail to unmap then we simply try to correct
|
| 1188 | * accounting and exit.
|
| 1189 |
|
| 1190 | static void unmap_source_vma(struct vma_remap_struct *vrm)
|
| 1191 | {
|
| 1192 | struct mm_struct *mm = current->mm;
|
| 1193 | unsigned long addr = vrm->addr;
|
| 1194 | unsigned long len = vrm->old_len;
|
| 1195 | struct vm_area_struct *vma = vrm->vma;
|
| 1196 | VMA_ITERATOR(vmi, mm, addr);
|
| 1197 | int err;
|
| 1198 | unsigned long vm_start;
|
| 1199 | unsigned long vm_end;
|
| 1200 | |
| 1201 | * It might seem odd that we check for MREMAP_DONTUNMAP here, given this
|
| 1202 | * function implies that we unmap the original VMA, which seems
|
| 1203 | * contradictory.
|
| 1204 | *
|
| 1205 | * However, this occurs when this operation was attempted and an error
|
| 1206 | * arose, in which case we _do_ wish to unmap the _new_ VMA, which means
|
| 1207 | * we actually _do_ want it be unaccounted.
|
| 1208 |
|
| 1209 | bool accountable_move = vma_test(vma, VMA_ACCOUNT_BIT) &&
|
| 1210 | !(vrm->flags & MREMAP_DONTUNMAP);
|
| 1211 |
|
| 1212 | |
| 1213 | * So we perform a trick here to prevent incorrect accounting. Any merge
|
| 1214 | * or new VMA allocation performed in copy_vma() does not adjust
|
| 1215 | * accounting, it is expected that callers handle this.
|
| 1216 | *
|
| 1217 | * And indeed we already have, accounting appropriately in the case of
|
| 1218 | * both in vrm_charge().
|
| 1219 | *
|
| 1220 | * However, when we unmap the existing VMA (to effect the move), this
|
| 1221 | * code will, if the VMA has VM_ACCOUNT set, attempt to unaccount
|
| 1222 | * removed pages.
|
| 1223 | *
|
| 1224 | * To avoid this we temporarily clear this flag, reinstating on any
|
| 1225 | * portions of the original VMA that remain.
|
| 1226 |
|
| 1227 | if (accountable_move) {
|
| 1228 | vma_clear_flags(vma, VMA_ACCOUNT_BIT);
|
| 1229 |
|
| 1230 | vm_start = vma->vm_start;
|
| 1231 | vm_end = vma->vm_end;
|
| 1232 | }
|
| 1233 |
|
| 1234 | err = do_vmi_munmap(&vmi, mm, addr, len, vrm->uf_unmap, false);
|
| 1235 | vrm->vma = NULL;
|
| 1236 | vrm->vmi_needs_invalidate = true;
|
| 1237 | if (err) {
|
| 1238 |
|
| 1239 | vm_acct_memory(len >> PAGE_SHIFT);
|
| 1240 | return;
|
| 1241 | }
|
| 1242 |
|
| 1243 | |
| 1244 | * If we mremap() from a VMA like this:
|
| 1245 | *
|
| 1246 | * addr end
|
| 1247 | * | |
|
| 1248 | * v v
|
| 1249 | * |-------------|
|
| 1250 | * | |
|
| 1251 | * |-------------|
|
| 1252 | *
|
| 1253 | * Having cleared VMA_ACCOUNT_BIT from the whole VMA, after we unmap
|
| 1254 | * above we'll end up with:
|
| 1255 | *
|
| 1256 | * addr end
|
| 1257 | * | |
|
| 1258 | * v v
|
| 1259 | * |---| |---|
|
| 1260 | * | A | | B |
|
| 1261 | * |---| |---|
|
| 1262 | *
|
| 1263 | * The VMI is still pointing at addr, so vma_prev() will give us A, and
|
| 1264 | * a subsequent or lone vma_next() will give as B.
|
| 1265 | *
|
| 1266 | * do_vmi_munmap() will have restored the VMI back to addr.
|
| 1267 |
|
| 1268 | if (accountable_move) {
|
| 1269 | unsigned long end = addr + len;
|
| 1270 |
|
| 1271 | if (vm_start < addr) {
|
| 1272 | struct vm_area_struct *prev = vma_prev(&vmi);
|
| 1273 |
|
| 1274 | vma_start_write(prev);
|
| 1275 | vma_set_flags(prev, VMA_ACCOUNT_BIT);
|
| 1276 | }
|
| 1277 |
|
| 1278 | if (vm_end > end) {
|
| 1279 | struct vm_area_struct *next = vma_next(&vmi);
|
| 1280 |
|
| 1281 | vma_start_write(next);
|
| 1282 | vma_set_flags(next, VMA_ACCOUNT_BIT);
|
| 1283 | }
|
| 1284 | }
|
| 1285 | }
|
| 1286 |
|
| 1287 | |
| 1288 | * Copy vrm->vma over to vrm->new_addr possibly adjusting size as part of the
|
| 1289 | * process. Additionally handle an error occurring on moving of page tables,
|
| 1290 | * where we reset vrm state to cause unmapping of the new VMA.
|
| 1291 | *
|
| 1292 | * Outputs the newly installed VMA to new_vma_ptr. Returns 0 on success or an
|
| 1293 | * error code.
|
| 1294 |
|
| 1295 | static int copy_vma_and_data(struct vma_remap_struct *vrm,
|
| 1296 | struct vm_area_struct **new_vma_ptr)
|
| 1297 | {
|
| 1298 | const pgoff_t new_pgoff = linear_page_index(vrm->vma, vrm->addr);
|
| 1299 | const pgoff_t new_anon_pgoff =
|
| 1300 | __linear_anon_page_index(vrm->vma, vrm->addr);
|
| 1301 | struct vm_area_struct *vma = vrm->vma;
|
| 1302 | struct vm_area_struct *new_vma;
|
| 1303 | unsigned long moved_len;
|
| 1304 | int err = 0;
|
| 1305 | PAGETABLE_MOVE(pmc, NULL, NULL, vrm->addr, vrm->new_addr, vrm->old_len);
|
| 1306 |
|
| 1307 | new_vma = copy_vma(&vma, vrm->new_addr, vrm->new_len, new_pgoff,
|
| 1308 | new_anon_pgoff, &pmc.need_rmap_locks,
|
| 1309 | vrm->flags & MREMAP_DONTUNMAP);
|
| 1310 | if (!new_vma) {
|
| 1311 | vrm_uncharge(vrm);
|
| 1312 | *new_vma_ptr = NULL;
|
| 1313 | return -ENOMEM;
|
| 1314 | }
|
| 1315 |
|
| 1316 | if (vma != vrm->vma)
|
| 1317 | vrm->vmi_needs_invalidate = true;
|
| 1318 |
|
| 1319 | vrm->vma = vma;
|
| 1320 | pmc.old = vma;
|
| 1321 | pmc.new = new_vma;
|
| 1322 |
|
| 1323 | moved_len = move_page_tables(&pmc);
|
| 1324 | if (moved_len < vrm->old_len)
|
| 1325 | err = -ENOMEM;
|
| 1326 | else if (vma->vm_ops && vma->vm_ops->mremap)
|
| 1327 | err = vma->vm_ops->mremap(new_vma);
|
| 1328 |
|
| 1329 | if (unlikely(err)) {
|
| 1330 | PAGETABLE_MOVE(pmc_revert, new_vma, vma, vrm->new_addr,
|
| 1331 | vrm->addr, moved_len);
|
| 1332 |
|
| 1333 | |
| 1334 | * On error, move entries back from new area to old,
|
| 1335 | * which will succeed since page tables still there,
|
| 1336 | * and then proceed to unmap new area instead of old.
|
| 1337 |
|
| 1338 | pmc_revert.need_rmap_locks = true;
|
| 1339 | move_page_tables(&pmc_revert);
|
| 1340 |
|
| 1341 | vrm->vma = new_vma;
|
| 1342 | vrm->old_len = vrm->new_len;
|
| 1343 | vrm->addr = vrm->new_addr;
|
| 1344 | } else {
|
| 1345 | mremap_userfaultfd_prep(new_vma, vrm->uf);
|
| 1346 | }
|
| 1347 |
|
| 1348 | fixup_hugetlb_reservations(vma);
|
| 1349 |
|
| 1350 | *new_vma_ptr = new_vma;
|
| 1351 | return err;
|
| 1352 | }
|
| 1353 |
|
| 1354 | |
| 1355 | * Perform final tasks for MADV_DONTUNMAP operation, clearing mlock() flag on
|
| 1356 | * remaining VMA by convention (it cannot be mlock()'d any longer, as pages in
|
| 1357 | * range are no longer mapped), and removing anon_vma_chain links from it if the
|
| 1358 | * entire VMA was copied over.
|
| 1359 |
|
| 1360 | static void dontunmap_complete(struct vma_remap_struct *vrm,
|
| 1361 | struct vm_area_struct *new_vma)
|
| 1362 | {
|
| 1363 | unsigned long start = vrm->addr;
|
| 1364 | unsigned long end = vrm->addr + vrm->old_len;
|
| 1365 | struct vm_area_struct *vma = vrm->vma;
|
| 1366 | unsigned long old_start = vma->vm_start;
|
| 1367 | unsigned long old_end = vma->vm_end;
|
| 1368 |
|
| 1369 |
|
| 1370 | VM_WARN_ON_ONCE(new_vma == vma);
|
| 1371 |
|
| 1372 |
|
| 1373 | vma_clear_flags_mask(vma, VMA_LOCKED_MASK);
|
| 1374 |
|
| 1375 | |
| 1376 | * anon_vma links of the old vma is no longer needed after its page
|
| 1377 | * table has been moved.
|
| 1378 |
|
| 1379 | if (start == old_start && end == old_end) {
|
| 1380 | const pgoff_t pgoff_unfaulted = vma->vm_start >> PAGE_SHIFT;
|
| 1381 |
|
| 1382 | unlink_anon_vmas(vma);
|
| 1383 | |
| 1384 | * The VMA is now unfaulted and it is an invariant that
|
| 1385 | * unfaulted anonymous VMAs have page offset equal to
|
| 1386 | * vma->vm_start >> PAGE_SHIFT.
|
| 1387 |
|
| 1388 | vma_set_anon_pgoff(vma, pgoff_unfaulted);
|
| 1389 | if (vma_is_anonymous(vma) && !vma->vm_file)
|
| 1390 | vma_set_pgoff(vma, pgoff_unfaulted);
|
| 1391 | }
|
| 1392 | }
|
| 1393 |
|
| 1394 | static unsigned long move_vma(struct vma_remap_struct *vrm)
|
| 1395 | {
|
| 1396 | const bool is_dontunmap = vrm->flags & MREMAP_DONTUNMAP;
|
| 1397 | struct mm_struct *mm = current->mm;
|
| 1398 | struct vm_area_struct *new_vma;
|
| 1399 | unsigned long hiwater_vm;
|
| 1400 | int err;
|
| 1401 |
|
| 1402 | err = prep_move_vma(vrm);
|
| 1403 | if (err)
|
| 1404 | return err;
|
| 1405 |
|
| 1406 | |
| 1407 | * If accounted, determine the number of bytes the operation will
|
| 1408 | * charge.
|
| 1409 |
|
| 1410 | if (!vrm_calc_charge(vrm))
|
| 1411 | return -ENOMEM;
|
| 1412 |
|
| 1413 |
|
| 1414 | vma_start_write(vrm->vma);
|
| 1415 |
|
| 1416 |
|
| 1417 | err = copy_vma_and_data(vrm, &new_vma);
|
| 1418 | |
| 1419 | * If we established the copied-to VMA, we attempt to recover from the
|
| 1420 | * error by setting the destination VMA to the source VMA and unmapping
|
| 1421 | * it below.
|
| 1422 |
|
| 1423 | if (err && !new_vma)
|
| 1424 | return err;
|
| 1425 |
|
| 1426 | |
| 1427 | * If we failed to move page tables we still do total_vm increment
|
| 1428 | * since do_munmap() will decrement it by old_len == new_len.
|
| 1429 | *
|
| 1430 | * Since total_vm is about to be raised artificially high for a
|
| 1431 | * moment, we need to restore high watermark afterwards: if stats
|
| 1432 | * are taken meanwhile, total_vm and hiwater_vm appear too high.
|
| 1433 | * If this were a serious issue, we'd add a flag to do_munmap().
|
| 1434 |
|
| 1435 | hiwater_vm = mm->hiwater_vm;
|
| 1436 |
|
| 1437 | if (unlikely(is_dontunmap && !err))
|
| 1438 | dontunmap_complete(vrm, new_vma);
|
| 1439 | vrm_stat_account(vrm, vrm->new_len);
|
| 1440 | if (!is_dontunmap || err)
|
| 1441 | unmap_source_vma(vrm);
|
| 1442 |
|
| 1443 | mm->hiwater_vm = hiwater_vm;
|
| 1444 |
|
| 1445 | return err ? (unsigned long)err : vrm->new_addr;
|
| 1446 | }
|
| 1447 |
|
| 1448 | |
| 1449 | * The user has requested that the VMA be shrunk (i.e., old_len > new_len), so
|
| 1450 | * execute this, optionally dropping the mmap lock when we do so.
|
| 1451 | *
|
| 1452 | * In both cases this invalidates the VMA, however if we don't drop the lock,
|
| 1453 | * then load the correct VMA into vrm->vma afterwards.
|
| 1454 |
|
| 1455 | static unsigned long shrink_vma(struct vma_remap_struct *vrm,
|
| 1456 | bool drop_lock)
|
| 1457 | {
|
| 1458 | struct mm_struct *mm = current->mm;
|
| 1459 | unsigned long unmap_start = vrm->addr + vrm->new_len;
|
| 1460 | unsigned long unmap_bytes = vrm->delta;
|
| 1461 | unsigned long res;
|
| 1462 | VMA_ITERATOR(vmi, mm, unmap_start);
|
| 1463 |
|
| 1464 | VM_BUG_ON(vrm->remap_type != MREMAP_SHRINK);
|
| 1465 |
|
| 1466 | res = do_vmi_munmap(&vmi, mm, unmap_start, unmap_bytes,
|
| 1467 | vrm->uf_unmap, drop_lock);
|
| 1468 | vrm->vma = NULL;
|
| 1469 | if (res)
|
| 1470 | return res;
|
| 1471 |
|
| 1472 | |
| 1473 | * If we've not dropped the lock, then we should reload the VMA to
|
| 1474 | * replace the invalidated VMA with the one that may have now been
|
| 1475 | * split.
|
| 1476 |
|
| 1477 | if (drop_lock) {
|
| 1478 | vrm->mmap_locked = false;
|
| 1479 | } else {
|
| 1480 | vrm->vma = vma_lookup(mm, vrm->addr);
|
| 1481 | if (!vrm->vma)
|
| 1482 | return -EFAULT;
|
| 1483 | }
|
| 1484 |
|
| 1485 | return 0;
|
| 1486 | }
|
| 1487 |
|
| 1488 | |
| 1489 | * mremap_to() - remap a vma to a new location.
|
| 1490 | * Returns: The new address of the vma or an error.
|
| 1491 |
|
| 1492 | static unsigned long mremap_to(struct vma_remap_struct *vrm)
|
| 1493 | {
|
| 1494 | struct mm_struct *mm = current->mm;
|
| 1495 | unsigned long err;
|
| 1496 |
|
| 1497 | if (vrm->flags & MREMAP_FIXED) {
|
| 1498 | |
| 1499 | * In mremap_to().
|
| 1500 | * VMA is moved to dst address, and munmap dst first.
|
| 1501 | * do_munmap will check if dst is sealed.
|
| 1502 |
|
| 1503 | err = do_munmap(mm, vrm->new_addr, vrm->new_len,
|
| 1504 | vrm->uf_unmap_early);
|
| 1505 | vrm->vma = NULL;
|
| 1506 | vrm->vmi_needs_invalidate = true;
|
| 1507 | if (err)
|
| 1508 | return err;
|
| 1509 |
|
| 1510 | |
| 1511 | * If we remap a portion of a VMA elsewhere in the same VMA,
|
| 1512 | * this can invalidate the old VMA. Reset.
|
| 1513 |
|
| 1514 | vrm->vma = vma_lookup(mm, vrm->addr);
|
| 1515 | if (!vrm->vma)
|
| 1516 | return -EFAULT;
|
| 1517 | }
|
| 1518 |
|
| 1519 | if (vrm->remap_type == MREMAP_SHRINK) {
|
| 1520 | err = shrink_vma(vrm, false);
|
| 1521 | if (err)
|
| 1522 | return err;
|
| 1523 |
|
| 1524 |
|
| 1525 | vrm->old_len = vrm->new_len;
|
| 1526 | }
|
| 1527 |
|
| 1528 |
|
| 1529 | if (vrm->flags & MREMAP_DONTUNMAP) {
|
| 1530 | vma_flags_t vma_flags = vrm->vma->flags;
|
| 1531 | unsigned long pages = vrm->old_len >> PAGE_SHIFT;
|
| 1532 |
|
| 1533 | if (!may_expand_vm(mm, &vma_flags, pages))
|
| 1534 | return -ENOMEM;
|
| 1535 | }
|
| 1536 |
|
| 1537 | err = vrm_set_new_addr(vrm);
|
| 1538 | if (err)
|
| 1539 | return err;
|
| 1540 |
|
| 1541 | return move_vma(vrm);
|
| 1542 | }
|
| 1543 |
|
| 1544 | static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
|
| 1545 | {
|
| 1546 | unsigned long end = vma->vm_end + delta;
|
| 1547 |
|
| 1548 | if (end < vma->vm_end)
|
| 1549 | return 0;
|
| 1550 | if (find_vma_intersection(vma->vm_mm, vma->vm_end, end))
|
| 1551 | return 0;
|
| 1552 | if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
|
| 1553 | 0, MAP_FIXED) & ~PAGE_MASK)
|
| 1554 | return 0;
|
| 1555 | return 1;
|
| 1556 | }
|
| 1557 |
|
| 1558 |
|
| 1559 | static bool vrm_can_expand_in_place(struct vma_remap_struct *vrm)
|
| 1560 | {
|
| 1561 |
|
| 1562 | unsigned long suffix_bytes = vrm->vma->vm_end - vrm->addr;
|
| 1563 |
|
| 1564 |
|
| 1565 | if (suffix_bytes != vrm->old_len)
|
| 1566 | return false;
|
| 1567 |
|
| 1568 |
|
| 1569 | if (!vma_expandable(vrm->vma, vrm->delta))
|
| 1570 | return false;
|
| 1571 |
|
| 1572 | return true;
|
| 1573 | }
|
| 1574 |
|
| 1575 | |
| 1576 | * We know we can expand the VMA in-place by delta pages, so do so.
|
| 1577 | *
|
| 1578 | * If we discover the VMA is locked, update mm_struct statistics accordingly and
|
| 1579 | * indicate so to the caller.
|
| 1580 |
|
| 1581 | static unsigned long expand_vma_in_place(struct vma_remap_struct *vrm)
|
| 1582 | {
|
| 1583 | struct mm_struct *mm = current->mm;
|
| 1584 | struct vm_area_struct *vma = vrm->vma;
|
| 1585 | VMA_ITERATOR(vmi, mm, vma->vm_end);
|
| 1586 |
|
| 1587 | if (!vrm_calc_charge(vrm))
|
| 1588 | return -ENOMEM;
|
| 1589 |
|
| 1590 | |
| 1591 | * Function vma_merge_extend() is called on the
|
| 1592 | * extension we are adding to the already existing vma,
|
| 1593 | * vma_merge_extend() will merge this extension with the
|
| 1594 | * already existing vma (expand operation itself) and
|
| 1595 | * possibly also with the next vma if it becomes
|
| 1596 | * adjacent to the expanded vma and otherwise
|
| 1597 | * compatible.
|
| 1598 |
|
| 1599 | vma = vma_merge_extend(&vmi, vma, vrm->delta);
|
| 1600 | if (!vma) {
|
| 1601 | vrm_uncharge(vrm);
|
| 1602 | return -ENOMEM;
|
| 1603 | }
|
| 1604 | vrm->vma = vma;
|
| 1605 |
|
| 1606 | vrm_stat_account(vrm, vrm->delta);
|
| 1607 |
|
| 1608 | return 0;
|
| 1609 | }
|
| 1610 |
|
| 1611 | static bool align_hugetlb(struct vma_remap_struct *vrm)
|
| 1612 | {
|
| 1613 | struct hstate *h __maybe_unused = hstate_vma(vrm->vma);
|
| 1614 |
|
| 1615 | vrm->old_len = ALIGN(vrm->old_len, huge_page_size(h));
|
| 1616 | vrm->new_len = ALIGN(vrm->new_len, huge_page_size(h));
|
| 1617 |
|
| 1618 |
|
| 1619 | if (vrm->addr & ~huge_page_mask(h))
|
| 1620 | return false;
|
| 1621 | if (vrm->new_addr & ~huge_page_mask(h))
|
| 1622 | return false;
|
| 1623 |
|
| 1624 | |
| 1625 | * Don't allow remap expansion, because the underlying hugetlb
|
| 1626 | * reservation is not yet capable to handle split reservation.
|
| 1627 |
|
| 1628 | if (vrm->new_len > vrm->old_len)
|
| 1629 | return false;
|
| 1630 |
|
| 1631 | return true;
|
| 1632 | }
|
| 1633 |
|
| 1634 | |
| 1635 | * We are mremap()'ing without specifying a fixed address to move to, but are
|
| 1636 | * requesting that the VMA's size be increased.
|
| 1637 | *
|
| 1638 | * Try to do so in-place, if this fails, then move the VMA to a new location to
|
| 1639 | * action the change.
|
| 1640 |
|
| 1641 | static unsigned long expand_vma(struct vma_remap_struct *vrm)
|
| 1642 | {
|
| 1643 | unsigned long err;
|
| 1644 |
|
| 1645 | |
| 1646 | * [addr, old_len) spans precisely to the end of the VMA, so try to
|
| 1647 | * expand it in-place.
|
| 1648 |
|
| 1649 | if (vrm_can_expand_in_place(vrm)) {
|
| 1650 | err = expand_vma_in_place(vrm);
|
| 1651 | if (err)
|
| 1652 | return err;
|
| 1653 |
|
| 1654 |
|
| 1655 | return vrm->addr;
|
| 1656 | }
|
| 1657 |
|
| 1658 | |
| 1659 | * We weren't able to just expand or shrink the area,
|
| 1660 | * we need to create a new one and move it.
|
| 1661 |
|
| 1662 |
|
| 1663 |
|
| 1664 | if (!(vrm->flags & MREMAP_MAYMOVE))
|
| 1665 | return -ENOMEM;
|
| 1666 |
|
| 1667 |
|
| 1668 | err = vrm_set_new_addr(vrm);
|
| 1669 | if (err)
|
| 1670 | return err;
|
| 1671 |
|
| 1672 | return move_vma(vrm);
|
| 1673 | }
|
| 1674 |
|
| 1675 | |
| 1676 | * Attempt to resize the VMA in-place, if we cannot, then move the VMA to the
|
| 1677 | * first available address to perform the operation.
|
| 1678 |
|
| 1679 | static unsigned long mremap_at(struct vma_remap_struct *vrm)
|
| 1680 | {
|
| 1681 | unsigned long res;
|
| 1682 |
|
| 1683 | switch (vrm->remap_type) {
|
| 1684 | case MREMAP_INVALID:
|
| 1685 | break;
|
| 1686 | case MREMAP_NO_RESIZE:
|
| 1687 |
|
| 1688 | return vrm->addr;
|
| 1689 | case MREMAP_SHRINK:
|
| 1690 | |
| 1691 | * SHRINK CASE. Can always be done in-place.
|
| 1692 | *
|
| 1693 | * Simply unmap the shrunken portion of the VMA. This does all
|
| 1694 | * the needed commit accounting, and we indicate that the mmap
|
| 1695 | * lock should be dropped.
|
| 1696 |
|
| 1697 | res = shrink_vma(vrm, true);
|
| 1698 | if (res)
|
| 1699 | return res;
|
| 1700 |
|
| 1701 | return vrm->addr;
|
| 1702 | case MREMAP_EXPAND:
|
| 1703 | return expand_vma(vrm);
|
| 1704 | }
|
| 1705 |
|
| 1706 |
|
| 1707 | WARN_ON_ONCE(1);
|
| 1708 | return -EINVAL;
|
| 1709 | }
|
| 1710 |
|
| 1711 | |
| 1712 | * Will this operation result in the VMA being expanded or moved and thus need
|
| 1713 | * to map a new portion of virtual address space?
|
| 1714 |
|
| 1715 | static bool vrm_will_map_new(struct vma_remap_struct *vrm)
|
| 1716 | {
|
| 1717 | if (vrm->remap_type == MREMAP_EXPAND)
|
| 1718 | return true;
|
| 1719 |
|
| 1720 | if (vrm_implies_new_addr(vrm))
|
| 1721 | return true;
|
| 1722 |
|
| 1723 | return false;
|
| 1724 | }
|
| 1725 |
|
| 1726 |
|
| 1727 | static bool vrm_move_only(struct vma_remap_struct *vrm)
|
| 1728 | {
|
| 1729 | if (!(vrm->flags & MREMAP_FIXED))
|
| 1730 | return false;
|
| 1731 |
|
| 1732 | if (vrm->old_len != vrm->new_len)
|
| 1733 | return false;
|
| 1734 |
|
| 1735 | return true;
|
| 1736 | }
|
| 1737 |
|
| 1738 | static void notify_uffd(struct vma_remap_struct *vrm, bool failed)
|
| 1739 | {
|
| 1740 | struct mm_struct *mm = current->mm;
|
| 1741 |
|
| 1742 |
|
| 1743 | userfaultfd_unmap_complete(mm, vrm->uf_unmap_early);
|
| 1744 | if (failed)
|
| 1745 | mremap_userfaultfd_fail(vrm->uf);
|
| 1746 | else
|
| 1747 | mremap_userfaultfd_complete(vrm->uf, vrm->addr,
|
| 1748 | vrm->new_addr, vrm->old_len);
|
| 1749 | userfaultfd_unmap_complete(mm, vrm->uf_unmap);
|
| 1750 | }
|
| 1751 |
|
| 1752 | static bool vma_multi_allowed(struct vm_area_struct *vma)
|
| 1753 | {
|
| 1754 | struct file *file = vma->vm_file;
|
| 1755 |
|
| 1756 | |
| 1757 | * We can't support moving multiple uffd VMAs as notify requires
|
| 1758 | * mmap lock to be dropped.
|
| 1759 |
|
| 1760 | if (userfaultfd_armed(vma))
|
| 1761 | return false;
|
| 1762 |
|
| 1763 | |
| 1764 | * Custom get unmapped area might result in MREMAP_FIXED not
|
| 1765 | * being obeyed.
|
| 1766 |
|
| 1767 | if (!file || !file->f_op->get_unmapped_area)
|
| 1768 | return true;
|
| 1769 |
|
| 1770 | if (vma_is_shmem(vma))
|
| 1771 | return true;
|
| 1772 | if (is_vm_hugetlb_page(vma))
|
| 1773 | return true;
|
| 1774 | if (file->f_op->get_unmapped_area == thp_get_unmapped_area)
|
| 1775 | return true;
|
| 1776 |
|
| 1777 | return false;
|
| 1778 | }
|
| 1779 |
|
| 1780 | static int check_prep_vma(struct vma_remap_struct *vrm)
|
| 1781 | {
|
| 1782 | struct vm_area_struct *vma = vrm->vma;
|
| 1783 | struct mm_struct *mm = current->mm;
|
| 1784 | unsigned long addr = vrm->addr;
|
| 1785 | unsigned long old_len, new_len, pgoff;
|
| 1786 |
|
| 1787 | if (!vma)
|
| 1788 | return -EFAULT;
|
| 1789 |
|
| 1790 |
|
| 1791 | if (vma_is_sealed(vma))
|
| 1792 | return -EPERM;
|
| 1793 |
|
| 1794 |
|
| 1795 | if (is_vm_hugetlb_page(vma) && !align_hugetlb(vrm))
|
| 1796 | return -EINVAL;
|
| 1797 |
|
| 1798 | vrm_set_delta(vrm);
|
| 1799 | vrm->remap_type = vrm_remap_type(vrm);
|
| 1800 |
|
| 1801 | if (!vrm_implies_new_addr(vrm))
|
| 1802 | vrm->new_addr = addr;
|
| 1803 |
|
| 1804 |
|
| 1805 | if (!vrm_will_map_new(vrm))
|
| 1806 | return 0;
|
| 1807 |
|
| 1808 | old_len = vrm->old_len;
|
| 1809 | new_len = vrm->new_len;
|
| 1810 |
|
| 1811 | |
| 1812 | * !old_len is a special case where an attempt is made to 'duplicate'
|
| 1813 | * a mapping. This makes no sense for private mappings as it will
|
| 1814 | * instead create a fresh/new mapping unrelated to the original. This
|
| 1815 | * is contrary to the basic idea of mremap which creates new mappings
|
| 1816 | * based on the original. There are no known use cases for this
|
| 1817 | * behavior. As a result, fail such attempts.
|
| 1818 |
|
| 1819 | if (!old_len && !vma_test_any(vma, VMA_SHARED_BIT, VMA_MAYSHARE_BIT)) {
|
| 1820 | pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n",
|
| 1821 | current->comm, current->pid);
|
| 1822 | return -EINVAL;
|
| 1823 | }
|
| 1824 |
|
| 1825 | if ((vrm->flags & MREMAP_DONTUNMAP) &&
|
| 1826 | vma_test_any(vma, VMA_DONTEXPAND_BIT, VMA_PFNMAP_BIT))
|
| 1827 | return -EINVAL;
|
| 1828 |
|
| 1829 | |
| 1830 | * We permit crossing of boundaries for the range being unmapped due to
|
| 1831 | * a shrink.
|
| 1832 |
|
| 1833 | if (vrm->remap_type == MREMAP_SHRINK)
|
| 1834 | old_len = new_len;
|
| 1835 |
|
| 1836 | |
| 1837 | * We can't remap across the end of VMAs, as another VMA may be
|
| 1838 | * adjacent:
|
| 1839 | *
|
| 1840 | * addr vma->vm_end
|
| 1841 | * |-----.----------|
|
| 1842 | * | . |
|
| 1843 | * |-----.----------|
|
| 1844 | * .<--------->xxx>
|
| 1845 | * old_len
|
| 1846 | *
|
| 1847 | * We also require that vma->vm_start <= addr < vma->vm_end.
|
| 1848 |
|
| 1849 | if (old_len > vma->vm_end - addr)
|
| 1850 | return -EFAULT;
|
| 1851 |
|
| 1852 | if (new_len == old_len)
|
| 1853 | return 0;
|
| 1854 |
|
| 1855 |
|
| 1856 | if (vma_test(vma, VMA_LOCKED_BIT))
|
| 1857 | vrm->populate_expand = true;
|
| 1858 |
|
| 1859 |
|
| 1860 | pgoff = linear_page_index(vma, addr);
|
| 1861 | if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
|
| 1862 | return -EINVAL;
|
| 1863 |
|
| 1864 | if (vma_test_any(vma, VMA_DONTEXPAND_BIT, VMA_PFNMAP_BIT))
|
| 1865 | return -EFAULT;
|
| 1866 |
|
| 1867 | if (!mlock_future_ok(mm, vma_test(vma, VMA_LOCKED_BIT), vrm->delta))
|
| 1868 | return -EAGAIN;
|
| 1869 |
|
| 1870 | if (!may_expand_vm(mm, &vma->flags, vrm->delta >> PAGE_SHIFT))
|
| 1871 | return -ENOMEM;
|
| 1872 |
|
| 1873 | return 0;
|
| 1874 | }
|
| 1875 |
|
| 1876 | |
| 1877 | * Are the parameters passed to mremap() valid? If so return 0, otherwise return
|
| 1878 | * error.
|
| 1879 |
|
| 1880 | static unsigned long check_mremap_params(struct vma_remap_struct *vrm)
|
| 1881 |
|
| 1882 | {
|
| 1883 | unsigned long addr = vrm->addr;
|
| 1884 | unsigned long flags = vrm->flags;
|
| 1885 |
|
| 1886 |
|
| 1887 | if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
|
| 1888 | return -EINVAL;
|
| 1889 |
|
| 1890 |
|
| 1891 | if (offset_in_page(addr))
|
| 1892 | return -EINVAL;
|
| 1893 |
|
| 1894 | |
| 1895 | * We allow a zero old-len as a special case
|
| 1896 | * for DOS-emu "duplicate shm area" thing. But
|
| 1897 | * a zero new-len is nonsensical.
|
| 1898 |
|
| 1899 | if (!vrm->new_len)
|
| 1900 | return -EINVAL;
|
| 1901 |
|
| 1902 |
|
| 1903 | if (vrm->new_len > TASK_SIZE)
|
| 1904 | return -EINVAL;
|
| 1905 |
|
| 1906 |
|
| 1907 | if (!vrm_implies_new_addr(vrm))
|
| 1908 | return 0;
|
| 1909 |
|
| 1910 |
|
| 1911 | if (vrm->new_addr > TASK_SIZE - vrm->new_len)
|
| 1912 | return -EINVAL;
|
| 1913 |
|
| 1914 |
|
| 1915 | if (offset_in_page(vrm->new_addr))
|
| 1916 | return -EINVAL;
|
| 1917 |
|
| 1918 |
|
| 1919 | if (!(flags & MREMAP_MAYMOVE))
|
| 1920 | return -EINVAL;
|
| 1921 |
|
| 1922 |
|
| 1923 | if (flags & MREMAP_DONTUNMAP && vrm->old_len != vrm->new_len)
|
| 1924 | return -EINVAL;
|
| 1925 |
|
| 1926 |
|
| 1927 | if (vrm_overlaps(vrm))
|
| 1928 | return -EINVAL;
|
| 1929 |
|
| 1930 | return 0;
|
| 1931 | }
|
| 1932 |
|
| 1933 | static unsigned long remap_move(struct vma_remap_struct *vrm)
|
| 1934 | {
|
| 1935 | struct vm_area_struct *vma;
|
| 1936 | unsigned long start = vrm->addr;
|
| 1937 | unsigned long end = vrm->addr + vrm->old_len;
|
| 1938 | unsigned long new_addr = vrm->new_addr;
|
| 1939 | unsigned long target_addr = new_addr;
|
| 1940 | unsigned long res = -EFAULT;
|
| 1941 | unsigned long last_end;
|
| 1942 | bool seen_vma = false;
|
| 1943 |
|
| 1944 | VMA_ITERATOR(vmi, current->mm, start);
|
| 1945 |
|
| 1946 | |
| 1947 | * When moving VMAs we allow for batched moves across multiple VMAs,
|
| 1948 | * with all VMAs in the input range [addr, addr + old_len) being moved
|
| 1949 | * (and split as necessary).
|
| 1950 |
|
| 1951 | for_each_vma_range(vmi, vma, end) {
|
| 1952 |
|
| 1953 | unsigned long addr = max(vma->vm_start, start);
|
| 1954 | unsigned long len = min(end, vma->vm_end) - addr;
|
| 1955 | unsigned long offset, res_vma;
|
| 1956 | bool multi_allowed;
|
| 1957 |
|
| 1958 |
|
| 1959 | if (!seen_vma && start < vma->vm_start)
|
| 1960 | return -EFAULT;
|
| 1961 |
|
| 1962 | |
| 1963 | * To sensibly move multiple VMAs, accounting for the fact that
|
| 1964 | * get_unmapped_area() may align even MAP_FIXED moves, we simply
|
| 1965 | * attempt to move such that the gaps between source VMAs remain
|
| 1966 | * consistent in destination VMAs, e.g.:
|
| 1967 | *
|
| 1968 | * X Y X Y
|
| 1969 | * <---> <-> <---> <->
|
| 1970 | * |-------| |-----| |-----| |-------| |-----| |-----|
|
| 1971 | * | A | | B | | C | ---> | A' | | B' | | C' |
|
| 1972 | * |-------| |-----| |-----| |-------| |-----| |-----|
|
| 1973 | * new_addr
|
| 1974 | *
|
| 1975 | * So we map B' at A'->vm_end + X, and C' at B'->vm_end + Y.
|
| 1976 |
|
| 1977 | offset = seen_vma ? vma->vm_start - last_end : 0;
|
| 1978 | last_end = vma->vm_end;
|
| 1979 |
|
| 1980 | vrm->vma = vma;
|
| 1981 | vrm->addr = addr;
|
| 1982 | vrm->new_addr = target_addr + offset;
|
| 1983 | vrm->old_len = vrm->new_len = len;
|
| 1984 |
|
| 1985 | multi_allowed = vma_multi_allowed(vma);
|
| 1986 | if (!multi_allowed) {
|
| 1987 |
|
| 1988 | if (seen_vma)
|
| 1989 | return -EFAULT;
|
| 1990 |
|
| 1991 | if (vma->vm_end < end)
|
| 1992 | return -EFAULT;
|
| 1993 | }
|
| 1994 |
|
| 1995 | res_vma = check_prep_vma(vrm);
|
| 1996 | if (!res_vma)
|
| 1997 | res_vma = mremap_to(vrm);
|
| 1998 | if (IS_ERR_VALUE(res_vma))
|
| 1999 | return res_vma;
|
| 2000 |
|
| 2001 | if (!seen_vma) {
|
| 2002 | VM_WARN_ON_ONCE(multi_allowed && res_vma != new_addr);
|
| 2003 | res = res_vma;
|
| 2004 | }
|
| 2005 |
|
| 2006 |
|
| 2007 | VM_WARN_ON_ONCE(!vrm->mmap_locked);
|
| 2008 |
|
| 2009 | VM_WARN_ON_ONCE(vrm->populate_expand);
|
| 2010 |
|
| 2011 | if (vrm->vmi_needs_invalidate) {
|
| 2012 | vma_iter_invalidate(&vmi);
|
| 2013 | vrm->vmi_needs_invalidate = false;
|
| 2014 | }
|
| 2015 | seen_vma = true;
|
| 2016 | target_addr = res_vma + vrm->new_len;
|
| 2017 | }
|
| 2018 |
|
| 2019 | return res;
|
| 2020 | }
|
| 2021 |
|
| 2022 | static unsigned long do_mremap(struct vma_remap_struct *vrm)
|
| 2023 | {
|
| 2024 | struct mm_struct *mm = current->mm;
|
| 2025 | unsigned long res;
|
| 2026 | bool failed;
|
| 2027 |
|
| 2028 | vrm->old_len = PAGE_ALIGN(vrm->old_len);
|
| 2029 | vrm->new_len = PAGE_ALIGN(vrm->new_len);
|
| 2030 |
|
| 2031 | res = check_mremap_params(vrm);
|
| 2032 | if (res)
|
| 2033 | return res;
|
| 2034 |
|
| 2035 | if (mmap_write_lock_killable(mm))
|
| 2036 | return -EINTR;
|
| 2037 | vrm->mmap_locked = true;
|
| 2038 |
|
| 2039 | if (!check_map_count_against_split_early(vrm)) {
|
| 2040 | mmap_write_unlock(mm);
|
| 2041 | return -ENOMEM;
|
| 2042 | }
|
| 2043 |
|
| 2044 | if (vrm_move_only(vrm)) {
|
| 2045 | res = remap_move(vrm);
|
| 2046 | } else {
|
| 2047 | vrm->vma = vma_lookup(current->mm, vrm->addr);
|
| 2048 | res = check_prep_vma(vrm);
|
| 2049 | if (res)
|
| 2050 | goto out;
|
| 2051 |
|
| 2052 |
|
| 2053 | res = vrm_implies_new_addr(vrm) ? mremap_to(vrm) : mremap_at(vrm);
|
| 2054 | }
|
| 2055 |
|
| 2056 | out:
|
| 2057 | failed = IS_ERR_VALUE(res);
|
| 2058 |
|
| 2059 | if (vrm->mmap_locked)
|
| 2060 | mmap_write_unlock(mm);
|
| 2061 |
|
| 2062 |
|
| 2063 | if (!failed && vrm->populate_expand)
|
| 2064 | mm_populate(vrm->new_addr + vrm->old_len, vrm->delta);
|
| 2065 |
|
| 2066 | notify_uffd(vrm, failed);
|
| 2067 | return res;
|
| 2068 | }
|
| 2069 |
|
| 2070 | |
| 2071 | * Expand (or shrink) an existing mapping, potentially moving it at the
|
| 2072 | * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
|
| 2073 | *
|
| 2074 | * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
|
| 2075 | * This option implies MREMAP_MAYMOVE.
|
| 2076 |
|
| 2077 | SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
|
| 2078 | unsigned long, new_len, unsigned long, flags,
|
| 2079 | unsigned long, new_addr)
|
| 2080 | {
|
| 2081 | struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
|
| 2082 | LIST_HEAD(uf_unmap_early);
|
| 2083 | LIST_HEAD(uf_unmap);
|
| 2084 | |
| 2085 | * There is a deliberate asymmetry here: we strip the pointer tag
|
| 2086 | * from the old address but leave the new address alone. This is
|
| 2087 | * for consistency with mmap(), where we prevent the creation of
|
| 2088 | * aliasing mappings in userspace by leaving the tag bits of the
|
| 2089 | * mapping address intact. A non-zero tag will cause the subsequent
|
| 2090 | * range checks to reject the address as invalid.
|
| 2091 | *
|
| 2092 | * See Documentation/arch/arm64/tagged-address-abi.rst for more
|
| 2093 | * information.
|
| 2094 |
|
| 2095 | struct vma_remap_struct vrm = {
|
| 2096 | .addr = untagged_addr(addr),
|
| 2097 | .old_len = old_len,
|
| 2098 | .new_len = new_len,
|
| 2099 | .flags = flags,
|
| 2100 | .new_addr = new_addr,
|
| 2101 |
|
| 2102 | .uf = &uf,
|
| 2103 | .uf_unmap_early = &uf_unmap_early,
|
| 2104 | .uf_unmap = &uf_unmap,
|
| 2105 |
|
| 2106 | .remap_type = MREMAP_INVALID,
|
| 2107 | };
|
| 2108 |
|
| 2109 | return do_mremap(&vrm);
|
| 2110 | }
|