| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * linux/mm/madvise.c |
| 4 | * |
| 5 | * Copyright (C) 1999 Linus Torvalds |
| 6 | * Copyright (C) 2002 Christoph Hellwig |
| 7 | */ |
| 8 | |
| 9 | #include <linux/mman.h> |
| 10 | #include <linux/pagemap.h> |
| 11 | #include <linux/syscalls.h> |
| 12 | #include <linux/mempolicy.h> |
| 13 | #include <linux/page-isolation.h> |
| 14 | #include <linux/userfaultfd_k.h> |
| 15 | #include <linux/hugetlb.h> |
| 16 | #include <linux/falloc.h> |
| 17 | #include <linux/fadvise.h> |
| 18 | #include <linux/sched.h> |
| 19 | #include <linux/ksm.h> |
| 20 | #include <linux/fs.h> |
| 21 | #include <linux/file.h> |
| 22 | #include <linux/blkdev.h> |
| 23 | #include <linux/backing-dev.h> |
| 24 | #include <linux/pagewalk.h> |
| 25 | #include <linux/swap.h> |
| 26 | #include <linux/swapops.h> |
| 27 | #include <linux/shmem_fs.h> |
| 28 | #include <linux/mmu_notifier.h> |
| 29 | |
| 30 | #include <asm/tlb.h> |
| 31 | |
| 32 | #include "internal.h" |
| 33 | |
| 34 | /* |
| 35 | * Any behaviour which results in changes to the vma->vm_flags needs to |
| 36 | * take mmap_sem for writing. Others, which simply traverse vmas, need |
| 37 | * to only take it for reading. |
| 38 | */ |
| 39 | static int madvise_need_mmap_write(int behavior) |
| 40 | { |
| 41 | switch (behavior) { |
| 42 | case MADV_REMOVE: |
| 43 | case MADV_WILLNEED: |
| 44 | case MADV_DONTNEED: |
| 45 | case MADV_FREE: |
| 46 | return 0; |
| 47 | default: |
| 48 | /* be safe, default to 1. list exceptions explicitly */ |
| 49 | return 1; |
| 50 | } |
| 51 | } |
| 52 | |
| 53 | /* |
| 54 | * We can potentially split a vm area into separate |
| 55 | * areas, each area with its own behavior. |
| 56 | */ |
| 57 | static long madvise_behavior(struct vm_area_struct *vma, |
| 58 | struct vm_area_struct **prev, |
| 59 | unsigned long start, unsigned long end, int behavior) |
| 60 | { |
| 61 | struct mm_struct *mm = vma->vm_mm; |
| 62 | int error = 0; |
| 63 | pgoff_t pgoff; |
| 64 | unsigned long new_flags = vma->vm_flags; |
| 65 | |
| 66 | switch (behavior) { |
| 67 | case MADV_NORMAL: |
| 68 | new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ; |
| 69 | break; |
| 70 | case MADV_SEQUENTIAL: |
| 71 | new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ; |
| 72 | break; |
| 73 | case MADV_RANDOM: |
| 74 | new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ; |
| 75 | break; |
| 76 | case MADV_DONTFORK: |
| 77 | new_flags |= VM_DONTCOPY; |
| 78 | break; |
| 79 | case MADV_DOFORK: |
| 80 | if (vma->vm_flags & VM_IO) { |
| 81 | error = -EINVAL; |
| 82 | goto out; |
| 83 | } |
| 84 | new_flags &= ~VM_DONTCOPY; |
| 85 | break; |
| 86 | case MADV_WIPEONFORK: |
| 87 | /* MADV_WIPEONFORK is only supported on anonymous memory. */ |
| 88 | if (vma->vm_file || vma->vm_flags & VM_SHARED) { |
| 89 | error = -EINVAL; |
| 90 | goto out; |
| 91 | } |
| 92 | new_flags |= VM_WIPEONFORK; |
| 93 | break; |
| 94 | case MADV_KEEPONFORK: |
| 95 | new_flags &= ~VM_WIPEONFORK; |
| 96 | break; |
| 97 | case MADV_DONTDUMP: |
| 98 | new_flags |= VM_DONTDUMP; |
| 99 | break; |
| 100 | case MADV_DODUMP: |
| 101 | if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) { |
| 102 | error = -EINVAL; |
| 103 | goto out; |
| 104 | } |
| 105 | new_flags &= ~VM_DONTDUMP; |
| 106 | break; |
| 107 | case MADV_MERGEABLE: |
| 108 | case MADV_UNMERGEABLE: |
| 109 | error = ksm_madvise(vma, start, end, behavior, &new_flags); |
| 110 | if (error) { |
| 111 | /* |
| 112 | * madvise() returns EAGAIN if kernel resources, such as |
| 113 | * slab, are temporarily unavailable. |
| 114 | */ |
| 115 | if (error == -ENOMEM) |
| 116 | error = -EAGAIN; |
| 117 | goto out; |
| 118 | } |
| 119 | break; |
| 120 | case MADV_HUGEPAGE: |
| 121 | case MADV_NOHUGEPAGE: |
| 122 | error = hugepage_madvise(vma, &new_flags, behavior); |
| 123 | if (error) { |
| 124 | /* |
| 125 | * madvise() returns EAGAIN if kernel resources, such as |
| 126 | * slab, are temporarily unavailable. |
| 127 | */ |
| 128 | if (error == -ENOMEM) |
| 129 | error = -EAGAIN; |
| 130 | goto out; |
| 131 | } |
| 132 | break; |
| 133 | } |
| 134 | |
| 135 | if (new_flags == vma->vm_flags) { |
| 136 | *prev = vma; |
| 137 | goto out; |
| 138 | } |
| 139 | |
| 140 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); |
| 141 | *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma, |
| 142 | vma->vm_file, pgoff, vma_policy(vma), |
| 143 | vma->vm_userfaultfd_ctx); |
| 144 | if (*prev) { |
| 145 | vma = *prev; |
| 146 | goto success; |
| 147 | } |
| 148 | |
| 149 | *prev = vma; |
| 150 | |
| 151 | if (start != vma->vm_start) { |
| 152 | if (unlikely(mm->map_count >= sysctl_max_map_count)) { |
| 153 | error = -ENOMEM; |
| 154 | goto out; |
| 155 | } |
| 156 | error = __split_vma(mm, vma, start, 1); |
| 157 | if (error) { |
| 158 | /* |
| 159 | * madvise() returns EAGAIN if kernel resources, such as |
| 160 | * slab, are temporarily unavailable. |
| 161 | */ |
| 162 | if (error == -ENOMEM) |
| 163 | error = -EAGAIN; |
| 164 | goto out; |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | if (end != vma->vm_end) { |
| 169 | if (unlikely(mm->map_count >= sysctl_max_map_count)) { |
| 170 | error = -ENOMEM; |
| 171 | goto out; |
| 172 | } |
| 173 | error = __split_vma(mm, vma, end, 0); |
| 174 | if (error) { |
| 175 | /* |
| 176 | * madvise() returns EAGAIN if kernel resources, such as |
| 177 | * slab, are temporarily unavailable. |
| 178 | */ |
| 179 | if (error == -ENOMEM) |
| 180 | error = -EAGAIN; |
| 181 | goto out; |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | success: |
| 186 | /* |
| 187 | * vm_flags is protected by the mmap_sem held in write mode. |
| 188 | */ |
| 189 | vma->vm_flags = new_flags; |
| 190 | out: |
| 191 | return error; |
| 192 | } |
| 193 | |
| 194 | #ifdef CONFIG_SWAP |
| 195 | static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start, |
| 196 | unsigned long end, struct mm_walk *walk) |
| 197 | { |
| 198 | pte_t *orig_pte; |
| 199 | struct vm_area_struct *vma = walk->private; |
| 200 | unsigned long index; |
| 201 | |
| 202 | if (pmd_none_or_trans_huge_or_clear_bad(pmd)) |
| 203 | return 0; |
| 204 | |
| 205 | for (index = start; index != end; index += PAGE_SIZE) { |
| 206 | pte_t pte; |
| 207 | swp_entry_t entry; |
| 208 | struct page *page; |
| 209 | spinlock_t *ptl; |
| 210 | |
| 211 | orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl); |
| 212 | pte = *(orig_pte + ((index - start) / PAGE_SIZE)); |
| 213 | pte_unmap_unlock(orig_pte, ptl); |
| 214 | |
| 215 | if (pte_present(pte) || pte_none(pte)) |
| 216 | continue; |
| 217 | entry = pte_to_swp_entry(pte); |
| 218 | if (unlikely(non_swap_entry(entry))) |
| 219 | continue; |
| 220 | |
| 221 | page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE, |
| 222 | vma, index, false); |
| 223 | if (page) |
| 224 | put_page(page); |
| 225 | } |
| 226 | |
| 227 | return 0; |
| 228 | } |
| 229 | |
| 230 | static const struct mm_walk_ops swapin_walk_ops = { |
| 231 | .pmd_entry = swapin_walk_pmd_entry, |
| 232 | }; |
| 233 | |
| 234 | static void force_shm_swapin_readahead(struct vm_area_struct *vma, |
| 235 | unsigned long start, unsigned long end, |
| 236 | struct address_space *mapping) |
| 237 | { |
| 238 | pgoff_t index; |
| 239 | struct page *page; |
| 240 | swp_entry_t swap; |
| 241 | |
| 242 | for (; start < end; start += PAGE_SIZE) { |
| 243 | index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; |
| 244 | |
| 245 | page = find_get_entry(mapping, index); |
| 246 | if (!xa_is_value(page)) { |
| 247 | if (page) |
| 248 | put_page(page); |
| 249 | continue; |
| 250 | } |
| 251 | swap = radix_to_swp_entry(page); |
| 252 | page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE, |
| 253 | NULL, 0, false); |
| 254 | if (page) |
| 255 | put_page(page); |
| 256 | } |
| 257 | |
| 258 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
| 259 | } |
| 260 | #endif /* CONFIG_SWAP */ |
| 261 | |
| 262 | /* |
| 263 | * Schedule all required I/O operations. Do not wait for completion. |
| 264 | */ |
| 265 | static long madvise_willneed(struct vm_area_struct *vma, |
| 266 | struct vm_area_struct **prev, |
| 267 | unsigned long start, unsigned long end) |
| 268 | { |
| 269 | struct file *file = vma->vm_file; |
| 270 | loff_t offset; |
| 271 | |
| 272 | *prev = vma; |
| 273 | #ifdef CONFIG_SWAP |
| 274 | if (!file) { |
| 275 | walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma); |
| 276 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
| 277 | return 0; |
| 278 | } |
| 279 | |
| 280 | if (shmem_mapping(file->f_mapping)) { |
| 281 | force_shm_swapin_readahead(vma, start, end, |
| 282 | file->f_mapping); |
| 283 | return 0; |
| 284 | } |
| 285 | #else |
| 286 | if (!file) |
| 287 | return -EBADF; |
| 288 | #endif |
| 289 | |
| 290 | if (IS_DAX(file_inode(file))) { |
| 291 | /* no bad return value, but ignore advice */ |
| 292 | return 0; |
| 293 | } |
| 294 | |
| 295 | /* |
| 296 | * Filesystem's fadvise may need to take various locks. We need to |
| 297 | * explicitly grab a reference because the vma (and hence the |
| 298 | * vma's reference to the file) can go away as soon as we drop |
| 299 | * mmap_sem. |
| 300 | */ |
| 301 | *prev = NULL; /* tell sys_madvise we drop mmap_sem */ |
| 302 | get_file(file); |
| 303 | up_read(¤t->mm->mmap_sem); |
| 304 | offset = (loff_t)(start - vma->vm_start) |
| 305 | + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); |
| 306 | vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED); |
| 307 | fput(file); |
| 308 | down_read(¤t->mm->mmap_sem); |
| 309 | return 0; |
| 310 | } |
| 311 | |
| 312 | static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr, |
| 313 | unsigned long end, struct mm_walk *walk) |
| 314 | |
| 315 | { |
| 316 | struct mmu_gather *tlb = walk->private; |
| 317 | struct mm_struct *mm = tlb->mm; |
| 318 | struct vm_area_struct *vma = walk->vma; |
| 319 | spinlock_t *ptl; |
| 320 | pte_t *orig_pte, *pte, ptent; |
| 321 | struct page *page; |
| 322 | int nr_swap = 0; |
| 323 | unsigned long next; |
| 324 | |
| 325 | next = pmd_addr_end(addr, end); |
| 326 | if (pmd_trans_huge(*pmd)) |
| 327 | if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next)) |
| 328 | goto next; |
| 329 | |
| 330 | if (pmd_trans_unstable(pmd)) |
| 331 | return 0; |
| 332 | |
| 333 | tlb_change_page_size(tlb, PAGE_SIZE); |
| 334 | orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl); |
| 335 | flush_tlb_batched_pending(mm); |
| 336 | arch_enter_lazy_mmu_mode(); |
| 337 | for (; addr != end; pte++, addr += PAGE_SIZE) { |
| 338 | ptent = *pte; |
| 339 | |
| 340 | if (pte_none(ptent)) |
| 341 | continue; |
| 342 | /* |
| 343 | * If the pte has swp_entry, just clear page table to |
| 344 | * prevent swap-in which is more expensive rather than |
| 345 | * (page allocation + zeroing). |
| 346 | */ |
| 347 | if (!pte_present(ptent)) { |
| 348 | swp_entry_t entry; |
| 349 | |
| 350 | entry = pte_to_swp_entry(ptent); |
| 351 | if (non_swap_entry(entry)) |
| 352 | continue; |
| 353 | nr_swap--; |
| 354 | free_swap_and_cache(entry); |
| 355 | pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); |
| 356 | continue; |
| 357 | } |
| 358 | |
| 359 | page = vm_normal_page(vma, addr, ptent); |
| 360 | if (!page) |
| 361 | continue; |
| 362 | |
| 363 | /* |
| 364 | * If pmd isn't transhuge but the page is THP and |
| 365 | * is owned by only this process, split it and |
| 366 | * deactivate all pages. |
| 367 | */ |
| 368 | if (PageTransCompound(page)) { |
| 369 | if (page_mapcount(page) != 1) |
| 370 | goto out; |
| 371 | get_page(page); |
| 372 | if (!trylock_page(page)) { |
| 373 | put_page(page); |
| 374 | goto out; |
| 375 | } |
| 376 | pte_unmap_unlock(orig_pte, ptl); |
| 377 | if (split_huge_page(page)) { |
| 378 | unlock_page(page); |
| 379 | put_page(page); |
| 380 | pte_offset_map_lock(mm, pmd, addr, &ptl); |
| 381 | goto out; |
| 382 | } |
| 383 | unlock_page(page); |
| 384 | put_page(page); |
| 385 | pte = pte_offset_map_lock(mm, pmd, addr, &ptl); |
| 386 | pte--; |
| 387 | addr -= PAGE_SIZE; |
| 388 | continue; |
| 389 | } |
| 390 | |
| 391 | VM_BUG_ON_PAGE(PageTransCompound(page), page); |
| 392 | |
| 393 | if (PageSwapCache(page) || PageDirty(page)) { |
| 394 | if (!trylock_page(page)) |
| 395 | continue; |
| 396 | /* |
| 397 | * If page is shared with others, we couldn't clear |
| 398 | * PG_dirty of the page. |
| 399 | */ |
| 400 | if (page_mapcount(page) != 1) { |
| 401 | unlock_page(page); |
| 402 | continue; |
| 403 | } |
| 404 | |
| 405 | if (PageSwapCache(page) && !try_to_free_swap(page)) { |
| 406 | unlock_page(page); |
| 407 | continue; |
| 408 | } |
| 409 | |
| 410 | ClearPageDirty(page); |
| 411 | unlock_page(page); |
| 412 | } |
| 413 | |
| 414 | if (pte_young(ptent) || pte_dirty(ptent)) { |
| 415 | /* |
| 416 | * Some of architecture(ex, PPC) don't update TLB |
| 417 | * with set_pte_at and tlb_remove_tlb_entry so for |
| 418 | * the portability, remap the pte with old|clean |
| 419 | * after pte clearing. |
| 420 | */ |
| 421 | ptent = ptep_get_and_clear_full(mm, addr, pte, |
| 422 | tlb->fullmm); |
| 423 | |
| 424 | ptent = pte_mkold(ptent); |
| 425 | ptent = pte_mkclean(ptent); |
| 426 | set_pte_at(mm, addr, pte, ptent); |
| 427 | tlb_remove_tlb_entry(tlb, pte, addr); |
| 428 | } |
| 429 | mark_page_lazyfree(page); |
| 430 | } |
| 431 | out: |
| 432 | if (nr_swap) { |
| 433 | if (current->mm == mm) |
| 434 | sync_mm_rss(mm); |
| 435 | |
| 436 | add_mm_counter(mm, MM_SWAPENTS, nr_swap); |
| 437 | } |
| 438 | arch_leave_lazy_mmu_mode(); |
| 439 | pte_unmap_unlock(orig_pte, ptl); |
| 440 | cond_resched(); |
| 441 | next: |
| 442 | return 0; |
| 443 | } |
| 444 | |
| 445 | static const struct mm_walk_ops madvise_free_walk_ops = { |
| 446 | .pmd_entry = madvise_free_pte_range, |
| 447 | }; |
| 448 | |
| 449 | static int madvise_free_single_vma(struct vm_area_struct *vma, |
| 450 | unsigned long start_addr, unsigned long end_addr) |
| 451 | { |
| 452 | struct mm_struct *mm = vma->vm_mm; |
| 453 | struct mmu_notifier_range range; |
| 454 | struct mmu_gather tlb; |
| 455 | |
| 456 | /* MADV_FREE works for only anon vma at the moment */ |
| 457 | if (!vma_is_anonymous(vma)) |
| 458 | return -EINVAL; |
| 459 | |
| 460 | range.start = max(vma->vm_start, start_addr); |
| 461 | if (range.start >= vma->vm_end) |
| 462 | return -EINVAL; |
| 463 | range.end = min(vma->vm_end, end_addr); |
| 464 | if (range.end <= vma->vm_start) |
| 465 | return -EINVAL; |
| 466 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, |
| 467 | range.start, range.end); |
| 468 | |
| 469 | lru_add_drain(); |
| 470 | tlb_gather_mmu(&tlb, mm, range.start, range.end); |
| 471 | update_hiwater_rss(mm); |
| 472 | |
| 473 | mmu_notifier_invalidate_range_start(&range); |
| 474 | tlb_start_vma(&tlb, vma); |
| 475 | walk_page_range(vma->vm_mm, range.start, range.end, |
| 476 | &madvise_free_walk_ops, &tlb); |
| 477 | tlb_end_vma(&tlb, vma); |
| 478 | mmu_notifier_invalidate_range_end(&range); |
| 479 | tlb_finish_mmu(&tlb, range.start, range.end); |
| 480 | |
| 481 | return 0; |
| 482 | } |
| 483 | |
| 484 | /* |
| 485 | * Application no longer needs these pages. If the pages are dirty, |
| 486 | * it's OK to just throw them away. The app will be more careful about |
| 487 | * data it wants to keep. Be sure to free swap resources too. The |
| 488 | * zap_page_range call sets things up for shrink_active_list to actually free |
| 489 | * these pages later if no one else has touched them in the meantime, |
| 490 | * although we could add these pages to a global reuse list for |
| 491 | * shrink_active_list to pick up before reclaiming other pages. |
| 492 | * |
| 493 | * NB: This interface discards data rather than pushes it out to swap, |
| 494 | * as some implementations do. This has performance implications for |
| 495 | * applications like large transactional databases which want to discard |
| 496 | * pages in anonymous maps after committing to backing store the data |
| 497 | * that was kept in them. There is no reason to write this data out to |
| 498 | * the swap area if the application is discarding it. |
| 499 | * |
| 500 | * An interface that causes the system to free clean pages and flush |
| 501 | * dirty pages is already available as msync(MS_INVALIDATE). |
| 502 | */ |
| 503 | static long madvise_dontneed_single_vma(struct vm_area_struct *vma, |
| 504 | unsigned long start, unsigned long end) |
| 505 | { |
| 506 | zap_page_range(vma, start, end - start); |
| 507 | return 0; |
| 508 | } |
| 509 | |
| 510 | static long madvise_dontneed_free(struct vm_area_struct *vma, |
| 511 | struct vm_area_struct **prev, |
| 512 | unsigned long start, unsigned long end, |
| 513 | int behavior) |
| 514 | { |
| 515 | *prev = vma; |
| 516 | if (!can_madv_dontneed_vma(vma)) |
| 517 | return -EINVAL; |
| 518 | |
| 519 | if (!userfaultfd_remove(vma, start, end)) { |
| 520 | *prev = NULL; /* mmap_sem has been dropped, prev is stale */ |
| 521 | |
| 522 | down_read(¤t->mm->mmap_sem); |
| 523 | vma = find_vma(current->mm, start); |
| 524 | if (!vma) |
| 525 | return -ENOMEM; |
| 526 | if (start < vma->vm_start) { |
| 527 | /* |
| 528 | * This "vma" under revalidation is the one |
| 529 | * with the lowest vma->vm_start where start |
| 530 | * is also < vma->vm_end. If start < |
| 531 | * vma->vm_start it means an hole materialized |
| 532 | * in the user address space within the |
| 533 | * virtual range passed to MADV_DONTNEED |
| 534 | * or MADV_FREE. |
| 535 | */ |
| 536 | return -ENOMEM; |
| 537 | } |
| 538 | if (!can_madv_dontneed_vma(vma)) |
| 539 | return -EINVAL; |
| 540 | if (end > vma->vm_end) { |
| 541 | /* |
| 542 | * Don't fail if end > vma->vm_end. If the old |
| 543 | * vma was splitted while the mmap_sem was |
| 544 | * released the effect of the concurrent |
| 545 | * operation may not cause madvise() to |
| 546 | * have an undefined result. There may be an |
| 547 | * adjacent next vma that we'll walk |
| 548 | * next. userfaultfd_remove() will generate an |
| 549 | * UFFD_EVENT_REMOVE repetition on the |
| 550 | * end-vma->vm_end range, but the manager can |
| 551 | * handle a repetition fine. |
| 552 | */ |
| 553 | end = vma->vm_end; |
| 554 | } |
| 555 | VM_WARN_ON(start >= end); |
| 556 | } |
| 557 | |
| 558 | if (behavior == MADV_DONTNEED) |
| 559 | return madvise_dontneed_single_vma(vma, start, end); |
| 560 | else if (behavior == MADV_FREE) |
| 561 | return madvise_free_single_vma(vma, start, end); |
| 562 | else |
| 563 | return -EINVAL; |
| 564 | } |
| 565 | |
| 566 | /* |
| 567 | * Application wants to free up the pages and associated backing store. |
| 568 | * This is effectively punching a hole into the middle of a file. |
| 569 | */ |
| 570 | static long madvise_remove(struct vm_area_struct *vma, |
| 571 | struct vm_area_struct **prev, |
| 572 | unsigned long start, unsigned long end) |
| 573 | { |
| 574 | loff_t offset; |
| 575 | int error; |
| 576 | struct file *f; |
| 577 | |
| 578 | *prev = NULL; /* tell sys_madvise we drop mmap_sem */ |
| 579 | |
| 580 | if (vma->vm_flags & VM_LOCKED) |
| 581 | return -EINVAL; |
| 582 | |
| 583 | f = vma->vm_file; |
| 584 | |
| 585 | if (!f || !f->f_mapping || !f->f_mapping->host) { |
| 586 | return -EINVAL; |
| 587 | } |
| 588 | |
| 589 | if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE)) |
| 590 | return -EACCES; |
| 591 | |
| 592 | offset = (loff_t)(start - vma->vm_start) |
| 593 | + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); |
| 594 | |
| 595 | /* |
| 596 | * Filesystem's fallocate may need to take i_mutex. We need to |
| 597 | * explicitly grab a reference because the vma (and hence the |
| 598 | * vma's reference to the file) can go away as soon as we drop |
| 599 | * mmap_sem. |
| 600 | */ |
| 601 | get_file(f); |
| 602 | if (userfaultfd_remove(vma, start, end)) { |
| 603 | /* mmap_sem was not released by userfaultfd_remove() */ |
| 604 | up_read(¤t->mm->mmap_sem); |
| 605 | } |
| 606 | error = vfs_fallocate(f, |
| 607 | FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, |
| 608 | offset, end - start); |
| 609 | fput(f); |
| 610 | down_read(¤t->mm->mmap_sem); |
| 611 | return error; |
| 612 | } |
| 613 | |
| 614 | #ifdef CONFIG_MEMORY_FAILURE |
| 615 | /* |
| 616 | * Error injection support for memory error handling. |
| 617 | */ |
| 618 | static int madvise_inject_error(int behavior, |
| 619 | unsigned long start, unsigned long end) |
| 620 | { |
| 621 | struct page *page; |
| 622 | struct zone *zone; |
| 623 | unsigned int order; |
| 624 | |
| 625 | if (!capable(CAP_SYS_ADMIN)) |
| 626 | return -EPERM; |
| 627 | |
| 628 | |
| 629 | for (; start < end; start += PAGE_SIZE << order) { |
| 630 | unsigned long pfn; |
| 631 | int ret; |
| 632 | |
| 633 | ret = get_user_pages_fast(start, 1, 0, &page); |
| 634 | if (ret != 1) |
| 635 | return ret; |
| 636 | pfn = page_to_pfn(page); |
| 637 | |
| 638 | /* |
| 639 | * When soft offlining hugepages, after migrating the page |
| 640 | * we dissolve it, therefore in the second loop "page" will |
| 641 | * no longer be a compound page, and order will be 0. |
| 642 | */ |
| 643 | order = compound_order(compound_head(page)); |
| 644 | |
| 645 | if (PageHWPoison(page)) { |
| 646 | put_page(page); |
| 647 | continue; |
| 648 | } |
| 649 | |
| 650 | if (behavior == MADV_SOFT_OFFLINE) { |
| 651 | pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n", |
| 652 | pfn, start); |
| 653 | |
| 654 | ret = soft_offline_page(page, MF_COUNT_INCREASED); |
| 655 | if (ret) |
| 656 | return ret; |
| 657 | continue; |
| 658 | } |
| 659 | |
| 660 | pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n", |
| 661 | pfn, start); |
| 662 | |
| 663 | /* |
| 664 | * Drop the page reference taken by get_user_pages_fast(). In |
| 665 | * the absence of MF_COUNT_INCREASED the memory_failure() |
| 666 | * routine is responsible for pinning the page to prevent it |
| 667 | * from being released back to the page allocator. |
| 668 | */ |
| 669 | put_page(page); |
| 670 | ret = memory_failure(pfn, 0); |
| 671 | if (ret) |
| 672 | return ret; |
| 673 | } |
| 674 | |
| 675 | /* Ensure that all poisoned pages are removed from per-cpu lists */ |
| 676 | for_each_populated_zone(zone) |
| 677 | drain_all_pages(zone); |
| 678 | |
| 679 | return 0; |
| 680 | } |
| 681 | #endif |
| 682 | |
| 683 | static long |
| 684 | madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev, |
| 685 | unsigned long start, unsigned long end, int behavior) |
| 686 | { |
| 687 | switch (behavior) { |
| 688 | case MADV_REMOVE: |
| 689 | return madvise_remove(vma, prev, start, end); |
| 690 | case MADV_WILLNEED: |
| 691 | return madvise_willneed(vma, prev, start, end); |
| 692 | case MADV_FREE: |
| 693 | case MADV_DONTNEED: |
| 694 | return madvise_dontneed_free(vma, prev, start, end, behavior); |
| 695 | default: |
| 696 | return madvise_behavior(vma, prev, start, end, behavior); |
| 697 | } |
| 698 | } |
| 699 | |
| 700 | static bool |
| 701 | madvise_behavior_valid(int behavior) |
| 702 | { |
| 703 | switch (behavior) { |
| 704 | case MADV_DOFORK: |
| 705 | case MADV_DONTFORK: |
| 706 | case MADV_NORMAL: |
| 707 | case MADV_SEQUENTIAL: |
| 708 | case MADV_RANDOM: |
| 709 | case MADV_REMOVE: |
| 710 | case MADV_WILLNEED: |
| 711 | case MADV_DONTNEED: |
| 712 | case MADV_FREE: |
| 713 | #ifdef CONFIG_KSM |
| 714 | case MADV_MERGEABLE: |
| 715 | case MADV_UNMERGEABLE: |
| 716 | #endif |
| 717 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 718 | case MADV_HUGEPAGE: |
| 719 | case MADV_NOHUGEPAGE: |
| 720 | #endif |
| 721 | case MADV_DONTDUMP: |
| 722 | case MADV_DODUMP: |
| 723 | case MADV_WIPEONFORK: |
| 724 | case MADV_KEEPONFORK: |
| 725 | #ifdef CONFIG_MEMORY_FAILURE |
| 726 | case MADV_SOFT_OFFLINE: |
| 727 | case MADV_HWPOISON: |
| 728 | #endif |
| 729 | return true; |
| 730 | |
| 731 | default: |
| 732 | return false; |
| 733 | } |
| 734 | } |
| 735 | |
| 736 | /* |
| 737 | * The madvise(2) system call. |
| 738 | * |
| 739 | * Applications can use madvise() to advise the kernel how it should |
| 740 | * handle paging I/O in this VM area. The idea is to help the kernel |
| 741 | * use appropriate read-ahead and caching techniques. The information |
| 742 | * provided is advisory only, and can be safely disregarded by the |
| 743 | * kernel without affecting the correct operation of the application. |
| 744 | * |
| 745 | * behavior values: |
| 746 | * MADV_NORMAL - the default behavior is to read clusters. This |
| 747 | * results in some read-ahead and read-behind. |
| 748 | * MADV_RANDOM - the system should read the minimum amount of data |
| 749 | * on any access, since it is unlikely that the appli- |
| 750 | * cation will need more than what it asks for. |
| 751 | * MADV_SEQUENTIAL - pages in the given range will probably be accessed |
| 752 | * once, so they can be aggressively read ahead, and |
| 753 | * can be freed soon after they are accessed. |
| 754 | * MADV_WILLNEED - the application is notifying the system to read |
| 755 | * some pages ahead. |
| 756 | * MADV_DONTNEED - the application is finished with the given range, |
| 757 | * so the kernel can free resources associated with it. |
| 758 | * MADV_FREE - the application marks pages in the given range as lazy free, |
| 759 | * where actual purges are postponed until memory pressure happens. |
| 760 | * MADV_REMOVE - the application wants to free up the given range of |
| 761 | * pages and associated backing store. |
| 762 | * MADV_DONTFORK - omit this area from child's address space when forking: |
| 763 | * typically, to avoid COWing pages pinned by get_user_pages(). |
| 764 | * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking. |
| 765 | * MADV_WIPEONFORK - present the child process with zero-filled memory in this |
| 766 | * range after a fork. |
| 767 | * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK |
| 768 | * MADV_HWPOISON - trigger memory error handler as if the given memory range |
| 769 | * were corrupted by unrecoverable hardware memory failure. |
| 770 | * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory. |
| 771 | * MADV_MERGEABLE - the application recommends that KSM try to merge pages in |
| 772 | * this area with pages of identical content from other such areas. |
| 773 | * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others. |
| 774 | * MADV_HUGEPAGE - the application wants to back the given range by transparent |
| 775 | * huge pages in the future. Existing pages might be coalesced and |
| 776 | * new pages might be allocated as THP. |
| 777 | * MADV_NOHUGEPAGE - mark the given range as not worth being backed by |
| 778 | * transparent huge pages so the existing pages will not be |
| 779 | * coalesced into THP and new pages will not be allocated as THP. |
| 780 | * MADV_DONTDUMP - the application wants to prevent pages in the given range |
| 781 | * from being included in its core dump. |
| 782 | * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump. |
| 783 | * |
| 784 | * return values: |
| 785 | * zero - success |
| 786 | * -EINVAL - start + len < 0, start is not page-aligned, |
| 787 | * "behavior" is not a valid value, or application |
| 788 | * is attempting to release locked or shared pages, |
| 789 | * or the specified address range includes file, Huge TLB, |
| 790 | * MAP_SHARED or VMPFNMAP range. |
| 791 | * -ENOMEM - addresses in the specified range are not currently |
| 792 | * mapped, or are outside the AS of the process. |
| 793 | * -EIO - an I/O error occurred while paging in data. |
| 794 | * -EBADF - map exists, but area maps something that isn't a file. |
| 795 | * -EAGAIN - a kernel resource was temporarily unavailable. |
| 796 | */ |
| 797 | SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior) |
| 798 | { |
| 799 | unsigned long end, tmp; |
| 800 | struct vm_area_struct *vma, *prev; |
| 801 | int unmapped_error = 0; |
| 802 | int error = -EINVAL; |
| 803 | int write; |
| 804 | size_t len; |
| 805 | struct blk_plug plug; |
| 806 | |
| 807 | if (!madvise_behavior_valid(behavior)) |
| 808 | return error; |
| 809 | |
| 810 | if (start & ~PAGE_MASK) |
| 811 | return error; |
| 812 | len = (len_in + ~PAGE_MASK) & PAGE_MASK; |
| 813 | |
| 814 | /* Check to see whether len was rounded up from small -ve to zero */ |
| 815 | if (len_in && !len) |
| 816 | return error; |
| 817 | |
| 818 | end = start + len; |
| 819 | if (end < start) |
| 820 | return error; |
| 821 | |
| 822 | error = 0; |
| 823 | if (end == start) |
| 824 | return error; |
| 825 | |
| 826 | #ifdef CONFIG_MEMORY_FAILURE |
| 827 | if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE) |
| 828 | return madvise_inject_error(behavior, start, start + len_in); |
| 829 | #endif |
| 830 | |
| 831 | write = madvise_need_mmap_write(behavior); |
| 832 | if (write) { |
| 833 | if (down_write_killable(¤t->mm->mmap_sem)) |
| 834 | return -EINTR; |
| 835 | } else { |
| 836 | down_read(¤t->mm->mmap_sem); |
| 837 | } |
| 838 | |
| 839 | /* |
| 840 | * If the interval [start,end) covers some unmapped address |
| 841 | * ranges, just ignore them, but return -ENOMEM at the end. |
| 842 | * - different from the way of handling in mlock etc. |
| 843 | */ |
| 844 | vma = find_vma_prev(current->mm, start, &prev); |
| 845 | if (vma && start > vma->vm_start) |
| 846 | prev = vma; |
| 847 | |
| 848 | blk_start_plug(&plug); |
| 849 | for (;;) { |
| 850 | /* Still start < end. */ |
| 851 | error = -ENOMEM; |
| 852 | if (!vma) |
| 853 | goto out; |
| 854 | |
| 855 | /* Here start < (end|vma->vm_end). */ |
| 856 | if (start < vma->vm_start) { |
| 857 | unmapped_error = -ENOMEM; |
| 858 | start = vma->vm_start; |
| 859 | if (start >= end) |
| 860 | goto out; |
| 861 | } |
| 862 | |
| 863 | /* Here vma->vm_start <= start < (end|vma->vm_end) */ |
| 864 | tmp = vma->vm_end; |
| 865 | if (end < tmp) |
| 866 | tmp = end; |
| 867 | |
| 868 | /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */ |
| 869 | error = madvise_vma(vma, &prev, start, tmp, behavior); |
| 870 | if (error) |
| 871 | goto out; |
| 872 | start = tmp; |
| 873 | if (prev && start < prev->vm_end) |
| 874 | start = prev->vm_end; |
| 875 | error = unmapped_error; |
| 876 | if (start >= end) |
| 877 | goto out; |
| 878 | if (prev) |
| 879 | vma = prev->vm_next; |
| 880 | else /* madvise_remove dropped mmap_sem */ |
| 881 | vma = find_vma(current->mm, start); |
| 882 | } |
| 883 | out: |
| 884 | blk_finish_plug(&plug); |
| 885 | if (write) |
| 886 | up_write(¤t->mm->mmap_sem); |
| 887 | else |
| 888 | up_read(¤t->mm->mmap_sem); |
| 889 | |
| 890 | return error; |
| 891 | } |