| 1 | /* |
| 2 | * linux/mm/madvise.c |
| 3 | * |
| 4 | * Copyright (C) 1999 Linus Torvalds |
| 5 | * Copyright (C) 2002 Christoph Hellwig |
| 6 | */ |
| 7 | |
| 8 | #include <linux/mman.h> |
| 9 | #include <linux/pagemap.h> |
| 10 | #include <linux/syscalls.h> |
| 11 | #include <linux/mempolicy.h> |
| 12 | #include <linux/page-isolation.h> |
| 13 | #include <linux/hugetlb.h> |
| 14 | #include <linux/falloc.h> |
| 15 | #include <linux/sched.h> |
| 16 | #include <linux/ksm.h> |
| 17 | #include <linux/fs.h> |
| 18 | #include <linux/file.h> |
| 19 | #include <linux/blkdev.h> |
| 20 | #include <linux/backing-dev.h> |
| 21 | #include <linux/swap.h> |
| 22 | #include <linux/swapops.h> |
| 23 | |
| 24 | /* |
| 25 | * Any behaviour which results in changes to the vma->vm_flags needs to |
| 26 | * take mmap_sem for writing. Others, which simply traverse vmas, need |
| 27 | * to only take it for reading. |
| 28 | */ |
| 29 | static int madvise_need_mmap_write(int behavior) |
| 30 | { |
| 31 | switch (behavior) { |
| 32 | case MADV_REMOVE: |
| 33 | case MADV_WILLNEED: |
| 34 | case MADV_DONTNEED: |
| 35 | return 0; |
| 36 | default: |
| 37 | /* be safe, default to 1. list exceptions explicitly */ |
| 38 | return 1; |
| 39 | } |
| 40 | } |
| 41 | |
| 42 | /* |
| 43 | * We can potentially split a vm area into separate |
| 44 | * areas, each area with its own behavior. |
| 45 | */ |
| 46 | static long madvise_behavior(struct vm_area_struct *vma, |
| 47 | struct vm_area_struct **prev, |
| 48 | unsigned long start, unsigned long end, int behavior) |
| 49 | { |
| 50 | struct mm_struct *mm = vma->vm_mm; |
| 51 | int error = 0; |
| 52 | pgoff_t pgoff; |
| 53 | unsigned long new_flags = vma->vm_flags; |
| 54 | |
| 55 | switch (behavior) { |
| 56 | case MADV_NORMAL: |
| 57 | new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ; |
| 58 | break; |
| 59 | case MADV_SEQUENTIAL: |
| 60 | new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ; |
| 61 | break; |
| 62 | case MADV_RANDOM: |
| 63 | new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ; |
| 64 | break; |
| 65 | case MADV_DONTFORK: |
| 66 | new_flags |= VM_DONTCOPY; |
| 67 | break; |
| 68 | case MADV_DOFORK: |
| 69 | if (vma->vm_flags & VM_IO) { |
| 70 | error = -EINVAL; |
| 71 | goto out; |
| 72 | } |
| 73 | new_flags &= ~VM_DONTCOPY; |
| 74 | break; |
| 75 | case MADV_DONTDUMP: |
| 76 | new_flags |= VM_DONTDUMP; |
| 77 | break; |
| 78 | case MADV_DODUMP: |
| 79 | if (new_flags & VM_SPECIAL) { |
| 80 | error = -EINVAL; |
| 81 | goto out; |
| 82 | } |
| 83 | new_flags &= ~VM_DONTDUMP; |
| 84 | break; |
| 85 | case MADV_MERGEABLE: |
| 86 | case MADV_UNMERGEABLE: |
| 87 | error = ksm_madvise(vma, start, end, behavior, &new_flags); |
| 88 | if (error) |
| 89 | goto out; |
| 90 | break; |
| 91 | case MADV_HUGEPAGE: |
| 92 | case MADV_NOHUGEPAGE: |
| 93 | error = hugepage_madvise(vma, &new_flags, behavior); |
| 94 | if (error) |
| 95 | goto out; |
| 96 | break; |
| 97 | } |
| 98 | |
| 99 | if (new_flags == vma->vm_flags) { |
| 100 | *prev = vma; |
| 101 | goto out; |
| 102 | } |
| 103 | |
| 104 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); |
| 105 | *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma, |
| 106 | vma->vm_file, pgoff, vma_policy(vma)); |
| 107 | if (*prev) { |
| 108 | vma = *prev; |
| 109 | goto success; |
| 110 | } |
| 111 | |
| 112 | *prev = vma; |
| 113 | |
| 114 | if (start != vma->vm_start) { |
| 115 | error = split_vma(mm, vma, start, 1); |
| 116 | if (error) |
| 117 | goto out; |
| 118 | } |
| 119 | |
| 120 | if (end != vma->vm_end) { |
| 121 | error = split_vma(mm, vma, end, 0); |
| 122 | if (error) |
| 123 | goto out; |
| 124 | } |
| 125 | |
| 126 | success: |
| 127 | /* |
| 128 | * vm_flags is protected by the mmap_sem held in write mode. |
| 129 | */ |
| 130 | vma->vm_flags = new_flags; |
| 131 | |
| 132 | out: |
| 133 | if (error == -ENOMEM) |
| 134 | error = -EAGAIN; |
| 135 | return error; |
| 136 | } |
| 137 | |
| 138 | #ifdef CONFIG_SWAP |
| 139 | static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start, |
| 140 | unsigned long end, struct mm_walk *walk) |
| 141 | { |
| 142 | pte_t *orig_pte; |
| 143 | struct vm_area_struct *vma = walk->private; |
| 144 | unsigned long index; |
| 145 | |
| 146 | if (pmd_none_or_trans_huge_or_clear_bad(pmd)) |
| 147 | return 0; |
| 148 | |
| 149 | for (index = start; index != end; index += PAGE_SIZE) { |
| 150 | pte_t pte; |
| 151 | swp_entry_t entry; |
| 152 | struct page *page; |
| 153 | spinlock_t *ptl; |
| 154 | |
| 155 | orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl); |
| 156 | pte = *(orig_pte + ((index - start) / PAGE_SIZE)); |
| 157 | pte_unmap_unlock(orig_pte, ptl); |
| 158 | |
| 159 | if (pte_present(pte) || pte_none(pte)) |
| 160 | continue; |
| 161 | entry = pte_to_swp_entry(pte); |
| 162 | if (unlikely(non_swap_entry(entry))) |
| 163 | continue; |
| 164 | |
| 165 | page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE, |
| 166 | vma, index); |
| 167 | if (page) |
| 168 | page_cache_release(page); |
| 169 | } |
| 170 | |
| 171 | return 0; |
| 172 | } |
| 173 | |
| 174 | static void force_swapin_readahead(struct vm_area_struct *vma, |
| 175 | unsigned long start, unsigned long end) |
| 176 | { |
| 177 | struct mm_walk walk = { |
| 178 | .mm = vma->vm_mm, |
| 179 | .pmd_entry = swapin_walk_pmd_entry, |
| 180 | .private = vma, |
| 181 | }; |
| 182 | |
| 183 | walk_page_range(start, end, &walk); |
| 184 | |
| 185 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
| 186 | } |
| 187 | |
| 188 | static void force_shm_swapin_readahead(struct vm_area_struct *vma, |
| 189 | unsigned long start, unsigned long end, |
| 190 | struct address_space *mapping) |
| 191 | { |
| 192 | pgoff_t index; |
| 193 | struct page *page; |
| 194 | swp_entry_t swap; |
| 195 | |
| 196 | for (; start < end; start += PAGE_SIZE) { |
| 197 | index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; |
| 198 | |
| 199 | page = find_get_entry(mapping, index); |
| 200 | if (!radix_tree_exceptional_entry(page)) { |
| 201 | if (page) |
| 202 | page_cache_release(page); |
| 203 | continue; |
| 204 | } |
| 205 | swap = radix_to_swp_entry(page); |
| 206 | page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE, |
| 207 | NULL, 0); |
| 208 | if (page) |
| 209 | page_cache_release(page); |
| 210 | } |
| 211 | |
| 212 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
| 213 | } |
| 214 | #endif /* CONFIG_SWAP */ |
| 215 | |
| 216 | /* |
| 217 | * Schedule all required I/O operations. Do not wait for completion. |
| 218 | */ |
| 219 | static long madvise_willneed(struct vm_area_struct *vma, |
| 220 | struct vm_area_struct **prev, |
| 221 | unsigned long start, unsigned long end) |
| 222 | { |
| 223 | struct file *file = vma->vm_file; |
| 224 | |
| 225 | #ifdef CONFIG_SWAP |
| 226 | if (!file) { |
| 227 | *prev = vma; |
| 228 | force_swapin_readahead(vma, start, end); |
| 229 | return 0; |
| 230 | } |
| 231 | |
| 232 | if (shmem_mapping(file->f_mapping)) { |
| 233 | *prev = vma; |
| 234 | force_shm_swapin_readahead(vma, start, end, |
| 235 | file->f_mapping); |
| 236 | return 0; |
| 237 | } |
| 238 | #else |
| 239 | if (!file) |
| 240 | return -EBADF; |
| 241 | #endif |
| 242 | |
| 243 | if (IS_DAX(file_inode(file))) { |
| 244 | /* no bad return value, but ignore advice */ |
| 245 | return 0; |
| 246 | } |
| 247 | |
| 248 | *prev = vma; |
| 249 | start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; |
| 250 | if (end > vma->vm_end) |
| 251 | end = vma->vm_end; |
| 252 | end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; |
| 253 | |
| 254 | force_page_cache_readahead(file->f_mapping, file, start, end - start); |
| 255 | return 0; |
| 256 | } |
| 257 | |
| 258 | /* |
| 259 | * Application no longer needs these pages. If the pages are dirty, |
| 260 | * it's OK to just throw them away. The app will be more careful about |
| 261 | * data it wants to keep. Be sure to free swap resources too. The |
| 262 | * zap_page_range call sets things up for shrink_active_list to actually free |
| 263 | * these pages later if no one else has touched them in the meantime, |
| 264 | * although we could add these pages to a global reuse list for |
| 265 | * shrink_active_list to pick up before reclaiming other pages. |
| 266 | * |
| 267 | * NB: This interface discards data rather than pushes it out to swap, |
| 268 | * as some implementations do. This has performance implications for |
| 269 | * applications like large transactional databases which want to discard |
| 270 | * pages in anonymous maps after committing to backing store the data |
| 271 | * that was kept in them. There is no reason to write this data out to |
| 272 | * the swap area if the application is discarding it. |
| 273 | * |
| 274 | * An interface that causes the system to free clean pages and flush |
| 275 | * dirty pages is already available as msync(MS_INVALIDATE). |
| 276 | */ |
| 277 | static long madvise_dontneed(struct vm_area_struct *vma, |
| 278 | struct vm_area_struct **prev, |
| 279 | unsigned long start, unsigned long end) |
| 280 | { |
| 281 | *prev = vma; |
| 282 | if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP)) |
| 283 | return -EINVAL; |
| 284 | |
| 285 | zap_page_range(vma, start, end - start, NULL); |
| 286 | return 0; |
| 287 | } |
| 288 | |
| 289 | /* |
| 290 | * Application wants to free up the pages and associated backing store. |
| 291 | * This is effectively punching a hole into the middle of a file. |
| 292 | */ |
| 293 | static long madvise_remove(struct vm_area_struct *vma, |
| 294 | struct vm_area_struct **prev, |
| 295 | unsigned long start, unsigned long end) |
| 296 | { |
| 297 | loff_t offset; |
| 298 | int error; |
| 299 | struct file *f; |
| 300 | |
| 301 | *prev = NULL; /* tell sys_madvise we drop mmap_sem */ |
| 302 | |
| 303 | if (vma->vm_flags & (VM_LOCKED | VM_HUGETLB)) |
| 304 | return -EINVAL; |
| 305 | |
| 306 | f = vma->vm_file; |
| 307 | |
| 308 | if (!f || !f->f_mapping || !f->f_mapping->host) { |
| 309 | return -EINVAL; |
| 310 | } |
| 311 | |
| 312 | if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE)) |
| 313 | return -EACCES; |
| 314 | |
| 315 | offset = (loff_t)(start - vma->vm_start) |
| 316 | + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); |
| 317 | |
| 318 | /* |
| 319 | * Filesystem's fallocate may need to take i_mutex. We need to |
| 320 | * explicitly grab a reference because the vma (and hence the |
| 321 | * vma's reference to the file) can go away as soon as we drop |
| 322 | * mmap_sem. |
| 323 | */ |
| 324 | get_file(f); |
| 325 | up_read(¤t->mm->mmap_sem); |
| 326 | error = vfs_fallocate(f, |
| 327 | FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, |
| 328 | offset, end - start); |
| 329 | fput(f); |
| 330 | down_read(¤t->mm->mmap_sem); |
| 331 | return error; |
| 332 | } |
| 333 | |
| 334 | #ifdef CONFIG_MEMORY_FAILURE |
| 335 | /* |
| 336 | * Error injection support for memory error handling. |
| 337 | */ |
| 338 | static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end) |
| 339 | { |
| 340 | struct page *p; |
| 341 | if (!capable(CAP_SYS_ADMIN)) |
| 342 | return -EPERM; |
| 343 | for (; start < end; start += PAGE_SIZE << |
| 344 | compound_order(compound_head(p))) { |
| 345 | int ret; |
| 346 | |
| 347 | ret = get_user_pages_fast(start, 1, 0, &p); |
| 348 | if (ret != 1) |
| 349 | return ret; |
| 350 | |
| 351 | if (PageHWPoison(p)) { |
| 352 | put_page(p); |
| 353 | continue; |
| 354 | } |
| 355 | if (bhv == MADV_SOFT_OFFLINE) { |
| 356 | pr_info("Soft offlining page %#lx at %#lx\n", |
| 357 | page_to_pfn(p), start); |
| 358 | ret = soft_offline_page(p, MF_COUNT_INCREASED); |
| 359 | if (ret) |
| 360 | return ret; |
| 361 | continue; |
| 362 | } |
| 363 | pr_info("Injecting memory failure for page %#lx at %#lx\n", |
| 364 | page_to_pfn(p), start); |
| 365 | /* Ignore return value for now */ |
| 366 | memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED); |
| 367 | } |
| 368 | return 0; |
| 369 | } |
| 370 | #endif |
| 371 | |
| 372 | static long |
| 373 | madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev, |
| 374 | unsigned long start, unsigned long end, int behavior) |
| 375 | { |
| 376 | switch (behavior) { |
| 377 | case MADV_REMOVE: |
| 378 | return madvise_remove(vma, prev, start, end); |
| 379 | case MADV_WILLNEED: |
| 380 | return madvise_willneed(vma, prev, start, end); |
| 381 | case MADV_DONTNEED: |
| 382 | return madvise_dontneed(vma, prev, start, end); |
| 383 | default: |
| 384 | return madvise_behavior(vma, prev, start, end, behavior); |
| 385 | } |
| 386 | } |
| 387 | |
| 388 | static int |
| 389 | madvise_behavior_valid(int behavior) |
| 390 | { |
| 391 | switch (behavior) { |
| 392 | case MADV_DOFORK: |
| 393 | case MADV_DONTFORK: |
| 394 | case MADV_NORMAL: |
| 395 | case MADV_SEQUENTIAL: |
| 396 | case MADV_RANDOM: |
| 397 | case MADV_REMOVE: |
| 398 | case MADV_WILLNEED: |
| 399 | case MADV_DONTNEED: |
| 400 | #ifdef CONFIG_KSM |
| 401 | case MADV_MERGEABLE: |
| 402 | case MADV_UNMERGEABLE: |
| 403 | #endif |
| 404 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 405 | case MADV_HUGEPAGE: |
| 406 | case MADV_NOHUGEPAGE: |
| 407 | #endif |
| 408 | case MADV_DONTDUMP: |
| 409 | case MADV_DODUMP: |
| 410 | return 1; |
| 411 | |
| 412 | default: |
| 413 | return 0; |
| 414 | } |
| 415 | } |
| 416 | |
| 417 | /* |
| 418 | * The madvise(2) system call. |
| 419 | * |
| 420 | * Applications can use madvise() to advise the kernel how it should |
| 421 | * handle paging I/O in this VM area. The idea is to help the kernel |
| 422 | * use appropriate read-ahead and caching techniques. The information |
| 423 | * provided is advisory only, and can be safely disregarded by the |
| 424 | * kernel without affecting the correct operation of the application. |
| 425 | * |
| 426 | * behavior values: |
| 427 | * MADV_NORMAL - the default behavior is to read clusters. This |
| 428 | * results in some read-ahead and read-behind. |
| 429 | * MADV_RANDOM - the system should read the minimum amount of data |
| 430 | * on any access, since it is unlikely that the appli- |
| 431 | * cation will need more than what it asks for. |
| 432 | * MADV_SEQUENTIAL - pages in the given range will probably be accessed |
| 433 | * once, so they can be aggressively read ahead, and |
| 434 | * can be freed soon after they are accessed. |
| 435 | * MADV_WILLNEED - the application is notifying the system to read |
| 436 | * some pages ahead. |
| 437 | * MADV_DONTNEED - the application is finished with the given range, |
| 438 | * so the kernel can free resources associated with it. |
| 439 | * MADV_REMOVE - the application wants to free up the given range of |
| 440 | * pages and associated backing store. |
| 441 | * MADV_DONTFORK - omit this area from child's address space when forking: |
| 442 | * typically, to avoid COWing pages pinned by get_user_pages(). |
| 443 | * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking. |
| 444 | * MADV_MERGEABLE - the application recommends that KSM try to merge pages in |
| 445 | * this area with pages of identical content from other such areas. |
| 446 | * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others. |
| 447 | * |
| 448 | * return values: |
| 449 | * zero - success |
| 450 | * -EINVAL - start + len < 0, start is not page-aligned, |
| 451 | * "behavior" is not a valid value, or application |
| 452 | * is attempting to release locked or shared pages. |
| 453 | * -ENOMEM - addresses in the specified range are not currently |
| 454 | * mapped, or are outside the AS of the process. |
| 455 | * -EIO - an I/O error occurred while paging in data. |
| 456 | * -EBADF - map exists, but area maps something that isn't a file. |
| 457 | * -EAGAIN - a kernel resource was temporarily unavailable. |
| 458 | */ |
| 459 | SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior) |
| 460 | { |
| 461 | unsigned long end, tmp; |
| 462 | struct vm_area_struct *vma, *prev; |
| 463 | int unmapped_error = 0; |
| 464 | int error = -EINVAL; |
| 465 | int write; |
| 466 | size_t len; |
| 467 | struct blk_plug plug; |
| 468 | |
| 469 | #ifdef CONFIG_MEMORY_FAILURE |
| 470 | if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE) |
| 471 | return madvise_hwpoison(behavior, start, start+len_in); |
| 472 | #endif |
| 473 | if (!madvise_behavior_valid(behavior)) |
| 474 | return error; |
| 475 | |
| 476 | if (start & ~PAGE_MASK) |
| 477 | return error; |
| 478 | len = (len_in + ~PAGE_MASK) & PAGE_MASK; |
| 479 | |
| 480 | /* Check to see whether len was rounded up from small -ve to zero */ |
| 481 | if (len_in && !len) |
| 482 | return error; |
| 483 | |
| 484 | end = start + len; |
| 485 | if (end < start) |
| 486 | return error; |
| 487 | |
| 488 | error = 0; |
| 489 | if (end == start) |
| 490 | return error; |
| 491 | |
| 492 | write = madvise_need_mmap_write(behavior); |
| 493 | if (write) |
| 494 | down_write(¤t->mm->mmap_sem); |
| 495 | else |
| 496 | down_read(¤t->mm->mmap_sem); |
| 497 | |
| 498 | /* |
| 499 | * If the interval [start,end) covers some unmapped address |
| 500 | * ranges, just ignore them, but return -ENOMEM at the end. |
| 501 | * - different from the way of handling in mlock etc. |
| 502 | */ |
| 503 | vma = find_vma_prev(current->mm, start, &prev); |
| 504 | if (vma && start > vma->vm_start) |
| 505 | prev = vma; |
| 506 | |
| 507 | blk_start_plug(&plug); |
| 508 | for (;;) { |
| 509 | /* Still start < end. */ |
| 510 | error = -ENOMEM; |
| 511 | if (!vma) |
| 512 | goto out; |
| 513 | |
| 514 | /* Here start < (end|vma->vm_end). */ |
| 515 | if (start < vma->vm_start) { |
| 516 | unmapped_error = -ENOMEM; |
| 517 | start = vma->vm_start; |
| 518 | if (start >= end) |
| 519 | goto out; |
| 520 | } |
| 521 | |
| 522 | /* Here vma->vm_start <= start < (end|vma->vm_end) */ |
| 523 | tmp = vma->vm_end; |
| 524 | if (end < tmp) |
| 525 | tmp = end; |
| 526 | |
| 527 | /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */ |
| 528 | error = madvise_vma(vma, &prev, start, tmp, behavior); |
| 529 | if (error) |
| 530 | goto out; |
| 531 | start = tmp; |
| 532 | if (prev && start < prev->vm_end) |
| 533 | start = prev->vm_end; |
| 534 | error = unmapped_error; |
| 535 | if (start >= end) |
| 536 | goto out; |
| 537 | if (prev) |
| 538 | vma = prev->vm_next; |
| 539 | else /* madvise_remove dropped mmap_sem */ |
| 540 | vma = find_vma(current->mm, start); |
| 541 | } |
| 542 | out: |
| 543 | blk_finish_plug(&plug); |
| 544 | if (write) |
| 545 | up_write(¤t->mm->mmap_sem); |
| 546 | else |
| 547 | up_read(¤t->mm->mmap_sem); |
| 548 | |
| 549 | return error; |
| 550 | } |