mm/madvise.c

   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                           vma->vm_userfaultfd_ctx);
 108         if (*prev) {
 109                 vma = *prev;
 110                 goto success;
 111         }
 112
 113         *prev = vma;
 114
 115         if (start != vma->vm_start) {
 116                 error = split_vma(mm, vma, start, 1);
 117                 if (error)
 118                         goto out;
 119         }
 120
 121         if (end != vma->vm_end) {
 122                 error = split_vma(mm, vma, end, 0);
 123                 if (error)
 124                         goto out;
 125         }
 126
 127 success:
 128         /*
 129          * vm_flags is protected by the mmap_sem held in write mode.
 130          */
 131         vma->vm_flags = new_flags;
 132
 133 out:
 134         if (error == -ENOMEM)
 135                 error = -EAGAIN;
 136         return error;
 137 }
 138
 139 #ifdef CONFIG_SWAP
 140 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 141         unsigned long end, struct mm_walk *walk)
 142 {
 143         pte_t *orig_pte;
 144         struct vm_area_struct *vma = walk->private;
 145         unsigned long index;
 146
 147         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 148                 return 0;
 149
 150         for (index = start; index != end; index += PAGE_SIZE) {
 151                 pte_t pte;
 152                 swp_entry_t entry;
 153                 struct page *page;
 154                 spinlock_t *ptl;
 155
 156                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 157                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 158                 pte_unmap_unlock(orig_pte, ptl);
 159
 160                 if (pte_present(pte) || pte_none(pte))
 161                         continue;
 162                 entry = pte_to_swp_entry(pte);
 163                 if (unlikely(non_swap_entry(entry)))
 164                         continue;
 165
 166                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 167                                                                 vma, index);
 168                 if (page)
 169                         page_cache_release(page);
 170         }
 171
 172         return 0;
 173 }
 174
 175 static void force_swapin_readahead(struct vm_area_struct *vma,
 176                 unsigned long start, unsigned long end)
 177 {
 178         struct mm_walk walk = {
 179                 .mm = vma->vm_mm,
 180                 .pmd_entry = swapin_walk_pmd_entry,
 181                 .private = vma,
 182         };
 183
 184         walk_page_range(start, end, &walk);
 185
 186         lru_add_drain();        /* Push any new pages onto the LRU now */
 187 }
 188
 189 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 190                 unsigned long start, unsigned long end,
 191                 struct address_space *mapping)
 192 {
 193         pgoff_t index;
 194         struct page *page;
 195         swp_entry_t swap;
 196
 197         for (; start < end; start += PAGE_SIZE) {
 198                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 199
 200                 page = find_get_entry(mapping, index);
 201                 if (!radix_tree_exceptional_entry(page)) {
 202                         if (page)
 203                                 page_cache_release(page);
 204                         continue;
 205                 }
 206                 swap = radix_to_swp_entry(page);
 207                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 208                                                                 NULL, 0);
 209                 if (page)
 210                         page_cache_release(page);
 211         }
 212
 213         lru_add_drain();        /* Push any new pages onto the LRU now */
 214 }
 215 #endif          /* CONFIG_SWAP */
 216
 217 /*
 218  * Schedule all required I/O operations.  Do not wait for completion.
 219  */
 220 static long madvise_willneed(struct vm_area_struct *vma,
 221                              struct vm_area_struct **prev,
 222                              unsigned long start, unsigned long end)
 223 {
 224         struct file *file = vma->vm_file;
 225
 226 #ifdef CONFIG_SWAP
 227         if (!file) {
 228                 *prev = vma;
 229                 force_swapin_readahead(vma, start, end);
 230                 return 0;
 231         }
 232
 233         if (shmem_mapping(file->f_mapping)) {
 234                 *prev = vma;
 235                 force_shm_swapin_readahead(vma, start, end,
 236                                         file->f_mapping);
 237                 return 0;
 238         }
 239 #else
 240         if (!file)
 241                 return -EBADF;
 242 #endif
 243
 244         if (IS_DAX(file_inode(file))) {
 245                 /* no bad return value, but ignore advice */
 246                 return 0;
 247         }
 248
 249         *prev = vma;
 250         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 251         if (end > vma->vm_end)
 252                 end = vma->vm_end;
 253         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 254
 255         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 256         return 0;
 257 }
 258
 259 /*
 260  * Application no longer needs these pages.  If the pages are dirty,
 261  * it's OK to just throw them away.  The app will be more careful about
 262  * data it wants to keep.  Be sure to free swap resources too.  The
 263  * zap_page_range call sets things up for shrink_active_list to actually free
 264  * these pages later if no one else has touched them in the meantime,
 265  * although we could add these pages to a global reuse list for
 266  * shrink_active_list to pick up before reclaiming other pages.
 267  *
 268  * NB: This interface discards data rather than pushes it out to swap,
 269  * as some implementations do.  This has performance implications for
 270  * applications like large transactional databases which want to discard
 271  * pages in anonymous maps after committing to backing store the data
 272  * that was kept in them.  There is no reason to write this data out to
 273  * the swap area if the application is discarding it.
 274  *
 275  * An interface that causes the system to free clean pages and flush
 276  * dirty pages is already available as msync(MS_INVALIDATE).
 277  */
 278 static long madvise_dontneed(struct vm_area_struct *vma,
 279                              struct vm_area_struct **prev,
 280                              unsigned long start, unsigned long end)
 281 {
 282         *prev = vma;
 283         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 284                 return -EINVAL;
 285
 286         zap_page_range(vma, start, end - start, NULL);
 287         return 0;
 288 }
 289
 290 /*
 291  * Application wants to free up the pages and associated backing store.
 292  * This is effectively punching a hole into the middle of a file.
 293  */
 294 static long madvise_remove(struct vm_area_struct *vma,
 295                                 struct vm_area_struct **prev,
 296                                 unsigned long start, unsigned long end)
 297 {
 298         loff_t offset;
 299         int error;
 300         struct file *f;
 301
 302         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 303
 304         if (vma->vm_flags & VM_LOCKED)
 305                 return -EINVAL;
 306
 307         f = vma->vm_file;
 308
 309         if (!f || !f->f_mapping || !f->f_mapping->host) {
 310                         return -EINVAL;
 311         }
 312
 313         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 314                 return -EACCES;
 315
 316         offset = (loff_t)(start - vma->vm_start)
 317                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 318
 319         /*
 320          * Filesystem's fallocate may need to take i_mutex.  We need to
 321          * explicitly grab a reference because the vma (and hence the
 322          * vma's reference to the file) can go away as soon as we drop
 323          * mmap_sem.
 324          */
 325         get_file(f);
 326         up_read(&current->mm->mmap_sem);
 327         error = vfs_fallocate(f,
 328                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 329                                 offset, end - start);
 330         fput(f);
 331         down_read(&current->mm->mmap_sem);
 332         return error;
 333 }
 334
 335 #ifdef CONFIG_MEMORY_FAILURE
 336 /*
 337  * Error injection support for memory error handling.
 338  */
 339 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 340 {
 341         struct page *p;
 342         if (!capable(CAP_SYS_ADMIN))
 343                 return -EPERM;
 344         for (; start < end; start += PAGE_SIZE <<
 345                                 compound_order(compound_head(p))) {
 346                 int ret;
 347
 348                 ret = get_user_pages_fast(start, 1, 0, &p);
 349                 if (ret != 1)
 350                         return ret;
 351
 352                 if (PageHWPoison(p)) {
 353                         put_page(p);
 354                         continue;
 355                 }
 356                 if (bhv == MADV_SOFT_OFFLINE) {
 357                         pr_info("Soft offlining page %#lx at %#lx\n",
 358                                 page_to_pfn(p), start);
 359                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 360                         if (ret)
 361                                 return ret;
 362                         continue;
 363                 }
 364                 pr_info("Injecting memory failure for page %#lx at %#lx\n",
 365                        page_to_pfn(p), start);
 366                 /* Ignore return value for now */
 367                 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 368         }
 369         return 0;
 370 }
 371 #endif
 372
 373 static long
 374 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 375                 unsigned long start, unsigned long end, int behavior)
 376 {
 377         switch (behavior) {
 378         case MADV_REMOVE:
 379                 return madvise_remove(vma, prev, start, end);
 380         case MADV_WILLNEED:
 381                 return madvise_willneed(vma, prev, start, end);
 382         case MADV_DONTNEED:
 383                 return madvise_dontneed(vma, prev, start, end);
 384         default:
 385                 return madvise_behavior(vma, prev, start, end, behavior);
 386         }
 387 }
 388
 389 static bool
 390 madvise_behavior_valid(int behavior)
 391 {
 392         switch (behavior) {
 393         case MADV_DOFORK:
 394         case MADV_DONTFORK:
 395         case MADV_NORMAL:
 396         case MADV_SEQUENTIAL:
 397         case MADV_RANDOM:
 398         case MADV_REMOVE:
 399         case MADV_WILLNEED:
 400         case MADV_DONTNEED:
 401 #ifdef CONFIG_KSM
 402         case MADV_MERGEABLE:
 403         case MADV_UNMERGEABLE:
 404 #endif
 405 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 406         case MADV_HUGEPAGE:
 407         case MADV_NOHUGEPAGE:
 408 #endif
 409         case MADV_DONTDUMP:
 410         case MADV_DODUMP:
 411                 return true;
 412
 413         default:
 414                 return false;
 415         }
 416 }
 417
 418 /*
 419  * The madvise(2) system call.
 420  *
 421  * Applications can use madvise() to advise the kernel how it should
 422  * handle paging I/O in this VM area.  The idea is to help the kernel
 423  * use appropriate read-ahead and caching techniques.  The information
 424  * provided is advisory only, and can be safely disregarded by the
 425  * kernel without affecting the correct operation of the application.
 426  *
 427  * behavior values:
 428  *  MADV_NORMAL - the default behavior is to read clusters.  This
 429  *              results in some read-ahead and read-behind.
 430  *  MADV_RANDOM - the system should read the minimum amount of data
 431  *              on any access, since it is unlikely that the appli-
 432  *              cation will need more than what it asks for.
 433  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 434  *              once, so they can be aggressively read ahead, and
 435  *              can be freed soon after they are accessed.
 436  *  MADV_WILLNEED - the application is notifying the system to read
 437  *              some pages ahead.
 438  *  MADV_DONTNEED - the application is finished with the given range,
 439  *              so the kernel can free resources associated with it.
 440  *  MADV_REMOVE - the application wants to free up the given range of
 441  *              pages and associated backing store.
 442  *  MADV_DONTFORK - omit this area from child's address space when forking:
 443  *              typically, to avoid COWing pages pinned by get_user_pages().
 444  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 445  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 446  *              this area with pages of identical content from other such areas.
 447  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 448  *
 449  * return values:
 450  *  zero    - success
 451  *  -EINVAL - start + len < 0, start is not page-aligned,
 452  *              "behavior" is not a valid value, or application
 453  *              is attempting to release locked or shared pages.
 454  *  -ENOMEM - addresses in the specified range are not currently
 455  *              mapped, or are outside the AS of the process.
 456  *  -EIO    - an I/O error occurred while paging in data.
 457  *  -EBADF  - map exists, but area maps something that isn't a file.
 458  *  -EAGAIN - a kernel resource was temporarily unavailable.
 459  */
 460 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 461 {
 462         unsigned long end, tmp;
 463         struct vm_area_struct *vma, *prev;
 464         int unmapped_error = 0;
 465         int error = -EINVAL;
 466         int write;
 467         size_t len;
 468         struct blk_plug plug;
 469
 470 #ifdef CONFIG_MEMORY_FAILURE
 471         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 472                 return madvise_hwpoison(behavior, start, start+len_in);
 473 #endif
 474         if (!madvise_behavior_valid(behavior))
 475                 return error;
 476
 477         if (start & ~PAGE_MASK)
 478                 return error;
 479         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 480
 481         /* Check to see whether len was rounded up from small -ve to zero */
 482         if (len_in && !len)
 483                 return error;
 484
 485         end = start + len;
 486         if (end < start)
 487                 return error;
 488
 489         error = 0;
 490         if (end == start)
 491                 return error;
 492
 493         write = madvise_need_mmap_write(behavior);
 494         if (write)
 495                 down_write(&current->mm->mmap_sem);
 496         else
 497                 down_read(&current->mm->mmap_sem);
 498
 499         /*
 500          * If the interval [start,end) covers some unmapped address
 501          * ranges, just ignore them, but return -ENOMEM at the end.
 502          * - different from the way of handling in mlock etc.
 503          */
 504         vma = find_vma_prev(current->mm, start, &prev);
 505         if (vma && start > vma->vm_start)
 506                 prev = vma;
 507
 508         blk_start_plug(&plug);
 509         for (;;) {
 510                 /* Still start < end. */
 511                 error = -ENOMEM;
 512                 if (!vma)
 513                         goto out;
 514
 515                 /* Here start < (end|vma->vm_end). */
 516                 if (start < vma->vm_start) {
 517                         unmapped_error = -ENOMEM;
 518                         start = vma->vm_start;
 519                         if (start >= end)
 520                                 goto out;
 521                 }
 522
 523                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 524                 tmp = vma->vm_end;
 525                 if (end < tmp)
 526                         tmp = end;
 527
 528                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 529                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 530                 if (error)
 531                         goto out;
 532                 start = tmp;
 533                 if (prev && start < prev->vm_end)
 534                         start = prev->vm_end;
 535                 error = unmapped_error;
 536                 if (start >= end)
 537                         goto out;
 538                 if (prev)
 539                         vma = prev->vm_next;
 540                 else    /* madvise_remove dropped mmap_sem */
 541                         vma = find_vma(current->mm, start);
 542         }
 543 out:
 544         blk_finish_plug(&plug);
 545         if (write)
 546                 up_write(&current->mm->mmap_sem);
 547         else
 548                 up_read(&current->mm->mmap_sem);
 549
 550         return error;
 551 }