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