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
  19 /*
  20  * Any behaviour which results in changes to the vma->vm_flags needs to
  21  * take mmap_sem for writing. Others, which simply traverse vmas, need
  22  * to only take it for reading.
  23  */
  24 static int madvise_need_mmap_write(int behavior)
  25 {
  26         switch (behavior) {
  27         case MADV_REMOVE:
  28         case MADV_WILLNEED:
  29         case MADV_DONTNEED:
  30                 return 0;
  31         default:
  32                 /* be safe, default to 1. list exceptions explicitly */
  33                 return 1;
  34         }
  35 }
  36
  37 /*
  38  * We can potentially split a vm area into separate
  39  * areas, each area with its own behavior.
  40  */
  41 static long madvise_behavior(struct vm_area_struct * vma,
  42                      struct vm_area_struct **prev,
  43                      unsigned long start, unsigned long end, int behavior)
  44 {
  45         struct mm_struct * mm = vma->vm_mm;
  46         int error = 0;
  47         pgoff_t pgoff;
  48         unsigned long new_flags = vma->vm_flags;
  49
  50         switch (behavior) {
  51         case MADV_NORMAL:
  52                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  53                 break;
  54         case MADV_SEQUENTIAL:
  55                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  56                 break;
  57         case MADV_RANDOM:
  58                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  59                 break;
  60         case MADV_DONTFORK:
  61                 new_flags |= VM_DONTCOPY;
  62                 break;
  63         case MADV_DOFORK:
  64                 if (vma->vm_flags & VM_IO) {
  65                         error = -EINVAL;
  66                         goto out;
  67                 }
  68                 new_flags &= ~VM_DONTCOPY;
  69                 break;
  70         case MADV_DONTDUMP:
  71                 new_flags |= VM_NODUMP;
  72                 break;
  73         case MADV_DODUMP:
  74                 new_flags &= ~VM_NODUMP;
  75                 break;
  76         case MADV_MERGEABLE:
  77         case MADV_UNMERGEABLE:
  78                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
  79                 if (error)
  80                         goto out;
  81                 break;
  82         case MADV_HUGEPAGE:
  83         case MADV_NOHUGEPAGE:
  84                 error = hugepage_madvise(vma, &new_flags, behavior);
  85                 if (error)
  86                         goto out;
  87                 break;
  88         }
  89
  90         if (new_flags == vma->vm_flags) {
  91                 *prev = vma;
  92                 goto out;
  93         }
  94
  95         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
  96         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
  97                                 vma->vm_file, pgoff, vma_policy(vma));
  98         if (*prev) {
  99                 vma = *prev;
 100                 goto success;
 101         }
 102
 103         *prev = vma;
 104
 105         if (start != vma->vm_start) {
 106                 error = split_vma(mm, vma, start, 1);
 107                 if (error)
 108                         goto out;
 109         }
 110
 111         if (end != vma->vm_end) {
 112                 error = split_vma(mm, vma, end, 0);
 113                 if (error)
 114                         goto out;
 115         }
 116
 117 success:
 118         /*
 119          * vm_flags is protected by the mmap_sem held in write mode.
 120          */
 121         vma->vm_flags = new_flags;
 122
 123 out:
 124         if (error == -ENOMEM)
 125                 error = -EAGAIN;
 126         return error;
 127 }
 128
 129 /*
 130  * Schedule all required I/O operations.  Do not wait for completion.
 131  */
 132 static long madvise_willneed(struct vm_area_struct * vma,
 133                              struct vm_area_struct ** prev,
 134                              unsigned long start, unsigned long end)
 135 {
 136         struct file *file = vma->vm_file;
 137
 138         if (!file)
 139                 return -EBADF;
 140
 141         if (file->f_mapping->a_ops->get_xip_mem) {
 142                 /* no bad return value, but ignore advice */
 143                 return 0;
 144         }
 145
 146         *prev = vma;
 147         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 148         if (end > vma->vm_end)
 149                 end = vma->vm_end;
 150         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 151
 152         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 153         return 0;
 154 }
 155
 156 /*
 157  * Application no longer needs these pages.  If the pages are dirty,
 158  * it's OK to just throw them away.  The app will be more careful about
 159  * data it wants to keep.  Be sure to free swap resources too.  The
 160  * zap_page_range call sets things up for shrink_active_list to actually free
 161  * these pages later if no one else has touched them in the meantime,
 162  * although we could add these pages to a global reuse list for
 163  * shrink_active_list to pick up before reclaiming other pages.
 164  *
 165  * NB: This interface discards data rather than pushes it out to swap,
 166  * as some implementations do.  This has performance implications for
 167  * applications like large transactional databases which want to discard
 168  * pages in anonymous maps after committing to backing store the data
 169  * that was kept in them.  There is no reason to write this data out to
 170  * the swap area if the application is discarding it.
 171  *
 172  * An interface that causes the system to free clean pages and flush
 173  * dirty pages is already available as msync(MS_INVALIDATE).
 174  */
 175 static long madvise_dontneed(struct vm_area_struct * vma,
 176                              struct vm_area_struct ** prev,
 177                              unsigned long start, unsigned long end)
 178 {
 179         *prev = vma;
 180         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 181                 return -EINVAL;
 182
 183         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 184                 struct zap_details details = {
 185                         .nonlinear_vma = vma,
 186                         .last_index = ULONG_MAX,
 187                 };
 188                 zap_page_range(vma, start, end - start, &details);
 189         } else
 190                 zap_page_range(vma, start, end - start, NULL);
 191         return 0;
 192 }
 193
 194 /*
 195  * Application wants to free up the pages and associated backing store.
 196  * This is effectively punching a hole into the middle of a file.
 197  *
 198  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 199  * Other filesystems return -ENOSYS.
 200  */
 201 static long madvise_remove(struct vm_area_struct *vma,
 202                                 struct vm_area_struct **prev,
 203                                 unsigned long start, unsigned long end)
 204 {
 205         loff_t offset;
 206         int error;
 207
 208         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 209
 210         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 211                 return -EINVAL;
 212
 213         if (!vma->vm_file || !vma->vm_file->f_mapping
 214                 || !vma->vm_file->f_mapping->host) {
 215                         return -EINVAL;
 216         }
 217
 218         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 219                 return -EACCES;
 220
 221         offset = (loff_t)(start - vma->vm_start)
 222                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 223
 224         /* filesystem's fallocate may need to take i_mutex */
 225         up_read(&current->mm->mmap_sem);
 226         error = do_fallocate(vma->vm_file,
 227                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 228                                 offset, end - start);
 229         down_read(&current->mm->mmap_sem);
 230         return error;
 231 }
 232
 233 #ifdef CONFIG_MEMORY_FAILURE
 234 /*
 235  * Error injection support for memory error handling.
 236  */
 237 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 238 {
 239         int ret = 0;
 240
 241         if (!capable(CAP_SYS_ADMIN))
 242                 return -EPERM;
 243         for (; start < end; start += PAGE_SIZE) {
 244                 struct page *p;
 245                 int ret = get_user_pages_fast(start, 1, 0, &p);
 246                 if (ret != 1)
 247                         return ret;
 248                 if (bhv == MADV_SOFT_OFFLINE) {
 249                         printk(KERN_INFO "Soft offlining page %lx at %lx\n",
 250                                 page_to_pfn(p), start);
 251                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 252                         if (ret)
 253                                 break;
 254                         continue;
 255                 }
 256                 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
 257                        page_to_pfn(p), start);
 258                 /* Ignore return value for now */
 259                 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 260         }
 261         return ret;
 262 }
 263 #endif
 264
 265 static long
 266 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 267                 unsigned long start, unsigned long end, int behavior)
 268 {
 269         switch (behavior) {
 270         case MADV_REMOVE:
 271                 return madvise_remove(vma, prev, start, end);
 272         case MADV_WILLNEED:
 273                 return madvise_willneed(vma, prev, start, end);
 274         case MADV_DONTNEED:
 275                 return madvise_dontneed(vma, prev, start, end);
 276         default:
 277                 return madvise_behavior(vma, prev, start, end, behavior);
 278         }
 279 }
 280
 281 static int
 282 madvise_behavior_valid(int behavior)
 283 {
 284         switch (behavior) {
 285         case MADV_DOFORK:
 286         case MADV_DONTFORK:
 287         case MADV_NORMAL:
 288         case MADV_SEQUENTIAL:
 289         case MADV_RANDOM:
 290         case MADV_REMOVE:
 291         case MADV_WILLNEED:
 292         case MADV_DONTNEED:
 293 #ifdef CONFIG_KSM
 294         case MADV_MERGEABLE:
 295         case MADV_UNMERGEABLE:
 296 #endif
 297 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 298         case MADV_HUGEPAGE:
 299         case MADV_NOHUGEPAGE:
 300 #endif
 301         case MADV_DONTDUMP:
 302         case MADV_DODUMP:
 303                 return 1;
 304
 305         default:
 306                 return 0;
 307         }
 308 }
 309
 310 /*
 311  * The madvise(2) system call.
 312  *
 313  * Applications can use madvise() to advise the kernel how it should
 314  * handle paging I/O in this VM area.  The idea is to help the kernel
 315  * use appropriate read-ahead and caching techniques.  The information
 316  * provided is advisory only, and can be safely disregarded by the
 317  * kernel without affecting the correct operation of the application.
 318  *
 319  * behavior values:
 320  *  MADV_NORMAL - the default behavior is to read clusters.  This
 321  *              results in some read-ahead and read-behind.
 322  *  MADV_RANDOM - the system should read the minimum amount of data
 323  *              on any access, since it is unlikely that the appli-
 324  *              cation will need more than what it asks for.
 325  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 326  *              once, so they can be aggressively read ahead, and
 327  *              can be freed soon after they are accessed.
 328  *  MADV_WILLNEED - the application is notifying the system to read
 329  *              some pages ahead.
 330  *  MADV_DONTNEED - the application is finished with the given range,
 331  *              so the kernel can free resources associated with it.
 332  *  MADV_REMOVE - the application wants to free up the given range of
 333  *              pages and associated backing store.
 334  *  MADV_DONTFORK - omit this area from child's address space when forking:
 335  *              typically, to avoid COWing pages pinned by get_user_pages().
 336  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 337  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 338  *              this area with pages of identical content from other such areas.
 339  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 340  *
 341  * return values:
 342  *  zero    - success
 343  *  -EINVAL - start + len < 0, start is not page-aligned,
 344  *              "behavior" is not a valid value, or application
 345  *              is attempting to release locked or shared pages.
 346  *  -ENOMEM - addresses in the specified range are not currently
 347  *              mapped, or are outside the AS of the process.
 348  *  -EIO    - an I/O error occurred while paging in data.
 349  *  -EBADF  - map exists, but area maps something that isn't a file.
 350  *  -EAGAIN - a kernel resource was temporarily unavailable.
 351  */
 352 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 353 {
 354         unsigned long end, tmp;
 355         struct vm_area_struct * vma, *prev;
 356         int unmapped_error = 0;
 357         int error = -EINVAL;
 358         int write;
 359         size_t len;
 360
 361 #ifdef CONFIG_MEMORY_FAILURE
 362         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 363                 return madvise_hwpoison(behavior, start, start+len_in);
 364 #endif
 365         if (!madvise_behavior_valid(behavior))
 366                 return error;
 367
 368         write = madvise_need_mmap_write(behavior);
 369         if (write)
 370                 down_write(&current->mm->mmap_sem);
 371         else
 372                 down_read(&current->mm->mmap_sem);
 373
 374         if (start & ~PAGE_MASK)
 375                 goto out;
 376         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 377
 378         /* Check to see whether len was rounded up from small -ve to zero */
 379         if (len_in && !len)
 380                 goto out;
 381
 382         end = start + len;
 383         if (end < start)
 384                 goto out;
 385
 386         error = 0;
 387         if (end == start)
 388                 goto out;
 389
 390         /*
 391          * If the interval [start,end) covers some unmapped address
 392          * ranges, just ignore them, but return -ENOMEM at the end.
 393          * - different from the way of handling in mlock etc.
 394          */
 395         vma = find_vma_prev(current->mm, start, &prev);
 396         if (vma && start > vma->vm_start)
 397                 prev = vma;
 398
 399         for (;;) {
 400                 /* Still start < end. */
 401                 error = -ENOMEM;
 402                 if (!vma)
 403                         goto out;
 404
 405                 /* Here start < (end|vma->vm_end). */
 406                 if (start < vma->vm_start) {
 407                         unmapped_error = -ENOMEM;
 408                         start = vma->vm_start;
 409                         if (start >= end)
 410                                 goto out;
 411                 }
 412
 413                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 414                 tmp = vma->vm_end;
 415                 if (end < tmp)
 416                         tmp = end;
 417
 418                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 419                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 420                 if (error)
 421                         goto out;
 422                 start = tmp;
 423                 if (prev && start < prev->vm_end)
 424                         start = prev->vm_end;
 425                 error = unmapped_error;
 426                 if (start >= end)
 427                         goto out;
 428                 if (prev)
 429                         vma = prev->vm_next;
 430                 else    /* madvise_remove dropped mmap_sem */
 431                         vma = find_vma(current->mm, start);
 432         }
 433 out:
 434         if (write)
 435                 up_write(&current->mm->mmap_sem);
 436         else
 437                 up_read(&current->mm->mmap_sem);
 438
 439         return error;
 440 }