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