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/hugetlb.h>
  13 #include <linux/sched.h>
  14 #include <linux/ksm.h>
  15
  16 /*
  17  * Any behaviour which results in changes to the vma->vm_flags needs to
  18  * take mmap_sem for writing. Others, which simply traverse vmas, need
  19  * to only take it for reading.
  20  */
  21 static int madvise_need_mmap_write(int behavior)
  22 {
  23         switch (behavior) {
  24         case MADV_REMOVE:
  25         case MADV_WILLNEED:
  26         case MADV_DONTNEED:
  27                 return 0;
  28         default:
  29                 /* be safe, default to 1. list exceptions explicitly */
  30                 return 1;
  31         }
  32 }
  33
  34 /*
  35  * We can potentially split a vm area into separate
  36  * areas, each area with its own behavior.
  37  */
  38 static long madvise_behavior(struct vm_area_struct * vma,
  39                      struct vm_area_struct **prev,
  40                      unsigned long start, unsigned long end, int behavior)
  41 {
  42         struct mm_struct * mm = vma->vm_mm;
  43         int error = 0;
  44         pgoff_t pgoff;
  45         unsigned long new_flags = vma->vm_flags;
  46
  47         switch (behavior) {
  48         case MADV_NORMAL:
  49                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  50                 break;
  51         case MADV_SEQUENTIAL:
  52                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  53                 break;
  54         case MADV_RANDOM:
  55                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  56                 break;
  57         case MADV_DONTFORK:
  58                 new_flags |= VM_DONTCOPY;
  59                 break;
  60         case MADV_DOFORK:
  61                 if (vma->vm_flags & VM_IO) {
  62                         error = -EINVAL;
  63                         goto out;
  64                 }
  65                 new_flags &= ~VM_DONTCOPY;
  66                 break;
  67         case MADV_MERGEABLE:
  68         case MADV_UNMERGEABLE:
  69                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
  70                 if (error)
  71                         goto out;
  72                 break;
  73         }
  74
  75         if (new_flags == vma->vm_flags) {
  76                 *prev = vma;
  77                 goto out;
  78         }
  79
  80         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
  81         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
  82                                 vma->vm_file, pgoff, vma_policy(vma));
  83         if (*prev) {
  84                 vma = *prev;
  85                 goto success;
  86         }
  87
  88         *prev = vma;
  89
  90         if (start != vma->vm_start) {
  91                 error = split_vma(mm, vma, start, 1);
  92                 if (error)
  93                         goto out;
  94         }
  95
  96         if (end != vma->vm_end) {
  97                 error = split_vma(mm, vma, end, 0);
  98                 if (error)
  99                         goto out;
 100         }
 101
 102 success:
 103         /*
 104          * vm_flags is protected by the mmap_sem held in write mode.
 105          */
 106         vma->vm_flags = new_flags;
 107
 108 out:
 109         if (error == -ENOMEM)
 110                 error = -EAGAIN;
 111         return error;
 112 }
 113
 114 /*
 115  * Schedule all required I/O operations.  Do not wait for completion.
 116  */
 117 static long madvise_willneed(struct vm_area_struct * vma,
 118                              struct vm_area_struct ** prev,
 119                              unsigned long start, unsigned long end)
 120 {
 121         struct file *file = vma->vm_file;
 122
 123         if (!file)
 124                 return -EBADF;
 125
 126         if (file->f_mapping->a_ops->get_xip_mem) {
 127                 /* no bad return value, but ignore advice */
 128                 return 0;
 129         }
 130
 131         *prev = vma;
 132         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 133         if (end > vma->vm_end)
 134                 end = vma->vm_end;
 135         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 136
 137         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 138         return 0;
 139 }
 140
 141 /*
 142  * Application no longer needs these pages.  If the pages are dirty,
 143  * it's OK to just throw them away.  The app will be more careful about
 144  * data it wants to keep.  Be sure to free swap resources too.  The
 145  * zap_page_range call sets things up for shrink_active_list to actually free
 146  * these pages later if no one else has touched them in the meantime,
 147  * although we could add these pages to a global reuse list for
 148  * shrink_active_list to pick up before reclaiming other pages.
 149  *
 150  * NB: This interface discards data rather than pushes it out to swap,
 151  * as some implementations do.  This has performance implications for
 152  * applications like large transactional databases which want to discard
 153  * pages in anonymous maps after committing to backing store the data
 154  * that was kept in them.  There is no reason to write this data out to
 155  * the swap area if the application is discarding it.
 156  *
 157  * An interface that causes the system to free clean pages and flush
 158  * dirty pages is already available as msync(MS_INVALIDATE).
 159  */
 160 static long madvise_dontneed(struct vm_area_struct * vma,
 161                              struct vm_area_struct ** prev,
 162                              unsigned long start, unsigned long end)
 163 {
 164         *prev = vma;
 165         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 166                 return -EINVAL;
 167
 168         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 169                 struct zap_details details = {
 170                         .nonlinear_vma = vma,
 171                         .last_index = ULONG_MAX,
 172                 };
 173                 zap_page_range(vma, start, end - start, &details);
 174         } else
 175                 zap_page_range(vma, start, end - start, NULL);
 176         return 0;
 177 }
 178
 179 /*
 180  * Application wants to free up the pages and associated backing store.
 181  * This is effectively punching a hole into the middle of a file.
 182  *
 183  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 184  * Other filesystems return -ENOSYS.
 185  */
 186 static long madvise_remove(struct vm_area_struct *vma,
 187                                 struct vm_area_struct **prev,
 188                                 unsigned long start, unsigned long end)
 189 {
 190         struct address_space *mapping;
 191         loff_t offset, endoff;
 192         int error;
 193
 194         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 195
 196         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 197                 return -EINVAL;
 198
 199         if (!vma->vm_file || !vma->vm_file->f_mapping
 200                 || !vma->vm_file->f_mapping->host) {
 201                         return -EINVAL;
 202         }
 203
 204         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 205                 return -EACCES;
 206
 207         mapping = vma->vm_file->f_mapping;
 208
 209         offset = (loff_t)(start - vma->vm_start)
 210                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 211         endoff = (loff_t)(end - vma->vm_start - 1)
 212                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 213
 214         /* vmtruncate_range needs to take i_mutex and i_alloc_sem */
 215         up_read(&current->mm->mmap_sem);
 216         error = vmtruncate_range(mapping->host, offset, endoff);
 217         down_read(&current->mm->mmap_sem);
 218         return error;
 219 }
 220
 221 #ifdef CONFIG_MEMORY_FAILURE
 222 /*
 223  * Error injection support for memory error handling.
 224  */
 225 static int madvise_hwpoison(unsigned long start, unsigned long end)
 226 {
 227         int ret = 0;
 228
 229         if (!capable(CAP_SYS_ADMIN))
 230                 return -EPERM;
 231         for (; start < end; start += PAGE_SIZE) {
 232                 struct page *p;
 233                 int ret = get_user_pages(current, current->mm, start, 1,
 234                                                 0, 0, &p, NULL);
 235                 if (ret != 1)
 236                         return ret;
 237                 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
 238                        page_to_pfn(p), start);
 239                 /* Ignore return value for now */
 240                 __memory_failure(page_to_pfn(p), 0, 1);
 241                 put_page(p);
 242         }
 243         return ret;
 244 }
 245 #endif
 246
 247 static long
 248 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 249                 unsigned long start, unsigned long end, int behavior)
 250 {
 251         switch (behavior) {
 252         case MADV_REMOVE:
 253                 return madvise_remove(vma, prev, start, end);
 254         case MADV_WILLNEED:
 255                 return madvise_willneed(vma, prev, start, end);
 256         case MADV_DONTNEED:
 257                 return madvise_dontneed(vma, prev, start, end);
 258         default:
 259                 return madvise_behavior(vma, prev, start, end, behavior);
 260         }
 261 }
 262
 263 static int
 264 madvise_behavior_valid(int behavior)
 265 {
 266         switch (behavior) {
 267         case MADV_DOFORK:
 268         case MADV_DONTFORK:
 269         case MADV_NORMAL:
 270         case MADV_SEQUENTIAL:
 271         case MADV_RANDOM:
 272         case MADV_REMOVE:
 273         case MADV_WILLNEED:
 274         case MADV_DONTNEED:
 275 #ifdef CONFIG_KSM
 276         case MADV_MERGEABLE:
 277         case MADV_UNMERGEABLE:
 278 #endif
 279                 return 1;
 280
 281         default:
 282                 return 0;
 283         }
 284 }
 285
 286 /*
 287  * The madvise(2) system call.
 288  *
 289  * Applications can use madvise() to advise the kernel how it should
 290  * handle paging I/O in this VM area.  The idea is to help the kernel
 291  * use appropriate read-ahead and caching techniques.  The information
 292  * provided is advisory only, and can be safely disregarded by the
 293  * kernel without affecting the correct operation of the application.
 294  *
 295  * behavior values:
 296  *  MADV_NORMAL - the default behavior is to read clusters.  This
 297  *              results in some read-ahead and read-behind.
 298  *  MADV_RANDOM - the system should read the minimum amount of data
 299  *              on any access, since it is unlikely that the appli-
 300  *              cation will need more than what it asks for.
 301  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 302  *              once, so they can be aggressively read ahead, and
 303  *              can be freed soon after they are accessed.
 304  *  MADV_WILLNEED - the application is notifying the system to read
 305  *              some pages ahead.
 306  *  MADV_DONTNEED - the application is finished with the given range,
 307  *              so the kernel can free resources associated with it.
 308  *  MADV_REMOVE - the application wants to free up the given range of
 309  *              pages and associated backing store.
 310  *  MADV_DONTFORK - omit this area from child's address space when forking:
 311  *              typically, to avoid COWing pages pinned by get_user_pages().
 312  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 313  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 314  *              this area with pages of identical content from other such areas.
 315  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 316  *
 317  * return values:
 318  *  zero    - success
 319  *  -EINVAL - start + len < 0, start is not page-aligned,
 320  *              "behavior" is not a valid value, or application
 321  *              is attempting to release locked or shared pages.
 322  *  -ENOMEM - addresses in the specified range are not currently
 323  *              mapped, or are outside the AS of the process.
 324  *  -EIO    - an I/O error occurred while paging in data.
 325  *  -EBADF  - map exists, but area maps something that isn't a file.
 326  *  -EAGAIN - a kernel resource was temporarily unavailable.
 327  */
 328 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 329 {
 330         unsigned long end, tmp;
 331         struct vm_area_struct * vma, *prev;
 332         int unmapped_error = 0;
 333         int error = -EINVAL;
 334         int write;
 335         size_t len;
 336
 337 #ifdef CONFIG_MEMORY_FAILURE
 338         if (behavior == MADV_HWPOISON)
 339                 return madvise_hwpoison(start, start+len_in);
 340 #endif
 341         if (!madvise_behavior_valid(behavior))
 342                 return error;
 343
 344         write = madvise_need_mmap_write(behavior);
 345         if (write)
 346                 down_write(&current->mm->mmap_sem);
 347         else
 348                 down_read(&current->mm->mmap_sem);
 349
 350         if (start & ~PAGE_MASK)
 351                 goto out;
 352         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 353
 354         /* Check to see whether len was rounded up from small -ve to zero */
 355         if (len_in && !len)
 356                 goto out;
 357
 358         end = start + len;
 359         if (end < start)
 360                 goto out;
 361
 362         error = 0;
 363         if (end == start)
 364                 goto out;
 365
 366         /*
 367          * If the interval [start,end) covers some unmapped address
 368          * ranges, just ignore them, but return -ENOMEM at the end.
 369          * - different from the way of handling in mlock etc.
 370          */
 371         vma = find_vma_prev(current->mm, start, &prev);
 372         if (vma && start > vma->vm_start)
 373                 prev = vma;
 374
 375         for (;;) {
 376                 /* Still start < end. */
 377                 error = -ENOMEM;
 378                 if (!vma)
 379                         goto out;
 380
 381                 /* Here start < (end|vma->vm_end). */
 382                 if (start < vma->vm_start) {
 383                         unmapped_error = -ENOMEM;
 384                         start = vma->vm_start;
 385                         if (start >= end)
 386                                 goto out;
 387                 }
 388
 389                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 390                 tmp = vma->vm_end;
 391                 if (end < tmp)
 392                         tmp = end;
 393
 394                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 395                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 396                 if (error)
 397                         goto out;
 398                 start = tmp;
 399                 if (prev && start < prev->vm_end)
 400                         start = prev->vm_end;
 401                 error = unmapped_error;
 402                 if (start >= end)
 403                         goto out;
 404                 if (prev)
 405                         vma = prev->vm_next;
 406                 else    /* madvise_remove dropped mmap_sem */
 407                         vma = find_vma(current->mm, start);
 408         }
 409 out:
 410         if (write)
 411                 up_write(&current->mm->mmap_sem);
 412         else
 413                 up_read(&current->mm->mmap_sem);
 414
 415         return error;
 416 }