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/sched.h>
  15 #include <linux/ksm.h>
  16
  17 /*
  18  * Any behaviour which results in changes to the vma->vm_flags needs to
  19  * take mmap_sem for writing. Others, which simply traverse vmas, need
  20  * to only take it for reading.
  21  */
  22 static int madvise_need_mmap_write(int behavior)
  23 {
  24         switch (behavior) {
  25         case MADV_REMOVE:
  26         case MADV_WILLNEED:
  27         case MADV_DONTNEED:
  28                 return 0;
  29         default:
  30                 /* be safe, default to 1. list exceptions explicitly */
  31                 return 1;
  32         }
  33 }
  34
  35 /*
  36  * We can potentially split a vm area into separate
  37  * areas, each area with its own behavior.
  38  */
  39 static long madvise_behavior(struct vm_area_struct * vma,
  40                      struct vm_area_struct **prev,
  41                      unsigned long start, unsigned long end, int behavior)
  42 {
  43         struct mm_struct * mm = vma->vm_mm;
  44         int error = 0;
  45         pgoff_t pgoff;
  46         unsigned long new_flags = vma->vm_flags;
  47
  48         switch (behavior) {
  49         case MADV_NORMAL:
  50                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  51                 break;
  52         case MADV_SEQUENTIAL:
  53                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  54                 break;
  55         case MADV_RANDOM:
  56                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  57                 break;
  58         case MADV_DONTFORK:
  59                 new_flags |= VM_DONTCOPY;
  60                 break;
  61         case MADV_DOFORK:
  62                 if (vma->vm_flags & VM_IO) {
  63                         error = -EINVAL;
  64                         goto out;
  65                 }
  66                 new_flags &= ~VM_DONTCOPY;
  67                 break;
  68         case MADV_MERGEABLE:
  69         case MADV_UNMERGEABLE:
  70                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
  71                 if (error)
  72                         goto out;
  73                 break;
  74         case MADV_HUGEPAGE:
  75         case MADV_NOHUGEPAGE:
  76                 error = hugepage_madvise(vma, &new_flags, behavior);
  77                 if (error)
  78                         goto out;
  79                 break;
  80         }
  81
  82         if (new_flags == vma->vm_flags) {
  83                 *prev = vma;
  84                 goto out;
  85         }
  86
  87         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
  88         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
  89                                 vma->vm_file, pgoff, vma_policy(vma));
  90         if (*prev) {
  91                 vma = *prev;
  92                 goto success;
  93         }
  94
  95         *prev = vma;
  96
  97         if (start != vma->vm_start) {
  98                 error = split_vma(mm, vma, start, 1);
  99                 if (error)
 100                         goto out;
 101         }
 102
 103         if (end != vma->vm_end) {
 104                 error = split_vma(mm, vma, end, 0);
 105                 if (error)
 106                         goto out;
 107         }
 108
 109 success:
 110         /*
 111          * vm_flags is protected by the mmap_sem held in write mode.
 112          */
 113         vma->vm_flags = new_flags;
 114
 115 out:
 116         if (error == -ENOMEM)
 117                 error = -EAGAIN;
 118         return error;
 119 }
 120
 121 /*
 122  * Schedule all required I/O operations.  Do not wait for completion.
 123  */
 124 static long madvise_willneed(struct vm_area_struct * vma,
 125                              struct vm_area_struct ** prev,
 126                              unsigned long start, unsigned long end)
 127 {
 128         struct file *file = vma->vm_file;
 129
 130         if (!file)
 131                 return -EBADF;
 132
 133         if (file->f_mapping->a_ops->get_xip_mem) {
 134                 /* no bad return value, but ignore advice */
 135                 return 0;
 136         }
 137
 138         *prev = vma;
 139         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 140         if (end > vma->vm_end)
 141                 end = vma->vm_end;
 142         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 143
 144         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 145         return 0;
 146 }
 147
 148 /*
 149  * Application no longer needs these pages.  If the pages are dirty,
 150  * it's OK to just throw them away.  The app will be more careful about
 151  * data it wants to keep.  Be sure to free swap resources too.  The
 152  * zap_page_range call sets things up for shrink_active_list to actually free
 153  * these pages later if no one else has touched them in the meantime,
 154  * although we could add these pages to a global reuse list for
 155  * shrink_active_list to pick up before reclaiming other pages.
 156  *
 157  * NB: This interface discards data rather than pushes it out to swap,
 158  * as some implementations do.  This has performance implications for
 159  * applications like large transactional databases which want to discard
 160  * pages in anonymous maps after committing to backing store the data
 161  * that was kept in them.  There is no reason to write this data out to
 162  * the swap area if the application is discarding it.
 163  *
 164  * An interface that causes the system to free clean pages and flush
 165  * dirty pages is already available as msync(MS_INVALIDATE).
 166  */
 167 static long madvise_dontneed(struct vm_area_struct * vma,
 168                              struct vm_area_struct ** prev,
 169                              unsigned long start, unsigned long end)
 170 {
 171         *prev = vma;
 172         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 173                 return -EINVAL;
 174
 175         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 176                 struct zap_details details = {
 177                         .nonlinear_vma = vma,
 178                         .last_index = ULONG_MAX,
 179                 };
 180                 zap_page_range(vma, start, end - start, &details);
 181         } else
 182                 zap_page_range(vma, start, end - start, NULL);
 183         return 0;
 184 }
 185
 186 /*
 187  * Application wants to free up the pages and associated backing store.
 188  * This is effectively punching a hole into the middle of a file.
 189  *
 190  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 191  * Other filesystems return -ENOSYS.
 192  */
 193 static long madvise_remove(struct vm_area_struct *vma,
 194                                 struct vm_area_struct **prev,
 195                                 unsigned long start, unsigned long end)
 196 {
 197         struct address_space *mapping;
 198         loff_t offset, endoff;
 199         int error;
 200
 201         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 202
 203         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 204                 return -EINVAL;
 205
 206         if (!vma->vm_file || !vma->vm_file->f_mapping
 207                 || !vma->vm_file->f_mapping->host) {
 208                         return -EINVAL;
 209         }
 210
 211         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 212                 return -EACCES;
 213
 214         mapping = vma->vm_file->f_mapping;
 215
 216         offset = (loff_t)(start - vma->vm_start)
 217                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 218         endoff = (loff_t)(end - vma->vm_start - 1)
 219                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 220
 221         /* vmtruncate_range needs to take i_mutex */
 222         up_read(&current->mm->mmap_sem);
 223         error = vmtruncate_range(mapping->host, offset, endoff);
 224         down_read(&current->mm->mmap_sem);
 225         return error;
 226 }
 227
 228 #ifdef CONFIG_MEMORY_FAILURE
 229 /*
 230  * Error injection support for memory error handling.
 231  */
 232 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 233 {
 234         int ret = 0;
 235
 236         if (!capable(CAP_SYS_ADMIN))
 237                 return -EPERM;
 238         for (; start < end; start += PAGE_SIZE) {
 239                 struct page *p;
 240                 int ret = get_user_pages_fast(start, 1, 0, &p);
 241                 if (ret != 1)
 242                         return ret;
 243                 if (bhv == MADV_SOFT_OFFLINE) {
 244                         printk(KERN_INFO "Soft offlining page %lx at %lx\n",
 245                                 page_to_pfn(p), start);
 246                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 247                         if (ret)
 248                                 break;
 249                         continue;
 250                 }
 251                 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
 252                        page_to_pfn(p), start);
 253                 /* Ignore return value for now */
 254                 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 255         }
 256         return ret;
 257 }
 258 #endif
 259
 260 static long
 261 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 262                 unsigned long start, unsigned long end, int behavior)
 263 {
 264         switch (behavior) {
 265         case MADV_REMOVE:
 266                 return madvise_remove(vma, prev, start, end);
 267         case MADV_WILLNEED:
 268                 return madvise_willneed(vma, prev, start, end);
 269         case MADV_DONTNEED:
 270                 return madvise_dontneed(vma, prev, start, end);
 271         default:
 272                 return madvise_behavior(vma, prev, start, end, behavior);
 273         }
 274 }
 275
 276 static int
 277 madvise_behavior_valid(int behavior)
 278 {
 279         switch (behavior) {
 280         case MADV_DOFORK:
 281         case MADV_DONTFORK:
 282         case MADV_NORMAL:
 283         case MADV_SEQUENTIAL:
 284         case MADV_RANDOM:
 285         case MADV_REMOVE:
 286         case MADV_WILLNEED:
 287         case MADV_DONTNEED:
 288 #ifdef CONFIG_KSM
 289         case MADV_MERGEABLE:
 290         case MADV_UNMERGEABLE:
 291 #endif
 292 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 293         case MADV_HUGEPAGE:
 294         case MADV_NOHUGEPAGE:
 295 #endif
 296                 return 1;
 297
 298         default:
 299                 return 0;
 300         }
 301 }
 302
 303 /*
 304  * The madvise(2) system call.
 305  *
 306  * Applications can use madvise() to advise the kernel how it should
 307  * handle paging I/O in this VM area.  The idea is to help the kernel
 308  * use appropriate read-ahead and caching techniques.  The information
 309  * provided is advisory only, and can be safely disregarded by the
 310  * kernel without affecting the correct operation of the application.
 311  *
 312  * behavior values:
 313  *  MADV_NORMAL - the default behavior is to read clusters.  This
 314  *              results in some read-ahead and read-behind.
 315  *  MADV_RANDOM - the system should read the minimum amount of data
 316  *              on any access, since it is unlikely that the appli-
 317  *              cation will need more than what it asks for.
 318  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 319  *              once, so they can be aggressively read ahead, and
 320  *              can be freed soon after they are accessed.
 321  *  MADV_WILLNEED - the application is notifying the system to read
 322  *              some pages ahead.
 323  *  MADV_DONTNEED - the application is finished with the given range,
 324  *              so the kernel can free resources associated with it.
 325  *  MADV_REMOVE - the application wants to free up the given range of
 326  *              pages and associated backing store.
 327  *  MADV_DONTFORK - omit this area from child's address space when forking:
 328  *              typically, to avoid COWing pages pinned by get_user_pages().
 329  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 330  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 331  *              this area with pages of identical content from other such areas.
 332  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 333  *
 334  * return values:
 335  *  zero    - success
 336  *  -EINVAL - start + len < 0, start is not page-aligned,
 337  *              "behavior" is not a valid value, or application
 338  *              is attempting to release locked or shared pages.
 339  *  -ENOMEM - addresses in the specified range are not currently
 340  *              mapped, or are outside the AS of the process.
 341  *  -EIO    - an I/O error occurred while paging in data.
 342  *  -EBADF  - map exists, but area maps something that isn't a file.
 343  *  -EAGAIN - a kernel resource was temporarily unavailable.
 344  */
 345 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 346 {
 347         unsigned long end, tmp;
 348         struct vm_area_struct * vma, *prev;
 349         int unmapped_error = 0;
 350         int error = -EINVAL;
 351         int write;
 352         size_t len;
 353
 354 #ifdef CONFIG_MEMORY_FAILURE
 355         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 356                 return madvise_hwpoison(behavior, start, start+len_in);
 357 #endif
 358         if (!madvise_behavior_valid(behavior))
 359                 return error;
 360
 361         write = madvise_need_mmap_write(behavior);
 362         if (write)
 363                 down_write(&current->mm->mmap_sem);
 364         else
 365                 down_read(&current->mm->mmap_sem);
 366
 367         if (start & ~PAGE_MASK)
 368                 goto out;
 369         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 370
 371         /* Check to see whether len was rounded up from small -ve to zero */
 372         if (len_in && !len)
 373                 goto out;
 374
 375         end = start + len;
 376         if (end < start)
 377                 goto out;
 378
 379         error = 0;
 380         if (end == start)
 381                 goto out;
 382
 383         /*
 384          * If the interval [start,end) covers some unmapped address
 385          * ranges, just ignore them, but return -ENOMEM at the end.
 386          * - different from the way of handling in mlock etc.
 387          */
 388         vma = find_vma_prev(current->mm, start, &prev);
 389         if (vma && start > vma->vm_start)
 390                 prev = vma;
 391
 392         for (;;) {
 393                 /* Still start < end. */
 394                 error = -ENOMEM;
 395                 if (!vma)
 396                         goto out;
 397
 398                 /* Here start < (end|vma->vm_end). */
 399                 if (start < vma->vm_start) {
 400                         unmapped_error = -ENOMEM;
 401                         start = vma->vm_start;
 402                         if (start >= end)
 403                                 goto out;
 404                 }
 405
 406                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 407                 tmp = vma->vm_end;
 408                 if (end < tmp)
 409                         tmp = end;
 410
 411                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 412                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 413                 if (error)
 414                         goto out;
 415                 start = tmp;
 416                 if (prev && start < prev->vm_end)
 417                         start = prev->vm_end;
 418                 error = unmapped_error;
 419                 if (start >= end)
 420                         goto out;
 421                 if (prev)
 422                         vma = prev->vm_next;
 423                 else    /* madvise_remove dropped mmap_sem */
 424                         vma = find_vma(current->mm, start);
 425         }
 426 out:
 427         if (write)
 428                 up_write(&current->mm->mmap_sem);
 429         else
 430                 up_read(&current->mm->mmap_sem);
 431
 432         return error;
 433 }