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