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 #include <linux/file.h>
  17
  18 /*
  19  * Any behaviour which results in changes to the vma->vm_flags needs to
  20  * take mmap_sem for writing. Others, which simply traverse vmas, need
  21  * to only take it for reading.
  22  */
  23 static int madvise_need_mmap_write(int behavior)
  24 {
  25         switch (behavior) {
  26         case MADV_REMOVE:
  27         case MADV_WILLNEED:
  28         case MADV_DONTNEED:
  29                 return 0;
  30         default:
  31                 /* be safe, default to 1. list exceptions explicitly */
  32                 return 1;
  33         }
  34 }
  35
  36 /*
  37  * We can potentially split a vm area into separate
  38  * areas, each area with its own behavior.
  39  */
  40 static long madvise_behavior(struct vm_area_struct * vma,
  41                      struct vm_area_struct **prev,
  42                      unsigned long start, unsigned long end, int behavior)
  43 {
  44         struct mm_struct * mm = vma->vm_mm;
  45         int error = 0;
  46         pgoff_t pgoff;
  47         unsigned long new_flags = vma->vm_flags;
  48
  49         switch (behavior) {
  50         case MADV_NORMAL:
  51                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  52                 break;
  53         case MADV_SEQUENTIAL:
  54                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  55                 break;
  56         case MADV_RANDOM:
  57                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  58                 break;
  59         case MADV_DONTFORK:
  60                 new_flags |= VM_DONTCOPY;
  61                 break;
  62         case MADV_DOFORK:
  63                 if (vma->vm_flags & VM_IO) {
  64                         error = -EINVAL;
  65                         goto out;
  66                 }
  67                 new_flags &= ~VM_DONTCOPY;
  68                 break;
  69         case MADV_DONTDUMP:
  70                 new_flags |= VM_NODUMP;
  71                 break;
  72         case MADV_DODUMP:
  73                 new_flags &= ~VM_NODUMP;
  74                 break;
  75         case MADV_MERGEABLE:
  76         case MADV_UNMERGEABLE:
  77                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
  78                 if (error)
  79                         goto out;
  80                 break;
  81         case MADV_HUGEPAGE:
  82         case MADV_NOHUGEPAGE:
  83                 error = hugepage_madvise(vma, &new_flags, behavior);
  84                 if (error)
  85                         goto out;
  86                 break;
  87         }
  88
  89         if (new_flags == vma->vm_flags) {
  90                 *prev = vma;
  91                 goto out;
  92         }
  93
  94         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
  95         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
  96                                 vma->vm_file, pgoff, vma_policy(vma));
  97         if (*prev) {
  98                 vma = *prev;
  99                 goto success;
 100         }
 101
 102         *prev = vma;
 103
 104         if (start != vma->vm_start) {
 105                 error = split_vma(mm, vma, start, 1);
 106                 if (error)
 107                         goto out;
 108         }
 109
 110         if (end != vma->vm_end) {
 111                 error = split_vma(mm, vma, end, 0);
 112                 if (error)
 113                         goto out;
 114         }
 115
 116 success:
 117         /*
 118          * vm_flags is protected by the mmap_sem held in write mode.
 119          */
 120         vma->vm_flags = new_flags;
 121
 122 out:
 123         if (error == -ENOMEM)
 124                 error = -EAGAIN;
 125         return error;
 126 }
 127
 128 /*
 129  * Schedule all required I/O operations.  Do not wait for completion.
 130  */
 131 static long madvise_willneed(struct vm_area_struct * vma,
 132                              struct vm_area_struct ** prev,
 133                              unsigned long start, unsigned long end)
 134 {
 135         struct file *file = vma->vm_file;
 136
 137         if (!file)
 138                 return -EBADF;
 139
 140         if (file->f_mapping->a_ops->get_xip_mem) {
 141                 /* no bad return value, but ignore advice */
 142                 return 0;
 143         }
 144
 145         *prev = vma;
 146         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 147         if (end > vma->vm_end)
 148                 end = vma->vm_end;
 149         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 150
 151         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 152         return 0;
 153 }
 154
 155 /*
 156  * Application no longer needs these pages.  If the pages are dirty,
 157  * it's OK to just throw them away.  The app will be more careful about
 158  * data it wants to keep.  Be sure to free swap resources too.  The
 159  * zap_page_range call sets things up for shrink_active_list to actually free
 160  * these pages later if no one else has touched them in the meantime,
 161  * although we could add these pages to a global reuse list for
 162  * shrink_active_list to pick up before reclaiming other pages.
 163  *
 164  * NB: This interface discards data rather than pushes it out to swap,
 165  * as some implementations do.  This has performance implications for
 166  * applications like large transactional databases which want to discard
 167  * pages in anonymous maps after committing to backing store the data
 168  * that was kept in them.  There is no reason to write this data out to
 169  * the swap area if the application is discarding it.
 170  *
 171  * An interface that causes the system to free clean pages and flush
 172  * dirty pages is already available as msync(MS_INVALIDATE).
 173  */
 174 static long madvise_dontneed(struct vm_area_struct * vma,
 175                              struct vm_area_struct ** prev,
 176                              unsigned long start, unsigned long end)
 177 {
 178         *prev = vma;
 179         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 180                 return -EINVAL;
 181
 182         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 183                 struct zap_details details = {
 184                         .nonlinear_vma = vma,
 185                         .last_index = ULONG_MAX,
 186                 };
 187                 zap_page_range(vma, start, end - start, &details);
 188         } else
 189                 zap_page_range(vma, start, end - start, NULL);
 190         return 0;
 191 }
 192
 193 /*
 194  * Application wants to free up the pages and associated backing store.
 195  * This is effectively punching a hole into the middle of a file.
 196  *
 197  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 198  * Other filesystems return -ENOSYS.
 199  */
 200 static long madvise_remove(struct vm_area_struct *vma,
 201                                 struct vm_area_struct **prev,
 202                                 unsigned long start, unsigned long end)
 203 {
 204         struct address_space *mapping;
 205         loff_t offset, endoff;
 206         int error;
 207         struct file *f;
 208
 209         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 210
 211         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 212                 return -EINVAL;
 213
 214         f = vma->vm_file;
 215
 216         if (!f || !f->f_mapping || !f->f_mapping->host) {
 217                         return -EINVAL;
 218         }
 219
 220         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 221                 return -EACCES;
 222
 223         mapping = vma->vm_file->f_mapping;
 224
 225         offset = (loff_t)(start - vma->vm_start)
 226                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 227         endoff = (loff_t)(end - vma->vm_start - 1)
 228                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 229
 230         /*
 231          * vmtruncate_range may need to take i_mutex.  We need to
 232          * explicitly grab a reference because the vma (and hence the
 233          * vma's reference to the file) can go away as soon as we drop
 234          * mmap_sem.
 235          */
 236         get_file(f);
 237         up_read(&current->mm->mmap_sem);
 238         error = vmtruncate_range(mapping->host, offset, endoff);
 239         fput(f);
 240         down_read(&current->mm->mmap_sem);
 241         return error;
 242 }
 243
 244 #ifdef CONFIG_MEMORY_FAILURE
 245 /*
 246  * Error injection support for memory error handling.
 247  */
 248 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 249 {
 250         int ret = 0;
 251
 252         if (!capable(CAP_SYS_ADMIN))
 253                 return -EPERM;
 254         for (; start < end; start += PAGE_SIZE) {
 255                 struct page *p;
 256                 int ret = get_user_pages_fast(start, 1, 0, &p);
 257                 if (ret != 1)
 258                         return ret;
 259                 if (bhv == MADV_SOFT_OFFLINE) {
 260                         printk(KERN_INFO "Soft offlining page %lx at %lx\n",
 261                                 page_to_pfn(p), start);
 262                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 263                         if (ret)
 264                                 break;
 265                         continue;
 266                 }
 267                 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
 268                        page_to_pfn(p), start);
 269                 /* Ignore return value for now */
 270                 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 271         }
 272         return ret;
 273 }
 274 #endif
 275
 276 static long
 277 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 278                 unsigned long start, unsigned long end, int behavior)
 279 {
 280         switch (behavior) {
 281         case MADV_REMOVE:
 282                 return madvise_remove(vma, prev, start, end);
 283         case MADV_WILLNEED:
 284                 return madvise_willneed(vma, prev, start, end);
 285         case MADV_DONTNEED:
 286                 return madvise_dontneed(vma, prev, start, end);
 287         default:
 288                 return madvise_behavior(vma, prev, start, end, behavior);
 289         }
 290 }
 291
 292 static int
 293 madvise_behavior_valid(int behavior)
 294 {
 295         switch (behavior) {
 296         case MADV_DOFORK:
 297         case MADV_DONTFORK:
 298         case MADV_NORMAL:
 299         case MADV_SEQUENTIAL:
 300         case MADV_RANDOM:
 301         case MADV_REMOVE:
 302         case MADV_WILLNEED:
 303         case MADV_DONTNEED:
 304 #ifdef CONFIG_KSM
 305         case MADV_MERGEABLE:
 306         case MADV_UNMERGEABLE:
 307 #endif
 308 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 309         case MADV_HUGEPAGE:
 310         case MADV_NOHUGEPAGE:
 311 #endif
 312         case MADV_DONTDUMP:
 313         case MADV_DODUMP:
 314                 return 1;
 315
 316         default:
 317                 return 0;
 318         }
 319 }
 320
 321 /*
 322  * The madvise(2) system call.
 323  *
 324  * Applications can use madvise() to advise the kernel how it should
 325  * handle paging I/O in this VM area.  The idea is to help the kernel
 326  * use appropriate read-ahead and caching techniques.  The information
 327  * provided is advisory only, and can be safely disregarded by the
 328  * kernel without affecting the correct operation of the application.
 329  *
 330  * behavior values:
 331  *  MADV_NORMAL - the default behavior is to read clusters.  This
 332  *              results in some read-ahead and read-behind.
 333  *  MADV_RANDOM - the system should read the minimum amount of data
 334  *              on any access, since it is unlikely that the appli-
 335  *              cation will need more than what it asks for.
 336  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 337  *              once, so they can be aggressively read ahead, and
 338  *              can be freed soon after they are accessed.
 339  *  MADV_WILLNEED - the application is notifying the system to read
 340  *              some pages ahead.
 341  *  MADV_DONTNEED - the application is finished with the given range,
 342  *              so the kernel can free resources associated with it.
 343  *  MADV_REMOVE - the application wants to free up the given range of
 344  *              pages and associated backing store.
 345  *  MADV_DONTFORK - omit this area from child's address space when forking:
 346  *              typically, to avoid COWing pages pinned by get_user_pages().
 347  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 348  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 349  *              this area with pages of identical content from other such areas.
 350  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 351  *
 352  * return values:
 353  *  zero    - success
 354  *  -EINVAL - start + len < 0, start is not page-aligned,
 355  *              "behavior" is not a valid value, or application
 356  *              is attempting to release locked or shared pages.
 357  *  -ENOMEM - addresses in the specified range are not currently
 358  *              mapped, or are outside the AS of the process.
 359  *  -EIO    - an I/O error occurred while paging in data.
 360  *  -EBADF  - map exists, but area maps something that isn't a file.
 361  *  -EAGAIN - a kernel resource was temporarily unavailable.
 362  */
 363 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 364 {
 365         unsigned long end, tmp;
 366         struct vm_area_struct * vma, *prev;
 367         int unmapped_error = 0;
 368         int error = -EINVAL;
 369         int write;
 370         size_t len;
 371
 372 #ifdef CONFIG_MEMORY_FAILURE
 373         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 374                 return madvise_hwpoison(behavior, start, start+len_in);
 375 #endif
 376         if (!madvise_behavior_valid(behavior))
 377                 return error;
 378
 379         write = madvise_need_mmap_write(behavior);
 380         if (write)
 381                 down_write(&current->mm->mmap_sem);
 382         else
 383                 down_read(&current->mm->mmap_sem);
 384
 385         if (start & ~PAGE_MASK)
 386                 goto out;
 387         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 388
 389         /* Check to see whether len was rounded up from small -ve to zero */
 390         if (len_in && !len)
 391                 goto out;
 392
 393         end = start + len;
 394         if (end < start)
 395                 goto out;
 396
 397         error = 0;
 398         if (end == start)
 399                 goto out;
 400
 401         /*
 402          * If the interval [start,end) covers some unmapped address
 403          * ranges, just ignore them, but return -ENOMEM at the end.
 404          * - different from the way of handling in mlock etc.
 405          */
 406         vma = find_vma_prev(current->mm, start, &prev);
 407         if (vma && start > vma->vm_start)
 408                 prev = vma;
 409
 410         for (;;) {
 411                 /* Still start < end. */
 412                 error = -ENOMEM;
 413                 if (!vma)
 414                         goto out;
 415
 416                 /* Here start < (end|vma->vm_end). */
 417                 if (start < vma->vm_start) {
 418                         unmapped_error = -ENOMEM;
 419                         start = vma->vm_start;
 420                         if (start >= end)
 421                                 goto out;
 422                 }
 423
 424                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 425                 tmp = vma->vm_end;
 426                 if (end < tmp)
 427                         tmp = end;
 428
 429                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 430                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 431                 if (error)
 432                         goto out;
 433                 start = tmp;
 434                 if (prev && start < prev->vm_end)
 435                         start = prev->vm_end;
 436                 error = unmapped_error;
 437                 if (start >= end)
 438                         goto out;
 439                 if (prev)
 440                         vma = prev->vm_next;
 441                 else    /* madvise_remove dropped mmap_sem */
 442                         vma = find_vma(current->mm, start);
 443         }
 444 out:
 445         if (write)
 446                 up_write(&current->mm->mmap_sem);
 447         else
 448                 up_read(&current->mm->mmap_sem);
 449
 450         return error;
 451 }