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