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