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