mm/shmem.c

   1 /*
   2  * Resizable virtual memory filesystem for Linux.
   3  *
   4  * Copyright (C) 2000 Linus Torvalds.
   5  *               2000 Transmeta Corp.
   6  *               2000-2001 Christoph Rohland
   7  *               2000-2001 SAP AG
   8  *               2002 Red Hat Inc.
   9  * Copyright (C) 2002-2011 Hugh Dickins.
  10  * Copyright (C) 2011 Google Inc.
  11  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  12  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  13  *
  14  * Extended attribute support for tmpfs:
  15  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
  16  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
  17  *
  18  * tiny-shmem:
  19  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
  20  *
  21  * This file is released under the GPL.
  22  */
  23
  24 #include <linux/fs.h>
  25 #include <linux/init.h>
  26 #include <linux/vfs.h>
  27 #include <linux/mount.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/file.h>
  30 #include <linux/mm.h>
  31 #include <linux/export.h>
  32 #include <linux/swap.h>
  33
  34 static struct vfsmount *shm_mnt;
  35
  36 #ifdef CONFIG_SHMEM
  37 /*
  38  * This virtual memory filesystem is heavily based on the ramfs. It
  39  * extends ramfs by the ability to use swap and honor resource limits
  40  * which makes it a completely usable filesystem.
  41  */
  42
  43 #include <linux/xattr.h>
  44 #include <linux/exportfs.h>
  45 #include <linux/posix_acl.h>
  46 #include <linux/generic_acl.h>
  47 #include <linux/mman.h>
  48 #include <linux/string.h>
  49 #include <linux/slab.h>
  50 #include <linux/backing-dev.h>
  51 #include <linux/shmem_fs.h>
  52 #include <linux/writeback.h>
  53 #include <linux/blkdev.h>
  54 #include <linux/pagevec.h>
  55 #include <linux/percpu_counter.h>
  56 #include <linux/falloc.h>
  57 #include <linux/splice.h>
  58 #include <linux/security.h>
  59 #include <linux/swapops.h>
  60 #include <linux/mempolicy.h>
  61 #include <linux/namei.h>
  62 #include <linux/ctype.h>
  63 #include <linux/migrate.h>
  64 #include <linux/highmem.h>
  65 #include <linux/seq_file.h>
  66 #include <linux/magic.h>
  67
  68 #include <asm/uaccess.h>
  69 #include <asm/pgtable.h>
  70
  71 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  72 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  73
  74 /* Pretend that each entry is of this size in directory's i_size */
  75 #define BOGO_DIRENT_SIZE 20
  76
  77 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  78 #define SHORT_SYMLINK_LEN 128
  79
  80 struct shmem_xattr {
  81         struct list_head list;  /* anchored by shmem_inode_info->xattr_list */
  82         char *name;             /* xattr name */
  83         size_t size;
  84         char value[0];
  85 };
  86
  87 /*
  88  * shmem_fallocate and shmem_writepage communicate via inode->i_private
  89  * (with i_mutex making sure that it has only one user at a time):
  90  * we would prefer not to enlarge the shmem inode just for that.
  91  */
  92 struct shmem_falloc {
  93         pgoff_t start;          /* start of range currently being fallocated */
  94         pgoff_t next;           /* the next page offset to be fallocated */
  95         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
  96         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
  97 };
  98
  99 /* Flag allocation requirements to shmem_getpage */
 100 enum sgp_type {
 101         SGP_READ,       /* don't exceed i_size, don't allocate page */
 102         SGP_CACHE,      /* don't exceed i_size, may allocate page */
 103         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
 104         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
 105         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
 106 };
 107
 108 #ifdef CONFIG_TMPFS
 109 static unsigned long shmem_default_max_blocks(void)
 110 {
 111         return totalram_pages / 2;
 112 }
 113
 114 static unsigned long shmem_default_max_inodes(void)
 115 {
 116         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 117 }
 118 #endif
 119
 120 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
 121 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 122                                 struct shmem_inode_info *info, pgoff_t index);
 123 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 124         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
 125
 126 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
 127         struct page **pagep, enum sgp_type sgp, int *fault_type)
 128 {
 129         return shmem_getpage_gfp(inode, index, pagep, sgp,
 130                         mapping_gfp_mask(inode->i_mapping), fault_type);
 131 }
 132
 133 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 134 {
 135         return sb->s_fs_info;
 136 }
 137
 138 /*
 139  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 140  * for shared memory and for shared anonymous (/dev/zero) mappings
 141  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 142  * consistent with the pre-accounting of private mappings ...
 143  */
 144 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 145 {
 146         return (flags & VM_NORESERVE) ?
 147                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 148 }
 149
 150 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 151 {
 152         if (!(flags & VM_NORESERVE))
 153                 vm_unacct_memory(VM_ACCT(size));
 154 }
 155
 156 /*
 157  * ... whereas tmpfs objects are accounted incrementally as
 158  * pages are allocated, in order to allow huge sparse files.
 159  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 160  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 161  */
 162 static inline int shmem_acct_block(unsigned long flags)
 163 {
 164         return (flags & VM_NORESERVE) ?
 165                 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 166 }
 167
 168 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 169 {
 170         if (flags & VM_NORESERVE)
 171                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 172 }
 173
 174 static const struct super_operations shmem_ops;
 175 static const struct address_space_operations shmem_aops;
 176 static const struct file_operations shmem_file_operations;
 177 static const struct inode_operations shmem_inode_operations;
 178 static const struct inode_operations shmem_dir_inode_operations;
 179 static const struct inode_operations shmem_special_inode_operations;
 180 static const struct vm_operations_struct shmem_vm_ops;
 181
 182 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 183         .ra_pages       = 0,    /* No readahead */
 184         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
 185 };
 186
 187 static LIST_HEAD(shmem_swaplist);
 188 static DEFINE_MUTEX(shmem_swaplist_mutex);
 189
 190 static int shmem_reserve_inode(struct super_block *sb)
 191 {
 192         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 193         if (sbinfo->max_inodes) {
 194                 spin_lock(&sbinfo->stat_lock);
 195                 if (!sbinfo->free_inodes) {
 196                         spin_unlock(&sbinfo->stat_lock);
 197                         return -ENOSPC;
 198                 }
 199                 sbinfo->free_inodes--;
 200                 spin_unlock(&sbinfo->stat_lock);
 201         }
 202         return 0;
 203 }
 204
 205 static void shmem_free_inode(struct super_block *sb)
 206 {
 207         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 208         if (sbinfo->max_inodes) {
 209                 spin_lock(&sbinfo->stat_lock);
 210                 sbinfo->free_inodes++;
 211                 spin_unlock(&sbinfo->stat_lock);
 212         }
 213 }
 214
 215 /**
 216  * shmem_recalc_inode - recalculate the block usage of an inode
 217  * @inode: inode to recalc
 218  *
 219  * We have to calculate the free blocks since the mm can drop
 220  * undirtied hole pages behind our back.
 221  *
 222  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 223  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 224  *
 225  * It has to be called with the spinlock held.
 226  */
 227 static void shmem_recalc_inode(struct inode *inode)
 228 {
 229         struct shmem_inode_info *info = SHMEM_I(inode);
 230         long freed;
 231
 232         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 233         if (freed > 0) {
 234                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 235                 if (sbinfo->max_blocks)
 236                         percpu_counter_add(&sbinfo->used_blocks, -freed);
 237                 info->alloced -= freed;
 238                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 239                 shmem_unacct_blocks(info->flags, freed);
 240         }
 241 }
 242
 243 /*
 244  * Replace item expected in radix tree by a new item, while holding tree lock.
 245  */
 246 static int shmem_radix_tree_replace(struct address_space *mapping,
 247                         pgoff_t index, void *expected, void *replacement)
 248 {
 249         void **pslot;
 250         void *item = NULL;
 251
 252         VM_BUG_ON(!expected);
 253         pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
 254         if (pslot)
 255                 item = radix_tree_deref_slot_protected(pslot,
 256                                                         &mapping->tree_lock);
 257         if (item != expected)
 258                 return -ENOENT;
 259         if (replacement)
 260                 radix_tree_replace_slot(pslot, replacement);
 261         else
 262                 radix_tree_delete(&mapping->page_tree, index);
 263         return 0;
 264 }
 265
 266 /*
 267  * Like add_to_page_cache_locked, but error if expected item has gone.
 268  */
 269 static int shmem_add_to_page_cache(struct page *page,
 270                                    struct address_space *mapping,
 271                                    pgoff_t index, gfp_t gfp, void *expected)
 272 {
 273         int error = 0;
 274
 275         VM_BUG_ON(!PageLocked(page));
 276         VM_BUG_ON(!PageSwapBacked(page));
 277
 278         if (!expected)
 279                 error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
 280         if (!error) {
 281                 page_cache_get(page);
 282                 page->mapping = mapping;
 283                 page->index = index;
 284
 285                 spin_lock_irq(&mapping->tree_lock);
 286                 if (!expected)
 287                         error = radix_tree_insert(&mapping->page_tree,
 288                                                         index, page);
 289                 else
 290                         error = shmem_radix_tree_replace(mapping, index,
 291                                                         expected, page);
 292                 if (!error) {
 293                         mapping->nrpages++;
 294                         __inc_zone_page_state(page, NR_FILE_PAGES);
 295                         __inc_zone_page_state(page, NR_SHMEM);
 296                         spin_unlock_irq(&mapping->tree_lock);
 297                 } else {
 298                         page->mapping = NULL;
 299                         spin_unlock_irq(&mapping->tree_lock);
 300                         page_cache_release(page);
 301                 }
 302                 if (!expected)
 303                         radix_tree_preload_end();
 304         }
 305         if (error)
 306                 mem_cgroup_uncharge_cache_page(page);
 307         return error;
 308 }
 309
 310 /*
 311  * Like delete_from_page_cache, but substitutes swap for page.
 312  */
 313 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 314 {
 315         struct address_space *mapping = page->mapping;
 316         int error;
 317
 318         spin_lock_irq(&mapping->tree_lock);
 319         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
 320         page->mapping = NULL;
 321         mapping->nrpages--;
 322         __dec_zone_page_state(page, NR_FILE_PAGES);
 323         __dec_zone_page_state(page, NR_SHMEM);
 324         spin_unlock_irq(&mapping->tree_lock);
 325         page_cache_release(page);
 326         BUG_ON(error);
 327 }
 328
 329 /*
 330  * Like find_get_pages, but collecting swap entries as well as pages.
 331  */
 332 static unsigned shmem_find_get_pages_and_swap(struct address_space *mapping,
 333                                         pgoff_t start, unsigned int nr_pages,
 334                                         struct page **pages, pgoff_t *indices)
 335 {
 336         unsigned int i;
 337         unsigned int ret;
 338         unsigned int nr_found;
 339
 340         rcu_read_lock();
 341 restart:
 342         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
 343                                 (void ***)pages, indices, start, nr_pages);
 344         ret = 0;
 345         for (i = 0; i < nr_found; i++) {
 346                 struct page *page;
 347 repeat:
 348                 page = radix_tree_deref_slot((void **)pages[i]);
 349                 if (unlikely(!page))
 350                         continue;
 351                 if (radix_tree_exception(page)) {
 352                         if (radix_tree_deref_retry(page))
 353                                 goto restart;
 354                         /*
 355                          * Otherwise, we must be storing a swap entry
 356                          * here as an exceptional entry: so return it
 357                          * without attempting to raise page count.
 358                          */
 359                         goto export;
 360                 }
 361                 if (!page_cache_get_speculative(page))
 362                         goto repeat;
 363
 364                 /* Has the page moved? */
 365                 if (unlikely(page != *((void **)pages[i]))) {
 366                         page_cache_release(page);
 367                         goto repeat;
 368                 }
 369 export:
 370                 indices[ret] = indices[i];
 371                 pages[ret] = page;
 372                 ret++;
 373         }
 374         if (unlikely(!ret && nr_found))
 375                 goto restart;
 376         rcu_read_unlock();
 377         return ret;
 378 }
 379
 380 /*
 381  * Remove swap entry from radix tree, free the swap and its page cache.
 382  */
 383 static int shmem_free_swap(struct address_space *mapping,
 384                            pgoff_t index, void *radswap)
 385 {
 386         int error;
 387
 388         spin_lock_irq(&mapping->tree_lock);
 389         error = shmem_radix_tree_replace(mapping, index, radswap, NULL);
 390         spin_unlock_irq(&mapping->tree_lock);
 391         if (!error)
 392                 free_swap_and_cache(radix_to_swp_entry(radswap));
 393         return error;
 394 }
 395
 396 /*
 397  * Pagevec may contain swap entries, so shuffle up pages before releasing.
 398  */
 399 static void shmem_deswap_pagevec(struct pagevec *pvec)
 400 {
 401         int i, j;
 402
 403         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
 404                 struct page *page = pvec->pages[i];
 405                 if (!radix_tree_exceptional_entry(page))
 406                         pvec->pages[j++] = page;
 407         }
 408         pvec->nr = j;
 409 }
 410
 411 /*
 412  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 413  */
 414 void shmem_unlock_mapping(struct address_space *mapping)
 415 {
 416         struct pagevec pvec;
 417         pgoff_t indices[PAGEVEC_SIZE];
 418         pgoff_t index = 0;
 419
 420         pagevec_init(&pvec, 0);
 421         /*
 422          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 423          */
 424         while (!mapping_unevictable(mapping)) {
 425                 /*
 426                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
 427                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
 428                  */
 429                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 430                                         PAGEVEC_SIZE, pvec.pages, indices);
 431                 if (!pvec.nr)
 432                         break;
 433                 index = indices[pvec.nr - 1] + 1;
 434                 shmem_deswap_pagevec(&pvec);
 435                 check_move_unevictable_pages(pvec.pages, pvec.nr);
 436                 pagevec_release(&pvec);
 437                 cond_resched();
 438         }
 439 }
 440
 441 /*
 442  * Remove range of pages and swap entries from radix tree, and free them.
 443  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 444  */
 445 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 446                                                                  bool unfalloc)
 447 {
 448         struct address_space *mapping = inode->i_mapping;
 449         struct shmem_inode_info *info = SHMEM_I(inode);
 450         pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 451         pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
 452         unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
 453         unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
 454         struct pagevec pvec;
 455         pgoff_t indices[PAGEVEC_SIZE];
 456         long nr_swaps_freed = 0;
 457         pgoff_t index;
 458         int i;
 459
 460         if (lend == -1)
 461                 end = -1;       /* unsigned, so actually very big */
 462
 463         pagevec_init(&pvec, 0);
 464         index = start;
 465         while (index < end) {
 466                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 467                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 468                                                         pvec.pages, indices);
 469                 if (!pvec.nr)
 470                         break;
 471                 mem_cgroup_uncharge_start();
 472                 for (i = 0; i < pagevec_count(&pvec); i++) {
 473                         struct page *page = pvec.pages[i];
 474
 475                         index = indices[i];
 476                         if (index >= end)
 477                                 break;
 478
 479                         if (radix_tree_exceptional_entry(page)) {
 480                                 if (unfalloc)
 481                                         continue;
 482                                 nr_swaps_freed += !shmem_free_swap(mapping,
 483                                                                 index, page);
 484                                 continue;
 485                         }
 486
 487                         if (!trylock_page(page))
 488                                 continue;
 489                         if (!unfalloc || !PageUptodate(page)) {
 490                                 if (page->mapping == mapping) {
 491                                         VM_BUG_ON(PageWriteback(page));
 492                                         truncate_inode_page(mapping, page);
 493                                 }
 494                         }
 495                         unlock_page(page);
 496                 }
 497                 shmem_deswap_pagevec(&pvec);
 498                 pagevec_release(&pvec);
 499                 mem_cgroup_uncharge_end();
 500                 cond_resched();
 501                 index++;
 502         }
 503
 504         if (partial_start) {
 505                 struct page *page = NULL;
 506                 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
 507                 if (page) {
 508                         unsigned int top = PAGE_CACHE_SIZE;
 509                         if (start > end) {
 510                                 top = partial_end;
 511                                 partial_end = 0;
 512                         }
 513                         zero_user_segment(page, partial_start, top);
 514                         set_page_dirty(page);
 515                         unlock_page(page);
 516                         page_cache_release(page);
 517                 }
 518         }
 519         if (partial_end) {
 520                 struct page *page = NULL;
 521                 shmem_getpage(inode, end, &page, SGP_READ, NULL);
 522                 if (page) {
 523                         zero_user_segment(page, 0, partial_end);
 524                         set_page_dirty(page);
 525                         unlock_page(page);
 526                         page_cache_release(page);
 527                 }
 528         }
 529         if (start >= end)
 530                 return;
 531
 532         index = start;
 533         for ( ; ; ) {
 534                 cond_resched();
 535                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 536                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 537                                                         pvec.pages, indices);
 538                 if (!pvec.nr) {
 539                         if (index == start || unfalloc)
 540                                 break;
 541                         index = start;
 542                         continue;
 543                 }
 544                 if ((index == start || unfalloc) && indices[0] >= end) {
 545                         shmem_deswap_pagevec(&pvec);
 546                         pagevec_release(&pvec);
 547                         break;
 548                 }
 549                 mem_cgroup_uncharge_start();
 550                 for (i = 0; i < pagevec_count(&pvec); i++) {
 551                         struct page *page = pvec.pages[i];
 552
 553                         index = indices[i];
 554                         if (index >= end)
 555                                 break;
 556
 557                         if (radix_tree_exceptional_entry(page)) {
 558                                 if (unfalloc)
 559                                         continue;
 560                                 nr_swaps_freed += !shmem_free_swap(mapping,
 561                                                                 index, page);
 562                                 continue;
 563                         }
 564
 565                         lock_page(page);
 566                         if (!unfalloc || !PageUptodate(page)) {
 567                                 if (page->mapping == mapping) {
 568                                         VM_BUG_ON(PageWriteback(page));
 569                                         truncate_inode_page(mapping, page);
 570                                 }
 571                         }
 572                         unlock_page(page);
 573                 }
 574                 shmem_deswap_pagevec(&pvec);
 575                 pagevec_release(&pvec);
 576                 mem_cgroup_uncharge_end();
 577                 index++;
 578         }
 579
 580         spin_lock(&info->lock);
 581         info->swapped -= nr_swaps_freed;
 582         shmem_recalc_inode(inode);
 583         spin_unlock(&info->lock);
 584 }
 585
 586 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
 587 {
 588         shmem_undo_range(inode, lstart, lend, false);
 589         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 590 }
 591 EXPORT_SYMBOL_GPL(shmem_truncate_range);
 592
 593 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
 594 {
 595         struct inode *inode = dentry->d_inode;
 596         int error;
 597
 598         error = inode_change_ok(inode, attr);
 599         if (error)
 600                 return error;
 601
 602         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
 603                 loff_t oldsize = inode->i_size;
 604                 loff_t newsize = attr->ia_size;
 605
 606                 if (newsize != oldsize) {
 607                         i_size_write(inode, newsize);
 608                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 609                 }
 610                 if (newsize < oldsize) {
 611                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
 612                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 613                         shmem_truncate_range(inode, newsize, (loff_t)-1);
 614                         /* unmap again to remove racily COWed private pages */
 615                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 616                 }
 617         }
 618
 619         setattr_copy(inode, attr);
 620 #ifdef CONFIG_TMPFS_POSIX_ACL
 621         if (attr->ia_valid & ATTR_MODE)
 622                 error = generic_acl_chmod(inode);
 623 #endif
 624         return error;
 625 }
 626
 627 static void shmem_evict_inode(struct inode *inode)
 628 {
 629         struct shmem_inode_info *info = SHMEM_I(inode);
 630         struct shmem_xattr *xattr, *nxattr;
 631
 632         if (inode->i_mapping->a_ops == &shmem_aops) {
 633                 shmem_unacct_size(info->flags, inode->i_size);
 634                 inode->i_size = 0;
 635                 shmem_truncate_range(inode, 0, (loff_t)-1);
 636                 if (!list_empty(&info->swaplist)) {
 637                         mutex_lock(&shmem_swaplist_mutex);
 638                         list_del_init(&info->swaplist);
 639                         mutex_unlock(&shmem_swaplist_mutex);
 640                 }
 641         } else
 642                 kfree(info->symlink);
 643
 644         list_for_each_entry_safe(xattr, nxattr, &info->xattr_list, list) {
 645                 kfree(xattr->name);
 646                 kfree(xattr);
 647         }
 648         BUG_ON(inode->i_blocks);
 649         shmem_free_inode(inode->i_sb);
 650         clear_inode(inode);
 651 }
 652
 653 /*
 654  * If swap found in inode, free it and move page from swapcache to filecache.
 655  */
 656 static int shmem_unuse_inode(struct shmem_inode_info *info,
 657                              swp_entry_t swap, struct page **pagep)
 658 {
 659         struct address_space *mapping = info->vfs_inode.i_mapping;
 660         void *radswap;
 661         pgoff_t index;
 662         gfp_t gfp;
 663         int error = 0;
 664
 665         radswap = swp_to_radix_entry(swap);
 666         index = radix_tree_locate_item(&mapping->page_tree, radswap);
 667         if (index == -1)
 668                 return 0;
 669
 670         /*
 671          * Move _head_ to start search for next from here.
 672          * But be careful: shmem_evict_inode checks list_empty without taking
 673          * mutex, and there's an instant in list_move_tail when info->swaplist
 674          * would appear empty, if it were the only one on shmem_swaplist.
 675          */
 676         if (shmem_swaplist.next != &info->swaplist)
 677                 list_move_tail(&shmem_swaplist, &info->swaplist);
 678
 679         gfp = mapping_gfp_mask(mapping);
 680         if (shmem_should_replace_page(*pagep, gfp)) {
 681                 mutex_unlock(&shmem_swaplist_mutex);
 682                 error = shmem_replace_page(pagep, gfp, info, index);
 683                 mutex_lock(&shmem_swaplist_mutex);
 684                 /*
 685                  * We needed to drop mutex to make that restrictive page
 686                  * allocation, but the inode might have been freed while we
 687                  * dropped it: although a racing shmem_evict_inode() cannot
 688                  * complete without emptying the radix_tree, our page lock
 689                  * on this swapcache page is not enough to prevent that -
 690                  * free_swap_and_cache() of our swap entry will only
 691                  * trylock_page(), removing swap from radix_tree whatever.
 692                  *
 693                  * We must not proceed to shmem_add_to_page_cache() if the
 694                  * inode has been freed, but of course we cannot rely on
 695                  * inode or mapping or info to check that.  However, we can
 696                  * safely check if our swap entry is still in use (and here
 697                  * it can't have got reused for another page): if it's still
 698                  * in use, then the inode cannot have been freed yet, and we
 699                  * can safely proceed (if it's no longer in use, that tells
 700                  * nothing about the inode, but we don't need to unuse swap).
 701                  */
 702                 if (!page_swapcount(*pagep))
 703                         error = -ENOENT;
 704         }
 705
 706         /*
 707          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 708          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 709          * beneath us (pagelock doesn't help until the page is in pagecache).
 710          */
 711         if (!error)
 712                 error = shmem_add_to_page_cache(*pagep, mapping, index,
 713                                                 GFP_NOWAIT, radswap);
 714         if (error != -ENOMEM) {
 715                 /*
 716                  * Truncation and eviction use free_swap_and_cache(), which
 717                  * only does trylock page: if we raced, best clean up here.
 718                  */
 719                 delete_from_swap_cache(*pagep);
 720                 set_page_dirty(*pagep);
 721                 if (!error) {
 722                         spin_lock(&info->lock);
 723                         info->swapped--;
 724                         spin_unlock(&info->lock);
 725                         swap_free(swap);
 726                 }
 727                 error = 1;      /* not an error, but entry was found */
 728         }
 729         return error;
 730 }
 731
 732 /*
 733  * Search through swapped inodes to find and replace swap by page.
 734  */
 735 int shmem_unuse(swp_entry_t swap, struct page *page)
 736 {
 737         struct list_head *this, *next;
 738         struct shmem_inode_info *info;
 739         int found = 0;
 740         int error = 0;
 741
 742         /*
 743          * There's a faint possibility that swap page was replaced before
 744          * caller locked it: caller will come back later with the right page.
 745          */
 746         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
 747                 goto out;
 748
 749         /*
 750          * Charge page using GFP_KERNEL while we can wait, before taking
 751          * the shmem_swaplist_mutex which might hold up shmem_writepage().
 752          * Charged back to the user (not to caller) when swap account is used.
 753          */
 754         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
 755         if (error)
 756                 goto out;
 757         /* No radix_tree_preload: swap entry keeps a place for page in tree */
 758
 759         mutex_lock(&shmem_swaplist_mutex);
 760         list_for_each_safe(this, next, &shmem_swaplist) {
 761                 info = list_entry(this, struct shmem_inode_info, swaplist);
 762                 if (info->swapped)
 763                         found = shmem_unuse_inode(info, swap, &page);
 764                 else
 765                         list_del_init(&info->swaplist);
 766                 cond_resched();
 767                 if (found)
 768                         break;
 769         }
 770         mutex_unlock(&shmem_swaplist_mutex);
 771
 772         if (found < 0)
 773                 error = found;
 774 out:
 775         unlock_page(page);
 776         page_cache_release(page);
 777         return error;
 778 }
 779
 780 /*
 781  * Move the page from the page cache to the swap cache.
 782  */
 783 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 784 {
 785         struct shmem_inode_info *info;
 786         struct address_space *mapping;
 787         struct inode *inode;
 788         swp_entry_t swap;
 789         pgoff_t index;
 790
 791         BUG_ON(!PageLocked(page));
 792         mapping = page->mapping;
 793         index = page->index;
 794         inode = mapping->host;
 795         info = SHMEM_I(inode);
 796         if (info->flags & VM_LOCKED)
 797                 goto redirty;
 798         if (!total_swap_pages)
 799                 goto redirty;
 800
 801         /*
 802          * shmem_backing_dev_info's capabilities prevent regular writeback or
 803          * sync from ever calling shmem_writepage; but a stacking filesystem
 804          * might use ->writepage of its underlying filesystem, in which case
 805          * tmpfs should write out to swap only in response to memory pressure,
 806          * and not for the writeback threads or sync.
 807          */
 808         if (!wbc->for_reclaim) {
 809                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
 810                 goto redirty;
 811         }
 812
 813         /*
 814          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
 815          * value into swapfile.c, the only way we can correctly account for a
 816          * fallocated page arriving here is now to initialize it and write it.
 817          *
 818          * That's okay for a page already fallocated earlier, but if we have
 819          * not yet completed the fallocation, then (a) we want to keep track
 820          * of this page in case we have to undo it, and (b) it may not be a
 821          * good idea to continue anyway, once we're pushing into swap.  So
 822          * reactivate the page, and let shmem_fallocate() quit when too many.
 823          */
 824         if (!PageUptodate(page)) {
 825                 if (inode->i_private) {
 826                         struct shmem_falloc *shmem_falloc;
 827                         spin_lock(&inode->i_lock);
 828                         shmem_falloc = inode->i_private;
 829                         if (shmem_falloc &&
 830                             index >= shmem_falloc->start &&
 831                             index < shmem_falloc->next)
 832                                 shmem_falloc->nr_unswapped++;
 833                         else
 834                                 shmem_falloc = NULL;
 835                         spin_unlock(&inode->i_lock);
 836                         if (shmem_falloc)
 837                                 goto redirty;
 838                 }
 839                 clear_highpage(page);
 840                 flush_dcache_page(page);
 841                 SetPageUptodate(page);
 842         }
 843
 844         swap = get_swap_page();
 845         if (!swap.val)
 846                 goto redirty;
 847
 848         /*
 849          * Add inode to shmem_unuse()'s list of swapped-out inodes,
 850          * if it's not already there.  Do it now before the page is
 851          * moved to swap cache, when its pagelock no longer protects
 852          * the inode from eviction.  But don't unlock the mutex until
 853          * we've incremented swapped, because shmem_unuse_inode() will
 854          * prune a !swapped inode from the swaplist under this mutex.
 855          */
 856         mutex_lock(&shmem_swaplist_mutex);
 857         if (list_empty(&info->swaplist))
 858                 list_add_tail(&info->swaplist, &shmem_swaplist);
 859
 860         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
 861                 swap_shmem_alloc(swap);
 862                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
 863
 864                 spin_lock(&info->lock);
 865                 info->swapped++;
 866                 shmem_recalc_inode(inode);
 867                 spin_unlock(&info->lock);
 868
 869                 mutex_unlock(&shmem_swaplist_mutex);
 870                 BUG_ON(page_mapped(page));
 871                 swap_writepage(page, wbc);
 872                 return 0;
 873         }
 874
 875         mutex_unlock(&shmem_swaplist_mutex);
 876         swapcache_free(swap, NULL);
 877 redirty:
 878         set_page_dirty(page);
 879         if (wbc->for_reclaim)
 880                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
 881         unlock_page(page);
 882         return 0;
 883 }
 884
 885 #ifdef CONFIG_NUMA
 886 #ifdef CONFIG_TMPFS
 887 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 888 {
 889         char buffer[64];
 890
 891         if (!mpol || mpol->mode == MPOL_DEFAULT)
 892                 return;         /* show nothing */
 893
 894         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
 895
 896         seq_printf(seq, ",mpol=%s", buffer);
 897 }
 898
 899 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 900 {
 901         struct mempolicy *mpol = NULL;
 902         if (sbinfo->mpol) {
 903                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
 904                 mpol = sbinfo->mpol;
 905                 mpol_get(mpol);
 906                 spin_unlock(&sbinfo->stat_lock);
 907         }
 908         return mpol;
 909 }
 910 #endif /* CONFIG_TMPFS */
 911
 912 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 913                         struct shmem_inode_info *info, pgoff_t index)
 914 {
 915         struct mempolicy mpol, *spol;
 916         struct vm_area_struct pvma;
 917
 918         spol = mpol_cond_copy(&mpol,
 919                         mpol_shared_policy_lookup(&info->policy, index));
 920
 921         /* Create a pseudo vma that just contains the policy */
 922         pvma.vm_start = 0;
 923         pvma.vm_pgoff = index;
 924         pvma.vm_ops = NULL;
 925         pvma.vm_policy = spol;
 926         return swapin_readahead(swap, gfp, &pvma, 0);
 927 }
 928
 929 static struct page *shmem_alloc_page(gfp_t gfp,
 930                         struct shmem_inode_info *info, pgoff_t index)
 931 {
 932         struct vm_area_struct pvma;
 933
 934         /* Create a pseudo vma that just contains the policy */
 935         pvma.vm_start = 0;
 936         pvma.vm_pgoff = index;
 937         pvma.vm_ops = NULL;
 938         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 939
 940         /*
 941          * alloc_page_vma() will drop the shared policy reference
 942          */
 943         return alloc_page_vma(gfp, &pvma, 0);
 944 }
 945 #else /* !CONFIG_NUMA */
 946 #ifdef CONFIG_TMPFS
 947 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 948 {
 949 }
 950 #endif /* CONFIG_TMPFS */
 951
 952 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 953                         struct shmem_inode_info *info, pgoff_t index)
 954 {
 955         return swapin_readahead(swap, gfp, NULL, 0);
 956 }
 957
 958 static inline struct page *shmem_alloc_page(gfp_t gfp,
 959                         struct shmem_inode_info *info, pgoff_t index)
 960 {
 961         return alloc_page(gfp);
 962 }
 963 #endif /* CONFIG_NUMA */
 964
 965 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
 966 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 967 {
 968         return NULL;
 969 }
 970 #endif
 971
 972 /*
 973  * When a page is moved from swapcache to shmem filecache (either by the
 974  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
 975  * shmem_unuse_inode()), it may have been read in earlier from swap, in
 976  * ignorance of the mapping it belongs to.  If that mapping has special
 977  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
 978  * we may need to copy to a suitable page before moving to filecache.
 979  *
 980  * In a future release, this may well be extended to respect cpuset and
 981  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
 982  * but for now it is a simple matter of zone.
 983  */
 984 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
 985 {
 986         return page_zonenum(page) > gfp_zone(gfp);
 987 }
 988
 989 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 990                                 struct shmem_inode_info *info, pgoff_t index)
 991 {
 992         struct page *oldpage, *newpage;
 993         struct address_space *swap_mapping;
 994         pgoff_t swap_index;
 995         int error;
 996
 997         oldpage = *pagep;
 998         swap_index = page_private(oldpage);
 999         swap_mapping = page_mapping(oldpage);
1000
1001         /*
1002          * We have arrived here because our zones are constrained, so don't
1003          * limit chance of success by further cpuset and node constraints.
1004          */
1005         gfp &= ~GFP_CONSTRAINT_MASK;
1006         newpage = shmem_alloc_page(gfp, info, index);
1007         if (!newpage)
1008                 return -ENOMEM;
1009
1010         page_cache_get(newpage);
1011         copy_highpage(newpage, oldpage);
1012         flush_dcache_page(newpage);
1013
1014         __set_page_locked(newpage);
1015         SetPageUptodate(newpage);
1016         SetPageSwapBacked(newpage);
1017         set_page_private(newpage, swap_index);
1018         SetPageSwapCache(newpage);
1019
1020         /*
1021          * Our caller will very soon move newpage out of swapcache, but it's
1022          * a nice clean interface for us to replace oldpage by newpage there.
1023          */
1024         spin_lock_irq(&swap_mapping->tree_lock);
1025         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1026                                                                    newpage);
1027         if (!error) {
1028                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1029                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1030         }
1031         spin_unlock_irq(&swap_mapping->tree_lock);
1032
1033         if (unlikely(error)) {
1034                 /*
1035                  * Is this possible?  I think not, now that our callers check
1036                  * both PageSwapCache and page_private after getting page lock;
1037                  * but be defensive.  Reverse old to newpage for clear and free.
1038                  */
1039                 oldpage = newpage;
1040         } else {
1041                 mem_cgroup_replace_page_cache(oldpage, newpage);
1042                 lru_cache_add_anon(newpage);
1043                 *pagep = newpage;
1044         }
1045
1046         ClearPageSwapCache(oldpage);
1047         set_page_private(oldpage, 0);
1048
1049         unlock_page(oldpage);
1050         page_cache_release(oldpage);
1051         page_cache_release(oldpage);
1052         return error;
1053 }
1054
1055 /*
1056  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1057  *
1058  * If we allocate a new one we do not mark it dirty. That's up to the
1059  * vm. If we swap it in we mark it dirty since we also free the swap
1060  * entry since a page cannot live in both the swap and page cache
1061  */
1062 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1063         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1064 {
1065         struct address_space *mapping = inode->i_mapping;
1066         struct shmem_inode_info *info;
1067         struct shmem_sb_info *sbinfo;
1068         struct page *page;
1069         swp_entry_t swap;
1070         int error;
1071         int once = 0;
1072         int alloced = 0;
1073
1074         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1075                 return -EFBIG;
1076 repeat:
1077         swap.val = 0;
1078         page = find_lock_page(mapping, index);
1079         if (radix_tree_exceptional_entry(page)) {
1080                 swap = radix_to_swp_entry(page);
1081                 page = NULL;
1082         }
1083
1084         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1085             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1086                 error = -EINVAL;
1087                 goto failed;
1088         }
1089
1090         /* fallocated page? */
1091         if (page && !PageUptodate(page)) {
1092                 if (sgp != SGP_READ)
1093                         goto clear;
1094                 unlock_page(page);
1095                 page_cache_release(page);
1096                 page = NULL;
1097         }
1098         if (page || (sgp == SGP_READ && !swap.val)) {
1099                 *pagep = page;
1100                 return 0;
1101         }
1102
1103         /*
1104          * Fast cache lookup did not find it:
1105          * bring it back from swap or allocate.
1106          */
1107         info = SHMEM_I(inode);
1108         sbinfo = SHMEM_SB(inode->i_sb);
1109
1110         if (swap.val) {
1111                 /* Look it up and read it in.. */
1112                 page = lookup_swap_cache(swap);
1113                 if (!page) {
1114                         /* here we actually do the io */
1115                         if (fault_type)
1116                                 *fault_type |= VM_FAULT_MAJOR;
1117                         page = shmem_swapin(swap, gfp, info, index);
1118                         if (!page) {
1119                                 error = -ENOMEM;
1120                                 goto failed;
1121                         }
1122                 }
1123
1124                 /* We have to do this with page locked to prevent races */
1125                 lock_page(page);
1126                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1127                     page->mapping) {
1128                         error = -EEXIST;        /* try again */
1129                         goto failed;
1130                 }
1131                 if (!PageUptodate(page)) {
1132                         error = -EIO;
1133                         goto failed;
1134                 }
1135                 wait_on_page_writeback(page);
1136
1137                 if (shmem_should_replace_page(page, gfp)) {
1138                         error = shmem_replace_page(&page, gfp, info, index);
1139                         if (error)
1140                                 goto failed;
1141                 }
1142
1143                 error = mem_cgroup_cache_charge(page, current->mm,
1144                                                 gfp & GFP_RECLAIM_MASK);
1145                 if (!error)
1146                         error = shmem_add_to_page_cache(page, mapping, index,
1147                                                 gfp, swp_to_radix_entry(swap));
1148                 if (error)
1149                         goto failed;
1150
1151                 spin_lock(&info->lock);
1152                 info->swapped--;
1153                 shmem_recalc_inode(inode);
1154                 spin_unlock(&info->lock);
1155
1156                 delete_from_swap_cache(page);
1157                 set_page_dirty(page);
1158                 swap_free(swap);
1159
1160         } else {
1161                 if (shmem_acct_block(info->flags)) {
1162                         error = -ENOSPC;
1163                         goto failed;
1164                 }
1165                 if (sbinfo->max_blocks) {
1166                         if (percpu_counter_compare(&sbinfo->used_blocks,
1167                                                 sbinfo->max_blocks) >= 0) {
1168                                 error = -ENOSPC;
1169                                 goto unacct;
1170                         }
1171                         percpu_counter_inc(&sbinfo->used_blocks);
1172                 }
1173
1174                 page = shmem_alloc_page(gfp, info, index);
1175                 if (!page) {
1176                         error = -ENOMEM;
1177                         goto decused;
1178                 }
1179
1180                 SetPageSwapBacked(page);
1181                 __set_page_locked(page);
1182                 error = mem_cgroup_cache_charge(page, current->mm,
1183                                                 gfp & GFP_RECLAIM_MASK);
1184                 if (!error)
1185                         error = shmem_add_to_page_cache(page, mapping, index,
1186                                                 gfp, NULL);
1187                 if (error)
1188                         goto decused;
1189                 lru_cache_add_anon(page);
1190
1191                 spin_lock(&info->lock);
1192                 info->alloced++;
1193                 inode->i_blocks += BLOCKS_PER_PAGE;
1194                 shmem_recalc_inode(inode);
1195                 spin_unlock(&info->lock);
1196                 alloced = true;
1197
1198                 /*
1199                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1200                  */
1201                 if (sgp == SGP_FALLOC)
1202                         sgp = SGP_WRITE;
1203 clear:
1204                 /*
1205                  * Let SGP_WRITE caller clear ends if write does not fill page;
1206                  * but SGP_FALLOC on a page fallocated earlier must initialize
1207                  * it now, lest undo on failure cancel our earlier guarantee.
1208                  */
1209                 if (sgp != SGP_WRITE) {
1210                         clear_highpage(page);
1211                         flush_dcache_page(page);
1212                         SetPageUptodate(page);
1213                 }
1214                 if (sgp == SGP_DIRTY)
1215                         set_page_dirty(page);
1216         }
1217
1218         /* Perhaps the file has been truncated since we checked */
1219         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1220             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1221                 error = -EINVAL;
1222                 if (alloced)
1223                         goto trunc;
1224                 else
1225                         goto failed;
1226         }
1227         *pagep = page;
1228         return 0;
1229
1230         /*
1231          * Error recovery.
1232          */
1233 trunc:
1234         info = SHMEM_I(inode);
1235         ClearPageDirty(page);
1236         delete_from_page_cache(page);
1237         spin_lock(&info->lock);
1238         info->alloced--;
1239         inode->i_blocks -= BLOCKS_PER_PAGE;
1240         spin_unlock(&info->lock);
1241 decused:
1242         sbinfo = SHMEM_SB(inode->i_sb);
1243         if (sbinfo->max_blocks)
1244                 percpu_counter_add(&sbinfo->used_blocks, -1);
1245 unacct:
1246         shmem_unacct_blocks(info->flags, 1);
1247 failed:
1248         if (swap.val && error != -EINVAL) {
1249                 struct page *test = find_get_page(mapping, index);
1250                 if (test && !radix_tree_exceptional_entry(test))
1251                         page_cache_release(test);
1252                 /* Have another try if the entry has changed */
1253                 if (test != swp_to_radix_entry(swap))
1254                         error = -EEXIST;
1255         }
1256         if (page) {
1257                 unlock_page(page);
1258                 page_cache_release(page);
1259         }
1260         if (error == -ENOSPC && !once++) {
1261                 info = SHMEM_I(inode);
1262                 spin_lock(&info->lock);
1263                 shmem_recalc_inode(inode);
1264                 spin_unlock(&info->lock);
1265                 goto repeat;
1266         }
1267         if (error == -EEXIST)
1268                 goto repeat;
1269         return error;
1270 }
1271
1272 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1273 {
1274         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1275         int error;
1276         int ret = VM_FAULT_LOCKED;
1277
1278         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1279         if (error)
1280                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1281
1282         if (ret & VM_FAULT_MAJOR) {
1283                 count_vm_event(PGMAJFAULT);
1284                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1285         }
1286         return ret;
1287 }
1288
1289 #ifdef CONFIG_NUMA
1290 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1291 {
1292         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1293         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1294 }
1295
1296 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1297                                           unsigned long addr)
1298 {
1299         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1300         pgoff_t index;
1301
1302         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1303         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1304 }
1305 #endif
1306
1307 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1308 {
1309         struct inode *inode = file->f_path.dentry->d_inode;
1310         struct shmem_inode_info *info = SHMEM_I(inode);
1311         int retval = -ENOMEM;
1312
1313         spin_lock(&info->lock);
1314         if (lock && !(info->flags & VM_LOCKED)) {
1315                 if (!user_shm_lock(inode->i_size, user))
1316                         goto out_nomem;
1317                 info->flags |= VM_LOCKED;
1318                 mapping_set_unevictable(file->f_mapping);
1319         }
1320         if (!lock && (info->flags & VM_LOCKED) && user) {
1321                 user_shm_unlock(inode->i_size, user);
1322                 info->flags &= ~VM_LOCKED;
1323                 mapping_clear_unevictable(file->f_mapping);
1324         }
1325         retval = 0;
1326
1327 out_nomem:
1328         spin_unlock(&info->lock);
1329         return retval;
1330 }
1331
1332 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1333 {
1334         file_accessed(file);
1335         vma->vm_ops = &shmem_vm_ops;
1336         vma->vm_flags |= VM_CAN_NONLINEAR;
1337         return 0;
1338 }
1339
1340 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1341                                      umode_t mode, dev_t dev, unsigned long flags)
1342 {
1343         struct inode *inode;
1344         struct shmem_inode_info *info;
1345         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1346
1347         if (shmem_reserve_inode(sb))
1348                 return NULL;
1349
1350         inode = new_inode(sb);
1351         if (inode) {
1352                 inode->i_ino = get_next_ino();
1353                 inode_init_owner(inode, dir, mode);
1354                 inode->i_blocks = 0;
1355                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1356                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1357                 inode->i_generation = get_seconds();
1358                 info = SHMEM_I(inode);
1359                 memset(info, 0, (char *)inode - (char *)info);
1360                 spin_lock_init(&info->lock);
1361                 info->flags = flags & VM_NORESERVE;
1362                 INIT_LIST_HEAD(&info->swaplist);
1363                 INIT_LIST_HEAD(&info->xattr_list);
1364                 cache_no_acl(inode);
1365
1366                 switch (mode & S_IFMT) {
1367                 default:
1368                         inode->i_op = &shmem_special_inode_operations;
1369                         init_special_inode(inode, mode, dev);
1370                         break;
1371                 case S_IFREG:
1372                         inode->i_mapping->a_ops = &shmem_aops;
1373                         inode->i_op = &shmem_inode_operations;
1374                         inode->i_fop = &shmem_file_operations;
1375                         mpol_shared_policy_init(&info->policy,
1376                                                  shmem_get_sbmpol(sbinfo));
1377                         break;
1378                 case S_IFDIR:
1379                         inc_nlink(inode);
1380                         /* Some things misbehave if size == 0 on a directory */
1381                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1382                         inode->i_op = &shmem_dir_inode_operations;
1383                         inode->i_fop = &simple_dir_operations;
1384                         break;
1385                 case S_IFLNK:
1386                         /*
1387                          * Must not load anything in the rbtree,
1388                          * mpol_free_shared_policy will not be called.
1389                          */
1390                         mpol_shared_policy_init(&info->policy, NULL);
1391                         break;
1392                 }
1393         } else
1394                 shmem_free_inode(sb);
1395         return inode;
1396 }
1397
1398 #ifdef CONFIG_TMPFS
1399 static const struct inode_operations shmem_symlink_inode_operations;
1400 static const struct inode_operations shmem_short_symlink_operations;
1401
1402 #ifdef CONFIG_TMPFS_XATTR
1403 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1404 #else
1405 #define shmem_initxattrs NULL
1406 #endif
1407
1408 static int
1409 shmem_write_begin(struct file *file, struct address_space *mapping,
1410                         loff_t pos, unsigned len, unsigned flags,
1411                         struct page **pagep, void **fsdata)
1412 {
1413         struct inode *inode = mapping->host;
1414         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1415         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1416 }
1417
1418 static int
1419 shmem_write_end(struct file *file, struct address_space *mapping,
1420                         loff_t pos, unsigned len, unsigned copied,
1421                         struct page *page, void *fsdata)
1422 {
1423         struct inode *inode = mapping->host;
1424
1425         if (pos + copied > inode->i_size)
1426                 i_size_write(inode, pos + copied);
1427
1428         if (!PageUptodate(page)) {
1429                 if (copied < PAGE_CACHE_SIZE) {
1430                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1431                         zero_user_segments(page, 0, from,
1432                                         from + copied, PAGE_CACHE_SIZE);
1433                 }
1434                 SetPageUptodate(page);
1435         }
1436         set_page_dirty(page);
1437         unlock_page(page);
1438         page_cache_release(page);
1439
1440         return copied;
1441 }
1442
1443 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1444 {
1445         struct inode *inode = filp->f_path.dentry->d_inode;
1446         struct address_space *mapping = inode->i_mapping;
1447         pgoff_t index;
1448         unsigned long offset;
1449         enum sgp_type sgp = SGP_READ;
1450
1451         /*
1452          * Might this read be for a stacking filesystem?  Then when reading
1453          * holes of a sparse file, we actually need to allocate those pages,
1454          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1455          */
1456         if (segment_eq(get_fs(), KERNEL_DS))
1457                 sgp = SGP_DIRTY;
1458
1459         index = *ppos >> PAGE_CACHE_SHIFT;
1460         offset = *ppos & ~PAGE_CACHE_MASK;
1461
1462         for (;;) {
1463                 struct page *page = NULL;
1464                 pgoff_t end_index;
1465                 unsigned long nr, ret;
1466                 loff_t i_size = i_size_read(inode);
1467
1468                 end_index = i_size >> PAGE_CACHE_SHIFT;
1469                 if (index > end_index)
1470                         break;
1471                 if (index == end_index) {
1472                         nr = i_size & ~PAGE_CACHE_MASK;
1473                         if (nr <= offset)
1474                                 break;
1475                 }
1476
1477                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1478                 if (desc->error) {
1479                         if (desc->error == -EINVAL)
1480                                 desc->error = 0;
1481                         break;
1482                 }
1483                 if (page)
1484                         unlock_page(page);
1485
1486                 /*
1487                  * We must evaluate after, since reads (unlike writes)
1488                  * are called without i_mutex protection against truncate
1489                  */
1490                 nr = PAGE_CACHE_SIZE;
1491                 i_size = i_size_read(inode);
1492                 end_index = i_size >> PAGE_CACHE_SHIFT;
1493                 if (index == end_index) {
1494                         nr = i_size & ~PAGE_CACHE_MASK;
1495                         if (nr <= offset) {
1496                                 if (page)
1497                                         page_cache_release(page);
1498                                 break;
1499                         }
1500                 }
1501                 nr -= offset;
1502
1503                 if (page) {
1504                         /*
1505                          * If users can be writing to this page using arbitrary
1506                          * virtual addresses, take care about potential aliasing
1507                          * before reading the page on the kernel side.
1508                          */
1509                         if (mapping_writably_mapped(mapping))
1510                                 flush_dcache_page(page);
1511                         /*
1512                          * Mark the page accessed if we read the beginning.
1513                          */
1514                         if (!offset)
1515                                 mark_page_accessed(page);
1516                 } else {
1517                         page = ZERO_PAGE(0);
1518                         page_cache_get(page);
1519                 }
1520
1521                 /*
1522                  * Ok, we have the page, and it's up-to-date, so
1523                  * now we can copy it to user space...
1524                  *
1525                  * The actor routine returns how many bytes were actually used..
1526                  * NOTE! This may not be the same as how much of a user buffer
1527                  * we filled up (we may be padding etc), so we can only update
1528                  * "pos" here (the actor routine has to update the user buffer
1529                  * pointers and the remaining count).
1530                  */
1531                 ret = actor(desc, page, offset, nr);
1532                 offset += ret;
1533                 index += offset >> PAGE_CACHE_SHIFT;
1534                 offset &= ~PAGE_CACHE_MASK;
1535
1536                 page_cache_release(page);
1537                 if (ret != nr || !desc->count)
1538                         break;
1539
1540                 cond_resched();
1541         }
1542
1543         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1544         file_accessed(filp);
1545 }
1546
1547 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1548                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1549 {
1550         struct file *filp = iocb->ki_filp;
1551         ssize_t retval;
1552         unsigned long seg;
1553         size_t count;
1554         loff_t *ppos = &iocb->ki_pos;
1555
1556         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1557         if (retval)
1558                 return retval;
1559
1560         for (seg = 0; seg < nr_segs; seg++) {
1561                 read_descriptor_t desc;
1562
1563                 desc.written = 0;
1564                 desc.arg.buf = iov[seg].iov_base;
1565                 desc.count = iov[seg].iov_len;
1566                 if (desc.count == 0)
1567                         continue;
1568                 desc.error = 0;
1569                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1570                 retval += desc.written;
1571                 if (desc.error) {
1572                         retval = retval ?: desc.error;
1573                         break;
1574                 }
1575                 if (desc.count > 0)
1576                         break;
1577         }
1578         return retval;
1579 }
1580
1581 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1582                                 struct pipe_inode_info *pipe, size_t len,
1583                                 unsigned int flags)
1584 {
1585         struct address_space *mapping = in->f_mapping;
1586         struct inode *inode = mapping->host;
1587         unsigned int loff, nr_pages, req_pages;
1588         struct page *pages[PIPE_DEF_BUFFERS];
1589         struct partial_page partial[PIPE_DEF_BUFFERS];
1590         struct page *page;
1591         pgoff_t index, end_index;
1592         loff_t isize, left;
1593         int error, page_nr;
1594         struct splice_pipe_desc spd = {
1595                 .pages = pages,
1596                 .partial = partial,
1597                 .nr_pages_max = PIPE_DEF_BUFFERS,
1598                 .flags = flags,
1599                 .ops = &page_cache_pipe_buf_ops,
1600                 .spd_release = spd_release_page,
1601         };
1602
1603         isize = i_size_read(inode);
1604         if (unlikely(*ppos >= isize))
1605                 return 0;
1606
1607         left = isize - *ppos;
1608         if (unlikely(left < len))
1609                 len = left;
1610
1611         if (splice_grow_spd(pipe, &spd))
1612                 return -ENOMEM;
1613
1614         index = *ppos >> PAGE_CACHE_SHIFT;
1615         loff = *ppos & ~PAGE_CACHE_MASK;
1616         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1617         nr_pages = min(req_pages, pipe->buffers);
1618
1619         spd.nr_pages = find_get_pages_contig(mapping, index,
1620                                                 nr_pages, spd.pages);
1621         index += spd.nr_pages;
1622         error = 0;
1623
1624         while (spd.nr_pages < nr_pages) {
1625                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1626                 if (error)
1627                         break;
1628                 unlock_page(page);
1629                 spd.pages[spd.nr_pages++] = page;
1630                 index++;
1631         }
1632
1633         index = *ppos >> PAGE_CACHE_SHIFT;
1634         nr_pages = spd.nr_pages;
1635         spd.nr_pages = 0;
1636
1637         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1638                 unsigned int this_len;
1639
1640                 if (!len)
1641                         break;
1642
1643                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1644                 page = spd.pages[page_nr];
1645
1646                 if (!PageUptodate(page) || page->mapping != mapping) {
1647                         error = shmem_getpage(inode, index, &page,
1648                                                         SGP_CACHE, NULL);
1649                         if (error)
1650                                 break;
1651                         unlock_page(page);
1652                         page_cache_release(spd.pages[page_nr]);
1653                         spd.pages[page_nr] = page;
1654                 }
1655
1656                 isize = i_size_read(inode);
1657                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1658                 if (unlikely(!isize || index > end_index))
1659                         break;
1660
1661                 if (end_index == index) {
1662                         unsigned int plen;
1663
1664                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1665                         if (plen <= loff)
1666                                 break;
1667
1668                         this_len = min(this_len, plen - loff);
1669                         len = this_len;
1670                 }
1671
1672                 spd.partial[page_nr].offset = loff;
1673                 spd.partial[page_nr].len = this_len;
1674                 len -= this_len;
1675                 loff = 0;
1676                 spd.nr_pages++;
1677                 index++;
1678         }
1679
1680         while (page_nr < nr_pages)
1681                 page_cache_release(spd.pages[page_nr++]);
1682
1683         if (spd.nr_pages)
1684                 error = splice_to_pipe(pipe, &spd);
1685
1686         splice_shrink_spd(&spd);
1687
1688         if (error > 0) {
1689                 *ppos += error;
1690                 file_accessed(in);
1691         }
1692         return error;
1693 }
1694
1695 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
1696                                                          loff_t len)
1697 {
1698         struct inode *inode = file->f_path.dentry->d_inode;
1699         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1700         struct shmem_falloc shmem_falloc;
1701         pgoff_t start, index, end;
1702         int error;
1703
1704         mutex_lock(&inode->i_mutex);
1705
1706         if (mode & FALLOC_FL_PUNCH_HOLE) {
1707                 struct address_space *mapping = file->f_mapping;
1708                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
1709                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
1710
1711                 if ((u64)unmap_end > (u64)unmap_start)
1712                         unmap_mapping_range(mapping, unmap_start,
1713                                             1 + unmap_end - unmap_start, 0);
1714                 shmem_truncate_range(inode, offset, offset + len - 1);
1715                 /* No need to unmap again: hole-punching leaves COWed pages */
1716                 error = 0;
1717                 goto out;
1718         }
1719
1720         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
1721         error = inode_newsize_ok(inode, offset + len);
1722         if (error)
1723                 goto out;
1724
1725         start = offset >> PAGE_CACHE_SHIFT;
1726         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1727         /* Try to avoid a swapstorm if len is impossible to satisfy */
1728         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
1729                 error = -ENOSPC;
1730                 goto out;
1731         }
1732
1733         shmem_falloc.start = start;
1734         shmem_falloc.next  = start;
1735         shmem_falloc.nr_falloced = 0;
1736         shmem_falloc.nr_unswapped = 0;
1737         spin_lock(&inode->i_lock);
1738         inode->i_private = &shmem_falloc;
1739         spin_unlock(&inode->i_lock);
1740
1741         for (index = start; index < end; index++) {
1742                 struct page *page;
1743
1744                 /*
1745                  * Good, the fallocate(2) manpage permits EINTR: we may have
1746                  * been interrupted because we are using up too much memory.
1747                  */
1748                 if (signal_pending(current))
1749                         error = -EINTR;
1750                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
1751                         error = -ENOMEM;
1752                 else
1753                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
1754                                                                         NULL);
1755                 if (error) {
1756                         /* Remove the !PageUptodate pages we added */
1757                         shmem_undo_range(inode,
1758                                 (loff_t)start << PAGE_CACHE_SHIFT,
1759                                 (loff_t)index << PAGE_CACHE_SHIFT, true);
1760                         goto undone;
1761                 }
1762
1763                 /*
1764                  * Inform shmem_writepage() how far we have reached.
1765                  * No need for lock or barrier: we have the page lock.
1766                  */
1767                 shmem_falloc.next++;
1768                 if (!PageUptodate(page))
1769                         shmem_falloc.nr_falloced++;
1770
1771                 /*
1772                  * If !PageUptodate, leave it that way so that freeable pages
1773                  * can be recognized if we need to rollback on error later.
1774                  * But set_page_dirty so that memory pressure will swap rather
1775                  * than free the pages we are allocating (and SGP_CACHE pages
1776                  * might still be clean: we now need to mark those dirty too).
1777                  */
1778                 set_page_dirty(page);
1779                 unlock_page(page);
1780                 page_cache_release(page);
1781                 cond_resched();
1782         }
1783
1784         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
1785                 i_size_write(inode, offset + len);
1786         inode->i_ctime = CURRENT_TIME;
1787 undone:
1788         spin_lock(&inode->i_lock);
1789         inode->i_private = NULL;
1790         spin_unlock(&inode->i_lock);
1791 out:
1792         mutex_unlock(&inode->i_mutex);
1793         return error;
1794 }
1795
1796 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1797 {
1798         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1799
1800         buf->f_type = TMPFS_MAGIC;
1801         buf->f_bsize = PAGE_CACHE_SIZE;
1802         buf->f_namelen = NAME_MAX;
1803         if (sbinfo->max_blocks) {
1804                 buf->f_blocks = sbinfo->max_blocks;
1805                 buf->f_bavail =
1806                 buf->f_bfree  = sbinfo->max_blocks -
1807                                 percpu_counter_sum(&sbinfo->used_blocks);
1808         }
1809         if (sbinfo->max_inodes) {
1810                 buf->f_files = sbinfo->max_inodes;
1811                 buf->f_ffree = sbinfo->free_inodes;
1812         }
1813         /* else leave those fields 0 like simple_statfs */
1814         return 0;
1815 }
1816
1817 /*
1818  * File creation. Allocate an inode, and we're done..
1819  */
1820 static int
1821 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
1822 {
1823         struct inode *inode;
1824         int error = -ENOSPC;
1825
1826         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1827         if (inode) {
1828                 error = security_inode_init_security(inode, dir,
1829                                                      &dentry->d_name,
1830                                                      shmem_initxattrs, NULL);
1831                 if (error) {
1832                         if (error != -EOPNOTSUPP) {
1833                                 iput(inode);
1834                                 return error;
1835                         }
1836                 }
1837 #ifdef CONFIG_TMPFS_POSIX_ACL
1838                 error = generic_acl_init(inode, dir);
1839                 if (error) {
1840                         iput(inode);
1841                         return error;
1842                 }
1843 #else
1844                 error = 0;
1845 #endif
1846                 dir->i_size += BOGO_DIRENT_SIZE;
1847                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1848                 d_instantiate(dentry, inode);
1849                 dget(dentry); /* Extra count - pin the dentry in core */
1850         }
1851         return error;
1852 }
1853
1854 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1855 {
1856         int error;
1857
1858         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1859                 return error;
1860         inc_nlink(dir);
1861         return 0;
1862 }
1863
1864 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
1865                 struct nameidata *nd)
1866 {
1867         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1868 }
1869
1870 /*
1871  * Link a file..
1872  */
1873 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1874 {
1875         struct inode *inode = old_dentry->d_inode;
1876         int ret;
1877
1878         /*
1879          * No ordinary (disk based) filesystem counts links as inodes;
1880          * but each new link needs a new dentry, pinning lowmem, and
1881          * tmpfs dentries cannot be pruned until they are unlinked.
1882          */
1883         ret = shmem_reserve_inode(inode->i_sb);
1884         if (ret)
1885                 goto out;
1886
1887         dir->i_size += BOGO_DIRENT_SIZE;
1888         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1889         inc_nlink(inode);
1890         ihold(inode);   /* New dentry reference */
1891         dget(dentry);           /* Extra pinning count for the created dentry */
1892         d_instantiate(dentry, inode);
1893 out:
1894         return ret;
1895 }
1896
1897 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1898 {
1899         struct inode *inode = dentry->d_inode;
1900
1901         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1902                 shmem_free_inode(inode->i_sb);
1903
1904         dir->i_size -= BOGO_DIRENT_SIZE;
1905         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1906         drop_nlink(inode);
1907         dput(dentry);   /* Undo the count from "create" - this does all the work */
1908         return 0;
1909 }
1910
1911 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1912 {
1913         if (!simple_empty(dentry))
1914                 return -ENOTEMPTY;
1915
1916         drop_nlink(dentry->d_inode);
1917         drop_nlink(dir);
1918         return shmem_unlink(dir, dentry);
1919 }
1920
1921 /*
1922  * The VFS layer already does all the dentry stuff for rename,
1923  * we just have to decrement the usage count for the target if
1924  * it exists so that the VFS layer correctly free's it when it
1925  * gets overwritten.
1926  */
1927 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1928 {
1929         struct inode *inode = old_dentry->d_inode;
1930         int they_are_dirs = S_ISDIR(inode->i_mode);
1931
1932         if (!simple_empty(new_dentry))
1933                 return -ENOTEMPTY;
1934
1935         if (new_dentry->d_inode) {
1936                 (void) shmem_unlink(new_dir, new_dentry);
1937                 if (they_are_dirs)
1938                         drop_nlink(old_dir);
1939         } else if (they_are_dirs) {
1940                 drop_nlink(old_dir);
1941                 inc_nlink(new_dir);
1942         }
1943
1944         old_dir->i_size -= BOGO_DIRENT_SIZE;
1945         new_dir->i_size += BOGO_DIRENT_SIZE;
1946         old_dir->i_ctime = old_dir->i_mtime =
1947         new_dir->i_ctime = new_dir->i_mtime =
1948         inode->i_ctime = CURRENT_TIME;
1949         return 0;
1950 }
1951
1952 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1953 {
1954         int error;
1955         int len;
1956         struct inode *inode;
1957         struct page *page;
1958         char *kaddr;
1959         struct shmem_inode_info *info;
1960
1961         len = strlen(symname) + 1;
1962         if (len > PAGE_CACHE_SIZE)
1963                 return -ENAMETOOLONG;
1964
1965         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
1966         if (!inode)
1967                 return -ENOSPC;
1968
1969         error = security_inode_init_security(inode, dir, &dentry->d_name,
1970                                              shmem_initxattrs, NULL);
1971         if (error) {
1972                 if (error != -EOPNOTSUPP) {
1973                         iput(inode);
1974                         return error;
1975                 }
1976                 error = 0;
1977         }
1978
1979         info = SHMEM_I(inode);
1980         inode->i_size = len-1;
1981         if (len <= SHORT_SYMLINK_LEN) {
1982                 info->symlink = kmemdup(symname, len, GFP_KERNEL);
1983                 if (!info->symlink) {
1984                         iput(inode);
1985                         return -ENOMEM;
1986                 }
1987                 inode->i_op = &shmem_short_symlink_operations;
1988         } else {
1989                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1990                 if (error) {
1991                         iput(inode);
1992                         return error;
1993                 }
1994                 inode->i_mapping->a_ops = &shmem_aops;
1995                 inode->i_op = &shmem_symlink_inode_operations;
1996                 kaddr = kmap_atomic(page);
1997                 memcpy(kaddr, symname, len);
1998                 kunmap_atomic(kaddr);
1999                 SetPageUptodate(page);
2000                 set_page_dirty(page);
2001                 unlock_page(page);
2002                 page_cache_release(page);
2003         }
2004         dir->i_size += BOGO_DIRENT_SIZE;
2005         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2006         d_instantiate(dentry, inode);
2007         dget(dentry);
2008         return 0;
2009 }
2010
2011 static void *shmem_follow_short_symlink(struct dentry *dentry, struct nameidata *nd)
2012 {
2013         nd_set_link(nd, SHMEM_I(dentry->d_inode)->symlink);
2014         return NULL;
2015 }
2016
2017 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2018 {
2019         struct page *page = NULL;
2020         int error = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2021         nd_set_link(nd, error ? ERR_PTR(error) : kmap(page));
2022         if (page)
2023                 unlock_page(page);
2024         return page;
2025 }
2026
2027 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2028 {
2029         if (!IS_ERR(nd_get_link(nd))) {
2030                 struct page *page = cookie;
2031                 kunmap(page);
2032                 mark_page_accessed(page);
2033                 page_cache_release(page);
2034         }
2035 }
2036
2037 #ifdef CONFIG_TMPFS_XATTR
2038 /*
2039  * Superblocks without xattr inode operations may get some security.* xattr
2040  * support from the LSM "for free". As soon as we have any other xattrs
2041  * like ACLs, we also need to implement the security.* handlers at
2042  * filesystem level, though.
2043  */
2044
2045 /*
2046  * Allocate new xattr and copy in the value; but leave the name to callers.
2047  */
2048 static struct shmem_xattr *shmem_xattr_alloc(const void *value, size_t size)
2049 {
2050         struct shmem_xattr *new_xattr;
2051         size_t len;
2052
2053         /* wrap around? */
2054         len = sizeof(*new_xattr) + size;
2055         if (len <= sizeof(*new_xattr))
2056                 return NULL;
2057
2058         new_xattr = kmalloc(len, GFP_KERNEL);
2059         if (!new_xattr)
2060                 return NULL;
2061
2062         new_xattr->size = size;
2063         memcpy(new_xattr->value, value, size);
2064         return new_xattr;
2065 }
2066
2067 /*
2068  * Callback for security_inode_init_security() for acquiring xattrs.
2069  */
2070 static int shmem_initxattrs(struct inode *inode,
2071                             const struct xattr *xattr_array,
2072                             void *fs_info)
2073 {
2074         struct shmem_inode_info *info = SHMEM_I(inode);
2075         const struct xattr *xattr;
2076         struct shmem_xattr *new_xattr;
2077         size_t len;
2078
2079         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2080                 new_xattr = shmem_xattr_alloc(xattr->value, xattr->value_len);
2081                 if (!new_xattr)
2082                         return -ENOMEM;
2083
2084                 len = strlen(xattr->name) + 1;
2085                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2086                                           GFP_KERNEL);
2087                 if (!new_xattr->name) {
2088                         kfree(new_xattr);
2089                         return -ENOMEM;
2090                 }
2091
2092                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2093                        XATTR_SECURITY_PREFIX_LEN);
2094                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2095                        xattr->name, len);
2096
2097                 spin_lock(&info->lock);
2098                 list_add(&new_xattr->list, &info->xattr_list);
2099                 spin_unlock(&info->lock);
2100         }
2101
2102         return 0;
2103 }
2104
2105 static int shmem_xattr_get(struct dentry *dentry, const char *name,
2106                            void *buffer, size_t size)
2107 {
2108         struct shmem_inode_info *info;
2109         struct shmem_xattr *xattr;
2110         int ret = -ENODATA;
2111
2112         info = SHMEM_I(dentry->d_inode);
2113
2114         spin_lock(&info->lock);
2115         list_for_each_entry(xattr, &info->xattr_list, list) {
2116                 if (strcmp(name, xattr->name))
2117                         continue;
2118
2119                 ret = xattr->size;
2120                 if (buffer) {
2121                         if (size < xattr->size)
2122                                 ret = -ERANGE;
2123                         else
2124                                 memcpy(buffer, xattr->value, xattr->size);
2125                 }
2126                 break;
2127         }
2128         spin_unlock(&info->lock);
2129         return ret;
2130 }
2131
2132 static int shmem_xattr_set(struct inode *inode, const char *name,
2133                            const void *value, size_t size, int flags)
2134 {
2135         struct shmem_inode_info *info = SHMEM_I(inode);
2136         struct shmem_xattr *xattr;
2137         struct shmem_xattr *new_xattr = NULL;
2138         int err = 0;
2139
2140         /* value == NULL means remove */
2141         if (value) {
2142                 new_xattr = shmem_xattr_alloc(value, size);
2143                 if (!new_xattr)
2144                         return -ENOMEM;
2145
2146                 new_xattr->name = kstrdup(name, GFP_KERNEL);
2147                 if (!new_xattr->name) {
2148                         kfree(new_xattr);
2149                         return -ENOMEM;
2150                 }
2151         }
2152
2153         spin_lock(&info->lock);
2154         list_for_each_entry(xattr, &info->xattr_list, list) {
2155                 if (!strcmp(name, xattr->name)) {
2156                         if (flags & XATTR_CREATE) {
2157                                 xattr = new_xattr;
2158                                 err = -EEXIST;
2159                         } else if (new_xattr) {
2160                                 list_replace(&xattr->list, &new_xattr->list);
2161                         } else {
2162                                 list_del(&xattr->list);
2163                         }
2164                         goto out;
2165                 }
2166         }
2167         if (flags & XATTR_REPLACE) {
2168                 xattr = new_xattr;
2169                 err = -ENODATA;
2170         } else {
2171                 list_add(&new_xattr->list, &info->xattr_list);
2172                 xattr = NULL;
2173         }
2174 out:
2175         spin_unlock(&info->lock);
2176         if (xattr)
2177                 kfree(xattr->name);
2178         kfree(xattr);
2179         return err;
2180 }
2181
2182 static const struct xattr_handler *shmem_xattr_handlers[] = {
2183 #ifdef CONFIG_TMPFS_POSIX_ACL
2184         &generic_acl_access_handler,
2185         &generic_acl_default_handler,
2186 #endif
2187         NULL
2188 };
2189
2190 static int shmem_xattr_validate(const char *name)
2191 {
2192         struct { const char *prefix; size_t len; } arr[] = {
2193                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2194                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2195         };
2196         int i;
2197
2198         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2199                 size_t preflen = arr[i].len;
2200                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2201                         if (!name[preflen])
2202                                 return -EINVAL;
2203                         return 0;
2204                 }
2205         }
2206         return -EOPNOTSUPP;
2207 }
2208
2209 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2210                               void *buffer, size_t size)
2211 {
2212         int err;
2213
2214         /*
2215          * If this is a request for a synthetic attribute in the system.*
2216          * namespace use the generic infrastructure to resolve a handler
2217          * for it via sb->s_xattr.
2218          */
2219         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2220                 return generic_getxattr(dentry, name, buffer, size);
2221
2222         err = shmem_xattr_validate(name);
2223         if (err)
2224                 return err;
2225
2226         return shmem_xattr_get(dentry, name, buffer, size);
2227 }
2228
2229 static int shmem_setxattr(struct dentry *dentry, const char *name,
2230                           const void *value, size_t size, int flags)
2231 {
2232         int err;
2233
2234         /*
2235          * If this is a request for a synthetic attribute in the system.*
2236          * namespace use the generic infrastructure to resolve a handler
2237          * for it via sb->s_xattr.
2238          */
2239         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2240                 return generic_setxattr(dentry, name, value, size, flags);
2241
2242         err = shmem_xattr_validate(name);
2243         if (err)
2244                 return err;
2245
2246         if (size == 0)
2247                 value = "";  /* empty EA, do not remove */
2248
2249         return shmem_xattr_set(dentry->d_inode, name, value, size, flags);
2250
2251 }
2252
2253 static int shmem_removexattr(struct dentry *dentry, const char *name)
2254 {
2255         int err;
2256
2257         /*
2258          * If this is a request for a synthetic attribute in the system.*
2259          * namespace use the generic infrastructure to resolve a handler
2260          * for it via sb->s_xattr.
2261          */
2262         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2263                 return generic_removexattr(dentry, name);
2264
2265         err = shmem_xattr_validate(name);
2266         if (err)
2267                 return err;
2268
2269         return shmem_xattr_set(dentry->d_inode, name, NULL, 0, XATTR_REPLACE);
2270 }
2271
2272 static bool xattr_is_trusted(const char *name)
2273 {
2274         return !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN);
2275 }
2276
2277 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2278 {
2279         bool trusted = capable(CAP_SYS_ADMIN);
2280         struct shmem_xattr *xattr;
2281         struct shmem_inode_info *info;
2282         size_t used = 0;
2283
2284         info = SHMEM_I(dentry->d_inode);
2285
2286         spin_lock(&info->lock);
2287         list_for_each_entry(xattr, &info->xattr_list, list) {
2288                 size_t len;
2289
2290                 /* skip "trusted." attributes for unprivileged callers */
2291                 if (!trusted && xattr_is_trusted(xattr->name))
2292                         continue;
2293
2294                 len = strlen(xattr->name) + 1;
2295                 used += len;
2296                 if (buffer) {
2297                         if (size < used) {
2298                                 used = -ERANGE;
2299                                 break;
2300                         }
2301                         memcpy(buffer, xattr->name, len);
2302                         buffer += len;
2303                 }
2304         }
2305         spin_unlock(&info->lock);
2306
2307         return used;
2308 }
2309 #endif /* CONFIG_TMPFS_XATTR */
2310
2311 static const struct inode_operations shmem_short_symlink_operations = {
2312         .readlink       = generic_readlink,
2313         .follow_link    = shmem_follow_short_symlink,
2314 #ifdef CONFIG_TMPFS_XATTR
2315         .setxattr       = shmem_setxattr,
2316         .getxattr       = shmem_getxattr,
2317         .listxattr      = shmem_listxattr,
2318         .removexattr    = shmem_removexattr,
2319 #endif
2320 };
2321
2322 static const struct inode_operations shmem_symlink_inode_operations = {
2323         .readlink       = generic_readlink,
2324         .follow_link    = shmem_follow_link,
2325         .put_link       = shmem_put_link,
2326 #ifdef CONFIG_TMPFS_XATTR
2327         .setxattr       = shmem_setxattr,
2328         .getxattr       = shmem_getxattr,
2329         .listxattr      = shmem_listxattr,
2330         .removexattr    = shmem_removexattr,
2331 #endif
2332 };
2333
2334 static struct dentry *shmem_get_parent(struct dentry *child)
2335 {
2336         return ERR_PTR(-ESTALE);
2337 }
2338
2339 static int shmem_match(struct inode *ino, void *vfh)
2340 {
2341         __u32 *fh = vfh;
2342         __u64 inum = fh[2];
2343         inum = (inum << 32) | fh[1];
2344         return ino->i_ino == inum && fh[0] == ino->i_generation;
2345 }
2346
2347 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2348                 struct fid *fid, int fh_len, int fh_type)
2349 {
2350         struct inode *inode;
2351         struct dentry *dentry = NULL;
2352         u64 inum = fid->raw[2];
2353         inum = (inum << 32) | fid->raw[1];
2354
2355         if (fh_len < 3)
2356                 return NULL;
2357
2358         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2359                         shmem_match, fid->raw);
2360         if (inode) {
2361                 dentry = d_find_alias(inode);
2362                 iput(inode);
2363         }
2364
2365         return dentry;
2366 }
2367
2368 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2369                                 struct inode *parent)
2370 {
2371         if (*len < 3) {
2372                 *len = 3;
2373                 return 255;
2374         }
2375
2376         if (inode_unhashed(inode)) {
2377                 /* Unfortunately insert_inode_hash is not idempotent,
2378                  * so as we hash inodes here rather than at creation
2379                  * time, we need a lock to ensure we only try
2380                  * to do it once
2381                  */
2382                 static DEFINE_SPINLOCK(lock);
2383                 spin_lock(&lock);
2384                 if (inode_unhashed(inode))
2385                         __insert_inode_hash(inode,
2386                                             inode->i_ino + inode->i_generation);
2387                 spin_unlock(&lock);
2388         }
2389
2390         fh[0] = inode->i_generation;
2391         fh[1] = inode->i_ino;
2392         fh[2] = ((__u64)inode->i_ino) >> 32;
2393
2394         *len = 3;
2395         return 1;
2396 }
2397
2398 static const struct export_operations shmem_export_ops = {
2399         .get_parent     = shmem_get_parent,
2400         .encode_fh      = shmem_encode_fh,
2401         .fh_to_dentry   = shmem_fh_to_dentry,
2402 };
2403
2404 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2405                                bool remount)
2406 {
2407         char *this_char, *value, *rest;
2408         uid_t uid;
2409         gid_t gid;
2410
2411         while (options != NULL) {
2412                 this_char = options;
2413                 for (;;) {
2414                         /*
2415                          * NUL-terminate this option: unfortunately,
2416                          * mount options form a comma-separated list,
2417                          * but mpol's nodelist may also contain commas.
2418                          */
2419                         options = strchr(options, ',');
2420                         if (options == NULL)
2421                                 break;
2422                         options++;
2423                         if (!isdigit(*options)) {
2424                                 options[-1] = '\0';
2425                                 break;
2426                         }
2427                 }
2428                 if (!*this_char)
2429                         continue;
2430                 if ((value = strchr(this_char,'=')) != NULL) {
2431                         *value++ = 0;
2432                 } else {
2433                         printk(KERN_ERR
2434                             "tmpfs: No value for mount option '%s'\n",
2435                             this_char);
2436                         return 1;
2437                 }
2438
2439                 if (!strcmp(this_char,"size")) {
2440                         unsigned long long size;
2441                         size = memparse(value,&rest);
2442                         if (*rest == '%') {
2443                                 size <<= PAGE_SHIFT;
2444                                 size *= totalram_pages;
2445                                 do_div(size, 100);
2446                                 rest++;
2447                         }
2448                         if (*rest)
2449                                 goto bad_val;
2450                         sbinfo->max_blocks =
2451                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2452                 } else if (!strcmp(this_char,"nr_blocks")) {
2453                         sbinfo->max_blocks = memparse(value, &rest);
2454                         if (*rest)
2455                                 goto bad_val;
2456                 } else if (!strcmp(this_char,"nr_inodes")) {
2457                         sbinfo->max_inodes = memparse(value, &rest);
2458                         if (*rest)
2459                                 goto bad_val;
2460                 } else if (!strcmp(this_char,"mode")) {
2461                         if (remount)
2462                                 continue;
2463                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2464                         if (*rest)
2465                                 goto bad_val;
2466                 } else if (!strcmp(this_char,"uid")) {
2467                         if (remount)
2468                                 continue;
2469                         uid = simple_strtoul(value, &rest, 0);
2470                         if (*rest)
2471                                 goto bad_val;
2472                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2473                         if (!uid_valid(sbinfo->uid))
2474                                 goto bad_val;
2475                 } else if (!strcmp(this_char,"gid")) {
2476                         if (remount)
2477                                 continue;
2478                         gid = simple_strtoul(value, &rest, 0);
2479                         if (*rest)
2480                                 goto bad_val;
2481                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2482                         if (!gid_valid(sbinfo->gid))
2483                                 goto bad_val;
2484                 } else if (!strcmp(this_char,"mpol")) {
2485                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2486                                 goto bad_val;
2487                 } else {
2488                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2489                                this_char);
2490                         return 1;
2491                 }
2492         }
2493         return 0;
2494
2495 bad_val:
2496         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2497                value, this_char);
2498         return 1;
2499
2500 }
2501
2502 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2503 {
2504         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2505         struct shmem_sb_info config = *sbinfo;
2506         unsigned long inodes;
2507         int error = -EINVAL;
2508
2509         if (shmem_parse_options(data, &config, true))
2510                 return error;
2511
2512         spin_lock(&sbinfo->stat_lock);
2513         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2514         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2515                 goto out;
2516         if (config.max_inodes < inodes)
2517                 goto out;
2518         /*
2519          * Those tests disallow limited->unlimited while any are in use;
2520          * but we must separately disallow unlimited->limited, because
2521          * in that case we have no record of how much is already in use.
2522          */
2523         if (config.max_blocks && !sbinfo->max_blocks)
2524                 goto out;
2525         if (config.max_inodes && !sbinfo->max_inodes)
2526                 goto out;
2527
2528         error = 0;
2529         sbinfo->max_blocks  = config.max_blocks;
2530         sbinfo->max_inodes  = config.max_inodes;
2531         sbinfo->free_inodes = config.max_inodes - inodes;
2532
2533         mpol_put(sbinfo->mpol);
2534         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2535 out:
2536         spin_unlock(&sbinfo->stat_lock);
2537         return error;
2538 }
2539
2540 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2541 {
2542         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2543
2544         if (sbinfo->max_blocks != shmem_default_max_blocks())
2545                 seq_printf(seq, ",size=%luk",
2546                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2547         if (sbinfo->max_inodes != shmem_default_max_inodes())
2548                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2549         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2550                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2551         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2552                 seq_printf(seq, ",uid=%u",
2553                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2554         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2555                 seq_printf(seq, ",gid=%u",
2556                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2557         shmem_show_mpol(seq, sbinfo->mpol);
2558         return 0;
2559 }
2560 #endif /* CONFIG_TMPFS */
2561
2562 static void shmem_put_super(struct super_block *sb)
2563 {
2564         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2565
2566         percpu_counter_destroy(&sbinfo->used_blocks);
2567         kfree(sbinfo);
2568         sb->s_fs_info = NULL;
2569 }
2570
2571 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2572 {
2573         struct inode *inode;
2574         struct shmem_sb_info *sbinfo;
2575         int err = -ENOMEM;
2576
2577         /* Round up to L1_CACHE_BYTES to resist false sharing */
2578         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2579                                 L1_CACHE_BYTES), GFP_KERNEL);
2580         if (!sbinfo)
2581                 return -ENOMEM;
2582
2583         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2584         sbinfo->uid = current_fsuid();
2585         sbinfo->gid = current_fsgid();
2586         sb->s_fs_info = sbinfo;
2587
2588 #ifdef CONFIG_TMPFS
2589         /*
2590          * Per default we only allow half of the physical ram per
2591          * tmpfs instance, limiting inodes to one per page of lowmem;
2592          * but the internal instance is left unlimited.
2593          */
2594         if (!(sb->s_flags & MS_NOUSER)) {
2595                 sbinfo->max_blocks = shmem_default_max_blocks();
2596                 sbinfo->max_inodes = shmem_default_max_inodes();
2597                 if (shmem_parse_options(data, sbinfo, false)) {
2598                         err = -EINVAL;
2599                         goto failed;
2600                 }
2601         }
2602         sb->s_export_op = &shmem_export_ops;
2603         sb->s_flags |= MS_NOSEC;
2604 #else
2605         sb->s_flags |= MS_NOUSER;
2606 #endif
2607
2608         spin_lock_init(&sbinfo->stat_lock);
2609         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2610                 goto failed;
2611         sbinfo->free_inodes = sbinfo->max_inodes;
2612
2613         sb->s_maxbytes = MAX_LFS_FILESIZE;
2614         sb->s_blocksize = PAGE_CACHE_SIZE;
2615         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2616         sb->s_magic = TMPFS_MAGIC;
2617         sb->s_op = &shmem_ops;
2618         sb->s_time_gran = 1;
2619 #ifdef CONFIG_TMPFS_XATTR
2620         sb->s_xattr = shmem_xattr_handlers;
2621 #endif
2622 #ifdef CONFIG_TMPFS_POSIX_ACL
2623         sb->s_flags |= MS_POSIXACL;
2624 #endif
2625
2626         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2627         if (!inode)
2628                 goto failed;
2629         inode->i_uid = sbinfo->uid;
2630         inode->i_gid = sbinfo->gid;
2631         sb->s_root = d_make_root(inode);
2632         if (!sb->s_root)
2633                 goto failed;
2634         return 0;
2635
2636 failed:
2637         shmem_put_super(sb);
2638         return err;
2639 }
2640
2641 static struct kmem_cache *shmem_inode_cachep;
2642
2643 static struct inode *shmem_alloc_inode(struct super_block *sb)
2644 {
2645         struct shmem_inode_info *info;
2646         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2647         if (!info)
2648                 return NULL;
2649         return &info->vfs_inode;
2650 }
2651
2652 static void shmem_destroy_callback(struct rcu_head *head)
2653 {
2654         struct inode *inode = container_of(head, struct inode, i_rcu);
2655         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2656 }
2657
2658 static void shmem_destroy_inode(struct inode *inode)
2659 {
2660         if (S_ISREG(inode->i_mode))
2661                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2662         call_rcu(&inode->i_rcu, shmem_destroy_callback);
2663 }
2664
2665 static void shmem_init_inode(void *foo)
2666 {
2667         struct shmem_inode_info *info = foo;
2668         inode_init_once(&info->vfs_inode);
2669 }
2670
2671 static int shmem_init_inodecache(void)
2672 {
2673         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2674                                 sizeof(struct shmem_inode_info),
2675                                 0, SLAB_PANIC, shmem_init_inode);
2676         return 0;
2677 }
2678
2679 static void shmem_destroy_inodecache(void)
2680 {
2681         kmem_cache_destroy(shmem_inode_cachep);
2682 }
2683
2684 static const struct address_space_operations shmem_aops = {
2685         .writepage      = shmem_writepage,
2686         .set_page_dirty = __set_page_dirty_no_writeback,
2687 #ifdef CONFIG_TMPFS
2688         .write_begin    = shmem_write_begin,
2689         .write_end      = shmem_write_end,
2690 #endif
2691         .migratepage    = migrate_page,
2692         .error_remove_page = generic_error_remove_page,
2693 };
2694
2695 static const struct file_operations shmem_file_operations = {
2696         .mmap           = shmem_mmap,
2697 #ifdef CONFIG_TMPFS
2698         .llseek         = generic_file_llseek,
2699         .read           = do_sync_read,
2700         .write          = do_sync_write,
2701         .aio_read       = shmem_file_aio_read,
2702         .aio_write      = generic_file_aio_write,
2703         .fsync          = noop_fsync,
2704         .splice_read    = shmem_file_splice_read,
2705         .splice_write   = generic_file_splice_write,
2706         .fallocate      = shmem_fallocate,
2707 #endif
2708 };
2709
2710 static const struct inode_operations shmem_inode_operations = {
2711         .setattr        = shmem_setattr,
2712 #ifdef CONFIG_TMPFS_XATTR
2713         .setxattr       = shmem_setxattr,
2714         .getxattr       = shmem_getxattr,
2715         .listxattr      = shmem_listxattr,
2716         .removexattr    = shmem_removexattr,
2717 #endif
2718 };
2719
2720 static const struct inode_operations shmem_dir_inode_operations = {
2721 #ifdef CONFIG_TMPFS
2722         .create         = shmem_create,
2723         .lookup         = simple_lookup,
2724         .link           = shmem_link,
2725         .unlink         = shmem_unlink,
2726         .symlink        = shmem_symlink,
2727         .mkdir          = shmem_mkdir,
2728         .rmdir          = shmem_rmdir,
2729         .mknod          = shmem_mknod,
2730         .rename         = shmem_rename,
2731 #endif
2732 #ifdef CONFIG_TMPFS_XATTR
2733         .setxattr       = shmem_setxattr,
2734         .getxattr       = shmem_getxattr,
2735         .listxattr      = shmem_listxattr,
2736         .removexattr    = shmem_removexattr,
2737 #endif
2738 #ifdef CONFIG_TMPFS_POSIX_ACL
2739         .setattr        = shmem_setattr,
2740 #endif
2741 };
2742
2743 static const struct inode_operations shmem_special_inode_operations = {
2744 #ifdef CONFIG_TMPFS_XATTR
2745         .setxattr       = shmem_setxattr,
2746         .getxattr       = shmem_getxattr,
2747         .listxattr      = shmem_listxattr,
2748         .removexattr    = shmem_removexattr,
2749 #endif
2750 #ifdef CONFIG_TMPFS_POSIX_ACL
2751         .setattr        = shmem_setattr,
2752 #endif
2753 };
2754
2755 static const struct super_operations shmem_ops = {
2756         .alloc_inode    = shmem_alloc_inode,
2757         .destroy_inode  = shmem_destroy_inode,
2758 #ifdef CONFIG_TMPFS
2759         .statfs         = shmem_statfs,
2760         .remount_fs     = shmem_remount_fs,
2761         .show_options   = shmem_show_options,
2762 #endif
2763         .evict_inode    = shmem_evict_inode,
2764         .drop_inode     = generic_delete_inode,
2765         .put_super      = shmem_put_super,
2766 };
2767
2768 static const struct vm_operations_struct shmem_vm_ops = {
2769         .fault          = shmem_fault,
2770 #ifdef CONFIG_NUMA
2771         .set_policy     = shmem_set_policy,
2772         .get_policy     = shmem_get_policy,
2773 #endif
2774 };
2775
2776 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2777         int flags, const char *dev_name, void *data)
2778 {
2779         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2780 }
2781
2782 static struct file_system_type shmem_fs_type = {
2783         .owner          = THIS_MODULE,
2784         .name           = "tmpfs",
2785         .mount          = shmem_mount,
2786         .kill_sb        = kill_litter_super,
2787 };
2788
2789 int __init shmem_init(void)
2790 {
2791         int error;
2792
2793         error = bdi_init(&shmem_backing_dev_info);
2794         if (error)
2795                 goto out4;
2796
2797         error = shmem_init_inodecache();
2798         if (error)
2799                 goto out3;
2800
2801         error = register_filesystem(&shmem_fs_type);
2802         if (error) {
2803                 printk(KERN_ERR "Could not register tmpfs\n");
2804                 goto out2;
2805         }
2806
2807         shm_mnt = vfs_kern_mount(&shmem_fs_type, MS_NOUSER,
2808                                  shmem_fs_type.name, NULL);
2809         if (IS_ERR(shm_mnt)) {
2810                 error = PTR_ERR(shm_mnt);
2811                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2812                 goto out1;
2813         }
2814         return 0;
2815
2816 out1:
2817         unregister_filesystem(&shmem_fs_type);
2818 out2:
2819         shmem_destroy_inodecache();
2820 out3:
2821         bdi_destroy(&shmem_backing_dev_info);
2822 out4:
2823         shm_mnt = ERR_PTR(error);
2824         return error;
2825 }
2826
2827 #else /* !CONFIG_SHMEM */
2828
2829 /*
2830  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2831  *
2832  * This is intended for small system where the benefits of the full
2833  * shmem code (swap-backed and resource-limited) are outweighed by
2834  * their complexity. On systems without swap this code should be
2835  * effectively equivalent, but much lighter weight.
2836  */
2837
2838 #include <linux/ramfs.h>
2839
2840 static struct file_system_type shmem_fs_type = {
2841         .name           = "tmpfs",
2842         .mount          = ramfs_mount,
2843         .kill_sb        = kill_litter_super,
2844 };
2845
2846 int __init shmem_init(void)
2847 {
2848         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
2849
2850         shm_mnt = kern_mount(&shmem_fs_type);
2851         BUG_ON(IS_ERR(shm_mnt));
2852
2853         return 0;
2854 }
2855
2856 int shmem_unuse(swp_entry_t swap, struct page *page)
2857 {
2858         return 0;
2859 }
2860
2861 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2862 {
2863         return 0;
2864 }
2865
2866 void shmem_unlock_mapping(struct address_space *mapping)
2867 {
2868 }
2869
2870 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
2871 {
2872         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
2873 }
2874 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2875
2876 #define shmem_vm_ops                            generic_file_vm_ops
2877 #define shmem_file_operations                   ramfs_file_operations
2878 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2879 #define shmem_acct_size(flags, size)            0
2880 #define shmem_unacct_size(flags, size)          do {} while (0)
2881
2882 #endif /* CONFIG_SHMEM */
2883
2884 /* common code */
2885
2886 /**
2887  * shmem_file_setup - get an unlinked file living in tmpfs
2888  * @name: name for dentry (to be seen in /proc/<pid>/maps
2889  * @size: size to be set for the file
2890  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2891  */
2892 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2893 {
2894         int error;
2895         struct file *file;
2896         struct inode *inode;
2897         struct path path;
2898         struct dentry *root;
2899         struct qstr this;
2900
2901         if (IS_ERR(shm_mnt))
2902                 return (void *)shm_mnt;
2903
2904         if (size < 0 || size > MAX_LFS_FILESIZE)
2905                 return ERR_PTR(-EINVAL);
2906
2907         if (shmem_acct_size(flags, size))
2908                 return ERR_PTR(-ENOMEM);
2909
2910         error = -ENOMEM;
2911         this.name = name;
2912         this.len = strlen(name);
2913         this.hash = 0; /* will go */
2914         root = shm_mnt->mnt_root;
2915         path.dentry = d_alloc(root, &this);
2916         if (!path.dentry)
2917                 goto put_memory;
2918         path.mnt = mntget(shm_mnt);
2919
2920         error = -ENOSPC;
2921         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2922         if (!inode)
2923                 goto put_dentry;
2924
2925         d_instantiate(path.dentry, inode);
2926         inode->i_size = size;
2927         clear_nlink(inode);     /* It is unlinked */
2928 #ifndef CONFIG_MMU
2929         error = ramfs_nommu_expand_for_mapping(inode, size);
2930         if (error)
2931                 goto put_dentry;
2932 #endif
2933
2934         error = -ENFILE;
2935         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2936                   &shmem_file_operations);
2937         if (!file)
2938                 goto put_dentry;
2939
2940         return file;
2941
2942 put_dentry:
2943         path_put(&path);
2944 put_memory:
2945         shmem_unacct_size(flags, size);
2946         return ERR_PTR(error);
2947 }
2948 EXPORT_SYMBOL_GPL(shmem_file_setup);
2949
2950 /**
2951  * shmem_zero_setup - setup a shared anonymous mapping
2952  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2953  */
2954 int shmem_zero_setup(struct vm_area_struct *vma)
2955 {
2956         struct file *file;
2957         loff_t size = vma->vm_end - vma->vm_start;
2958
2959         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2960         if (IS_ERR(file))
2961                 return PTR_ERR(file);
2962
2963         if (vma->vm_file)
2964                 fput(vma->vm_file);
2965         vma->vm_file = file;
2966         vma->vm_ops = &shmem_vm_ops;
2967         vma->vm_flags |= VM_CAN_NONLINEAR;
2968         return 0;
2969 }
2970
2971 /**
2972  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
2973  * @mapping:    the page's address_space
2974  * @index:      the page index
2975  * @gfp:        the page allocator flags to use if allocating
2976  *
2977  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
2978  * with any new page allocations done using the specified allocation flags.
2979  * But read_cache_page_gfp() uses the ->readpage() method: which does not
2980  * suit tmpfs, since it may have pages in swapcache, and needs to find those
2981  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
2982  *
2983  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
2984  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
2985  */
2986 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
2987                                          pgoff_t index, gfp_t gfp)
2988 {
2989 #ifdef CONFIG_SHMEM
2990         struct inode *inode = mapping->host;
2991         struct page *page;
2992         int error;
2993
2994         BUG_ON(mapping->a_ops != &shmem_aops);
2995         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
2996         if (error)
2997                 page = ERR_PTR(error);
2998         else
2999                 unlock_page(page);
3000         return page;
3001 #else
3002         /*
3003          * The tiny !SHMEM case uses ramfs without swap
3004          */
3005         return read_cache_page_gfp(mapping, index, gfp);
3006 #endif
3007 }
3008 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);