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/ramfs.h>
  29 #include <linux/pagemap.h>
  30 #include <linux/file.h>
  31 #include <linux/mm.h>
  32 #include <linux/export.h>
  33 #include <linux/swap.h>
  34
  35 static struct vfsmount *shm_mnt;
  36
  37 #ifdef CONFIG_SHMEM
  38 /*
  39  * This virtual memory filesystem is heavily based on the ramfs. It
  40  * extends ramfs by the ability to use swap and honor resource limits
  41  * which makes it a completely usable filesystem.
  42  */
  43
  44 #include <linux/xattr.h>
  45 #include <linux/exportfs.h>
  46 #include <linux/posix_acl.h>
  47 #include <linux/generic_acl.h>
  48 #include <linux/mman.h>
  49 #include <linux/string.h>
  50 #include <linux/slab.h>
  51 #include <linux/backing-dev.h>
  52 #include <linux/shmem_fs.h>
  53 #include <linux/writeback.h>
  54 #include <linux/blkdev.h>
  55 #include <linux/pagevec.h>
  56 #include <linux/percpu_counter.h>
  57 #include <linux/falloc.h>
  58 #include <linux/splice.h>
  59 #include <linux/security.h>
  60 #include <linux/swapops.h>
  61 #include <linux/mempolicy.h>
  62 #include <linux/namei.h>
  63 #include <linux/ctype.h>
  64 #include <linux/migrate.h>
  65 #include <linux/highmem.h>
  66 #include <linux/seq_file.h>
  67 #include <linux/magic.h>
  68
  69 #include <asm/uaccess.h>
  70 #include <asm/pgtable.h>
  71
  72 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  73 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  74
  75 /* Pretend that each entry is of this size in directory's i_size */
  76 #define BOGO_DIRENT_SIZE 20
  77
  78 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  79 #define SHORT_SYMLINK_LEN 128
  80
  81 /*
  82  * shmem_fallocate and shmem_writepage communicate via inode->i_private
  83  * (with i_mutex making sure that it has only one user at a time):
  84  * we would prefer not to enlarge the shmem inode just for that.
  85  */
  86 struct shmem_falloc {
  87         pgoff_t start;          /* start of range currently being fallocated */
  88         pgoff_t next;           /* the next page offset to be fallocated */
  89         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
  90         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
  91 };
  92
  93 /* Flag allocation requirements to shmem_getpage */
  94 enum sgp_type {
  95         SGP_READ,       /* don't exceed i_size, don't allocate page */
  96         SGP_CACHE,      /* don't exceed i_size, may allocate page */
  97         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
  98         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
  99         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
 100 };
 101
 102 #ifdef CONFIG_TMPFS
 103 static unsigned long shmem_default_max_blocks(void)
 104 {
 105         return totalram_pages / 2;
 106 }
 107
 108 static unsigned long shmem_default_max_inodes(void)
 109 {
 110         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 111 }
 112 #endif
 113
 114 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
 115 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 116                                 struct shmem_inode_info *info, pgoff_t index);
 117 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 118         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
 119
 120 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
 121         struct page **pagep, enum sgp_type sgp, int *fault_type)
 122 {
 123         return shmem_getpage_gfp(inode, index, pagep, sgp,
 124                         mapping_gfp_mask(inode->i_mapping), fault_type);
 125 }
 126
 127 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 128 {
 129         return sb->s_fs_info;
 130 }
 131
 132 /*
 133  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 134  * for shared memory and for shared anonymous (/dev/zero) mappings
 135  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 136  * consistent with the pre-accounting of private mappings ...
 137  */
 138 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 139 {
 140         return (flags & VM_NORESERVE) ?
 141                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 142 }
 143
 144 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 145 {
 146         if (!(flags & VM_NORESERVE))
 147                 vm_unacct_memory(VM_ACCT(size));
 148 }
 149
 150 /*
 151  * ... whereas tmpfs objects are accounted incrementally as
 152  * pages are allocated, in order to allow huge sparse files.
 153  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 154  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 155  */
 156 static inline int shmem_acct_block(unsigned long flags)
 157 {
 158         return (flags & VM_NORESERVE) ?
 159                 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 160 }
 161
 162 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 163 {
 164         if (flags & VM_NORESERVE)
 165                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 166 }
 167
 168 static const struct super_operations shmem_ops;
 169 static const struct address_space_operations shmem_aops;
 170 static const struct file_operations shmem_file_operations;
 171 static const struct inode_operations shmem_inode_operations;
 172 static const struct inode_operations shmem_dir_inode_operations;
 173 static const struct inode_operations shmem_special_inode_operations;
 174 static const struct vm_operations_struct shmem_vm_ops;
 175
 176 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 177         .ra_pages       = 0,    /* No readahead */
 178         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
 179 };
 180
 181 static LIST_HEAD(shmem_swaplist);
 182 static DEFINE_MUTEX(shmem_swaplist_mutex);
 183
 184 static int shmem_reserve_inode(struct super_block *sb)
 185 {
 186         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 187         if (sbinfo->max_inodes) {
 188                 spin_lock(&sbinfo->stat_lock);
 189                 if (!sbinfo->free_inodes) {
 190                         spin_unlock(&sbinfo->stat_lock);
 191                         return -ENOSPC;
 192                 }
 193                 sbinfo->free_inodes--;
 194                 spin_unlock(&sbinfo->stat_lock);
 195         }
 196         return 0;
 197 }
 198
 199 static void shmem_free_inode(struct super_block *sb)
 200 {
 201         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 202         if (sbinfo->max_inodes) {
 203                 spin_lock(&sbinfo->stat_lock);
 204                 sbinfo->free_inodes++;
 205                 spin_unlock(&sbinfo->stat_lock);
 206         }
 207 }
 208
 209 /**
 210  * shmem_recalc_inode - recalculate the block usage of an inode
 211  * @inode: inode to recalc
 212  *
 213  * We have to calculate the free blocks since the mm can drop
 214  * undirtied hole pages behind our back.
 215  *
 216  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 217  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 218  *
 219  * It has to be called with the spinlock held.
 220  */
 221 static void shmem_recalc_inode(struct inode *inode)
 222 {
 223         struct shmem_inode_info *info = SHMEM_I(inode);
 224         long freed;
 225
 226         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 227         if (freed > 0) {
 228                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 229                 if (sbinfo->max_blocks)
 230                         percpu_counter_add(&sbinfo->used_blocks, -freed);
 231                 info->alloced -= freed;
 232                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 233                 shmem_unacct_blocks(info->flags, freed);
 234         }
 235 }
 236
 237 /*
 238  * Replace item expected in radix tree by a new item, while holding tree lock.
 239  */
 240 static int shmem_radix_tree_replace(struct address_space *mapping,
 241                         pgoff_t index, void *expected, void *replacement)
 242 {
 243         void **pslot;
 244         void *item = NULL;
 245
 246         VM_BUG_ON(!expected);
 247         pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
 248         if (pslot)
 249                 item = radix_tree_deref_slot_protected(pslot,
 250                                                         &mapping->tree_lock);
 251         if (item != expected)
 252                 return -ENOENT;
 253         if (replacement)
 254                 radix_tree_replace_slot(pslot, replacement);
 255         else
 256                 radix_tree_delete(&mapping->page_tree, index);
 257         return 0;
 258 }
 259
 260 /*
 261  * Sometimes, before we decide whether to proceed or to fail, we must check
 262  * that an entry was not already brought back from swap by a racing thread.
 263  *
 264  * Checking page is not enough: by the time a SwapCache page is locked, it
 265  * might be reused, and again be SwapCache, using the same swap as before.
 266  */
 267 static bool shmem_confirm_swap(struct address_space *mapping,
 268                                pgoff_t index, swp_entry_t swap)
 269 {
 270         void *item;
 271
 272         rcu_read_lock();
 273         item = radix_tree_lookup(&mapping->page_tree, index);
 274         rcu_read_unlock();
 275         return item == swp_to_radix_entry(swap);
 276 }
 277
 278 /*
 279  * Like add_to_page_cache_locked, but error if expected item has gone.
 280  */
 281 static int shmem_add_to_page_cache(struct page *page,
 282                                    struct address_space *mapping,
 283                                    pgoff_t index, gfp_t gfp, void *expected)
 284 {
 285         int error;
 286
 287         VM_BUG_ON(!PageLocked(page));
 288         VM_BUG_ON(!PageSwapBacked(page));
 289
 290         page_cache_get(page);
 291         page->mapping = mapping;
 292         page->index = index;
 293
 294         spin_lock_irq(&mapping->tree_lock);
 295         if (!expected)
 296                 error = radix_tree_insert(&mapping->page_tree, index, page);
 297         else
 298                 error = shmem_radix_tree_replace(mapping, index, expected,
 299                                                                  page);
 300         if (!error) {
 301                 mapping->nrpages++;
 302                 __inc_zone_page_state(page, NR_FILE_PAGES);
 303                 __inc_zone_page_state(page, NR_SHMEM);
 304                 spin_unlock_irq(&mapping->tree_lock);
 305         } else {
 306                 page->mapping = NULL;
 307                 spin_unlock_irq(&mapping->tree_lock);
 308                 page_cache_release(page);
 309         }
 310         return error;
 311 }
 312
 313 /*
 314  * Like delete_from_page_cache, but substitutes swap for page.
 315  */
 316 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 317 {
 318         struct address_space *mapping = page->mapping;
 319         int error;
 320
 321         spin_lock_irq(&mapping->tree_lock);
 322         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
 323         page->mapping = NULL;
 324         mapping->nrpages--;
 325         __dec_zone_page_state(page, NR_FILE_PAGES);
 326         __dec_zone_page_state(page, NR_SHMEM);
 327         spin_unlock_irq(&mapping->tree_lock);
 328         page_cache_release(page);
 329         BUG_ON(error);
 330 }
 331
 332 /*
 333  * Like find_get_pages, but collecting swap entries as well as pages.
 334  */
 335 static unsigned shmem_find_get_pages_and_swap(struct address_space *mapping,
 336                                         pgoff_t start, unsigned int nr_pages,
 337                                         struct page **pages, pgoff_t *indices)
 338 {
 339         void **slot;
 340         unsigned int ret = 0;
 341         struct radix_tree_iter iter;
 342
 343         if (!nr_pages)
 344                 return 0;
 345
 346         rcu_read_lock();
 347 restart:
 348         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
 349                 struct page *page;
 350 repeat:
 351                 page = radix_tree_deref_slot(slot);
 352                 if (unlikely(!page))
 353                         continue;
 354                 if (radix_tree_exception(page)) {
 355                         if (radix_tree_deref_retry(page))
 356                                 goto restart;
 357                         /*
 358                          * Otherwise, we must be storing a swap entry
 359                          * here as an exceptional entry: so return it
 360                          * without attempting to raise page count.
 361                          */
 362                         goto export;
 363                 }
 364                 if (!page_cache_get_speculative(page))
 365                         goto repeat;
 366
 367                 /* Has the page moved? */
 368                 if (unlikely(page != *slot)) {
 369                         page_cache_release(page);
 370                         goto repeat;
 371                 }
 372 export:
 373                 indices[ret] = iter.index;
 374                 pages[ret] = page;
 375                 if (++ret == nr_pages)
 376                         break;
 377         }
 378         rcu_read_unlock();
 379         return ret;
 380 }
 381
 382 /*
 383  * Remove swap entry from radix tree, free the swap and its page cache.
 384  */
 385 static int shmem_free_swap(struct address_space *mapping,
 386                            pgoff_t index, void *radswap)
 387 {
 388         int error;
 389
 390         spin_lock_irq(&mapping->tree_lock);
 391         error = shmem_radix_tree_replace(mapping, index, radswap, NULL);
 392         spin_unlock_irq(&mapping->tree_lock);
 393         if (!error)
 394                 free_swap_and_cache(radix_to_swp_entry(radswap));
 395         return error;
 396 }
 397
 398 /*
 399  * Pagevec may contain swap entries, so shuffle up pages before releasing.
 400  */
 401 static void shmem_deswap_pagevec(struct pagevec *pvec)
 402 {
 403         int i, j;
 404
 405         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
 406                 struct page *page = pvec->pages[i];
 407                 if (!radix_tree_exceptional_entry(page))
 408                         pvec->pages[j++] = page;
 409         }
 410         pvec->nr = j;
 411 }
 412
 413 /*
 414  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 415  */
 416 void shmem_unlock_mapping(struct address_space *mapping)
 417 {
 418         struct pagevec pvec;
 419         pgoff_t indices[PAGEVEC_SIZE];
 420         pgoff_t index = 0;
 421
 422         pagevec_init(&pvec, 0);
 423         /*
 424          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 425          */
 426         while (!mapping_unevictable(mapping)) {
 427                 /*
 428                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
 429                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
 430                  */
 431                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 432                                         PAGEVEC_SIZE, pvec.pages, indices);
 433                 if (!pvec.nr)
 434                         break;
 435                 index = indices[pvec.nr - 1] + 1;
 436                 shmem_deswap_pagevec(&pvec);
 437                 check_move_unevictable_pages(pvec.pages, pvec.nr);
 438                 pagevec_release(&pvec);
 439                 cond_resched();
 440         }
 441 }
 442
 443 /*
 444  * Remove range of pages and swap entries from radix tree, and free them.
 445  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 446  */
 447 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 448                                                                  bool unfalloc)
 449 {
 450         struct address_space *mapping = inode->i_mapping;
 451         struct shmem_inode_info *info = SHMEM_I(inode);
 452         pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 453         pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
 454         unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
 455         unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
 456         struct pagevec pvec;
 457         pgoff_t indices[PAGEVEC_SIZE];
 458         long nr_swaps_freed = 0;
 459         pgoff_t index;
 460         int i;
 461
 462         if (lend == -1)
 463                 end = -1;       /* unsigned, so actually very big */
 464
 465         pagevec_init(&pvec, 0);
 466         index = start;
 467         while (index < end) {
 468                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 469                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 470                                                         pvec.pages, indices);
 471                 if (!pvec.nr)
 472                         break;
 473                 mem_cgroup_uncharge_start();
 474                 for (i = 0; i < pagevec_count(&pvec); i++) {
 475                         struct page *page = pvec.pages[i];
 476
 477                         index = indices[i];
 478                         if (index >= end)
 479                                 break;
 480
 481                         if (radix_tree_exceptional_entry(page)) {
 482                                 if (unfalloc)
 483                                         continue;
 484                                 nr_swaps_freed += !shmem_free_swap(mapping,
 485                                                                 index, page);
 486                                 continue;
 487                         }
 488
 489                         if (!trylock_page(page))
 490                                 continue;
 491                         if (!unfalloc || !PageUptodate(page)) {
 492                                 if (page->mapping == mapping) {
 493                                         VM_BUG_ON(PageWriteback(page));
 494                                         truncate_inode_page(mapping, page);
 495                                 }
 496                         }
 497                         unlock_page(page);
 498                 }
 499                 shmem_deswap_pagevec(&pvec);
 500                 pagevec_release(&pvec);
 501                 mem_cgroup_uncharge_end();
 502                 cond_resched();
 503                 index++;
 504         }
 505
 506         if (partial_start) {
 507                 struct page *page = NULL;
 508                 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
 509                 if (page) {
 510                         unsigned int top = PAGE_CACHE_SIZE;
 511                         if (start > end) {
 512                                 top = partial_end;
 513                                 partial_end = 0;
 514                         }
 515                         zero_user_segment(page, partial_start, top);
 516                         set_page_dirty(page);
 517                         unlock_page(page);
 518                         page_cache_release(page);
 519                 }
 520         }
 521         if (partial_end) {
 522                 struct page *page = NULL;
 523                 shmem_getpage(inode, end, &page, SGP_READ, NULL);
 524                 if (page) {
 525                         zero_user_segment(page, 0, partial_end);
 526                         set_page_dirty(page);
 527                         unlock_page(page);
 528                         page_cache_release(page);
 529                 }
 530         }
 531         if (start >= end)
 532                 return;
 533
 534         index = start;
 535         for ( ; ; ) {
 536                 cond_resched();
 537                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 538                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 539                                                         pvec.pages, indices);
 540                 if (!pvec.nr) {
 541                         if (index == start || unfalloc)
 542                                 break;
 543                         index = start;
 544                         continue;
 545                 }
 546                 if ((index == start || unfalloc) && indices[0] >= end) {
 547                         shmem_deswap_pagevec(&pvec);
 548                         pagevec_release(&pvec);
 549                         break;
 550                 }
 551                 mem_cgroup_uncharge_start();
 552                 for (i = 0; i < pagevec_count(&pvec); i++) {
 553                         struct page *page = pvec.pages[i];
 554
 555                         index = indices[i];
 556                         if (index >= end)
 557                                 break;
 558
 559                         if (radix_tree_exceptional_entry(page)) {
 560                                 if (unfalloc)
 561                                         continue;
 562                                 nr_swaps_freed += !shmem_free_swap(mapping,
 563                                                                 index, page);
 564                                 continue;
 565                         }
 566
 567                         lock_page(page);
 568                         if (!unfalloc || !PageUptodate(page)) {
 569                                 if (page->mapping == mapping) {
 570                                         VM_BUG_ON(PageWriteback(page));
 571                                         truncate_inode_page(mapping, page);
 572                                 }
 573                         }
 574                         unlock_page(page);
 575                 }
 576                 shmem_deswap_pagevec(&pvec);
 577                 pagevec_release(&pvec);
 578                 mem_cgroup_uncharge_end();
 579                 index++;
 580         }
 581
 582         spin_lock(&info->lock);
 583         info->swapped -= nr_swaps_freed;
 584         shmem_recalc_inode(inode);
 585         spin_unlock(&info->lock);
 586 }
 587
 588 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
 589 {
 590         shmem_undo_range(inode, lstart, lend, false);
 591         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 592 }
 593 EXPORT_SYMBOL_GPL(shmem_truncate_range);
 594
 595 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
 596 {
 597         struct inode *inode = dentry->d_inode;
 598         int error;
 599
 600         error = inode_change_ok(inode, attr);
 601         if (error)
 602                 return error;
 603
 604         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
 605                 loff_t oldsize = inode->i_size;
 606                 loff_t newsize = attr->ia_size;
 607
 608                 if (newsize != oldsize) {
 609                         i_size_write(inode, newsize);
 610                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 611                 }
 612                 if (newsize < oldsize) {
 613                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
 614                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 615                         shmem_truncate_range(inode, newsize, (loff_t)-1);
 616                         /* unmap again to remove racily COWed private pages */
 617                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 618                 }
 619         }
 620
 621         setattr_copy(inode, attr);
 622 #ifdef CONFIG_TMPFS_POSIX_ACL
 623         if (attr->ia_valid & ATTR_MODE)
 624                 error = generic_acl_chmod(inode);
 625 #endif
 626         return error;
 627 }
 628
 629 static void shmem_evict_inode(struct inode *inode)
 630 {
 631         struct shmem_inode_info *info = SHMEM_I(inode);
 632
 633         if (inode->i_mapping->a_ops == &shmem_aops) {
 634                 shmem_unacct_size(info->flags, inode->i_size);
 635                 inode->i_size = 0;
 636                 shmem_truncate_range(inode, 0, (loff_t)-1);
 637                 if (!list_empty(&info->swaplist)) {
 638                         mutex_lock(&shmem_swaplist_mutex);
 639                         list_del_init(&info->swaplist);
 640                         mutex_unlock(&shmem_swaplist_mutex);
 641                 }
 642         } else
 643                 kfree(info->symlink);
 644
 645         simple_xattrs_free(&info->xattrs);
 646         WARN_ON(inode->i_blocks);
 647         shmem_free_inode(inode->i_sb);
 648         clear_inode(inode);
 649 }
 650
 651 /*
 652  * If swap found in inode, free it and move page from swapcache to filecache.
 653  */
 654 static int shmem_unuse_inode(struct shmem_inode_info *info,
 655                              swp_entry_t swap, struct page **pagep)
 656 {
 657         struct address_space *mapping = info->vfs_inode.i_mapping;
 658         void *radswap;
 659         pgoff_t index;
 660         gfp_t gfp;
 661         int error = 0;
 662
 663         radswap = swp_to_radix_entry(swap);
 664         index = radix_tree_locate_item(&mapping->page_tree, radswap);
 665         if (index == -1)
 666                 return 0;
 667
 668         /*
 669          * Move _head_ to start search for next from here.
 670          * But be careful: shmem_evict_inode checks list_empty without taking
 671          * mutex, and there's an instant in list_move_tail when info->swaplist
 672          * would appear empty, if it were the only one on shmem_swaplist.
 673          */
 674         if (shmem_swaplist.next != &info->swaplist)
 675                 list_move_tail(&shmem_swaplist, &info->swaplist);
 676
 677         gfp = mapping_gfp_mask(mapping);
 678         if (shmem_should_replace_page(*pagep, gfp)) {
 679                 mutex_unlock(&shmem_swaplist_mutex);
 680                 error = shmem_replace_page(pagep, gfp, info, index);
 681                 mutex_lock(&shmem_swaplist_mutex);
 682                 /*
 683                  * We needed to drop mutex to make that restrictive page
 684                  * allocation, but the inode might have been freed while we
 685                  * dropped it: although a racing shmem_evict_inode() cannot
 686                  * complete without emptying the radix_tree, our page lock
 687                  * on this swapcache page is not enough to prevent that -
 688                  * free_swap_and_cache() of our swap entry will only
 689                  * trylock_page(), removing swap from radix_tree whatever.
 690                  *
 691                  * We must not proceed to shmem_add_to_page_cache() if the
 692                  * inode has been freed, but of course we cannot rely on
 693                  * inode or mapping or info to check that.  However, we can
 694                  * safely check if our swap entry is still in use (and here
 695                  * it can't have got reused for another page): if it's still
 696                  * in use, then the inode cannot have been freed yet, and we
 697                  * can safely proceed (if it's no longer in use, that tells
 698                  * nothing about the inode, but we don't need to unuse swap).
 699                  */
 700                 if (!page_swapcount(*pagep))
 701                         error = -ENOENT;
 702         }
 703
 704         /*
 705          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 706          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 707          * beneath us (pagelock doesn't help until the page is in pagecache).
 708          */
 709         if (!error)
 710                 error = shmem_add_to_page_cache(*pagep, mapping, index,
 711                                                 GFP_NOWAIT, radswap);
 712         if (error != -ENOMEM) {
 713                 /*
 714                  * Truncation and eviction use free_swap_and_cache(), which
 715                  * only does trylock page: if we raced, best clean up here.
 716                  */
 717                 delete_from_swap_cache(*pagep);
 718                 set_page_dirty(*pagep);
 719                 if (!error) {
 720                         spin_lock(&info->lock);
 721                         info->swapped--;
 722                         spin_unlock(&info->lock);
 723                         swap_free(swap);
 724                 }
 725                 error = 1;      /* not an error, but entry was found */
 726         }
 727         return error;
 728 }
 729
 730 /*
 731  * Search through swapped inodes to find and replace swap by page.
 732  */
 733 int shmem_unuse(swp_entry_t swap, struct page *page)
 734 {
 735         struct list_head *this, *next;
 736         struct shmem_inode_info *info;
 737         int found = 0;
 738         int error = 0;
 739
 740         /*
 741          * There's a faint possibility that swap page was replaced before
 742          * caller locked it: caller will come back later with the right page.
 743          */
 744         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
 745                 goto out;
 746
 747         /*
 748          * Charge page using GFP_KERNEL while we can wait, before taking
 749          * the shmem_swaplist_mutex which might hold up shmem_writepage().
 750          * Charged back to the user (not to caller) when swap account is used.
 751          */
 752         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
 753         if (error)
 754                 goto out;
 755         /* No radix_tree_preload: swap entry keeps a place for page in tree */
 756
 757         mutex_lock(&shmem_swaplist_mutex);
 758         list_for_each_safe(this, next, &shmem_swaplist) {
 759                 info = list_entry(this, struct shmem_inode_info, swaplist);
 760                 if (info->swapped)
 761                         found = shmem_unuse_inode(info, swap, &page);
 762                 else
 763                         list_del_init(&info->swaplist);
 764                 cond_resched();
 765                 if (found)
 766                         break;
 767         }
 768         mutex_unlock(&shmem_swaplist_mutex);
 769
 770         if (found < 0)
 771                 error = found;
 772 out:
 773         unlock_page(page);
 774         page_cache_release(page);
 775         return error;
 776 }
 777
 778 /*
 779  * Move the page from the page cache to the swap cache.
 780  */
 781 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 782 {
 783         struct shmem_inode_info *info;
 784         struct address_space *mapping;
 785         struct inode *inode;
 786         swp_entry_t swap;
 787         pgoff_t index;
 788
 789         BUG_ON(!PageLocked(page));
 790         mapping = page->mapping;
 791         index = page->index;
 792         inode = mapping->host;
 793         info = SHMEM_I(inode);
 794         if (info->flags & VM_LOCKED)
 795                 goto redirty;
 796         if (!total_swap_pages)
 797                 goto redirty;
 798
 799         /*
 800          * shmem_backing_dev_info's capabilities prevent regular writeback or
 801          * sync from ever calling shmem_writepage; but a stacking filesystem
 802          * might use ->writepage of its underlying filesystem, in which case
 803          * tmpfs should write out to swap only in response to memory pressure,
 804          * and not for the writeback threads or sync.
 805          */
 806         if (!wbc->for_reclaim) {
 807                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
 808                 goto redirty;
 809         }
 810
 811         /*
 812          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
 813          * value into swapfile.c, the only way we can correctly account for a
 814          * fallocated page arriving here is now to initialize it and write it.
 815          *
 816          * That's okay for a page already fallocated earlier, but if we have
 817          * not yet completed the fallocation, then (a) we want to keep track
 818          * of this page in case we have to undo it, and (b) it may not be a
 819          * good idea to continue anyway, once we're pushing into swap.  So
 820          * reactivate the page, and let shmem_fallocate() quit when too many.
 821          */
 822         if (!PageUptodate(page)) {
 823                 if (inode->i_private) {
 824                         struct shmem_falloc *shmem_falloc;
 825                         spin_lock(&inode->i_lock);
 826                         shmem_falloc = inode->i_private;
 827                         if (shmem_falloc &&
 828                             index >= shmem_falloc->start &&
 829                             index < shmem_falloc->next)
 830                                 shmem_falloc->nr_unswapped++;
 831                         else
 832                                 shmem_falloc = NULL;
 833                         spin_unlock(&inode->i_lock);
 834                         if (shmem_falloc)
 835                                 goto redirty;
 836                 }
 837                 clear_highpage(page);
 838                 flush_dcache_page(page);
 839                 SetPageUptodate(page);
 840         }
 841
 842         swap = get_swap_page();
 843         if (!swap.val)
 844                 goto redirty;
 845
 846         /*
 847          * Add inode to shmem_unuse()'s list of swapped-out inodes,
 848          * if it's not already there.  Do it now before the page is
 849          * moved to swap cache, when its pagelock no longer protects
 850          * the inode from eviction.  But don't unlock the mutex until
 851          * we've incremented swapped, because shmem_unuse_inode() will
 852          * prune a !swapped inode from the swaplist under this mutex.
 853          */
 854         mutex_lock(&shmem_swaplist_mutex);
 855         if (list_empty(&info->swaplist))
 856                 list_add_tail(&info->swaplist, &shmem_swaplist);
 857
 858         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
 859                 swap_shmem_alloc(swap);
 860                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
 861
 862                 spin_lock(&info->lock);
 863                 info->swapped++;
 864                 shmem_recalc_inode(inode);
 865                 spin_unlock(&info->lock);
 866
 867                 mutex_unlock(&shmem_swaplist_mutex);
 868                 BUG_ON(page_mapped(page));
 869                 swap_writepage(page, wbc);
 870                 return 0;
 871         }
 872
 873         mutex_unlock(&shmem_swaplist_mutex);
 874         swapcache_free(swap, NULL);
 875 redirty:
 876         set_page_dirty(page);
 877         if (wbc->for_reclaim)
 878                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
 879         unlock_page(page);
 880         return 0;
 881 }
 882
 883 #ifdef CONFIG_NUMA
 884 #ifdef CONFIG_TMPFS
 885 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 886 {
 887         char buffer[64];
 888
 889         if (!mpol || mpol->mode == MPOL_DEFAULT)
 890                 return;         /* show nothing */
 891
 892         mpol_to_str(buffer, sizeof(buffer), mpol);
 893
 894         seq_printf(seq, ",mpol=%s", buffer);
 895 }
 896
 897 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 898 {
 899         struct mempolicy *mpol = NULL;
 900         if (sbinfo->mpol) {
 901                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
 902                 mpol = sbinfo->mpol;
 903                 mpol_get(mpol);
 904                 spin_unlock(&sbinfo->stat_lock);
 905         }
 906         return mpol;
 907 }
 908 #endif /* CONFIG_TMPFS */
 909
 910 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 911                         struct shmem_inode_info *info, pgoff_t index)
 912 {
 913         struct vm_area_struct pvma;
 914         struct page *page;
 915
 916         /* Create a pseudo vma that just contains the policy */
 917         pvma.vm_start = 0;
 918         /* Bias interleave by inode number to distribute better across nodes */
 919         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 920         pvma.vm_ops = NULL;
 921         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 922
 923         page = swapin_readahead(swap, gfp, &pvma, 0);
 924
 925         /* Drop reference taken by mpol_shared_policy_lookup() */
 926         mpol_cond_put(pvma.vm_policy);
 927
 928         return page;
 929 }
 930
 931 static struct page *shmem_alloc_page(gfp_t gfp,
 932                         struct shmem_inode_info *info, pgoff_t index)
 933 {
 934         struct vm_area_struct pvma;
 935         struct page *page;
 936
 937         /* Create a pseudo vma that just contains the policy */
 938         pvma.vm_start = 0;
 939         /* Bias interleave by inode number to distribute better across nodes */
 940         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 941         pvma.vm_ops = NULL;
 942         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 943
 944         page = alloc_page_vma(gfp, &pvma, 0);
 945
 946         /* Drop reference taken by mpol_shared_policy_lookup() */
 947         mpol_cond_put(pvma.vm_policy);
 948
 949         return page;
 950 }
 951 #else /* !CONFIG_NUMA */
 952 #ifdef CONFIG_TMPFS
 953 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 954 {
 955 }
 956 #endif /* CONFIG_TMPFS */
 957
 958 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 959                         struct shmem_inode_info *info, pgoff_t index)
 960 {
 961         return swapin_readahead(swap, gfp, NULL, 0);
 962 }
 963
 964 static inline struct page *shmem_alloc_page(gfp_t gfp,
 965                         struct shmem_inode_info *info, pgoff_t index)
 966 {
 967         return alloc_page(gfp);
 968 }
 969 #endif /* CONFIG_NUMA */
 970
 971 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
 972 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 973 {
 974         return NULL;
 975 }
 976 #endif
 977
 978 /*
 979  * When a page is moved from swapcache to shmem filecache (either by the
 980  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
 981  * shmem_unuse_inode()), it may have been read in earlier from swap, in
 982  * ignorance of the mapping it belongs to.  If that mapping has special
 983  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
 984  * we may need to copy to a suitable page before moving to filecache.
 985  *
 986  * In a future release, this may well be extended to respect cpuset and
 987  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
 988  * but for now it is a simple matter of zone.
 989  */
 990 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
 991 {
 992         return page_zonenum(page) > gfp_zone(gfp);
 993 }
 994
 995 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 996                                 struct shmem_inode_info *info, pgoff_t index)
 997 {
 998         struct page *oldpage, *newpage;
 999         struct address_space *swap_mapping;
1000         pgoff_t swap_index;
1001         int error;
1002
1003         oldpage = *pagep;
1004         swap_index = page_private(oldpage);
1005         swap_mapping = page_mapping(oldpage);
1006
1007         /*
1008          * We have arrived here because our zones are constrained, so don't
1009          * limit chance of success by further cpuset and node constraints.
1010          */
1011         gfp &= ~GFP_CONSTRAINT_MASK;
1012         newpage = shmem_alloc_page(gfp, info, index);
1013         if (!newpage)
1014                 return -ENOMEM;
1015
1016         page_cache_get(newpage);
1017         copy_highpage(newpage, oldpage);
1018         flush_dcache_page(newpage);
1019
1020         __set_page_locked(newpage);
1021         SetPageUptodate(newpage);
1022         SetPageSwapBacked(newpage);
1023         set_page_private(newpage, swap_index);
1024         SetPageSwapCache(newpage);
1025
1026         /*
1027          * Our caller will very soon move newpage out of swapcache, but it's
1028          * a nice clean interface for us to replace oldpage by newpage there.
1029          */
1030         spin_lock_irq(&swap_mapping->tree_lock);
1031         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1032                                                                    newpage);
1033         if (!error) {
1034                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1035                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1036         }
1037         spin_unlock_irq(&swap_mapping->tree_lock);
1038
1039         if (unlikely(error)) {
1040                 /*
1041                  * Is this possible?  I think not, now that our callers check
1042                  * both PageSwapCache and page_private after getting page lock;
1043                  * but be defensive.  Reverse old to newpage for clear and free.
1044                  */
1045                 oldpage = newpage;
1046         } else {
1047                 mem_cgroup_replace_page_cache(oldpage, newpage);
1048                 lru_cache_add_anon(newpage);
1049                 *pagep = newpage;
1050         }
1051
1052         ClearPageSwapCache(oldpage);
1053         set_page_private(oldpage, 0);
1054
1055         unlock_page(oldpage);
1056         page_cache_release(oldpage);
1057         page_cache_release(oldpage);
1058         return error;
1059 }
1060
1061 /*
1062  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1063  *
1064  * If we allocate a new one we do not mark it dirty. That's up to the
1065  * vm. If we swap it in we mark it dirty since we also free the swap
1066  * entry since a page cannot live in both the swap and page cache
1067  */
1068 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1069         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1070 {
1071         struct address_space *mapping = inode->i_mapping;
1072         struct shmem_inode_info *info;
1073         struct shmem_sb_info *sbinfo;
1074         struct page *page;
1075         swp_entry_t swap;
1076         int error;
1077         int once = 0;
1078         int alloced = 0;
1079
1080         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1081                 return -EFBIG;
1082 repeat:
1083         swap.val = 0;
1084         page = find_lock_page(mapping, index);
1085         if (radix_tree_exceptional_entry(page)) {
1086                 swap = radix_to_swp_entry(page);
1087                 page = NULL;
1088         }
1089
1090         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1091             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1092                 error = -EINVAL;
1093                 goto failed;
1094         }
1095
1096         /* fallocated page? */
1097         if (page && !PageUptodate(page)) {
1098                 if (sgp != SGP_READ)
1099                         goto clear;
1100                 unlock_page(page);
1101                 page_cache_release(page);
1102                 page = NULL;
1103         }
1104         if (page || (sgp == SGP_READ && !swap.val)) {
1105                 *pagep = page;
1106                 return 0;
1107         }
1108
1109         /*
1110          * Fast cache lookup did not find it:
1111          * bring it back from swap or allocate.
1112          */
1113         info = SHMEM_I(inode);
1114         sbinfo = SHMEM_SB(inode->i_sb);
1115
1116         if (swap.val) {
1117                 /* Look it up and read it in.. */
1118                 page = lookup_swap_cache(swap);
1119                 if (!page) {
1120                         /* here we actually do the io */
1121                         if (fault_type)
1122                                 *fault_type |= VM_FAULT_MAJOR;
1123                         page = shmem_swapin(swap, gfp, info, index);
1124                         if (!page) {
1125                                 error = -ENOMEM;
1126                                 goto failed;
1127                         }
1128                 }
1129
1130                 /* We have to do this with page locked to prevent races */
1131                 lock_page(page);
1132                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1133                     !shmem_confirm_swap(mapping, index, swap)) {
1134                         error = -EEXIST;        /* try again */
1135                         goto unlock;
1136                 }
1137                 if (!PageUptodate(page)) {
1138                         error = -EIO;
1139                         goto failed;
1140                 }
1141                 wait_on_page_writeback(page);
1142
1143                 if (shmem_should_replace_page(page, gfp)) {
1144                         error = shmem_replace_page(&page, gfp, info, index);
1145                         if (error)
1146                                 goto failed;
1147                 }
1148
1149                 error = mem_cgroup_cache_charge(page, current->mm,
1150                                                 gfp & GFP_RECLAIM_MASK);
1151                 if (!error) {
1152                         error = shmem_add_to_page_cache(page, mapping, index,
1153                                                 gfp, swp_to_radix_entry(swap));
1154                         /*
1155                          * We already confirmed swap under page lock, and make
1156                          * no memory allocation here, so usually no possibility
1157                          * of error; but free_swap_and_cache() only trylocks a
1158                          * page, so it is just possible that the entry has been
1159                          * truncated or holepunched since swap was confirmed.
1160                          * shmem_undo_range() will have done some of the
1161                          * unaccounting, now delete_from_swap_cache() will do
1162                          * the rest (including mem_cgroup_uncharge_swapcache).
1163                          * Reset swap.val? No, leave it so "failed" goes back to
1164                          * "repeat": reading a hole and writing should succeed.
1165                          */
1166                         if (error)
1167                                 delete_from_swap_cache(page);
1168                 }
1169                 if (error)
1170                         goto failed;
1171
1172                 spin_lock(&info->lock);
1173                 info->swapped--;
1174                 shmem_recalc_inode(inode);
1175                 spin_unlock(&info->lock);
1176
1177                 delete_from_swap_cache(page);
1178                 set_page_dirty(page);
1179                 swap_free(swap);
1180
1181         } else {
1182                 if (shmem_acct_block(info->flags)) {
1183                         error = -ENOSPC;
1184                         goto failed;
1185                 }
1186                 if (sbinfo->max_blocks) {
1187                         if (percpu_counter_compare(&sbinfo->used_blocks,
1188                                                 sbinfo->max_blocks) >= 0) {
1189                                 error = -ENOSPC;
1190                                 goto unacct;
1191                         }
1192                         percpu_counter_inc(&sbinfo->used_blocks);
1193                 }
1194
1195                 page = shmem_alloc_page(gfp, info, index);
1196                 if (!page) {
1197                         error = -ENOMEM;
1198                         goto decused;
1199                 }
1200
1201                 SetPageSwapBacked(page);
1202                 __set_page_locked(page);
1203                 error = mem_cgroup_cache_charge(page, current->mm,
1204                                                 gfp & GFP_RECLAIM_MASK);
1205                 if (error)
1206                         goto decused;
1207                 error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
1208                 if (!error) {
1209                         error = shmem_add_to_page_cache(page, mapping, index,
1210                                                         gfp, NULL);
1211                         radix_tree_preload_end();
1212                 }
1213                 if (error) {
1214                         mem_cgroup_uncharge_cache_page(page);
1215                         goto decused;
1216                 }
1217                 lru_cache_add_anon(page);
1218
1219                 spin_lock(&info->lock);
1220                 info->alloced++;
1221                 inode->i_blocks += BLOCKS_PER_PAGE;
1222                 shmem_recalc_inode(inode);
1223                 spin_unlock(&info->lock);
1224                 alloced = true;
1225
1226                 /*
1227                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1228                  */
1229                 if (sgp == SGP_FALLOC)
1230                         sgp = SGP_WRITE;
1231 clear:
1232                 /*
1233                  * Let SGP_WRITE caller clear ends if write does not fill page;
1234                  * but SGP_FALLOC on a page fallocated earlier must initialize
1235                  * it now, lest undo on failure cancel our earlier guarantee.
1236                  */
1237                 if (sgp != SGP_WRITE) {
1238                         clear_highpage(page);
1239                         flush_dcache_page(page);
1240                         SetPageUptodate(page);
1241                 }
1242                 if (sgp == SGP_DIRTY)
1243                         set_page_dirty(page);
1244         }
1245
1246         /* Perhaps the file has been truncated since we checked */
1247         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1248             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1249                 error = -EINVAL;
1250                 if (alloced)
1251                         goto trunc;
1252                 else
1253                         goto failed;
1254         }
1255         *pagep = page;
1256         return 0;
1257
1258         /*
1259          * Error recovery.
1260          */
1261 trunc:
1262         info = SHMEM_I(inode);
1263         ClearPageDirty(page);
1264         delete_from_page_cache(page);
1265         spin_lock(&info->lock);
1266         info->alloced--;
1267         inode->i_blocks -= BLOCKS_PER_PAGE;
1268         spin_unlock(&info->lock);
1269 decused:
1270         sbinfo = SHMEM_SB(inode->i_sb);
1271         if (sbinfo->max_blocks)
1272                 percpu_counter_add(&sbinfo->used_blocks, -1);
1273 unacct:
1274         shmem_unacct_blocks(info->flags, 1);
1275 failed:
1276         if (swap.val && error != -EINVAL &&
1277             !shmem_confirm_swap(mapping, index, swap))
1278                 error = -EEXIST;
1279 unlock:
1280         if (page) {
1281                 unlock_page(page);
1282                 page_cache_release(page);
1283         }
1284         if (error == -ENOSPC && !once++) {
1285                 info = SHMEM_I(inode);
1286                 spin_lock(&info->lock);
1287                 shmem_recalc_inode(inode);
1288                 spin_unlock(&info->lock);
1289                 goto repeat;
1290         }
1291         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1292                 goto repeat;
1293         return error;
1294 }
1295
1296 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1297 {
1298         struct inode *inode = file_inode(vma->vm_file);
1299         int error;
1300         int ret = VM_FAULT_LOCKED;
1301
1302         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1303         if (error)
1304                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1305
1306         if (ret & VM_FAULT_MAJOR) {
1307                 count_vm_event(PGMAJFAULT);
1308                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1309         }
1310         return ret;
1311 }
1312
1313 #ifdef CONFIG_NUMA
1314 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1315 {
1316         struct inode *inode = file_inode(vma->vm_file);
1317         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1318 }
1319
1320 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1321                                           unsigned long addr)
1322 {
1323         struct inode *inode = file_inode(vma->vm_file);
1324         pgoff_t index;
1325
1326         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1327         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1328 }
1329 #endif
1330
1331 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1332 {
1333         struct inode *inode = file_inode(file);
1334         struct shmem_inode_info *info = SHMEM_I(inode);
1335         int retval = -ENOMEM;
1336
1337         spin_lock(&info->lock);
1338         if (lock && !(info->flags & VM_LOCKED)) {
1339                 if (!user_shm_lock(inode->i_size, user))
1340                         goto out_nomem;
1341                 info->flags |= VM_LOCKED;
1342                 mapping_set_unevictable(file->f_mapping);
1343         }
1344         if (!lock && (info->flags & VM_LOCKED) && user) {
1345                 user_shm_unlock(inode->i_size, user);
1346                 info->flags &= ~VM_LOCKED;
1347                 mapping_clear_unevictable(file->f_mapping);
1348         }
1349         retval = 0;
1350
1351 out_nomem:
1352         spin_unlock(&info->lock);
1353         return retval;
1354 }
1355
1356 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1357 {
1358         file_accessed(file);
1359         vma->vm_ops = &shmem_vm_ops;
1360         return 0;
1361 }
1362
1363 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1364                                      umode_t mode, dev_t dev, unsigned long flags)
1365 {
1366         struct inode *inode;
1367         struct shmem_inode_info *info;
1368         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1369
1370         if (shmem_reserve_inode(sb))
1371                 return NULL;
1372
1373         inode = new_inode(sb);
1374         if (inode) {
1375                 inode->i_ino = get_next_ino();
1376                 inode_init_owner(inode, dir, mode);
1377                 inode->i_blocks = 0;
1378                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1379                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1380                 inode->i_generation = get_seconds();
1381                 info = SHMEM_I(inode);
1382                 memset(info, 0, (char *)inode - (char *)info);
1383                 spin_lock_init(&info->lock);
1384                 info->flags = flags & VM_NORESERVE;
1385                 INIT_LIST_HEAD(&info->swaplist);
1386                 simple_xattrs_init(&info->xattrs);
1387                 cache_no_acl(inode);
1388
1389                 switch (mode & S_IFMT) {
1390                 default:
1391                         inode->i_op = &shmem_special_inode_operations;
1392                         init_special_inode(inode, mode, dev);
1393                         break;
1394                 case S_IFREG:
1395                         inode->i_mapping->a_ops = &shmem_aops;
1396                         inode->i_op = &shmem_inode_operations;
1397                         inode->i_fop = &shmem_file_operations;
1398                         mpol_shared_policy_init(&info->policy,
1399                                                  shmem_get_sbmpol(sbinfo));
1400                         break;
1401                 case S_IFDIR:
1402                         inc_nlink(inode);
1403                         /* Some things misbehave if size == 0 on a directory */
1404                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1405                         inode->i_op = &shmem_dir_inode_operations;
1406                         inode->i_fop = &simple_dir_operations;
1407                         break;
1408                 case S_IFLNK:
1409                         /*
1410                          * Must not load anything in the rbtree,
1411                          * mpol_free_shared_policy will not be called.
1412                          */
1413                         mpol_shared_policy_init(&info->policy, NULL);
1414                         break;
1415                 }
1416         } else
1417                 shmem_free_inode(sb);
1418         return inode;
1419 }
1420
1421 #ifdef CONFIG_TMPFS
1422 static const struct inode_operations shmem_symlink_inode_operations;
1423 static const struct inode_operations shmem_short_symlink_operations;
1424
1425 #ifdef CONFIG_TMPFS_XATTR
1426 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1427 #else
1428 #define shmem_initxattrs NULL
1429 #endif
1430
1431 static int
1432 shmem_write_begin(struct file *file, struct address_space *mapping,
1433                         loff_t pos, unsigned len, unsigned flags,
1434                         struct page **pagep, void **fsdata)
1435 {
1436         struct inode *inode = mapping->host;
1437         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1438         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1439 }
1440
1441 static int
1442 shmem_write_end(struct file *file, struct address_space *mapping,
1443                         loff_t pos, unsigned len, unsigned copied,
1444                         struct page *page, void *fsdata)
1445 {
1446         struct inode *inode = mapping->host;
1447
1448         if (pos + copied > inode->i_size)
1449                 i_size_write(inode, pos + copied);
1450
1451         if (!PageUptodate(page)) {
1452                 if (copied < PAGE_CACHE_SIZE) {
1453                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1454                         zero_user_segments(page, 0, from,
1455                                         from + copied, PAGE_CACHE_SIZE);
1456                 }
1457                 SetPageUptodate(page);
1458         }
1459         set_page_dirty(page);
1460         unlock_page(page);
1461         page_cache_release(page);
1462
1463         return copied;
1464 }
1465
1466 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1467 {
1468         struct inode *inode = file_inode(filp);
1469         struct address_space *mapping = inode->i_mapping;
1470         pgoff_t index;
1471         unsigned long offset;
1472         enum sgp_type sgp = SGP_READ;
1473
1474         /*
1475          * Might this read be for a stacking filesystem?  Then when reading
1476          * holes of a sparse file, we actually need to allocate those pages,
1477          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1478          */
1479         if (segment_eq(get_fs(), KERNEL_DS))
1480                 sgp = SGP_DIRTY;
1481
1482         index = *ppos >> PAGE_CACHE_SHIFT;
1483         offset = *ppos & ~PAGE_CACHE_MASK;
1484
1485         for (;;) {
1486                 struct page *page = NULL;
1487                 pgoff_t end_index;
1488                 unsigned long nr, ret;
1489                 loff_t i_size = i_size_read(inode);
1490
1491                 end_index = i_size >> PAGE_CACHE_SHIFT;
1492                 if (index > end_index)
1493                         break;
1494                 if (index == end_index) {
1495                         nr = i_size & ~PAGE_CACHE_MASK;
1496                         if (nr <= offset)
1497                                 break;
1498                 }
1499
1500                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1501                 if (desc->error) {
1502                         if (desc->error == -EINVAL)
1503                                 desc->error = 0;
1504                         break;
1505                 }
1506                 if (page)
1507                         unlock_page(page);
1508
1509                 /*
1510                  * We must evaluate after, since reads (unlike writes)
1511                  * are called without i_mutex protection against truncate
1512                  */
1513                 nr = PAGE_CACHE_SIZE;
1514                 i_size = i_size_read(inode);
1515                 end_index = i_size >> PAGE_CACHE_SHIFT;
1516                 if (index == end_index) {
1517                         nr = i_size & ~PAGE_CACHE_MASK;
1518                         if (nr <= offset) {
1519                                 if (page)
1520                                         page_cache_release(page);
1521                                 break;
1522                         }
1523                 }
1524                 nr -= offset;
1525
1526                 if (page) {
1527                         /*
1528                          * If users can be writing to this page using arbitrary
1529                          * virtual addresses, take care about potential aliasing
1530                          * before reading the page on the kernel side.
1531                          */
1532                         if (mapping_writably_mapped(mapping))
1533                                 flush_dcache_page(page);
1534                         /*
1535                          * Mark the page accessed if we read the beginning.
1536                          */
1537                         if (!offset)
1538                                 mark_page_accessed(page);
1539                 } else {
1540                         page = ZERO_PAGE(0);
1541                         page_cache_get(page);
1542                 }
1543
1544                 /*
1545                  * Ok, we have the page, and it's up-to-date, so
1546                  * now we can copy it to user space...
1547                  *
1548                  * The actor routine returns how many bytes were actually used..
1549                  * NOTE! This may not be the same as how much of a user buffer
1550                  * we filled up (we may be padding etc), so we can only update
1551                  * "pos" here (the actor routine has to update the user buffer
1552                  * pointers and the remaining count).
1553                  */
1554                 ret = actor(desc, page, offset, nr);
1555                 offset += ret;
1556                 index += offset >> PAGE_CACHE_SHIFT;
1557                 offset &= ~PAGE_CACHE_MASK;
1558
1559                 page_cache_release(page);
1560                 if (ret != nr || !desc->count)
1561                         break;
1562
1563                 cond_resched();
1564         }
1565
1566         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1567         file_accessed(filp);
1568 }
1569
1570 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1571                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1572 {
1573         struct file *filp = iocb->ki_filp;
1574         ssize_t retval;
1575         unsigned long seg;
1576         size_t count;
1577         loff_t *ppos = &iocb->ki_pos;
1578
1579         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1580         if (retval)
1581                 return retval;
1582
1583         for (seg = 0; seg < nr_segs; seg++) {
1584                 read_descriptor_t desc;
1585
1586                 desc.written = 0;
1587                 desc.arg.buf = iov[seg].iov_base;
1588                 desc.count = iov[seg].iov_len;
1589                 if (desc.count == 0)
1590                         continue;
1591                 desc.error = 0;
1592                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1593                 retval += desc.written;
1594                 if (desc.error) {
1595                         retval = retval ?: desc.error;
1596                         break;
1597                 }
1598                 if (desc.count > 0)
1599                         break;
1600         }
1601         return retval;
1602 }
1603
1604 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1605                                 struct pipe_inode_info *pipe, size_t len,
1606                                 unsigned int flags)
1607 {
1608         struct address_space *mapping = in->f_mapping;
1609         struct inode *inode = mapping->host;
1610         unsigned int loff, nr_pages, req_pages;
1611         struct page *pages[PIPE_DEF_BUFFERS];
1612         struct partial_page partial[PIPE_DEF_BUFFERS];
1613         struct page *page;
1614         pgoff_t index, end_index;
1615         loff_t isize, left;
1616         int error, page_nr;
1617         struct splice_pipe_desc spd = {
1618                 .pages = pages,
1619                 .partial = partial,
1620                 .nr_pages_max = PIPE_DEF_BUFFERS,
1621                 .flags = flags,
1622                 .ops = &page_cache_pipe_buf_ops,
1623                 .spd_release = spd_release_page,
1624         };
1625
1626         isize = i_size_read(inode);
1627         if (unlikely(*ppos >= isize))
1628                 return 0;
1629
1630         left = isize - *ppos;
1631         if (unlikely(left < len))
1632                 len = left;
1633
1634         if (splice_grow_spd(pipe, &spd))
1635                 return -ENOMEM;
1636
1637         index = *ppos >> PAGE_CACHE_SHIFT;
1638         loff = *ppos & ~PAGE_CACHE_MASK;
1639         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1640         nr_pages = min(req_pages, pipe->buffers);
1641
1642         spd.nr_pages = find_get_pages_contig(mapping, index,
1643                                                 nr_pages, spd.pages);
1644         index += spd.nr_pages;
1645         error = 0;
1646
1647         while (spd.nr_pages < nr_pages) {
1648                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1649                 if (error)
1650                         break;
1651                 unlock_page(page);
1652                 spd.pages[spd.nr_pages++] = page;
1653                 index++;
1654         }
1655
1656         index = *ppos >> PAGE_CACHE_SHIFT;
1657         nr_pages = spd.nr_pages;
1658         spd.nr_pages = 0;
1659
1660         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1661                 unsigned int this_len;
1662
1663                 if (!len)
1664                         break;
1665
1666                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1667                 page = spd.pages[page_nr];
1668
1669                 if (!PageUptodate(page) || page->mapping != mapping) {
1670                         error = shmem_getpage(inode, index, &page,
1671                                                         SGP_CACHE, NULL);
1672                         if (error)
1673                                 break;
1674                         unlock_page(page);
1675                         page_cache_release(spd.pages[page_nr]);
1676                         spd.pages[page_nr] = page;
1677                 }
1678
1679                 isize = i_size_read(inode);
1680                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1681                 if (unlikely(!isize || index > end_index))
1682                         break;
1683
1684                 if (end_index == index) {
1685                         unsigned int plen;
1686
1687                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1688                         if (plen <= loff)
1689                                 break;
1690
1691                         this_len = min(this_len, plen - loff);
1692                         len = this_len;
1693                 }
1694
1695                 spd.partial[page_nr].offset = loff;
1696                 spd.partial[page_nr].len = this_len;
1697                 len -= this_len;
1698                 loff = 0;
1699                 spd.nr_pages++;
1700                 index++;
1701         }
1702
1703         while (page_nr < nr_pages)
1704                 page_cache_release(spd.pages[page_nr++]);
1705
1706         if (spd.nr_pages)
1707                 error = splice_to_pipe(pipe, &spd);
1708
1709         splice_shrink_spd(&spd);
1710
1711         if (error > 0) {
1712                 *ppos += error;
1713                 file_accessed(in);
1714         }
1715         return error;
1716 }
1717
1718 /*
1719  * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
1720  */
1721 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
1722                                     pgoff_t index, pgoff_t end, int whence)
1723 {
1724         struct page *page;
1725         struct pagevec pvec;
1726         pgoff_t indices[PAGEVEC_SIZE];
1727         bool done = false;
1728         int i;
1729
1730         pagevec_init(&pvec, 0);
1731         pvec.nr = 1;            /* start small: we may be there already */
1732         while (!done) {
1733                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
1734                                         pvec.nr, pvec.pages, indices);
1735                 if (!pvec.nr) {
1736                         if (whence == SEEK_DATA)
1737                                 index = end;
1738                         break;
1739                 }
1740                 for (i = 0; i < pvec.nr; i++, index++) {
1741                         if (index < indices[i]) {
1742                                 if (whence == SEEK_HOLE) {
1743                                         done = true;
1744                                         break;
1745                                 }
1746                                 index = indices[i];
1747                         }
1748                         page = pvec.pages[i];
1749                         if (page && !radix_tree_exceptional_entry(page)) {
1750                                 if (!PageUptodate(page))
1751                                         page = NULL;
1752                         }
1753                         if (index >= end ||
1754                             (page && whence == SEEK_DATA) ||
1755                             (!page && whence == SEEK_HOLE)) {
1756                                 done = true;
1757                                 break;
1758                         }
1759                 }
1760                 shmem_deswap_pagevec(&pvec);
1761                 pagevec_release(&pvec);
1762                 pvec.nr = PAGEVEC_SIZE;
1763                 cond_resched();
1764         }
1765         return index;
1766 }
1767
1768 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
1769 {
1770         struct address_space *mapping = file->f_mapping;
1771         struct inode *inode = mapping->host;
1772         pgoff_t start, end;
1773         loff_t new_offset;
1774
1775         if (whence != SEEK_DATA && whence != SEEK_HOLE)
1776                 return generic_file_llseek_size(file, offset, whence,
1777                                         MAX_LFS_FILESIZE, i_size_read(inode));
1778         mutex_lock(&inode->i_mutex);
1779         /* We're holding i_mutex so we can access i_size directly */
1780
1781         if (offset < 0)
1782                 offset = -EINVAL;
1783         else if (offset >= inode->i_size)
1784                 offset = -ENXIO;
1785         else {
1786                 start = offset >> PAGE_CACHE_SHIFT;
1787                 end = (inode->i_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1788                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
1789                 new_offset <<= PAGE_CACHE_SHIFT;
1790                 if (new_offset > offset) {
1791                         if (new_offset < inode->i_size)
1792                                 offset = new_offset;
1793                         else if (whence == SEEK_DATA)
1794                                 offset = -ENXIO;
1795                         else
1796                                 offset = inode->i_size;
1797                 }
1798         }
1799
1800         if (offset >= 0 && offset != file->f_pos) {
1801                 file->f_pos = offset;
1802                 file->f_version = 0;
1803         }
1804         mutex_unlock(&inode->i_mutex);
1805         return offset;
1806 }
1807
1808 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
1809                                                          loff_t len)
1810 {
1811         struct inode *inode = file_inode(file);
1812         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1813         struct shmem_falloc shmem_falloc;
1814         pgoff_t start, index, end;
1815         int error;
1816
1817         mutex_lock(&inode->i_mutex);
1818
1819         if (mode & FALLOC_FL_PUNCH_HOLE) {
1820                 struct address_space *mapping = file->f_mapping;
1821                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
1822                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
1823
1824                 if ((u64)unmap_end > (u64)unmap_start)
1825                         unmap_mapping_range(mapping, unmap_start,
1826                                             1 + unmap_end - unmap_start, 0);
1827                 shmem_truncate_range(inode, offset, offset + len - 1);
1828                 /* No need to unmap again: hole-punching leaves COWed pages */
1829                 error = 0;
1830                 goto out;
1831         }
1832
1833         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
1834         error = inode_newsize_ok(inode, offset + len);
1835         if (error)
1836                 goto out;
1837
1838         start = offset >> PAGE_CACHE_SHIFT;
1839         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1840         /* Try to avoid a swapstorm if len is impossible to satisfy */
1841         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
1842                 error = -ENOSPC;
1843                 goto out;
1844         }
1845
1846         shmem_falloc.start = start;
1847         shmem_falloc.next  = start;
1848         shmem_falloc.nr_falloced = 0;
1849         shmem_falloc.nr_unswapped = 0;
1850         spin_lock(&inode->i_lock);
1851         inode->i_private = &shmem_falloc;
1852         spin_unlock(&inode->i_lock);
1853
1854         for (index = start; index < end; index++) {
1855                 struct page *page;
1856
1857                 /*
1858                  * Good, the fallocate(2) manpage permits EINTR: we may have
1859                  * been interrupted because we are using up too much memory.
1860                  */
1861                 if (signal_pending(current))
1862                         error = -EINTR;
1863                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
1864                         error = -ENOMEM;
1865                 else
1866                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
1867                                                                         NULL);
1868                 if (error) {
1869                         /* Remove the !PageUptodate pages we added */
1870                         shmem_undo_range(inode,
1871                                 (loff_t)start << PAGE_CACHE_SHIFT,
1872                                 (loff_t)index << PAGE_CACHE_SHIFT, true);
1873                         goto undone;
1874                 }
1875
1876                 /*
1877                  * Inform shmem_writepage() how far we have reached.
1878                  * No need for lock or barrier: we have the page lock.
1879                  */
1880                 shmem_falloc.next++;
1881                 if (!PageUptodate(page))
1882                         shmem_falloc.nr_falloced++;
1883
1884                 /*
1885                  * If !PageUptodate, leave it that way so that freeable pages
1886                  * can be recognized if we need to rollback on error later.
1887                  * But set_page_dirty so that memory pressure will swap rather
1888                  * than free the pages we are allocating (and SGP_CACHE pages
1889                  * might still be clean: we now need to mark those dirty too).
1890                  */
1891                 set_page_dirty(page);
1892                 unlock_page(page);
1893                 page_cache_release(page);
1894                 cond_resched();
1895         }
1896
1897         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
1898                 i_size_write(inode, offset + len);
1899         inode->i_ctime = CURRENT_TIME;
1900 undone:
1901         spin_lock(&inode->i_lock);
1902         inode->i_private = NULL;
1903         spin_unlock(&inode->i_lock);
1904 out:
1905         mutex_unlock(&inode->i_mutex);
1906         return error;
1907 }
1908
1909 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1910 {
1911         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1912
1913         buf->f_type = TMPFS_MAGIC;
1914         buf->f_bsize = PAGE_CACHE_SIZE;
1915         buf->f_namelen = NAME_MAX;
1916         if (sbinfo->max_blocks) {
1917                 buf->f_blocks = sbinfo->max_blocks;
1918                 buf->f_bavail =
1919                 buf->f_bfree  = sbinfo->max_blocks -
1920                                 percpu_counter_sum(&sbinfo->used_blocks);
1921         }
1922         if (sbinfo->max_inodes) {
1923                 buf->f_files = sbinfo->max_inodes;
1924                 buf->f_ffree = sbinfo->free_inodes;
1925         }
1926         /* else leave those fields 0 like simple_statfs */
1927         return 0;
1928 }
1929
1930 /*
1931  * File creation. Allocate an inode, and we're done..
1932  */
1933 static int
1934 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
1935 {
1936         struct inode *inode;
1937         int error = -ENOSPC;
1938
1939         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1940         if (inode) {
1941                 error = security_inode_init_security(inode, dir,
1942                                                      &dentry->d_name,
1943                                                      shmem_initxattrs, NULL);
1944                 if (error) {
1945                         if (error != -EOPNOTSUPP) {
1946                                 iput(inode);
1947                                 return error;
1948                         }
1949                 }
1950 #ifdef CONFIG_TMPFS_POSIX_ACL
1951                 error = generic_acl_init(inode, dir);
1952                 if (error) {
1953                         iput(inode);
1954                         return error;
1955                 }
1956 #else
1957                 error = 0;
1958 #endif
1959                 dir->i_size += BOGO_DIRENT_SIZE;
1960                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1961                 d_instantiate(dentry, inode);
1962                 dget(dentry); /* Extra count - pin the dentry in core */
1963         }
1964         return error;
1965 }
1966
1967 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1968 {
1969         int error;
1970
1971         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1972                 return error;
1973         inc_nlink(dir);
1974         return 0;
1975 }
1976
1977 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
1978                 bool excl)
1979 {
1980         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1981 }
1982
1983 /*
1984  * Link a file..
1985  */
1986 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1987 {
1988         struct inode *inode = old_dentry->d_inode;
1989         int ret;
1990
1991         /*
1992          * No ordinary (disk based) filesystem counts links as inodes;
1993          * but each new link needs a new dentry, pinning lowmem, and
1994          * tmpfs dentries cannot be pruned until they are unlinked.
1995          */
1996         ret = shmem_reserve_inode(inode->i_sb);
1997         if (ret)
1998                 goto out;
1999
2000         dir->i_size += BOGO_DIRENT_SIZE;
2001         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2002         inc_nlink(inode);
2003         ihold(inode);   /* New dentry reference */
2004         dget(dentry);           /* Extra pinning count for the created dentry */
2005         d_instantiate(dentry, inode);
2006 out:
2007         return ret;
2008 }
2009
2010 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2011 {
2012         struct inode *inode = dentry->d_inode;
2013
2014         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2015                 shmem_free_inode(inode->i_sb);
2016
2017         dir->i_size -= BOGO_DIRENT_SIZE;
2018         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2019         drop_nlink(inode);
2020         dput(dentry);   /* Undo the count from "create" - this does all the work */
2021         return 0;
2022 }
2023
2024 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2025 {
2026         if (!simple_empty(dentry))
2027                 return -ENOTEMPTY;
2028
2029         drop_nlink(dentry->d_inode);
2030         drop_nlink(dir);
2031         return shmem_unlink(dir, dentry);
2032 }
2033
2034 /*
2035  * The VFS layer already does all the dentry stuff for rename,
2036  * we just have to decrement the usage count for the target if
2037  * it exists so that the VFS layer correctly free's it when it
2038  * gets overwritten.
2039  */
2040 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2041 {
2042         struct inode *inode = old_dentry->d_inode;
2043         int they_are_dirs = S_ISDIR(inode->i_mode);
2044
2045         if (!simple_empty(new_dentry))
2046                 return -ENOTEMPTY;
2047
2048         if (new_dentry->d_inode) {
2049                 (void) shmem_unlink(new_dir, new_dentry);
2050                 if (they_are_dirs)
2051                         drop_nlink(old_dir);
2052         } else if (they_are_dirs) {
2053                 drop_nlink(old_dir);
2054                 inc_nlink(new_dir);
2055         }
2056
2057         old_dir->i_size -= BOGO_DIRENT_SIZE;
2058         new_dir->i_size += BOGO_DIRENT_SIZE;
2059         old_dir->i_ctime = old_dir->i_mtime =
2060         new_dir->i_ctime = new_dir->i_mtime =
2061         inode->i_ctime = CURRENT_TIME;
2062         return 0;
2063 }
2064
2065 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2066 {
2067         int error;
2068         int len;
2069         struct inode *inode;
2070         struct page *page;
2071         char *kaddr;
2072         struct shmem_inode_info *info;
2073
2074         len = strlen(symname) + 1;
2075         if (len > PAGE_CACHE_SIZE)
2076                 return -ENAMETOOLONG;
2077
2078         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2079         if (!inode)
2080                 return -ENOSPC;
2081
2082         error = security_inode_init_security(inode, dir, &dentry->d_name,
2083                                              shmem_initxattrs, NULL);
2084         if (error) {
2085                 if (error != -EOPNOTSUPP) {
2086                         iput(inode);
2087                         return error;
2088                 }
2089                 error = 0;
2090         }
2091
2092         info = SHMEM_I(inode);
2093         inode->i_size = len-1;
2094         if (len <= SHORT_SYMLINK_LEN) {
2095                 info->symlink = kmemdup(symname, len, GFP_KERNEL);
2096                 if (!info->symlink) {
2097                         iput(inode);
2098                         return -ENOMEM;
2099                 }
2100                 inode->i_op = &shmem_short_symlink_operations;
2101         } else {
2102                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2103                 if (error) {
2104                         iput(inode);
2105                         return error;
2106                 }
2107                 inode->i_mapping->a_ops = &shmem_aops;
2108                 inode->i_op = &shmem_symlink_inode_operations;
2109                 kaddr = kmap_atomic(page);
2110                 memcpy(kaddr, symname, len);
2111                 kunmap_atomic(kaddr);
2112                 SetPageUptodate(page);
2113                 set_page_dirty(page);
2114                 unlock_page(page);
2115                 page_cache_release(page);
2116         }
2117         dir->i_size += BOGO_DIRENT_SIZE;
2118         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2119         d_instantiate(dentry, inode);
2120         dget(dentry);
2121         return 0;
2122 }
2123
2124 static void *shmem_follow_short_symlink(struct dentry *dentry, struct nameidata *nd)
2125 {
2126         nd_set_link(nd, SHMEM_I(dentry->d_inode)->symlink);
2127         return NULL;
2128 }
2129
2130 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2131 {
2132         struct page *page = NULL;
2133         int error = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2134         nd_set_link(nd, error ? ERR_PTR(error) : kmap(page));
2135         if (page)
2136                 unlock_page(page);
2137         return page;
2138 }
2139
2140 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2141 {
2142         if (!IS_ERR(nd_get_link(nd))) {
2143                 struct page *page = cookie;
2144                 kunmap(page);
2145                 mark_page_accessed(page);
2146                 page_cache_release(page);
2147         }
2148 }
2149
2150 #ifdef CONFIG_TMPFS_XATTR
2151 /*
2152  * Superblocks without xattr inode operations may get some security.* xattr
2153  * support from the LSM "for free". As soon as we have any other xattrs
2154  * like ACLs, we also need to implement the security.* handlers at
2155  * filesystem level, though.
2156  */
2157
2158 /*
2159  * Callback for security_inode_init_security() for acquiring xattrs.
2160  */
2161 static int shmem_initxattrs(struct inode *inode,
2162                             const struct xattr *xattr_array,
2163                             void *fs_info)
2164 {
2165         struct shmem_inode_info *info = SHMEM_I(inode);
2166         const struct xattr *xattr;
2167         struct simple_xattr *new_xattr;
2168         size_t len;
2169
2170         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2171                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2172                 if (!new_xattr)
2173                         return -ENOMEM;
2174
2175                 len = strlen(xattr->name) + 1;
2176                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2177                                           GFP_KERNEL);
2178                 if (!new_xattr->name) {
2179                         kfree(new_xattr);
2180                         return -ENOMEM;
2181                 }
2182
2183                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2184                        XATTR_SECURITY_PREFIX_LEN);
2185                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2186                        xattr->name, len);
2187
2188                 simple_xattr_list_add(&info->xattrs, new_xattr);
2189         }
2190
2191         return 0;
2192 }
2193
2194 static const struct xattr_handler *shmem_xattr_handlers[] = {
2195 #ifdef CONFIG_TMPFS_POSIX_ACL
2196         &generic_acl_access_handler,
2197         &generic_acl_default_handler,
2198 #endif
2199         NULL
2200 };
2201
2202 static int shmem_xattr_validate(const char *name)
2203 {
2204         struct { const char *prefix; size_t len; } arr[] = {
2205                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2206                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2207         };
2208         int i;
2209
2210         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2211                 size_t preflen = arr[i].len;
2212                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2213                         if (!name[preflen])
2214                                 return -EINVAL;
2215                         return 0;
2216                 }
2217         }
2218         return -EOPNOTSUPP;
2219 }
2220
2221 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2222                               void *buffer, size_t size)
2223 {
2224         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2225         int err;
2226
2227         /*
2228          * If this is a request for a synthetic attribute in the system.*
2229          * namespace use the generic infrastructure to resolve a handler
2230          * for it via sb->s_xattr.
2231          */
2232         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2233                 return generic_getxattr(dentry, name, buffer, size);
2234
2235         err = shmem_xattr_validate(name);
2236         if (err)
2237                 return err;
2238
2239         return simple_xattr_get(&info->xattrs, name, buffer, size);
2240 }
2241
2242 static int shmem_setxattr(struct dentry *dentry, const char *name,
2243                           const void *value, size_t size, int flags)
2244 {
2245         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2246         int err;
2247
2248         /*
2249          * If this is a request for a synthetic attribute in the system.*
2250          * namespace use the generic infrastructure to resolve a handler
2251          * for it via sb->s_xattr.
2252          */
2253         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2254                 return generic_setxattr(dentry, name, value, size, flags);
2255
2256         err = shmem_xattr_validate(name);
2257         if (err)
2258                 return err;
2259
2260         return simple_xattr_set(&info->xattrs, name, value, size, flags);
2261 }
2262
2263 static int shmem_removexattr(struct dentry *dentry, const char *name)
2264 {
2265         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2266         int err;
2267
2268         /*
2269          * If this is a request for a synthetic attribute in the system.*
2270          * namespace use the generic infrastructure to resolve a handler
2271          * for it via sb->s_xattr.
2272          */
2273         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2274                 return generic_removexattr(dentry, name);
2275
2276         err = shmem_xattr_validate(name);
2277         if (err)
2278                 return err;
2279
2280         return simple_xattr_remove(&info->xattrs, name);
2281 }
2282
2283 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2284 {
2285         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2286         return simple_xattr_list(&info->xattrs, buffer, size);
2287 }
2288 #endif /* CONFIG_TMPFS_XATTR */
2289
2290 static const struct inode_operations shmem_short_symlink_operations = {
2291         .readlink       = generic_readlink,
2292         .follow_link    = shmem_follow_short_symlink,
2293 #ifdef CONFIG_TMPFS_XATTR
2294         .setxattr       = shmem_setxattr,
2295         .getxattr       = shmem_getxattr,
2296         .listxattr      = shmem_listxattr,
2297         .removexattr    = shmem_removexattr,
2298 #endif
2299 };
2300
2301 static const struct inode_operations shmem_symlink_inode_operations = {
2302         .readlink       = generic_readlink,
2303         .follow_link    = shmem_follow_link,
2304         .put_link       = shmem_put_link,
2305 #ifdef CONFIG_TMPFS_XATTR
2306         .setxattr       = shmem_setxattr,
2307         .getxattr       = shmem_getxattr,
2308         .listxattr      = shmem_listxattr,
2309         .removexattr    = shmem_removexattr,
2310 #endif
2311 };
2312
2313 static struct dentry *shmem_get_parent(struct dentry *child)
2314 {
2315         return ERR_PTR(-ESTALE);
2316 }
2317
2318 static int shmem_match(struct inode *ino, void *vfh)
2319 {
2320         __u32 *fh = vfh;
2321         __u64 inum = fh[2];
2322         inum = (inum << 32) | fh[1];
2323         return ino->i_ino == inum && fh[0] == ino->i_generation;
2324 }
2325
2326 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2327                 struct fid *fid, int fh_len, int fh_type)
2328 {
2329         struct inode *inode;
2330         struct dentry *dentry = NULL;
2331         u64 inum;
2332
2333         if (fh_len < 3)
2334                 return NULL;
2335
2336         inum = fid->raw[2];
2337         inum = (inum << 32) | fid->raw[1];
2338
2339         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2340                         shmem_match, fid->raw);
2341         if (inode) {
2342                 dentry = d_find_alias(inode);
2343                 iput(inode);
2344         }
2345
2346         return dentry;
2347 }
2348
2349 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2350                                 struct inode *parent)
2351 {
2352         if (*len < 3) {
2353                 *len = 3;
2354                 return FILEID_INVALID;
2355         }
2356
2357         if (inode_unhashed(inode)) {
2358                 /* Unfortunately insert_inode_hash is not idempotent,
2359                  * so as we hash inodes here rather than at creation
2360                  * time, we need a lock to ensure we only try
2361                  * to do it once
2362                  */
2363                 static DEFINE_SPINLOCK(lock);
2364                 spin_lock(&lock);
2365                 if (inode_unhashed(inode))
2366                         __insert_inode_hash(inode,
2367                                             inode->i_ino + inode->i_generation);
2368                 spin_unlock(&lock);
2369         }
2370
2371         fh[0] = inode->i_generation;
2372         fh[1] = inode->i_ino;
2373         fh[2] = ((__u64)inode->i_ino) >> 32;
2374
2375         *len = 3;
2376         return 1;
2377 }
2378
2379 static const struct export_operations shmem_export_ops = {
2380         .get_parent     = shmem_get_parent,
2381         .encode_fh      = shmem_encode_fh,
2382         .fh_to_dentry   = shmem_fh_to_dentry,
2383 };
2384
2385 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2386                                bool remount)
2387 {
2388         char *this_char, *value, *rest;
2389         struct mempolicy *mpol = NULL;
2390         uid_t uid;
2391         gid_t gid;
2392
2393         while (options != NULL) {
2394                 this_char = options;
2395                 for (;;) {
2396                         /*
2397                          * NUL-terminate this option: unfortunately,
2398                          * mount options form a comma-separated list,
2399                          * but mpol's nodelist may also contain commas.
2400                          */
2401                         options = strchr(options, ',');
2402                         if (options == NULL)
2403                                 break;
2404                         options++;
2405                         if (!isdigit(*options)) {
2406                                 options[-1] = '\0';
2407                                 break;
2408                         }
2409                 }
2410                 if (!*this_char)
2411                         continue;
2412                 if ((value = strchr(this_char,'=')) != NULL) {
2413                         *value++ = 0;
2414                 } else {
2415                         printk(KERN_ERR
2416                             "tmpfs: No value for mount option '%s'\n",
2417                             this_char);
2418                         goto error;
2419                 }
2420
2421                 if (!strcmp(this_char,"size")) {
2422                         unsigned long long size;
2423                         size = memparse(value,&rest);
2424                         if (*rest == '%') {
2425                                 size <<= PAGE_SHIFT;
2426                                 size *= totalram_pages;
2427                                 do_div(size, 100);
2428                                 rest++;
2429                         }
2430                         if (*rest)
2431                                 goto bad_val;
2432                         sbinfo->max_blocks =
2433                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2434                 } else if (!strcmp(this_char,"nr_blocks")) {
2435                         sbinfo->max_blocks = memparse(value, &rest);
2436                         if (*rest)
2437                                 goto bad_val;
2438                 } else if (!strcmp(this_char,"nr_inodes")) {
2439                         sbinfo->max_inodes = memparse(value, &rest);
2440                         if (*rest)
2441                                 goto bad_val;
2442                 } else if (!strcmp(this_char,"mode")) {
2443                         if (remount)
2444                                 continue;
2445                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2446                         if (*rest)
2447                                 goto bad_val;
2448                 } else if (!strcmp(this_char,"uid")) {
2449                         if (remount)
2450                                 continue;
2451                         uid = simple_strtoul(value, &rest, 0);
2452                         if (*rest)
2453                                 goto bad_val;
2454                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2455                         if (!uid_valid(sbinfo->uid))
2456                                 goto bad_val;
2457                 } else if (!strcmp(this_char,"gid")) {
2458                         if (remount)
2459                                 continue;
2460                         gid = simple_strtoul(value, &rest, 0);
2461                         if (*rest)
2462                                 goto bad_val;
2463                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2464                         if (!gid_valid(sbinfo->gid))
2465                                 goto bad_val;
2466                 } else if (!strcmp(this_char,"mpol")) {
2467                         mpol_put(mpol);
2468                         mpol = NULL;
2469                         if (mpol_parse_str(value, &mpol))
2470                                 goto bad_val;
2471                 } else {
2472                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2473                                this_char);
2474                         goto error;
2475                 }
2476         }
2477         sbinfo->mpol = mpol;
2478         return 0;
2479
2480 bad_val:
2481         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2482                value, this_char);
2483 error:
2484         mpol_put(mpol);
2485         return 1;
2486
2487 }
2488
2489 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2490 {
2491         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2492         struct shmem_sb_info config = *sbinfo;
2493         unsigned long inodes;
2494         int error = -EINVAL;
2495
2496         config.mpol = NULL;
2497         if (shmem_parse_options(data, &config, true))
2498                 return error;
2499
2500         spin_lock(&sbinfo->stat_lock);
2501         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2502         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2503                 goto out;
2504         if (config.max_inodes < inodes)
2505                 goto out;
2506         /*
2507          * Those tests disallow limited->unlimited while any are in use;
2508          * but we must separately disallow unlimited->limited, because
2509          * in that case we have no record of how much is already in use.
2510          */
2511         if (config.max_blocks && !sbinfo->max_blocks)
2512                 goto out;
2513         if (config.max_inodes && !sbinfo->max_inodes)
2514                 goto out;
2515
2516         error = 0;
2517         sbinfo->max_blocks  = config.max_blocks;
2518         sbinfo->max_inodes  = config.max_inodes;
2519         sbinfo->free_inodes = config.max_inodes - inodes;
2520
2521         /*
2522          * Preserve previous mempolicy unless mpol remount option was specified.
2523          */
2524         if (config.mpol) {
2525                 mpol_put(sbinfo->mpol);
2526                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
2527         }
2528 out:
2529         spin_unlock(&sbinfo->stat_lock);
2530         return error;
2531 }
2532
2533 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2534 {
2535         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2536
2537         if (sbinfo->max_blocks != shmem_default_max_blocks())
2538                 seq_printf(seq, ",size=%luk",
2539                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2540         if (sbinfo->max_inodes != shmem_default_max_inodes())
2541                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2542         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2543                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2544         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2545                 seq_printf(seq, ",uid=%u",
2546                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2547         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2548                 seq_printf(seq, ",gid=%u",
2549                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2550         shmem_show_mpol(seq, sbinfo->mpol);
2551         return 0;
2552 }
2553 #endif /* CONFIG_TMPFS */
2554
2555 static void shmem_put_super(struct super_block *sb)
2556 {
2557         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2558
2559         percpu_counter_destroy(&sbinfo->used_blocks);
2560         mpol_put(sbinfo->mpol);
2561         kfree(sbinfo);
2562         sb->s_fs_info = NULL;
2563 }
2564
2565 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2566 {
2567         struct inode *inode;
2568         struct shmem_sb_info *sbinfo;
2569         int err = -ENOMEM;
2570
2571         /* Round up to L1_CACHE_BYTES to resist false sharing */
2572         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2573                                 L1_CACHE_BYTES), GFP_KERNEL);
2574         if (!sbinfo)
2575                 return -ENOMEM;
2576
2577         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2578         sbinfo->uid = current_fsuid();
2579         sbinfo->gid = current_fsgid();
2580         sb->s_fs_info = sbinfo;
2581
2582 #ifdef CONFIG_TMPFS
2583         /*
2584          * Per default we only allow half of the physical ram per
2585          * tmpfs instance, limiting inodes to one per page of lowmem;
2586          * but the internal instance is left unlimited.
2587          */
2588         if (!(sb->s_flags & MS_NOUSER)) {
2589                 sbinfo->max_blocks = shmem_default_max_blocks();
2590                 sbinfo->max_inodes = shmem_default_max_inodes();
2591                 if (shmem_parse_options(data, sbinfo, false)) {
2592                         err = -EINVAL;
2593                         goto failed;
2594                 }
2595         }
2596         sb->s_export_op = &shmem_export_ops;
2597         sb->s_flags |= MS_NOSEC;
2598 #else
2599         sb->s_flags |= MS_NOUSER;
2600 #endif
2601
2602         spin_lock_init(&sbinfo->stat_lock);
2603         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2604                 goto failed;
2605         sbinfo->free_inodes = sbinfo->max_inodes;
2606
2607         sb->s_maxbytes = MAX_LFS_FILESIZE;
2608         sb->s_blocksize = PAGE_CACHE_SIZE;
2609         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2610         sb->s_magic = TMPFS_MAGIC;
2611         sb->s_op = &shmem_ops;
2612         sb->s_time_gran = 1;
2613 #ifdef CONFIG_TMPFS_XATTR
2614         sb->s_xattr = shmem_xattr_handlers;
2615 #endif
2616 #ifdef CONFIG_TMPFS_POSIX_ACL
2617         sb->s_flags |= MS_POSIXACL;
2618 #endif
2619
2620         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2621         if (!inode)
2622                 goto failed;
2623         inode->i_uid = sbinfo->uid;
2624         inode->i_gid = sbinfo->gid;
2625         sb->s_root = d_make_root(inode);
2626         if (!sb->s_root)
2627                 goto failed;
2628         return 0;
2629
2630 failed:
2631         shmem_put_super(sb);
2632         return err;
2633 }
2634
2635 static struct kmem_cache *shmem_inode_cachep;
2636
2637 static struct inode *shmem_alloc_inode(struct super_block *sb)
2638 {
2639         struct shmem_inode_info *info;
2640         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2641         if (!info)
2642                 return NULL;
2643         return &info->vfs_inode;
2644 }
2645
2646 static void shmem_destroy_callback(struct rcu_head *head)
2647 {
2648         struct inode *inode = container_of(head, struct inode, i_rcu);
2649         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2650 }
2651
2652 static void shmem_destroy_inode(struct inode *inode)
2653 {
2654         if (S_ISREG(inode->i_mode))
2655                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2656         call_rcu(&inode->i_rcu, shmem_destroy_callback);
2657 }
2658
2659 static void shmem_init_inode(void *foo)
2660 {
2661         struct shmem_inode_info *info = foo;
2662         inode_init_once(&info->vfs_inode);
2663 }
2664
2665 static int shmem_init_inodecache(void)
2666 {
2667         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2668                                 sizeof(struct shmem_inode_info),
2669                                 0, SLAB_PANIC, shmem_init_inode);
2670         return 0;
2671 }
2672
2673 static void shmem_destroy_inodecache(void)
2674 {
2675         kmem_cache_destroy(shmem_inode_cachep);
2676 }
2677
2678 static const struct address_space_operations shmem_aops = {
2679         .writepage      = shmem_writepage,
2680         .set_page_dirty = __set_page_dirty_no_writeback,
2681 #ifdef CONFIG_TMPFS
2682         .write_begin    = shmem_write_begin,
2683         .write_end      = shmem_write_end,
2684 #endif
2685         .migratepage    = migrate_page,
2686         .error_remove_page = generic_error_remove_page,
2687 };
2688
2689 static const struct file_operations shmem_file_operations = {
2690         .mmap           = shmem_mmap,
2691 #ifdef CONFIG_TMPFS
2692         .llseek         = shmem_file_llseek,
2693         .read           = do_sync_read,
2694         .write          = do_sync_write,
2695         .aio_read       = shmem_file_aio_read,
2696         .aio_write      = generic_file_aio_write,
2697         .fsync          = noop_fsync,
2698         .splice_read    = shmem_file_splice_read,
2699         .splice_write   = generic_file_splice_write,
2700         .fallocate      = shmem_fallocate,
2701 #endif
2702 };
2703
2704 static const struct inode_operations shmem_inode_operations = {
2705         .setattr        = shmem_setattr,
2706 #ifdef CONFIG_TMPFS_XATTR
2707         .setxattr       = shmem_setxattr,
2708         .getxattr       = shmem_getxattr,
2709         .listxattr      = shmem_listxattr,
2710         .removexattr    = shmem_removexattr,
2711 #endif
2712 };
2713
2714 static const struct inode_operations shmem_dir_inode_operations = {
2715 #ifdef CONFIG_TMPFS
2716         .create         = shmem_create,
2717         .lookup         = simple_lookup,
2718         .link           = shmem_link,
2719         .unlink         = shmem_unlink,
2720         .symlink        = shmem_symlink,
2721         .mkdir          = shmem_mkdir,
2722         .rmdir          = shmem_rmdir,
2723         .mknod          = shmem_mknod,
2724         .rename         = shmem_rename,
2725 #endif
2726 #ifdef CONFIG_TMPFS_XATTR
2727         .setxattr       = shmem_setxattr,
2728         .getxattr       = shmem_getxattr,
2729         .listxattr      = shmem_listxattr,
2730         .removexattr    = shmem_removexattr,
2731 #endif
2732 #ifdef CONFIG_TMPFS_POSIX_ACL
2733         .setattr        = shmem_setattr,
2734 #endif
2735 };
2736
2737 static const struct inode_operations shmem_special_inode_operations = {
2738 #ifdef CONFIG_TMPFS_XATTR
2739         .setxattr       = shmem_setxattr,
2740         .getxattr       = shmem_getxattr,
2741         .listxattr      = shmem_listxattr,
2742         .removexattr    = shmem_removexattr,
2743 #endif
2744 #ifdef CONFIG_TMPFS_POSIX_ACL
2745         .setattr        = shmem_setattr,
2746 #endif
2747 };
2748
2749 static const struct super_operations shmem_ops = {
2750         .alloc_inode    = shmem_alloc_inode,
2751         .destroy_inode  = shmem_destroy_inode,
2752 #ifdef CONFIG_TMPFS
2753         .statfs         = shmem_statfs,
2754         .remount_fs     = shmem_remount_fs,
2755         .show_options   = shmem_show_options,
2756 #endif
2757         .evict_inode    = shmem_evict_inode,
2758         .drop_inode     = generic_delete_inode,
2759         .put_super      = shmem_put_super,
2760 };
2761
2762 static const struct vm_operations_struct shmem_vm_ops = {
2763         .fault          = shmem_fault,
2764 #ifdef CONFIG_NUMA
2765         .set_policy     = shmem_set_policy,
2766         .get_policy     = shmem_get_policy,
2767 #endif
2768         .remap_pages    = generic_file_remap_pages,
2769 };
2770
2771 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2772         int flags, const char *dev_name, void *data)
2773 {
2774         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2775 }
2776
2777 static struct file_system_type shmem_fs_type = {
2778         .owner          = THIS_MODULE,
2779         .name           = "tmpfs",
2780         .mount          = shmem_mount,
2781         .kill_sb        = kill_litter_super,
2782         .fs_flags       = FS_USERNS_MOUNT,
2783 };
2784
2785 int __init shmem_init(void)
2786 {
2787         int error;
2788
2789         error = bdi_init(&shmem_backing_dev_info);
2790         if (error)
2791                 goto out4;
2792
2793         error = shmem_init_inodecache();
2794         if (error)
2795                 goto out3;
2796
2797         error = register_filesystem(&shmem_fs_type);
2798         if (error) {
2799                 printk(KERN_ERR "Could not register tmpfs\n");
2800                 goto out2;
2801         }
2802
2803         shm_mnt = vfs_kern_mount(&shmem_fs_type, MS_NOUSER,
2804                                  shmem_fs_type.name, NULL);
2805         if (IS_ERR(shm_mnt)) {
2806                 error = PTR_ERR(shm_mnt);
2807                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2808                 goto out1;
2809         }
2810         return 0;
2811
2812 out1:
2813         unregister_filesystem(&shmem_fs_type);
2814 out2:
2815         shmem_destroy_inodecache();
2816 out3:
2817         bdi_destroy(&shmem_backing_dev_info);
2818 out4:
2819         shm_mnt = ERR_PTR(error);
2820         return error;
2821 }
2822
2823 #else /* !CONFIG_SHMEM */
2824
2825 /*
2826  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2827  *
2828  * This is intended for small system where the benefits of the full
2829  * shmem code (swap-backed and resource-limited) are outweighed by
2830  * their complexity. On systems without swap this code should be
2831  * effectively equivalent, but much lighter weight.
2832  */
2833
2834 static struct file_system_type shmem_fs_type = {
2835         .name           = "tmpfs",
2836         .mount          = ramfs_mount,
2837         .kill_sb        = kill_litter_super,
2838         .fs_flags       = FS_USERNS_MOUNT,
2839 };
2840
2841 int __init shmem_init(void)
2842 {
2843         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
2844
2845         shm_mnt = kern_mount(&shmem_fs_type);
2846         BUG_ON(IS_ERR(shm_mnt));
2847
2848         return 0;
2849 }
2850
2851 int shmem_unuse(swp_entry_t swap, struct page *page)
2852 {
2853         return 0;
2854 }
2855
2856 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2857 {
2858         return 0;
2859 }
2860
2861 void shmem_unlock_mapping(struct address_space *mapping)
2862 {
2863 }
2864
2865 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
2866 {
2867         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
2868 }
2869 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2870
2871 #define shmem_vm_ops                            generic_file_vm_ops
2872 #define shmem_file_operations                   ramfs_file_operations
2873 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2874 #define shmem_acct_size(flags, size)            0
2875 #define shmem_unacct_size(flags, size)          do {} while (0)
2876
2877 #endif /* CONFIG_SHMEM */
2878
2879 /* common code */
2880
2881 static char *shmem_dname(struct dentry *dentry, char *buffer, int buflen)
2882 {
2883         return dynamic_dname(dentry, buffer, buflen, "/%s (deleted)",
2884                                 dentry->d_name.name);
2885 }
2886
2887 static struct dentry_operations anon_ops = {
2888         .d_dname = shmem_dname
2889 };
2890
2891 /**
2892  * shmem_file_setup - get an unlinked file living in tmpfs
2893  * @name: name for dentry (to be seen in /proc/<pid>/maps
2894  * @size: size to be set for the file
2895  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2896  */
2897 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2898 {
2899         struct file *res;
2900         struct inode *inode;
2901         struct path path;
2902         struct super_block *sb;
2903         struct qstr this;
2904
2905         if (IS_ERR(shm_mnt))
2906                 return ERR_CAST(shm_mnt);
2907
2908         if (size < 0 || size > MAX_LFS_FILESIZE)
2909                 return ERR_PTR(-EINVAL);
2910
2911         if (shmem_acct_size(flags, size))
2912                 return ERR_PTR(-ENOMEM);
2913
2914         res = ERR_PTR(-ENOMEM);
2915         this.name = name;
2916         this.len = strlen(name);
2917         this.hash = 0; /* will go */
2918         sb = shm_mnt->mnt_sb;
2919         path.dentry = d_alloc_pseudo(sb, &this);
2920         if (!path.dentry)
2921                 goto put_memory;
2922         d_set_d_op(path.dentry, &anon_ops);
2923         path.mnt = mntget(shm_mnt);
2924
2925         res = ERR_PTR(-ENOSPC);
2926         inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2927         if (!inode)
2928                 goto put_dentry;
2929
2930         d_instantiate(path.dentry, inode);
2931         inode->i_size = size;
2932         clear_nlink(inode);     /* It is unlinked */
2933         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
2934         if (IS_ERR(res))
2935                 goto put_dentry;
2936
2937         res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2938                   &shmem_file_operations);
2939         if (IS_ERR(res))
2940                 goto put_dentry;
2941
2942         return res;
2943
2944 put_dentry:
2945         path_put(&path);
2946 put_memory:
2947         shmem_unacct_size(flags, size);
2948         return res;
2949 }
2950 EXPORT_SYMBOL_GPL(shmem_file_setup);
2951
2952 /**
2953  * shmem_zero_setup - setup a shared anonymous mapping
2954  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2955  */
2956 int shmem_zero_setup(struct vm_area_struct *vma)
2957 {
2958         struct file *file;
2959         loff_t size = vma->vm_end - vma->vm_start;
2960
2961         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2962         if (IS_ERR(file))
2963                 return PTR_ERR(file);
2964
2965         if (vma->vm_file)
2966                 fput(vma->vm_file);
2967         vma->vm_file = file;
2968         vma->vm_ops = &shmem_vm_ops;
2969         return 0;
2970 }
2971
2972 /**
2973  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
2974  * @mapping:    the page's address_space
2975  * @index:      the page index
2976  * @gfp:        the page allocator flags to use if allocating
2977  *
2978  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
2979  * with any new page allocations done using the specified allocation flags.
2980  * But read_cache_page_gfp() uses the ->readpage() method: which does not
2981  * suit tmpfs, since it may have pages in swapcache, and needs to find those
2982  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
2983  *
2984  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
2985  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
2986  */
2987 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
2988                                          pgoff_t index, gfp_t gfp)
2989 {
2990 #ifdef CONFIG_SHMEM
2991         struct inode *inode = mapping->host;
2992         struct page *page;
2993         int error;
2994
2995         BUG_ON(mapping->a_ops != &shmem_aops);
2996         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
2997         if (error)
2998                 page = ERR_PTR(error);
2999         else
3000                 unlock_page(page);
3001         return page;
3002 #else
3003         /*
3004          * The tiny !SHMEM case uses ramfs without swap
3005          */
3006         return read_cache_page_gfp(mapping, index, gfp);
3007 #endif
3008 }
3009 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);