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 #include <linux/aio.h>
  35
  36 static struct vfsmount *shm_mnt;
  37
  38 #ifdef CONFIG_SHMEM
  39 /*
  40  * This virtual memory filesystem is heavily based on the ramfs. It
  41  * extends ramfs by the ability to use swap and honor resource limits
  42  * which makes it a completely usable filesystem.
  43  */
  44
  45 #include <linux/xattr.h>
  46 #include <linux/exportfs.h>
  47 #include <linux/posix_acl.h>
  48 #include <linux/posix_acl_xattr.h>
  49 #include <linux/mman.h>
  50 #include <linux/string.h>
  51 #include <linux/slab.h>
  52 #include <linux/backing-dev.h>
  53 #include <linux/shmem_fs.h>
  54 #include <linux/writeback.h>
  55 #include <linux/blkdev.h>
  56 #include <linux/pagevec.h>
  57 #include <linux/percpu_counter.h>
  58 #include <linux/falloc.h>
  59 #include <linux/splice.h>
  60 #include <linux/security.h>
  61 #include <linux/swapops.h>
  62 #include <linux/mempolicy.h>
  63 #include <linux/namei.h>
  64 #include <linux/ctype.h>
  65 #include <linux/migrate.h>
  66 #include <linux/highmem.h>
  67 #include <linux/seq_file.h>
  68 #include <linux/magic.h>
  69
  70 #include <asm/uaccess.h>
  71 #include <asm/pgtable.h>
  72
  73 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  74 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  75
  76 /* Pretend that each entry is of this size in directory's i_size */
  77 #define BOGO_DIRENT_SIZE 20
  78
  79 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  80 #define SHORT_SYMLINK_LEN 128
  81
  82 /*
  83  * shmem_fallocate and shmem_writepage communicate via inode->i_private
  84  * (with i_mutex making sure that it has only one user at a time):
  85  * we would prefer not to enlarge the shmem inode just for that.
  86  */
  87 struct shmem_falloc {
  88         pgoff_t start;          /* start of range currently being fallocated */
  89         pgoff_t next;           /* the next page offset to be fallocated */
  90         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
  91         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
  92 };
  93
  94 /* Flag allocation requirements to shmem_getpage */
  95 enum sgp_type {
  96         SGP_READ,       /* don't exceed i_size, don't allocate page */
  97         SGP_CACHE,      /* don't exceed i_size, may allocate page */
  98         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
  99         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
 100         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
 101 };
 102
 103 #ifdef CONFIG_TMPFS
 104 static unsigned long shmem_default_max_blocks(void)
 105 {
 106         return totalram_pages / 2;
 107 }
 108
 109 static unsigned long shmem_default_max_inodes(void)
 110 {
 111         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 112 }
 113 #endif
 114
 115 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
 116 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 117                                 struct shmem_inode_info *info, pgoff_t index);
 118 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 119         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
 120
 121 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
 122         struct page **pagep, enum sgp_type sgp, int *fault_type)
 123 {
 124         return shmem_getpage_gfp(inode, index, pagep, sgp,
 125                         mapping_gfp_mask(inode->i_mapping), fault_type);
 126 }
 127
 128 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 129 {
 130         return sb->s_fs_info;
 131 }
 132
 133 /*
 134  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 135  * for shared memory and for shared anonymous (/dev/zero) mappings
 136  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 137  * consistent with the pre-accounting of private mappings ...
 138  */
 139 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 140 {
 141         return (flags & VM_NORESERVE) ?
 142                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 143 }
 144
 145 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 146 {
 147         if (!(flags & VM_NORESERVE))
 148                 vm_unacct_memory(VM_ACCT(size));
 149 }
 150
 151 /*
 152  * ... whereas tmpfs objects are accounted incrementally as
 153  * pages are allocated, in order to allow huge sparse files.
 154  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 155  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 156  */
 157 static inline int shmem_acct_block(unsigned long flags)
 158 {
 159         return (flags & VM_NORESERVE) ?
 160                 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 161 }
 162
 163 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 164 {
 165         if (flags & VM_NORESERVE)
 166                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 167 }
 168
 169 static const struct super_operations shmem_ops;
 170 static const struct address_space_operations shmem_aops;
 171 static const struct file_operations shmem_file_operations;
 172 static const struct inode_operations shmem_inode_operations;
 173 static const struct inode_operations shmem_dir_inode_operations;
 174 static const struct inode_operations shmem_special_inode_operations;
 175 static const struct vm_operations_struct shmem_vm_ops;
 176
 177 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 178         .ra_pages       = 0,    /* No readahead */
 179         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
 180 };
 181
 182 static LIST_HEAD(shmem_swaplist);
 183 static DEFINE_MUTEX(shmem_swaplist_mutex);
 184
 185 static int shmem_reserve_inode(struct super_block *sb)
 186 {
 187         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 188         if (sbinfo->max_inodes) {
 189                 spin_lock(&sbinfo->stat_lock);
 190                 if (!sbinfo->free_inodes) {
 191                         spin_unlock(&sbinfo->stat_lock);
 192                         return -ENOSPC;
 193                 }
 194                 sbinfo->free_inodes--;
 195                 spin_unlock(&sbinfo->stat_lock);
 196         }
 197         return 0;
 198 }
 199
 200 static void shmem_free_inode(struct super_block *sb)
 201 {
 202         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 203         if (sbinfo->max_inodes) {
 204                 spin_lock(&sbinfo->stat_lock);
 205                 sbinfo->free_inodes++;
 206                 spin_unlock(&sbinfo->stat_lock);
 207         }
 208 }
 209
 210 /**
 211  * shmem_recalc_inode - recalculate the block usage of an inode
 212  * @inode: inode to recalc
 213  *
 214  * We have to calculate the free blocks since the mm can drop
 215  * undirtied hole pages behind our back.
 216  *
 217  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 218  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 219  *
 220  * It has to be called with the spinlock held.
 221  */
 222 static void shmem_recalc_inode(struct inode *inode)
 223 {
 224         struct shmem_inode_info *info = SHMEM_I(inode);
 225         long freed;
 226
 227         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 228         if (freed > 0) {
 229                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 230                 if (sbinfo->max_blocks)
 231                         percpu_counter_add(&sbinfo->used_blocks, -freed);
 232                 info->alloced -= freed;
 233                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 234                 shmem_unacct_blocks(info->flags, freed);
 235         }
 236 }
 237
 238 /*
 239  * Replace item expected in radix tree by a new item, while holding tree lock.
 240  */
 241 static int shmem_radix_tree_replace(struct address_space *mapping,
 242                         pgoff_t index, void *expected, void *replacement)
 243 {
 244         void **pslot;
 245         void *item;
 246
 247         VM_BUG_ON(!expected);
 248         VM_BUG_ON(!replacement);
 249         pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
 250         if (!pslot)
 251                 return -ENOENT;
 252         item = radix_tree_deref_slot_protected(pslot, &mapping->tree_lock);
 253         if (item != expected)
 254                 return -ENOENT;
 255         radix_tree_replace_slot(pslot, replacement);
 256         return 0;
 257 }
 258
 259 /*
 260  * Sometimes, before we decide whether to proceed or to fail, we must check
 261  * that an entry was not already brought back from swap by a racing thread.
 262  *
 263  * Checking page is not enough: by the time a SwapCache page is locked, it
 264  * might be reused, and again be SwapCache, using the same swap as before.
 265  */
 266 static bool shmem_confirm_swap(struct address_space *mapping,
 267                                pgoff_t index, swp_entry_t swap)
 268 {
 269         void *item;
 270
 271         rcu_read_lock();
 272         item = radix_tree_lookup(&mapping->page_tree, index);
 273         rcu_read_unlock();
 274         return item == swp_to_radix_entry(swap);
 275 }
 276
 277 /*
 278  * Like add_to_page_cache_locked, but error if expected item has gone.
 279  */
 280 static int shmem_add_to_page_cache(struct page *page,
 281                                    struct address_space *mapping,
 282                                    pgoff_t index, gfp_t gfp, void *expected)
 283 {
 284         int error;
 285
 286         VM_BUG_ON_PAGE(!PageLocked(page), page);
 287         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 288
 289         page_cache_get(page);
 290         page->mapping = mapping;
 291         page->index = index;
 292
 293         spin_lock_irq(&mapping->tree_lock);
 294         if (!expected)
 295                 error = radix_tree_insert(&mapping->page_tree, index, page);
 296         else
 297                 error = shmem_radix_tree_replace(mapping, index, expected,
 298                                                                  page);
 299         if (!error) {
 300                 mapping->nrpages++;
 301                 __inc_zone_page_state(page, NR_FILE_PAGES);
 302                 __inc_zone_page_state(page, NR_SHMEM);
 303                 spin_unlock_irq(&mapping->tree_lock);
 304         } else {
 305                 page->mapping = NULL;
 306                 spin_unlock_irq(&mapping->tree_lock);
 307                 page_cache_release(page);
 308         }
 309         return error;
 310 }
 311
 312 /*
 313  * Like delete_from_page_cache, but substitutes swap for page.
 314  */
 315 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 316 {
 317         struct address_space *mapping = page->mapping;
 318         int error;
 319
 320         spin_lock_irq(&mapping->tree_lock);
 321         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
 322         page->mapping = NULL;
 323         mapping->nrpages--;
 324         __dec_zone_page_state(page, NR_FILE_PAGES);
 325         __dec_zone_page_state(page, NR_SHMEM);
 326         spin_unlock_irq(&mapping->tree_lock);
 327         page_cache_release(page);
 328         BUG_ON(error);
 329 }
 330
 331 /*
 332  * Like find_get_pages, but collecting swap entries as well as pages.
 333  */
 334 static unsigned shmem_find_get_pages_and_swap(struct address_space *mapping,
 335                                         pgoff_t start, unsigned int nr_pages,
 336                                         struct page **pages, pgoff_t *indices)
 337 {
 338         void **slot;
 339         unsigned int ret = 0;
 340         struct radix_tree_iter iter;
 341
 342         if (!nr_pages)
 343                 return 0;
 344
 345         rcu_read_lock();
 346 restart:
 347         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
 348                 struct page *page;
 349 repeat:
 350                 page = radix_tree_deref_slot(slot);
 351                 if (unlikely(!page))
 352                         continue;
 353                 if (radix_tree_exception(page)) {
 354                         if (radix_tree_deref_retry(page))
 355                                 goto restart;
 356                         /*
 357                          * Otherwise, we must be storing a swap entry
 358                          * here as an exceptional entry: so return it
 359                          * without attempting to raise page count.
 360                          */
 361                         goto export;
 362                 }
 363                 if (!page_cache_get_speculative(page))
 364                         goto repeat;
 365
 366                 /* Has the page moved? */
 367                 if (unlikely(page != *slot)) {
 368                         page_cache_release(page);
 369                         goto repeat;
 370                 }
 371 export:
 372                 indices[ret] = iter.index;
 373                 pages[ret] = page;
 374                 if (++ret == nr_pages)
 375                         break;
 376         }
 377         rcu_read_unlock();
 378         return ret;
 379 }
 380
 381 /*
 382  * Remove swap entry from radix tree, free the swap and its page cache.
 383  */
 384 static int shmem_free_swap(struct address_space *mapping,
 385                            pgoff_t index, void *radswap)
 386 {
 387         void *old;
 388
 389         spin_lock_irq(&mapping->tree_lock);
 390         old = radix_tree_delete_item(&mapping->page_tree, index, radswap);
 391         spin_unlock_irq(&mapping->tree_lock);
 392         if (old != radswap)
 393                 return -ENOENT;
 394         free_swap_and_cache(radix_to_swp_entry(radswap));
 395         return 0;
 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_PAGE(PageWriteback(page), 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_PAGE(PageWriteback(page), 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         if (attr->ia_valid & ATTR_MODE)
 623                 error = posix_acl_chmod(inode, inode->i_mode);
 624         return error;
 625 }
 626
 627 static void shmem_evict_inode(struct inode *inode)
 628 {
 629         struct shmem_inode_info *info = SHMEM_I(inode);
 630
 631         if (inode->i_mapping->a_ops == &shmem_aops) {
 632                 shmem_unacct_size(info->flags, inode->i_size);
 633                 inode->i_size = 0;
 634                 shmem_truncate_range(inode, 0, (loff_t)-1);
 635                 if (!list_empty(&info->swaplist)) {
 636                         mutex_lock(&shmem_swaplist_mutex);
 637                         list_del_init(&info->swaplist);
 638                         mutex_unlock(&shmem_swaplist_mutex);
 639                 }
 640         } else
 641                 kfree(info->symlink);
 642
 643         simple_xattrs_free(&info->xattrs);
 644         WARN_ON(inode->i_blocks);
 645         shmem_free_inode(inode->i_sb);
 646         clear_inode(inode);
 647 }
 648
 649 /*
 650  * If swap found in inode, free it and move page from swapcache to filecache.
 651  */
 652 static int shmem_unuse_inode(struct shmem_inode_info *info,
 653                              swp_entry_t swap, struct page **pagep)
 654 {
 655         struct address_space *mapping = info->vfs_inode.i_mapping;
 656         void *radswap;
 657         pgoff_t index;
 658         gfp_t gfp;
 659         int error = 0;
 660
 661         radswap = swp_to_radix_entry(swap);
 662         index = radix_tree_locate_item(&mapping->page_tree, radswap);
 663         if (index == -1)
 664                 return 0;
 665
 666         /*
 667          * Move _head_ to start search for next from here.
 668          * But be careful: shmem_evict_inode checks list_empty without taking
 669          * mutex, and there's an instant in list_move_tail when info->swaplist
 670          * would appear empty, if it were the only one on shmem_swaplist.
 671          */
 672         if (shmem_swaplist.next != &info->swaplist)
 673                 list_move_tail(&shmem_swaplist, &info->swaplist);
 674
 675         gfp = mapping_gfp_mask(mapping);
 676         if (shmem_should_replace_page(*pagep, gfp)) {
 677                 mutex_unlock(&shmem_swaplist_mutex);
 678                 error = shmem_replace_page(pagep, gfp, info, index);
 679                 mutex_lock(&shmem_swaplist_mutex);
 680                 /*
 681                  * We needed to drop mutex to make that restrictive page
 682                  * allocation, but the inode might have been freed while we
 683                  * dropped it: although a racing shmem_evict_inode() cannot
 684                  * complete without emptying the radix_tree, our page lock
 685                  * on this swapcache page is not enough to prevent that -
 686                  * free_swap_and_cache() of our swap entry will only
 687                  * trylock_page(), removing swap from radix_tree whatever.
 688                  *
 689                  * We must not proceed to shmem_add_to_page_cache() if the
 690                  * inode has been freed, but of course we cannot rely on
 691                  * inode or mapping or info to check that.  However, we can
 692                  * safely check if our swap entry is still in use (and here
 693                  * it can't have got reused for another page): if it's still
 694                  * in use, then the inode cannot have been freed yet, and we
 695                  * can safely proceed (if it's no longer in use, that tells
 696                  * nothing about the inode, but we don't need to unuse swap).
 697                  */
 698                 if (!page_swapcount(*pagep))
 699                         error = -ENOENT;
 700         }
 701
 702         /*
 703          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 704          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 705          * beneath us (pagelock doesn't help until the page is in pagecache).
 706          */
 707         if (!error)
 708                 error = shmem_add_to_page_cache(*pagep, mapping, index,
 709                                                 GFP_NOWAIT, radswap);
 710         if (error != -ENOMEM) {
 711                 /*
 712                  * Truncation and eviction use free_swap_and_cache(), which
 713                  * only does trylock page: if we raced, best clean up here.
 714                  */
 715                 delete_from_swap_cache(*pagep);
 716                 set_page_dirty(*pagep);
 717                 if (!error) {
 718                         spin_lock(&info->lock);
 719                         info->swapped--;
 720                         spin_unlock(&info->lock);
 721                         swap_free(swap);
 722                 }
 723                 error = 1;      /* not an error, but entry was found */
 724         }
 725         return error;
 726 }
 727
 728 /*
 729  * Search through swapped inodes to find and replace swap by page.
 730  */
 731 int shmem_unuse(swp_entry_t swap, struct page *page)
 732 {
 733         struct list_head *this, *next;
 734         struct shmem_inode_info *info;
 735         int found = 0;
 736         int error = 0;
 737
 738         /*
 739          * There's a faint possibility that swap page was replaced before
 740          * caller locked it: caller will come back later with the right page.
 741          */
 742         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
 743                 goto out;
 744
 745         /*
 746          * Charge page using GFP_KERNEL while we can wait, before taking
 747          * the shmem_swaplist_mutex which might hold up shmem_writepage().
 748          * Charged back to the user (not to caller) when swap account is used.
 749          */
 750         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
 751         if (error)
 752                 goto out;
 753         /* No radix_tree_preload: swap entry keeps a place for page in tree */
 754
 755         mutex_lock(&shmem_swaplist_mutex);
 756         list_for_each_safe(this, next, &shmem_swaplist) {
 757                 info = list_entry(this, struct shmem_inode_info, swaplist);
 758                 if (info->swapped)
 759                         found = shmem_unuse_inode(info, swap, &page);
 760                 else
 761                         list_del_init(&info->swaplist);
 762                 cond_resched();
 763                 if (found)
 764                         break;
 765         }
 766         mutex_unlock(&shmem_swaplist_mutex);
 767
 768         if (found < 0)
 769                 error = found;
 770 out:
 771         unlock_page(page);
 772         page_cache_release(page);
 773         return error;
 774 }
 775
 776 /*
 777  * Move the page from the page cache to the swap cache.
 778  */
 779 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 780 {
 781         struct shmem_inode_info *info;
 782         struct address_space *mapping;
 783         struct inode *inode;
 784         swp_entry_t swap;
 785         pgoff_t index;
 786
 787         BUG_ON(!PageLocked(page));
 788         mapping = page->mapping;
 789         index = page->index;
 790         inode = mapping->host;
 791         info = SHMEM_I(inode);
 792         if (info->flags & VM_LOCKED)
 793                 goto redirty;
 794         if (!total_swap_pages)
 795                 goto redirty;
 796
 797         /*
 798          * shmem_backing_dev_info's capabilities prevent regular writeback or
 799          * sync from ever calling shmem_writepage; but a stacking filesystem
 800          * might use ->writepage of its underlying filesystem, in which case
 801          * tmpfs should write out to swap only in response to memory pressure,
 802          * and not for the writeback threads or sync.
 803          */
 804         if (!wbc->for_reclaim) {
 805                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
 806                 goto redirty;
 807         }
 808
 809         /*
 810          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
 811          * value into swapfile.c, the only way we can correctly account for a
 812          * fallocated page arriving here is now to initialize it and write it.
 813          *
 814          * That's okay for a page already fallocated earlier, but if we have
 815          * not yet completed the fallocation, then (a) we want to keep track
 816          * of this page in case we have to undo it, and (b) it may not be a
 817          * good idea to continue anyway, once we're pushing into swap.  So
 818          * reactivate the page, and let shmem_fallocate() quit when too many.
 819          */
 820         if (!PageUptodate(page)) {
 821                 if (inode->i_private) {
 822                         struct shmem_falloc *shmem_falloc;
 823                         spin_lock(&inode->i_lock);
 824                         shmem_falloc = inode->i_private;
 825                         if (shmem_falloc &&
 826                             index >= shmem_falloc->start &&
 827                             index < shmem_falloc->next)
 828                                 shmem_falloc->nr_unswapped++;
 829                         else
 830                                 shmem_falloc = NULL;
 831                         spin_unlock(&inode->i_lock);
 832                         if (shmem_falloc)
 833                                 goto redirty;
 834                 }
 835                 clear_highpage(page);
 836                 flush_dcache_page(page);
 837                 SetPageUptodate(page);
 838         }
 839
 840         swap = get_swap_page();
 841         if (!swap.val)
 842                 goto redirty;
 843
 844         /*
 845          * Add inode to shmem_unuse()'s list of swapped-out inodes,
 846          * if it's not already there.  Do it now before the page is
 847          * moved to swap cache, when its pagelock no longer protects
 848          * the inode from eviction.  But don't unlock the mutex until
 849          * we've incremented swapped, because shmem_unuse_inode() will
 850          * prune a !swapped inode from the swaplist under this mutex.
 851          */
 852         mutex_lock(&shmem_swaplist_mutex);
 853         if (list_empty(&info->swaplist))
 854                 list_add_tail(&info->swaplist, &shmem_swaplist);
 855
 856         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
 857                 swap_shmem_alloc(swap);
 858                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
 859
 860                 spin_lock(&info->lock);
 861                 info->swapped++;
 862                 shmem_recalc_inode(inode);
 863                 spin_unlock(&info->lock);
 864
 865                 mutex_unlock(&shmem_swaplist_mutex);
 866                 BUG_ON(page_mapped(page));
 867                 swap_writepage(page, wbc);
 868                 return 0;
 869         }
 870
 871         mutex_unlock(&shmem_swaplist_mutex);
 872         swapcache_free(swap, NULL);
 873 redirty:
 874         set_page_dirty(page);
 875         if (wbc->for_reclaim)
 876                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
 877         unlock_page(page);
 878         return 0;
 879 }
 880
 881 #ifdef CONFIG_NUMA
 882 #ifdef CONFIG_TMPFS
 883 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 884 {
 885         char buffer[64];
 886
 887         if (!mpol || mpol->mode == MPOL_DEFAULT)
 888                 return;         /* show nothing */
 889
 890         mpol_to_str(buffer, sizeof(buffer), mpol);
 891
 892         seq_printf(seq, ",mpol=%s", buffer);
 893 }
 894
 895 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 896 {
 897         struct mempolicy *mpol = NULL;
 898         if (sbinfo->mpol) {
 899                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
 900                 mpol = sbinfo->mpol;
 901                 mpol_get(mpol);
 902                 spin_unlock(&sbinfo->stat_lock);
 903         }
 904         return mpol;
 905 }
 906 #endif /* CONFIG_TMPFS */
 907
 908 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 909                         struct shmem_inode_info *info, pgoff_t index)
 910 {
 911         struct vm_area_struct pvma;
 912         struct page *page;
 913
 914         /* Create a pseudo vma that just contains the policy */
 915         pvma.vm_start = 0;
 916         /* Bias interleave by inode number to distribute better across nodes */
 917         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 918         pvma.vm_ops = NULL;
 919         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 920
 921         page = swapin_readahead(swap, gfp, &pvma, 0);
 922
 923         /* Drop reference taken by mpol_shared_policy_lookup() */
 924         mpol_cond_put(pvma.vm_policy);
 925
 926         return page;
 927 }
 928
 929 static struct page *shmem_alloc_page(gfp_t gfp,
 930                         struct shmem_inode_info *info, pgoff_t index)
 931 {
 932         struct vm_area_struct pvma;
 933         struct page *page;
 934
 935         /* Create a pseudo vma that just contains the policy */
 936         pvma.vm_start = 0;
 937         /* Bias interleave by inode number to distribute better across nodes */
 938         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 939         pvma.vm_ops = NULL;
 940         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 941
 942         page = alloc_page_vma(gfp, &pvma, 0);
 943
 944         /* Drop reference taken by mpol_shared_policy_lookup() */
 945         mpol_cond_put(pvma.vm_policy);
 946
 947         return page;
 948 }
 949 #else /* !CONFIG_NUMA */
 950 #ifdef CONFIG_TMPFS
 951 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 952 {
 953 }
 954 #endif /* CONFIG_TMPFS */
 955
 956 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 957                         struct shmem_inode_info *info, pgoff_t index)
 958 {
 959         return swapin_readahead(swap, gfp, NULL, 0);
 960 }
 961
 962 static inline struct page *shmem_alloc_page(gfp_t gfp,
 963                         struct shmem_inode_info *info, pgoff_t index)
 964 {
 965         return alloc_page(gfp);
 966 }
 967 #endif /* CONFIG_NUMA */
 968
 969 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
 970 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 971 {
 972         return NULL;
 973 }
 974 #endif
 975
 976 /*
 977  * When a page is moved from swapcache to shmem filecache (either by the
 978  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
 979  * shmem_unuse_inode()), it may have been read in earlier from swap, in
 980  * ignorance of the mapping it belongs to.  If that mapping has special
 981  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
 982  * we may need to copy to a suitable page before moving to filecache.
 983  *
 984  * In a future release, this may well be extended to respect cpuset and
 985  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
 986  * but for now it is a simple matter of zone.
 987  */
 988 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
 989 {
 990         return page_zonenum(page) > gfp_zone(gfp);
 991 }
 992
 993 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 994                                 struct shmem_inode_info *info, pgoff_t index)
 995 {
 996         struct page *oldpage, *newpage;
 997         struct address_space *swap_mapping;
 998         pgoff_t swap_index;
 999         int error;
1000
1001         oldpage = *pagep;
1002         swap_index = page_private(oldpage);
1003         swap_mapping = page_mapping(oldpage);
1004
1005         /*
1006          * We have arrived here because our zones are constrained, so don't
1007          * limit chance of success by further cpuset and node constraints.
1008          */
1009         gfp &= ~GFP_CONSTRAINT_MASK;
1010         newpage = shmem_alloc_page(gfp, info, index);
1011         if (!newpage)
1012                 return -ENOMEM;
1013
1014         page_cache_get(newpage);
1015         copy_highpage(newpage, oldpage);
1016         flush_dcache_page(newpage);
1017
1018         __set_page_locked(newpage);
1019         SetPageUptodate(newpage);
1020         SetPageSwapBacked(newpage);
1021         set_page_private(newpage, swap_index);
1022         SetPageSwapCache(newpage);
1023
1024         /*
1025          * Our caller will very soon move newpage out of swapcache, but it's
1026          * a nice clean interface for us to replace oldpage by newpage there.
1027          */
1028         spin_lock_irq(&swap_mapping->tree_lock);
1029         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1030                                                                    newpage);
1031         if (!error) {
1032                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1033                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1034         }
1035         spin_unlock_irq(&swap_mapping->tree_lock);
1036
1037         if (unlikely(error)) {
1038                 /*
1039                  * Is this possible?  I think not, now that our callers check
1040                  * both PageSwapCache and page_private after getting page lock;
1041                  * but be defensive.  Reverse old to newpage for clear and free.
1042                  */
1043                 oldpage = newpage;
1044         } else {
1045                 mem_cgroup_replace_page_cache(oldpage, newpage);
1046                 lru_cache_add_anon(newpage);
1047                 *pagep = newpage;
1048         }
1049
1050         ClearPageSwapCache(oldpage);
1051         set_page_private(oldpage, 0);
1052
1053         unlock_page(oldpage);
1054         page_cache_release(oldpage);
1055         page_cache_release(oldpage);
1056         return error;
1057 }
1058
1059 /*
1060  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1061  *
1062  * If we allocate a new one we do not mark it dirty. That's up to the
1063  * vm. If we swap it in we mark it dirty since we also free the swap
1064  * entry since a page cannot live in both the swap and page cache
1065  */
1066 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1067         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1068 {
1069         struct address_space *mapping = inode->i_mapping;
1070         struct shmem_inode_info *info;
1071         struct shmem_sb_info *sbinfo;
1072         struct page *page;
1073         swp_entry_t swap;
1074         int error;
1075         int once = 0;
1076         int alloced = 0;
1077
1078         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1079                 return -EFBIG;
1080 repeat:
1081         swap.val = 0;
1082         page = find_lock_page(mapping, index);
1083         if (radix_tree_exceptional_entry(page)) {
1084                 swap = radix_to_swp_entry(page);
1085                 page = NULL;
1086         }
1087
1088         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1089             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1090                 error = -EINVAL;
1091                 goto failed;
1092         }
1093
1094         /* fallocated page? */
1095         if (page && !PageUptodate(page)) {
1096                 if (sgp != SGP_READ)
1097                         goto clear;
1098                 unlock_page(page);
1099                 page_cache_release(page);
1100                 page = NULL;
1101         }
1102         if (page || (sgp == SGP_READ && !swap.val)) {
1103                 *pagep = page;
1104                 return 0;
1105         }
1106
1107         /*
1108          * Fast cache lookup did not find it:
1109          * bring it back from swap or allocate.
1110          */
1111         info = SHMEM_I(inode);
1112         sbinfo = SHMEM_SB(inode->i_sb);
1113
1114         if (swap.val) {
1115                 /* Look it up and read it in.. */
1116                 page = lookup_swap_cache(swap);
1117                 if (!page) {
1118                         /* here we actually do the io */
1119                         if (fault_type)
1120                                 *fault_type |= VM_FAULT_MAJOR;
1121                         page = shmem_swapin(swap, gfp, info, index);
1122                         if (!page) {
1123                                 error = -ENOMEM;
1124                                 goto failed;
1125                         }
1126                 }
1127
1128                 /* We have to do this with page locked to prevent races */
1129                 lock_page(page);
1130                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1131                     !shmem_confirm_swap(mapping, index, swap)) {
1132                         error = -EEXIST;        /* try again */
1133                         goto unlock;
1134                 }
1135                 if (!PageUptodate(page)) {
1136                         error = -EIO;
1137                         goto failed;
1138                 }
1139                 wait_on_page_writeback(page);
1140
1141                 if (shmem_should_replace_page(page, gfp)) {
1142                         error = shmem_replace_page(&page, gfp, info, index);
1143                         if (error)
1144                                 goto failed;
1145                 }
1146
1147                 error = mem_cgroup_cache_charge(page, current->mm,
1148                                                 gfp & GFP_RECLAIM_MASK);
1149                 if (!error) {
1150                         error = shmem_add_to_page_cache(page, mapping, index,
1151                                                 gfp, swp_to_radix_entry(swap));
1152                         /*
1153                          * We already confirmed swap under page lock, and make
1154                          * no memory allocation here, so usually no possibility
1155                          * of error; but free_swap_and_cache() only trylocks a
1156                          * page, so it is just possible that the entry has been
1157                          * truncated or holepunched since swap was confirmed.
1158                          * shmem_undo_range() will have done some of the
1159                          * unaccounting, now delete_from_swap_cache() will do
1160                          * the rest (including mem_cgroup_uncharge_swapcache).
1161                          * Reset swap.val? No, leave it so "failed" goes back to
1162                          * "repeat": reading a hole and writing should succeed.
1163                          */
1164                         if (error)
1165                                 delete_from_swap_cache(page);
1166                 }
1167                 if (error)
1168                         goto failed;
1169
1170                 spin_lock(&info->lock);
1171                 info->swapped--;
1172                 shmem_recalc_inode(inode);
1173                 spin_unlock(&info->lock);
1174
1175                 delete_from_swap_cache(page);
1176                 set_page_dirty(page);
1177                 swap_free(swap);
1178
1179         } else {
1180                 if (shmem_acct_block(info->flags)) {
1181                         error = -ENOSPC;
1182                         goto failed;
1183                 }
1184                 if (sbinfo->max_blocks) {
1185                         if (percpu_counter_compare(&sbinfo->used_blocks,
1186                                                 sbinfo->max_blocks) >= 0) {
1187                                 error = -ENOSPC;
1188                                 goto unacct;
1189                         }
1190                         percpu_counter_inc(&sbinfo->used_blocks);
1191                 }
1192
1193                 page = shmem_alloc_page(gfp, info, index);
1194                 if (!page) {
1195                         error = -ENOMEM;
1196                         goto decused;
1197                 }
1198
1199                 SetPageSwapBacked(page);
1200                 __set_page_locked(page);
1201                 error = mem_cgroup_cache_charge(page, current->mm,
1202                                                 gfp & GFP_RECLAIM_MASK);
1203                 if (error)
1204                         goto decused;
1205                 error = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
1206                 if (!error) {
1207                         error = shmem_add_to_page_cache(page, mapping, index,
1208                                                         gfp, NULL);
1209                         radix_tree_preload_end();
1210                 }
1211                 if (error) {
1212                         mem_cgroup_uncharge_cache_page(page);
1213                         goto decused;
1214                 }
1215                 lru_cache_add_anon(page);
1216
1217                 spin_lock(&info->lock);
1218                 info->alloced++;
1219                 inode->i_blocks += BLOCKS_PER_PAGE;
1220                 shmem_recalc_inode(inode);
1221                 spin_unlock(&info->lock);
1222                 alloced = true;
1223
1224                 /*
1225                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1226                  */
1227                 if (sgp == SGP_FALLOC)
1228                         sgp = SGP_WRITE;
1229 clear:
1230                 /*
1231                  * Let SGP_WRITE caller clear ends if write does not fill page;
1232                  * but SGP_FALLOC on a page fallocated earlier must initialize
1233                  * it now, lest undo on failure cancel our earlier guarantee.
1234                  */
1235                 if (sgp != SGP_WRITE) {
1236                         clear_highpage(page);
1237                         flush_dcache_page(page);
1238                         SetPageUptodate(page);
1239                 }
1240                 if (sgp == SGP_DIRTY)
1241                         set_page_dirty(page);
1242         }
1243
1244         /* Perhaps the file has been truncated since we checked */
1245         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1246             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1247                 error = -EINVAL;
1248                 if (alloced)
1249                         goto trunc;
1250                 else
1251                         goto failed;
1252         }
1253         *pagep = page;
1254         return 0;
1255
1256         /*
1257          * Error recovery.
1258          */
1259 trunc:
1260         info = SHMEM_I(inode);
1261         ClearPageDirty(page);
1262         delete_from_page_cache(page);
1263         spin_lock(&info->lock);
1264         info->alloced--;
1265         inode->i_blocks -= BLOCKS_PER_PAGE;
1266         spin_unlock(&info->lock);
1267 decused:
1268         sbinfo = SHMEM_SB(inode->i_sb);
1269         if (sbinfo->max_blocks)
1270                 percpu_counter_add(&sbinfo->used_blocks, -1);
1271 unacct:
1272         shmem_unacct_blocks(info->flags, 1);
1273 failed:
1274         if (swap.val && error != -EINVAL &&
1275             !shmem_confirm_swap(mapping, index, swap))
1276                 error = -EEXIST;
1277 unlock:
1278         if (page) {
1279                 unlock_page(page);
1280                 page_cache_release(page);
1281         }
1282         if (error == -ENOSPC && !once++) {
1283                 info = SHMEM_I(inode);
1284                 spin_lock(&info->lock);
1285                 shmem_recalc_inode(inode);
1286                 spin_unlock(&info->lock);
1287                 goto repeat;
1288         }
1289         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1290                 goto repeat;
1291         return error;
1292 }
1293
1294 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1295 {
1296         struct inode *inode = file_inode(vma->vm_file);
1297         int error;
1298         int ret = VM_FAULT_LOCKED;
1299
1300         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1301         if (error)
1302                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1303
1304         if (ret & VM_FAULT_MAJOR) {
1305                 count_vm_event(PGMAJFAULT);
1306                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1307         }
1308         return ret;
1309 }
1310
1311 #ifdef CONFIG_NUMA
1312 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1313 {
1314         struct inode *inode = file_inode(vma->vm_file);
1315         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1316 }
1317
1318 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1319                                           unsigned long addr)
1320 {
1321         struct inode *inode = file_inode(vma->vm_file);
1322         pgoff_t index;
1323
1324         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1325         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1326 }
1327 #endif
1328
1329 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1330 {
1331         struct inode *inode = file_inode(file);
1332         struct shmem_inode_info *info = SHMEM_I(inode);
1333         int retval = -ENOMEM;
1334
1335         spin_lock(&info->lock);
1336         if (lock && !(info->flags & VM_LOCKED)) {
1337                 if (!user_shm_lock(inode->i_size, user))
1338                         goto out_nomem;
1339                 info->flags |= VM_LOCKED;
1340                 mapping_set_unevictable(file->f_mapping);
1341         }
1342         if (!lock && (info->flags & VM_LOCKED) && user) {
1343                 user_shm_unlock(inode->i_size, user);
1344                 info->flags &= ~VM_LOCKED;
1345                 mapping_clear_unevictable(file->f_mapping);
1346         }
1347         retval = 0;
1348
1349 out_nomem:
1350         spin_unlock(&info->lock);
1351         return retval;
1352 }
1353
1354 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1355 {
1356         file_accessed(file);
1357         vma->vm_ops = &shmem_vm_ops;
1358         return 0;
1359 }
1360
1361 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1362                                      umode_t mode, dev_t dev, unsigned long flags)
1363 {
1364         struct inode *inode;
1365         struct shmem_inode_info *info;
1366         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1367
1368         if (shmem_reserve_inode(sb))
1369                 return NULL;
1370
1371         inode = new_inode(sb);
1372         if (inode) {
1373                 inode->i_ino = get_next_ino();
1374                 inode_init_owner(inode, dir, mode);
1375                 inode->i_blocks = 0;
1376                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1377                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1378                 inode->i_generation = get_seconds();
1379                 info = SHMEM_I(inode);
1380                 memset(info, 0, (char *)inode - (char *)info);
1381                 spin_lock_init(&info->lock);
1382                 info->flags = flags & VM_NORESERVE;
1383                 INIT_LIST_HEAD(&info->swaplist);
1384                 simple_xattrs_init(&info->xattrs);
1385                 cache_no_acl(inode);
1386
1387                 switch (mode & S_IFMT) {
1388                 default:
1389                         inode->i_op = &shmem_special_inode_operations;
1390                         init_special_inode(inode, mode, dev);
1391                         break;
1392                 case S_IFREG:
1393                         inode->i_mapping->a_ops = &shmem_aops;
1394                         inode->i_op = &shmem_inode_operations;
1395                         inode->i_fop = &shmem_file_operations;
1396                         mpol_shared_policy_init(&info->policy,
1397                                                  shmem_get_sbmpol(sbinfo));
1398                         break;
1399                 case S_IFDIR:
1400                         inc_nlink(inode);
1401                         /* Some things misbehave if size == 0 on a directory */
1402                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1403                         inode->i_op = &shmem_dir_inode_operations;
1404                         inode->i_fop = &simple_dir_operations;
1405                         break;
1406                 case S_IFLNK:
1407                         /*
1408                          * Must not load anything in the rbtree,
1409                          * mpol_free_shared_policy will not be called.
1410                          */
1411                         mpol_shared_policy_init(&info->policy, NULL);
1412                         break;
1413                 }
1414         } else
1415                 shmem_free_inode(sb);
1416         return inode;
1417 }
1418
1419 #ifdef CONFIG_TMPFS
1420 static const struct inode_operations shmem_symlink_inode_operations;
1421 static const struct inode_operations shmem_short_symlink_operations;
1422
1423 #ifdef CONFIG_TMPFS_XATTR
1424 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1425 #else
1426 #define shmem_initxattrs NULL
1427 #endif
1428
1429 static int
1430 shmem_write_begin(struct file *file, struct address_space *mapping,
1431                         loff_t pos, unsigned len, unsigned flags,
1432                         struct page **pagep, void **fsdata)
1433 {
1434         struct inode *inode = mapping->host;
1435         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1436         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1437 }
1438
1439 static int
1440 shmem_write_end(struct file *file, struct address_space *mapping,
1441                         loff_t pos, unsigned len, unsigned copied,
1442                         struct page *page, void *fsdata)
1443 {
1444         struct inode *inode = mapping->host;
1445
1446         if (pos + copied > inode->i_size)
1447                 i_size_write(inode, pos + copied);
1448
1449         if (!PageUptodate(page)) {
1450                 if (copied < PAGE_CACHE_SIZE) {
1451                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1452                         zero_user_segments(page, 0, from,
1453                                         from + copied, PAGE_CACHE_SIZE);
1454                 }
1455                 SetPageUptodate(page);
1456         }
1457         set_page_dirty(page);
1458         unlock_page(page);
1459         page_cache_release(page);
1460
1461         return copied;
1462 }
1463
1464 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1465 {
1466         struct inode *inode = file_inode(filp);
1467         struct address_space *mapping = inode->i_mapping;
1468         pgoff_t index;
1469         unsigned long offset;
1470         enum sgp_type sgp = SGP_READ;
1471
1472         /*
1473          * Might this read be for a stacking filesystem?  Then when reading
1474          * holes of a sparse file, we actually need to allocate those pages,
1475          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1476          */
1477         if (segment_eq(get_fs(), KERNEL_DS))
1478                 sgp = SGP_DIRTY;
1479
1480         index = *ppos >> PAGE_CACHE_SHIFT;
1481         offset = *ppos & ~PAGE_CACHE_MASK;
1482
1483         for (;;) {
1484                 struct page *page = NULL;
1485                 pgoff_t end_index;
1486                 unsigned long nr, ret;
1487                 loff_t i_size = i_size_read(inode);
1488
1489                 end_index = i_size >> PAGE_CACHE_SHIFT;
1490                 if (index > end_index)
1491                         break;
1492                 if (index == end_index) {
1493                         nr = i_size & ~PAGE_CACHE_MASK;
1494                         if (nr <= offset)
1495                                 break;
1496                 }
1497
1498                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1499                 if (desc->error) {
1500                         if (desc->error == -EINVAL)
1501                                 desc->error = 0;
1502                         break;
1503                 }
1504                 if (page)
1505                         unlock_page(page);
1506
1507                 /*
1508                  * We must evaluate after, since reads (unlike writes)
1509                  * are called without i_mutex protection against truncate
1510                  */
1511                 nr = PAGE_CACHE_SIZE;
1512                 i_size = i_size_read(inode);
1513                 end_index = i_size >> PAGE_CACHE_SHIFT;
1514                 if (index == end_index) {
1515                         nr = i_size & ~PAGE_CACHE_MASK;
1516                         if (nr <= offset) {
1517                                 if (page)
1518                                         page_cache_release(page);
1519                                 break;
1520                         }
1521                 }
1522                 nr -= offset;
1523
1524                 if (page) {
1525                         /*
1526                          * If users can be writing to this page using arbitrary
1527                          * virtual addresses, take care about potential aliasing
1528                          * before reading the page on the kernel side.
1529                          */
1530                         if (mapping_writably_mapped(mapping))
1531                                 flush_dcache_page(page);
1532                         /*
1533                          * Mark the page accessed if we read the beginning.
1534                          */
1535                         if (!offset)
1536                                 mark_page_accessed(page);
1537                 } else {
1538                         page = ZERO_PAGE(0);
1539                         page_cache_get(page);
1540                 }
1541
1542                 /*
1543                  * Ok, we have the page, and it's up-to-date, so
1544                  * now we can copy it to user space...
1545                  *
1546                  * The actor routine returns how many bytes were actually used..
1547                  * NOTE! This may not be the same as how much of a user buffer
1548                  * we filled up (we may be padding etc), so we can only update
1549                  * "pos" here (the actor routine has to update the user buffer
1550                  * pointers and the remaining count).
1551                  */
1552                 ret = actor(desc, page, offset, nr);
1553                 offset += ret;
1554                 index += offset >> PAGE_CACHE_SHIFT;
1555                 offset &= ~PAGE_CACHE_MASK;
1556
1557                 page_cache_release(page);
1558                 if (ret != nr || !desc->count)
1559                         break;
1560
1561                 cond_resched();
1562         }
1563
1564         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1565         file_accessed(filp);
1566 }
1567
1568 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1569                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1570 {
1571         struct file *filp = iocb->ki_filp;
1572         ssize_t retval;
1573         unsigned long seg;
1574         size_t count;
1575         loff_t *ppos = &iocb->ki_pos;
1576
1577         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1578         if (retval)
1579                 return retval;
1580
1581         for (seg = 0; seg < nr_segs; seg++) {
1582                 read_descriptor_t desc;
1583
1584                 desc.written = 0;
1585                 desc.arg.buf = iov[seg].iov_base;
1586                 desc.count = iov[seg].iov_len;
1587                 if (desc.count == 0)
1588                         continue;
1589                 desc.error = 0;
1590                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1591                 retval += desc.written;
1592                 if (desc.error) {
1593                         retval = retval ?: desc.error;
1594                         break;
1595                 }
1596                 if (desc.count > 0)
1597                         break;
1598         }
1599         return retval;
1600 }
1601
1602 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1603                                 struct pipe_inode_info *pipe, size_t len,
1604                                 unsigned int flags)
1605 {
1606         struct address_space *mapping = in->f_mapping;
1607         struct inode *inode = mapping->host;
1608         unsigned int loff, nr_pages, req_pages;
1609         struct page *pages[PIPE_DEF_BUFFERS];
1610         struct partial_page partial[PIPE_DEF_BUFFERS];
1611         struct page *page;
1612         pgoff_t index, end_index;
1613         loff_t isize, left;
1614         int error, page_nr;
1615         struct splice_pipe_desc spd = {
1616                 .pages = pages,
1617                 .partial = partial,
1618                 .nr_pages_max = PIPE_DEF_BUFFERS,
1619                 .flags = flags,
1620                 .ops = &page_cache_pipe_buf_ops,
1621                 .spd_release = spd_release_page,
1622         };
1623
1624         isize = i_size_read(inode);
1625         if (unlikely(*ppos >= isize))
1626                 return 0;
1627
1628         left = isize - *ppos;
1629         if (unlikely(left < len))
1630                 len = left;
1631
1632         if (splice_grow_spd(pipe, &spd))
1633                 return -ENOMEM;
1634
1635         index = *ppos >> PAGE_CACHE_SHIFT;
1636         loff = *ppos & ~PAGE_CACHE_MASK;
1637         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1638         nr_pages = min(req_pages, pipe->buffers);
1639
1640         spd.nr_pages = find_get_pages_contig(mapping, index,
1641                                                 nr_pages, spd.pages);
1642         index += spd.nr_pages;
1643         error = 0;
1644
1645         while (spd.nr_pages < nr_pages) {
1646                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1647                 if (error)
1648                         break;
1649                 unlock_page(page);
1650                 spd.pages[spd.nr_pages++] = page;
1651                 index++;
1652         }
1653
1654         index = *ppos >> PAGE_CACHE_SHIFT;
1655         nr_pages = spd.nr_pages;
1656         spd.nr_pages = 0;
1657
1658         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1659                 unsigned int this_len;
1660
1661                 if (!len)
1662                         break;
1663
1664                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1665                 page = spd.pages[page_nr];
1666
1667                 if (!PageUptodate(page) || page->mapping != mapping) {
1668                         error = shmem_getpage(inode, index, &page,
1669                                                         SGP_CACHE, NULL);
1670                         if (error)
1671                                 break;
1672                         unlock_page(page);
1673                         page_cache_release(spd.pages[page_nr]);
1674                         spd.pages[page_nr] = page;
1675                 }
1676
1677                 isize = i_size_read(inode);
1678                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1679                 if (unlikely(!isize || index > end_index))
1680                         break;
1681
1682                 if (end_index == index) {
1683                         unsigned int plen;
1684
1685                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1686                         if (plen <= loff)
1687                                 break;
1688
1689                         this_len = min(this_len, plen - loff);
1690                         len = this_len;
1691                 }
1692
1693                 spd.partial[page_nr].offset = loff;
1694                 spd.partial[page_nr].len = this_len;
1695                 len -= this_len;
1696                 loff = 0;
1697                 spd.nr_pages++;
1698                 index++;
1699         }
1700
1701         while (page_nr < nr_pages)
1702                 page_cache_release(spd.pages[page_nr++]);
1703
1704         if (spd.nr_pages)
1705                 error = splice_to_pipe(pipe, &spd);
1706
1707         splice_shrink_spd(&spd);
1708
1709         if (error > 0) {
1710                 *ppos += error;
1711                 file_accessed(in);
1712         }
1713         return error;
1714 }
1715
1716 /*
1717  * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
1718  */
1719 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
1720                                     pgoff_t index, pgoff_t end, int whence)
1721 {
1722         struct page *page;
1723         struct pagevec pvec;
1724         pgoff_t indices[PAGEVEC_SIZE];
1725         bool done = false;
1726         int i;
1727
1728         pagevec_init(&pvec, 0);
1729         pvec.nr = 1;            /* start small: we may be there already */
1730         while (!done) {
1731                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
1732                                         pvec.nr, pvec.pages, indices);
1733                 if (!pvec.nr) {
1734                         if (whence == SEEK_DATA)
1735                                 index = end;
1736                         break;
1737                 }
1738                 for (i = 0; i < pvec.nr; i++, index++) {
1739                         if (index < indices[i]) {
1740                                 if (whence == SEEK_HOLE) {
1741                                         done = true;
1742                                         break;
1743                                 }
1744                                 index = indices[i];
1745                         }
1746                         page = pvec.pages[i];
1747                         if (page && !radix_tree_exceptional_entry(page)) {
1748                                 if (!PageUptodate(page))
1749                                         page = NULL;
1750                         }
1751                         if (index >= end ||
1752                             (page && whence == SEEK_DATA) ||
1753                             (!page && whence == SEEK_HOLE)) {
1754                                 done = true;
1755                                 break;
1756                         }
1757                 }
1758                 shmem_deswap_pagevec(&pvec);
1759                 pagevec_release(&pvec);
1760                 pvec.nr = PAGEVEC_SIZE;
1761                 cond_resched();
1762         }
1763         return index;
1764 }
1765
1766 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
1767 {
1768         struct address_space *mapping = file->f_mapping;
1769         struct inode *inode = mapping->host;
1770         pgoff_t start, end;
1771         loff_t new_offset;
1772
1773         if (whence != SEEK_DATA && whence != SEEK_HOLE)
1774                 return generic_file_llseek_size(file, offset, whence,
1775                                         MAX_LFS_FILESIZE, i_size_read(inode));
1776         mutex_lock(&inode->i_mutex);
1777         /* We're holding i_mutex so we can access i_size directly */
1778
1779         if (offset < 0)
1780                 offset = -EINVAL;
1781         else if (offset >= inode->i_size)
1782                 offset = -ENXIO;
1783         else {
1784                 start = offset >> PAGE_CACHE_SHIFT;
1785                 end = (inode->i_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1786                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
1787                 new_offset <<= PAGE_CACHE_SHIFT;
1788                 if (new_offset > offset) {
1789                         if (new_offset < inode->i_size)
1790                                 offset = new_offset;
1791                         else if (whence == SEEK_DATA)
1792                                 offset = -ENXIO;
1793                         else
1794                                 offset = inode->i_size;
1795                 }
1796         }
1797
1798         if (offset >= 0)
1799                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
1800         mutex_unlock(&inode->i_mutex);
1801         return offset;
1802 }
1803
1804 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
1805                                                          loff_t len)
1806 {
1807         struct inode *inode = file_inode(file);
1808         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1809         struct shmem_falloc shmem_falloc;
1810         pgoff_t start, index, end;
1811         int error;
1812
1813         mutex_lock(&inode->i_mutex);
1814
1815         if (mode & FALLOC_FL_PUNCH_HOLE) {
1816                 struct address_space *mapping = file->f_mapping;
1817                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
1818                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
1819
1820                 if ((u64)unmap_end > (u64)unmap_start)
1821                         unmap_mapping_range(mapping, unmap_start,
1822                                             1 + unmap_end - unmap_start, 0);
1823                 shmem_truncate_range(inode, offset, offset + len - 1);
1824                 /* No need to unmap again: hole-punching leaves COWed pages */
1825                 error = 0;
1826                 goto out;
1827         }
1828
1829         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
1830         error = inode_newsize_ok(inode, offset + len);
1831         if (error)
1832                 goto out;
1833
1834         start = offset >> PAGE_CACHE_SHIFT;
1835         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1836         /* Try to avoid a swapstorm if len is impossible to satisfy */
1837         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
1838                 error = -ENOSPC;
1839                 goto out;
1840         }
1841
1842         shmem_falloc.start = start;
1843         shmem_falloc.next  = start;
1844         shmem_falloc.nr_falloced = 0;
1845         shmem_falloc.nr_unswapped = 0;
1846         spin_lock(&inode->i_lock);
1847         inode->i_private = &shmem_falloc;
1848         spin_unlock(&inode->i_lock);
1849
1850         for (index = start; index < end; index++) {
1851                 struct page *page;
1852
1853                 /*
1854                  * Good, the fallocate(2) manpage permits EINTR: we may have
1855                  * been interrupted because we are using up too much memory.
1856                  */
1857                 if (signal_pending(current))
1858                         error = -EINTR;
1859                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
1860                         error = -ENOMEM;
1861                 else
1862                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
1863                                                                         NULL);
1864                 if (error) {
1865                         /* Remove the !PageUptodate pages we added */
1866                         shmem_undo_range(inode,
1867                                 (loff_t)start << PAGE_CACHE_SHIFT,
1868                                 (loff_t)index << PAGE_CACHE_SHIFT, true);
1869                         goto undone;
1870                 }
1871
1872                 /*
1873                  * Inform shmem_writepage() how far we have reached.
1874                  * No need for lock or barrier: we have the page lock.
1875                  */
1876                 shmem_falloc.next++;
1877                 if (!PageUptodate(page))
1878                         shmem_falloc.nr_falloced++;
1879
1880                 /*
1881                  * If !PageUptodate, leave it that way so that freeable pages
1882                  * can be recognized if we need to rollback on error later.
1883                  * But set_page_dirty so that memory pressure will swap rather
1884                  * than free the pages we are allocating (and SGP_CACHE pages
1885                  * might still be clean: we now need to mark those dirty too).
1886                  */
1887                 set_page_dirty(page);
1888                 unlock_page(page);
1889                 page_cache_release(page);
1890                 cond_resched();
1891         }
1892
1893         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
1894                 i_size_write(inode, offset + len);
1895         inode->i_ctime = CURRENT_TIME;
1896 undone:
1897         spin_lock(&inode->i_lock);
1898         inode->i_private = NULL;
1899         spin_unlock(&inode->i_lock);
1900 out:
1901         mutex_unlock(&inode->i_mutex);
1902         return error;
1903 }
1904
1905 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1906 {
1907         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1908
1909         buf->f_type = TMPFS_MAGIC;
1910         buf->f_bsize = PAGE_CACHE_SIZE;
1911         buf->f_namelen = NAME_MAX;
1912         if (sbinfo->max_blocks) {
1913                 buf->f_blocks = sbinfo->max_blocks;
1914                 buf->f_bavail =
1915                 buf->f_bfree  = sbinfo->max_blocks -
1916                                 percpu_counter_sum(&sbinfo->used_blocks);
1917         }
1918         if (sbinfo->max_inodes) {
1919                 buf->f_files = sbinfo->max_inodes;
1920                 buf->f_ffree = sbinfo->free_inodes;
1921         }
1922         /* else leave those fields 0 like simple_statfs */
1923         return 0;
1924 }
1925
1926 /*
1927  * File creation. Allocate an inode, and we're done..
1928  */
1929 static int
1930 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
1931 {
1932         struct inode *inode;
1933         int error = -ENOSPC;
1934
1935         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1936         if (inode) {
1937                 error = simple_acl_create(dir, inode);
1938                 if (error)
1939                         goto out_iput;
1940                 error = security_inode_init_security(inode, dir,
1941                                                      &dentry->d_name,
1942                                                      shmem_initxattrs, NULL);
1943                 if (error && error != -EOPNOTSUPP)
1944                         goto out_iput;
1945
1946                 error = 0;
1947                 dir->i_size += BOGO_DIRENT_SIZE;
1948                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1949                 d_instantiate(dentry, inode);
1950                 dget(dentry); /* Extra count - pin the dentry in core */
1951         }
1952         return error;
1953 out_iput:
1954         iput(inode);
1955         return error;
1956 }
1957
1958 static int
1959 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
1960 {
1961         struct inode *inode;
1962         int error = -ENOSPC;
1963
1964         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
1965         if (inode) {
1966                 error = security_inode_init_security(inode, dir,
1967                                                      NULL,
1968                                                      shmem_initxattrs, NULL);
1969                 if (error && error != -EOPNOTSUPP)
1970                         goto out_iput;
1971                 error = simple_acl_create(dir, inode);
1972                 if (error)
1973                         goto out_iput;
1974                 d_tmpfile(dentry, inode);
1975         }
1976         return error;
1977 out_iput:
1978         iput(inode);
1979         return error;
1980 }
1981
1982 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1983 {
1984         int error;
1985
1986         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1987                 return error;
1988         inc_nlink(dir);
1989         return 0;
1990 }
1991
1992 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
1993                 bool excl)
1994 {
1995         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1996 }
1997
1998 /*
1999  * Link a file..
2000  */
2001 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2002 {
2003         struct inode *inode = old_dentry->d_inode;
2004         int ret;
2005
2006         /*
2007          * No ordinary (disk based) filesystem counts links as inodes;
2008          * but each new link needs a new dentry, pinning lowmem, and
2009          * tmpfs dentries cannot be pruned until they are unlinked.
2010          */
2011         ret = shmem_reserve_inode(inode->i_sb);
2012         if (ret)
2013                 goto out;
2014
2015         dir->i_size += BOGO_DIRENT_SIZE;
2016         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2017         inc_nlink(inode);
2018         ihold(inode);   /* New dentry reference */
2019         dget(dentry);           /* Extra pinning count for the created dentry */
2020         d_instantiate(dentry, inode);
2021 out:
2022         return ret;
2023 }
2024
2025 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2026 {
2027         struct inode *inode = dentry->d_inode;
2028
2029         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2030                 shmem_free_inode(inode->i_sb);
2031
2032         dir->i_size -= BOGO_DIRENT_SIZE;
2033         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2034         drop_nlink(inode);
2035         dput(dentry);   /* Undo the count from "create" - this does all the work */
2036         return 0;
2037 }
2038
2039 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2040 {
2041         if (!simple_empty(dentry))
2042                 return -ENOTEMPTY;
2043
2044         drop_nlink(dentry->d_inode);
2045         drop_nlink(dir);
2046         return shmem_unlink(dir, dentry);
2047 }
2048
2049 /*
2050  * The VFS layer already does all the dentry stuff for rename,
2051  * we just have to decrement the usage count for the target if
2052  * it exists so that the VFS layer correctly free's it when it
2053  * gets overwritten.
2054  */
2055 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2056 {
2057         struct inode *inode = old_dentry->d_inode;
2058         int they_are_dirs = S_ISDIR(inode->i_mode);
2059
2060         if (!simple_empty(new_dentry))
2061                 return -ENOTEMPTY;
2062
2063         if (new_dentry->d_inode) {
2064                 (void) shmem_unlink(new_dir, new_dentry);
2065                 if (they_are_dirs)
2066                         drop_nlink(old_dir);
2067         } else if (they_are_dirs) {
2068                 drop_nlink(old_dir);
2069                 inc_nlink(new_dir);
2070         }
2071
2072         old_dir->i_size -= BOGO_DIRENT_SIZE;
2073         new_dir->i_size += BOGO_DIRENT_SIZE;
2074         old_dir->i_ctime = old_dir->i_mtime =
2075         new_dir->i_ctime = new_dir->i_mtime =
2076         inode->i_ctime = CURRENT_TIME;
2077         return 0;
2078 }
2079
2080 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2081 {
2082         int error;
2083         int len;
2084         struct inode *inode;
2085         struct page *page;
2086         char *kaddr;
2087         struct shmem_inode_info *info;
2088
2089         len = strlen(symname) + 1;
2090         if (len > PAGE_CACHE_SIZE)
2091                 return -ENAMETOOLONG;
2092
2093         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2094         if (!inode)
2095                 return -ENOSPC;
2096
2097         error = security_inode_init_security(inode, dir, &dentry->d_name,
2098                                              shmem_initxattrs, NULL);
2099         if (error) {
2100                 if (error != -EOPNOTSUPP) {
2101                         iput(inode);
2102                         return error;
2103                 }
2104                 error = 0;
2105         }
2106
2107         info = SHMEM_I(inode);
2108         inode->i_size = len-1;
2109         if (len <= SHORT_SYMLINK_LEN) {
2110                 info->symlink = kmemdup(symname, len, GFP_KERNEL);
2111                 if (!info->symlink) {
2112                         iput(inode);
2113                         return -ENOMEM;
2114                 }
2115                 inode->i_op = &shmem_short_symlink_operations;
2116         } else {
2117                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2118                 if (error) {
2119                         iput(inode);
2120                         return error;
2121                 }
2122                 inode->i_mapping->a_ops = &shmem_aops;
2123                 inode->i_op = &shmem_symlink_inode_operations;
2124                 kaddr = kmap_atomic(page);
2125                 memcpy(kaddr, symname, len);
2126                 kunmap_atomic(kaddr);
2127                 SetPageUptodate(page);
2128                 set_page_dirty(page);
2129                 unlock_page(page);
2130                 page_cache_release(page);
2131         }
2132         dir->i_size += BOGO_DIRENT_SIZE;
2133         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2134         d_instantiate(dentry, inode);
2135         dget(dentry);
2136         return 0;
2137 }
2138
2139 static void *shmem_follow_short_symlink(struct dentry *dentry, struct nameidata *nd)
2140 {
2141         nd_set_link(nd, SHMEM_I(dentry->d_inode)->symlink);
2142         return NULL;
2143 }
2144
2145 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2146 {
2147         struct page *page = NULL;
2148         int error = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2149         nd_set_link(nd, error ? ERR_PTR(error) : kmap(page));
2150         if (page)
2151                 unlock_page(page);
2152         return page;
2153 }
2154
2155 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2156 {
2157         if (!IS_ERR(nd_get_link(nd))) {
2158                 struct page *page = cookie;
2159                 kunmap(page);
2160                 mark_page_accessed(page);
2161                 page_cache_release(page);
2162         }
2163 }
2164
2165 #ifdef CONFIG_TMPFS_XATTR
2166 /*
2167  * Superblocks without xattr inode operations may get some security.* xattr
2168  * support from the LSM "for free". As soon as we have any other xattrs
2169  * like ACLs, we also need to implement the security.* handlers at
2170  * filesystem level, though.
2171  */
2172
2173 /*
2174  * Callback for security_inode_init_security() for acquiring xattrs.
2175  */
2176 static int shmem_initxattrs(struct inode *inode,
2177                             const struct xattr *xattr_array,
2178                             void *fs_info)
2179 {
2180         struct shmem_inode_info *info = SHMEM_I(inode);
2181         const struct xattr *xattr;
2182         struct simple_xattr *new_xattr;
2183         size_t len;
2184
2185         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2186                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2187                 if (!new_xattr)
2188                         return -ENOMEM;
2189
2190                 len = strlen(xattr->name) + 1;
2191                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2192                                           GFP_KERNEL);
2193                 if (!new_xattr->name) {
2194                         kfree(new_xattr);
2195                         return -ENOMEM;
2196                 }
2197
2198                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2199                        XATTR_SECURITY_PREFIX_LEN);
2200                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2201                        xattr->name, len);
2202
2203                 simple_xattr_list_add(&info->xattrs, new_xattr);
2204         }
2205
2206         return 0;
2207 }
2208
2209 static const struct xattr_handler *shmem_xattr_handlers[] = {
2210 #ifdef CONFIG_TMPFS_POSIX_ACL
2211         &posix_acl_access_xattr_handler,
2212         &posix_acl_default_xattr_handler,
2213 #endif
2214         NULL
2215 };
2216
2217 static int shmem_xattr_validate(const char *name)
2218 {
2219         struct { const char *prefix; size_t len; } arr[] = {
2220                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2221                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2222         };
2223         int i;
2224
2225         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2226                 size_t preflen = arr[i].len;
2227                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2228                         if (!name[preflen])
2229                                 return -EINVAL;
2230                         return 0;
2231                 }
2232         }
2233         return -EOPNOTSUPP;
2234 }
2235
2236 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2237                               void *buffer, size_t size)
2238 {
2239         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2240         int err;
2241
2242         /*
2243          * If this is a request for a synthetic attribute in the system.*
2244          * namespace use the generic infrastructure to resolve a handler
2245          * for it via sb->s_xattr.
2246          */
2247         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2248                 return generic_getxattr(dentry, name, buffer, size);
2249
2250         err = shmem_xattr_validate(name);
2251         if (err)
2252                 return err;
2253
2254         return simple_xattr_get(&info->xattrs, name, buffer, size);
2255 }
2256
2257 static int shmem_setxattr(struct dentry *dentry, const char *name,
2258                           const void *value, size_t size, int flags)
2259 {
2260         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2261         int err;
2262
2263         /*
2264          * If this is a request for a synthetic attribute in the system.*
2265          * namespace use the generic infrastructure to resolve a handler
2266          * for it via sb->s_xattr.
2267          */
2268         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2269                 return generic_setxattr(dentry, name, value, size, flags);
2270
2271         err = shmem_xattr_validate(name);
2272         if (err)
2273                 return err;
2274
2275         return simple_xattr_set(&info->xattrs, name, value, size, flags);
2276 }
2277
2278 static int shmem_removexattr(struct dentry *dentry, const char *name)
2279 {
2280         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2281         int err;
2282
2283         /*
2284          * If this is a request for a synthetic attribute in the system.*
2285          * namespace use the generic infrastructure to resolve a handler
2286          * for it via sb->s_xattr.
2287          */
2288         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2289                 return generic_removexattr(dentry, name);
2290
2291         err = shmem_xattr_validate(name);
2292         if (err)
2293                 return err;
2294
2295         return simple_xattr_remove(&info->xattrs, name);
2296 }
2297
2298 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2299 {
2300         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2301         return simple_xattr_list(&info->xattrs, buffer, size);
2302 }
2303 #endif /* CONFIG_TMPFS_XATTR */
2304
2305 static const struct inode_operations shmem_short_symlink_operations = {
2306         .readlink       = generic_readlink,
2307         .follow_link    = shmem_follow_short_symlink,
2308 #ifdef CONFIG_TMPFS_XATTR
2309         .setxattr       = shmem_setxattr,
2310         .getxattr       = shmem_getxattr,
2311         .listxattr      = shmem_listxattr,
2312         .removexattr    = shmem_removexattr,
2313 #endif
2314 };
2315
2316 static const struct inode_operations shmem_symlink_inode_operations = {
2317         .readlink       = generic_readlink,
2318         .follow_link    = shmem_follow_link,
2319         .put_link       = shmem_put_link,
2320 #ifdef CONFIG_TMPFS_XATTR
2321         .setxattr       = shmem_setxattr,
2322         .getxattr       = shmem_getxattr,
2323         .listxattr      = shmem_listxattr,
2324         .removexattr    = shmem_removexattr,
2325 #endif
2326 };
2327
2328 static struct dentry *shmem_get_parent(struct dentry *child)
2329 {
2330         return ERR_PTR(-ESTALE);
2331 }
2332
2333 static int shmem_match(struct inode *ino, void *vfh)
2334 {
2335         __u32 *fh = vfh;
2336         __u64 inum = fh[2];
2337         inum = (inum << 32) | fh[1];
2338         return ino->i_ino == inum && fh[0] == ino->i_generation;
2339 }
2340
2341 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2342                 struct fid *fid, int fh_len, int fh_type)
2343 {
2344         struct inode *inode;
2345         struct dentry *dentry = NULL;
2346         u64 inum;
2347
2348         if (fh_len < 3)
2349                 return NULL;
2350
2351         inum = fid->raw[2];
2352         inum = (inum << 32) | fid->raw[1];
2353
2354         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2355                         shmem_match, fid->raw);
2356         if (inode) {
2357                 dentry = d_find_alias(inode);
2358                 iput(inode);
2359         }
2360
2361         return dentry;
2362 }
2363
2364 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2365                                 struct inode *parent)
2366 {
2367         if (*len < 3) {
2368                 *len = 3;
2369                 return FILEID_INVALID;
2370         }
2371
2372         if (inode_unhashed(inode)) {
2373                 /* Unfortunately insert_inode_hash is not idempotent,
2374                  * so as we hash inodes here rather than at creation
2375                  * time, we need a lock to ensure we only try
2376                  * to do it once
2377                  */
2378                 static DEFINE_SPINLOCK(lock);
2379                 spin_lock(&lock);
2380                 if (inode_unhashed(inode))
2381                         __insert_inode_hash(inode,
2382                                             inode->i_ino + inode->i_generation);
2383                 spin_unlock(&lock);
2384         }
2385
2386         fh[0] = inode->i_generation;
2387         fh[1] = inode->i_ino;
2388         fh[2] = ((__u64)inode->i_ino) >> 32;
2389
2390         *len = 3;
2391         return 1;
2392 }
2393
2394 static const struct export_operations shmem_export_ops = {
2395         .get_parent     = shmem_get_parent,
2396         .encode_fh      = shmem_encode_fh,
2397         .fh_to_dentry   = shmem_fh_to_dentry,
2398 };
2399
2400 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2401                                bool remount)
2402 {
2403         char *this_char, *value, *rest;
2404         struct mempolicy *mpol = NULL;
2405         uid_t uid;
2406         gid_t gid;
2407
2408         while (options != NULL) {
2409                 this_char = options;
2410                 for (;;) {
2411                         /*
2412                          * NUL-terminate this option: unfortunately,
2413                          * mount options form a comma-separated list,
2414                          * but mpol's nodelist may also contain commas.
2415                          */
2416                         options = strchr(options, ',');
2417                         if (options == NULL)
2418                                 break;
2419                         options++;
2420                         if (!isdigit(*options)) {
2421                                 options[-1] = '\0';
2422                                 break;
2423                         }
2424                 }
2425                 if (!*this_char)
2426                         continue;
2427                 if ((value = strchr(this_char,'=')) != NULL) {
2428                         *value++ = 0;
2429                 } else {
2430                         printk(KERN_ERR
2431                             "tmpfs: No value for mount option '%s'\n",
2432                             this_char);
2433                         goto error;
2434                 }
2435
2436                 if (!strcmp(this_char,"size")) {
2437                         unsigned long long size;
2438                         size = memparse(value,&rest);
2439                         if (*rest == '%') {
2440                                 size <<= PAGE_SHIFT;
2441                                 size *= totalram_pages;
2442                                 do_div(size, 100);
2443                                 rest++;
2444                         }
2445                         if (*rest)
2446                                 goto bad_val;
2447                         sbinfo->max_blocks =
2448                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2449                 } else if (!strcmp(this_char,"nr_blocks")) {
2450                         sbinfo->max_blocks = memparse(value, &rest);
2451                         if (*rest)
2452                                 goto bad_val;
2453                 } else if (!strcmp(this_char,"nr_inodes")) {
2454                         sbinfo->max_inodes = memparse(value, &rest);
2455                         if (*rest)
2456                                 goto bad_val;
2457                 } else if (!strcmp(this_char,"mode")) {
2458                         if (remount)
2459                                 continue;
2460                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2461                         if (*rest)
2462                                 goto bad_val;
2463                 } else if (!strcmp(this_char,"uid")) {
2464                         if (remount)
2465                                 continue;
2466                         uid = simple_strtoul(value, &rest, 0);
2467                         if (*rest)
2468                                 goto bad_val;
2469                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2470                         if (!uid_valid(sbinfo->uid))
2471                                 goto bad_val;
2472                 } else if (!strcmp(this_char,"gid")) {
2473                         if (remount)
2474                                 continue;
2475                         gid = simple_strtoul(value, &rest, 0);
2476                         if (*rest)
2477                                 goto bad_val;
2478                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2479                         if (!gid_valid(sbinfo->gid))
2480                                 goto bad_val;
2481                 } else if (!strcmp(this_char,"mpol")) {
2482                         mpol_put(mpol);
2483                         mpol = NULL;
2484                         if (mpol_parse_str(value, &mpol))
2485                                 goto bad_val;
2486                 } else {
2487                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2488                                this_char);
2489                         goto error;
2490                 }
2491         }
2492         sbinfo->mpol = mpol;
2493         return 0;
2494
2495 bad_val:
2496         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2497                value, this_char);
2498 error:
2499         mpol_put(mpol);
2500         return 1;
2501
2502 }
2503
2504 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2505 {
2506         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2507         struct shmem_sb_info config = *sbinfo;
2508         unsigned long inodes;
2509         int error = -EINVAL;
2510
2511         config.mpol = NULL;
2512         if (shmem_parse_options(data, &config, true))
2513                 return error;
2514
2515         spin_lock(&sbinfo->stat_lock);
2516         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2517         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2518                 goto out;
2519         if (config.max_inodes < inodes)
2520                 goto out;
2521         /*
2522          * Those tests disallow limited->unlimited while any are in use;
2523          * but we must separately disallow unlimited->limited, because
2524          * in that case we have no record of how much is already in use.
2525          */
2526         if (config.max_blocks && !sbinfo->max_blocks)
2527                 goto out;
2528         if (config.max_inodes && !sbinfo->max_inodes)
2529                 goto out;
2530
2531         error = 0;
2532         sbinfo->max_blocks  = config.max_blocks;
2533         sbinfo->max_inodes  = config.max_inodes;
2534         sbinfo->free_inodes = config.max_inodes - inodes;
2535
2536         /*
2537          * Preserve previous mempolicy unless mpol remount option was specified.
2538          */
2539         if (config.mpol) {
2540                 mpol_put(sbinfo->mpol);
2541                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
2542         }
2543 out:
2544         spin_unlock(&sbinfo->stat_lock);
2545         return error;
2546 }
2547
2548 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2549 {
2550         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2551
2552         if (sbinfo->max_blocks != shmem_default_max_blocks())
2553                 seq_printf(seq, ",size=%luk",
2554                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2555         if (sbinfo->max_inodes != shmem_default_max_inodes())
2556                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2557         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2558                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2559         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2560                 seq_printf(seq, ",uid=%u",
2561                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2562         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2563                 seq_printf(seq, ",gid=%u",
2564                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2565         shmem_show_mpol(seq, sbinfo->mpol);
2566         return 0;
2567 }
2568 #endif /* CONFIG_TMPFS */
2569
2570 static void shmem_put_super(struct super_block *sb)
2571 {
2572         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2573
2574         percpu_counter_destroy(&sbinfo->used_blocks);
2575         mpol_put(sbinfo->mpol);
2576         kfree(sbinfo);
2577         sb->s_fs_info = NULL;
2578 }
2579
2580 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2581 {
2582         struct inode *inode;
2583         struct shmem_sb_info *sbinfo;
2584         int err = -ENOMEM;
2585
2586         /* Round up to L1_CACHE_BYTES to resist false sharing */
2587         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2588                                 L1_CACHE_BYTES), GFP_KERNEL);
2589         if (!sbinfo)
2590                 return -ENOMEM;
2591
2592         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2593         sbinfo->uid = current_fsuid();
2594         sbinfo->gid = current_fsgid();
2595         sb->s_fs_info = sbinfo;
2596
2597 #ifdef CONFIG_TMPFS
2598         /*
2599          * Per default we only allow half of the physical ram per
2600          * tmpfs instance, limiting inodes to one per page of lowmem;
2601          * but the internal instance is left unlimited.
2602          */
2603         if (!(sb->s_flags & MS_KERNMOUNT)) {
2604                 sbinfo->max_blocks = shmem_default_max_blocks();
2605                 sbinfo->max_inodes = shmem_default_max_inodes();
2606                 if (shmem_parse_options(data, sbinfo, false)) {
2607                         err = -EINVAL;
2608                         goto failed;
2609                 }
2610         } else {
2611                 sb->s_flags |= MS_NOUSER;
2612         }
2613         sb->s_export_op = &shmem_export_ops;
2614         sb->s_flags |= MS_NOSEC;
2615 #else
2616         sb->s_flags |= MS_NOUSER;
2617 #endif
2618
2619         spin_lock_init(&sbinfo->stat_lock);
2620         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2621                 goto failed;
2622         sbinfo->free_inodes = sbinfo->max_inodes;
2623
2624         sb->s_maxbytes = MAX_LFS_FILESIZE;
2625         sb->s_blocksize = PAGE_CACHE_SIZE;
2626         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2627         sb->s_magic = TMPFS_MAGIC;
2628         sb->s_op = &shmem_ops;
2629         sb->s_time_gran = 1;
2630 #ifdef CONFIG_TMPFS_XATTR
2631         sb->s_xattr = shmem_xattr_handlers;
2632 #endif
2633 #ifdef CONFIG_TMPFS_POSIX_ACL
2634         sb->s_flags |= MS_POSIXACL;
2635 #endif
2636
2637         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2638         if (!inode)
2639                 goto failed;
2640         inode->i_uid = sbinfo->uid;
2641         inode->i_gid = sbinfo->gid;
2642         sb->s_root = d_make_root(inode);
2643         if (!sb->s_root)
2644                 goto failed;
2645         return 0;
2646
2647 failed:
2648         shmem_put_super(sb);
2649         return err;
2650 }
2651
2652 static struct kmem_cache *shmem_inode_cachep;
2653
2654 static struct inode *shmem_alloc_inode(struct super_block *sb)
2655 {
2656         struct shmem_inode_info *info;
2657         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2658         if (!info)
2659                 return NULL;
2660         return &info->vfs_inode;
2661 }
2662
2663 static void shmem_destroy_callback(struct rcu_head *head)
2664 {
2665         struct inode *inode = container_of(head, struct inode, i_rcu);
2666         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2667 }
2668
2669 static void shmem_destroy_inode(struct inode *inode)
2670 {
2671         if (S_ISREG(inode->i_mode))
2672                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2673         call_rcu(&inode->i_rcu, shmem_destroy_callback);
2674 }
2675
2676 static void shmem_init_inode(void *foo)
2677 {
2678         struct shmem_inode_info *info = foo;
2679         inode_init_once(&info->vfs_inode);
2680 }
2681
2682 static int shmem_init_inodecache(void)
2683 {
2684         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2685                                 sizeof(struct shmem_inode_info),
2686                                 0, SLAB_PANIC, shmem_init_inode);
2687         return 0;
2688 }
2689
2690 static void shmem_destroy_inodecache(void)
2691 {
2692         kmem_cache_destroy(shmem_inode_cachep);
2693 }
2694
2695 static const struct address_space_operations shmem_aops = {
2696         .writepage      = shmem_writepage,
2697         .set_page_dirty = __set_page_dirty_no_writeback,
2698 #ifdef CONFIG_TMPFS
2699         .write_begin    = shmem_write_begin,
2700         .write_end      = shmem_write_end,
2701 #endif
2702         .migratepage    = migrate_page,
2703         .error_remove_page = generic_error_remove_page,
2704 };
2705
2706 static const struct file_operations shmem_file_operations = {
2707         .mmap           = shmem_mmap,
2708 #ifdef CONFIG_TMPFS
2709         .llseek         = shmem_file_llseek,
2710         .read           = do_sync_read,
2711         .write          = do_sync_write,
2712         .aio_read       = shmem_file_aio_read,
2713         .aio_write      = generic_file_aio_write,
2714         .fsync          = noop_fsync,
2715         .splice_read    = shmem_file_splice_read,
2716         .splice_write   = generic_file_splice_write,
2717         .fallocate      = shmem_fallocate,
2718 #endif
2719 };
2720
2721 static const struct inode_operations shmem_inode_operations = {
2722         .setattr        = shmem_setattr,
2723 #ifdef CONFIG_TMPFS_XATTR
2724         .setxattr       = shmem_setxattr,
2725         .getxattr       = shmem_getxattr,
2726         .listxattr      = shmem_listxattr,
2727         .removexattr    = shmem_removexattr,
2728         .set_acl        = simple_set_acl,
2729 #endif
2730 };
2731
2732 static const struct inode_operations shmem_dir_inode_operations = {
2733 #ifdef CONFIG_TMPFS
2734         .create         = shmem_create,
2735         .lookup         = simple_lookup,
2736         .link           = shmem_link,
2737         .unlink         = shmem_unlink,
2738         .symlink        = shmem_symlink,
2739         .mkdir          = shmem_mkdir,
2740         .rmdir          = shmem_rmdir,
2741         .mknod          = shmem_mknod,
2742         .rename         = shmem_rename,
2743         .tmpfile        = shmem_tmpfile,
2744 #endif
2745 #ifdef CONFIG_TMPFS_XATTR
2746         .setxattr       = shmem_setxattr,
2747         .getxattr       = shmem_getxattr,
2748         .listxattr      = shmem_listxattr,
2749         .removexattr    = shmem_removexattr,
2750 #endif
2751 #ifdef CONFIG_TMPFS_POSIX_ACL
2752         .setattr        = shmem_setattr,
2753         .set_acl        = simple_set_acl,
2754 #endif
2755 };
2756
2757 static const struct inode_operations shmem_special_inode_operations = {
2758 #ifdef CONFIG_TMPFS_XATTR
2759         .setxattr       = shmem_setxattr,
2760         .getxattr       = shmem_getxattr,
2761         .listxattr      = shmem_listxattr,
2762         .removexattr    = shmem_removexattr,
2763 #endif
2764 #ifdef CONFIG_TMPFS_POSIX_ACL
2765         .setattr        = shmem_setattr,
2766         .set_acl        = simple_set_acl,
2767 #endif
2768 };
2769
2770 static const struct super_operations shmem_ops = {
2771         .alloc_inode    = shmem_alloc_inode,
2772         .destroy_inode  = shmem_destroy_inode,
2773 #ifdef CONFIG_TMPFS
2774         .statfs         = shmem_statfs,
2775         .remount_fs     = shmem_remount_fs,
2776         .show_options   = shmem_show_options,
2777 #endif
2778         .evict_inode    = shmem_evict_inode,
2779         .drop_inode     = generic_delete_inode,
2780         .put_super      = shmem_put_super,
2781 };
2782
2783 static const struct vm_operations_struct shmem_vm_ops = {
2784         .fault          = shmem_fault,
2785 #ifdef CONFIG_NUMA
2786         .set_policy     = shmem_set_policy,
2787         .get_policy     = shmem_get_policy,
2788 #endif
2789         .remap_pages    = generic_file_remap_pages,
2790 };
2791
2792 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2793         int flags, const char *dev_name, void *data)
2794 {
2795         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2796 }
2797
2798 static struct file_system_type shmem_fs_type = {
2799         .owner          = THIS_MODULE,
2800         .name           = "tmpfs",
2801         .mount          = shmem_mount,
2802         .kill_sb        = kill_litter_super,
2803         .fs_flags       = FS_USERNS_MOUNT,
2804 };
2805
2806 int __init shmem_init(void)
2807 {
2808         int error;
2809
2810         /* If rootfs called this, don't re-init */
2811         if (shmem_inode_cachep)
2812                 return 0;
2813
2814         error = bdi_init(&shmem_backing_dev_info);
2815         if (error)
2816                 goto out4;
2817
2818         error = shmem_init_inodecache();
2819         if (error)
2820                 goto out3;
2821
2822         error = register_filesystem(&shmem_fs_type);
2823         if (error) {
2824                 printk(KERN_ERR "Could not register tmpfs\n");
2825                 goto out2;
2826         }
2827
2828         shm_mnt = kern_mount(&shmem_fs_type);
2829         if (IS_ERR(shm_mnt)) {
2830                 error = PTR_ERR(shm_mnt);
2831                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2832                 goto out1;
2833         }
2834         return 0;
2835
2836 out1:
2837         unregister_filesystem(&shmem_fs_type);
2838 out2:
2839         shmem_destroy_inodecache();
2840 out3:
2841         bdi_destroy(&shmem_backing_dev_info);
2842 out4:
2843         shm_mnt = ERR_PTR(error);
2844         return error;
2845 }
2846
2847 #else /* !CONFIG_SHMEM */
2848
2849 /*
2850  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2851  *
2852  * This is intended for small system where the benefits of the full
2853  * shmem code (swap-backed and resource-limited) are outweighed by
2854  * their complexity. On systems without swap this code should be
2855  * effectively equivalent, but much lighter weight.
2856  */
2857
2858 static struct file_system_type shmem_fs_type = {
2859         .name           = "tmpfs",
2860         .mount          = ramfs_mount,
2861         .kill_sb        = kill_litter_super,
2862         .fs_flags       = FS_USERNS_MOUNT,
2863 };
2864
2865 int __init shmem_init(void)
2866 {
2867         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
2868
2869         shm_mnt = kern_mount(&shmem_fs_type);
2870         BUG_ON(IS_ERR(shm_mnt));
2871
2872         return 0;
2873 }
2874
2875 int shmem_unuse(swp_entry_t swap, struct page *page)
2876 {
2877         return 0;
2878 }
2879
2880 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2881 {
2882         return 0;
2883 }
2884
2885 void shmem_unlock_mapping(struct address_space *mapping)
2886 {
2887 }
2888
2889 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
2890 {
2891         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
2892 }
2893 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2894
2895 #define shmem_vm_ops                            generic_file_vm_ops
2896 #define shmem_file_operations                   ramfs_file_operations
2897 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2898 #define shmem_acct_size(flags, size)            0
2899 #define shmem_unacct_size(flags, size)          do {} while (0)
2900
2901 #endif /* CONFIG_SHMEM */
2902
2903 /* common code */
2904
2905 static struct dentry_operations anon_ops = {
2906         .d_dname = simple_dname
2907 };
2908
2909 static struct file *__shmem_file_setup(const char *name, loff_t size,
2910                                        unsigned long flags, unsigned int i_flags)
2911 {
2912         struct file *res;
2913         struct inode *inode;
2914         struct path path;
2915         struct super_block *sb;
2916         struct qstr this;
2917
2918         if (IS_ERR(shm_mnt))
2919                 return ERR_CAST(shm_mnt);
2920
2921         if (size < 0 || size > MAX_LFS_FILESIZE)
2922                 return ERR_PTR(-EINVAL);
2923
2924         if (shmem_acct_size(flags, size))
2925                 return ERR_PTR(-ENOMEM);
2926
2927         res = ERR_PTR(-ENOMEM);
2928         this.name = name;
2929         this.len = strlen(name);
2930         this.hash = 0; /* will go */
2931         sb = shm_mnt->mnt_sb;
2932         path.dentry = d_alloc_pseudo(sb, &this);
2933         if (!path.dentry)
2934                 goto put_memory;
2935         d_set_d_op(path.dentry, &anon_ops);
2936         path.mnt = mntget(shm_mnt);
2937
2938         res = ERR_PTR(-ENOSPC);
2939         inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2940         if (!inode)
2941                 goto put_dentry;
2942
2943         inode->i_flags |= i_flags;
2944         d_instantiate(path.dentry, inode);
2945         inode->i_size = size;
2946         clear_nlink(inode);     /* It is unlinked */
2947         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
2948         if (IS_ERR(res))
2949                 goto put_dentry;
2950
2951         res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2952                   &shmem_file_operations);
2953         if (IS_ERR(res))
2954                 goto put_dentry;
2955
2956         return res;
2957
2958 put_dentry:
2959         path_put(&path);
2960 put_memory:
2961         shmem_unacct_size(flags, size);
2962         return res;
2963 }
2964
2965 /**
2966  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
2967  *      kernel internal.  There will be NO LSM permission checks against the
2968  *      underlying inode.  So users of this interface must do LSM checks at a
2969  *      higher layer.  The one user is the big_key implementation.  LSM checks
2970  *      are provided at the key level rather than the inode level.
2971  * @name: name for dentry (to be seen in /proc/<pid>/maps
2972  * @size: size to be set for the file
2973  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2974  */
2975 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
2976 {
2977         return __shmem_file_setup(name, size, flags, S_PRIVATE);
2978 }
2979
2980 /**
2981  * shmem_file_setup - get an unlinked file living in tmpfs
2982  * @name: name for dentry (to be seen in /proc/<pid>/maps
2983  * @size: size to be set for the file
2984  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2985  */
2986 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2987 {
2988         return __shmem_file_setup(name, size, flags, 0);
2989 }
2990 EXPORT_SYMBOL_GPL(shmem_file_setup);
2991
2992 /**
2993  * shmem_zero_setup - setup a shared anonymous mapping
2994  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2995  */
2996 int shmem_zero_setup(struct vm_area_struct *vma)
2997 {
2998         struct file *file;
2999         loff_t size = vma->vm_end - vma->vm_start;
3000
3001         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
3002         if (IS_ERR(file))
3003                 return PTR_ERR(file);
3004
3005         if (vma->vm_file)
3006                 fput(vma->vm_file);
3007         vma->vm_file = file;
3008         vma->vm_ops = &shmem_vm_ops;
3009         return 0;
3010 }
3011
3012 /**
3013  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3014  * @mapping:    the page's address_space
3015  * @index:      the page index
3016  * @gfp:        the page allocator flags to use if allocating
3017  *
3018  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3019  * with any new page allocations done using the specified allocation flags.
3020  * But read_cache_page_gfp() uses the ->readpage() method: which does not
3021  * suit tmpfs, since it may have pages in swapcache, and needs to find those
3022  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3023  *
3024  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
3025  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
3026  */
3027 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3028                                          pgoff_t index, gfp_t gfp)
3029 {
3030 #ifdef CONFIG_SHMEM
3031         struct inode *inode = mapping->host;
3032         struct page *page;
3033         int error;
3034
3035         BUG_ON(mapping->a_ops != &shmem_aops);
3036         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3037         if (error)
3038                 page = ERR_PTR(error);
3039         else
3040                 unlock_page(page);
3041         return page;
3042 #else
3043         /*
3044          * The tiny !SHMEM case uses ramfs without swap
3045          */
3046         return read_cache_page_gfp(mapping, index, gfp);
3047 #endif
3048 }
3049 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);