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) 2002-2005 Hugh Dickins.
  10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  12  *
  13  * Extended attribute support for tmpfs:
  14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
  15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
  16  *
  17  * tiny-shmem:
  18  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
  19  *
  20  * This file is released under the GPL.
  21  */
  22
  23 #include <linux/fs.h>
  24 #include <linux/init.h>
  25 #include <linux/vfs.h>
  26 #include <linux/mount.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/file.h>
  29 #include <linux/mm.h>
  30 #include <linux/module.h>
  31 #include <linux/percpu_counter.h>
  32 #include <linux/swap.h>
  33
  34 static struct vfsmount *shm_mnt;
  35
  36 #ifdef CONFIG_SHMEM
  37 /*
  38  * This virtual memory filesystem is heavily based on the ramfs. It
  39  * extends ramfs by the ability to use swap and honor resource limits
  40  * which makes it a completely usable filesystem.
  41  */
  42
  43 #include <linux/xattr.h>
  44 #include <linux/exportfs.h>
  45 #include <linux/posix_acl.h>
  46 #include <linux/generic_acl.h>
  47 #include <linux/mman.h>
  48 #include <linux/string.h>
  49 #include <linux/slab.h>
  50 #include <linux/backing-dev.h>
  51 #include <linux/shmem_fs.h>
  52 #include <linux/writeback.h>
  53 #include <linux/blkdev.h>
  54 #include <linux/splice.h>
  55 #include <linux/security.h>
  56 #include <linux/swapops.h>
  57 #include <linux/mempolicy.h>
  58 #include <linux/namei.h>
  59 #include <linux/ctype.h>
  60 #include <linux/migrate.h>
  61 #include <linux/highmem.h>
  62 #include <linux/seq_file.h>
  63 #include <linux/magic.h>
  64
  65 #include <asm/uaccess.h>
  66 #include <asm/div64.h>
  67 #include <asm/pgtable.h>
  68
  69 /*
  70  * The maximum size of a shmem/tmpfs file is limited by the maximum size of
  71  * its triple-indirect swap vector - see illustration at shmem_swp_entry().
  72  *
  73  * With 4kB page size, maximum file size is just over 2TB on a 32-bit kernel,
  74  * but one eighth of that on a 64-bit kernel.  With 8kB page size, maximum
  75  * file size is just over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
  76  * MAX_LFS_FILESIZE being then more restrictive than swap vector layout.
  77  *
  78  * We use / and * instead of shifts in the definitions below, so that the swap
  79  * vector can be tested with small even values (e.g. 20) for ENTRIES_PER_PAGE.
  80  */
  81 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
  82 #define ENTRIES_PER_PAGEPAGE ((unsigned long long)ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
  83
  84 #define SHMSWP_MAX_INDEX (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
  85 #define SHMSWP_MAX_BYTES (SHMSWP_MAX_INDEX << PAGE_CACHE_SHIFT)
  86
  87 #define SHMEM_MAX_BYTES  min_t(unsigned long long, SHMSWP_MAX_BYTES, MAX_LFS_FILESIZE)
  88 #define SHMEM_MAX_INDEX  ((unsigned long)((SHMEM_MAX_BYTES+1) >> PAGE_CACHE_SHIFT))
  89
  90 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  91 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  92
  93 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
  94 #define SHMEM_PAGEIN     VM_READ
  95 #define SHMEM_TRUNCATE   VM_WRITE
  96
  97 /* Definition to limit shmem_truncate's steps between cond_rescheds */
  98 #define LATENCY_LIMIT    64
  99
 100 /* Pretend that each entry is of this size in directory's i_size */
 101 #define BOGO_DIRENT_SIZE 20
 102
 103 struct shmem_xattr {
 104         struct list_head list;  /* anchored by shmem_inode_info->xattr_list */
 105         char *name;             /* xattr name */
 106         size_t size;
 107         char value[0];
 108 };
 109
 110 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
 111 enum sgp_type {
 112         SGP_READ,       /* don't exceed i_size, don't allocate page */
 113         SGP_CACHE,      /* don't exceed i_size, may allocate page */
 114         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
 115         SGP_WRITE,      /* may exceed i_size, may allocate page */
 116 };
 117
 118 #ifdef CONFIG_TMPFS
 119 static unsigned long shmem_default_max_blocks(void)
 120 {
 121         return totalram_pages / 2;
 122 }
 123
 124 static unsigned long shmem_default_max_inodes(void)
 125 {
 126         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 127 }
 128 #endif
 129
 130 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 131         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
 132
 133 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
 134         struct page **pagep, enum sgp_type sgp, int *fault_type)
 135 {
 136         return shmem_getpage_gfp(inode, index, pagep, sgp,
 137                         mapping_gfp_mask(inode->i_mapping), fault_type);
 138 }
 139
 140 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
 141 {
 142         /*
 143          * The above definition of ENTRIES_PER_PAGE, and the use of
 144          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
 145          * might be reconsidered if it ever diverges from PAGE_SIZE.
 146          *
 147          * Mobility flags are masked out as swap vectors cannot move
 148          */
 149         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
 150                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
 151 }
 152
 153 static inline void shmem_dir_free(struct page *page)
 154 {
 155         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
 156 }
 157
 158 static struct page **shmem_dir_map(struct page *page)
 159 {
 160         return (struct page **)kmap_atomic(page, KM_USER0);
 161 }
 162
 163 static inline void shmem_dir_unmap(struct page **dir)
 164 {
 165         kunmap_atomic(dir, KM_USER0);
 166 }
 167
 168 static swp_entry_t *shmem_swp_map(struct page *page)
 169 {
 170         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
 171 }
 172
 173 static inline void shmem_swp_balance_unmap(void)
 174 {
 175         /*
 176          * When passing a pointer to an i_direct entry, to code which
 177          * also handles indirect entries and so will shmem_swp_unmap,
 178          * we must arrange for the preempt count to remain in balance.
 179          * What kmap_atomic of a lowmem page does depends on config
 180          * and architecture, so pretend to kmap_atomic some lowmem page.
 181          */
 182         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
 183 }
 184
 185 static inline void shmem_swp_unmap(swp_entry_t *entry)
 186 {
 187         kunmap_atomic(entry, KM_USER1);
 188 }
 189
 190 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 191 {
 192         return sb->s_fs_info;
 193 }
 194
 195 /*
 196  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 197  * for shared memory and for shared anonymous (/dev/zero) mappings
 198  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 199  * consistent with the pre-accounting of private mappings ...
 200  */
 201 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 202 {
 203         return (flags & VM_NORESERVE) ?
 204                 0 : security_vm_enough_memory_kern(VM_ACCT(size));
 205 }
 206
 207 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 208 {
 209         if (!(flags & VM_NORESERVE))
 210                 vm_unacct_memory(VM_ACCT(size));
 211 }
 212
 213 /*
 214  * ... whereas tmpfs objects are accounted incrementally as
 215  * pages are allocated, in order to allow huge sparse files.
 216  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 217  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 218  */
 219 static inline int shmem_acct_block(unsigned long flags)
 220 {
 221         return (flags & VM_NORESERVE) ?
 222                 security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 223 }
 224
 225 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 226 {
 227         if (flags & VM_NORESERVE)
 228                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 229 }
 230
 231 static const struct super_operations shmem_ops;
 232 static const struct address_space_operations shmem_aops;
 233 static const struct file_operations shmem_file_operations;
 234 static const struct inode_operations shmem_inode_operations;
 235 static const struct inode_operations shmem_dir_inode_operations;
 236 static const struct inode_operations shmem_special_inode_operations;
 237 static const struct vm_operations_struct shmem_vm_ops;
 238
 239 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 240         .ra_pages       = 0,    /* No readahead */
 241         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
 242 };
 243
 244 static LIST_HEAD(shmem_swaplist);
 245 static DEFINE_MUTEX(shmem_swaplist_mutex);
 246
 247 static void shmem_free_blocks(struct inode *inode, long pages)
 248 {
 249         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 250         if (sbinfo->max_blocks) {
 251                 percpu_counter_add(&sbinfo->used_blocks, -pages);
 252                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
 253         }
 254 }
 255
 256 static int shmem_reserve_inode(struct super_block *sb)
 257 {
 258         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 259         if (sbinfo->max_inodes) {
 260                 spin_lock(&sbinfo->stat_lock);
 261                 if (!sbinfo->free_inodes) {
 262                         spin_unlock(&sbinfo->stat_lock);
 263                         return -ENOSPC;
 264                 }
 265                 sbinfo->free_inodes--;
 266                 spin_unlock(&sbinfo->stat_lock);
 267         }
 268         return 0;
 269 }
 270
 271 static void shmem_free_inode(struct super_block *sb)
 272 {
 273         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 274         if (sbinfo->max_inodes) {
 275                 spin_lock(&sbinfo->stat_lock);
 276                 sbinfo->free_inodes++;
 277                 spin_unlock(&sbinfo->stat_lock);
 278         }
 279 }
 280
 281 /**
 282  * shmem_recalc_inode - recalculate the size of an inode
 283  * @inode: inode to recalc
 284  *
 285  * We have to calculate the free blocks since the mm can drop
 286  * undirtied hole pages behind our back.
 287  *
 288  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 289  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 290  *
 291  * It has to be called with the spinlock held.
 292  */
 293 static void shmem_recalc_inode(struct inode *inode)
 294 {
 295         struct shmem_inode_info *info = SHMEM_I(inode);
 296         long freed;
 297
 298         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 299         if (freed > 0) {
 300                 info->alloced -= freed;
 301                 shmem_unacct_blocks(info->flags, freed);
 302                 shmem_free_blocks(inode, freed);
 303         }
 304 }
 305
 306 /**
 307  * shmem_swp_entry - find the swap vector position in the info structure
 308  * @info:  info structure for the inode
 309  * @index: index of the page to find
 310  * @page:  optional page to add to the structure. Has to be preset to
 311  *         all zeros
 312  *
 313  * If there is no space allocated yet it will return NULL when
 314  * page is NULL, else it will use the page for the needed block,
 315  * setting it to NULL on return to indicate that it has been used.
 316  *
 317  * The swap vector is organized the following way:
 318  *
 319  * There are SHMEM_NR_DIRECT entries directly stored in the
 320  * shmem_inode_info structure. So small files do not need an addional
 321  * allocation.
 322  *
 323  * For pages with index > SHMEM_NR_DIRECT there is the pointer
 324  * i_indirect which points to a page which holds in the first half
 325  * doubly indirect blocks, in the second half triple indirect blocks:
 326  *
 327  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
 328  * following layout (for SHMEM_NR_DIRECT == 16):
 329  *
 330  * i_indirect -> dir --> 16-19
 331  *            |      +-> 20-23
 332  *            |
 333  *            +-->dir2 --> 24-27
 334  *            |        +-> 28-31
 335  *            |        +-> 32-35
 336  *            |        +-> 36-39
 337  *            |
 338  *            +-->dir3 --> 40-43
 339  *                     +-> 44-47
 340  *                     +-> 48-51
 341  *                     +-> 52-55
 342  */
 343 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
 344 {
 345         unsigned long offset;
 346         struct page **dir;
 347         struct page *subdir;
 348
 349         if (index < SHMEM_NR_DIRECT) {
 350                 shmem_swp_balance_unmap();
 351                 return info->i_direct+index;
 352         }
 353         if (!info->i_indirect) {
 354                 if (page) {
 355                         info->i_indirect = *page;
 356                         *page = NULL;
 357                 }
 358                 return NULL;                    /* need another page */
 359         }
 360
 361         index -= SHMEM_NR_DIRECT;
 362         offset = index % ENTRIES_PER_PAGE;
 363         index /= ENTRIES_PER_PAGE;
 364         dir = shmem_dir_map(info->i_indirect);
 365
 366         if (index >= ENTRIES_PER_PAGE/2) {
 367                 index -= ENTRIES_PER_PAGE/2;
 368                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
 369                 index %= ENTRIES_PER_PAGE;
 370                 subdir = *dir;
 371                 if (!subdir) {
 372                         if (page) {
 373                                 *dir = *page;
 374                                 *page = NULL;
 375                         }
 376                         shmem_dir_unmap(dir);
 377                         return NULL;            /* need another page */
 378                 }
 379                 shmem_dir_unmap(dir);
 380                 dir = shmem_dir_map(subdir);
 381         }
 382
 383         dir += index;
 384         subdir = *dir;
 385         if (!subdir) {
 386                 if (!page || !(subdir = *page)) {
 387                         shmem_dir_unmap(dir);
 388                         return NULL;            /* need a page */
 389                 }
 390                 *dir = subdir;
 391                 *page = NULL;
 392         }
 393         shmem_dir_unmap(dir);
 394         return shmem_swp_map(subdir) + offset;
 395 }
 396
 397 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
 398 {
 399         long incdec = value? 1: -1;
 400
 401         entry->val = value;
 402         info->swapped += incdec;
 403         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
 404                 struct page *page = kmap_atomic_to_page(entry);
 405                 set_page_private(page, page_private(page) + incdec);
 406         }
 407 }
 408
 409 /**
 410  * shmem_swp_alloc - get the position of the swap entry for the page.
 411  * @info:       info structure for the inode
 412  * @index:      index of the page to find
 413  * @sgp:        check and recheck i_size? skip allocation?
 414  * @gfp:        gfp mask to use for any page allocation
 415  *
 416  * If the entry does not exist, allocate it.
 417  */
 418 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info,
 419                         unsigned long index, enum sgp_type sgp, gfp_t gfp)
 420 {
 421         struct inode *inode = &info->vfs_inode;
 422         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 423         struct page *page = NULL;
 424         swp_entry_t *entry;
 425
 426         if (sgp != SGP_WRITE &&
 427             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
 428                 return ERR_PTR(-EINVAL);
 429
 430         while (!(entry = shmem_swp_entry(info, index, &page))) {
 431                 if (sgp == SGP_READ)
 432                         return shmem_swp_map(ZERO_PAGE(0));
 433                 /*
 434                  * Test used_blocks against 1 less max_blocks, since we have 1 data
 435                  * page (and perhaps indirect index pages) yet to allocate:
 436                  * a waste to allocate index if we cannot allocate data.
 437                  */
 438                 if (sbinfo->max_blocks) {
 439                         if (percpu_counter_compare(&sbinfo->used_blocks,
 440                                                 sbinfo->max_blocks - 1) >= 0)
 441                                 return ERR_PTR(-ENOSPC);
 442                         percpu_counter_inc(&sbinfo->used_blocks);
 443                         inode->i_blocks += BLOCKS_PER_PAGE;
 444                 }
 445
 446                 spin_unlock(&info->lock);
 447                 page = shmem_dir_alloc(gfp);
 448                 spin_lock(&info->lock);
 449
 450                 if (!page) {
 451                         shmem_free_blocks(inode, 1);
 452                         return ERR_PTR(-ENOMEM);
 453                 }
 454                 if (sgp != SGP_WRITE &&
 455                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
 456                         entry = ERR_PTR(-EINVAL);
 457                         break;
 458                 }
 459                 if (info->next_index <= index)
 460                         info->next_index = index + 1;
 461         }
 462         if (page) {
 463                 /* another task gave its page, or truncated the file */
 464                 shmem_free_blocks(inode, 1);
 465                 shmem_dir_free(page);
 466         }
 467         if (info->next_index <= index && !IS_ERR(entry))
 468                 info->next_index = index + 1;
 469         return entry;
 470 }
 471
 472 /**
 473  * shmem_free_swp - free some swap entries in a directory
 474  * @dir:        pointer to the directory
 475  * @edir:       pointer after last entry of the directory
 476  * @punch_lock: pointer to spinlock when needed for the holepunch case
 477  */
 478 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
 479                                                 spinlock_t *punch_lock)
 480 {
 481         spinlock_t *punch_unlock = NULL;
 482         swp_entry_t *ptr;
 483         int freed = 0;
 484
 485         for (ptr = dir; ptr < edir; ptr++) {
 486                 if (ptr->val) {
 487                         if (unlikely(punch_lock)) {
 488                                 punch_unlock = punch_lock;
 489                                 punch_lock = NULL;
 490                                 spin_lock(punch_unlock);
 491                                 if (!ptr->val)
 492                                         continue;
 493                         }
 494                         free_swap_and_cache(*ptr);
 495                         *ptr = (swp_entry_t){0};
 496                         freed++;
 497                 }
 498         }
 499         if (punch_unlock)
 500                 spin_unlock(punch_unlock);
 501         return freed;
 502 }
 503
 504 static int shmem_map_and_free_swp(struct page *subdir, int offset,
 505                 int limit, struct page ***dir, spinlock_t *punch_lock)
 506 {
 507         swp_entry_t *ptr;
 508         int freed = 0;
 509
 510         ptr = shmem_swp_map(subdir);
 511         for (; offset < limit; offset += LATENCY_LIMIT) {
 512                 int size = limit - offset;
 513                 if (size > LATENCY_LIMIT)
 514                         size = LATENCY_LIMIT;
 515                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
 516                                                         punch_lock);
 517                 if (need_resched()) {
 518                         shmem_swp_unmap(ptr);
 519                         if (*dir) {
 520                                 shmem_dir_unmap(*dir);
 521                                 *dir = NULL;
 522                         }
 523                         cond_resched();
 524                         ptr = shmem_swp_map(subdir);
 525                 }
 526         }
 527         shmem_swp_unmap(ptr);
 528         return freed;
 529 }
 530
 531 static void shmem_free_pages(struct list_head *next)
 532 {
 533         struct page *page;
 534         int freed = 0;
 535
 536         do {
 537                 page = container_of(next, struct page, lru);
 538                 next = next->next;
 539                 shmem_dir_free(page);
 540                 freed++;
 541                 if (freed >= LATENCY_LIMIT) {
 542                         cond_resched();
 543                         freed = 0;
 544                 }
 545         } while (next);
 546 }
 547
 548 void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
 549 {
 550         struct shmem_inode_info *info = SHMEM_I(inode);
 551         unsigned long idx;
 552         unsigned long size;
 553         unsigned long limit;
 554         unsigned long stage;
 555         unsigned long diroff;
 556         struct page **dir;
 557         struct page *topdir;
 558         struct page *middir;
 559         struct page *subdir;
 560         swp_entry_t *ptr;
 561         LIST_HEAD(pages_to_free);
 562         long nr_pages_to_free = 0;
 563         long nr_swaps_freed = 0;
 564         int offset;
 565         int freed;
 566         int punch_hole;
 567         spinlock_t *needs_lock;
 568         spinlock_t *punch_lock;
 569         unsigned long upper_limit;
 570
 571         truncate_inode_pages_range(inode->i_mapping, start, end);
 572
 573         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 574         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 575         if (idx >= info->next_index)
 576                 return;
 577
 578         spin_lock(&info->lock);
 579         info->flags |= SHMEM_TRUNCATE;
 580         if (likely(end == (loff_t) -1)) {
 581                 limit = info->next_index;
 582                 upper_limit = SHMEM_MAX_INDEX;
 583                 info->next_index = idx;
 584                 needs_lock = NULL;
 585                 punch_hole = 0;
 586         } else {
 587                 if (end + 1 >= inode->i_size) { /* we may free a little more */
 588                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
 589                                                         PAGE_CACHE_SHIFT;
 590                         upper_limit = SHMEM_MAX_INDEX;
 591                 } else {
 592                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
 593                         upper_limit = limit;
 594                 }
 595                 needs_lock = &info->lock;
 596                 punch_hole = 1;
 597         }
 598
 599         topdir = info->i_indirect;
 600         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
 601                 info->i_indirect = NULL;
 602                 nr_pages_to_free++;
 603                 list_add(&topdir->lru, &pages_to_free);
 604         }
 605         spin_unlock(&info->lock);
 606
 607         if (info->swapped && idx < SHMEM_NR_DIRECT) {
 608                 ptr = info->i_direct;
 609                 size = limit;
 610                 if (size > SHMEM_NR_DIRECT)
 611                         size = SHMEM_NR_DIRECT;
 612                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
 613         }
 614
 615         /*
 616          * If there are no indirect blocks or we are punching a hole
 617          * below indirect blocks, nothing to be done.
 618          */
 619         if (!topdir || limit <= SHMEM_NR_DIRECT)
 620                 goto done2;
 621
 622         /*
 623          * The truncation case has already dropped info->lock, and we're safe
 624          * because i_size and next_index have already been lowered, preventing
 625          * access beyond.  But in the punch_hole case, we still need to take
 626          * the lock when updating the swap directory, because there might be
 627          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
 628          * shmem_writepage.  However, whenever we find we can remove a whole
 629          * directory page (not at the misaligned start or end of the range),
 630          * we first NULLify its pointer in the level above, and then have no
 631          * need to take the lock when updating its contents: needs_lock and
 632          * punch_lock (either pointing to info->lock or NULL) manage this.
 633          */
 634
 635         upper_limit -= SHMEM_NR_DIRECT;
 636         limit -= SHMEM_NR_DIRECT;
 637         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
 638         offset = idx % ENTRIES_PER_PAGE;
 639         idx -= offset;
 640
 641         dir = shmem_dir_map(topdir);
 642         stage = ENTRIES_PER_PAGEPAGE/2;
 643         if (idx < ENTRIES_PER_PAGEPAGE/2) {
 644                 middir = topdir;
 645                 diroff = idx/ENTRIES_PER_PAGE;
 646         } else {
 647                 dir += ENTRIES_PER_PAGE/2;
 648                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
 649                 while (stage <= idx)
 650                         stage += ENTRIES_PER_PAGEPAGE;
 651                 middir = *dir;
 652                 if (*dir) {
 653                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
 654                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
 655                         if (!diroff && !offset && upper_limit >= stage) {
 656                                 if (needs_lock) {
 657                                         spin_lock(needs_lock);
 658                                         *dir = NULL;
 659                                         spin_unlock(needs_lock);
 660                                         needs_lock = NULL;
 661                                 } else
 662                                         *dir = NULL;
 663                                 nr_pages_to_free++;
 664                                 list_add(&middir->lru, &pages_to_free);
 665                         }
 666                         shmem_dir_unmap(dir);
 667                         dir = shmem_dir_map(middir);
 668                 } else {
 669                         diroff = 0;
 670                         offset = 0;
 671                         idx = stage;
 672                 }
 673         }
 674
 675         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
 676                 if (unlikely(idx == stage)) {
 677                         shmem_dir_unmap(dir);
 678                         dir = shmem_dir_map(topdir) +
 679                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 680                         while (!*dir) {
 681                                 dir++;
 682                                 idx += ENTRIES_PER_PAGEPAGE;
 683                                 if (idx >= limit)
 684                                         goto done1;
 685                         }
 686                         stage = idx + ENTRIES_PER_PAGEPAGE;
 687                         middir = *dir;
 688                         if (punch_hole)
 689                                 needs_lock = &info->lock;
 690                         if (upper_limit >= stage) {
 691                                 if (needs_lock) {
 692                                         spin_lock(needs_lock);
 693                                         *dir = NULL;
 694                                         spin_unlock(needs_lock);
 695                                         needs_lock = NULL;
 696                                 } else
 697                                         *dir = NULL;
 698                                 nr_pages_to_free++;
 699                                 list_add(&middir->lru, &pages_to_free);
 700                         }
 701                         shmem_dir_unmap(dir);
 702                         cond_resched();
 703                         dir = shmem_dir_map(middir);
 704                         diroff = 0;
 705                 }
 706                 punch_lock = needs_lock;
 707                 subdir = dir[diroff];
 708                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
 709                         if (needs_lock) {
 710                                 spin_lock(needs_lock);
 711                                 dir[diroff] = NULL;
 712                                 spin_unlock(needs_lock);
 713                                 punch_lock = NULL;
 714                         } else
 715                                 dir[diroff] = NULL;
 716                         nr_pages_to_free++;
 717                         list_add(&subdir->lru, &pages_to_free);
 718                 }
 719                 if (subdir && page_private(subdir) /* has swap entries */) {
 720                         size = limit - idx;
 721                         if (size > ENTRIES_PER_PAGE)
 722                                 size = ENTRIES_PER_PAGE;
 723                         freed = shmem_map_and_free_swp(subdir,
 724                                         offset, size, &dir, punch_lock);
 725                         if (!dir)
 726                                 dir = shmem_dir_map(middir);
 727                         nr_swaps_freed += freed;
 728                         if (offset || punch_lock) {
 729                                 spin_lock(&info->lock);
 730                                 set_page_private(subdir,
 731                                         page_private(subdir) - freed);
 732                                 spin_unlock(&info->lock);
 733                         } else
 734                                 BUG_ON(page_private(subdir) != freed);
 735                 }
 736                 offset = 0;
 737         }
 738 done1:
 739         shmem_dir_unmap(dir);
 740 done2:
 741         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
 742                 /*
 743                  * Call truncate_inode_pages again: racing shmem_unuse_inode
 744                  * may have swizzled a page in from swap since
 745                  * truncate_pagecache or generic_delete_inode did it, before we
 746                  * lowered next_index.  Also, though shmem_getpage checks
 747                  * i_size before adding to cache, no recheck after: so fix the
 748                  * narrow window there too.
 749                  */
 750                 truncate_inode_pages_range(inode->i_mapping, start, end);
 751         }
 752
 753         spin_lock(&info->lock);
 754         info->flags &= ~SHMEM_TRUNCATE;
 755         info->swapped -= nr_swaps_freed;
 756         if (nr_pages_to_free)
 757                 shmem_free_blocks(inode, nr_pages_to_free);
 758         shmem_recalc_inode(inode);
 759         spin_unlock(&info->lock);
 760
 761         /*
 762          * Empty swap vector directory pages to be freed?
 763          */
 764         if (!list_empty(&pages_to_free)) {
 765                 pages_to_free.prev->next = NULL;
 766                 shmem_free_pages(pages_to_free.next);
 767         }
 768 }
 769 EXPORT_SYMBOL_GPL(shmem_truncate_range);
 770
 771 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
 772 {
 773         struct inode *inode = dentry->d_inode;
 774         int error;
 775
 776         error = inode_change_ok(inode, attr);
 777         if (error)
 778                 return error;
 779
 780         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
 781                 loff_t oldsize = inode->i_size;
 782                 loff_t newsize = attr->ia_size;
 783                 struct page *page = NULL;
 784
 785                 if (newsize < oldsize) {
 786                         /*
 787                          * If truncating down to a partial page, then
 788                          * if that page is already allocated, hold it
 789                          * in memory until the truncation is over, so
 790                          * truncate_partial_page cannot miss it were
 791                          * it assigned to swap.
 792                          */
 793                         if (newsize & (PAGE_CACHE_SIZE-1)) {
 794                                 (void) shmem_getpage(inode,
 795                                         newsize >> PAGE_CACHE_SHIFT,
 796                                                 &page, SGP_READ, NULL);
 797                                 if (page)
 798                                         unlock_page(page);
 799                         }
 800                         /*
 801                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
 802                          * detect if any pages might have been added to cache
 803                          * after truncate_inode_pages.  But we needn't bother
 804                          * if it's being fully truncated to zero-length: the
 805                          * nrpages check is efficient enough in that case.
 806                          */
 807                         if (newsize) {
 808                                 struct shmem_inode_info *info = SHMEM_I(inode);
 809                                 spin_lock(&info->lock);
 810                                 info->flags &= ~SHMEM_PAGEIN;
 811                                 spin_unlock(&info->lock);
 812                         }
 813                 }
 814                 if (newsize != oldsize) {
 815                         i_size_write(inode, newsize);
 816                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 817                 }
 818                 if (newsize < oldsize) {
 819                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
 820                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 821                         shmem_truncate_range(inode, newsize, (loff_t)-1);
 822                         /* unmap again to remove racily COWed private pages */
 823                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 824                 }
 825                 if (page)
 826                         page_cache_release(page);
 827         }
 828
 829         setattr_copy(inode, attr);
 830 #ifdef CONFIG_TMPFS_POSIX_ACL
 831         if (attr->ia_valid & ATTR_MODE)
 832                 error = generic_acl_chmod(inode);
 833 #endif
 834         return error;
 835 }
 836
 837 static void shmem_evict_inode(struct inode *inode)
 838 {
 839         struct shmem_inode_info *info = SHMEM_I(inode);
 840         struct shmem_xattr *xattr, *nxattr;
 841
 842         if (inode->i_mapping->a_ops == &shmem_aops) {
 843                 shmem_unacct_size(info->flags, inode->i_size);
 844                 inode->i_size = 0;
 845                 shmem_truncate_range(inode, 0, (loff_t)-1);
 846                 if (!list_empty(&info->swaplist)) {
 847                         mutex_lock(&shmem_swaplist_mutex);
 848                         list_del_init(&info->swaplist);
 849                         mutex_unlock(&shmem_swaplist_mutex);
 850                 }
 851         }
 852
 853         list_for_each_entry_safe(xattr, nxattr, &info->xattr_list, list) {
 854                 kfree(xattr->name);
 855                 kfree(xattr);
 856         }
 857         BUG_ON(inode->i_blocks);
 858         shmem_free_inode(inode->i_sb);
 859         end_writeback(inode);
 860 }
 861
 862 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
 863 {
 864         swp_entry_t *ptr;
 865
 866         for (ptr = dir; ptr < edir; ptr++) {
 867                 if (ptr->val == entry.val)
 868                         return ptr - dir;
 869         }
 870         return -1;
 871 }
 872
 873 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
 874 {
 875         struct address_space *mapping;
 876         unsigned long idx;
 877         unsigned long size;
 878         unsigned long limit;
 879         unsigned long stage;
 880         struct page **dir;
 881         struct page *subdir;
 882         swp_entry_t *ptr;
 883         int offset;
 884         int error;
 885
 886         idx = 0;
 887         ptr = info->i_direct;
 888         spin_lock(&info->lock);
 889         if (!info->swapped) {
 890                 list_del_init(&info->swaplist);
 891                 goto lost2;
 892         }
 893         limit = info->next_index;
 894         size = limit;
 895         if (size > SHMEM_NR_DIRECT)
 896                 size = SHMEM_NR_DIRECT;
 897         offset = shmem_find_swp(entry, ptr, ptr+size);
 898         if (offset >= 0) {
 899                 shmem_swp_balance_unmap();
 900                 goto found;
 901         }
 902         if (!info->i_indirect)
 903                 goto lost2;
 904
 905         dir = shmem_dir_map(info->i_indirect);
 906         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
 907
 908         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
 909                 if (unlikely(idx == stage)) {
 910                         shmem_dir_unmap(dir-1);
 911                         if (cond_resched_lock(&info->lock)) {
 912                                 /* check it has not been truncated */
 913                                 if (limit > info->next_index) {
 914                                         limit = info->next_index;
 915                                         if (idx >= limit)
 916                                                 goto lost2;
 917                                 }
 918                         }
 919                         dir = shmem_dir_map(info->i_indirect) +
 920                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 921                         while (!*dir) {
 922                                 dir++;
 923                                 idx += ENTRIES_PER_PAGEPAGE;
 924                                 if (idx >= limit)
 925                                         goto lost1;
 926                         }
 927                         stage = idx + ENTRIES_PER_PAGEPAGE;
 928                         subdir = *dir;
 929                         shmem_dir_unmap(dir);
 930                         dir = shmem_dir_map(subdir);
 931                 }
 932                 subdir = *dir;
 933                 if (subdir && page_private(subdir)) {
 934                         ptr = shmem_swp_map(subdir);
 935                         size = limit - idx;
 936                         if (size > ENTRIES_PER_PAGE)
 937                                 size = ENTRIES_PER_PAGE;
 938                         offset = shmem_find_swp(entry, ptr, ptr+size);
 939                         shmem_swp_unmap(ptr);
 940                         if (offset >= 0) {
 941                                 shmem_dir_unmap(dir);
 942                                 ptr = shmem_swp_map(subdir);
 943                                 goto found;
 944                         }
 945                 }
 946         }
 947 lost1:
 948         shmem_dir_unmap(dir-1);
 949 lost2:
 950         spin_unlock(&info->lock);
 951         return 0;
 952 found:
 953         idx += offset;
 954         ptr += offset;
 955
 956         /*
 957          * Move _head_ to start search for next from here.
 958          * But be careful: shmem_evict_inode checks list_empty without taking
 959          * mutex, and there's an instant in list_move_tail when info->swaplist
 960          * would appear empty, if it were the only one on shmem_swaplist.  We
 961          * could avoid doing it if inode NULL; or use this minor optimization.
 962          */
 963         if (shmem_swaplist.next != &info->swaplist)
 964                 list_move_tail(&shmem_swaplist, &info->swaplist);
 965
 966         /*
 967          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 968          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 969          * beneath us (pagelock doesn't help until the page is in pagecache).
 970          */
 971         mapping = info->vfs_inode.i_mapping;
 972         error = add_to_page_cache_locked(page, mapping, idx, GFP_NOWAIT);
 973         /* which does mem_cgroup_uncharge_cache_page on error */
 974
 975         if (error == -EEXIST) {
 976                 struct page *filepage = find_get_page(mapping, idx);
 977                 error = 1;
 978                 if (filepage) {
 979                         /*
 980                          * There might be a more uptodate page coming down
 981                          * from a stacked writepage: forget our swappage if so.
 982                          */
 983                         if (PageUptodate(filepage))
 984                                 error = 0;
 985                         page_cache_release(filepage);
 986                 }
 987         }
 988         if (!error) {
 989                 delete_from_swap_cache(page);
 990                 set_page_dirty(page);
 991                 info->flags |= SHMEM_PAGEIN;
 992                 shmem_swp_set(info, ptr, 0);
 993                 swap_free(entry);
 994                 error = 1;      /* not an error, but entry was found */
 995         }
 996         shmem_swp_unmap(ptr);
 997         spin_unlock(&info->lock);
 998         return error;
 999 }
1000
1001 /*
1002  * shmem_unuse() search for an eventually swapped out shmem page.
1003  */
1004 int shmem_unuse(swp_entry_t entry, struct page *page)
1005 {
1006         struct list_head *p, *next;
1007         struct shmem_inode_info *info;
1008         int found = 0;
1009         int error;
1010
1011         /*
1012          * Charge page using GFP_KERNEL while we can wait, before taking
1013          * the shmem_swaplist_mutex which might hold up shmem_writepage().
1014          * Charged back to the user (not to caller) when swap account is used.
1015          * add_to_page_cache() will be called with GFP_NOWAIT.
1016          */
1017         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
1018         if (error)
1019                 goto out;
1020         /*
1021          * Try to preload while we can wait, to not make a habit of
1022          * draining atomic reserves; but don't latch on to this cpu,
1023          * it's okay if sometimes we get rescheduled after this.
1024          */
1025         error = radix_tree_preload(GFP_KERNEL);
1026         if (error)
1027                 goto uncharge;
1028         radix_tree_preload_end();
1029
1030         mutex_lock(&shmem_swaplist_mutex);
1031         list_for_each_safe(p, next, &shmem_swaplist) {
1032                 info = list_entry(p, struct shmem_inode_info, swaplist);
1033                 found = shmem_unuse_inode(info, entry, page);
1034                 cond_resched();
1035                 if (found)
1036                         break;
1037         }
1038         mutex_unlock(&shmem_swaplist_mutex);
1039
1040 uncharge:
1041         if (!found)
1042                 mem_cgroup_uncharge_cache_page(page);
1043         if (found < 0)
1044                 error = found;
1045 out:
1046         unlock_page(page);
1047         page_cache_release(page);
1048         return error;
1049 }
1050
1051 /*
1052  * Move the page from the page cache to the swap cache.
1053  */
1054 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1055 {
1056         struct shmem_inode_info *info;
1057         swp_entry_t *entry, swap;
1058         struct address_space *mapping;
1059         unsigned long index;
1060         struct inode *inode;
1061
1062         BUG_ON(!PageLocked(page));
1063         mapping = page->mapping;
1064         index = page->index;
1065         inode = mapping->host;
1066         info = SHMEM_I(inode);
1067         if (info->flags & VM_LOCKED)
1068                 goto redirty;
1069         if (!total_swap_pages)
1070                 goto redirty;
1071
1072         /*
1073          * shmem_backing_dev_info's capabilities prevent regular writeback or
1074          * sync from ever calling shmem_writepage; but a stacking filesystem
1075          * may use the ->writepage of its underlying filesystem, in which case
1076          * tmpfs should write out to swap only in response to memory pressure,
1077          * and not for the writeback threads or sync.  However, in those cases,
1078          * we do still want to check if there's a redundant swappage to be
1079          * discarded.
1080          */
1081         if (wbc->for_reclaim)
1082                 swap = get_swap_page();
1083         else
1084                 swap.val = 0;
1085
1086         /*
1087          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1088          * if it's not already there.  Do it now because we cannot take
1089          * mutex while holding spinlock, and must do so before the page
1090          * is moved to swap cache, when its pagelock no longer protects
1091          * the inode from eviction.  But don't unlock the mutex until
1092          * we've taken the spinlock, because shmem_unuse_inode() will
1093          * prune a !swapped inode from the swaplist under both locks.
1094          */
1095         if (swap.val) {
1096                 mutex_lock(&shmem_swaplist_mutex);
1097                 if (list_empty(&info->swaplist))
1098                         list_add_tail(&info->swaplist, &shmem_swaplist);
1099         }
1100
1101         spin_lock(&info->lock);
1102         if (swap.val)
1103                 mutex_unlock(&shmem_swaplist_mutex);
1104
1105         if (index >= info->next_index) {
1106                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1107                 goto unlock;
1108         }
1109         entry = shmem_swp_entry(info, index, NULL);
1110         if (entry->val) {
1111                 /*
1112                  * The more uptodate page coming down from a stacked
1113                  * writepage should replace our old swappage.
1114                  */
1115                 free_swap_and_cache(*entry);
1116                 shmem_swp_set(info, entry, 0);
1117         }
1118         shmem_recalc_inode(inode);
1119
1120         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1121                 delete_from_page_cache(page);
1122                 shmem_swp_set(info, entry, swap.val);
1123                 shmem_swp_unmap(entry);
1124                 swap_shmem_alloc(swap);
1125                 spin_unlock(&info->lock);
1126                 BUG_ON(page_mapped(page));
1127                 swap_writepage(page, wbc);
1128                 return 0;
1129         }
1130
1131         shmem_swp_unmap(entry);
1132 unlock:
1133         spin_unlock(&info->lock);
1134         /*
1135          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1136          * clear SWAP_HAS_CACHE flag.
1137          */
1138         swapcache_free(swap, NULL);
1139 redirty:
1140         set_page_dirty(page);
1141         if (wbc->for_reclaim)
1142                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1143         unlock_page(page);
1144         return 0;
1145 }
1146
1147 #ifdef CONFIG_NUMA
1148 #ifdef CONFIG_TMPFS
1149 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1150 {
1151         char buffer[64];
1152
1153         if (!mpol || mpol->mode == MPOL_DEFAULT)
1154                 return;         /* show nothing */
1155
1156         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1157
1158         seq_printf(seq, ",mpol=%s", buffer);
1159 }
1160
1161 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1162 {
1163         struct mempolicy *mpol = NULL;
1164         if (sbinfo->mpol) {
1165                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1166                 mpol = sbinfo->mpol;
1167                 mpol_get(mpol);
1168                 spin_unlock(&sbinfo->stat_lock);
1169         }
1170         return mpol;
1171 }
1172 #endif /* CONFIG_TMPFS */
1173
1174 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1175                         struct shmem_inode_info *info, unsigned long idx)
1176 {
1177         struct mempolicy mpol, *spol;
1178         struct vm_area_struct pvma;
1179         struct page *page;
1180
1181         spol = mpol_cond_copy(&mpol,
1182                                 mpol_shared_policy_lookup(&info->policy, idx));
1183
1184         /* Create a pseudo vma that just contains the policy */
1185         pvma.vm_start = 0;
1186         pvma.vm_pgoff = idx;
1187         pvma.vm_ops = NULL;
1188         pvma.vm_policy = spol;
1189         page = swapin_readahead(entry, gfp, &pvma, 0);
1190         return page;
1191 }
1192
1193 static struct page *shmem_alloc_page(gfp_t gfp,
1194                         struct shmem_inode_info *info, unsigned long idx)
1195 {
1196         struct vm_area_struct pvma;
1197
1198         /* Create a pseudo vma that just contains the policy */
1199         pvma.vm_start = 0;
1200         pvma.vm_pgoff = idx;
1201         pvma.vm_ops = NULL;
1202         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1203
1204         /*
1205          * alloc_page_vma() will drop the shared policy reference
1206          */
1207         return alloc_page_vma(gfp, &pvma, 0);
1208 }
1209 #else /* !CONFIG_NUMA */
1210 #ifdef CONFIG_TMPFS
1211 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1212 {
1213 }
1214 #endif /* CONFIG_TMPFS */
1215
1216 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1217                         struct shmem_inode_info *info, unsigned long idx)
1218 {
1219         return swapin_readahead(entry, gfp, NULL, 0);
1220 }
1221
1222 static inline struct page *shmem_alloc_page(gfp_t gfp,
1223                         struct shmem_inode_info *info, unsigned long idx)
1224 {
1225         return alloc_page(gfp);
1226 }
1227 #endif /* CONFIG_NUMA */
1228
1229 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1230 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1231 {
1232         return NULL;
1233 }
1234 #endif
1235
1236 /*
1237  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1238  *
1239  * If we allocate a new one we do not mark it dirty. That's up to the
1240  * vm. If we swap it in we mark it dirty since we also free the swap
1241  * entry since a page cannot live in both the swap and page cache
1242  */
1243 static int shmem_getpage_gfp(struct inode *inode, pgoff_t idx,
1244         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1245 {
1246         struct address_space *mapping = inode->i_mapping;
1247         struct shmem_inode_info *info = SHMEM_I(inode);
1248         struct shmem_sb_info *sbinfo;
1249         struct page *filepage;
1250         struct page *swappage;
1251         struct page *prealloc_page = NULL;
1252         swp_entry_t *entry;
1253         swp_entry_t swap;
1254         int error;
1255
1256         if (idx >= SHMEM_MAX_INDEX)
1257                 return -EFBIG;
1258 repeat:
1259         filepage = find_lock_page(mapping, idx);
1260         if (filepage && PageUptodate(filepage))
1261                 goto done;
1262         if (!filepage) {
1263                 /*
1264                  * Try to preload while we can wait, to not make a habit of
1265                  * draining atomic reserves; but don't latch on to this cpu.
1266                  */
1267                 error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
1268                 if (error)
1269                         goto failed;
1270                 radix_tree_preload_end();
1271                 if (sgp != SGP_READ && !prealloc_page) {
1272                         /* We don't care if this fails */
1273                         prealloc_page = shmem_alloc_page(gfp, info, idx);
1274                         if (prealloc_page) {
1275                                 if (mem_cgroup_cache_charge(prealloc_page,
1276                                                 current->mm, GFP_KERNEL)) {
1277                                         page_cache_release(prealloc_page);
1278                                         prealloc_page = NULL;
1279                                 }
1280                         }
1281                 }
1282         }
1283         error = 0;
1284
1285         spin_lock(&info->lock);
1286         shmem_recalc_inode(inode);
1287         entry = shmem_swp_alloc(info, idx, sgp, gfp);
1288         if (IS_ERR(entry)) {
1289                 spin_unlock(&info->lock);
1290                 error = PTR_ERR(entry);
1291                 goto failed;
1292         }
1293         swap = *entry;
1294
1295         if (swap.val) {
1296                 /* Look it up and read it in.. */
1297                 swappage = lookup_swap_cache(swap);
1298                 if (!swappage) {
1299                         shmem_swp_unmap(entry);
1300                         spin_unlock(&info->lock);
1301                         /* here we actually do the io */
1302                         if (fault_type)
1303                                 *fault_type |= VM_FAULT_MAJOR;
1304                         swappage = shmem_swapin(swap, gfp, info, idx);
1305                         if (!swappage) {
1306                                 spin_lock(&info->lock);
1307                                 entry = shmem_swp_alloc(info, idx, sgp, gfp);
1308                                 if (IS_ERR(entry))
1309                                         error = PTR_ERR(entry);
1310                                 else {
1311                                         if (entry->val == swap.val)
1312                                                 error = -ENOMEM;
1313                                         shmem_swp_unmap(entry);
1314                                 }
1315                                 spin_unlock(&info->lock);
1316                                 if (error)
1317                                         goto failed;
1318                                 goto repeat;
1319                         }
1320                         wait_on_page_locked(swappage);
1321                         page_cache_release(swappage);
1322                         goto repeat;
1323                 }
1324
1325                 /* We have to do this with page locked to prevent races */
1326                 if (!trylock_page(swappage)) {
1327                         shmem_swp_unmap(entry);
1328                         spin_unlock(&info->lock);
1329                         wait_on_page_locked(swappage);
1330                         page_cache_release(swappage);
1331                         goto repeat;
1332                 }
1333                 if (PageWriteback(swappage)) {
1334                         shmem_swp_unmap(entry);
1335                         spin_unlock(&info->lock);
1336                         wait_on_page_writeback(swappage);
1337                         unlock_page(swappage);
1338                         page_cache_release(swappage);
1339                         goto repeat;
1340                 }
1341                 if (!PageUptodate(swappage)) {
1342                         shmem_swp_unmap(entry);
1343                         spin_unlock(&info->lock);
1344                         unlock_page(swappage);
1345                         page_cache_release(swappage);
1346                         error = -EIO;
1347                         goto failed;
1348                 }
1349
1350                 if (filepage) {
1351                         shmem_swp_set(info, entry, 0);
1352                         shmem_swp_unmap(entry);
1353                         delete_from_swap_cache(swappage);
1354                         spin_unlock(&info->lock);
1355                         copy_highpage(filepage, swappage);
1356                         unlock_page(swappage);
1357                         page_cache_release(swappage);
1358                         flush_dcache_page(filepage);
1359                         SetPageUptodate(filepage);
1360                         set_page_dirty(filepage);
1361                         swap_free(swap);
1362                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1363                                         idx, GFP_NOWAIT))) {
1364                         info->flags |= SHMEM_PAGEIN;
1365                         shmem_swp_set(info, entry, 0);
1366                         shmem_swp_unmap(entry);
1367                         delete_from_swap_cache(swappage);
1368                         spin_unlock(&info->lock);
1369                         filepage = swappage;
1370                         set_page_dirty(filepage);
1371                         swap_free(swap);
1372                 } else {
1373                         shmem_swp_unmap(entry);
1374                         spin_unlock(&info->lock);
1375                         if (error == -ENOMEM) {
1376                                 /*
1377                                  * reclaim from proper memory cgroup and
1378                                  * call memcg's OOM if needed.
1379                                  */
1380                                 error = mem_cgroup_shmem_charge_fallback(
1381                                                                 swappage,
1382                                                                 current->mm,
1383                                                                 gfp);
1384                                 if (error) {
1385                                         unlock_page(swappage);
1386                                         page_cache_release(swappage);
1387                                         goto failed;
1388                                 }
1389                         }
1390                         unlock_page(swappage);
1391                         page_cache_release(swappage);
1392                         goto repeat;
1393                 }
1394         } else if (sgp == SGP_READ && !filepage) {
1395                 shmem_swp_unmap(entry);
1396                 filepage = find_get_page(mapping, idx);
1397                 if (filepage &&
1398                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1399                         spin_unlock(&info->lock);
1400                         wait_on_page_locked(filepage);
1401                         page_cache_release(filepage);
1402                         filepage = NULL;
1403                         goto repeat;
1404                 }
1405                 spin_unlock(&info->lock);
1406         } else {
1407                 shmem_swp_unmap(entry);
1408                 sbinfo = SHMEM_SB(inode->i_sb);
1409                 if (sbinfo->max_blocks) {
1410                         if (percpu_counter_compare(&sbinfo->used_blocks,
1411                                                 sbinfo->max_blocks) >= 0 ||
1412                             shmem_acct_block(info->flags))
1413                                 goto nospace;
1414                         percpu_counter_inc(&sbinfo->used_blocks);
1415                         inode->i_blocks += BLOCKS_PER_PAGE;
1416                 } else if (shmem_acct_block(info->flags))
1417                         goto nospace;
1418
1419                 if (!filepage) {
1420                         int ret;
1421
1422                         if (!prealloc_page) {
1423                                 spin_unlock(&info->lock);
1424                                 filepage = shmem_alloc_page(gfp, info, idx);
1425                                 if (!filepage) {
1426                                         spin_lock(&info->lock);
1427                                         shmem_unacct_blocks(info->flags, 1);
1428                                         shmem_free_blocks(inode, 1);
1429                                         spin_unlock(&info->lock);
1430                                         error = -ENOMEM;
1431                                         goto failed;
1432                                 }
1433                                 SetPageSwapBacked(filepage);
1434
1435                                 /*
1436                                  * Precharge page while we can wait, compensate
1437                                  * after
1438                                  */
1439                                 error = mem_cgroup_cache_charge(filepage,
1440                                         current->mm, GFP_KERNEL);
1441                                 if (error) {
1442                                         page_cache_release(filepage);
1443                                         spin_lock(&info->lock);
1444                                         shmem_unacct_blocks(info->flags, 1);
1445                                         shmem_free_blocks(inode, 1);
1446                                         spin_unlock(&info->lock);
1447                                         filepage = NULL;
1448                                         goto failed;
1449                                 }
1450
1451                                 spin_lock(&info->lock);
1452                         } else {
1453                                 filepage = prealloc_page;
1454                                 prealloc_page = NULL;
1455                                 SetPageSwapBacked(filepage);
1456                         }
1457
1458                         entry = shmem_swp_alloc(info, idx, sgp, gfp);
1459                         if (IS_ERR(entry))
1460                                 error = PTR_ERR(entry);
1461                         else {
1462                                 swap = *entry;
1463                                 shmem_swp_unmap(entry);
1464                         }
1465                         ret = error || swap.val;
1466                         if (ret)
1467                                 mem_cgroup_uncharge_cache_page(filepage);
1468                         else
1469                                 ret = add_to_page_cache_lru(filepage, mapping,
1470                                                 idx, GFP_NOWAIT);
1471                         /*
1472                          * At add_to_page_cache_lru() failure, uncharge will
1473                          * be done automatically.
1474                          */
1475                         if (ret) {
1476                                 shmem_unacct_blocks(info->flags, 1);
1477                                 shmem_free_blocks(inode, 1);
1478                                 spin_unlock(&info->lock);
1479                                 page_cache_release(filepage);
1480                                 filepage = NULL;
1481                                 if (error)
1482                                         goto failed;
1483                                 goto repeat;
1484                         }
1485                         info->flags |= SHMEM_PAGEIN;
1486                 }
1487
1488                 info->alloced++;
1489                 spin_unlock(&info->lock);
1490                 clear_highpage(filepage);
1491                 flush_dcache_page(filepage);
1492                 SetPageUptodate(filepage);
1493                 if (sgp == SGP_DIRTY)
1494                         set_page_dirty(filepage);
1495         }
1496 done:
1497         *pagep = filepage;
1498         error = 0;
1499         goto out;
1500
1501 nospace:
1502         /*
1503          * Perhaps the page was brought in from swap between find_lock_page
1504          * and taking info->lock?  We allow for that at add_to_page_cache_lru,
1505          * but must also avoid reporting a spurious ENOSPC while working on a
1506          * full tmpfs.
1507          */
1508         if (!filepage) {
1509                 struct page *page = find_get_page(mapping, idx);
1510                 if (page) {
1511                         spin_unlock(&info->lock);
1512                         page_cache_release(page);
1513                         goto repeat;
1514                 }
1515         }
1516         spin_unlock(&info->lock);
1517         error = -ENOSPC;
1518 failed:
1519         if (filepage) {
1520                 unlock_page(filepage);
1521                 page_cache_release(filepage);
1522         }
1523 out:
1524         if (prealloc_page) {
1525                 mem_cgroup_uncharge_cache_page(prealloc_page);
1526                 page_cache_release(prealloc_page);
1527         }
1528         return error;
1529 }
1530
1531 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1532 {
1533         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1534         int error;
1535         int ret = VM_FAULT_LOCKED;
1536
1537         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1538                 return VM_FAULT_SIGBUS;
1539
1540         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1541         if (error)
1542                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1543
1544         if (ret & VM_FAULT_MAJOR) {
1545                 count_vm_event(PGMAJFAULT);
1546                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1547         }
1548         return ret;
1549 }
1550
1551 #ifdef CONFIG_NUMA
1552 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1553 {
1554         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1555         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1556 }
1557
1558 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1559                                           unsigned long addr)
1560 {
1561         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1562         unsigned long idx;
1563
1564         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1565         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1566 }
1567 #endif
1568
1569 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1570 {
1571         struct inode *inode = file->f_path.dentry->d_inode;
1572         struct shmem_inode_info *info = SHMEM_I(inode);
1573         int retval = -ENOMEM;
1574
1575         spin_lock(&info->lock);
1576         if (lock && !(info->flags & VM_LOCKED)) {
1577                 if (!user_shm_lock(inode->i_size, user))
1578                         goto out_nomem;
1579                 info->flags |= VM_LOCKED;
1580                 mapping_set_unevictable(file->f_mapping);
1581         }
1582         if (!lock && (info->flags & VM_LOCKED) && user) {
1583                 user_shm_unlock(inode->i_size, user);
1584                 info->flags &= ~VM_LOCKED;
1585                 mapping_clear_unevictable(file->f_mapping);
1586                 scan_mapping_unevictable_pages(file->f_mapping);
1587         }
1588         retval = 0;
1589
1590 out_nomem:
1591         spin_unlock(&info->lock);
1592         return retval;
1593 }
1594
1595 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1596 {
1597         file_accessed(file);
1598         vma->vm_ops = &shmem_vm_ops;
1599         vma->vm_flags |= VM_CAN_NONLINEAR;
1600         return 0;
1601 }
1602
1603 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1604                                      int mode, dev_t dev, unsigned long flags)
1605 {
1606         struct inode *inode;
1607         struct shmem_inode_info *info;
1608         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1609
1610         if (shmem_reserve_inode(sb))
1611                 return NULL;
1612
1613         inode = new_inode(sb);
1614         if (inode) {
1615                 inode->i_ino = get_next_ino();
1616                 inode_init_owner(inode, dir, mode);
1617                 inode->i_blocks = 0;
1618                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1619                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1620                 inode->i_generation = get_seconds();
1621                 info = SHMEM_I(inode);
1622                 memset(info, 0, (char *)inode - (char *)info);
1623                 spin_lock_init(&info->lock);
1624                 info->flags = flags & VM_NORESERVE;
1625                 INIT_LIST_HEAD(&info->swaplist);
1626                 INIT_LIST_HEAD(&info->xattr_list);
1627                 cache_no_acl(inode);
1628
1629                 switch (mode & S_IFMT) {
1630                 default:
1631                         inode->i_op = &shmem_special_inode_operations;
1632                         init_special_inode(inode, mode, dev);
1633                         break;
1634                 case S_IFREG:
1635                         inode->i_mapping->a_ops = &shmem_aops;
1636                         inode->i_op = &shmem_inode_operations;
1637                         inode->i_fop = &shmem_file_operations;
1638                         mpol_shared_policy_init(&info->policy,
1639                                                  shmem_get_sbmpol(sbinfo));
1640                         break;
1641                 case S_IFDIR:
1642                         inc_nlink(inode);
1643                         /* Some things misbehave if size == 0 on a directory */
1644                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1645                         inode->i_op = &shmem_dir_inode_operations;
1646                         inode->i_fop = &simple_dir_operations;
1647                         break;
1648                 case S_IFLNK:
1649                         /*
1650                          * Must not load anything in the rbtree,
1651                          * mpol_free_shared_policy will not be called.
1652                          */
1653                         mpol_shared_policy_init(&info->policy, NULL);
1654                         break;
1655                 }
1656         } else
1657                 shmem_free_inode(sb);
1658         return inode;
1659 }
1660
1661 #ifdef CONFIG_TMPFS
1662 static const struct inode_operations shmem_symlink_inode_operations;
1663 static const struct inode_operations shmem_symlink_inline_operations;
1664
1665 static int
1666 shmem_write_begin(struct file *file, struct address_space *mapping,
1667                         loff_t pos, unsigned len, unsigned flags,
1668                         struct page **pagep, void **fsdata)
1669 {
1670         struct inode *inode = mapping->host;
1671         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1672         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1673 }
1674
1675 static int
1676 shmem_write_end(struct file *file, struct address_space *mapping,
1677                         loff_t pos, unsigned len, unsigned copied,
1678                         struct page *page, void *fsdata)
1679 {
1680         struct inode *inode = mapping->host;
1681
1682         if (pos + copied > inode->i_size)
1683                 i_size_write(inode, pos + copied);
1684
1685         set_page_dirty(page);
1686         unlock_page(page);
1687         page_cache_release(page);
1688
1689         return copied;
1690 }
1691
1692 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1693 {
1694         struct inode *inode = filp->f_path.dentry->d_inode;
1695         struct address_space *mapping = inode->i_mapping;
1696         unsigned long index, offset;
1697         enum sgp_type sgp = SGP_READ;
1698
1699         /*
1700          * Might this read be for a stacking filesystem?  Then when reading
1701          * holes of a sparse file, we actually need to allocate those pages,
1702          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1703          */
1704         if (segment_eq(get_fs(), KERNEL_DS))
1705                 sgp = SGP_DIRTY;
1706
1707         index = *ppos >> PAGE_CACHE_SHIFT;
1708         offset = *ppos & ~PAGE_CACHE_MASK;
1709
1710         for (;;) {
1711                 struct page *page = NULL;
1712                 unsigned long end_index, nr, ret;
1713                 loff_t i_size = i_size_read(inode);
1714
1715                 end_index = i_size >> PAGE_CACHE_SHIFT;
1716                 if (index > end_index)
1717                         break;
1718                 if (index == end_index) {
1719                         nr = i_size & ~PAGE_CACHE_MASK;
1720                         if (nr <= offset)
1721                                 break;
1722                 }
1723
1724                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1725                 if (desc->error) {
1726                         if (desc->error == -EINVAL)
1727                                 desc->error = 0;
1728                         break;
1729                 }
1730                 if (page)
1731                         unlock_page(page);
1732
1733                 /*
1734                  * We must evaluate after, since reads (unlike writes)
1735                  * are called without i_mutex protection against truncate
1736                  */
1737                 nr = PAGE_CACHE_SIZE;
1738                 i_size = i_size_read(inode);
1739                 end_index = i_size >> PAGE_CACHE_SHIFT;
1740                 if (index == end_index) {
1741                         nr = i_size & ~PAGE_CACHE_MASK;
1742                         if (nr <= offset) {
1743                                 if (page)
1744                                         page_cache_release(page);
1745                                 break;
1746                         }
1747                 }
1748                 nr -= offset;
1749
1750                 if (page) {
1751                         /*
1752                          * If users can be writing to this page using arbitrary
1753                          * virtual addresses, take care about potential aliasing
1754                          * before reading the page on the kernel side.
1755                          */
1756                         if (mapping_writably_mapped(mapping))
1757                                 flush_dcache_page(page);
1758                         /*
1759                          * Mark the page accessed if we read the beginning.
1760                          */
1761                         if (!offset)
1762                                 mark_page_accessed(page);
1763                 } else {
1764                         page = ZERO_PAGE(0);
1765                         page_cache_get(page);
1766                 }
1767
1768                 /*
1769                  * Ok, we have the page, and it's up-to-date, so
1770                  * now we can copy it to user space...
1771                  *
1772                  * The actor routine returns how many bytes were actually used..
1773                  * NOTE! This may not be the same as how much of a user buffer
1774                  * we filled up (we may be padding etc), so we can only update
1775                  * "pos" here (the actor routine has to update the user buffer
1776                  * pointers and the remaining count).
1777                  */
1778                 ret = actor(desc, page, offset, nr);
1779                 offset += ret;
1780                 index += offset >> PAGE_CACHE_SHIFT;
1781                 offset &= ~PAGE_CACHE_MASK;
1782
1783                 page_cache_release(page);
1784                 if (ret != nr || !desc->count)
1785                         break;
1786
1787                 cond_resched();
1788         }
1789
1790         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1791         file_accessed(filp);
1792 }
1793
1794 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1795                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1796 {
1797         struct file *filp = iocb->ki_filp;
1798         ssize_t retval;
1799         unsigned long seg;
1800         size_t count;
1801         loff_t *ppos = &iocb->ki_pos;
1802
1803         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1804         if (retval)
1805                 return retval;
1806
1807         for (seg = 0; seg < nr_segs; seg++) {
1808                 read_descriptor_t desc;
1809
1810                 desc.written = 0;
1811                 desc.arg.buf = iov[seg].iov_base;
1812                 desc.count = iov[seg].iov_len;
1813                 if (desc.count == 0)
1814                         continue;
1815                 desc.error = 0;
1816                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1817                 retval += desc.written;
1818                 if (desc.error) {
1819                         retval = retval ?: desc.error;
1820                         break;
1821                 }
1822                 if (desc.count > 0)
1823                         break;
1824         }
1825         return retval;
1826 }
1827
1828 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1829                                 struct pipe_inode_info *pipe, size_t len,
1830                                 unsigned int flags)
1831 {
1832         struct address_space *mapping = in->f_mapping;
1833         struct inode *inode = mapping->host;
1834         unsigned int loff, nr_pages, req_pages;
1835         struct page *pages[PIPE_DEF_BUFFERS];
1836         struct partial_page partial[PIPE_DEF_BUFFERS];
1837         struct page *page;
1838         pgoff_t index, end_index;
1839         loff_t isize, left;
1840         int error, page_nr;
1841         struct splice_pipe_desc spd = {
1842                 .pages = pages,
1843                 .partial = partial,
1844                 .flags = flags,
1845                 .ops = &page_cache_pipe_buf_ops,
1846                 .spd_release = spd_release_page,
1847         };
1848
1849         isize = i_size_read(inode);
1850         if (unlikely(*ppos >= isize))
1851                 return 0;
1852
1853         left = isize - *ppos;
1854         if (unlikely(left < len))
1855                 len = left;
1856
1857         if (splice_grow_spd(pipe, &spd))
1858                 return -ENOMEM;
1859
1860         index = *ppos >> PAGE_CACHE_SHIFT;
1861         loff = *ppos & ~PAGE_CACHE_MASK;
1862         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1863         nr_pages = min(req_pages, pipe->buffers);
1864
1865         spd.nr_pages = find_get_pages_contig(mapping, index,
1866                                                 nr_pages, spd.pages);
1867         index += spd.nr_pages;
1868         error = 0;
1869
1870         while (spd.nr_pages < nr_pages) {
1871                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1872                 if (error)
1873                         break;
1874                 unlock_page(page);
1875                 spd.pages[spd.nr_pages++] = page;
1876                 index++;
1877         }
1878
1879         index = *ppos >> PAGE_CACHE_SHIFT;
1880         nr_pages = spd.nr_pages;
1881         spd.nr_pages = 0;
1882
1883         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1884                 unsigned int this_len;
1885
1886                 if (!len)
1887                         break;
1888
1889                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1890                 page = spd.pages[page_nr];
1891
1892                 if (!PageUptodate(page) || page->mapping != mapping) {
1893                         error = shmem_getpage(inode, index, &page,
1894                                                         SGP_CACHE, NULL);
1895                         if (error)
1896                                 break;
1897                         unlock_page(page);
1898                         page_cache_release(spd.pages[page_nr]);
1899                         spd.pages[page_nr] = page;
1900                 }
1901
1902                 isize = i_size_read(inode);
1903                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1904                 if (unlikely(!isize || index > end_index))
1905                         break;
1906
1907                 if (end_index == index) {
1908                         unsigned int plen;
1909
1910                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1911                         if (plen <= loff)
1912                                 break;
1913
1914                         this_len = min(this_len, plen - loff);
1915                         len = this_len;
1916                 }
1917
1918                 spd.partial[page_nr].offset = loff;
1919                 spd.partial[page_nr].len = this_len;
1920                 len -= this_len;
1921                 loff = 0;
1922                 spd.nr_pages++;
1923                 index++;
1924         }
1925
1926         while (page_nr < nr_pages)
1927                 page_cache_release(spd.pages[page_nr++]);
1928
1929         if (spd.nr_pages)
1930                 error = splice_to_pipe(pipe, &spd);
1931
1932         splice_shrink_spd(pipe, &spd);
1933
1934         if (error > 0) {
1935                 *ppos += error;
1936                 file_accessed(in);
1937         }
1938         return error;
1939 }
1940
1941 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1942 {
1943         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1944
1945         buf->f_type = TMPFS_MAGIC;
1946         buf->f_bsize = PAGE_CACHE_SIZE;
1947         buf->f_namelen = NAME_MAX;
1948         if (sbinfo->max_blocks) {
1949                 buf->f_blocks = sbinfo->max_blocks;
1950                 buf->f_bavail = buf->f_bfree =
1951                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1952         }
1953         if (sbinfo->max_inodes) {
1954                 buf->f_files = sbinfo->max_inodes;
1955                 buf->f_ffree = sbinfo->free_inodes;
1956         }
1957         /* else leave those fields 0 like simple_statfs */
1958         return 0;
1959 }
1960
1961 /*
1962  * File creation. Allocate an inode, and we're done..
1963  */
1964 static int
1965 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1966 {
1967         struct inode *inode;
1968         int error = -ENOSPC;
1969
1970         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1971         if (inode) {
1972                 error = security_inode_init_security(inode, dir,
1973                                                      &dentry->d_name, NULL,
1974                                                      NULL, NULL);
1975                 if (error) {
1976                         if (error != -EOPNOTSUPP) {
1977                                 iput(inode);
1978                                 return error;
1979                         }
1980                 }
1981 #ifdef CONFIG_TMPFS_POSIX_ACL
1982                 error = generic_acl_init(inode, dir);
1983                 if (error) {
1984                         iput(inode);
1985                         return error;
1986                 }
1987 #else
1988                 error = 0;
1989 #endif
1990                 dir->i_size += BOGO_DIRENT_SIZE;
1991                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1992                 d_instantiate(dentry, inode);
1993                 dget(dentry); /* Extra count - pin the dentry in core */
1994         }
1995         return error;
1996 }
1997
1998 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1999 {
2000         int error;
2001
2002         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2003                 return error;
2004         inc_nlink(dir);
2005         return 0;
2006 }
2007
2008 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
2009                 struct nameidata *nd)
2010 {
2011         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2012 }
2013
2014 /*
2015  * Link a file..
2016  */
2017 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2018 {
2019         struct inode *inode = old_dentry->d_inode;
2020         int ret;
2021
2022         /*
2023          * No ordinary (disk based) filesystem counts links as inodes;
2024          * but each new link needs a new dentry, pinning lowmem, and
2025          * tmpfs dentries cannot be pruned until they are unlinked.
2026          */
2027         ret = shmem_reserve_inode(inode->i_sb);
2028         if (ret)
2029                 goto out;
2030
2031         dir->i_size += BOGO_DIRENT_SIZE;
2032         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2033         inc_nlink(inode);
2034         ihold(inode);   /* New dentry reference */
2035         dget(dentry);           /* Extra pinning count for the created dentry */
2036         d_instantiate(dentry, inode);
2037 out:
2038         return ret;
2039 }
2040
2041 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2042 {
2043         struct inode *inode = dentry->d_inode;
2044
2045         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2046                 shmem_free_inode(inode->i_sb);
2047
2048         dir->i_size -= BOGO_DIRENT_SIZE;
2049         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2050         drop_nlink(inode);
2051         dput(dentry);   /* Undo the count from "create" - this does all the work */
2052         return 0;
2053 }
2054
2055 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2056 {
2057         if (!simple_empty(dentry))
2058                 return -ENOTEMPTY;
2059
2060         drop_nlink(dentry->d_inode);
2061         drop_nlink(dir);
2062         return shmem_unlink(dir, dentry);
2063 }
2064
2065 /*
2066  * The VFS layer already does all the dentry stuff for rename,
2067  * we just have to decrement the usage count for the target if
2068  * it exists so that the VFS layer correctly free's it when it
2069  * gets overwritten.
2070  */
2071 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2072 {
2073         struct inode *inode = old_dentry->d_inode;
2074         int they_are_dirs = S_ISDIR(inode->i_mode);
2075
2076         if (!simple_empty(new_dentry))
2077                 return -ENOTEMPTY;
2078
2079         if (new_dentry->d_inode) {
2080                 (void) shmem_unlink(new_dir, new_dentry);
2081                 if (they_are_dirs)
2082                         drop_nlink(old_dir);
2083         } else if (they_are_dirs) {
2084                 drop_nlink(old_dir);
2085                 inc_nlink(new_dir);
2086         }
2087
2088         old_dir->i_size -= BOGO_DIRENT_SIZE;
2089         new_dir->i_size += BOGO_DIRENT_SIZE;
2090         old_dir->i_ctime = old_dir->i_mtime =
2091         new_dir->i_ctime = new_dir->i_mtime =
2092         inode->i_ctime = CURRENT_TIME;
2093         return 0;
2094 }
2095
2096 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2097 {
2098         int error;
2099         int len;
2100         struct inode *inode;
2101         struct page *page;
2102         char *kaddr;
2103         struct shmem_inode_info *info;
2104
2105         len = strlen(symname) + 1;
2106         if (len > PAGE_CACHE_SIZE)
2107                 return -ENAMETOOLONG;
2108
2109         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2110         if (!inode)
2111                 return -ENOSPC;
2112
2113         error = security_inode_init_security(inode, dir, &dentry->d_name, NULL,
2114                                              NULL, NULL);
2115         if (error) {
2116                 if (error != -EOPNOTSUPP) {
2117                         iput(inode);
2118                         return error;
2119                 }
2120                 error = 0;
2121         }
2122
2123         info = SHMEM_I(inode);
2124         inode->i_size = len-1;
2125         if (len <= SHMEM_SYMLINK_INLINE_LEN) {
2126                 /* do it inline */
2127                 memcpy(info->inline_symlink, symname, len);
2128                 inode->i_op = &shmem_symlink_inline_operations;
2129         } else {
2130                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2131                 if (error) {
2132                         iput(inode);
2133                         return error;
2134                 }
2135                 inode->i_mapping->a_ops = &shmem_aops;
2136                 inode->i_op = &shmem_symlink_inode_operations;
2137                 kaddr = kmap_atomic(page, KM_USER0);
2138                 memcpy(kaddr, symname, len);
2139                 kunmap_atomic(kaddr, KM_USER0);
2140                 set_page_dirty(page);
2141                 unlock_page(page);
2142                 page_cache_release(page);
2143         }
2144         dir->i_size += BOGO_DIRENT_SIZE;
2145         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2146         d_instantiate(dentry, inode);
2147         dget(dentry);
2148         return 0;
2149 }
2150
2151 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2152 {
2153         nd_set_link(nd, SHMEM_I(dentry->d_inode)->inline_symlink);
2154         return NULL;
2155 }
2156
2157 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2158 {
2159         struct page *page = NULL;
2160         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2161         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2162         if (page)
2163                 unlock_page(page);
2164         return page;
2165 }
2166
2167 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2168 {
2169         if (!IS_ERR(nd_get_link(nd))) {
2170                 struct page *page = cookie;
2171                 kunmap(page);
2172                 mark_page_accessed(page);
2173                 page_cache_release(page);
2174         }
2175 }
2176
2177 #ifdef CONFIG_TMPFS_XATTR
2178 /*
2179  * Superblocks without xattr inode operations may get some security.* xattr
2180  * support from the LSM "for free". As soon as we have any other xattrs
2181  * like ACLs, we also need to implement the security.* handlers at
2182  * filesystem level, though.
2183  */
2184
2185 static int shmem_xattr_get(struct dentry *dentry, const char *name,
2186                            void *buffer, size_t size)
2187 {
2188         struct shmem_inode_info *info;
2189         struct shmem_xattr *xattr;
2190         int ret = -ENODATA;
2191
2192         info = SHMEM_I(dentry->d_inode);
2193
2194         spin_lock(&info->lock);
2195         list_for_each_entry(xattr, &info->xattr_list, list) {
2196                 if (strcmp(name, xattr->name))
2197                         continue;
2198
2199                 ret = xattr->size;
2200                 if (buffer) {
2201                         if (size < xattr->size)
2202                                 ret = -ERANGE;
2203                         else
2204                                 memcpy(buffer, xattr->value, xattr->size);
2205                 }
2206                 break;
2207         }
2208         spin_unlock(&info->lock);
2209         return ret;
2210 }
2211
2212 static int shmem_xattr_set(struct dentry *dentry, const char *name,
2213                            const void *value, size_t size, int flags)
2214 {
2215         struct inode *inode = dentry->d_inode;
2216         struct shmem_inode_info *info = SHMEM_I(inode);
2217         struct shmem_xattr *xattr;
2218         struct shmem_xattr *new_xattr = NULL;
2219         size_t len;
2220         int err = 0;
2221
2222         /* value == NULL means remove */
2223         if (value) {
2224                 /* wrap around? */
2225                 len = sizeof(*new_xattr) + size;
2226                 if (len <= sizeof(*new_xattr))
2227                         return -ENOMEM;
2228
2229                 new_xattr = kmalloc(len, GFP_KERNEL);
2230                 if (!new_xattr)
2231                         return -ENOMEM;
2232
2233                 new_xattr->name = kstrdup(name, GFP_KERNEL);
2234                 if (!new_xattr->name) {
2235                         kfree(new_xattr);
2236                         return -ENOMEM;
2237                 }
2238
2239                 new_xattr->size = size;
2240                 memcpy(new_xattr->value, value, size);
2241         }
2242
2243         spin_lock(&info->lock);
2244         list_for_each_entry(xattr, &info->xattr_list, list) {
2245                 if (!strcmp(name, xattr->name)) {
2246                         if (flags & XATTR_CREATE) {
2247                                 xattr = new_xattr;
2248                                 err = -EEXIST;
2249                         } else if (new_xattr) {
2250                                 list_replace(&xattr->list, &new_xattr->list);
2251                         } else {
2252                                 list_del(&xattr->list);
2253                         }
2254                         goto out;
2255                 }
2256         }
2257         if (flags & XATTR_REPLACE) {
2258                 xattr = new_xattr;
2259                 err = -ENODATA;
2260         } else {
2261                 list_add(&new_xattr->list, &info->xattr_list);
2262                 xattr = NULL;
2263         }
2264 out:
2265         spin_unlock(&info->lock);
2266         if (xattr)
2267                 kfree(xattr->name);
2268         kfree(xattr);
2269         return err;
2270 }
2271
2272
2273 static const struct xattr_handler *shmem_xattr_handlers[] = {
2274 #ifdef CONFIG_TMPFS_POSIX_ACL
2275         &generic_acl_access_handler,
2276         &generic_acl_default_handler,
2277 #endif
2278         NULL
2279 };
2280
2281 static int shmem_xattr_validate(const char *name)
2282 {
2283         struct { const char *prefix; size_t len; } arr[] = {
2284                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2285                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2286         };
2287         int i;
2288
2289         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2290                 size_t preflen = arr[i].len;
2291                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2292                         if (!name[preflen])
2293                                 return -EINVAL;
2294                         return 0;
2295                 }
2296         }
2297         return -EOPNOTSUPP;
2298 }
2299
2300 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2301                               void *buffer, size_t size)
2302 {
2303         int err;
2304
2305         /*
2306          * If this is a request for a synthetic attribute in the system.*
2307          * namespace use the generic infrastructure to resolve a handler
2308          * for it via sb->s_xattr.
2309          */
2310         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2311                 return generic_getxattr(dentry, name, buffer, size);
2312
2313         err = shmem_xattr_validate(name);
2314         if (err)
2315                 return err;
2316
2317         return shmem_xattr_get(dentry, name, buffer, size);
2318 }
2319
2320 static int shmem_setxattr(struct dentry *dentry, const char *name,
2321                           const void *value, size_t size, int flags)
2322 {
2323         int err;
2324
2325         /*
2326          * If this is a request for a synthetic attribute in the system.*
2327          * namespace use the generic infrastructure to resolve a handler
2328          * for it via sb->s_xattr.
2329          */
2330         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2331                 return generic_setxattr(dentry, name, value, size, flags);
2332
2333         err = shmem_xattr_validate(name);
2334         if (err)
2335                 return err;
2336
2337         if (size == 0)
2338                 value = "";  /* empty EA, do not remove */
2339
2340         return shmem_xattr_set(dentry, name, value, size, flags);
2341
2342 }
2343
2344 static int shmem_removexattr(struct dentry *dentry, const char *name)
2345 {
2346         int err;
2347
2348         /*
2349          * If this is a request for a synthetic attribute in the system.*
2350          * namespace use the generic infrastructure to resolve a handler
2351          * for it via sb->s_xattr.
2352          */
2353         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2354                 return generic_removexattr(dentry, name);
2355
2356         err = shmem_xattr_validate(name);
2357         if (err)
2358                 return err;
2359
2360         return shmem_xattr_set(dentry, name, NULL, 0, XATTR_REPLACE);
2361 }
2362
2363 static bool xattr_is_trusted(const char *name)
2364 {
2365         return !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN);
2366 }
2367
2368 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2369 {
2370         bool trusted = capable(CAP_SYS_ADMIN);
2371         struct shmem_xattr *xattr;
2372         struct shmem_inode_info *info;
2373         size_t used = 0;
2374
2375         info = SHMEM_I(dentry->d_inode);
2376
2377         spin_lock(&info->lock);
2378         list_for_each_entry(xattr, &info->xattr_list, list) {
2379                 size_t len;
2380
2381                 /* skip "trusted." attributes for unprivileged callers */
2382                 if (!trusted && xattr_is_trusted(xattr->name))
2383                         continue;
2384
2385                 len = strlen(xattr->name) + 1;
2386                 used += len;
2387                 if (buffer) {
2388                         if (size < used) {
2389                                 used = -ERANGE;
2390                                 break;
2391                         }
2392                         memcpy(buffer, xattr->name, len);
2393                         buffer += len;
2394                 }
2395         }
2396         spin_unlock(&info->lock);
2397
2398         return used;
2399 }
2400 #endif /* CONFIG_TMPFS_XATTR */
2401
2402 static const struct inode_operations shmem_symlink_inline_operations = {
2403         .readlink       = generic_readlink,
2404         .follow_link    = shmem_follow_link_inline,
2405 #ifdef CONFIG_TMPFS_XATTR
2406         .setxattr       = shmem_setxattr,
2407         .getxattr       = shmem_getxattr,
2408         .listxattr      = shmem_listxattr,
2409         .removexattr    = shmem_removexattr,
2410 #endif
2411 };
2412
2413 static const struct inode_operations shmem_symlink_inode_operations = {
2414         .readlink       = generic_readlink,
2415         .follow_link    = shmem_follow_link,
2416         .put_link       = shmem_put_link,
2417 #ifdef CONFIG_TMPFS_XATTR
2418         .setxattr       = shmem_setxattr,
2419         .getxattr       = shmem_getxattr,
2420         .listxattr      = shmem_listxattr,
2421         .removexattr    = shmem_removexattr,
2422 #endif
2423 };
2424
2425 static struct dentry *shmem_get_parent(struct dentry *child)
2426 {
2427         return ERR_PTR(-ESTALE);
2428 }
2429
2430 static int shmem_match(struct inode *ino, void *vfh)
2431 {
2432         __u32 *fh = vfh;
2433         __u64 inum = fh[2];
2434         inum = (inum << 32) | fh[1];
2435         return ino->i_ino == inum && fh[0] == ino->i_generation;
2436 }
2437
2438 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2439                 struct fid *fid, int fh_len, int fh_type)
2440 {
2441         struct inode *inode;
2442         struct dentry *dentry = NULL;
2443         u64 inum = fid->raw[2];
2444         inum = (inum << 32) | fid->raw[1];
2445
2446         if (fh_len < 3)
2447                 return NULL;
2448
2449         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2450                         shmem_match, fid->raw);
2451         if (inode) {
2452                 dentry = d_find_alias(inode);
2453                 iput(inode);
2454         }
2455
2456         return dentry;
2457 }
2458
2459 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2460                                 int connectable)
2461 {
2462         struct inode *inode = dentry->d_inode;
2463
2464         if (*len < 3) {
2465                 *len = 3;
2466                 return 255;
2467         }
2468
2469         if (inode_unhashed(inode)) {
2470                 /* Unfortunately insert_inode_hash is not idempotent,
2471                  * so as we hash inodes here rather than at creation
2472                  * time, we need a lock to ensure we only try
2473                  * to do it once
2474                  */
2475                 static DEFINE_SPINLOCK(lock);
2476                 spin_lock(&lock);
2477                 if (inode_unhashed(inode))
2478                         __insert_inode_hash(inode,
2479                                             inode->i_ino + inode->i_generation);
2480                 spin_unlock(&lock);
2481         }
2482
2483         fh[0] = inode->i_generation;
2484         fh[1] = inode->i_ino;
2485         fh[2] = ((__u64)inode->i_ino) >> 32;
2486
2487         *len = 3;
2488         return 1;
2489 }
2490
2491 static const struct export_operations shmem_export_ops = {
2492         .get_parent     = shmem_get_parent,
2493         .encode_fh      = shmem_encode_fh,
2494         .fh_to_dentry   = shmem_fh_to_dentry,
2495 };
2496
2497 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2498                                bool remount)
2499 {
2500         char *this_char, *value, *rest;
2501
2502         while (options != NULL) {
2503                 this_char = options;
2504                 for (;;) {
2505                         /*
2506                          * NUL-terminate this option: unfortunately,
2507                          * mount options form a comma-separated list,
2508                          * but mpol's nodelist may also contain commas.
2509                          */
2510                         options = strchr(options, ',');
2511                         if (options == NULL)
2512                                 break;
2513                         options++;
2514                         if (!isdigit(*options)) {
2515                                 options[-1] = '\0';
2516                                 break;
2517                         }
2518                 }
2519                 if (!*this_char)
2520                         continue;
2521                 if ((value = strchr(this_char,'=')) != NULL) {
2522                         *value++ = 0;
2523                 } else {
2524                         printk(KERN_ERR
2525                             "tmpfs: No value for mount option '%s'\n",
2526                             this_char);
2527                         return 1;
2528                 }
2529
2530                 if (!strcmp(this_char,"size")) {
2531                         unsigned long long size;
2532                         size = memparse(value,&rest);
2533                         if (*rest == '%') {
2534                                 size <<= PAGE_SHIFT;
2535                                 size *= totalram_pages;
2536                                 do_div(size, 100);
2537                                 rest++;
2538                         }
2539                         if (*rest)
2540                                 goto bad_val;
2541                         sbinfo->max_blocks =
2542                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2543                 } else if (!strcmp(this_char,"nr_blocks")) {
2544                         sbinfo->max_blocks = memparse(value, &rest);
2545                         if (*rest)
2546                                 goto bad_val;
2547                 } else if (!strcmp(this_char,"nr_inodes")) {
2548                         sbinfo->max_inodes = memparse(value, &rest);
2549                         if (*rest)
2550                                 goto bad_val;
2551                 } else if (!strcmp(this_char,"mode")) {
2552                         if (remount)
2553                                 continue;
2554                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2555                         if (*rest)
2556                                 goto bad_val;
2557                 } else if (!strcmp(this_char,"uid")) {
2558                         if (remount)
2559                                 continue;
2560                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2561                         if (*rest)
2562                                 goto bad_val;
2563                 } else if (!strcmp(this_char,"gid")) {
2564                         if (remount)
2565                                 continue;
2566                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2567                         if (*rest)
2568                                 goto bad_val;
2569                 } else if (!strcmp(this_char,"mpol")) {
2570                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2571                                 goto bad_val;
2572                 } else {
2573                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2574                                this_char);
2575                         return 1;
2576                 }
2577         }
2578         return 0;
2579
2580 bad_val:
2581         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2582                value, this_char);
2583         return 1;
2584
2585 }
2586
2587 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2588 {
2589         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2590         struct shmem_sb_info config = *sbinfo;
2591         unsigned long inodes;
2592         int error = -EINVAL;
2593
2594         if (shmem_parse_options(data, &config, true))
2595                 return error;
2596
2597         spin_lock(&sbinfo->stat_lock);
2598         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2599         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2600                 goto out;
2601         if (config.max_inodes < inodes)
2602                 goto out;
2603         /*
2604          * Those tests also disallow limited->unlimited while any are in
2605          * use, so i_blocks will always be zero when max_blocks is zero;
2606          * but we must separately disallow unlimited->limited, because
2607          * in that case we have no record of how much is already in use.
2608          */
2609         if (config.max_blocks && !sbinfo->max_blocks)
2610                 goto out;
2611         if (config.max_inodes && !sbinfo->max_inodes)
2612                 goto out;
2613
2614         error = 0;
2615         sbinfo->max_blocks  = config.max_blocks;
2616         sbinfo->max_inodes  = config.max_inodes;
2617         sbinfo->free_inodes = config.max_inodes - inodes;
2618
2619         mpol_put(sbinfo->mpol);
2620         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2621 out:
2622         spin_unlock(&sbinfo->stat_lock);
2623         return error;
2624 }
2625
2626 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2627 {
2628         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2629
2630         if (sbinfo->max_blocks != shmem_default_max_blocks())
2631                 seq_printf(seq, ",size=%luk",
2632                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2633         if (sbinfo->max_inodes != shmem_default_max_inodes())
2634                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2635         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2636                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2637         if (sbinfo->uid != 0)
2638                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2639         if (sbinfo->gid != 0)
2640                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2641         shmem_show_mpol(seq, sbinfo->mpol);
2642         return 0;
2643 }
2644 #endif /* CONFIG_TMPFS */
2645
2646 static void shmem_put_super(struct super_block *sb)
2647 {
2648         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2649
2650         percpu_counter_destroy(&sbinfo->used_blocks);
2651         kfree(sbinfo);
2652         sb->s_fs_info = NULL;
2653 }
2654
2655 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2656 {
2657         struct inode *inode;
2658         struct dentry *root;
2659         struct shmem_sb_info *sbinfo;
2660         int err = -ENOMEM;
2661
2662         /* Round up to L1_CACHE_BYTES to resist false sharing */
2663         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2664                                 L1_CACHE_BYTES), GFP_KERNEL);
2665         if (!sbinfo)
2666                 return -ENOMEM;
2667
2668         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2669         sbinfo->uid = current_fsuid();
2670         sbinfo->gid = current_fsgid();
2671         sb->s_fs_info = sbinfo;
2672
2673 #ifdef CONFIG_TMPFS
2674         /*
2675          * Per default we only allow half of the physical ram per
2676          * tmpfs instance, limiting inodes to one per page of lowmem;
2677          * but the internal instance is left unlimited.
2678          */
2679         if (!(sb->s_flags & MS_NOUSER)) {
2680                 sbinfo->max_blocks = shmem_default_max_blocks();
2681                 sbinfo->max_inodes = shmem_default_max_inodes();
2682                 if (shmem_parse_options(data, sbinfo, false)) {
2683                         err = -EINVAL;
2684                         goto failed;
2685                 }
2686         }
2687         sb->s_export_op = &shmem_export_ops;
2688 #else
2689         sb->s_flags |= MS_NOUSER;
2690 #endif
2691
2692         spin_lock_init(&sbinfo->stat_lock);
2693         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2694                 goto failed;
2695         sbinfo->free_inodes = sbinfo->max_inodes;
2696
2697         sb->s_maxbytes = SHMEM_MAX_BYTES;
2698         sb->s_blocksize = PAGE_CACHE_SIZE;
2699         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2700         sb->s_magic = TMPFS_MAGIC;
2701         sb->s_op = &shmem_ops;
2702         sb->s_time_gran = 1;
2703 #ifdef CONFIG_TMPFS_XATTR
2704         sb->s_xattr = shmem_xattr_handlers;
2705 #endif
2706 #ifdef CONFIG_TMPFS_POSIX_ACL
2707         sb->s_flags |= MS_POSIXACL;
2708 #endif
2709
2710         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2711         if (!inode)
2712                 goto failed;
2713         inode->i_uid = sbinfo->uid;
2714         inode->i_gid = sbinfo->gid;
2715         root = d_alloc_root(inode);
2716         if (!root)
2717                 goto failed_iput;
2718         sb->s_root = root;
2719         return 0;
2720
2721 failed_iput:
2722         iput(inode);
2723 failed:
2724         shmem_put_super(sb);
2725         return err;
2726 }
2727
2728 static struct kmem_cache *shmem_inode_cachep;
2729
2730 static struct inode *shmem_alloc_inode(struct super_block *sb)
2731 {
2732         struct shmem_inode_info *p;
2733         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2734         if (!p)
2735                 return NULL;
2736         return &p->vfs_inode;
2737 }
2738
2739 static void shmem_i_callback(struct rcu_head *head)
2740 {
2741         struct inode *inode = container_of(head, struct inode, i_rcu);
2742         INIT_LIST_HEAD(&inode->i_dentry);
2743         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2744 }
2745
2746 static void shmem_destroy_inode(struct inode *inode)
2747 {
2748         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2749                 /* only struct inode is valid if it's an inline symlink */
2750                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2751         }
2752         call_rcu(&inode->i_rcu, shmem_i_callback);
2753 }
2754
2755 static void init_once(void *foo)
2756 {
2757         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2758
2759         inode_init_once(&p->vfs_inode);
2760 }
2761
2762 static int init_inodecache(void)
2763 {
2764         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2765                                 sizeof(struct shmem_inode_info),
2766                                 0, SLAB_PANIC, init_once);
2767         return 0;
2768 }
2769
2770 static void destroy_inodecache(void)
2771 {
2772         kmem_cache_destroy(shmem_inode_cachep);
2773 }
2774
2775 static const struct address_space_operations shmem_aops = {
2776         .writepage      = shmem_writepage,
2777         .set_page_dirty = __set_page_dirty_no_writeback,
2778 #ifdef CONFIG_TMPFS
2779         .write_begin    = shmem_write_begin,
2780         .write_end      = shmem_write_end,
2781 #endif
2782         .migratepage    = migrate_page,
2783         .error_remove_page = generic_error_remove_page,
2784 };
2785
2786 static const struct file_operations shmem_file_operations = {
2787         .mmap           = shmem_mmap,
2788 #ifdef CONFIG_TMPFS
2789         .llseek         = generic_file_llseek,
2790         .read           = do_sync_read,
2791         .write          = do_sync_write,
2792         .aio_read       = shmem_file_aio_read,
2793         .aio_write      = generic_file_aio_write,
2794         .fsync          = noop_fsync,
2795         .splice_read    = shmem_file_splice_read,
2796         .splice_write   = generic_file_splice_write,
2797 #endif
2798 };
2799
2800 static const struct inode_operations shmem_inode_operations = {
2801         .setattr        = shmem_setattr,
2802         .truncate_range = shmem_truncate_range,
2803 #ifdef CONFIG_TMPFS_XATTR
2804         .setxattr       = shmem_setxattr,
2805         .getxattr       = shmem_getxattr,
2806         .listxattr      = shmem_listxattr,
2807         .removexattr    = shmem_removexattr,
2808 #endif
2809 #ifdef CONFIG_TMPFS_POSIX_ACL
2810         .check_acl      = generic_check_acl,
2811 #endif
2812
2813 };
2814
2815 static const struct inode_operations shmem_dir_inode_operations = {
2816 #ifdef CONFIG_TMPFS
2817         .create         = shmem_create,
2818         .lookup         = simple_lookup,
2819         .link           = shmem_link,
2820         .unlink         = shmem_unlink,
2821         .symlink        = shmem_symlink,
2822         .mkdir          = shmem_mkdir,
2823         .rmdir          = shmem_rmdir,
2824         .mknod          = shmem_mknod,
2825         .rename         = shmem_rename,
2826 #endif
2827 #ifdef CONFIG_TMPFS_XATTR
2828         .setxattr       = shmem_setxattr,
2829         .getxattr       = shmem_getxattr,
2830         .listxattr      = shmem_listxattr,
2831         .removexattr    = shmem_removexattr,
2832 #endif
2833 #ifdef CONFIG_TMPFS_POSIX_ACL
2834         .setattr        = shmem_setattr,
2835         .check_acl      = generic_check_acl,
2836 #endif
2837 };
2838
2839 static const struct inode_operations shmem_special_inode_operations = {
2840 #ifdef CONFIG_TMPFS_XATTR
2841         .setxattr       = shmem_setxattr,
2842         .getxattr       = shmem_getxattr,
2843         .listxattr      = shmem_listxattr,
2844         .removexattr    = shmem_removexattr,
2845 #endif
2846 #ifdef CONFIG_TMPFS_POSIX_ACL
2847         .setattr        = shmem_setattr,
2848         .check_acl      = generic_check_acl,
2849 #endif
2850 };
2851
2852 static const struct super_operations shmem_ops = {
2853         .alloc_inode    = shmem_alloc_inode,
2854         .destroy_inode  = shmem_destroy_inode,
2855 #ifdef CONFIG_TMPFS
2856         .statfs         = shmem_statfs,
2857         .remount_fs     = shmem_remount_fs,
2858         .show_options   = shmem_show_options,
2859 #endif
2860         .evict_inode    = shmem_evict_inode,
2861         .drop_inode     = generic_delete_inode,
2862         .put_super      = shmem_put_super,
2863 };
2864
2865 static const struct vm_operations_struct shmem_vm_ops = {
2866         .fault          = shmem_fault,
2867 #ifdef CONFIG_NUMA
2868         .set_policy     = shmem_set_policy,
2869         .get_policy     = shmem_get_policy,
2870 #endif
2871 };
2872
2873
2874 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2875         int flags, const char *dev_name, void *data)
2876 {
2877         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2878 }
2879
2880 static struct file_system_type tmpfs_fs_type = {
2881         .owner          = THIS_MODULE,
2882         .name           = "tmpfs",
2883         .mount          = shmem_mount,
2884         .kill_sb        = kill_litter_super,
2885 };
2886
2887 int __init init_tmpfs(void)
2888 {
2889         int error;
2890
2891         error = bdi_init(&shmem_backing_dev_info);
2892         if (error)
2893                 goto out4;
2894
2895         error = init_inodecache();
2896         if (error)
2897                 goto out3;
2898
2899         error = register_filesystem(&tmpfs_fs_type);
2900         if (error) {
2901                 printk(KERN_ERR "Could not register tmpfs\n");
2902                 goto out2;
2903         }
2904
2905         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2906                                 tmpfs_fs_type.name, NULL);
2907         if (IS_ERR(shm_mnt)) {
2908                 error = PTR_ERR(shm_mnt);
2909                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2910                 goto out1;
2911         }
2912         return 0;
2913
2914 out1:
2915         unregister_filesystem(&tmpfs_fs_type);
2916 out2:
2917         destroy_inodecache();
2918 out3:
2919         bdi_destroy(&shmem_backing_dev_info);
2920 out4:
2921         shm_mnt = ERR_PTR(error);
2922         return error;
2923 }
2924
2925 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2926 /**
2927  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2928  * @inode: the inode to be searched
2929  * @pgoff: the offset to be searched
2930  * @pagep: the pointer for the found page to be stored
2931  * @ent: the pointer for the found swap entry to be stored
2932  *
2933  * If a page is found, refcount of it is incremented. Callers should handle
2934  * these refcount.
2935  */
2936 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2937                                         struct page **pagep, swp_entry_t *ent)
2938 {
2939         swp_entry_t entry = { .val = 0 }, *ptr;
2940         struct page *page = NULL;
2941         struct shmem_inode_info *info = SHMEM_I(inode);
2942
2943         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2944                 goto out;
2945
2946         spin_lock(&info->lock);
2947         ptr = shmem_swp_entry(info, pgoff, NULL);
2948 #ifdef CONFIG_SWAP
2949         if (ptr && ptr->val) {
2950                 entry.val = ptr->val;
2951                 page = find_get_page(&swapper_space, entry.val);
2952         } else
2953 #endif
2954                 page = find_get_page(inode->i_mapping, pgoff);
2955         if (ptr)
2956                 shmem_swp_unmap(ptr);
2957         spin_unlock(&info->lock);
2958 out:
2959         *pagep = page;
2960         *ent = entry;
2961 }
2962 #endif
2963
2964 #else /* !CONFIG_SHMEM */
2965
2966 /*
2967  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2968  *
2969  * This is intended for small system where the benefits of the full
2970  * shmem code (swap-backed and resource-limited) are outweighed by
2971  * their complexity. On systems without swap this code should be
2972  * effectively equivalent, but much lighter weight.
2973  */
2974
2975 #include <linux/ramfs.h>
2976
2977 static struct file_system_type tmpfs_fs_type = {
2978         .name           = "tmpfs",
2979         .mount          = ramfs_mount,
2980         .kill_sb        = kill_litter_super,
2981 };
2982
2983 int __init init_tmpfs(void)
2984 {
2985         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2986
2987         shm_mnt = kern_mount(&tmpfs_fs_type);
2988         BUG_ON(IS_ERR(shm_mnt));
2989
2990         return 0;
2991 }
2992
2993 int shmem_unuse(swp_entry_t entry, struct page *page)
2994 {
2995         return 0;
2996 }
2997
2998 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2999 {
3000         return 0;
3001 }
3002
3003 void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
3004 {
3005         truncate_inode_pages_range(inode->i_mapping, start, end);
3006 }
3007 EXPORT_SYMBOL_GPL(shmem_truncate_range);
3008
3009 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
3010 /**
3011  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
3012  * @inode: the inode to be searched
3013  * @pgoff: the offset to be searched
3014  * @pagep: the pointer for the found page to be stored
3015  * @ent: the pointer for the found swap entry to be stored
3016  *
3017  * If a page is found, refcount of it is incremented. Callers should handle
3018  * these refcount.
3019  */
3020 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
3021                                         struct page **pagep, swp_entry_t *ent)
3022 {
3023         struct page *page = NULL;
3024
3025         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
3026                 goto out;
3027         page = find_get_page(inode->i_mapping, pgoff);
3028 out:
3029         *pagep = page;
3030         *ent = (swp_entry_t){ .val = 0 };
3031 }
3032 #endif
3033
3034 #define shmem_vm_ops                            generic_file_vm_ops
3035 #define shmem_file_operations                   ramfs_file_operations
3036 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
3037 #define shmem_acct_size(flags, size)            0
3038 #define shmem_unacct_size(flags, size)          do {} while (0)
3039 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
3040
3041 #endif /* CONFIG_SHMEM */
3042
3043 /* common code */
3044
3045 /**
3046  * shmem_file_setup - get an unlinked file living in tmpfs
3047  * @name: name for dentry (to be seen in /proc/<pid>/maps
3048  * @size: size to be set for the file
3049  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3050  */
3051 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
3052 {
3053         int error;
3054         struct file *file;
3055         struct inode *inode;
3056         struct path path;
3057         struct dentry *root;
3058         struct qstr this;
3059
3060         if (IS_ERR(shm_mnt))
3061                 return (void *)shm_mnt;
3062
3063         if (size < 0 || size > SHMEM_MAX_BYTES)
3064                 return ERR_PTR(-EINVAL);
3065
3066         if (shmem_acct_size(flags, size))
3067                 return ERR_PTR(-ENOMEM);
3068
3069         error = -ENOMEM;
3070         this.name = name;
3071         this.len = strlen(name);
3072         this.hash = 0; /* will go */
3073         root = shm_mnt->mnt_root;
3074         path.dentry = d_alloc(root, &this);
3075         if (!path.dentry)
3076                 goto put_memory;
3077         path.mnt = mntget(shm_mnt);
3078
3079         error = -ENOSPC;
3080         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
3081         if (!inode)
3082                 goto put_dentry;
3083
3084         d_instantiate(path.dentry, inode);
3085         inode->i_size = size;
3086         inode->i_nlink = 0;     /* It is unlinked */
3087 #ifndef CONFIG_MMU
3088         error = ramfs_nommu_expand_for_mapping(inode, size);
3089         if (error)
3090                 goto put_dentry;
3091 #endif
3092
3093         error = -ENFILE;
3094         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
3095                   &shmem_file_operations);
3096         if (!file)
3097                 goto put_dentry;
3098
3099         return file;
3100
3101 put_dentry:
3102         path_put(&path);
3103 put_memory:
3104         shmem_unacct_size(flags, size);
3105         return ERR_PTR(error);
3106 }
3107 EXPORT_SYMBOL_GPL(shmem_file_setup);
3108
3109 /**
3110  * shmem_zero_setup - setup a shared anonymous mapping
3111  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3112  */
3113 int shmem_zero_setup(struct vm_area_struct *vma)
3114 {
3115         struct file *file;
3116         loff_t size = vma->vm_end - vma->vm_start;
3117
3118         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
3119         if (IS_ERR(file))
3120                 return PTR_ERR(file);
3121
3122         if (vma->vm_file)
3123                 fput(vma->vm_file);
3124         vma->vm_file = file;
3125         vma->vm_ops = &shmem_vm_ops;
3126         vma->vm_flags |= VM_CAN_NONLINEAR;
3127         return 0;
3128 }
3129
3130 /**
3131  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3132  * @mapping:    the page's address_space
3133  * @index:      the page index
3134  * @gfp:        the page allocator flags to use if allocating
3135  *
3136  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3137  * with any new page allocations done using the specified allocation flags.
3138  * But read_cache_page_gfp() uses the ->readpage() method: which does not
3139  * suit tmpfs, since it may have pages in swapcache, and needs to find those
3140  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3141  *
3142  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
3143  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
3144  */
3145 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3146                                          pgoff_t index, gfp_t gfp)
3147 {
3148 #ifdef CONFIG_SHMEM
3149         struct inode *inode = mapping->host;
3150         struct page *page;
3151         int error;
3152
3153         BUG_ON(mapping->a_ops != &shmem_aops);
3154         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3155         if (error)
3156                 page = ERR_PTR(error);
3157         else
3158                 unlock_page(page);
3159         return page;
3160 #else
3161         /*
3162          * The tiny !SHMEM case uses ramfs without swap
3163          */
3164         return read_cache_page_gfp(mapping, index, gfp);
3165 #endif
3166 }
3167 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);