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  * This file is released under the GPL.
  18  */
  19
  20 /*
  21  * This virtual memory filesystem is heavily based on the ramfs. It
  22  * extends ramfs by the ability to use swap and honor resource limits
  23  * which makes it a completely usable filesystem.
  24  */
  25
  26 #include <linux/config.h>
  27 #include <linux/module.h>
  28 #include <linux/init.h>
  29 #include <linux/devfs_fs_kernel.h>
  30 #include <linux/fs.h>
  31 #include <linux/mm.h>
  32 #include <linux/mman.h>
  33 #include <linux/file.h>
  34 #include <linux/swap.h>
  35 #include <linux/pagemap.h>
  36 #include <linux/string.h>
  37 #include <linux/slab.h>
  38 #include <linux/backing-dev.h>
  39 #include <linux/shmem_fs.h>
  40 #include <linux/mount.h>
  41 #include <linux/writeback.h>
  42 #include <linux/vfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/security.h>
  45 #include <linux/swapops.h>
  46 #include <linux/mempolicy.h>
  47 #include <linux/namei.h>
  48 #include <linux/ctype.h>
  49 #include <asm/uaccess.h>
  50 #include <asm/div64.h>
  51 #include <asm/pgtable.h>
  52
  53 /* This magic number is used in glibc for posix shared memory */
  54 #define TMPFS_MAGIC     0x01021994
  55
  56 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
  57 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
  58 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  59
  60 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
  61 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
  62
  63 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  64
  65 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
  66 #define SHMEM_PAGEIN     VM_READ
  67 #define SHMEM_TRUNCATE   VM_WRITE
  68
  69 /* Definition to limit shmem_truncate's steps between cond_rescheds */
  70 #define LATENCY_LIMIT    64
  71
  72 /* Pretend that each entry is of this size in directory's i_size */
  73 #define BOGO_DIRENT_SIZE 20
  74
  75 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
  76 enum sgp_type {
  77         SGP_QUICK,      /* don't try more than file page cache lookup */
  78         SGP_READ,       /* don't exceed i_size, don't allocate page */
  79         SGP_CACHE,      /* don't exceed i_size, may allocate page */
  80         SGP_WRITE,      /* may exceed i_size, may allocate page */
  81 };
  82
  83 static int shmem_getpage(struct inode *inode, unsigned long idx,
  84                          struct page **pagep, enum sgp_type sgp, int *type);
  85
  86 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
  87 {
  88         /*
  89          * The above definition of ENTRIES_PER_PAGE, and the use of
  90          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
  91          * might be reconsidered if it ever diverges from PAGE_SIZE.
  92          */
  93         return alloc_pages(gfp_mask, PAGE_CACHE_SHIFT-PAGE_SHIFT);
  94 }
  95
  96 static inline void shmem_dir_free(struct page *page)
  97 {
  98         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
  99 }
 100
 101 static struct page **shmem_dir_map(struct page *page)
 102 {
 103         return (struct page **)kmap_atomic(page, KM_USER0);
 104 }
 105
 106 static inline void shmem_dir_unmap(struct page **dir)
 107 {
 108         kunmap_atomic(dir, KM_USER0);
 109 }
 110
 111 static swp_entry_t *shmem_swp_map(struct page *page)
 112 {
 113         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
 114 }
 115
 116 static inline void shmem_swp_balance_unmap(void)
 117 {
 118         /*
 119          * When passing a pointer to an i_direct entry, to code which
 120          * also handles indirect entries and so will shmem_swp_unmap,
 121          * we must arrange for the preempt count to remain in balance.
 122          * What kmap_atomic of a lowmem page does depends on config
 123          * and architecture, so pretend to kmap_atomic some lowmem page.
 124          */
 125         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
 126 }
 127
 128 static inline void shmem_swp_unmap(swp_entry_t *entry)
 129 {
 130         kunmap_atomic(entry, KM_USER1);
 131 }
 132
 133 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 134 {
 135         return sb->s_fs_info;
 136 }
 137
 138 /*
 139  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 140  * for shared memory and for shared anonymous (/dev/zero) mappings
 141  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 142  * consistent with the pre-accounting of private mappings ...
 143  */
 144 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 145 {
 146         return (flags & VM_ACCOUNT)?
 147                 security_vm_enough_memory(VM_ACCT(size)): 0;
 148 }
 149
 150 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 151 {
 152         if (flags & VM_ACCOUNT)
 153                 vm_unacct_memory(VM_ACCT(size));
 154 }
 155
 156 /*
 157  * ... whereas tmpfs objects are accounted incrementally as
 158  * pages are allocated, in order to allow huge sparse files.
 159  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 160  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 161  */
 162 static inline int shmem_acct_block(unsigned long flags)
 163 {
 164         return (flags & VM_ACCOUNT)?
 165                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
 166 }
 167
 168 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 169 {
 170         if (!(flags & VM_ACCOUNT))
 171                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 172 }
 173
 174 static struct super_operations shmem_ops;
 175 static struct address_space_operations shmem_aops;
 176 static struct file_operations shmem_file_operations;
 177 static struct inode_operations shmem_inode_operations;
 178 static struct inode_operations shmem_dir_inode_operations;
 179 static struct vm_operations_struct shmem_vm_ops;
 180
 181 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 182         .ra_pages       = 0,    /* No readahead */
 183         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
 184         .unplug_io_fn   = default_unplug_io_fn,
 185 };
 186
 187 static LIST_HEAD(shmem_swaplist);
 188 static DEFINE_SPINLOCK(shmem_swaplist_lock);
 189
 190 static void shmem_free_blocks(struct inode *inode, long pages)
 191 {
 192         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 193         if (sbinfo->max_blocks) {
 194                 spin_lock(&sbinfo->stat_lock);
 195                 sbinfo->free_blocks += pages;
 196                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
 197                 spin_unlock(&sbinfo->stat_lock);
 198         }
 199 }
 200
 201 /*
 202  * shmem_recalc_inode - recalculate the size of an inode
 203  *
 204  * @inode: inode to recalc
 205  *
 206  * We have to calculate the free blocks since the mm can drop
 207  * undirtied hole pages behind our back.
 208  *
 209  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 210  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 211  *
 212  * It has to be called with the spinlock held.
 213  */
 214 static void shmem_recalc_inode(struct inode *inode)
 215 {
 216         struct shmem_inode_info *info = SHMEM_I(inode);
 217         long freed;
 218
 219         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 220         if (freed > 0) {
 221                 info->alloced -= freed;
 222                 shmem_unacct_blocks(info->flags, freed);
 223                 shmem_free_blocks(inode, freed);
 224         }
 225 }
 226
 227 /*
 228  * shmem_swp_entry - find the swap vector position in the info structure
 229  *
 230  * @info:  info structure for the inode
 231  * @index: index of the page to find
 232  * @page:  optional page to add to the structure. Has to be preset to
 233  *         all zeros
 234  *
 235  * If there is no space allocated yet it will return NULL when
 236  * page is NULL, else it will use the page for the needed block,
 237  * setting it to NULL on return to indicate that it has been used.
 238  *
 239  * The swap vector is organized the following way:
 240  *
 241  * There are SHMEM_NR_DIRECT entries directly stored in the
 242  * shmem_inode_info structure. So small files do not need an addional
 243  * allocation.
 244  *
 245  * For pages with index > SHMEM_NR_DIRECT there is the pointer
 246  * i_indirect which points to a page which holds in the first half
 247  * doubly indirect blocks, in the second half triple indirect blocks:
 248  *
 249  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
 250  * following layout (for SHMEM_NR_DIRECT == 16):
 251  *
 252  * i_indirect -> dir --> 16-19
 253  *            |      +-> 20-23
 254  *            |
 255  *            +-->dir2 --> 24-27
 256  *            |        +-> 28-31
 257  *            |        +-> 32-35
 258  *            |        +-> 36-39
 259  *            |
 260  *            +-->dir3 --> 40-43
 261  *                     +-> 44-47
 262  *                     +-> 48-51
 263  *                     +-> 52-55
 264  */
 265 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
 266 {
 267         unsigned long offset;
 268         struct page **dir;
 269         struct page *subdir;
 270
 271         if (index < SHMEM_NR_DIRECT) {
 272                 shmem_swp_balance_unmap();
 273                 return info->i_direct+index;
 274         }
 275         if (!info->i_indirect) {
 276                 if (page) {
 277                         info->i_indirect = *page;
 278                         *page = NULL;
 279                 }
 280                 return NULL;                    /* need another page */
 281         }
 282
 283         index -= SHMEM_NR_DIRECT;
 284         offset = index % ENTRIES_PER_PAGE;
 285         index /= ENTRIES_PER_PAGE;
 286         dir = shmem_dir_map(info->i_indirect);
 287
 288         if (index >= ENTRIES_PER_PAGE/2) {
 289                 index -= ENTRIES_PER_PAGE/2;
 290                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
 291                 index %= ENTRIES_PER_PAGE;
 292                 subdir = *dir;
 293                 if (!subdir) {
 294                         if (page) {
 295                                 *dir = *page;
 296                                 *page = NULL;
 297                         }
 298                         shmem_dir_unmap(dir);
 299                         return NULL;            /* need another page */
 300                 }
 301                 shmem_dir_unmap(dir);
 302                 dir = shmem_dir_map(subdir);
 303         }
 304
 305         dir += index;
 306         subdir = *dir;
 307         if (!subdir) {
 308                 if (!page || !(subdir = *page)) {
 309                         shmem_dir_unmap(dir);
 310                         return NULL;            /* need a page */
 311                 }
 312                 *dir = subdir;
 313                 *page = NULL;
 314         }
 315         shmem_dir_unmap(dir);
 316         return shmem_swp_map(subdir) + offset;
 317 }
 318
 319 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
 320 {
 321         long incdec = value? 1: -1;
 322
 323         entry->val = value;
 324         info->swapped += incdec;
 325         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
 326                 struct page *page = kmap_atomic_to_page(entry);
 327                 set_page_private(page, page_private(page) + incdec);
 328         }
 329 }
 330
 331 /*
 332  * shmem_swp_alloc - get the position of the swap entry for the page.
 333  *                   If it does not exist allocate the entry.
 334  *
 335  * @info:       info structure for the inode
 336  * @index:      index of the page to find
 337  * @sgp:        check and recheck i_size? skip allocation?
 338  */
 339 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
 340 {
 341         struct inode *inode = &info->vfs_inode;
 342         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 343         struct page *page = NULL;
 344         swp_entry_t *entry;
 345
 346         if (sgp != SGP_WRITE &&
 347             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
 348                 return ERR_PTR(-EINVAL);
 349
 350         while (!(entry = shmem_swp_entry(info, index, &page))) {
 351                 if (sgp == SGP_READ)
 352                         return shmem_swp_map(ZERO_PAGE(0));
 353                 /*
 354                  * Test free_blocks against 1 not 0, since we have 1 data
 355                  * page (and perhaps indirect index pages) yet to allocate:
 356                  * a waste to allocate index if we cannot allocate data.
 357                  */
 358                 if (sbinfo->max_blocks) {
 359                         spin_lock(&sbinfo->stat_lock);
 360                         if (sbinfo->free_blocks <= 1) {
 361                                 spin_unlock(&sbinfo->stat_lock);
 362                                 return ERR_PTR(-ENOSPC);
 363                         }
 364                         sbinfo->free_blocks--;
 365                         inode->i_blocks += BLOCKS_PER_PAGE;
 366                         spin_unlock(&sbinfo->stat_lock);
 367                 }
 368
 369                 spin_unlock(&info->lock);
 370                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping) | __GFP_ZERO);
 371                 if (page)
 372                         set_page_private(page, 0);
 373                 spin_lock(&info->lock);
 374
 375                 if (!page) {
 376                         shmem_free_blocks(inode, 1);
 377                         return ERR_PTR(-ENOMEM);
 378                 }
 379                 if (sgp != SGP_WRITE &&
 380                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
 381                         entry = ERR_PTR(-EINVAL);
 382                         break;
 383                 }
 384                 if (info->next_index <= index)
 385                         info->next_index = index + 1;
 386         }
 387         if (page) {
 388                 /* another task gave its page, or truncated the file */
 389                 shmem_free_blocks(inode, 1);
 390                 shmem_dir_free(page);
 391         }
 392         if (info->next_index <= index && !IS_ERR(entry))
 393                 info->next_index = index + 1;
 394         return entry;
 395 }
 396
 397 /*
 398  * shmem_free_swp - free some swap entries in a directory
 399  *
 400  * @dir:   pointer to the directory
 401  * @edir:  pointer after last entry of the directory
 402  */
 403 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir)
 404 {
 405         swp_entry_t *ptr;
 406         int freed = 0;
 407
 408         for (ptr = dir; ptr < edir; ptr++) {
 409                 if (ptr->val) {
 410                         free_swap_and_cache(*ptr);
 411                         *ptr = (swp_entry_t){0};
 412                         freed++;
 413                 }
 414         }
 415         return freed;
 416 }
 417
 418 static int shmem_map_and_free_swp(struct page *subdir,
 419                 int offset, int limit, struct page ***dir)
 420 {
 421         swp_entry_t *ptr;
 422         int freed = 0;
 423
 424         ptr = shmem_swp_map(subdir);
 425         for (; offset < limit; offset += LATENCY_LIMIT) {
 426                 int size = limit - offset;
 427                 if (size > LATENCY_LIMIT)
 428                         size = LATENCY_LIMIT;
 429                 freed += shmem_free_swp(ptr+offset, ptr+offset+size);
 430                 if (need_resched()) {
 431                         shmem_swp_unmap(ptr);
 432                         if (*dir) {
 433                                 shmem_dir_unmap(*dir);
 434                                 *dir = NULL;
 435                         }
 436                         cond_resched();
 437                         ptr = shmem_swp_map(subdir);
 438                 }
 439         }
 440         shmem_swp_unmap(ptr);
 441         return freed;
 442 }
 443
 444 static void shmem_free_pages(struct list_head *next)
 445 {
 446         struct page *page;
 447         int freed = 0;
 448
 449         do {
 450                 page = container_of(next, struct page, lru);
 451                 next = next->next;
 452                 shmem_dir_free(page);
 453                 freed++;
 454                 if (freed >= LATENCY_LIMIT) {
 455                         cond_resched();
 456                         freed = 0;
 457                 }
 458         } while (next);
 459 }
 460
 461 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
 462 {
 463         struct shmem_inode_info *info = SHMEM_I(inode);
 464         unsigned long idx;
 465         unsigned long size;
 466         unsigned long limit;
 467         unsigned long stage;
 468         unsigned long diroff;
 469         struct page **dir;
 470         struct page *topdir;
 471         struct page *middir;
 472         struct page *subdir;
 473         swp_entry_t *ptr;
 474         LIST_HEAD(pages_to_free);
 475         long nr_pages_to_free = 0;
 476         long nr_swaps_freed = 0;
 477         int offset;
 478         int freed;
 479         int punch_hole = 0;
 480
 481         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 482         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 483         if (idx >= info->next_index)
 484                 return;
 485
 486         spin_lock(&info->lock);
 487         info->flags |= SHMEM_TRUNCATE;
 488         if (likely(end == (loff_t) -1)) {
 489                 limit = info->next_index;
 490                 info->next_index = idx;
 491         } else {
 492                 limit = (end + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 493                 if (limit > info->next_index)
 494                         limit = info->next_index;
 495                 punch_hole = 1;
 496         }
 497
 498         topdir = info->i_indirect;
 499         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
 500                 info->i_indirect = NULL;
 501                 nr_pages_to_free++;
 502                 list_add(&topdir->lru, &pages_to_free);
 503         }
 504         spin_unlock(&info->lock);
 505
 506         if (info->swapped && idx < SHMEM_NR_DIRECT) {
 507                 ptr = info->i_direct;
 508                 size = limit;
 509                 if (size > SHMEM_NR_DIRECT)
 510                         size = SHMEM_NR_DIRECT;
 511                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size);
 512         }
 513         if (!topdir)
 514                 goto done2;
 515
 516         BUG_ON(limit <= SHMEM_NR_DIRECT);
 517         limit -= SHMEM_NR_DIRECT;
 518         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
 519         offset = idx % ENTRIES_PER_PAGE;
 520         idx -= offset;
 521
 522         dir = shmem_dir_map(topdir);
 523         stage = ENTRIES_PER_PAGEPAGE/2;
 524         if (idx < ENTRIES_PER_PAGEPAGE/2) {
 525                 middir = topdir;
 526                 diroff = idx/ENTRIES_PER_PAGE;
 527         } else {
 528                 dir += ENTRIES_PER_PAGE/2;
 529                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
 530                 while (stage <= idx)
 531                         stage += ENTRIES_PER_PAGEPAGE;
 532                 middir = *dir;
 533                 if (*dir) {
 534                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
 535                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
 536                         if (!diroff && !offset) {
 537                                 *dir = NULL;
 538                                 nr_pages_to_free++;
 539                                 list_add(&middir->lru, &pages_to_free);
 540                         }
 541                         shmem_dir_unmap(dir);
 542                         dir = shmem_dir_map(middir);
 543                 } else {
 544                         diroff = 0;
 545                         offset = 0;
 546                         idx = stage;
 547                 }
 548         }
 549
 550         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
 551                 if (unlikely(idx == stage)) {
 552                         shmem_dir_unmap(dir);
 553                         dir = shmem_dir_map(topdir) +
 554                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 555                         while (!*dir) {
 556                                 dir++;
 557                                 idx += ENTRIES_PER_PAGEPAGE;
 558                                 if (idx >= limit)
 559                                         goto done1;
 560                         }
 561                         stage = idx + ENTRIES_PER_PAGEPAGE;
 562                         middir = *dir;
 563                         *dir = NULL;
 564                         nr_pages_to_free++;
 565                         list_add(&middir->lru, &pages_to_free);
 566                         shmem_dir_unmap(dir);
 567                         cond_resched();
 568                         dir = shmem_dir_map(middir);
 569                         diroff = 0;
 570                 }
 571                 subdir = dir[diroff];
 572                 if (subdir && page_private(subdir)) {
 573                         size = limit - idx;
 574                         if (size > ENTRIES_PER_PAGE)
 575                                 size = ENTRIES_PER_PAGE;
 576                         freed = shmem_map_and_free_swp(subdir,
 577                                                 offset, size, &dir);
 578                         if (!dir)
 579                                 dir = shmem_dir_map(middir);
 580                         nr_swaps_freed += freed;
 581                         if (offset)
 582                                 spin_lock(&info->lock);
 583                         set_page_private(subdir, page_private(subdir) - freed);
 584                         if (offset)
 585                                 spin_unlock(&info->lock);
 586                         if (!punch_hole)
 587                                 BUG_ON(page_private(subdir) > offset);
 588                 }
 589                 if (offset)
 590                         offset = 0;
 591                 else if (subdir && !page_private(subdir)) {
 592                         dir[diroff] = NULL;
 593                         nr_pages_to_free++;
 594                         list_add(&subdir->lru, &pages_to_free);
 595                 }
 596         }
 597 done1:
 598         shmem_dir_unmap(dir);
 599 done2:
 600         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
 601                 /*
 602                  * Call truncate_inode_pages again: racing shmem_unuse_inode
 603                  * may have swizzled a page in from swap since vmtruncate or
 604                  * generic_delete_inode did it, before we lowered next_index.
 605                  * Also, though shmem_getpage checks i_size before adding to
 606                  * cache, no recheck after: so fix the narrow window there too.
 607                  */
 608                 truncate_inode_pages_range(inode->i_mapping, start, end);
 609         }
 610
 611         spin_lock(&info->lock);
 612         info->flags &= ~SHMEM_TRUNCATE;
 613         info->swapped -= nr_swaps_freed;
 614         if (nr_pages_to_free)
 615                 shmem_free_blocks(inode, nr_pages_to_free);
 616         shmem_recalc_inode(inode);
 617         spin_unlock(&info->lock);
 618
 619         /*
 620          * Empty swap vector directory pages to be freed?
 621          */
 622         if (!list_empty(&pages_to_free)) {
 623                 pages_to_free.prev->next = NULL;
 624                 shmem_free_pages(pages_to_free.next);
 625         }
 626 }
 627
 628 static void shmem_truncate(struct inode *inode)
 629 {
 630         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
 631 }
 632
 633 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
 634 {
 635         struct inode *inode = dentry->d_inode;
 636         struct page *page = NULL;
 637         int error;
 638
 639         if (attr->ia_valid & ATTR_SIZE) {
 640                 if (attr->ia_size < inode->i_size) {
 641                         /*
 642                          * If truncating down to a partial page, then
 643                          * if that page is already allocated, hold it
 644                          * in memory until the truncation is over, so
 645                          * truncate_partial_page cannnot miss it were
 646                          * it assigned to swap.
 647                          */
 648                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
 649                                 (void) shmem_getpage(inode,
 650                                         attr->ia_size>>PAGE_CACHE_SHIFT,
 651                                                 &page, SGP_READ, NULL);
 652                         }
 653                         /*
 654                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
 655                          * detect if any pages might have been added to cache
 656                          * after truncate_inode_pages.  But we needn't bother
 657                          * if it's being fully truncated to zero-length: the
 658                          * nrpages check is efficient enough in that case.
 659                          */
 660                         if (attr->ia_size) {
 661                                 struct shmem_inode_info *info = SHMEM_I(inode);
 662                                 spin_lock(&info->lock);
 663                                 info->flags &= ~SHMEM_PAGEIN;
 664                                 spin_unlock(&info->lock);
 665                         }
 666                 }
 667         }
 668
 669         error = inode_change_ok(inode, attr);
 670         if (!error)
 671                 error = inode_setattr(inode, attr);
 672         if (page)
 673                 page_cache_release(page);
 674         return error;
 675 }
 676
 677 static void shmem_delete_inode(struct inode *inode)
 678 {
 679         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 680         struct shmem_inode_info *info = SHMEM_I(inode);
 681
 682         if (inode->i_op->truncate == shmem_truncate) {
 683                 truncate_inode_pages(inode->i_mapping, 0);
 684                 shmem_unacct_size(info->flags, inode->i_size);
 685                 inode->i_size = 0;
 686                 shmem_truncate(inode);
 687                 if (!list_empty(&info->swaplist)) {
 688                         spin_lock(&shmem_swaplist_lock);
 689                         list_del_init(&info->swaplist);
 690                         spin_unlock(&shmem_swaplist_lock);
 691                 }
 692         }
 693         BUG_ON(inode->i_blocks);
 694         if (sbinfo->max_inodes) {
 695                 spin_lock(&sbinfo->stat_lock);
 696                 sbinfo->free_inodes++;
 697                 spin_unlock(&sbinfo->stat_lock);
 698         }
 699         clear_inode(inode);
 700 }
 701
 702 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
 703 {
 704         swp_entry_t *ptr;
 705
 706         for (ptr = dir; ptr < edir; ptr++) {
 707                 if (ptr->val == entry.val)
 708                         return ptr - dir;
 709         }
 710         return -1;
 711 }
 712
 713 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
 714 {
 715         struct inode *inode;
 716         unsigned long idx;
 717         unsigned long size;
 718         unsigned long limit;
 719         unsigned long stage;
 720         struct page **dir;
 721         struct page *subdir;
 722         swp_entry_t *ptr;
 723         int offset;
 724
 725         idx = 0;
 726         ptr = info->i_direct;
 727         spin_lock(&info->lock);
 728         limit = info->next_index;
 729         size = limit;
 730         if (size > SHMEM_NR_DIRECT)
 731                 size = SHMEM_NR_DIRECT;
 732         offset = shmem_find_swp(entry, ptr, ptr+size);
 733         if (offset >= 0) {
 734                 shmem_swp_balance_unmap();
 735                 goto found;
 736         }
 737         if (!info->i_indirect)
 738                 goto lost2;
 739
 740         dir = shmem_dir_map(info->i_indirect);
 741         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
 742
 743         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
 744                 if (unlikely(idx == stage)) {
 745                         shmem_dir_unmap(dir-1);
 746                         dir = shmem_dir_map(info->i_indirect) +
 747                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 748                         while (!*dir) {
 749                                 dir++;
 750                                 idx += ENTRIES_PER_PAGEPAGE;
 751                                 if (idx >= limit)
 752                                         goto lost1;
 753                         }
 754                         stage = idx + ENTRIES_PER_PAGEPAGE;
 755                         subdir = *dir;
 756                         shmem_dir_unmap(dir);
 757                         dir = shmem_dir_map(subdir);
 758                 }
 759                 subdir = *dir;
 760                 if (subdir && page_private(subdir)) {
 761                         ptr = shmem_swp_map(subdir);
 762                         size = limit - idx;
 763                         if (size > ENTRIES_PER_PAGE)
 764                                 size = ENTRIES_PER_PAGE;
 765                         offset = shmem_find_swp(entry, ptr, ptr+size);
 766                         if (offset >= 0) {
 767                                 shmem_dir_unmap(dir);
 768                                 goto found;
 769                         }
 770                         shmem_swp_unmap(ptr);
 771                 }
 772         }
 773 lost1:
 774         shmem_dir_unmap(dir-1);
 775 lost2:
 776         spin_unlock(&info->lock);
 777         return 0;
 778 found:
 779         idx += offset;
 780         inode = &info->vfs_inode;
 781         if (move_from_swap_cache(page, idx, inode->i_mapping) == 0) {
 782                 info->flags |= SHMEM_PAGEIN;
 783                 shmem_swp_set(info, ptr + offset, 0);
 784         }
 785         shmem_swp_unmap(ptr);
 786         spin_unlock(&info->lock);
 787         /*
 788          * Decrement swap count even when the entry is left behind:
 789          * try_to_unuse will skip over mms, then reincrement count.
 790          */
 791         swap_free(entry);
 792         return 1;
 793 }
 794
 795 /*
 796  * shmem_unuse() search for an eventually swapped out shmem page.
 797  */
 798 int shmem_unuse(swp_entry_t entry, struct page *page)
 799 {
 800         struct list_head *p, *next;
 801         struct shmem_inode_info *info;
 802         int found = 0;
 803
 804         spin_lock(&shmem_swaplist_lock);
 805         list_for_each_safe(p, next, &shmem_swaplist) {
 806                 info = list_entry(p, struct shmem_inode_info, swaplist);
 807                 if (!info->swapped)
 808                         list_del_init(&info->swaplist);
 809                 else if (shmem_unuse_inode(info, entry, page)) {
 810                         /* move head to start search for next from here */
 811                         list_move_tail(&shmem_swaplist, &info->swaplist);
 812                         found = 1;
 813                         break;
 814                 }
 815         }
 816         spin_unlock(&shmem_swaplist_lock);
 817         return found;
 818 }
 819
 820 /*
 821  * Move the page from the page cache to the swap cache.
 822  */
 823 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 824 {
 825         struct shmem_inode_info *info;
 826         swp_entry_t *entry, swap;
 827         struct address_space *mapping;
 828         unsigned long index;
 829         struct inode *inode;
 830
 831         BUG_ON(!PageLocked(page));
 832         BUG_ON(page_mapped(page));
 833
 834         mapping = page->mapping;
 835         index = page->index;
 836         inode = mapping->host;
 837         info = SHMEM_I(inode);
 838         if (info->flags & VM_LOCKED)
 839                 goto redirty;
 840         swap = get_swap_page();
 841         if (!swap.val)
 842                 goto redirty;
 843
 844         spin_lock(&info->lock);
 845         shmem_recalc_inode(inode);
 846         if (index >= info->next_index) {
 847                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
 848                 goto unlock;
 849         }
 850         entry = shmem_swp_entry(info, index, NULL);
 851         BUG_ON(!entry);
 852         BUG_ON(entry->val);
 853
 854         if (move_to_swap_cache(page, swap) == 0) {
 855                 shmem_swp_set(info, entry, swap.val);
 856                 shmem_swp_unmap(entry);
 857                 spin_unlock(&info->lock);
 858                 if (list_empty(&info->swaplist)) {
 859                         spin_lock(&shmem_swaplist_lock);
 860                         /* move instead of add in case we're racing */
 861                         list_move_tail(&info->swaplist, &shmem_swaplist);
 862                         spin_unlock(&shmem_swaplist_lock);
 863                 }
 864                 unlock_page(page);
 865                 return 0;
 866         }
 867
 868         shmem_swp_unmap(entry);
 869 unlock:
 870         spin_unlock(&info->lock);
 871         swap_free(swap);
 872 redirty:
 873         set_page_dirty(page);
 874         return AOP_WRITEPAGE_ACTIVATE;  /* Return with the page locked */
 875 }
 876
 877 #ifdef CONFIG_NUMA
 878 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
 879 {
 880         char *nodelist = strchr(value, ':');
 881         int err = 1;
 882
 883         if (nodelist) {
 884                 /* NUL-terminate policy string */
 885                 *nodelist++ = '\0';
 886                 if (nodelist_parse(nodelist, *policy_nodes))
 887                         goto out;
 888         }
 889         if (!strcmp(value, "default")) {
 890                 *policy = MPOL_DEFAULT;
 891                 /* Don't allow a nodelist */
 892                 if (!nodelist)
 893                         err = 0;
 894         } else if (!strcmp(value, "prefer")) {
 895                 *policy = MPOL_PREFERRED;
 896                 /* Insist on a nodelist of one node only */
 897                 if (nodelist) {
 898                         char *rest = nodelist;
 899                         while (isdigit(*rest))
 900                                 rest++;
 901                         if (!*rest)
 902                                 err = 0;
 903                 }
 904         } else if (!strcmp(value, "bind")) {
 905                 *policy = MPOL_BIND;
 906                 /* Insist on a nodelist */
 907                 if (nodelist)
 908                         err = 0;
 909         } else if (!strcmp(value, "interleave")) {
 910                 *policy = MPOL_INTERLEAVE;
 911                 /* Default to nodes online if no nodelist */
 912                 if (!nodelist)
 913                         *policy_nodes = node_online_map;
 914                 err = 0;
 915         }
 916 out:
 917         /* Restore string for error message */
 918         if (nodelist)
 919                 *--nodelist = ':';
 920         return err;
 921 }
 922
 923 static struct page *shmem_swapin_async(struct shared_policy *p,
 924                                        swp_entry_t entry, unsigned long idx)
 925 {
 926         struct page *page;
 927         struct vm_area_struct pvma;
 928
 929         /* Create a pseudo vma that just contains the policy */
 930         memset(&pvma, 0, sizeof(struct vm_area_struct));
 931         pvma.vm_end = PAGE_SIZE;
 932         pvma.vm_pgoff = idx;
 933         pvma.vm_policy = mpol_shared_policy_lookup(p, idx);
 934         page = read_swap_cache_async(entry, &pvma, 0);
 935         mpol_free(pvma.vm_policy);
 936         return page;
 937 }
 938
 939 struct page *shmem_swapin(struct shmem_inode_info *info, swp_entry_t entry,
 940                           unsigned long idx)
 941 {
 942         struct shared_policy *p = &info->policy;
 943         int i, num;
 944         struct page *page;
 945         unsigned long offset;
 946
 947         num = valid_swaphandles(entry, &offset);
 948         for (i = 0; i < num; offset++, i++) {
 949                 page = shmem_swapin_async(p,
 950                                 swp_entry(swp_type(entry), offset), idx);
 951                 if (!page)
 952                         break;
 953                 page_cache_release(page);
 954         }
 955         lru_add_drain();        /* Push any new pages onto the LRU now */
 956         return shmem_swapin_async(p, entry, idx);
 957 }
 958
 959 static struct page *
 960 shmem_alloc_page(gfp_t gfp, struct shmem_inode_info *info,
 961                  unsigned long idx)
 962 {
 963         struct vm_area_struct pvma;
 964         struct page *page;
 965
 966         memset(&pvma, 0, sizeof(struct vm_area_struct));
 967         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
 968         pvma.vm_pgoff = idx;
 969         pvma.vm_end = PAGE_SIZE;
 970         page = alloc_page_vma(gfp | __GFP_ZERO, &pvma, 0);
 971         mpol_free(pvma.vm_policy);
 972         return page;
 973 }
 974 #else
 975 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
 976 {
 977         return 1;
 978 }
 979
 980 static inline struct page *
 981 shmem_swapin(struct shmem_inode_info *info,swp_entry_t entry,unsigned long idx)
 982 {
 983         swapin_readahead(entry, 0, NULL);
 984         return read_swap_cache_async(entry, NULL, 0);
 985 }
 986
 987 static inline struct page *
 988 shmem_alloc_page(gfp_t gfp,struct shmem_inode_info *info, unsigned long idx)
 989 {
 990         return alloc_page(gfp | __GFP_ZERO);
 991 }
 992 #endif
 993
 994 /*
 995  * shmem_getpage - either get the page from swap or allocate a new one
 996  *
 997  * If we allocate a new one we do not mark it dirty. That's up to the
 998  * vm. If we swap it in we mark it dirty since we also free the swap
 999  * entry since a page cannot live in both the swap and page cache
1000  */
1001 static int shmem_getpage(struct inode *inode, unsigned long idx,
1002                         struct page **pagep, enum sgp_type sgp, int *type)
1003 {
1004         struct address_space *mapping = inode->i_mapping;
1005         struct shmem_inode_info *info = SHMEM_I(inode);
1006         struct shmem_sb_info *sbinfo;
1007         struct page *filepage = *pagep;
1008         struct page *swappage;
1009         swp_entry_t *entry;
1010         swp_entry_t swap;
1011         int error;
1012
1013         if (idx >= SHMEM_MAX_INDEX)
1014                 return -EFBIG;
1015         /*
1016          * Normally, filepage is NULL on entry, and either found
1017          * uptodate immediately, or allocated and zeroed, or read
1018          * in under swappage, which is then assigned to filepage.
1019          * But shmem_prepare_write passes in a locked filepage,
1020          * which may be found not uptodate by other callers too,
1021          * and may need to be copied from the swappage read in.
1022          */
1023 repeat:
1024         if (!filepage)
1025                 filepage = find_lock_page(mapping, idx);
1026         if (filepage && PageUptodate(filepage))
1027                 goto done;
1028         error = 0;
1029         if (sgp == SGP_QUICK)
1030                 goto failed;
1031
1032         spin_lock(&info->lock);
1033         shmem_recalc_inode(inode);
1034         entry = shmem_swp_alloc(info, idx, sgp);
1035         if (IS_ERR(entry)) {
1036                 spin_unlock(&info->lock);
1037                 error = PTR_ERR(entry);
1038                 goto failed;
1039         }
1040         swap = *entry;
1041
1042         if (swap.val) {
1043                 /* Look it up and read it in.. */
1044                 swappage = lookup_swap_cache(swap);
1045                 if (!swappage) {
1046                         shmem_swp_unmap(entry);
1047                         spin_unlock(&info->lock);
1048                         /* here we actually do the io */
1049                         if (type && *type == VM_FAULT_MINOR) {
1050                                 inc_page_state(pgmajfault);
1051                                 *type = VM_FAULT_MAJOR;
1052                         }
1053                         swappage = shmem_swapin(info, swap, idx);
1054                         if (!swappage) {
1055                                 spin_lock(&info->lock);
1056                                 entry = shmem_swp_alloc(info, idx, sgp);
1057                                 if (IS_ERR(entry))
1058                                         error = PTR_ERR(entry);
1059                                 else {
1060                                         if (entry->val == swap.val)
1061                                                 error = -ENOMEM;
1062                                         shmem_swp_unmap(entry);
1063                                 }
1064                                 spin_unlock(&info->lock);
1065                                 if (error)
1066                                         goto failed;
1067                                 goto repeat;
1068                         }
1069                         wait_on_page_locked(swappage);
1070                         page_cache_release(swappage);
1071                         goto repeat;
1072                 }
1073
1074                 /* We have to do this with page locked to prevent races */
1075                 if (TestSetPageLocked(swappage)) {
1076                         shmem_swp_unmap(entry);
1077                         spin_unlock(&info->lock);
1078                         wait_on_page_locked(swappage);
1079                         page_cache_release(swappage);
1080                         goto repeat;
1081                 }
1082                 if (!PageSwapCache(swappage)) {
1083                         /* Page migration has occured */
1084                         shmem_swp_unmap(entry);
1085                         spin_unlock(&info->lock);
1086                         unlock_page(swappage);
1087                         page_cache_release(swappage);
1088                         goto repeat;
1089                 }
1090                 if (PageWriteback(swappage)) {
1091                         shmem_swp_unmap(entry);
1092                         spin_unlock(&info->lock);
1093                         wait_on_page_writeback(swappage);
1094                         unlock_page(swappage);
1095                         page_cache_release(swappage);
1096                         goto repeat;
1097                 }
1098                 if (!PageUptodate(swappage)) {
1099                         shmem_swp_unmap(entry);
1100                         spin_unlock(&info->lock);
1101                         unlock_page(swappage);
1102                         page_cache_release(swappage);
1103                         error = -EIO;
1104                         goto failed;
1105                 }
1106
1107                 if (filepage) {
1108                         shmem_swp_set(info, entry, 0);
1109                         shmem_swp_unmap(entry);
1110                         delete_from_swap_cache(swappage);
1111                         spin_unlock(&info->lock);
1112                         copy_highpage(filepage, swappage);
1113                         unlock_page(swappage);
1114                         page_cache_release(swappage);
1115                         flush_dcache_page(filepage);
1116                         SetPageUptodate(filepage);
1117                         set_page_dirty(filepage);
1118                         swap_free(swap);
1119                 } else if (!(error = move_from_swap_cache(
1120                                 swappage, idx, mapping))) {
1121                         info->flags |= SHMEM_PAGEIN;
1122                         shmem_swp_set(info, entry, 0);
1123                         shmem_swp_unmap(entry);
1124                         spin_unlock(&info->lock);
1125                         filepage = swappage;
1126                         swap_free(swap);
1127                 } else {
1128                         shmem_swp_unmap(entry);
1129                         spin_unlock(&info->lock);
1130                         unlock_page(swappage);
1131                         page_cache_release(swappage);
1132                         if (error == -ENOMEM) {
1133                                 /* let kswapd refresh zone for GFP_ATOMICs */
1134                                 blk_congestion_wait(WRITE, HZ/50);
1135                         }
1136                         goto repeat;
1137                 }
1138         } else if (sgp == SGP_READ && !filepage) {
1139                 shmem_swp_unmap(entry);
1140                 filepage = find_get_page(mapping, idx);
1141                 if (filepage &&
1142                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1143                         spin_unlock(&info->lock);
1144                         wait_on_page_locked(filepage);
1145                         page_cache_release(filepage);
1146                         filepage = NULL;
1147                         goto repeat;
1148                 }
1149                 spin_unlock(&info->lock);
1150         } else {
1151                 shmem_swp_unmap(entry);
1152                 sbinfo = SHMEM_SB(inode->i_sb);
1153                 if (sbinfo->max_blocks) {
1154                         spin_lock(&sbinfo->stat_lock);
1155                         if (sbinfo->free_blocks == 0 ||
1156                             shmem_acct_block(info->flags)) {
1157                                 spin_unlock(&sbinfo->stat_lock);
1158                                 spin_unlock(&info->lock);
1159                                 error = -ENOSPC;
1160                                 goto failed;
1161                         }
1162                         sbinfo->free_blocks--;
1163                         inode->i_blocks += BLOCKS_PER_PAGE;
1164                         spin_unlock(&sbinfo->stat_lock);
1165                 } else if (shmem_acct_block(info->flags)) {
1166                         spin_unlock(&info->lock);
1167                         error = -ENOSPC;
1168                         goto failed;
1169                 }
1170
1171                 if (!filepage) {
1172                         spin_unlock(&info->lock);
1173                         filepage = shmem_alloc_page(mapping_gfp_mask(mapping),
1174                                                     info,
1175                                                     idx);
1176                         if (!filepage) {
1177                                 shmem_unacct_blocks(info->flags, 1);
1178                                 shmem_free_blocks(inode, 1);
1179                                 error = -ENOMEM;
1180                                 goto failed;
1181                         }
1182
1183                         spin_lock(&info->lock);
1184                         entry = shmem_swp_alloc(info, idx, sgp);
1185                         if (IS_ERR(entry))
1186                                 error = PTR_ERR(entry);
1187                         else {
1188                                 swap = *entry;
1189                                 shmem_swp_unmap(entry);
1190                         }
1191                         if (error || swap.val || 0 != add_to_page_cache_lru(
1192                                         filepage, mapping, idx, GFP_ATOMIC)) {
1193                                 spin_unlock(&info->lock);
1194                                 page_cache_release(filepage);
1195                                 shmem_unacct_blocks(info->flags, 1);
1196                                 shmem_free_blocks(inode, 1);
1197                                 filepage = NULL;
1198                                 if (error)
1199                                         goto failed;
1200                                 goto repeat;
1201                         }
1202                         info->flags |= SHMEM_PAGEIN;
1203                 }
1204
1205                 info->alloced++;
1206                 spin_unlock(&info->lock);
1207                 flush_dcache_page(filepage);
1208                 SetPageUptodate(filepage);
1209         }
1210 done:
1211         if (*pagep != filepage) {
1212                 unlock_page(filepage);
1213                 *pagep = filepage;
1214         }
1215         return 0;
1216
1217 failed:
1218         if (*pagep != filepage) {
1219                 unlock_page(filepage);
1220                 page_cache_release(filepage);
1221         }
1222         return error;
1223 }
1224
1225 struct page *shmem_nopage(struct vm_area_struct *vma, unsigned long address, int *type)
1226 {
1227         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1228         struct page *page = NULL;
1229         unsigned long idx;
1230         int error;
1231
1232         idx = (address - vma->vm_start) >> PAGE_SHIFT;
1233         idx += vma->vm_pgoff;
1234         idx >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
1235         if (((loff_t) idx << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1236                 return NOPAGE_SIGBUS;
1237
1238         error = shmem_getpage(inode, idx, &page, SGP_CACHE, type);
1239         if (error)
1240                 return (error == -ENOMEM)? NOPAGE_OOM: NOPAGE_SIGBUS;
1241
1242         mark_page_accessed(page);
1243         return page;
1244 }
1245
1246 static int shmem_populate(struct vm_area_struct *vma,
1247         unsigned long addr, unsigned long len,
1248         pgprot_t prot, unsigned long pgoff, int nonblock)
1249 {
1250         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1251         struct mm_struct *mm = vma->vm_mm;
1252         enum sgp_type sgp = nonblock? SGP_QUICK: SGP_CACHE;
1253         unsigned long size;
1254
1255         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1256         if (pgoff >= size || pgoff + (len >> PAGE_SHIFT) > size)
1257                 return -EINVAL;
1258
1259         while ((long) len > 0) {
1260                 struct page *page = NULL;
1261                 int err;
1262                 /*
1263                  * Will need changing if PAGE_CACHE_SIZE != PAGE_SIZE
1264                  */
1265                 err = shmem_getpage(inode, pgoff, &page, sgp, NULL);
1266                 if (err)
1267                         return err;
1268                 /* Page may still be null, but only if nonblock was set. */
1269                 if (page) {
1270                         mark_page_accessed(page);
1271                         err = install_page(mm, vma, addr, page, prot);
1272                         if (err) {
1273                                 page_cache_release(page);
1274                                 return err;
1275                         }
1276                 } else if (vma->vm_flags & VM_NONLINEAR) {
1277                         /* No page was found just because we can't read it in
1278                          * now (being here implies nonblock != 0), but the page
1279                          * may exist, so set the PTE to fault it in later. */
1280                         err = install_file_pte(mm, vma, addr, pgoff, prot);
1281                         if (err)
1282                                 return err;
1283                 }
1284
1285                 len -= PAGE_SIZE;
1286                 addr += PAGE_SIZE;
1287                 pgoff++;
1288         }
1289         return 0;
1290 }
1291
1292 #ifdef CONFIG_NUMA
1293 int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1294 {
1295         struct inode *i = vma->vm_file->f_dentry->d_inode;
1296         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1297 }
1298
1299 struct mempolicy *
1300 shmem_get_policy(struct vm_area_struct *vma, unsigned long addr)
1301 {
1302         struct inode *i = vma->vm_file->f_dentry->d_inode;
1303         unsigned long idx;
1304
1305         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1306         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1307 }
1308 #endif
1309
1310 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1311 {
1312         struct inode *inode = file->f_dentry->d_inode;
1313         struct shmem_inode_info *info = SHMEM_I(inode);
1314         int retval = -ENOMEM;
1315
1316         spin_lock(&info->lock);
1317         if (lock && !(info->flags & VM_LOCKED)) {
1318                 if (!user_shm_lock(inode->i_size, user))
1319                         goto out_nomem;
1320                 info->flags |= VM_LOCKED;
1321         }
1322         if (!lock && (info->flags & VM_LOCKED) && user) {
1323                 user_shm_unlock(inode->i_size, user);
1324                 info->flags &= ~VM_LOCKED;
1325         }
1326         retval = 0;
1327 out_nomem:
1328         spin_unlock(&info->lock);
1329         return retval;
1330 }
1331
1332 int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1333 {
1334         file_accessed(file);
1335         vma->vm_ops = &shmem_vm_ops;
1336         return 0;
1337 }
1338
1339 static struct inode *
1340 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1341 {
1342         struct inode *inode;
1343         struct shmem_inode_info *info;
1344         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1345
1346         if (sbinfo->max_inodes) {
1347                 spin_lock(&sbinfo->stat_lock);
1348                 if (!sbinfo->free_inodes) {
1349                         spin_unlock(&sbinfo->stat_lock);
1350                         return NULL;
1351                 }
1352                 sbinfo->free_inodes--;
1353                 spin_unlock(&sbinfo->stat_lock);
1354         }
1355
1356         inode = new_inode(sb);
1357         if (inode) {
1358                 inode->i_mode = mode;
1359                 inode->i_uid = current->fsuid;
1360                 inode->i_gid = current->fsgid;
1361                 inode->i_blksize = PAGE_CACHE_SIZE;
1362                 inode->i_blocks = 0;
1363                 inode->i_mapping->a_ops = &shmem_aops;
1364                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1365                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1366                 info = SHMEM_I(inode);
1367                 memset(info, 0, (char *)inode - (char *)info);
1368                 spin_lock_init(&info->lock);
1369                 INIT_LIST_HEAD(&info->swaplist);
1370
1371                 switch (mode & S_IFMT) {
1372                 default:
1373                         init_special_inode(inode, mode, dev);
1374                         break;
1375                 case S_IFREG:
1376                         inode->i_op = &shmem_inode_operations;
1377                         inode->i_fop = &shmem_file_operations;
1378                         mpol_shared_policy_init(&info->policy, sbinfo->policy,
1379                                                         &sbinfo->policy_nodes);
1380                         break;
1381                 case S_IFDIR:
1382                         inode->i_nlink++;
1383                         /* Some things misbehave if size == 0 on a directory */
1384                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1385                         inode->i_op = &shmem_dir_inode_operations;
1386                         inode->i_fop = &simple_dir_operations;
1387                         break;
1388                 case S_IFLNK:
1389                         /*
1390                          * Must not load anything in the rbtree,
1391                          * mpol_free_shared_policy will not be called.
1392                          */
1393                         mpol_shared_policy_init(&info->policy, MPOL_DEFAULT,
1394                                                 NULL);
1395                         break;
1396                 }
1397         } else if (sbinfo->max_inodes) {
1398                 spin_lock(&sbinfo->stat_lock);
1399                 sbinfo->free_inodes++;
1400                 spin_unlock(&sbinfo->stat_lock);
1401         }
1402         return inode;
1403 }
1404
1405 #ifdef CONFIG_TMPFS
1406 static struct inode_operations shmem_symlink_inode_operations;
1407 static struct inode_operations shmem_symlink_inline_operations;
1408
1409 /*
1410  * Normally tmpfs makes no use of shmem_prepare_write, but it
1411  * lets a tmpfs file be used read-write below the loop driver.
1412  */
1413 static int
1414 shmem_prepare_write(struct file *file, struct page *page, unsigned offset, unsigned to)
1415 {
1416         struct inode *inode = page->mapping->host;
1417         return shmem_getpage(inode, page->index, &page, SGP_WRITE, NULL);
1418 }
1419
1420 static ssize_t
1421 shmem_file_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
1422 {
1423         struct inode    *inode = file->f_dentry->d_inode;
1424         loff_t          pos;
1425         unsigned long   written;
1426         ssize_t         err;
1427
1428         if ((ssize_t) count < 0)
1429                 return -EINVAL;
1430
1431         if (!access_ok(VERIFY_READ, buf, count))
1432                 return -EFAULT;
1433
1434         mutex_lock(&inode->i_mutex);
1435
1436         pos = *ppos;
1437         written = 0;
1438
1439         err = generic_write_checks(file, &pos, &count, 0);
1440         if (err || !count)
1441                 goto out;
1442
1443         err = remove_suid(file->f_dentry);
1444         if (err)
1445                 goto out;
1446
1447         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1448
1449         do {
1450                 struct page *page = NULL;
1451                 unsigned long bytes, index, offset;
1452                 char *kaddr;
1453                 int left;
1454
1455                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1456                 index = pos >> PAGE_CACHE_SHIFT;
1457                 bytes = PAGE_CACHE_SIZE - offset;
1458                 if (bytes > count)
1459                         bytes = count;
1460
1461                 /*
1462                  * We don't hold page lock across copy from user -
1463                  * what would it guard against? - so no deadlock here.
1464                  * But it still may be a good idea to prefault below.
1465                  */
1466
1467                 err = shmem_getpage(inode, index, &page, SGP_WRITE, NULL);
1468                 if (err)
1469                         break;
1470
1471                 left = bytes;
1472                 if (PageHighMem(page)) {
1473                         volatile unsigned char dummy;
1474                         __get_user(dummy, buf);
1475                         __get_user(dummy, buf + bytes - 1);
1476
1477                         kaddr = kmap_atomic(page, KM_USER0);
1478                         left = __copy_from_user_inatomic(kaddr + offset,
1479                                                         buf, bytes);
1480                         kunmap_atomic(kaddr, KM_USER0);
1481                 }
1482                 if (left) {
1483                         kaddr = kmap(page);
1484                         left = __copy_from_user(kaddr + offset, buf, bytes);
1485                         kunmap(page);
1486                 }
1487
1488                 written += bytes;
1489                 count -= bytes;
1490                 pos += bytes;
1491                 buf += bytes;
1492                 if (pos > inode->i_size)
1493                         i_size_write(inode, pos);
1494
1495                 flush_dcache_page(page);
1496                 set_page_dirty(page);
1497                 mark_page_accessed(page);
1498                 page_cache_release(page);
1499
1500                 if (left) {
1501                         pos -= left;
1502                         written -= left;
1503                         err = -EFAULT;
1504                         break;
1505                 }
1506
1507                 /*
1508                  * Our dirty pages are not counted in nr_dirty,
1509                  * and we do not attempt to balance dirty pages.
1510                  */
1511
1512                 cond_resched();
1513         } while (count);
1514
1515         *ppos = pos;
1516         if (written)
1517                 err = written;
1518 out:
1519         mutex_unlock(&inode->i_mutex);
1520         return err;
1521 }
1522
1523 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1524 {
1525         struct inode *inode = filp->f_dentry->d_inode;
1526         struct address_space *mapping = inode->i_mapping;
1527         unsigned long index, offset;
1528
1529         index = *ppos >> PAGE_CACHE_SHIFT;
1530         offset = *ppos & ~PAGE_CACHE_MASK;
1531
1532         for (;;) {
1533                 struct page *page = NULL;
1534                 unsigned long end_index, nr, ret;
1535                 loff_t i_size = i_size_read(inode);
1536
1537                 end_index = i_size >> PAGE_CACHE_SHIFT;
1538                 if (index > end_index)
1539                         break;
1540                 if (index == end_index) {
1541                         nr = i_size & ~PAGE_CACHE_MASK;
1542                         if (nr <= offset)
1543                                 break;
1544                 }
1545
1546                 desc->error = shmem_getpage(inode, index, &page, SGP_READ, NULL);
1547                 if (desc->error) {
1548                         if (desc->error == -EINVAL)
1549                                 desc->error = 0;
1550                         break;
1551                 }
1552
1553                 /*
1554                  * We must evaluate after, since reads (unlike writes)
1555                  * are called without i_mutex protection against truncate
1556                  */
1557                 nr = PAGE_CACHE_SIZE;
1558                 i_size = i_size_read(inode);
1559                 end_index = i_size >> PAGE_CACHE_SHIFT;
1560                 if (index == end_index) {
1561                         nr = i_size & ~PAGE_CACHE_MASK;
1562                         if (nr <= offset) {
1563                                 if (page)
1564                                         page_cache_release(page);
1565                                 break;
1566                         }
1567                 }
1568                 nr -= offset;
1569
1570                 if (page) {
1571                         /*
1572                          * If users can be writing to this page using arbitrary
1573                          * virtual addresses, take care about potential aliasing
1574                          * before reading the page on the kernel side.
1575                          */
1576                         if (mapping_writably_mapped(mapping))
1577                                 flush_dcache_page(page);
1578                         /*
1579                          * Mark the page accessed if we read the beginning.
1580                          */
1581                         if (!offset)
1582                                 mark_page_accessed(page);
1583                 } else {
1584                         page = ZERO_PAGE(0);
1585                         page_cache_get(page);
1586                 }
1587
1588                 /*
1589                  * Ok, we have the page, and it's up-to-date, so
1590                  * now we can copy it to user space...
1591                  *
1592                  * The actor routine returns how many bytes were actually used..
1593                  * NOTE! This may not be the same as how much of a user buffer
1594                  * we filled up (we may be padding etc), so we can only update
1595                  * "pos" here (the actor routine has to update the user buffer
1596                  * pointers and the remaining count).
1597                  */
1598                 ret = actor(desc, page, offset, nr);
1599                 offset += ret;
1600                 index += offset >> PAGE_CACHE_SHIFT;
1601                 offset &= ~PAGE_CACHE_MASK;
1602
1603                 page_cache_release(page);
1604                 if (ret != nr || !desc->count)
1605                         break;
1606
1607                 cond_resched();
1608         }
1609
1610         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1611         file_accessed(filp);
1612 }
1613
1614 static ssize_t shmem_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1615 {
1616         read_descriptor_t desc;
1617
1618         if ((ssize_t) count < 0)
1619                 return -EINVAL;
1620         if (!access_ok(VERIFY_WRITE, buf, count))
1621                 return -EFAULT;
1622         if (!count)
1623                 return 0;
1624
1625         desc.written = 0;
1626         desc.count = count;
1627         desc.arg.buf = buf;
1628         desc.error = 0;
1629
1630         do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1631         if (desc.written)
1632                 return desc.written;
1633         return desc.error;
1634 }
1635
1636 static ssize_t shmem_file_sendfile(struct file *in_file, loff_t *ppos,
1637                          size_t count, read_actor_t actor, void *target)
1638 {
1639         read_descriptor_t desc;
1640
1641         if (!count)
1642                 return 0;
1643
1644         desc.written = 0;
1645         desc.count = count;
1646         desc.arg.data = target;
1647         desc.error = 0;
1648
1649         do_shmem_file_read(in_file, ppos, &desc, actor);
1650         if (desc.written)
1651                 return desc.written;
1652         return desc.error;
1653 }
1654
1655 static int shmem_statfs(struct super_block *sb, struct kstatfs *buf)
1656 {
1657         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1658
1659         buf->f_type = TMPFS_MAGIC;
1660         buf->f_bsize = PAGE_CACHE_SIZE;
1661         buf->f_namelen = NAME_MAX;
1662         spin_lock(&sbinfo->stat_lock);
1663         if (sbinfo->max_blocks) {
1664                 buf->f_blocks = sbinfo->max_blocks;
1665                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1666         }
1667         if (sbinfo->max_inodes) {
1668                 buf->f_files = sbinfo->max_inodes;
1669                 buf->f_ffree = sbinfo->free_inodes;
1670         }
1671         /* else leave those fields 0 like simple_statfs */
1672         spin_unlock(&sbinfo->stat_lock);
1673         return 0;
1674 }
1675
1676 /*
1677  * File creation. Allocate an inode, and we're done..
1678  */
1679 static int
1680 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1681 {
1682         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1683         int error = -ENOSPC;
1684
1685         if (inode) {
1686                 error = security_inode_init_security(inode, dir, NULL, NULL,
1687                                                      NULL);
1688                 if (error) {
1689                         if (error != -EOPNOTSUPP) {
1690                                 iput(inode);
1691                                 return error;
1692                         }
1693                         error = 0;
1694                 }
1695                 if (dir->i_mode & S_ISGID) {
1696                         inode->i_gid = dir->i_gid;
1697                         if (S_ISDIR(mode))
1698                                 inode->i_mode |= S_ISGID;
1699                 }
1700                 dir->i_size += BOGO_DIRENT_SIZE;
1701                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1702                 d_instantiate(dentry, inode);
1703                 dget(dentry); /* Extra count - pin the dentry in core */
1704         }
1705         return error;
1706 }
1707
1708 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1709 {
1710         int error;
1711
1712         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1713                 return error;
1714         dir->i_nlink++;
1715         return 0;
1716 }
1717
1718 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1719                 struct nameidata *nd)
1720 {
1721         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1722 }
1723
1724 /*
1725  * Link a file..
1726  */
1727 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1728 {
1729         struct inode *inode = old_dentry->d_inode;
1730         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1731
1732         /*
1733          * No ordinary (disk based) filesystem counts links as inodes;
1734          * but each new link needs a new dentry, pinning lowmem, and
1735          * tmpfs dentries cannot be pruned until they are unlinked.
1736          */
1737         if (sbinfo->max_inodes) {
1738                 spin_lock(&sbinfo->stat_lock);
1739                 if (!sbinfo->free_inodes) {
1740                         spin_unlock(&sbinfo->stat_lock);
1741                         return -ENOSPC;
1742                 }
1743                 sbinfo->free_inodes--;
1744                 spin_unlock(&sbinfo->stat_lock);
1745         }
1746
1747         dir->i_size += BOGO_DIRENT_SIZE;
1748         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1749         inode->i_nlink++;
1750         atomic_inc(&inode->i_count);    /* New dentry reference */
1751         dget(dentry);           /* Extra pinning count for the created dentry */
1752         d_instantiate(dentry, inode);
1753         return 0;
1754 }
1755
1756 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1757 {
1758         struct inode *inode = dentry->d_inode;
1759
1760         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode)) {
1761                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1762                 if (sbinfo->max_inodes) {
1763                         spin_lock(&sbinfo->stat_lock);
1764                         sbinfo->free_inodes++;
1765                         spin_unlock(&sbinfo->stat_lock);
1766                 }
1767         }
1768
1769         dir->i_size -= BOGO_DIRENT_SIZE;
1770         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1771         inode->i_nlink--;
1772         dput(dentry);   /* Undo the count from "create" - this does all the work */
1773         return 0;
1774 }
1775
1776 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1777 {
1778         if (!simple_empty(dentry))
1779                 return -ENOTEMPTY;
1780
1781         dir->i_nlink--;
1782         return shmem_unlink(dir, dentry);
1783 }
1784
1785 /*
1786  * The VFS layer already does all the dentry stuff for rename,
1787  * we just have to decrement the usage count for the target if
1788  * it exists so that the VFS layer correctly free's it when it
1789  * gets overwritten.
1790  */
1791 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1792 {
1793         struct inode *inode = old_dentry->d_inode;
1794         int they_are_dirs = S_ISDIR(inode->i_mode);
1795
1796         if (!simple_empty(new_dentry))
1797                 return -ENOTEMPTY;
1798
1799         if (new_dentry->d_inode) {
1800                 (void) shmem_unlink(new_dir, new_dentry);
1801                 if (they_are_dirs)
1802                         old_dir->i_nlink--;
1803         } else if (they_are_dirs) {
1804                 old_dir->i_nlink--;
1805                 new_dir->i_nlink++;
1806         }
1807
1808         old_dir->i_size -= BOGO_DIRENT_SIZE;
1809         new_dir->i_size += BOGO_DIRENT_SIZE;
1810         old_dir->i_ctime = old_dir->i_mtime =
1811         new_dir->i_ctime = new_dir->i_mtime =
1812         inode->i_ctime = CURRENT_TIME;
1813         return 0;
1814 }
1815
1816 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1817 {
1818         int error;
1819         int len;
1820         struct inode *inode;
1821         struct page *page = NULL;
1822         char *kaddr;
1823         struct shmem_inode_info *info;
1824
1825         len = strlen(symname) + 1;
1826         if (len > PAGE_CACHE_SIZE)
1827                 return -ENAMETOOLONG;
1828
1829         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1830         if (!inode)
1831                 return -ENOSPC;
1832
1833         error = security_inode_init_security(inode, dir, NULL, NULL,
1834                                              NULL);
1835         if (error) {
1836                 if (error != -EOPNOTSUPP) {
1837                         iput(inode);
1838                         return error;
1839                 }
1840                 error = 0;
1841         }
1842
1843         info = SHMEM_I(inode);
1844         inode->i_size = len-1;
1845         if (len <= (char *)inode - (char *)info) {
1846                 /* do it inline */
1847                 memcpy(info, symname, len);
1848                 inode->i_op = &shmem_symlink_inline_operations;
1849         } else {
1850                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1851                 if (error) {
1852                         iput(inode);
1853                         return error;
1854                 }
1855                 inode->i_op = &shmem_symlink_inode_operations;
1856                 kaddr = kmap_atomic(page, KM_USER0);
1857                 memcpy(kaddr, symname, len);
1858                 kunmap_atomic(kaddr, KM_USER0);
1859                 set_page_dirty(page);
1860                 page_cache_release(page);
1861         }
1862         if (dir->i_mode & S_ISGID)
1863                 inode->i_gid = dir->i_gid;
1864         dir->i_size += BOGO_DIRENT_SIZE;
1865         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1866         d_instantiate(dentry, inode);
1867         dget(dentry);
1868         return 0;
1869 }
1870
1871 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1872 {
1873         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1874         return NULL;
1875 }
1876
1877 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1878 {
1879         struct page *page = NULL;
1880         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1881         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1882         return page;
1883 }
1884
1885 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1886 {
1887         if (!IS_ERR(nd_get_link(nd))) {
1888                 struct page *page = cookie;
1889                 kunmap(page);
1890                 mark_page_accessed(page);
1891                 page_cache_release(page);
1892         }
1893 }
1894
1895 static struct inode_operations shmem_symlink_inline_operations = {
1896         .readlink       = generic_readlink,
1897         .follow_link    = shmem_follow_link_inline,
1898 };
1899
1900 static struct inode_operations shmem_symlink_inode_operations = {
1901         .truncate       = shmem_truncate,
1902         .readlink       = generic_readlink,
1903         .follow_link    = shmem_follow_link,
1904         .put_link       = shmem_put_link,
1905 };
1906
1907 static int shmem_parse_options(char *options, int *mode, uid_t *uid,
1908         gid_t *gid, unsigned long *blocks, unsigned long *inodes,
1909         int *policy, nodemask_t *policy_nodes)
1910 {
1911         char *this_char, *value, *rest;
1912
1913         while (options != NULL) {
1914                 this_char = options;
1915                 for (;;) {
1916                         /*
1917                          * NUL-terminate this option: unfortunately,
1918                          * mount options form a comma-separated list,
1919                          * but mpol's nodelist may also contain commas.
1920                          */
1921                         options = strchr(options, ',');
1922                         if (options == NULL)
1923                                 break;
1924                         options++;
1925                         if (!isdigit(*options)) {
1926                                 options[-1] = '\0';
1927                                 break;
1928                         }
1929                 }
1930                 if (!*this_char)
1931                         continue;
1932                 if ((value = strchr(this_char,'=')) != NULL) {
1933                         *value++ = 0;
1934                 } else {
1935                         printk(KERN_ERR
1936                             "tmpfs: No value for mount option '%s'\n",
1937                             this_char);
1938                         return 1;
1939                 }
1940
1941                 if (!strcmp(this_char,"size")) {
1942                         unsigned long long size;
1943                         size = memparse(value,&rest);
1944                         if (*rest == '%') {
1945                                 size <<= PAGE_SHIFT;
1946                                 size *= totalram_pages;
1947                                 do_div(size, 100);
1948                                 rest++;
1949                         }
1950                         if (*rest)
1951                                 goto bad_val;
1952                         *blocks = size >> PAGE_CACHE_SHIFT;
1953                 } else if (!strcmp(this_char,"nr_blocks")) {
1954                         *blocks = memparse(value,&rest);
1955                         if (*rest)
1956                                 goto bad_val;
1957                 } else if (!strcmp(this_char,"nr_inodes")) {
1958                         *inodes = memparse(value,&rest);
1959                         if (*rest)
1960                                 goto bad_val;
1961                 } else if (!strcmp(this_char,"mode")) {
1962                         if (!mode)
1963                                 continue;
1964                         *mode = simple_strtoul(value,&rest,8);
1965                         if (*rest)
1966                                 goto bad_val;
1967                 } else if (!strcmp(this_char,"uid")) {
1968                         if (!uid)
1969                                 continue;
1970                         *uid = simple_strtoul(value,&rest,0);
1971                         if (*rest)
1972                                 goto bad_val;
1973                 } else if (!strcmp(this_char,"gid")) {
1974                         if (!gid)
1975                                 continue;
1976                         *gid = simple_strtoul(value,&rest,0);
1977                         if (*rest)
1978                                 goto bad_val;
1979                 } else if (!strcmp(this_char,"mpol")) {
1980                         if (shmem_parse_mpol(value,policy,policy_nodes))
1981                                 goto bad_val;
1982                 } else {
1983                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
1984                                this_char);
1985                         return 1;
1986                 }
1987         }
1988         return 0;
1989
1990 bad_val:
1991         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
1992                value, this_char);
1993         return 1;
1994
1995 }
1996
1997 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
1998 {
1999         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2000         unsigned long max_blocks = sbinfo->max_blocks;
2001         unsigned long max_inodes = sbinfo->max_inodes;
2002         int policy = sbinfo->policy;
2003         nodemask_t policy_nodes = sbinfo->policy_nodes;
2004         unsigned long blocks;
2005         unsigned long inodes;
2006         int error = -EINVAL;
2007
2008         if (shmem_parse_options(data, NULL, NULL, NULL, &max_blocks,
2009                                 &max_inodes, &policy, &policy_nodes))
2010                 return error;
2011
2012         spin_lock(&sbinfo->stat_lock);
2013         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2014         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2015         if (max_blocks < blocks)
2016                 goto out;
2017         if (max_inodes < inodes)
2018                 goto out;
2019         /*
2020          * Those tests also disallow limited->unlimited while any are in
2021          * use, so i_blocks will always be zero when max_blocks is zero;
2022          * but we must separately disallow unlimited->limited, because
2023          * in that case we have no record of how much is already in use.
2024          */
2025         if (max_blocks && !sbinfo->max_blocks)
2026                 goto out;
2027         if (max_inodes && !sbinfo->max_inodes)
2028                 goto out;
2029
2030         error = 0;
2031         sbinfo->max_blocks  = max_blocks;
2032         sbinfo->free_blocks = max_blocks - blocks;
2033         sbinfo->max_inodes  = max_inodes;
2034         sbinfo->free_inodes = max_inodes - inodes;
2035         sbinfo->policy = policy;
2036         sbinfo->policy_nodes = policy_nodes;
2037 out:
2038         spin_unlock(&sbinfo->stat_lock);
2039         return error;
2040 }
2041 #endif
2042
2043 static void shmem_put_super(struct super_block *sb)
2044 {
2045         kfree(sb->s_fs_info);
2046         sb->s_fs_info = NULL;
2047 }
2048
2049 static int shmem_fill_super(struct super_block *sb,
2050                             void *data, int silent)
2051 {
2052         struct inode *inode;
2053         struct dentry *root;
2054         int mode   = S_IRWXUGO | S_ISVTX;
2055         uid_t uid = current->fsuid;
2056         gid_t gid = current->fsgid;
2057         int err = -ENOMEM;
2058         struct shmem_sb_info *sbinfo;
2059         unsigned long blocks = 0;
2060         unsigned long inodes = 0;
2061         int policy = MPOL_DEFAULT;
2062         nodemask_t policy_nodes = node_online_map;
2063
2064 #ifdef CONFIG_TMPFS
2065         /*
2066          * Per default we only allow half of the physical ram per
2067          * tmpfs instance, limiting inodes to one per page of lowmem;
2068          * but the internal instance is left unlimited.
2069          */
2070         if (!(sb->s_flags & MS_NOUSER)) {
2071                 blocks = totalram_pages / 2;
2072                 inodes = totalram_pages - totalhigh_pages;
2073                 if (inodes > blocks)
2074                         inodes = blocks;
2075                 if (shmem_parse_options(data, &mode, &uid, &gid, &blocks,
2076                                         &inodes, &policy, &policy_nodes))
2077                         return -EINVAL;
2078         }
2079 #else
2080         sb->s_flags |= MS_NOUSER;
2081 #endif
2082
2083         /* Round up to L1_CACHE_BYTES to resist false sharing */
2084         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2085                                 L1_CACHE_BYTES), GFP_KERNEL);
2086         if (!sbinfo)
2087                 return -ENOMEM;
2088
2089         spin_lock_init(&sbinfo->stat_lock);
2090         sbinfo->max_blocks = blocks;
2091         sbinfo->free_blocks = blocks;
2092         sbinfo->max_inodes = inodes;
2093         sbinfo->free_inodes = inodes;
2094         sbinfo->policy = policy;
2095         sbinfo->policy_nodes = policy_nodes;
2096
2097         sb->s_fs_info = sbinfo;
2098         sb->s_maxbytes = SHMEM_MAX_BYTES;
2099         sb->s_blocksize = PAGE_CACHE_SIZE;
2100         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2101         sb->s_magic = TMPFS_MAGIC;
2102         sb->s_op = &shmem_ops;
2103
2104         inode = shmem_get_inode(sb, S_IFDIR | mode, 0);
2105         if (!inode)
2106                 goto failed;
2107         inode->i_uid = uid;
2108         inode->i_gid = gid;
2109         root = d_alloc_root(inode);
2110         if (!root)
2111                 goto failed_iput;
2112         sb->s_root = root;
2113         return 0;
2114
2115 failed_iput:
2116         iput(inode);
2117 failed:
2118         shmem_put_super(sb);
2119         return err;
2120 }
2121
2122 static struct kmem_cache *shmem_inode_cachep;
2123
2124 static struct inode *shmem_alloc_inode(struct super_block *sb)
2125 {
2126         struct shmem_inode_info *p;
2127         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, SLAB_KERNEL);
2128         if (!p)
2129                 return NULL;
2130         return &p->vfs_inode;
2131 }
2132
2133 static void shmem_destroy_inode(struct inode *inode)
2134 {
2135         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2136                 /* only struct inode is valid if it's an inline symlink */
2137                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2138         }
2139         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2140 }
2141
2142 static void init_once(void *foo, struct kmem_cache *cachep,
2143                       unsigned long flags)
2144 {
2145         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2146
2147         if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
2148             SLAB_CTOR_CONSTRUCTOR) {
2149                 inode_init_once(&p->vfs_inode);
2150         }
2151 }
2152
2153 static int init_inodecache(void)
2154 {
2155         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2156                                 sizeof(struct shmem_inode_info),
2157                                 0, 0, init_once, NULL);
2158         if (shmem_inode_cachep == NULL)
2159                 return -ENOMEM;
2160         return 0;
2161 }
2162
2163 static void destroy_inodecache(void)
2164 {
2165         if (kmem_cache_destroy(shmem_inode_cachep))
2166                 printk(KERN_INFO "shmem_inode_cache: not all structures were freed\n");
2167 }
2168
2169 static struct address_space_operations shmem_aops = {
2170         .writepage      = shmem_writepage,
2171         .set_page_dirty = __set_page_dirty_nobuffers,
2172 #ifdef CONFIG_TMPFS
2173         .prepare_write  = shmem_prepare_write,
2174         .commit_write   = simple_commit_write,
2175 #endif
2176 };
2177
2178 static struct file_operations shmem_file_operations = {
2179         .mmap           = shmem_mmap,
2180 #ifdef CONFIG_TMPFS
2181         .llseek         = generic_file_llseek,
2182         .read           = shmem_file_read,
2183         .write          = shmem_file_write,
2184         .fsync          = simple_sync_file,
2185         .sendfile       = shmem_file_sendfile,
2186 #endif
2187 };
2188
2189 static struct inode_operations shmem_inode_operations = {
2190         .truncate       = shmem_truncate,
2191         .setattr        = shmem_notify_change,
2192         .truncate_range = shmem_truncate_range,
2193 };
2194
2195 static struct inode_operations shmem_dir_inode_operations = {
2196 #ifdef CONFIG_TMPFS
2197         .create         = shmem_create,
2198         .lookup         = simple_lookup,
2199         .link           = shmem_link,
2200         .unlink         = shmem_unlink,
2201         .symlink        = shmem_symlink,
2202         .mkdir          = shmem_mkdir,
2203         .rmdir          = shmem_rmdir,
2204         .mknod          = shmem_mknod,
2205         .rename         = shmem_rename,
2206 #endif
2207 };
2208
2209 static struct super_operations shmem_ops = {
2210         .alloc_inode    = shmem_alloc_inode,
2211         .destroy_inode  = shmem_destroy_inode,
2212 #ifdef CONFIG_TMPFS
2213         .statfs         = shmem_statfs,
2214         .remount_fs     = shmem_remount_fs,
2215 #endif
2216         .delete_inode   = shmem_delete_inode,
2217         .drop_inode     = generic_delete_inode,
2218         .put_super      = shmem_put_super,
2219 };
2220
2221 static struct vm_operations_struct shmem_vm_ops = {
2222         .nopage         = shmem_nopage,
2223         .populate       = shmem_populate,
2224 #ifdef CONFIG_NUMA
2225         .set_policy     = shmem_set_policy,
2226         .get_policy     = shmem_get_policy,
2227 #endif
2228 };
2229
2230
2231 static struct super_block *shmem_get_sb(struct file_system_type *fs_type,
2232         int flags, const char *dev_name, void *data)
2233 {
2234         return get_sb_nodev(fs_type, flags, data, shmem_fill_super);
2235 }
2236
2237 static struct file_system_type tmpfs_fs_type = {
2238         .owner          = THIS_MODULE,
2239         .name           = "tmpfs",
2240         .get_sb         = shmem_get_sb,
2241         .kill_sb        = kill_litter_super,
2242 };
2243 static struct vfsmount *shm_mnt;
2244
2245 static int __init init_tmpfs(void)
2246 {
2247         int error;
2248
2249         error = init_inodecache();
2250         if (error)
2251                 goto out3;
2252
2253         error = register_filesystem(&tmpfs_fs_type);
2254         if (error) {
2255                 printk(KERN_ERR "Could not register tmpfs\n");
2256                 goto out2;
2257         }
2258 #ifdef CONFIG_TMPFS
2259         devfs_mk_dir("shm");
2260 #endif
2261         shm_mnt = do_kern_mount(tmpfs_fs_type.name, MS_NOUSER,
2262                                 tmpfs_fs_type.name, NULL);
2263         if (IS_ERR(shm_mnt)) {
2264                 error = PTR_ERR(shm_mnt);
2265                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2266                 goto out1;
2267         }
2268         return 0;
2269
2270 out1:
2271         unregister_filesystem(&tmpfs_fs_type);
2272 out2:
2273         destroy_inodecache();
2274 out3:
2275         shm_mnt = ERR_PTR(error);
2276         return error;
2277 }
2278 module_init(init_tmpfs)
2279
2280 /*
2281  * shmem_file_setup - get an unlinked file living in tmpfs
2282  *
2283  * @name: name for dentry (to be seen in /proc/<pid>/maps
2284  * @size: size to be set for the file
2285  *
2286  */
2287 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2288 {
2289         int error;
2290         struct file *file;
2291         struct inode *inode;
2292         struct dentry *dentry, *root;
2293         struct qstr this;
2294
2295         if (IS_ERR(shm_mnt))
2296                 return (void *)shm_mnt;
2297
2298         if (size < 0 || size > SHMEM_MAX_BYTES)
2299                 return ERR_PTR(-EINVAL);
2300
2301         if (shmem_acct_size(flags, size))
2302                 return ERR_PTR(-ENOMEM);
2303
2304         error = -ENOMEM;
2305         this.name = name;
2306         this.len = strlen(name);
2307         this.hash = 0; /* will go */
2308         root = shm_mnt->mnt_root;
2309         dentry = d_alloc(root, &this);
2310         if (!dentry)
2311                 goto put_memory;
2312
2313         error = -ENFILE;
2314         file = get_empty_filp();
2315         if (!file)
2316                 goto put_dentry;
2317
2318         error = -ENOSPC;
2319         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2320         if (!inode)
2321                 goto close_file;
2322
2323         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2324         d_instantiate(dentry, inode);
2325         inode->i_size = size;
2326         inode->i_nlink = 0;     /* It is unlinked */
2327         file->f_vfsmnt = mntget(shm_mnt);
2328         file->f_dentry = dentry;
2329         file->f_mapping = inode->i_mapping;
2330         file->f_op = &shmem_file_operations;
2331         file->f_mode = FMODE_WRITE | FMODE_READ;
2332         return file;
2333
2334 close_file:
2335         put_filp(file);
2336 put_dentry:
2337         dput(dentry);
2338 put_memory:
2339         shmem_unacct_size(flags, size);
2340         return ERR_PTR(error);
2341 }
2342
2343 /*
2344  * shmem_zero_setup - setup a shared anonymous mapping
2345  *
2346  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2347  */
2348 int shmem_zero_setup(struct vm_area_struct *vma)
2349 {
2350         struct file *file;
2351         loff_t size = vma->vm_end - vma->vm_start;
2352
2353         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2354         if (IS_ERR(file))
2355                 return PTR_ERR(file);
2356
2357         if (vma->vm_file)
2358                 fput(vma->vm_file);
2359         vma->vm_file = file;
2360         vma->vm_ops = &shmem_vm_ops;
2361         return 0;
2362 }