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