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