mm/shmem.c

   1 /*
   2  * Resizable virtual memory filesystem for Linux.
   3  *
   4  * Copyright (C) 2000 Linus Torvalds.
   5  *               2000 Transmeta Corp.
   6  *               2000-2001 Christoph Rohland
   7  *               2000-2001 SAP AG
   8  *               2002 Red Hat Inc.
   9  * Copyright (C) 2002-2005 Hugh Dickins.
  10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  12  *
  13  * Extended attribute support for tmpfs:
  14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
  15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
  16  *
  17  * This file is released under the GPL.
  18  */
  19
  20 /*
  21  * This virtual memory filesystem is heavily based on the ramfs. It
  22  * extends ramfs by the ability to use swap and honor resource limits
  23  * which makes it a completely usable filesystem.
  24  */
  25
  26 #include <linux/config.h>
  27 #include <linux/module.h>
  28 #include <linux/init.h>
  29 #include <linux/devfs_fs_kernel.h>
  30 #include <linux/fs.h>
  31 #include <linux/mm.h>
  32 #include <linux/mman.h>
  33 #include <linux/file.h>
  34 #include <linux/swap.h>
  35 #include <linux/pagemap.h>
  36 #include <linux/string.h>
  37 #include <linux/slab.h>
  38 #include <linux/backing-dev.h>
  39 #include <linux/shmem_fs.h>
  40 #include <linux/mount.h>
  41 #include <linux/writeback.h>
  42 #include <linux/vfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/security.h>
  45 #include <linux/swapops.h>
  46 #include <linux/mempolicy.h>
  47 #include <linux/namei.h>
  48 #include <linux/ctype.h>
  49 #include <linux/migrate.h>
  50
  51 #include <asm/uaccess.h>
  52 #include <asm/div64.h>
  53 #include <asm/pgtable.h>
  54
  55 /* This magic number is used in glibc for posix shared memory */
  56 #define TMPFS_MAGIC     0x01021994
  57
  58 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
  59 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
  60 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  61
  62 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
  63 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
  64
  65 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  66
  67 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
  68 #define SHMEM_PAGEIN     VM_READ
  69 #define SHMEM_TRUNCATE   VM_WRITE
  70
  71 /* Definition to limit shmem_truncate's steps between cond_rescheds */
  72 #define LATENCY_LIMIT    64
  73
  74 /* Pretend that each entry is of this size in directory's i_size */
  75 #define BOGO_DIRENT_SIZE 20
  76
  77 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
  78 enum sgp_type {
  79         SGP_QUICK,      /* don't try more than file page cache lookup */
  80         SGP_READ,       /* don't exceed i_size, don't allocate page */
  81         SGP_CACHE,      /* don't exceed i_size, may allocate page */
  82         SGP_WRITE,      /* may exceed i_size, may allocate page */
  83 };
  84
  85 static int shmem_getpage(struct inode *inode, unsigned long idx,
  86                          struct page **pagep, enum sgp_type sgp, int *type);
  87
  88 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
  89 {
  90         /*
  91          * The above definition of ENTRIES_PER_PAGE, and the use of
  92          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
  93          * might be reconsidered if it ever diverges from PAGE_SIZE.
  94          */
  95         return alloc_pages(gfp_mask, PAGE_CACHE_SHIFT-PAGE_SHIFT);
  96 }
  97
  98 static inline void shmem_dir_free(struct page *page)
  99 {
 100         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
 101 }
 102
 103 static struct page **shmem_dir_map(struct page *page)
 104 {
 105         return (struct page **)kmap_atomic(page, KM_USER0);
 106 }
 107
 108 static inline void shmem_dir_unmap(struct page **dir)
 109 {
 110         kunmap_atomic(dir, KM_USER0);
 111 }
 112
 113 static swp_entry_t *shmem_swp_map(struct page *page)
 114 {
 115         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
 116 }
 117
 118 static inline void shmem_swp_balance_unmap(void)
 119 {
 120         /*
 121          * When passing a pointer to an i_direct entry, to code which
 122          * also handles indirect entries and so will shmem_swp_unmap,
 123          * we must arrange for the preempt count to remain in balance.
 124          * What kmap_atomic of a lowmem page does depends on config
 125          * and architecture, so pretend to kmap_atomic some lowmem page.
 126          */
 127         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
 128 }
 129
 130 static inline void shmem_swp_unmap(swp_entry_t *entry)
 131 {
 132         kunmap_atomic(entry, KM_USER1);
 133 }
 134
 135 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 136 {
 137         return sb->s_fs_info;
 138 }
 139
 140 /*
 141  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 142  * for shared memory and for shared anonymous (/dev/zero) mappings
 143  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 144  * consistent with the pre-accounting of private mappings ...
 145  */
 146 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 147 {
 148         return (flags & VM_ACCOUNT)?
 149                 security_vm_enough_memory(VM_ACCT(size)): 0;
 150 }
 151
 152 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 153 {
 154         if (flags & VM_ACCOUNT)
 155                 vm_unacct_memory(VM_ACCT(size));
 156 }
 157
 158 /*
 159  * ... whereas tmpfs objects are accounted incrementally as
 160  * pages are allocated, in order to allow huge sparse files.
 161  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 162  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 163  */
 164 static inline int shmem_acct_block(unsigned long flags)
 165 {
 166         return (flags & VM_ACCOUNT)?
 167                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
 168 }
 169
 170 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 171 {
 172         if (!(flags & VM_ACCOUNT))
 173                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 174 }
 175
 176 static struct super_operations shmem_ops;
 177 static struct address_space_operations shmem_aops;
 178 static struct file_operations shmem_file_operations;
 179 static struct inode_operations shmem_inode_operations;
 180 static struct inode_operations shmem_dir_inode_operations;
 181 static struct vm_operations_struct shmem_vm_ops;
 182
 183 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 184         .ra_pages       = 0,    /* No readahead */
 185         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
 186         .unplug_io_fn   = default_unplug_io_fn,
 187 };
 188
 189 static LIST_HEAD(shmem_swaplist);
 190 static DEFINE_SPINLOCK(shmem_swaplist_lock);
 191
 192 static void shmem_free_blocks(struct inode *inode, long pages)
 193 {
 194         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 195         if (sbinfo->max_blocks) {
 196                 spin_lock(&sbinfo->stat_lock);
 197                 sbinfo->free_blocks += pages;
 198                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
 199                 spin_unlock(&sbinfo->stat_lock);
 200         }
 201 }
 202
 203 /*
 204  * shmem_recalc_inode - recalculate the size of an inode
 205  *
 206  * @inode: inode to recalc
 207  *
 208  * We have to calculate the free blocks since the mm can drop
 209  * undirtied hole pages behind our back.
 210  *
 211  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 212  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 213  *
 214  * It has to be called with the spinlock held.
 215  */
 216 static void shmem_recalc_inode(struct inode *inode)
 217 {
 218         struct shmem_inode_info *info = SHMEM_I(inode);
 219         long freed;
 220
 221         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 222         if (freed > 0) {
 223                 info->alloced -= freed;
 224                 shmem_unacct_blocks(info->flags, freed);
 225                 shmem_free_blocks(inode, freed);
 226         }
 227 }
 228
 229 /*
 230  * shmem_swp_entry - find the swap vector position in the info structure
 231  *
 232  * @info:  info structure for the inode
 233  * @index: index of the page to find
 234  * @page:  optional page to add to the structure. Has to be preset to
 235  *         all zeros
 236  *
 237  * If there is no space allocated yet it will return NULL when
 238  * page is NULL, else it will use the page for the needed block,
 239  * setting it to NULL on return to indicate that it has been used.
 240  *
 241  * The swap vector is organized the following way:
 242  *
 243  * There are SHMEM_NR_DIRECT entries directly stored in the
 244  * shmem_inode_info structure. So small files do not need an addional
 245  * allocation.
 246  *
 247  * For pages with index > SHMEM_NR_DIRECT there is the pointer
 248  * i_indirect which points to a page which holds in the first half
 249  * doubly indirect blocks, in the second half triple indirect blocks:
 250  *
 251  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
 252  * following layout (for SHMEM_NR_DIRECT == 16):
 253  *
 254  * i_indirect -> dir --> 16-19
 255  *            |      +-> 20-23
 256  *            |
 257  *            +-->dir2 --> 24-27
 258  *            |        +-> 28-31
 259  *            |        +-> 32-35
 260  *            |        +-> 36-39
 261  *            |
 262  *            +-->dir3 --> 40-43
 263  *                     +-> 44-47
 264  *                     +-> 48-51
 265  *                     +-> 52-55
 266  */
 267 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
 268 {
 269         unsigned long offset;
 270         struct page **dir;
 271         struct page *subdir;
 272
 273         if (index < SHMEM_NR_DIRECT) {
 274                 shmem_swp_balance_unmap();
 275                 return info->i_direct+index;
 276         }
 277         if (!info->i_indirect) {
 278                 if (page) {
 279                         info->i_indirect = *page;
 280                         *page = NULL;
 281                 }
 282                 return NULL;                    /* need another page */
 283         }
 284
 285         index -= SHMEM_NR_DIRECT;
 286         offset = index % ENTRIES_PER_PAGE;
 287         index /= ENTRIES_PER_PAGE;
 288         dir = shmem_dir_map(info->i_indirect);
 289
 290         if (index >= ENTRIES_PER_PAGE/2) {
 291                 index -= ENTRIES_PER_PAGE/2;
 292                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
 293                 index %= ENTRIES_PER_PAGE;
 294                 subdir = *dir;
 295                 if (!subdir) {
 296                         if (page) {
 297                                 *dir = *page;
 298                                 *page = NULL;
 299                         }
 300                         shmem_dir_unmap(dir);
 301                         return NULL;            /* need another page */
 302                 }
 303                 shmem_dir_unmap(dir);
 304                 dir = shmem_dir_map(subdir);
 305         }
 306
 307         dir += index;
 308         subdir = *dir;
 309         if (!subdir) {
 310                 if (!page || !(subdir = *page)) {
 311                         shmem_dir_unmap(dir);
 312                         return NULL;            /* need a page */
 313                 }
 314                 *dir = subdir;
 315                 *page = NULL;
 316         }
 317         shmem_dir_unmap(dir);
 318         return shmem_swp_map(subdir) + offset;
 319 }
 320
 321 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
 322 {
 323         long incdec = value? 1: -1;
 324
 325         entry->val = value;
 326         info->swapped += incdec;
 327         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
 328                 struct page *page = kmap_atomic_to_page(entry);
 329                 set_page_private(page, page_private(page) + incdec);
 330         }
 331 }
 332
 333 /*
 334  * shmem_swp_alloc - get the position of the swap entry for the page.
 335  *                   If it does not exist allocate the entry.
 336  *
 337  * @info:       info structure for the inode
 338  * @index:      index of the page to find
 339  * @sgp:        check and recheck i_size? skip allocation?
 340  */
 341 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
 342 {
 343         struct inode *inode = &info->vfs_inode;
 344         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 345         struct page *page = NULL;
 346         swp_entry_t *entry;
 347
 348         if (sgp != SGP_WRITE &&
 349             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
 350                 return ERR_PTR(-EINVAL);
 351
 352         while (!(entry = shmem_swp_entry(info, index, &page))) {
 353                 if (sgp == SGP_READ)
 354                         return shmem_swp_map(ZERO_PAGE(0));
 355                 /*
 356                  * Test free_blocks against 1 not 0, since we have 1 data
 357                  * page (and perhaps indirect index pages) yet to allocate:
 358                  * a waste to allocate index if we cannot allocate data.
 359                  */
 360                 if (sbinfo->max_blocks) {
 361                         spin_lock(&sbinfo->stat_lock);
 362                         if (sbinfo->free_blocks <= 1) {
 363                                 spin_unlock(&sbinfo->stat_lock);
 364                                 return ERR_PTR(-ENOSPC);
 365                         }
 366                         sbinfo->free_blocks--;
 367                         inode->i_blocks += BLOCKS_PER_PAGE;
 368                         spin_unlock(&sbinfo->stat_lock);
 369                 }
 370
 371                 spin_unlock(&info->lock);
 372                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping) | __GFP_ZERO);
 373                 if (page)
 374                         set_page_private(page, 0);
 375                 spin_lock(&info->lock);
 376
 377                 if (!page) {
 378                         shmem_free_blocks(inode, 1);
 379                         return ERR_PTR(-ENOMEM);
 380                 }
 381                 if (sgp != SGP_WRITE &&
 382                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
 383                         entry = ERR_PTR(-EINVAL);
 384                         break;
 385                 }
 386                 if (info->next_index <= index)
 387                         info->next_index = index + 1;
 388         }
 389         if (page) {
 390                 /* another task gave its page, or truncated the file */
 391                 shmem_free_blocks(inode, 1);
 392                 shmem_dir_free(page);
 393         }
 394         if (info->next_index <= index && !IS_ERR(entry))
 395                 info->next_index = index + 1;
 396         return entry;
 397 }
 398
 399 /*
 400  * shmem_free_swp - free some swap entries in a directory
 401  *
 402  * @dir:   pointer to the directory
 403  * @edir:  pointer after last entry of the directory
 404  */
 405 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir)
 406 {
 407         swp_entry_t *ptr;
 408         int freed = 0;
 409
 410         for (ptr = dir; ptr < edir; ptr++) {
 411                 if (ptr->val) {
 412                         free_swap_and_cache(*ptr);
 413                         *ptr = (swp_entry_t){0};
 414                         freed++;
 415                 }
 416         }
 417         return freed;
 418 }
 419
 420 static int shmem_map_and_free_swp(struct page *subdir,
 421                 int offset, int limit, struct page ***dir)
 422 {
 423         swp_entry_t *ptr;
 424         int freed = 0;
 425
 426         ptr = shmem_swp_map(subdir);
 427         for (; offset < limit; offset += LATENCY_LIMIT) {
 428                 int size = limit - offset;
 429                 if (size > LATENCY_LIMIT)
 430                         size = LATENCY_LIMIT;
 431                 freed += shmem_free_swp(ptr+offset, ptr+offset+size);
 432                 if (need_resched()) {
 433                         shmem_swp_unmap(ptr);
 434                         if (*dir) {
 435                                 shmem_dir_unmap(*dir);
 436                                 *dir = NULL;
 437                         }
 438                         cond_resched();
 439                         ptr = shmem_swp_map(subdir);
 440                 }
 441         }
 442         shmem_swp_unmap(ptr);
 443         return freed;
 444 }
 445
 446 static void shmem_free_pages(struct list_head *next)
 447 {
 448         struct page *page;
 449         int freed = 0;
 450
 451         do {
 452                 page = container_of(next, struct page, lru);
 453                 next = next->next;
 454                 shmem_dir_free(page);
 455                 freed++;
 456                 if (freed >= LATENCY_LIMIT) {
 457                         cond_resched();
 458                         freed = 0;
 459                 }
 460         } while (next);
 461 }
 462
 463 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
 464 {
 465         struct shmem_inode_info *info = SHMEM_I(inode);
 466         unsigned long idx;
 467         unsigned long size;
 468         unsigned long limit;
 469         unsigned long stage;
 470         unsigned long diroff;
 471         struct page **dir;
 472         struct page *topdir;
 473         struct page *middir;
 474         struct page *subdir;
 475         swp_entry_t *ptr;
 476         LIST_HEAD(pages_to_free);
 477         long nr_pages_to_free = 0;
 478         long nr_swaps_freed = 0;
 479         int offset;
 480         int freed;
 481         int punch_hole = 0;
 482
 483         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 484         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 485         if (idx >= info->next_index)
 486                 return;
 487
 488         spin_lock(&info->lock);
 489         info->flags |= SHMEM_TRUNCATE;
 490         if (likely(end == (loff_t) -1)) {
 491                 limit = info->next_index;
 492                 info->next_index = idx;
 493         } else {
 494                 limit = (end + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 495                 if (limit > info->next_index)
 496                         limit = info->next_index;
 497                 punch_hole = 1;
 498         }
 499
 500         topdir = info->i_indirect;
 501         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
 502                 info->i_indirect = NULL;
 503                 nr_pages_to_free++;
 504                 list_add(&topdir->lru, &pages_to_free);
 505         }
 506         spin_unlock(&info->lock);
 507
 508         if (info->swapped && idx < SHMEM_NR_DIRECT) {
 509                 ptr = info->i_direct;
 510                 size = limit;
 511                 if (size > SHMEM_NR_DIRECT)
 512                         size = SHMEM_NR_DIRECT;
 513                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size);
 514         }
 515         if (!topdir)
 516                 goto done2;
 517
 518         BUG_ON(limit <= SHMEM_NR_DIRECT);
 519         limit -= SHMEM_NR_DIRECT;
 520         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
 521         offset = idx % ENTRIES_PER_PAGE;
 522         idx -= offset;
 523
 524         dir = shmem_dir_map(topdir);
 525         stage = ENTRIES_PER_PAGEPAGE/2;
 526         if (idx < ENTRIES_PER_PAGEPAGE/2) {
 527                 middir = topdir;
 528                 diroff = idx/ENTRIES_PER_PAGE;
 529         } else {
 530                 dir += ENTRIES_PER_PAGE/2;
 531                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
 532                 while (stage <= idx)
 533                         stage += ENTRIES_PER_PAGEPAGE;
 534                 middir = *dir;
 535                 if (*dir) {
 536                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
 537                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
 538                         if (!diroff && !offset) {
 539                                 *dir = NULL;
 540                                 nr_pages_to_free++;
 541                                 list_add(&middir->lru, &pages_to_free);
 542                         }
 543                         shmem_dir_unmap(dir);
 544                         dir = shmem_dir_map(middir);
 545                 } else {
 546                         diroff = 0;
 547                         offset = 0;
 548                         idx = stage;
 549                 }
 550         }
 551
 552         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
 553                 if (unlikely(idx == stage)) {
 554                         shmem_dir_unmap(dir);
 555                         dir = shmem_dir_map(topdir) +
 556                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 557                         while (!*dir) {
 558                                 dir++;
 559                                 idx += ENTRIES_PER_PAGEPAGE;
 560                                 if (idx >= limit)
 561                                         goto done1;
 562                         }
 563                         stage = idx + ENTRIES_PER_PAGEPAGE;
 564                         middir = *dir;
 565                         *dir = NULL;
 566                         nr_pages_to_free++;
 567                         list_add(&middir->lru, &pages_to_free);
 568                         shmem_dir_unmap(dir);
 569                         cond_resched();
 570                         dir = shmem_dir_map(middir);
 571                         diroff = 0;
 572                 }
 573                 subdir = dir[diroff];
 574                 if (subdir && page_private(subdir)) {
 575                         size = limit - idx;
 576                         if (size > ENTRIES_PER_PAGE)
 577                                 size = ENTRIES_PER_PAGE;
 578                         freed = shmem_map_and_free_swp(subdir,
 579                                                 offset, size, &dir);
 580                         if (!dir)
 581                                 dir = shmem_dir_map(middir);
 582                         nr_swaps_freed += freed;
 583                         if (offset)
 584                                 spin_lock(&info->lock);
 585                         set_page_private(subdir, page_private(subdir) - freed);
 586                         if (offset)
 587                                 spin_unlock(&info->lock);
 588                         if (!punch_hole)
 589                                 BUG_ON(page_private(subdir) > offset);
 590                 }
 591                 if (offset)
 592                         offset = 0;
 593                 else if (subdir && !page_private(subdir)) {
 594                         dir[diroff] = NULL;
 595                         nr_pages_to_free++;
 596                         list_add(&subdir->lru, &pages_to_free);
 597                 }
 598         }
 599 done1:
 600         shmem_dir_unmap(dir);
 601 done2:
 602         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
 603                 /*
 604                  * Call truncate_inode_pages again: racing shmem_unuse_inode
 605                  * may have swizzled a page in from swap since vmtruncate or
 606                  * generic_delete_inode did it, before we lowered next_index.
 607                  * Also, though shmem_getpage checks i_size before adding to
 608                  * cache, no recheck after: so fix the narrow window there too.
 609                  */
 610                 truncate_inode_pages_range(inode->i_mapping, start, end);
 611         }
 612
 613         spin_lock(&info->lock);
 614         info->flags &= ~SHMEM_TRUNCATE;
 615         info->swapped -= nr_swaps_freed;
 616         if (nr_pages_to_free)
 617                 shmem_free_blocks(inode, nr_pages_to_free);
 618         shmem_recalc_inode(inode);
 619         spin_unlock(&info->lock);
 620
 621         /*
 622          * Empty swap vector directory pages to be freed?
 623          */
 624         if (!list_empty(&pages_to_free)) {
 625                 pages_to_free.prev->next = NULL;
 626                 shmem_free_pages(pages_to_free.next);
 627         }
 628 }
 629
 630 static void shmem_truncate(struct inode *inode)
 631 {
 632         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
 633 }
 634
 635 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
 636 {
 637         struct inode *inode = dentry->d_inode;
 638         struct page *page = NULL;
 639         int error;
 640
 641         if (attr->ia_valid & ATTR_SIZE) {
 642                 if (attr->ia_size < inode->i_size) {
 643                         /*
 644                          * If truncating down to a partial page, then
 645                          * if that page is already allocated, hold it
 646                          * in memory until the truncation is over, so
 647                          * truncate_partial_page cannnot miss it were
 648                          * it assigned to swap.
 649                          */
 650                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
 651                                 (void) shmem_getpage(inode,
 652                                         attr->ia_size>>PAGE_CACHE_SHIFT,
 653                                                 &page, SGP_READ, NULL);
 654                         }
 655                         /*
 656                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
 657                          * detect if any pages might have been added to cache
 658                          * after truncate_inode_pages.  But we needn't bother
 659                          * if it's being fully truncated to zero-length: the
 660                          * nrpages check is efficient enough in that case.
 661                          */
 662                         if (attr->ia_size) {
 663                                 struct shmem_inode_info *info = SHMEM_I(inode);
 664                                 spin_lock(&info->lock);
 665                                 info->flags &= ~SHMEM_PAGEIN;
 666                                 spin_unlock(&info->lock);
 667                         }
 668                 }
 669         }
 670
 671         error = inode_change_ok(inode, attr);
 672         if (!error)
 673                 error = inode_setattr(inode, attr);
 674         if (page)
 675                 page_cache_release(page);
 676         return error;
 677 }
 678
 679 static void shmem_delete_inode(struct inode *inode)
 680 {
 681         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 682         struct shmem_inode_info *info = SHMEM_I(inode);
 683
 684         if (inode->i_op->truncate == shmem_truncate) {
 685                 truncate_inode_pages(inode->i_mapping, 0);
 686                 shmem_unacct_size(info->flags, inode->i_size);
 687                 inode->i_size = 0;
 688                 shmem_truncate(inode);
 689                 if (!list_empty(&info->swaplist)) {
 690                         spin_lock(&shmem_swaplist_lock);
 691                         list_del_init(&info->swaplist);
 692                         spin_unlock(&shmem_swaplist_lock);
 693                 }
 694         }
 695         BUG_ON(inode->i_blocks);
 696         if (sbinfo->max_inodes) {
 697                 spin_lock(&sbinfo->stat_lock);
 698                 sbinfo->free_inodes++;
 699                 spin_unlock(&sbinfo->stat_lock);
 700         }
 701         clear_inode(inode);
 702 }
 703
 704 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
 705 {
 706         swp_entry_t *ptr;
 707
 708         for (ptr = dir; ptr < edir; ptr++) {
 709                 if (ptr->val == entry.val)
 710                         return ptr - dir;
 711         }
 712         return -1;
 713 }
 714
 715 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
 716 {
 717         struct inode *inode;
 718         unsigned long idx;
 719         unsigned long size;
 720         unsigned long limit;
 721         unsigned long stage;
 722         struct page **dir;
 723         struct page *subdir;
 724         swp_entry_t *ptr;
 725         int offset;
 726
 727         idx = 0;
 728         ptr = info->i_direct;
 729         spin_lock(&info->lock);
 730         limit = info->next_index;
 731         size = limit;
 732         if (size > SHMEM_NR_DIRECT)
 733                 size = SHMEM_NR_DIRECT;
 734         offset = shmem_find_swp(entry, ptr, ptr+size);
 735         if (offset >= 0) {
 736                 shmem_swp_balance_unmap();
 737                 goto found;
 738         }
 739         if (!info->i_indirect)
 740                 goto lost2;
 741
 742         dir = shmem_dir_map(info->i_indirect);
 743         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
 744
 745         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
 746                 if (unlikely(idx == stage)) {
 747                         shmem_dir_unmap(dir-1);
 748                         dir = shmem_dir_map(info->i_indirect) +
 749                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 750                         while (!*dir) {
 751                                 dir++;
 752                                 idx += ENTRIES_PER_PAGEPAGE;
 753                                 if (idx >= limit)
 754                                         goto lost1;
 755                         }
 756                         stage = idx + ENTRIES_PER_PAGEPAGE;
 757                         subdir = *dir;
 758                         shmem_dir_unmap(dir);
 759                         dir = shmem_dir_map(subdir);
 760                 }
 761                 subdir = *dir;
 762                 if (subdir && page_private(subdir)) {
 763                         ptr = shmem_swp_map(subdir);
 764                         size = limit - idx;
 765                         if (size > ENTRIES_PER_PAGE)
 766                                 size = ENTRIES_PER_PAGE;
 767                         offset = shmem_find_swp(entry, ptr, ptr+size);
 768                         if (offset >= 0) {
 769                                 shmem_dir_unmap(dir);
 770                                 goto found;
 771                         }
 772                         shmem_swp_unmap(ptr);
 773                 }
 774         }
 775 lost1:
 776         shmem_dir_unmap(dir-1);
 777 lost2:
 778         spin_unlock(&info->lock);
 779         return 0;
 780 found:
 781         idx += offset;
 782         inode = &info->vfs_inode;
 783         if (move_from_swap_cache(page, idx, inode->i_mapping) == 0) {
 784                 info->flags |= SHMEM_PAGEIN;
 785                 shmem_swp_set(info, ptr + offset, 0);
 786         }
 787         shmem_swp_unmap(ptr);
 788         spin_unlock(&info->lock);
 789         /*
 790          * Decrement swap count even when the entry is left behind:
 791          * try_to_unuse will skip over mms, then reincrement count.
 792          */
 793         swap_free(entry);
 794         return 1;
 795 }
 796
 797 /*
 798  * shmem_unuse() search for an eventually swapped out shmem page.
 799  */
 800 int shmem_unuse(swp_entry_t entry, struct page *page)
 801 {
 802         struct list_head *p, *next;
 803         struct shmem_inode_info *info;
 804         int found = 0;
 805
 806         spin_lock(&shmem_swaplist_lock);
 807         list_for_each_safe(p, next, &shmem_swaplist) {
 808                 info = list_entry(p, struct shmem_inode_info, swaplist);
 809                 if (!info->swapped)
 810                         list_del_init(&info->swaplist);
 811                 else if (shmem_unuse_inode(info, entry, page)) {
 812                         /* move head to start search for next from here */
 813                         list_move_tail(&shmem_swaplist, &info->swaplist);
 814                         found = 1;
 815                         break;
 816                 }
 817         }
 818         spin_unlock(&shmem_swaplist_lock);
 819         return found;
 820 }
 821
 822 /*
 823  * Move the page from the page cache to the swap cache.
 824  */
 825 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 826 {
 827         struct shmem_inode_info *info;
 828         swp_entry_t *entry, swap;
 829         struct address_space *mapping;
 830         unsigned long index;
 831         struct inode *inode;
 832
 833         BUG_ON(!PageLocked(page));
 834         BUG_ON(page_mapped(page));
 835
 836         mapping = page->mapping;
 837         index = page->index;
 838         inode = mapping->host;
 839         info = SHMEM_I(inode);
 840         if (info->flags & VM_LOCKED)
 841                 goto redirty;
 842         swap = get_swap_page();
 843         if (!swap.val)
 844                 goto redirty;
 845
 846         spin_lock(&info->lock);
 847         shmem_recalc_inode(inode);
 848         if (index >= info->next_index) {
 849                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
 850                 goto unlock;
 851         }
 852         entry = shmem_swp_entry(info, index, NULL);
 853         BUG_ON(!entry);
 854         BUG_ON(entry->val);
 855
 856         if (move_to_swap_cache(page, swap) == 0) {
 857                 shmem_swp_set(info, entry, swap.val);
 858                 shmem_swp_unmap(entry);
 859                 spin_unlock(&info->lock);
 860                 if (list_empty(&info->swaplist)) {
 861                         spin_lock(&shmem_swaplist_lock);
 862                         /* move instead of add in case we're racing */
 863                         list_move_tail(&info->swaplist, &shmem_swaplist);
 864                         spin_unlock(&shmem_swaplist_lock);
 865                 }
 866                 unlock_page(page);
 867                 return 0;
 868         }
 869
 870         shmem_swp_unmap(entry);
 871 unlock:
 872         spin_unlock(&info->lock);
 873         swap_free(swap);
 874 redirty:
 875         set_page_dirty(page);
 876         return AOP_WRITEPAGE_ACTIVATE;  /* Return with the page locked */
 877 }
 878
 879 #ifdef CONFIG_NUMA
 880 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
 881 {
 882         char *nodelist = strchr(value, ':');
 883         int err = 1;
 884
 885         if (nodelist) {
 886                 /* NUL-terminate policy string */
 887                 *nodelist++ = '\0';
 888                 if (nodelist_parse(nodelist, *policy_nodes))
 889                         goto out;
 890         }
 891         if (!strcmp(value, "default")) {
 892                 *policy = MPOL_DEFAULT;
 893                 /* Don't allow a nodelist */
 894                 if (!nodelist)
 895                         err = 0;
 896         } else if (!strcmp(value, "prefer")) {
 897                 *policy = MPOL_PREFERRED;
 898                 /* Insist on a nodelist of one node only */
 899                 if (nodelist) {
 900                         char *rest = nodelist;
 901                         while (isdigit(*rest))
 902                                 rest++;
 903                         if (!*rest)
 904                                 err = 0;
 905                 }
 906         } else if (!strcmp(value, "bind")) {
 907                 *policy = MPOL_BIND;
 908                 /* Insist on a nodelist */
 909                 if (nodelist)
 910                         err = 0;
 911         } else if (!strcmp(value, "interleave")) {
 912                 *policy = MPOL_INTERLEAVE;
 913                 /* Default to nodes online if no nodelist */
 914                 if (!nodelist)
 915                         *policy_nodes = node_online_map;
 916                 err = 0;
 917         }
 918 out:
 919         /* Restore string for error message */
 920         if (nodelist)
 921                 *--nodelist = ':';
 922         return err;
 923 }
 924
 925 static struct page *shmem_swapin_async(struct shared_policy *p,
 926                                        swp_entry_t entry, unsigned long idx)
 927 {
 928         struct page *page;
 929         struct vm_area_struct pvma;
 930
 931         /* Create a pseudo vma that just contains the policy */
 932         memset(&pvma, 0, sizeof(struct vm_area_struct));
 933         pvma.vm_end = PAGE_SIZE;
 934         pvma.vm_pgoff = idx;
 935         pvma.vm_policy = mpol_shared_policy_lookup(p, idx);
 936         page = read_swap_cache_async(entry, &pvma, 0);
 937         mpol_free(pvma.vm_policy);
 938         return page;
 939 }
 940
 941 struct page *shmem_swapin(struct shmem_inode_info *info, swp_entry_t entry,
 942                           unsigned long idx)
 943 {
 944         struct shared_policy *p = &info->policy;
 945         int i, num;
 946         struct page *page;
 947         unsigned long offset;
 948
 949         num = valid_swaphandles(entry, &offset);
 950         for (i = 0; i < num; offset++, i++) {
 951                 page = shmem_swapin_async(p,
 952                                 swp_entry(swp_type(entry), offset), idx);
 953                 if (!page)
 954                         break;
 955                 page_cache_release(page);
 956         }
 957         lru_add_drain();        /* Push any new pages onto the LRU now */
 958         return shmem_swapin_async(p, entry, idx);
 959 }
 960
 961 static struct page *
 962 shmem_alloc_page(gfp_t gfp, struct shmem_inode_info *info,
 963                  unsigned long idx)
 964 {
 965         struct vm_area_struct pvma;
 966         struct page *page;
 967
 968         memset(&pvma, 0, sizeof(struct vm_area_struct));
 969         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
 970         pvma.vm_pgoff = idx;
 971         pvma.vm_end = PAGE_SIZE;
 972         page = alloc_page_vma(gfp | __GFP_ZERO, &pvma, 0);
 973         mpol_free(pvma.vm_policy);
 974         return page;
 975 }
 976 #else
 977 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
 978 {
 979         return 1;
 980 }
 981
 982 static inline struct page *
 983 shmem_swapin(struct shmem_inode_info *info,swp_entry_t entry,unsigned long idx)
 984 {
 985         swapin_readahead(entry, 0, NULL);
 986         return read_swap_cache_async(entry, NULL, 0);
 987 }
 988
 989 static inline struct page *
 990 shmem_alloc_page(gfp_t gfp,struct shmem_inode_info *info, unsigned long idx)
 991 {
 992         return alloc_page(gfp | __GFP_ZERO);
 993 }
 994 #endif
 995
 996 /*
 997  * shmem_getpage - either get the page from swap or allocate a new one
 998  *
 999  * If we allocate a new one we do not mark it dirty. That's up to the
1000  * vm. If we swap it in we mark it dirty since we also free the swap
1001  * entry since a page cannot live in both the swap and page cache
1002  */
1003 static int shmem_getpage(struct inode *inode, unsigned long idx,
1004                         struct page **pagep, enum sgp_type sgp, int *type)
1005 {
1006         struct address_space *mapping = inode->i_mapping;
1007         struct shmem_inode_info *info = SHMEM_I(inode);
1008         struct shmem_sb_info *sbinfo;
1009         struct page *filepage = *pagep;
1010         struct page *swappage;
1011         swp_entry_t *entry;
1012         swp_entry_t swap;
1013         int error;
1014
1015         if (idx >= SHMEM_MAX_INDEX)
1016                 return -EFBIG;
1017         /*
1018          * Normally, filepage is NULL on entry, and either found
1019          * uptodate immediately, or allocated and zeroed, or read
1020          * in under swappage, which is then assigned to filepage.
1021          * But shmem_prepare_write passes in a locked filepage,
1022          * which may be found not uptodate by other callers too,
1023          * and may need to be copied from the swappage read in.
1024          */
1025 repeat:
1026         if (!filepage)
1027                 filepage = find_lock_page(mapping, idx);
1028         if (filepage && PageUptodate(filepage))
1029                 goto done;
1030         error = 0;
1031         if (sgp == SGP_QUICK)
1032                 goto failed;
1033
1034         spin_lock(&info->lock);
1035         shmem_recalc_inode(inode);
1036         entry = shmem_swp_alloc(info, idx, sgp);
1037         if (IS_ERR(entry)) {
1038                 spin_unlock(&info->lock);
1039                 error = PTR_ERR(entry);
1040                 goto failed;
1041         }
1042         swap = *entry;
1043
1044         if (swap.val) {
1045                 /* Look it up and read it in.. */
1046                 swappage = lookup_swap_cache(swap);
1047                 if (!swappage) {
1048                         shmem_swp_unmap(entry);
1049                         spin_unlock(&info->lock);
1050                         /* here we actually do the io */
1051                         if (type && *type == VM_FAULT_MINOR) {
1052                                 inc_page_state(pgmajfault);
1053                                 *type = VM_FAULT_MAJOR;
1054                         }
1055                         swappage = shmem_swapin(info, swap, idx);
1056                         if (!swappage) {
1057                                 spin_lock(&info->lock);
1058                                 entry = shmem_swp_alloc(info, idx, sgp);
1059                                 if (IS_ERR(entry))
1060                                         error = PTR_ERR(entry);
1061                                 else {
1062                                         if (entry->val == swap.val)
1063                                                 error = -ENOMEM;
1064                                         shmem_swp_unmap(entry);
1065                                 }
1066                                 spin_unlock(&info->lock);
1067                                 if (error)
1068                                         goto failed;
1069                                 goto repeat;
1070                         }
1071                         wait_on_page_locked(swappage);
1072                         page_cache_release(swappage);
1073                         goto repeat;
1074                 }
1075
1076                 /* We have to do this with page locked to prevent races */
1077                 if (TestSetPageLocked(swappage)) {
1078                         shmem_swp_unmap(entry);
1079                         spin_unlock(&info->lock);
1080                         wait_on_page_locked(swappage);
1081                         page_cache_release(swappage);
1082                         goto repeat;
1083                 }
1084                 if (!PageSwapCache(swappage)) {
1085                         /* Page migration has occured */
1086                         shmem_swp_unmap(entry);
1087                         spin_unlock(&info->lock);
1088                         unlock_page(swappage);
1089                         page_cache_release(swappage);
1090                         goto repeat;
1091                 }
1092                 if (PageWriteback(swappage)) {
1093                         shmem_swp_unmap(entry);
1094                         spin_unlock(&info->lock);
1095                         wait_on_page_writeback(swappage);
1096                         unlock_page(swappage);
1097                         page_cache_release(swappage);
1098                         goto repeat;
1099                 }
1100                 if (!PageUptodate(swappage)) {
1101                         shmem_swp_unmap(entry);
1102                         spin_unlock(&info->lock);
1103                         unlock_page(swappage);
1104                         page_cache_release(swappage);
1105                         error = -EIO;
1106                         goto failed;
1107                 }
1108
1109                 if (filepage) {
1110                         shmem_swp_set(info, entry, 0);
1111                         shmem_swp_unmap(entry);
1112                         delete_from_swap_cache(swappage);
1113                         spin_unlock(&info->lock);
1114                         copy_highpage(filepage, swappage);
1115                         unlock_page(swappage);
1116                         page_cache_release(swappage);
1117                         flush_dcache_page(filepage);
1118                         SetPageUptodate(filepage);
1119                         set_page_dirty(filepage);
1120                         swap_free(swap);
1121                 } else if (!(error = move_from_swap_cache(
1122                                 swappage, idx, mapping))) {
1123                         info->flags |= SHMEM_PAGEIN;
1124                         shmem_swp_set(info, entry, 0);
1125                         shmem_swp_unmap(entry);
1126                         spin_unlock(&info->lock);
1127                         filepage = swappage;
1128                         swap_free(swap);
1129                 } else {
1130                         shmem_swp_unmap(entry);
1131                         spin_unlock(&info->lock);
1132                         unlock_page(swappage);
1133                         page_cache_release(swappage);
1134                         if (error == -ENOMEM) {
1135                                 /* let kswapd refresh zone for GFP_ATOMICs */
1136                                 blk_congestion_wait(WRITE, HZ/50);
1137                         }
1138                         goto repeat;
1139                 }
1140         } else if (sgp == SGP_READ && !filepage) {
1141                 shmem_swp_unmap(entry);
1142                 filepage = find_get_page(mapping, idx);
1143                 if (filepage &&
1144                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1145                         spin_unlock(&info->lock);
1146                         wait_on_page_locked(filepage);
1147                         page_cache_release(filepage);
1148                         filepage = NULL;
1149                         goto repeat;
1150                 }
1151                 spin_unlock(&info->lock);
1152         } else {
1153                 shmem_swp_unmap(entry);
1154                 sbinfo = SHMEM_SB(inode->i_sb);
1155                 if (sbinfo->max_blocks) {
1156                         spin_lock(&sbinfo->stat_lock);
1157                         if (sbinfo->free_blocks == 0 ||
1158                             shmem_acct_block(info->flags)) {
1159                                 spin_unlock(&sbinfo->stat_lock);
1160                                 spin_unlock(&info->lock);
1161                                 error = -ENOSPC;
1162                                 goto failed;
1163                         }
1164                         sbinfo->free_blocks--;
1165                         inode->i_blocks += BLOCKS_PER_PAGE;
1166                         spin_unlock(&sbinfo->stat_lock);
1167                 } else if (shmem_acct_block(info->flags)) {
1168                         spin_unlock(&info->lock);
1169                         error = -ENOSPC;
1170                         goto failed;
1171                 }
1172
1173                 if (!filepage) {
1174                         spin_unlock(&info->lock);
1175                         filepage = shmem_alloc_page(mapping_gfp_mask(mapping),
1176                                                     info,
1177                                                     idx);
1178                         if (!filepage) {
1179                                 shmem_unacct_blocks(info->flags, 1);
1180                                 shmem_free_blocks(inode, 1);
1181                                 error = -ENOMEM;
1182                                 goto failed;
1183                         }
1184
1185                         spin_lock(&info->lock);
1186                         entry = shmem_swp_alloc(info, idx, sgp);
1187                         if (IS_ERR(entry))
1188                                 error = PTR_ERR(entry);
1189                         else {
1190                                 swap = *entry;
1191                                 shmem_swp_unmap(entry);
1192                         }
1193                         if (error || swap.val || 0 != add_to_page_cache_lru(
1194                                         filepage, mapping, idx, GFP_ATOMIC)) {
1195                                 spin_unlock(&info->lock);
1196                                 page_cache_release(filepage);
1197                                 shmem_unacct_blocks(info->flags, 1);
1198                                 shmem_free_blocks(inode, 1);
1199                                 filepage = NULL;
1200                                 if (error)
1201                                         goto failed;
1202                                 goto repeat;
1203                         }
1204                         info->flags |= SHMEM_PAGEIN;
1205                 }
1206
1207                 info->alloced++;
1208                 spin_unlock(&info->lock);
1209                 flush_dcache_page(filepage);
1210                 SetPageUptodate(filepage);
1211         }
1212 done:
1213         if (*pagep != filepage) {
1214                 unlock_page(filepage);
1215                 *pagep = filepage;
1216         }
1217         return 0;
1218
1219 failed:
1220         if (*pagep != filepage) {
1221                 unlock_page(filepage);
1222                 page_cache_release(filepage);
1223         }
1224         return error;
1225 }
1226
1227 struct page *shmem_nopage(struct vm_area_struct *vma, unsigned long address, int *type)
1228 {
1229         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1230         struct page *page = NULL;
1231         unsigned long idx;
1232         int error;
1233
1234         idx = (address - vma->vm_start) >> PAGE_SHIFT;
1235         idx += vma->vm_pgoff;
1236         idx >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
1237         if (((loff_t) idx << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1238                 return NOPAGE_SIGBUS;
1239
1240         error = shmem_getpage(inode, idx, &page, SGP_CACHE, type);
1241         if (error)
1242                 return (error == -ENOMEM)? NOPAGE_OOM: NOPAGE_SIGBUS;
1243
1244         mark_page_accessed(page);
1245         return page;
1246 }
1247
1248 static int shmem_populate(struct vm_area_struct *vma,
1249         unsigned long addr, unsigned long len,
1250         pgprot_t prot, unsigned long pgoff, int nonblock)
1251 {
1252         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1253         struct mm_struct *mm = vma->vm_mm;
1254         enum sgp_type sgp = nonblock? SGP_QUICK: SGP_CACHE;
1255         unsigned long size;
1256
1257         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1258         if (pgoff >= size || pgoff + (len >> PAGE_SHIFT) > size)
1259                 return -EINVAL;
1260
1261         while ((long) len > 0) {
1262                 struct page *page = NULL;
1263                 int err;
1264                 /*
1265                  * Will need changing if PAGE_CACHE_SIZE != PAGE_SIZE
1266                  */
1267                 err = shmem_getpage(inode, pgoff, &page, sgp, NULL);
1268                 if (err)
1269                         return err;
1270                 /* Page may still be null, but only if nonblock was set. */
1271                 if (page) {
1272                         mark_page_accessed(page);
1273                         err = install_page(mm, vma, addr, page, prot);
1274                         if (err) {
1275                                 page_cache_release(page);
1276                                 return err;
1277                         }
1278                 } else if (vma->vm_flags & VM_NONLINEAR) {
1279                         /* No page was found just because we can't read it in
1280                          * now (being here implies nonblock != 0), but the page
1281                          * may exist, so set the PTE to fault it in later. */
1282                         err = install_file_pte(mm, vma, addr, pgoff, prot);
1283                         if (err)
1284                                 return err;
1285                 }
1286
1287                 len -= PAGE_SIZE;
1288                 addr += PAGE_SIZE;
1289                 pgoff++;
1290         }
1291         return 0;
1292 }
1293
1294 #ifdef CONFIG_NUMA
1295 int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1296 {
1297         struct inode *i = vma->vm_file->f_dentry->d_inode;
1298         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1299 }
1300
1301 struct mempolicy *
1302 shmem_get_policy(struct vm_area_struct *vma, unsigned long addr)
1303 {
1304         struct inode *i = vma->vm_file->f_dentry->d_inode;
1305         unsigned long idx;
1306
1307         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1308         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1309 }
1310 #endif
1311
1312 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1313 {
1314         struct inode *inode = file->f_dentry->d_inode;
1315         struct shmem_inode_info *info = SHMEM_I(inode);
1316         int retval = -ENOMEM;
1317
1318         spin_lock(&info->lock);
1319         if (lock && !(info->flags & VM_LOCKED)) {
1320                 if (!user_shm_lock(inode->i_size, user))
1321                         goto out_nomem;
1322                 info->flags |= VM_LOCKED;
1323         }
1324         if (!lock && (info->flags & VM_LOCKED) && user) {
1325                 user_shm_unlock(inode->i_size, user);
1326                 info->flags &= ~VM_LOCKED;
1327         }
1328         retval = 0;
1329 out_nomem:
1330         spin_unlock(&info->lock);
1331         return retval;
1332 }
1333
1334 int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1335 {
1336         file_accessed(file);
1337         vma->vm_ops = &shmem_vm_ops;
1338         return 0;
1339 }
1340
1341 static struct inode *
1342 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1343 {
1344         struct inode *inode;
1345         struct shmem_inode_info *info;
1346         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1347
1348         if (sbinfo->max_inodes) {
1349                 spin_lock(&sbinfo->stat_lock);
1350                 if (!sbinfo->free_inodes) {
1351                         spin_unlock(&sbinfo->stat_lock);
1352                         return NULL;
1353                 }
1354                 sbinfo->free_inodes--;
1355                 spin_unlock(&sbinfo->stat_lock);
1356         }
1357
1358         inode = new_inode(sb);
1359         if (inode) {
1360                 inode->i_mode = mode;
1361                 inode->i_uid = current->fsuid;
1362                 inode->i_gid = current->fsgid;
1363                 inode->i_blksize = PAGE_CACHE_SIZE;
1364                 inode->i_blocks = 0;
1365                 inode->i_mapping->a_ops = &shmem_aops;
1366                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1367                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1368                 info = SHMEM_I(inode);
1369                 memset(info, 0, (char *)inode - (char *)info);
1370                 spin_lock_init(&info->lock);
1371                 INIT_LIST_HEAD(&info->swaplist);
1372
1373                 switch (mode & S_IFMT) {
1374                 default:
1375                         init_special_inode(inode, mode, dev);
1376                         break;
1377                 case S_IFREG:
1378                         inode->i_op = &shmem_inode_operations;
1379                         inode->i_fop = &shmem_file_operations;
1380                         mpol_shared_policy_init(&info->policy, sbinfo->policy,
1381                                                         &sbinfo->policy_nodes);
1382                         break;
1383                 case S_IFDIR:
1384                         inode->i_nlink++;
1385                         /* Some things misbehave if size == 0 on a directory */
1386                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1387                         inode->i_op = &shmem_dir_inode_operations;
1388                         inode->i_fop = &simple_dir_operations;
1389                         break;
1390                 case S_IFLNK:
1391                         /*
1392                          * Must not load anything in the rbtree,
1393                          * mpol_free_shared_policy will not be called.
1394                          */
1395                         mpol_shared_policy_init(&info->policy, MPOL_DEFAULT,
1396                                                 NULL);
1397                         break;
1398                 }
1399         } else if (sbinfo->max_inodes) {
1400                 spin_lock(&sbinfo->stat_lock);
1401                 sbinfo->free_inodes++;
1402                 spin_unlock(&sbinfo->stat_lock);
1403         }
1404         return inode;
1405 }
1406
1407 #ifdef CONFIG_TMPFS
1408 static struct inode_operations shmem_symlink_inode_operations;
1409 static struct inode_operations shmem_symlink_inline_operations;
1410
1411 /*
1412  * Normally tmpfs makes no use of shmem_prepare_write, but it
1413  * lets a tmpfs file be used read-write below the loop driver.
1414  */
1415 static int
1416 shmem_prepare_write(struct file *file, struct page *page, unsigned offset, unsigned to)
1417 {
1418         struct inode *inode = page->mapping->host;
1419         return shmem_getpage(inode, page->index, &page, SGP_WRITE, NULL);
1420 }
1421
1422 static ssize_t
1423 shmem_file_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
1424 {
1425         struct inode    *inode = file->f_dentry->d_inode;
1426         loff_t          pos;
1427         unsigned long   written;
1428         ssize_t         err;
1429
1430         if ((ssize_t) count < 0)
1431                 return -EINVAL;
1432
1433         if (!access_ok(VERIFY_READ, buf, count))
1434                 return -EFAULT;
1435
1436         mutex_lock(&inode->i_mutex);
1437
1438         pos = *ppos;
1439         written = 0;
1440
1441         err = generic_write_checks(file, &pos, &count, 0);
1442         if (err || !count)
1443                 goto out;
1444
1445         err = remove_suid(file->f_dentry);
1446         if (err)
1447                 goto out;
1448
1449         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1450
1451         do {
1452                 struct page *page = NULL;
1453                 unsigned long bytes, index, offset;
1454                 char *kaddr;
1455                 int left;
1456
1457                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1458                 index = pos >> PAGE_CACHE_SHIFT;
1459                 bytes = PAGE_CACHE_SIZE - offset;
1460                 if (bytes > count)
1461                         bytes = count;
1462
1463                 /*
1464                  * We don't hold page lock across copy from user -
1465                  * what would it guard against? - so no deadlock here.
1466                  * But it still may be a good idea to prefault below.
1467                  */
1468
1469                 err = shmem_getpage(inode, index, &page, SGP_WRITE, NULL);
1470                 if (err)
1471                         break;
1472
1473                 left = bytes;
1474                 if (PageHighMem(page)) {
1475                         volatile unsigned char dummy;
1476                         __get_user(dummy, buf);
1477                         __get_user(dummy, buf + bytes - 1);
1478
1479                         kaddr = kmap_atomic(page, KM_USER0);
1480                         left = __copy_from_user_inatomic(kaddr + offset,
1481                                                         buf, bytes);
1482                         kunmap_atomic(kaddr, KM_USER0);
1483                 }
1484                 if (left) {
1485                         kaddr = kmap(page);
1486                         left = __copy_from_user(kaddr + offset, buf, bytes);
1487                         kunmap(page);
1488                 }
1489
1490                 written += bytes;
1491                 count -= bytes;
1492                 pos += bytes;
1493                 buf += bytes;
1494                 if (pos > inode->i_size)
1495                         i_size_write(inode, pos);
1496
1497                 flush_dcache_page(page);
1498                 set_page_dirty(page);
1499                 mark_page_accessed(page);
1500                 page_cache_release(page);
1501
1502                 if (left) {
1503                         pos -= left;
1504                         written -= left;
1505                         err = -EFAULT;
1506                         break;
1507                 }
1508
1509                 /*
1510                  * Our dirty pages are not counted in nr_dirty,
1511                  * and we do not attempt to balance dirty pages.
1512                  */
1513
1514                 cond_resched();
1515         } while (count);
1516
1517         *ppos = pos;
1518         if (written)
1519                 err = written;
1520 out:
1521         mutex_unlock(&inode->i_mutex);
1522         return err;
1523 }
1524
1525 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1526 {
1527         struct inode *inode = filp->f_dentry->d_inode;
1528         struct address_space *mapping = inode->i_mapping;
1529         unsigned long index, offset;
1530
1531         index = *ppos >> PAGE_CACHE_SHIFT;
1532         offset = *ppos & ~PAGE_CACHE_MASK;
1533
1534         for (;;) {
1535                 struct page *page = NULL;
1536                 unsigned long end_index, nr, ret;
1537                 loff_t i_size = i_size_read(inode);
1538
1539                 end_index = i_size >> PAGE_CACHE_SHIFT;
1540                 if (index > end_index)
1541                         break;
1542                 if (index == end_index) {
1543                         nr = i_size & ~PAGE_CACHE_MASK;
1544                         if (nr <= offset)
1545                                 break;
1546                 }
1547
1548                 desc->error = shmem_getpage(inode, index, &page, SGP_READ, NULL);
1549                 if (desc->error) {
1550                         if (desc->error == -EINVAL)
1551                                 desc->error = 0;
1552                         break;
1553                 }
1554
1555                 /*
1556                  * We must evaluate after, since reads (unlike writes)
1557                  * are called without i_mutex protection against truncate
1558                  */
1559                 nr = PAGE_CACHE_SIZE;
1560                 i_size = i_size_read(inode);
1561                 end_index = i_size >> PAGE_CACHE_SHIFT;
1562                 if (index == end_index) {
1563                         nr = i_size & ~PAGE_CACHE_MASK;
1564                         if (nr <= offset) {
1565                                 if (page)
1566                                         page_cache_release(page);
1567                                 break;
1568                         }
1569                 }
1570                 nr -= offset;
1571
1572                 if (page) {
1573                         /*
1574                          * If users can be writing to this page using arbitrary
1575                          * virtual addresses, take care about potential aliasing
1576                          * before reading the page on the kernel side.
1577                          */
1578                         if (mapping_writably_mapped(mapping))
1579                                 flush_dcache_page(page);
1580                         /*
1581                          * Mark the page accessed if we read the beginning.
1582                          */
1583                         if (!offset)
1584                                 mark_page_accessed(page);
1585                 } else {
1586                         page = ZERO_PAGE(0);
1587                         page_cache_get(page);
1588                 }
1589
1590                 /*
1591                  * Ok, we have the page, and it's up-to-date, so
1592                  * now we can copy it to user space...
1593                  *
1594                  * The actor routine returns how many bytes were actually used..
1595                  * NOTE! This may not be the same as how much of a user buffer
1596                  * we filled up (we may be padding etc), so we can only update
1597                  * "pos" here (the actor routine has to update the user buffer
1598                  * pointers and the remaining count).
1599                  */
1600                 ret = actor(desc, page, offset, nr);
1601                 offset += ret;
1602                 index += offset >> PAGE_CACHE_SHIFT;
1603                 offset &= ~PAGE_CACHE_MASK;
1604
1605                 page_cache_release(page);
1606                 if (ret != nr || !desc->count)
1607                         break;
1608
1609                 cond_resched();
1610         }
1611
1612         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1613         file_accessed(filp);
1614 }
1615
1616 static ssize_t shmem_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1617 {
1618         read_descriptor_t desc;
1619
1620         if ((ssize_t) count < 0)
1621                 return -EINVAL;
1622         if (!access_ok(VERIFY_WRITE, buf, count))
1623                 return -EFAULT;
1624         if (!count)
1625                 return 0;
1626
1627         desc.written = 0;
1628         desc.count = count;
1629         desc.arg.buf = buf;
1630         desc.error = 0;
1631
1632         do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1633         if (desc.written)
1634                 return desc.written;
1635         return desc.error;
1636 }
1637
1638 static ssize_t shmem_file_sendfile(struct file *in_file, loff_t *ppos,
1639                          size_t count, read_actor_t actor, void *target)
1640 {
1641         read_descriptor_t desc;
1642
1643         if (!count)
1644                 return 0;
1645
1646         desc.written = 0;
1647         desc.count = count;
1648         desc.arg.data = target;
1649         desc.error = 0;
1650
1651         do_shmem_file_read(in_file, ppos, &desc, actor);
1652         if (desc.written)
1653                 return desc.written;
1654         return desc.error;
1655 }
1656
1657 static int shmem_statfs(struct super_block *sb, struct kstatfs *buf)
1658 {
1659         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1660
1661         buf->f_type = TMPFS_MAGIC;
1662         buf->f_bsize = PAGE_CACHE_SIZE;
1663         buf->f_namelen = NAME_MAX;
1664         spin_lock(&sbinfo->stat_lock);
1665         if (sbinfo->max_blocks) {
1666                 buf->f_blocks = sbinfo->max_blocks;
1667                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1668         }
1669         if (sbinfo->max_inodes) {
1670                 buf->f_files = sbinfo->max_inodes;
1671                 buf->f_ffree = sbinfo->free_inodes;
1672         }
1673         /* else leave those fields 0 like simple_statfs */
1674         spin_unlock(&sbinfo->stat_lock);
1675         return 0;
1676 }
1677
1678 /*
1679  * File creation. Allocate an inode, and we're done..
1680  */
1681 static int
1682 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1683 {
1684         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1685         int error = -ENOSPC;
1686
1687         if (inode) {
1688                 error = security_inode_init_security(inode, dir, NULL, NULL,
1689                                                      NULL);
1690                 if (error) {
1691                         if (error != -EOPNOTSUPP) {
1692                                 iput(inode);
1693                                 return error;
1694                         }
1695                         error = 0;
1696                 }
1697                 if (dir->i_mode & S_ISGID) {
1698                         inode->i_gid = dir->i_gid;
1699                         if (S_ISDIR(mode))
1700                                 inode->i_mode |= S_ISGID;
1701                 }
1702                 dir->i_size += BOGO_DIRENT_SIZE;
1703                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1704                 d_instantiate(dentry, inode);
1705                 dget(dentry); /* Extra count - pin the dentry in core */
1706         }
1707         return error;
1708 }
1709
1710 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1711 {
1712         int error;
1713
1714         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1715                 return error;
1716         dir->i_nlink++;
1717         return 0;
1718 }
1719
1720 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1721                 struct nameidata *nd)
1722 {
1723         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1724 }
1725
1726 /*
1727  * Link a file..
1728  */
1729 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1730 {
1731         struct inode *inode = old_dentry->d_inode;
1732         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1733
1734         /*
1735          * No ordinary (disk based) filesystem counts links as inodes;
1736          * but each new link needs a new dentry, pinning lowmem, and
1737          * tmpfs dentries cannot be pruned until they are unlinked.
1738          */
1739         if (sbinfo->max_inodes) {
1740                 spin_lock(&sbinfo->stat_lock);
1741                 if (!sbinfo->free_inodes) {
1742                         spin_unlock(&sbinfo->stat_lock);
1743                         return -ENOSPC;
1744                 }
1745                 sbinfo->free_inodes--;
1746                 spin_unlock(&sbinfo->stat_lock);
1747         }
1748
1749         dir->i_size += BOGO_DIRENT_SIZE;
1750         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1751         inode->i_nlink++;
1752         atomic_inc(&inode->i_count);    /* New dentry reference */
1753         dget(dentry);           /* Extra pinning count for the created dentry */
1754         d_instantiate(dentry, inode);
1755         return 0;
1756 }
1757
1758 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1759 {
1760         struct inode *inode = dentry->d_inode;
1761
1762         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode)) {
1763                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1764                 if (sbinfo->max_inodes) {
1765                         spin_lock(&sbinfo->stat_lock);
1766                         sbinfo->free_inodes++;
1767                         spin_unlock(&sbinfo->stat_lock);
1768                 }
1769         }
1770
1771         dir->i_size -= BOGO_DIRENT_SIZE;
1772         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1773         inode->i_nlink--;
1774         dput(dentry);   /* Undo the count from "create" - this does all the work */
1775         return 0;
1776 }
1777
1778 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1779 {
1780         if (!simple_empty(dentry))
1781                 return -ENOTEMPTY;
1782
1783         dentry->d_inode->i_nlink--;
1784         dir->i_nlink--;
1785         return shmem_unlink(dir, dentry);
1786 }
1787
1788 /*
1789  * The VFS layer already does all the dentry stuff for rename,
1790  * we just have to decrement the usage count for the target if
1791  * it exists so that the VFS layer correctly free's it when it
1792  * gets overwritten.
1793  */
1794 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1795 {
1796         struct inode *inode = old_dentry->d_inode;
1797         int they_are_dirs = S_ISDIR(inode->i_mode);
1798
1799         if (!simple_empty(new_dentry))
1800                 return -ENOTEMPTY;
1801
1802         if (new_dentry->d_inode) {
1803                 (void) shmem_unlink(new_dir, new_dentry);
1804                 if (they_are_dirs)
1805                         old_dir->i_nlink--;
1806         } else if (they_are_dirs) {
1807                 old_dir->i_nlink--;
1808                 new_dir->i_nlink++;
1809         }
1810
1811         old_dir->i_size -= BOGO_DIRENT_SIZE;
1812         new_dir->i_size += BOGO_DIRENT_SIZE;
1813         old_dir->i_ctime = old_dir->i_mtime =
1814         new_dir->i_ctime = new_dir->i_mtime =
1815         inode->i_ctime = CURRENT_TIME;
1816         return 0;
1817 }
1818
1819 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1820 {
1821         int error;
1822         int len;
1823         struct inode *inode;
1824         struct page *page = NULL;
1825         char *kaddr;
1826         struct shmem_inode_info *info;
1827
1828         len = strlen(symname) + 1;
1829         if (len > PAGE_CACHE_SIZE)
1830                 return -ENAMETOOLONG;
1831
1832         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1833         if (!inode)
1834                 return -ENOSPC;
1835
1836         error = security_inode_init_security(inode, dir, NULL, NULL,
1837                                              NULL);
1838         if (error) {
1839                 if (error != -EOPNOTSUPP) {
1840                         iput(inode);
1841                         return error;
1842                 }
1843                 error = 0;
1844         }
1845
1846         info = SHMEM_I(inode);
1847         inode->i_size = len-1;
1848         if (len <= (char *)inode - (char *)info) {
1849                 /* do it inline */
1850                 memcpy(info, symname, len);
1851                 inode->i_op = &shmem_symlink_inline_operations;
1852         } else {
1853                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1854                 if (error) {
1855                         iput(inode);
1856                         return error;
1857                 }
1858                 inode->i_op = &shmem_symlink_inode_operations;
1859                 kaddr = kmap_atomic(page, KM_USER0);
1860                 memcpy(kaddr, symname, len);
1861                 kunmap_atomic(kaddr, KM_USER0);
1862                 set_page_dirty(page);
1863                 page_cache_release(page);
1864         }
1865         if (dir->i_mode & S_ISGID)
1866                 inode->i_gid = dir->i_gid;
1867         dir->i_size += BOGO_DIRENT_SIZE;
1868         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1869         d_instantiate(dentry, inode);
1870         dget(dentry);
1871         return 0;
1872 }
1873
1874 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1875 {
1876         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1877         return NULL;
1878 }
1879
1880 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1881 {
1882         struct page *page = NULL;
1883         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1884         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1885         return page;
1886 }
1887
1888 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1889 {
1890         if (!IS_ERR(nd_get_link(nd))) {
1891                 struct page *page = cookie;
1892                 kunmap(page);
1893                 mark_page_accessed(page);
1894                 page_cache_release(page);
1895         }
1896 }
1897
1898 static struct inode_operations shmem_symlink_inline_operations = {
1899         .readlink       = generic_readlink,
1900         .follow_link    = shmem_follow_link_inline,
1901 };
1902
1903 static struct inode_operations shmem_symlink_inode_operations = {
1904         .truncate       = shmem_truncate,
1905         .readlink       = generic_readlink,
1906         .follow_link    = shmem_follow_link,
1907         .put_link       = shmem_put_link,
1908 };
1909
1910 static int shmem_parse_options(char *options, int *mode, uid_t *uid,
1911         gid_t *gid, unsigned long *blocks, unsigned long *inodes,
1912         int *policy, nodemask_t *policy_nodes)
1913 {
1914         char *this_char, *value, *rest;
1915
1916         while (options != NULL) {
1917                 this_char = options;
1918                 for (;;) {
1919                         /*
1920                          * NUL-terminate this option: unfortunately,
1921                          * mount options form a comma-separated list,
1922                          * but mpol's nodelist may also contain commas.
1923                          */
1924                         options = strchr(options, ',');
1925                         if (options == NULL)
1926                                 break;
1927                         options++;
1928                         if (!isdigit(*options)) {
1929                                 options[-1] = '\0';
1930                                 break;
1931                         }
1932                 }
1933                 if (!*this_char)
1934                         continue;
1935                 if ((value = strchr(this_char,'=')) != NULL) {
1936                         *value++ = 0;
1937                 } else {
1938                         printk(KERN_ERR
1939                             "tmpfs: No value for mount option '%s'\n",
1940                             this_char);
1941                         return 1;
1942                 }
1943
1944                 if (!strcmp(this_char,"size")) {
1945                         unsigned long long size;
1946                         size = memparse(value,&rest);
1947                         if (*rest == '%') {
1948                                 size <<= PAGE_SHIFT;
1949                                 size *= totalram_pages;
1950                                 do_div(size, 100);
1951                                 rest++;
1952                         }
1953                         if (*rest)
1954                                 goto bad_val;
1955                         *blocks = size >> PAGE_CACHE_SHIFT;
1956                 } else if (!strcmp(this_char,"nr_blocks")) {
1957                         *blocks = memparse(value,&rest);
1958                         if (*rest)
1959                                 goto bad_val;
1960                 } else if (!strcmp(this_char,"nr_inodes")) {
1961                         *inodes = memparse(value,&rest);
1962                         if (*rest)
1963                                 goto bad_val;
1964                 } else if (!strcmp(this_char,"mode")) {
1965                         if (!mode)
1966                                 continue;
1967                         *mode = simple_strtoul(value,&rest,8);
1968                         if (*rest)
1969                                 goto bad_val;
1970                 } else if (!strcmp(this_char,"uid")) {
1971                         if (!uid)
1972                                 continue;
1973                         *uid = simple_strtoul(value,&rest,0);
1974                         if (*rest)
1975                                 goto bad_val;
1976                 } else if (!strcmp(this_char,"gid")) {
1977                         if (!gid)
1978                                 continue;
1979                         *gid = simple_strtoul(value,&rest,0);
1980                         if (*rest)
1981                                 goto bad_val;
1982                 } else if (!strcmp(this_char,"mpol")) {
1983                         if (shmem_parse_mpol(value,policy,policy_nodes))
1984                                 goto bad_val;
1985                 } else {
1986                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
1987                                this_char);
1988                         return 1;
1989                 }
1990         }
1991         return 0;
1992
1993 bad_val:
1994         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
1995                value, this_char);
1996         return 1;
1997
1998 }
1999
2000 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2001 {
2002         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2003         unsigned long max_blocks = sbinfo->max_blocks;
2004         unsigned long max_inodes = sbinfo->max_inodes;
2005         int policy = sbinfo->policy;
2006         nodemask_t policy_nodes = sbinfo->policy_nodes;
2007         unsigned long blocks;
2008         unsigned long inodes;
2009         int error = -EINVAL;
2010
2011         if (shmem_parse_options(data, NULL, NULL, NULL, &max_blocks,
2012                                 &max_inodes, &policy, &policy_nodes))
2013                 return error;
2014
2015         spin_lock(&sbinfo->stat_lock);
2016         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2017         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2018         if (max_blocks < blocks)
2019                 goto out;
2020         if (max_inodes < inodes)
2021                 goto out;
2022         /*
2023          * Those tests also disallow limited->unlimited while any are in
2024          * use, so i_blocks will always be zero when max_blocks is zero;
2025          * but we must separately disallow unlimited->limited, because
2026          * in that case we have no record of how much is already in use.
2027          */
2028         if (max_blocks && !sbinfo->max_blocks)
2029                 goto out;
2030         if (max_inodes && !sbinfo->max_inodes)
2031                 goto out;
2032
2033         error = 0;
2034         sbinfo->max_blocks  = max_blocks;
2035         sbinfo->free_blocks = max_blocks - blocks;
2036         sbinfo->max_inodes  = max_inodes;
2037         sbinfo->free_inodes = max_inodes - inodes;
2038         sbinfo->policy = policy;
2039         sbinfo->policy_nodes = policy_nodes;
2040 out:
2041         spin_unlock(&sbinfo->stat_lock);
2042         return error;
2043 }
2044 #endif
2045
2046 static void shmem_put_super(struct super_block *sb)
2047 {
2048         kfree(sb->s_fs_info);
2049         sb->s_fs_info = NULL;
2050 }
2051
2052 static int shmem_fill_super(struct super_block *sb,
2053                             void *data, int silent)
2054 {
2055         struct inode *inode;
2056         struct dentry *root;
2057         int mode   = S_IRWXUGO | S_ISVTX;
2058         uid_t uid = current->fsuid;
2059         gid_t gid = current->fsgid;
2060         int err = -ENOMEM;
2061         struct shmem_sb_info *sbinfo;
2062         unsigned long blocks = 0;
2063         unsigned long inodes = 0;
2064         int policy = MPOL_DEFAULT;
2065         nodemask_t policy_nodes = node_online_map;
2066
2067 #ifdef CONFIG_TMPFS
2068         /*
2069          * Per default we only allow half of the physical ram per
2070          * tmpfs instance, limiting inodes to one per page of lowmem;
2071          * but the internal instance is left unlimited.
2072          */
2073         if (!(sb->s_flags & MS_NOUSER)) {
2074                 blocks = totalram_pages / 2;
2075                 inodes = totalram_pages - totalhigh_pages;
2076                 if (inodes > blocks)
2077                         inodes = blocks;
2078                 if (shmem_parse_options(data, &mode, &uid, &gid, &blocks,
2079                                         &inodes, &policy, &policy_nodes))
2080                         return -EINVAL;
2081         }
2082 #else
2083         sb->s_flags |= MS_NOUSER;
2084 #endif
2085
2086         /* Round up to L1_CACHE_BYTES to resist false sharing */
2087         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2088                                 L1_CACHE_BYTES), GFP_KERNEL);
2089         if (!sbinfo)
2090                 return -ENOMEM;
2091
2092         spin_lock_init(&sbinfo->stat_lock);
2093         sbinfo->max_blocks = blocks;
2094         sbinfo->free_blocks = blocks;
2095         sbinfo->max_inodes = inodes;
2096         sbinfo->free_inodes = inodes;
2097         sbinfo->policy = policy;
2098         sbinfo->policy_nodes = policy_nodes;
2099
2100         sb->s_fs_info = sbinfo;
2101         sb->s_maxbytes = SHMEM_MAX_BYTES;
2102         sb->s_blocksize = PAGE_CACHE_SIZE;
2103         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2104         sb->s_magic = TMPFS_MAGIC;
2105         sb->s_op = &shmem_ops;
2106         sb->s_time_gran = 1;
2107
2108         inode = shmem_get_inode(sb, S_IFDIR | mode, 0);
2109         if (!inode)
2110                 goto failed;
2111         inode->i_uid = uid;
2112         inode->i_gid = gid;
2113         root = d_alloc_root(inode);
2114         if (!root)
2115                 goto failed_iput;
2116         sb->s_root = root;
2117         return 0;
2118
2119 failed_iput:
2120         iput(inode);
2121 failed:
2122         shmem_put_super(sb);
2123         return err;
2124 }
2125
2126 static struct kmem_cache *shmem_inode_cachep;
2127
2128 static struct inode *shmem_alloc_inode(struct super_block *sb)
2129 {
2130         struct shmem_inode_info *p;
2131         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, SLAB_KERNEL);
2132         if (!p)
2133                 return NULL;
2134         return &p->vfs_inode;
2135 }
2136
2137 static void shmem_destroy_inode(struct inode *inode)
2138 {
2139         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2140                 /* only struct inode is valid if it's an inline symlink */
2141                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2142         }
2143         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2144 }
2145
2146 static void init_once(void *foo, struct kmem_cache *cachep,
2147                       unsigned long flags)
2148 {
2149         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2150
2151         if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
2152             SLAB_CTOR_CONSTRUCTOR) {
2153                 inode_init_once(&p->vfs_inode);
2154         }
2155 }
2156
2157 static int init_inodecache(void)
2158 {
2159         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2160                                 sizeof(struct shmem_inode_info),
2161                                 0, 0, init_once, NULL);
2162         if (shmem_inode_cachep == NULL)
2163                 return -ENOMEM;
2164         return 0;
2165 }
2166
2167 static void destroy_inodecache(void)
2168 {
2169         if (kmem_cache_destroy(shmem_inode_cachep))
2170                 printk(KERN_INFO "shmem_inode_cache: not all structures were freed\n");
2171 }
2172
2173 static struct address_space_operations shmem_aops = {
2174         .writepage      = shmem_writepage,
2175         .set_page_dirty = __set_page_dirty_nobuffers,
2176 #ifdef CONFIG_TMPFS
2177         .prepare_write  = shmem_prepare_write,
2178         .commit_write   = simple_commit_write,
2179 #endif
2180         .migratepage    = migrate_page,
2181 };
2182
2183 static struct file_operations shmem_file_operations = {
2184         .mmap           = shmem_mmap,
2185 #ifdef CONFIG_TMPFS
2186         .llseek         = generic_file_llseek,
2187         .read           = shmem_file_read,
2188         .write          = shmem_file_write,
2189         .fsync          = simple_sync_file,
2190         .sendfile       = shmem_file_sendfile,
2191 #endif
2192 };
2193
2194 static struct inode_operations shmem_inode_operations = {
2195         .truncate       = shmem_truncate,
2196         .setattr        = shmem_notify_change,
2197         .truncate_range = shmem_truncate_range,
2198 };
2199
2200 static struct inode_operations shmem_dir_inode_operations = {
2201 #ifdef CONFIG_TMPFS
2202         .create         = shmem_create,
2203         .lookup         = simple_lookup,
2204         .link           = shmem_link,
2205         .unlink         = shmem_unlink,
2206         .symlink        = shmem_symlink,
2207         .mkdir          = shmem_mkdir,
2208         .rmdir          = shmem_rmdir,
2209         .mknod          = shmem_mknod,
2210         .rename         = shmem_rename,
2211 #endif
2212 };
2213
2214 static struct super_operations shmem_ops = {
2215         .alloc_inode    = shmem_alloc_inode,
2216         .destroy_inode  = shmem_destroy_inode,
2217 #ifdef CONFIG_TMPFS
2218         .statfs         = shmem_statfs,
2219         .remount_fs     = shmem_remount_fs,
2220 #endif
2221         .delete_inode   = shmem_delete_inode,
2222         .drop_inode     = generic_delete_inode,
2223         .put_super      = shmem_put_super,
2224 };
2225
2226 static struct vm_operations_struct shmem_vm_ops = {
2227         .nopage         = shmem_nopage,
2228         .populate       = shmem_populate,
2229 #ifdef CONFIG_NUMA
2230         .set_policy     = shmem_set_policy,
2231         .get_policy     = shmem_get_policy,
2232 #endif
2233 };
2234
2235
2236 static struct super_block *shmem_get_sb(struct file_system_type *fs_type,
2237         int flags, const char *dev_name, void *data)
2238 {
2239         return get_sb_nodev(fs_type, flags, data, shmem_fill_super);
2240 }
2241
2242 static struct file_system_type tmpfs_fs_type = {
2243         .owner          = THIS_MODULE,
2244         .name           = "tmpfs",
2245         .get_sb         = shmem_get_sb,
2246         .kill_sb        = kill_litter_super,
2247 };
2248 static struct vfsmount *shm_mnt;
2249
2250 static int __init init_tmpfs(void)
2251 {
2252         int error;
2253
2254         error = init_inodecache();
2255         if (error)
2256                 goto out3;
2257
2258         error = register_filesystem(&tmpfs_fs_type);
2259         if (error) {
2260                 printk(KERN_ERR "Could not register tmpfs\n");
2261                 goto out2;
2262         }
2263 #ifdef CONFIG_TMPFS
2264         devfs_mk_dir("shm");
2265 #endif
2266         shm_mnt = do_kern_mount(tmpfs_fs_type.name, MS_NOUSER,
2267                                 tmpfs_fs_type.name, NULL);
2268         if (IS_ERR(shm_mnt)) {
2269                 error = PTR_ERR(shm_mnt);
2270                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2271                 goto out1;
2272         }
2273         return 0;
2274
2275 out1:
2276         unregister_filesystem(&tmpfs_fs_type);
2277 out2:
2278         destroy_inodecache();
2279 out3:
2280         shm_mnt = ERR_PTR(error);
2281         return error;
2282 }
2283 module_init(init_tmpfs)
2284
2285 /*
2286  * shmem_file_setup - get an unlinked file living in tmpfs
2287  *
2288  * @name: name for dentry (to be seen in /proc/<pid>/maps
2289  * @size: size to be set for the file
2290  *
2291  */
2292 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2293 {
2294         int error;
2295         struct file *file;
2296         struct inode *inode;
2297         struct dentry *dentry, *root;
2298         struct qstr this;
2299
2300         if (IS_ERR(shm_mnt))
2301                 return (void *)shm_mnt;
2302
2303         if (size < 0 || size > SHMEM_MAX_BYTES)
2304                 return ERR_PTR(-EINVAL);
2305
2306         if (shmem_acct_size(flags, size))
2307                 return ERR_PTR(-ENOMEM);
2308
2309         error = -ENOMEM;
2310         this.name = name;
2311         this.len = strlen(name);
2312         this.hash = 0; /* will go */
2313         root = shm_mnt->mnt_root;
2314         dentry = d_alloc(root, &this);
2315         if (!dentry)
2316                 goto put_memory;
2317
2318         error = -ENFILE;
2319         file = get_empty_filp();
2320         if (!file)
2321                 goto put_dentry;
2322
2323         error = -ENOSPC;
2324         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2325         if (!inode)
2326                 goto close_file;
2327
2328         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2329         d_instantiate(dentry, inode);
2330         inode->i_size = size;
2331         inode->i_nlink = 0;     /* It is unlinked */
2332         file->f_vfsmnt = mntget(shm_mnt);
2333         file->f_dentry = dentry;
2334         file->f_mapping = inode->i_mapping;
2335         file->f_op = &shmem_file_operations;
2336         file->f_mode = FMODE_WRITE | FMODE_READ;
2337         return file;
2338
2339 close_file:
2340         put_filp(file);
2341 put_dentry:
2342         dput(dentry);
2343 put_memory:
2344         shmem_unacct_size(flags, size);
2345         return ERR_PTR(error);
2346 }
2347
2348 /*
2349  * shmem_zero_setup - setup a shared anonymous mapping
2350  *
2351  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2352  */
2353 int shmem_zero_setup(struct vm_area_struct *vma)
2354 {
2355         struct file *file;
2356         loff_t size = vma->vm_end - vma->vm_start;
2357
2358         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2359         if (IS_ERR(file))
2360                 return PTR_ERR(file);
2361
2362         if (vma->vm_file)
2363                 fput(vma->vm_file);
2364         vma->vm_file = file;
2365         vma->vm_ops = &shmem_vm_ops;
2366         return 0;
2367 }