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