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/ramfs.h>
  29 #include <linux/pagemap.h>
  30 #include <linux/file.h>
  31 #include <linux/mm.h>
  32 #include <linux/export.h>
  33 #include <linux/swap.h>
  34 #include <linux/uio.h>
  35
  36 static struct vfsmount *shm_mnt;
  37
  38 #ifdef CONFIG_SHMEM
  39 /*
  40  * This virtual memory filesystem is heavily based on the ramfs. It
  41  * extends ramfs by the ability to use swap and honor resource limits
  42  * which makes it a completely usable filesystem.
  43  */
  44
  45 #include <linux/xattr.h>
  46 #include <linux/exportfs.h>
  47 #include <linux/posix_acl.h>
  48 #include <linux/posix_acl_xattr.h>
  49 #include <linux/mman.h>
  50 #include <linux/string.h>
  51 #include <linux/slab.h>
  52 #include <linux/backing-dev.h>
  53 #include <linux/shmem_fs.h>
  54 #include <linux/writeback.h>
  55 #include <linux/blkdev.h>
  56 #include <linux/pagevec.h>
  57 #include <linux/percpu_counter.h>
  58 #include <linux/falloc.h>
  59 #include <linux/splice.h>
  60 #include <linux/security.h>
  61 #include <linux/swapops.h>
  62 #include <linux/mempolicy.h>
  63 #include <linux/namei.h>
  64 #include <linux/ctype.h>
  65 #include <linux/migrate.h>
  66 #include <linux/highmem.h>
  67 #include <linux/seq_file.h>
  68 #include <linux/magic.h>
  69 #include <linux/syscalls.h>
  70 #include <linux/fcntl.h>
  71 #include <uapi/linux/memfd.h>
  72
  73 #include <asm/uaccess.h>
  74 #include <asm/pgtable.h>
  75
  76 #include "internal.h"
  77
  78 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  79 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  80
  81 /* Pretend that each entry is of this size in directory's i_size */
  82 #define BOGO_DIRENT_SIZE 20
  83
  84 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  85 #define SHORT_SYMLINK_LEN 128
  86
  87 /*
  88  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
  89  * inode->i_private (with i_mutex making sure that it has only one user at
  90  * a time): we would prefer not to enlarge the shmem inode just for that.
  91  */
  92 struct shmem_falloc {
  93         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
  94         pgoff_t start;          /* start of range currently being fallocated */
  95         pgoff_t next;           /* the next page offset to be fallocated */
  96         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
  97         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
  98 };
  99
 100 /* Flag allocation requirements to shmem_getpage */
 101 enum sgp_type {
 102         SGP_READ,       /* don't exceed i_size, don't allocate page */
 103         SGP_CACHE,      /* don't exceed i_size, may allocate page */
 104         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
 105         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
 106         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
 107 };
 108
 109 #ifdef CONFIG_TMPFS
 110 static unsigned long shmem_default_max_blocks(void)
 111 {
 112         return totalram_pages / 2;
 113 }
 114
 115 static unsigned long shmem_default_max_inodes(void)
 116 {
 117         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 118 }
 119 #endif
 120
 121 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
 122 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 123                                 struct shmem_inode_info *info, pgoff_t index);
 124 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 125         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
 126
 127 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
 128         struct page **pagep, enum sgp_type sgp, int *fault_type)
 129 {
 130         return shmem_getpage_gfp(inode, index, pagep, sgp,
 131                         mapping_gfp_mask(inode->i_mapping), fault_type);
 132 }
 133
 134 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 135 {
 136         return sb->s_fs_info;
 137 }
 138
 139 /*
 140  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 141  * for shared memory and for shared anonymous (/dev/zero) mappings
 142  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 143  * consistent with the pre-accounting of private mappings ...
 144  */
 145 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 146 {
 147         return (flags & VM_NORESERVE) ?
 148                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 149 }
 150
 151 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 152 {
 153         if (!(flags & VM_NORESERVE))
 154                 vm_unacct_memory(VM_ACCT(size));
 155 }
 156
 157 static inline int shmem_reacct_size(unsigned long flags,
 158                 loff_t oldsize, loff_t newsize)
 159 {
 160         if (!(flags & VM_NORESERVE)) {
 161                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
 162                         return security_vm_enough_memory_mm(current->mm,
 163                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
 164                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
 165                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
 166         }
 167         return 0;
 168 }
 169
 170 /*
 171  * ... whereas tmpfs objects are accounted incrementally as
 172  * pages are allocated, in order to allow huge sparse files.
 173  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 174  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 175  */
 176 static inline int shmem_acct_block(unsigned long flags)
 177 {
 178         return (flags & VM_NORESERVE) ?
 179                 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 180 }
 181
 182 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 183 {
 184         if (flags & VM_NORESERVE)
 185                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 186 }
 187
 188 static const struct super_operations shmem_ops;
 189 static const struct address_space_operations shmem_aops;
 190 static const struct file_operations shmem_file_operations;
 191 static const struct inode_operations shmem_inode_operations;
 192 static const struct inode_operations shmem_dir_inode_operations;
 193 static const struct inode_operations shmem_special_inode_operations;
 194 static const struct vm_operations_struct shmem_vm_ops;
 195
 196 static LIST_HEAD(shmem_swaplist);
 197 static DEFINE_MUTEX(shmem_swaplist_mutex);
 198
 199 static int shmem_reserve_inode(struct super_block *sb)
 200 {
 201         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 202         if (sbinfo->max_inodes) {
 203                 spin_lock(&sbinfo->stat_lock);
 204                 if (!sbinfo->free_inodes) {
 205                         spin_unlock(&sbinfo->stat_lock);
 206                         return -ENOSPC;
 207                 }
 208                 sbinfo->free_inodes--;
 209                 spin_unlock(&sbinfo->stat_lock);
 210         }
 211         return 0;
 212 }
 213
 214 static void shmem_free_inode(struct super_block *sb)
 215 {
 216         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 217         if (sbinfo->max_inodes) {
 218                 spin_lock(&sbinfo->stat_lock);
 219                 sbinfo->free_inodes++;
 220                 spin_unlock(&sbinfo->stat_lock);
 221         }
 222 }
 223
 224 /**
 225  * shmem_recalc_inode - recalculate the block usage of an inode
 226  * @inode: inode to recalc
 227  *
 228  * We have to calculate the free blocks since the mm can drop
 229  * undirtied hole pages behind our back.
 230  *
 231  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 232  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 233  *
 234  * It has to be called with the spinlock held.
 235  */
 236 static void shmem_recalc_inode(struct inode *inode)
 237 {
 238         struct shmem_inode_info *info = SHMEM_I(inode);
 239         long freed;
 240
 241         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 242         if (freed > 0) {
 243                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 244                 if (sbinfo->max_blocks)
 245                         percpu_counter_add(&sbinfo->used_blocks, -freed);
 246                 info->alloced -= freed;
 247                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 248                 shmem_unacct_blocks(info->flags, freed);
 249         }
 250 }
 251
 252 /*
 253  * Replace item expected in radix tree by a new item, while holding tree lock.
 254  */
 255 static int shmem_radix_tree_replace(struct address_space *mapping,
 256                         pgoff_t index, void *expected, void *replacement)
 257 {
 258         void **pslot;
 259         void *item;
 260
 261         VM_BUG_ON(!expected);
 262         VM_BUG_ON(!replacement);
 263         pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
 264         if (!pslot)
 265                 return -ENOENT;
 266         item = radix_tree_deref_slot_protected(pslot, &mapping->tree_lock);
 267         if (item != expected)
 268                 return -ENOENT;
 269         radix_tree_replace_slot(pslot, replacement);
 270         return 0;
 271 }
 272
 273 /*
 274  * Sometimes, before we decide whether to proceed or to fail, we must check
 275  * that an entry was not already brought back from swap by a racing thread.
 276  *
 277  * Checking page is not enough: by the time a SwapCache page is locked, it
 278  * might be reused, and again be SwapCache, using the same swap as before.
 279  */
 280 static bool shmem_confirm_swap(struct address_space *mapping,
 281                                pgoff_t index, swp_entry_t swap)
 282 {
 283         void *item;
 284
 285         rcu_read_lock();
 286         item = radix_tree_lookup(&mapping->page_tree, index);
 287         rcu_read_unlock();
 288         return item == swp_to_radix_entry(swap);
 289 }
 290
 291 /*
 292  * Like add_to_page_cache_locked, but error if expected item has gone.
 293  */
 294 static int shmem_add_to_page_cache(struct page *page,
 295                                    struct address_space *mapping,
 296                                    pgoff_t index, void *expected)
 297 {
 298         int error;
 299
 300         VM_BUG_ON_PAGE(!PageLocked(page), page);
 301         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 302
 303         page_cache_get(page);
 304         page->mapping = mapping;
 305         page->index = index;
 306
 307         spin_lock_irq(&mapping->tree_lock);
 308         if (!expected)
 309                 error = radix_tree_insert(&mapping->page_tree, index, page);
 310         else
 311                 error = shmem_radix_tree_replace(mapping, index, expected,
 312                                                                  page);
 313         if (!error) {
 314                 mapping->nrpages++;
 315                 __inc_zone_page_state(page, NR_FILE_PAGES);
 316                 __inc_zone_page_state(page, NR_SHMEM);
 317                 spin_unlock_irq(&mapping->tree_lock);
 318         } else {
 319                 page->mapping = NULL;
 320                 spin_unlock_irq(&mapping->tree_lock);
 321                 page_cache_release(page);
 322         }
 323         return error;
 324 }
 325
 326 /*
 327  * Like delete_from_page_cache, but substitutes swap for page.
 328  */
 329 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 330 {
 331         struct address_space *mapping = page->mapping;
 332         int error;
 333
 334         spin_lock_irq(&mapping->tree_lock);
 335         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
 336         page->mapping = NULL;
 337         mapping->nrpages--;
 338         __dec_zone_page_state(page, NR_FILE_PAGES);
 339         __dec_zone_page_state(page, NR_SHMEM);
 340         spin_unlock_irq(&mapping->tree_lock);
 341         page_cache_release(page);
 342         BUG_ON(error);
 343 }
 344
 345 /*
 346  * Remove swap entry from radix tree, free the swap and its page cache.
 347  */
 348 static int shmem_free_swap(struct address_space *mapping,
 349                            pgoff_t index, void *radswap)
 350 {
 351         void *old;
 352
 353         spin_lock_irq(&mapping->tree_lock);
 354         old = radix_tree_delete_item(&mapping->page_tree, index, radswap);
 355         spin_unlock_irq(&mapping->tree_lock);
 356         if (old != radswap)
 357                 return -ENOENT;
 358         free_swap_and_cache(radix_to_swp_entry(radswap));
 359         return 0;
 360 }
 361
 362 /*
 363  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 364  */
 365 void shmem_unlock_mapping(struct address_space *mapping)
 366 {
 367         struct pagevec pvec;
 368         pgoff_t indices[PAGEVEC_SIZE];
 369         pgoff_t index = 0;
 370
 371         pagevec_init(&pvec, 0);
 372         /*
 373          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 374          */
 375         while (!mapping_unevictable(mapping)) {
 376                 /*
 377                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
 378                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
 379                  */
 380                 pvec.nr = find_get_entries(mapping, index,
 381                                            PAGEVEC_SIZE, pvec.pages, indices);
 382                 if (!pvec.nr)
 383                         break;
 384                 index = indices[pvec.nr - 1] + 1;
 385                 pagevec_remove_exceptionals(&pvec);
 386                 check_move_unevictable_pages(pvec.pages, pvec.nr);
 387                 pagevec_release(&pvec);
 388                 cond_resched();
 389         }
 390 }
 391
 392 /*
 393  * Remove range of pages and swap entries from radix tree, and free them.
 394  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 395  */
 396 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 397                                                                  bool unfalloc)
 398 {
 399         struct address_space *mapping = inode->i_mapping;
 400         struct shmem_inode_info *info = SHMEM_I(inode);
 401         pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 402         pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
 403         unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
 404         unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
 405         struct pagevec pvec;
 406         pgoff_t indices[PAGEVEC_SIZE];
 407         long nr_swaps_freed = 0;
 408         pgoff_t index;
 409         int i;
 410
 411         if (lend == -1)
 412                 end = -1;       /* unsigned, so actually very big */
 413
 414         pagevec_init(&pvec, 0);
 415         index = start;
 416         while (index < end) {
 417                 pvec.nr = find_get_entries(mapping, index,
 418                         min(end - index, (pgoff_t)PAGEVEC_SIZE),
 419                         pvec.pages, indices);
 420                 if (!pvec.nr)
 421                         break;
 422                 for (i = 0; i < pagevec_count(&pvec); i++) {
 423                         struct page *page = pvec.pages[i];
 424
 425                         index = indices[i];
 426                         if (index >= end)
 427                                 break;
 428
 429                         if (radix_tree_exceptional_entry(page)) {
 430                                 if (unfalloc)
 431                                         continue;
 432                                 nr_swaps_freed += !shmem_free_swap(mapping,
 433                                                                 index, page);
 434                                 continue;
 435                         }
 436
 437                         if (!trylock_page(page))
 438                                 continue;
 439                         if (!unfalloc || !PageUptodate(page)) {
 440                                 if (page->mapping == mapping) {
 441                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
 442                                         truncate_inode_page(mapping, page);
 443                                 }
 444                         }
 445                         unlock_page(page);
 446                 }
 447                 pagevec_remove_exceptionals(&pvec);
 448                 pagevec_release(&pvec);
 449                 cond_resched();
 450                 index++;
 451         }
 452
 453         if (partial_start) {
 454                 struct page *page = NULL;
 455                 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
 456                 if (page) {
 457                         unsigned int top = PAGE_CACHE_SIZE;
 458                         if (start > end) {
 459                                 top = partial_end;
 460                                 partial_end = 0;
 461                         }
 462                         zero_user_segment(page, partial_start, top);
 463                         set_page_dirty(page);
 464                         unlock_page(page);
 465                         page_cache_release(page);
 466                 }
 467         }
 468         if (partial_end) {
 469                 struct page *page = NULL;
 470                 shmem_getpage(inode, end, &page, SGP_READ, NULL);
 471                 if (page) {
 472                         zero_user_segment(page, 0, partial_end);
 473                         set_page_dirty(page);
 474                         unlock_page(page);
 475                         page_cache_release(page);
 476                 }
 477         }
 478         if (start >= end)
 479                 return;
 480
 481         index = start;
 482         while (index < end) {
 483                 cond_resched();
 484
 485                 pvec.nr = find_get_entries(mapping, index,
 486                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 487                                 pvec.pages, indices);
 488                 if (!pvec.nr) {
 489                         /* If all gone or hole-punch or unfalloc, we're done */
 490                         if (index == start || end != -1)
 491                                 break;
 492                         /* But if truncating, restart to make sure all gone */
 493                         index = start;
 494                         continue;
 495                 }
 496                 for (i = 0; i < pagevec_count(&pvec); i++) {
 497                         struct page *page = pvec.pages[i];
 498
 499                         index = indices[i];
 500                         if (index >= end)
 501                                 break;
 502
 503                         if (radix_tree_exceptional_entry(page)) {
 504                                 if (unfalloc)
 505                                         continue;
 506                                 if (shmem_free_swap(mapping, index, page)) {
 507                                         /* Swap was replaced by page: retry */
 508                                         index--;
 509                                         break;
 510                                 }
 511                                 nr_swaps_freed++;
 512                                 continue;
 513                         }
 514
 515                         lock_page(page);
 516                         if (!unfalloc || !PageUptodate(page)) {
 517                                 if (page->mapping == mapping) {
 518                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
 519                                         truncate_inode_page(mapping, page);
 520                                 } else {
 521                                         /* Page was replaced by swap: retry */
 522                                         unlock_page(page);
 523                                         index--;
 524                                         break;
 525                                 }
 526                         }
 527                         unlock_page(page);
 528                 }
 529                 pagevec_remove_exceptionals(&pvec);
 530                 pagevec_release(&pvec);
 531                 index++;
 532         }
 533
 534         spin_lock(&info->lock);
 535         info->swapped -= nr_swaps_freed;
 536         shmem_recalc_inode(inode);
 537         spin_unlock(&info->lock);
 538 }
 539
 540 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
 541 {
 542         shmem_undo_range(inode, lstart, lend, false);
 543         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 544 }
 545 EXPORT_SYMBOL_GPL(shmem_truncate_range);
 546
 547 static int shmem_getattr(struct vfsmount *mnt, struct dentry *dentry,
 548                          struct kstat *stat)
 549 {
 550         struct inode *inode = dentry->d_inode;
 551         struct shmem_inode_info *info = SHMEM_I(inode);
 552
 553         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
 554                 spin_lock(&info->lock);
 555                 shmem_recalc_inode(inode);
 556                 spin_unlock(&info->lock);
 557         }
 558         generic_fillattr(inode, stat);
 559         return 0;
 560 }
 561
 562 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
 563 {
 564         struct inode *inode = d_inode(dentry);
 565         struct shmem_inode_info *info = SHMEM_I(inode);
 566         int error;
 567
 568         error = inode_change_ok(inode, attr);
 569         if (error)
 570                 return error;
 571
 572         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
 573                 loff_t oldsize = inode->i_size;
 574                 loff_t newsize = attr->ia_size;
 575
 576                 /* protected by i_mutex */
 577                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 578                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 579                         return -EPERM;
 580
 581                 if (newsize != oldsize) {
 582                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
 583                                         oldsize, newsize);
 584                         if (error)
 585                                 return error;
 586                         i_size_write(inode, newsize);
 587                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 588                 }
 589                 if (newsize <= oldsize) {
 590                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
 591                         if (oldsize > holebegin)
 592                                 unmap_mapping_range(inode->i_mapping,
 593                                                         holebegin, 0, 1);
 594                         if (info->alloced)
 595                                 shmem_truncate_range(inode,
 596                                                         newsize, (loff_t)-1);
 597                         /* unmap again to remove racily COWed private pages */
 598                         if (oldsize > holebegin)
 599                                 unmap_mapping_range(inode->i_mapping,
 600                                                         holebegin, 0, 1);
 601                 }
 602         }
 603
 604         setattr_copy(inode, attr);
 605         if (attr->ia_valid & ATTR_MODE)
 606                 error = posix_acl_chmod(inode, inode->i_mode);
 607         return error;
 608 }
 609
 610 static void shmem_evict_inode(struct inode *inode)
 611 {
 612         struct shmem_inode_info *info = SHMEM_I(inode);
 613
 614         if (inode->i_mapping->a_ops == &shmem_aops) {
 615                 shmem_unacct_size(info->flags, inode->i_size);
 616                 inode->i_size = 0;
 617                 shmem_truncate_range(inode, 0, (loff_t)-1);
 618                 if (!list_empty(&info->swaplist)) {
 619                         mutex_lock(&shmem_swaplist_mutex);
 620                         list_del_init(&info->swaplist);
 621                         mutex_unlock(&shmem_swaplist_mutex);
 622                 }
 623         }
 624
 625         simple_xattrs_free(&info->xattrs);
 626         WARN_ON(inode->i_blocks);
 627         shmem_free_inode(inode->i_sb);
 628         clear_inode(inode);
 629 }
 630
 631 /*
 632  * If swap found in inode, free it and move page from swapcache to filecache.
 633  */
 634 static int shmem_unuse_inode(struct shmem_inode_info *info,
 635                              swp_entry_t swap, struct page **pagep)
 636 {
 637         struct address_space *mapping = info->vfs_inode.i_mapping;
 638         void *radswap;
 639         pgoff_t index;
 640         gfp_t gfp;
 641         int error = 0;
 642
 643         radswap = swp_to_radix_entry(swap);
 644         index = radix_tree_locate_item(&mapping->page_tree, radswap);
 645         if (index == -1)
 646                 return -EAGAIN; /* tell shmem_unuse we found nothing */
 647
 648         /*
 649          * Move _head_ to start search for next from here.
 650          * But be careful: shmem_evict_inode checks list_empty without taking
 651          * mutex, and there's an instant in list_move_tail when info->swaplist
 652          * would appear empty, if it were the only one on shmem_swaplist.
 653          */
 654         if (shmem_swaplist.next != &info->swaplist)
 655                 list_move_tail(&shmem_swaplist, &info->swaplist);
 656
 657         gfp = mapping_gfp_mask(mapping);
 658         if (shmem_should_replace_page(*pagep, gfp)) {
 659                 mutex_unlock(&shmem_swaplist_mutex);
 660                 error = shmem_replace_page(pagep, gfp, info, index);
 661                 mutex_lock(&shmem_swaplist_mutex);
 662                 /*
 663                  * We needed to drop mutex to make that restrictive page
 664                  * allocation, but the inode might have been freed while we
 665                  * dropped it: although a racing shmem_evict_inode() cannot
 666                  * complete without emptying the radix_tree, our page lock
 667                  * on this swapcache page is not enough to prevent that -
 668                  * free_swap_and_cache() of our swap entry will only
 669                  * trylock_page(), removing swap from radix_tree whatever.
 670                  *
 671                  * We must not proceed to shmem_add_to_page_cache() if the
 672                  * inode has been freed, but of course we cannot rely on
 673                  * inode or mapping or info to check that.  However, we can
 674                  * safely check if our swap entry is still in use (and here
 675                  * it can't have got reused for another page): if it's still
 676                  * in use, then the inode cannot have been freed yet, and we
 677                  * can safely proceed (if it's no longer in use, that tells
 678                  * nothing about the inode, but we don't need to unuse swap).
 679                  */
 680                 if (!page_swapcount(*pagep))
 681                         error = -ENOENT;
 682         }
 683
 684         /*
 685          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 686          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 687          * beneath us (pagelock doesn't help until the page is in pagecache).
 688          */
 689         if (!error)
 690                 error = shmem_add_to_page_cache(*pagep, mapping, index,
 691                                                 radswap);
 692         if (error != -ENOMEM) {
 693                 /*
 694                  * Truncation and eviction use free_swap_and_cache(), which
 695                  * only does trylock page: if we raced, best clean up here.
 696                  */
 697                 delete_from_swap_cache(*pagep);
 698                 set_page_dirty(*pagep);
 699                 if (!error) {
 700                         spin_lock(&info->lock);
 701                         info->swapped--;
 702                         spin_unlock(&info->lock);
 703                         swap_free(swap);
 704                 }
 705         }
 706         return error;
 707 }
 708
 709 /*
 710  * Search through swapped inodes to find and replace swap by page.
 711  */
 712 int shmem_unuse(swp_entry_t swap, struct page *page)
 713 {
 714         struct list_head *this, *next;
 715         struct shmem_inode_info *info;
 716         struct mem_cgroup *memcg;
 717         int error = 0;
 718
 719         /*
 720          * There's a faint possibility that swap page was replaced before
 721          * caller locked it: caller will come back later with the right page.
 722          */
 723         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
 724                 goto out;
 725
 726         /*
 727          * Charge page using GFP_KERNEL while we can wait, before taking
 728          * the shmem_swaplist_mutex which might hold up shmem_writepage().
 729          * Charged back to the user (not to caller) when swap account is used.
 730          */
 731         error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg);
 732         if (error)
 733                 goto out;
 734         /* No radix_tree_preload: swap entry keeps a place for page in tree */
 735         error = -EAGAIN;
 736
 737         mutex_lock(&shmem_swaplist_mutex);
 738         list_for_each_safe(this, next, &shmem_swaplist) {
 739                 info = list_entry(this, struct shmem_inode_info, swaplist);
 740                 if (info->swapped)
 741                         error = shmem_unuse_inode(info, swap, &page);
 742                 else
 743                         list_del_init(&info->swaplist);
 744                 cond_resched();
 745                 if (error != -EAGAIN)
 746                         break;
 747                 /* found nothing in this: move on to search the next */
 748         }
 749         mutex_unlock(&shmem_swaplist_mutex);
 750
 751         if (error) {
 752                 if (error != -ENOMEM)
 753                         error = 0;
 754                 mem_cgroup_cancel_charge(page, memcg);
 755         } else
 756                 mem_cgroup_commit_charge(page, memcg, true);
 757 out:
 758         unlock_page(page);
 759         page_cache_release(page);
 760         return error;
 761 }
 762
 763 /*
 764  * Move the page from the page cache to the swap cache.
 765  */
 766 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 767 {
 768         struct shmem_inode_info *info;
 769         struct address_space *mapping;
 770         struct inode *inode;
 771         swp_entry_t swap;
 772         pgoff_t index;
 773
 774         BUG_ON(!PageLocked(page));
 775         mapping = page->mapping;
 776         index = page->index;
 777         inode = mapping->host;
 778         info = SHMEM_I(inode);
 779         if (info->flags & VM_LOCKED)
 780                 goto redirty;
 781         if (!total_swap_pages)
 782                 goto redirty;
 783
 784         /*
 785          * Our capabilities prevent regular writeback or sync from ever calling
 786          * shmem_writepage; but a stacking filesystem might use ->writepage of
 787          * its underlying filesystem, in which case tmpfs should write out to
 788          * swap only in response to memory pressure, and not for the writeback
 789          * threads or sync.
 790          */
 791         if (!wbc->for_reclaim) {
 792                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
 793                 goto redirty;
 794         }
 795
 796         /*
 797          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
 798          * value into swapfile.c, the only way we can correctly account for a
 799          * fallocated page arriving here is now to initialize it and write it.
 800          *
 801          * That's okay for a page already fallocated earlier, but if we have
 802          * not yet completed the fallocation, then (a) we want to keep track
 803          * of this page in case we have to undo it, and (b) it may not be a
 804          * good idea to continue anyway, once we're pushing into swap.  So
 805          * reactivate the page, and let shmem_fallocate() quit when too many.
 806          */
 807         if (!PageUptodate(page)) {
 808                 if (inode->i_private) {
 809                         struct shmem_falloc *shmem_falloc;
 810                         spin_lock(&inode->i_lock);
 811                         shmem_falloc = inode->i_private;
 812                         if (shmem_falloc &&
 813                             !shmem_falloc->waitq &&
 814                             index >= shmem_falloc->start &&
 815                             index < shmem_falloc->next)
 816                                 shmem_falloc->nr_unswapped++;
 817                         else
 818                                 shmem_falloc = NULL;
 819                         spin_unlock(&inode->i_lock);
 820                         if (shmem_falloc)
 821                                 goto redirty;
 822                 }
 823                 clear_highpage(page);
 824                 flush_dcache_page(page);
 825                 SetPageUptodate(page);
 826         }
 827
 828         swap = get_swap_page();
 829         if (!swap.val)
 830                 goto redirty;
 831
 832         /*
 833          * Add inode to shmem_unuse()'s list of swapped-out inodes,
 834          * if it's not already there.  Do it now before the page is
 835          * moved to swap cache, when its pagelock no longer protects
 836          * the inode from eviction.  But don't unlock the mutex until
 837          * we've incremented swapped, because shmem_unuse_inode() will
 838          * prune a !swapped inode from the swaplist under this mutex.
 839          */
 840         mutex_lock(&shmem_swaplist_mutex);
 841         if (list_empty(&info->swaplist))
 842                 list_add_tail(&info->swaplist, &shmem_swaplist);
 843
 844         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
 845                 spin_lock(&info->lock);
 846                 shmem_recalc_inode(inode);
 847                 info->swapped++;
 848                 spin_unlock(&info->lock);
 849
 850                 swap_shmem_alloc(swap);
 851                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
 852
 853                 mutex_unlock(&shmem_swaplist_mutex);
 854                 BUG_ON(page_mapped(page));
 855                 swap_writepage(page, wbc);
 856                 return 0;
 857         }
 858
 859         mutex_unlock(&shmem_swaplist_mutex);
 860         swapcache_free(swap);
 861 redirty:
 862         set_page_dirty(page);
 863         if (wbc->for_reclaim)
 864                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
 865         unlock_page(page);
 866         return 0;
 867 }
 868
 869 #ifdef CONFIG_NUMA
 870 #ifdef CONFIG_TMPFS
 871 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 872 {
 873         char buffer[64];
 874
 875         if (!mpol || mpol->mode == MPOL_DEFAULT)
 876                 return;         /* show nothing */
 877
 878         mpol_to_str(buffer, sizeof(buffer), mpol);
 879
 880         seq_printf(seq, ",mpol=%s", buffer);
 881 }
 882
 883 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 884 {
 885         struct mempolicy *mpol = NULL;
 886         if (sbinfo->mpol) {
 887                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
 888                 mpol = sbinfo->mpol;
 889                 mpol_get(mpol);
 890                 spin_unlock(&sbinfo->stat_lock);
 891         }
 892         return mpol;
 893 }
 894 #endif /* CONFIG_TMPFS */
 895
 896 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 897                         struct shmem_inode_info *info, pgoff_t index)
 898 {
 899         struct vm_area_struct pvma;
 900         struct page *page;
 901
 902         /* Create a pseudo vma that just contains the policy */
 903         pvma.vm_start = 0;
 904         /* Bias interleave by inode number to distribute better across nodes */
 905         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 906         pvma.vm_ops = NULL;
 907         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 908
 909         page = swapin_readahead(swap, gfp, &pvma, 0);
 910
 911         /* Drop reference taken by mpol_shared_policy_lookup() */
 912         mpol_cond_put(pvma.vm_policy);
 913
 914         return page;
 915 }
 916
 917 static struct page *shmem_alloc_page(gfp_t gfp,
 918                         struct shmem_inode_info *info, pgoff_t index)
 919 {
 920         struct vm_area_struct pvma;
 921         struct page *page;
 922
 923         /* Create a pseudo vma that just contains the policy */
 924         pvma.vm_start = 0;
 925         /* Bias interleave by inode number to distribute better across nodes */
 926         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 927         pvma.vm_ops = NULL;
 928         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 929
 930         page = alloc_page_vma(gfp, &pvma, 0);
 931
 932         /* Drop reference taken by mpol_shared_policy_lookup() */
 933         mpol_cond_put(pvma.vm_policy);
 934
 935         return page;
 936 }
 937 #else /* !CONFIG_NUMA */
 938 #ifdef CONFIG_TMPFS
 939 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 940 {
 941 }
 942 #endif /* CONFIG_TMPFS */
 943
 944 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 945                         struct shmem_inode_info *info, pgoff_t index)
 946 {
 947         return swapin_readahead(swap, gfp, NULL, 0);
 948 }
 949
 950 static inline struct page *shmem_alloc_page(gfp_t gfp,
 951                         struct shmem_inode_info *info, pgoff_t index)
 952 {
 953         return alloc_page(gfp);
 954 }
 955 #endif /* CONFIG_NUMA */
 956
 957 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
 958 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 959 {
 960         return NULL;
 961 }
 962 #endif
 963
 964 /*
 965  * When a page is moved from swapcache to shmem filecache (either by the
 966  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
 967  * shmem_unuse_inode()), it may have been read in earlier from swap, in
 968  * ignorance of the mapping it belongs to.  If that mapping has special
 969  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
 970  * we may need to copy to a suitable page before moving to filecache.
 971  *
 972  * In a future release, this may well be extended to respect cpuset and
 973  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
 974  * but for now it is a simple matter of zone.
 975  */
 976 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
 977 {
 978         return page_zonenum(page) > gfp_zone(gfp);
 979 }
 980
 981 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 982                                 struct shmem_inode_info *info, pgoff_t index)
 983 {
 984         struct page *oldpage, *newpage;
 985         struct address_space *swap_mapping;
 986         pgoff_t swap_index;
 987         int error;
 988
 989         oldpage = *pagep;
 990         swap_index = page_private(oldpage);
 991         swap_mapping = page_mapping(oldpage);
 992
 993         /*
 994          * We have arrived here because our zones are constrained, so don't
 995          * limit chance of success by further cpuset and node constraints.
 996          */
 997         gfp &= ~GFP_CONSTRAINT_MASK;
 998         newpage = shmem_alloc_page(gfp, info, index);
 999         if (!newpage)
1000                 return -ENOMEM;
1001
1002         page_cache_get(newpage);
1003         copy_highpage(newpage, oldpage);
1004         flush_dcache_page(newpage);
1005
1006         __set_page_locked(newpage);
1007         SetPageUptodate(newpage);
1008         SetPageSwapBacked(newpage);
1009         set_page_private(newpage, swap_index);
1010         SetPageSwapCache(newpage);
1011
1012         /*
1013          * Our caller will very soon move newpage out of swapcache, but it's
1014          * a nice clean interface for us to replace oldpage by newpage there.
1015          */
1016         spin_lock_irq(&swap_mapping->tree_lock);
1017         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1018                                                                    newpage);
1019         if (!error) {
1020                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1021                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1022         }
1023         spin_unlock_irq(&swap_mapping->tree_lock);
1024
1025         if (unlikely(error)) {
1026                 /*
1027                  * Is this possible?  I think not, now that our callers check
1028                  * both PageSwapCache and page_private after getting page lock;
1029                  * but be defensive.  Reverse old to newpage for clear and free.
1030                  */
1031                 oldpage = newpage;
1032         } else {
1033                 mem_cgroup_replace_page(oldpage, newpage);
1034                 lru_cache_add_anon(newpage);
1035                 *pagep = newpage;
1036         }
1037
1038         ClearPageSwapCache(oldpage);
1039         set_page_private(oldpage, 0);
1040
1041         unlock_page(oldpage);
1042         page_cache_release(oldpage);
1043         page_cache_release(oldpage);
1044         return error;
1045 }
1046
1047 /*
1048  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1049  *
1050  * If we allocate a new one we do not mark it dirty. That's up to the
1051  * vm. If we swap it in we mark it dirty since we also free the swap
1052  * entry since a page cannot live in both the swap and page cache
1053  */
1054 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1055         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1056 {
1057         struct address_space *mapping = inode->i_mapping;
1058         struct shmem_inode_info *info;
1059         struct shmem_sb_info *sbinfo;
1060         struct mem_cgroup *memcg;
1061         struct page *page;
1062         swp_entry_t swap;
1063         int error;
1064         int once = 0;
1065         int alloced = 0;
1066
1067         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1068                 return -EFBIG;
1069 repeat:
1070         swap.val = 0;
1071         page = find_lock_entry(mapping, index);
1072         if (radix_tree_exceptional_entry(page)) {
1073                 swap = radix_to_swp_entry(page);
1074                 page = NULL;
1075         }
1076
1077         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1078             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1079                 error = -EINVAL;
1080                 goto unlock;
1081         }
1082
1083         if (page && sgp == SGP_WRITE)
1084                 mark_page_accessed(page);
1085
1086         /* fallocated page? */
1087         if (page && !PageUptodate(page)) {
1088                 if (sgp != SGP_READ)
1089                         goto clear;
1090                 unlock_page(page);
1091                 page_cache_release(page);
1092                 page = NULL;
1093         }
1094         if (page || (sgp == SGP_READ && !swap.val)) {
1095                 *pagep = page;
1096                 return 0;
1097         }
1098
1099         /*
1100          * Fast cache lookup did not find it:
1101          * bring it back from swap or allocate.
1102          */
1103         info = SHMEM_I(inode);
1104         sbinfo = SHMEM_SB(inode->i_sb);
1105
1106         if (swap.val) {
1107                 /* Look it up and read it in.. */
1108                 page = lookup_swap_cache(swap);
1109                 if (!page) {
1110                         /* here we actually do the io */
1111                         if (fault_type)
1112                                 *fault_type |= VM_FAULT_MAJOR;
1113                         page = shmem_swapin(swap, gfp, info, index);
1114                         if (!page) {
1115                                 error = -ENOMEM;
1116                                 goto failed;
1117                         }
1118                 }
1119
1120                 /* We have to do this with page locked to prevent races */
1121                 lock_page(page);
1122                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1123                     !shmem_confirm_swap(mapping, index, swap)) {
1124                         error = -EEXIST;        /* try again */
1125                         goto unlock;
1126                 }
1127                 if (!PageUptodate(page)) {
1128                         error = -EIO;
1129                         goto failed;
1130                 }
1131                 wait_on_page_writeback(page);
1132
1133                 if (shmem_should_replace_page(page, gfp)) {
1134                         error = shmem_replace_page(&page, gfp, info, index);
1135                         if (error)
1136                                 goto failed;
1137                 }
1138
1139                 error = mem_cgroup_try_charge(page, current->mm, gfp, &memcg);
1140                 if (!error) {
1141                         error = shmem_add_to_page_cache(page, mapping, index,
1142                                                 swp_to_radix_entry(swap));
1143                         /*
1144                          * We already confirmed swap under page lock, and make
1145                          * no memory allocation here, so usually no possibility
1146                          * of error; but free_swap_and_cache() only trylocks a
1147                          * page, so it is just possible that the entry has been
1148                          * truncated or holepunched since swap was confirmed.
1149                          * shmem_undo_range() will have done some of the
1150                          * unaccounting, now delete_from_swap_cache() will do
1151                          * the rest.
1152                          * Reset swap.val? No, leave it so "failed" goes back to
1153                          * "repeat": reading a hole and writing should succeed.
1154                          */
1155                         if (error) {
1156                                 mem_cgroup_cancel_charge(page, memcg);
1157                                 delete_from_swap_cache(page);
1158                         }
1159                 }
1160                 if (error)
1161                         goto failed;
1162
1163                 mem_cgroup_commit_charge(page, memcg, true);
1164
1165                 spin_lock(&info->lock);
1166                 info->swapped--;
1167                 shmem_recalc_inode(inode);
1168                 spin_unlock(&info->lock);
1169
1170                 if (sgp == SGP_WRITE)
1171                         mark_page_accessed(page);
1172
1173                 delete_from_swap_cache(page);
1174                 set_page_dirty(page);
1175                 swap_free(swap);
1176
1177         } else {
1178                 if (shmem_acct_block(info->flags)) {
1179                         error = -ENOSPC;
1180                         goto failed;
1181                 }
1182                 if (sbinfo->max_blocks) {
1183                         if (percpu_counter_compare(&sbinfo->used_blocks,
1184                                                 sbinfo->max_blocks) >= 0) {
1185                                 error = -ENOSPC;
1186                                 goto unacct;
1187                         }
1188                         percpu_counter_inc(&sbinfo->used_blocks);
1189                 }
1190
1191                 page = shmem_alloc_page(gfp, info, index);
1192                 if (!page) {
1193                         error = -ENOMEM;
1194                         goto decused;
1195                 }
1196
1197                 __SetPageSwapBacked(page);
1198                 __set_page_locked(page);
1199                 if (sgp == SGP_WRITE)
1200                         __SetPageReferenced(page);
1201
1202                 error = mem_cgroup_try_charge(page, current->mm, gfp, &memcg);
1203                 if (error)
1204                         goto decused;
1205                 error = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
1206                 if (!error) {
1207                         error = shmem_add_to_page_cache(page, mapping, index,
1208                                                         NULL);
1209                         radix_tree_preload_end();
1210                 }
1211                 if (error) {
1212                         mem_cgroup_cancel_charge(page, memcg);
1213                         goto decused;
1214                 }
1215                 mem_cgroup_commit_charge(page, memcg, false);
1216                 lru_cache_add_anon(page);
1217
1218                 spin_lock(&info->lock);
1219                 info->alloced++;
1220                 inode->i_blocks += BLOCKS_PER_PAGE;
1221                 shmem_recalc_inode(inode);
1222                 spin_unlock(&info->lock);
1223                 alloced = true;
1224
1225                 /*
1226                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1227                  */
1228                 if (sgp == SGP_FALLOC)
1229                         sgp = SGP_WRITE;
1230 clear:
1231                 /*
1232                  * Let SGP_WRITE caller clear ends if write does not fill page;
1233                  * but SGP_FALLOC on a page fallocated earlier must initialize
1234                  * it now, lest undo on failure cancel our earlier guarantee.
1235                  */
1236                 if (sgp != SGP_WRITE) {
1237                         clear_highpage(page);
1238                         flush_dcache_page(page);
1239                         SetPageUptodate(page);
1240                 }
1241                 if (sgp == SGP_DIRTY)
1242                         set_page_dirty(page);
1243         }
1244
1245         /* Perhaps the file has been truncated since we checked */
1246         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1247             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1248                 if (alloced) {
1249                         ClearPageDirty(page);
1250                         delete_from_page_cache(page);
1251                         spin_lock(&info->lock);
1252                         shmem_recalc_inode(inode);
1253                         spin_unlock(&info->lock);
1254                 }
1255                 error = -EINVAL;
1256                 goto unlock;
1257         }
1258         *pagep = page;
1259         return 0;
1260
1261         /*
1262          * Error recovery.
1263          */
1264 decused:
1265         if (sbinfo->max_blocks)
1266                 percpu_counter_add(&sbinfo->used_blocks, -1);
1267 unacct:
1268         shmem_unacct_blocks(info->flags, 1);
1269 failed:
1270         if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1271                 error = -EEXIST;
1272 unlock:
1273         if (page) {
1274                 unlock_page(page);
1275                 page_cache_release(page);
1276         }
1277         if (error == -ENOSPC && !once++) {
1278                 info = SHMEM_I(inode);
1279                 spin_lock(&info->lock);
1280                 shmem_recalc_inode(inode);
1281                 spin_unlock(&info->lock);
1282                 goto repeat;
1283         }
1284         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1285                 goto repeat;
1286         return error;
1287 }
1288
1289 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1290 {
1291         struct inode *inode = file_inode(vma->vm_file);
1292         int error;
1293         int ret = VM_FAULT_LOCKED;
1294
1295         /*
1296          * Trinity finds that probing a hole which tmpfs is punching can
1297          * prevent the hole-punch from ever completing: which in turn
1298          * locks writers out with its hold on i_mutex.  So refrain from
1299          * faulting pages into the hole while it's being punched.  Although
1300          * shmem_undo_range() does remove the additions, it may be unable to
1301          * keep up, as each new page needs its own unmap_mapping_range() call,
1302          * and the i_mmap tree grows ever slower to scan if new vmas are added.
1303          *
1304          * It does not matter if we sometimes reach this check just before the
1305          * hole-punch begins, so that one fault then races with the punch:
1306          * we just need to make racing faults a rare case.
1307          *
1308          * The implementation below would be much simpler if we just used a
1309          * standard mutex or completion: but we cannot take i_mutex in fault,
1310          * and bloating every shmem inode for this unlikely case would be sad.
1311          */
1312         if (unlikely(inode->i_private)) {
1313                 struct shmem_falloc *shmem_falloc;
1314
1315                 spin_lock(&inode->i_lock);
1316                 shmem_falloc = inode->i_private;
1317                 if (shmem_falloc &&
1318                     shmem_falloc->waitq &&
1319                     vmf->pgoff >= shmem_falloc->start &&
1320                     vmf->pgoff < shmem_falloc->next) {
1321                         wait_queue_head_t *shmem_falloc_waitq;
1322                         DEFINE_WAIT(shmem_fault_wait);
1323
1324                         ret = VM_FAULT_NOPAGE;
1325                         if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1326                            !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1327                                 /* It's polite to up mmap_sem if we can */
1328                                 up_read(&vma->vm_mm->mmap_sem);
1329                                 ret = VM_FAULT_RETRY;
1330                         }
1331
1332                         shmem_falloc_waitq = shmem_falloc->waitq;
1333                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1334                                         TASK_UNINTERRUPTIBLE);
1335                         spin_unlock(&inode->i_lock);
1336                         schedule();
1337
1338                         /*
1339                          * shmem_falloc_waitq points into the shmem_fallocate()
1340                          * stack of the hole-punching task: shmem_falloc_waitq
1341                          * is usually invalid by the time we reach here, but
1342                          * finish_wait() does not dereference it in that case;
1343                          * though i_lock needed lest racing with wake_up_all().
1344                          */
1345                         spin_lock(&inode->i_lock);
1346                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1347                         spin_unlock(&inode->i_lock);
1348                         return ret;
1349                 }
1350                 spin_unlock(&inode->i_lock);
1351         }
1352
1353         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1354         if (error)
1355                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1356
1357         if (ret & VM_FAULT_MAJOR) {
1358                 count_vm_event(PGMAJFAULT);
1359                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1360         }
1361         return ret;
1362 }
1363
1364 #ifdef CONFIG_NUMA
1365 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1366 {
1367         struct inode *inode = file_inode(vma->vm_file);
1368         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1369 }
1370
1371 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1372                                           unsigned long addr)
1373 {
1374         struct inode *inode = file_inode(vma->vm_file);
1375         pgoff_t index;
1376
1377         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1378         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1379 }
1380 #endif
1381
1382 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1383 {
1384         struct inode *inode = file_inode(file);
1385         struct shmem_inode_info *info = SHMEM_I(inode);
1386         int retval = -ENOMEM;
1387
1388         spin_lock(&info->lock);
1389         if (lock && !(info->flags & VM_LOCKED)) {
1390                 if (!user_shm_lock(inode->i_size, user))
1391                         goto out_nomem;
1392                 info->flags |= VM_LOCKED;
1393                 mapping_set_unevictable(file->f_mapping);
1394         }
1395         if (!lock && (info->flags & VM_LOCKED) && user) {
1396                 user_shm_unlock(inode->i_size, user);
1397                 info->flags &= ~VM_LOCKED;
1398                 mapping_clear_unevictable(file->f_mapping);
1399         }
1400         retval = 0;
1401
1402 out_nomem:
1403         spin_unlock(&info->lock);
1404         return retval;
1405 }
1406
1407 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1408 {
1409         file_accessed(file);
1410         vma->vm_ops = &shmem_vm_ops;
1411         return 0;
1412 }
1413
1414 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1415                                      umode_t mode, dev_t dev, unsigned long flags)
1416 {
1417         struct inode *inode;
1418         struct shmem_inode_info *info;
1419         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1420
1421         if (shmem_reserve_inode(sb))
1422                 return NULL;
1423
1424         inode = new_inode(sb);
1425         if (inode) {
1426                 inode->i_ino = get_next_ino();
1427                 inode_init_owner(inode, dir, mode);
1428                 inode->i_blocks = 0;
1429                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1430                 inode->i_generation = get_seconds();
1431                 info = SHMEM_I(inode);
1432                 memset(info, 0, (char *)inode - (char *)info);
1433                 spin_lock_init(&info->lock);
1434                 info->seals = F_SEAL_SEAL;
1435                 info->flags = flags & VM_NORESERVE;
1436                 INIT_LIST_HEAD(&info->swaplist);
1437                 simple_xattrs_init(&info->xattrs);
1438                 cache_no_acl(inode);
1439
1440                 switch (mode & S_IFMT) {
1441                 default:
1442                         inode->i_op = &shmem_special_inode_operations;
1443                         init_special_inode(inode, mode, dev);
1444                         break;
1445                 case S_IFREG:
1446                         inode->i_mapping->a_ops = &shmem_aops;
1447                         inode->i_op = &shmem_inode_operations;
1448                         inode->i_fop = &shmem_file_operations;
1449                         mpol_shared_policy_init(&info->policy,
1450                                                  shmem_get_sbmpol(sbinfo));
1451                         break;
1452                 case S_IFDIR:
1453                         inc_nlink(inode);
1454                         /* Some things misbehave if size == 0 on a directory */
1455                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1456                         inode->i_op = &shmem_dir_inode_operations;
1457                         inode->i_fop = &simple_dir_operations;
1458                         break;
1459                 case S_IFLNK:
1460                         /*
1461                          * Must not load anything in the rbtree,
1462                          * mpol_free_shared_policy will not be called.
1463                          */
1464                         mpol_shared_policy_init(&info->policy, NULL);
1465                         break;
1466                 }
1467
1468                 lockdep_annotate_inode_mutex_key(inode);
1469         } else
1470                 shmem_free_inode(sb);
1471         return inode;
1472 }
1473
1474 bool shmem_mapping(struct address_space *mapping)
1475 {
1476         if (!mapping->host)
1477                 return false;
1478
1479         return mapping->host->i_sb->s_op == &shmem_ops;
1480 }
1481
1482 #ifdef CONFIG_TMPFS
1483 static const struct inode_operations shmem_symlink_inode_operations;
1484 static const struct inode_operations shmem_short_symlink_operations;
1485
1486 #ifdef CONFIG_TMPFS_XATTR
1487 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1488 #else
1489 #define shmem_initxattrs NULL
1490 #endif
1491
1492 static int
1493 shmem_write_begin(struct file *file, struct address_space *mapping,
1494                         loff_t pos, unsigned len, unsigned flags,
1495                         struct page **pagep, void **fsdata)
1496 {
1497         struct inode *inode = mapping->host;
1498         struct shmem_inode_info *info = SHMEM_I(inode);
1499         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1500
1501         /* i_mutex is held by caller */
1502         if (unlikely(info->seals)) {
1503                 if (info->seals & F_SEAL_WRITE)
1504                         return -EPERM;
1505                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
1506                         return -EPERM;
1507         }
1508
1509         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1510 }
1511
1512 static int
1513 shmem_write_end(struct file *file, struct address_space *mapping,
1514                         loff_t pos, unsigned len, unsigned copied,
1515                         struct page *page, void *fsdata)
1516 {
1517         struct inode *inode = mapping->host;
1518
1519         if (pos + copied > inode->i_size)
1520                 i_size_write(inode, pos + copied);
1521
1522         if (!PageUptodate(page)) {
1523                 if (copied < PAGE_CACHE_SIZE) {
1524                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1525                         zero_user_segments(page, 0, from,
1526                                         from + copied, PAGE_CACHE_SIZE);
1527                 }
1528                 SetPageUptodate(page);
1529         }
1530         set_page_dirty(page);
1531         unlock_page(page);
1532         page_cache_release(page);
1533
1534         return copied;
1535 }
1536
1537 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1538 {
1539         struct file *file = iocb->ki_filp;
1540         struct inode *inode = file_inode(file);
1541         struct address_space *mapping = inode->i_mapping;
1542         pgoff_t index;
1543         unsigned long offset;
1544         enum sgp_type sgp = SGP_READ;
1545         int error = 0;
1546         ssize_t retval = 0;
1547         loff_t *ppos = &iocb->ki_pos;
1548
1549         /*
1550          * Might this read be for a stacking filesystem?  Then when reading
1551          * holes of a sparse file, we actually need to allocate those pages,
1552          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1553          */
1554         if (!iter_is_iovec(to))
1555                 sgp = SGP_DIRTY;
1556
1557         index = *ppos >> PAGE_CACHE_SHIFT;
1558         offset = *ppos & ~PAGE_CACHE_MASK;
1559
1560         for (;;) {
1561                 struct page *page = NULL;
1562                 pgoff_t end_index;
1563                 unsigned long nr, ret;
1564                 loff_t i_size = i_size_read(inode);
1565
1566                 end_index = i_size >> PAGE_CACHE_SHIFT;
1567                 if (index > end_index)
1568                         break;
1569                 if (index == end_index) {
1570                         nr = i_size & ~PAGE_CACHE_MASK;
1571                         if (nr <= offset)
1572                                 break;
1573                 }
1574
1575                 error = shmem_getpage(inode, index, &page, sgp, NULL);
1576                 if (error) {
1577                         if (error == -EINVAL)
1578                                 error = 0;
1579                         break;
1580                 }
1581                 if (page)
1582                         unlock_page(page);
1583
1584                 /*
1585                  * We must evaluate after, since reads (unlike writes)
1586                  * are called without i_mutex protection against truncate
1587                  */
1588                 nr = PAGE_CACHE_SIZE;
1589                 i_size = i_size_read(inode);
1590                 end_index = i_size >> PAGE_CACHE_SHIFT;
1591                 if (index == end_index) {
1592                         nr = i_size & ~PAGE_CACHE_MASK;
1593                         if (nr <= offset) {
1594                                 if (page)
1595                                         page_cache_release(page);
1596                                 break;
1597                         }
1598                 }
1599                 nr -= offset;
1600
1601                 if (page) {
1602                         /*
1603                          * If users can be writing to this page using arbitrary
1604                          * virtual addresses, take care about potential aliasing
1605                          * before reading the page on the kernel side.
1606                          */
1607                         if (mapping_writably_mapped(mapping))
1608                                 flush_dcache_page(page);
1609                         /*
1610                          * Mark the page accessed if we read the beginning.
1611                          */
1612                         if (!offset)
1613                                 mark_page_accessed(page);
1614                 } else {
1615                         page = ZERO_PAGE(0);
1616                         page_cache_get(page);
1617                 }
1618
1619                 /*
1620                  * Ok, we have the page, and it's up-to-date, so
1621                  * now we can copy it to user space...
1622                  */
1623                 ret = copy_page_to_iter(page, offset, nr, to);
1624                 retval += ret;
1625                 offset += ret;
1626                 index += offset >> PAGE_CACHE_SHIFT;
1627                 offset &= ~PAGE_CACHE_MASK;
1628
1629                 page_cache_release(page);
1630                 if (!iov_iter_count(to))
1631                         break;
1632                 if (ret < nr) {
1633                         error = -EFAULT;
1634                         break;
1635                 }
1636                 cond_resched();
1637         }
1638
1639         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1640         file_accessed(file);
1641         return retval ? retval : error;
1642 }
1643
1644 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1645                                 struct pipe_inode_info *pipe, size_t len,
1646                                 unsigned int flags)
1647 {
1648         struct address_space *mapping = in->f_mapping;
1649         struct inode *inode = mapping->host;
1650         unsigned int loff, nr_pages, req_pages;
1651         struct page *pages[PIPE_DEF_BUFFERS];
1652         struct partial_page partial[PIPE_DEF_BUFFERS];
1653         struct page *page;
1654         pgoff_t index, end_index;
1655         loff_t isize, left;
1656         int error, page_nr;
1657         struct splice_pipe_desc spd = {
1658                 .pages = pages,
1659                 .partial = partial,
1660                 .nr_pages_max = PIPE_DEF_BUFFERS,
1661                 .flags = flags,
1662                 .ops = &page_cache_pipe_buf_ops,
1663                 .spd_release = spd_release_page,
1664         };
1665
1666         isize = i_size_read(inode);
1667         if (unlikely(*ppos >= isize))
1668                 return 0;
1669
1670         left = isize - *ppos;
1671         if (unlikely(left < len))
1672                 len = left;
1673
1674         if (splice_grow_spd(pipe, &spd))
1675                 return -ENOMEM;
1676
1677         index = *ppos >> PAGE_CACHE_SHIFT;
1678         loff = *ppos & ~PAGE_CACHE_MASK;
1679         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1680         nr_pages = min(req_pages, spd.nr_pages_max);
1681
1682         spd.nr_pages = find_get_pages_contig(mapping, index,
1683                                                 nr_pages, spd.pages);
1684         index += spd.nr_pages;
1685         error = 0;
1686
1687         while (spd.nr_pages < nr_pages) {
1688                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1689                 if (error)
1690                         break;
1691                 unlock_page(page);
1692                 spd.pages[spd.nr_pages++] = page;
1693                 index++;
1694         }
1695
1696         index = *ppos >> PAGE_CACHE_SHIFT;
1697         nr_pages = spd.nr_pages;
1698         spd.nr_pages = 0;
1699
1700         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1701                 unsigned int this_len;
1702
1703                 if (!len)
1704                         break;
1705
1706                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1707                 page = spd.pages[page_nr];
1708
1709                 if (!PageUptodate(page) || page->mapping != mapping) {
1710                         error = shmem_getpage(inode, index, &page,
1711                                                         SGP_CACHE, NULL);
1712                         if (error)
1713                                 break;
1714                         unlock_page(page);
1715                         page_cache_release(spd.pages[page_nr]);
1716                         spd.pages[page_nr] = page;
1717                 }
1718
1719                 isize = i_size_read(inode);
1720                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1721                 if (unlikely(!isize || index > end_index))
1722                         break;
1723
1724                 if (end_index == index) {
1725                         unsigned int plen;
1726
1727                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1728                         if (plen <= loff)
1729                                 break;
1730
1731                         this_len = min(this_len, plen - loff);
1732                         len = this_len;
1733                 }
1734
1735                 spd.partial[page_nr].offset = loff;
1736                 spd.partial[page_nr].len = this_len;
1737                 len -= this_len;
1738                 loff = 0;
1739                 spd.nr_pages++;
1740                 index++;
1741         }
1742
1743         while (page_nr < nr_pages)
1744                 page_cache_release(spd.pages[page_nr++]);
1745
1746         if (spd.nr_pages)
1747                 error = splice_to_pipe(pipe, &spd);
1748
1749         splice_shrink_spd(&spd);
1750
1751         if (error > 0) {
1752                 *ppos += error;
1753                 file_accessed(in);
1754         }
1755         return error;
1756 }
1757
1758 /*
1759  * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
1760  */
1761 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
1762                                     pgoff_t index, pgoff_t end, int whence)
1763 {
1764         struct page *page;
1765         struct pagevec pvec;
1766         pgoff_t indices[PAGEVEC_SIZE];
1767         bool done = false;
1768         int i;
1769
1770         pagevec_init(&pvec, 0);
1771         pvec.nr = 1;            /* start small: we may be there already */
1772         while (!done) {
1773                 pvec.nr = find_get_entries(mapping, index,
1774                                         pvec.nr, pvec.pages, indices);
1775                 if (!pvec.nr) {
1776                         if (whence == SEEK_DATA)
1777                                 index = end;
1778                         break;
1779                 }
1780                 for (i = 0; i < pvec.nr; i++, index++) {
1781                         if (index < indices[i]) {
1782                                 if (whence == SEEK_HOLE) {
1783                                         done = true;
1784                                         break;
1785                                 }
1786                                 index = indices[i];
1787                         }
1788                         page = pvec.pages[i];
1789                         if (page && !radix_tree_exceptional_entry(page)) {
1790                                 if (!PageUptodate(page))
1791                                         page = NULL;
1792                         }
1793                         if (index >= end ||
1794                             (page && whence == SEEK_DATA) ||
1795                             (!page && whence == SEEK_HOLE)) {
1796                                 done = true;
1797                                 break;
1798                         }
1799                 }
1800                 pagevec_remove_exceptionals(&pvec);
1801                 pagevec_release(&pvec);
1802                 pvec.nr = PAGEVEC_SIZE;
1803                 cond_resched();
1804         }
1805         return index;
1806 }
1807
1808 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
1809 {
1810         struct address_space *mapping = file->f_mapping;
1811         struct inode *inode = mapping->host;
1812         pgoff_t start, end;
1813         loff_t new_offset;
1814
1815         if (whence != SEEK_DATA && whence != SEEK_HOLE)
1816                 return generic_file_llseek_size(file, offset, whence,
1817                                         MAX_LFS_FILESIZE, i_size_read(inode));
1818         mutex_lock(&inode->i_mutex);
1819         /* We're holding i_mutex so we can access i_size directly */
1820
1821         if (offset < 0 || offset >= inode->i_size)
1822                 offset = -ENXIO;
1823         else {
1824                 start = offset >> PAGE_CACHE_SHIFT;
1825                 end = (inode->i_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1826                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
1827                 new_offset <<= PAGE_CACHE_SHIFT;
1828                 if (new_offset > offset) {
1829                         if (new_offset < inode->i_size)
1830                                 offset = new_offset;
1831                         else if (whence == SEEK_DATA)
1832                                 offset = -ENXIO;
1833                         else
1834                                 offset = inode->i_size;
1835                 }
1836         }
1837
1838         if (offset >= 0)
1839                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
1840         mutex_unlock(&inode->i_mutex);
1841         return offset;
1842 }
1843
1844 /*
1845  * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
1846  * so reuse a tag which we firmly believe is never set or cleared on shmem.
1847  */
1848 #define SHMEM_TAG_PINNED        PAGECACHE_TAG_TOWRITE
1849 #define LAST_SCAN               4       /* about 150ms max */
1850
1851 static void shmem_tag_pins(struct address_space *mapping)
1852 {
1853         struct radix_tree_iter iter;
1854         void **slot;
1855         pgoff_t start;
1856         struct page *page;
1857         unsigned int tagged = 0;
1858
1859         lru_add_drain();
1860         start = 0;
1861
1862         spin_lock_irq(&mapping->tree_lock);
1863 restart:
1864         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1865                 page = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
1866                 if (!page || radix_tree_exception(page)) {
1867                         if (radix_tree_deref_retry(page))
1868                                 goto restart;
1869                 } else if (page_count(page) - page_mapcount(page) > 1) {
1870                         radix_tree_tag_set(&mapping->page_tree, iter.index,
1871                                            SHMEM_TAG_PINNED);
1872                 }
1873
1874                 if (++tagged % 1024)
1875                         continue;
1876
1877                 spin_unlock_irq(&mapping->tree_lock);
1878                 cond_resched();
1879                 start = iter.index + 1;
1880                 spin_lock_irq(&mapping->tree_lock);
1881                 goto restart;
1882         }
1883         spin_unlock_irq(&mapping->tree_lock);
1884 }
1885
1886 /*
1887  * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
1888  * via get_user_pages(), drivers might have some pending I/O without any active
1889  * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
1890  * and see whether it has an elevated ref-count. If so, we tag them and wait for
1891  * them to be dropped.
1892  * The caller must guarantee that no new user will acquire writable references
1893  * to those pages to avoid races.
1894  */
1895 static int shmem_wait_for_pins(struct address_space *mapping)
1896 {
1897         struct radix_tree_iter iter;
1898         void **slot;
1899         pgoff_t start;
1900         struct page *page;
1901         int error, scan;
1902
1903         shmem_tag_pins(mapping);
1904
1905         error = 0;
1906         for (scan = 0; scan <= LAST_SCAN; scan++) {
1907                 if (!radix_tree_tagged(&mapping->page_tree, SHMEM_TAG_PINNED))
1908                         break;
1909
1910                 if (!scan)
1911                         lru_add_drain_all();
1912                 else if (schedule_timeout_killable((HZ << scan) / 200))
1913                         scan = LAST_SCAN;
1914
1915                 start = 0;
1916                 rcu_read_lock();
1917 restart:
1918                 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter,
1919                                            start, SHMEM_TAG_PINNED) {
1920
1921                         page = radix_tree_deref_slot(slot);
1922                         if (radix_tree_exception(page)) {
1923                                 if (radix_tree_deref_retry(page))
1924                                         goto restart;
1925
1926                                 page = NULL;
1927                         }
1928
1929                         if (page &&
1930                             page_count(page) - page_mapcount(page) != 1) {
1931                                 if (scan < LAST_SCAN)
1932                                         goto continue_resched;
1933
1934                                 /*
1935                                  * On the last scan, we clean up all those tags
1936                                  * we inserted; but make a note that we still
1937                                  * found pages pinned.
1938                                  */
1939                                 error = -EBUSY;
1940                         }
1941
1942                         spin_lock_irq(&mapping->tree_lock);
1943                         radix_tree_tag_clear(&mapping->page_tree,
1944                                              iter.index, SHMEM_TAG_PINNED);
1945                         spin_unlock_irq(&mapping->tree_lock);
1946 continue_resched:
1947                         if (need_resched()) {
1948                                 cond_resched_rcu();
1949                                 start = iter.index + 1;
1950                                 goto restart;
1951                         }
1952                 }
1953                 rcu_read_unlock();
1954         }
1955
1956         return error;
1957 }
1958
1959 #define F_ALL_SEALS (F_SEAL_SEAL | \
1960                      F_SEAL_SHRINK | \
1961                      F_SEAL_GROW | \
1962                      F_SEAL_WRITE)
1963
1964 int shmem_add_seals(struct file *file, unsigned int seals)
1965 {
1966         struct inode *inode = file_inode(file);
1967         struct shmem_inode_info *info = SHMEM_I(inode);
1968         int error;
1969
1970         /*
1971          * SEALING
1972          * Sealing allows multiple parties to share a shmem-file but restrict
1973          * access to a specific subset of file operations. Seals can only be
1974          * added, but never removed. This way, mutually untrusted parties can
1975          * share common memory regions with a well-defined policy. A malicious
1976          * peer can thus never perform unwanted operations on a shared object.
1977          *
1978          * Seals are only supported on special shmem-files and always affect
1979          * the whole underlying inode. Once a seal is set, it may prevent some
1980          * kinds of access to the file. Currently, the following seals are
1981          * defined:
1982          *   SEAL_SEAL: Prevent further seals from being set on this file
1983          *   SEAL_SHRINK: Prevent the file from shrinking
1984          *   SEAL_GROW: Prevent the file from growing
1985          *   SEAL_WRITE: Prevent write access to the file
1986          *
1987          * As we don't require any trust relationship between two parties, we
1988          * must prevent seals from being removed. Therefore, sealing a file
1989          * only adds a given set of seals to the file, it never touches
1990          * existing seals. Furthermore, the "setting seals"-operation can be
1991          * sealed itself, which basically prevents any further seal from being
1992          * added.
1993          *
1994          * Semantics of sealing are only defined on volatile files. Only
1995          * anonymous shmem files support sealing. More importantly, seals are
1996          * never written to disk. Therefore, there's no plan to support it on
1997          * other file types.
1998          */
1999
2000         if (file->f_op != &shmem_file_operations)
2001                 return -EINVAL;
2002         if (!(file->f_mode & FMODE_WRITE))
2003                 return -EPERM;
2004         if (seals & ~(unsigned int)F_ALL_SEALS)
2005                 return -EINVAL;
2006
2007         mutex_lock(&inode->i_mutex);
2008
2009         if (info->seals & F_SEAL_SEAL) {
2010                 error = -EPERM;
2011                 goto unlock;
2012         }
2013
2014         if ((seals & F_SEAL_WRITE) && !(info->seals & F_SEAL_WRITE)) {
2015                 error = mapping_deny_writable(file->f_mapping);
2016                 if (error)
2017                         goto unlock;
2018
2019                 error = shmem_wait_for_pins(file->f_mapping);
2020                 if (error) {
2021                         mapping_allow_writable(file->f_mapping);
2022                         goto unlock;
2023                 }
2024         }
2025
2026         info->seals |= seals;
2027         error = 0;
2028
2029 unlock:
2030         mutex_unlock(&inode->i_mutex);
2031         return error;
2032 }
2033 EXPORT_SYMBOL_GPL(shmem_add_seals);
2034
2035 int shmem_get_seals(struct file *file)
2036 {
2037         if (file->f_op != &shmem_file_operations)
2038                 return -EINVAL;
2039
2040         return SHMEM_I(file_inode(file))->seals;
2041 }
2042 EXPORT_SYMBOL_GPL(shmem_get_seals);
2043
2044 long shmem_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2045 {
2046         long error;
2047
2048         switch (cmd) {
2049         case F_ADD_SEALS:
2050                 /* disallow upper 32bit */
2051                 if (arg > UINT_MAX)
2052                         return -EINVAL;
2053
2054                 error = shmem_add_seals(file, arg);
2055                 break;
2056         case F_GET_SEALS:
2057                 error = shmem_get_seals(file);
2058                 break;
2059         default:
2060                 error = -EINVAL;
2061                 break;
2062         }
2063
2064         return error;
2065 }
2066
2067 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2068                                                          loff_t len)
2069 {
2070         struct inode *inode = file_inode(file);
2071         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2072         struct shmem_inode_info *info = SHMEM_I(inode);
2073         struct shmem_falloc shmem_falloc;
2074         pgoff_t start, index, end;
2075         int error;
2076
2077         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2078                 return -EOPNOTSUPP;
2079
2080         mutex_lock(&inode->i_mutex);
2081
2082         if (mode & FALLOC_FL_PUNCH_HOLE) {
2083                 struct address_space *mapping = file->f_mapping;
2084                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2085                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2086                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2087
2088                 /* protected by i_mutex */
2089                 if (info->seals & F_SEAL_WRITE) {
2090                         error = -EPERM;
2091                         goto out;
2092                 }
2093
2094                 shmem_falloc.waitq = &shmem_falloc_waitq;
2095                 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2096                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2097                 spin_lock(&inode->i_lock);
2098                 inode->i_private = &shmem_falloc;
2099                 spin_unlock(&inode->i_lock);
2100
2101                 if ((u64)unmap_end > (u64)unmap_start)
2102                         unmap_mapping_range(mapping, unmap_start,
2103                                             1 + unmap_end - unmap_start, 0);
2104                 shmem_truncate_range(inode, offset, offset + len - 1);
2105                 /* No need to unmap again: hole-punching leaves COWed pages */
2106
2107                 spin_lock(&inode->i_lock);
2108                 inode->i_private = NULL;
2109                 wake_up_all(&shmem_falloc_waitq);
2110                 spin_unlock(&inode->i_lock);
2111                 error = 0;
2112                 goto out;
2113         }
2114
2115         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2116         error = inode_newsize_ok(inode, offset + len);
2117         if (error)
2118                 goto out;
2119
2120         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2121                 error = -EPERM;
2122                 goto out;
2123         }
2124
2125         start = offset >> PAGE_CACHE_SHIFT;
2126         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2127         /* Try to avoid a swapstorm if len is impossible to satisfy */
2128         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2129                 error = -ENOSPC;
2130                 goto out;
2131         }
2132
2133         shmem_falloc.waitq = NULL;
2134         shmem_falloc.start = start;
2135         shmem_falloc.next  = start;
2136         shmem_falloc.nr_falloced = 0;
2137         shmem_falloc.nr_unswapped = 0;
2138         spin_lock(&inode->i_lock);
2139         inode->i_private = &shmem_falloc;
2140         spin_unlock(&inode->i_lock);
2141
2142         for (index = start; index < end; index++) {
2143                 struct page *page;
2144
2145                 /*
2146                  * Good, the fallocate(2) manpage permits EINTR: we may have
2147                  * been interrupted because we are using up too much memory.
2148                  */
2149                 if (signal_pending(current))
2150                         error = -EINTR;
2151                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2152                         error = -ENOMEM;
2153                 else
2154                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
2155                                                                         NULL);
2156                 if (error) {
2157                         /* Remove the !PageUptodate pages we added */
2158                         if (index > start) {
2159                                 shmem_undo_range(inode,
2160                                  (loff_t)start << PAGE_CACHE_SHIFT,
2161                                  ((loff_t)index << PAGE_CACHE_SHIFT) - 1, true);
2162                         }
2163                         goto undone;
2164                 }
2165
2166                 /*
2167                  * Inform shmem_writepage() how far we have reached.
2168                  * No need for lock or barrier: we have the page lock.
2169                  */
2170                 shmem_falloc.next++;
2171                 if (!PageUptodate(page))
2172                         shmem_falloc.nr_falloced++;
2173
2174                 /*
2175                  * If !PageUptodate, leave it that way so that freeable pages
2176                  * can be recognized if we need to rollback on error later.
2177                  * But set_page_dirty so that memory pressure will swap rather
2178                  * than free the pages we are allocating (and SGP_CACHE pages
2179                  * might still be clean: we now need to mark those dirty too).
2180                  */
2181                 set_page_dirty(page);
2182                 unlock_page(page);
2183                 page_cache_release(page);
2184                 cond_resched();
2185         }
2186
2187         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2188                 i_size_write(inode, offset + len);
2189         inode->i_ctime = CURRENT_TIME;
2190 undone:
2191         spin_lock(&inode->i_lock);
2192         inode->i_private = NULL;
2193         spin_unlock(&inode->i_lock);
2194 out:
2195         mutex_unlock(&inode->i_mutex);
2196         return error;
2197 }
2198
2199 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2200 {
2201         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2202
2203         buf->f_type = TMPFS_MAGIC;
2204         buf->f_bsize = PAGE_CACHE_SIZE;
2205         buf->f_namelen = NAME_MAX;
2206         if (sbinfo->max_blocks) {
2207                 buf->f_blocks = sbinfo->max_blocks;
2208                 buf->f_bavail =
2209                 buf->f_bfree  = sbinfo->max_blocks -
2210                                 percpu_counter_sum(&sbinfo->used_blocks);
2211         }
2212         if (sbinfo->max_inodes) {
2213                 buf->f_files = sbinfo->max_inodes;
2214                 buf->f_ffree = sbinfo->free_inodes;
2215         }
2216         /* else leave those fields 0 like simple_statfs */
2217         return 0;
2218 }
2219
2220 /*
2221  * File creation. Allocate an inode, and we're done..
2222  */
2223 static int
2224 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2225 {
2226         struct inode *inode;
2227         int error = -ENOSPC;
2228
2229         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2230         if (inode) {
2231                 error = simple_acl_create(dir, inode);
2232                 if (error)
2233                         goto out_iput;
2234                 error = security_inode_init_security(inode, dir,
2235                                                      &dentry->d_name,
2236                                                      shmem_initxattrs, NULL);
2237                 if (error && error != -EOPNOTSUPP)
2238                         goto out_iput;
2239
2240                 error = 0;
2241                 dir->i_size += BOGO_DIRENT_SIZE;
2242                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2243                 d_instantiate(dentry, inode);
2244                 dget(dentry); /* Extra count - pin the dentry in core */
2245         }
2246         return error;
2247 out_iput:
2248         iput(inode);
2249         return error;
2250 }
2251
2252 static int
2253 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2254 {
2255         struct inode *inode;
2256         int error = -ENOSPC;
2257
2258         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2259         if (inode) {
2260                 error = security_inode_init_security(inode, dir,
2261                                                      NULL,
2262                                                      shmem_initxattrs, NULL);
2263                 if (error && error != -EOPNOTSUPP)
2264                         goto out_iput;
2265                 error = simple_acl_create(dir, inode);
2266                 if (error)
2267                         goto out_iput;
2268                 d_tmpfile(dentry, inode);
2269         }
2270         return error;
2271 out_iput:
2272         iput(inode);
2273         return error;
2274 }
2275
2276 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2277 {
2278         int error;
2279
2280         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2281                 return error;
2282         inc_nlink(dir);
2283         return 0;
2284 }
2285
2286 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2287                 bool excl)
2288 {
2289         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2290 }
2291
2292 /*
2293  * Link a file..
2294  */
2295 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2296 {
2297         struct inode *inode = d_inode(old_dentry);
2298         int ret = 0;
2299
2300         /*
2301          * No ordinary (disk based) filesystem counts links as inodes;
2302          * but each new link needs a new dentry, pinning lowmem, and
2303          * tmpfs dentries cannot be pruned until they are unlinked.
2304          * But if an O_TMPFILE file is linked into the tmpfs, the
2305          * first link must skip that, to get the accounting right.
2306          */
2307         if (inode->i_nlink) {
2308                 ret = shmem_reserve_inode(inode->i_sb);
2309                 if (ret)
2310                         goto out;
2311         }
2312
2313         dir->i_size += BOGO_DIRENT_SIZE;
2314         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2315         inc_nlink(inode);
2316         ihold(inode);   /* New dentry reference */
2317         dget(dentry);           /* Extra pinning count for the created dentry */
2318         d_instantiate(dentry, inode);
2319 out:
2320         return ret;
2321 }
2322
2323 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2324 {
2325         struct inode *inode = d_inode(dentry);
2326
2327         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2328                 shmem_free_inode(inode->i_sb);
2329
2330         dir->i_size -= BOGO_DIRENT_SIZE;
2331         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2332         drop_nlink(inode);
2333         dput(dentry);   /* Undo the count from "create" - this does all the work */
2334         return 0;
2335 }
2336
2337 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2338 {
2339         if (!simple_empty(dentry))
2340                 return -ENOTEMPTY;
2341
2342         drop_nlink(d_inode(dentry));
2343         drop_nlink(dir);
2344         return shmem_unlink(dir, dentry);
2345 }
2346
2347 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2348 {
2349         bool old_is_dir = d_is_dir(old_dentry);
2350         bool new_is_dir = d_is_dir(new_dentry);
2351
2352         if (old_dir != new_dir && old_is_dir != new_is_dir) {
2353                 if (old_is_dir) {
2354                         drop_nlink(old_dir);
2355                         inc_nlink(new_dir);
2356                 } else {
2357                         drop_nlink(new_dir);
2358                         inc_nlink(old_dir);
2359                 }
2360         }
2361         old_dir->i_ctime = old_dir->i_mtime =
2362         new_dir->i_ctime = new_dir->i_mtime =
2363         d_inode(old_dentry)->i_ctime =
2364         d_inode(new_dentry)->i_ctime = CURRENT_TIME;
2365
2366         return 0;
2367 }
2368
2369 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
2370 {
2371         struct dentry *whiteout;
2372         int error;
2373
2374         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2375         if (!whiteout)
2376                 return -ENOMEM;
2377
2378         error = shmem_mknod(old_dir, whiteout,
2379                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2380         dput(whiteout);
2381         if (error)
2382                 return error;
2383
2384         /*
2385          * Cheat and hash the whiteout while the old dentry is still in
2386          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
2387          *
2388          * d_lookup() will consistently find one of them at this point,
2389          * not sure which one, but that isn't even important.
2390          */
2391         d_rehash(whiteout);
2392         return 0;
2393 }
2394
2395 /*
2396  * The VFS layer already does all the dentry stuff for rename,
2397  * we just have to decrement the usage count for the target if
2398  * it exists so that the VFS layer correctly free's it when it
2399  * gets overwritten.
2400  */
2401 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
2402 {
2403         struct inode *inode = d_inode(old_dentry);
2404         int they_are_dirs = S_ISDIR(inode->i_mode);
2405
2406         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
2407                 return -EINVAL;
2408
2409         if (flags & RENAME_EXCHANGE)
2410                 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
2411
2412         if (!simple_empty(new_dentry))
2413                 return -ENOTEMPTY;
2414
2415         if (flags & RENAME_WHITEOUT) {
2416                 int error;
2417
2418                 error = shmem_whiteout(old_dir, old_dentry);
2419                 if (error)
2420                         return error;
2421         }
2422
2423         if (d_really_is_positive(new_dentry)) {
2424                 (void) shmem_unlink(new_dir, new_dentry);
2425                 if (they_are_dirs) {
2426                         drop_nlink(d_inode(new_dentry));
2427                         drop_nlink(old_dir);
2428                 }
2429         } else if (they_are_dirs) {
2430                 drop_nlink(old_dir);
2431                 inc_nlink(new_dir);
2432         }
2433
2434         old_dir->i_size -= BOGO_DIRENT_SIZE;
2435         new_dir->i_size += BOGO_DIRENT_SIZE;
2436         old_dir->i_ctime = old_dir->i_mtime =
2437         new_dir->i_ctime = new_dir->i_mtime =
2438         inode->i_ctime = CURRENT_TIME;
2439         return 0;
2440 }
2441
2442 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2443 {
2444         int error;
2445         int len;
2446         struct inode *inode;
2447         struct page *page;
2448         char *kaddr;
2449         struct shmem_inode_info *info;
2450
2451         len = strlen(symname) + 1;
2452         if (len > PAGE_CACHE_SIZE)
2453                 return -ENAMETOOLONG;
2454
2455         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2456         if (!inode)
2457                 return -ENOSPC;
2458
2459         error = security_inode_init_security(inode, dir, &dentry->d_name,
2460                                              shmem_initxattrs, NULL);
2461         if (error) {
2462                 if (error != -EOPNOTSUPP) {
2463                         iput(inode);
2464                         return error;
2465                 }
2466                 error = 0;
2467         }
2468
2469         info = SHMEM_I(inode);
2470         inode->i_size = len-1;
2471         if (len <= SHORT_SYMLINK_LEN) {
2472                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
2473                 if (!inode->i_link) {
2474                         iput(inode);
2475                         return -ENOMEM;
2476                 }
2477                 inode->i_op = &shmem_short_symlink_operations;
2478         } else {
2479                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2480                 if (error) {
2481                         iput(inode);
2482                         return error;
2483                 }
2484                 inode->i_mapping->a_ops = &shmem_aops;
2485                 inode->i_op = &shmem_symlink_inode_operations;
2486                 kaddr = kmap_atomic(page);
2487                 memcpy(kaddr, symname, len);
2488                 kunmap_atomic(kaddr);
2489                 SetPageUptodate(page);
2490                 set_page_dirty(page);
2491                 unlock_page(page);
2492                 page_cache_release(page);
2493         }
2494         dir->i_size += BOGO_DIRENT_SIZE;
2495         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2496         d_instantiate(dentry, inode);
2497         dget(dentry);
2498         return 0;
2499 }
2500
2501 static const char *shmem_follow_link(struct dentry *dentry, void **cookie)
2502 {
2503         struct page *page = NULL;
2504         int error = shmem_getpage(d_inode(dentry), 0, &page, SGP_READ, NULL);
2505         if (error)
2506                 return ERR_PTR(error);
2507         unlock_page(page);
2508         *cookie = page;
2509         return kmap(page);
2510 }
2511
2512 static void shmem_put_link(struct inode *unused, void *cookie)
2513 {
2514         struct page *page = cookie;
2515         kunmap(page);
2516         mark_page_accessed(page);
2517         page_cache_release(page);
2518 }
2519
2520 #ifdef CONFIG_TMPFS_XATTR
2521 /*
2522  * Superblocks without xattr inode operations may get some security.* xattr
2523  * support from the LSM "for free". As soon as we have any other xattrs
2524  * like ACLs, we also need to implement the security.* handlers at
2525  * filesystem level, though.
2526  */
2527
2528 /*
2529  * Callback for security_inode_init_security() for acquiring xattrs.
2530  */
2531 static int shmem_initxattrs(struct inode *inode,
2532                             const struct xattr *xattr_array,
2533                             void *fs_info)
2534 {
2535         struct shmem_inode_info *info = SHMEM_I(inode);
2536         const struct xattr *xattr;
2537         struct simple_xattr *new_xattr;
2538         size_t len;
2539
2540         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2541                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2542                 if (!new_xattr)
2543                         return -ENOMEM;
2544
2545                 len = strlen(xattr->name) + 1;
2546                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2547                                           GFP_KERNEL);
2548                 if (!new_xattr->name) {
2549                         kfree(new_xattr);
2550                         return -ENOMEM;
2551                 }
2552
2553                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2554                        XATTR_SECURITY_PREFIX_LEN);
2555                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2556                        xattr->name, len);
2557
2558                 simple_xattr_list_add(&info->xattrs, new_xattr);
2559         }
2560
2561         return 0;
2562 }
2563
2564 static const struct xattr_handler *shmem_xattr_handlers[] = {
2565 #ifdef CONFIG_TMPFS_POSIX_ACL
2566         &posix_acl_access_xattr_handler,
2567         &posix_acl_default_xattr_handler,
2568 #endif
2569         NULL
2570 };
2571
2572 static int shmem_xattr_validate(const char *name)
2573 {
2574         struct { const char *prefix; size_t len; } arr[] = {
2575                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2576                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2577         };
2578         int i;
2579
2580         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2581                 size_t preflen = arr[i].len;
2582                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2583                         if (!name[preflen])
2584                                 return -EINVAL;
2585                         return 0;
2586                 }
2587         }
2588         return -EOPNOTSUPP;
2589 }
2590
2591 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2592                               void *buffer, size_t size)
2593 {
2594         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2595         int err;
2596
2597         /*
2598          * If this is a request for a synthetic attribute in the system.*
2599          * namespace use the generic infrastructure to resolve a handler
2600          * for it via sb->s_xattr.
2601          */
2602         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2603                 return generic_getxattr(dentry, name, buffer, size);
2604
2605         err = shmem_xattr_validate(name);
2606         if (err)
2607                 return err;
2608
2609         return simple_xattr_get(&info->xattrs, name, buffer, size);
2610 }
2611
2612 static int shmem_setxattr(struct dentry *dentry, const char *name,
2613                           const void *value, size_t size, int flags)
2614 {
2615         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2616         int err;
2617
2618         /*
2619          * If this is a request for a synthetic attribute in the system.*
2620          * namespace use the generic infrastructure to resolve a handler
2621          * for it via sb->s_xattr.
2622          */
2623         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2624                 return generic_setxattr(dentry, name, value, size, flags);
2625
2626         err = shmem_xattr_validate(name);
2627         if (err)
2628                 return err;
2629
2630         return simple_xattr_set(&info->xattrs, name, value, size, flags);
2631 }
2632
2633 static int shmem_removexattr(struct dentry *dentry, const char *name)
2634 {
2635         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2636         int err;
2637
2638         /*
2639          * If this is a request for a synthetic attribute in the system.*
2640          * namespace use the generic infrastructure to resolve a handler
2641          * for it via sb->s_xattr.
2642          */
2643         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2644                 return generic_removexattr(dentry, name);
2645
2646         err = shmem_xattr_validate(name);
2647         if (err)
2648                 return err;
2649
2650         return simple_xattr_remove(&info->xattrs, name);
2651 }
2652
2653 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2654 {
2655         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2656         return simple_xattr_list(&info->xattrs, buffer, size);
2657 }
2658 #endif /* CONFIG_TMPFS_XATTR */
2659
2660 static const struct inode_operations shmem_short_symlink_operations = {
2661         .readlink       = generic_readlink,
2662         .follow_link    = simple_follow_link,
2663 #ifdef CONFIG_TMPFS_XATTR
2664         .setxattr       = shmem_setxattr,
2665         .getxattr       = shmem_getxattr,
2666         .listxattr      = shmem_listxattr,
2667         .removexattr    = shmem_removexattr,
2668 #endif
2669 };
2670
2671 static const struct inode_operations shmem_symlink_inode_operations = {
2672         .readlink       = generic_readlink,
2673         .follow_link    = shmem_follow_link,
2674         .put_link       = shmem_put_link,
2675 #ifdef CONFIG_TMPFS_XATTR
2676         .setxattr       = shmem_setxattr,
2677         .getxattr       = shmem_getxattr,
2678         .listxattr      = shmem_listxattr,
2679         .removexattr    = shmem_removexattr,
2680 #endif
2681 };
2682
2683 static struct dentry *shmem_get_parent(struct dentry *child)
2684 {
2685         return ERR_PTR(-ESTALE);
2686 }
2687
2688 static int shmem_match(struct inode *ino, void *vfh)
2689 {
2690         __u32 *fh = vfh;
2691         __u64 inum = fh[2];
2692         inum = (inum << 32) | fh[1];
2693         return ino->i_ino == inum && fh[0] == ino->i_generation;
2694 }
2695
2696 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2697                 struct fid *fid, int fh_len, int fh_type)
2698 {
2699         struct inode *inode;
2700         struct dentry *dentry = NULL;
2701         u64 inum;
2702
2703         if (fh_len < 3)
2704                 return NULL;
2705
2706         inum = fid->raw[2];
2707         inum = (inum << 32) | fid->raw[1];
2708
2709         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2710                         shmem_match, fid->raw);
2711         if (inode) {
2712                 dentry = d_find_alias(inode);
2713                 iput(inode);
2714         }
2715
2716         return dentry;
2717 }
2718
2719 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2720                                 struct inode *parent)
2721 {
2722         if (*len < 3) {
2723                 *len = 3;
2724                 return FILEID_INVALID;
2725         }
2726
2727         if (inode_unhashed(inode)) {
2728                 /* Unfortunately insert_inode_hash is not idempotent,
2729                  * so as we hash inodes here rather than at creation
2730                  * time, we need a lock to ensure we only try
2731                  * to do it once
2732                  */
2733                 static DEFINE_SPINLOCK(lock);
2734                 spin_lock(&lock);
2735                 if (inode_unhashed(inode))
2736                         __insert_inode_hash(inode,
2737                                             inode->i_ino + inode->i_generation);
2738                 spin_unlock(&lock);
2739         }
2740
2741         fh[0] = inode->i_generation;
2742         fh[1] = inode->i_ino;
2743         fh[2] = ((__u64)inode->i_ino) >> 32;
2744
2745         *len = 3;
2746         return 1;
2747 }
2748
2749 static const struct export_operations shmem_export_ops = {
2750         .get_parent     = shmem_get_parent,
2751         .encode_fh      = shmem_encode_fh,
2752         .fh_to_dentry   = shmem_fh_to_dentry,
2753 };
2754
2755 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2756                                bool remount)
2757 {
2758         char *this_char, *value, *rest;
2759         struct mempolicy *mpol = NULL;
2760         uid_t uid;
2761         gid_t gid;
2762
2763         while (options != NULL) {
2764                 this_char = options;
2765                 for (;;) {
2766                         /*
2767                          * NUL-terminate this option: unfortunately,
2768                          * mount options form a comma-separated list,
2769                          * but mpol's nodelist may also contain commas.
2770                          */
2771                         options = strchr(options, ',');
2772                         if (options == NULL)
2773                                 break;
2774                         options++;
2775                         if (!isdigit(*options)) {
2776                                 options[-1] = '\0';
2777                                 break;
2778                         }
2779                 }
2780                 if (!*this_char)
2781                         continue;
2782                 if ((value = strchr(this_char,'=')) != NULL) {
2783                         *value++ = 0;
2784                 } else {
2785                         printk(KERN_ERR
2786                             "tmpfs: No value for mount option '%s'\n",
2787                             this_char);
2788                         goto error;
2789                 }
2790
2791                 if (!strcmp(this_char,"size")) {
2792                         unsigned long long size;
2793                         size = memparse(value,&rest);
2794                         if (*rest == '%') {
2795                                 size <<= PAGE_SHIFT;
2796                                 size *= totalram_pages;
2797                                 do_div(size, 100);
2798                                 rest++;
2799                         }
2800                         if (*rest)
2801                                 goto bad_val;
2802                         sbinfo->max_blocks =
2803                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2804                 } else if (!strcmp(this_char,"nr_blocks")) {
2805                         sbinfo->max_blocks = memparse(value, &rest);
2806                         if (*rest)
2807                                 goto bad_val;
2808                 } else if (!strcmp(this_char,"nr_inodes")) {
2809                         sbinfo->max_inodes = memparse(value, &rest);
2810                         if (*rest)
2811                                 goto bad_val;
2812                 } else if (!strcmp(this_char,"mode")) {
2813                         if (remount)
2814                                 continue;
2815                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2816                         if (*rest)
2817                                 goto bad_val;
2818                 } else if (!strcmp(this_char,"uid")) {
2819                         if (remount)
2820                                 continue;
2821                         uid = simple_strtoul(value, &rest, 0);
2822                         if (*rest)
2823                                 goto bad_val;
2824                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2825                         if (!uid_valid(sbinfo->uid))
2826                                 goto bad_val;
2827                 } else if (!strcmp(this_char,"gid")) {
2828                         if (remount)
2829                                 continue;
2830                         gid = simple_strtoul(value, &rest, 0);
2831                         if (*rest)
2832                                 goto bad_val;
2833                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2834                         if (!gid_valid(sbinfo->gid))
2835                                 goto bad_val;
2836                 } else if (!strcmp(this_char,"mpol")) {
2837                         mpol_put(mpol);
2838                         mpol = NULL;
2839                         if (mpol_parse_str(value, &mpol))
2840                                 goto bad_val;
2841                 } else {
2842                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2843                                this_char);
2844                         goto error;
2845                 }
2846         }
2847         sbinfo->mpol = mpol;
2848         return 0;
2849
2850 bad_val:
2851         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2852                value, this_char);
2853 error:
2854         mpol_put(mpol);
2855         return 1;
2856
2857 }
2858
2859 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2860 {
2861         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2862         struct shmem_sb_info config = *sbinfo;
2863         unsigned long inodes;
2864         int error = -EINVAL;
2865
2866         config.mpol = NULL;
2867         if (shmem_parse_options(data, &config, true))
2868                 return error;
2869
2870         spin_lock(&sbinfo->stat_lock);
2871         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2872         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2873                 goto out;
2874         if (config.max_inodes < inodes)
2875                 goto out;
2876         /*
2877          * Those tests disallow limited->unlimited while any are in use;
2878          * but we must separately disallow unlimited->limited, because
2879          * in that case we have no record of how much is already in use.
2880          */
2881         if (config.max_blocks && !sbinfo->max_blocks)
2882                 goto out;
2883         if (config.max_inodes && !sbinfo->max_inodes)
2884                 goto out;
2885
2886         error = 0;
2887         sbinfo->max_blocks  = config.max_blocks;
2888         sbinfo->max_inodes  = config.max_inodes;
2889         sbinfo->free_inodes = config.max_inodes - inodes;
2890
2891         /*
2892          * Preserve previous mempolicy unless mpol remount option was specified.
2893          */
2894         if (config.mpol) {
2895                 mpol_put(sbinfo->mpol);
2896                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
2897         }
2898 out:
2899         spin_unlock(&sbinfo->stat_lock);
2900         return error;
2901 }
2902
2903 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2904 {
2905         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2906
2907         if (sbinfo->max_blocks != shmem_default_max_blocks())
2908                 seq_printf(seq, ",size=%luk",
2909                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2910         if (sbinfo->max_inodes != shmem_default_max_inodes())
2911                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2912         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2913                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2914         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2915                 seq_printf(seq, ",uid=%u",
2916                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2917         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2918                 seq_printf(seq, ",gid=%u",
2919                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2920         shmem_show_mpol(seq, sbinfo->mpol);
2921         return 0;
2922 }
2923
2924 #define MFD_NAME_PREFIX "memfd:"
2925 #define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
2926 #define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
2927
2928 #define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING)
2929
2930 SYSCALL_DEFINE2(memfd_create,
2931                 const char __user *, uname,
2932                 unsigned int, flags)
2933 {
2934         struct shmem_inode_info *info;
2935         struct file *file;
2936         int fd, error;
2937         char *name;
2938         long len;
2939
2940         if (flags & ~(unsigned int)MFD_ALL_FLAGS)
2941                 return -EINVAL;
2942
2943         /* length includes terminating zero */
2944         len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
2945         if (len <= 0)
2946                 return -EFAULT;
2947         if (len > MFD_NAME_MAX_LEN + 1)
2948                 return -EINVAL;
2949
2950         name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_TEMPORARY);
2951         if (!name)
2952                 return -ENOMEM;
2953
2954         strcpy(name, MFD_NAME_PREFIX);
2955         if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
2956                 error = -EFAULT;
2957                 goto err_name;
2958         }
2959
2960         /* terminating-zero may have changed after strnlen_user() returned */
2961         if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
2962                 error = -EFAULT;
2963                 goto err_name;
2964         }
2965
2966         fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
2967         if (fd < 0) {
2968                 error = fd;
2969                 goto err_name;
2970         }
2971
2972         file = shmem_file_setup(name, 0, VM_NORESERVE);
2973         if (IS_ERR(file)) {
2974                 error = PTR_ERR(file);
2975                 goto err_fd;
2976         }
2977         info = SHMEM_I(file_inode(file));
2978         file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
2979         file->f_flags |= O_RDWR | O_LARGEFILE;
2980         if (flags & MFD_ALLOW_SEALING)
2981                 info->seals &= ~F_SEAL_SEAL;
2982
2983         fd_install(fd, file);
2984         kfree(name);
2985         return fd;
2986
2987 err_fd:
2988         put_unused_fd(fd);
2989 err_name:
2990         kfree(name);
2991         return error;
2992 }
2993
2994 #endif /* CONFIG_TMPFS */
2995
2996 static void shmem_put_super(struct super_block *sb)
2997 {
2998         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2999
3000         percpu_counter_destroy(&sbinfo->used_blocks);
3001         mpol_put(sbinfo->mpol);
3002         kfree(sbinfo);
3003         sb->s_fs_info = NULL;
3004 }
3005
3006 int shmem_fill_super(struct super_block *sb, void *data, int silent)
3007 {
3008         struct inode *inode;
3009         struct shmem_sb_info *sbinfo;
3010         int err = -ENOMEM;
3011
3012         /* Round up to L1_CACHE_BYTES to resist false sharing */
3013         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3014                                 L1_CACHE_BYTES), GFP_KERNEL);
3015         if (!sbinfo)
3016                 return -ENOMEM;
3017
3018         sbinfo->mode = S_IRWXUGO | S_ISVTX;
3019         sbinfo->uid = current_fsuid();
3020         sbinfo->gid = current_fsgid();
3021         sb->s_fs_info = sbinfo;
3022
3023 #ifdef CONFIG_TMPFS
3024         /*
3025          * Per default we only allow half of the physical ram per
3026          * tmpfs instance, limiting inodes to one per page of lowmem;
3027          * but the internal instance is left unlimited.
3028          */
3029         if (!(sb->s_flags & MS_KERNMOUNT)) {
3030                 sbinfo->max_blocks = shmem_default_max_blocks();
3031                 sbinfo->max_inodes = shmem_default_max_inodes();
3032                 if (shmem_parse_options(data, sbinfo, false)) {
3033                         err = -EINVAL;
3034                         goto failed;
3035                 }
3036         } else {
3037                 sb->s_flags |= MS_NOUSER;
3038         }
3039         sb->s_export_op = &shmem_export_ops;
3040         sb->s_flags |= MS_NOSEC;
3041 #else
3042         sb->s_flags |= MS_NOUSER;
3043 #endif
3044
3045         spin_lock_init(&sbinfo->stat_lock);
3046         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3047                 goto failed;
3048         sbinfo->free_inodes = sbinfo->max_inodes;
3049
3050         sb->s_maxbytes = MAX_LFS_FILESIZE;
3051         sb->s_blocksize = PAGE_CACHE_SIZE;
3052         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
3053         sb->s_magic = TMPFS_MAGIC;
3054         sb->s_op = &shmem_ops;
3055         sb->s_time_gran = 1;
3056 #ifdef CONFIG_TMPFS_XATTR
3057         sb->s_xattr = shmem_xattr_handlers;
3058 #endif
3059 #ifdef CONFIG_TMPFS_POSIX_ACL
3060         sb->s_flags |= MS_POSIXACL;
3061 #endif
3062
3063         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3064         if (!inode)
3065                 goto failed;
3066         inode->i_uid = sbinfo->uid;
3067         inode->i_gid = sbinfo->gid;
3068         sb->s_root = d_make_root(inode);
3069         if (!sb->s_root)
3070                 goto failed;
3071         return 0;
3072
3073 failed:
3074         shmem_put_super(sb);
3075         return err;
3076 }
3077
3078 static struct kmem_cache *shmem_inode_cachep;
3079
3080 static struct inode *shmem_alloc_inode(struct super_block *sb)
3081 {
3082         struct shmem_inode_info *info;
3083         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3084         if (!info)
3085                 return NULL;
3086         return &info->vfs_inode;
3087 }
3088
3089 static void shmem_destroy_callback(struct rcu_head *head)
3090 {
3091         struct inode *inode = container_of(head, struct inode, i_rcu);
3092         kfree(inode->i_link);
3093         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3094 }
3095
3096 static void shmem_destroy_inode(struct inode *inode)
3097 {
3098         if (S_ISREG(inode->i_mode))
3099                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3100         call_rcu(&inode->i_rcu, shmem_destroy_callback);
3101 }
3102
3103 static void shmem_init_inode(void *foo)
3104 {
3105         struct shmem_inode_info *info = foo;
3106         inode_init_once(&info->vfs_inode);
3107 }
3108
3109 static int shmem_init_inodecache(void)
3110 {
3111         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3112                                 sizeof(struct shmem_inode_info),
3113                                 0, SLAB_PANIC, shmem_init_inode);
3114         return 0;
3115 }
3116
3117 static void shmem_destroy_inodecache(void)
3118 {
3119         kmem_cache_destroy(shmem_inode_cachep);
3120 }
3121
3122 static const struct address_space_operations shmem_aops = {
3123         .writepage      = shmem_writepage,
3124         .set_page_dirty = __set_page_dirty_no_writeback,
3125 #ifdef CONFIG_TMPFS
3126         .write_begin    = shmem_write_begin,
3127         .write_end      = shmem_write_end,
3128 #endif
3129 #ifdef CONFIG_MIGRATION
3130         .migratepage    = migrate_page,
3131 #endif
3132         .error_remove_page = generic_error_remove_page,
3133 };
3134
3135 static const struct file_operations shmem_file_operations = {
3136         .mmap           = shmem_mmap,
3137 #ifdef CONFIG_TMPFS
3138         .llseek         = shmem_file_llseek,
3139         .read_iter      = shmem_file_read_iter,
3140         .write_iter     = generic_file_write_iter,
3141         .fsync          = noop_fsync,
3142         .splice_read    = shmem_file_splice_read,
3143         .splice_write   = iter_file_splice_write,
3144         .fallocate      = shmem_fallocate,
3145 #endif
3146 };
3147
3148 static const struct inode_operations shmem_inode_operations = {
3149         .getattr        = shmem_getattr,
3150         .setattr        = shmem_setattr,
3151 #ifdef CONFIG_TMPFS_XATTR
3152         .setxattr       = shmem_setxattr,
3153         .getxattr       = shmem_getxattr,
3154         .listxattr      = shmem_listxattr,
3155         .removexattr    = shmem_removexattr,
3156         .set_acl        = simple_set_acl,
3157 #endif
3158 };
3159
3160 static const struct inode_operations shmem_dir_inode_operations = {
3161 #ifdef CONFIG_TMPFS
3162         .create         = shmem_create,
3163         .lookup         = simple_lookup,
3164         .link           = shmem_link,
3165         .unlink         = shmem_unlink,
3166         .symlink        = shmem_symlink,
3167         .mkdir          = shmem_mkdir,
3168         .rmdir          = shmem_rmdir,
3169         .mknod          = shmem_mknod,
3170         .rename2        = shmem_rename2,
3171         .tmpfile        = shmem_tmpfile,
3172 #endif
3173 #ifdef CONFIG_TMPFS_XATTR
3174         .setxattr       = shmem_setxattr,
3175         .getxattr       = shmem_getxattr,
3176         .listxattr      = shmem_listxattr,
3177         .removexattr    = shmem_removexattr,
3178 #endif
3179 #ifdef CONFIG_TMPFS_POSIX_ACL
3180         .setattr        = shmem_setattr,
3181         .set_acl        = simple_set_acl,
3182 #endif
3183 };
3184
3185 static const struct inode_operations shmem_special_inode_operations = {
3186 #ifdef CONFIG_TMPFS_XATTR
3187         .setxattr       = shmem_setxattr,
3188         .getxattr       = shmem_getxattr,
3189         .listxattr      = shmem_listxattr,
3190         .removexattr    = shmem_removexattr,
3191 #endif
3192 #ifdef CONFIG_TMPFS_POSIX_ACL
3193         .setattr        = shmem_setattr,
3194         .set_acl        = simple_set_acl,
3195 #endif
3196 };
3197
3198 static const struct super_operations shmem_ops = {
3199         .alloc_inode    = shmem_alloc_inode,
3200         .destroy_inode  = shmem_destroy_inode,
3201 #ifdef CONFIG_TMPFS
3202         .statfs         = shmem_statfs,
3203         .remount_fs     = shmem_remount_fs,
3204         .show_options   = shmem_show_options,
3205 #endif
3206         .evict_inode    = shmem_evict_inode,
3207         .drop_inode     = generic_delete_inode,
3208         .put_super      = shmem_put_super,
3209 };
3210
3211 static const struct vm_operations_struct shmem_vm_ops = {
3212         .fault          = shmem_fault,
3213         .map_pages      = filemap_map_pages,
3214 #ifdef CONFIG_NUMA
3215         .set_policy     = shmem_set_policy,
3216         .get_policy     = shmem_get_policy,
3217 #endif
3218 };
3219
3220 static struct dentry *shmem_mount(struct file_system_type *fs_type,
3221         int flags, const char *dev_name, void *data)
3222 {
3223         return mount_nodev(fs_type, flags, data, shmem_fill_super);
3224 }
3225
3226 static struct file_system_type shmem_fs_type = {
3227         .owner          = THIS_MODULE,
3228         .name           = "tmpfs",
3229         .mount          = shmem_mount,
3230         .kill_sb        = kill_litter_super,
3231         .fs_flags       = FS_USERNS_MOUNT,
3232 };
3233
3234 int __init shmem_init(void)
3235 {
3236         int error;
3237
3238         /* If rootfs called this, don't re-init */
3239         if (shmem_inode_cachep)
3240                 return 0;
3241
3242         error = shmem_init_inodecache();
3243         if (error)
3244                 goto out3;
3245
3246         error = register_filesystem(&shmem_fs_type);
3247         if (error) {
3248                 printk(KERN_ERR "Could not register tmpfs\n");
3249                 goto out2;
3250         }
3251
3252         shm_mnt = kern_mount(&shmem_fs_type);
3253         if (IS_ERR(shm_mnt)) {
3254                 error = PTR_ERR(shm_mnt);
3255                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
3256                 goto out1;
3257         }
3258         return 0;
3259
3260 out1:
3261         unregister_filesystem(&shmem_fs_type);
3262 out2:
3263         shmem_destroy_inodecache();
3264 out3:
3265         shm_mnt = ERR_PTR(error);
3266         return error;
3267 }
3268
3269 #else /* !CONFIG_SHMEM */
3270
3271 /*
3272  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3273  *
3274  * This is intended for small system where the benefits of the full
3275  * shmem code (swap-backed and resource-limited) are outweighed by
3276  * their complexity. On systems without swap this code should be
3277  * effectively equivalent, but much lighter weight.
3278  */
3279
3280 static struct file_system_type shmem_fs_type = {
3281         .name           = "tmpfs",
3282         .mount          = ramfs_mount,
3283         .kill_sb        = kill_litter_super,
3284         .fs_flags       = FS_USERNS_MOUNT,
3285 };
3286
3287 int __init shmem_init(void)
3288 {
3289         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3290
3291         shm_mnt = kern_mount(&shmem_fs_type);
3292         BUG_ON(IS_ERR(shm_mnt));
3293
3294         return 0;
3295 }
3296
3297 int shmem_unuse(swp_entry_t swap, struct page *page)
3298 {
3299         return 0;
3300 }
3301
3302 int shmem_lock(struct file *file, int lock, struct user_struct *user)
3303 {
3304         return 0;
3305 }
3306
3307 void shmem_unlock_mapping(struct address_space *mapping)
3308 {
3309 }
3310
3311 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3312 {
3313         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3314 }
3315 EXPORT_SYMBOL_GPL(shmem_truncate_range);
3316
3317 #define shmem_vm_ops                            generic_file_vm_ops
3318 #define shmem_file_operations                   ramfs_file_operations
3319 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
3320 #define shmem_acct_size(flags, size)            0
3321 #define shmem_unacct_size(flags, size)          do {} while (0)
3322
3323 #endif /* CONFIG_SHMEM */
3324
3325 /* common code */
3326
3327 static struct dentry_operations anon_ops = {
3328         .d_dname = simple_dname
3329 };
3330
3331 static struct file *__shmem_file_setup(const char *name, loff_t size,
3332                                        unsigned long flags, unsigned int i_flags)
3333 {
3334         struct file *res;
3335         struct inode *inode;
3336         struct path path;
3337         struct super_block *sb;
3338         struct qstr this;
3339
3340         if (IS_ERR(shm_mnt))
3341                 return ERR_CAST(shm_mnt);
3342
3343         if (size < 0 || size > MAX_LFS_FILESIZE)
3344                 return ERR_PTR(-EINVAL);
3345
3346         if (shmem_acct_size(flags, size))
3347                 return ERR_PTR(-ENOMEM);
3348
3349         res = ERR_PTR(-ENOMEM);
3350         this.name = name;
3351         this.len = strlen(name);
3352         this.hash = 0; /* will go */
3353         sb = shm_mnt->mnt_sb;
3354         path.mnt = mntget(shm_mnt);
3355         path.dentry = d_alloc_pseudo(sb, &this);
3356         if (!path.dentry)
3357                 goto put_memory;
3358         d_set_d_op(path.dentry, &anon_ops);
3359
3360         res = ERR_PTR(-ENOSPC);
3361         inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
3362         if (!inode)
3363                 goto put_memory;
3364
3365         inode->i_flags |= i_flags;
3366         d_instantiate(path.dentry, inode);
3367         inode->i_size = size;
3368         clear_nlink(inode);     /* It is unlinked */
3369         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
3370         if (IS_ERR(res))
3371                 goto put_path;
3372
3373         res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
3374                   &shmem_file_operations);
3375         if (IS_ERR(res))
3376                 goto put_path;
3377
3378         return res;
3379
3380 put_memory:
3381         shmem_unacct_size(flags, size);
3382 put_path:
3383         path_put(&path);
3384         return res;
3385 }
3386
3387 /**
3388  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
3389  *      kernel internal.  There will be NO LSM permission checks against the
3390  *      underlying inode.  So users of this interface must do LSM checks at a
3391  *      higher layer.  The users are the big_key and shm implementations.  LSM
3392  *      checks are provided at the key or shm level rather than the inode.
3393  * @name: name for dentry (to be seen in /proc/<pid>/maps
3394  * @size: size to be set for the file
3395  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3396  */
3397 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
3398 {
3399         return __shmem_file_setup(name, size, flags, S_PRIVATE);
3400 }
3401
3402 /**
3403  * shmem_file_setup - get an unlinked file living in tmpfs
3404  * @name: name for dentry (to be seen in /proc/<pid>/maps
3405  * @size: size to be set for the file
3406  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3407  */
3408 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
3409 {
3410         return __shmem_file_setup(name, size, flags, 0);
3411 }
3412 EXPORT_SYMBOL_GPL(shmem_file_setup);
3413
3414 /**
3415  * shmem_zero_setup - setup a shared anonymous mapping
3416  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3417  */
3418 int shmem_zero_setup(struct vm_area_struct *vma)
3419 {
3420         struct file *file;
3421         loff_t size = vma->vm_end - vma->vm_start;
3422
3423         /*
3424          * Cloning a new file under mmap_sem leads to a lock ordering conflict
3425          * between XFS directory reading and selinux: since this file is only
3426          * accessible to the user through its mapping, use S_PRIVATE flag to
3427          * bypass file security, in the same way as shmem_kernel_file_setup().
3428          */
3429         file = __shmem_file_setup("dev/zero", size, vma->vm_flags, S_PRIVATE);
3430         if (IS_ERR(file))
3431                 return PTR_ERR(file);
3432
3433         if (vma->vm_file)
3434                 fput(vma->vm_file);
3435         vma->vm_file = file;
3436         vma->vm_ops = &shmem_vm_ops;
3437         return 0;
3438 }
3439
3440 /**
3441  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3442  * @mapping:    the page's address_space
3443  * @index:      the page index
3444  * @gfp:        the page allocator flags to use if allocating
3445  *
3446  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3447  * with any new page allocations done using the specified allocation flags.
3448  * But read_cache_page_gfp() uses the ->readpage() method: which does not
3449  * suit tmpfs, since it may have pages in swapcache, and needs to find those
3450  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3451  *
3452  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
3453  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
3454  */
3455 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3456                                          pgoff_t index, gfp_t gfp)
3457 {
3458 #ifdef CONFIG_SHMEM
3459         struct inode *inode = mapping->host;
3460         struct page *page;
3461         int error;
3462
3463         BUG_ON(mapping->a_ops != &shmem_aops);
3464         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3465         if (error)
3466                 page = ERR_PTR(error);
3467         else
3468                 unlock_page(page);
3469         return page;
3470 #else
3471         /*
3472          * The tiny !SHMEM case uses ramfs without swap
3473          */
3474         return read_cache_page_gfp(mapping, index, gfp);
3475 #endif
3476 }
3477 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);