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