mm/shmem.c

   1 /*
   2  * Resizable virtual memory filesystem for Linux.
   3  *
   4  * Copyright (C) 2000 Linus Torvalds.
   5  *               2000 Transmeta Corp.
   6  *               2000-2001 Christoph Rohland
   7  *               2000-2001 SAP AG
   8  *               2002 Red Hat Inc.
   9  * Copyright (C) 2002-2011 Hugh Dickins.
  10  * Copyright (C) 2011 Google Inc.
  11  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  12  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  13  *
  14  * Extended attribute support for tmpfs:
  15  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
  16  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
  17  *
  18  * tiny-shmem:
  19  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
  20  *
  21  * This file is released under the GPL.
  22  */
  23
  24 #include <linux/fs.h>
  25 #include <linux/init.h>
  26 #include <linux/vfs.h>
  27 #include <linux/mount.h>
  28 #include <linux/ramfs.h>
  29 #include <linux/pagemap.h>
  30 #include <linux/file.h>
  31 #include <linux/mm.h>
  32 #include <linux/export.h>
  33 #include <linux/swap.h>
  34 #include <linux/uio.h>
  35
  36 static struct vfsmount *shm_mnt;
  37
  38 #ifdef CONFIG_SHMEM
  39 /*
  40  * This virtual memory filesystem is heavily based on the ramfs. It
  41  * extends ramfs by the ability to use swap and honor resource limits
  42  * which makes it a completely usable filesystem.
  43  */
  44
  45 #include <linux/xattr.h>
  46 #include <linux/exportfs.h>
  47 #include <linux/posix_acl.h>
  48 #include <linux/posix_acl_xattr.h>
  49 #include <linux/mman.h>
  50 #include <linux/string.h>
  51 #include <linux/slab.h>
  52 #include <linux/backing-dev.h>
  53 #include <linux/shmem_fs.h>
  54 #include <linux/writeback.h>
  55 #include <linux/blkdev.h>
  56 #include <linux/pagevec.h>
  57 #include <linux/percpu_counter.h>
  58 #include <linux/falloc.h>
  59 #include <linux/splice.h>
  60 #include <linux/security.h>
  61 #include <linux/swapops.h>
  62 #include <linux/mempolicy.h>
  63 #include <linux/namei.h>
  64 #include <linux/ctype.h>
  65 #include <linux/migrate.h>
  66 #include <linux/highmem.h>
  67 #include <linux/seq_file.h>
  68 #include <linux/magic.h>
  69 #include <linux/syscalls.h>
  70 #include <linux/fcntl.h>
  71 #include <uapi/linux/memfd.h>
  72
  73 #include <asm/uaccess.h>
  74 #include <asm/pgtable.h>
  75
  76 #include "internal.h"
  77
  78 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  79 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  80
  81 /* Pretend that each entry is of this size in directory's i_size */
  82 #define BOGO_DIRENT_SIZE 20
  83
  84 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  85 #define SHORT_SYMLINK_LEN 128
  86
  87 /*
  88  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
  89  * inode->i_private (with i_mutex making sure that it has only one user at
  90  * a time): we would prefer not to enlarge the shmem inode just for that.
  91  */
  92 struct shmem_falloc {
  93         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
  94         pgoff_t start;          /* start of range currently being fallocated */
  95         pgoff_t next;           /* the next page offset to be fallocated */
  96         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
  97         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
  98 };
  99
 100 /* Flag allocation requirements to shmem_getpage */
 101 enum sgp_type {
 102         SGP_READ,       /* don't exceed i_size, don't allocate page */
 103         SGP_CACHE,      /* don't exceed i_size, may allocate page */
 104         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
 105         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
 106         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
 107 };
 108
 109 #ifdef CONFIG_TMPFS
 110 static unsigned long shmem_default_max_blocks(void)
 111 {
 112         return totalram_pages / 2;
 113 }
 114
 115 static unsigned long shmem_default_max_inodes(void)
 116 {
 117         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 118 }
 119 #endif
 120
 121 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
 122 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 123                                 struct shmem_inode_info *info, pgoff_t index);
 124 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 125         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
 126
 127 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
 128         struct page **pagep, enum sgp_type sgp, int *fault_type)
 129 {
 130         return shmem_getpage_gfp(inode, index, pagep, sgp,
 131                         mapping_gfp_mask(inode->i_mapping), fault_type);
 132 }
 133
 134 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 135 {
 136         return sb->s_fs_info;
 137 }
 138
 139 /*
 140  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 141  * for shared memory and for shared anonymous (/dev/zero) mappings
 142  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 143  * consistent with the pre-accounting of private mappings ...
 144  */
 145 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 146 {
 147         return (flags & VM_NORESERVE) ?
 148                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 149 }
 150
 151 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 152 {
 153         if (!(flags & VM_NORESERVE))
 154                 vm_unacct_memory(VM_ACCT(size));
 155 }
 156
 157 static inline int shmem_reacct_size(unsigned long flags,
 158                 loff_t oldsize, loff_t newsize)
 159 {
 160         if (!(flags & VM_NORESERVE)) {
 161                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
 162                         return security_vm_enough_memory_mm(current->mm,
 163                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
 164                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
 165                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
 166         }
 167         return 0;
 168 }
 169
 170 /*
 171  * ... whereas tmpfs objects are accounted incrementally as
 172  * pages are allocated, in order to allow huge sparse files.
 173  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 174  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 175  */
 176 static inline int shmem_acct_block(unsigned long flags)
 177 {
 178         return (flags & VM_NORESERVE) ?
 179                 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 180 }
 181
 182 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 183 {
 184         if (flags & VM_NORESERVE)
 185                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 186 }
 187
 188 static const struct super_operations shmem_ops;
 189 static const struct address_space_operations shmem_aops;
 190 static const struct file_operations shmem_file_operations;
 191 static const struct inode_operations shmem_inode_operations;
 192 static const struct inode_operations shmem_dir_inode_operations;
 193 static const struct inode_operations shmem_special_inode_operations;
 194 static const struct vm_operations_struct shmem_vm_ops;
 195
 196 static LIST_HEAD(shmem_swaplist);
 197 static DEFINE_MUTEX(shmem_swaplist_mutex);
 198
 199 static int shmem_reserve_inode(struct super_block *sb)
 200 {
 201         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 202         if (sbinfo->max_inodes) {
 203                 spin_lock(&sbinfo->stat_lock);
 204                 if (!sbinfo->free_inodes) {
 205                         spin_unlock(&sbinfo->stat_lock);
 206                         return -ENOSPC;
 207                 }
 208                 sbinfo->free_inodes--;
 209                 spin_unlock(&sbinfo->stat_lock);
 210         }
 211         return 0;
 212 }
 213
 214 static void shmem_free_inode(struct super_block *sb)
 215 {
 216         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 217         if (sbinfo->max_inodes) {
 218                 spin_lock(&sbinfo->stat_lock);
 219                 sbinfo->free_inodes++;
 220                 spin_unlock(&sbinfo->stat_lock);
 221         }
 222 }
 223
 224 /**
 225  * shmem_recalc_inode - recalculate the block usage of an inode
 226  * @inode: inode to recalc
 227  *
 228  * We have to calculate the free blocks since the mm can drop
 229  * undirtied hole pages behind our back.
 230  *
 231  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 232  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 233  *
 234  * It has to be called with the spinlock held.
 235  */
 236 static void shmem_recalc_inode(struct inode *inode)
 237 {
 238         struct shmem_inode_info *info = SHMEM_I(inode);
 239         long freed;
 240
 241         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 242         if (freed > 0) {
 243                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 244                 if (sbinfo->max_blocks)
 245                         percpu_counter_add(&sbinfo->used_blocks, -freed);
 246                 info->alloced -= freed;
 247                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 248                 shmem_unacct_blocks(info->flags, freed);
 249         }
 250 }
 251
 252 /*
 253  * Replace item expected in radix tree by a new item, while holding tree lock.
 254  */
 255 static int shmem_radix_tree_replace(struct address_space *mapping,
 256                         pgoff_t index, void *expected, void *replacement)
 257 {
 258         void **pslot;
 259         void *item;
 260
 261         VM_BUG_ON(!expected);
 262         VM_BUG_ON(!replacement);
 263         pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
 264         if (!pslot)
 265                 return -ENOENT;
 266         item = radix_tree_deref_slot_protected(pslot, &mapping->tree_lock);
 267         if (item != expected)
 268                 return -ENOENT;
 269         radix_tree_replace_slot(pslot, replacement);
 270         return 0;
 271 }
 272
 273 /*
 274  * Sometimes, before we decide whether to proceed or to fail, we must check
 275  * that an entry was not already brought back from swap by a racing thread.
 276  *
 277  * Checking page is not enough: by the time a SwapCache page is locked, it
 278  * might be reused, and again be SwapCache, using the same swap as before.
 279  */
 280 static bool shmem_confirm_swap(struct address_space *mapping,
 281                                pgoff_t index, swp_entry_t swap)
 282 {
 283         void *item;
 284
 285         rcu_read_lock();
 286         item = radix_tree_lookup(&mapping->page_tree, index);
 287         rcu_read_unlock();
 288         return item == swp_to_radix_entry(swap);
 289 }
 290
 291 /*
 292  * Like add_to_page_cache_locked, but error if expected item has gone.
 293  */
 294 static int shmem_add_to_page_cache(struct page *page,
 295                                    struct address_space *mapping,
 296                                    pgoff_t index, void *expected)
 297 {
 298         int error;
 299
 300         VM_BUG_ON_PAGE(!PageLocked(page), page);
 301         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 302
 303         page_cache_get(page);
 304         page->mapping = mapping;
 305         page->index = index;
 306
 307         spin_lock_irq(&mapping->tree_lock);
 308         if (!expected)
 309                 error = radix_tree_insert(&mapping->page_tree, index, page);
 310         else
 311                 error = shmem_radix_tree_replace(mapping, index, expected,
 312                                                                  page);
 313         if (!error) {
 314                 mapping->nrpages++;
 315                 __inc_zone_page_state(page, NR_FILE_PAGES);
 316                 __inc_zone_page_state(page, NR_SHMEM);
 317                 spin_unlock_irq(&mapping->tree_lock);
 318         } else {
 319                 page->mapping = NULL;
 320                 spin_unlock_irq(&mapping->tree_lock);
 321                 page_cache_release(page);
 322         }
 323         return error;
 324 }
 325
 326 /*
 327  * Like delete_from_page_cache, but substitutes swap for page.
 328  */
 329 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 330 {
 331         struct address_space *mapping = page->mapping;
 332         int error;
 333
 334         spin_lock_irq(&mapping->tree_lock);
 335         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
 336         page->mapping = NULL;
 337         mapping->nrpages--;
 338         __dec_zone_page_state(page, NR_FILE_PAGES);
 339         __dec_zone_page_state(page, NR_SHMEM);
 340         spin_unlock_irq(&mapping->tree_lock);
 341         page_cache_release(page);
 342         BUG_ON(error);
 343 }
 344
 345 /*
 346  * Remove swap entry from radix tree, free the swap and its page cache.
 347  */
 348 static int shmem_free_swap(struct address_space *mapping,
 349                            pgoff_t index, void *radswap)
 350 {
 351         void *old;
 352
 353         spin_lock_irq(&mapping->tree_lock);
 354         old = radix_tree_delete_item(&mapping->page_tree, index, radswap);
 355         spin_unlock_irq(&mapping->tree_lock);
 356         if (old != radswap)
 357                 return -ENOENT;
 358         free_swap_and_cache(radix_to_swp_entry(radswap));
 359         return 0;
 360 }
 361
 362 /*
 363  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 364  */
 365 void shmem_unlock_mapping(struct address_space *mapping)
 366 {
 367         struct pagevec pvec;
 368         pgoff_t indices[PAGEVEC_SIZE];
 369         pgoff_t index = 0;
 370
 371         pagevec_init(&pvec, 0);
 372         /*
 373          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 374          */
 375         while (!mapping_unevictable(mapping)) {
 376                 /*
 377                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
 378                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
 379                  */
 380                 pvec.nr = find_get_entries(mapping, index,
 381                                            PAGEVEC_SIZE, pvec.pages, indices);
 382                 if (!pvec.nr)
 383                         break;
 384                 index = indices[pvec.nr - 1] + 1;
 385                 pagevec_remove_exceptionals(&pvec);
 386                 check_move_unevictable_pages(pvec.pages, pvec.nr);
 387                 pagevec_release(&pvec);
 388                 cond_resched();
 389         }
 390 }
 391
 392 /*
 393  * Remove range of pages and swap entries from radix tree, and free them.
 394  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 395  */
 396 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 397                                                                  bool unfalloc)
 398 {
 399         struct address_space *mapping = inode->i_mapping;
 400         struct shmem_inode_info *info = SHMEM_I(inode);
 401         pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 402         pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
 403         unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
 404         unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
 405         struct pagevec pvec;
 406         pgoff_t indices[PAGEVEC_SIZE];
 407         long nr_swaps_freed = 0;
 408         pgoff_t index;
 409         int i;
 410
 411         if (lend == -1)
 412                 end = -1;       /* unsigned, so actually very big */
 413
 414         pagevec_init(&pvec, 0);
 415         index = start;
 416         while (index < end) {
 417                 pvec.nr = find_get_entries(mapping, index,
 418                         min(end - index, (pgoff_t)PAGEVEC_SIZE),
 419                         pvec.pages, indices);
 420                 if (!pvec.nr)
 421                         break;
 422                 for (i = 0; i < pagevec_count(&pvec); i++) {
 423                         struct page *page = pvec.pages[i];
 424
 425                         index = indices[i];
 426                         if (index >= end)
 427                                 break;
 428
 429                         if (radix_tree_exceptional_entry(page)) {
 430                                 if (unfalloc)
 431                                         continue;
 432                                 nr_swaps_freed += !shmem_free_swap(mapping,
 433                                                                 index, page);
 434                                 continue;
 435                         }
 436
 437                         if (!trylock_page(page))
 438                                 continue;
 439                         if (!unfalloc || !PageUptodate(page)) {
 440                                 if (page->mapping == mapping) {
 441                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
 442                                         truncate_inode_page(mapping, page);
 443                                 }
 444                         }
 445                         unlock_page(page);
 446                 }
 447                 pagevec_remove_exceptionals(&pvec);
 448                 pagevec_release(&pvec);
 449                 cond_resched();
 450                 index++;
 451         }
 452
 453         if (partial_start) {
 454                 struct page *page = NULL;
 455                 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
 456                 if (page) {
 457                         unsigned int top = PAGE_CACHE_SIZE;
 458                         if (start > end) {
 459                                 top = partial_end;
 460                                 partial_end = 0;
 461                         }
 462                         zero_user_segment(page, partial_start, top);
 463                         set_page_dirty(page);
 464                         unlock_page(page);
 465                         page_cache_release(page);
 466                 }
 467         }
 468         if (partial_end) {
 469                 struct page *page = NULL;
 470                 shmem_getpage(inode, end, &page, SGP_READ, NULL);
 471                 if (page) {
 472                         zero_user_segment(page, 0, partial_end);
 473                         set_page_dirty(page);
 474                         unlock_page(page);
 475                         page_cache_release(page);
 476                 }
 477         }
 478         if (start >= end)
 479                 return;
 480
 481         index = start;
 482         while (index < end) {
 483                 cond_resched();
 484
 485                 pvec.nr = find_get_entries(mapping, index,
 486                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 487                                 pvec.pages, indices);
 488                 if (!pvec.nr) {
 489                         /* If all gone or hole-punch or unfalloc, we're done */
 490                         if (index == start || end != -1)
 491                                 break;
 492                         /* But if truncating, restart to make sure all gone */
 493                         index = start;
 494                         continue;
 495                 }
 496                 for (i = 0; i < pagevec_count(&pvec); i++) {
 497                         struct page *page = pvec.pages[i];
 498
 499                         index = indices[i];
 500                         if (index >= end)
 501                                 break;
 502
 503                         if (radix_tree_exceptional_entry(page)) {
 504                                 if (unfalloc)
 505                                         continue;
 506                                 if (shmem_free_swap(mapping, index, page)) {
 507                                         /* Swap was replaced by page: retry */
 508                                         index--;
 509                                         break;
 510                                 }
 511                                 nr_swaps_freed++;
 512                                 continue;
 513                         }
 514
 515                         lock_page(page);
 516                         if (!unfalloc || !PageUptodate(page)) {
 517                                 if (page->mapping == mapping) {
 518                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
 519                                         truncate_inode_page(mapping, page);
 520                                 } else {
 521                                         /* Page was replaced by swap: retry */
 522                                         unlock_page(page);
 523                                         index--;
 524                                         break;
 525                                 }
 526                         }
 527                         unlock_page(page);
 528                 }
 529                 pagevec_remove_exceptionals(&pvec);
 530                 pagevec_release(&pvec);
 531                 index++;
 532         }
 533
 534         spin_lock(&info->lock);
 535         info->swapped -= nr_swaps_freed;
 536         shmem_recalc_inode(inode);
 537         spin_unlock(&info->lock);
 538 }
 539
 540 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
 541 {
 542         shmem_undo_range(inode, lstart, lend, false);
 543         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 544 }
 545 EXPORT_SYMBOL_GPL(shmem_truncate_range);
 546
 547 static int shmem_getattr(struct vfsmount *mnt, struct dentry *dentry,
 548                          struct kstat *stat)
 549 {
 550         struct inode *inode = dentry->d_inode;
 551         struct shmem_inode_info *info = SHMEM_I(inode);
 552
 553         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
 554                 spin_lock(&info->lock);
 555                 shmem_recalc_inode(inode);
 556                 spin_unlock(&info->lock);
 557         }
 558         generic_fillattr(inode, stat);
 559         return 0;
 560 }
 561
 562 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
 563 {
 564         struct inode *inode = d_inode(dentry);
 565         struct shmem_inode_info *info = SHMEM_I(inode);
 566         int error;
 567
 568         error = inode_change_ok(inode, attr);
 569         if (error)
 570                 return error;
 571
 572         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
 573                 loff_t oldsize = inode->i_size;
 574                 loff_t newsize = attr->ia_size;
 575
 576                 /* protected by i_mutex */
 577                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 578                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 579                         return -EPERM;
 580
 581                 if (newsize != oldsize) {
 582                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
 583                                         oldsize, newsize);
 584                         if (error)
 585                                 return error;
 586                         i_size_write(inode, newsize);
 587                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 588                 }
 589                 if (newsize <= oldsize) {
 590                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
 591                         if (oldsize > holebegin)
 592                                 unmap_mapping_range(inode->i_mapping,
 593                                                         holebegin, 0, 1);
 594                         if (info->alloced)
 595                                 shmem_truncate_range(inode,
 596                                                         newsize, (loff_t)-1);
 597                         /* unmap again to remove racily COWed private pages */
 598                         if (oldsize > holebegin)
 599                                 unmap_mapping_range(inode->i_mapping,
 600                                                         holebegin, 0, 1);
 601                 }
 602         }
 603
 604         setattr_copy(inode, attr);
 605         if (attr->ia_valid & ATTR_MODE)
 606                 error = posix_acl_chmod(inode, inode->i_mode);
 607         return error;
 608 }
 609
 610 static void shmem_evict_inode(struct inode *inode)
 611 {
 612         struct shmem_inode_info *info = SHMEM_I(inode);
 613
 614         if (inode->i_mapping->a_ops == &shmem_aops) {
 615                 shmem_unacct_size(info->flags, inode->i_size);
 616                 inode->i_size = 0;
 617                 shmem_truncate_range(inode, 0, (loff_t)-1);
 618                 if (!list_empty(&info->swaplist)) {
 619                         mutex_lock(&shmem_swaplist_mutex);
 620                         list_del_init(&info->swaplist);
 621                         mutex_unlock(&shmem_swaplist_mutex);
 622                 }
 623         } else
 624                 kfree(info->symlink);
 625
 626         simple_xattrs_free(&info->xattrs);
 627         WARN_ON(inode->i_blocks);
 628         shmem_free_inode(inode->i_sb);
 629         clear_inode(inode);
 630 }
 631
 632 /*
 633  * If swap found in inode, free it and move page from swapcache to filecache.
 634  */
 635 static int shmem_unuse_inode(struct shmem_inode_info *info,
 636                              swp_entry_t swap, struct page **pagep)
 637 {
 638         struct address_space *mapping = info->vfs_inode.i_mapping;
 639         void *radswap;
 640         pgoff_t index;
 641         gfp_t gfp;
 642         int error = 0;
 643
 644         radswap = swp_to_radix_entry(swap);
 645         index = radix_tree_locate_item(&mapping->page_tree, radswap);
 646         if (index == -1)
 647                 return -EAGAIN; /* tell shmem_unuse we found nothing */
 648
 649         /*
 650          * Move _head_ to start search for next from here.
 651          * But be careful: shmem_evict_inode checks list_empty without taking
 652          * mutex, and there's an instant in list_move_tail when info->swaplist
 653          * would appear empty, if it were the only one on shmem_swaplist.
 654          */
 655         if (shmem_swaplist.next != &info->swaplist)
 656                 list_move_tail(&shmem_swaplist, &info->swaplist);
 657
 658         gfp = mapping_gfp_mask(mapping);
 659         if (shmem_should_replace_page(*pagep, gfp)) {
 660                 mutex_unlock(&shmem_swaplist_mutex);
 661                 error = shmem_replace_page(pagep, gfp, info, index);
 662                 mutex_lock(&shmem_swaplist_mutex);
 663                 /*
 664                  * We needed to drop mutex to make that restrictive page
 665                  * allocation, but the inode might have been freed while we
 666                  * dropped it: although a racing shmem_evict_inode() cannot
 667                  * complete without emptying the radix_tree, our page lock
 668                  * on this swapcache page is not enough to prevent that -
 669                  * free_swap_and_cache() of our swap entry will only
 670                  * trylock_page(), removing swap from radix_tree whatever.
 671                  *
 672                  * We must not proceed to shmem_add_to_page_cache() if the
 673                  * inode has been freed, but of course we cannot rely on
 674                  * inode or mapping or info to check that.  However, we can
 675                  * safely check if our swap entry is still in use (and here
 676                  * it can't have got reused for another page): if it's still
 677                  * in use, then the inode cannot have been freed yet, and we
 678                  * can safely proceed (if it's no longer in use, that tells
 679                  * nothing about the inode, but we don't need to unuse swap).
 680                  */
 681                 if (!page_swapcount(*pagep))
 682                         error = -ENOENT;
 683         }
 684
 685         /*
 686          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 687          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 688          * beneath us (pagelock doesn't help until the page is in pagecache).
 689          */
 690         if (!error)
 691                 error = shmem_add_to_page_cache(*pagep, mapping, index,
 692                                                 radswap);
 693         if (error != -ENOMEM) {
 694                 /*
 695                  * Truncation and eviction use free_swap_and_cache(), which
 696                  * only does trylock page: if we raced, best clean up here.
 697                  */
 698                 delete_from_swap_cache(*pagep);
 699                 set_page_dirty(*pagep);
 700                 if (!error) {
 701                         spin_lock(&info->lock);
 702                         info->swapped--;
 703                         spin_unlock(&info->lock);
 704                         swap_free(swap);
 705                 }
 706         }
 707         return error;
 708 }
 709
 710 /*
 711  * Search through swapped inodes to find and replace swap by page.
 712  */
 713 int shmem_unuse(swp_entry_t swap, struct page *page)
 714 {
 715         struct list_head *this, *next;
 716         struct shmem_inode_info *info;
 717         struct mem_cgroup *memcg;
 718         int error = 0;
 719
 720         /*
 721          * There's a faint possibility that swap page was replaced before
 722          * caller locked it: caller will come back later with the right page.
 723          */
 724         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
 725                 goto out;
 726
 727         /*
 728          * Charge page using GFP_KERNEL while we can wait, before taking
 729          * the shmem_swaplist_mutex which might hold up shmem_writepage().
 730          * Charged back to the user (not to caller) when swap account is used.
 731          */
 732         error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg);
 733         if (error)
 734                 goto out;
 735         /* No radix_tree_preload: swap entry keeps a place for page in tree */
 736         error = -EAGAIN;
 737
 738         mutex_lock(&shmem_swaplist_mutex);
 739         list_for_each_safe(this, next, &shmem_swaplist) {
 740                 info = list_entry(this, struct shmem_inode_info, swaplist);
 741                 if (info->swapped)
 742                         error = shmem_unuse_inode(info, swap, &page);
 743                 else
 744                         list_del_init(&info->swaplist);
 745                 cond_resched();
 746                 if (error != -EAGAIN)
 747                         break;
 748                 /* found nothing in this: move on to search the next */
 749         }
 750         mutex_unlock(&shmem_swaplist_mutex);
 751
 752         if (error) {
 753                 if (error != -ENOMEM)
 754                         error = 0;
 755                 mem_cgroup_cancel_charge(page, memcg);
 756         } else
 757                 mem_cgroup_commit_charge(page, memcg, true);
 758 out:
 759         unlock_page(page);
 760         page_cache_release(page);
 761         return error;
 762 }
 763
 764 /*
 765  * Move the page from the page cache to the swap cache.
 766  */
 767 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 768 {
 769         struct shmem_inode_info *info;
 770         struct address_space *mapping;
 771         struct inode *inode;
 772         swp_entry_t swap;
 773         pgoff_t index;
 774
 775         BUG_ON(!PageLocked(page));
 776         mapping = page->mapping;
 777         index = page->index;
 778         inode = mapping->host;
 779         info = SHMEM_I(inode);
 780         if (info->flags & VM_LOCKED)
 781                 goto redirty;
 782         if (!total_swap_pages)
 783                 goto redirty;
 784
 785         /*
 786          * Our capabilities prevent regular writeback or sync from ever calling
 787          * shmem_writepage; but a stacking filesystem might use ->writepage of
 788          * its underlying filesystem, in which case tmpfs should write out to
 789          * swap only in response to memory pressure, and not for the writeback
 790          * threads or sync.
 791          */
 792         if (!wbc->for_reclaim) {
 793                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
 794                 goto redirty;
 795         }
 796
 797         /*
 798          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
 799          * value into swapfile.c, the only way we can correctly account for a
 800          * fallocated page arriving here is now to initialize it and write it.
 801          *
 802          * That's okay for a page already fallocated earlier, but if we have
 803          * not yet completed the fallocation, then (a) we want to keep track
 804          * of this page in case we have to undo it, and (b) it may not be a
 805          * good idea to continue anyway, once we're pushing into swap.  So
 806          * reactivate the page, and let shmem_fallocate() quit when too many.
 807          */
 808         if (!PageUptodate(page)) {
 809                 if (inode->i_private) {
 810                         struct shmem_falloc *shmem_falloc;
 811                         spin_lock(&inode->i_lock);
 812                         shmem_falloc = inode->i_private;
 813                         if (shmem_falloc &&
 814                             !shmem_falloc->waitq &&
 815                             index >= shmem_falloc->start &&
 816                             index < shmem_falloc->next)
 817                                 shmem_falloc->nr_unswapped++;
 818                         else
 819                                 shmem_falloc = NULL;
 820                         spin_unlock(&inode->i_lock);
 821                         if (shmem_falloc)
 822                                 goto redirty;
 823                 }
 824                 clear_highpage(page);
 825                 flush_dcache_page(page);
 826                 SetPageUptodate(page);
 827         }
 828
 829         swap = get_swap_page();
 830         if (!swap.val)
 831                 goto redirty;
 832
 833         /*
 834          * Add inode to shmem_unuse()'s list of swapped-out inodes,
 835          * if it's not already there.  Do it now before the page is
 836          * moved to swap cache, when its pagelock no longer protects
 837          * the inode from eviction.  But don't unlock the mutex until
 838          * we've incremented swapped, because shmem_unuse_inode() will
 839          * prune a !swapped inode from the swaplist under this mutex.
 840          */
 841         mutex_lock(&shmem_swaplist_mutex);
 842         if (list_empty(&info->swaplist))
 843                 list_add_tail(&info->swaplist, &shmem_swaplist);
 844
 845         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
 846                 spin_lock(&info->lock);
 847                 shmem_recalc_inode(inode);
 848                 info->swapped++;
 849                 spin_unlock(&info->lock);
 850
 851                 swap_shmem_alloc(swap);
 852                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
 853
 854                 mutex_unlock(&shmem_swaplist_mutex);
 855                 BUG_ON(page_mapped(page));
 856                 swap_writepage(page, wbc);
 857                 return 0;
 858         }
 859
 860         mutex_unlock(&shmem_swaplist_mutex);
 861         swapcache_free(swap);
 862 redirty:
 863         set_page_dirty(page);
 864         if (wbc->for_reclaim)
 865                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
 866         unlock_page(page);
 867         return 0;
 868 }
 869
 870 #ifdef CONFIG_NUMA
 871 #ifdef CONFIG_TMPFS
 872 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 873 {
 874         char buffer[64];
 875
 876         if (!mpol || mpol->mode == MPOL_DEFAULT)
 877                 return;         /* show nothing */
 878
 879         mpol_to_str(buffer, sizeof(buffer), mpol);
 880
 881         seq_printf(seq, ",mpol=%s", buffer);
 882 }
 883
 884 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 885 {
 886         struct mempolicy *mpol = NULL;
 887         if (sbinfo->mpol) {
 888                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
 889                 mpol = sbinfo->mpol;
 890                 mpol_get(mpol);
 891                 spin_unlock(&sbinfo->stat_lock);
 892         }
 893         return mpol;
 894 }
 895 #endif /* CONFIG_TMPFS */
 896
 897 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 898                         struct shmem_inode_info *info, pgoff_t index)
 899 {
 900         struct vm_area_struct pvma;
 901         struct page *page;
 902
 903         /* Create a pseudo vma that just contains the policy */
 904         pvma.vm_start = 0;
 905         /* Bias interleave by inode number to distribute better across nodes */
 906         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 907         pvma.vm_ops = NULL;
 908         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 909
 910         page = swapin_readahead(swap, gfp, &pvma, 0);
 911
 912         /* Drop reference taken by mpol_shared_policy_lookup() */
 913         mpol_cond_put(pvma.vm_policy);
 914
 915         return page;
 916 }
 917
 918 static struct page *shmem_alloc_page(gfp_t gfp,
 919                         struct shmem_inode_info *info, pgoff_t index)
 920 {
 921         struct vm_area_struct pvma;
 922         struct page *page;
 923
 924         /* Create a pseudo vma that just contains the policy */
 925         pvma.vm_start = 0;
 926         /* Bias interleave by inode number to distribute better across nodes */
 927         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 928         pvma.vm_ops = NULL;
 929         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 930
 931         page = alloc_page_vma(gfp, &pvma, 0);
 932
 933         /* Drop reference taken by mpol_shared_policy_lookup() */
 934         mpol_cond_put(pvma.vm_policy);
 935
 936         return page;
 937 }
 938 #else /* !CONFIG_NUMA */
 939 #ifdef CONFIG_TMPFS
 940 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 941 {
 942 }
 943 #endif /* CONFIG_TMPFS */
 944
 945 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 946                         struct shmem_inode_info *info, pgoff_t index)
 947 {
 948         return swapin_readahead(swap, gfp, NULL, 0);
 949 }
 950
 951 static inline struct page *shmem_alloc_page(gfp_t gfp,
 952                         struct shmem_inode_info *info, pgoff_t index)
 953 {
 954         return alloc_page(gfp);
 955 }
 956 #endif /* CONFIG_NUMA */
 957
 958 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
 959 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 960 {
 961         return NULL;
 962 }
 963 #endif
 964
 965 /*
 966  * When a page is moved from swapcache to shmem filecache (either by the
 967  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
 968  * shmem_unuse_inode()), it may have been read in earlier from swap, in
 969  * ignorance of the mapping it belongs to.  If that mapping has special
 970  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
 971  * we may need to copy to a suitable page before moving to filecache.
 972  *
 973  * In a future release, this may well be extended to respect cpuset and
 974  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
 975  * but for now it is a simple matter of zone.
 976  */
 977 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
 978 {
 979         return page_zonenum(page) > gfp_zone(gfp);
 980 }
 981
 982 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 983                                 struct shmem_inode_info *info, pgoff_t index)
 984 {
 985         struct page *oldpage, *newpage;
 986         struct address_space *swap_mapping;
 987         pgoff_t swap_index;
 988         int error;
 989
 990         oldpage = *pagep;
 991         swap_index = page_private(oldpage);
 992         swap_mapping = page_mapping(oldpage);
 993
 994         /*
 995          * We have arrived here because our zones are constrained, so don't
 996          * limit chance of success by further cpuset and node constraints.
 997          */
 998         gfp &= ~GFP_CONSTRAINT_MASK;
 999         newpage = shmem_alloc_page(gfp, info, index);
1000         if (!newpage)
1001                 return -ENOMEM;
1002
1003         page_cache_get(newpage);
1004         copy_highpage(newpage, oldpage);
1005         flush_dcache_page(newpage);
1006
1007         __set_page_locked(newpage);
1008         SetPageUptodate(newpage);
1009         SetPageSwapBacked(newpage);
1010         set_page_private(newpage, swap_index);
1011         SetPageSwapCache(newpage);
1012
1013         /*
1014          * Our caller will very soon move newpage out of swapcache, but it's
1015          * a nice clean interface for us to replace oldpage by newpage there.
1016          */
1017         spin_lock_irq(&swap_mapping->tree_lock);
1018         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1019                                                                    newpage);
1020         if (!error) {
1021                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1022                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1023         }
1024         spin_unlock_irq(&swap_mapping->tree_lock);
1025
1026         if (unlikely(error)) {
1027                 /*
1028                  * Is this possible?  I think not, now that our callers check
1029                  * both PageSwapCache and page_private after getting page lock;
1030                  * but be defensive.  Reverse old to newpage for clear and free.
1031                  */
1032                 oldpage = newpage;
1033         } else {
1034                 mem_cgroup_replace_page(oldpage, newpage);
1035                 lru_cache_add_anon(newpage);
1036                 *pagep = newpage;
1037         }
1038
1039         ClearPageSwapCache(oldpage);
1040         set_page_private(oldpage, 0);
1041
1042         unlock_page(oldpage);
1043         page_cache_release(oldpage);
1044         page_cache_release(oldpage);
1045         return error;
1046 }
1047
1048 /*
1049  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1050  *
1051  * If we allocate a new one we do not mark it dirty. That's up to the
1052  * vm. If we swap it in we mark it dirty since we also free the swap
1053  * entry since a page cannot live in both the swap and page cache
1054  */
1055 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1056         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1057 {
1058         struct address_space *mapping = inode->i_mapping;
1059         struct shmem_inode_info *info;
1060         struct shmem_sb_info *sbinfo;
1061         struct mem_cgroup *memcg;
1062         struct page *page;
1063         swp_entry_t swap;
1064         int error;
1065         int once = 0;
1066         int alloced = 0;
1067
1068         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1069                 return -EFBIG;
1070 repeat:
1071         swap.val = 0;
1072         page = find_lock_entry(mapping, index);
1073         if (radix_tree_exceptional_entry(page)) {
1074                 swap = radix_to_swp_entry(page);
1075                 page = NULL;
1076         }
1077
1078         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1079             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1080                 error = -EINVAL;
1081                 goto unlock;
1082         }
1083
1084         if (page && sgp == SGP_WRITE)
1085                 mark_page_accessed(page);
1086
1087         /* fallocated page? */
1088         if (page && !PageUptodate(page)) {
1089                 if (sgp != SGP_READ)
1090                         goto clear;
1091                 unlock_page(page);
1092                 page_cache_release(page);
1093                 page = NULL;
1094         }
1095         if (page || (sgp == SGP_READ && !swap.val)) {
1096                 *pagep = page;
1097                 return 0;
1098         }
1099
1100         /*
1101          * Fast cache lookup did not find it:
1102          * bring it back from swap or allocate.
1103          */
1104         info = SHMEM_I(inode);
1105         sbinfo = SHMEM_SB(inode->i_sb);
1106
1107         if (swap.val) {
1108                 /* Look it up and read it in.. */
1109                 page = lookup_swap_cache(swap);
1110                 if (!page) {
1111                         /* here we actually do the io */
1112                         if (fault_type)
1113                                 *fault_type |= VM_FAULT_MAJOR;
1114                         page = shmem_swapin(swap, gfp, info, index);
1115                         if (!page) {
1116                                 error = -ENOMEM;
1117                                 goto failed;
1118                         }
1119                 }
1120
1121                 /* We have to do this with page locked to prevent races */
1122                 lock_page(page);
1123                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1124                     !shmem_confirm_swap(mapping, index, swap)) {
1125                         error = -EEXIST;        /* try again */
1126                         goto unlock;
1127                 }
1128                 if (!PageUptodate(page)) {
1129                         error = -EIO;
1130                         goto failed;
1131                 }
1132                 wait_on_page_writeback(page);
1133
1134                 if (shmem_should_replace_page(page, gfp)) {
1135                         error = shmem_replace_page(&page, gfp, info, index);
1136                         if (error)
1137                                 goto failed;
1138                 }
1139
1140                 error = mem_cgroup_try_charge(page, current->mm, gfp, &memcg);
1141                 if (!error) {
1142                         error = shmem_add_to_page_cache(page, mapping, index,
1143                                                 swp_to_radix_entry(swap));
1144                         /*
1145                          * We already confirmed swap under page lock, and make
1146                          * no memory allocation here, so usually no possibility
1147                          * of error; but free_swap_and_cache() only trylocks a
1148                          * page, so it is just possible that the entry has been
1149                          * truncated or holepunched since swap was confirmed.
1150                          * shmem_undo_range() will have done some of the
1151                          * unaccounting, now delete_from_swap_cache() will do
1152                          * the rest.
1153                          * Reset swap.val? No, leave it so "failed" goes back to
1154                          * "repeat": reading a hole and writing should succeed.
1155                          */
1156                         if (error) {
1157                                 mem_cgroup_cancel_charge(page, memcg);
1158                                 delete_from_swap_cache(page);
1159                         }
1160                 }
1161                 if (error)
1162                         goto failed;
1163
1164                 mem_cgroup_commit_charge(page, memcg, true);
1165
1166                 spin_lock(&info->lock);
1167                 info->swapped--;
1168                 shmem_recalc_inode(inode);
1169                 spin_unlock(&info->lock);
1170
1171                 if (sgp == SGP_WRITE)
1172                         mark_page_accessed(page);
1173
1174                 delete_from_swap_cache(page);
1175                 set_page_dirty(page);
1176                 swap_free(swap);
1177
1178         } else {
1179                 if (shmem_acct_block(info->flags)) {
1180                         error = -ENOSPC;
1181                         goto failed;
1182                 }
1183                 if (sbinfo->max_blocks) {
1184                         if (percpu_counter_compare(&sbinfo->used_blocks,
1185                                                 sbinfo->max_blocks) >= 0) {
1186                                 error = -ENOSPC;
1187                                 goto unacct;
1188                         }
1189                         percpu_counter_inc(&sbinfo->used_blocks);
1190                 }
1191
1192                 page = shmem_alloc_page(gfp, info, index);
1193                 if (!page) {
1194                         error = -ENOMEM;
1195                         goto decused;
1196                 }
1197
1198                 __SetPageSwapBacked(page);
1199                 __set_page_locked(page);
1200                 if (sgp == SGP_WRITE)
1201                         __SetPageReferenced(page);
1202
1203                 error = mem_cgroup_try_charge(page, current->mm, gfp, &memcg);
1204                 if (error)
1205                         goto decused;
1206                 error = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
1207                 if (!error) {
1208                         error = shmem_add_to_page_cache(page, mapping, index,
1209                                                         NULL);
1210                         radix_tree_preload_end();
1211                 }
1212                 if (error) {
1213                         mem_cgroup_cancel_charge(page, memcg);
1214                         goto decused;
1215                 }
1216                 mem_cgroup_commit_charge(page, memcg, false);
1217                 lru_cache_add_anon(page);
1218
1219                 spin_lock(&info->lock);
1220                 info->alloced++;
1221                 inode->i_blocks += BLOCKS_PER_PAGE;
1222                 shmem_recalc_inode(inode);
1223                 spin_unlock(&info->lock);
1224                 alloced = true;
1225
1226                 /*
1227                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1228                  */
1229                 if (sgp == SGP_FALLOC)
1230                         sgp = SGP_WRITE;
1231 clear:
1232                 /*
1233                  * Let SGP_WRITE caller clear ends if write does not fill page;
1234                  * but SGP_FALLOC on a page fallocated earlier must initialize
1235                  * it now, lest undo on failure cancel our earlier guarantee.
1236                  */
1237                 if (sgp != SGP_WRITE) {
1238                         clear_highpage(page);
1239                         flush_dcache_page(page);
1240                         SetPageUptodate(page);
1241                 }
1242                 if (sgp == SGP_DIRTY)
1243                         set_page_dirty(page);
1244         }
1245
1246         /* Perhaps the file has been truncated since we checked */
1247         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1248             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1249                 if (alloced) {
1250                         ClearPageDirty(page);
1251                         delete_from_page_cache(page);
1252                         spin_lock(&info->lock);
1253                         shmem_recalc_inode(inode);
1254                         spin_unlock(&info->lock);
1255                 }
1256                 error = -EINVAL;
1257                 goto unlock;
1258         }
1259         *pagep = page;
1260         return 0;
1261
1262         /*
1263          * Error recovery.
1264          */
1265 decused:
1266         if (sbinfo->max_blocks)
1267                 percpu_counter_add(&sbinfo->used_blocks, -1);
1268 unacct:
1269         shmem_unacct_blocks(info->flags, 1);
1270 failed:
1271         if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1272                 error = -EEXIST;
1273 unlock:
1274         if (page) {
1275                 unlock_page(page);
1276                 page_cache_release(page);
1277         }
1278         if (error == -ENOSPC && !once++) {
1279                 info = SHMEM_I(inode);
1280                 spin_lock(&info->lock);
1281                 shmem_recalc_inode(inode);
1282                 spin_unlock(&info->lock);
1283                 goto repeat;
1284         }
1285         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1286                 goto repeat;
1287         return error;
1288 }
1289
1290 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1291 {
1292         struct inode *inode = file_inode(vma->vm_file);
1293         int error;
1294         int ret = VM_FAULT_LOCKED;
1295
1296         /*
1297          * Trinity finds that probing a hole which tmpfs is punching can
1298          * prevent the hole-punch from ever completing: which in turn
1299          * locks writers out with its hold on i_mutex.  So refrain from
1300          * faulting pages into the hole while it's being punched.  Although
1301          * shmem_undo_range() does remove the additions, it may be unable to
1302          * keep up, as each new page needs its own unmap_mapping_range() call,
1303          * and the i_mmap tree grows ever slower to scan if new vmas are added.
1304          *
1305          * It does not matter if we sometimes reach this check just before the
1306          * hole-punch begins, so that one fault then races with the punch:
1307          * we just need to make racing faults a rare case.
1308          *
1309          * The implementation below would be much simpler if we just used a
1310          * standard mutex or completion: but we cannot take i_mutex in fault,
1311          * and bloating every shmem inode for this unlikely case would be sad.
1312          */
1313         if (unlikely(inode->i_private)) {
1314                 struct shmem_falloc *shmem_falloc;
1315
1316                 spin_lock(&inode->i_lock);
1317                 shmem_falloc = inode->i_private;
1318                 if (shmem_falloc &&
1319                     shmem_falloc->waitq &&
1320                     vmf->pgoff >= shmem_falloc->start &&
1321                     vmf->pgoff < shmem_falloc->next) {
1322                         wait_queue_head_t *shmem_falloc_waitq;
1323                         DEFINE_WAIT(shmem_fault_wait);
1324
1325                         ret = VM_FAULT_NOPAGE;
1326                         if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1327                            !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1328                                 /* It's polite to up mmap_sem if we can */
1329                                 up_read(&vma->vm_mm->mmap_sem);
1330                                 ret = VM_FAULT_RETRY;
1331                         }
1332
1333                         shmem_falloc_waitq = shmem_falloc->waitq;
1334                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1335                                         TASK_UNINTERRUPTIBLE);
1336                         spin_unlock(&inode->i_lock);
1337                         schedule();
1338
1339                         /*
1340                          * shmem_falloc_waitq points into the shmem_fallocate()
1341                          * stack of the hole-punching task: shmem_falloc_waitq
1342                          * is usually invalid by the time we reach here, but
1343                          * finish_wait() does not dereference it in that case;
1344                          * though i_lock needed lest racing with wake_up_all().
1345                          */
1346                         spin_lock(&inode->i_lock);
1347                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1348                         spin_unlock(&inode->i_lock);
1349                         return ret;
1350                 }
1351                 spin_unlock(&inode->i_lock);
1352         }
1353
1354         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1355         if (error)
1356                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1357
1358         if (ret & VM_FAULT_MAJOR) {
1359                 count_vm_event(PGMAJFAULT);
1360                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1361         }
1362         return ret;
1363 }
1364
1365 #ifdef CONFIG_NUMA
1366 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1367 {
1368         struct inode *inode = file_inode(vma->vm_file);
1369         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1370 }
1371
1372 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1373                                           unsigned long addr)
1374 {
1375         struct inode *inode = file_inode(vma->vm_file);
1376         pgoff_t index;
1377
1378         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1379         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1380 }
1381 #endif
1382
1383 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1384 {
1385         struct inode *inode = file_inode(file);
1386         struct shmem_inode_info *info = SHMEM_I(inode);
1387         int retval = -ENOMEM;
1388
1389         spin_lock(&info->lock);
1390         if (lock && !(info->flags & VM_LOCKED)) {
1391                 if (!user_shm_lock(inode->i_size, user))
1392                         goto out_nomem;
1393                 info->flags |= VM_LOCKED;
1394                 mapping_set_unevictable(file->f_mapping);
1395         }
1396         if (!lock && (info->flags & VM_LOCKED) && user) {
1397                 user_shm_unlock(inode->i_size, user);
1398                 info->flags &= ~VM_LOCKED;
1399                 mapping_clear_unevictable(file->f_mapping);
1400         }
1401         retval = 0;
1402
1403 out_nomem:
1404         spin_unlock(&info->lock);
1405         return retval;
1406 }
1407
1408 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1409 {
1410         file_accessed(file);
1411         vma->vm_ops = &shmem_vm_ops;
1412         return 0;
1413 }
1414
1415 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1416                                      umode_t mode, dev_t dev, unsigned long flags)
1417 {
1418         struct inode *inode;
1419         struct shmem_inode_info *info;
1420         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1421
1422         if (shmem_reserve_inode(sb))
1423                 return NULL;
1424
1425         inode = new_inode(sb);
1426         if (inode) {
1427                 inode->i_ino = get_next_ino();
1428                 inode_init_owner(inode, dir, mode);
1429                 inode->i_blocks = 0;
1430                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1431                 inode->i_generation = get_seconds();
1432                 info = SHMEM_I(inode);
1433                 memset(info, 0, (char *)inode - (char *)info);
1434                 spin_lock_init(&info->lock);
1435                 info->seals = F_SEAL_SEAL;
1436                 info->flags = flags & VM_NORESERVE;
1437                 INIT_LIST_HEAD(&info->swaplist);
1438                 simple_xattrs_init(&info->xattrs);
1439                 cache_no_acl(inode);
1440
1441                 switch (mode & S_IFMT) {
1442                 default:
1443                         inode->i_op = &shmem_special_inode_operations;
1444                         init_special_inode(inode, mode, dev);
1445                         break;
1446                 case S_IFREG:
1447                         inode->i_mapping->a_ops = &shmem_aops;
1448                         inode->i_op = &shmem_inode_operations;
1449                         inode->i_fop = &shmem_file_operations;
1450                         mpol_shared_policy_init(&info->policy,
1451                                                  shmem_get_sbmpol(sbinfo));
1452                         break;
1453                 case S_IFDIR:
1454                         inc_nlink(inode);
1455                         /* Some things misbehave if size == 0 on a directory */
1456                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1457                         inode->i_op = &shmem_dir_inode_operations;
1458                         inode->i_fop = &simple_dir_operations;
1459                         break;
1460                 case S_IFLNK:
1461                         /*
1462                          * Must not load anything in the rbtree,
1463                          * mpol_free_shared_policy will not be called.
1464                          */
1465                         mpol_shared_policy_init(&info->policy, NULL);
1466                         break;
1467                 }
1468         } else
1469                 shmem_free_inode(sb);
1470         return inode;
1471 }
1472
1473 bool shmem_mapping(struct address_space *mapping)
1474 {
1475         if (!mapping->host)
1476                 return false;
1477
1478         return mapping->host->i_sb->s_op == &shmem_ops;
1479 }
1480
1481 #ifdef CONFIG_TMPFS
1482 static const struct inode_operations shmem_symlink_inode_operations;
1483 static const struct inode_operations shmem_short_symlink_operations;
1484
1485 #ifdef CONFIG_TMPFS_XATTR
1486 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1487 #else
1488 #define shmem_initxattrs NULL
1489 #endif
1490
1491 static int
1492 shmem_write_begin(struct file *file, struct address_space *mapping,
1493                         loff_t pos, unsigned len, unsigned flags,
1494                         struct page **pagep, void **fsdata)
1495 {
1496         struct inode *inode = mapping->host;
1497         struct shmem_inode_info *info = SHMEM_I(inode);
1498         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1499
1500         /* i_mutex is held by caller */
1501         if (unlikely(info->seals)) {
1502                 if (info->seals & F_SEAL_WRITE)
1503                         return -EPERM;
1504                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
1505                         return -EPERM;
1506         }
1507
1508         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1509 }
1510
1511 static int
1512 shmem_write_end(struct file *file, struct address_space *mapping,
1513                         loff_t pos, unsigned len, unsigned copied,
1514                         struct page *page, void *fsdata)
1515 {
1516         struct inode *inode = mapping->host;
1517
1518         if (pos + copied > inode->i_size)
1519                 i_size_write(inode, pos + copied);
1520
1521         if (!PageUptodate(page)) {
1522                 if (copied < PAGE_CACHE_SIZE) {
1523                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1524                         zero_user_segments(page, 0, from,
1525                                         from + copied, PAGE_CACHE_SIZE);
1526                 }
1527                 SetPageUptodate(page);
1528         }
1529         set_page_dirty(page);
1530         unlock_page(page);
1531         page_cache_release(page);
1532
1533         return copied;
1534 }
1535
1536 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1537 {
1538         struct file *file = iocb->ki_filp;
1539         struct inode *inode = file_inode(file);
1540         struct address_space *mapping = inode->i_mapping;
1541         pgoff_t index;
1542         unsigned long offset;
1543         enum sgp_type sgp = SGP_READ;
1544         int error = 0;
1545         ssize_t retval = 0;
1546         loff_t *ppos = &iocb->ki_pos;
1547
1548         /*
1549          * Might this read be for a stacking filesystem?  Then when reading
1550          * holes of a sparse file, we actually need to allocate those pages,
1551          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1552          */
1553         if (!iter_is_iovec(to))
1554                 sgp = SGP_DIRTY;
1555
1556         index = *ppos >> PAGE_CACHE_SHIFT;
1557         offset = *ppos & ~PAGE_CACHE_MASK;
1558
1559         for (;;) {
1560                 struct page *page = NULL;
1561                 pgoff_t end_index;
1562                 unsigned long nr, ret;
1563                 loff_t i_size = i_size_read(inode);
1564
1565                 end_index = i_size >> PAGE_CACHE_SHIFT;
1566                 if (index > end_index)
1567                         break;
1568                 if (index == end_index) {
1569                         nr = i_size & ~PAGE_CACHE_MASK;
1570                         if (nr <= offset)
1571                                 break;
1572                 }
1573
1574                 error = shmem_getpage(inode, index, &page, sgp, NULL);
1575                 if (error) {
1576                         if (error == -EINVAL)
1577                                 error = 0;
1578                         break;
1579                 }
1580                 if (page)
1581                         unlock_page(page);
1582
1583                 /*
1584                  * We must evaluate after, since reads (unlike writes)
1585                  * are called without i_mutex protection against truncate
1586                  */
1587                 nr = PAGE_CACHE_SIZE;
1588                 i_size = i_size_read(inode);
1589                 end_index = i_size >> PAGE_CACHE_SHIFT;
1590                 if (index == end_index) {
1591                         nr = i_size & ~PAGE_CACHE_MASK;
1592                         if (nr <= offset) {
1593                                 if (page)
1594                                         page_cache_release(page);
1595                                 break;
1596                         }
1597                 }
1598                 nr -= offset;
1599
1600                 if (page) {
1601                         /*
1602                          * If users can be writing to this page using arbitrary
1603                          * virtual addresses, take care about potential aliasing
1604                          * before reading the page on the kernel side.
1605                          */
1606                         if (mapping_writably_mapped(mapping))
1607                                 flush_dcache_page(page);
1608                         /*
1609                          * Mark the page accessed if we read the beginning.
1610                          */
1611                         if (!offset)
1612                                 mark_page_accessed(page);
1613                 } else {
1614                         page = ZERO_PAGE(0);
1615                         page_cache_get(page);
1616                 }
1617
1618                 /*
1619                  * Ok, we have the page, and it's up-to-date, so
1620                  * now we can copy it to user space...
1621                  */
1622                 ret = copy_page_to_iter(page, offset, nr, to);
1623                 retval += ret;
1624                 offset += ret;
1625                 index += offset >> PAGE_CACHE_SHIFT;
1626                 offset &= ~PAGE_CACHE_MASK;
1627
1628                 page_cache_release(page);
1629                 if (!iov_iter_count(to))
1630                         break;
1631                 if (ret < nr) {
1632                         error = -EFAULT;
1633                         break;
1634                 }
1635                 cond_resched();
1636         }
1637
1638         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1639         file_accessed(file);
1640         return retval ? retval : error;
1641 }
1642
1643 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1644                                 struct pipe_inode_info *pipe, size_t len,
1645                                 unsigned int flags)
1646 {
1647         struct address_space *mapping = in->f_mapping;
1648         struct inode *inode = mapping->host;
1649         unsigned int loff, nr_pages, req_pages;
1650         struct page *pages[PIPE_DEF_BUFFERS];
1651         struct partial_page partial[PIPE_DEF_BUFFERS];
1652         struct page *page;
1653         pgoff_t index, end_index;
1654         loff_t isize, left;
1655         int error, page_nr;
1656         struct splice_pipe_desc spd = {
1657                 .pages = pages,
1658                 .partial = partial,
1659                 .nr_pages_max = PIPE_DEF_BUFFERS,
1660                 .flags = flags,
1661                 .ops = &page_cache_pipe_buf_ops,
1662                 .spd_release = spd_release_page,
1663         };
1664
1665         isize = i_size_read(inode);
1666         if (unlikely(*ppos >= isize))
1667                 return 0;
1668
1669         left = isize - *ppos;
1670         if (unlikely(left < len))
1671                 len = left;
1672
1673         if (splice_grow_spd(pipe, &spd))
1674                 return -ENOMEM;
1675
1676         index = *ppos >> PAGE_CACHE_SHIFT;
1677         loff = *ppos & ~PAGE_CACHE_MASK;
1678         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1679         nr_pages = min(req_pages, spd.nr_pages_max);
1680
1681         spd.nr_pages = find_get_pages_contig(mapping, index,
1682                                                 nr_pages, spd.pages);
1683         index += spd.nr_pages;
1684         error = 0;
1685
1686         while (spd.nr_pages < nr_pages) {
1687                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1688                 if (error)
1689                         break;
1690                 unlock_page(page);
1691                 spd.pages[spd.nr_pages++] = page;
1692                 index++;
1693         }
1694
1695         index = *ppos >> PAGE_CACHE_SHIFT;
1696         nr_pages = spd.nr_pages;
1697         spd.nr_pages = 0;
1698
1699         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1700                 unsigned int this_len;
1701
1702                 if (!len)
1703                         break;
1704
1705                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1706                 page = spd.pages[page_nr];
1707
1708                 if (!PageUptodate(page) || page->mapping != mapping) {
1709                         error = shmem_getpage(inode, index, &page,
1710                                                         SGP_CACHE, NULL);
1711                         if (error)
1712                                 break;
1713                         unlock_page(page);
1714                         page_cache_release(spd.pages[page_nr]);
1715                         spd.pages[page_nr] = page;
1716                 }
1717
1718                 isize = i_size_read(inode);
1719                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1720                 if (unlikely(!isize || index > end_index))
1721                         break;
1722
1723                 if (end_index == index) {
1724                         unsigned int plen;
1725
1726                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1727                         if (plen <= loff)
1728                                 break;
1729
1730                         this_len = min(this_len, plen - loff);
1731                         len = this_len;
1732                 }
1733
1734                 spd.partial[page_nr].offset = loff;
1735                 spd.partial[page_nr].len = this_len;
1736                 len -= this_len;
1737                 loff = 0;
1738                 spd.nr_pages++;
1739                 index++;
1740         }
1741
1742         while (page_nr < nr_pages)
1743                 page_cache_release(spd.pages[page_nr++]);
1744
1745         if (spd.nr_pages)
1746                 error = splice_to_pipe(pipe, &spd);
1747
1748         splice_shrink_spd(&spd);
1749
1750         if (error > 0) {
1751                 *ppos += error;
1752                 file_accessed(in);
1753         }
1754         return error;
1755 }
1756
1757 /*
1758  * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
1759  */
1760 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
1761                                     pgoff_t index, pgoff_t end, int whence)
1762 {
1763         struct page *page;
1764         struct pagevec pvec;
1765         pgoff_t indices[PAGEVEC_SIZE];
1766         bool done = false;
1767         int i;
1768
1769         pagevec_init(&pvec, 0);
1770         pvec.nr = 1;            /* start small: we may be there already */
1771         while (!done) {
1772                 pvec.nr = find_get_entries(mapping, index,
1773                                         pvec.nr, pvec.pages, indices);
1774                 if (!pvec.nr) {
1775                         if (whence == SEEK_DATA)
1776                                 index = end;
1777                         break;
1778                 }
1779                 for (i = 0; i < pvec.nr; i++, index++) {
1780                         if (index < indices[i]) {
1781                                 if (whence == SEEK_HOLE) {
1782                                         done = true;
1783                                         break;
1784                                 }
1785                                 index = indices[i];
1786                         }
1787                         page = pvec.pages[i];
1788                         if (page && !radix_tree_exceptional_entry(page)) {
1789                                 if (!PageUptodate(page))
1790                                         page = NULL;
1791                         }
1792                         if (index >= end ||
1793                             (page && whence == SEEK_DATA) ||
1794                             (!page && whence == SEEK_HOLE)) {
1795                                 done = true;
1796                                 break;
1797                         }
1798                 }
1799                 pagevec_remove_exceptionals(&pvec);
1800                 pagevec_release(&pvec);
1801                 pvec.nr = PAGEVEC_SIZE;
1802                 cond_resched();
1803         }
1804         return index;
1805 }
1806
1807 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
1808 {
1809         struct address_space *mapping = file->f_mapping;
1810         struct inode *inode = mapping->host;
1811         pgoff_t start, end;
1812         loff_t new_offset;
1813
1814         if (whence != SEEK_DATA && whence != SEEK_HOLE)
1815                 return generic_file_llseek_size(file, offset, whence,
1816                                         MAX_LFS_FILESIZE, i_size_read(inode));
1817         mutex_lock(&inode->i_mutex);
1818         /* We're holding i_mutex so we can access i_size directly */
1819
1820         if (offset < 0)
1821                 offset = -EINVAL;
1822         else if (offset >= inode->i_size)
1823                 offset = -ENXIO;
1824         else {
1825                 start = offset >> PAGE_CACHE_SHIFT;
1826                 end = (inode->i_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1827                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
1828                 new_offset <<= PAGE_CACHE_SHIFT;
1829                 if (new_offset > offset) {
1830                         if (new_offset < inode->i_size)
1831                                 offset = new_offset;
1832                         else if (whence == SEEK_DATA)
1833                                 offset = -ENXIO;
1834                         else
1835                                 offset = inode->i_size;
1836                 }
1837         }
1838
1839         if (offset >= 0)
1840                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
1841         mutex_unlock(&inode->i_mutex);
1842         return offset;
1843 }
1844
1845 /*
1846  * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
1847  * so reuse a tag which we firmly believe is never set or cleared on shmem.
1848  */
1849 #define SHMEM_TAG_PINNED        PAGECACHE_TAG_TOWRITE
1850 #define LAST_SCAN               4       /* about 150ms max */
1851
1852 static void shmem_tag_pins(struct address_space *mapping)
1853 {
1854         struct radix_tree_iter iter;
1855         void **slot;
1856         pgoff_t start;
1857         struct page *page;
1858
1859         lru_add_drain();
1860         start = 0;
1861         rcu_read_lock();
1862
1863 restart:
1864         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1865                 page = radix_tree_deref_slot(slot);
1866                 if (!page || radix_tree_exception(page)) {
1867                         if (radix_tree_deref_retry(page))
1868                                 goto restart;
1869                 } else if (page_count(page) - page_mapcount(page) > 1) {
1870                         spin_lock_irq(&mapping->tree_lock);
1871                         radix_tree_tag_set(&mapping->page_tree, iter.index,
1872                                            SHMEM_TAG_PINNED);
1873                         spin_unlock_irq(&mapping->tree_lock);
1874                 }
1875
1876                 if (need_resched()) {
1877                         cond_resched_rcu();
1878                         start = iter.index + 1;
1879                         goto restart;
1880                 }
1881         }
1882         rcu_read_unlock();
1883 }
1884
1885 /*
1886  * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
1887  * via get_user_pages(), drivers might have some pending I/O without any active
1888  * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
1889  * and see whether it has an elevated ref-count. If so, we tag them and wait for
1890  * them to be dropped.
1891  * The caller must guarantee that no new user will acquire writable references
1892  * to those pages to avoid races.
1893  */
1894 static int shmem_wait_for_pins(struct address_space *mapping)
1895 {
1896         struct radix_tree_iter iter;
1897         void **slot;
1898         pgoff_t start;
1899         struct page *page;
1900         int error, scan;
1901
1902         shmem_tag_pins(mapping);
1903
1904         error = 0;
1905         for (scan = 0; scan <= LAST_SCAN; scan++) {
1906                 if (!radix_tree_tagged(&mapping->page_tree, SHMEM_TAG_PINNED))
1907                         break;
1908
1909                 if (!scan)
1910                         lru_add_drain_all();
1911                 else if (schedule_timeout_killable((HZ << scan) / 200))
1912                         scan = LAST_SCAN;
1913
1914                 start = 0;
1915                 rcu_read_lock();
1916 restart:
1917                 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter,
1918                                            start, SHMEM_TAG_PINNED) {
1919
1920                         page = radix_tree_deref_slot(slot);
1921                         if (radix_tree_exception(page)) {
1922                                 if (radix_tree_deref_retry(page))
1923                                         goto restart;
1924
1925                                 page = NULL;
1926                         }
1927
1928                         if (page &&
1929                             page_count(page) - page_mapcount(page) != 1) {
1930                                 if (scan < LAST_SCAN)
1931                                         goto continue_resched;
1932
1933                                 /*
1934                                  * On the last scan, we clean up all those tags
1935                                  * we inserted; but make a note that we still
1936                                  * found pages pinned.
1937                                  */
1938                                 error = -EBUSY;
1939                         }
1940
1941                         spin_lock_irq(&mapping->tree_lock);
1942                         radix_tree_tag_clear(&mapping->page_tree,
1943                                              iter.index, SHMEM_TAG_PINNED);
1944                         spin_unlock_irq(&mapping->tree_lock);
1945 continue_resched:
1946                         if (need_resched()) {
1947                                 cond_resched_rcu();
1948                                 start = iter.index + 1;
1949                                 goto restart;
1950                         }
1951                 }
1952                 rcu_read_unlock();
1953         }
1954
1955         return error;
1956 }
1957
1958 #define F_ALL_SEALS (F_SEAL_SEAL | \
1959                      F_SEAL_SHRINK | \
1960                      F_SEAL_GROW | \
1961                      F_SEAL_WRITE)
1962
1963 int shmem_add_seals(struct file *file, unsigned int seals)
1964 {
1965         struct inode *inode = file_inode(file);
1966         struct shmem_inode_info *info = SHMEM_I(inode);
1967         int error;
1968
1969         /*
1970          * SEALING
1971          * Sealing allows multiple parties to share a shmem-file but restrict
1972          * access to a specific subset of file operations. Seals can only be
1973          * added, but never removed. This way, mutually untrusted parties can
1974          * share common memory regions with a well-defined policy. A malicious
1975          * peer can thus never perform unwanted operations on a shared object.
1976          *
1977          * Seals are only supported on special shmem-files and always affect
1978          * the whole underlying inode. Once a seal is set, it may prevent some
1979          * kinds of access to the file. Currently, the following seals are
1980          * defined:
1981          *   SEAL_SEAL: Prevent further seals from being set on this file
1982          *   SEAL_SHRINK: Prevent the file from shrinking
1983          *   SEAL_GROW: Prevent the file from growing
1984          *   SEAL_WRITE: Prevent write access to the file
1985          *
1986          * As we don't require any trust relationship between two parties, we
1987          * must prevent seals from being removed. Therefore, sealing a file
1988          * only adds a given set of seals to the file, it never touches
1989          * existing seals. Furthermore, the "setting seals"-operation can be
1990          * sealed itself, which basically prevents any further seal from being
1991          * added.
1992          *
1993          * Semantics of sealing are only defined on volatile files. Only
1994          * anonymous shmem files support sealing. More importantly, seals are
1995          * never written to disk. Therefore, there's no plan to support it on
1996          * other file types.
1997          */
1998
1999         if (file->f_op != &shmem_file_operations)
2000                 return -EINVAL;
2001         if (!(file->f_mode & FMODE_WRITE))
2002                 return -EPERM;
2003         if (seals & ~(unsigned int)F_ALL_SEALS)
2004                 return -EINVAL;
2005
2006         mutex_lock(&inode->i_mutex);
2007
2008         if (info->seals & F_SEAL_SEAL) {
2009                 error = -EPERM;
2010                 goto unlock;
2011         }
2012
2013         if ((seals & F_SEAL_WRITE) && !(info->seals & F_SEAL_WRITE)) {
2014                 error = mapping_deny_writable(file->f_mapping);
2015                 if (error)
2016                         goto unlock;
2017
2018                 error = shmem_wait_for_pins(file->f_mapping);
2019                 if (error) {
2020                         mapping_allow_writable(file->f_mapping);
2021                         goto unlock;
2022                 }
2023         }
2024
2025         info->seals |= seals;
2026         error = 0;
2027
2028 unlock:
2029         mutex_unlock(&inode->i_mutex);
2030         return error;
2031 }
2032 EXPORT_SYMBOL_GPL(shmem_add_seals);
2033
2034 int shmem_get_seals(struct file *file)
2035 {
2036         if (file->f_op != &shmem_file_operations)
2037                 return -EINVAL;
2038
2039         return SHMEM_I(file_inode(file))->seals;
2040 }
2041 EXPORT_SYMBOL_GPL(shmem_get_seals);
2042
2043 long shmem_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2044 {
2045         long error;
2046
2047         switch (cmd) {
2048         case F_ADD_SEALS:
2049                 /* disallow upper 32bit */
2050                 if (arg > UINT_MAX)
2051                         return -EINVAL;
2052
2053                 error = shmem_add_seals(file, arg);
2054                 break;
2055         case F_GET_SEALS:
2056                 error = shmem_get_seals(file);
2057                 break;
2058         default:
2059                 error = -EINVAL;
2060                 break;
2061         }
2062
2063         return error;
2064 }
2065
2066 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2067                                                          loff_t len)
2068 {
2069         struct inode *inode = file_inode(file);
2070         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2071         struct shmem_inode_info *info = SHMEM_I(inode);
2072         struct shmem_falloc shmem_falloc;
2073         pgoff_t start, index, end;
2074         int error;
2075
2076         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2077                 return -EOPNOTSUPP;
2078
2079         mutex_lock(&inode->i_mutex);
2080
2081         if (mode & FALLOC_FL_PUNCH_HOLE) {
2082                 struct address_space *mapping = file->f_mapping;
2083                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2084                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2085                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2086
2087                 /* protected by i_mutex */
2088                 if (info->seals & F_SEAL_WRITE) {
2089                         error = -EPERM;
2090                         goto out;
2091                 }
2092
2093                 shmem_falloc.waitq = &shmem_falloc_waitq;
2094                 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2095                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2096                 spin_lock(&inode->i_lock);
2097                 inode->i_private = &shmem_falloc;
2098                 spin_unlock(&inode->i_lock);
2099
2100                 if ((u64)unmap_end > (u64)unmap_start)
2101                         unmap_mapping_range(mapping, unmap_start,
2102                                             1 + unmap_end - unmap_start, 0);
2103                 shmem_truncate_range(inode, offset, offset + len - 1);
2104                 /* No need to unmap again: hole-punching leaves COWed pages */
2105
2106                 spin_lock(&inode->i_lock);
2107                 inode->i_private = NULL;
2108                 wake_up_all(&shmem_falloc_waitq);
2109                 spin_unlock(&inode->i_lock);
2110                 error = 0;
2111                 goto out;
2112         }
2113
2114         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2115         error = inode_newsize_ok(inode, offset + len);
2116         if (error)
2117                 goto out;
2118
2119         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2120                 error = -EPERM;
2121                 goto out;
2122         }
2123
2124         start = offset >> PAGE_CACHE_SHIFT;
2125         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2126         /* Try to avoid a swapstorm if len is impossible to satisfy */
2127         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2128                 error = -ENOSPC;
2129                 goto out;
2130         }
2131
2132         shmem_falloc.waitq = NULL;
2133         shmem_falloc.start = start;
2134         shmem_falloc.next  = start;
2135         shmem_falloc.nr_falloced = 0;
2136         shmem_falloc.nr_unswapped = 0;
2137         spin_lock(&inode->i_lock);
2138         inode->i_private = &shmem_falloc;
2139         spin_unlock(&inode->i_lock);
2140
2141         for (index = start; index < end; index++) {
2142                 struct page *page;
2143
2144                 /*
2145                  * Good, the fallocate(2) manpage permits EINTR: we may have
2146                  * been interrupted because we are using up too much memory.
2147                  */
2148                 if (signal_pending(current))
2149                         error = -EINTR;
2150                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2151                         error = -ENOMEM;
2152                 else
2153                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
2154                                                                         NULL);
2155                 if (error) {
2156                         /* Remove the !PageUptodate pages we added */
2157                         shmem_undo_range(inode,
2158                                 (loff_t)start << PAGE_CACHE_SHIFT,
2159                                 (loff_t)index << PAGE_CACHE_SHIFT, true);
2160                         goto undone;
2161                 }
2162
2163                 /*
2164                  * Inform shmem_writepage() how far we have reached.
2165                  * No need for lock or barrier: we have the page lock.
2166                  */
2167                 shmem_falloc.next++;
2168                 if (!PageUptodate(page))
2169                         shmem_falloc.nr_falloced++;
2170
2171                 /*
2172                  * If !PageUptodate, leave it that way so that freeable pages
2173                  * can be recognized if we need to rollback on error later.
2174                  * But set_page_dirty so that memory pressure will swap rather
2175                  * than free the pages we are allocating (and SGP_CACHE pages
2176                  * might still be clean: we now need to mark those dirty too).
2177                  */
2178                 set_page_dirty(page);
2179                 unlock_page(page);
2180                 page_cache_release(page);
2181                 cond_resched();
2182         }
2183
2184         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2185                 i_size_write(inode, offset + len);
2186         inode->i_ctime = CURRENT_TIME;
2187 undone:
2188         spin_lock(&inode->i_lock);
2189         inode->i_private = NULL;
2190         spin_unlock(&inode->i_lock);
2191 out:
2192         mutex_unlock(&inode->i_mutex);
2193         return error;
2194 }
2195
2196 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2197 {
2198         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2199
2200         buf->f_type = TMPFS_MAGIC;
2201         buf->f_bsize = PAGE_CACHE_SIZE;
2202         buf->f_namelen = NAME_MAX;
2203         if (sbinfo->max_blocks) {
2204                 buf->f_blocks = sbinfo->max_blocks;
2205                 buf->f_bavail =
2206                 buf->f_bfree  = sbinfo->max_blocks -
2207                                 percpu_counter_sum(&sbinfo->used_blocks);
2208         }
2209         if (sbinfo->max_inodes) {
2210                 buf->f_files = sbinfo->max_inodes;
2211                 buf->f_ffree = sbinfo->free_inodes;
2212         }
2213         /* else leave those fields 0 like simple_statfs */
2214         return 0;
2215 }
2216
2217 /*
2218  * File creation. Allocate an inode, and we're done..
2219  */
2220 static int
2221 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2222 {
2223         struct inode *inode;
2224         int error = -ENOSPC;
2225
2226         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2227         if (inode) {
2228                 error = simple_acl_create(dir, inode);
2229                 if (error)
2230                         goto out_iput;
2231                 error = security_inode_init_security(inode, dir,
2232                                                      &dentry->d_name,
2233                                                      shmem_initxattrs, NULL);
2234                 if (error && error != -EOPNOTSUPP)
2235                         goto out_iput;
2236
2237                 error = 0;
2238                 dir->i_size += BOGO_DIRENT_SIZE;
2239                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2240                 d_instantiate(dentry, inode);
2241                 dget(dentry); /* Extra count - pin the dentry in core */
2242         }
2243         return error;
2244 out_iput:
2245         iput(inode);
2246         return error;
2247 }
2248
2249 static int
2250 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2251 {
2252         struct inode *inode;
2253         int error = -ENOSPC;
2254
2255         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2256         if (inode) {
2257                 error = security_inode_init_security(inode, dir,
2258                                                      NULL,
2259                                                      shmem_initxattrs, NULL);
2260                 if (error && error != -EOPNOTSUPP)
2261                         goto out_iput;
2262                 error = simple_acl_create(dir, inode);
2263                 if (error)
2264                         goto out_iput;
2265                 d_tmpfile(dentry, inode);
2266         }
2267         return error;
2268 out_iput:
2269         iput(inode);
2270         return error;
2271 }
2272
2273 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2274 {
2275         int error;
2276
2277         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2278                 return error;
2279         inc_nlink(dir);
2280         return 0;
2281 }
2282
2283 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2284                 bool excl)
2285 {
2286         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2287 }
2288
2289 /*
2290  * Link a file..
2291  */
2292 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2293 {
2294         struct inode *inode = d_inode(old_dentry);
2295         int ret;
2296
2297         /*
2298          * No ordinary (disk based) filesystem counts links as inodes;
2299          * but each new link needs a new dentry, pinning lowmem, and
2300          * tmpfs dentries cannot be pruned until they are unlinked.
2301          */
2302         ret = shmem_reserve_inode(inode->i_sb);
2303         if (ret)
2304                 goto out;
2305
2306         dir->i_size += BOGO_DIRENT_SIZE;
2307         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2308         inc_nlink(inode);
2309         ihold(inode);   /* New dentry reference */
2310         dget(dentry);           /* Extra pinning count for the created dentry */
2311         d_instantiate(dentry, inode);
2312 out:
2313         return ret;
2314 }
2315
2316 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2317 {
2318         struct inode *inode = d_inode(dentry);
2319
2320         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2321                 shmem_free_inode(inode->i_sb);
2322
2323         dir->i_size -= BOGO_DIRENT_SIZE;
2324         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2325         drop_nlink(inode);
2326         dput(dentry);   /* Undo the count from "create" - this does all the work */
2327         return 0;
2328 }
2329
2330 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2331 {
2332         if (!simple_empty(dentry))
2333                 return -ENOTEMPTY;
2334
2335         drop_nlink(d_inode(dentry));
2336         drop_nlink(dir);
2337         return shmem_unlink(dir, dentry);
2338 }
2339
2340 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2341 {
2342         bool old_is_dir = d_is_dir(old_dentry);
2343         bool new_is_dir = d_is_dir(new_dentry);
2344
2345         if (old_dir != new_dir && old_is_dir != new_is_dir) {
2346                 if (old_is_dir) {
2347                         drop_nlink(old_dir);
2348                         inc_nlink(new_dir);
2349                 } else {
2350                         drop_nlink(new_dir);
2351                         inc_nlink(old_dir);
2352                 }
2353         }
2354         old_dir->i_ctime = old_dir->i_mtime =
2355         new_dir->i_ctime = new_dir->i_mtime =
2356         d_inode(old_dentry)->i_ctime =
2357         d_inode(new_dentry)->i_ctime = CURRENT_TIME;
2358
2359         return 0;
2360 }
2361
2362 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
2363 {
2364         struct dentry *whiteout;
2365         int error;
2366
2367         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2368         if (!whiteout)
2369                 return -ENOMEM;
2370
2371         error = shmem_mknod(old_dir, whiteout,
2372                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2373         dput(whiteout);
2374         if (error)
2375                 return error;
2376
2377         /*
2378          * Cheat and hash the whiteout while the old dentry is still in
2379          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
2380          *
2381          * d_lookup() will consistently find one of them at this point,
2382          * not sure which one, but that isn't even important.
2383          */
2384         d_rehash(whiteout);
2385         return 0;
2386 }
2387
2388 /*
2389  * The VFS layer already does all the dentry stuff for rename,
2390  * we just have to decrement the usage count for the target if
2391  * it exists so that the VFS layer correctly free's it when it
2392  * gets overwritten.
2393  */
2394 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
2395 {
2396         struct inode *inode = d_inode(old_dentry);
2397         int they_are_dirs = S_ISDIR(inode->i_mode);
2398
2399         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
2400                 return -EINVAL;
2401
2402         if (flags & RENAME_EXCHANGE)
2403                 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
2404
2405         if (!simple_empty(new_dentry))
2406                 return -ENOTEMPTY;
2407
2408         if (flags & RENAME_WHITEOUT) {
2409                 int error;
2410
2411                 error = shmem_whiteout(old_dir, old_dentry);
2412                 if (error)
2413                         return error;
2414         }
2415
2416         if (d_really_is_positive(new_dentry)) {
2417                 (void) shmem_unlink(new_dir, new_dentry);
2418                 if (they_are_dirs) {
2419                         drop_nlink(d_inode(new_dentry));
2420                         drop_nlink(old_dir);
2421                 }
2422         } else if (they_are_dirs) {
2423                 drop_nlink(old_dir);
2424                 inc_nlink(new_dir);
2425         }
2426
2427         old_dir->i_size -= BOGO_DIRENT_SIZE;
2428         new_dir->i_size += BOGO_DIRENT_SIZE;
2429         old_dir->i_ctime = old_dir->i_mtime =
2430         new_dir->i_ctime = new_dir->i_mtime =
2431         inode->i_ctime = CURRENT_TIME;
2432         return 0;
2433 }
2434
2435 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2436 {
2437         int error;
2438         int len;
2439         struct inode *inode;
2440         struct page *page;
2441         char *kaddr;
2442         struct shmem_inode_info *info;
2443
2444         len = strlen(symname) + 1;
2445         if (len > PAGE_CACHE_SIZE)
2446                 return -ENAMETOOLONG;
2447
2448         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2449         if (!inode)
2450                 return -ENOSPC;
2451
2452         error = security_inode_init_security(inode, dir, &dentry->d_name,
2453                                              shmem_initxattrs, NULL);
2454         if (error) {
2455                 if (error != -EOPNOTSUPP) {
2456                         iput(inode);
2457                         return error;
2458                 }
2459                 error = 0;
2460         }
2461
2462         info = SHMEM_I(inode);
2463         inode->i_size = len-1;
2464         if (len <= SHORT_SYMLINK_LEN) {
2465                 info->symlink = kmemdup(symname, len, GFP_KERNEL);
2466                 if (!info->symlink) {
2467                         iput(inode);
2468                         return -ENOMEM;
2469                 }
2470                 inode->i_op = &shmem_short_symlink_operations;
2471                 inode->i_link = info->symlink;
2472         } else {
2473                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2474                 if (error) {
2475                         iput(inode);
2476                         return error;
2477                 }
2478                 inode->i_mapping->a_ops = &shmem_aops;
2479                 inode->i_op = &shmem_symlink_inode_operations;
2480                 kaddr = kmap_atomic(page);
2481                 memcpy(kaddr, symname, len);
2482                 kunmap_atomic(kaddr);
2483                 SetPageUptodate(page);
2484                 set_page_dirty(page);
2485                 unlock_page(page);
2486                 page_cache_release(page);
2487         }
2488         dir->i_size += BOGO_DIRENT_SIZE;
2489         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2490         d_instantiate(dentry, inode);
2491         dget(dentry);
2492         return 0;
2493 }
2494
2495 static const char *shmem_follow_link(struct dentry *dentry, void **cookie)
2496 {
2497         struct page *page = NULL;
2498         int error = shmem_getpage(d_inode(dentry), 0, &page, SGP_READ, NULL);
2499         if (error)
2500                 return ERR_PTR(error);
2501         unlock_page(page);
2502         *cookie = page;
2503         return kmap(page);
2504 }
2505
2506 static void shmem_put_link(struct inode *unused, void *cookie)
2507 {
2508         struct page *page = cookie;
2509         kunmap(page);
2510         mark_page_accessed(page);
2511         page_cache_release(page);
2512 }
2513
2514 #ifdef CONFIG_TMPFS_XATTR
2515 /*
2516  * Superblocks without xattr inode operations may get some security.* xattr
2517  * support from the LSM "for free". As soon as we have any other xattrs
2518  * like ACLs, we also need to implement the security.* handlers at
2519  * filesystem level, though.
2520  */
2521
2522 /*
2523  * Callback for security_inode_init_security() for acquiring xattrs.
2524  */
2525 static int shmem_initxattrs(struct inode *inode,
2526                             const struct xattr *xattr_array,
2527                             void *fs_info)
2528 {
2529         struct shmem_inode_info *info = SHMEM_I(inode);
2530         const struct xattr *xattr;
2531         struct simple_xattr *new_xattr;
2532         size_t len;
2533
2534         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2535                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2536                 if (!new_xattr)
2537                         return -ENOMEM;
2538
2539                 len = strlen(xattr->name) + 1;
2540                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2541                                           GFP_KERNEL);
2542                 if (!new_xattr->name) {
2543                         kfree(new_xattr);
2544                         return -ENOMEM;
2545                 }
2546
2547                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2548                        XATTR_SECURITY_PREFIX_LEN);
2549                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2550                        xattr->name, len);
2551
2552                 simple_xattr_list_add(&info->xattrs, new_xattr);
2553         }
2554
2555         return 0;
2556 }
2557
2558 static const struct xattr_handler *shmem_xattr_handlers[] = {
2559 #ifdef CONFIG_TMPFS_POSIX_ACL
2560         &posix_acl_access_xattr_handler,
2561         &posix_acl_default_xattr_handler,
2562 #endif
2563         NULL
2564 };
2565
2566 static int shmem_xattr_validate(const char *name)
2567 {
2568         struct { const char *prefix; size_t len; } arr[] = {
2569                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2570                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2571         };
2572         int i;
2573
2574         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2575                 size_t preflen = arr[i].len;
2576                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2577                         if (!name[preflen])
2578                                 return -EINVAL;
2579                         return 0;
2580                 }
2581         }
2582         return -EOPNOTSUPP;
2583 }
2584
2585 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2586                               void *buffer, size_t size)
2587 {
2588         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2589         int err;
2590
2591         /*
2592          * If this is a request for a synthetic attribute in the system.*
2593          * namespace use the generic infrastructure to resolve a handler
2594          * for it via sb->s_xattr.
2595          */
2596         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2597                 return generic_getxattr(dentry, name, buffer, size);
2598
2599         err = shmem_xattr_validate(name);
2600         if (err)
2601                 return err;
2602
2603         return simple_xattr_get(&info->xattrs, name, buffer, size);
2604 }
2605
2606 static int shmem_setxattr(struct dentry *dentry, const char *name,
2607                           const void *value, size_t size, int flags)
2608 {
2609         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2610         int err;
2611
2612         /*
2613          * If this is a request for a synthetic attribute in the system.*
2614          * namespace use the generic infrastructure to resolve a handler
2615          * for it via sb->s_xattr.
2616          */
2617         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2618                 return generic_setxattr(dentry, name, value, size, flags);
2619
2620         err = shmem_xattr_validate(name);
2621         if (err)
2622                 return err;
2623
2624         return simple_xattr_set(&info->xattrs, name, value, size, flags);
2625 }
2626
2627 static int shmem_removexattr(struct dentry *dentry, const char *name)
2628 {
2629         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2630         int err;
2631
2632         /*
2633          * If this is a request for a synthetic attribute in the system.*
2634          * namespace use the generic infrastructure to resolve a handler
2635          * for it via sb->s_xattr.
2636          */
2637         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2638                 return generic_removexattr(dentry, name);
2639
2640         err = shmem_xattr_validate(name);
2641         if (err)
2642                 return err;
2643
2644         return simple_xattr_remove(&info->xattrs, name);
2645 }
2646
2647 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2648 {
2649         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2650         return simple_xattr_list(&info->xattrs, buffer, size);
2651 }
2652 #endif /* CONFIG_TMPFS_XATTR */
2653
2654 static const struct inode_operations shmem_short_symlink_operations = {
2655         .readlink       = generic_readlink,
2656         .follow_link    = simple_follow_link,
2657 #ifdef CONFIG_TMPFS_XATTR
2658         .setxattr       = shmem_setxattr,
2659         .getxattr       = shmem_getxattr,
2660         .listxattr      = shmem_listxattr,
2661         .removexattr    = shmem_removexattr,
2662 #endif
2663 };
2664
2665 static const struct inode_operations shmem_symlink_inode_operations = {
2666         .readlink       = generic_readlink,
2667         .follow_link    = shmem_follow_link,
2668         .put_link       = shmem_put_link,
2669 #ifdef CONFIG_TMPFS_XATTR
2670         .setxattr       = shmem_setxattr,
2671         .getxattr       = shmem_getxattr,
2672         .listxattr      = shmem_listxattr,
2673         .removexattr    = shmem_removexattr,
2674 #endif
2675 };
2676
2677 static struct dentry *shmem_get_parent(struct dentry *child)
2678 {
2679         return ERR_PTR(-ESTALE);
2680 }
2681
2682 static int shmem_match(struct inode *ino, void *vfh)
2683 {
2684         __u32 *fh = vfh;
2685         __u64 inum = fh[2];
2686         inum = (inum << 32) | fh[1];
2687         return ino->i_ino == inum && fh[0] == ino->i_generation;
2688 }
2689
2690 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2691                 struct fid *fid, int fh_len, int fh_type)
2692 {
2693         struct inode *inode;
2694         struct dentry *dentry = NULL;
2695         u64 inum;
2696
2697         if (fh_len < 3)
2698                 return NULL;
2699
2700         inum = fid->raw[2];
2701         inum = (inum << 32) | fid->raw[1];
2702
2703         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2704                         shmem_match, fid->raw);
2705         if (inode) {
2706                 dentry = d_find_alias(inode);
2707                 iput(inode);
2708         }
2709
2710         return dentry;
2711 }
2712
2713 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2714                                 struct inode *parent)
2715 {
2716         if (*len < 3) {
2717                 *len = 3;
2718                 return FILEID_INVALID;
2719         }
2720
2721         if (inode_unhashed(inode)) {
2722                 /* Unfortunately insert_inode_hash is not idempotent,
2723                  * so as we hash inodes here rather than at creation
2724                  * time, we need a lock to ensure we only try
2725                  * to do it once
2726                  */
2727                 static DEFINE_SPINLOCK(lock);
2728                 spin_lock(&lock);
2729                 if (inode_unhashed(inode))
2730                         __insert_inode_hash(inode,
2731                                             inode->i_ino + inode->i_generation);
2732                 spin_unlock(&lock);
2733         }
2734
2735         fh[0] = inode->i_generation;
2736         fh[1] = inode->i_ino;
2737         fh[2] = ((__u64)inode->i_ino) >> 32;
2738
2739         *len = 3;
2740         return 1;
2741 }
2742
2743 static const struct export_operations shmem_export_ops = {
2744         .get_parent     = shmem_get_parent,
2745         .encode_fh      = shmem_encode_fh,
2746         .fh_to_dentry   = shmem_fh_to_dentry,
2747 };
2748
2749 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2750                                bool remount)
2751 {
2752         char *this_char, *value, *rest;
2753         struct mempolicy *mpol = NULL;
2754         uid_t uid;
2755         gid_t gid;
2756
2757         while (options != NULL) {
2758                 this_char = options;
2759                 for (;;) {
2760                         /*
2761                          * NUL-terminate this option: unfortunately,
2762                          * mount options form a comma-separated list,
2763                          * but mpol's nodelist may also contain commas.
2764                          */
2765                         options = strchr(options, ',');
2766                         if (options == NULL)
2767                                 break;
2768                         options++;
2769                         if (!isdigit(*options)) {
2770                                 options[-1] = '\0';
2771                                 break;
2772                         }
2773                 }
2774                 if (!*this_char)
2775                         continue;
2776                 if ((value = strchr(this_char,'=')) != NULL) {
2777                         *value++ = 0;
2778                 } else {
2779                         printk(KERN_ERR
2780                             "tmpfs: No value for mount option '%s'\n",
2781                             this_char);
2782                         goto error;
2783                 }
2784
2785                 if (!strcmp(this_char,"size")) {
2786                         unsigned long long size;
2787                         size = memparse(value,&rest);
2788                         if (*rest == '%') {
2789                                 size <<= PAGE_SHIFT;
2790                                 size *= totalram_pages;
2791                                 do_div(size, 100);
2792                                 rest++;
2793                         }
2794                         if (*rest)
2795                                 goto bad_val;
2796                         sbinfo->max_blocks =
2797                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2798                 } else if (!strcmp(this_char,"nr_blocks")) {
2799                         sbinfo->max_blocks = memparse(value, &rest);
2800                         if (*rest)
2801                                 goto bad_val;
2802                 } else if (!strcmp(this_char,"nr_inodes")) {
2803                         sbinfo->max_inodes = memparse(value, &rest);
2804                         if (*rest)
2805                                 goto bad_val;
2806                 } else if (!strcmp(this_char,"mode")) {
2807                         if (remount)
2808                                 continue;
2809                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2810                         if (*rest)
2811                                 goto bad_val;
2812                 } else if (!strcmp(this_char,"uid")) {
2813                         if (remount)
2814                                 continue;
2815                         uid = simple_strtoul(value, &rest, 0);
2816                         if (*rest)
2817                                 goto bad_val;
2818                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2819                         if (!uid_valid(sbinfo->uid))
2820                                 goto bad_val;
2821                 } else if (!strcmp(this_char,"gid")) {
2822                         if (remount)
2823                                 continue;
2824                         gid = simple_strtoul(value, &rest, 0);
2825                         if (*rest)
2826                                 goto bad_val;
2827                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2828                         if (!gid_valid(sbinfo->gid))
2829                                 goto bad_val;
2830                 } else if (!strcmp(this_char,"mpol")) {
2831                         mpol_put(mpol);
2832                         mpol = NULL;
2833                         if (mpol_parse_str(value, &mpol))
2834                                 goto bad_val;
2835                 } else {
2836                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2837                                this_char);
2838                         goto error;
2839                 }
2840         }
2841         sbinfo->mpol = mpol;
2842         return 0;
2843
2844 bad_val:
2845         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2846                value, this_char);
2847 error:
2848         mpol_put(mpol);
2849         return 1;
2850
2851 }
2852
2853 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2854 {
2855         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2856         struct shmem_sb_info config = *sbinfo;
2857         unsigned long inodes;
2858         int error = -EINVAL;
2859
2860         config.mpol = NULL;
2861         if (shmem_parse_options(data, &config, true))
2862                 return error;
2863
2864         spin_lock(&sbinfo->stat_lock);
2865         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2866         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2867                 goto out;
2868         if (config.max_inodes < inodes)
2869                 goto out;
2870         /*
2871          * Those tests disallow limited->unlimited while any are in use;
2872          * but we must separately disallow unlimited->limited, because
2873          * in that case we have no record of how much is already in use.
2874          */
2875         if (config.max_blocks && !sbinfo->max_blocks)
2876                 goto out;
2877         if (config.max_inodes && !sbinfo->max_inodes)
2878                 goto out;
2879
2880         error = 0;
2881         sbinfo->max_blocks  = config.max_blocks;
2882         sbinfo->max_inodes  = config.max_inodes;
2883         sbinfo->free_inodes = config.max_inodes - inodes;
2884
2885         /*
2886          * Preserve previous mempolicy unless mpol remount option was specified.
2887          */
2888         if (config.mpol) {
2889                 mpol_put(sbinfo->mpol);
2890                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
2891         }
2892 out:
2893         spin_unlock(&sbinfo->stat_lock);
2894         return error;
2895 }
2896
2897 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2898 {
2899         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2900
2901         if (sbinfo->max_blocks != shmem_default_max_blocks())
2902                 seq_printf(seq, ",size=%luk",
2903                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2904         if (sbinfo->max_inodes != shmem_default_max_inodes())
2905                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2906         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2907                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2908         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2909                 seq_printf(seq, ",uid=%u",
2910                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2911         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2912                 seq_printf(seq, ",gid=%u",
2913                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2914         shmem_show_mpol(seq, sbinfo->mpol);
2915         return 0;
2916 }
2917
2918 #define MFD_NAME_PREFIX "memfd:"
2919 #define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
2920 #define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
2921
2922 #define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING)
2923
2924 SYSCALL_DEFINE2(memfd_create,
2925                 const char __user *, uname,
2926                 unsigned int, flags)
2927 {
2928         struct shmem_inode_info *info;
2929         struct file *file;
2930         int fd, error;
2931         char *name;
2932         long len;
2933
2934         if (flags & ~(unsigned int)MFD_ALL_FLAGS)
2935                 return -EINVAL;
2936
2937         /* length includes terminating zero */
2938         len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
2939         if (len <= 0)
2940                 return -EFAULT;
2941         if (len > MFD_NAME_MAX_LEN + 1)
2942                 return -EINVAL;
2943
2944         name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_TEMPORARY);
2945         if (!name)
2946                 return -ENOMEM;
2947
2948         strcpy(name, MFD_NAME_PREFIX);
2949         if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
2950                 error = -EFAULT;
2951                 goto err_name;
2952         }
2953
2954         /* terminating-zero may have changed after strnlen_user() returned */
2955         if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
2956                 error = -EFAULT;
2957                 goto err_name;
2958         }
2959
2960         fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
2961         if (fd < 0) {
2962                 error = fd;
2963                 goto err_name;
2964         }
2965
2966         file = shmem_file_setup(name, 0, VM_NORESERVE);
2967         if (IS_ERR(file)) {
2968                 error = PTR_ERR(file);
2969                 goto err_fd;
2970         }
2971         info = SHMEM_I(file_inode(file));
2972         file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
2973         file->f_flags |= O_RDWR | O_LARGEFILE;
2974         if (flags & MFD_ALLOW_SEALING)
2975                 info->seals &= ~F_SEAL_SEAL;
2976
2977         fd_install(fd, file);
2978         kfree(name);
2979         return fd;
2980
2981 err_fd:
2982         put_unused_fd(fd);
2983 err_name:
2984         kfree(name);
2985         return error;
2986 }
2987
2988 #endif /* CONFIG_TMPFS */
2989
2990 static void shmem_put_super(struct super_block *sb)
2991 {
2992         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2993
2994         percpu_counter_destroy(&sbinfo->used_blocks);
2995         mpol_put(sbinfo->mpol);
2996         kfree(sbinfo);
2997         sb->s_fs_info = NULL;
2998 }
2999
3000 int shmem_fill_super(struct super_block *sb, void *data, int silent)
3001 {
3002         struct inode *inode;
3003         struct shmem_sb_info *sbinfo;
3004         int err = -ENOMEM;
3005
3006         /* Round up to L1_CACHE_BYTES to resist false sharing */
3007         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3008                                 L1_CACHE_BYTES), GFP_KERNEL);
3009         if (!sbinfo)
3010                 return -ENOMEM;
3011
3012         sbinfo->mode = S_IRWXUGO | S_ISVTX;
3013         sbinfo->uid = current_fsuid();
3014         sbinfo->gid = current_fsgid();
3015         sb->s_fs_info = sbinfo;
3016
3017 #ifdef CONFIG_TMPFS
3018         /*
3019          * Per default we only allow half of the physical ram per
3020          * tmpfs instance, limiting inodes to one per page of lowmem;
3021          * but the internal instance is left unlimited.
3022          */
3023         if (!(sb->s_flags & MS_KERNMOUNT)) {
3024                 sbinfo->max_blocks = shmem_default_max_blocks();
3025                 sbinfo->max_inodes = shmem_default_max_inodes();
3026                 if (shmem_parse_options(data, sbinfo, false)) {
3027                         err = -EINVAL;
3028                         goto failed;
3029                 }
3030         } else {
3031                 sb->s_flags |= MS_NOUSER;
3032         }
3033         sb->s_export_op = &shmem_export_ops;
3034         sb->s_flags |= MS_NOSEC;
3035 #else
3036         sb->s_flags |= MS_NOUSER;
3037 #endif
3038
3039         spin_lock_init(&sbinfo->stat_lock);
3040         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3041                 goto failed;
3042         sbinfo->free_inodes = sbinfo->max_inodes;
3043
3044         sb->s_maxbytes = MAX_LFS_FILESIZE;
3045         sb->s_blocksize = PAGE_CACHE_SIZE;
3046         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
3047         sb->s_magic = TMPFS_MAGIC;
3048         sb->s_op = &shmem_ops;
3049         sb->s_time_gran = 1;
3050 #ifdef CONFIG_TMPFS_XATTR
3051         sb->s_xattr = shmem_xattr_handlers;
3052 #endif
3053 #ifdef CONFIG_TMPFS_POSIX_ACL
3054         sb->s_flags |= MS_POSIXACL;
3055 #endif
3056
3057         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3058         if (!inode)
3059                 goto failed;
3060         inode->i_uid = sbinfo->uid;
3061         inode->i_gid = sbinfo->gid;
3062         sb->s_root = d_make_root(inode);
3063         if (!sb->s_root)
3064                 goto failed;
3065         return 0;
3066
3067 failed:
3068         shmem_put_super(sb);
3069         return err;
3070 }
3071
3072 static struct kmem_cache *shmem_inode_cachep;
3073
3074 static struct inode *shmem_alloc_inode(struct super_block *sb)
3075 {
3076         struct shmem_inode_info *info;
3077         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3078         if (!info)
3079                 return NULL;
3080         return &info->vfs_inode;
3081 }
3082
3083 static void shmem_destroy_callback(struct rcu_head *head)
3084 {
3085         struct inode *inode = container_of(head, struct inode, i_rcu);
3086         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3087 }
3088
3089 static void shmem_destroy_inode(struct inode *inode)
3090 {
3091         if (S_ISREG(inode->i_mode))
3092                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3093         call_rcu(&inode->i_rcu, shmem_destroy_callback);
3094 }
3095
3096 static void shmem_init_inode(void *foo)
3097 {
3098         struct shmem_inode_info *info = foo;
3099         inode_init_once(&info->vfs_inode);
3100 }
3101
3102 static int shmem_init_inodecache(void)
3103 {
3104         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3105                                 sizeof(struct shmem_inode_info),
3106                                 0, SLAB_PANIC, shmem_init_inode);
3107         return 0;
3108 }
3109
3110 static void shmem_destroy_inodecache(void)
3111 {
3112         kmem_cache_destroy(shmem_inode_cachep);
3113 }
3114
3115 static const struct address_space_operations shmem_aops = {
3116         .writepage      = shmem_writepage,
3117         .set_page_dirty = __set_page_dirty_no_writeback,
3118 #ifdef CONFIG_TMPFS
3119         .write_begin    = shmem_write_begin,
3120         .write_end      = shmem_write_end,
3121 #endif
3122 #ifdef CONFIG_MIGRATION
3123         .migratepage    = migrate_page,
3124 #endif
3125         .error_remove_page = generic_error_remove_page,
3126 };
3127
3128 static const struct file_operations shmem_file_operations = {
3129         .mmap           = shmem_mmap,
3130 #ifdef CONFIG_TMPFS
3131         .llseek         = shmem_file_llseek,
3132         .read_iter      = shmem_file_read_iter,
3133         .write_iter     = generic_file_write_iter,
3134         .fsync          = noop_fsync,
3135         .splice_read    = shmem_file_splice_read,
3136         .splice_write   = iter_file_splice_write,
3137         .fallocate      = shmem_fallocate,
3138 #endif
3139 };
3140
3141 static const struct inode_operations shmem_inode_operations = {
3142         .getattr        = shmem_getattr,
3143         .setattr        = shmem_setattr,
3144 #ifdef CONFIG_TMPFS_XATTR
3145         .setxattr       = shmem_setxattr,
3146         .getxattr       = shmem_getxattr,
3147         .listxattr      = shmem_listxattr,
3148         .removexattr    = shmem_removexattr,
3149         .set_acl        = simple_set_acl,
3150 #endif
3151 };
3152
3153 static const struct inode_operations shmem_dir_inode_operations = {
3154 #ifdef CONFIG_TMPFS
3155         .create         = shmem_create,
3156         .lookup         = simple_lookup,
3157         .link           = shmem_link,
3158         .unlink         = shmem_unlink,
3159         .symlink        = shmem_symlink,
3160         .mkdir          = shmem_mkdir,
3161         .rmdir          = shmem_rmdir,
3162         .mknod          = shmem_mknod,
3163         .rename2        = shmem_rename2,
3164         .tmpfile        = shmem_tmpfile,
3165 #endif
3166 #ifdef CONFIG_TMPFS_XATTR
3167         .setxattr       = shmem_setxattr,
3168         .getxattr       = shmem_getxattr,
3169         .listxattr      = shmem_listxattr,
3170         .removexattr    = shmem_removexattr,
3171 #endif
3172 #ifdef CONFIG_TMPFS_POSIX_ACL
3173         .setattr        = shmem_setattr,
3174         .set_acl        = simple_set_acl,
3175 #endif
3176 };
3177
3178 static const struct inode_operations shmem_special_inode_operations = {
3179 #ifdef CONFIG_TMPFS_XATTR
3180         .setxattr       = shmem_setxattr,
3181         .getxattr       = shmem_getxattr,
3182         .listxattr      = shmem_listxattr,
3183         .removexattr    = shmem_removexattr,
3184 #endif
3185 #ifdef CONFIG_TMPFS_POSIX_ACL
3186         .setattr        = shmem_setattr,
3187         .set_acl        = simple_set_acl,
3188 #endif
3189 };
3190
3191 static const struct super_operations shmem_ops = {
3192         .alloc_inode    = shmem_alloc_inode,
3193         .destroy_inode  = shmem_destroy_inode,
3194 #ifdef CONFIG_TMPFS
3195         .statfs         = shmem_statfs,
3196         .remount_fs     = shmem_remount_fs,
3197         .show_options   = shmem_show_options,
3198 #endif
3199         .evict_inode    = shmem_evict_inode,
3200         .drop_inode     = generic_delete_inode,
3201         .put_super      = shmem_put_super,
3202 };
3203
3204 static const struct vm_operations_struct shmem_vm_ops = {
3205         .fault          = shmem_fault,
3206         .map_pages      = filemap_map_pages,
3207 #ifdef CONFIG_NUMA
3208         .set_policy     = shmem_set_policy,
3209         .get_policy     = shmem_get_policy,
3210 #endif
3211 };
3212
3213 static struct dentry *shmem_mount(struct file_system_type *fs_type,
3214         int flags, const char *dev_name, void *data)
3215 {
3216         return mount_nodev(fs_type, flags, data, shmem_fill_super);
3217 }
3218
3219 static struct file_system_type shmem_fs_type = {
3220         .owner          = THIS_MODULE,
3221         .name           = "tmpfs",
3222         .mount          = shmem_mount,
3223         .kill_sb        = kill_litter_super,
3224         .fs_flags       = FS_USERNS_MOUNT,
3225 };
3226
3227 int __init shmem_init(void)
3228 {
3229         int error;
3230
3231         /* If rootfs called this, don't re-init */
3232         if (shmem_inode_cachep)
3233                 return 0;
3234
3235         error = shmem_init_inodecache();
3236         if (error)
3237                 goto out3;
3238
3239         error = register_filesystem(&shmem_fs_type);
3240         if (error) {
3241                 printk(KERN_ERR "Could not register tmpfs\n");
3242                 goto out2;
3243         }
3244
3245         shm_mnt = kern_mount(&shmem_fs_type);
3246         if (IS_ERR(shm_mnt)) {
3247                 error = PTR_ERR(shm_mnt);
3248                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
3249                 goto out1;
3250         }
3251         return 0;
3252
3253 out1:
3254         unregister_filesystem(&shmem_fs_type);
3255 out2:
3256         shmem_destroy_inodecache();
3257 out3:
3258         shm_mnt = ERR_PTR(error);
3259         return error;
3260 }
3261
3262 #else /* !CONFIG_SHMEM */
3263
3264 /*
3265  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3266  *
3267  * This is intended for small system where the benefits of the full
3268  * shmem code (swap-backed and resource-limited) are outweighed by
3269  * their complexity. On systems without swap this code should be
3270  * effectively equivalent, but much lighter weight.
3271  */
3272
3273 static struct file_system_type shmem_fs_type = {
3274         .name           = "tmpfs",
3275         .mount          = ramfs_mount,
3276         .kill_sb        = kill_litter_super,
3277         .fs_flags       = FS_USERNS_MOUNT,
3278 };
3279
3280 int __init shmem_init(void)
3281 {
3282         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3283
3284         shm_mnt = kern_mount(&shmem_fs_type);
3285         BUG_ON(IS_ERR(shm_mnt));
3286
3287         return 0;
3288 }
3289
3290 int shmem_unuse(swp_entry_t swap, struct page *page)
3291 {
3292         return 0;
3293 }
3294
3295 int shmem_lock(struct file *file, int lock, struct user_struct *user)
3296 {
3297         return 0;
3298 }
3299
3300 void shmem_unlock_mapping(struct address_space *mapping)
3301 {
3302 }
3303
3304 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3305 {
3306         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3307 }
3308 EXPORT_SYMBOL_GPL(shmem_truncate_range);
3309
3310 #define shmem_vm_ops                            generic_file_vm_ops
3311 #define shmem_file_operations                   ramfs_file_operations
3312 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
3313 #define shmem_acct_size(flags, size)            0
3314 #define shmem_unacct_size(flags, size)          do {} while (0)
3315
3316 #endif /* CONFIG_SHMEM */
3317
3318 /* common code */
3319
3320 static struct dentry_operations anon_ops = {
3321         .d_dname = simple_dname
3322 };
3323
3324 static struct file *__shmem_file_setup(const char *name, loff_t size,
3325                                        unsigned long flags, unsigned int i_flags)
3326 {
3327         struct file *res;
3328         struct inode *inode;
3329         struct path path;
3330         struct super_block *sb;
3331         struct qstr this;
3332
3333         if (IS_ERR(shm_mnt))
3334                 return ERR_CAST(shm_mnt);
3335
3336         if (size < 0 || size > MAX_LFS_FILESIZE)
3337                 return ERR_PTR(-EINVAL);
3338
3339         if (shmem_acct_size(flags, size))
3340                 return ERR_PTR(-ENOMEM);
3341
3342         res = ERR_PTR(-ENOMEM);
3343         this.name = name;
3344         this.len = strlen(name);
3345         this.hash = 0; /* will go */
3346         sb = shm_mnt->mnt_sb;
3347         path.mnt = mntget(shm_mnt);
3348         path.dentry = d_alloc_pseudo(sb, &this);
3349         if (!path.dentry)
3350                 goto put_memory;
3351         d_set_d_op(path.dentry, &anon_ops);
3352
3353         res = ERR_PTR(-ENOSPC);
3354         inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
3355         if (!inode)
3356                 goto put_memory;
3357
3358         inode->i_flags |= i_flags;
3359         d_instantiate(path.dentry, inode);
3360         inode->i_size = size;
3361         clear_nlink(inode);     /* It is unlinked */
3362         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
3363         if (IS_ERR(res))
3364                 goto put_path;
3365
3366         res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
3367                   &shmem_file_operations);
3368         if (IS_ERR(res))
3369                 goto put_path;
3370
3371         return res;
3372
3373 put_memory:
3374         shmem_unacct_size(flags, size);
3375 put_path:
3376         path_put(&path);
3377         return res;
3378 }
3379
3380 /**
3381  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
3382  *      kernel internal.  There will be NO LSM permission checks against the
3383  *      underlying inode.  So users of this interface must do LSM checks at a
3384  *      higher layer.  The users are the big_key and shm implementations.  LSM
3385  *      checks are provided at the key or shm level rather than the inode.
3386  * @name: name for dentry (to be seen in /proc/<pid>/maps
3387  * @size: size to be set for the file
3388  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3389  */
3390 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
3391 {
3392         return __shmem_file_setup(name, size, flags, S_PRIVATE);
3393 }
3394
3395 /**
3396  * shmem_file_setup - get an unlinked file living in tmpfs
3397  * @name: name for dentry (to be seen in /proc/<pid>/maps
3398  * @size: size to be set for the file
3399  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3400  */
3401 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
3402 {
3403         return __shmem_file_setup(name, size, flags, 0);
3404 }
3405 EXPORT_SYMBOL_GPL(shmem_file_setup);
3406
3407 /**
3408  * shmem_zero_setup - setup a shared anonymous mapping
3409  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3410  */
3411 int shmem_zero_setup(struct vm_area_struct *vma)
3412 {
3413         struct file *file;
3414         loff_t size = vma->vm_end - vma->vm_start;
3415
3416         /*
3417          * Cloning a new file under mmap_sem leads to a lock ordering conflict
3418          * between XFS directory reading and selinux: since this file is only
3419          * accessible to the user through its mapping, use S_PRIVATE flag to
3420          * bypass file security, in the same way as shmem_kernel_file_setup().
3421          */
3422         file = __shmem_file_setup("dev/zero", size, vma->vm_flags, S_PRIVATE);
3423         if (IS_ERR(file))
3424                 return PTR_ERR(file);
3425
3426         if (vma->vm_file)
3427                 fput(vma->vm_file);
3428         vma->vm_file = file;
3429         vma->vm_ops = &shmem_vm_ops;
3430         return 0;
3431 }
3432
3433 /**
3434  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3435  * @mapping:    the page's address_space
3436  * @index:      the page index
3437  * @gfp:        the page allocator flags to use if allocating
3438  *
3439  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3440  * with any new page allocations done using the specified allocation flags.
3441  * But read_cache_page_gfp() uses the ->readpage() method: which does not
3442  * suit tmpfs, since it may have pages in swapcache, and needs to find those
3443  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3444  *
3445  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
3446  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
3447  */
3448 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3449                                          pgoff_t index, gfp_t gfp)
3450 {
3451 #ifdef CONFIG_SHMEM
3452         struct inode *inode = mapping->host;
3453         struct page *page;
3454         int error;
3455
3456         BUG_ON(mapping->a_ops != &shmem_aops);
3457         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3458         if (error)
3459                 page = ERR_PTR(error);
3460         else
3461                 unlock_page(page);
3462         return page;
3463 #else
3464         /*
3465          * The tiny !SHMEM case uses ramfs without swap
3466          */
3467         return read_cache_page_gfp(mapping, index, gfp);
3468 #endif
3469 }
3470 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);