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/random.h>
  33 #include <linux/sched/signal.h>
  34 #include <linux/export.h>
  35 #include <linux/swap.h>
  36 #include <linux/uio.h>
  37 #include <linux/khugepaged.h>
  38 #include <linux/hugetlb.h>
  39 #include <linux/frontswap.h>
  40
  41 #include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
  42
  43 static struct vfsmount *shm_mnt;
  44
  45 #ifdef CONFIG_SHMEM
  46 /*
  47  * This virtual memory filesystem is heavily based on the ramfs. It
  48  * extends ramfs by the ability to use swap and honor resource limits
  49  * which makes it a completely usable filesystem.
  50  */
  51
  52 #include <linux/xattr.h>
  53 #include <linux/exportfs.h>
  54 #include <linux/posix_acl.h>
  55 #include <linux/posix_acl_xattr.h>
  56 #include <linux/mman.h>
  57 #include <linux/string.h>
  58 #include <linux/slab.h>
  59 #include <linux/backing-dev.h>
  60 #include <linux/shmem_fs.h>
  61 #include <linux/writeback.h>
  62 #include <linux/blkdev.h>
  63 #include <linux/pagevec.h>
  64 #include <linux/percpu_counter.h>
  65 #include <linux/falloc.h>
  66 #include <linux/splice.h>
  67 #include <linux/security.h>
  68 #include <linux/swapops.h>
  69 #include <linux/mempolicy.h>
  70 #include <linux/namei.h>
  71 #include <linux/ctype.h>
  72 #include <linux/migrate.h>
  73 #include <linux/highmem.h>
  74 #include <linux/seq_file.h>
  75 #include <linux/magic.h>
  76 #include <linux/syscalls.h>
  77 #include <linux/fcntl.h>
  78 #include <uapi/linux/memfd.h>
  79 #include <linux/userfaultfd_k.h>
  80 #include <linux/rmap.h>
  81 #include <linux/uuid.h>
  82
  83 #include <linux/uaccess.h>
  84 #include <asm/pgtable.h>
  85
  86 #include "internal.h"
  87
  88 #define BLOCKS_PER_PAGE  (PAGE_SIZE/512)
  89 #define VM_ACCT(size)    (PAGE_ALIGN(size) >> PAGE_SHIFT)
  90
  91 /* Pretend that each entry is of this size in directory's i_size */
  92 #define BOGO_DIRENT_SIZE 20
  93
  94 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  95 #define SHORT_SYMLINK_LEN 128
  96
  97 /*
  98  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
  99  * inode->i_private (with i_mutex making sure that it has only one user at
 100  * a time): we would prefer not to enlarge the shmem inode just for that.
 101  */
 102 struct shmem_falloc {
 103         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
 104         pgoff_t start;          /* start of range currently being fallocated */
 105         pgoff_t next;           /* the next page offset to be fallocated */
 106         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
 107         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
 108 };
 109
 110 #ifdef CONFIG_TMPFS
 111 static unsigned long shmem_default_max_blocks(void)
 112 {
 113         return totalram_pages() / 2;
 114 }
 115
 116 static unsigned long shmem_default_max_inodes(void)
 117 {
 118         unsigned long nr_pages = totalram_pages();
 119
 120         return min(nr_pages - totalhigh_pages(), nr_pages / 2);
 121 }
 122 #endif
 123
 124 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
 125 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 126                                 struct shmem_inode_info *info, pgoff_t index);
 127 static int shmem_swapin_page(struct inode *inode, pgoff_t index,
 128                              struct page **pagep, enum sgp_type sgp,
 129                              gfp_t gfp, struct vm_area_struct *vma,
 130                              vm_fault_t *fault_type);
 131 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 132                 struct page **pagep, enum sgp_type sgp,
 133                 gfp_t gfp, struct vm_area_struct *vma,
 134                 struct vm_fault *vmf, vm_fault_t *fault_type);
 135
 136 int shmem_getpage(struct inode *inode, pgoff_t index,
 137                 struct page **pagep, enum sgp_type sgp)
 138 {
 139         return shmem_getpage_gfp(inode, index, pagep, sgp,
 140                 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
 141 }
 142
 143 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 144 {
 145         return sb->s_fs_info;
 146 }
 147
 148 /*
 149  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 150  * for shared memory and for shared anonymous (/dev/zero) mappings
 151  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 152  * consistent with the pre-accounting of private mappings ...
 153  */
 154 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 155 {
 156         return (flags & VM_NORESERVE) ?
 157                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 158 }
 159
 160 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 161 {
 162         if (!(flags & VM_NORESERVE))
 163                 vm_unacct_memory(VM_ACCT(size));
 164 }
 165
 166 static inline int shmem_reacct_size(unsigned long flags,
 167                 loff_t oldsize, loff_t newsize)
 168 {
 169         if (!(flags & VM_NORESERVE)) {
 170                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
 171                         return security_vm_enough_memory_mm(current->mm,
 172                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
 173                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
 174                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
 175         }
 176         return 0;
 177 }
 178
 179 /*
 180  * ... whereas tmpfs objects are accounted incrementally as
 181  * pages are allocated, in order to allow large sparse files.
 182  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 183  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 184  */
 185 static inline int shmem_acct_block(unsigned long flags, long pages)
 186 {
 187         if (!(flags & VM_NORESERVE))
 188                 return 0;
 189
 190         return security_vm_enough_memory_mm(current->mm,
 191                         pages * VM_ACCT(PAGE_SIZE));
 192 }
 193
 194 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 195 {
 196         if (flags & VM_NORESERVE)
 197                 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
 198 }
 199
 200 static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
 201 {
 202         struct shmem_inode_info *info = SHMEM_I(inode);
 203         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 204
 205         if (shmem_acct_block(info->flags, pages))
 206                 return false;
 207
 208         if (sbinfo->max_blocks) {
 209                 if (percpu_counter_compare(&sbinfo->used_blocks,
 210                                            sbinfo->max_blocks - pages) > 0)
 211                         goto unacct;
 212                 percpu_counter_add(&sbinfo->used_blocks, pages);
 213         }
 214
 215         return true;
 216
 217 unacct:
 218         shmem_unacct_blocks(info->flags, pages);
 219         return false;
 220 }
 221
 222 static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
 223 {
 224         struct shmem_inode_info *info = SHMEM_I(inode);
 225         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 226
 227         if (sbinfo->max_blocks)
 228                 percpu_counter_sub(&sbinfo->used_blocks, pages);
 229         shmem_unacct_blocks(info->flags, pages);
 230 }
 231
 232 static const struct super_operations shmem_ops;
 233 static const struct address_space_operations shmem_aops;
 234 static const struct file_operations shmem_file_operations;
 235 static const struct inode_operations shmem_inode_operations;
 236 static const struct inode_operations shmem_dir_inode_operations;
 237 static const struct inode_operations shmem_special_inode_operations;
 238 static const struct vm_operations_struct shmem_vm_ops;
 239 static struct file_system_type shmem_fs_type;
 240
 241 bool vma_is_shmem(struct vm_area_struct *vma)
 242 {
 243         return vma->vm_ops == &shmem_vm_ops;
 244 }
 245
 246 static LIST_HEAD(shmem_swaplist);
 247 static DEFINE_MUTEX(shmem_swaplist_mutex);
 248
 249 static int shmem_reserve_inode(struct super_block *sb)
 250 {
 251         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 252         if (sbinfo->max_inodes) {
 253                 spin_lock(&sbinfo->stat_lock);
 254                 if (!sbinfo->free_inodes) {
 255                         spin_unlock(&sbinfo->stat_lock);
 256                         return -ENOSPC;
 257                 }
 258                 sbinfo->free_inodes--;
 259                 spin_unlock(&sbinfo->stat_lock);
 260         }
 261         return 0;
 262 }
 263
 264 static void shmem_free_inode(struct super_block *sb)
 265 {
 266         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 267         if (sbinfo->max_inodes) {
 268                 spin_lock(&sbinfo->stat_lock);
 269                 sbinfo->free_inodes++;
 270                 spin_unlock(&sbinfo->stat_lock);
 271         }
 272 }
 273
 274 /**
 275  * shmem_recalc_inode - recalculate the block usage of an inode
 276  * @inode: inode to recalc
 277  *
 278  * We have to calculate the free blocks since the mm can drop
 279  * undirtied hole pages behind our back.
 280  *
 281  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 282  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 283  *
 284  * It has to be called with the spinlock held.
 285  */
 286 static void shmem_recalc_inode(struct inode *inode)
 287 {
 288         struct shmem_inode_info *info = SHMEM_I(inode);
 289         long freed;
 290
 291         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 292         if (freed > 0) {
 293                 info->alloced -= freed;
 294                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 295                 shmem_inode_unacct_blocks(inode, freed);
 296         }
 297 }
 298
 299 bool shmem_charge(struct inode *inode, long pages)
 300 {
 301         struct shmem_inode_info *info = SHMEM_I(inode);
 302         unsigned long flags;
 303
 304         if (!shmem_inode_acct_block(inode, pages))
 305                 return false;
 306
 307         /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
 308         inode->i_mapping->nrpages += pages;
 309
 310         spin_lock_irqsave(&info->lock, flags);
 311         info->alloced += pages;
 312         inode->i_blocks += pages * BLOCKS_PER_PAGE;
 313         shmem_recalc_inode(inode);
 314         spin_unlock_irqrestore(&info->lock, flags);
 315
 316         return true;
 317 }
 318
 319 void shmem_uncharge(struct inode *inode, long pages)
 320 {
 321         struct shmem_inode_info *info = SHMEM_I(inode);
 322         unsigned long flags;
 323
 324         /* nrpages adjustment done by __delete_from_page_cache() or caller */
 325
 326         spin_lock_irqsave(&info->lock, flags);
 327         info->alloced -= pages;
 328         inode->i_blocks -= pages * BLOCKS_PER_PAGE;
 329         shmem_recalc_inode(inode);
 330         spin_unlock_irqrestore(&info->lock, flags);
 331
 332         shmem_inode_unacct_blocks(inode, pages);
 333 }
 334
 335 /*
 336  * Replace item expected in xarray by a new item, while holding xa_lock.
 337  */
 338 static int shmem_replace_entry(struct address_space *mapping,
 339                         pgoff_t index, void *expected, void *replacement)
 340 {
 341         XA_STATE(xas, &mapping->i_pages, index);
 342         void *item;
 343
 344         VM_BUG_ON(!expected);
 345         VM_BUG_ON(!replacement);
 346         item = xas_load(&xas);
 347         if (item != expected)
 348                 return -ENOENT;
 349         xas_store(&xas, replacement);
 350         return 0;
 351 }
 352
 353 /*
 354  * Sometimes, before we decide whether to proceed or to fail, we must check
 355  * that an entry was not already brought back from swap by a racing thread.
 356  *
 357  * Checking page is not enough: by the time a SwapCache page is locked, it
 358  * might be reused, and again be SwapCache, using the same swap as before.
 359  */
 360 static bool shmem_confirm_swap(struct address_space *mapping,
 361                                pgoff_t index, swp_entry_t swap)
 362 {
 363         return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
 364 }
 365
 366 /*
 367  * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
 368  *
 369  * SHMEM_HUGE_NEVER:
 370  *      disables huge pages for the mount;
 371  * SHMEM_HUGE_ALWAYS:
 372  *      enables huge pages for the mount;
 373  * SHMEM_HUGE_WITHIN_SIZE:
 374  *      only allocate huge pages if the page will be fully within i_size,
 375  *      also respect fadvise()/madvise() hints;
 376  * SHMEM_HUGE_ADVISE:
 377  *      only allocate huge pages if requested with fadvise()/madvise();
 378  */
 379
 380 #define SHMEM_HUGE_NEVER        0
 381 #define SHMEM_HUGE_ALWAYS       1
 382 #define SHMEM_HUGE_WITHIN_SIZE  2
 383 #define SHMEM_HUGE_ADVISE       3
 384
 385 /*
 386  * Special values.
 387  * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
 388  *
 389  * SHMEM_HUGE_DENY:
 390  *      disables huge on shm_mnt and all mounts, for emergency use;
 391  * SHMEM_HUGE_FORCE:
 392  *      enables huge on shm_mnt and all mounts, w/o needing option, for testing;
 393  *
 394  */
 395 #define SHMEM_HUGE_DENY         (-1)
 396 #define SHMEM_HUGE_FORCE        (-2)
 397
 398 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
 399 /* ifdef here to avoid bloating shmem.o when not necessary */
 400
 401 static int shmem_huge __read_mostly;
 402
 403 #if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
 404 static int shmem_parse_huge(const char *str)
 405 {
 406         if (!strcmp(str, "never"))
 407                 return SHMEM_HUGE_NEVER;
 408         if (!strcmp(str, "always"))
 409                 return SHMEM_HUGE_ALWAYS;
 410         if (!strcmp(str, "within_size"))
 411                 return SHMEM_HUGE_WITHIN_SIZE;
 412         if (!strcmp(str, "advise"))
 413                 return SHMEM_HUGE_ADVISE;
 414         if (!strcmp(str, "deny"))
 415                 return SHMEM_HUGE_DENY;
 416         if (!strcmp(str, "force"))
 417                 return SHMEM_HUGE_FORCE;
 418         return -EINVAL;
 419 }
 420
 421 static const char *shmem_format_huge(int huge)
 422 {
 423         switch (huge) {
 424         case SHMEM_HUGE_NEVER:
 425                 return "never";
 426         case SHMEM_HUGE_ALWAYS:
 427                 return "always";
 428         case SHMEM_HUGE_WITHIN_SIZE:
 429                 return "within_size";
 430         case SHMEM_HUGE_ADVISE:
 431                 return "advise";
 432         case SHMEM_HUGE_DENY:
 433                 return "deny";
 434         case SHMEM_HUGE_FORCE:
 435                 return "force";
 436         default:
 437                 VM_BUG_ON(1);
 438                 return "bad_val";
 439         }
 440 }
 441 #endif
 442
 443 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 444                 struct shrink_control *sc, unsigned long nr_to_split)
 445 {
 446         LIST_HEAD(list), *pos, *next;
 447         LIST_HEAD(to_remove);
 448         struct inode *inode;
 449         struct shmem_inode_info *info;
 450         struct page *page;
 451         unsigned long batch = sc ? sc->nr_to_scan : 128;
 452         int removed = 0, split = 0;
 453
 454         if (list_empty(&sbinfo->shrinklist))
 455                 return SHRINK_STOP;
 456
 457         spin_lock(&sbinfo->shrinklist_lock);
 458         list_for_each_safe(pos, next, &sbinfo->shrinklist) {
 459                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 460
 461                 /* pin the inode */
 462                 inode = igrab(&info->vfs_inode);
 463
 464                 /* inode is about to be evicted */
 465                 if (!inode) {
 466                         list_del_init(&info->shrinklist);
 467                         removed++;
 468                         goto next;
 469                 }
 470
 471                 /* Check if there's anything to gain */
 472                 if (round_up(inode->i_size, PAGE_SIZE) ==
 473                                 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
 474                         list_move(&info->shrinklist, &to_remove);
 475                         removed++;
 476                         goto next;
 477                 }
 478
 479                 list_move(&info->shrinklist, &list);
 480 next:
 481                 if (!--batch)
 482                         break;
 483         }
 484         spin_unlock(&sbinfo->shrinklist_lock);
 485
 486         list_for_each_safe(pos, next, &to_remove) {
 487                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 488                 inode = &info->vfs_inode;
 489                 list_del_init(&info->shrinklist);
 490                 iput(inode);
 491         }
 492
 493         list_for_each_safe(pos, next, &list) {
 494                 int ret;
 495
 496                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 497                 inode = &info->vfs_inode;
 498
 499                 if (nr_to_split && split >= nr_to_split)
 500                         goto leave;
 501
 502                 page = find_get_page(inode->i_mapping,
 503                                 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
 504                 if (!page)
 505                         goto drop;
 506
 507                 /* No huge page at the end of the file: nothing to split */
 508                 if (!PageTransHuge(page)) {
 509                         put_page(page);
 510                         goto drop;
 511                 }
 512
 513                 /*
 514                  * Leave the inode on the list if we failed to lock
 515                  * the page at this time.
 516                  *
 517                  * Waiting for the lock may lead to deadlock in the
 518                  * reclaim path.
 519                  */
 520                 if (!trylock_page(page)) {
 521                         put_page(page);
 522                         goto leave;
 523                 }
 524
 525                 ret = split_huge_page(page);
 526                 unlock_page(page);
 527                 put_page(page);
 528
 529                 /* If split failed leave the inode on the list */
 530                 if (ret)
 531                         goto leave;
 532
 533                 split++;
 534 drop:
 535                 list_del_init(&info->shrinklist);
 536                 removed++;
 537 leave:
 538                 iput(inode);
 539         }
 540
 541         spin_lock(&sbinfo->shrinklist_lock);
 542         list_splice_tail(&list, &sbinfo->shrinklist);
 543         sbinfo->shrinklist_len -= removed;
 544         spin_unlock(&sbinfo->shrinklist_lock);
 545
 546         return split;
 547 }
 548
 549 static long shmem_unused_huge_scan(struct super_block *sb,
 550                 struct shrink_control *sc)
 551 {
 552         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 553
 554         if (!READ_ONCE(sbinfo->shrinklist_len))
 555                 return SHRINK_STOP;
 556
 557         return shmem_unused_huge_shrink(sbinfo, sc, 0);
 558 }
 559
 560 static long shmem_unused_huge_count(struct super_block *sb,
 561                 struct shrink_control *sc)
 562 {
 563         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 564         return READ_ONCE(sbinfo->shrinklist_len);
 565 }
 566 #else /* !CONFIG_TRANSPARENT_HUGE_PAGECACHE */
 567
 568 #define shmem_huge SHMEM_HUGE_DENY
 569
 570 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 571                 struct shrink_control *sc, unsigned long nr_to_split)
 572 {
 573         return 0;
 574 }
 575 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
 576
 577 static inline bool is_huge_enabled(struct shmem_sb_info *sbinfo)
 578 {
 579         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
 580             (shmem_huge == SHMEM_HUGE_FORCE || sbinfo->huge) &&
 581             shmem_huge != SHMEM_HUGE_DENY)
 582                 return true;
 583         return false;
 584 }
 585
 586 /*
 587  * Like add_to_page_cache_locked, but error if expected item has gone.
 588  */
 589 static int shmem_add_to_page_cache(struct page *page,
 590                                    struct address_space *mapping,
 591                                    pgoff_t index, void *expected, gfp_t gfp)
 592 {
 593         XA_STATE_ORDER(xas, &mapping->i_pages, index, compound_order(page));
 594         unsigned long i = 0;
 595         unsigned long nr = 1UL << compound_order(page);
 596
 597         VM_BUG_ON_PAGE(PageTail(page), page);
 598         VM_BUG_ON_PAGE(index != round_down(index, nr), page);
 599         VM_BUG_ON_PAGE(!PageLocked(page), page);
 600         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 601         VM_BUG_ON(expected && PageTransHuge(page));
 602
 603         page_ref_add(page, nr);
 604         page->mapping = mapping;
 605         page->index = index;
 606
 607         do {
 608                 void *entry;
 609                 xas_lock_irq(&xas);
 610                 entry = xas_find_conflict(&xas);
 611                 if (entry != expected)
 612                         xas_set_err(&xas, -EEXIST);
 613                 xas_create_range(&xas);
 614                 if (xas_error(&xas))
 615                         goto unlock;
 616 next:
 617                 xas_store(&xas, page + i);
 618                 if (++i < nr) {
 619                         xas_next(&xas);
 620                         goto next;
 621                 }
 622                 if (PageTransHuge(page)) {
 623                         count_vm_event(THP_FILE_ALLOC);
 624                         __inc_node_page_state(page, NR_SHMEM_THPS);
 625                 }
 626                 mapping->nrpages += nr;
 627                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
 628                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, nr);
 629 unlock:
 630                 xas_unlock_irq(&xas);
 631         } while (xas_nomem(&xas, gfp));
 632
 633         if (xas_error(&xas)) {
 634                 page->mapping = NULL;
 635                 page_ref_sub(page, nr);
 636                 return xas_error(&xas);
 637         }
 638
 639         return 0;
 640 }
 641
 642 /*
 643  * Like delete_from_page_cache, but substitutes swap for page.
 644  */
 645 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 646 {
 647         struct address_space *mapping = page->mapping;
 648         int error;
 649
 650         VM_BUG_ON_PAGE(PageCompound(page), page);
 651
 652         xa_lock_irq(&mapping->i_pages);
 653         error = shmem_replace_entry(mapping, page->index, page, radswap);
 654         page->mapping = NULL;
 655         mapping->nrpages--;
 656         __dec_node_page_state(page, NR_FILE_PAGES);
 657         __dec_node_page_state(page, NR_SHMEM);
 658         xa_unlock_irq(&mapping->i_pages);
 659         put_page(page);
 660         BUG_ON(error);
 661 }
 662
 663 /*
 664  * Remove swap entry from page cache, free the swap and its page cache.
 665  */
 666 static int shmem_free_swap(struct address_space *mapping,
 667                            pgoff_t index, void *radswap)
 668 {
 669         void *old;
 670
 671         old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
 672         if (old != radswap)
 673                 return -ENOENT;
 674         free_swap_and_cache(radix_to_swp_entry(radswap));
 675         return 0;
 676 }
 677
 678 /*
 679  * Determine (in bytes) how many of the shmem object's pages mapped by the
 680  * given offsets are swapped out.
 681  *
 682  * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
 683  * as long as the inode doesn't go away and racy results are not a problem.
 684  */
 685 unsigned long shmem_partial_swap_usage(struct address_space *mapping,
 686                                                 pgoff_t start, pgoff_t end)
 687 {
 688         XA_STATE(xas, &mapping->i_pages, start);
 689         struct page *page;
 690         unsigned long swapped = 0;
 691
 692         rcu_read_lock();
 693         xas_for_each(&xas, page, end - 1) {
 694                 if (xas_retry(&xas, page))
 695                         continue;
 696                 if (xa_is_value(page))
 697                         swapped++;
 698
 699                 if (need_resched()) {
 700                         xas_pause(&xas);
 701                         cond_resched_rcu();
 702                 }
 703         }
 704
 705         rcu_read_unlock();
 706
 707         return swapped << PAGE_SHIFT;
 708 }
 709
 710 /*
 711  * Determine (in bytes) how many of the shmem object's pages mapped by the
 712  * given vma is swapped out.
 713  *
 714  * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
 715  * as long as the inode doesn't go away and racy results are not a problem.
 716  */
 717 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
 718 {
 719         struct inode *inode = file_inode(vma->vm_file);
 720         struct shmem_inode_info *info = SHMEM_I(inode);
 721         struct address_space *mapping = inode->i_mapping;
 722         unsigned long swapped;
 723
 724         /* Be careful as we don't hold info->lock */
 725         swapped = READ_ONCE(info->swapped);
 726
 727         /*
 728          * The easier cases are when the shmem object has nothing in swap, or
 729          * the vma maps it whole. Then we can simply use the stats that we
 730          * already track.
 731          */
 732         if (!swapped)
 733                 return 0;
 734
 735         if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
 736                 return swapped << PAGE_SHIFT;
 737
 738         /* Here comes the more involved part */
 739         return shmem_partial_swap_usage(mapping,
 740                         linear_page_index(vma, vma->vm_start),
 741                         linear_page_index(vma, vma->vm_end));
 742 }
 743
 744 /*
 745  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 746  */
 747 void shmem_unlock_mapping(struct address_space *mapping)
 748 {
 749         struct pagevec pvec;
 750         pgoff_t indices[PAGEVEC_SIZE];
 751         pgoff_t index = 0;
 752
 753         pagevec_init(&pvec);
 754         /*
 755          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 756          */
 757         while (!mapping_unevictable(mapping)) {
 758                 /*
 759                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
 760                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
 761                  */
 762                 pvec.nr = find_get_entries(mapping, index,
 763                                            PAGEVEC_SIZE, pvec.pages, indices);
 764                 if (!pvec.nr)
 765                         break;
 766                 index = indices[pvec.nr - 1] + 1;
 767                 pagevec_remove_exceptionals(&pvec);
 768                 check_move_unevictable_pages(&pvec);
 769                 pagevec_release(&pvec);
 770                 cond_resched();
 771         }
 772 }
 773
 774 /*
 775  * Remove range of pages and swap entries from page cache, and free them.
 776  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 777  */
 778 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 779                                                                  bool unfalloc)
 780 {
 781         struct address_space *mapping = inode->i_mapping;
 782         struct shmem_inode_info *info = SHMEM_I(inode);
 783         pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
 784         pgoff_t end = (lend + 1) >> PAGE_SHIFT;
 785         unsigned int partial_start = lstart & (PAGE_SIZE - 1);
 786         unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
 787         struct pagevec pvec;
 788         pgoff_t indices[PAGEVEC_SIZE];
 789         long nr_swaps_freed = 0;
 790         pgoff_t index;
 791         int i;
 792
 793         if (lend == -1)
 794                 end = -1;       /* unsigned, so actually very big */
 795
 796         pagevec_init(&pvec);
 797         index = start;
 798         while (index < end) {
 799                 pvec.nr = find_get_entries(mapping, index,
 800                         min(end - index, (pgoff_t)PAGEVEC_SIZE),
 801                         pvec.pages, indices);
 802                 if (!pvec.nr)
 803                         break;
 804                 for (i = 0; i < pagevec_count(&pvec); i++) {
 805                         struct page *page = pvec.pages[i];
 806
 807                         index = indices[i];
 808                         if (index >= end)
 809                                 break;
 810
 811                         if (xa_is_value(page)) {
 812                                 if (unfalloc)
 813                                         continue;
 814                                 nr_swaps_freed += !shmem_free_swap(mapping,
 815                                                                 index, page);
 816                                 continue;
 817                         }
 818
 819                         VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
 820
 821                         if (!trylock_page(page))
 822                                 continue;
 823
 824                         if (PageTransTail(page)) {
 825                                 /* Middle of THP: zero out the page */
 826                                 clear_highpage(page);
 827                                 unlock_page(page);
 828                                 continue;
 829                         } else if (PageTransHuge(page)) {
 830                                 if (index == round_down(end, HPAGE_PMD_NR)) {
 831                                         /*
 832                                          * Range ends in the middle of THP:
 833                                          * zero out the page
 834                                          */
 835                                         clear_highpage(page);
 836                                         unlock_page(page);
 837                                         continue;
 838                                 }
 839                                 index += HPAGE_PMD_NR - 1;
 840                                 i += HPAGE_PMD_NR - 1;
 841                         }
 842
 843                         if (!unfalloc || !PageUptodate(page)) {
 844                                 VM_BUG_ON_PAGE(PageTail(page), page);
 845                                 if (page_mapping(page) == mapping) {
 846                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
 847                                         truncate_inode_page(mapping, page);
 848                                 }
 849                         }
 850                         unlock_page(page);
 851                 }
 852                 pagevec_remove_exceptionals(&pvec);
 853                 pagevec_release(&pvec);
 854                 cond_resched();
 855                 index++;
 856         }
 857
 858         if (partial_start) {
 859                 struct page *page = NULL;
 860                 shmem_getpage(inode, start - 1, &page, SGP_READ);
 861                 if (page) {
 862                         unsigned int top = PAGE_SIZE;
 863                         if (start > end) {
 864                                 top = partial_end;
 865                                 partial_end = 0;
 866                         }
 867                         zero_user_segment(page, partial_start, top);
 868                         set_page_dirty(page);
 869                         unlock_page(page);
 870                         put_page(page);
 871                 }
 872         }
 873         if (partial_end) {
 874                 struct page *page = NULL;
 875                 shmem_getpage(inode, end, &page, SGP_READ);
 876                 if (page) {
 877                         zero_user_segment(page, 0, partial_end);
 878                         set_page_dirty(page);
 879                         unlock_page(page);
 880                         put_page(page);
 881                 }
 882         }
 883         if (start >= end)
 884                 return;
 885
 886         index = start;
 887         while (index < end) {
 888                 cond_resched();
 889
 890                 pvec.nr = find_get_entries(mapping, index,
 891                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 892                                 pvec.pages, indices);
 893                 if (!pvec.nr) {
 894                         /* If all gone or hole-punch or unfalloc, we're done */
 895                         if (index == start || end != -1)
 896                                 break;
 897                         /* But if truncating, restart to make sure all gone */
 898                         index = start;
 899                         continue;
 900                 }
 901                 for (i = 0; i < pagevec_count(&pvec); i++) {
 902                         struct page *page = pvec.pages[i];
 903
 904                         index = indices[i];
 905                         if (index >= end)
 906                                 break;
 907
 908                         if (xa_is_value(page)) {
 909                                 if (unfalloc)
 910                                         continue;
 911                                 if (shmem_free_swap(mapping, index, page)) {
 912                                         /* Swap was replaced by page: retry */
 913                                         index--;
 914                                         break;
 915                                 }
 916                                 nr_swaps_freed++;
 917                                 continue;
 918                         }
 919
 920                         lock_page(page);
 921
 922                         if (PageTransTail(page)) {
 923                                 /* Middle of THP: zero out the page */
 924                                 clear_highpage(page);
 925                                 unlock_page(page);
 926                                 /*
 927                                  * Partial thp truncate due 'start' in middle
 928                                  * of THP: don't need to look on these pages
 929                                  * again on !pvec.nr restart.
 930                                  */
 931                                 if (index != round_down(end, HPAGE_PMD_NR))
 932                                         start++;
 933                                 continue;
 934                         } else if (PageTransHuge(page)) {
 935                                 if (index == round_down(end, HPAGE_PMD_NR)) {
 936                                         /*
 937                                          * Range ends in the middle of THP:
 938                                          * zero out the page
 939                                          */
 940                                         clear_highpage(page);
 941                                         unlock_page(page);
 942                                         continue;
 943                                 }
 944                                 index += HPAGE_PMD_NR - 1;
 945                                 i += HPAGE_PMD_NR - 1;
 946                         }
 947
 948                         if (!unfalloc || !PageUptodate(page)) {
 949                                 VM_BUG_ON_PAGE(PageTail(page), page);
 950                                 if (page_mapping(page) == mapping) {
 951                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
 952                                         truncate_inode_page(mapping, page);
 953                                 } else {
 954                                         /* Page was replaced by swap: retry */
 955                                         unlock_page(page);
 956                                         index--;
 957                                         break;
 958                                 }
 959                         }
 960                         unlock_page(page);
 961                 }
 962                 pagevec_remove_exceptionals(&pvec);
 963                 pagevec_release(&pvec);
 964                 index++;
 965         }
 966
 967         spin_lock_irq(&info->lock);
 968         info->swapped -= nr_swaps_freed;
 969         shmem_recalc_inode(inode);
 970         spin_unlock_irq(&info->lock);
 971 }
 972
 973 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
 974 {
 975         shmem_undo_range(inode, lstart, lend, false);
 976         inode->i_ctime = inode->i_mtime = current_time(inode);
 977 }
 978 EXPORT_SYMBOL_GPL(shmem_truncate_range);
 979
 980 static int shmem_getattr(const struct path *path, struct kstat *stat,
 981                          u32 request_mask, unsigned int query_flags)
 982 {
 983         struct inode *inode = path->dentry->d_inode;
 984         struct shmem_inode_info *info = SHMEM_I(inode);
 985         struct shmem_sb_info *sb_info = SHMEM_SB(inode->i_sb);
 986
 987         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
 988                 spin_lock_irq(&info->lock);
 989                 shmem_recalc_inode(inode);
 990                 spin_unlock_irq(&info->lock);
 991         }
 992         generic_fillattr(inode, stat);
 993
 994         if (is_huge_enabled(sb_info))
 995                 stat->blksize = HPAGE_PMD_SIZE;
 996
 997         return 0;
 998 }
 999
1000 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1001 {
1002         struct inode *inode = d_inode(dentry);
1003         struct shmem_inode_info *info = SHMEM_I(inode);
1004         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1005         int error;
1006
1007         error = setattr_prepare(dentry, attr);
1008         if (error)
1009                 return error;
1010
1011         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1012                 loff_t oldsize = inode->i_size;
1013                 loff_t newsize = attr->ia_size;
1014
1015                 /* protected by i_mutex */
1016                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1017                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1018                         return -EPERM;
1019
1020                 if (newsize != oldsize) {
1021                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1022                                         oldsize, newsize);
1023                         if (error)
1024                                 return error;
1025                         i_size_write(inode, newsize);
1026                         inode->i_ctime = inode->i_mtime = current_time(inode);
1027                 }
1028                 if (newsize <= oldsize) {
1029                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1030                         if (oldsize > holebegin)
1031                                 unmap_mapping_range(inode->i_mapping,
1032                                                         holebegin, 0, 1);
1033                         if (info->alloced)
1034                                 shmem_truncate_range(inode,
1035                                                         newsize, (loff_t)-1);
1036                         /* unmap again to remove racily COWed private pages */
1037                         if (oldsize > holebegin)
1038                                 unmap_mapping_range(inode->i_mapping,
1039                                                         holebegin, 0, 1);
1040
1041                         /*
1042                          * Part of the huge page can be beyond i_size: subject
1043                          * to shrink under memory pressure.
1044                          */
1045                         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
1046                                 spin_lock(&sbinfo->shrinklist_lock);
1047                                 /*
1048                                  * _careful to defend against unlocked access to
1049                                  * ->shrink_list in shmem_unused_huge_shrink()
1050                                  */
1051                                 if (list_empty_careful(&info->shrinklist)) {
1052                                         list_add_tail(&info->shrinklist,
1053                                                         &sbinfo->shrinklist);
1054                                         sbinfo->shrinklist_len++;
1055                                 }
1056                                 spin_unlock(&sbinfo->shrinklist_lock);
1057                         }
1058                 }
1059         }
1060
1061         setattr_copy(inode, attr);
1062         if (attr->ia_valid & ATTR_MODE)
1063                 error = posix_acl_chmod(inode, inode->i_mode);
1064         return error;
1065 }
1066
1067 static void shmem_evict_inode(struct inode *inode)
1068 {
1069         struct shmem_inode_info *info = SHMEM_I(inode);
1070         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1071
1072         if (inode->i_mapping->a_ops == &shmem_aops) {
1073                 shmem_unacct_size(info->flags, inode->i_size);
1074                 inode->i_size = 0;
1075                 shmem_truncate_range(inode, 0, (loff_t)-1);
1076                 if (!list_empty(&info->shrinklist)) {
1077                         spin_lock(&sbinfo->shrinklist_lock);
1078                         if (!list_empty(&info->shrinklist)) {
1079                                 list_del_init(&info->shrinklist);
1080                                 sbinfo->shrinklist_len--;
1081                         }
1082                         spin_unlock(&sbinfo->shrinklist_lock);
1083                 }
1084                 if (!list_empty(&info->swaplist)) {
1085                         mutex_lock(&shmem_swaplist_mutex);
1086                         list_del_init(&info->swaplist);
1087                         mutex_unlock(&shmem_swaplist_mutex);
1088                 }
1089         }
1090
1091         simple_xattrs_free(&info->xattrs);
1092         WARN_ON(inode->i_blocks);
1093         shmem_free_inode(inode->i_sb);
1094         clear_inode(inode);
1095 }
1096
1097 extern struct swap_info_struct *swap_info[];
1098
1099 static int shmem_find_swap_entries(struct address_space *mapping,
1100                                    pgoff_t start, unsigned int nr_entries,
1101                                    struct page **entries, pgoff_t *indices,
1102                                    unsigned int type, bool frontswap)
1103 {
1104         XA_STATE(xas, &mapping->i_pages, start);
1105         struct page *page;
1106         swp_entry_t entry;
1107         unsigned int ret = 0;
1108
1109         if (!nr_entries)
1110                 return 0;
1111
1112         rcu_read_lock();
1113         xas_for_each(&xas, page, ULONG_MAX) {
1114                 if (xas_retry(&xas, page))
1115                         continue;
1116
1117                 if (!xa_is_value(page))
1118                         continue;
1119
1120                 entry = radix_to_swp_entry(page);
1121                 if (swp_type(entry) != type)
1122                         continue;
1123                 if (frontswap &&
1124                     !frontswap_test(swap_info[type], swp_offset(entry)))
1125                         continue;
1126
1127                 indices[ret] = xas.xa_index;
1128                 entries[ret] = page;
1129
1130                 if (need_resched()) {
1131                         xas_pause(&xas);
1132                         cond_resched_rcu();
1133                 }
1134                 if (++ret == nr_entries)
1135                         break;
1136         }
1137         rcu_read_unlock();
1138
1139         return ret;
1140 }
1141
1142 /*
1143  * Move the swapped pages for an inode to page cache. Returns the count
1144  * of pages swapped in, or the error in case of failure.
1145  */
1146 static int shmem_unuse_swap_entries(struct inode *inode, struct pagevec pvec,
1147                                     pgoff_t *indices)
1148 {
1149         int i = 0;
1150         int ret = 0;
1151         int error = 0;
1152         struct address_space *mapping = inode->i_mapping;
1153
1154         for (i = 0; i < pvec.nr; i++) {
1155                 struct page *page = pvec.pages[i];
1156
1157                 if (!xa_is_value(page))
1158                         continue;
1159                 error = shmem_swapin_page(inode, indices[i],
1160                                           &page, SGP_CACHE,
1161                                           mapping_gfp_mask(mapping),
1162                                           NULL, NULL);
1163                 if (error == 0) {
1164                         unlock_page(page);
1165                         put_page(page);
1166                         ret++;
1167                 }
1168                 if (error == -ENOMEM)
1169                         break;
1170                 error = 0;
1171         }
1172         return error ? error : ret;
1173 }
1174
1175 /*
1176  * If swap found in inode, free it and move page from swapcache to filecache.
1177  */
1178 static int shmem_unuse_inode(struct inode *inode, unsigned int type,
1179                              bool frontswap, unsigned long *fs_pages_to_unuse)
1180 {
1181         struct address_space *mapping = inode->i_mapping;
1182         pgoff_t start = 0;
1183         struct pagevec pvec;
1184         pgoff_t indices[PAGEVEC_SIZE];
1185         bool frontswap_partial = (frontswap && *fs_pages_to_unuse > 0);
1186         int ret = 0;
1187
1188         pagevec_init(&pvec);
1189         do {
1190                 unsigned int nr_entries = PAGEVEC_SIZE;
1191
1192                 if (frontswap_partial && *fs_pages_to_unuse < PAGEVEC_SIZE)
1193                         nr_entries = *fs_pages_to_unuse;
1194
1195                 pvec.nr = shmem_find_swap_entries(mapping, start, nr_entries,
1196                                                   pvec.pages, indices,
1197                                                   type, frontswap);
1198                 if (pvec.nr == 0) {
1199                         ret = 0;
1200                         break;
1201                 }
1202
1203                 ret = shmem_unuse_swap_entries(inode, pvec, indices);
1204                 if (ret < 0)
1205                         break;
1206
1207                 if (frontswap_partial) {
1208                         *fs_pages_to_unuse -= ret;
1209                         if (*fs_pages_to_unuse == 0) {
1210                                 ret = FRONTSWAP_PAGES_UNUSED;
1211                                 break;
1212                         }
1213                 }
1214
1215                 start = indices[pvec.nr - 1];
1216         } while (true);
1217
1218         return ret;
1219 }
1220
1221 /*
1222  * Read all the shared memory data that resides in the swap
1223  * device 'type' back into memory, so the swap device can be
1224  * unused.
1225  */
1226 int shmem_unuse(unsigned int type, bool frontswap,
1227                 unsigned long *fs_pages_to_unuse)
1228 {
1229         struct shmem_inode_info *info, *next;
1230         struct inode *inode;
1231         struct inode *prev_inode = NULL;
1232         int error = 0;
1233
1234         if (list_empty(&shmem_swaplist))
1235                 return 0;
1236
1237         mutex_lock(&shmem_swaplist_mutex);
1238
1239         /*
1240          * The extra refcount on the inode is necessary to safely dereference
1241          * p->next after re-acquiring the lock. New shmem inodes with swap
1242          * get added to the end of the list and we will scan them all.
1243          */
1244         list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1245                 if (!info->swapped) {
1246                         list_del_init(&info->swaplist);
1247                         continue;
1248                 }
1249
1250                 inode = igrab(&info->vfs_inode);
1251                 if (!inode)
1252                         continue;
1253
1254                 mutex_unlock(&shmem_swaplist_mutex);
1255                 if (prev_inode)
1256                         iput(prev_inode);
1257                 prev_inode = inode;
1258
1259                 error = shmem_unuse_inode(inode, type, frontswap,
1260                                           fs_pages_to_unuse);
1261                 cond_resched();
1262
1263                 mutex_lock(&shmem_swaplist_mutex);
1264                 next = list_next_entry(info, swaplist);
1265                 if (!info->swapped)
1266                         list_del_init(&info->swaplist);
1267                 if (error)
1268                         break;
1269         }
1270         mutex_unlock(&shmem_swaplist_mutex);
1271
1272         if (prev_inode)
1273                 iput(prev_inode);
1274
1275         return error;
1276 }
1277
1278 /*
1279  * Move the page from the page cache to the swap cache.
1280  */
1281 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1282 {
1283         struct shmem_inode_info *info;
1284         struct address_space *mapping;
1285         struct inode *inode;
1286         swp_entry_t swap;
1287         pgoff_t index;
1288
1289         VM_BUG_ON_PAGE(PageCompound(page), page);
1290         BUG_ON(!PageLocked(page));
1291         mapping = page->mapping;
1292         index = page->index;
1293         inode = mapping->host;
1294         info = SHMEM_I(inode);
1295         if (info->flags & VM_LOCKED)
1296                 goto redirty;
1297         if (!total_swap_pages)
1298                 goto redirty;
1299
1300         /*
1301          * Our capabilities prevent regular writeback or sync from ever calling
1302          * shmem_writepage; but a stacking filesystem might use ->writepage of
1303          * its underlying filesystem, in which case tmpfs should write out to
1304          * swap only in response to memory pressure, and not for the writeback
1305          * threads or sync.
1306          */
1307         if (!wbc->for_reclaim) {
1308                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1309                 goto redirty;
1310         }
1311
1312         /*
1313          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1314          * value into swapfile.c, the only way we can correctly account for a
1315          * fallocated page arriving here is now to initialize it and write it.
1316          *
1317          * That's okay for a page already fallocated earlier, but if we have
1318          * not yet completed the fallocation, then (a) we want to keep track
1319          * of this page in case we have to undo it, and (b) it may not be a
1320          * good idea to continue anyway, once we're pushing into swap.  So
1321          * reactivate the page, and let shmem_fallocate() quit when too many.
1322          */
1323         if (!PageUptodate(page)) {
1324                 if (inode->i_private) {
1325                         struct shmem_falloc *shmem_falloc;
1326                         spin_lock(&inode->i_lock);
1327                         shmem_falloc = inode->i_private;
1328                         if (shmem_falloc &&
1329                             !shmem_falloc->waitq &&
1330                             index >= shmem_falloc->start &&
1331                             index < shmem_falloc->next)
1332                                 shmem_falloc->nr_unswapped++;
1333                         else
1334                                 shmem_falloc = NULL;
1335                         spin_unlock(&inode->i_lock);
1336                         if (shmem_falloc)
1337                                 goto redirty;
1338                 }
1339                 clear_highpage(page);
1340                 flush_dcache_page(page);
1341                 SetPageUptodate(page);
1342         }
1343
1344         swap = get_swap_page(page);
1345         if (!swap.val)
1346                 goto redirty;
1347
1348         /*
1349          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1350          * if it's not already there.  Do it now before the page is
1351          * moved to swap cache, when its pagelock no longer protects
1352          * the inode from eviction.  But don't unlock the mutex until
1353          * we've incremented swapped, because shmem_unuse_inode() will
1354          * prune a !swapped inode from the swaplist under this mutex.
1355          */
1356         mutex_lock(&shmem_swaplist_mutex);
1357         if (list_empty(&info->swaplist))
1358                 list_add(&info->swaplist, &shmem_swaplist);
1359
1360         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1361                 spin_lock_irq(&info->lock);
1362                 shmem_recalc_inode(inode);
1363                 info->swapped++;
1364                 spin_unlock_irq(&info->lock);
1365
1366                 swap_shmem_alloc(swap);
1367                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1368
1369                 mutex_unlock(&shmem_swaplist_mutex);
1370                 BUG_ON(page_mapped(page));
1371                 swap_writepage(page, wbc);
1372                 return 0;
1373         }
1374
1375         mutex_unlock(&shmem_swaplist_mutex);
1376         put_swap_page(page, swap);
1377 redirty:
1378         set_page_dirty(page);
1379         if (wbc->for_reclaim)
1380                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1381         unlock_page(page);
1382         return 0;
1383 }
1384
1385 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1386 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1387 {
1388         char buffer[64];
1389
1390         if (!mpol || mpol->mode == MPOL_DEFAULT)
1391                 return;         /* show nothing */
1392
1393         mpol_to_str(buffer, sizeof(buffer), mpol);
1394
1395         seq_printf(seq, ",mpol=%s", buffer);
1396 }
1397
1398 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1399 {
1400         struct mempolicy *mpol = NULL;
1401         if (sbinfo->mpol) {
1402                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1403                 mpol = sbinfo->mpol;
1404                 mpol_get(mpol);
1405                 spin_unlock(&sbinfo->stat_lock);
1406         }
1407         return mpol;
1408 }
1409 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1410 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1411 {
1412 }
1413 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1414 {
1415         return NULL;
1416 }
1417 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1418 #ifndef CONFIG_NUMA
1419 #define vm_policy vm_private_data
1420 #endif
1421
1422 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1423                 struct shmem_inode_info *info, pgoff_t index)
1424 {
1425         /* Create a pseudo vma that just contains the policy */
1426         vma_init(vma, NULL);
1427         /* Bias interleave by inode number to distribute better across nodes */
1428         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1429         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1430 }
1431
1432 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1433 {
1434         /* Drop reference taken by mpol_shared_policy_lookup() */
1435         mpol_cond_put(vma->vm_policy);
1436 }
1437
1438 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1439                         struct shmem_inode_info *info, pgoff_t index)
1440 {
1441         struct vm_area_struct pvma;
1442         struct page *page;
1443         struct vm_fault vmf;
1444
1445         shmem_pseudo_vma_init(&pvma, info, index);
1446         vmf.vma = &pvma;
1447         vmf.address = 0;
1448         page = swap_cluster_readahead(swap, gfp, &vmf);
1449         shmem_pseudo_vma_destroy(&pvma);
1450
1451         return page;
1452 }
1453
1454 static struct page *shmem_alloc_hugepage(gfp_t gfp,
1455                 struct shmem_inode_info *info, pgoff_t index)
1456 {
1457         struct vm_area_struct pvma;
1458         struct address_space *mapping = info->vfs_inode.i_mapping;
1459         pgoff_t hindex;
1460         struct page *page;
1461
1462         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1463                 return NULL;
1464
1465         hindex = round_down(index, HPAGE_PMD_NR);
1466         if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1467                                                                 XA_PRESENT))
1468                 return NULL;
1469
1470         shmem_pseudo_vma_init(&pvma, info, hindex);
1471         page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1472                         HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1473         shmem_pseudo_vma_destroy(&pvma);
1474         if (page)
1475                 prep_transhuge_page(page);
1476         return page;
1477 }
1478
1479 static struct page *shmem_alloc_page(gfp_t gfp,
1480                         struct shmem_inode_info *info, pgoff_t index)
1481 {
1482         struct vm_area_struct pvma;
1483         struct page *page;
1484
1485         shmem_pseudo_vma_init(&pvma, info, index);
1486         page = alloc_page_vma(gfp, &pvma, 0);
1487         shmem_pseudo_vma_destroy(&pvma);
1488
1489         return page;
1490 }
1491
1492 static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1493                 struct inode *inode,
1494                 pgoff_t index, bool huge)
1495 {
1496         struct shmem_inode_info *info = SHMEM_I(inode);
1497         struct page *page;
1498         int nr;
1499         int err = -ENOSPC;
1500
1501         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1502                 huge = false;
1503         nr = huge ? HPAGE_PMD_NR : 1;
1504
1505         if (!shmem_inode_acct_block(inode, nr))
1506                 goto failed;
1507
1508         if (huge)
1509                 page = shmem_alloc_hugepage(gfp, info, index);
1510         else
1511                 page = shmem_alloc_page(gfp, info, index);
1512         if (page) {
1513                 __SetPageLocked(page);
1514                 __SetPageSwapBacked(page);
1515                 return page;
1516         }
1517
1518         err = -ENOMEM;
1519         shmem_inode_unacct_blocks(inode, nr);
1520 failed:
1521         return ERR_PTR(err);
1522 }
1523
1524 /*
1525  * When a page is moved from swapcache to shmem filecache (either by the
1526  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1527  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1528  * ignorance of the mapping it belongs to.  If that mapping has special
1529  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1530  * we may need to copy to a suitable page before moving to filecache.
1531  *
1532  * In a future release, this may well be extended to respect cpuset and
1533  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1534  * but for now it is a simple matter of zone.
1535  */
1536 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1537 {
1538         return page_zonenum(page) > gfp_zone(gfp);
1539 }
1540
1541 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1542                                 struct shmem_inode_info *info, pgoff_t index)
1543 {
1544         struct page *oldpage, *newpage;
1545         struct address_space *swap_mapping;
1546         swp_entry_t entry;
1547         pgoff_t swap_index;
1548         int error;
1549
1550         oldpage = *pagep;
1551         entry.val = page_private(oldpage);
1552         swap_index = swp_offset(entry);
1553         swap_mapping = page_mapping(oldpage);
1554
1555         /*
1556          * We have arrived here because our zones are constrained, so don't
1557          * limit chance of success by further cpuset and node constraints.
1558          */
1559         gfp &= ~GFP_CONSTRAINT_MASK;
1560         newpage = shmem_alloc_page(gfp, info, index);
1561         if (!newpage)
1562                 return -ENOMEM;
1563
1564         get_page(newpage);
1565         copy_highpage(newpage, oldpage);
1566         flush_dcache_page(newpage);
1567
1568         __SetPageLocked(newpage);
1569         __SetPageSwapBacked(newpage);
1570         SetPageUptodate(newpage);
1571         set_page_private(newpage, entry.val);
1572         SetPageSwapCache(newpage);
1573
1574         /*
1575          * Our caller will very soon move newpage out of swapcache, but it's
1576          * a nice clean interface for us to replace oldpage by newpage there.
1577          */
1578         xa_lock_irq(&swap_mapping->i_pages);
1579         error = shmem_replace_entry(swap_mapping, swap_index, oldpage, newpage);
1580         if (!error) {
1581                 __inc_node_page_state(newpage, NR_FILE_PAGES);
1582                 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1583         }
1584         xa_unlock_irq(&swap_mapping->i_pages);
1585
1586         if (unlikely(error)) {
1587                 /*
1588                  * Is this possible?  I think not, now that our callers check
1589                  * both PageSwapCache and page_private after getting page lock;
1590                  * but be defensive.  Reverse old to newpage for clear and free.
1591                  */
1592                 oldpage = newpage;
1593         } else {
1594                 mem_cgroup_migrate(oldpage, newpage);
1595                 lru_cache_add_anon(newpage);
1596                 *pagep = newpage;
1597         }
1598
1599         ClearPageSwapCache(oldpage);
1600         set_page_private(oldpage, 0);
1601
1602         unlock_page(oldpage);
1603         put_page(oldpage);
1604         put_page(oldpage);
1605         return error;
1606 }
1607
1608 /*
1609  * Swap in the page pointed to by *pagep.
1610  * Caller has to make sure that *pagep contains a valid swapped page.
1611  * Returns 0 and the page in pagep if success. On failure, returns the
1612  * the error code and NULL in *pagep.
1613  */
1614 static int shmem_swapin_page(struct inode *inode, pgoff_t index,
1615                              struct page **pagep, enum sgp_type sgp,
1616                              gfp_t gfp, struct vm_area_struct *vma,
1617                              vm_fault_t *fault_type)
1618 {
1619         struct address_space *mapping = inode->i_mapping;
1620         struct shmem_inode_info *info = SHMEM_I(inode);
1621         struct mm_struct *charge_mm = vma ? vma->vm_mm : current->mm;
1622         struct mem_cgroup *memcg;
1623         struct page *page;
1624         swp_entry_t swap;
1625         int error;
1626
1627         VM_BUG_ON(!*pagep || !xa_is_value(*pagep));
1628         swap = radix_to_swp_entry(*pagep);
1629         *pagep = NULL;
1630
1631         /* Look it up and read it in.. */
1632         page = lookup_swap_cache(swap, NULL, 0);
1633         if (!page) {
1634                 /* Or update major stats only when swapin succeeds?? */
1635                 if (fault_type) {
1636                         *fault_type |= VM_FAULT_MAJOR;
1637                         count_vm_event(PGMAJFAULT);
1638                         count_memcg_event_mm(charge_mm, PGMAJFAULT);
1639                 }
1640                 /* Here we actually start the io */
1641                 page = shmem_swapin(swap, gfp, info, index);
1642                 if (!page) {
1643                         error = -ENOMEM;
1644                         goto failed;
1645                 }
1646         }
1647
1648         /* We have to do this with page locked to prevent races */
1649         lock_page(page);
1650         if (!PageSwapCache(page) || page_private(page) != swap.val ||
1651             !shmem_confirm_swap(mapping, index, swap)) {
1652                 error = -EEXIST;
1653                 goto unlock;
1654         }
1655         if (!PageUptodate(page)) {
1656                 error = -EIO;
1657                 goto failed;
1658         }
1659         wait_on_page_writeback(page);
1660
1661         if (shmem_should_replace_page(page, gfp)) {
1662                 error = shmem_replace_page(&page, gfp, info, index);
1663                 if (error)
1664                         goto failed;
1665         }
1666
1667         error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
1668                                             false);
1669         if (!error) {
1670                 error = shmem_add_to_page_cache(page, mapping, index,
1671                                                 swp_to_radix_entry(swap), gfp);
1672                 /*
1673                  * We already confirmed swap under page lock, and make
1674                  * no memory allocation here, so usually no possibility
1675                  * of error; but free_swap_and_cache() only trylocks a
1676                  * page, so it is just possible that the entry has been
1677                  * truncated or holepunched since swap was confirmed.
1678                  * shmem_undo_range() will have done some of the
1679                  * unaccounting, now delete_from_swap_cache() will do
1680                  * the rest.
1681                  */
1682                 if (error) {
1683                         mem_cgroup_cancel_charge(page, memcg, false);
1684                         delete_from_swap_cache(page);
1685                 }
1686         }
1687         if (error)
1688                 goto failed;
1689
1690         mem_cgroup_commit_charge(page, memcg, true, false);
1691
1692         spin_lock_irq(&info->lock);
1693         info->swapped--;
1694         shmem_recalc_inode(inode);
1695         spin_unlock_irq(&info->lock);
1696
1697         if (sgp == SGP_WRITE)
1698                 mark_page_accessed(page);
1699
1700         delete_from_swap_cache(page);
1701         set_page_dirty(page);
1702         swap_free(swap);
1703
1704         *pagep = page;
1705         return 0;
1706 failed:
1707         if (!shmem_confirm_swap(mapping, index, swap))
1708                 error = -EEXIST;
1709 unlock:
1710         if (page) {
1711                 unlock_page(page);
1712                 put_page(page);
1713         }
1714
1715         return error;
1716 }
1717
1718 /*
1719  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1720  *
1721  * If we allocate a new one we do not mark it dirty. That's up to the
1722  * vm. If we swap it in we mark it dirty since we also free the swap
1723  * entry since a page cannot live in both the swap and page cache.
1724  *
1725  * fault_mm and fault_type are only supplied by shmem_fault:
1726  * otherwise they are NULL.
1727  */
1728 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1729         struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1730         struct vm_area_struct *vma, struct vm_fault *vmf,
1731                         vm_fault_t *fault_type)
1732 {
1733         struct address_space *mapping = inode->i_mapping;
1734         struct shmem_inode_info *info = SHMEM_I(inode);
1735         struct shmem_sb_info *sbinfo;
1736         struct mm_struct *charge_mm;
1737         struct mem_cgroup *memcg;
1738         struct page *page;
1739         enum sgp_type sgp_huge = sgp;
1740         pgoff_t hindex = index;
1741         int error;
1742         int once = 0;
1743         int alloced = 0;
1744
1745         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1746                 return -EFBIG;
1747         if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1748                 sgp = SGP_CACHE;
1749 repeat:
1750         if (sgp <= SGP_CACHE &&
1751             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1752                 return -EINVAL;
1753         }
1754
1755         sbinfo = SHMEM_SB(inode->i_sb);
1756         charge_mm = vma ? vma->vm_mm : current->mm;
1757
1758         page = find_lock_entry(mapping, index);
1759         if (xa_is_value(page)) {
1760                 error = shmem_swapin_page(inode, index, &page,
1761                                           sgp, gfp, vma, fault_type);
1762                 if (error == -EEXIST)
1763                         goto repeat;
1764
1765                 *pagep = page;
1766                 return error;
1767         }
1768
1769         if (page && sgp == SGP_WRITE)
1770                 mark_page_accessed(page);
1771
1772         /* fallocated page? */
1773         if (page && !PageUptodate(page)) {
1774                 if (sgp != SGP_READ)
1775                         goto clear;
1776                 unlock_page(page);
1777                 put_page(page);
1778                 page = NULL;
1779         }
1780         if (page || sgp == SGP_READ) {
1781                 *pagep = page;
1782                 return 0;
1783         }
1784
1785         /*
1786          * Fast cache lookup did not find it:
1787          * bring it back from swap or allocate.
1788          */
1789
1790         if (vma && userfaultfd_missing(vma)) {
1791                 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1792                 return 0;
1793         }
1794
1795         /* shmem_symlink() */
1796         if (mapping->a_ops != &shmem_aops)
1797                 goto alloc_nohuge;
1798         if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1799                 goto alloc_nohuge;
1800         if (shmem_huge == SHMEM_HUGE_FORCE)
1801                 goto alloc_huge;
1802         switch (sbinfo->huge) {
1803                 loff_t i_size;
1804                 pgoff_t off;
1805         case SHMEM_HUGE_NEVER:
1806                 goto alloc_nohuge;
1807         case SHMEM_HUGE_WITHIN_SIZE:
1808                 off = round_up(index, HPAGE_PMD_NR);
1809                 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1810                 if (i_size >= HPAGE_PMD_SIZE &&
1811                     i_size >> PAGE_SHIFT >= off)
1812                         goto alloc_huge;
1813                 /* fallthrough */
1814         case SHMEM_HUGE_ADVISE:
1815                 if (sgp_huge == SGP_HUGE)
1816                         goto alloc_huge;
1817                 /* TODO: implement fadvise() hints */
1818                 goto alloc_nohuge;
1819         }
1820
1821 alloc_huge:
1822         page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1823         if (IS_ERR(page)) {
1824 alloc_nohuge:
1825                 page = shmem_alloc_and_acct_page(gfp, inode,
1826                                                  index, false);
1827         }
1828         if (IS_ERR(page)) {
1829                 int retry = 5;
1830
1831                 error = PTR_ERR(page);
1832                 page = NULL;
1833                 if (error != -ENOSPC)
1834                         goto unlock;
1835                 /*
1836                  * Try to reclaim some space by splitting a huge page
1837                  * beyond i_size on the filesystem.
1838                  */
1839                 while (retry--) {
1840                         int ret;
1841
1842                         ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1843                         if (ret == SHRINK_STOP)
1844                                 break;
1845                         if (ret)
1846                                 goto alloc_nohuge;
1847                 }
1848                 goto unlock;
1849         }
1850
1851         if (PageTransHuge(page))
1852                 hindex = round_down(index, HPAGE_PMD_NR);
1853         else
1854                 hindex = index;
1855
1856         if (sgp == SGP_WRITE)
1857                 __SetPageReferenced(page);
1858
1859         error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
1860                                             PageTransHuge(page));
1861         if (error)
1862                 goto unacct;
1863         error = shmem_add_to_page_cache(page, mapping, hindex,
1864                                         NULL, gfp & GFP_RECLAIM_MASK);
1865         if (error) {
1866                 mem_cgroup_cancel_charge(page, memcg,
1867                                          PageTransHuge(page));
1868                 goto unacct;
1869         }
1870         mem_cgroup_commit_charge(page, memcg, false,
1871                                  PageTransHuge(page));
1872         lru_cache_add_anon(page);
1873
1874         spin_lock_irq(&info->lock);
1875         info->alloced += 1 << compound_order(page);
1876         inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1877         shmem_recalc_inode(inode);
1878         spin_unlock_irq(&info->lock);
1879         alloced = true;
1880
1881         if (PageTransHuge(page) &&
1882             DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1883                         hindex + HPAGE_PMD_NR - 1) {
1884                 /*
1885                  * Part of the huge page is beyond i_size: subject
1886                  * to shrink under memory pressure.
1887                  */
1888                 spin_lock(&sbinfo->shrinklist_lock);
1889                 /*
1890                  * _careful to defend against unlocked access to
1891                  * ->shrink_list in shmem_unused_huge_shrink()
1892                  */
1893                 if (list_empty_careful(&info->shrinklist)) {
1894                         list_add_tail(&info->shrinklist,
1895                                       &sbinfo->shrinklist);
1896                         sbinfo->shrinklist_len++;
1897                 }
1898                 spin_unlock(&sbinfo->shrinklist_lock);
1899         }
1900
1901         /*
1902          * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1903          */
1904         if (sgp == SGP_FALLOC)
1905                 sgp = SGP_WRITE;
1906 clear:
1907         /*
1908          * Let SGP_WRITE caller clear ends if write does not fill page;
1909          * but SGP_FALLOC on a page fallocated earlier must initialize
1910          * it now, lest undo on failure cancel our earlier guarantee.
1911          */
1912         if (sgp != SGP_WRITE && !PageUptodate(page)) {
1913                 struct page *head = compound_head(page);
1914                 int i;
1915
1916                 for (i = 0; i < (1 << compound_order(head)); i++) {
1917                         clear_highpage(head + i);
1918                         flush_dcache_page(head + i);
1919                 }
1920                 SetPageUptodate(head);
1921         }
1922
1923         /* Perhaps the file has been truncated since we checked */
1924         if (sgp <= SGP_CACHE &&
1925             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1926                 if (alloced) {
1927                         ClearPageDirty(page);
1928                         delete_from_page_cache(page);
1929                         spin_lock_irq(&info->lock);
1930                         shmem_recalc_inode(inode);
1931                         spin_unlock_irq(&info->lock);
1932                 }
1933                 error = -EINVAL;
1934                 goto unlock;
1935         }
1936         *pagep = page + index - hindex;
1937         return 0;
1938
1939         /*
1940          * Error recovery.
1941          */
1942 unacct:
1943         shmem_inode_unacct_blocks(inode, 1 << compound_order(page));
1944
1945         if (PageTransHuge(page)) {
1946                 unlock_page(page);
1947                 put_page(page);
1948                 goto alloc_nohuge;
1949         }
1950 unlock:
1951         if (page) {
1952                 unlock_page(page);
1953                 put_page(page);
1954         }
1955         if (error == -ENOSPC && !once++) {
1956                 spin_lock_irq(&info->lock);
1957                 shmem_recalc_inode(inode);
1958                 spin_unlock_irq(&info->lock);
1959                 goto repeat;
1960         }
1961         if (error == -EEXIST)
1962                 goto repeat;
1963         return error;
1964 }
1965
1966 /*
1967  * This is like autoremove_wake_function, but it removes the wait queue
1968  * entry unconditionally - even if something else had already woken the
1969  * target.
1970  */
1971 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1972 {
1973         int ret = default_wake_function(wait, mode, sync, key);
1974         list_del_init(&wait->entry);
1975         return ret;
1976 }
1977
1978 static vm_fault_t shmem_fault(struct vm_fault *vmf)
1979 {
1980         struct vm_area_struct *vma = vmf->vma;
1981         struct inode *inode = file_inode(vma->vm_file);
1982         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1983         enum sgp_type sgp;
1984         int err;
1985         vm_fault_t ret = VM_FAULT_LOCKED;
1986
1987         /*
1988          * Trinity finds that probing a hole which tmpfs is punching can
1989          * prevent the hole-punch from ever completing: which in turn
1990          * locks writers out with its hold on i_mutex.  So refrain from
1991          * faulting pages into the hole while it's being punched.  Although
1992          * shmem_undo_range() does remove the additions, it may be unable to
1993          * keep up, as each new page needs its own unmap_mapping_range() call,
1994          * and the i_mmap tree grows ever slower to scan if new vmas are added.
1995          *
1996          * It does not matter if we sometimes reach this check just before the
1997          * hole-punch begins, so that one fault then races with the punch:
1998          * we just need to make racing faults a rare case.
1999          *
2000          * The implementation below would be much simpler if we just used a
2001          * standard mutex or completion: but we cannot take i_mutex in fault,
2002          * and bloating every shmem inode for this unlikely case would be sad.
2003          */
2004         if (unlikely(inode->i_private)) {
2005                 struct shmem_falloc *shmem_falloc;
2006
2007                 spin_lock(&inode->i_lock);
2008                 shmem_falloc = inode->i_private;
2009                 if (shmem_falloc &&
2010                     shmem_falloc->waitq &&
2011                     vmf->pgoff >= shmem_falloc->start &&
2012                     vmf->pgoff < shmem_falloc->next) {
2013                         wait_queue_head_t *shmem_falloc_waitq;
2014                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2015
2016                         ret = VM_FAULT_NOPAGE;
2017                         if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
2018                            !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
2019                                 /* It's polite to up mmap_sem if we can */
2020                                 up_read(&vma->vm_mm->mmap_sem);
2021                                 ret = VM_FAULT_RETRY;
2022                         }
2023
2024                         shmem_falloc_waitq = shmem_falloc->waitq;
2025                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2026                                         TASK_UNINTERRUPTIBLE);
2027                         spin_unlock(&inode->i_lock);
2028                         schedule();
2029
2030                         /*
2031                          * shmem_falloc_waitq points into the shmem_fallocate()
2032                          * stack of the hole-punching task: shmem_falloc_waitq
2033                          * is usually invalid by the time we reach here, but
2034                          * finish_wait() does not dereference it in that case;
2035                          * though i_lock needed lest racing with wake_up_all().
2036                          */
2037                         spin_lock(&inode->i_lock);
2038                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2039                         spin_unlock(&inode->i_lock);
2040                         return ret;
2041                 }
2042                 spin_unlock(&inode->i_lock);
2043         }
2044
2045         sgp = SGP_CACHE;
2046
2047         if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2048             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2049                 sgp = SGP_NOHUGE;
2050         else if (vma->vm_flags & VM_HUGEPAGE)
2051                 sgp = SGP_HUGE;
2052
2053         err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2054                                   gfp, vma, vmf, &ret);
2055         if (err)
2056                 return vmf_error(err);
2057         return ret;
2058 }
2059
2060 unsigned long shmem_get_unmapped_area(struct file *file,
2061                                       unsigned long uaddr, unsigned long len,
2062                                       unsigned long pgoff, unsigned long flags)
2063 {
2064         unsigned long (*get_area)(struct file *,
2065                 unsigned long, unsigned long, unsigned long, unsigned long);
2066         unsigned long addr;
2067         unsigned long offset;
2068         unsigned long inflated_len;
2069         unsigned long inflated_addr;
2070         unsigned long inflated_offset;
2071
2072         if (len > TASK_SIZE)
2073                 return -ENOMEM;
2074
2075         get_area = current->mm->get_unmapped_area;
2076         addr = get_area(file, uaddr, len, pgoff, flags);
2077
2078         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
2079                 return addr;
2080         if (IS_ERR_VALUE(addr))
2081                 return addr;
2082         if (addr & ~PAGE_MASK)
2083                 return addr;
2084         if (addr > TASK_SIZE - len)
2085                 return addr;
2086
2087         if (shmem_huge == SHMEM_HUGE_DENY)
2088                 return addr;
2089         if (len < HPAGE_PMD_SIZE)
2090                 return addr;
2091         if (flags & MAP_FIXED)
2092                 return addr;
2093         /*
2094          * Our priority is to support MAP_SHARED mapped hugely;
2095          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2096          * But if caller specified an address hint, respect that as before.
2097          */
2098         if (uaddr)
2099                 return addr;
2100
2101         if (shmem_huge != SHMEM_HUGE_FORCE) {
2102                 struct super_block *sb;
2103
2104                 if (file) {
2105                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2106                         sb = file_inode(file)->i_sb;
2107                 } else {
2108                         /*
2109                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2110                          * for "/dev/zero", to create a shared anonymous object.
2111                          */
2112                         if (IS_ERR(shm_mnt))
2113                                 return addr;
2114                         sb = shm_mnt->mnt_sb;
2115                 }
2116                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2117                         return addr;
2118         }
2119
2120         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2121         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2122                 return addr;
2123         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2124                 return addr;
2125
2126         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2127         if (inflated_len > TASK_SIZE)
2128                 return addr;
2129         if (inflated_len < len)
2130                 return addr;
2131
2132         inflated_addr = get_area(NULL, 0, inflated_len, 0, flags);
2133         if (IS_ERR_VALUE(inflated_addr))
2134                 return addr;
2135         if (inflated_addr & ~PAGE_MASK)
2136                 return addr;
2137
2138         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2139         inflated_addr += offset - inflated_offset;
2140         if (inflated_offset > offset)
2141                 inflated_addr += HPAGE_PMD_SIZE;
2142
2143         if (inflated_addr > TASK_SIZE - len)
2144                 return addr;
2145         return inflated_addr;
2146 }
2147
2148 #ifdef CONFIG_NUMA
2149 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2150 {
2151         struct inode *inode = file_inode(vma->vm_file);
2152         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2153 }
2154
2155 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2156                                           unsigned long addr)
2157 {
2158         struct inode *inode = file_inode(vma->vm_file);
2159         pgoff_t index;
2160
2161         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2162         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2163 }
2164 #endif
2165
2166 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2167 {
2168         struct inode *inode = file_inode(file);
2169         struct shmem_inode_info *info = SHMEM_I(inode);
2170         int retval = -ENOMEM;
2171
2172         spin_lock_irq(&info->lock);
2173         if (lock && !(info->flags & VM_LOCKED)) {
2174                 if (!user_shm_lock(inode->i_size, user))
2175                         goto out_nomem;
2176                 info->flags |= VM_LOCKED;
2177                 mapping_set_unevictable(file->f_mapping);
2178         }
2179         if (!lock && (info->flags & VM_LOCKED) && user) {
2180                 user_shm_unlock(inode->i_size, user);
2181                 info->flags &= ~VM_LOCKED;
2182                 mapping_clear_unevictable(file->f_mapping);
2183         }
2184         retval = 0;
2185
2186 out_nomem:
2187         spin_unlock_irq(&info->lock);
2188         return retval;
2189 }
2190
2191 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2192 {
2193         struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2194
2195         if (info->seals & F_SEAL_FUTURE_WRITE) {
2196                 /*
2197                  * New PROT_WRITE and MAP_SHARED mmaps are not allowed when
2198                  * "future write" seal active.
2199                  */
2200                 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
2201                         return -EPERM;
2202
2203                 /*
2204                  * Since the F_SEAL_FUTURE_WRITE seals allow for a MAP_SHARED
2205                  * read-only mapping, take care to not allow mprotect to revert
2206                  * protections.
2207                  */
2208                 vma->vm_flags &= ~(VM_MAYWRITE);
2209         }
2210
2211         file_accessed(file);
2212         vma->vm_ops = &shmem_vm_ops;
2213         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2214                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2215                         (vma->vm_end & HPAGE_PMD_MASK)) {
2216                 khugepaged_enter(vma, vma->vm_flags);
2217         }
2218         return 0;
2219 }
2220
2221 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2222                                      umode_t mode, dev_t dev, unsigned long flags)
2223 {
2224         struct inode *inode;
2225         struct shmem_inode_info *info;
2226         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2227
2228         if (shmem_reserve_inode(sb))
2229                 return NULL;
2230
2231         inode = new_inode(sb);
2232         if (inode) {
2233                 inode->i_ino = get_next_ino();
2234                 inode_init_owner(inode, dir, mode);
2235                 inode->i_blocks = 0;
2236                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2237                 inode->i_generation = prandom_u32();
2238                 info = SHMEM_I(inode);
2239                 memset(info, 0, (char *)inode - (char *)info);
2240                 spin_lock_init(&info->lock);
2241                 info->seals = F_SEAL_SEAL;
2242                 info->flags = flags & VM_NORESERVE;
2243                 INIT_LIST_HEAD(&info->shrinklist);
2244                 INIT_LIST_HEAD(&info->swaplist);
2245                 simple_xattrs_init(&info->xattrs);
2246                 cache_no_acl(inode);
2247
2248                 switch (mode & S_IFMT) {
2249                 default:
2250                         inode->i_op = &shmem_special_inode_operations;
2251                         init_special_inode(inode, mode, dev);
2252                         break;
2253                 case S_IFREG:
2254                         inode->i_mapping->a_ops = &shmem_aops;
2255                         inode->i_op = &shmem_inode_operations;
2256                         inode->i_fop = &shmem_file_operations;
2257                         mpol_shared_policy_init(&info->policy,
2258                                                  shmem_get_sbmpol(sbinfo));
2259                         break;
2260                 case S_IFDIR:
2261                         inc_nlink(inode);
2262                         /* Some things misbehave if size == 0 on a directory */
2263                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2264                         inode->i_op = &shmem_dir_inode_operations;
2265                         inode->i_fop = &simple_dir_operations;
2266                         break;
2267                 case S_IFLNK:
2268                         /*
2269                          * Must not load anything in the rbtree,
2270                          * mpol_free_shared_policy will not be called.
2271                          */
2272                         mpol_shared_policy_init(&info->policy, NULL);
2273                         break;
2274                 }
2275
2276                 lockdep_annotate_inode_mutex_key(inode);
2277         } else
2278                 shmem_free_inode(sb);
2279         return inode;
2280 }
2281
2282 bool shmem_mapping(struct address_space *mapping)
2283 {
2284         return mapping->a_ops == &shmem_aops;
2285 }
2286
2287 static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2288                                   pmd_t *dst_pmd,
2289                                   struct vm_area_struct *dst_vma,
2290                                   unsigned long dst_addr,
2291                                   unsigned long src_addr,
2292                                   bool zeropage,
2293                                   struct page **pagep)
2294 {
2295         struct inode *inode = file_inode(dst_vma->vm_file);
2296         struct shmem_inode_info *info = SHMEM_I(inode);
2297         struct address_space *mapping = inode->i_mapping;
2298         gfp_t gfp = mapping_gfp_mask(mapping);
2299         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2300         struct mem_cgroup *memcg;
2301         spinlock_t *ptl;
2302         void *page_kaddr;
2303         struct page *page;
2304         pte_t _dst_pte, *dst_pte;
2305         int ret;
2306         pgoff_t offset, max_off;
2307
2308         ret = -ENOMEM;
2309         if (!shmem_inode_acct_block(inode, 1))
2310                 goto out;
2311
2312         if (!*pagep) {
2313                 page = shmem_alloc_page(gfp, info, pgoff);
2314                 if (!page)
2315                         goto out_unacct_blocks;
2316
2317                 if (!zeropage) {        /* mcopy_atomic */
2318                         page_kaddr = kmap_atomic(page);
2319                         ret = copy_from_user(page_kaddr,
2320                                              (const void __user *)src_addr,
2321                                              PAGE_SIZE);
2322                         kunmap_atomic(page_kaddr);
2323
2324                         /* fallback to copy_from_user outside mmap_sem */
2325                         if (unlikely(ret)) {
2326                                 *pagep = page;
2327                                 shmem_inode_unacct_blocks(inode, 1);
2328                                 /* don't free the page */
2329                                 return -ENOENT;
2330                         }
2331                 } else {                /* mfill_zeropage_atomic */
2332                         clear_highpage(page);
2333                 }
2334         } else {
2335                 page = *pagep;
2336                 *pagep = NULL;
2337         }
2338
2339         VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2340         __SetPageLocked(page);
2341         __SetPageSwapBacked(page);
2342         __SetPageUptodate(page);
2343
2344         ret = -EFAULT;
2345         offset = linear_page_index(dst_vma, dst_addr);
2346         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2347         if (unlikely(offset >= max_off))
2348                 goto out_release;
2349
2350         ret = mem_cgroup_try_charge_delay(page, dst_mm, gfp, &memcg, false);
2351         if (ret)
2352                 goto out_release;
2353
2354         ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL,
2355                                                 gfp & GFP_RECLAIM_MASK);
2356         if (ret)
2357                 goto out_release_uncharge;
2358
2359         mem_cgroup_commit_charge(page, memcg, false, false);
2360
2361         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2362         if (dst_vma->vm_flags & VM_WRITE)
2363                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2364         else {
2365                 /*
2366                  * We don't set the pte dirty if the vma has no
2367                  * VM_WRITE permission, so mark the page dirty or it
2368                  * could be freed from under us. We could do it
2369                  * unconditionally before unlock_page(), but doing it
2370                  * only if VM_WRITE is not set is faster.
2371                  */
2372                 set_page_dirty(page);
2373         }
2374
2375         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2376
2377         ret = -EFAULT;
2378         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2379         if (unlikely(offset >= max_off))
2380                 goto out_release_uncharge_unlock;
2381
2382         ret = -EEXIST;
2383         if (!pte_none(*dst_pte))
2384                 goto out_release_uncharge_unlock;
2385
2386         lru_cache_add_anon(page);
2387
2388         spin_lock(&info->lock);
2389         info->alloced++;
2390         inode->i_blocks += BLOCKS_PER_PAGE;
2391         shmem_recalc_inode(inode);
2392         spin_unlock(&info->lock);
2393
2394         inc_mm_counter(dst_mm, mm_counter_file(page));
2395         page_add_file_rmap(page, false);
2396         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2397
2398         /* No need to invalidate - it was non-present before */
2399         update_mmu_cache(dst_vma, dst_addr, dst_pte);
2400         pte_unmap_unlock(dst_pte, ptl);
2401         unlock_page(page);
2402         ret = 0;
2403 out:
2404         return ret;
2405 out_release_uncharge_unlock:
2406         pte_unmap_unlock(dst_pte, ptl);
2407         ClearPageDirty(page);
2408         delete_from_page_cache(page);
2409 out_release_uncharge:
2410         mem_cgroup_cancel_charge(page, memcg, false);
2411 out_release:
2412         unlock_page(page);
2413         put_page(page);
2414 out_unacct_blocks:
2415         shmem_inode_unacct_blocks(inode, 1);
2416         goto out;
2417 }
2418
2419 int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2420                            pmd_t *dst_pmd,
2421                            struct vm_area_struct *dst_vma,
2422                            unsigned long dst_addr,
2423                            unsigned long src_addr,
2424                            struct page **pagep)
2425 {
2426         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2427                                       dst_addr, src_addr, false, pagep);
2428 }
2429
2430 int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2431                              pmd_t *dst_pmd,
2432                              struct vm_area_struct *dst_vma,
2433                              unsigned long dst_addr)
2434 {
2435         struct page *page = NULL;
2436
2437         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2438                                       dst_addr, 0, true, &page);
2439 }
2440
2441 #ifdef CONFIG_TMPFS
2442 static const struct inode_operations shmem_symlink_inode_operations;
2443 static const struct inode_operations shmem_short_symlink_operations;
2444
2445 #ifdef CONFIG_TMPFS_XATTR
2446 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2447 #else
2448 #define shmem_initxattrs NULL
2449 #endif
2450
2451 static int
2452 shmem_write_begin(struct file *file, struct address_space *mapping,
2453                         loff_t pos, unsigned len, unsigned flags,
2454                         struct page **pagep, void **fsdata)
2455 {
2456         struct inode *inode = mapping->host;
2457         struct shmem_inode_info *info = SHMEM_I(inode);
2458         pgoff_t index = pos >> PAGE_SHIFT;
2459
2460         /* i_mutex is held by caller */
2461         if (unlikely(info->seals & (F_SEAL_GROW |
2462                                    F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2463                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2464                         return -EPERM;
2465                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2466                         return -EPERM;
2467         }
2468
2469         return shmem_getpage(inode, index, pagep, SGP_WRITE);
2470 }
2471
2472 static int
2473 shmem_write_end(struct file *file, struct address_space *mapping,
2474                         loff_t pos, unsigned len, unsigned copied,
2475                         struct page *page, void *fsdata)
2476 {
2477         struct inode *inode = mapping->host;
2478
2479         if (pos + copied > inode->i_size)
2480                 i_size_write(inode, pos + copied);
2481
2482         if (!PageUptodate(page)) {
2483                 struct page *head = compound_head(page);
2484                 if (PageTransCompound(page)) {
2485                         int i;
2486
2487                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2488                                 if (head + i == page)
2489                                         continue;
2490                                 clear_highpage(head + i);
2491                                 flush_dcache_page(head + i);
2492                         }
2493                 }
2494                 if (copied < PAGE_SIZE) {
2495                         unsigned from = pos & (PAGE_SIZE - 1);
2496                         zero_user_segments(page, 0, from,
2497                                         from + copied, PAGE_SIZE);
2498                 }
2499                 SetPageUptodate(head);
2500         }
2501         set_page_dirty(page);
2502         unlock_page(page);
2503         put_page(page);
2504
2505         return copied;
2506 }
2507
2508 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2509 {
2510         struct file *file = iocb->ki_filp;
2511         struct inode *inode = file_inode(file);
2512         struct address_space *mapping = inode->i_mapping;
2513         pgoff_t index;
2514         unsigned long offset;
2515         enum sgp_type sgp = SGP_READ;
2516         int error = 0;
2517         ssize_t retval = 0;
2518         loff_t *ppos = &iocb->ki_pos;
2519
2520         /*
2521          * Might this read be for a stacking filesystem?  Then when reading
2522          * holes of a sparse file, we actually need to allocate those pages,
2523          * and even mark them dirty, so it cannot exceed the max_blocks limit.
2524          */
2525         if (!iter_is_iovec(to))
2526                 sgp = SGP_CACHE;
2527
2528         index = *ppos >> PAGE_SHIFT;
2529         offset = *ppos & ~PAGE_MASK;
2530
2531         for (;;) {
2532                 struct page *page = NULL;
2533                 pgoff_t end_index;
2534                 unsigned long nr, ret;
2535                 loff_t i_size = i_size_read(inode);
2536
2537                 end_index = i_size >> PAGE_SHIFT;
2538                 if (index > end_index)
2539                         break;
2540                 if (index == end_index) {
2541                         nr = i_size & ~PAGE_MASK;
2542                         if (nr <= offset)
2543                                 break;
2544                 }
2545
2546                 error = shmem_getpage(inode, index, &page, sgp);
2547                 if (error) {
2548                         if (error == -EINVAL)
2549                                 error = 0;
2550                         break;
2551                 }
2552                 if (page) {
2553                         if (sgp == SGP_CACHE)
2554                                 set_page_dirty(page);
2555                         unlock_page(page);
2556                 }
2557
2558                 /*
2559                  * We must evaluate after, since reads (unlike writes)
2560                  * are called without i_mutex protection against truncate
2561                  */
2562                 nr = PAGE_SIZE;
2563                 i_size = i_size_read(inode);
2564                 end_index = i_size >> PAGE_SHIFT;
2565                 if (index == end_index) {
2566                         nr = i_size & ~PAGE_MASK;
2567                         if (nr <= offset) {
2568                                 if (page)
2569                                         put_page(page);
2570                                 break;
2571                         }
2572                 }
2573                 nr -= offset;
2574
2575                 if (page) {
2576                         /*
2577                          * If users can be writing to this page using arbitrary
2578                          * virtual addresses, take care about potential aliasing
2579                          * before reading the page on the kernel side.
2580                          */
2581                         if (mapping_writably_mapped(mapping))
2582                                 flush_dcache_page(page);
2583                         /*
2584                          * Mark the page accessed if we read the beginning.
2585                          */
2586                         if (!offset)
2587                                 mark_page_accessed(page);
2588                 } else {
2589                         page = ZERO_PAGE(0);
2590                         get_page(page);
2591                 }
2592
2593                 /*
2594                  * Ok, we have the page, and it's up-to-date, so
2595                  * now we can copy it to user space...
2596                  */
2597                 ret = copy_page_to_iter(page, offset, nr, to);
2598                 retval += ret;
2599                 offset += ret;
2600                 index += offset >> PAGE_SHIFT;
2601                 offset &= ~PAGE_MASK;
2602
2603                 put_page(page);
2604                 if (!iov_iter_count(to))
2605                         break;
2606                 if (ret < nr) {
2607                         error = -EFAULT;
2608                         break;
2609                 }
2610                 cond_resched();
2611         }
2612
2613         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2614         file_accessed(file);
2615         return retval ? retval : error;
2616 }
2617
2618 /*
2619  * llseek SEEK_DATA or SEEK_HOLE through the page cache.
2620  */
2621 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2622                                     pgoff_t index, pgoff_t end, int whence)
2623 {
2624         struct page *page;
2625         struct pagevec pvec;
2626         pgoff_t indices[PAGEVEC_SIZE];
2627         bool done = false;
2628         int i;
2629
2630         pagevec_init(&pvec);
2631         pvec.nr = 1;            /* start small: we may be there already */
2632         while (!done) {
2633                 pvec.nr = find_get_entries(mapping, index,
2634                                         pvec.nr, pvec.pages, indices);
2635                 if (!pvec.nr) {
2636                         if (whence == SEEK_DATA)
2637                                 index = end;
2638                         break;
2639                 }
2640                 for (i = 0; i < pvec.nr; i++, index++) {
2641                         if (index < indices[i]) {
2642                                 if (whence == SEEK_HOLE) {
2643                                         done = true;
2644                                         break;
2645                                 }
2646                                 index = indices[i];
2647                         }
2648                         page = pvec.pages[i];
2649                         if (page && !xa_is_value(page)) {
2650                                 if (!PageUptodate(page))
2651                                         page = NULL;
2652                         }
2653                         if (index >= end ||
2654                             (page && whence == SEEK_DATA) ||
2655                             (!page && whence == SEEK_HOLE)) {
2656                                 done = true;
2657                                 break;
2658                         }
2659                 }
2660                 pagevec_remove_exceptionals(&pvec);
2661                 pagevec_release(&pvec);
2662                 pvec.nr = PAGEVEC_SIZE;
2663                 cond_resched();
2664         }
2665         return index;
2666 }
2667
2668 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2669 {
2670         struct address_space *mapping = file->f_mapping;
2671         struct inode *inode = mapping->host;
2672         pgoff_t start, end;
2673         loff_t new_offset;
2674
2675         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2676                 return generic_file_llseek_size(file, offset, whence,
2677                                         MAX_LFS_FILESIZE, i_size_read(inode));
2678         inode_lock(inode);
2679         /* We're holding i_mutex so we can access i_size directly */
2680
2681         if (offset < 0 || offset >= inode->i_size)
2682                 offset = -ENXIO;
2683         else {
2684                 start = offset >> PAGE_SHIFT;
2685                 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2686                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2687                 new_offset <<= PAGE_SHIFT;
2688                 if (new_offset > offset) {
2689                         if (new_offset < inode->i_size)
2690                                 offset = new_offset;
2691                         else if (whence == SEEK_DATA)
2692                                 offset = -ENXIO;
2693                         else
2694                                 offset = inode->i_size;
2695                 }
2696         }
2697
2698         if (offset >= 0)
2699                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2700         inode_unlock(inode);
2701         return offset;
2702 }
2703
2704 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2705                                                          loff_t len)
2706 {
2707         struct inode *inode = file_inode(file);
2708         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2709         struct shmem_inode_info *info = SHMEM_I(inode);
2710         struct shmem_falloc shmem_falloc;
2711         pgoff_t start, index, end;
2712         int error;
2713
2714         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2715                 return -EOPNOTSUPP;
2716
2717         inode_lock(inode);
2718
2719         if (mode & FALLOC_FL_PUNCH_HOLE) {
2720                 struct address_space *mapping = file->f_mapping;
2721                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2722                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2723                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2724
2725                 /* protected by i_mutex */
2726                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2727                         error = -EPERM;
2728                         goto out;
2729                 }
2730
2731                 shmem_falloc.waitq = &shmem_falloc_waitq;
2732                 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2733                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2734                 spin_lock(&inode->i_lock);
2735                 inode->i_private = &shmem_falloc;
2736                 spin_unlock(&inode->i_lock);
2737
2738                 if ((u64)unmap_end > (u64)unmap_start)
2739                         unmap_mapping_range(mapping, unmap_start,
2740                                             1 + unmap_end - unmap_start, 0);
2741                 shmem_truncate_range(inode, offset, offset + len - 1);
2742                 /* No need to unmap again: hole-punching leaves COWed pages */
2743
2744                 spin_lock(&inode->i_lock);
2745                 inode->i_private = NULL;
2746                 wake_up_all(&shmem_falloc_waitq);
2747                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2748                 spin_unlock(&inode->i_lock);
2749                 error = 0;
2750                 goto out;
2751         }
2752
2753         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2754         error = inode_newsize_ok(inode, offset + len);
2755         if (error)
2756                 goto out;
2757
2758         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2759                 error = -EPERM;
2760                 goto out;
2761         }
2762
2763         start = offset >> PAGE_SHIFT;
2764         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2765         /* Try to avoid a swapstorm if len is impossible to satisfy */
2766         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2767                 error = -ENOSPC;
2768                 goto out;
2769         }
2770
2771         shmem_falloc.waitq = NULL;
2772         shmem_falloc.start = start;
2773         shmem_falloc.next  = start;
2774         shmem_falloc.nr_falloced = 0;
2775         shmem_falloc.nr_unswapped = 0;
2776         spin_lock(&inode->i_lock);
2777         inode->i_private = &shmem_falloc;
2778         spin_unlock(&inode->i_lock);
2779
2780         for (index = start; index < end; index++) {
2781                 struct page *page;
2782
2783                 /*
2784                  * Good, the fallocate(2) manpage permits EINTR: we may have
2785                  * been interrupted because we are using up too much memory.
2786                  */
2787                 if (signal_pending(current))
2788                         error = -EINTR;
2789                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2790                         error = -ENOMEM;
2791                 else
2792                         error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2793                 if (error) {
2794                         /* Remove the !PageUptodate pages we added */
2795                         if (index > start) {
2796                                 shmem_undo_range(inode,
2797                                     (loff_t)start << PAGE_SHIFT,
2798                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2799                         }
2800                         goto undone;
2801                 }
2802
2803                 /*
2804                  * Inform shmem_writepage() how far we have reached.
2805                  * No need for lock or barrier: we have the page lock.
2806                  */
2807                 shmem_falloc.next++;
2808                 if (!PageUptodate(page))
2809                         shmem_falloc.nr_falloced++;
2810
2811                 /*
2812                  * If !PageUptodate, leave it that way so that freeable pages
2813                  * can be recognized if we need to rollback on error later.
2814                  * But set_page_dirty so that memory pressure will swap rather
2815                  * than free the pages we are allocating (and SGP_CACHE pages
2816                  * might still be clean: we now need to mark those dirty too).
2817                  */
2818                 set_page_dirty(page);
2819                 unlock_page(page);
2820                 put_page(page);
2821                 cond_resched();
2822         }
2823
2824         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2825                 i_size_write(inode, offset + len);
2826         inode->i_ctime = current_time(inode);
2827 undone:
2828         spin_lock(&inode->i_lock);
2829         inode->i_private = NULL;
2830         spin_unlock(&inode->i_lock);
2831 out:
2832         inode_unlock(inode);
2833         return error;
2834 }
2835
2836 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2837 {
2838         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2839
2840         buf->f_type = TMPFS_MAGIC;
2841         buf->f_bsize = PAGE_SIZE;
2842         buf->f_namelen = NAME_MAX;
2843         if (sbinfo->max_blocks) {
2844                 buf->f_blocks = sbinfo->max_blocks;
2845                 buf->f_bavail =
2846                 buf->f_bfree  = sbinfo->max_blocks -
2847                                 percpu_counter_sum(&sbinfo->used_blocks);
2848         }
2849         if (sbinfo->max_inodes) {
2850                 buf->f_files = sbinfo->max_inodes;
2851                 buf->f_ffree = sbinfo->free_inodes;
2852         }
2853         /* else leave those fields 0 like simple_statfs */
2854         return 0;
2855 }
2856
2857 /*
2858  * File creation. Allocate an inode, and we're done..
2859  */
2860 static int
2861 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2862 {
2863         struct inode *inode;
2864         int error = -ENOSPC;
2865
2866         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2867         if (inode) {
2868                 error = simple_acl_create(dir, inode);
2869                 if (error)
2870                         goto out_iput;
2871                 error = security_inode_init_security(inode, dir,
2872                                                      &dentry->d_name,
2873                                                      shmem_initxattrs, NULL);
2874                 if (error && error != -EOPNOTSUPP)
2875                         goto out_iput;
2876
2877                 error = 0;
2878                 dir->i_size += BOGO_DIRENT_SIZE;
2879                 dir->i_ctime = dir->i_mtime = current_time(dir);
2880                 d_instantiate(dentry, inode);
2881                 dget(dentry); /* Extra count - pin the dentry in core */
2882         }
2883         return error;
2884 out_iput:
2885         iput(inode);
2886         return error;
2887 }
2888
2889 static int
2890 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2891 {
2892         struct inode *inode;
2893         int error = -ENOSPC;
2894
2895         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2896         if (inode) {
2897                 error = security_inode_init_security(inode, dir,
2898                                                      NULL,
2899                                                      shmem_initxattrs, NULL);
2900                 if (error && error != -EOPNOTSUPP)
2901                         goto out_iput;
2902                 error = simple_acl_create(dir, inode);
2903                 if (error)
2904                         goto out_iput;
2905                 d_tmpfile(dentry, inode);
2906         }
2907         return error;
2908 out_iput:
2909         iput(inode);
2910         return error;
2911 }
2912
2913 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2914 {
2915         int error;
2916
2917         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2918                 return error;
2919         inc_nlink(dir);
2920         return 0;
2921 }
2922
2923 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2924                 bool excl)
2925 {
2926         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2927 }
2928
2929 /*
2930  * Link a file..
2931  */
2932 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2933 {
2934         struct inode *inode = d_inode(old_dentry);
2935         int ret = 0;
2936
2937         /*
2938          * No ordinary (disk based) filesystem counts links as inodes;
2939          * but each new link needs a new dentry, pinning lowmem, and
2940          * tmpfs dentries cannot be pruned until they are unlinked.
2941          * But if an O_TMPFILE file is linked into the tmpfs, the
2942          * first link must skip that, to get the accounting right.
2943          */
2944         if (inode->i_nlink) {
2945                 ret = shmem_reserve_inode(inode->i_sb);
2946                 if (ret)
2947                         goto out;
2948         }
2949
2950         dir->i_size += BOGO_DIRENT_SIZE;
2951         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2952         inc_nlink(inode);
2953         ihold(inode);   /* New dentry reference */
2954         dget(dentry);           /* Extra pinning count for the created dentry */
2955         d_instantiate(dentry, inode);
2956 out:
2957         return ret;
2958 }
2959
2960 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2961 {
2962         struct inode *inode = d_inode(dentry);
2963
2964         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2965                 shmem_free_inode(inode->i_sb);
2966
2967         dir->i_size -= BOGO_DIRENT_SIZE;
2968         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2969         drop_nlink(inode);
2970         dput(dentry);   /* Undo the count from "create" - this does all the work */
2971         return 0;
2972 }
2973
2974 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2975 {
2976         if (!simple_empty(dentry))
2977                 return -ENOTEMPTY;
2978
2979         drop_nlink(d_inode(dentry));
2980         drop_nlink(dir);
2981         return shmem_unlink(dir, dentry);
2982 }
2983
2984 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2985 {
2986         bool old_is_dir = d_is_dir(old_dentry);
2987         bool new_is_dir = d_is_dir(new_dentry);
2988
2989         if (old_dir != new_dir && old_is_dir != new_is_dir) {
2990                 if (old_is_dir) {
2991                         drop_nlink(old_dir);
2992                         inc_nlink(new_dir);
2993                 } else {
2994                         drop_nlink(new_dir);
2995                         inc_nlink(old_dir);
2996                 }
2997         }
2998         old_dir->i_ctime = old_dir->i_mtime =
2999         new_dir->i_ctime = new_dir->i_mtime =
3000         d_inode(old_dentry)->i_ctime =
3001         d_inode(new_dentry)->i_ctime = current_time(old_dir);
3002
3003         return 0;
3004 }
3005
3006 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3007 {
3008         struct dentry *whiteout;
3009         int error;
3010
3011         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3012         if (!whiteout)
3013                 return -ENOMEM;
3014
3015         error = shmem_mknod(old_dir, whiteout,
3016                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3017         dput(whiteout);
3018         if (error)
3019                 return error;
3020
3021         /*
3022          * Cheat and hash the whiteout while the old dentry is still in
3023          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3024          *
3025          * d_lookup() will consistently find one of them at this point,
3026          * not sure which one, but that isn't even important.
3027          */
3028         d_rehash(whiteout);
3029         return 0;
3030 }
3031
3032 /*
3033  * The VFS layer already does all the dentry stuff for rename,
3034  * we just have to decrement the usage count for the target if
3035  * it exists so that the VFS layer correctly free's it when it
3036  * gets overwritten.
3037  */
3038 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3039 {
3040         struct inode *inode = d_inode(old_dentry);
3041         int they_are_dirs = S_ISDIR(inode->i_mode);
3042
3043         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3044                 return -EINVAL;
3045
3046         if (flags & RENAME_EXCHANGE)
3047                 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3048
3049         if (!simple_empty(new_dentry))
3050                 return -ENOTEMPTY;
3051
3052         if (flags & RENAME_WHITEOUT) {
3053                 int error;
3054
3055                 error = shmem_whiteout(old_dir, old_dentry);
3056                 if (error)
3057                         return error;
3058         }
3059
3060         if (d_really_is_positive(new_dentry)) {
3061                 (void) shmem_unlink(new_dir, new_dentry);
3062                 if (they_are_dirs) {
3063                         drop_nlink(d_inode(new_dentry));
3064                         drop_nlink(old_dir);
3065                 }
3066         } else if (they_are_dirs) {
3067                 drop_nlink(old_dir);
3068                 inc_nlink(new_dir);
3069         }
3070
3071         old_dir->i_size -= BOGO_DIRENT_SIZE;
3072         new_dir->i_size += BOGO_DIRENT_SIZE;
3073         old_dir->i_ctime = old_dir->i_mtime =
3074         new_dir->i_ctime = new_dir->i_mtime =
3075         inode->i_ctime = current_time(old_dir);
3076         return 0;
3077 }
3078
3079 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3080 {
3081         int error;
3082         int len;
3083         struct inode *inode;
3084         struct page *page;
3085
3086         len = strlen(symname) + 1;
3087         if (len > PAGE_SIZE)
3088                 return -ENAMETOOLONG;
3089
3090         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3091                                 VM_NORESERVE);
3092         if (!inode)
3093                 return -ENOSPC;
3094
3095         error = security_inode_init_security(inode, dir, &dentry->d_name,
3096                                              shmem_initxattrs, NULL);
3097         if (error) {
3098                 if (error != -EOPNOTSUPP) {
3099                         iput(inode);
3100                         return error;
3101                 }
3102                 error = 0;
3103         }
3104
3105         inode->i_size = len-1;
3106         if (len <= SHORT_SYMLINK_LEN) {
3107                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3108                 if (!inode->i_link) {
3109                         iput(inode);
3110                         return -ENOMEM;
3111                 }
3112                 inode->i_op = &shmem_short_symlink_operations;
3113         } else {
3114                 inode_nohighmem(inode);
3115                 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3116                 if (error) {
3117                         iput(inode);
3118                         return error;
3119                 }
3120                 inode->i_mapping->a_ops = &shmem_aops;
3121                 inode->i_op = &shmem_symlink_inode_operations;
3122                 memcpy(page_address(page), symname, len);
3123                 SetPageUptodate(page);
3124                 set_page_dirty(page);
3125                 unlock_page(page);
3126                 put_page(page);
3127         }
3128         dir->i_size += BOGO_DIRENT_SIZE;
3129         dir->i_ctime = dir->i_mtime = current_time(dir);
3130         d_instantiate(dentry, inode);
3131         dget(dentry);
3132         return 0;
3133 }
3134
3135 static void shmem_put_link(void *arg)
3136 {
3137         mark_page_accessed(arg);
3138         put_page(arg);
3139 }
3140
3141 static const char *shmem_get_link(struct dentry *dentry,
3142                                   struct inode *inode,
3143                                   struct delayed_call *done)
3144 {
3145         struct page *page = NULL;
3146         int error;
3147         if (!dentry) {
3148                 page = find_get_page(inode->i_mapping, 0);
3149                 if (!page)
3150                         return ERR_PTR(-ECHILD);
3151                 if (!PageUptodate(page)) {
3152                         put_page(page);
3153                         return ERR_PTR(-ECHILD);
3154                 }
3155         } else {
3156                 error = shmem_getpage(inode, 0, &page, SGP_READ);
3157                 if (error)
3158                         return ERR_PTR(error);
3159                 unlock_page(page);
3160         }
3161         set_delayed_call(done, shmem_put_link, page);
3162         return page_address(page);
3163 }
3164
3165 #ifdef CONFIG_TMPFS_XATTR
3166 /*
3167  * Superblocks without xattr inode operations may get some security.* xattr
3168  * support from the LSM "for free". As soon as we have any other xattrs
3169  * like ACLs, we also need to implement the security.* handlers at
3170  * filesystem level, though.
3171  */
3172
3173 /*
3174  * Callback for security_inode_init_security() for acquiring xattrs.
3175  */
3176 static int shmem_initxattrs(struct inode *inode,
3177                             const struct xattr *xattr_array,
3178                             void *fs_info)
3179 {
3180         struct shmem_inode_info *info = SHMEM_I(inode);
3181         const struct xattr *xattr;
3182         struct simple_xattr *new_xattr;
3183         size_t len;
3184
3185         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3186                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3187                 if (!new_xattr)
3188                         return -ENOMEM;
3189
3190                 len = strlen(xattr->name) + 1;
3191                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3192                                           GFP_KERNEL);
3193                 if (!new_xattr->name) {
3194                         kfree(new_xattr);
3195                         return -ENOMEM;
3196                 }
3197
3198                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3199                        XATTR_SECURITY_PREFIX_LEN);
3200                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3201                        xattr->name, len);
3202
3203                 simple_xattr_list_add(&info->xattrs, new_xattr);
3204         }
3205
3206         return 0;
3207 }
3208
3209 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3210                                    struct dentry *unused, struct inode *inode,
3211                                    const char *name, void *buffer, size_t size)
3212 {
3213         struct shmem_inode_info *info = SHMEM_I(inode);
3214
3215         name = xattr_full_name(handler, name);
3216         return simple_xattr_get(&info->xattrs, name, buffer, size);
3217 }
3218
3219 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3220                                    struct dentry *unused, struct inode *inode,
3221                                    const char *name, const void *value,
3222                                    size_t size, int flags)
3223 {
3224         struct shmem_inode_info *info = SHMEM_I(inode);
3225
3226         name = xattr_full_name(handler, name);
3227         return simple_xattr_set(&info->xattrs, name, value, size, flags);
3228 }
3229
3230 static const struct xattr_handler shmem_security_xattr_handler = {
3231         .prefix = XATTR_SECURITY_PREFIX,
3232         .get = shmem_xattr_handler_get,
3233         .set = shmem_xattr_handler_set,
3234 };
3235
3236 static const struct xattr_handler shmem_trusted_xattr_handler = {
3237         .prefix = XATTR_TRUSTED_PREFIX,
3238         .get = shmem_xattr_handler_get,
3239         .set = shmem_xattr_handler_set,
3240 };
3241
3242 static const struct xattr_handler *shmem_xattr_handlers[] = {
3243 #ifdef CONFIG_TMPFS_POSIX_ACL
3244         &posix_acl_access_xattr_handler,
3245         &posix_acl_default_xattr_handler,
3246 #endif
3247         &shmem_security_xattr_handler,
3248         &shmem_trusted_xattr_handler,
3249         NULL
3250 };
3251
3252 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3253 {
3254         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3255         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3256 }
3257 #endif /* CONFIG_TMPFS_XATTR */
3258
3259 static const struct inode_operations shmem_short_symlink_operations = {
3260         .get_link       = simple_get_link,
3261 #ifdef CONFIG_TMPFS_XATTR
3262         .listxattr      = shmem_listxattr,
3263 #endif
3264 };
3265
3266 static const struct inode_operations shmem_symlink_inode_operations = {
3267         .get_link       = shmem_get_link,
3268 #ifdef CONFIG_TMPFS_XATTR
3269         .listxattr      = shmem_listxattr,
3270 #endif
3271 };
3272
3273 static struct dentry *shmem_get_parent(struct dentry *child)
3274 {
3275         return ERR_PTR(-ESTALE);
3276 }
3277
3278 static int shmem_match(struct inode *ino, void *vfh)
3279 {
3280         __u32 *fh = vfh;
3281         __u64 inum = fh[2];
3282         inum = (inum << 32) | fh[1];
3283         return ino->i_ino == inum && fh[0] == ino->i_generation;
3284 }
3285
3286 /* Find any alias of inode, but prefer a hashed alias */
3287 static struct dentry *shmem_find_alias(struct inode *inode)
3288 {
3289         struct dentry *alias = d_find_alias(inode);
3290
3291         return alias ?: d_find_any_alias(inode);
3292 }
3293
3294
3295 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3296                 struct fid *fid, int fh_len, int fh_type)
3297 {
3298         struct inode *inode;
3299         struct dentry *dentry = NULL;
3300         u64 inum;
3301
3302         if (fh_len < 3)
3303                 return NULL;
3304
3305         inum = fid->raw[2];
3306         inum = (inum << 32) | fid->raw[1];
3307
3308         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3309                         shmem_match, fid->raw);
3310         if (inode) {
3311                 dentry = shmem_find_alias(inode);
3312                 iput(inode);
3313         }
3314
3315         return dentry;
3316 }
3317
3318 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3319                                 struct inode *parent)
3320 {
3321         if (*len < 3) {
3322                 *len = 3;
3323                 return FILEID_INVALID;
3324         }
3325
3326         if (inode_unhashed(inode)) {
3327                 /* Unfortunately insert_inode_hash is not idempotent,
3328                  * so as we hash inodes here rather than at creation
3329                  * time, we need a lock to ensure we only try
3330                  * to do it once
3331                  */
3332                 static DEFINE_SPINLOCK(lock);
3333                 spin_lock(&lock);
3334                 if (inode_unhashed(inode))
3335                         __insert_inode_hash(inode,
3336                                             inode->i_ino + inode->i_generation);
3337                 spin_unlock(&lock);
3338         }
3339
3340         fh[0] = inode->i_generation;
3341         fh[1] = inode->i_ino;
3342         fh[2] = ((__u64)inode->i_ino) >> 32;
3343
3344         *len = 3;
3345         return 1;
3346 }
3347
3348 static const struct export_operations shmem_export_ops = {
3349         .get_parent     = shmem_get_parent,
3350         .encode_fh      = shmem_encode_fh,
3351         .fh_to_dentry   = shmem_fh_to_dentry,
3352 };
3353
3354 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3355                                bool remount)
3356 {
3357         char *this_char, *value, *rest;
3358         struct mempolicy *mpol = NULL;
3359         uid_t uid;
3360         gid_t gid;
3361
3362         while (options != NULL) {
3363                 this_char = options;
3364                 for (;;) {
3365                         /*
3366                          * NUL-terminate this option: unfortunately,
3367                          * mount options form a comma-separated list,
3368                          * but mpol's nodelist may also contain commas.
3369                          */
3370                         options = strchr(options, ',');
3371                         if (options == NULL)
3372                                 break;
3373                         options++;
3374                         if (!isdigit(*options)) {
3375                                 options[-1] = '\0';
3376                                 break;
3377                         }
3378                 }
3379                 if (!*this_char)
3380                         continue;
3381                 if ((value = strchr(this_char,'=')) != NULL) {
3382                         *value++ = 0;
3383                 } else {
3384                         pr_err("tmpfs: No value for mount option '%s'\n",
3385                                this_char);
3386                         goto error;
3387                 }
3388
3389                 if (!strcmp(this_char,"size")) {
3390                         unsigned long long size;
3391                         size = memparse(value,&rest);
3392                         if (*rest == '%') {
3393                                 size <<= PAGE_SHIFT;
3394                                 size *= totalram_pages();
3395                                 do_div(size, 100);
3396                                 rest++;
3397                         }
3398                         if (*rest)
3399                                 goto bad_val;
3400                         sbinfo->max_blocks =
3401                                 DIV_ROUND_UP(size, PAGE_SIZE);
3402                 } else if (!strcmp(this_char,"nr_blocks")) {
3403                         sbinfo->max_blocks = memparse(value, &rest);
3404                         if (*rest)
3405                                 goto bad_val;
3406                 } else if (!strcmp(this_char,"nr_inodes")) {
3407                         sbinfo->max_inodes = memparse(value, &rest);
3408                         if (*rest)
3409                                 goto bad_val;
3410                 } else if (!strcmp(this_char,"mode")) {
3411                         if (remount)
3412                                 continue;
3413                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3414                         if (*rest)
3415                                 goto bad_val;
3416                 } else if (!strcmp(this_char,"uid")) {
3417                         if (remount)
3418                                 continue;
3419                         uid = simple_strtoul(value, &rest, 0);
3420                         if (*rest)
3421                                 goto bad_val;
3422                         sbinfo->uid = make_kuid(current_user_ns(), uid);
3423                         if (!uid_valid(sbinfo->uid))
3424                                 goto bad_val;
3425                 } else if (!strcmp(this_char,"gid")) {
3426                         if (remount)
3427                                 continue;
3428                         gid = simple_strtoul(value, &rest, 0);
3429                         if (*rest)
3430                                 goto bad_val;
3431                         sbinfo->gid = make_kgid(current_user_ns(), gid);
3432                         if (!gid_valid(sbinfo->gid))
3433                                 goto bad_val;
3434 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3435                 } else if (!strcmp(this_char, "huge")) {
3436                         int huge;
3437                         huge = shmem_parse_huge(value);
3438                         if (huge < 0)
3439                                 goto bad_val;
3440                         if (!has_transparent_hugepage() &&
3441                                         huge != SHMEM_HUGE_NEVER)
3442                                 goto bad_val;
3443                         sbinfo->huge = huge;
3444 #endif
3445 #ifdef CONFIG_NUMA
3446                 } else if (!strcmp(this_char,"mpol")) {
3447                         mpol_put(mpol);
3448                         mpol = NULL;
3449                         if (mpol_parse_str(value, &mpol))
3450                                 goto bad_val;
3451 #endif
3452                 } else {
3453                         pr_err("tmpfs: Bad mount option %s\n", this_char);
3454                         goto error;
3455                 }
3456         }
3457         sbinfo->mpol = mpol;
3458         return 0;
3459
3460 bad_val:
3461         pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3462                value, this_char);
3463 error:
3464         mpol_put(mpol);
3465         return 1;
3466
3467 }
3468
3469 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3470 {
3471         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3472         struct shmem_sb_info config = *sbinfo;
3473         unsigned long inodes;
3474         int error = -EINVAL;
3475
3476         config.mpol = NULL;
3477         if (shmem_parse_options(data, &config, true))
3478                 return error;
3479
3480         spin_lock(&sbinfo->stat_lock);
3481         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3482         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3483                 goto out;
3484         if (config.max_inodes < inodes)
3485                 goto out;
3486         /*
3487          * Those tests disallow limited->unlimited while any are in use;
3488          * but we must separately disallow unlimited->limited, because
3489          * in that case we have no record of how much is already in use.
3490          */
3491         if (config.max_blocks && !sbinfo->max_blocks)
3492                 goto out;
3493         if (config.max_inodes && !sbinfo->max_inodes)
3494                 goto out;
3495
3496         error = 0;
3497         sbinfo->huge = config.huge;
3498         sbinfo->max_blocks  = config.max_blocks;
3499         sbinfo->max_inodes  = config.max_inodes;
3500         sbinfo->free_inodes = config.max_inodes - inodes;
3501
3502         /*
3503          * Preserve previous mempolicy unless mpol remount option was specified.
3504          */
3505         if (config.mpol) {
3506                 mpol_put(sbinfo->mpol);
3507                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
3508         }
3509 out:
3510         spin_unlock(&sbinfo->stat_lock);
3511         return error;
3512 }
3513
3514 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3515 {
3516         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3517
3518         if (sbinfo->max_blocks != shmem_default_max_blocks())
3519                 seq_printf(seq, ",size=%luk",
3520                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3521         if (sbinfo->max_inodes != shmem_default_max_inodes())
3522                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3523         if (sbinfo->mode != (0777 | S_ISVTX))
3524                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3525         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3526                 seq_printf(seq, ",uid=%u",
3527                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3528         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3529                 seq_printf(seq, ",gid=%u",
3530                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3531 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3532         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3533         if (sbinfo->huge)
3534                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3535 #endif
3536         shmem_show_mpol(seq, sbinfo->mpol);
3537         return 0;
3538 }
3539
3540 #endif /* CONFIG_TMPFS */
3541
3542 static void shmem_put_super(struct super_block *sb)
3543 {
3544         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3545
3546         percpu_counter_destroy(&sbinfo->used_blocks);
3547         mpol_put(sbinfo->mpol);
3548         kfree(sbinfo);
3549         sb->s_fs_info = NULL;
3550 }
3551
3552 int shmem_fill_super(struct super_block *sb, void *data, int silent)
3553 {
3554         struct inode *inode;
3555         struct shmem_sb_info *sbinfo;
3556         int err = -ENOMEM;
3557
3558         /* Round up to L1_CACHE_BYTES to resist false sharing */
3559         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3560                                 L1_CACHE_BYTES), GFP_KERNEL);
3561         if (!sbinfo)
3562                 return -ENOMEM;
3563
3564         sbinfo->mode = 0777 | S_ISVTX;
3565         sbinfo->uid = current_fsuid();
3566         sbinfo->gid = current_fsgid();
3567         sb->s_fs_info = sbinfo;
3568
3569 #ifdef CONFIG_TMPFS
3570         /*
3571          * Per default we only allow half of the physical ram per
3572          * tmpfs instance, limiting inodes to one per page of lowmem;
3573          * but the internal instance is left unlimited.
3574          */
3575         if (!(sb->s_flags & SB_KERNMOUNT)) {
3576                 sbinfo->max_blocks = shmem_default_max_blocks();
3577                 sbinfo->max_inodes = shmem_default_max_inodes();
3578                 if (shmem_parse_options(data, sbinfo, false)) {
3579                         err = -EINVAL;
3580                         goto failed;
3581                 }
3582         } else {
3583                 sb->s_flags |= SB_NOUSER;
3584         }
3585         sb->s_export_op = &shmem_export_ops;
3586         sb->s_flags |= SB_NOSEC;
3587 #else
3588         sb->s_flags |= SB_NOUSER;
3589 #endif
3590
3591         spin_lock_init(&sbinfo->stat_lock);
3592         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3593                 goto failed;
3594         sbinfo->free_inodes = sbinfo->max_inodes;
3595         spin_lock_init(&sbinfo->shrinklist_lock);
3596         INIT_LIST_HEAD(&sbinfo->shrinklist);
3597
3598         sb->s_maxbytes = MAX_LFS_FILESIZE;
3599         sb->s_blocksize = PAGE_SIZE;
3600         sb->s_blocksize_bits = PAGE_SHIFT;
3601         sb->s_magic = TMPFS_MAGIC;
3602         sb->s_op = &shmem_ops;
3603         sb->s_time_gran = 1;
3604 #ifdef CONFIG_TMPFS_XATTR
3605         sb->s_xattr = shmem_xattr_handlers;
3606 #endif
3607 #ifdef CONFIG_TMPFS_POSIX_ACL
3608         sb->s_flags |= SB_POSIXACL;
3609 #endif
3610         uuid_gen(&sb->s_uuid);
3611
3612         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3613         if (!inode)
3614                 goto failed;
3615         inode->i_uid = sbinfo->uid;
3616         inode->i_gid = sbinfo->gid;
3617         sb->s_root = d_make_root(inode);
3618         if (!sb->s_root)
3619                 goto failed;
3620         return 0;
3621
3622 failed:
3623         shmem_put_super(sb);
3624         return err;
3625 }
3626
3627 static struct kmem_cache *shmem_inode_cachep;
3628
3629 static struct inode *shmem_alloc_inode(struct super_block *sb)
3630 {
3631         struct shmem_inode_info *info;
3632         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3633         if (!info)
3634                 return NULL;
3635         return &info->vfs_inode;
3636 }
3637
3638 static void shmem_destroy_callback(struct rcu_head *head)
3639 {
3640         struct inode *inode = container_of(head, struct inode, i_rcu);
3641         if (S_ISLNK(inode->i_mode))
3642                 kfree(inode->i_link);
3643         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3644 }
3645
3646 static void shmem_destroy_inode(struct inode *inode)
3647 {
3648         if (S_ISREG(inode->i_mode))
3649                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3650         call_rcu(&inode->i_rcu, shmem_destroy_callback);
3651 }
3652
3653 static void shmem_init_inode(void *foo)
3654 {
3655         struct shmem_inode_info *info = foo;
3656         inode_init_once(&info->vfs_inode);
3657 }
3658
3659 static void shmem_init_inodecache(void)
3660 {
3661         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3662                                 sizeof(struct shmem_inode_info),
3663                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3664 }
3665
3666 static void shmem_destroy_inodecache(void)
3667 {
3668         kmem_cache_destroy(shmem_inode_cachep);
3669 }
3670
3671 static const struct address_space_operations shmem_aops = {
3672         .writepage      = shmem_writepage,
3673         .set_page_dirty = __set_page_dirty_no_writeback,
3674 #ifdef CONFIG_TMPFS
3675         .write_begin    = shmem_write_begin,
3676         .write_end      = shmem_write_end,
3677 #endif
3678 #ifdef CONFIG_MIGRATION
3679         .migratepage    = migrate_page,
3680 #endif
3681         .error_remove_page = generic_error_remove_page,
3682 };
3683
3684 static const struct file_operations shmem_file_operations = {
3685         .mmap           = shmem_mmap,
3686         .get_unmapped_area = shmem_get_unmapped_area,
3687 #ifdef CONFIG_TMPFS
3688         .llseek         = shmem_file_llseek,
3689         .read_iter      = shmem_file_read_iter,
3690         .write_iter     = generic_file_write_iter,
3691         .fsync          = noop_fsync,
3692         .splice_read    = generic_file_splice_read,
3693         .splice_write   = iter_file_splice_write,
3694         .fallocate      = shmem_fallocate,
3695 #endif
3696 };
3697
3698 static const struct inode_operations shmem_inode_operations = {
3699         .getattr        = shmem_getattr,
3700         .setattr        = shmem_setattr,
3701 #ifdef CONFIG_TMPFS_XATTR
3702         .listxattr      = shmem_listxattr,
3703         .set_acl        = simple_set_acl,
3704 #endif
3705 };
3706
3707 static const struct inode_operations shmem_dir_inode_operations = {
3708 #ifdef CONFIG_TMPFS
3709         .create         = shmem_create,
3710         .lookup         = simple_lookup,
3711         .link           = shmem_link,
3712         .unlink         = shmem_unlink,
3713         .symlink        = shmem_symlink,
3714         .mkdir          = shmem_mkdir,
3715         .rmdir          = shmem_rmdir,
3716         .mknod          = shmem_mknod,
3717         .rename         = shmem_rename2,
3718         .tmpfile        = shmem_tmpfile,
3719 #endif
3720 #ifdef CONFIG_TMPFS_XATTR
3721         .listxattr      = shmem_listxattr,
3722 #endif
3723 #ifdef CONFIG_TMPFS_POSIX_ACL
3724         .setattr        = shmem_setattr,
3725         .set_acl        = simple_set_acl,
3726 #endif
3727 };
3728
3729 static const struct inode_operations shmem_special_inode_operations = {
3730 #ifdef CONFIG_TMPFS_XATTR
3731         .listxattr      = shmem_listxattr,
3732 #endif
3733 #ifdef CONFIG_TMPFS_POSIX_ACL
3734         .setattr        = shmem_setattr,
3735         .set_acl        = simple_set_acl,
3736 #endif
3737 };
3738
3739 static const struct super_operations shmem_ops = {
3740         .alloc_inode    = shmem_alloc_inode,
3741         .destroy_inode  = shmem_destroy_inode,
3742 #ifdef CONFIG_TMPFS
3743         .statfs         = shmem_statfs,
3744         .remount_fs     = shmem_remount_fs,
3745         .show_options   = shmem_show_options,
3746 #endif
3747         .evict_inode    = shmem_evict_inode,
3748         .drop_inode     = generic_delete_inode,
3749         .put_super      = shmem_put_super,
3750 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3751         .nr_cached_objects      = shmem_unused_huge_count,
3752         .free_cached_objects    = shmem_unused_huge_scan,
3753 #endif
3754 };
3755
3756 static const struct vm_operations_struct shmem_vm_ops = {
3757         .fault          = shmem_fault,
3758         .map_pages      = filemap_map_pages,
3759 #ifdef CONFIG_NUMA
3760         .set_policy     = shmem_set_policy,
3761         .get_policy     = shmem_get_policy,
3762 #endif
3763 };
3764
3765 static struct dentry *shmem_mount(struct file_system_type *fs_type,
3766         int flags, const char *dev_name, void *data)
3767 {
3768         return mount_nodev(fs_type, flags, data, shmem_fill_super);
3769 }
3770
3771 static struct file_system_type shmem_fs_type = {
3772         .owner          = THIS_MODULE,
3773         .name           = "tmpfs",
3774         .mount          = shmem_mount,
3775         .kill_sb        = kill_litter_super,
3776         .fs_flags       = FS_USERNS_MOUNT,
3777 };
3778
3779 int __init shmem_init(void)
3780 {
3781         int error;
3782
3783         /* If rootfs called this, don't re-init */
3784         if (shmem_inode_cachep)
3785                 return 0;
3786
3787         shmem_init_inodecache();
3788
3789         error = register_filesystem(&shmem_fs_type);
3790         if (error) {
3791                 pr_err("Could not register tmpfs\n");
3792                 goto out2;
3793         }
3794
3795         shm_mnt = kern_mount(&shmem_fs_type);
3796         if (IS_ERR(shm_mnt)) {
3797                 error = PTR_ERR(shm_mnt);
3798                 pr_err("Could not kern_mount tmpfs\n");
3799                 goto out1;
3800         }
3801
3802 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3803         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
3804                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3805         else
3806                 shmem_huge = 0; /* just in case it was patched */
3807 #endif
3808         return 0;
3809
3810 out1:
3811         unregister_filesystem(&shmem_fs_type);
3812 out2:
3813         shmem_destroy_inodecache();
3814         shm_mnt = ERR_PTR(error);
3815         return error;
3816 }
3817
3818 #if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
3819 static ssize_t shmem_enabled_show(struct kobject *kobj,
3820                 struct kobj_attribute *attr, char *buf)
3821 {
3822         int values[] = {
3823                 SHMEM_HUGE_ALWAYS,
3824                 SHMEM_HUGE_WITHIN_SIZE,
3825                 SHMEM_HUGE_ADVISE,
3826                 SHMEM_HUGE_NEVER,
3827                 SHMEM_HUGE_DENY,
3828                 SHMEM_HUGE_FORCE,
3829         };
3830         int i, count;
3831
3832         for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
3833                 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
3834
3835                 count += sprintf(buf + count, fmt,
3836                                 shmem_format_huge(values[i]));
3837         }
3838         buf[count - 1] = '\n';
3839         return count;
3840 }
3841
3842 static ssize_t shmem_enabled_store(struct kobject *kobj,
3843                 struct kobj_attribute *attr, const char *buf, size_t count)
3844 {
3845         char tmp[16];
3846         int huge;
3847
3848         if (count + 1 > sizeof(tmp))
3849                 return -EINVAL;
3850         memcpy(tmp, buf, count);
3851         tmp[count] = '\0';
3852         if (count && tmp[count - 1] == '\n')
3853                 tmp[count - 1] = '\0';
3854
3855         huge = shmem_parse_huge(tmp);
3856         if (huge == -EINVAL)
3857                 return -EINVAL;
3858         if (!has_transparent_hugepage() &&
3859                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
3860                 return -EINVAL;
3861
3862         shmem_huge = huge;
3863         if (shmem_huge > SHMEM_HUGE_DENY)
3864                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3865         return count;
3866 }
3867
3868 struct kobj_attribute shmem_enabled_attr =
3869         __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
3870 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE && CONFIG_SYSFS */
3871
3872 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3873 bool shmem_huge_enabled(struct vm_area_struct *vma)
3874 {
3875         struct inode *inode = file_inode(vma->vm_file);
3876         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3877         loff_t i_size;
3878         pgoff_t off;
3879
3880         if (shmem_huge == SHMEM_HUGE_FORCE)
3881                 return true;
3882         if (shmem_huge == SHMEM_HUGE_DENY)
3883                 return false;
3884         switch (sbinfo->huge) {
3885                 case SHMEM_HUGE_NEVER:
3886                         return false;
3887                 case SHMEM_HUGE_ALWAYS:
3888                         return true;
3889                 case SHMEM_HUGE_WITHIN_SIZE:
3890                         off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
3891                         i_size = round_up(i_size_read(inode), PAGE_SIZE);
3892                         if (i_size >= HPAGE_PMD_SIZE &&
3893                                         i_size >> PAGE_SHIFT >= off)
3894                                 return true;
3895                         /* fall through */
3896                 case SHMEM_HUGE_ADVISE:
3897                         /* TODO: implement fadvise() hints */
3898                         return (vma->vm_flags & VM_HUGEPAGE);
3899                 default:
3900                         VM_BUG_ON(1);
3901                         return false;
3902         }
3903 }
3904 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
3905
3906 #else /* !CONFIG_SHMEM */
3907
3908 /*
3909  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3910  *
3911  * This is intended for small system where the benefits of the full
3912  * shmem code (swap-backed and resource-limited) are outweighed by
3913  * their complexity. On systems without swap this code should be
3914  * effectively equivalent, but much lighter weight.
3915  */
3916
3917 static struct file_system_type shmem_fs_type = {
3918         .name           = "tmpfs",
3919         .mount          = ramfs_mount,
3920         .kill_sb        = kill_litter_super,
3921         .fs_flags       = FS_USERNS_MOUNT,
3922 };
3923
3924 int __init shmem_init(void)
3925 {
3926         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3927
3928         shm_mnt = kern_mount(&shmem_fs_type);
3929         BUG_ON(IS_ERR(shm_mnt));
3930
3931         return 0;
3932 }
3933
3934 int shmem_unuse(unsigned int type, bool frontswap,
3935                 unsigned long *fs_pages_to_unuse)
3936 {
3937         return 0;
3938 }
3939
3940 int shmem_lock(struct file *file, int lock, struct user_struct *user)
3941 {
3942         return 0;
3943 }
3944
3945 void shmem_unlock_mapping(struct address_space *mapping)
3946 {
3947 }
3948
3949 #ifdef CONFIG_MMU
3950 unsigned long shmem_get_unmapped_area(struct file *file,
3951                                       unsigned long addr, unsigned long len,
3952                                       unsigned long pgoff, unsigned long flags)
3953 {
3954         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
3955 }
3956 #endif
3957
3958 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3959 {
3960         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3961 }
3962 EXPORT_SYMBOL_GPL(shmem_truncate_range);
3963
3964 #define shmem_vm_ops                            generic_file_vm_ops
3965 #define shmem_file_operations                   ramfs_file_operations
3966 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
3967 #define shmem_acct_size(flags, size)            0
3968 #define shmem_unacct_size(flags, size)          do {} while (0)
3969
3970 #endif /* CONFIG_SHMEM */
3971
3972 /* common code */
3973
3974 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
3975                                        unsigned long flags, unsigned int i_flags)
3976 {
3977         struct inode *inode;
3978         struct file *res;
3979
3980         if (IS_ERR(mnt))
3981                 return ERR_CAST(mnt);
3982
3983         if (size < 0 || size > MAX_LFS_FILESIZE)
3984                 return ERR_PTR(-EINVAL);
3985
3986         if (shmem_acct_size(flags, size))
3987                 return ERR_PTR(-ENOMEM);
3988
3989         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
3990                                 flags);
3991         if (unlikely(!inode)) {
3992                 shmem_unacct_size(flags, size);
3993                 return ERR_PTR(-ENOSPC);
3994         }
3995         inode->i_flags |= i_flags;
3996         inode->i_size = size;
3997         clear_nlink(inode);     /* It is unlinked */
3998         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
3999         if (!IS_ERR(res))
4000                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4001                                 &shmem_file_operations);
4002         if (IS_ERR(res))
4003                 iput(inode);
4004         return res;
4005 }
4006
4007 /**
4008  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4009  *      kernel internal.  There will be NO LSM permission checks against the
4010  *      underlying inode.  So users of this interface must do LSM checks at a
4011  *      higher layer.  The users are the big_key and shm implementations.  LSM
4012  *      checks are provided at the key or shm level rather than the inode.
4013  * @name: name for dentry (to be seen in /proc/<pid>/maps
4014  * @size: size to be set for the file
4015  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4016  */
4017 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4018 {
4019         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4020 }
4021
4022 /**
4023  * shmem_file_setup - get an unlinked file living in tmpfs
4024  * @name: name for dentry (to be seen in /proc/<pid>/maps
4025  * @size: size to be set for the file
4026  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4027  */
4028 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4029 {
4030         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4031 }
4032 EXPORT_SYMBOL_GPL(shmem_file_setup);
4033
4034 /**
4035  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4036  * @mnt: the tmpfs mount where the file will be created
4037  * @name: name for dentry (to be seen in /proc/<pid>/maps
4038  * @size: size to be set for the file
4039  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4040  */
4041 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4042                                        loff_t size, unsigned long flags)
4043 {
4044         return __shmem_file_setup(mnt, name, size, flags, 0);
4045 }
4046 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4047
4048 /**
4049  * shmem_zero_setup - setup a shared anonymous mapping
4050  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
4051  */
4052 int shmem_zero_setup(struct vm_area_struct *vma)
4053 {
4054         struct file *file;
4055         loff_t size = vma->vm_end - vma->vm_start;
4056
4057         /*
4058          * Cloning a new file under mmap_sem leads to a lock ordering conflict
4059          * between XFS directory reading and selinux: since this file is only
4060          * accessible to the user through its mapping, use S_PRIVATE flag to
4061          * bypass file security, in the same way as shmem_kernel_file_setup().
4062          */
4063         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4064         if (IS_ERR(file))
4065                 return PTR_ERR(file);
4066
4067         if (vma->vm_file)
4068                 fput(vma->vm_file);
4069         vma->vm_file = file;
4070         vma->vm_ops = &shmem_vm_ops;
4071
4072         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
4073                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4074                         (vma->vm_end & HPAGE_PMD_MASK)) {
4075                 khugepaged_enter(vma, vma->vm_flags);
4076         }
4077
4078         return 0;
4079 }
4080
4081 /**
4082  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4083  * @mapping:    the page's address_space
4084  * @index:      the page index
4085  * @gfp:        the page allocator flags to use if allocating
4086  *
4087  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4088  * with any new page allocations done using the specified allocation flags.
4089  * But read_cache_page_gfp() uses the ->readpage() method: which does not
4090  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4091  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4092  *
4093  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4094  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4095  */
4096 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4097                                          pgoff_t index, gfp_t gfp)
4098 {
4099 #ifdef CONFIG_SHMEM
4100         struct inode *inode = mapping->host;
4101         struct page *page;
4102         int error;
4103
4104         BUG_ON(mapping->a_ops != &shmem_aops);
4105         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4106                                   gfp, NULL, NULL, NULL);
4107         if (error)
4108                 page = ERR_PTR(error);
4109         else
4110                 unlock_page(page);
4111         return page;
4112 #else
4113         /*
4114          * The tiny !SHMEM case uses ramfs without swap
4115          */
4116         return read_cache_page_gfp(mapping, index, gfp);
4117 #endif
4118 }
4119 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);