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                 while (!list_empty(&info->swaplist)) {
1085                         /* Wait while shmem_unuse() is scanning this inode... */
1086                         wait_var_event(&info->stop_eviction,
1087                                        !atomic_read(&info->stop_eviction));
1088                         mutex_lock(&shmem_swaplist_mutex);
1089                         /* ...but beware of the race if we peeked too early */
1090                         if (!atomic_read(&info->stop_eviction))
1091                                 list_del_init(&info->swaplist);
1092                         mutex_unlock(&shmem_swaplist_mutex);
1093                 }
1094         }
1095
1096         simple_xattrs_free(&info->xattrs);
1097         WARN_ON(inode->i_blocks);
1098         shmem_free_inode(inode->i_sb);
1099         clear_inode(inode);
1100 }
1101
1102 extern struct swap_info_struct *swap_info[];
1103
1104 static int shmem_find_swap_entries(struct address_space *mapping,
1105                                    pgoff_t start, unsigned int nr_entries,
1106                                    struct page **entries, pgoff_t *indices,
1107                                    unsigned int type, bool frontswap)
1108 {
1109         XA_STATE(xas, &mapping->i_pages, start);
1110         struct page *page;
1111         swp_entry_t entry;
1112         unsigned int ret = 0;
1113
1114         if (!nr_entries)
1115                 return 0;
1116
1117         rcu_read_lock();
1118         xas_for_each(&xas, page, ULONG_MAX) {
1119                 if (xas_retry(&xas, page))
1120                         continue;
1121
1122                 if (!xa_is_value(page))
1123                         continue;
1124
1125                 entry = radix_to_swp_entry(page);
1126                 if (swp_type(entry) != type)
1127                         continue;
1128                 if (frontswap &&
1129                     !frontswap_test(swap_info[type], swp_offset(entry)))
1130                         continue;
1131
1132                 indices[ret] = xas.xa_index;
1133                 entries[ret] = page;
1134
1135                 if (need_resched()) {
1136                         xas_pause(&xas);
1137                         cond_resched_rcu();
1138                 }
1139                 if (++ret == nr_entries)
1140                         break;
1141         }
1142         rcu_read_unlock();
1143
1144         return ret;
1145 }
1146
1147 /*
1148  * Move the swapped pages for an inode to page cache. Returns the count
1149  * of pages swapped in, or the error in case of failure.
1150  */
1151 static int shmem_unuse_swap_entries(struct inode *inode, struct pagevec pvec,
1152                                     pgoff_t *indices)
1153 {
1154         int i = 0;
1155         int ret = 0;
1156         int error = 0;
1157         struct address_space *mapping = inode->i_mapping;
1158
1159         for (i = 0; i < pvec.nr; i++) {
1160                 struct page *page = pvec.pages[i];
1161
1162                 if (!xa_is_value(page))
1163                         continue;
1164                 error = shmem_swapin_page(inode, indices[i],
1165                                           &page, SGP_CACHE,
1166                                           mapping_gfp_mask(mapping),
1167                                           NULL, NULL);
1168                 if (error == 0) {
1169                         unlock_page(page);
1170                         put_page(page);
1171                         ret++;
1172                 }
1173                 if (error == -ENOMEM)
1174                         break;
1175                 error = 0;
1176         }
1177         return error ? error : ret;
1178 }
1179
1180 /*
1181  * If swap found in inode, free it and move page from swapcache to filecache.
1182  */
1183 static int shmem_unuse_inode(struct inode *inode, unsigned int type,
1184                              bool frontswap, unsigned long *fs_pages_to_unuse)
1185 {
1186         struct address_space *mapping = inode->i_mapping;
1187         pgoff_t start = 0;
1188         struct pagevec pvec;
1189         pgoff_t indices[PAGEVEC_SIZE];
1190         bool frontswap_partial = (frontswap && *fs_pages_to_unuse > 0);
1191         int ret = 0;
1192
1193         pagevec_init(&pvec);
1194         do {
1195                 unsigned int nr_entries = PAGEVEC_SIZE;
1196
1197                 if (frontswap_partial && *fs_pages_to_unuse < PAGEVEC_SIZE)
1198                         nr_entries = *fs_pages_to_unuse;
1199
1200                 pvec.nr = shmem_find_swap_entries(mapping, start, nr_entries,
1201                                                   pvec.pages, indices,
1202                                                   type, frontswap);
1203                 if (pvec.nr == 0) {
1204                         ret = 0;
1205                         break;
1206                 }
1207
1208                 ret = shmem_unuse_swap_entries(inode, pvec, indices);
1209                 if (ret < 0)
1210                         break;
1211
1212                 if (frontswap_partial) {
1213                         *fs_pages_to_unuse -= ret;
1214                         if (*fs_pages_to_unuse == 0) {
1215                                 ret = FRONTSWAP_PAGES_UNUSED;
1216                                 break;
1217                         }
1218                 }
1219
1220                 start = indices[pvec.nr - 1];
1221         } while (true);
1222
1223         return ret;
1224 }
1225
1226 /*
1227  * Read all the shared memory data that resides in the swap
1228  * device 'type' back into memory, so the swap device can be
1229  * unused.
1230  */
1231 int shmem_unuse(unsigned int type, bool frontswap,
1232                 unsigned long *fs_pages_to_unuse)
1233 {
1234         struct shmem_inode_info *info, *next;
1235         int error = 0;
1236
1237         if (list_empty(&shmem_swaplist))
1238                 return 0;
1239
1240         mutex_lock(&shmem_swaplist_mutex);
1241         list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1242                 if (!info->swapped) {
1243                         list_del_init(&info->swaplist);
1244                         continue;
1245                 }
1246                 /*
1247                  * Drop the swaplist mutex while searching the inode for swap;
1248                  * but before doing so, make sure shmem_evict_inode() will not
1249                  * remove placeholder inode from swaplist, nor let it be freed
1250                  * (igrab() would protect from unlink, but not from unmount).
1251                  */
1252                 atomic_inc(&info->stop_eviction);
1253                 mutex_unlock(&shmem_swaplist_mutex);
1254
1255                 error = shmem_unuse_inode(&info->vfs_inode, type, frontswap,
1256                                           fs_pages_to_unuse);
1257                 cond_resched();
1258
1259                 mutex_lock(&shmem_swaplist_mutex);
1260                 next = list_next_entry(info, swaplist);
1261                 if (!info->swapped)
1262                         list_del_init(&info->swaplist);
1263                 if (atomic_dec_and_test(&info->stop_eviction))
1264                         wake_up_var(&info->stop_eviction);
1265                 if (error)
1266                         break;
1267         }
1268         mutex_unlock(&shmem_swaplist_mutex);
1269
1270         return error;
1271 }
1272
1273 /*
1274  * Move the page from the page cache to the swap cache.
1275  */
1276 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1277 {
1278         struct shmem_inode_info *info;
1279         struct address_space *mapping;
1280         struct inode *inode;
1281         swp_entry_t swap;
1282         pgoff_t index;
1283
1284         VM_BUG_ON_PAGE(PageCompound(page), page);
1285         BUG_ON(!PageLocked(page));
1286         mapping = page->mapping;
1287         index = page->index;
1288         inode = mapping->host;
1289         info = SHMEM_I(inode);
1290         if (info->flags & VM_LOCKED)
1291                 goto redirty;
1292         if (!total_swap_pages)
1293                 goto redirty;
1294
1295         /*
1296          * Our capabilities prevent regular writeback or sync from ever calling
1297          * shmem_writepage; but a stacking filesystem might use ->writepage of
1298          * its underlying filesystem, in which case tmpfs should write out to
1299          * swap only in response to memory pressure, and not for the writeback
1300          * threads or sync.
1301          */
1302         if (!wbc->for_reclaim) {
1303                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1304                 goto redirty;
1305         }
1306
1307         /*
1308          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1309          * value into swapfile.c, the only way we can correctly account for a
1310          * fallocated page arriving here is now to initialize it and write it.
1311          *
1312          * That's okay for a page already fallocated earlier, but if we have
1313          * not yet completed the fallocation, then (a) we want to keep track
1314          * of this page in case we have to undo it, and (b) it may not be a
1315          * good idea to continue anyway, once we're pushing into swap.  So
1316          * reactivate the page, and let shmem_fallocate() quit when too many.
1317          */
1318         if (!PageUptodate(page)) {
1319                 if (inode->i_private) {
1320                         struct shmem_falloc *shmem_falloc;
1321                         spin_lock(&inode->i_lock);
1322                         shmem_falloc = inode->i_private;
1323                         if (shmem_falloc &&
1324                             !shmem_falloc->waitq &&
1325                             index >= shmem_falloc->start &&
1326                             index < shmem_falloc->next)
1327                                 shmem_falloc->nr_unswapped++;
1328                         else
1329                                 shmem_falloc = NULL;
1330                         spin_unlock(&inode->i_lock);
1331                         if (shmem_falloc)
1332                                 goto redirty;
1333                 }
1334                 clear_highpage(page);
1335                 flush_dcache_page(page);
1336                 SetPageUptodate(page);
1337         }
1338
1339         swap = get_swap_page(page);
1340         if (!swap.val)
1341                 goto redirty;
1342
1343         /*
1344          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1345          * if it's not already there.  Do it now before the page is
1346          * moved to swap cache, when its pagelock no longer protects
1347          * the inode from eviction.  But don't unlock the mutex until
1348          * we've incremented swapped, because shmem_unuse_inode() will
1349          * prune a !swapped inode from the swaplist under this mutex.
1350          */
1351         mutex_lock(&shmem_swaplist_mutex);
1352         if (list_empty(&info->swaplist))
1353                 list_add(&info->swaplist, &shmem_swaplist);
1354
1355         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1356                 spin_lock_irq(&info->lock);
1357                 shmem_recalc_inode(inode);
1358                 info->swapped++;
1359                 spin_unlock_irq(&info->lock);
1360
1361                 swap_shmem_alloc(swap);
1362                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1363
1364                 mutex_unlock(&shmem_swaplist_mutex);
1365                 BUG_ON(page_mapped(page));
1366                 swap_writepage(page, wbc);
1367                 return 0;
1368         }
1369
1370         mutex_unlock(&shmem_swaplist_mutex);
1371         put_swap_page(page, swap);
1372 redirty:
1373         set_page_dirty(page);
1374         if (wbc->for_reclaim)
1375                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1376         unlock_page(page);
1377         return 0;
1378 }
1379
1380 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1381 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1382 {
1383         char buffer[64];
1384
1385         if (!mpol || mpol->mode == MPOL_DEFAULT)
1386                 return;         /* show nothing */
1387
1388         mpol_to_str(buffer, sizeof(buffer), mpol);
1389
1390         seq_printf(seq, ",mpol=%s", buffer);
1391 }
1392
1393 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1394 {
1395         struct mempolicy *mpol = NULL;
1396         if (sbinfo->mpol) {
1397                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1398                 mpol = sbinfo->mpol;
1399                 mpol_get(mpol);
1400                 spin_unlock(&sbinfo->stat_lock);
1401         }
1402         return mpol;
1403 }
1404 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1405 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1406 {
1407 }
1408 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1409 {
1410         return NULL;
1411 }
1412 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1413 #ifndef CONFIG_NUMA
1414 #define vm_policy vm_private_data
1415 #endif
1416
1417 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1418                 struct shmem_inode_info *info, pgoff_t index)
1419 {
1420         /* Create a pseudo vma that just contains the policy */
1421         vma_init(vma, NULL);
1422         /* Bias interleave by inode number to distribute better across nodes */
1423         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1424         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1425 }
1426
1427 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1428 {
1429         /* Drop reference taken by mpol_shared_policy_lookup() */
1430         mpol_cond_put(vma->vm_policy);
1431 }
1432
1433 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1434                         struct shmem_inode_info *info, pgoff_t index)
1435 {
1436         struct vm_area_struct pvma;
1437         struct page *page;
1438         struct vm_fault vmf;
1439
1440         shmem_pseudo_vma_init(&pvma, info, index);
1441         vmf.vma = &pvma;
1442         vmf.address = 0;
1443         page = swap_cluster_readahead(swap, gfp, &vmf);
1444         shmem_pseudo_vma_destroy(&pvma);
1445
1446         return page;
1447 }
1448
1449 static struct page *shmem_alloc_hugepage(gfp_t gfp,
1450                 struct shmem_inode_info *info, pgoff_t index)
1451 {
1452         struct vm_area_struct pvma;
1453         struct address_space *mapping = info->vfs_inode.i_mapping;
1454         pgoff_t hindex;
1455         struct page *page;
1456
1457         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1458                 return NULL;
1459
1460         hindex = round_down(index, HPAGE_PMD_NR);
1461         if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1462                                                                 XA_PRESENT))
1463                 return NULL;
1464
1465         shmem_pseudo_vma_init(&pvma, info, hindex);
1466         page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1467                         HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1468         shmem_pseudo_vma_destroy(&pvma);
1469         if (page)
1470                 prep_transhuge_page(page);
1471         return page;
1472 }
1473
1474 static struct page *shmem_alloc_page(gfp_t gfp,
1475                         struct shmem_inode_info *info, pgoff_t index)
1476 {
1477         struct vm_area_struct pvma;
1478         struct page *page;
1479
1480         shmem_pseudo_vma_init(&pvma, info, index);
1481         page = alloc_page_vma(gfp, &pvma, 0);
1482         shmem_pseudo_vma_destroy(&pvma);
1483
1484         return page;
1485 }
1486
1487 static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1488                 struct inode *inode,
1489                 pgoff_t index, bool huge)
1490 {
1491         struct shmem_inode_info *info = SHMEM_I(inode);
1492         struct page *page;
1493         int nr;
1494         int err = -ENOSPC;
1495
1496         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1497                 huge = false;
1498         nr = huge ? HPAGE_PMD_NR : 1;
1499
1500         if (!shmem_inode_acct_block(inode, nr))
1501                 goto failed;
1502
1503         if (huge)
1504                 page = shmem_alloc_hugepage(gfp, info, index);
1505         else
1506                 page = shmem_alloc_page(gfp, info, index);
1507         if (page) {
1508                 __SetPageLocked(page);
1509                 __SetPageSwapBacked(page);
1510                 return page;
1511         }
1512
1513         err = -ENOMEM;
1514         shmem_inode_unacct_blocks(inode, nr);
1515 failed:
1516         return ERR_PTR(err);
1517 }
1518
1519 /*
1520  * When a page is moved from swapcache to shmem filecache (either by the
1521  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1522  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1523  * ignorance of the mapping it belongs to.  If that mapping has special
1524  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1525  * we may need to copy to a suitable page before moving to filecache.
1526  *
1527  * In a future release, this may well be extended to respect cpuset and
1528  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1529  * but for now it is a simple matter of zone.
1530  */
1531 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1532 {
1533         return page_zonenum(page) > gfp_zone(gfp);
1534 }
1535
1536 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1537                                 struct shmem_inode_info *info, pgoff_t index)
1538 {
1539         struct page *oldpage, *newpage;
1540         struct address_space *swap_mapping;
1541         swp_entry_t entry;
1542         pgoff_t swap_index;
1543         int error;
1544
1545         oldpage = *pagep;
1546         entry.val = page_private(oldpage);
1547         swap_index = swp_offset(entry);
1548         swap_mapping = page_mapping(oldpage);
1549
1550         /*
1551          * We have arrived here because our zones are constrained, so don't
1552          * limit chance of success by further cpuset and node constraints.
1553          */
1554         gfp &= ~GFP_CONSTRAINT_MASK;
1555         newpage = shmem_alloc_page(gfp, info, index);
1556         if (!newpage)
1557                 return -ENOMEM;
1558
1559         get_page(newpage);
1560         copy_highpage(newpage, oldpage);
1561         flush_dcache_page(newpage);
1562
1563         __SetPageLocked(newpage);
1564         __SetPageSwapBacked(newpage);
1565         SetPageUptodate(newpage);
1566         set_page_private(newpage, entry.val);
1567         SetPageSwapCache(newpage);
1568
1569         /*
1570          * Our caller will very soon move newpage out of swapcache, but it's
1571          * a nice clean interface for us to replace oldpage by newpage there.
1572          */
1573         xa_lock_irq(&swap_mapping->i_pages);
1574         error = shmem_replace_entry(swap_mapping, swap_index, oldpage, newpage);
1575         if (!error) {
1576                 __inc_node_page_state(newpage, NR_FILE_PAGES);
1577                 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1578         }
1579         xa_unlock_irq(&swap_mapping->i_pages);
1580
1581         if (unlikely(error)) {
1582                 /*
1583                  * Is this possible?  I think not, now that our callers check
1584                  * both PageSwapCache and page_private after getting page lock;
1585                  * but be defensive.  Reverse old to newpage for clear and free.
1586                  */
1587                 oldpage = newpage;
1588         } else {
1589                 mem_cgroup_migrate(oldpage, newpage);
1590                 lru_cache_add_anon(newpage);
1591                 *pagep = newpage;
1592         }
1593
1594         ClearPageSwapCache(oldpage);
1595         set_page_private(oldpage, 0);
1596
1597         unlock_page(oldpage);
1598         put_page(oldpage);
1599         put_page(oldpage);
1600         return error;
1601 }
1602
1603 /*
1604  * Swap in the page pointed to by *pagep.
1605  * Caller has to make sure that *pagep contains a valid swapped page.
1606  * Returns 0 and the page in pagep if success. On failure, returns the
1607  * the error code and NULL in *pagep.
1608  */
1609 static int shmem_swapin_page(struct inode *inode, pgoff_t index,
1610                              struct page **pagep, enum sgp_type sgp,
1611                              gfp_t gfp, struct vm_area_struct *vma,
1612                              vm_fault_t *fault_type)
1613 {
1614         struct address_space *mapping = inode->i_mapping;
1615         struct shmem_inode_info *info = SHMEM_I(inode);
1616         struct mm_struct *charge_mm = vma ? vma->vm_mm : current->mm;
1617         struct mem_cgroup *memcg;
1618         struct page *page;
1619         swp_entry_t swap;
1620         int error;
1621
1622         VM_BUG_ON(!*pagep || !xa_is_value(*pagep));
1623         swap = radix_to_swp_entry(*pagep);
1624         *pagep = NULL;
1625
1626         /* Look it up and read it in.. */
1627         page = lookup_swap_cache(swap, NULL, 0);
1628         if (!page) {
1629                 /* Or update major stats only when swapin succeeds?? */
1630                 if (fault_type) {
1631                         *fault_type |= VM_FAULT_MAJOR;
1632                         count_vm_event(PGMAJFAULT);
1633                         count_memcg_event_mm(charge_mm, PGMAJFAULT);
1634                 }
1635                 /* Here we actually start the io */
1636                 page = shmem_swapin(swap, gfp, info, index);
1637                 if (!page) {
1638                         error = -ENOMEM;
1639                         goto failed;
1640                 }
1641         }
1642
1643         /* We have to do this with page locked to prevent races */
1644         lock_page(page);
1645         if (!PageSwapCache(page) || page_private(page) != swap.val ||
1646             !shmem_confirm_swap(mapping, index, swap)) {
1647                 error = -EEXIST;
1648                 goto unlock;
1649         }
1650         if (!PageUptodate(page)) {
1651                 error = -EIO;
1652                 goto failed;
1653         }
1654         wait_on_page_writeback(page);
1655
1656         if (shmem_should_replace_page(page, gfp)) {
1657                 error = shmem_replace_page(&page, gfp, info, index);
1658                 if (error)
1659                         goto failed;
1660         }
1661
1662         error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
1663                                             false);
1664         if (!error) {
1665                 error = shmem_add_to_page_cache(page, mapping, index,
1666                                                 swp_to_radix_entry(swap), gfp);
1667                 /*
1668                  * We already confirmed swap under page lock, and make
1669                  * no memory allocation here, so usually no possibility
1670                  * of error; but free_swap_and_cache() only trylocks a
1671                  * page, so it is just possible that the entry has been
1672                  * truncated or holepunched since swap was confirmed.
1673                  * shmem_undo_range() will have done some of the
1674                  * unaccounting, now delete_from_swap_cache() will do
1675                  * the rest.
1676                  */
1677                 if (error) {
1678                         mem_cgroup_cancel_charge(page, memcg, false);
1679                         delete_from_swap_cache(page);
1680                 }
1681         }
1682         if (error)
1683                 goto failed;
1684
1685         mem_cgroup_commit_charge(page, memcg, true, false);
1686
1687         spin_lock_irq(&info->lock);
1688         info->swapped--;
1689         shmem_recalc_inode(inode);
1690         spin_unlock_irq(&info->lock);
1691
1692         if (sgp == SGP_WRITE)
1693                 mark_page_accessed(page);
1694
1695         delete_from_swap_cache(page);
1696         set_page_dirty(page);
1697         swap_free(swap);
1698
1699         *pagep = page;
1700         return 0;
1701 failed:
1702         if (!shmem_confirm_swap(mapping, index, swap))
1703                 error = -EEXIST;
1704 unlock:
1705         if (page) {
1706                 unlock_page(page);
1707                 put_page(page);
1708         }
1709
1710         return error;
1711 }
1712
1713 /*
1714  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1715  *
1716  * If we allocate a new one we do not mark it dirty. That's up to the
1717  * vm. If we swap it in we mark it dirty since we also free the swap
1718  * entry since a page cannot live in both the swap and page cache.
1719  *
1720  * fault_mm and fault_type are only supplied by shmem_fault:
1721  * otherwise they are NULL.
1722  */
1723 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1724         struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1725         struct vm_area_struct *vma, struct vm_fault *vmf,
1726                         vm_fault_t *fault_type)
1727 {
1728         struct address_space *mapping = inode->i_mapping;
1729         struct shmem_inode_info *info = SHMEM_I(inode);
1730         struct shmem_sb_info *sbinfo;
1731         struct mm_struct *charge_mm;
1732         struct mem_cgroup *memcg;
1733         struct page *page;
1734         enum sgp_type sgp_huge = sgp;
1735         pgoff_t hindex = index;
1736         int error;
1737         int once = 0;
1738         int alloced = 0;
1739
1740         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1741                 return -EFBIG;
1742         if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1743                 sgp = SGP_CACHE;
1744 repeat:
1745         if (sgp <= SGP_CACHE &&
1746             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1747                 return -EINVAL;
1748         }
1749
1750         sbinfo = SHMEM_SB(inode->i_sb);
1751         charge_mm = vma ? vma->vm_mm : current->mm;
1752
1753         page = find_lock_entry(mapping, index);
1754         if (xa_is_value(page)) {
1755                 error = shmem_swapin_page(inode, index, &page,
1756                                           sgp, gfp, vma, fault_type);
1757                 if (error == -EEXIST)
1758                         goto repeat;
1759
1760                 *pagep = page;
1761                 return error;
1762         }
1763
1764         if (page && sgp == SGP_WRITE)
1765                 mark_page_accessed(page);
1766
1767         /* fallocated page? */
1768         if (page && !PageUptodate(page)) {
1769                 if (sgp != SGP_READ)
1770                         goto clear;
1771                 unlock_page(page);
1772                 put_page(page);
1773                 page = NULL;
1774         }
1775         if (page || sgp == SGP_READ) {
1776                 *pagep = page;
1777                 return 0;
1778         }
1779
1780         /*
1781          * Fast cache lookup did not find it:
1782          * bring it back from swap or allocate.
1783          */
1784
1785         if (vma && userfaultfd_missing(vma)) {
1786                 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1787                 return 0;
1788         }
1789
1790         /* shmem_symlink() */
1791         if (mapping->a_ops != &shmem_aops)
1792                 goto alloc_nohuge;
1793         if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1794                 goto alloc_nohuge;
1795         if (shmem_huge == SHMEM_HUGE_FORCE)
1796                 goto alloc_huge;
1797         switch (sbinfo->huge) {
1798                 loff_t i_size;
1799                 pgoff_t off;
1800         case SHMEM_HUGE_NEVER:
1801                 goto alloc_nohuge;
1802         case SHMEM_HUGE_WITHIN_SIZE:
1803                 off = round_up(index, HPAGE_PMD_NR);
1804                 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1805                 if (i_size >= HPAGE_PMD_SIZE &&
1806                     i_size >> PAGE_SHIFT >= off)
1807                         goto alloc_huge;
1808                 /* fallthrough */
1809         case SHMEM_HUGE_ADVISE:
1810                 if (sgp_huge == SGP_HUGE)
1811                         goto alloc_huge;
1812                 /* TODO: implement fadvise() hints */
1813                 goto alloc_nohuge;
1814         }
1815
1816 alloc_huge:
1817         page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1818         if (IS_ERR(page)) {
1819 alloc_nohuge:
1820                 page = shmem_alloc_and_acct_page(gfp, inode,
1821                                                  index, false);
1822         }
1823         if (IS_ERR(page)) {
1824                 int retry = 5;
1825
1826                 error = PTR_ERR(page);
1827                 page = NULL;
1828                 if (error != -ENOSPC)
1829                         goto unlock;
1830                 /*
1831                  * Try to reclaim some space by splitting a huge page
1832                  * beyond i_size on the filesystem.
1833                  */
1834                 while (retry--) {
1835                         int ret;
1836
1837                         ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1838                         if (ret == SHRINK_STOP)
1839                                 break;
1840                         if (ret)
1841                                 goto alloc_nohuge;
1842                 }
1843                 goto unlock;
1844         }
1845
1846         if (PageTransHuge(page))
1847                 hindex = round_down(index, HPAGE_PMD_NR);
1848         else
1849                 hindex = index;
1850
1851         if (sgp == SGP_WRITE)
1852                 __SetPageReferenced(page);
1853
1854         error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
1855                                             PageTransHuge(page));
1856         if (error)
1857                 goto unacct;
1858         error = shmem_add_to_page_cache(page, mapping, hindex,
1859                                         NULL, gfp & GFP_RECLAIM_MASK);
1860         if (error) {
1861                 mem_cgroup_cancel_charge(page, memcg,
1862                                          PageTransHuge(page));
1863                 goto unacct;
1864         }
1865         mem_cgroup_commit_charge(page, memcg, false,
1866                                  PageTransHuge(page));
1867         lru_cache_add_anon(page);
1868
1869         spin_lock_irq(&info->lock);
1870         info->alloced += 1 << compound_order(page);
1871         inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1872         shmem_recalc_inode(inode);
1873         spin_unlock_irq(&info->lock);
1874         alloced = true;
1875
1876         if (PageTransHuge(page) &&
1877             DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1878                         hindex + HPAGE_PMD_NR - 1) {
1879                 /*
1880                  * Part of the huge page is beyond i_size: subject
1881                  * to shrink under memory pressure.
1882                  */
1883                 spin_lock(&sbinfo->shrinklist_lock);
1884                 /*
1885                  * _careful to defend against unlocked access to
1886                  * ->shrink_list in shmem_unused_huge_shrink()
1887                  */
1888                 if (list_empty_careful(&info->shrinklist)) {
1889                         list_add_tail(&info->shrinklist,
1890                                       &sbinfo->shrinklist);
1891                         sbinfo->shrinklist_len++;
1892                 }
1893                 spin_unlock(&sbinfo->shrinklist_lock);
1894         }
1895
1896         /*
1897          * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1898          */
1899         if (sgp == SGP_FALLOC)
1900                 sgp = SGP_WRITE;
1901 clear:
1902         /*
1903          * Let SGP_WRITE caller clear ends if write does not fill page;
1904          * but SGP_FALLOC on a page fallocated earlier must initialize
1905          * it now, lest undo on failure cancel our earlier guarantee.
1906          */
1907         if (sgp != SGP_WRITE && !PageUptodate(page)) {
1908                 struct page *head = compound_head(page);
1909                 int i;
1910
1911                 for (i = 0; i < (1 << compound_order(head)); i++) {
1912                         clear_highpage(head + i);
1913                         flush_dcache_page(head + i);
1914                 }
1915                 SetPageUptodate(head);
1916         }
1917
1918         /* Perhaps the file has been truncated since we checked */
1919         if (sgp <= SGP_CACHE &&
1920             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1921                 if (alloced) {
1922                         ClearPageDirty(page);
1923                         delete_from_page_cache(page);
1924                         spin_lock_irq(&info->lock);
1925                         shmem_recalc_inode(inode);
1926                         spin_unlock_irq(&info->lock);
1927                 }
1928                 error = -EINVAL;
1929                 goto unlock;
1930         }
1931         *pagep = page + index - hindex;
1932         return 0;
1933
1934         /*
1935          * Error recovery.
1936          */
1937 unacct:
1938         shmem_inode_unacct_blocks(inode, 1 << compound_order(page));
1939
1940         if (PageTransHuge(page)) {
1941                 unlock_page(page);
1942                 put_page(page);
1943                 goto alloc_nohuge;
1944         }
1945 unlock:
1946         if (page) {
1947                 unlock_page(page);
1948                 put_page(page);
1949         }
1950         if (error == -ENOSPC && !once++) {
1951                 spin_lock_irq(&info->lock);
1952                 shmem_recalc_inode(inode);
1953                 spin_unlock_irq(&info->lock);
1954                 goto repeat;
1955         }
1956         if (error == -EEXIST)
1957                 goto repeat;
1958         return error;
1959 }
1960
1961 /*
1962  * This is like autoremove_wake_function, but it removes the wait queue
1963  * entry unconditionally - even if something else had already woken the
1964  * target.
1965  */
1966 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1967 {
1968         int ret = default_wake_function(wait, mode, sync, key);
1969         list_del_init(&wait->entry);
1970         return ret;
1971 }
1972
1973 static vm_fault_t shmem_fault(struct vm_fault *vmf)
1974 {
1975         struct vm_area_struct *vma = vmf->vma;
1976         struct inode *inode = file_inode(vma->vm_file);
1977         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1978         enum sgp_type sgp;
1979         int err;
1980         vm_fault_t ret = VM_FAULT_LOCKED;
1981
1982         /*
1983          * Trinity finds that probing a hole which tmpfs is punching can
1984          * prevent the hole-punch from ever completing: which in turn
1985          * locks writers out with its hold on i_mutex.  So refrain from
1986          * faulting pages into the hole while it's being punched.  Although
1987          * shmem_undo_range() does remove the additions, it may be unable to
1988          * keep up, as each new page needs its own unmap_mapping_range() call,
1989          * and the i_mmap tree grows ever slower to scan if new vmas are added.
1990          *
1991          * It does not matter if we sometimes reach this check just before the
1992          * hole-punch begins, so that one fault then races with the punch:
1993          * we just need to make racing faults a rare case.
1994          *
1995          * The implementation below would be much simpler if we just used a
1996          * standard mutex or completion: but we cannot take i_mutex in fault,
1997          * and bloating every shmem inode for this unlikely case would be sad.
1998          */
1999         if (unlikely(inode->i_private)) {
2000                 struct shmem_falloc *shmem_falloc;
2001
2002                 spin_lock(&inode->i_lock);
2003                 shmem_falloc = inode->i_private;
2004                 if (shmem_falloc &&
2005                     shmem_falloc->waitq &&
2006                     vmf->pgoff >= shmem_falloc->start &&
2007                     vmf->pgoff < shmem_falloc->next) {
2008                         wait_queue_head_t *shmem_falloc_waitq;
2009                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2010
2011                         ret = VM_FAULT_NOPAGE;
2012                         if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
2013                            !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
2014                                 /* It's polite to up mmap_sem if we can */
2015                                 up_read(&vma->vm_mm->mmap_sem);
2016                                 ret = VM_FAULT_RETRY;
2017                         }
2018
2019                         shmem_falloc_waitq = shmem_falloc->waitq;
2020                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2021                                         TASK_UNINTERRUPTIBLE);
2022                         spin_unlock(&inode->i_lock);
2023                         schedule();
2024
2025                         /*
2026                          * shmem_falloc_waitq points into the shmem_fallocate()
2027                          * stack of the hole-punching task: shmem_falloc_waitq
2028                          * is usually invalid by the time we reach here, but
2029                          * finish_wait() does not dereference it in that case;
2030                          * though i_lock needed lest racing with wake_up_all().
2031                          */
2032                         spin_lock(&inode->i_lock);
2033                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2034                         spin_unlock(&inode->i_lock);
2035                         return ret;
2036                 }
2037                 spin_unlock(&inode->i_lock);
2038         }
2039
2040         sgp = SGP_CACHE;
2041
2042         if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2043             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2044                 sgp = SGP_NOHUGE;
2045         else if (vma->vm_flags & VM_HUGEPAGE)
2046                 sgp = SGP_HUGE;
2047
2048         err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2049                                   gfp, vma, vmf, &ret);
2050         if (err)
2051                 return vmf_error(err);
2052         return ret;
2053 }
2054
2055 unsigned long shmem_get_unmapped_area(struct file *file,
2056                                       unsigned long uaddr, unsigned long len,
2057                                       unsigned long pgoff, unsigned long flags)
2058 {
2059         unsigned long (*get_area)(struct file *,
2060                 unsigned long, unsigned long, unsigned long, unsigned long);
2061         unsigned long addr;
2062         unsigned long offset;
2063         unsigned long inflated_len;
2064         unsigned long inflated_addr;
2065         unsigned long inflated_offset;
2066
2067         if (len > TASK_SIZE)
2068                 return -ENOMEM;
2069
2070         get_area = current->mm->get_unmapped_area;
2071         addr = get_area(file, uaddr, len, pgoff, flags);
2072
2073         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
2074                 return addr;
2075         if (IS_ERR_VALUE(addr))
2076                 return addr;
2077         if (addr & ~PAGE_MASK)
2078                 return addr;
2079         if (addr > TASK_SIZE - len)
2080                 return addr;
2081
2082         if (shmem_huge == SHMEM_HUGE_DENY)
2083                 return addr;
2084         if (len < HPAGE_PMD_SIZE)
2085                 return addr;
2086         if (flags & MAP_FIXED)
2087                 return addr;
2088         /*
2089          * Our priority is to support MAP_SHARED mapped hugely;
2090          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2091          * But if caller specified an address hint, respect that as before.
2092          */
2093         if (uaddr)
2094                 return addr;
2095
2096         if (shmem_huge != SHMEM_HUGE_FORCE) {
2097                 struct super_block *sb;
2098
2099                 if (file) {
2100                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2101                         sb = file_inode(file)->i_sb;
2102                 } else {
2103                         /*
2104                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2105                          * for "/dev/zero", to create a shared anonymous object.
2106                          */
2107                         if (IS_ERR(shm_mnt))
2108                                 return addr;
2109                         sb = shm_mnt->mnt_sb;
2110                 }
2111                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2112                         return addr;
2113         }
2114
2115         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2116         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2117                 return addr;
2118         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2119                 return addr;
2120
2121         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2122         if (inflated_len > TASK_SIZE)
2123                 return addr;
2124         if (inflated_len < len)
2125                 return addr;
2126
2127         inflated_addr = get_area(NULL, 0, inflated_len, 0, flags);
2128         if (IS_ERR_VALUE(inflated_addr))
2129                 return addr;
2130         if (inflated_addr & ~PAGE_MASK)
2131                 return addr;
2132
2133         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2134         inflated_addr += offset - inflated_offset;
2135         if (inflated_offset > offset)
2136                 inflated_addr += HPAGE_PMD_SIZE;
2137
2138         if (inflated_addr > TASK_SIZE - len)
2139                 return addr;
2140         return inflated_addr;
2141 }
2142
2143 #ifdef CONFIG_NUMA
2144 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2145 {
2146         struct inode *inode = file_inode(vma->vm_file);
2147         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2148 }
2149
2150 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2151                                           unsigned long addr)
2152 {
2153         struct inode *inode = file_inode(vma->vm_file);
2154         pgoff_t index;
2155
2156         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2157         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2158 }
2159 #endif
2160
2161 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2162 {
2163         struct inode *inode = file_inode(file);
2164         struct shmem_inode_info *info = SHMEM_I(inode);
2165         int retval = -ENOMEM;
2166
2167         spin_lock_irq(&info->lock);
2168         if (lock && !(info->flags & VM_LOCKED)) {
2169                 if (!user_shm_lock(inode->i_size, user))
2170                         goto out_nomem;
2171                 info->flags |= VM_LOCKED;
2172                 mapping_set_unevictable(file->f_mapping);
2173         }
2174         if (!lock && (info->flags & VM_LOCKED) && user) {
2175                 user_shm_unlock(inode->i_size, user);
2176                 info->flags &= ~VM_LOCKED;
2177                 mapping_clear_unevictable(file->f_mapping);
2178         }
2179         retval = 0;
2180
2181 out_nomem:
2182         spin_unlock_irq(&info->lock);
2183         return retval;
2184 }
2185
2186 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2187 {
2188         struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2189
2190         if (info->seals & F_SEAL_FUTURE_WRITE) {
2191                 /*
2192                  * New PROT_WRITE and MAP_SHARED mmaps are not allowed when
2193                  * "future write" seal active.
2194                  */
2195                 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
2196                         return -EPERM;
2197
2198                 /*
2199                  * Since the F_SEAL_FUTURE_WRITE seals allow for a MAP_SHARED
2200                  * read-only mapping, take care to not allow mprotect to revert
2201                  * protections.
2202                  */
2203                 vma->vm_flags &= ~(VM_MAYWRITE);
2204         }
2205
2206         file_accessed(file);
2207         vma->vm_ops = &shmem_vm_ops;
2208         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2209                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2210                         (vma->vm_end & HPAGE_PMD_MASK)) {
2211                 khugepaged_enter(vma, vma->vm_flags);
2212         }
2213         return 0;
2214 }
2215
2216 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2217                                      umode_t mode, dev_t dev, unsigned long flags)
2218 {
2219         struct inode *inode;
2220         struct shmem_inode_info *info;
2221         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2222
2223         if (shmem_reserve_inode(sb))
2224                 return NULL;
2225
2226         inode = new_inode(sb);
2227         if (inode) {
2228                 inode->i_ino = get_next_ino();
2229                 inode_init_owner(inode, dir, mode);
2230                 inode->i_blocks = 0;
2231                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2232                 inode->i_generation = prandom_u32();
2233                 info = SHMEM_I(inode);
2234                 memset(info, 0, (char *)inode - (char *)info);
2235                 spin_lock_init(&info->lock);
2236                 atomic_set(&info->stop_eviction, 0);
2237                 info->seals = F_SEAL_SEAL;
2238                 info->flags = flags & VM_NORESERVE;
2239                 INIT_LIST_HEAD(&info->shrinklist);
2240                 INIT_LIST_HEAD(&info->swaplist);
2241                 simple_xattrs_init(&info->xattrs);
2242                 cache_no_acl(inode);
2243
2244                 switch (mode & S_IFMT) {
2245                 default:
2246                         inode->i_op = &shmem_special_inode_operations;
2247                         init_special_inode(inode, mode, dev);
2248                         break;
2249                 case S_IFREG:
2250                         inode->i_mapping->a_ops = &shmem_aops;
2251                         inode->i_op = &shmem_inode_operations;
2252                         inode->i_fop = &shmem_file_operations;
2253                         mpol_shared_policy_init(&info->policy,
2254                                                  shmem_get_sbmpol(sbinfo));
2255                         break;
2256                 case S_IFDIR:
2257                         inc_nlink(inode);
2258                         /* Some things misbehave if size == 0 on a directory */
2259                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2260                         inode->i_op = &shmem_dir_inode_operations;
2261                         inode->i_fop = &simple_dir_operations;
2262                         break;
2263                 case S_IFLNK:
2264                         /*
2265                          * Must not load anything in the rbtree,
2266                          * mpol_free_shared_policy will not be called.
2267                          */
2268                         mpol_shared_policy_init(&info->policy, NULL);
2269                         break;
2270                 }
2271
2272                 lockdep_annotate_inode_mutex_key(inode);
2273         } else
2274                 shmem_free_inode(sb);
2275         return inode;
2276 }
2277
2278 bool shmem_mapping(struct address_space *mapping)
2279 {
2280         return mapping->a_ops == &shmem_aops;
2281 }
2282
2283 static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2284                                   pmd_t *dst_pmd,
2285                                   struct vm_area_struct *dst_vma,
2286                                   unsigned long dst_addr,
2287                                   unsigned long src_addr,
2288                                   bool zeropage,
2289                                   struct page **pagep)
2290 {
2291         struct inode *inode = file_inode(dst_vma->vm_file);
2292         struct shmem_inode_info *info = SHMEM_I(inode);
2293         struct address_space *mapping = inode->i_mapping;
2294         gfp_t gfp = mapping_gfp_mask(mapping);
2295         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2296         struct mem_cgroup *memcg;
2297         spinlock_t *ptl;
2298         void *page_kaddr;
2299         struct page *page;
2300         pte_t _dst_pte, *dst_pte;
2301         int ret;
2302         pgoff_t offset, max_off;
2303
2304         ret = -ENOMEM;
2305         if (!shmem_inode_acct_block(inode, 1))
2306                 goto out;
2307
2308         if (!*pagep) {
2309                 page = shmem_alloc_page(gfp, info, pgoff);
2310                 if (!page)
2311                         goto out_unacct_blocks;
2312
2313                 if (!zeropage) {        /* mcopy_atomic */
2314                         page_kaddr = kmap_atomic(page);
2315                         ret = copy_from_user(page_kaddr,
2316                                              (const void __user *)src_addr,
2317                                              PAGE_SIZE);
2318                         kunmap_atomic(page_kaddr);
2319
2320                         /* fallback to copy_from_user outside mmap_sem */
2321                         if (unlikely(ret)) {
2322                                 *pagep = page;
2323                                 shmem_inode_unacct_blocks(inode, 1);
2324                                 /* don't free the page */
2325                                 return -ENOENT;
2326                         }
2327                 } else {                /* mfill_zeropage_atomic */
2328                         clear_highpage(page);
2329                 }
2330         } else {
2331                 page = *pagep;
2332                 *pagep = NULL;
2333         }
2334
2335         VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2336         __SetPageLocked(page);
2337         __SetPageSwapBacked(page);
2338         __SetPageUptodate(page);
2339
2340         ret = -EFAULT;
2341         offset = linear_page_index(dst_vma, dst_addr);
2342         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2343         if (unlikely(offset >= max_off))
2344                 goto out_release;
2345
2346         ret = mem_cgroup_try_charge_delay(page, dst_mm, gfp, &memcg, false);
2347         if (ret)
2348                 goto out_release;
2349
2350         ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL,
2351                                                 gfp & GFP_RECLAIM_MASK);
2352         if (ret)
2353                 goto out_release_uncharge;
2354
2355         mem_cgroup_commit_charge(page, memcg, false, false);
2356
2357         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2358         if (dst_vma->vm_flags & VM_WRITE)
2359                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2360         else {
2361                 /*
2362                  * We don't set the pte dirty if the vma has no
2363                  * VM_WRITE permission, so mark the page dirty or it
2364                  * could be freed from under us. We could do it
2365                  * unconditionally before unlock_page(), but doing it
2366                  * only if VM_WRITE is not set is faster.
2367                  */
2368                 set_page_dirty(page);
2369         }
2370
2371         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2372
2373         ret = -EFAULT;
2374         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2375         if (unlikely(offset >= max_off))
2376                 goto out_release_uncharge_unlock;
2377
2378         ret = -EEXIST;
2379         if (!pte_none(*dst_pte))
2380                 goto out_release_uncharge_unlock;
2381
2382         lru_cache_add_anon(page);
2383
2384         spin_lock(&info->lock);
2385         info->alloced++;
2386         inode->i_blocks += BLOCKS_PER_PAGE;
2387         shmem_recalc_inode(inode);
2388         spin_unlock(&info->lock);
2389
2390         inc_mm_counter(dst_mm, mm_counter_file(page));
2391         page_add_file_rmap(page, false);
2392         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2393
2394         /* No need to invalidate - it was non-present before */
2395         update_mmu_cache(dst_vma, dst_addr, dst_pte);
2396         pte_unmap_unlock(dst_pte, ptl);
2397         unlock_page(page);
2398         ret = 0;
2399 out:
2400         return ret;
2401 out_release_uncharge_unlock:
2402         pte_unmap_unlock(dst_pte, ptl);
2403         ClearPageDirty(page);
2404         delete_from_page_cache(page);
2405 out_release_uncharge:
2406         mem_cgroup_cancel_charge(page, memcg, false);
2407 out_release:
2408         unlock_page(page);
2409         put_page(page);
2410 out_unacct_blocks:
2411         shmem_inode_unacct_blocks(inode, 1);
2412         goto out;
2413 }
2414
2415 int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2416                            pmd_t *dst_pmd,
2417                            struct vm_area_struct *dst_vma,
2418                            unsigned long dst_addr,
2419                            unsigned long src_addr,
2420                            struct page **pagep)
2421 {
2422         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2423                                       dst_addr, src_addr, false, pagep);
2424 }
2425
2426 int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2427                              pmd_t *dst_pmd,
2428                              struct vm_area_struct *dst_vma,
2429                              unsigned long dst_addr)
2430 {
2431         struct page *page = NULL;
2432
2433         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2434                                       dst_addr, 0, true, &page);
2435 }
2436
2437 #ifdef CONFIG_TMPFS
2438 static const struct inode_operations shmem_symlink_inode_operations;
2439 static const struct inode_operations shmem_short_symlink_operations;
2440
2441 #ifdef CONFIG_TMPFS_XATTR
2442 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2443 #else
2444 #define shmem_initxattrs NULL
2445 #endif
2446
2447 static int
2448 shmem_write_begin(struct file *file, struct address_space *mapping,
2449                         loff_t pos, unsigned len, unsigned flags,
2450                         struct page **pagep, void **fsdata)
2451 {
2452         struct inode *inode = mapping->host;
2453         struct shmem_inode_info *info = SHMEM_I(inode);
2454         pgoff_t index = pos >> PAGE_SHIFT;
2455
2456         /* i_mutex is held by caller */
2457         if (unlikely(info->seals & (F_SEAL_GROW |
2458                                    F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2459                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2460                         return -EPERM;
2461                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2462                         return -EPERM;
2463         }
2464
2465         return shmem_getpage(inode, index, pagep, SGP_WRITE);
2466 }
2467
2468 static int
2469 shmem_write_end(struct file *file, struct address_space *mapping,
2470                         loff_t pos, unsigned len, unsigned copied,
2471                         struct page *page, void *fsdata)
2472 {
2473         struct inode *inode = mapping->host;
2474
2475         if (pos + copied > inode->i_size)
2476                 i_size_write(inode, pos + copied);
2477
2478         if (!PageUptodate(page)) {
2479                 struct page *head = compound_head(page);
2480                 if (PageTransCompound(page)) {
2481                         int i;
2482
2483                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2484                                 if (head + i == page)
2485                                         continue;
2486                                 clear_highpage(head + i);
2487                                 flush_dcache_page(head + i);
2488                         }
2489                 }
2490                 if (copied < PAGE_SIZE) {
2491                         unsigned from = pos & (PAGE_SIZE - 1);
2492                         zero_user_segments(page, 0, from,
2493                                         from + copied, PAGE_SIZE);
2494                 }
2495                 SetPageUptodate(head);
2496         }
2497         set_page_dirty(page);
2498         unlock_page(page);
2499         put_page(page);
2500
2501         return copied;
2502 }
2503
2504 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2505 {
2506         struct file *file = iocb->ki_filp;
2507         struct inode *inode = file_inode(file);
2508         struct address_space *mapping = inode->i_mapping;
2509         pgoff_t index;
2510         unsigned long offset;
2511         enum sgp_type sgp = SGP_READ;
2512         int error = 0;
2513         ssize_t retval = 0;
2514         loff_t *ppos = &iocb->ki_pos;
2515
2516         /*
2517          * Might this read be for a stacking filesystem?  Then when reading
2518          * holes of a sparse file, we actually need to allocate those pages,
2519          * and even mark them dirty, so it cannot exceed the max_blocks limit.
2520          */
2521         if (!iter_is_iovec(to))
2522                 sgp = SGP_CACHE;
2523
2524         index = *ppos >> PAGE_SHIFT;
2525         offset = *ppos & ~PAGE_MASK;
2526
2527         for (;;) {
2528                 struct page *page = NULL;
2529                 pgoff_t end_index;
2530                 unsigned long nr, ret;
2531                 loff_t i_size = i_size_read(inode);
2532
2533                 end_index = i_size >> PAGE_SHIFT;
2534                 if (index > end_index)
2535                         break;
2536                 if (index == end_index) {
2537                         nr = i_size & ~PAGE_MASK;
2538                         if (nr <= offset)
2539                                 break;
2540                 }
2541
2542                 error = shmem_getpage(inode, index, &page, sgp);
2543                 if (error) {
2544                         if (error == -EINVAL)
2545                                 error = 0;
2546                         break;
2547                 }
2548                 if (page) {
2549                         if (sgp == SGP_CACHE)
2550                                 set_page_dirty(page);
2551                         unlock_page(page);
2552                 }
2553
2554                 /*
2555                  * We must evaluate after, since reads (unlike writes)
2556                  * are called without i_mutex protection against truncate
2557                  */
2558                 nr = PAGE_SIZE;
2559                 i_size = i_size_read(inode);
2560                 end_index = i_size >> PAGE_SHIFT;
2561                 if (index == end_index) {
2562                         nr = i_size & ~PAGE_MASK;
2563                         if (nr <= offset) {
2564                                 if (page)
2565                                         put_page(page);
2566                                 break;
2567                         }
2568                 }
2569                 nr -= offset;
2570
2571                 if (page) {
2572                         /*
2573                          * If users can be writing to this page using arbitrary
2574                          * virtual addresses, take care about potential aliasing
2575                          * before reading the page on the kernel side.
2576                          */
2577                         if (mapping_writably_mapped(mapping))
2578                                 flush_dcache_page(page);
2579                         /*
2580                          * Mark the page accessed if we read the beginning.
2581                          */
2582                         if (!offset)
2583                                 mark_page_accessed(page);
2584                 } else {
2585                         page = ZERO_PAGE(0);
2586                         get_page(page);
2587                 }
2588
2589                 /*
2590                  * Ok, we have the page, and it's up-to-date, so
2591                  * now we can copy it to user space...
2592                  */
2593                 ret = copy_page_to_iter(page, offset, nr, to);
2594                 retval += ret;
2595                 offset += ret;
2596                 index += offset >> PAGE_SHIFT;
2597                 offset &= ~PAGE_MASK;
2598
2599                 put_page(page);
2600                 if (!iov_iter_count(to))
2601                         break;
2602                 if (ret < nr) {
2603                         error = -EFAULT;
2604                         break;
2605                 }
2606                 cond_resched();
2607         }
2608
2609         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2610         file_accessed(file);
2611         return retval ? retval : error;
2612 }
2613
2614 /*
2615  * llseek SEEK_DATA or SEEK_HOLE through the page cache.
2616  */
2617 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2618                                     pgoff_t index, pgoff_t end, int whence)
2619 {
2620         struct page *page;
2621         struct pagevec pvec;
2622         pgoff_t indices[PAGEVEC_SIZE];
2623         bool done = false;
2624         int i;
2625
2626         pagevec_init(&pvec);
2627         pvec.nr = 1;            /* start small: we may be there already */
2628         while (!done) {
2629                 pvec.nr = find_get_entries(mapping, index,
2630                                         pvec.nr, pvec.pages, indices);
2631                 if (!pvec.nr) {
2632                         if (whence == SEEK_DATA)
2633                                 index = end;
2634                         break;
2635                 }
2636                 for (i = 0; i < pvec.nr; i++, index++) {
2637                         if (index < indices[i]) {
2638                                 if (whence == SEEK_HOLE) {
2639                                         done = true;
2640                                         break;
2641                                 }
2642                                 index = indices[i];
2643                         }
2644                         page = pvec.pages[i];
2645                         if (page && !xa_is_value(page)) {
2646                                 if (!PageUptodate(page))
2647                                         page = NULL;
2648                         }
2649                         if (index >= end ||
2650                             (page && whence == SEEK_DATA) ||
2651                             (!page && whence == SEEK_HOLE)) {
2652                                 done = true;
2653                                 break;
2654                         }
2655                 }
2656                 pagevec_remove_exceptionals(&pvec);
2657                 pagevec_release(&pvec);
2658                 pvec.nr = PAGEVEC_SIZE;
2659                 cond_resched();
2660         }
2661         return index;
2662 }
2663
2664 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2665 {
2666         struct address_space *mapping = file->f_mapping;
2667         struct inode *inode = mapping->host;
2668         pgoff_t start, end;
2669         loff_t new_offset;
2670
2671         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2672                 return generic_file_llseek_size(file, offset, whence,
2673                                         MAX_LFS_FILESIZE, i_size_read(inode));
2674         inode_lock(inode);
2675         /* We're holding i_mutex so we can access i_size directly */
2676
2677         if (offset < 0 || offset >= inode->i_size)
2678                 offset = -ENXIO;
2679         else {
2680                 start = offset >> PAGE_SHIFT;
2681                 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2682                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2683                 new_offset <<= PAGE_SHIFT;
2684                 if (new_offset > offset) {
2685                         if (new_offset < inode->i_size)
2686                                 offset = new_offset;
2687                         else if (whence == SEEK_DATA)
2688                                 offset = -ENXIO;
2689                         else
2690                                 offset = inode->i_size;
2691                 }
2692         }
2693
2694         if (offset >= 0)
2695                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2696         inode_unlock(inode);
2697         return offset;
2698 }
2699
2700 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2701                                                          loff_t len)
2702 {
2703         struct inode *inode = file_inode(file);
2704         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2705         struct shmem_inode_info *info = SHMEM_I(inode);
2706         struct shmem_falloc shmem_falloc;
2707         pgoff_t start, index, end;
2708         int error;
2709
2710         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2711                 return -EOPNOTSUPP;
2712
2713         inode_lock(inode);
2714
2715         if (mode & FALLOC_FL_PUNCH_HOLE) {
2716                 struct address_space *mapping = file->f_mapping;
2717                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2718                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2719                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2720
2721                 /* protected by i_mutex */
2722                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2723                         error = -EPERM;
2724                         goto out;
2725                 }
2726
2727                 shmem_falloc.waitq = &shmem_falloc_waitq;
2728                 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2729                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2730                 spin_lock(&inode->i_lock);
2731                 inode->i_private = &shmem_falloc;
2732                 spin_unlock(&inode->i_lock);
2733
2734                 if ((u64)unmap_end > (u64)unmap_start)
2735                         unmap_mapping_range(mapping, unmap_start,
2736                                             1 + unmap_end - unmap_start, 0);
2737                 shmem_truncate_range(inode, offset, offset + len - 1);
2738                 /* No need to unmap again: hole-punching leaves COWed pages */
2739
2740                 spin_lock(&inode->i_lock);
2741                 inode->i_private = NULL;
2742                 wake_up_all(&shmem_falloc_waitq);
2743                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2744                 spin_unlock(&inode->i_lock);
2745                 error = 0;
2746                 goto out;
2747         }
2748
2749         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2750         error = inode_newsize_ok(inode, offset + len);
2751         if (error)
2752                 goto out;
2753
2754         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2755                 error = -EPERM;
2756                 goto out;
2757         }
2758
2759         start = offset >> PAGE_SHIFT;
2760         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2761         /* Try to avoid a swapstorm if len is impossible to satisfy */
2762         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2763                 error = -ENOSPC;
2764                 goto out;
2765         }
2766
2767         shmem_falloc.waitq = NULL;
2768         shmem_falloc.start = start;
2769         shmem_falloc.next  = start;
2770         shmem_falloc.nr_falloced = 0;
2771         shmem_falloc.nr_unswapped = 0;
2772         spin_lock(&inode->i_lock);
2773         inode->i_private = &shmem_falloc;
2774         spin_unlock(&inode->i_lock);
2775
2776         for (index = start; index < end; index++) {
2777                 struct page *page;
2778
2779                 /*
2780                  * Good, the fallocate(2) manpage permits EINTR: we may have
2781                  * been interrupted because we are using up too much memory.
2782                  */
2783                 if (signal_pending(current))
2784                         error = -EINTR;
2785                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2786                         error = -ENOMEM;
2787                 else
2788                         error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2789                 if (error) {
2790                         /* Remove the !PageUptodate pages we added */
2791                         if (index > start) {
2792                                 shmem_undo_range(inode,
2793                                     (loff_t)start << PAGE_SHIFT,
2794                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2795                         }
2796                         goto undone;
2797                 }
2798
2799                 /*
2800                  * Inform shmem_writepage() how far we have reached.
2801                  * No need for lock or barrier: we have the page lock.
2802                  */
2803                 shmem_falloc.next++;
2804                 if (!PageUptodate(page))
2805                         shmem_falloc.nr_falloced++;
2806
2807                 /*
2808                  * If !PageUptodate, leave it that way so that freeable pages
2809                  * can be recognized if we need to rollback on error later.
2810                  * But set_page_dirty so that memory pressure will swap rather
2811                  * than free the pages we are allocating (and SGP_CACHE pages
2812                  * might still be clean: we now need to mark those dirty too).
2813                  */
2814                 set_page_dirty(page);
2815                 unlock_page(page);
2816                 put_page(page);
2817                 cond_resched();
2818         }
2819
2820         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2821                 i_size_write(inode, offset + len);
2822         inode->i_ctime = current_time(inode);
2823 undone:
2824         spin_lock(&inode->i_lock);
2825         inode->i_private = NULL;
2826         spin_unlock(&inode->i_lock);
2827 out:
2828         inode_unlock(inode);
2829         return error;
2830 }
2831
2832 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2833 {
2834         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2835
2836         buf->f_type = TMPFS_MAGIC;
2837         buf->f_bsize = PAGE_SIZE;
2838         buf->f_namelen = NAME_MAX;
2839         if (sbinfo->max_blocks) {
2840                 buf->f_blocks = sbinfo->max_blocks;
2841                 buf->f_bavail =
2842                 buf->f_bfree  = sbinfo->max_blocks -
2843                                 percpu_counter_sum(&sbinfo->used_blocks);
2844         }
2845         if (sbinfo->max_inodes) {
2846                 buf->f_files = sbinfo->max_inodes;
2847                 buf->f_ffree = sbinfo->free_inodes;
2848         }
2849         /* else leave those fields 0 like simple_statfs */
2850         return 0;
2851 }
2852
2853 /*
2854  * File creation. Allocate an inode, and we're done..
2855  */
2856 static int
2857 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2858 {
2859         struct inode *inode;
2860         int error = -ENOSPC;
2861
2862         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2863         if (inode) {
2864                 error = simple_acl_create(dir, inode);
2865                 if (error)
2866                         goto out_iput;
2867                 error = security_inode_init_security(inode, dir,
2868                                                      &dentry->d_name,
2869                                                      shmem_initxattrs, NULL);
2870                 if (error && error != -EOPNOTSUPP)
2871                         goto out_iput;
2872
2873                 error = 0;
2874                 dir->i_size += BOGO_DIRENT_SIZE;
2875                 dir->i_ctime = dir->i_mtime = current_time(dir);
2876                 d_instantiate(dentry, inode);
2877                 dget(dentry); /* Extra count - pin the dentry in core */
2878         }
2879         return error;
2880 out_iput:
2881         iput(inode);
2882         return error;
2883 }
2884
2885 static int
2886 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2887 {
2888         struct inode *inode;
2889         int error = -ENOSPC;
2890
2891         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2892         if (inode) {
2893                 error = security_inode_init_security(inode, dir,
2894                                                      NULL,
2895                                                      shmem_initxattrs, NULL);
2896                 if (error && error != -EOPNOTSUPP)
2897                         goto out_iput;
2898                 error = simple_acl_create(dir, inode);
2899                 if (error)
2900                         goto out_iput;
2901                 d_tmpfile(dentry, inode);
2902         }
2903         return error;
2904 out_iput:
2905         iput(inode);
2906         return error;
2907 }
2908
2909 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2910 {
2911         int error;
2912
2913         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2914                 return error;
2915         inc_nlink(dir);
2916         return 0;
2917 }
2918
2919 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2920                 bool excl)
2921 {
2922         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2923 }
2924
2925 /*
2926  * Link a file..
2927  */
2928 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2929 {
2930         struct inode *inode = d_inode(old_dentry);
2931         int ret = 0;
2932
2933         /*
2934          * No ordinary (disk based) filesystem counts links as inodes;
2935          * but each new link needs a new dentry, pinning lowmem, and
2936          * tmpfs dentries cannot be pruned until they are unlinked.
2937          * But if an O_TMPFILE file is linked into the tmpfs, the
2938          * first link must skip that, to get the accounting right.
2939          */
2940         if (inode->i_nlink) {
2941                 ret = shmem_reserve_inode(inode->i_sb);
2942                 if (ret)
2943                         goto out;
2944         }
2945
2946         dir->i_size += BOGO_DIRENT_SIZE;
2947         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2948         inc_nlink(inode);
2949         ihold(inode);   /* New dentry reference */
2950         dget(dentry);           /* Extra pinning count for the created dentry */
2951         d_instantiate(dentry, inode);
2952 out:
2953         return ret;
2954 }
2955
2956 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2957 {
2958         struct inode *inode = d_inode(dentry);
2959
2960         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2961                 shmem_free_inode(inode->i_sb);
2962
2963         dir->i_size -= BOGO_DIRENT_SIZE;
2964         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2965         drop_nlink(inode);
2966         dput(dentry);   /* Undo the count from "create" - this does all the work */
2967         return 0;
2968 }
2969
2970 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2971 {
2972         if (!simple_empty(dentry))
2973                 return -ENOTEMPTY;
2974
2975         drop_nlink(d_inode(dentry));
2976         drop_nlink(dir);
2977         return shmem_unlink(dir, dentry);
2978 }
2979
2980 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2981 {
2982         bool old_is_dir = d_is_dir(old_dentry);
2983         bool new_is_dir = d_is_dir(new_dentry);
2984
2985         if (old_dir != new_dir && old_is_dir != new_is_dir) {
2986                 if (old_is_dir) {
2987                         drop_nlink(old_dir);
2988                         inc_nlink(new_dir);
2989                 } else {
2990                         drop_nlink(new_dir);
2991                         inc_nlink(old_dir);
2992                 }
2993         }
2994         old_dir->i_ctime = old_dir->i_mtime =
2995         new_dir->i_ctime = new_dir->i_mtime =
2996         d_inode(old_dentry)->i_ctime =
2997         d_inode(new_dentry)->i_ctime = current_time(old_dir);
2998
2999         return 0;
3000 }
3001
3002 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3003 {
3004         struct dentry *whiteout;
3005         int error;
3006
3007         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3008         if (!whiteout)
3009                 return -ENOMEM;
3010
3011         error = shmem_mknod(old_dir, whiteout,
3012                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3013         dput(whiteout);
3014         if (error)
3015                 return error;
3016
3017         /*
3018          * Cheat and hash the whiteout while the old dentry is still in
3019          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3020          *
3021          * d_lookup() will consistently find one of them at this point,
3022          * not sure which one, but that isn't even important.
3023          */
3024         d_rehash(whiteout);
3025         return 0;
3026 }
3027
3028 /*
3029  * The VFS layer already does all the dentry stuff for rename,
3030  * we just have to decrement the usage count for the target if
3031  * it exists so that the VFS layer correctly free's it when it
3032  * gets overwritten.
3033  */
3034 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3035 {
3036         struct inode *inode = d_inode(old_dentry);
3037         int they_are_dirs = S_ISDIR(inode->i_mode);
3038
3039         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3040                 return -EINVAL;
3041
3042         if (flags & RENAME_EXCHANGE)
3043                 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3044
3045         if (!simple_empty(new_dentry))
3046                 return -ENOTEMPTY;
3047
3048         if (flags & RENAME_WHITEOUT) {
3049                 int error;
3050
3051                 error = shmem_whiteout(old_dir, old_dentry);
3052                 if (error)
3053                         return error;
3054         }
3055
3056         if (d_really_is_positive(new_dentry)) {
3057                 (void) shmem_unlink(new_dir, new_dentry);
3058                 if (they_are_dirs) {
3059                         drop_nlink(d_inode(new_dentry));
3060                         drop_nlink(old_dir);
3061                 }
3062         } else if (they_are_dirs) {
3063                 drop_nlink(old_dir);
3064                 inc_nlink(new_dir);
3065         }
3066
3067         old_dir->i_size -= BOGO_DIRENT_SIZE;
3068         new_dir->i_size += BOGO_DIRENT_SIZE;
3069         old_dir->i_ctime = old_dir->i_mtime =
3070         new_dir->i_ctime = new_dir->i_mtime =
3071         inode->i_ctime = current_time(old_dir);
3072         return 0;
3073 }
3074
3075 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3076 {
3077         int error;
3078         int len;
3079         struct inode *inode;
3080         struct page *page;
3081
3082         len = strlen(symname) + 1;
3083         if (len > PAGE_SIZE)
3084                 return -ENAMETOOLONG;
3085
3086         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3087                                 VM_NORESERVE);
3088         if (!inode)
3089                 return -ENOSPC;
3090
3091         error = security_inode_init_security(inode, dir, &dentry->d_name,
3092                                              shmem_initxattrs, NULL);
3093         if (error) {
3094                 if (error != -EOPNOTSUPP) {
3095                         iput(inode);
3096                         return error;
3097                 }
3098                 error = 0;
3099         }
3100
3101         inode->i_size = len-1;
3102         if (len <= SHORT_SYMLINK_LEN) {
3103                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3104                 if (!inode->i_link) {
3105                         iput(inode);
3106                         return -ENOMEM;
3107                 }
3108                 inode->i_op = &shmem_short_symlink_operations;
3109         } else {
3110                 inode_nohighmem(inode);
3111                 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3112                 if (error) {
3113                         iput(inode);
3114                         return error;
3115                 }
3116                 inode->i_mapping->a_ops = &shmem_aops;
3117                 inode->i_op = &shmem_symlink_inode_operations;
3118                 memcpy(page_address(page), symname, len);
3119                 SetPageUptodate(page);
3120                 set_page_dirty(page);
3121                 unlock_page(page);
3122                 put_page(page);
3123         }
3124         dir->i_size += BOGO_DIRENT_SIZE;
3125         dir->i_ctime = dir->i_mtime = current_time(dir);
3126         d_instantiate(dentry, inode);
3127         dget(dentry);
3128         return 0;
3129 }
3130
3131 static void shmem_put_link(void *arg)
3132 {
3133         mark_page_accessed(arg);
3134         put_page(arg);
3135 }
3136
3137 static const char *shmem_get_link(struct dentry *dentry,
3138                                   struct inode *inode,
3139                                   struct delayed_call *done)
3140 {
3141         struct page *page = NULL;
3142         int error;
3143         if (!dentry) {
3144                 page = find_get_page(inode->i_mapping, 0);
3145                 if (!page)
3146                         return ERR_PTR(-ECHILD);
3147                 if (!PageUptodate(page)) {
3148                         put_page(page);
3149                         return ERR_PTR(-ECHILD);
3150                 }
3151         } else {
3152                 error = shmem_getpage(inode, 0, &page, SGP_READ);
3153                 if (error)
3154                         return ERR_PTR(error);
3155                 unlock_page(page);
3156         }
3157         set_delayed_call(done, shmem_put_link, page);
3158         return page_address(page);
3159 }
3160
3161 #ifdef CONFIG_TMPFS_XATTR
3162 /*
3163  * Superblocks without xattr inode operations may get some security.* xattr
3164  * support from the LSM "for free". As soon as we have any other xattrs
3165  * like ACLs, we also need to implement the security.* handlers at
3166  * filesystem level, though.
3167  */
3168
3169 /*
3170  * Callback for security_inode_init_security() for acquiring xattrs.
3171  */
3172 static int shmem_initxattrs(struct inode *inode,
3173                             const struct xattr *xattr_array,
3174                             void *fs_info)
3175 {
3176         struct shmem_inode_info *info = SHMEM_I(inode);
3177         const struct xattr *xattr;
3178         struct simple_xattr *new_xattr;
3179         size_t len;
3180
3181         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3182                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3183                 if (!new_xattr)
3184                         return -ENOMEM;
3185
3186                 len = strlen(xattr->name) + 1;
3187                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3188                                           GFP_KERNEL);
3189                 if (!new_xattr->name) {
3190                         kfree(new_xattr);
3191                         return -ENOMEM;
3192                 }
3193
3194                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3195                        XATTR_SECURITY_PREFIX_LEN);
3196                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3197                        xattr->name, len);
3198
3199                 simple_xattr_list_add(&info->xattrs, new_xattr);
3200         }
3201
3202         return 0;
3203 }
3204
3205 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3206                                    struct dentry *unused, struct inode *inode,
3207                                    const char *name, void *buffer, size_t size)
3208 {
3209         struct shmem_inode_info *info = SHMEM_I(inode);
3210
3211         name = xattr_full_name(handler, name);
3212         return simple_xattr_get(&info->xattrs, name, buffer, size);
3213 }
3214
3215 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3216                                    struct dentry *unused, struct inode *inode,
3217                                    const char *name, const void *value,
3218                                    size_t size, int flags)
3219 {
3220         struct shmem_inode_info *info = SHMEM_I(inode);
3221
3222         name = xattr_full_name(handler, name);
3223         return simple_xattr_set(&info->xattrs, name, value, size, flags);
3224 }
3225
3226 static const struct xattr_handler shmem_security_xattr_handler = {
3227         .prefix = XATTR_SECURITY_PREFIX,
3228         .get = shmem_xattr_handler_get,
3229         .set = shmem_xattr_handler_set,
3230 };
3231
3232 static const struct xattr_handler shmem_trusted_xattr_handler = {
3233         .prefix = XATTR_TRUSTED_PREFIX,
3234         .get = shmem_xattr_handler_get,
3235         .set = shmem_xattr_handler_set,
3236 };
3237
3238 static const struct xattr_handler *shmem_xattr_handlers[] = {
3239 #ifdef CONFIG_TMPFS_POSIX_ACL
3240         &posix_acl_access_xattr_handler,
3241         &posix_acl_default_xattr_handler,
3242 #endif
3243         &shmem_security_xattr_handler,
3244         &shmem_trusted_xattr_handler,
3245         NULL
3246 };
3247
3248 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3249 {
3250         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3251         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3252 }
3253 #endif /* CONFIG_TMPFS_XATTR */
3254
3255 static const struct inode_operations shmem_short_symlink_operations = {
3256         .get_link       = simple_get_link,
3257 #ifdef CONFIG_TMPFS_XATTR
3258         .listxattr      = shmem_listxattr,
3259 #endif
3260 };
3261
3262 static const struct inode_operations shmem_symlink_inode_operations = {
3263         .get_link       = shmem_get_link,
3264 #ifdef CONFIG_TMPFS_XATTR
3265         .listxattr      = shmem_listxattr,
3266 #endif
3267 };
3268
3269 static struct dentry *shmem_get_parent(struct dentry *child)
3270 {
3271         return ERR_PTR(-ESTALE);
3272 }
3273
3274 static int shmem_match(struct inode *ino, void *vfh)
3275 {
3276         __u32 *fh = vfh;
3277         __u64 inum = fh[2];
3278         inum = (inum << 32) | fh[1];
3279         return ino->i_ino == inum && fh[0] == ino->i_generation;
3280 }
3281
3282 /* Find any alias of inode, but prefer a hashed alias */
3283 static struct dentry *shmem_find_alias(struct inode *inode)
3284 {
3285         struct dentry *alias = d_find_alias(inode);
3286
3287         return alias ?: d_find_any_alias(inode);
3288 }
3289
3290
3291 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3292                 struct fid *fid, int fh_len, int fh_type)
3293 {
3294         struct inode *inode;
3295         struct dentry *dentry = NULL;
3296         u64 inum;
3297
3298         if (fh_len < 3)
3299                 return NULL;
3300
3301         inum = fid->raw[2];
3302         inum = (inum << 32) | fid->raw[1];
3303
3304         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3305                         shmem_match, fid->raw);
3306         if (inode) {
3307                 dentry = shmem_find_alias(inode);
3308                 iput(inode);
3309         }
3310
3311         return dentry;
3312 }
3313
3314 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3315                                 struct inode *parent)
3316 {
3317         if (*len < 3) {
3318                 *len = 3;
3319                 return FILEID_INVALID;
3320         }
3321
3322         if (inode_unhashed(inode)) {
3323                 /* Unfortunately insert_inode_hash is not idempotent,
3324                  * so as we hash inodes here rather than at creation
3325                  * time, we need a lock to ensure we only try
3326                  * to do it once
3327                  */
3328                 static DEFINE_SPINLOCK(lock);
3329                 spin_lock(&lock);
3330                 if (inode_unhashed(inode))
3331                         __insert_inode_hash(inode,
3332                                             inode->i_ino + inode->i_generation);
3333                 spin_unlock(&lock);
3334         }
3335
3336         fh[0] = inode->i_generation;
3337         fh[1] = inode->i_ino;
3338         fh[2] = ((__u64)inode->i_ino) >> 32;
3339
3340         *len = 3;
3341         return 1;
3342 }
3343
3344 static const struct export_operations shmem_export_ops = {
3345         .get_parent     = shmem_get_parent,
3346         .encode_fh      = shmem_encode_fh,
3347         .fh_to_dentry   = shmem_fh_to_dentry,
3348 };
3349
3350 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3351                                bool remount)
3352 {
3353         char *this_char, *value, *rest;
3354         struct mempolicy *mpol = NULL;
3355         uid_t uid;
3356         gid_t gid;
3357
3358         while (options != NULL) {
3359                 this_char = options;
3360                 for (;;) {
3361                         /*
3362                          * NUL-terminate this option: unfortunately,
3363                          * mount options form a comma-separated list,
3364                          * but mpol's nodelist may also contain commas.
3365                          */
3366                         options = strchr(options, ',');
3367                         if (options == NULL)
3368                                 break;
3369                         options++;
3370                         if (!isdigit(*options)) {
3371                                 options[-1] = '\0';
3372                                 break;
3373                         }
3374                 }
3375                 if (!*this_char)
3376                         continue;
3377                 if ((value = strchr(this_char,'=')) != NULL) {
3378                         *value++ = 0;
3379                 } else {
3380                         pr_err("tmpfs: No value for mount option '%s'\n",
3381                                this_char);
3382                         goto error;
3383                 }
3384
3385                 if (!strcmp(this_char,"size")) {
3386                         unsigned long long size;
3387                         size = memparse(value,&rest);
3388                         if (*rest == '%') {
3389                                 size <<= PAGE_SHIFT;
3390                                 size *= totalram_pages();
3391                                 do_div(size, 100);
3392                                 rest++;
3393                         }
3394                         if (*rest)
3395                                 goto bad_val;
3396                         sbinfo->max_blocks =
3397                                 DIV_ROUND_UP(size, PAGE_SIZE);
3398                 } else if (!strcmp(this_char,"nr_blocks")) {
3399                         sbinfo->max_blocks = memparse(value, &rest);
3400                         if (*rest)
3401                                 goto bad_val;
3402                 } else if (!strcmp(this_char,"nr_inodes")) {
3403                         sbinfo->max_inodes = memparse(value, &rest);
3404                         if (*rest)
3405                                 goto bad_val;
3406                 } else if (!strcmp(this_char,"mode")) {
3407                         if (remount)
3408                                 continue;
3409                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3410                         if (*rest)
3411                                 goto bad_val;
3412                 } else if (!strcmp(this_char,"uid")) {
3413                         if (remount)
3414                                 continue;
3415                         uid = simple_strtoul(value, &rest, 0);
3416                         if (*rest)
3417                                 goto bad_val;
3418                         sbinfo->uid = make_kuid(current_user_ns(), uid);
3419                         if (!uid_valid(sbinfo->uid))
3420                                 goto bad_val;
3421                 } else if (!strcmp(this_char,"gid")) {
3422                         if (remount)
3423                                 continue;
3424                         gid = simple_strtoul(value, &rest, 0);
3425                         if (*rest)
3426                                 goto bad_val;
3427                         sbinfo->gid = make_kgid(current_user_ns(), gid);
3428                         if (!gid_valid(sbinfo->gid))
3429                                 goto bad_val;
3430 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3431                 } else if (!strcmp(this_char, "huge")) {
3432                         int huge;
3433                         huge = shmem_parse_huge(value);
3434                         if (huge < 0)
3435                                 goto bad_val;
3436                         if (!has_transparent_hugepage() &&
3437                                         huge != SHMEM_HUGE_NEVER)
3438                                 goto bad_val;
3439                         sbinfo->huge = huge;
3440 #endif
3441 #ifdef CONFIG_NUMA
3442                 } else if (!strcmp(this_char,"mpol")) {
3443                         mpol_put(mpol);
3444                         mpol = NULL;
3445                         if (mpol_parse_str(value, &mpol))
3446                                 goto bad_val;
3447 #endif
3448                 } else {
3449                         pr_err("tmpfs: Bad mount option %s\n", this_char);
3450                         goto error;
3451                 }
3452         }
3453         sbinfo->mpol = mpol;
3454         return 0;
3455
3456 bad_val:
3457         pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3458                value, this_char);
3459 error:
3460         mpol_put(mpol);
3461         return 1;
3462
3463 }
3464
3465 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3466 {
3467         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3468         struct shmem_sb_info config = *sbinfo;
3469         unsigned long inodes;
3470         int error = -EINVAL;
3471
3472         config.mpol = NULL;
3473         if (shmem_parse_options(data, &config, true))
3474                 return error;
3475
3476         spin_lock(&sbinfo->stat_lock);
3477         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3478         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3479                 goto out;
3480         if (config.max_inodes < inodes)
3481                 goto out;
3482         /*
3483          * Those tests disallow limited->unlimited while any are in use;
3484          * but we must separately disallow unlimited->limited, because
3485          * in that case we have no record of how much is already in use.
3486          */
3487         if (config.max_blocks && !sbinfo->max_blocks)
3488                 goto out;
3489         if (config.max_inodes && !sbinfo->max_inodes)
3490                 goto out;
3491
3492         error = 0;
3493         sbinfo->huge = config.huge;
3494         sbinfo->max_blocks  = config.max_blocks;
3495         sbinfo->max_inodes  = config.max_inodes;
3496         sbinfo->free_inodes = config.max_inodes - inodes;
3497
3498         /*
3499          * Preserve previous mempolicy unless mpol remount option was specified.
3500          */
3501         if (config.mpol) {
3502                 mpol_put(sbinfo->mpol);
3503                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
3504         }
3505 out:
3506         spin_unlock(&sbinfo->stat_lock);
3507         return error;
3508 }
3509
3510 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3511 {
3512         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3513
3514         if (sbinfo->max_blocks != shmem_default_max_blocks())
3515                 seq_printf(seq, ",size=%luk",
3516                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3517         if (sbinfo->max_inodes != shmem_default_max_inodes())
3518                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3519         if (sbinfo->mode != (0777 | S_ISVTX))
3520                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3521         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3522                 seq_printf(seq, ",uid=%u",
3523                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3524         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3525                 seq_printf(seq, ",gid=%u",
3526                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3527 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3528         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3529         if (sbinfo->huge)
3530                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3531 #endif
3532         shmem_show_mpol(seq, sbinfo->mpol);
3533         return 0;
3534 }
3535
3536 #endif /* CONFIG_TMPFS */
3537
3538 static void shmem_put_super(struct super_block *sb)
3539 {
3540         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3541
3542         percpu_counter_destroy(&sbinfo->used_blocks);
3543         mpol_put(sbinfo->mpol);
3544         kfree(sbinfo);
3545         sb->s_fs_info = NULL;
3546 }
3547
3548 int shmem_fill_super(struct super_block *sb, void *data, int silent)
3549 {
3550         struct inode *inode;
3551         struct shmem_sb_info *sbinfo;
3552         int err = -ENOMEM;
3553
3554         /* Round up to L1_CACHE_BYTES to resist false sharing */
3555         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3556                                 L1_CACHE_BYTES), GFP_KERNEL);
3557         if (!sbinfo)
3558                 return -ENOMEM;
3559
3560         sbinfo->mode = 0777 | S_ISVTX;
3561         sbinfo->uid = current_fsuid();
3562         sbinfo->gid = current_fsgid();
3563         sb->s_fs_info = sbinfo;
3564
3565 #ifdef CONFIG_TMPFS
3566         /*
3567          * Per default we only allow half of the physical ram per
3568          * tmpfs instance, limiting inodes to one per page of lowmem;
3569          * but the internal instance is left unlimited.
3570          */
3571         if (!(sb->s_flags & SB_KERNMOUNT)) {
3572                 sbinfo->max_blocks = shmem_default_max_blocks();
3573                 sbinfo->max_inodes = shmem_default_max_inodes();
3574                 if (shmem_parse_options(data, sbinfo, false)) {
3575                         err = -EINVAL;
3576                         goto failed;
3577                 }
3578         } else {
3579                 sb->s_flags |= SB_NOUSER;
3580         }
3581         sb->s_export_op = &shmem_export_ops;
3582         sb->s_flags |= SB_NOSEC;
3583 #else
3584         sb->s_flags |= SB_NOUSER;
3585 #endif
3586
3587         spin_lock_init(&sbinfo->stat_lock);
3588         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3589                 goto failed;
3590         sbinfo->free_inodes = sbinfo->max_inodes;
3591         spin_lock_init(&sbinfo->shrinklist_lock);
3592         INIT_LIST_HEAD(&sbinfo->shrinklist);
3593
3594         sb->s_maxbytes = MAX_LFS_FILESIZE;
3595         sb->s_blocksize = PAGE_SIZE;
3596         sb->s_blocksize_bits = PAGE_SHIFT;
3597         sb->s_magic = TMPFS_MAGIC;
3598         sb->s_op = &shmem_ops;
3599         sb->s_time_gran = 1;
3600 #ifdef CONFIG_TMPFS_XATTR
3601         sb->s_xattr = shmem_xattr_handlers;
3602 #endif
3603 #ifdef CONFIG_TMPFS_POSIX_ACL
3604         sb->s_flags |= SB_POSIXACL;
3605 #endif
3606         uuid_gen(&sb->s_uuid);
3607
3608         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3609         if (!inode)
3610                 goto failed;
3611         inode->i_uid = sbinfo->uid;
3612         inode->i_gid = sbinfo->gid;
3613         sb->s_root = d_make_root(inode);
3614         if (!sb->s_root)
3615                 goto failed;
3616         return 0;
3617
3618 failed:
3619         shmem_put_super(sb);
3620         return err;
3621 }
3622
3623 static struct kmem_cache *shmem_inode_cachep;
3624
3625 static struct inode *shmem_alloc_inode(struct super_block *sb)
3626 {
3627         struct shmem_inode_info *info;
3628         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3629         if (!info)
3630                 return NULL;
3631         return &info->vfs_inode;
3632 }
3633
3634 static void shmem_destroy_callback(struct rcu_head *head)
3635 {
3636         struct inode *inode = container_of(head, struct inode, i_rcu);
3637         if (S_ISLNK(inode->i_mode))
3638                 kfree(inode->i_link);
3639         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3640 }
3641
3642 static void shmem_destroy_inode(struct inode *inode)
3643 {
3644         if (S_ISREG(inode->i_mode))
3645                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3646         call_rcu(&inode->i_rcu, shmem_destroy_callback);
3647 }
3648
3649 static void shmem_init_inode(void *foo)
3650 {
3651         struct shmem_inode_info *info = foo;
3652         inode_init_once(&info->vfs_inode);
3653 }
3654
3655 static void shmem_init_inodecache(void)
3656 {
3657         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3658                                 sizeof(struct shmem_inode_info),
3659                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3660 }
3661
3662 static void shmem_destroy_inodecache(void)
3663 {
3664         kmem_cache_destroy(shmem_inode_cachep);
3665 }
3666
3667 static const struct address_space_operations shmem_aops = {
3668         .writepage      = shmem_writepage,
3669         .set_page_dirty = __set_page_dirty_no_writeback,
3670 #ifdef CONFIG_TMPFS
3671         .write_begin    = shmem_write_begin,
3672         .write_end      = shmem_write_end,
3673 #endif
3674 #ifdef CONFIG_MIGRATION
3675         .migratepage    = migrate_page,
3676 #endif
3677         .error_remove_page = generic_error_remove_page,
3678 };
3679
3680 static const struct file_operations shmem_file_operations = {
3681         .mmap           = shmem_mmap,
3682         .get_unmapped_area = shmem_get_unmapped_area,
3683 #ifdef CONFIG_TMPFS
3684         .llseek         = shmem_file_llseek,
3685         .read_iter      = shmem_file_read_iter,
3686         .write_iter     = generic_file_write_iter,
3687         .fsync          = noop_fsync,
3688         .splice_read    = generic_file_splice_read,
3689         .splice_write   = iter_file_splice_write,
3690         .fallocate      = shmem_fallocate,
3691 #endif
3692 };
3693
3694 static const struct inode_operations shmem_inode_operations = {
3695         .getattr        = shmem_getattr,
3696         .setattr        = shmem_setattr,
3697 #ifdef CONFIG_TMPFS_XATTR
3698         .listxattr      = shmem_listxattr,
3699         .set_acl        = simple_set_acl,
3700 #endif
3701 };
3702
3703 static const struct inode_operations shmem_dir_inode_operations = {
3704 #ifdef CONFIG_TMPFS
3705         .create         = shmem_create,
3706         .lookup         = simple_lookup,
3707         .link           = shmem_link,
3708         .unlink         = shmem_unlink,
3709         .symlink        = shmem_symlink,
3710         .mkdir          = shmem_mkdir,
3711         .rmdir          = shmem_rmdir,
3712         .mknod          = shmem_mknod,
3713         .rename         = shmem_rename2,
3714         .tmpfile        = shmem_tmpfile,
3715 #endif
3716 #ifdef CONFIG_TMPFS_XATTR
3717         .listxattr      = shmem_listxattr,
3718 #endif
3719 #ifdef CONFIG_TMPFS_POSIX_ACL
3720         .setattr        = shmem_setattr,
3721         .set_acl        = simple_set_acl,
3722 #endif
3723 };
3724
3725 static const struct inode_operations shmem_special_inode_operations = {
3726 #ifdef CONFIG_TMPFS_XATTR
3727         .listxattr      = shmem_listxattr,
3728 #endif
3729 #ifdef CONFIG_TMPFS_POSIX_ACL
3730         .setattr        = shmem_setattr,
3731         .set_acl        = simple_set_acl,
3732 #endif
3733 };
3734
3735 static const struct super_operations shmem_ops = {
3736         .alloc_inode    = shmem_alloc_inode,
3737         .destroy_inode  = shmem_destroy_inode,
3738 #ifdef CONFIG_TMPFS
3739         .statfs         = shmem_statfs,
3740         .remount_fs     = shmem_remount_fs,
3741         .show_options   = shmem_show_options,
3742 #endif
3743         .evict_inode    = shmem_evict_inode,
3744         .drop_inode     = generic_delete_inode,
3745         .put_super      = shmem_put_super,
3746 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3747         .nr_cached_objects      = shmem_unused_huge_count,
3748         .free_cached_objects    = shmem_unused_huge_scan,
3749 #endif
3750 };
3751
3752 static const struct vm_operations_struct shmem_vm_ops = {
3753         .fault          = shmem_fault,
3754         .map_pages      = filemap_map_pages,
3755 #ifdef CONFIG_NUMA
3756         .set_policy     = shmem_set_policy,
3757         .get_policy     = shmem_get_policy,
3758 #endif
3759 };
3760
3761 static struct dentry *shmem_mount(struct file_system_type *fs_type,
3762         int flags, const char *dev_name, void *data)
3763 {
3764         return mount_nodev(fs_type, flags, data, shmem_fill_super);
3765 }
3766
3767 static struct file_system_type shmem_fs_type = {
3768         .owner          = THIS_MODULE,
3769         .name           = "tmpfs",
3770         .mount          = shmem_mount,
3771         .kill_sb        = kill_litter_super,
3772         .fs_flags       = FS_USERNS_MOUNT,
3773 };
3774
3775 int __init shmem_init(void)
3776 {
3777         int error;
3778
3779         /* If rootfs called this, don't re-init */
3780         if (shmem_inode_cachep)
3781                 return 0;
3782
3783         shmem_init_inodecache();
3784
3785         error = register_filesystem(&shmem_fs_type);
3786         if (error) {
3787                 pr_err("Could not register tmpfs\n");
3788                 goto out2;
3789         }
3790
3791         shm_mnt = kern_mount(&shmem_fs_type);
3792         if (IS_ERR(shm_mnt)) {
3793                 error = PTR_ERR(shm_mnt);
3794                 pr_err("Could not kern_mount tmpfs\n");
3795                 goto out1;
3796         }
3797
3798 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3799         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
3800                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3801         else
3802                 shmem_huge = 0; /* just in case it was patched */
3803 #endif
3804         return 0;
3805
3806 out1:
3807         unregister_filesystem(&shmem_fs_type);
3808 out2:
3809         shmem_destroy_inodecache();
3810         shm_mnt = ERR_PTR(error);
3811         return error;
3812 }
3813
3814 #if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
3815 static ssize_t shmem_enabled_show(struct kobject *kobj,
3816                 struct kobj_attribute *attr, char *buf)
3817 {
3818         int values[] = {
3819                 SHMEM_HUGE_ALWAYS,
3820                 SHMEM_HUGE_WITHIN_SIZE,
3821                 SHMEM_HUGE_ADVISE,
3822                 SHMEM_HUGE_NEVER,
3823                 SHMEM_HUGE_DENY,
3824                 SHMEM_HUGE_FORCE,
3825         };
3826         int i, count;
3827
3828         for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
3829                 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
3830
3831                 count += sprintf(buf + count, fmt,
3832                                 shmem_format_huge(values[i]));
3833         }
3834         buf[count - 1] = '\n';
3835         return count;
3836 }
3837
3838 static ssize_t shmem_enabled_store(struct kobject *kobj,
3839                 struct kobj_attribute *attr, const char *buf, size_t count)
3840 {
3841         char tmp[16];
3842         int huge;
3843
3844         if (count + 1 > sizeof(tmp))
3845                 return -EINVAL;
3846         memcpy(tmp, buf, count);
3847         tmp[count] = '\0';
3848         if (count && tmp[count - 1] == '\n')
3849                 tmp[count - 1] = '\0';
3850
3851         huge = shmem_parse_huge(tmp);
3852         if (huge == -EINVAL)
3853                 return -EINVAL;
3854         if (!has_transparent_hugepage() &&
3855                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
3856                 return -EINVAL;
3857
3858         shmem_huge = huge;
3859         if (shmem_huge > SHMEM_HUGE_DENY)
3860                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3861         return count;
3862 }
3863
3864 struct kobj_attribute shmem_enabled_attr =
3865         __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
3866 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE && CONFIG_SYSFS */
3867
3868 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3869 bool shmem_huge_enabled(struct vm_area_struct *vma)
3870 {
3871         struct inode *inode = file_inode(vma->vm_file);
3872         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3873         loff_t i_size;
3874         pgoff_t off;
3875
3876         if (shmem_huge == SHMEM_HUGE_FORCE)
3877                 return true;
3878         if (shmem_huge == SHMEM_HUGE_DENY)
3879                 return false;
3880         switch (sbinfo->huge) {
3881                 case SHMEM_HUGE_NEVER:
3882                         return false;
3883                 case SHMEM_HUGE_ALWAYS:
3884                         return true;
3885                 case SHMEM_HUGE_WITHIN_SIZE:
3886                         off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
3887                         i_size = round_up(i_size_read(inode), PAGE_SIZE);
3888                         if (i_size >= HPAGE_PMD_SIZE &&
3889                                         i_size >> PAGE_SHIFT >= off)
3890                                 return true;
3891                         /* fall through */
3892                 case SHMEM_HUGE_ADVISE:
3893                         /* TODO: implement fadvise() hints */
3894                         return (vma->vm_flags & VM_HUGEPAGE);
3895                 default:
3896                         VM_BUG_ON(1);
3897                         return false;
3898         }
3899 }
3900 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
3901
3902 #else /* !CONFIG_SHMEM */
3903
3904 /*
3905  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3906  *
3907  * This is intended for small system where the benefits of the full
3908  * shmem code (swap-backed and resource-limited) are outweighed by
3909  * their complexity. On systems without swap this code should be
3910  * effectively equivalent, but much lighter weight.
3911  */
3912
3913 static struct file_system_type shmem_fs_type = {
3914         .name           = "tmpfs",
3915         .mount          = ramfs_mount,
3916         .kill_sb        = kill_litter_super,
3917         .fs_flags       = FS_USERNS_MOUNT,
3918 };
3919
3920 int __init shmem_init(void)
3921 {
3922         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3923
3924         shm_mnt = kern_mount(&shmem_fs_type);
3925         BUG_ON(IS_ERR(shm_mnt));
3926
3927         return 0;
3928 }
3929
3930 int shmem_unuse(unsigned int type, bool frontswap,
3931                 unsigned long *fs_pages_to_unuse)
3932 {
3933         return 0;
3934 }
3935
3936 int shmem_lock(struct file *file, int lock, struct user_struct *user)
3937 {
3938         return 0;
3939 }
3940
3941 void shmem_unlock_mapping(struct address_space *mapping)
3942 {
3943 }
3944
3945 #ifdef CONFIG_MMU
3946 unsigned long shmem_get_unmapped_area(struct file *file,
3947                                       unsigned long addr, unsigned long len,
3948                                       unsigned long pgoff, unsigned long flags)
3949 {
3950         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
3951 }
3952 #endif
3953
3954 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3955 {
3956         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3957 }
3958 EXPORT_SYMBOL_GPL(shmem_truncate_range);
3959
3960 #define shmem_vm_ops                            generic_file_vm_ops
3961 #define shmem_file_operations                   ramfs_file_operations
3962 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
3963 #define shmem_acct_size(flags, size)            0
3964 #define shmem_unacct_size(flags, size)          do {} while (0)
3965
3966 #endif /* CONFIG_SHMEM */
3967
3968 /* common code */
3969
3970 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
3971                                        unsigned long flags, unsigned int i_flags)
3972 {
3973         struct inode *inode;
3974         struct file *res;
3975
3976         if (IS_ERR(mnt))
3977                 return ERR_CAST(mnt);
3978
3979         if (size < 0 || size > MAX_LFS_FILESIZE)
3980                 return ERR_PTR(-EINVAL);
3981
3982         if (shmem_acct_size(flags, size))
3983                 return ERR_PTR(-ENOMEM);
3984
3985         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
3986                                 flags);
3987         if (unlikely(!inode)) {
3988                 shmem_unacct_size(flags, size);
3989                 return ERR_PTR(-ENOSPC);
3990         }
3991         inode->i_flags |= i_flags;
3992         inode->i_size = size;
3993         clear_nlink(inode);     /* It is unlinked */
3994         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
3995         if (!IS_ERR(res))
3996                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
3997                                 &shmem_file_operations);
3998         if (IS_ERR(res))
3999                 iput(inode);
4000         return res;
4001 }
4002
4003 /**
4004  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4005  *      kernel internal.  There will be NO LSM permission checks against the
4006  *      underlying inode.  So users of this interface must do LSM checks at a
4007  *      higher layer.  The users are the big_key and shm implementations.  LSM
4008  *      checks are provided at the key or shm level rather than the inode.
4009  * @name: name for dentry (to be seen in /proc/<pid>/maps
4010  * @size: size to be set for the file
4011  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4012  */
4013 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4014 {
4015         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4016 }
4017
4018 /**
4019  * shmem_file_setup - get an unlinked file living in tmpfs
4020  * @name: name for dentry (to be seen in /proc/<pid>/maps
4021  * @size: size to be set for the file
4022  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4023  */
4024 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4025 {
4026         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4027 }
4028 EXPORT_SYMBOL_GPL(shmem_file_setup);
4029
4030 /**
4031  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4032  * @mnt: the tmpfs mount where the file will be created
4033  * @name: name for dentry (to be seen in /proc/<pid>/maps
4034  * @size: size to be set for the file
4035  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4036  */
4037 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4038                                        loff_t size, unsigned long flags)
4039 {
4040         return __shmem_file_setup(mnt, name, size, flags, 0);
4041 }
4042 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4043
4044 /**
4045  * shmem_zero_setup - setup a shared anonymous mapping
4046  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
4047  */
4048 int shmem_zero_setup(struct vm_area_struct *vma)
4049 {
4050         struct file *file;
4051         loff_t size = vma->vm_end - vma->vm_start;
4052
4053         /*
4054          * Cloning a new file under mmap_sem leads to a lock ordering conflict
4055          * between XFS directory reading and selinux: since this file is only
4056          * accessible to the user through its mapping, use S_PRIVATE flag to
4057          * bypass file security, in the same way as shmem_kernel_file_setup().
4058          */
4059         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4060         if (IS_ERR(file))
4061                 return PTR_ERR(file);
4062
4063         if (vma->vm_file)
4064                 fput(vma->vm_file);
4065         vma->vm_file = file;
4066         vma->vm_ops = &shmem_vm_ops;
4067
4068         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
4069                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4070                         (vma->vm_end & HPAGE_PMD_MASK)) {
4071                 khugepaged_enter(vma, vma->vm_flags);
4072         }
4073
4074         return 0;
4075 }
4076
4077 /**
4078  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4079  * @mapping:    the page's address_space
4080  * @index:      the page index
4081  * @gfp:        the page allocator flags to use if allocating
4082  *
4083  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4084  * with any new page allocations done using the specified allocation flags.
4085  * But read_cache_page_gfp() uses the ->readpage() method: which does not
4086  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4087  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4088  *
4089  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4090  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4091  */
4092 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4093                                          pgoff_t index, gfp_t gfp)
4094 {
4095 #ifdef CONFIG_SHMEM
4096         struct inode *inode = mapping->host;
4097         struct page *page;
4098         int error;
4099
4100         BUG_ON(mapping->a_ops != &shmem_aops);
4101         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4102                                   gfp, NULL, NULL, NULL);
4103         if (error)
4104                 page = ERR_PTR(error);
4105         else
4106                 unlock_page(page);
4107         return page;
4108 #else
4109         /*
4110          * The tiny !SHMEM case uses ramfs without swap
4111          */
4112         return read_cache_page_gfp(mapping, index, gfp);
4113 #endif
4114 }
4115 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);