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/hugetlb.h>
  38 #include <linux/fs_parser.h>
  39 #include <linux/swapfile.h>
  40 #include "swap.h"
  41
  42 static struct vfsmount *shm_mnt;
  43
  44 #ifdef CONFIG_SHMEM
  45 /*
  46  * This virtual memory filesystem is heavily based on the ramfs. It
  47  * extends ramfs by the ability to use swap and honor resource limits
  48  * which makes it a completely usable filesystem.
  49  */
  50
  51 #include <linux/xattr.h>
  52 #include <linux/exportfs.h>
  53 #include <linux/posix_acl.h>
  54 #include <linux/posix_acl_xattr.h>
  55 #include <linux/mman.h>
  56 #include <linux/string.h>
  57 #include <linux/slab.h>
  58 #include <linux/backing-dev.h>
  59 #include <linux/shmem_fs.h>
  60 #include <linux/writeback.h>
  61 #include <linux/pagevec.h>
  62 #include <linux/percpu_counter.h>
  63 #include <linux/falloc.h>
  64 #include <linux/splice.h>
  65 #include <linux/security.h>
  66 #include <linux/swapops.h>
  67 #include <linux/mempolicy.h>
  68 #include <linux/namei.h>
  69 #include <linux/ctype.h>
  70 #include <linux/migrate.h>
  71 #include <linux/highmem.h>
  72 #include <linux/seq_file.h>
  73 #include <linux/magic.h>
  74 #include <linux/syscalls.h>
  75 #include <linux/fcntl.h>
  76 #include <uapi/linux/memfd.h>
  77 #include <linux/userfaultfd_k.h>
  78 #include <linux/rmap.h>
  79 #include <linux/uuid.h>
  80
  81 #include <linux/uaccess.h>
  82
  83 #include "internal.h"
  84
  85 #define BLOCKS_PER_PAGE  (PAGE_SIZE/512)
  86 #define VM_ACCT(size)    (PAGE_ALIGN(size) >> PAGE_SHIFT)
  87
  88 /* Pretend that each entry is of this size in directory's i_size */
  89 #define BOGO_DIRENT_SIZE 20
  90
  91 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  92 #define SHORT_SYMLINK_LEN 128
  93
  94 /*
  95  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
  96  * inode->i_private (with i_rwsem making sure that it has only one user at
  97  * a time): we would prefer not to enlarge the shmem inode just for that.
  98  */
  99 struct shmem_falloc {
 100         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
 101         pgoff_t start;          /* start of range currently being fallocated */
 102         pgoff_t next;           /* the next page offset to be fallocated */
 103         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
 104         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
 105 };
 106
 107 struct shmem_options {
 108         unsigned long long blocks;
 109         unsigned long long inodes;
 110         struct mempolicy *mpol;
 111         kuid_t uid;
 112         kgid_t gid;
 113         umode_t mode;
 114         bool full_inums;
 115         int huge;
 116         int seen;
 117 #define SHMEM_SEEN_BLOCKS 1
 118 #define SHMEM_SEEN_INODES 2
 119 #define SHMEM_SEEN_HUGE 4
 120 #define SHMEM_SEEN_INUMS 8
 121 };
 122
 123 #ifdef CONFIG_TMPFS
 124 static unsigned long shmem_default_max_blocks(void)
 125 {
 126         return totalram_pages() / 2;
 127 }
 128
 129 static unsigned long shmem_default_max_inodes(void)
 130 {
 131         unsigned long nr_pages = totalram_pages();
 132
 133         return min(nr_pages - totalhigh_pages(), nr_pages / 2);
 134 }
 135 #endif
 136
 137 static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
 138                              struct folio **foliop, enum sgp_type sgp,
 139                              gfp_t gfp, struct vm_area_struct *vma,
 140                              vm_fault_t *fault_type);
 141 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 142                 struct page **pagep, enum sgp_type sgp,
 143                 gfp_t gfp, struct vm_area_struct *vma,
 144                 struct vm_fault *vmf, vm_fault_t *fault_type);
 145
 146 int shmem_getpage(struct inode *inode, pgoff_t index,
 147                 struct page **pagep, enum sgp_type sgp)
 148 {
 149         return shmem_getpage_gfp(inode, index, pagep, sgp,
 150                 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
 151 }
 152
 153 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 154 {
 155         return sb->s_fs_info;
 156 }
 157
 158 /*
 159  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 160  * for shared memory and for shared anonymous (/dev/zero) mappings
 161  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 162  * consistent with the pre-accounting of private mappings ...
 163  */
 164 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 165 {
 166         return (flags & VM_NORESERVE) ?
 167                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 168 }
 169
 170 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 171 {
 172         if (!(flags & VM_NORESERVE))
 173                 vm_unacct_memory(VM_ACCT(size));
 174 }
 175
 176 static inline int shmem_reacct_size(unsigned long flags,
 177                 loff_t oldsize, loff_t newsize)
 178 {
 179         if (!(flags & VM_NORESERVE)) {
 180                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
 181                         return security_vm_enough_memory_mm(current->mm,
 182                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
 183                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
 184                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
 185         }
 186         return 0;
 187 }
 188
 189 /*
 190  * ... whereas tmpfs objects are accounted incrementally as
 191  * pages are allocated, in order to allow large sparse files.
 192  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 193  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 194  */
 195 static inline int shmem_acct_block(unsigned long flags, long pages)
 196 {
 197         if (!(flags & VM_NORESERVE))
 198                 return 0;
 199
 200         return security_vm_enough_memory_mm(current->mm,
 201                         pages * VM_ACCT(PAGE_SIZE));
 202 }
 203
 204 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 205 {
 206         if (flags & VM_NORESERVE)
 207                 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
 208 }
 209
 210 static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
 211 {
 212         struct shmem_inode_info *info = SHMEM_I(inode);
 213         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 214
 215         if (shmem_acct_block(info->flags, pages))
 216                 return false;
 217
 218         if (sbinfo->max_blocks) {
 219                 if (percpu_counter_compare(&sbinfo->used_blocks,
 220                                            sbinfo->max_blocks - pages) > 0)
 221                         goto unacct;
 222                 percpu_counter_add(&sbinfo->used_blocks, pages);
 223         }
 224
 225         return true;
 226
 227 unacct:
 228         shmem_unacct_blocks(info->flags, pages);
 229         return false;
 230 }
 231
 232 static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
 233 {
 234         struct shmem_inode_info *info = SHMEM_I(inode);
 235         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 236
 237         if (sbinfo->max_blocks)
 238                 percpu_counter_sub(&sbinfo->used_blocks, pages);
 239         shmem_unacct_blocks(info->flags, pages);
 240 }
 241
 242 static const struct super_operations shmem_ops;
 243 const struct address_space_operations shmem_aops;
 244 static const struct file_operations shmem_file_operations;
 245 static const struct inode_operations shmem_inode_operations;
 246 static const struct inode_operations shmem_dir_inode_operations;
 247 static const struct inode_operations shmem_special_inode_operations;
 248 static const struct vm_operations_struct shmem_vm_ops;
 249 static struct file_system_type shmem_fs_type;
 250
 251 bool vma_is_shmem(struct vm_area_struct *vma)
 252 {
 253         return vma->vm_ops == &shmem_vm_ops;
 254 }
 255
 256 static LIST_HEAD(shmem_swaplist);
 257 static DEFINE_MUTEX(shmem_swaplist_mutex);
 258
 259 /*
 260  * shmem_reserve_inode() performs bookkeeping to reserve a shmem inode, and
 261  * produces a novel ino for the newly allocated inode.
 262  *
 263  * It may also be called when making a hard link to permit the space needed by
 264  * each dentry. However, in that case, no new inode number is needed since that
 265  * internally draws from another pool of inode numbers (currently global
 266  * get_next_ino()). This case is indicated by passing NULL as inop.
 267  */
 268 #define SHMEM_INO_BATCH 1024
 269 static int shmem_reserve_inode(struct super_block *sb, ino_t *inop)
 270 {
 271         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 272         ino_t ino;
 273
 274         if (!(sb->s_flags & SB_KERNMOUNT)) {
 275                 raw_spin_lock(&sbinfo->stat_lock);
 276                 if (sbinfo->max_inodes) {
 277                         if (!sbinfo->free_inodes) {
 278                                 raw_spin_unlock(&sbinfo->stat_lock);
 279                                 return -ENOSPC;
 280                         }
 281                         sbinfo->free_inodes--;
 282                 }
 283                 if (inop) {
 284                         ino = sbinfo->next_ino++;
 285                         if (unlikely(is_zero_ino(ino)))
 286                                 ino = sbinfo->next_ino++;
 287                         if (unlikely(!sbinfo->full_inums &&
 288                                      ino > UINT_MAX)) {
 289                                 /*
 290                                  * Emulate get_next_ino uint wraparound for
 291                                  * compatibility
 292                                  */
 293                                 if (IS_ENABLED(CONFIG_64BIT))
 294                                         pr_warn("%s: inode number overflow on device %d, consider using inode64 mount option\n",
 295                                                 __func__, MINOR(sb->s_dev));
 296                                 sbinfo->next_ino = 1;
 297                                 ino = sbinfo->next_ino++;
 298                         }
 299                         *inop = ino;
 300                 }
 301                 raw_spin_unlock(&sbinfo->stat_lock);
 302         } else if (inop) {
 303                 /*
 304                  * __shmem_file_setup, one of our callers, is lock-free: it
 305                  * doesn't hold stat_lock in shmem_reserve_inode since
 306                  * max_inodes is always 0, and is called from potentially
 307                  * unknown contexts. As such, use a per-cpu batched allocator
 308                  * which doesn't require the per-sb stat_lock unless we are at
 309                  * the batch boundary.
 310                  *
 311                  * We don't need to worry about inode{32,64} since SB_KERNMOUNT
 312                  * shmem mounts are not exposed to userspace, so we don't need
 313                  * to worry about things like glibc compatibility.
 314                  */
 315                 ino_t *next_ino;
 316
 317                 next_ino = per_cpu_ptr(sbinfo->ino_batch, get_cpu());
 318                 ino = *next_ino;
 319                 if (unlikely(ino % SHMEM_INO_BATCH == 0)) {
 320                         raw_spin_lock(&sbinfo->stat_lock);
 321                         ino = sbinfo->next_ino;
 322                         sbinfo->next_ino += SHMEM_INO_BATCH;
 323                         raw_spin_unlock(&sbinfo->stat_lock);
 324                         if (unlikely(is_zero_ino(ino)))
 325                                 ino++;
 326                 }
 327                 *inop = ino;
 328                 *next_ino = ++ino;
 329                 put_cpu();
 330         }
 331
 332         return 0;
 333 }
 334
 335 static void shmem_free_inode(struct super_block *sb)
 336 {
 337         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 338         if (sbinfo->max_inodes) {
 339                 raw_spin_lock(&sbinfo->stat_lock);
 340                 sbinfo->free_inodes++;
 341                 raw_spin_unlock(&sbinfo->stat_lock);
 342         }
 343 }
 344
 345 /**
 346  * shmem_recalc_inode - recalculate the block usage of an inode
 347  * @inode: inode to recalc
 348  *
 349  * We have to calculate the free blocks since the mm can drop
 350  * undirtied hole pages behind our back.
 351  *
 352  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 353  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 354  *
 355  * It has to be called with the spinlock held.
 356  */
 357 static void shmem_recalc_inode(struct inode *inode)
 358 {
 359         struct shmem_inode_info *info = SHMEM_I(inode);
 360         long freed;
 361
 362         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 363         if (freed > 0) {
 364                 info->alloced -= freed;
 365                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 366                 shmem_inode_unacct_blocks(inode, freed);
 367         }
 368 }
 369
 370 bool shmem_charge(struct inode *inode, long pages)
 371 {
 372         struct shmem_inode_info *info = SHMEM_I(inode);
 373         unsigned long flags;
 374
 375         if (!shmem_inode_acct_block(inode, pages))
 376                 return false;
 377
 378         /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
 379         inode->i_mapping->nrpages += pages;
 380
 381         spin_lock_irqsave(&info->lock, flags);
 382         info->alloced += pages;
 383         inode->i_blocks += pages * BLOCKS_PER_PAGE;
 384         shmem_recalc_inode(inode);
 385         spin_unlock_irqrestore(&info->lock, flags);
 386
 387         return true;
 388 }
 389
 390 void shmem_uncharge(struct inode *inode, long pages)
 391 {
 392         struct shmem_inode_info *info = SHMEM_I(inode);
 393         unsigned long flags;
 394
 395         /* nrpages adjustment done by __delete_from_page_cache() or caller */
 396
 397         spin_lock_irqsave(&info->lock, flags);
 398         info->alloced -= pages;
 399         inode->i_blocks -= pages * BLOCKS_PER_PAGE;
 400         shmem_recalc_inode(inode);
 401         spin_unlock_irqrestore(&info->lock, flags);
 402
 403         shmem_inode_unacct_blocks(inode, pages);
 404 }
 405
 406 /*
 407  * Replace item expected in xarray by a new item, while holding xa_lock.
 408  */
 409 static int shmem_replace_entry(struct address_space *mapping,
 410                         pgoff_t index, void *expected, void *replacement)
 411 {
 412         XA_STATE(xas, &mapping->i_pages, index);
 413         void *item;
 414
 415         VM_BUG_ON(!expected);
 416         VM_BUG_ON(!replacement);
 417         item = xas_load(&xas);
 418         if (item != expected)
 419                 return -ENOENT;
 420         xas_store(&xas, replacement);
 421         return 0;
 422 }
 423
 424 /*
 425  * Sometimes, before we decide whether to proceed or to fail, we must check
 426  * that an entry was not already brought back from swap by a racing thread.
 427  *
 428  * Checking page is not enough: by the time a SwapCache page is locked, it
 429  * might be reused, and again be SwapCache, using the same swap as before.
 430  */
 431 static bool shmem_confirm_swap(struct address_space *mapping,
 432                                pgoff_t index, swp_entry_t swap)
 433 {
 434         return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
 435 }
 436
 437 /*
 438  * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
 439  *
 440  * SHMEM_HUGE_NEVER:
 441  *      disables huge pages for the mount;
 442  * SHMEM_HUGE_ALWAYS:
 443  *      enables huge pages for the mount;
 444  * SHMEM_HUGE_WITHIN_SIZE:
 445  *      only allocate huge pages if the page will be fully within i_size,
 446  *      also respect fadvise()/madvise() hints;
 447  * SHMEM_HUGE_ADVISE:
 448  *      only allocate huge pages if requested with fadvise()/madvise();
 449  */
 450
 451 #define SHMEM_HUGE_NEVER        0
 452 #define SHMEM_HUGE_ALWAYS       1
 453 #define SHMEM_HUGE_WITHIN_SIZE  2
 454 #define SHMEM_HUGE_ADVISE       3
 455
 456 /*
 457  * Special values.
 458  * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
 459  *
 460  * SHMEM_HUGE_DENY:
 461  *      disables huge on shm_mnt and all mounts, for emergency use;
 462  * SHMEM_HUGE_FORCE:
 463  *      enables huge on shm_mnt and all mounts, w/o needing option, for testing;
 464  *
 465  */
 466 #define SHMEM_HUGE_DENY         (-1)
 467 #define SHMEM_HUGE_FORCE        (-2)
 468
 469 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 470 /* ifdef here to avoid bloating shmem.o when not necessary */
 471
 472 static int shmem_huge __read_mostly = SHMEM_HUGE_NEVER;
 473
 474 bool shmem_is_huge(struct vm_area_struct *vma,
 475                    struct inode *inode, pgoff_t index)
 476 {
 477         loff_t i_size;
 478
 479         if (!S_ISREG(inode->i_mode))
 480                 return false;
 481         if (shmem_huge == SHMEM_HUGE_DENY)
 482                 return false;
 483         if (vma && ((vma->vm_flags & VM_NOHUGEPAGE) ||
 484             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)))
 485                 return false;
 486         if (shmem_huge == SHMEM_HUGE_FORCE)
 487                 return true;
 488
 489         switch (SHMEM_SB(inode->i_sb)->huge) {
 490         case SHMEM_HUGE_ALWAYS:
 491                 return true;
 492         case SHMEM_HUGE_WITHIN_SIZE:
 493                 index = round_up(index + 1, HPAGE_PMD_NR);
 494                 i_size = round_up(i_size_read(inode), PAGE_SIZE);
 495                 if (i_size >> PAGE_SHIFT >= index)
 496                         return true;
 497                 fallthrough;
 498         case SHMEM_HUGE_ADVISE:
 499                 if (vma && (vma->vm_flags & VM_HUGEPAGE))
 500                         return true;
 501                 fallthrough;
 502         default:
 503                 return false;
 504         }
 505 }
 506
 507 #if defined(CONFIG_SYSFS)
 508 static int shmem_parse_huge(const char *str)
 509 {
 510         if (!strcmp(str, "never"))
 511                 return SHMEM_HUGE_NEVER;
 512         if (!strcmp(str, "always"))
 513                 return SHMEM_HUGE_ALWAYS;
 514         if (!strcmp(str, "within_size"))
 515                 return SHMEM_HUGE_WITHIN_SIZE;
 516         if (!strcmp(str, "advise"))
 517                 return SHMEM_HUGE_ADVISE;
 518         if (!strcmp(str, "deny"))
 519                 return SHMEM_HUGE_DENY;
 520         if (!strcmp(str, "force"))
 521                 return SHMEM_HUGE_FORCE;
 522         return -EINVAL;
 523 }
 524 #endif
 525
 526 #if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
 527 static const char *shmem_format_huge(int huge)
 528 {
 529         switch (huge) {
 530         case SHMEM_HUGE_NEVER:
 531                 return "never";
 532         case SHMEM_HUGE_ALWAYS:
 533                 return "always";
 534         case SHMEM_HUGE_WITHIN_SIZE:
 535                 return "within_size";
 536         case SHMEM_HUGE_ADVISE:
 537                 return "advise";
 538         case SHMEM_HUGE_DENY:
 539                 return "deny";
 540         case SHMEM_HUGE_FORCE:
 541                 return "force";
 542         default:
 543                 VM_BUG_ON(1);
 544                 return "bad_val";
 545         }
 546 }
 547 #endif
 548
 549 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 550                 struct shrink_control *sc, unsigned long nr_to_split)
 551 {
 552         LIST_HEAD(list), *pos, *next;
 553         LIST_HEAD(to_remove);
 554         struct inode *inode;
 555         struct shmem_inode_info *info;
 556         struct folio *folio;
 557         unsigned long batch = sc ? sc->nr_to_scan : 128;
 558         int split = 0;
 559
 560         if (list_empty(&sbinfo->shrinklist))
 561                 return SHRINK_STOP;
 562
 563         spin_lock(&sbinfo->shrinklist_lock);
 564         list_for_each_safe(pos, next, &sbinfo->shrinklist) {
 565                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 566
 567                 /* pin the inode */
 568                 inode = igrab(&info->vfs_inode);
 569
 570                 /* inode is about to be evicted */
 571                 if (!inode) {
 572                         list_del_init(&info->shrinklist);
 573                         goto next;
 574                 }
 575
 576                 /* Check if there's anything to gain */
 577                 if (round_up(inode->i_size, PAGE_SIZE) ==
 578                                 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
 579                         list_move(&info->shrinklist, &to_remove);
 580                         goto next;
 581                 }
 582
 583                 list_move(&info->shrinklist, &list);
 584 next:
 585                 sbinfo->shrinklist_len--;
 586                 if (!--batch)
 587                         break;
 588         }
 589         spin_unlock(&sbinfo->shrinklist_lock);
 590
 591         list_for_each_safe(pos, next, &to_remove) {
 592                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 593                 inode = &info->vfs_inode;
 594                 list_del_init(&info->shrinklist);
 595                 iput(inode);
 596         }
 597
 598         list_for_each_safe(pos, next, &list) {
 599                 int ret;
 600                 pgoff_t index;
 601
 602                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 603                 inode = &info->vfs_inode;
 604
 605                 if (nr_to_split && split >= nr_to_split)
 606                         goto move_back;
 607
 608                 index = (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT;
 609                 folio = filemap_get_folio(inode->i_mapping, index);
 610                 if (!folio)
 611                         goto drop;
 612
 613                 /* No huge page at the end of the file: nothing to split */
 614                 if (!folio_test_large(folio)) {
 615                         folio_put(folio);
 616                         goto drop;
 617                 }
 618
 619                 /*
 620                  * Move the inode on the list back to shrinklist if we failed
 621                  * to lock the page at this time.
 622                  *
 623                  * Waiting for the lock may lead to deadlock in the
 624                  * reclaim path.
 625                  */
 626                 if (!folio_trylock(folio)) {
 627                         folio_put(folio);
 628                         goto move_back;
 629                 }
 630
 631                 ret = split_huge_page(&folio->page);
 632                 folio_unlock(folio);
 633                 folio_put(folio);
 634
 635                 /* If split failed move the inode on the list back to shrinklist */
 636                 if (ret)
 637                         goto move_back;
 638
 639                 split++;
 640 drop:
 641                 list_del_init(&info->shrinklist);
 642                 goto put;
 643 move_back:
 644                 /*
 645                  * Make sure the inode is either on the global list or deleted
 646                  * from any local list before iput() since it could be deleted
 647                  * in another thread once we put the inode (then the local list
 648                  * is corrupted).
 649                  */
 650                 spin_lock(&sbinfo->shrinklist_lock);
 651                 list_move(&info->shrinklist, &sbinfo->shrinklist);
 652                 sbinfo->shrinklist_len++;
 653                 spin_unlock(&sbinfo->shrinklist_lock);
 654 put:
 655                 iput(inode);
 656         }
 657
 658         return split;
 659 }
 660
 661 static long shmem_unused_huge_scan(struct super_block *sb,
 662                 struct shrink_control *sc)
 663 {
 664         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 665
 666         if (!READ_ONCE(sbinfo->shrinklist_len))
 667                 return SHRINK_STOP;
 668
 669         return shmem_unused_huge_shrink(sbinfo, sc, 0);
 670 }
 671
 672 static long shmem_unused_huge_count(struct super_block *sb,
 673                 struct shrink_control *sc)
 674 {
 675         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 676         return READ_ONCE(sbinfo->shrinklist_len);
 677 }
 678 #else /* !CONFIG_TRANSPARENT_HUGEPAGE */
 679
 680 #define shmem_huge SHMEM_HUGE_DENY
 681
 682 bool shmem_is_huge(struct vm_area_struct *vma,
 683                    struct inode *inode, pgoff_t index)
 684 {
 685         return false;
 686 }
 687
 688 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 689                 struct shrink_control *sc, unsigned long nr_to_split)
 690 {
 691         return 0;
 692 }
 693 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 694
 695 /*
 696  * Like add_to_page_cache_locked, but error if expected item has gone.
 697  */
 698 static int shmem_add_to_page_cache(struct folio *folio,
 699                                    struct address_space *mapping,
 700                                    pgoff_t index, void *expected, gfp_t gfp,
 701                                    struct mm_struct *charge_mm)
 702 {
 703         XA_STATE_ORDER(xas, &mapping->i_pages, index, folio_order(folio));
 704         long nr = folio_nr_pages(folio);
 705         int error;
 706
 707         VM_BUG_ON_FOLIO(index != round_down(index, nr), folio);
 708         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 709         VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio);
 710         VM_BUG_ON(expected && folio_test_large(folio));
 711
 712         folio_ref_add(folio, nr);
 713         folio->mapping = mapping;
 714         folio->index = index;
 715
 716         if (!folio_test_swapcache(folio)) {
 717                 error = mem_cgroup_charge(folio, charge_mm, gfp);
 718                 if (error) {
 719                         if (folio_test_pmd_mappable(folio)) {
 720                                 count_vm_event(THP_FILE_FALLBACK);
 721                                 count_vm_event(THP_FILE_FALLBACK_CHARGE);
 722                         }
 723                         goto error;
 724                 }
 725         }
 726         folio_throttle_swaprate(folio, gfp);
 727
 728         do {
 729                 xas_lock_irq(&xas);
 730                 if (expected != xas_find_conflict(&xas)) {
 731                         xas_set_err(&xas, -EEXIST);
 732                         goto unlock;
 733                 }
 734                 if (expected && xas_find_conflict(&xas)) {
 735                         xas_set_err(&xas, -EEXIST);
 736                         goto unlock;
 737                 }
 738                 xas_store(&xas, folio);
 739                 if (xas_error(&xas))
 740                         goto unlock;
 741                 if (folio_test_pmd_mappable(folio)) {
 742                         count_vm_event(THP_FILE_ALLOC);
 743                         __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr);
 744                 }
 745                 mapping->nrpages += nr;
 746                 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
 747                 __lruvec_stat_mod_folio(folio, NR_SHMEM, nr);
 748 unlock:
 749                 xas_unlock_irq(&xas);
 750         } while (xas_nomem(&xas, gfp));
 751
 752         if (xas_error(&xas)) {
 753                 error = xas_error(&xas);
 754                 goto error;
 755         }
 756
 757         return 0;
 758 error:
 759         folio->mapping = NULL;
 760         folio_ref_sub(folio, nr);
 761         return error;
 762 }
 763
 764 /*
 765  * Like delete_from_page_cache, but substitutes swap for page.
 766  */
 767 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 768 {
 769         struct address_space *mapping = page->mapping;
 770         int error;
 771
 772         VM_BUG_ON_PAGE(PageCompound(page), page);
 773
 774         xa_lock_irq(&mapping->i_pages);
 775         error = shmem_replace_entry(mapping, page->index, page, radswap);
 776         page->mapping = NULL;
 777         mapping->nrpages--;
 778         __dec_lruvec_page_state(page, NR_FILE_PAGES);
 779         __dec_lruvec_page_state(page, NR_SHMEM);
 780         xa_unlock_irq(&mapping->i_pages);
 781         put_page(page);
 782         BUG_ON(error);
 783 }
 784
 785 /*
 786  * Remove swap entry from page cache, free the swap and its page cache.
 787  */
 788 static int shmem_free_swap(struct address_space *mapping,
 789                            pgoff_t index, void *radswap)
 790 {
 791         void *old;
 792
 793         old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
 794         if (old != radswap)
 795                 return -ENOENT;
 796         free_swap_and_cache(radix_to_swp_entry(radswap));
 797         return 0;
 798 }
 799
 800 /*
 801  * Determine (in bytes) how many of the shmem object's pages mapped by the
 802  * given offsets are swapped out.
 803  *
 804  * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
 805  * as long as the inode doesn't go away and racy results are not a problem.
 806  */
 807 unsigned long shmem_partial_swap_usage(struct address_space *mapping,
 808                                                 pgoff_t start, pgoff_t end)
 809 {
 810         XA_STATE(xas, &mapping->i_pages, start);
 811         struct page *page;
 812         unsigned long swapped = 0;
 813
 814         rcu_read_lock();
 815         xas_for_each(&xas, page, end - 1) {
 816                 if (xas_retry(&xas, page))
 817                         continue;
 818                 if (xa_is_value(page))
 819                         swapped++;
 820
 821                 if (need_resched()) {
 822                         xas_pause(&xas);
 823                         cond_resched_rcu();
 824                 }
 825         }
 826
 827         rcu_read_unlock();
 828
 829         return swapped << PAGE_SHIFT;
 830 }
 831
 832 /*
 833  * Determine (in bytes) how many of the shmem object's pages mapped by the
 834  * given vma is swapped out.
 835  *
 836  * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
 837  * as long as the inode doesn't go away and racy results are not a problem.
 838  */
 839 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
 840 {
 841         struct inode *inode = file_inode(vma->vm_file);
 842         struct shmem_inode_info *info = SHMEM_I(inode);
 843         struct address_space *mapping = inode->i_mapping;
 844         unsigned long swapped;
 845
 846         /* Be careful as we don't hold info->lock */
 847         swapped = READ_ONCE(info->swapped);
 848
 849         /*
 850          * The easier cases are when the shmem object has nothing in swap, or
 851          * the vma maps it whole. Then we can simply use the stats that we
 852          * already track.
 853          */
 854         if (!swapped)
 855                 return 0;
 856
 857         if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
 858                 return swapped << PAGE_SHIFT;
 859
 860         /* Here comes the more involved part */
 861         return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
 862                                         vma->vm_pgoff + vma_pages(vma));
 863 }
 864
 865 /*
 866  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 867  */
 868 void shmem_unlock_mapping(struct address_space *mapping)
 869 {
 870         struct pagevec pvec;
 871         pgoff_t index = 0;
 872
 873         pagevec_init(&pvec);
 874         /*
 875          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 876          */
 877         while (!mapping_unevictable(mapping)) {
 878                 if (!pagevec_lookup(&pvec, mapping, &index))
 879                         break;
 880                 check_move_unevictable_pages(&pvec);
 881                 pagevec_release(&pvec);
 882                 cond_resched();
 883         }
 884 }
 885
 886 static struct folio *shmem_get_partial_folio(struct inode *inode, pgoff_t index)
 887 {
 888         struct folio *folio;
 889         struct page *page;
 890
 891         /*
 892          * At first avoid shmem_getpage(,,,SGP_READ): that fails
 893          * beyond i_size, and reports fallocated pages as holes.
 894          */
 895         folio = __filemap_get_folio(inode->i_mapping, index,
 896                                         FGP_ENTRY | FGP_LOCK, 0);
 897         if (!xa_is_value(folio))
 898                 return folio;
 899         /*
 900          * But read a page back from swap if any of it is within i_size
 901          * (although in some cases this is just a waste of time).
 902          */
 903         page = NULL;
 904         shmem_getpage(inode, index, &page, SGP_READ);
 905         return page ? page_folio(page) : NULL;
 906 }
 907
 908 /*
 909  * Remove range of pages and swap entries from page cache, and free them.
 910  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 911  */
 912 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 913                                                                  bool unfalloc)
 914 {
 915         struct address_space *mapping = inode->i_mapping;
 916         struct shmem_inode_info *info = SHMEM_I(inode);
 917         pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
 918         pgoff_t end = (lend + 1) >> PAGE_SHIFT;
 919         struct folio_batch fbatch;
 920         pgoff_t indices[PAGEVEC_SIZE];
 921         struct folio *folio;
 922         bool same_folio;
 923         long nr_swaps_freed = 0;
 924         pgoff_t index;
 925         int i;
 926
 927         if (lend == -1)
 928                 end = -1;       /* unsigned, so actually very big */
 929
 930         if (info->fallocend > start && info->fallocend <= end && !unfalloc)
 931                 info->fallocend = start;
 932
 933         folio_batch_init(&fbatch);
 934         index = start;
 935         while (index < end && find_lock_entries(mapping, index, end - 1,
 936                         &fbatch, indices)) {
 937                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 938                         folio = fbatch.folios[i];
 939
 940                         index = indices[i];
 941
 942                         if (xa_is_value(folio)) {
 943                                 if (unfalloc)
 944                                         continue;
 945                                 nr_swaps_freed += !shmem_free_swap(mapping,
 946                                                                 index, folio);
 947                                 continue;
 948                         }
 949                         index += folio_nr_pages(folio) - 1;
 950
 951                         if (!unfalloc || !folio_test_uptodate(folio))
 952                                 truncate_inode_folio(mapping, folio);
 953                         folio_unlock(folio);
 954                 }
 955                 folio_batch_remove_exceptionals(&fbatch);
 956                 folio_batch_release(&fbatch);
 957                 cond_resched();
 958                 index++;
 959         }
 960
 961         same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
 962         folio = shmem_get_partial_folio(inode, lstart >> PAGE_SHIFT);
 963         if (folio) {
 964                 same_folio = lend < folio_pos(folio) + folio_size(folio);
 965                 folio_mark_dirty(folio);
 966                 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
 967                         start = folio->index + folio_nr_pages(folio);
 968                         if (same_folio)
 969                                 end = folio->index;
 970                 }
 971                 folio_unlock(folio);
 972                 folio_put(folio);
 973                 folio = NULL;
 974         }
 975
 976         if (!same_folio)
 977                 folio = shmem_get_partial_folio(inode, lend >> PAGE_SHIFT);
 978         if (folio) {
 979                 folio_mark_dirty(folio);
 980                 if (!truncate_inode_partial_folio(folio, lstart, lend))
 981                         end = folio->index;
 982                 folio_unlock(folio);
 983                 folio_put(folio);
 984         }
 985
 986         index = start;
 987         while (index < end) {
 988                 cond_resched();
 989
 990                 if (!find_get_entries(mapping, index, end - 1, &fbatch,
 991                                 indices)) {
 992                         /* If all gone or hole-punch or unfalloc, we're done */
 993                         if (index == start || end != -1)
 994                                 break;
 995                         /* But if truncating, restart to make sure all gone */
 996                         index = start;
 997                         continue;
 998                 }
 999                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
1000                         folio = fbatch.folios[i];
1001
1002                         index = indices[i];
1003                         if (xa_is_value(folio)) {
1004                                 if (unfalloc)
1005                                         continue;
1006                                 if (shmem_free_swap(mapping, index, folio)) {
1007                                         /* Swap was replaced by page: retry */
1008                                         index--;
1009                                         break;
1010                                 }
1011                                 nr_swaps_freed++;
1012                                 continue;
1013                         }
1014
1015                         folio_lock(folio);
1016
1017                         if (!unfalloc || !folio_test_uptodate(folio)) {
1018                                 if (folio_mapping(folio) != mapping) {
1019                                         /* Page was replaced by swap: retry */
1020                                         folio_unlock(folio);
1021                                         index--;
1022                                         break;
1023                                 }
1024                                 VM_BUG_ON_FOLIO(folio_test_writeback(folio),
1025                                                 folio);
1026                                 truncate_inode_folio(mapping, folio);
1027                         }
1028                         index = folio->index + folio_nr_pages(folio) - 1;
1029                         folio_unlock(folio);
1030                 }
1031                 folio_batch_remove_exceptionals(&fbatch);
1032                 folio_batch_release(&fbatch);
1033                 index++;
1034         }
1035
1036         spin_lock_irq(&info->lock);
1037         info->swapped -= nr_swaps_freed;
1038         shmem_recalc_inode(inode);
1039         spin_unlock_irq(&info->lock);
1040 }
1041
1042 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1043 {
1044         shmem_undo_range(inode, lstart, lend, false);
1045         inode->i_ctime = inode->i_mtime = current_time(inode);
1046 }
1047 EXPORT_SYMBOL_GPL(shmem_truncate_range);
1048
1049 static int shmem_getattr(struct user_namespace *mnt_userns,
1050                          const struct path *path, struct kstat *stat,
1051                          u32 request_mask, unsigned int query_flags)
1052 {
1053         struct inode *inode = path->dentry->d_inode;
1054         struct shmem_inode_info *info = SHMEM_I(inode);
1055
1056         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1057                 spin_lock_irq(&info->lock);
1058                 shmem_recalc_inode(inode);
1059                 spin_unlock_irq(&info->lock);
1060         }
1061         generic_fillattr(&init_user_ns, inode, stat);
1062
1063         if (shmem_is_huge(NULL, inode, 0))
1064                 stat->blksize = HPAGE_PMD_SIZE;
1065
1066         if (request_mask & STATX_BTIME) {
1067                 stat->result_mask |= STATX_BTIME;
1068                 stat->btime.tv_sec = info->i_crtime.tv_sec;
1069                 stat->btime.tv_nsec = info->i_crtime.tv_nsec;
1070         }
1071
1072         return 0;
1073 }
1074
1075 static int shmem_setattr(struct user_namespace *mnt_userns,
1076                          struct dentry *dentry, struct iattr *attr)
1077 {
1078         struct inode *inode = d_inode(dentry);
1079         struct shmem_inode_info *info = SHMEM_I(inode);
1080         int error;
1081
1082         error = setattr_prepare(&init_user_ns, dentry, attr);
1083         if (error)
1084                 return error;
1085
1086         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1087                 loff_t oldsize = inode->i_size;
1088                 loff_t newsize = attr->ia_size;
1089
1090                 /* protected by i_rwsem */
1091                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1092                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1093                         return -EPERM;
1094
1095                 if (newsize != oldsize) {
1096                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1097                                         oldsize, newsize);
1098                         if (error)
1099                                 return error;
1100                         i_size_write(inode, newsize);
1101                         inode->i_ctime = inode->i_mtime = current_time(inode);
1102                 }
1103                 if (newsize <= oldsize) {
1104                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1105                         if (oldsize > holebegin)
1106                                 unmap_mapping_range(inode->i_mapping,
1107                                                         holebegin, 0, 1);
1108                         if (info->alloced)
1109                                 shmem_truncate_range(inode,
1110                                                         newsize, (loff_t)-1);
1111                         /* unmap again to remove racily COWed private pages */
1112                         if (oldsize > holebegin)
1113                                 unmap_mapping_range(inode->i_mapping,
1114                                                         holebegin, 0, 1);
1115                 }
1116         }
1117
1118         setattr_copy(&init_user_ns, inode, attr);
1119         if (attr->ia_valid & ATTR_MODE)
1120                 error = posix_acl_chmod(&init_user_ns, inode, inode->i_mode);
1121         return error;
1122 }
1123
1124 static void shmem_evict_inode(struct inode *inode)
1125 {
1126         struct shmem_inode_info *info = SHMEM_I(inode);
1127         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1128
1129         if (shmem_mapping(inode->i_mapping)) {
1130                 shmem_unacct_size(info->flags, inode->i_size);
1131                 inode->i_size = 0;
1132                 mapping_set_exiting(inode->i_mapping);
1133                 shmem_truncate_range(inode, 0, (loff_t)-1);
1134                 if (!list_empty(&info->shrinklist)) {
1135                         spin_lock(&sbinfo->shrinklist_lock);
1136                         if (!list_empty(&info->shrinklist)) {
1137                                 list_del_init(&info->shrinklist);
1138                                 sbinfo->shrinklist_len--;
1139                         }
1140                         spin_unlock(&sbinfo->shrinklist_lock);
1141                 }
1142                 while (!list_empty(&info->swaplist)) {
1143                         /* Wait while shmem_unuse() is scanning this inode... */
1144                         wait_var_event(&info->stop_eviction,
1145                                        !atomic_read(&info->stop_eviction));
1146                         mutex_lock(&shmem_swaplist_mutex);
1147                         /* ...but beware of the race if we peeked too early */
1148                         if (!atomic_read(&info->stop_eviction))
1149                                 list_del_init(&info->swaplist);
1150                         mutex_unlock(&shmem_swaplist_mutex);
1151                 }
1152         }
1153
1154         simple_xattrs_free(&info->xattrs);
1155         WARN_ON(inode->i_blocks);
1156         shmem_free_inode(inode->i_sb);
1157         clear_inode(inode);
1158 }
1159
1160 static int shmem_find_swap_entries(struct address_space *mapping,
1161                                    pgoff_t start, struct folio_batch *fbatch,
1162                                    pgoff_t *indices, unsigned int type)
1163 {
1164         XA_STATE(xas, &mapping->i_pages, start);
1165         struct folio *folio;
1166         swp_entry_t entry;
1167
1168         rcu_read_lock();
1169         xas_for_each(&xas, folio, ULONG_MAX) {
1170                 if (xas_retry(&xas, folio))
1171                         continue;
1172
1173                 if (!xa_is_value(folio))
1174                         continue;
1175
1176                 entry = radix_to_swp_entry(folio);
1177                 /*
1178                  * swapin error entries can be found in the mapping. But they're
1179                  * deliberately ignored here as we've done everything we can do.
1180                  */
1181                 if (swp_type(entry) != type)
1182                         continue;
1183
1184                 indices[folio_batch_count(fbatch)] = xas.xa_index;
1185                 if (!folio_batch_add(fbatch, folio))
1186                         break;
1187
1188                 if (need_resched()) {
1189                         xas_pause(&xas);
1190                         cond_resched_rcu();
1191                 }
1192         }
1193         rcu_read_unlock();
1194
1195         return xas.xa_index;
1196 }
1197
1198 /*
1199  * Move the swapped pages for an inode to page cache. Returns the count
1200  * of pages swapped in, or the error in case of failure.
1201  */
1202 static int shmem_unuse_swap_entries(struct inode *inode,
1203                 struct folio_batch *fbatch, pgoff_t *indices)
1204 {
1205         int i = 0;
1206         int ret = 0;
1207         int error = 0;
1208         struct address_space *mapping = inode->i_mapping;
1209
1210         for (i = 0; i < folio_batch_count(fbatch); i++) {
1211                 struct folio *folio = fbatch->folios[i];
1212
1213                 if (!xa_is_value(folio))
1214                         continue;
1215                 error = shmem_swapin_folio(inode, indices[i],
1216                                           &folio, SGP_CACHE,
1217                                           mapping_gfp_mask(mapping),
1218                                           NULL, NULL);
1219                 if (error == 0) {
1220                         folio_unlock(folio);
1221                         folio_put(folio);
1222                         ret++;
1223                 }
1224                 if (error == -ENOMEM)
1225                         break;
1226                 error = 0;
1227         }
1228         return error ? error : ret;
1229 }
1230
1231 /*
1232  * If swap found in inode, free it and move page from swapcache to filecache.
1233  */
1234 static int shmem_unuse_inode(struct inode *inode, unsigned int type)
1235 {
1236         struct address_space *mapping = inode->i_mapping;
1237         pgoff_t start = 0;
1238         struct folio_batch fbatch;
1239         pgoff_t indices[PAGEVEC_SIZE];
1240         int ret = 0;
1241
1242         do {
1243                 folio_batch_init(&fbatch);
1244                 shmem_find_swap_entries(mapping, start, &fbatch, indices, type);
1245                 if (folio_batch_count(&fbatch) == 0) {
1246                         ret = 0;
1247                         break;
1248                 }
1249
1250                 ret = shmem_unuse_swap_entries(inode, &fbatch, indices);
1251                 if (ret < 0)
1252                         break;
1253
1254                 start = indices[folio_batch_count(&fbatch) - 1];
1255         } while (true);
1256
1257         return ret;
1258 }
1259
1260 /*
1261  * Read all the shared memory data that resides in the swap
1262  * device 'type' back into memory, so the swap device can be
1263  * unused.
1264  */
1265 int shmem_unuse(unsigned int type)
1266 {
1267         struct shmem_inode_info *info, *next;
1268         int error = 0;
1269
1270         if (list_empty(&shmem_swaplist))
1271                 return 0;
1272
1273         mutex_lock(&shmem_swaplist_mutex);
1274         list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1275                 if (!info->swapped) {
1276                         list_del_init(&info->swaplist);
1277                         continue;
1278                 }
1279                 /*
1280                  * Drop the swaplist mutex while searching the inode for swap;
1281                  * but before doing so, make sure shmem_evict_inode() will not
1282                  * remove placeholder inode from swaplist, nor let it be freed
1283                  * (igrab() would protect from unlink, but not from unmount).
1284                  */
1285                 atomic_inc(&info->stop_eviction);
1286                 mutex_unlock(&shmem_swaplist_mutex);
1287
1288                 error = shmem_unuse_inode(&info->vfs_inode, type);
1289                 cond_resched();
1290
1291                 mutex_lock(&shmem_swaplist_mutex);
1292                 next = list_next_entry(info, swaplist);
1293                 if (!info->swapped)
1294                         list_del_init(&info->swaplist);
1295                 if (atomic_dec_and_test(&info->stop_eviction))
1296                         wake_up_var(&info->stop_eviction);
1297                 if (error)
1298                         break;
1299         }
1300         mutex_unlock(&shmem_swaplist_mutex);
1301
1302         return error;
1303 }
1304
1305 /*
1306  * Move the page from the page cache to the swap cache.
1307  */
1308 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1309 {
1310         struct folio *folio = page_folio(page);
1311         struct shmem_inode_info *info;
1312         struct address_space *mapping;
1313         struct inode *inode;
1314         swp_entry_t swap;
1315         pgoff_t index;
1316
1317         /*
1318          * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "always" or
1319          * "force", drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
1320          * and its shmem_writeback() needs them to be split when swapping.
1321          */
1322         if (PageTransCompound(page)) {
1323                 /* Ensure the subpages are still dirty */
1324                 SetPageDirty(page);
1325                 if (split_huge_page(page) < 0)
1326                         goto redirty;
1327                 ClearPageDirty(page);
1328         }
1329
1330         BUG_ON(!PageLocked(page));
1331         mapping = page->mapping;
1332         index = page->index;
1333         inode = mapping->host;
1334         info = SHMEM_I(inode);
1335         if (info->flags & VM_LOCKED)
1336                 goto redirty;
1337         if (!total_swap_pages)
1338                 goto redirty;
1339
1340         /*
1341          * Our capabilities prevent regular writeback or sync from ever calling
1342          * shmem_writepage; but a stacking filesystem might use ->writepage of
1343          * its underlying filesystem, in which case tmpfs should write out to
1344          * swap only in response to memory pressure, and not for the writeback
1345          * threads or sync.
1346          */
1347         if (!wbc->for_reclaim) {
1348                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1349                 goto redirty;
1350         }
1351
1352         /*
1353          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1354          * value into swapfile.c, the only way we can correctly account for a
1355          * fallocated page arriving here is now to initialize it and write it.
1356          *
1357          * That's okay for a page already fallocated earlier, but if we have
1358          * not yet completed the fallocation, then (a) we want to keep track
1359          * of this page in case we have to undo it, and (b) it may not be a
1360          * good idea to continue anyway, once we're pushing into swap.  So
1361          * reactivate the page, and let shmem_fallocate() quit when too many.
1362          */
1363         if (!PageUptodate(page)) {
1364                 if (inode->i_private) {
1365                         struct shmem_falloc *shmem_falloc;
1366                         spin_lock(&inode->i_lock);
1367                         shmem_falloc = inode->i_private;
1368                         if (shmem_falloc &&
1369                             !shmem_falloc->waitq &&
1370                             index >= shmem_falloc->start &&
1371                             index < shmem_falloc->next)
1372                                 shmem_falloc->nr_unswapped++;
1373                         else
1374                                 shmem_falloc = NULL;
1375                         spin_unlock(&inode->i_lock);
1376                         if (shmem_falloc)
1377                                 goto redirty;
1378                 }
1379                 clear_highpage(page);
1380                 flush_dcache_page(page);
1381                 SetPageUptodate(page);
1382         }
1383
1384         swap = folio_alloc_swap(folio);
1385         if (!swap.val)
1386                 goto redirty;
1387
1388         /*
1389          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1390          * if it's not already there.  Do it now before the page is
1391          * moved to swap cache, when its pagelock no longer protects
1392          * the inode from eviction.  But don't unlock the mutex until
1393          * we've incremented swapped, because shmem_unuse_inode() will
1394          * prune a !swapped inode from the swaplist under this mutex.
1395          */
1396         mutex_lock(&shmem_swaplist_mutex);
1397         if (list_empty(&info->swaplist))
1398                 list_add(&info->swaplist, &shmem_swaplist);
1399
1400         if (add_to_swap_cache(page, swap,
1401                         __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1402                         NULL) == 0) {
1403                 spin_lock_irq(&info->lock);
1404                 shmem_recalc_inode(inode);
1405                 info->swapped++;
1406                 spin_unlock_irq(&info->lock);
1407
1408                 swap_shmem_alloc(swap);
1409                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1410
1411                 mutex_unlock(&shmem_swaplist_mutex);
1412                 BUG_ON(page_mapped(page));
1413                 swap_writepage(page, wbc);
1414                 return 0;
1415         }
1416
1417         mutex_unlock(&shmem_swaplist_mutex);
1418         put_swap_page(page, swap);
1419 redirty:
1420         set_page_dirty(page);
1421         if (wbc->for_reclaim)
1422                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1423         unlock_page(page);
1424         return 0;
1425 }
1426
1427 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1428 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1429 {
1430         char buffer[64];
1431
1432         if (!mpol || mpol->mode == MPOL_DEFAULT)
1433                 return;         /* show nothing */
1434
1435         mpol_to_str(buffer, sizeof(buffer), mpol);
1436
1437         seq_printf(seq, ",mpol=%s", buffer);
1438 }
1439
1440 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1441 {
1442         struct mempolicy *mpol = NULL;
1443         if (sbinfo->mpol) {
1444                 raw_spin_lock(&sbinfo->stat_lock);      /* prevent replace/use races */
1445                 mpol = sbinfo->mpol;
1446                 mpol_get(mpol);
1447                 raw_spin_unlock(&sbinfo->stat_lock);
1448         }
1449         return mpol;
1450 }
1451 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1452 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1453 {
1454 }
1455 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1456 {
1457         return NULL;
1458 }
1459 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1460 #ifndef CONFIG_NUMA
1461 #define vm_policy vm_private_data
1462 #endif
1463
1464 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1465                 struct shmem_inode_info *info, pgoff_t index)
1466 {
1467         /* Create a pseudo vma that just contains the policy */
1468         vma_init(vma, NULL);
1469         /* Bias interleave by inode number to distribute better across nodes */
1470         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1471         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1472 }
1473
1474 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1475 {
1476         /* Drop reference taken by mpol_shared_policy_lookup() */
1477         mpol_cond_put(vma->vm_policy);
1478 }
1479
1480 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1481                         struct shmem_inode_info *info, pgoff_t index)
1482 {
1483         struct vm_area_struct pvma;
1484         struct page *page;
1485         struct vm_fault vmf = {
1486                 .vma = &pvma,
1487         };
1488
1489         shmem_pseudo_vma_init(&pvma, info, index);
1490         page = swap_cluster_readahead(swap, gfp, &vmf);
1491         shmem_pseudo_vma_destroy(&pvma);
1492
1493         return page;
1494 }
1495
1496 /*
1497  * Make sure huge_gfp is always more limited than limit_gfp.
1498  * Some of the flags set permissions, while others set limitations.
1499  */
1500 static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t limit_gfp)
1501 {
1502         gfp_t allowflags = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
1503         gfp_t denyflags = __GFP_NOWARN | __GFP_NORETRY;
1504         gfp_t zoneflags = limit_gfp & GFP_ZONEMASK;
1505         gfp_t result = huge_gfp & ~(allowflags | GFP_ZONEMASK);
1506
1507         /* Allow allocations only from the originally specified zones. */
1508         result |= zoneflags;
1509
1510         /*
1511          * Minimize the result gfp by taking the union with the deny flags,
1512          * and the intersection of the allow flags.
1513          */
1514         result |= (limit_gfp & denyflags);
1515         result |= (huge_gfp & limit_gfp) & allowflags;
1516
1517         return result;
1518 }
1519
1520 static struct folio *shmem_alloc_hugefolio(gfp_t gfp,
1521                 struct shmem_inode_info *info, pgoff_t index)
1522 {
1523         struct vm_area_struct pvma;
1524         struct address_space *mapping = info->vfs_inode.i_mapping;
1525         pgoff_t hindex;
1526         struct folio *folio;
1527
1528         hindex = round_down(index, HPAGE_PMD_NR);
1529         if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1530                                                                 XA_PRESENT))
1531                 return NULL;
1532
1533         shmem_pseudo_vma_init(&pvma, info, hindex);
1534         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, &pvma, 0, true);
1535         shmem_pseudo_vma_destroy(&pvma);
1536         if (!folio)
1537                 count_vm_event(THP_FILE_FALLBACK);
1538         return folio;
1539 }
1540
1541 static struct folio *shmem_alloc_folio(gfp_t gfp,
1542                         struct shmem_inode_info *info, pgoff_t index)
1543 {
1544         struct vm_area_struct pvma;
1545         struct folio *folio;
1546
1547         shmem_pseudo_vma_init(&pvma, info, index);
1548         folio = vma_alloc_folio(gfp, 0, &pvma, 0, false);
1549         shmem_pseudo_vma_destroy(&pvma);
1550
1551         return folio;
1552 }
1553
1554 static struct page *shmem_alloc_page(gfp_t gfp,
1555                         struct shmem_inode_info *info, pgoff_t index)
1556 {
1557         return &shmem_alloc_folio(gfp, info, index)->page;
1558 }
1559
1560 static struct folio *shmem_alloc_and_acct_folio(gfp_t gfp, struct inode *inode,
1561                 pgoff_t index, bool huge)
1562 {
1563         struct shmem_inode_info *info = SHMEM_I(inode);
1564         struct folio *folio;
1565         int nr;
1566         int err = -ENOSPC;
1567
1568         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1569                 huge = false;
1570         nr = huge ? HPAGE_PMD_NR : 1;
1571
1572         if (!shmem_inode_acct_block(inode, nr))
1573                 goto failed;
1574
1575         if (huge)
1576                 folio = shmem_alloc_hugefolio(gfp, info, index);
1577         else
1578                 folio = shmem_alloc_folio(gfp, info, index);
1579         if (folio) {
1580                 __folio_set_locked(folio);
1581                 __folio_set_swapbacked(folio);
1582                 return folio;
1583         }
1584
1585         err = -ENOMEM;
1586         shmem_inode_unacct_blocks(inode, nr);
1587 failed:
1588         return ERR_PTR(err);
1589 }
1590
1591 /*
1592  * When a page is moved from swapcache to shmem filecache (either by the
1593  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1594  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1595  * ignorance of the mapping it belongs to.  If that mapping has special
1596  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1597  * we may need to copy to a suitable page before moving to filecache.
1598  *
1599  * In a future release, this may well be extended to respect cpuset and
1600  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1601  * but for now it is a simple matter of zone.
1602  */
1603 static bool shmem_should_replace_folio(struct folio *folio, gfp_t gfp)
1604 {
1605         return folio_zonenum(folio) > gfp_zone(gfp);
1606 }
1607
1608 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1609                                 struct shmem_inode_info *info, pgoff_t index)
1610 {
1611         struct page *oldpage, *newpage;
1612         struct folio *old, *new;
1613         struct address_space *swap_mapping;
1614         swp_entry_t entry;
1615         pgoff_t swap_index;
1616         int error;
1617
1618         oldpage = *pagep;
1619         entry.val = page_private(oldpage);
1620         swap_index = swp_offset(entry);
1621         swap_mapping = page_mapping(oldpage);
1622
1623         /*
1624          * We have arrived here because our zones are constrained, so don't
1625          * limit chance of success by further cpuset and node constraints.
1626          */
1627         gfp &= ~GFP_CONSTRAINT_MASK;
1628         newpage = shmem_alloc_page(gfp, info, index);
1629         if (!newpage)
1630                 return -ENOMEM;
1631
1632         get_page(newpage);
1633         copy_highpage(newpage, oldpage);
1634         flush_dcache_page(newpage);
1635
1636         __SetPageLocked(newpage);
1637         __SetPageSwapBacked(newpage);
1638         SetPageUptodate(newpage);
1639         set_page_private(newpage, entry.val);
1640         SetPageSwapCache(newpage);
1641
1642         /*
1643          * Our caller will very soon move newpage out of swapcache, but it's
1644          * a nice clean interface for us to replace oldpage by newpage there.
1645          */
1646         xa_lock_irq(&swap_mapping->i_pages);
1647         error = shmem_replace_entry(swap_mapping, swap_index, oldpage, newpage);
1648         if (!error) {
1649                 old = page_folio(oldpage);
1650                 new = page_folio(newpage);
1651                 mem_cgroup_migrate(old, new);
1652                 __inc_lruvec_page_state(newpage, NR_FILE_PAGES);
1653                 __dec_lruvec_page_state(oldpage, NR_FILE_PAGES);
1654         }
1655         xa_unlock_irq(&swap_mapping->i_pages);
1656
1657         if (unlikely(error)) {
1658                 /*
1659                  * Is this possible?  I think not, now that our callers check
1660                  * both PageSwapCache and page_private after getting page lock;
1661                  * but be defensive.  Reverse old to newpage for clear and free.
1662                  */
1663                 oldpage = newpage;
1664         } else {
1665                 lru_cache_add(newpage);
1666                 *pagep = newpage;
1667         }
1668
1669         ClearPageSwapCache(oldpage);
1670         set_page_private(oldpage, 0);
1671
1672         unlock_page(oldpage);
1673         put_page(oldpage);
1674         put_page(oldpage);
1675         return error;
1676 }
1677
1678 static void shmem_set_folio_swapin_error(struct inode *inode, pgoff_t index,
1679                                          struct folio *folio, swp_entry_t swap)
1680 {
1681         struct address_space *mapping = inode->i_mapping;
1682         struct shmem_inode_info *info = SHMEM_I(inode);
1683         swp_entry_t swapin_error;
1684         void *old;
1685
1686         swapin_error = make_swapin_error_entry(&folio->page);
1687         old = xa_cmpxchg_irq(&mapping->i_pages, index,
1688                              swp_to_radix_entry(swap),
1689                              swp_to_radix_entry(swapin_error), 0);
1690         if (old != swp_to_radix_entry(swap))
1691                 return;
1692
1693         folio_wait_writeback(folio);
1694         delete_from_swap_cache(&folio->page);
1695         spin_lock_irq(&info->lock);
1696         /*
1697          * Don't treat swapin error folio as alloced. Otherwise inode->i_blocks won't
1698          * be 0 when inode is released and thus trigger WARN_ON(inode->i_blocks) in
1699          * shmem_evict_inode.
1700          */
1701         info->alloced--;
1702         info->swapped--;
1703         shmem_recalc_inode(inode);
1704         spin_unlock_irq(&info->lock);
1705         swap_free(swap);
1706 }
1707
1708 /*
1709  * Swap in the page pointed to by *pagep.
1710  * Caller has to make sure that *pagep contains a valid swapped page.
1711  * Returns 0 and the page in pagep if success. On failure, returns the
1712  * error code and NULL in *pagep.
1713  */
1714 static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
1715                              struct folio **foliop, enum sgp_type sgp,
1716                              gfp_t gfp, struct vm_area_struct *vma,
1717                              vm_fault_t *fault_type)
1718 {
1719         struct address_space *mapping = inode->i_mapping;
1720         struct shmem_inode_info *info = SHMEM_I(inode);
1721         struct mm_struct *charge_mm = vma ? vma->vm_mm : NULL;
1722         struct page *page;
1723         struct folio *folio = NULL;
1724         swp_entry_t swap;
1725         int error;
1726
1727         VM_BUG_ON(!*foliop || !xa_is_value(*foliop));
1728         swap = radix_to_swp_entry(*foliop);
1729         *foliop = NULL;
1730
1731         if (is_swapin_error_entry(swap))
1732                 return -EIO;
1733
1734         /* Look it up and read it in.. */
1735         page = lookup_swap_cache(swap, NULL, 0);
1736         if (!page) {
1737                 /* Or update major stats only when swapin succeeds?? */
1738                 if (fault_type) {
1739                         *fault_type |= VM_FAULT_MAJOR;
1740                         count_vm_event(PGMAJFAULT);
1741                         count_memcg_event_mm(charge_mm, PGMAJFAULT);
1742                 }
1743                 /* Here we actually start the io */
1744                 page = shmem_swapin(swap, gfp, info, index);
1745                 if (!page) {
1746                         error = -ENOMEM;
1747                         goto failed;
1748                 }
1749         }
1750         folio = page_folio(page);
1751
1752         /* We have to do this with page locked to prevent races */
1753         folio_lock(folio);
1754         if (!folio_test_swapcache(folio) ||
1755             folio_swap_entry(folio).val != swap.val ||
1756             !shmem_confirm_swap(mapping, index, swap)) {
1757                 error = -EEXIST;
1758                 goto unlock;
1759         }
1760         if (!folio_test_uptodate(folio)) {
1761                 error = -EIO;
1762                 goto failed;
1763         }
1764         folio_wait_writeback(folio);
1765
1766         /*
1767          * Some architectures may have to restore extra metadata to the
1768          * folio after reading from swap.
1769          */
1770         arch_swap_restore(swap, folio);
1771
1772         if (shmem_should_replace_folio(folio, gfp)) {
1773                 error = shmem_replace_page(&page, gfp, info, index);
1774                 if (error)
1775                         goto failed;
1776         }
1777
1778         error = shmem_add_to_page_cache(folio, mapping, index,
1779                                         swp_to_radix_entry(swap), gfp,
1780                                         charge_mm);
1781         if (error)
1782                 goto failed;
1783
1784         spin_lock_irq(&info->lock);
1785         info->swapped--;
1786         shmem_recalc_inode(inode);
1787         spin_unlock_irq(&info->lock);
1788
1789         if (sgp == SGP_WRITE)
1790                 folio_mark_accessed(folio);
1791
1792         delete_from_swap_cache(&folio->page);
1793         folio_mark_dirty(folio);
1794         swap_free(swap);
1795
1796         *foliop = folio;
1797         return 0;
1798 failed:
1799         if (!shmem_confirm_swap(mapping, index, swap))
1800                 error = -EEXIST;
1801         if (error == -EIO)
1802                 shmem_set_folio_swapin_error(inode, index, folio, swap);
1803 unlock:
1804         if (folio) {
1805                 folio_unlock(folio);
1806                 folio_put(folio);
1807         }
1808
1809         return error;
1810 }
1811
1812 /*
1813  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1814  *
1815  * If we allocate a new one we do not mark it dirty. That's up to the
1816  * vm. If we swap it in we mark it dirty since we also free the swap
1817  * entry since a page cannot live in both the swap and page cache.
1818  *
1819  * vma, vmf, and fault_type are only supplied by shmem_fault:
1820  * otherwise they are NULL.
1821  */
1822 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1823         struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1824         struct vm_area_struct *vma, struct vm_fault *vmf,
1825                         vm_fault_t *fault_type)
1826 {
1827         struct address_space *mapping = inode->i_mapping;
1828         struct shmem_inode_info *info = SHMEM_I(inode);
1829         struct shmem_sb_info *sbinfo;
1830         struct mm_struct *charge_mm;
1831         struct folio *folio;
1832         pgoff_t hindex = index;
1833         gfp_t huge_gfp;
1834         int error;
1835         int once = 0;
1836         int alloced = 0;
1837
1838         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1839                 return -EFBIG;
1840 repeat:
1841         if (sgp <= SGP_CACHE &&
1842             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1843                 return -EINVAL;
1844         }
1845
1846         sbinfo = SHMEM_SB(inode->i_sb);
1847         charge_mm = vma ? vma->vm_mm : NULL;
1848
1849         folio = __filemap_get_folio(mapping, index, FGP_ENTRY | FGP_LOCK, 0);
1850         if (folio && vma && userfaultfd_minor(vma)) {
1851                 if (!xa_is_value(folio)) {
1852                         folio_unlock(folio);
1853                         folio_put(folio);
1854                 }
1855                 *fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
1856                 return 0;
1857         }
1858
1859         if (xa_is_value(folio)) {
1860                 error = shmem_swapin_folio(inode, index, &folio,
1861                                           sgp, gfp, vma, fault_type);
1862                 if (error == -EEXIST)
1863                         goto repeat;
1864
1865                 *pagep = &folio->page;
1866                 return error;
1867         }
1868
1869         if (folio) {
1870                 hindex = folio->index;
1871                 if (sgp == SGP_WRITE)
1872                         folio_mark_accessed(folio);
1873                 if (folio_test_uptodate(folio))
1874                         goto out;
1875                 /* fallocated page */
1876                 if (sgp != SGP_READ)
1877                         goto clear;
1878                 folio_unlock(folio);
1879                 folio_put(folio);
1880         }
1881
1882         /*
1883          * SGP_READ: succeed on hole, with NULL page, letting caller zero.
1884          * SGP_NOALLOC: fail on hole, with NULL page, letting caller fail.
1885          */
1886         *pagep = NULL;
1887         if (sgp == SGP_READ)
1888                 return 0;
1889         if (sgp == SGP_NOALLOC)
1890                 return -ENOENT;
1891
1892         /*
1893          * Fast cache lookup and swap lookup did not find it: allocate.
1894          */
1895
1896         if (vma && userfaultfd_missing(vma)) {
1897                 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1898                 return 0;
1899         }
1900
1901         if (!shmem_is_huge(vma, inode, index))
1902                 goto alloc_nohuge;
1903
1904         huge_gfp = vma_thp_gfp_mask(vma);
1905         huge_gfp = limit_gfp_mask(huge_gfp, gfp);
1906         folio = shmem_alloc_and_acct_folio(huge_gfp, inode, index, true);
1907         if (IS_ERR(folio)) {
1908 alloc_nohuge:
1909                 folio = shmem_alloc_and_acct_folio(gfp, inode, index, false);
1910         }
1911         if (IS_ERR(folio)) {
1912                 int retry = 5;
1913
1914                 error = PTR_ERR(folio);
1915                 folio = NULL;
1916                 if (error != -ENOSPC)
1917                         goto unlock;
1918                 /*
1919                  * Try to reclaim some space by splitting a huge page
1920                  * beyond i_size on the filesystem.
1921                  */
1922                 while (retry--) {
1923                         int ret;
1924
1925                         ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1926                         if (ret == SHRINK_STOP)
1927                                 break;
1928                         if (ret)
1929                                 goto alloc_nohuge;
1930                 }
1931                 goto unlock;
1932         }
1933
1934         hindex = round_down(index, folio_nr_pages(folio));
1935
1936         if (sgp == SGP_WRITE)
1937                 __folio_set_referenced(folio);
1938
1939         error = shmem_add_to_page_cache(folio, mapping, hindex,
1940                                         NULL, gfp & GFP_RECLAIM_MASK,
1941                                         charge_mm);
1942         if (error)
1943                 goto unacct;
1944         folio_add_lru(folio);
1945
1946         spin_lock_irq(&info->lock);
1947         info->alloced += folio_nr_pages(folio);
1948         inode->i_blocks += (blkcnt_t)BLOCKS_PER_PAGE << folio_order(folio);
1949         shmem_recalc_inode(inode);
1950         spin_unlock_irq(&info->lock);
1951         alloced = true;
1952
1953         if (folio_test_pmd_mappable(folio) &&
1954             DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1955                         hindex + HPAGE_PMD_NR - 1) {
1956                 /*
1957                  * Part of the huge page is beyond i_size: subject
1958                  * to shrink under memory pressure.
1959                  */
1960                 spin_lock(&sbinfo->shrinklist_lock);
1961                 /*
1962                  * _careful to defend against unlocked access to
1963                  * ->shrink_list in shmem_unused_huge_shrink()
1964                  */
1965                 if (list_empty_careful(&info->shrinklist)) {
1966                         list_add_tail(&info->shrinklist,
1967                                       &sbinfo->shrinklist);
1968                         sbinfo->shrinklist_len++;
1969                 }
1970                 spin_unlock(&sbinfo->shrinklist_lock);
1971         }
1972
1973         /*
1974          * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1975          */
1976         if (sgp == SGP_FALLOC)
1977                 sgp = SGP_WRITE;
1978 clear:
1979         /*
1980          * Let SGP_WRITE caller clear ends if write does not fill page;
1981          * but SGP_FALLOC on a page fallocated earlier must initialize
1982          * it now, lest undo on failure cancel our earlier guarantee.
1983          */
1984         if (sgp != SGP_WRITE && !folio_test_uptodate(folio)) {
1985                 long i, n = folio_nr_pages(folio);
1986
1987                 for (i = 0; i < n; i++)
1988                         clear_highpage(folio_page(folio, i));
1989                 flush_dcache_folio(folio);
1990                 folio_mark_uptodate(folio);
1991         }
1992
1993         /* Perhaps the file has been truncated since we checked */
1994         if (sgp <= SGP_CACHE &&
1995             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1996                 if (alloced) {
1997                         folio_clear_dirty(folio);
1998                         filemap_remove_folio(folio);
1999                         spin_lock_irq(&info->lock);
2000                         shmem_recalc_inode(inode);
2001                         spin_unlock_irq(&info->lock);
2002                 }
2003                 error = -EINVAL;
2004                 goto unlock;
2005         }
2006 out:
2007         *pagep = folio_page(folio, index - hindex);
2008         return 0;
2009
2010         /*
2011          * Error recovery.
2012          */
2013 unacct:
2014         shmem_inode_unacct_blocks(inode, folio_nr_pages(folio));
2015
2016         if (folio_test_large(folio)) {
2017                 folio_unlock(folio);
2018                 folio_put(folio);
2019                 goto alloc_nohuge;
2020         }
2021 unlock:
2022         if (folio) {
2023                 folio_unlock(folio);
2024                 folio_put(folio);
2025         }
2026         if (error == -ENOSPC && !once++) {
2027                 spin_lock_irq(&info->lock);
2028                 shmem_recalc_inode(inode);
2029                 spin_unlock_irq(&info->lock);
2030                 goto repeat;
2031         }
2032         if (error == -EEXIST)
2033                 goto repeat;
2034         return error;
2035 }
2036
2037 /*
2038  * This is like autoremove_wake_function, but it removes the wait queue
2039  * entry unconditionally - even if something else had already woken the
2040  * target.
2041  */
2042 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2043 {
2044         int ret = default_wake_function(wait, mode, sync, key);
2045         list_del_init(&wait->entry);
2046         return ret;
2047 }
2048
2049 static vm_fault_t shmem_fault(struct vm_fault *vmf)
2050 {
2051         struct vm_area_struct *vma = vmf->vma;
2052         struct inode *inode = file_inode(vma->vm_file);
2053         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2054         int err;
2055         vm_fault_t ret = VM_FAULT_LOCKED;
2056
2057         /*
2058          * Trinity finds that probing a hole which tmpfs is punching can
2059          * prevent the hole-punch from ever completing: which in turn
2060          * locks writers out with its hold on i_rwsem.  So refrain from
2061          * faulting pages into the hole while it's being punched.  Although
2062          * shmem_undo_range() does remove the additions, it may be unable to
2063          * keep up, as each new page needs its own unmap_mapping_range() call,
2064          * and the i_mmap tree grows ever slower to scan if new vmas are added.
2065          *
2066          * It does not matter if we sometimes reach this check just before the
2067          * hole-punch begins, so that one fault then races with the punch:
2068          * we just need to make racing faults a rare case.
2069          *
2070          * The implementation below would be much simpler if we just used a
2071          * standard mutex or completion: but we cannot take i_rwsem in fault,
2072          * and bloating every shmem inode for this unlikely case would be sad.
2073          */
2074         if (unlikely(inode->i_private)) {
2075                 struct shmem_falloc *shmem_falloc;
2076
2077                 spin_lock(&inode->i_lock);
2078                 shmem_falloc = inode->i_private;
2079                 if (shmem_falloc &&
2080                     shmem_falloc->waitq &&
2081                     vmf->pgoff >= shmem_falloc->start &&
2082                     vmf->pgoff < shmem_falloc->next) {
2083                         struct file *fpin;
2084                         wait_queue_head_t *shmem_falloc_waitq;
2085                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2086
2087                         ret = VM_FAULT_NOPAGE;
2088                         fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2089                         if (fpin)
2090                                 ret = VM_FAULT_RETRY;
2091
2092                         shmem_falloc_waitq = shmem_falloc->waitq;
2093                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2094                                         TASK_UNINTERRUPTIBLE);
2095                         spin_unlock(&inode->i_lock);
2096                         schedule();
2097
2098                         /*
2099                          * shmem_falloc_waitq points into the shmem_fallocate()
2100                          * stack of the hole-punching task: shmem_falloc_waitq
2101                          * is usually invalid by the time we reach here, but
2102                          * finish_wait() does not dereference it in that case;
2103                          * though i_lock needed lest racing with wake_up_all().
2104                          */
2105                         spin_lock(&inode->i_lock);
2106                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2107                         spin_unlock(&inode->i_lock);
2108
2109                         if (fpin)
2110                                 fput(fpin);
2111                         return ret;
2112                 }
2113                 spin_unlock(&inode->i_lock);
2114         }
2115
2116         err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, SGP_CACHE,
2117                                   gfp, vma, vmf, &ret);
2118         if (err)
2119                 return vmf_error(err);
2120         return ret;
2121 }
2122
2123 unsigned long shmem_get_unmapped_area(struct file *file,
2124                                       unsigned long uaddr, unsigned long len,
2125                                       unsigned long pgoff, unsigned long flags)
2126 {
2127         unsigned long (*get_area)(struct file *,
2128                 unsigned long, unsigned long, unsigned long, unsigned long);
2129         unsigned long addr;
2130         unsigned long offset;
2131         unsigned long inflated_len;
2132         unsigned long inflated_addr;
2133         unsigned long inflated_offset;
2134
2135         if (len > TASK_SIZE)
2136                 return -ENOMEM;
2137
2138         get_area = current->mm->get_unmapped_area;
2139         addr = get_area(file, uaddr, len, pgoff, flags);
2140
2141         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2142                 return addr;
2143         if (IS_ERR_VALUE(addr))
2144                 return addr;
2145         if (addr & ~PAGE_MASK)
2146                 return addr;
2147         if (addr > TASK_SIZE - len)
2148                 return addr;
2149
2150         if (shmem_huge == SHMEM_HUGE_DENY)
2151                 return addr;
2152         if (len < HPAGE_PMD_SIZE)
2153                 return addr;
2154         if (flags & MAP_FIXED)
2155                 return addr;
2156         /*
2157          * Our priority is to support MAP_SHARED mapped hugely;
2158          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2159          * But if caller specified an address hint and we allocated area there
2160          * successfully, respect that as before.
2161          */
2162         if (uaddr == addr)
2163                 return addr;
2164
2165         if (shmem_huge != SHMEM_HUGE_FORCE) {
2166                 struct super_block *sb;
2167
2168                 if (file) {
2169                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2170                         sb = file_inode(file)->i_sb;
2171                 } else {
2172                         /*
2173                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2174                          * for "/dev/zero", to create a shared anonymous object.
2175                          */
2176                         if (IS_ERR(shm_mnt))
2177                                 return addr;
2178                         sb = shm_mnt->mnt_sb;
2179                 }
2180                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2181                         return addr;
2182         }
2183
2184         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2185         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2186                 return addr;
2187         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2188                 return addr;
2189
2190         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2191         if (inflated_len > TASK_SIZE)
2192                 return addr;
2193         if (inflated_len < len)
2194                 return addr;
2195
2196         inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2197         if (IS_ERR_VALUE(inflated_addr))
2198                 return addr;
2199         if (inflated_addr & ~PAGE_MASK)
2200                 return addr;
2201
2202         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2203         inflated_addr += offset - inflated_offset;
2204         if (inflated_offset > offset)
2205                 inflated_addr += HPAGE_PMD_SIZE;
2206
2207         if (inflated_addr > TASK_SIZE - len)
2208                 return addr;
2209         return inflated_addr;
2210 }
2211
2212 #ifdef CONFIG_NUMA
2213 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2214 {
2215         struct inode *inode = file_inode(vma->vm_file);
2216         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2217 }
2218
2219 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2220                                           unsigned long addr)
2221 {
2222         struct inode *inode = file_inode(vma->vm_file);
2223         pgoff_t index;
2224
2225         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2226         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2227 }
2228 #endif
2229
2230 int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
2231 {
2232         struct inode *inode = file_inode(file);
2233         struct shmem_inode_info *info = SHMEM_I(inode);
2234         int retval = -ENOMEM;
2235
2236         /*
2237          * What serializes the accesses to info->flags?
2238          * ipc_lock_object() when called from shmctl_do_lock(),
2239          * no serialization needed when called from shm_destroy().
2240          */
2241         if (lock && !(info->flags & VM_LOCKED)) {
2242                 if (!user_shm_lock(inode->i_size, ucounts))
2243                         goto out_nomem;
2244                 info->flags |= VM_LOCKED;
2245                 mapping_set_unevictable(file->f_mapping);
2246         }
2247         if (!lock && (info->flags & VM_LOCKED) && ucounts) {
2248                 user_shm_unlock(inode->i_size, ucounts);
2249                 info->flags &= ~VM_LOCKED;
2250                 mapping_clear_unevictable(file->f_mapping);
2251         }
2252         retval = 0;
2253
2254 out_nomem:
2255         return retval;
2256 }
2257
2258 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2259 {
2260         struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2261         int ret;
2262
2263         ret = seal_check_future_write(info->seals, vma);
2264         if (ret)
2265                 return ret;
2266
2267         /* arm64 - allow memory tagging on RAM-based files */
2268         vma->vm_flags |= VM_MTE_ALLOWED;
2269
2270         file_accessed(file);
2271         vma->vm_ops = &shmem_vm_ops;
2272         return 0;
2273 }
2274
2275 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2276                                      umode_t mode, dev_t dev, unsigned long flags)
2277 {
2278         struct inode *inode;
2279         struct shmem_inode_info *info;
2280         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2281         ino_t ino;
2282
2283         if (shmem_reserve_inode(sb, &ino))
2284                 return NULL;
2285
2286         inode = new_inode(sb);
2287         if (inode) {
2288                 inode->i_ino = ino;
2289                 inode_init_owner(&init_user_ns, inode, dir, mode);
2290                 inode->i_blocks = 0;
2291                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2292                 inode->i_generation = prandom_u32();
2293                 info = SHMEM_I(inode);
2294                 memset(info, 0, (char *)inode - (char *)info);
2295                 spin_lock_init(&info->lock);
2296                 atomic_set(&info->stop_eviction, 0);
2297                 info->seals = F_SEAL_SEAL;
2298                 info->flags = flags & VM_NORESERVE;
2299                 info->i_crtime = inode->i_mtime;
2300                 INIT_LIST_HEAD(&info->shrinklist);
2301                 INIT_LIST_HEAD(&info->swaplist);
2302                 simple_xattrs_init(&info->xattrs);
2303                 cache_no_acl(inode);
2304                 mapping_set_large_folios(inode->i_mapping);
2305
2306                 switch (mode & S_IFMT) {
2307                 default:
2308                         inode->i_op = &shmem_special_inode_operations;
2309                         init_special_inode(inode, mode, dev);
2310                         break;
2311                 case S_IFREG:
2312                         inode->i_mapping->a_ops = &shmem_aops;
2313                         inode->i_op = &shmem_inode_operations;
2314                         inode->i_fop = &shmem_file_operations;
2315                         mpol_shared_policy_init(&info->policy,
2316                                                  shmem_get_sbmpol(sbinfo));
2317                         break;
2318                 case S_IFDIR:
2319                         inc_nlink(inode);
2320                         /* Some things misbehave if size == 0 on a directory */
2321                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2322                         inode->i_op = &shmem_dir_inode_operations;
2323                         inode->i_fop = &simple_dir_operations;
2324                         break;
2325                 case S_IFLNK:
2326                         /*
2327                          * Must not load anything in the rbtree,
2328                          * mpol_free_shared_policy will not be called.
2329                          */
2330                         mpol_shared_policy_init(&info->policy, NULL);
2331                         break;
2332                 }
2333
2334                 lockdep_annotate_inode_mutex_key(inode);
2335         } else
2336                 shmem_free_inode(sb);
2337         return inode;
2338 }
2339
2340 #ifdef CONFIG_USERFAULTFD
2341 int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2342                            pmd_t *dst_pmd,
2343                            struct vm_area_struct *dst_vma,
2344                            unsigned long dst_addr,
2345                            unsigned long src_addr,
2346                            bool zeropage, bool wp_copy,
2347                            struct page **pagep)
2348 {
2349         struct inode *inode = file_inode(dst_vma->vm_file);
2350         struct shmem_inode_info *info = SHMEM_I(inode);
2351         struct address_space *mapping = inode->i_mapping;
2352         gfp_t gfp = mapping_gfp_mask(mapping);
2353         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2354         void *page_kaddr;
2355         struct folio *folio;
2356         struct page *page;
2357         int ret;
2358         pgoff_t max_off;
2359
2360         if (!shmem_inode_acct_block(inode, 1)) {
2361                 /*
2362                  * We may have got a page, returned -ENOENT triggering a retry,
2363                  * and now we find ourselves with -ENOMEM. Release the page, to
2364                  * avoid a BUG_ON in our caller.
2365                  */
2366                 if (unlikely(*pagep)) {
2367                         put_page(*pagep);
2368                         *pagep = NULL;
2369                 }
2370                 return -ENOMEM;
2371         }
2372
2373         if (!*pagep) {
2374                 ret = -ENOMEM;
2375                 page = shmem_alloc_page(gfp, info, pgoff);
2376                 if (!page)
2377                         goto out_unacct_blocks;
2378
2379                 if (!zeropage) {        /* COPY */
2380                         page_kaddr = kmap_atomic(page);
2381                         ret = copy_from_user(page_kaddr,
2382                                              (const void __user *)src_addr,
2383                                              PAGE_SIZE);
2384                         kunmap_atomic(page_kaddr);
2385
2386                         /* fallback to copy_from_user outside mmap_lock */
2387                         if (unlikely(ret)) {
2388                                 *pagep = page;
2389                                 ret = -ENOENT;
2390                                 /* don't free the page */
2391                                 goto out_unacct_blocks;
2392                         }
2393
2394                         flush_dcache_page(page);
2395                 } else {                /* ZEROPAGE */
2396                         clear_user_highpage(page, dst_addr);
2397                 }
2398         } else {
2399                 page = *pagep;
2400                 *pagep = NULL;
2401         }
2402
2403         VM_BUG_ON(PageLocked(page));
2404         VM_BUG_ON(PageSwapBacked(page));
2405         __SetPageLocked(page);
2406         __SetPageSwapBacked(page);
2407         __SetPageUptodate(page);
2408
2409         ret = -EFAULT;
2410         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2411         if (unlikely(pgoff >= max_off))
2412                 goto out_release;
2413
2414         folio = page_folio(page);
2415         ret = shmem_add_to_page_cache(folio, mapping, pgoff, NULL,
2416                                       gfp & GFP_RECLAIM_MASK, dst_mm);
2417         if (ret)
2418                 goto out_release;
2419
2420         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
2421                                        page, true, wp_copy);
2422         if (ret)
2423                 goto out_delete_from_cache;
2424
2425         spin_lock_irq(&info->lock);
2426         info->alloced++;
2427         inode->i_blocks += BLOCKS_PER_PAGE;
2428         shmem_recalc_inode(inode);
2429         spin_unlock_irq(&info->lock);
2430
2431         unlock_page(page);
2432         return 0;
2433 out_delete_from_cache:
2434         delete_from_page_cache(page);
2435 out_release:
2436         unlock_page(page);
2437         put_page(page);
2438 out_unacct_blocks:
2439         shmem_inode_unacct_blocks(inode, 1);
2440         return ret;
2441 }
2442 #endif /* CONFIG_USERFAULTFD */
2443
2444 #ifdef CONFIG_TMPFS
2445 static const struct inode_operations shmem_symlink_inode_operations;
2446 static const struct inode_operations shmem_short_symlink_operations;
2447
2448 #ifdef CONFIG_TMPFS_XATTR
2449 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2450 #else
2451 #define shmem_initxattrs NULL
2452 #endif
2453
2454 static int
2455 shmem_write_begin(struct file *file, struct address_space *mapping,
2456                         loff_t pos, unsigned len,
2457                         struct page **pagep, void **fsdata)
2458 {
2459         struct inode *inode = mapping->host;
2460         struct shmem_inode_info *info = SHMEM_I(inode);
2461         pgoff_t index = pos >> PAGE_SHIFT;
2462         int ret = 0;
2463
2464         /* i_rwsem is held by caller */
2465         if (unlikely(info->seals & (F_SEAL_GROW |
2466                                    F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2467                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2468                         return -EPERM;
2469                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2470                         return -EPERM;
2471         }
2472
2473         ret = shmem_getpage(inode, index, pagep, SGP_WRITE);
2474
2475         if (ret)
2476                 return ret;
2477
2478         if (PageHWPoison(*pagep)) {
2479                 unlock_page(*pagep);
2480                 put_page(*pagep);
2481                 *pagep = NULL;
2482                 return -EIO;
2483         }
2484
2485         return 0;
2486 }
2487
2488 static int
2489 shmem_write_end(struct file *file, struct address_space *mapping,
2490                         loff_t pos, unsigned len, unsigned copied,
2491                         struct page *page, void *fsdata)
2492 {
2493         struct inode *inode = mapping->host;
2494
2495         if (pos + copied > inode->i_size)
2496                 i_size_write(inode, pos + copied);
2497
2498         if (!PageUptodate(page)) {
2499                 struct page *head = compound_head(page);
2500                 if (PageTransCompound(page)) {
2501                         int i;
2502
2503                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2504                                 if (head + i == page)
2505                                         continue;
2506                                 clear_highpage(head + i);
2507                                 flush_dcache_page(head + i);
2508                         }
2509                 }
2510                 if (copied < PAGE_SIZE) {
2511                         unsigned from = pos & (PAGE_SIZE - 1);
2512                         zero_user_segments(page, 0, from,
2513                                         from + copied, PAGE_SIZE);
2514                 }
2515                 SetPageUptodate(head);
2516         }
2517         set_page_dirty(page);
2518         unlock_page(page);
2519         put_page(page);
2520
2521         return copied;
2522 }
2523
2524 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2525 {
2526         struct file *file = iocb->ki_filp;
2527         struct inode *inode = file_inode(file);
2528         struct address_space *mapping = inode->i_mapping;
2529         pgoff_t index;
2530         unsigned long offset;
2531         int error = 0;
2532         ssize_t retval = 0;
2533         loff_t *ppos = &iocb->ki_pos;
2534
2535         index = *ppos >> PAGE_SHIFT;
2536         offset = *ppos & ~PAGE_MASK;
2537
2538         for (;;) {
2539                 struct page *page = NULL;
2540                 pgoff_t end_index;
2541                 unsigned long nr, ret;
2542                 loff_t i_size = i_size_read(inode);
2543
2544                 end_index = i_size >> PAGE_SHIFT;
2545                 if (index > end_index)
2546                         break;
2547                 if (index == end_index) {
2548                         nr = i_size & ~PAGE_MASK;
2549                         if (nr <= offset)
2550                                 break;
2551                 }
2552
2553                 error = shmem_getpage(inode, index, &page, SGP_READ);
2554                 if (error) {
2555                         if (error == -EINVAL)
2556                                 error = 0;
2557                         break;
2558                 }
2559                 if (page) {
2560                         unlock_page(page);
2561
2562                         if (PageHWPoison(page)) {
2563                                 put_page(page);
2564                                 error = -EIO;
2565                                 break;
2566                         }
2567                 }
2568
2569                 /*
2570                  * We must evaluate after, since reads (unlike writes)
2571                  * are called without i_rwsem protection against truncate
2572                  */
2573                 nr = PAGE_SIZE;
2574                 i_size = i_size_read(inode);
2575                 end_index = i_size >> PAGE_SHIFT;
2576                 if (index == end_index) {
2577                         nr = i_size & ~PAGE_MASK;
2578                         if (nr <= offset) {
2579                                 if (page)
2580                                         put_page(page);
2581                                 break;
2582                         }
2583                 }
2584                 nr -= offset;
2585
2586                 if (page) {
2587                         /*
2588                          * If users can be writing to this page using arbitrary
2589                          * virtual addresses, take care about potential aliasing
2590                          * before reading the page on the kernel side.
2591                          */
2592                         if (mapping_writably_mapped(mapping))
2593                                 flush_dcache_page(page);
2594                         /*
2595                          * Mark the page accessed if we read the beginning.
2596                          */
2597                         if (!offset)
2598                                 mark_page_accessed(page);
2599                         /*
2600                          * Ok, we have the page, and it's up-to-date, so
2601                          * now we can copy it to user space...
2602                          */
2603                         ret = copy_page_to_iter(page, offset, nr, to);
2604                         put_page(page);
2605
2606                 } else if (iter_is_iovec(to)) {
2607                         /*
2608                          * Copy to user tends to be so well optimized, but
2609                          * clear_user() not so much, that it is noticeably
2610                          * faster to copy the zero page instead of clearing.
2611                          */
2612                         ret = copy_page_to_iter(ZERO_PAGE(0), offset, nr, to);
2613                 } else {
2614                         /*
2615                          * But submitting the same page twice in a row to
2616                          * splice() - or others? - can result in confusion:
2617                          * so don't attempt that optimization on pipes etc.
2618                          */
2619                         ret = iov_iter_zero(nr, to);
2620                 }
2621
2622                 retval += ret;
2623                 offset += ret;
2624                 index += offset >> PAGE_SHIFT;
2625                 offset &= ~PAGE_MASK;
2626
2627                 if (!iov_iter_count(to))
2628                         break;
2629                 if (ret < nr) {
2630                         error = -EFAULT;
2631                         break;
2632                 }
2633                 cond_resched();
2634         }
2635
2636         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2637         file_accessed(file);
2638         return retval ? retval : error;
2639 }
2640
2641 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2642 {
2643         struct address_space *mapping = file->f_mapping;
2644         struct inode *inode = mapping->host;
2645
2646         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2647                 return generic_file_llseek_size(file, offset, whence,
2648                                         MAX_LFS_FILESIZE, i_size_read(inode));
2649         if (offset < 0)
2650                 return -ENXIO;
2651
2652         inode_lock(inode);
2653         /* We're holding i_rwsem so we can access i_size directly */
2654         offset = mapping_seek_hole_data(mapping, offset, inode->i_size, whence);
2655         if (offset >= 0)
2656                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2657         inode_unlock(inode);
2658         return offset;
2659 }
2660
2661 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2662                                                          loff_t len)
2663 {
2664         struct inode *inode = file_inode(file);
2665         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2666         struct shmem_inode_info *info = SHMEM_I(inode);
2667         struct shmem_falloc shmem_falloc;
2668         pgoff_t start, index, end, undo_fallocend;
2669         int error;
2670
2671         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2672                 return -EOPNOTSUPP;
2673
2674         inode_lock(inode);
2675
2676         if (mode & FALLOC_FL_PUNCH_HOLE) {
2677                 struct address_space *mapping = file->f_mapping;
2678                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2679                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2680                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2681
2682                 /* protected by i_rwsem */
2683                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2684                         error = -EPERM;
2685                         goto out;
2686                 }
2687
2688                 shmem_falloc.waitq = &shmem_falloc_waitq;
2689                 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2690                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2691                 spin_lock(&inode->i_lock);
2692                 inode->i_private = &shmem_falloc;
2693                 spin_unlock(&inode->i_lock);
2694
2695                 if ((u64)unmap_end > (u64)unmap_start)
2696                         unmap_mapping_range(mapping, unmap_start,
2697                                             1 + unmap_end - unmap_start, 0);
2698                 shmem_truncate_range(inode, offset, offset + len - 1);
2699                 /* No need to unmap again: hole-punching leaves COWed pages */
2700
2701                 spin_lock(&inode->i_lock);
2702                 inode->i_private = NULL;
2703                 wake_up_all(&shmem_falloc_waitq);
2704                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2705                 spin_unlock(&inode->i_lock);
2706                 error = 0;
2707                 goto out;
2708         }
2709
2710         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2711         error = inode_newsize_ok(inode, offset + len);
2712         if (error)
2713                 goto out;
2714
2715         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2716                 error = -EPERM;
2717                 goto out;
2718         }
2719
2720         start = offset >> PAGE_SHIFT;
2721         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2722         /* Try to avoid a swapstorm if len is impossible to satisfy */
2723         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2724                 error = -ENOSPC;
2725                 goto out;
2726         }
2727
2728         shmem_falloc.waitq = NULL;
2729         shmem_falloc.start = start;
2730         shmem_falloc.next  = start;
2731         shmem_falloc.nr_falloced = 0;
2732         shmem_falloc.nr_unswapped = 0;
2733         spin_lock(&inode->i_lock);
2734         inode->i_private = &shmem_falloc;
2735         spin_unlock(&inode->i_lock);
2736
2737         /*
2738          * info->fallocend is only relevant when huge pages might be
2739          * involved: to prevent split_huge_page() freeing fallocated
2740          * pages when FALLOC_FL_KEEP_SIZE committed beyond i_size.
2741          */
2742         undo_fallocend = info->fallocend;
2743         if (info->fallocend < end)
2744                 info->fallocend = end;
2745
2746         for (index = start; index < end; ) {
2747                 struct page *page;
2748
2749                 /*
2750                  * Good, the fallocate(2) manpage permits EINTR: we may have
2751                  * been interrupted because we are using up too much memory.
2752                  */
2753                 if (signal_pending(current))
2754                         error = -EINTR;
2755                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2756                         error = -ENOMEM;
2757                 else
2758                         error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2759                 if (error) {
2760                         info->fallocend = undo_fallocend;
2761                         /* Remove the !PageUptodate pages we added */
2762                         if (index > start) {
2763                                 shmem_undo_range(inode,
2764                                     (loff_t)start << PAGE_SHIFT,
2765                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2766                         }
2767                         goto undone;
2768                 }
2769
2770                 index++;
2771                 /*
2772                  * Here is a more important optimization than it appears:
2773                  * a second SGP_FALLOC on the same huge page will clear it,
2774                  * making it PageUptodate and un-undoable if we fail later.
2775                  */
2776                 if (PageTransCompound(page)) {
2777                         index = round_up(index, HPAGE_PMD_NR);
2778                         /* Beware 32-bit wraparound */
2779                         if (!index)
2780                                 index--;
2781                 }
2782
2783                 /*
2784                  * Inform shmem_writepage() how far we have reached.
2785                  * No need for lock or barrier: we have the page lock.
2786                  */
2787                 if (!PageUptodate(page))
2788                         shmem_falloc.nr_falloced += index - shmem_falloc.next;
2789                 shmem_falloc.next = index;
2790
2791                 /*
2792                  * If !PageUptodate, leave it that way so that freeable pages
2793                  * can be recognized if we need to rollback on error later.
2794                  * But set_page_dirty so that memory pressure will swap rather
2795                  * than free the pages we are allocating (and SGP_CACHE pages
2796                  * might still be clean: we now need to mark those dirty too).
2797                  */
2798                 set_page_dirty(page);
2799                 unlock_page(page);
2800                 put_page(page);
2801                 cond_resched();
2802         }
2803
2804         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2805                 i_size_write(inode, offset + len);
2806         inode->i_ctime = current_time(inode);
2807 undone:
2808         spin_lock(&inode->i_lock);
2809         inode->i_private = NULL;
2810         spin_unlock(&inode->i_lock);
2811 out:
2812         inode_unlock(inode);
2813         return error;
2814 }
2815
2816 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2817 {
2818         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2819
2820         buf->f_type = TMPFS_MAGIC;
2821         buf->f_bsize = PAGE_SIZE;
2822         buf->f_namelen = NAME_MAX;
2823         if (sbinfo->max_blocks) {
2824                 buf->f_blocks = sbinfo->max_blocks;
2825                 buf->f_bavail =
2826                 buf->f_bfree  = sbinfo->max_blocks -
2827                                 percpu_counter_sum(&sbinfo->used_blocks);
2828         }
2829         if (sbinfo->max_inodes) {
2830                 buf->f_files = sbinfo->max_inodes;
2831                 buf->f_ffree = sbinfo->free_inodes;
2832         }
2833         /* else leave those fields 0 like simple_statfs */
2834
2835         buf->f_fsid = uuid_to_fsid(dentry->d_sb->s_uuid.b);
2836
2837         return 0;
2838 }
2839
2840 /*
2841  * File creation. Allocate an inode, and we're done..
2842  */
2843 static int
2844 shmem_mknod(struct user_namespace *mnt_userns, struct inode *dir,
2845             struct dentry *dentry, umode_t mode, dev_t dev)
2846 {
2847         struct inode *inode;
2848         int error = -ENOSPC;
2849
2850         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2851         if (inode) {
2852                 error = simple_acl_create(dir, inode);
2853                 if (error)
2854                         goto out_iput;
2855                 error = security_inode_init_security(inode, dir,
2856                                                      &dentry->d_name,
2857                                                      shmem_initxattrs, NULL);
2858                 if (error && error != -EOPNOTSUPP)
2859                         goto out_iput;
2860
2861                 error = 0;
2862                 dir->i_size += BOGO_DIRENT_SIZE;
2863                 dir->i_ctime = dir->i_mtime = current_time(dir);
2864                 d_instantiate(dentry, inode);
2865                 dget(dentry); /* Extra count - pin the dentry in core */
2866         }
2867         return error;
2868 out_iput:
2869         iput(inode);
2870         return error;
2871 }
2872
2873 static int
2874 shmem_tmpfile(struct user_namespace *mnt_userns, struct inode *dir,
2875               struct dentry *dentry, umode_t mode)
2876 {
2877         struct inode *inode;
2878         int error = -ENOSPC;
2879
2880         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2881         if (inode) {
2882                 error = security_inode_init_security(inode, dir,
2883                                                      NULL,
2884                                                      shmem_initxattrs, NULL);
2885                 if (error && error != -EOPNOTSUPP)
2886                         goto out_iput;
2887                 error = simple_acl_create(dir, inode);
2888                 if (error)
2889                         goto out_iput;
2890                 d_tmpfile(dentry, inode);
2891         }
2892         return error;
2893 out_iput:
2894         iput(inode);
2895         return error;
2896 }
2897
2898 static int shmem_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
2899                        struct dentry *dentry, umode_t mode)
2900 {
2901         int error;
2902
2903         if ((error = shmem_mknod(&init_user_ns, dir, dentry,
2904                                  mode | S_IFDIR, 0)))
2905                 return error;
2906         inc_nlink(dir);
2907         return 0;
2908 }
2909
2910 static int shmem_create(struct user_namespace *mnt_userns, struct inode *dir,
2911                         struct dentry *dentry, umode_t mode, bool excl)
2912 {
2913         return shmem_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
2914 }
2915
2916 /*
2917  * Link a file..
2918  */
2919 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2920 {
2921         struct inode *inode = d_inode(old_dentry);
2922         int ret = 0;
2923
2924         /*
2925          * No ordinary (disk based) filesystem counts links as inodes;
2926          * but each new link needs a new dentry, pinning lowmem, and
2927          * tmpfs dentries cannot be pruned until they are unlinked.
2928          * But if an O_TMPFILE file is linked into the tmpfs, the
2929          * first link must skip that, to get the accounting right.
2930          */
2931         if (inode->i_nlink) {
2932                 ret = shmem_reserve_inode(inode->i_sb, NULL);
2933                 if (ret)
2934                         goto out;
2935         }
2936
2937         dir->i_size += BOGO_DIRENT_SIZE;
2938         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2939         inc_nlink(inode);
2940         ihold(inode);   /* New dentry reference */
2941         dget(dentry);           /* Extra pinning count for the created dentry */
2942         d_instantiate(dentry, inode);
2943 out:
2944         return ret;
2945 }
2946
2947 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2948 {
2949         struct inode *inode = d_inode(dentry);
2950
2951         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2952                 shmem_free_inode(inode->i_sb);
2953
2954         dir->i_size -= BOGO_DIRENT_SIZE;
2955         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2956         drop_nlink(inode);
2957         dput(dentry);   /* Undo the count from "create" - this does all the work */
2958         return 0;
2959 }
2960
2961 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2962 {
2963         if (!simple_empty(dentry))
2964                 return -ENOTEMPTY;
2965
2966         drop_nlink(d_inode(dentry));
2967         drop_nlink(dir);
2968         return shmem_unlink(dir, dentry);
2969 }
2970
2971 static int shmem_whiteout(struct user_namespace *mnt_userns,
2972                           struct inode *old_dir, struct dentry *old_dentry)
2973 {
2974         struct dentry *whiteout;
2975         int error;
2976
2977         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2978         if (!whiteout)
2979                 return -ENOMEM;
2980
2981         error = shmem_mknod(&init_user_ns, old_dir, whiteout,
2982                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2983         dput(whiteout);
2984         if (error)
2985                 return error;
2986
2987         /*
2988          * Cheat and hash the whiteout while the old dentry is still in
2989          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
2990          *
2991          * d_lookup() will consistently find one of them at this point,
2992          * not sure which one, but that isn't even important.
2993          */
2994         d_rehash(whiteout);
2995         return 0;
2996 }
2997
2998 /*
2999  * The VFS layer already does all the dentry stuff for rename,
3000  * we just have to decrement the usage count for the target if
3001  * it exists so that the VFS layer correctly free's it when it
3002  * gets overwritten.
3003  */
3004 static int shmem_rename2(struct user_namespace *mnt_userns,
3005                          struct inode *old_dir, struct dentry *old_dentry,
3006                          struct inode *new_dir, struct dentry *new_dentry,
3007                          unsigned int flags)
3008 {
3009         struct inode *inode = d_inode(old_dentry);
3010         int they_are_dirs = S_ISDIR(inode->i_mode);
3011
3012         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3013                 return -EINVAL;
3014
3015         if (flags & RENAME_EXCHANGE)
3016                 return simple_rename_exchange(old_dir, old_dentry, new_dir, new_dentry);
3017
3018         if (!simple_empty(new_dentry))
3019                 return -ENOTEMPTY;
3020
3021         if (flags & RENAME_WHITEOUT) {
3022                 int error;
3023
3024                 error = shmem_whiteout(&init_user_ns, old_dir, old_dentry);
3025                 if (error)
3026                         return error;
3027         }
3028
3029         if (d_really_is_positive(new_dentry)) {
3030                 (void) shmem_unlink(new_dir, new_dentry);
3031                 if (they_are_dirs) {
3032                         drop_nlink(d_inode(new_dentry));
3033                         drop_nlink(old_dir);
3034                 }
3035         } else if (they_are_dirs) {
3036                 drop_nlink(old_dir);
3037                 inc_nlink(new_dir);
3038         }
3039
3040         old_dir->i_size -= BOGO_DIRENT_SIZE;
3041         new_dir->i_size += BOGO_DIRENT_SIZE;
3042         old_dir->i_ctime = old_dir->i_mtime =
3043         new_dir->i_ctime = new_dir->i_mtime =
3044         inode->i_ctime = current_time(old_dir);
3045         return 0;
3046 }
3047
3048 static int shmem_symlink(struct user_namespace *mnt_userns, struct inode *dir,
3049                          struct dentry *dentry, const char *symname)
3050 {
3051         int error;
3052         int len;
3053         struct inode *inode;
3054         struct page *page;
3055
3056         len = strlen(symname) + 1;
3057         if (len > PAGE_SIZE)
3058                 return -ENAMETOOLONG;
3059
3060         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3061                                 VM_NORESERVE);
3062         if (!inode)
3063                 return -ENOSPC;
3064
3065         error = security_inode_init_security(inode, dir, &dentry->d_name,
3066                                              shmem_initxattrs, NULL);
3067         if (error && error != -EOPNOTSUPP) {
3068                 iput(inode);
3069                 return error;
3070         }
3071
3072         inode->i_size = len-1;
3073         if (len <= SHORT_SYMLINK_LEN) {
3074                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3075                 if (!inode->i_link) {
3076                         iput(inode);
3077                         return -ENOMEM;
3078                 }
3079                 inode->i_op = &shmem_short_symlink_operations;
3080         } else {
3081                 inode_nohighmem(inode);
3082                 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3083                 if (error) {
3084                         iput(inode);
3085                         return error;
3086                 }
3087                 inode->i_mapping->a_ops = &shmem_aops;
3088                 inode->i_op = &shmem_symlink_inode_operations;
3089                 memcpy(page_address(page), symname, len);
3090                 SetPageUptodate(page);
3091                 set_page_dirty(page);
3092                 unlock_page(page);
3093                 put_page(page);
3094         }
3095         dir->i_size += BOGO_DIRENT_SIZE;
3096         dir->i_ctime = dir->i_mtime = current_time(dir);
3097         d_instantiate(dentry, inode);
3098         dget(dentry);
3099         return 0;
3100 }
3101
3102 static void shmem_put_link(void *arg)
3103 {
3104         mark_page_accessed(arg);
3105         put_page(arg);
3106 }
3107
3108 static const char *shmem_get_link(struct dentry *dentry,
3109                                   struct inode *inode,
3110                                   struct delayed_call *done)
3111 {
3112         struct page *page = NULL;
3113         int error;
3114         if (!dentry) {
3115                 page = find_get_page(inode->i_mapping, 0);
3116                 if (!page)
3117                         return ERR_PTR(-ECHILD);
3118                 if (PageHWPoison(page) ||
3119                     !PageUptodate(page)) {
3120                         put_page(page);
3121                         return ERR_PTR(-ECHILD);
3122                 }
3123         } else {
3124                 error = shmem_getpage(inode, 0, &page, SGP_READ);
3125                 if (error)
3126                         return ERR_PTR(error);
3127                 if (!page)
3128                         return ERR_PTR(-ECHILD);
3129                 if (PageHWPoison(page)) {
3130                         unlock_page(page);
3131                         put_page(page);
3132                         return ERR_PTR(-ECHILD);
3133                 }
3134                 unlock_page(page);
3135         }
3136         set_delayed_call(done, shmem_put_link, page);
3137         return page_address(page);
3138 }
3139
3140 #ifdef CONFIG_TMPFS_XATTR
3141 /*
3142  * Superblocks without xattr inode operations may get some security.* xattr
3143  * support from the LSM "for free". As soon as we have any other xattrs
3144  * like ACLs, we also need to implement the security.* handlers at
3145  * filesystem level, though.
3146  */
3147
3148 /*
3149  * Callback for security_inode_init_security() for acquiring xattrs.
3150  */
3151 static int shmem_initxattrs(struct inode *inode,
3152                             const struct xattr *xattr_array,
3153                             void *fs_info)
3154 {
3155         struct shmem_inode_info *info = SHMEM_I(inode);
3156         const struct xattr *xattr;
3157         struct simple_xattr *new_xattr;
3158         size_t len;
3159
3160         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3161                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3162                 if (!new_xattr)
3163                         return -ENOMEM;
3164
3165                 len = strlen(xattr->name) + 1;
3166                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3167                                           GFP_KERNEL);
3168                 if (!new_xattr->name) {
3169                         kvfree(new_xattr);
3170                         return -ENOMEM;
3171                 }
3172
3173                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3174                        XATTR_SECURITY_PREFIX_LEN);
3175                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3176                        xattr->name, len);
3177
3178                 simple_xattr_list_add(&info->xattrs, new_xattr);
3179         }
3180
3181         return 0;
3182 }
3183
3184 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3185                                    struct dentry *unused, struct inode *inode,
3186                                    const char *name, void *buffer, size_t size)
3187 {
3188         struct shmem_inode_info *info = SHMEM_I(inode);
3189
3190         name = xattr_full_name(handler, name);
3191         return simple_xattr_get(&info->xattrs, name, buffer, size);
3192 }
3193
3194 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3195                                    struct user_namespace *mnt_userns,
3196                                    struct dentry *unused, struct inode *inode,
3197                                    const char *name, const void *value,
3198                                    size_t size, int flags)
3199 {
3200         struct shmem_inode_info *info = SHMEM_I(inode);
3201
3202         name = xattr_full_name(handler, name);
3203         return simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3204 }
3205
3206 static const struct xattr_handler shmem_security_xattr_handler = {
3207         .prefix = XATTR_SECURITY_PREFIX,
3208         .get = shmem_xattr_handler_get,
3209         .set = shmem_xattr_handler_set,
3210 };
3211
3212 static const struct xattr_handler shmem_trusted_xattr_handler = {
3213         .prefix = XATTR_TRUSTED_PREFIX,
3214         .get = shmem_xattr_handler_get,
3215         .set = shmem_xattr_handler_set,
3216 };
3217
3218 static const struct xattr_handler *shmem_xattr_handlers[] = {
3219 #ifdef CONFIG_TMPFS_POSIX_ACL
3220         &posix_acl_access_xattr_handler,
3221         &posix_acl_default_xattr_handler,
3222 #endif
3223         &shmem_security_xattr_handler,
3224         &shmem_trusted_xattr_handler,
3225         NULL
3226 };
3227
3228 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3229 {
3230         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3231         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3232 }
3233 #endif /* CONFIG_TMPFS_XATTR */
3234
3235 static const struct inode_operations shmem_short_symlink_operations = {
3236         .getattr        = shmem_getattr,
3237         .get_link       = simple_get_link,
3238 #ifdef CONFIG_TMPFS_XATTR
3239         .listxattr      = shmem_listxattr,
3240 #endif
3241 };
3242
3243 static const struct inode_operations shmem_symlink_inode_operations = {
3244         .getattr        = shmem_getattr,
3245         .get_link       = shmem_get_link,
3246 #ifdef CONFIG_TMPFS_XATTR
3247         .listxattr      = shmem_listxattr,
3248 #endif
3249 };
3250
3251 static struct dentry *shmem_get_parent(struct dentry *child)
3252 {
3253         return ERR_PTR(-ESTALE);
3254 }
3255
3256 static int shmem_match(struct inode *ino, void *vfh)
3257 {
3258         __u32 *fh = vfh;
3259         __u64 inum = fh[2];
3260         inum = (inum << 32) | fh[1];
3261         return ino->i_ino == inum && fh[0] == ino->i_generation;
3262 }
3263
3264 /* Find any alias of inode, but prefer a hashed alias */
3265 static struct dentry *shmem_find_alias(struct inode *inode)
3266 {
3267         struct dentry *alias = d_find_alias(inode);
3268
3269         return alias ?: d_find_any_alias(inode);
3270 }
3271
3272
3273 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3274                 struct fid *fid, int fh_len, int fh_type)
3275 {
3276         struct inode *inode;
3277         struct dentry *dentry = NULL;
3278         u64 inum;
3279
3280         if (fh_len < 3)
3281                 return NULL;
3282
3283         inum = fid->raw[2];
3284         inum = (inum << 32) | fid->raw[1];
3285
3286         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3287                         shmem_match, fid->raw);
3288         if (inode) {
3289                 dentry = shmem_find_alias(inode);
3290                 iput(inode);
3291         }
3292
3293         return dentry;
3294 }
3295
3296 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3297                                 struct inode *parent)
3298 {
3299         if (*len < 3) {
3300                 *len = 3;
3301                 return FILEID_INVALID;
3302         }
3303
3304         if (inode_unhashed(inode)) {
3305                 /* Unfortunately insert_inode_hash is not idempotent,
3306                  * so as we hash inodes here rather than at creation
3307                  * time, we need a lock to ensure we only try
3308                  * to do it once
3309                  */
3310                 static DEFINE_SPINLOCK(lock);
3311                 spin_lock(&lock);
3312                 if (inode_unhashed(inode))
3313                         __insert_inode_hash(inode,
3314                                             inode->i_ino + inode->i_generation);
3315                 spin_unlock(&lock);
3316         }
3317
3318         fh[0] = inode->i_generation;
3319         fh[1] = inode->i_ino;
3320         fh[2] = ((__u64)inode->i_ino) >> 32;
3321
3322         *len = 3;
3323         return 1;
3324 }
3325
3326 static const struct export_operations shmem_export_ops = {
3327         .get_parent     = shmem_get_parent,
3328         .encode_fh      = shmem_encode_fh,
3329         .fh_to_dentry   = shmem_fh_to_dentry,
3330 };
3331
3332 enum shmem_param {
3333         Opt_gid,
3334         Opt_huge,
3335         Opt_mode,
3336         Opt_mpol,
3337         Opt_nr_blocks,
3338         Opt_nr_inodes,
3339         Opt_size,
3340         Opt_uid,
3341         Opt_inode32,
3342         Opt_inode64,
3343 };
3344
3345 static const struct constant_table shmem_param_enums_huge[] = {
3346         {"never",       SHMEM_HUGE_NEVER },
3347         {"always",      SHMEM_HUGE_ALWAYS },
3348         {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3349         {"advise",      SHMEM_HUGE_ADVISE },
3350         {}
3351 };
3352
3353 const struct fs_parameter_spec shmem_fs_parameters[] = {
3354         fsparam_u32   ("gid",           Opt_gid),
3355         fsparam_enum  ("huge",          Opt_huge,  shmem_param_enums_huge),
3356         fsparam_u32oct("mode",          Opt_mode),
3357         fsparam_string("mpol",          Opt_mpol),
3358         fsparam_string("nr_blocks",     Opt_nr_blocks),
3359         fsparam_string("nr_inodes",     Opt_nr_inodes),
3360         fsparam_string("size",          Opt_size),
3361         fsparam_u32   ("uid",           Opt_uid),
3362         fsparam_flag  ("inode32",       Opt_inode32),
3363         fsparam_flag  ("inode64",       Opt_inode64),
3364         {}
3365 };
3366
3367 static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3368 {
3369         struct shmem_options *ctx = fc->fs_private;
3370         struct fs_parse_result result;
3371         unsigned long long size;
3372         char *rest;
3373         int opt;
3374
3375         opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3376         if (opt < 0)
3377                 return opt;
3378
3379         switch (opt) {
3380         case Opt_size:
3381                 size = memparse(param->string, &rest);
3382                 if (*rest == '%') {
3383                         size <<= PAGE_SHIFT;
3384                         size *= totalram_pages();
3385                         do_div(size, 100);
3386                         rest++;
3387                 }
3388                 if (*rest)
3389                         goto bad_value;
3390                 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3391                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3392                 break;
3393         case Opt_nr_blocks:
3394                 ctx->blocks = memparse(param->string, &rest);
3395                 if (*rest)
3396                         goto bad_value;
3397                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3398                 break;
3399         case Opt_nr_inodes:
3400                 ctx->inodes = memparse(param->string, &rest);
3401                 if (*rest)
3402                         goto bad_value;
3403                 ctx->seen |= SHMEM_SEEN_INODES;
3404                 break;
3405         case Opt_mode:
3406                 ctx->mode = result.uint_32 & 07777;
3407                 break;
3408         case Opt_uid:
3409                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
3410                 if (!uid_valid(ctx->uid))
3411                         goto bad_value;
3412                 break;
3413         case Opt_gid:
3414                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
3415                 if (!gid_valid(ctx->gid))
3416                         goto bad_value;
3417                 break;
3418         case Opt_huge:
3419                 ctx->huge = result.uint_32;
3420                 if (ctx->huge != SHMEM_HUGE_NEVER &&
3421                     !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3422                       has_transparent_hugepage()))
3423                         goto unsupported_parameter;
3424                 ctx->seen |= SHMEM_SEEN_HUGE;
3425                 break;
3426         case Opt_mpol:
3427                 if (IS_ENABLED(CONFIG_NUMA)) {
3428                         mpol_put(ctx->mpol);
3429                         ctx->mpol = NULL;
3430                         if (mpol_parse_str(param->string, &ctx->mpol))
3431                                 goto bad_value;
3432                         break;
3433                 }
3434                 goto unsupported_parameter;
3435         case Opt_inode32:
3436                 ctx->full_inums = false;
3437                 ctx->seen |= SHMEM_SEEN_INUMS;
3438                 break;
3439         case Opt_inode64:
3440                 if (sizeof(ino_t) < 8) {
3441                         return invalfc(fc,
3442                                        "Cannot use inode64 with <64bit inums in kernel\n");
3443                 }
3444                 ctx->full_inums = true;
3445                 ctx->seen |= SHMEM_SEEN_INUMS;
3446                 break;
3447         }
3448         return 0;
3449
3450 unsupported_parameter:
3451         return invalfc(fc, "Unsupported parameter '%s'", param->key);
3452 bad_value:
3453         return invalfc(fc, "Bad value for '%s'", param->key);
3454 }
3455
3456 static int shmem_parse_options(struct fs_context *fc, void *data)
3457 {
3458         char *options = data;
3459
3460         if (options) {
3461                 int err = security_sb_eat_lsm_opts(options, &fc->security);
3462                 if (err)
3463                         return err;
3464         }
3465
3466         while (options != NULL) {
3467                 char *this_char = options;
3468                 for (;;) {
3469                         /*
3470                          * NUL-terminate this option: unfortunately,
3471                          * mount options form a comma-separated list,
3472                          * but mpol's nodelist may also contain commas.
3473                          */
3474                         options = strchr(options, ',');
3475                         if (options == NULL)
3476                                 break;
3477                         options++;
3478                         if (!isdigit(*options)) {
3479                                 options[-1] = '\0';
3480                                 break;
3481                         }
3482                 }
3483                 if (*this_char) {
3484                         char *value = strchr(this_char, '=');
3485                         size_t len = 0;
3486                         int err;
3487
3488                         if (value) {
3489                                 *value++ = '\0';
3490                                 len = strlen(value);
3491                         }
3492                         err = vfs_parse_fs_string(fc, this_char, value, len);
3493                         if (err < 0)
3494                                 return err;
3495                 }
3496         }
3497         return 0;
3498 }
3499
3500 /*
3501  * Reconfigure a shmem filesystem.
3502  *
3503  * Note that we disallow change from limited->unlimited blocks/inodes while any
3504  * are in use; but we must separately disallow unlimited->limited, because in
3505  * that case we have no record of how much is already in use.
3506  */
3507 static int shmem_reconfigure(struct fs_context *fc)
3508 {
3509         struct shmem_options *ctx = fc->fs_private;
3510         struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3511         unsigned long inodes;
3512         struct mempolicy *mpol = NULL;
3513         const char *err;
3514
3515         raw_spin_lock(&sbinfo->stat_lock);
3516         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3517         if (ctx->blocks > S64_MAX) {
3518                 err = "Number of blocks too large";
3519                 goto out;
3520         }
3521         if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3522                 if (!sbinfo->max_blocks) {
3523                         err = "Cannot retroactively limit size";
3524                         goto out;
3525                 }
3526                 if (percpu_counter_compare(&sbinfo->used_blocks,
3527                                            ctx->blocks) > 0) {
3528                         err = "Too small a size for current use";
3529                         goto out;
3530                 }
3531         }
3532         if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3533                 if (!sbinfo->max_inodes) {
3534                         err = "Cannot retroactively limit inodes";
3535                         goto out;
3536                 }
3537                 if (ctx->inodes < inodes) {
3538                         err = "Too few inodes for current use";
3539                         goto out;
3540                 }
3541         }
3542
3543         if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3544             sbinfo->next_ino > UINT_MAX) {
3545                 err = "Current inum too high to switch to 32-bit inums";
3546                 goto out;
3547         }
3548
3549         if (ctx->seen & SHMEM_SEEN_HUGE)
3550                 sbinfo->huge = ctx->huge;
3551         if (ctx->seen & SHMEM_SEEN_INUMS)
3552                 sbinfo->full_inums = ctx->full_inums;
3553         if (ctx->seen & SHMEM_SEEN_BLOCKS)
3554                 sbinfo->max_blocks  = ctx->blocks;
3555         if (ctx->seen & SHMEM_SEEN_INODES) {
3556                 sbinfo->max_inodes  = ctx->inodes;
3557                 sbinfo->free_inodes = ctx->inodes - inodes;
3558         }
3559
3560         /*
3561          * Preserve previous mempolicy unless mpol remount option was specified.
3562          */
3563         if (ctx->mpol) {
3564                 mpol = sbinfo->mpol;
3565                 sbinfo->mpol = ctx->mpol;       /* transfers initial ref */
3566                 ctx->mpol = NULL;
3567         }
3568         raw_spin_unlock(&sbinfo->stat_lock);
3569         mpol_put(mpol);
3570         return 0;
3571 out:
3572         raw_spin_unlock(&sbinfo->stat_lock);
3573         return invalfc(fc, "%s", err);
3574 }
3575
3576 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3577 {
3578         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3579
3580         if (sbinfo->max_blocks != shmem_default_max_blocks())
3581                 seq_printf(seq, ",size=%luk",
3582                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3583         if (sbinfo->max_inodes != shmem_default_max_inodes())
3584                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3585         if (sbinfo->mode != (0777 | S_ISVTX))
3586                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3587         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3588                 seq_printf(seq, ",uid=%u",
3589                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3590         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3591                 seq_printf(seq, ",gid=%u",
3592                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3593
3594         /*
3595          * Showing inode{64,32} might be useful even if it's the system default,
3596          * since then people don't have to resort to checking both here and
3597          * /proc/config.gz to confirm 64-bit inums were successfully applied
3598          * (which may not even exist if IKCONFIG_PROC isn't enabled).
3599          *
3600          * We hide it when inode64 isn't the default and we are using 32-bit
3601          * inodes, since that probably just means the feature isn't even under
3602          * consideration.
3603          *
3604          * As such:
3605          *
3606          *                     +-----------------+-----------------+
3607          *                     | TMPFS_INODE64=y | TMPFS_INODE64=n |
3608          *  +------------------+-----------------+-----------------+
3609          *  | full_inums=true  | show            | show            |
3610          *  | full_inums=false | show            | hide            |
3611          *  +------------------+-----------------+-----------------+
3612          *
3613          */
3614         if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3615                 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3616 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3617         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3618         if (sbinfo->huge)
3619                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3620 #endif
3621         shmem_show_mpol(seq, sbinfo->mpol);
3622         return 0;
3623 }
3624
3625 #endif /* CONFIG_TMPFS */
3626
3627 static void shmem_put_super(struct super_block *sb)
3628 {
3629         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3630
3631         free_percpu(sbinfo->ino_batch);
3632         percpu_counter_destroy(&sbinfo->used_blocks);
3633         mpol_put(sbinfo->mpol);
3634         kfree(sbinfo);
3635         sb->s_fs_info = NULL;
3636 }
3637
3638 static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3639 {
3640         struct shmem_options *ctx = fc->fs_private;
3641         struct inode *inode;
3642         struct shmem_sb_info *sbinfo;
3643
3644         /* Round up to L1_CACHE_BYTES to resist false sharing */
3645         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3646                                 L1_CACHE_BYTES), GFP_KERNEL);
3647         if (!sbinfo)
3648                 return -ENOMEM;
3649
3650         sb->s_fs_info = sbinfo;
3651
3652 #ifdef CONFIG_TMPFS
3653         /*
3654          * Per default we only allow half of the physical ram per
3655          * tmpfs instance, limiting inodes to one per page of lowmem;
3656          * but the internal instance is left unlimited.
3657          */
3658         if (!(sb->s_flags & SB_KERNMOUNT)) {
3659                 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3660                         ctx->blocks = shmem_default_max_blocks();
3661                 if (!(ctx->seen & SHMEM_SEEN_INODES))
3662                         ctx->inodes = shmem_default_max_inodes();
3663                 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3664                         ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3665         } else {
3666                 sb->s_flags |= SB_NOUSER;
3667         }
3668         sb->s_export_op = &shmem_export_ops;
3669         sb->s_flags |= SB_NOSEC;
3670 #else
3671         sb->s_flags |= SB_NOUSER;
3672 #endif
3673         sbinfo->max_blocks = ctx->blocks;
3674         sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3675         if (sb->s_flags & SB_KERNMOUNT) {
3676                 sbinfo->ino_batch = alloc_percpu(ino_t);
3677                 if (!sbinfo->ino_batch)
3678                         goto failed;
3679         }
3680         sbinfo->uid = ctx->uid;
3681         sbinfo->gid = ctx->gid;
3682         sbinfo->full_inums = ctx->full_inums;
3683         sbinfo->mode = ctx->mode;
3684         sbinfo->huge = ctx->huge;
3685         sbinfo->mpol = ctx->mpol;
3686         ctx->mpol = NULL;
3687
3688         raw_spin_lock_init(&sbinfo->stat_lock);
3689         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3690                 goto failed;
3691         spin_lock_init(&sbinfo->shrinklist_lock);
3692         INIT_LIST_HEAD(&sbinfo->shrinklist);
3693
3694         sb->s_maxbytes = MAX_LFS_FILESIZE;
3695         sb->s_blocksize = PAGE_SIZE;
3696         sb->s_blocksize_bits = PAGE_SHIFT;
3697         sb->s_magic = TMPFS_MAGIC;
3698         sb->s_op = &shmem_ops;
3699         sb->s_time_gran = 1;
3700 #ifdef CONFIG_TMPFS_XATTR
3701         sb->s_xattr = shmem_xattr_handlers;
3702 #endif
3703 #ifdef CONFIG_TMPFS_POSIX_ACL
3704         sb->s_flags |= SB_POSIXACL;
3705 #endif
3706         uuid_gen(&sb->s_uuid);
3707
3708         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3709         if (!inode)
3710                 goto failed;
3711         inode->i_uid = sbinfo->uid;
3712         inode->i_gid = sbinfo->gid;
3713         sb->s_root = d_make_root(inode);
3714         if (!sb->s_root)
3715                 goto failed;
3716         return 0;
3717
3718 failed:
3719         shmem_put_super(sb);
3720         return -ENOMEM;
3721 }
3722
3723 static int shmem_get_tree(struct fs_context *fc)
3724 {
3725         return get_tree_nodev(fc, shmem_fill_super);
3726 }
3727
3728 static void shmem_free_fc(struct fs_context *fc)
3729 {
3730         struct shmem_options *ctx = fc->fs_private;
3731
3732         if (ctx) {
3733                 mpol_put(ctx->mpol);
3734                 kfree(ctx);
3735         }
3736 }
3737
3738 static const struct fs_context_operations shmem_fs_context_ops = {
3739         .free                   = shmem_free_fc,
3740         .get_tree               = shmem_get_tree,
3741 #ifdef CONFIG_TMPFS
3742         .parse_monolithic       = shmem_parse_options,
3743         .parse_param            = shmem_parse_one,
3744         .reconfigure            = shmem_reconfigure,
3745 #endif
3746 };
3747
3748 static struct kmem_cache *shmem_inode_cachep;
3749
3750 static struct inode *shmem_alloc_inode(struct super_block *sb)
3751 {
3752         struct shmem_inode_info *info;
3753         info = alloc_inode_sb(sb, shmem_inode_cachep, GFP_KERNEL);
3754         if (!info)
3755                 return NULL;
3756         return &info->vfs_inode;
3757 }
3758
3759 static void shmem_free_in_core_inode(struct inode *inode)
3760 {
3761         if (S_ISLNK(inode->i_mode))
3762                 kfree(inode->i_link);
3763         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3764 }
3765
3766 static void shmem_destroy_inode(struct inode *inode)
3767 {
3768         if (S_ISREG(inode->i_mode))
3769                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3770 }
3771
3772 static void shmem_init_inode(void *foo)
3773 {
3774         struct shmem_inode_info *info = foo;
3775         inode_init_once(&info->vfs_inode);
3776 }
3777
3778 static void shmem_init_inodecache(void)
3779 {
3780         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3781                                 sizeof(struct shmem_inode_info),
3782                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3783 }
3784
3785 static void shmem_destroy_inodecache(void)
3786 {
3787         kmem_cache_destroy(shmem_inode_cachep);
3788 }
3789
3790 /* Keep the page in page cache instead of truncating it */
3791 static int shmem_error_remove_page(struct address_space *mapping,
3792                                    struct page *page)
3793 {
3794         return 0;
3795 }
3796
3797 const struct address_space_operations shmem_aops = {
3798         .writepage      = shmem_writepage,
3799         .dirty_folio    = noop_dirty_folio,
3800 #ifdef CONFIG_TMPFS
3801         .write_begin    = shmem_write_begin,
3802         .write_end      = shmem_write_end,
3803 #endif
3804 #ifdef CONFIG_MIGRATION
3805         .migratepage    = migrate_page,
3806 #endif
3807         .error_remove_page = shmem_error_remove_page,
3808 };
3809 EXPORT_SYMBOL(shmem_aops);
3810
3811 static const struct file_operations shmem_file_operations = {
3812         .mmap           = shmem_mmap,
3813         .get_unmapped_area = shmem_get_unmapped_area,
3814 #ifdef CONFIG_TMPFS
3815         .llseek         = shmem_file_llseek,
3816         .read_iter      = shmem_file_read_iter,
3817         .write_iter     = generic_file_write_iter,
3818         .fsync          = noop_fsync,
3819         .splice_read    = generic_file_splice_read,
3820         .splice_write   = iter_file_splice_write,
3821         .fallocate      = shmem_fallocate,
3822 #endif
3823 };
3824
3825 static const struct inode_operations shmem_inode_operations = {
3826         .getattr        = shmem_getattr,
3827         .setattr        = shmem_setattr,
3828 #ifdef CONFIG_TMPFS_XATTR
3829         .listxattr      = shmem_listxattr,
3830         .set_acl        = simple_set_acl,
3831 #endif
3832 };
3833
3834 static const struct inode_operations shmem_dir_inode_operations = {
3835 #ifdef CONFIG_TMPFS
3836         .getattr        = shmem_getattr,
3837         .create         = shmem_create,
3838         .lookup         = simple_lookup,
3839         .link           = shmem_link,
3840         .unlink         = shmem_unlink,
3841         .symlink        = shmem_symlink,
3842         .mkdir          = shmem_mkdir,
3843         .rmdir          = shmem_rmdir,
3844         .mknod          = shmem_mknod,
3845         .rename         = shmem_rename2,
3846         .tmpfile        = shmem_tmpfile,
3847 #endif
3848 #ifdef CONFIG_TMPFS_XATTR
3849         .listxattr      = shmem_listxattr,
3850 #endif
3851 #ifdef CONFIG_TMPFS_POSIX_ACL
3852         .setattr        = shmem_setattr,
3853         .set_acl        = simple_set_acl,
3854 #endif
3855 };
3856
3857 static const struct inode_operations shmem_special_inode_operations = {
3858         .getattr        = shmem_getattr,
3859 #ifdef CONFIG_TMPFS_XATTR
3860         .listxattr      = shmem_listxattr,
3861 #endif
3862 #ifdef CONFIG_TMPFS_POSIX_ACL
3863         .setattr        = shmem_setattr,
3864         .set_acl        = simple_set_acl,
3865 #endif
3866 };
3867
3868 static const struct super_operations shmem_ops = {
3869         .alloc_inode    = shmem_alloc_inode,
3870         .free_inode     = shmem_free_in_core_inode,
3871         .destroy_inode  = shmem_destroy_inode,
3872 #ifdef CONFIG_TMPFS
3873         .statfs         = shmem_statfs,
3874         .show_options   = shmem_show_options,
3875 #endif
3876         .evict_inode    = shmem_evict_inode,
3877         .drop_inode     = generic_delete_inode,
3878         .put_super      = shmem_put_super,
3879 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3880         .nr_cached_objects      = shmem_unused_huge_count,
3881         .free_cached_objects    = shmem_unused_huge_scan,
3882 #endif
3883 };
3884
3885 static const struct vm_operations_struct shmem_vm_ops = {
3886         .fault          = shmem_fault,
3887         .map_pages      = filemap_map_pages,
3888 #ifdef CONFIG_NUMA
3889         .set_policy     = shmem_set_policy,
3890         .get_policy     = shmem_get_policy,
3891 #endif
3892 };
3893
3894 int shmem_init_fs_context(struct fs_context *fc)
3895 {
3896         struct shmem_options *ctx;
3897
3898         ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
3899         if (!ctx)
3900                 return -ENOMEM;
3901
3902         ctx->mode = 0777 | S_ISVTX;
3903         ctx->uid = current_fsuid();
3904         ctx->gid = current_fsgid();
3905
3906         fc->fs_private = ctx;
3907         fc->ops = &shmem_fs_context_ops;
3908         return 0;
3909 }
3910
3911 static struct file_system_type shmem_fs_type = {
3912         .owner          = THIS_MODULE,
3913         .name           = "tmpfs",
3914         .init_fs_context = shmem_init_fs_context,
3915 #ifdef CONFIG_TMPFS
3916         .parameters     = shmem_fs_parameters,
3917 #endif
3918         .kill_sb        = kill_litter_super,
3919         .fs_flags       = FS_USERNS_MOUNT,
3920 };
3921
3922 void __init shmem_init(void)
3923 {
3924         int error;
3925
3926         shmem_init_inodecache();
3927
3928         error = register_filesystem(&shmem_fs_type);
3929         if (error) {
3930                 pr_err("Could not register tmpfs\n");
3931                 goto out2;
3932         }
3933
3934         shm_mnt = kern_mount(&shmem_fs_type);
3935         if (IS_ERR(shm_mnt)) {
3936                 error = PTR_ERR(shm_mnt);
3937                 pr_err("Could not kern_mount tmpfs\n");
3938                 goto out1;
3939         }
3940
3941 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3942         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
3943                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3944         else
3945                 shmem_huge = SHMEM_HUGE_NEVER; /* just in case it was patched */
3946 #endif
3947         return;
3948
3949 out1:
3950         unregister_filesystem(&shmem_fs_type);
3951 out2:
3952         shmem_destroy_inodecache();
3953         shm_mnt = ERR_PTR(error);
3954 }
3955
3956 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
3957 static ssize_t shmem_enabled_show(struct kobject *kobj,
3958                                   struct kobj_attribute *attr, char *buf)
3959 {
3960         static const int values[] = {
3961                 SHMEM_HUGE_ALWAYS,
3962                 SHMEM_HUGE_WITHIN_SIZE,
3963                 SHMEM_HUGE_ADVISE,
3964                 SHMEM_HUGE_NEVER,
3965                 SHMEM_HUGE_DENY,
3966                 SHMEM_HUGE_FORCE,
3967         };
3968         int len = 0;
3969         int i;
3970
3971         for (i = 0; i < ARRAY_SIZE(values); i++) {
3972                 len += sysfs_emit_at(buf, len,
3973                                      shmem_huge == values[i] ? "%s[%s]" : "%s%s",
3974                                      i ? " " : "",
3975                                      shmem_format_huge(values[i]));
3976         }
3977
3978         len += sysfs_emit_at(buf, len, "\n");
3979
3980         return len;
3981 }
3982
3983 static ssize_t shmem_enabled_store(struct kobject *kobj,
3984                 struct kobj_attribute *attr, const char *buf, size_t count)
3985 {
3986         char tmp[16];
3987         int huge;
3988
3989         if (count + 1 > sizeof(tmp))
3990                 return -EINVAL;
3991         memcpy(tmp, buf, count);
3992         tmp[count] = '\0';
3993         if (count && tmp[count - 1] == '\n')
3994                 tmp[count - 1] = '\0';
3995
3996         huge = shmem_parse_huge(tmp);
3997         if (huge == -EINVAL)
3998                 return -EINVAL;
3999         if (!has_transparent_hugepage() &&
4000                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4001                 return -EINVAL;
4002
4003         shmem_huge = huge;
4004         if (shmem_huge > SHMEM_HUGE_DENY)
4005                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4006         return count;
4007 }
4008
4009 struct kobj_attribute shmem_enabled_attr = __ATTR_RW(shmem_enabled);
4010 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4011
4012 #else /* !CONFIG_SHMEM */
4013
4014 /*
4015  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4016  *
4017  * This is intended for small system where the benefits of the full
4018  * shmem code (swap-backed and resource-limited) are outweighed by
4019  * their complexity. On systems without swap this code should be
4020  * effectively equivalent, but much lighter weight.
4021  */
4022
4023 static struct file_system_type shmem_fs_type = {
4024         .name           = "tmpfs",
4025         .init_fs_context = ramfs_init_fs_context,
4026         .parameters     = ramfs_fs_parameters,
4027         .kill_sb        = kill_litter_super,
4028         .fs_flags       = FS_USERNS_MOUNT,
4029 };
4030
4031 void __init shmem_init(void)
4032 {
4033         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4034
4035         shm_mnt = kern_mount(&shmem_fs_type);
4036         BUG_ON(IS_ERR(shm_mnt));
4037 }
4038
4039 int shmem_unuse(unsigned int type)
4040 {
4041         return 0;
4042 }
4043
4044 int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
4045 {
4046         return 0;
4047 }
4048
4049 void shmem_unlock_mapping(struct address_space *mapping)
4050 {
4051 }
4052
4053 #ifdef CONFIG_MMU
4054 unsigned long shmem_get_unmapped_area(struct file *file,
4055                                       unsigned long addr, unsigned long len,
4056                                       unsigned long pgoff, unsigned long flags)
4057 {
4058         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4059 }
4060 #endif
4061
4062 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4063 {
4064         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4065 }
4066 EXPORT_SYMBOL_GPL(shmem_truncate_range);
4067
4068 #define shmem_vm_ops                            generic_file_vm_ops
4069 #define shmem_file_operations                   ramfs_file_operations
4070 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
4071 #define shmem_acct_size(flags, size)            0
4072 #define shmem_unacct_size(flags, size)          do {} while (0)
4073
4074 #endif /* CONFIG_SHMEM */
4075
4076 /* common code */
4077
4078 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4079                                        unsigned long flags, unsigned int i_flags)
4080 {
4081         struct inode *inode;
4082         struct file *res;
4083
4084         if (IS_ERR(mnt))
4085                 return ERR_CAST(mnt);
4086
4087         if (size < 0 || size > MAX_LFS_FILESIZE)
4088                 return ERR_PTR(-EINVAL);
4089
4090         if (shmem_acct_size(flags, size))
4091                 return ERR_PTR(-ENOMEM);
4092
4093         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4094                                 flags);
4095         if (unlikely(!inode)) {
4096                 shmem_unacct_size(flags, size);
4097                 return ERR_PTR(-ENOSPC);
4098         }
4099         inode->i_flags |= i_flags;
4100         inode->i_size = size;
4101         clear_nlink(inode);     /* It is unlinked */
4102         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4103         if (!IS_ERR(res))
4104                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4105                                 &shmem_file_operations);
4106         if (IS_ERR(res))
4107                 iput(inode);
4108         return res;
4109 }
4110
4111 /**
4112  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4113  *      kernel internal.  There will be NO LSM permission checks against the
4114  *      underlying inode.  So users of this interface must do LSM checks at a
4115  *      higher layer.  The users are the big_key and shm implementations.  LSM
4116  *      checks are provided at the key or shm level rather than the inode.
4117  * @name: name for dentry (to be seen in /proc/<pid>/maps
4118  * @size: size to be set for the file
4119  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4120  */
4121 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4122 {
4123         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4124 }
4125
4126 /**
4127  * shmem_file_setup - get an unlinked file living in tmpfs
4128  * @name: name for dentry (to be seen in /proc/<pid>/maps
4129  * @size: size to be set for the file
4130  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4131  */
4132 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4133 {
4134         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4135 }
4136 EXPORT_SYMBOL_GPL(shmem_file_setup);
4137
4138 /**
4139  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4140  * @mnt: the tmpfs mount where the file will be created
4141  * @name: name for dentry (to be seen in /proc/<pid>/maps
4142  * @size: size to be set for the file
4143  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4144  */
4145 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4146                                        loff_t size, unsigned long flags)
4147 {
4148         return __shmem_file_setup(mnt, name, size, flags, 0);
4149 }
4150 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4151
4152 /**
4153  * shmem_zero_setup - setup a shared anonymous mapping
4154  * @vma: the vma to be mmapped is prepared by do_mmap
4155  */
4156 int shmem_zero_setup(struct vm_area_struct *vma)
4157 {
4158         struct file *file;
4159         loff_t size = vma->vm_end - vma->vm_start;
4160
4161         /*
4162          * Cloning a new file under mmap_lock leads to a lock ordering conflict
4163          * between XFS directory reading and selinux: since this file is only
4164          * accessible to the user through its mapping, use S_PRIVATE flag to
4165          * bypass file security, in the same way as shmem_kernel_file_setup().
4166          */
4167         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4168         if (IS_ERR(file))
4169                 return PTR_ERR(file);
4170
4171         if (vma->vm_file)
4172                 fput(vma->vm_file);
4173         vma->vm_file = file;
4174         vma->vm_ops = &shmem_vm_ops;
4175
4176         return 0;
4177 }
4178
4179 /**
4180  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4181  * @mapping:    the page's address_space
4182  * @index:      the page index
4183  * @gfp:        the page allocator flags to use if allocating
4184  *
4185  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4186  * with any new page allocations done using the specified allocation flags.
4187  * But read_cache_page_gfp() uses the ->read_folio() method: which does not
4188  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4189  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4190  *
4191  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4192  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4193  */
4194 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4195                                          pgoff_t index, gfp_t gfp)
4196 {
4197 #ifdef CONFIG_SHMEM
4198         struct inode *inode = mapping->host;
4199         struct page *page;
4200         int error;
4201
4202         BUG_ON(!shmem_mapping(mapping));
4203         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4204                                   gfp, NULL, NULL, NULL);
4205         if (error)
4206                 return ERR_PTR(error);
4207
4208         unlock_page(page);
4209         if (PageHWPoison(page)) {
4210                 put_page(page);
4211                 return ERR_PTR(-EIO);
4212         }
4213
4214         return page;
4215 #else
4216         /*
4217          * The tiny !SHMEM case uses ramfs without swap
4218          */
4219         return read_cache_page_gfp(mapping, index, gfp);
4220 #endif
4221 }
4222 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);