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