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/fileattr.h>
  32 #include <linux/mm.h>
  33 #include <linux/random.h>
  34 #include <linux/sched/signal.h>
  35 #include <linux/export.h>
  36 #include <linux/swap.h>
  37 #include <linux/uio.h>
  38 #include <linux/hugetlb.h>
  39 #include <linux/fs_parser.h>
  40 #include <linux/swapfile.h>
  41 #include <linux/iversion.h>
  42 #include "swap.h"
  43
  44 static struct vfsmount *shm_mnt;
  45
  46 #ifdef CONFIG_SHMEM
  47 /*
  48  * This virtual memory filesystem is heavily based on the ramfs. It
  49  * extends ramfs by the ability to use swap and honor resource limits
  50  * which makes it a completely usable filesystem.
  51  */
  52
  53 #include <linux/xattr.h>
  54 #include <linux/exportfs.h>
  55 #include <linux/posix_acl.h>
  56 #include <linux/posix_acl_xattr.h>
  57 #include <linux/mman.h>
  58 #include <linux/string.h>
  59 #include <linux/slab.h>
  60 #include <linux/backing-dev.h>
  61 #include <linux/shmem_fs.h>
  62 #include <linux/writeback.h>
  63 #include <linux/pagevec.h>
  64 #include <linux/percpu_counter.h>
  65 #include <linux/falloc.h>
  66 #include <linux/splice.h>
  67 #include <linux/security.h>
  68 #include <linux/swapops.h>
  69 #include <linux/mempolicy.h>
  70 #include <linux/namei.h>
  71 #include <linux/ctype.h>
  72 #include <linux/migrate.h>
  73 #include <linux/highmem.h>
  74 #include <linux/seq_file.h>
  75 #include <linux/magic.h>
  76 #include <linux/syscalls.h>
  77 #include <linux/fcntl.h>
  78 #include <uapi/linux/memfd.h>
  79 #include <linux/userfaultfd_k.h>
  80 #include <linux/rmap.h>
  81 #include <linux/uuid.h>
  82
  83 #include <linux/uaccess.h>
  84
  85 #include "internal.h"
  86
  87 #define BLOCKS_PER_PAGE  (PAGE_SIZE/512)
  88 #define VM_ACCT(size)    (PAGE_ALIGN(size) >> PAGE_SHIFT)
  89
  90 /* Pretend that each entry is of this size in directory's i_size */
  91 #define BOGO_DIRENT_SIZE 20
  92
  93 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  94 #define SHORT_SYMLINK_LEN 128
  95
  96 /*
  97  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
  98  * inode->i_private (with i_rwsem making sure that it has only one user at
  99  * a time): we would prefer not to enlarge the shmem inode just for that.
 100  */
 101 struct shmem_falloc {
 102         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
 103         pgoff_t start;          /* start of range currently being fallocated */
 104         pgoff_t next;           /* the next page offset to be fallocated */
 105         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
 106         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
 107 };
 108
 109 struct shmem_options {
 110         unsigned long long blocks;
 111         unsigned long long inodes;
 112         struct mempolicy *mpol;
 113         kuid_t uid;
 114         kgid_t gid;
 115         umode_t mode;
 116         bool full_inums;
 117         int huge;
 118         int seen;
 119 #define SHMEM_SEEN_BLOCKS 1
 120 #define SHMEM_SEEN_INODES 2
 121 #define SHMEM_SEEN_HUGE 4
 122 #define SHMEM_SEEN_INUMS 8
 123 };
 124
 125 #ifdef CONFIG_TMPFS
 126 static unsigned long shmem_default_max_blocks(void)
 127 {
 128         return totalram_pages() / 2;
 129 }
 130
 131 static unsigned long shmem_default_max_inodes(void)
 132 {
 133         unsigned long nr_pages = totalram_pages();
 134
 135         return min(nr_pages - totalhigh_pages(), nr_pages / 2);
 136 }
 137 #endif
 138
 139 static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
 140                              struct folio **foliop, enum sgp_type sgp,
 141                              gfp_t gfp, struct vm_area_struct *vma,
 142                              vm_fault_t *fault_type);
 143
 144 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 145 {
 146         return sb->s_fs_info;
 147 }
 148
 149 /*
 150  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 151  * for shared memory and for shared anonymous (/dev/zero) mappings
 152  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 153  * consistent with the pre-accounting of private mappings ...
 154  */
 155 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 156 {
 157         return (flags & VM_NORESERVE) ?
 158                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 159 }
 160
 161 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 162 {
 163         if (!(flags & VM_NORESERVE))
 164                 vm_unacct_memory(VM_ACCT(size));
 165 }
 166
 167 static inline int shmem_reacct_size(unsigned long flags,
 168                 loff_t oldsize, loff_t newsize)
 169 {
 170         if (!(flags & VM_NORESERVE)) {
 171                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
 172                         return security_vm_enough_memory_mm(current->mm,
 173                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
 174                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
 175                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
 176         }
 177         return 0;
 178 }
 179
 180 /*
 181  * ... whereas tmpfs objects are accounted incrementally as
 182  * pages are allocated, in order to allow large sparse files.
 183  * shmem_get_folio reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 184  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 185  */
 186 static inline int shmem_acct_block(unsigned long flags, long pages)
 187 {
 188         if (!(flags & VM_NORESERVE))
 189                 return 0;
 190
 191         return security_vm_enough_memory_mm(current->mm,
 192                         pages * VM_ACCT(PAGE_SIZE));
 193 }
 194
 195 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 196 {
 197         if (flags & VM_NORESERVE)
 198                 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
 199 }
 200
 201 static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
 202 {
 203         struct shmem_inode_info *info = SHMEM_I(inode);
 204         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 205
 206         if (shmem_acct_block(info->flags, pages))
 207                 return false;
 208
 209         if (sbinfo->max_blocks) {
 210                 if (percpu_counter_compare(&sbinfo->used_blocks,
 211                                            sbinfo->max_blocks - pages) > 0)
 212                         goto unacct;
 213                 percpu_counter_add(&sbinfo->used_blocks, pages);
 214         }
 215
 216         return true;
 217
 218 unacct:
 219         shmem_unacct_blocks(info->flags, pages);
 220         return false;
 221 }
 222
 223 static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
 224 {
 225         struct shmem_inode_info *info = SHMEM_I(inode);
 226         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 227
 228         if (sbinfo->max_blocks)
 229                 percpu_counter_sub(&sbinfo->used_blocks, pages);
 230         shmem_unacct_blocks(info->flags, pages);
 231 }
 232
 233 static const struct super_operations shmem_ops;
 234 const struct address_space_operations shmem_aops;
 235 static const struct file_operations shmem_file_operations;
 236 static const struct inode_operations shmem_inode_operations;
 237 static const struct inode_operations shmem_dir_inode_operations;
 238 static const struct inode_operations shmem_special_inode_operations;
 239 static const struct vm_operations_struct shmem_vm_ops;
 240 static struct file_system_type shmem_fs_type;
 241
 242 bool vma_is_shmem(struct vm_area_struct *vma)
 243 {
 244         return vma->vm_ops == &shmem_vm_ops;
 245 }
 246
 247 static LIST_HEAD(shmem_swaplist);
 248 static DEFINE_MUTEX(shmem_swaplist_mutex);
 249
 250 /*
 251  * shmem_reserve_inode() performs bookkeeping to reserve a shmem inode, and
 252  * produces a novel ino for the newly allocated inode.
 253  *
 254  * It may also be called when making a hard link to permit the space needed by
 255  * each dentry. However, in that case, no new inode number is needed since that
 256  * internally draws from another pool of inode numbers (currently global
 257  * get_next_ino()). This case is indicated by passing NULL as inop.
 258  */
 259 #define SHMEM_INO_BATCH 1024
 260 static int shmem_reserve_inode(struct super_block *sb, ino_t *inop)
 261 {
 262         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 263         ino_t ino;
 264
 265         if (!(sb->s_flags & SB_KERNMOUNT)) {
 266                 raw_spin_lock(&sbinfo->stat_lock);
 267                 if (sbinfo->max_inodes) {
 268                         if (!sbinfo->free_inodes) {
 269                                 raw_spin_unlock(&sbinfo->stat_lock);
 270                                 return -ENOSPC;
 271                         }
 272                         sbinfo->free_inodes--;
 273                 }
 274                 if (inop) {
 275                         ino = sbinfo->next_ino++;
 276                         if (unlikely(is_zero_ino(ino)))
 277                                 ino = sbinfo->next_ino++;
 278                         if (unlikely(!sbinfo->full_inums &&
 279                                      ino > UINT_MAX)) {
 280                                 /*
 281                                  * Emulate get_next_ino uint wraparound for
 282                                  * compatibility
 283                                  */
 284                                 if (IS_ENABLED(CONFIG_64BIT))
 285                                         pr_warn("%s: inode number overflow on device %d, consider using inode64 mount option\n",
 286                                                 __func__, MINOR(sb->s_dev));
 287                                 sbinfo->next_ino = 1;
 288                                 ino = sbinfo->next_ino++;
 289                         }
 290                         *inop = ino;
 291                 }
 292                 raw_spin_unlock(&sbinfo->stat_lock);
 293         } else if (inop) {
 294                 /*
 295                  * __shmem_file_setup, one of our callers, is lock-free: it
 296                  * doesn't hold stat_lock in shmem_reserve_inode since
 297                  * max_inodes is always 0, and is called from potentially
 298                  * unknown contexts. As such, use a per-cpu batched allocator
 299                  * which doesn't require the per-sb stat_lock unless we are at
 300                  * the batch boundary.
 301                  *
 302                  * We don't need to worry about inode{32,64} since SB_KERNMOUNT
 303                  * shmem mounts are not exposed to userspace, so we don't need
 304                  * to worry about things like glibc compatibility.
 305                  */
 306                 ino_t *next_ino;
 307
 308                 next_ino = per_cpu_ptr(sbinfo->ino_batch, get_cpu());
 309                 ino = *next_ino;
 310                 if (unlikely(ino % SHMEM_INO_BATCH == 0)) {
 311                         raw_spin_lock(&sbinfo->stat_lock);
 312                         ino = sbinfo->next_ino;
 313                         sbinfo->next_ino += SHMEM_INO_BATCH;
 314                         raw_spin_unlock(&sbinfo->stat_lock);
 315                         if (unlikely(is_zero_ino(ino)))
 316                                 ino++;
 317                 }
 318                 *inop = ino;
 319                 *next_ino = ++ino;
 320                 put_cpu();
 321         }
 322
 323         return 0;
 324 }
 325
 326 static void shmem_free_inode(struct super_block *sb)
 327 {
 328         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 329         if (sbinfo->max_inodes) {
 330                 raw_spin_lock(&sbinfo->stat_lock);
 331                 sbinfo->free_inodes++;
 332                 raw_spin_unlock(&sbinfo->stat_lock);
 333         }
 334 }
 335
 336 /**
 337  * shmem_recalc_inode - recalculate the block usage of an inode
 338  * @inode: inode to recalc
 339  *
 340  * We have to calculate the free blocks since the mm can drop
 341  * undirtied hole pages behind our back.
 342  *
 343  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 344  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 345  *
 346  * It has to be called with the spinlock held.
 347  */
 348 static void shmem_recalc_inode(struct inode *inode)
 349 {
 350         struct shmem_inode_info *info = SHMEM_I(inode);
 351         long freed;
 352
 353         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 354         if (freed > 0) {
 355                 info->alloced -= freed;
 356                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 357                 shmem_inode_unacct_blocks(inode, freed);
 358         }
 359 }
 360
 361 bool shmem_charge(struct inode *inode, long pages)
 362 {
 363         struct shmem_inode_info *info = SHMEM_I(inode);
 364         unsigned long flags;
 365
 366         if (!shmem_inode_acct_block(inode, pages))
 367                 return false;
 368
 369         /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
 370         inode->i_mapping->nrpages += pages;
 371
 372         spin_lock_irqsave(&info->lock, flags);
 373         info->alloced += pages;
 374         inode->i_blocks += pages * BLOCKS_PER_PAGE;
 375         shmem_recalc_inode(inode);
 376         spin_unlock_irqrestore(&info->lock, flags);
 377
 378         return true;
 379 }
 380
 381 void shmem_uncharge(struct inode *inode, long pages)
 382 {
 383         struct shmem_inode_info *info = SHMEM_I(inode);
 384         unsigned long flags;
 385
 386         /* nrpages adjustment done by __filemap_remove_folio() or caller */
 387
 388         spin_lock_irqsave(&info->lock, flags);
 389         info->alloced -= pages;
 390         inode->i_blocks -= pages * BLOCKS_PER_PAGE;
 391         shmem_recalc_inode(inode);
 392         spin_unlock_irqrestore(&info->lock, flags);
 393
 394         shmem_inode_unacct_blocks(inode, pages);
 395 }
 396
 397 /*
 398  * Replace item expected in xarray by a new item, while holding xa_lock.
 399  */
 400 static int shmem_replace_entry(struct address_space *mapping,
 401                         pgoff_t index, void *expected, void *replacement)
 402 {
 403         XA_STATE(xas, &mapping->i_pages, index);
 404         void *item;
 405
 406         VM_BUG_ON(!expected);
 407         VM_BUG_ON(!replacement);
 408         item = xas_load(&xas);
 409         if (item != expected)
 410                 return -ENOENT;
 411         xas_store(&xas, replacement);
 412         return 0;
 413 }
 414
 415 /*
 416  * Sometimes, before we decide whether to proceed or to fail, we must check
 417  * that an entry was not already brought back from swap by a racing thread.
 418  *
 419  * Checking page is not enough: by the time a SwapCache page is locked, it
 420  * might be reused, and again be SwapCache, using the same swap as before.
 421  */
 422 static bool shmem_confirm_swap(struct address_space *mapping,
 423                                pgoff_t index, swp_entry_t swap)
 424 {
 425         return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
 426 }
 427
 428 /*
 429  * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
 430  *
 431  * SHMEM_HUGE_NEVER:
 432  *      disables huge pages for the mount;
 433  * SHMEM_HUGE_ALWAYS:
 434  *      enables huge pages for the mount;
 435  * SHMEM_HUGE_WITHIN_SIZE:
 436  *      only allocate huge pages if the page will be fully within i_size,
 437  *      also respect fadvise()/madvise() hints;
 438  * SHMEM_HUGE_ADVISE:
 439  *      only allocate huge pages if requested with fadvise()/madvise();
 440  */
 441
 442 #define SHMEM_HUGE_NEVER        0
 443 #define SHMEM_HUGE_ALWAYS       1
 444 #define SHMEM_HUGE_WITHIN_SIZE  2
 445 #define SHMEM_HUGE_ADVISE       3
 446
 447 /*
 448  * Special values.
 449  * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
 450  *
 451  * SHMEM_HUGE_DENY:
 452  *      disables huge on shm_mnt and all mounts, for emergency use;
 453  * SHMEM_HUGE_FORCE:
 454  *      enables huge on shm_mnt and all mounts, w/o needing option, for testing;
 455  *
 456  */
 457 #define SHMEM_HUGE_DENY         (-1)
 458 #define SHMEM_HUGE_FORCE        (-2)
 459
 460 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 461 /* ifdef here to avoid bloating shmem.o when not necessary */
 462
 463 static int shmem_huge __read_mostly = SHMEM_HUGE_NEVER;
 464
 465 bool shmem_is_huge(struct vm_area_struct *vma, struct inode *inode,
 466                    pgoff_t index, bool shmem_huge_force)
 467 {
 468         loff_t i_size;
 469
 470         if (!S_ISREG(inode->i_mode))
 471                 return false;
 472         if (vma && ((vma->vm_flags & VM_NOHUGEPAGE) ||
 473             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)))
 474                 return false;
 475         if (shmem_huge == SHMEM_HUGE_DENY)
 476                 return false;
 477         if (shmem_huge_force || shmem_huge == SHMEM_HUGE_FORCE)
 478                 return true;
 479
 480         switch (SHMEM_SB(inode->i_sb)->huge) {
 481         case SHMEM_HUGE_ALWAYS:
 482                 return true;
 483         case SHMEM_HUGE_WITHIN_SIZE:
 484                 index = round_up(index + 1, HPAGE_PMD_NR);
 485                 i_size = round_up(i_size_read(inode), PAGE_SIZE);
 486                 if (i_size >> PAGE_SHIFT >= index)
 487                         return true;
 488                 fallthrough;
 489         case SHMEM_HUGE_ADVISE:
 490                 if (vma && (vma->vm_flags & VM_HUGEPAGE))
 491                         return true;
 492                 fallthrough;
 493         default:
 494                 return false;
 495         }
 496 }
 497
 498 #if defined(CONFIG_SYSFS)
 499 static int shmem_parse_huge(const char *str)
 500 {
 501         if (!strcmp(str, "never"))
 502                 return SHMEM_HUGE_NEVER;
 503         if (!strcmp(str, "always"))
 504                 return SHMEM_HUGE_ALWAYS;
 505         if (!strcmp(str, "within_size"))
 506                 return SHMEM_HUGE_WITHIN_SIZE;
 507         if (!strcmp(str, "advise"))
 508                 return SHMEM_HUGE_ADVISE;
 509         if (!strcmp(str, "deny"))
 510                 return SHMEM_HUGE_DENY;
 511         if (!strcmp(str, "force"))
 512                 return SHMEM_HUGE_FORCE;
 513         return -EINVAL;
 514 }
 515 #endif
 516
 517 #if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
 518 static const char *shmem_format_huge(int huge)
 519 {
 520         switch (huge) {
 521         case SHMEM_HUGE_NEVER:
 522                 return "never";
 523         case SHMEM_HUGE_ALWAYS:
 524                 return "always";
 525         case SHMEM_HUGE_WITHIN_SIZE:
 526                 return "within_size";
 527         case SHMEM_HUGE_ADVISE:
 528                 return "advise";
 529         case SHMEM_HUGE_DENY:
 530                 return "deny";
 531         case SHMEM_HUGE_FORCE:
 532                 return "force";
 533         default:
 534                 VM_BUG_ON(1);
 535                 return "bad_val";
 536         }
 537 }
 538 #endif
 539
 540 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 541                 struct shrink_control *sc, unsigned long nr_to_split)
 542 {
 543         LIST_HEAD(list), *pos, *next;
 544         LIST_HEAD(to_remove);
 545         struct inode *inode;
 546         struct shmem_inode_info *info;
 547         struct folio *folio;
 548         unsigned long batch = sc ? sc->nr_to_scan : 128;
 549         int split = 0;
 550
 551         if (list_empty(&sbinfo->shrinklist))
 552                 return SHRINK_STOP;
 553
 554         spin_lock(&sbinfo->shrinklist_lock);
 555         list_for_each_safe(pos, next, &sbinfo->shrinklist) {
 556                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 557
 558                 /* pin the inode */
 559                 inode = igrab(&info->vfs_inode);
 560
 561                 /* inode is about to be evicted */
 562                 if (!inode) {
 563                         list_del_init(&info->shrinklist);
 564                         goto next;
 565                 }
 566
 567                 /* Check if there's anything to gain */
 568                 if (round_up(inode->i_size, PAGE_SIZE) ==
 569                                 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
 570                         list_move(&info->shrinklist, &to_remove);
 571                         goto next;
 572                 }
 573
 574                 list_move(&info->shrinklist, &list);
 575 next:
 576                 sbinfo->shrinklist_len--;
 577                 if (!--batch)
 578                         break;
 579         }
 580         spin_unlock(&sbinfo->shrinklist_lock);
 581
 582         list_for_each_safe(pos, next, &to_remove) {
 583                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 584                 inode = &info->vfs_inode;
 585                 list_del_init(&info->shrinklist);
 586                 iput(inode);
 587         }
 588
 589         list_for_each_safe(pos, next, &list) {
 590                 int ret;
 591                 pgoff_t index;
 592
 593                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 594                 inode = &info->vfs_inode;
 595
 596                 if (nr_to_split && split >= nr_to_split)
 597                         goto move_back;
 598
 599                 index = (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT;
 600                 folio = filemap_get_folio(inode->i_mapping, index);
 601                 if (!folio)
 602                         goto drop;
 603
 604                 /* No huge page at the end of the file: nothing to split */
 605                 if (!folio_test_large(folio)) {
 606                         folio_put(folio);
 607                         goto drop;
 608                 }
 609
 610                 /*
 611                  * Move the inode on the list back to shrinklist if we failed
 612                  * to lock the page at this time.
 613                  *
 614                  * Waiting for the lock may lead to deadlock in the
 615                  * reclaim path.
 616                  */
 617                 if (!folio_trylock(folio)) {
 618                         folio_put(folio);
 619                         goto move_back;
 620                 }
 621
 622                 ret = split_folio(folio);
 623                 folio_unlock(folio);
 624                 folio_put(folio);
 625
 626                 /* If split failed move the inode on the list back to shrinklist */
 627                 if (ret)
 628                         goto move_back;
 629
 630                 split++;
 631 drop:
 632                 list_del_init(&info->shrinklist);
 633                 goto put;
 634 move_back:
 635                 /*
 636                  * Make sure the inode is either on the global list or deleted
 637                  * from any local list before iput() since it could be deleted
 638                  * in another thread once we put the inode (then the local list
 639                  * is corrupted).
 640                  */
 641                 spin_lock(&sbinfo->shrinklist_lock);
 642                 list_move(&info->shrinklist, &sbinfo->shrinklist);
 643                 sbinfo->shrinklist_len++;
 644                 spin_unlock(&sbinfo->shrinklist_lock);
 645 put:
 646                 iput(inode);
 647         }
 648
 649         return split;
 650 }
 651
 652 static long shmem_unused_huge_scan(struct super_block *sb,
 653                 struct shrink_control *sc)
 654 {
 655         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 656
 657         if (!READ_ONCE(sbinfo->shrinklist_len))
 658                 return SHRINK_STOP;
 659
 660         return shmem_unused_huge_shrink(sbinfo, sc, 0);
 661 }
 662
 663 static long shmem_unused_huge_count(struct super_block *sb,
 664                 struct shrink_control *sc)
 665 {
 666         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 667         return READ_ONCE(sbinfo->shrinklist_len);
 668 }
 669 #else /* !CONFIG_TRANSPARENT_HUGEPAGE */
 670
 671 #define shmem_huge SHMEM_HUGE_DENY
 672
 673 bool shmem_is_huge(struct vm_area_struct *vma, struct inode *inode,
 674                    pgoff_t index, bool shmem_huge_force)
 675 {
 676         return false;
 677 }
 678
 679 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 680                 struct shrink_control *sc, unsigned long nr_to_split)
 681 {
 682         return 0;
 683 }
 684 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 685
 686 /*
 687  * Like filemap_add_folio, but error if expected item has gone.
 688  */
 689 static int shmem_add_to_page_cache(struct folio *folio,
 690                                    struct address_space *mapping,
 691                                    pgoff_t index, void *expected, gfp_t gfp,
 692                                    struct mm_struct *charge_mm)
 693 {
 694         XA_STATE_ORDER(xas, &mapping->i_pages, index, folio_order(folio));
 695         long nr = folio_nr_pages(folio);
 696         int error;
 697
 698         VM_BUG_ON_FOLIO(index != round_down(index, nr), folio);
 699         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 700         VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio);
 701         VM_BUG_ON(expected && folio_test_large(folio));
 702
 703         folio_ref_add(folio, nr);
 704         folio->mapping = mapping;
 705         folio->index = index;
 706
 707         if (!folio_test_swapcache(folio)) {
 708                 error = mem_cgroup_charge(folio, charge_mm, gfp);
 709                 if (error) {
 710                         if (folio_test_pmd_mappable(folio)) {
 711                                 count_vm_event(THP_FILE_FALLBACK);
 712                                 count_vm_event(THP_FILE_FALLBACK_CHARGE);
 713                         }
 714                         goto error;
 715                 }
 716         }
 717         folio_throttle_swaprate(folio, gfp);
 718
 719         do {
 720                 xas_lock_irq(&xas);
 721                 if (expected != xas_find_conflict(&xas)) {
 722                         xas_set_err(&xas, -EEXIST);
 723                         goto unlock;
 724                 }
 725                 if (expected && xas_find_conflict(&xas)) {
 726                         xas_set_err(&xas, -EEXIST);
 727                         goto unlock;
 728                 }
 729                 xas_store(&xas, folio);
 730                 if (xas_error(&xas))
 731                         goto unlock;
 732                 if (folio_test_pmd_mappable(folio)) {
 733                         count_vm_event(THP_FILE_ALLOC);
 734                         __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr);
 735                 }
 736                 mapping->nrpages += nr;
 737                 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
 738                 __lruvec_stat_mod_folio(folio, NR_SHMEM, nr);
 739 unlock:
 740                 xas_unlock_irq(&xas);
 741         } while (xas_nomem(&xas, gfp));
 742
 743         if (xas_error(&xas)) {
 744                 error = xas_error(&xas);
 745                 goto error;
 746         }
 747
 748         return 0;
 749 error:
 750         folio->mapping = NULL;
 751         folio_ref_sub(folio, nr);
 752         return error;
 753 }
 754
 755 /*
 756  * Like delete_from_page_cache, but substitutes swap for @folio.
 757  */
 758 static void shmem_delete_from_page_cache(struct folio *folio, void *radswap)
 759 {
 760         struct address_space *mapping = folio->mapping;
 761         long nr = folio_nr_pages(folio);
 762         int error;
 763
 764         xa_lock_irq(&mapping->i_pages);
 765         error = shmem_replace_entry(mapping, folio->index, folio, radswap);
 766         folio->mapping = NULL;
 767         mapping->nrpages -= nr;
 768         __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
 769         __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
 770         xa_unlock_irq(&mapping->i_pages);
 771         folio_put(folio);
 772         BUG_ON(error);
 773 }
 774
 775 /*
 776  * Remove swap entry from page cache, free the swap and its page cache.
 777  */
 778 static int shmem_free_swap(struct address_space *mapping,
 779                            pgoff_t index, void *radswap)
 780 {
 781         void *old;
 782
 783         old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
 784         if (old != radswap)
 785                 return -ENOENT;
 786         free_swap_and_cache(radix_to_swp_entry(radswap));
 787         return 0;
 788 }
 789
 790 /*
 791  * Determine (in bytes) how many of the shmem object's pages mapped by the
 792  * given offsets are swapped out.
 793  *
 794  * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
 795  * as long as the inode doesn't go away and racy results are not a problem.
 796  */
 797 unsigned long shmem_partial_swap_usage(struct address_space *mapping,
 798                                                 pgoff_t start, pgoff_t end)
 799 {
 800         XA_STATE(xas, &mapping->i_pages, start);
 801         struct page *page;
 802         unsigned long swapped = 0;
 803         unsigned long max = end - 1;
 804
 805         rcu_read_lock();
 806         xas_for_each(&xas, page, max) {
 807                 if (xas_retry(&xas, page))
 808                         continue;
 809                 if (xa_is_value(page))
 810                         swapped++;
 811                 if (xas.xa_index == max)
 812                         break;
 813                 if (need_resched()) {
 814                         xas_pause(&xas);
 815                         cond_resched_rcu();
 816                 }
 817         }
 818
 819         rcu_read_unlock();
 820
 821         return swapped << PAGE_SHIFT;
 822 }
 823
 824 /*
 825  * Determine (in bytes) how many of the shmem object's pages mapped by the
 826  * given vma is swapped out.
 827  *
 828  * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
 829  * as long as the inode doesn't go away and racy results are not a problem.
 830  */
 831 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
 832 {
 833         struct inode *inode = file_inode(vma->vm_file);
 834         struct shmem_inode_info *info = SHMEM_I(inode);
 835         struct address_space *mapping = inode->i_mapping;
 836         unsigned long swapped;
 837
 838         /* Be careful as we don't hold info->lock */
 839         swapped = READ_ONCE(info->swapped);
 840
 841         /*
 842          * The easier cases are when the shmem object has nothing in swap, or
 843          * the vma maps it whole. Then we can simply use the stats that we
 844          * already track.
 845          */
 846         if (!swapped)
 847                 return 0;
 848
 849         if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
 850                 return swapped << PAGE_SHIFT;
 851
 852         /* Here comes the more involved part */
 853         return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
 854                                         vma->vm_pgoff + vma_pages(vma));
 855 }
 856
 857 /*
 858  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 859  */
 860 void shmem_unlock_mapping(struct address_space *mapping)
 861 {
 862         struct folio_batch fbatch;
 863         pgoff_t index = 0;
 864
 865         folio_batch_init(&fbatch);
 866         /*
 867          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 868          */
 869         while (!mapping_unevictable(mapping) &&
 870                filemap_get_folios(mapping, &index, ~0UL, &fbatch)) {
 871                 check_move_unevictable_folios(&fbatch);
 872                 folio_batch_release(&fbatch);
 873                 cond_resched();
 874         }
 875 }
 876
 877 static struct folio *shmem_get_partial_folio(struct inode *inode, pgoff_t index)
 878 {
 879         struct folio *folio;
 880
 881         /*
 882          * At first avoid shmem_get_folio(,,,SGP_READ): that fails
 883          * beyond i_size, and reports fallocated pages as holes.
 884          */
 885         folio = __filemap_get_folio(inode->i_mapping, index,
 886                                         FGP_ENTRY | FGP_LOCK, 0);
 887         if (!xa_is_value(folio))
 888                 return folio;
 889         /*
 890          * But read a page back from swap if any of it is within i_size
 891          * (although in some cases this is just a waste of time).
 892          */
 893         folio = NULL;
 894         shmem_get_folio(inode, index, &folio, SGP_READ);
 895         return folio;
 896 }
 897
 898 /*
 899  * Remove range of pages and swap entries from page cache, and free them.
 900  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 901  */
 902 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 903                                                                  bool unfalloc)
 904 {
 905         struct address_space *mapping = inode->i_mapping;
 906         struct shmem_inode_info *info = SHMEM_I(inode);
 907         pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
 908         pgoff_t end = (lend + 1) >> PAGE_SHIFT;
 909         struct folio_batch fbatch;
 910         pgoff_t indices[PAGEVEC_SIZE];
 911         struct folio *folio;
 912         bool same_folio;
 913         long nr_swaps_freed = 0;
 914         pgoff_t index;
 915         int i;
 916
 917         if (lend == -1)
 918                 end = -1;       /* unsigned, so actually very big */
 919
 920         if (info->fallocend > start && info->fallocend <= end && !unfalloc)
 921                 info->fallocend = start;
 922
 923         folio_batch_init(&fbatch);
 924         index = start;
 925         while (index < end && find_lock_entries(mapping, index, end - 1,
 926                         &fbatch, indices)) {
 927                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 928                         folio = fbatch.folios[i];
 929
 930                         index = indices[i];
 931
 932                         if (xa_is_value(folio)) {
 933                                 if (unfalloc)
 934                                         continue;
 935                                 nr_swaps_freed += !shmem_free_swap(mapping,
 936                                                                 index, folio);
 937                                 continue;
 938                         }
 939                         index += folio_nr_pages(folio) - 1;
 940
 941                         if (!unfalloc || !folio_test_uptodate(folio))
 942                                 truncate_inode_folio(mapping, folio);
 943                         folio_unlock(folio);
 944                 }
 945                 folio_batch_remove_exceptionals(&fbatch);
 946                 folio_batch_release(&fbatch);
 947                 cond_resched();
 948                 index++;
 949         }
 950
 951         /*
 952          * When undoing a failed fallocate, we want none of the partial folio
 953          * zeroing and splitting below, but shall want to truncate the whole
 954          * folio when !uptodate indicates that it was added by this fallocate,
 955          * even when [lstart, lend] covers only a part of the folio.
 956          */
 957         if (unfalloc)
 958                 goto whole_folios;
 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 whole_folios:
 986
 987         index = start;
 988         while (index < end) {
 989                 cond_resched();
 990
 991                 if (!find_get_entries(mapping, index, end - 1, &fbatch,
 992                                 indices)) {
 993                         /* If all gone or hole-punch or unfalloc, we're done */
 994                         if (index == start || end != -1)
 995                                 break;
 996                         /* But if truncating, restart to make sure all gone */
 997                         index = start;
 998                         continue;
 999                 }
1000                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
1001                         folio = fbatch.folios[i];
1002
1003                         index = indices[i];
1004                         if (xa_is_value(folio)) {
1005                                 if (unfalloc)
1006                                         continue;
1007                                 if (shmem_free_swap(mapping, index, folio)) {
1008                                         /* Swap was replaced by page: retry */
1009                                         index--;
1010                                         break;
1011                                 }
1012                                 nr_swaps_freed++;
1013                                 continue;
1014                         }
1015
1016                         folio_lock(folio);
1017
1018                         if (!unfalloc || !folio_test_uptodate(folio)) {
1019                                 if (folio_mapping(folio) != mapping) {
1020                                         /* Page was replaced by swap: retry */
1021                                         folio_unlock(folio);
1022                                         index--;
1023                                         break;
1024                                 }
1025                                 VM_BUG_ON_FOLIO(folio_test_writeback(folio),
1026                                                 folio);
1027                                 truncate_inode_folio(mapping, folio);
1028                         }
1029                         index = folio->index + folio_nr_pages(folio) - 1;
1030                         folio_unlock(folio);
1031                 }
1032                 folio_batch_remove_exceptionals(&fbatch);
1033                 folio_batch_release(&fbatch);
1034                 index++;
1035         }
1036
1037         spin_lock_irq(&info->lock);
1038         info->swapped -= nr_swaps_freed;
1039         shmem_recalc_inode(inode);
1040         spin_unlock_irq(&info->lock);
1041 }
1042
1043 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1044 {
1045         shmem_undo_range(inode, lstart, lend, false);
1046         inode->i_ctime = inode->i_mtime = current_time(inode);
1047         inode_inc_iversion(inode);
1048 }
1049 EXPORT_SYMBOL_GPL(shmem_truncate_range);
1050
1051 static int shmem_getattr(struct user_namespace *mnt_userns,
1052                          const struct path *path, struct kstat *stat,
1053                          u32 request_mask, unsigned int query_flags)
1054 {
1055         struct inode *inode = path->dentry->d_inode;
1056         struct shmem_inode_info *info = SHMEM_I(inode);
1057
1058         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1059                 spin_lock_irq(&info->lock);
1060                 shmem_recalc_inode(inode);
1061                 spin_unlock_irq(&info->lock);
1062         }
1063         if (info->fsflags & FS_APPEND_FL)
1064                 stat->attributes |= STATX_ATTR_APPEND;
1065         if (info->fsflags & FS_IMMUTABLE_FL)
1066                 stat->attributes |= STATX_ATTR_IMMUTABLE;
1067         if (info->fsflags & FS_NODUMP_FL)
1068                 stat->attributes |= STATX_ATTR_NODUMP;
1069         stat->attributes_mask |= (STATX_ATTR_APPEND |
1070                         STATX_ATTR_IMMUTABLE |
1071                         STATX_ATTR_NODUMP);
1072         generic_fillattr(&init_user_ns, inode, stat);
1073
1074         if (shmem_is_huge(NULL, inode, 0, false))
1075                 stat->blksize = HPAGE_PMD_SIZE;
1076
1077         if (request_mask & STATX_BTIME) {
1078                 stat->result_mask |= STATX_BTIME;
1079                 stat->btime.tv_sec = info->i_crtime.tv_sec;
1080                 stat->btime.tv_nsec = info->i_crtime.tv_nsec;
1081         }
1082
1083         return 0;
1084 }
1085
1086 static int shmem_setattr(struct user_namespace *mnt_userns,
1087                          struct dentry *dentry, struct iattr *attr)
1088 {
1089         struct inode *inode = d_inode(dentry);
1090         struct shmem_inode_info *info = SHMEM_I(inode);
1091         int error;
1092         bool update_mtime = false;
1093         bool update_ctime = true;
1094
1095         error = setattr_prepare(&init_user_ns, dentry, attr);
1096         if (error)
1097                 return error;
1098
1099         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1100                 loff_t oldsize = inode->i_size;
1101                 loff_t newsize = attr->ia_size;
1102
1103                 /* protected by i_rwsem */
1104                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1105                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1106                         return -EPERM;
1107
1108                 if (newsize != oldsize) {
1109                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1110                                         oldsize, newsize);
1111                         if (error)
1112                                 return error;
1113                         i_size_write(inode, newsize);
1114                         update_mtime = true;
1115                 } else {
1116                         update_ctime = false;
1117                 }
1118                 if (newsize <= oldsize) {
1119                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1120                         if (oldsize > holebegin)
1121                                 unmap_mapping_range(inode->i_mapping,
1122                                                         holebegin, 0, 1);
1123                         if (info->alloced)
1124                                 shmem_truncate_range(inode,
1125                                                         newsize, (loff_t)-1);
1126                         /* unmap again to remove racily COWed private pages */
1127                         if (oldsize > holebegin)
1128                                 unmap_mapping_range(inode->i_mapping,
1129                                                         holebegin, 0, 1);
1130                 }
1131         }
1132
1133         setattr_copy(&init_user_ns, inode, attr);
1134         if (attr->ia_valid & ATTR_MODE)
1135                 error = posix_acl_chmod(&init_user_ns, inode, inode->i_mode);
1136         if (!error && update_ctime) {
1137                 inode->i_ctime = current_time(inode);
1138                 if (update_mtime)
1139                         inode->i_mtime = inode->i_ctime;
1140                 inode_inc_iversion(inode);
1141         }
1142         return error;
1143 }
1144
1145 static void shmem_evict_inode(struct inode *inode)
1146 {
1147         struct shmem_inode_info *info = SHMEM_I(inode);
1148         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1149
1150         if (shmem_mapping(inode->i_mapping)) {
1151                 shmem_unacct_size(info->flags, inode->i_size);
1152                 inode->i_size = 0;
1153                 mapping_set_exiting(inode->i_mapping);
1154                 shmem_truncate_range(inode, 0, (loff_t)-1);
1155                 if (!list_empty(&info->shrinklist)) {
1156                         spin_lock(&sbinfo->shrinklist_lock);
1157                         if (!list_empty(&info->shrinklist)) {
1158                                 list_del_init(&info->shrinklist);
1159                                 sbinfo->shrinklist_len--;
1160                         }
1161                         spin_unlock(&sbinfo->shrinklist_lock);
1162                 }
1163                 while (!list_empty(&info->swaplist)) {
1164                         /* Wait while shmem_unuse() is scanning this inode... */
1165                         wait_var_event(&info->stop_eviction,
1166                                        !atomic_read(&info->stop_eviction));
1167                         mutex_lock(&shmem_swaplist_mutex);
1168                         /* ...but beware of the race if we peeked too early */
1169                         if (!atomic_read(&info->stop_eviction))
1170                                 list_del_init(&info->swaplist);
1171                         mutex_unlock(&shmem_swaplist_mutex);
1172                 }
1173         }
1174
1175         simple_xattrs_free(&info->xattrs);
1176         WARN_ON(inode->i_blocks);
1177         shmem_free_inode(inode->i_sb);
1178         clear_inode(inode);
1179 }
1180
1181 static int shmem_find_swap_entries(struct address_space *mapping,
1182                                    pgoff_t start, struct folio_batch *fbatch,
1183                                    pgoff_t *indices, unsigned int type)
1184 {
1185         XA_STATE(xas, &mapping->i_pages, start);
1186         struct folio *folio;
1187         swp_entry_t entry;
1188
1189         rcu_read_lock();
1190         xas_for_each(&xas, folio, ULONG_MAX) {
1191                 if (xas_retry(&xas, folio))
1192                         continue;
1193
1194                 if (!xa_is_value(folio))
1195                         continue;
1196
1197                 entry = radix_to_swp_entry(folio);
1198                 /*
1199                  * swapin error entries can be found in the mapping. But they're
1200                  * deliberately ignored here as we've done everything we can do.
1201                  */
1202                 if (swp_type(entry) != type)
1203                         continue;
1204
1205                 indices[folio_batch_count(fbatch)] = xas.xa_index;
1206                 if (!folio_batch_add(fbatch, folio))
1207                         break;
1208
1209                 if (need_resched()) {
1210                         xas_pause(&xas);
1211                         cond_resched_rcu();
1212                 }
1213         }
1214         rcu_read_unlock();
1215
1216         return xas.xa_index;
1217 }
1218
1219 /*
1220  * Move the swapped pages for an inode to page cache. Returns the count
1221  * of pages swapped in, or the error in case of failure.
1222  */
1223 static int shmem_unuse_swap_entries(struct inode *inode,
1224                 struct folio_batch *fbatch, pgoff_t *indices)
1225 {
1226         int i = 0;
1227         int ret = 0;
1228         int error = 0;
1229         struct address_space *mapping = inode->i_mapping;
1230
1231         for (i = 0; i < folio_batch_count(fbatch); i++) {
1232                 struct folio *folio = fbatch->folios[i];
1233
1234                 if (!xa_is_value(folio))
1235                         continue;
1236                 error = shmem_swapin_folio(inode, indices[i],
1237                                           &folio, SGP_CACHE,
1238                                           mapping_gfp_mask(mapping),
1239                                           NULL, NULL);
1240                 if (error == 0) {
1241                         folio_unlock(folio);
1242                         folio_put(folio);
1243                         ret++;
1244                 }
1245                 if (error == -ENOMEM)
1246                         break;
1247                 error = 0;
1248         }
1249         return error ? error : ret;
1250 }
1251
1252 /*
1253  * If swap found in inode, free it and move page from swapcache to filecache.
1254  */
1255 static int shmem_unuse_inode(struct inode *inode, unsigned int type)
1256 {
1257         struct address_space *mapping = inode->i_mapping;
1258         pgoff_t start = 0;
1259         struct folio_batch fbatch;
1260         pgoff_t indices[PAGEVEC_SIZE];
1261         int ret = 0;
1262
1263         do {
1264                 folio_batch_init(&fbatch);
1265                 shmem_find_swap_entries(mapping, start, &fbatch, indices, type);
1266                 if (folio_batch_count(&fbatch) == 0) {
1267                         ret = 0;
1268                         break;
1269                 }
1270
1271                 ret = shmem_unuse_swap_entries(inode, &fbatch, indices);
1272                 if (ret < 0)
1273                         break;
1274
1275                 start = indices[folio_batch_count(&fbatch) - 1];
1276         } while (true);
1277
1278         return ret;
1279 }
1280
1281 /*
1282  * Read all the shared memory data that resides in the swap
1283  * device 'type' back into memory, so the swap device can be
1284  * unused.
1285  */
1286 int shmem_unuse(unsigned int type)
1287 {
1288         struct shmem_inode_info *info, *next;
1289         int error = 0;
1290
1291         if (list_empty(&shmem_swaplist))
1292                 return 0;
1293
1294         mutex_lock(&shmem_swaplist_mutex);
1295         list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1296                 if (!info->swapped) {
1297                         list_del_init(&info->swaplist);
1298                         continue;
1299                 }
1300                 /*
1301                  * Drop the swaplist mutex while searching the inode for swap;
1302                  * but before doing so, make sure shmem_evict_inode() will not
1303                  * remove placeholder inode from swaplist, nor let it be freed
1304                  * (igrab() would protect from unlink, but not from unmount).
1305                  */
1306                 atomic_inc(&info->stop_eviction);
1307                 mutex_unlock(&shmem_swaplist_mutex);
1308
1309                 error = shmem_unuse_inode(&info->vfs_inode, type);
1310                 cond_resched();
1311
1312                 mutex_lock(&shmem_swaplist_mutex);
1313                 next = list_next_entry(info, swaplist);
1314                 if (!info->swapped)
1315                         list_del_init(&info->swaplist);
1316                 if (atomic_dec_and_test(&info->stop_eviction))
1317                         wake_up_var(&info->stop_eviction);
1318                 if (error)
1319                         break;
1320         }
1321         mutex_unlock(&shmem_swaplist_mutex);
1322
1323         return error;
1324 }
1325
1326 /*
1327  * Move the page from the page cache to the swap cache.
1328  */
1329 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1330 {
1331         struct folio *folio = page_folio(page);
1332         struct shmem_inode_info *info;
1333         struct address_space *mapping;
1334         struct inode *inode;
1335         swp_entry_t swap;
1336         pgoff_t index;
1337
1338         /*
1339          * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "always" or
1340          * "force", drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
1341          * and its shmem_writeback() needs them to be split when swapping.
1342          */
1343         if (folio_test_large(folio)) {
1344                 /* Ensure the subpages are still dirty */
1345                 folio_test_set_dirty(folio);
1346                 if (split_huge_page(page) < 0)
1347                         goto redirty;
1348                 folio = page_folio(page);
1349                 folio_clear_dirty(folio);
1350         }
1351
1352         BUG_ON(!folio_test_locked(folio));
1353         mapping = folio->mapping;
1354         index = folio->index;
1355         inode = mapping->host;
1356         info = SHMEM_I(inode);
1357         if (info->flags & VM_LOCKED)
1358                 goto redirty;
1359         if (!total_swap_pages)
1360                 goto redirty;
1361
1362         /*
1363          * Our capabilities prevent regular writeback or sync from ever calling
1364          * shmem_writepage; but a stacking filesystem might use ->writepage of
1365          * its underlying filesystem, in which case tmpfs should write out to
1366          * swap only in response to memory pressure, and not for the writeback
1367          * threads or sync.
1368          */
1369         if (!wbc->for_reclaim) {
1370                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1371                 goto redirty;
1372         }
1373
1374         /*
1375          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1376          * value into swapfile.c, the only way we can correctly account for a
1377          * fallocated folio arriving here is now to initialize it and write it.
1378          *
1379          * That's okay for a folio already fallocated earlier, but if we have
1380          * not yet completed the fallocation, then (a) we want to keep track
1381          * of this folio in case we have to undo it, and (b) it may not be a
1382          * good idea to continue anyway, once we're pushing into swap.  So
1383          * reactivate the folio, and let shmem_fallocate() quit when too many.
1384          */
1385         if (!folio_test_uptodate(folio)) {
1386                 if (inode->i_private) {
1387                         struct shmem_falloc *shmem_falloc;
1388                         spin_lock(&inode->i_lock);
1389                         shmem_falloc = inode->i_private;
1390                         if (shmem_falloc &&
1391                             !shmem_falloc->waitq &&
1392                             index >= shmem_falloc->start &&
1393                             index < shmem_falloc->next)
1394                                 shmem_falloc->nr_unswapped++;
1395                         else
1396                                 shmem_falloc = NULL;
1397                         spin_unlock(&inode->i_lock);
1398                         if (shmem_falloc)
1399                                 goto redirty;
1400                 }
1401                 folio_zero_range(folio, 0, folio_size(folio));
1402                 flush_dcache_folio(folio);
1403                 folio_mark_uptodate(folio);
1404         }
1405
1406         swap = folio_alloc_swap(folio);
1407         if (!swap.val)
1408                 goto redirty;
1409
1410         /*
1411          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1412          * if it's not already there.  Do it now before the folio is
1413          * moved to swap cache, when its pagelock no longer protects
1414          * the inode from eviction.  But don't unlock the mutex until
1415          * we've incremented swapped, because shmem_unuse_inode() will
1416          * prune a !swapped inode from the swaplist under this mutex.
1417          */
1418         mutex_lock(&shmem_swaplist_mutex);
1419         if (list_empty(&info->swaplist))
1420                 list_add(&info->swaplist, &shmem_swaplist);
1421
1422         if (add_to_swap_cache(folio, swap,
1423                         __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1424                         NULL) == 0) {
1425                 spin_lock_irq(&info->lock);
1426                 shmem_recalc_inode(inode);
1427                 info->swapped++;
1428                 spin_unlock_irq(&info->lock);
1429
1430                 swap_shmem_alloc(swap);
1431                 shmem_delete_from_page_cache(folio, swp_to_radix_entry(swap));
1432
1433                 mutex_unlock(&shmem_swaplist_mutex);
1434                 BUG_ON(folio_mapped(folio));
1435                 swap_writepage(&folio->page, wbc);
1436                 return 0;
1437         }
1438
1439         mutex_unlock(&shmem_swaplist_mutex);
1440         put_swap_folio(folio, swap);
1441 redirty:
1442         folio_mark_dirty(folio);
1443         if (wbc->for_reclaim)
1444                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with folio locked */
1445         folio_unlock(folio);
1446         return 0;
1447 }
1448
1449 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1450 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1451 {
1452         char buffer[64];
1453
1454         if (!mpol || mpol->mode == MPOL_DEFAULT)
1455                 return;         /* show nothing */
1456
1457         mpol_to_str(buffer, sizeof(buffer), mpol);
1458
1459         seq_printf(seq, ",mpol=%s", buffer);
1460 }
1461
1462 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1463 {
1464         struct mempolicy *mpol = NULL;
1465         if (sbinfo->mpol) {
1466                 raw_spin_lock(&sbinfo->stat_lock);      /* prevent replace/use races */
1467                 mpol = sbinfo->mpol;
1468                 mpol_get(mpol);
1469                 raw_spin_unlock(&sbinfo->stat_lock);
1470         }
1471         return mpol;
1472 }
1473 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1474 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1475 {
1476 }
1477 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1478 {
1479         return NULL;
1480 }
1481 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1482 #ifndef CONFIG_NUMA
1483 #define vm_policy vm_private_data
1484 #endif
1485
1486 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1487                 struct shmem_inode_info *info, pgoff_t index)
1488 {
1489         /* Create a pseudo vma that just contains the policy */
1490         vma_init(vma, NULL);
1491         /* Bias interleave by inode number to distribute better across nodes */
1492         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1493         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1494 }
1495
1496 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1497 {
1498         /* Drop reference taken by mpol_shared_policy_lookup() */
1499         mpol_cond_put(vma->vm_policy);
1500 }
1501
1502 static struct folio *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1503                         struct shmem_inode_info *info, pgoff_t index)
1504 {
1505         struct vm_area_struct pvma;
1506         struct page *page;
1507         struct vm_fault vmf = {
1508                 .vma = &pvma,
1509         };
1510
1511         shmem_pseudo_vma_init(&pvma, info, index);
1512         page = swap_cluster_readahead(swap, gfp, &vmf);
1513         shmem_pseudo_vma_destroy(&pvma);
1514
1515         if (!page)
1516                 return NULL;
1517         return page_folio(page);
1518 }
1519
1520 /*
1521  * Make sure huge_gfp is always more limited than limit_gfp.
1522  * Some of the flags set permissions, while others set limitations.
1523  */
1524 static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t limit_gfp)
1525 {
1526         gfp_t allowflags = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
1527         gfp_t denyflags = __GFP_NOWARN | __GFP_NORETRY;
1528         gfp_t zoneflags = limit_gfp & GFP_ZONEMASK;
1529         gfp_t result = huge_gfp & ~(allowflags | GFP_ZONEMASK);
1530
1531         /* Allow allocations only from the originally specified zones. */
1532         result |= zoneflags;
1533
1534         /*
1535          * Minimize the result gfp by taking the union with the deny flags,
1536          * and the intersection of the allow flags.
1537          */
1538         result |= (limit_gfp & denyflags);
1539         result |= (huge_gfp & limit_gfp) & allowflags;
1540
1541         return result;
1542 }
1543
1544 static struct folio *shmem_alloc_hugefolio(gfp_t gfp,
1545                 struct shmem_inode_info *info, pgoff_t index)
1546 {
1547         struct vm_area_struct pvma;
1548         struct address_space *mapping = info->vfs_inode.i_mapping;
1549         pgoff_t hindex;
1550         struct folio *folio;
1551
1552         hindex = round_down(index, HPAGE_PMD_NR);
1553         if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1554                                                                 XA_PRESENT))
1555                 return NULL;
1556
1557         shmem_pseudo_vma_init(&pvma, info, hindex);
1558         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, &pvma, 0, true);
1559         shmem_pseudo_vma_destroy(&pvma);
1560         if (!folio)
1561                 count_vm_event(THP_FILE_FALLBACK);
1562         return folio;
1563 }
1564
1565 static struct folio *shmem_alloc_folio(gfp_t gfp,
1566                         struct shmem_inode_info *info, pgoff_t index)
1567 {
1568         struct vm_area_struct pvma;
1569         struct folio *folio;
1570
1571         shmem_pseudo_vma_init(&pvma, info, index);
1572         folio = vma_alloc_folio(gfp, 0, &pvma, 0, false);
1573         shmem_pseudo_vma_destroy(&pvma);
1574
1575         return folio;
1576 }
1577
1578 static struct folio *shmem_alloc_and_acct_folio(gfp_t gfp, struct inode *inode,
1579                 pgoff_t index, bool huge)
1580 {
1581         struct shmem_inode_info *info = SHMEM_I(inode);
1582         struct folio *folio;
1583         int nr;
1584         int err = -ENOSPC;
1585
1586         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1587                 huge = false;
1588         nr = huge ? HPAGE_PMD_NR : 1;
1589
1590         if (!shmem_inode_acct_block(inode, nr))
1591                 goto failed;
1592
1593         if (huge)
1594                 folio = shmem_alloc_hugefolio(gfp, info, index);
1595         else
1596                 folio = shmem_alloc_folio(gfp, info, index);
1597         if (folio) {
1598                 __folio_set_locked(folio);
1599                 __folio_set_swapbacked(folio);
1600                 return folio;
1601         }
1602
1603         err = -ENOMEM;
1604         shmem_inode_unacct_blocks(inode, nr);
1605 failed:
1606         return ERR_PTR(err);
1607 }
1608
1609 /*
1610  * When a page is moved from swapcache to shmem filecache (either by the
1611  * usual swapin of shmem_get_folio_gfp(), or by the less common swapoff of
1612  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1613  * ignorance of the mapping it belongs to.  If that mapping has special
1614  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1615  * we may need to copy to a suitable page before moving to filecache.
1616  *
1617  * In a future release, this may well be extended to respect cpuset and
1618  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1619  * but for now it is a simple matter of zone.
1620  */
1621 static bool shmem_should_replace_folio(struct folio *folio, gfp_t gfp)
1622 {
1623         return folio_zonenum(folio) > gfp_zone(gfp);
1624 }
1625
1626 static int shmem_replace_folio(struct folio **foliop, gfp_t gfp,
1627                                 struct shmem_inode_info *info, pgoff_t index)
1628 {
1629         struct folio *old, *new;
1630         struct address_space *swap_mapping;
1631         swp_entry_t entry;
1632         pgoff_t swap_index;
1633         int error;
1634
1635         old = *foliop;
1636         entry = folio_swap_entry(old);
1637         swap_index = swp_offset(entry);
1638         swap_mapping = swap_address_space(entry);
1639
1640         /*
1641          * We have arrived here because our zones are constrained, so don't
1642          * limit chance of success by further cpuset and node constraints.
1643          */
1644         gfp &= ~GFP_CONSTRAINT_MASK;
1645         VM_BUG_ON_FOLIO(folio_test_large(old), old);
1646         new = shmem_alloc_folio(gfp, info, index);
1647         if (!new)
1648                 return -ENOMEM;
1649
1650         folio_get(new);
1651         folio_copy(new, old);
1652         flush_dcache_folio(new);
1653
1654         __folio_set_locked(new);
1655         __folio_set_swapbacked(new);
1656         folio_mark_uptodate(new);
1657         folio_set_swap_entry(new, entry);
1658         folio_set_swapcache(new);
1659
1660         /*
1661          * Our caller will very soon move newpage out of swapcache, but it's
1662          * a nice clean interface for us to replace oldpage by newpage there.
1663          */
1664         xa_lock_irq(&swap_mapping->i_pages);
1665         error = shmem_replace_entry(swap_mapping, swap_index, old, new);
1666         if (!error) {
1667                 mem_cgroup_migrate(old, new);
1668                 __lruvec_stat_mod_folio(new, NR_FILE_PAGES, 1);
1669                 __lruvec_stat_mod_folio(new, NR_SHMEM, 1);
1670                 __lruvec_stat_mod_folio(old, NR_FILE_PAGES, -1);
1671                 __lruvec_stat_mod_folio(old, NR_SHMEM, -1);
1672         }
1673         xa_unlock_irq(&swap_mapping->i_pages);
1674
1675         if (unlikely(error)) {
1676                 /*
1677                  * Is this possible?  I think not, now that our callers check
1678                  * both PageSwapCache and page_private after getting page lock;
1679                  * but be defensive.  Reverse old to newpage for clear and free.
1680                  */
1681                 old = new;
1682         } else {
1683                 folio_add_lru(new);
1684                 *foliop = new;
1685         }
1686
1687         folio_clear_swapcache(old);
1688         old->private = NULL;
1689
1690         folio_unlock(old);
1691         folio_put_refs(old, 2);
1692         return error;
1693 }
1694
1695 static void shmem_set_folio_swapin_error(struct inode *inode, pgoff_t index,
1696                                          struct folio *folio, swp_entry_t swap)
1697 {
1698         struct address_space *mapping = inode->i_mapping;
1699         struct shmem_inode_info *info = SHMEM_I(inode);
1700         swp_entry_t swapin_error;
1701         void *old;
1702
1703         swapin_error = make_swapin_error_entry(&folio->page);
1704         old = xa_cmpxchg_irq(&mapping->i_pages, index,
1705                              swp_to_radix_entry(swap),
1706                              swp_to_radix_entry(swapin_error), 0);
1707         if (old != swp_to_radix_entry(swap))
1708                 return;
1709
1710         folio_wait_writeback(folio);
1711         delete_from_swap_cache(folio);
1712         spin_lock_irq(&info->lock);
1713         /*
1714          * Don't treat swapin error folio as alloced. Otherwise inode->i_blocks won't
1715          * be 0 when inode is released and thus trigger WARN_ON(inode->i_blocks) in
1716          * shmem_evict_inode.
1717          */
1718         info->alloced--;
1719         info->swapped--;
1720         shmem_recalc_inode(inode);
1721         spin_unlock_irq(&info->lock);
1722         swap_free(swap);
1723 }
1724
1725 /*
1726  * Swap in the folio pointed to by *foliop.
1727  * Caller has to make sure that *foliop contains a valid swapped folio.
1728  * Returns 0 and the folio in foliop if success. On failure, returns the
1729  * error code and NULL in *foliop.
1730  */
1731 static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
1732                              struct folio **foliop, enum sgp_type sgp,
1733                              gfp_t gfp, struct vm_area_struct *vma,
1734                              vm_fault_t *fault_type)
1735 {
1736         struct address_space *mapping = inode->i_mapping;
1737         struct shmem_inode_info *info = SHMEM_I(inode);
1738         struct mm_struct *charge_mm = vma ? vma->vm_mm : NULL;
1739         struct folio *folio = NULL;
1740         swp_entry_t swap;
1741         int error;
1742
1743         VM_BUG_ON(!*foliop || !xa_is_value(*foliop));
1744         swap = radix_to_swp_entry(*foliop);
1745         *foliop = NULL;
1746
1747         if (is_swapin_error_entry(swap))
1748                 return -EIO;
1749
1750         /* Look it up and read it in.. */
1751         folio = swap_cache_get_folio(swap, NULL, 0);
1752         if (!folio) {
1753                 /* Or update major stats only when swapin succeeds?? */
1754                 if (fault_type) {
1755                         *fault_type |= VM_FAULT_MAJOR;
1756                         count_vm_event(PGMAJFAULT);
1757                         count_memcg_event_mm(charge_mm, PGMAJFAULT);
1758                 }
1759                 /* Here we actually start the io */
1760                 folio = shmem_swapin(swap, gfp, info, index);
1761                 if (!folio) {
1762                         error = -ENOMEM;
1763                         goto failed;
1764                 }
1765         }
1766
1767         /* We have to do this with folio locked to prevent races */
1768         folio_lock(folio);
1769         if (!folio_test_swapcache(folio) ||
1770             folio_swap_entry(folio).val != swap.val ||
1771             !shmem_confirm_swap(mapping, index, swap)) {
1772                 error = -EEXIST;
1773                 goto unlock;
1774         }
1775         if (!folio_test_uptodate(folio)) {
1776                 error = -EIO;
1777                 goto failed;
1778         }
1779         folio_wait_writeback(folio);
1780
1781         /*
1782          * Some architectures may have to restore extra metadata to the
1783          * folio after reading from swap.
1784          */
1785         arch_swap_restore(swap, folio);
1786
1787         if (shmem_should_replace_folio(folio, gfp)) {
1788                 error = shmem_replace_folio(&folio, gfp, info, index);
1789                 if (error)
1790                         goto failed;
1791         }
1792
1793         error = shmem_add_to_page_cache(folio, mapping, index,
1794                                         swp_to_radix_entry(swap), gfp,
1795                                         charge_mm);
1796         if (error)
1797                 goto failed;
1798
1799         spin_lock_irq(&info->lock);
1800         info->swapped--;
1801         shmem_recalc_inode(inode);
1802         spin_unlock_irq(&info->lock);
1803
1804         if (sgp == SGP_WRITE)
1805                 folio_mark_accessed(folio);
1806
1807         delete_from_swap_cache(folio);
1808         folio_mark_dirty(folio);
1809         swap_free(swap);
1810
1811         *foliop = folio;
1812         return 0;
1813 failed:
1814         if (!shmem_confirm_swap(mapping, index, swap))
1815                 error = -EEXIST;
1816         if (error == -EIO)
1817                 shmem_set_folio_swapin_error(inode, index, folio, swap);
1818 unlock:
1819         if (folio) {
1820                 folio_unlock(folio);
1821                 folio_put(folio);
1822         }
1823
1824         return error;
1825 }
1826
1827 /*
1828  * shmem_get_folio_gfp - find page in cache, or get from swap, or allocate
1829  *
1830  * If we allocate a new one we do not mark it dirty. That's up to the
1831  * vm. If we swap it in we mark it dirty since we also free the swap
1832  * entry since a page cannot live in both the swap and page cache.
1833  *
1834  * vma, vmf, and fault_type are only supplied by shmem_fault:
1835  * otherwise they are NULL.
1836  */
1837 static int shmem_get_folio_gfp(struct inode *inode, pgoff_t index,
1838                 struct folio **foliop, enum sgp_type sgp, gfp_t gfp,
1839                 struct vm_area_struct *vma, struct vm_fault *vmf,
1840                 vm_fault_t *fault_type)
1841 {
1842         struct address_space *mapping = inode->i_mapping;
1843         struct shmem_inode_info *info = SHMEM_I(inode);
1844         struct shmem_sb_info *sbinfo;
1845         struct mm_struct *charge_mm;
1846         struct folio *folio;
1847         pgoff_t hindex = index;
1848         gfp_t huge_gfp;
1849         int error;
1850         int once = 0;
1851         int alloced = 0;
1852
1853         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1854                 return -EFBIG;
1855 repeat:
1856         if (sgp <= SGP_CACHE &&
1857             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1858                 return -EINVAL;
1859         }
1860
1861         sbinfo = SHMEM_SB(inode->i_sb);
1862         charge_mm = vma ? vma->vm_mm : NULL;
1863
1864         folio = __filemap_get_folio(mapping, index, FGP_ENTRY | FGP_LOCK, 0);
1865         if (folio && vma && userfaultfd_minor(vma)) {
1866                 if (!xa_is_value(folio)) {
1867                         folio_unlock(folio);
1868                         folio_put(folio);
1869                 }
1870                 *fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
1871                 return 0;
1872         }
1873
1874         if (xa_is_value(folio)) {
1875                 error = shmem_swapin_folio(inode, index, &folio,
1876                                           sgp, gfp, vma, fault_type);
1877                 if (error == -EEXIST)
1878                         goto repeat;
1879
1880                 *foliop = folio;
1881                 return error;
1882         }
1883
1884         if (folio) {
1885                 hindex = folio->index;
1886                 if (sgp == SGP_WRITE)
1887                         folio_mark_accessed(folio);
1888                 if (folio_test_uptodate(folio))
1889                         goto out;
1890                 /* fallocated folio */
1891                 if (sgp != SGP_READ)
1892                         goto clear;
1893                 folio_unlock(folio);
1894                 folio_put(folio);
1895         }
1896
1897         /*
1898          * SGP_READ: succeed on hole, with NULL folio, letting caller zero.
1899          * SGP_NOALLOC: fail on hole, with NULL folio, letting caller fail.
1900          */
1901         *foliop = NULL;
1902         if (sgp == SGP_READ)
1903                 return 0;
1904         if (sgp == SGP_NOALLOC)
1905                 return -ENOENT;
1906
1907         /*
1908          * Fast cache lookup and swap lookup did not find it: allocate.
1909          */
1910
1911         if (vma && userfaultfd_missing(vma)) {
1912                 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1913                 return 0;
1914         }
1915
1916         if (!shmem_is_huge(vma, inode, index, false))
1917                 goto alloc_nohuge;
1918
1919         huge_gfp = vma_thp_gfp_mask(vma);
1920         huge_gfp = limit_gfp_mask(huge_gfp, gfp);
1921         folio = shmem_alloc_and_acct_folio(huge_gfp, inode, index, true);
1922         if (IS_ERR(folio)) {
1923 alloc_nohuge:
1924                 folio = shmem_alloc_and_acct_folio(gfp, inode, index, false);
1925         }
1926         if (IS_ERR(folio)) {
1927                 int retry = 5;
1928
1929                 error = PTR_ERR(folio);
1930                 folio = NULL;
1931                 if (error != -ENOSPC)
1932                         goto unlock;
1933                 /*
1934                  * Try to reclaim some space by splitting a large folio
1935                  * beyond i_size on the filesystem.
1936                  */
1937                 while (retry--) {
1938                         int ret;
1939
1940                         ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1941                         if (ret == SHRINK_STOP)
1942                                 break;
1943                         if (ret)
1944                                 goto alloc_nohuge;
1945                 }
1946                 goto unlock;
1947         }
1948
1949         hindex = round_down(index, folio_nr_pages(folio));
1950
1951         if (sgp == SGP_WRITE)
1952                 __folio_set_referenced(folio);
1953
1954         error = shmem_add_to_page_cache(folio, mapping, hindex,
1955                                         NULL, gfp & GFP_RECLAIM_MASK,
1956                                         charge_mm);
1957         if (error)
1958                 goto unacct;
1959         folio_add_lru(folio);
1960
1961         spin_lock_irq(&info->lock);
1962         info->alloced += folio_nr_pages(folio);
1963         inode->i_blocks += (blkcnt_t)BLOCKS_PER_PAGE << folio_order(folio);
1964         shmem_recalc_inode(inode);
1965         spin_unlock_irq(&info->lock);
1966         alloced = true;
1967
1968         if (folio_test_pmd_mappable(folio) &&
1969             DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1970                                         folio_next_index(folio) - 1) {
1971                 /*
1972                  * Part of the large folio is beyond i_size: subject
1973                  * to shrink under memory pressure.
1974                  */
1975                 spin_lock(&sbinfo->shrinklist_lock);
1976                 /*
1977                  * _careful to defend against unlocked access to
1978                  * ->shrink_list in shmem_unused_huge_shrink()
1979                  */
1980                 if (list_empty_careful(&info->shrinklist)) {
1981                         list_add_tail(&info->shrinklist,
1982                                       &sbinfo->shrinklist);
1983                         sbinfo->shrinklist_len++;
1984                 }
1985                 spin_unlock(&sbinfo->shrinklist_lock);
1986         }
1987
1988         /*
1989          * Let SGP_FALLOC use the SGP_WRITE optimization on a new folio.
1990          */
1991         if (sgp == SGP_FALLOC)
1992                 sgp = SGP_WRITE;
1993 clear:
1994         /*
1995          * Let SGP_WRITE caller clear ends if write does not fill folio;
1996          * but SGP_FALLOC on a folio fallocated earlier must initialize
1997          * it now, lest undo on failure cancel our earlier guarantee.
1998          */
1999         if (sgp != SGP_WRITE && !folio_test_uptodate(folio)) {
2000                 long i, n = folio_nr_pages(folio);
2001
2002                 for (i = 0; i < n; i++)
2003                         clear_highpage(folio_page(folio, i));
2004                 flush_dcache_folio(folio);
2005                 folio_mark_uptodate(folio);
2006         }
2007
2008         /* Perhaps the file has been truncated since we checked */
2009         if (sgp <= SGP_CACHE &&
2010             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
2011                 if (alloced) {
2012                         folio_clear_dirty(folio);
2013                         filemap_remove_folio(folio);
2014                         spin_lock_irq(&info->lock);
2015                         shmem_recalc_inode(inode);
2016                         spin_unlock_irq(&info->lock);
2017                 }
2018                 error = -EINVAL;
2019                 goto unlock;
2020         }
2021 out:
2022         *foliop = folio;
2023         return 0;
2024
2025         /*
2026          * Error recovery.
2027          */
2028 unacct:
2029         shmem_inode_unacct_blocks(inode, folio_nr_pages(folio));
2030
2031         if (folio_test_large(folio)) {
2032                 folio_unlock(folio);
2033                 folio_put(folio);
2034                 goto alloc_nohuge;
2035         }
2036 unlock:
2037         if (folio) {
2038                 folio_unlock(folio);
2039                 folio_put(folio);
2040         }
2041         if (error == -ENOSPC && !once++) {
2042                 spin_lock_irq(&info->lock);
2043                 shmem_recalc_inode(inode);
2044                 spin_unlock_irq(&info->lock);
2045                 goto repeat;
2046         }
2047         if (error == -EEXIST)
2048                 goto repeat;
2049         return error;
2050 }
2051
2052 int shmem_get_folio(struct inode *inode, pgoff_t index, struct folio **foliop,
2053                 enum sgp_type sgp)
2054 {
2055         return shmem_get_folio_gfp(inode, index, foliop, sgp,
2056                         mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
2057 }
2058
2059 /*
2060  * This is like autoremove_wake_function, but it removes the wait queue
2061  * entry unconditionally - even if something else had already woken the
2062  * target.
2063  */
2064 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2065 {
2066         int ret = default_wake_function(wait, mode, sync, key);
2067         list_del_init(&wait->entry);
2068         return ret;
2069 }
2070
2071 static vm_fault_t shmem_fault(struct vm_fault *vmf)
2072 {
2073         struct vm_area_struct *vma = vmf->vma;
2074         struct inode *inode = file_inode(vma->vm_file);
2075         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2076         struct folio *folio = NULL;
2077         int err;
2078         vm_fault_t ret = VM_FAULT_LOCKED;
2079
2080         /*
2081          * Trinity finds that probing a hole which tmpfs is punching can
2082          * prevent the hole-punch from ever completing: which in turn
2083          * locks writers out with its hold on i_rwsem.  So refrain from
2084          * faulting pages into the hole while it's being punched.  Although
2085          * shmem_undo_range() does remove the additions, it may be unable to
2086          * keep up, as each new page needs its own unmap_mapping_range() call,
2087          * and the i_mmap tree grows ever slower to scan if new vmas are added.
2088          *
2089          * It does not matter if we sometimes reach this check just before the
2090          * hole-punch begins, so that one fault then races with the punch:
2091          * we just need to make racing faults a rare case.
2092          *
2093          * The implementation below would be much simpler if we just used a
2094          * standard mutex or completion: but we cannot take i_rwsem in fault,
2095          * and bloating every shmem inode for this unlikely case would be sad.
2096          */
2097         if (unlikely(inode->i_private)) {
2098                 struct shmem_falloc *shmem_falloc;
2099
2100                 spin_lock(&inode->i_lock);
2101                 shmem_falloc = inode->i_private;
2102                 if (shmem_falloc &&
2103                     shmem_falloc->waitq &&
2104                     vmf->pgoff >= shmem_falloc->start &&
2105                     vmf->pgoff < shmem_falloc->next) {
2106                         struct file *fpin;
2107                         wait_queue_head_t *shmem_falloc_waitq;
2108                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2109
2110                         ret = VM_FAULT_NOPAGE;
2111                         fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2112                         if (fpin)
2113                                 ret = VM_FAULT_RETRY;
2114
2115                         shmem_falloc_waitq = shmem_falloc->waitq;
2116                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2117                                         TASK_UNINTERRUPTIBLE);
2118                         spin_unlock(&inode->i_lock);
2119                         schedule();
2120
2121                         /*
2122                          * shmem_falloc_waitq points into the shmem_fallocate()
2123                          * stack of the hole-punching task: shmem_falloc_waitq
2124                          * is usually invalid by the time we reach here, but
2125                          * finish_wait() does not dereference it in that case;
2126                          * though i_lock needed lest racing with wake_up_all().
2127                          */
2128                         spin_lock(&inode->i_lock);
2129                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2130                         spin_unlock(&inode->i_lock);
2131
2132                         if (fpin)
2133                                 fput(fpin);
2134                         return ret;
2135                 }
2136                 spin_unlock(&inode->i_lock);
2137         }
2138
2139         err = shmem_get_folio_gfp(inode, vmf->pgoff, &folio, SGP_CACHE,
2140                                   gfp, vma, vmf, &ret);
2141         if (err)
2142                 return vmf_error(err);
2143         if (folio)
2144                 vmf->page = folio_file_page(folio, vmf->pgoff);
2145         return ret;
2146 }
2147
2148 unsigned long shmem_get_unmapped_area(struct file *file,
2149                                       unsigned long uaddr, unsigned long len,
2150                                       unsigned long pgoff, unsigned long flags)
2151 {
2152         unsigned long (*get_area)(struct file *,
2153                 unsigned long, unsigned long, unsigned long, unsigned long);
2154         unsigned long addr;
2155         unsigned long offset;
2156         unsigned long inflated_len;
2157         unsigned long inflated_addr;
2158         unsigned long inflated_offset;
2159
2160         if (len > TASK_SIZE)
2161                 return -ENOMEM;
2162
2163         get_area = current->mm->get_unmapped_area;
2164         addr = get_area(file, uaddr, len, pgoff, flags);
2165
2166         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2167                 return addr;
2168         if (IS_ERR_VALUE(addr))
2169                 return addr;
2170         if (addr & ~PAGE_MASK)
2171                 return addr;
2172         if (addr > TASK_SIZE - len)
2173                 return addr;
2174
2175         if (shmem_huge == SHMEM_HUGE_DENY)
2176                 return addr;
2177         if (len < HPAGE_PMD_SIZE)
2178                 return addr;
2179         if (flags & MAP_FIXED)
2180                 return addr;
2181         /*
2182          * Our priority is to support MAP_SHARED mapped hugely;
2183          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2184          * But if caller specified an address hint and we allocated area there
2185          * successfully, respect that as before.
2186          */
2187         if (uaddr == addr)
2188                 return addr;
2189
2190         if (shmem_huge != SHMEM_HUGE_FORCE) {
2191                 struct super_block *sb;
2192
2193                 if (file) {
2194                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2195                         sb = file_inode(file)->i_sb;
2196                 } else {
2197                         /*
2198                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2199                          * for "/dev/zero", to create a shared anonymous object.
2200                          */
2201                         if (IS_ERR(shm_mnt))
2202                                 return addr;
2203                         sb = shm_mnt->mnt_sb;
2204                 }
2205                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2206                         return addr;
2207         }
2208
2209         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2210         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2211                 return addr;
2212         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2213                 return addr;
2214
2215         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2216         if (inflated_len > TASK_SIZE)
2217                 return addr;
2218         if (inflated_len < len)
2219                 return addr;
2220
2221         inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2222         if (IS_ERR_VALUE(inflated_addr))
2223                 return addr;
2224         if (inflated_addr & ~PAGE_MASK)
2225                 return addr;
2226
2227         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2228         inflated_addr += offset - inflated_offset;
2229         if (inflated_offset > offset)
2230                 inflated_addr += HPAGE_PMD_SIZE;
2231
2232         if (inflated_addr > TASK_SIZE - len)
2233                 return addr;
2234         return inflated_addr;
2235 }
2236
2237 #ifdef CONFIG_NUMA
2238 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2239 {
2240         struct inode *inode = file_inode(vma->vm_file);
2241         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2242 }
2243
2244 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2245                                           unsigned long addr)
2246 {
2247         struct inode *inode = file_inode(vma->vm_file);
2248         pgoff_t index;
2249
2250         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2251         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2252 }
2253 #endif
2254
2255 int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
2256 {
2257         struct inode *inode = file_inode(file);
2258         struct shmem_inode_info *info = SHMEM_I(inode);
2259         int retval = -ENOMEM;
2260
2261         /*
2262          * What serializes the accesses to info->flags?
2263          * ipc_lock_object() when called from shmctl_do_lock(),
2264          * no serialization needed when called from shm_destroy().
2265          */
2266         if (lock && !(info->flags & VM_LOCKED)) {
2267                 if (!user_shm_lock(inode->i_size, ucounts))
2268                         goto out_nomem;
2269                 info->flags |= VM_LOCKED;
2270                 mapping_set_unevictable(file->f_mapping);
2271         }
2272         if (!lock && (info->flags & VM_LOCKED) && ucounts) {
2273                 user_shm_unlock(inode->i_size, ucounts);
2274                 info->flags &= ~VM_LOCKED;
2275                 mapping_clear_unevictable(file->f_mapping);
2276         }
2277         retval = 0;
2278
2279 out_nomem:
2280         return retval;
2281 }
2282
2283 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2284 {
2285         struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2286         int ret;
2287
2288         ret = seal_check_future_write(info->seals, vma);
2289         if (ret)
2290                 return ret;
2291
2292         /* arm64 - allow memory tagging on RAM-based files */
2293         vma->vm_flags |= VM_MTE_ALLOWED;
2294
2295         file_accessed(file);
2296         vma->vm_ops = &shmem_vm_ops;
2297         return 0;
2298 }
2299
2300 #ifdef CONFIG_TMPFS_XATTR
2301 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2302
2303 /*
2304  * chattr's fsflags are unrelated to extended attributes,
2305  * but tmpfs has chosen to enable them under the same config option.
2306  */
2307 static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2308 {
2309         unsigned int i_flags = 0;
2310
2311         if (fsflags & FS_NOATIME_FL)
2312                 i_flags |= S_NOATIME;
2313         if (fsflags & FS_APPEND_FL)
2314                 i_flags |= S_APPEND;
2315         if (fsflags & FS_IMMUTABLE_FL)
2316                 i_flags |= S_IMMUTABLE;
2317         /*
2318          * But FS_NODUMP_FL does not require any action in i_flags.
2319          */
2320         inode_set_flags(inode, i_flags, S_NOATIME | S_APPEND | S_IMMUTABLE);
2321 }
2322 #else
2323 static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2324 {
2325 }
2326 #define shmem_initxattrs NULL
2327 #endif
2328
2329 static struct inode *shmem_get_inode(struct super_block *sb, struct inode *dir,
2330                                      umode_t mode, dev_t dev, unsigned long flags)
2331 {
2332         struct inode *inode;
2333         struct shmem_inode_info *info;
2334         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2335         ino_t ino;
2336
2337         if (shmem_reserve_inode(sb, &ino))
2338                 return NULL;
2339
2340         inode = new_inode(sb);
2341         if (inode) {
2342                 inode->i_ino = ino;
2343                 inode_init_owner(&init_user_ns, inode, dir, mode);
2344                 inode->i_blocks = 0;
2345                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2346                 inode->i_generation = get_random_u32();
2347                 info = SHMEM_I(inode);
2348                 memset(info, 0, (char *)inode - (char *)info);
2349                 spin_lock_init(&info->lock);
2350                 atomic_set(&info->stop_eviction, 0);
2351                 info->seals = F_SEAL_SEAL;
2352                 info->flags = flags & VM_NORESERVE;
2353                 info->i_crtime = inode->i_mtime;
2354                 info->fsflags = (dir == NULL) ? 0 :
2355                         SHMEM_I(dir)->fsflags & SHMEM_FL_INHERITED;
2356                 if (info->fsflags)
2357                         shmem_set_inode_flags(inode, info->fsflags);
2358                 INIT_LIST_HEAD(&info->shrinklist);
2359                 INIT_LIST_HEAD(&info->swaplist);
2360                 simple_xattrs_init(&info->xattrs);
2361                 cache_no_acl(inode);
2362                 mapping_set_large_folios(inode->i_mapping);
2363
2364                 switch (mode & S_IFMT) {
2365                 default:
2366                         inode->i_op = &shmem_special_inode_operations;
2367                         init_special_inode(inode, mode, dev);
2368                         break;
2369                 case S_IFREG:
2370                         inode->i_mapping->a_ops = &shmem_aops;
2371                         inode->i_op = &shmem_inode_operations;
2372                         inode->i_fop = &shmem_file_operations;
2373                         mpol_shared_policy_init(&info->policy,
2374                                                  shmem_get_sbmpol(sbinfo));
2375                         break;
2376                 case S_IFDIR:
2377                         inc_nlink(inode);
2378                         /* Some things misbehave if size == 0 on a directory */
2379                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2380                         inode->i_op = &shmem_dir_inode_operations;
2381                         inode->i_fop = &simple_dir_operations;
2382                         break;
2383                 case S_IFLNK:
2384                         /*
2385                          * Must not load anything in the rbtree,
2386                          * mpol_free_shared_policy will not be called.
2387                          */
2388                         mpol_shared_policy_init(&info->policy, NULL);
2389                         break;
2390                 }
2391
2392                 lockdep_annotate_inode_mutex_key(inode);
2393         } else
2394                 shmem_free_inode(sb);
2395         return inode;
2396 }
2397
2398 #ifdef CONFIG_USERFAULTFD
2399 int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2400                            pmd_t *dst_pmd,
2401                            struct vm_area_struct *dst_vma,
2402                            unsigned long dst_addr,
2403                            unsigned long src_addr,
2404                            bool zeropage, bool wp_copy,
2405                            struct page **pagep)
2406 {
2407         struct inode *inode = file_inode(dst_vma->vm_file);
2408         struct shmem_inode_info *info = SHMEM_I(inode);
2409         struct address_space *mapping = inode->i_mapping;
2410         gfp_t gfp = mapping_gfp_mask(mapping);
2411         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2412         void *page_kaddr;
2413         struct folio *folio;
2414         int ret;
2415         pgoff_t max_off;
2416
2417         if (!shmem_inode_acct_block(inode, 1)) {
2418                 /*
2419                  * We may have got a page, returned -ENOENT triggering a retry,
2420                  * and now we find ourselves with -ENOMEM. Release the page, to
2421                  * avoid a BUG_ON in our caller.
2422                  */
2423                 if (unlikely(*pagep)) {
2424                         put_page(*pagep);
2425                         *pagep = NULL;
2426                 }
2427                 return -ENOMEM;
2428         }
2429
2430         if (!*pagep) {
2431                 ret = -ENOMEM;
2432                 folio = shmem_alloc_folio(gfp, info, pgoff);
2433                 if (!folio)
2434                         goto out_unacct_blocks;
2435
2436                 if (!zeropage) {        /* COPY */
2437                         page_kaddr = kmap_local_folio(folio, 0);
2438                         /*
2439                          * The read mmap_lock is held here.  Despite the
2440                          * mmap_lock being read recursive a deadlock is still
2441                          * possible if a writer has taken a lock.  For example:
2442                          *
2443                          * process A thread 1 takes read lock on own mmap_lock
2444                          * process A thread 2 calls mmap, blocks taking write lock
2445                          * process B thread 1 takes page fault, read lock on own mmap lock
2446                          * process B thread 2 calls mmap, blocks taking write lock
2447                          * process A thread 1 blocks taking read lock on process B
2448                          * process B thread 1 blocks taking read lock on process A
2449                          *
2450                          * Disable page faults to prevent potential deadlock
2451                          * and retry the copy outside the mmap_lock.
2452                          */
2453                         pagefault_disable();
2454                         ret = copy_from_user(page_kaddr,
2455                                              (const void __user *)src_addr,
2456                                              PAGE_SIZE);
2457                         pagefault_enable();
2458                         kunmap_local(page_kaddr);
2459
2460                         /* fallback to copy_from_user outside mmap_lock */
2461                         if (unlikely(ret)) {
2462                                 *pagep = &folio->page;
2463                                 ret = -ENOENT;
2464                                 /* don't free the page */
2465                                 goto out_unacct_blocks;
2466                         }
2467
2468                         flush_dcache_folio(folio);
2469                 } else {                /* ZEROPAGE */
2470                         clear_user_highpage(&folio->page, dst_addr);
2471                 }
2472         } else {
2473                 folio = page_folio(*pagep);
2474                 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2475                 *pagep = NULL;
2476         }
2477
2478         VM_BUG_ON(folio_test_locked(folio));
2479         VM_BUG_ON(folio_test_swapbacked(folio));
2480         __folio_set_locked(folio);
2481         __folio_set_swapbacked(folio);
2482         __folio_mark_uptodate(folio);
2483
2484         ret = -EFAULT;
2485         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2486         if (unlikely(pgoff >= max_off))
2487                 goto out_release;
2488
2489         ret = shmem_add_to_page_cache(folio, mapping, pgoff, NULL,
2490                                       gfp & GFP_RECLAIM_MASK, dst_mm);
2491         if (ret)
2492                 goto out_release;
2493
2494         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
2495                                        &folio->page, true, wp_copy);
2496         if (ret)
2497                 goto out_delete_from_cache;
2498
2499         spin_lock_irq(&info->lock);
2500         info->alloced++;
2501         inode->i_blocks += BLOCKS_PER_PAGE;
2502         shmem_recalc_inode(inode);
2503         spin_unlock_irq(&info->lock);
2504
2505         folio_unlock(folio);
2506         return 0;
2507 out_delete_from_cache:
2508         filemap_remove_folio(folio);
2509 out_release:
2510         folio_unlock(folio);
2511         folio_put(folio);
2512 out_unacct_blocks:
2513         shmem_inode_unacct_blocks(inode, 1);
2514         return ret;
2515 }
2516 #endif /* CONFIG_USERFAULTFD */
2517
2518 #ifdef CONFIG_TMPFS
2519 static const struct inode_operations shmem_symlink_inode_operations;
2520 static const struct inode_operations shmem_short_symlink_operations;
2521
2522 static int
2523 shmem_write_begin(struct file *file, struct address_space *mapping,
2524                         loff_t pos, unsigned len,
2525                         struct page **pagep, void **fsdata)
2526 {
2527         struct inode *inode = mapping->host;
2528         struct shmem_inode_info *info = SHMEM_I(inode);
2529         pgoff_t index = pos >> PAGE_SHIFT;
2530         struct folio *folio;
2531         int ret = 0;
2532
2533         /* i_rwsem is held by caller */
2534         if (unlikely(info->seals & (F_SEAL_GROW |
2535                                    F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2536                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2537                         return -EPERM;
2538                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2539                         return -EPERM;
2540         }
2541
2542         ret = shmem_get_folio(inode, index, &folio, SGP_WRITE);
2543
2544         if (ret)
2545                 return ret;
2546
2547         *pagep = folio_file_page(folio, index);
2548         if (PageHWPoison(*pagep)) {
2549                 folio_unlock(folio);
2550                 folio_put(folio);
2551                 *pagep = NULL;
2552                 return -EIO;
2553         }
2554
2555         return 0;
2556 }
2557
2558 static int
2559 shmem_write_end(struct file *file, struct address_space *mapping,
2560                         loff_t pos, unsigned len, unsigned copied,
2561                         struct page *page, void *fsdata)
2562 {
2563         struct inode *inode = mapping->host;
2564
2565         if (pos + copied > inode->i_size)
2566                 i_size_write(inode, pos + copied);
2567
2568         if (!PageUptodate(page)) {
2569                 struct page *head = compound_head(page);
2570                 if (PageTransCompound(page)) {
2571                         int i;
2572
2573                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2574                                 if (head + i == page)
2575                                         continue;
2576                                 clear_highpage(head + i);
2577                                 flush_dcache_page(head + i);
2578                         }
2579                 }
2580                 if (copied < PAGE_SIZE) {
2581                         unsigned from = pos & (PAGE_SIZE - 1);
2582                         zero_user_segments(page, 0, from,
2583                                         from + copied, PAGE_SIZE);
2584                 }
2585                 SetPageUptodate(head);
2586         }
2587         set_page_dirty(page);
2588         unlock_page(page);
2589         put_page(page);
2590
2591         return copied;
2592 }
2593
2594 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2595 {
2596         struct file *file = iocb->ki_filp;
2597         struct inode *inode = file_inode(file);
2598         struct address_space *mapping = inode->i_mapping;
2599         pgoff_t index;
2600         unsigned long offset;
2601         int error = 0;
2602         ssize_t retval = 0;
2603         loff_t *ppos = &iocb->ki_pos;
2604
2605         index = *ppos >> PAGE_SHIFT;
2606         offset = *ppos & ~PAGE_MASK;
2607
2608         for (;;) {
2609                 struct folio *folio = NULL;
2610                 struct page *page = NULL;
2611                 pgoff_t end_index;
2612                 unsigned long nr, ret;
2613                 loff_t i_size = i_size_read(inode);
2614
2615                 end_index = i_size >> PAGE_SHIFT;
2616                 if (index > end_index)
2617                         break;
2618                 if (index == end_index) {
2619                         nr = i_size & ~PAGE_MASK;
2620                         if (nr <= offset)
2621                                 break;
2622                 }
2623
2624                 error = shmem_get_folio(inode, index, &folio, SGP_READ);
2625                 if (error) {
2626                         if (error == -EINVAL)
2627                                 error = 0;
2628                         break;
2629                 }
2630                 if (folio) {
2631                         folio_unlock(folio);
2632
2633                         page = folio_file_page(folio, index);
2634                         if (PageHWPoison(page)) {
2635                                 folio_put(folio);
2636                                 error = -EIO;
2637                                 break;
2638                         }
2639                 }
2640
2641                 /*
2642                  * We must evaluate after, since reads (unlike writes)
2643                  * are called without i_rwsem protection against truncate
2644                  */
2645                 nr = PAGE_SIZE;
2646                 i_size = i_size_read(inode);
2647                 end_index = i_size >> PAGE_SHIFT;
2648                 if (index == end_index) {
2649                         nr = i_size & ~PAGE_MASK;
2650                         if (nr <= offset) {
2651                                 if (folio)
2652                                         folio_put(folio);
2653                                 break;
2654                         }
2655                 }
2656                 nr -= offset;
2657
2658                 if (folio) {
2659                         /*
2660                          * If users can be writing to this page using arbitrary
2661                          * virtual addresses, take care about potential aliasing
2662                          * before reading the page on the kernel side.
2663                          */
2664                         if (mapping_writably_mapped(mapping))
2665                                 flush_dcache_page(page);
2666                         /*
2667                          * Mark the page accessed if we read the beginning.
2668                          */
2669                         if (!offset)
2670                                 folio_mark_accessed(folio);
2671                         /*
2672                          * Ok, we have the page, and it's up-to-date, so
2673                          * now we can copy it to user space...
2674                          */
2675                         ret = copy_page_to_iter(page, offset, nr, to);
2676                         folio_put(folio);
2677
2678                 } else if (user_backed_iter(to)) {
2679                         /*
2680                          * Copy to user tends to be so well optimized, but
2681                          * clear_user() not so much, that it is noticeably
2682                          * faster to copy the zero page instead of clearing.
2683                          */
2684                         ret = copy_page_to_iter(ZERO_PAGE(0), offset, nr, to);
2685                 } else {
2686                         /*
2687                          * But submitting the same page twice in a row to
2688                          * splice() - or others? - can result in confusion:
2689                          * so don't attempt that optimization on pipes etc.
2690                          */
2691                         ret = iov_iter_zero(nr, to);
2692                 }
2693
2694                 retval += ret;
2695                 offset += ret;
2696                 index += offset >> PAGE_SHIFT;
2697                 offset &= ~PAGE_MASK;
2698
2699                 if (!iov_iter_count(to))
2700                         break;
2701                 if (ret < nr) {
2702                         error = -EFAULT;
2703                         break;
2704                 }
2705                 cond_resched();
2706         }
2707
2708         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2709         file_accessed(file);
2710         return retval ? retval : error;
2711 }
2712
2713 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2714 {
2715         struct address_space *mapping = file->f_mapping;
2716         struct inode *inode = mapping->host;
2717
2718         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2719                 return generic_file_llseek_size(file, offset, whence,
2720                                         MAX_LFS_FILESIZE, i_size_read(inode));
2721         if (offset < 0)
2722                 return -ENXIO;
2723
2724         inode_lock(inode);
2725         /* We're holding i_rwsem so we can access i_size directly */
2726         offset = mapping_seek_hole_data(mapping, offset, inode->i_size, whence);
2727         if (offset >= 0)
2728                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2729         inode_unlock(inode);
2730         return offset;
2731 }
2732
2733 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2734                                                          loff_t len)
2735 {
2736         struct inode *inode = file_inode(file);
2737         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2738         struct shmem_inode_info *info = SHMEM_I(inode);
2739         struct shmem_falloc shmem_falloc;
2740         pgoff_t start, index, end, undo_fallocend;
2741         int error;
2742
2743         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2744                 return -EOPNOTSUPP;
2745
2746         inode_lock(inode);
2747
2748         if (mode & FALLOC_FL_PUNCH_HOLE) {
2749                 struct address_space *mapping = file->f_mapping;
2750                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2751                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2752                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2753
2754                 /* protected by i_rwsem */
2755                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2756                         error = -EPERM;
2757                         goto out;
2758                 }
2759
2760                 shmem_falloc.waitq = &shmem_falloc_waitq;
2761                 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2762                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2763                 spin_lock(&inode->i_lock);
2764                 inode->i_private = &shmem_falloc;
2765                 spin_unlock(&inode->i_lock);
2766
2767                 if ((u64)unmap_end > (u64)unmap_start)
2768                         unmap_mapping_range(mapping, unmap_start,
2769                                             1 + unmap_end - unmap_start, 0);
2770                 shmem_truncate_range(inode, offset, offset + len - 1);
2771                 /* No need to unmap again: hole-punching leaves COWed pages */
2772
2773                 spin_lock(&inode->i_lock);
2774                 inode->i_private = NULL;
2775                 wake_up_all(&shmem_falloc_waitq);
2776                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2777                 spin_unlock(&inode->i_lock);
2778                 error = 0;
2779                 goto out;
2780         }
2781
2782         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2783         error = inode_newsize_ok(inode, offset + len);
2784         if (error)
2785                 goto out;
2786
2787         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2788                 error = -EPERM;
2789                 goto out;
2790         }
2791
2792         start = offset >> PAGE_SHIFT;
2793         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2794         /* Try to avoid a swapstorm if len is impossible to satisfy */
2795         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2796                 error = -ENOSPC;
2797                 goto out;
2798         }
2799
2800         shmem_falloc.waitq = NULL;
2801         shmem_falloc.start = start;
2802         shmem_falloc.next  = start;
2803         shmem_falloc.nr_falloced = 0;
2804         shmem_falloc.nr_unswapped = 0;
2805         spin_lock(&inode->i_lock);
2806         inode->i_private = &shmem_falloc;
2807         spin_unlock(&inode->i_lock);
2808
2809         /*
2810          * info->fallocend is only relevant when huge pages might be
2811          * involved: to prevent split_huge_page() freeing fallocated
2812          * pages when FALLOC_FL_KEEP_SIZE committed beyond i_size.
2813          */
2814         undo_fallocend = info->fallocend;
2815         if (info->fallocend < end)
2816                 info->fallocend = end;
2817
2818         for (index = start; index < end; ) {
2819                 struct folio *folio;
2820
2821                 /*
2822                  * Good, the fallocate(2) manpage permits EINTR: we may have
2823                  * been interrupted because we are using up too much memory.
2824                  */
2825                 if (signal_pending(current))
2826                         error = -EINTR;
2827                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2828                         error = -ENOMEM;
2829                 else
2830                         error = shmem_get_folio(inode, index, &folio,
2831                                                 SGP_FALLOC);
2832                 if (error) {
2833                         info->fallocend = undo_fallocend;
2834                         /* Remove the !uptodate folios we added */
2835                         if (index > start) {
2836                                 shmem_undo_range(inode,
2837                                     (loff_t)start << PAGE_SHIFT,
2838                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2839                         }
2840                         goto undone;
2841                 }
2842
2843                 /*
2844                  * Here is a more important optimization than it appears:
2845                  * a second SGP_FALLOC on the same large folio will clear it,
2846                  * making it uptodate and un-undoable if we fail later.
2847                  */
2848                 index = folio_next_index(folio);
2849                 /* Beware 32-bit wraparound */
2850                 if (!index)
2851                         index--;
2852
2853                 /*
2854                  * Inform shmem_writepage() how far we have reached.
2855                  * No need for lock or barrier: we have the page lock.
2856                  */
2857                 if (!folio_test_uptodate(folio))
2858                         shmem_falloc.nr_falloced += index - shmem_falloc.next;
2859                 shmem_falloc.next = index;
2860
2861                 /*
2862                  * If !uptodate, leave it that way so that freeable folios
2863                  * can be recognized if we need to rollback on error later.
2864                  * But mark it dirty so that memory pressure will swap rather
2865                  * than free the folios we are allocating (and SGP_CACHE folios
2866                  * might still be clean: we now need to mark those dirty too).
2867                  */
2868                 folio_mark_dirty(folio);
2869                 folio_unlock(folio);
2870                 folio_put(folio);
2871                 cond_resched();
2872         }
2873
2874         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2875                 i_size_write(inode, offset + len);
2876 undone:
2877         spin_lock(&inode->i_lock);
2878         inode->i_private = NULL;
2879         spin_unlock(&inode->i_lock);
2880 out:
2881         if (!error)
2882                 file_modified(file);
2883         inode_unlock(inode);
2884         return error;
2885 }
2886
2887 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2888 {
2889         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2890
2891         buf->f_type = TMPFS_MAGIC;
2892         buf->f_bsize = PAGE_SIZE;
2893         buf->f_namelen = NAME_MAX;
2894         if (sbinfo->max_blocks) {
2895                 buf->f_blocks = sbinfo->max_blocks;
2896                 buf->f_bavail =
2897                 buf->f_bfree  = sbinfo->max_blocks -
2898                                 percpu_counter_sum(&sbinfo->used_blocks);
2899         }
2900         if (sbinfo->max_inodes) {
2901                 buf->f_files = sbinfo->max_inodes;
2902                 buf->f_ffree = sbinfo->free_inodes;
2903         }
2904         /* else leave those fields 0 like simple_statfs */
2905
2906         buf->f_fsid = uuid_to_fsid(dentry->d_sb->s_uuid.b);
2907
2908         return 0;
2909 }
2910
2911 /*
2912  * File creation. Allocate an inode, and we're done..
2913  */
2914 static int
2915 shmem_mknod(struct user_namespace *mnt_userns, struct inode *dir,
2916             struct dentry *dentry, umode_t mode, dev_t dev)
2917 {
2918         struct inode *inode;
2919         int error = -ENOSPC;
2920
2921         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2922         if (inode) {
2923                 error = simple_acl_create(dir, inode);
2924                 if (error)
2925                         goto out_iput;
2926                 error = security_inode_init_security(inode, dir,
2927                                                      &dentry->d_name,
2928                                                      shmem_initxattrs, NULL);
2929                 if (error && error != -EOPNOTSUPP)
2930                         goto out_iput;
2931
2932                 error = 0;
2933                 dir->i_size += BOGO_DIRENT_SIZE;
2934                 dir->i_ctime = dir->i_mtime = current_time(dir);
2935                 inode_inc_iversion(dir);
2936                 d_instantiate(dentry, inode);
2937                 dget(dentry); /* Extra count - pin the dentry in core */
2938         }
2939         return error;
2940 out_iput:
2941         iput(inode);
2942         return error;
2943 }
2944
2945 static int
2946 shmem_tmpfile(struct user_namespace *mnt_userns, struct inode *dir,
2947               struct file *file, umode_t mode)
2948 {
2949         struct inode *inode;
2950         int error = -ENOSPC;
2951
2952         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2953         if (inode) {
2954                 error = security_inode_init_security(inode, dir,
2955                                                      NULL,
2956                                                      shmem_initxattrs, NULL);
2957                 if (error && error != -EOPNOTSUPP)
2958                         goto out_iput;
2959                 error = simple_acl_create(dir, inode);
2960                 if (error)
2961                         goto out_iput;
2962                 d_tmpfile(file, inode);
2963         }
2964         return finish_open_simple(file, error);
2965 out_iput:
2966         iput(inode);
2967         return error;
2968 }
2969
2970 static int shmem_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
2971                        struct dentry *dentry, umode_t mode)
2972 {
2973         int error;
2974
2975         if ((error = shmem_mknod(&init_user_ns, dir, dentry,
2976                                  mode | S_IFDIR, 0)))
2977                 return error;
2978         inc_nlink(dir);
2979         return 0;
2980 }
2981
2982 static int shmem_create(struct user_namespace *mnt_userns, struct inode *dir,
2983                         struct dentry *dentry, umode_t mode, bool excl)
2984 {
2985         return shmem_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
2986 }
2987
2988 /*
2989  * Link a file..
2990  */
2991 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2992 {
2993         struct inode *inode = d_inode(old_dentry);
2994         int ret = 0;
2995
2996         /*
2997          * No ordinary (disk based) filesystem counts links as inodes;
2998          * but each new link needs a new dentry, pinning lowmem, and
2999          * tmpfs dentries cannot be pruned until they are unlinked.
3000          * But if an O_TMPFILE file is linked into the tmpfs, the
3001          * first link must skip that, to get the accounting right.
3002          */
3003         if (inode->i_nlink) {
3004                 ret = shmem_reserve_inode(inode->i_sb, NULL);
3005                 if (ret)
3006                         goto out;
3007         }
3008
3009         dir->i_size += BOGO_DIRENT_SIZE;
3010         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3011         inode_inc_iversion(dir);
3012         inc_nlink(inode);
3013         ihold(inode);   /* New dentry reference */
3014         dget(dentry);           /* Extra pinning count for the created dentry */
3015         d_instantiate(dentry, inode);
3016 out:
3017         return ret;
3018 }
3019
3020 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3021 {
3022         struct inode *inode = d_inode(dentry);
3023
3024         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3025                 shmem_free_inode(inode->i_sb);
3026
3027         dir->i_size -= BOGO_DIRENT_SIZE;
3028         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3029         inode_inc_iversion(dir);
3030         drop_nlink(inode);
3031         dput(dentry);   /* Undo the count from "create" - this does all the work */
3032         return 0;
3033 }
3034
3035 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3036 {
3037         if (!simple_empty(dentry))
3038                 return -ENOTEMPTY;
3039
3040         drop_nlink(d_inode(dentry));
3041         drop_nlink(dir);
3042         return shmem_unlink(dir, dentry);
3043 }
3044
3045 static int shmem_whiteout(struct user_namespace *mnt_userns,
3046                           struct inode *old_dir, struct dentry *old_dentry)
3047 {
3048         struct dentry *whiteout;
3049         int error;
3050
3051         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3052         if (!whiteout)
3053                 return -ENOMEM;
3054
3055         error = shmem_mknod(&init_user_ns, old_dir, whiteout,
3056                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3057         dput(whiteout);
3058         if (error)
3059                 return error;
3060
3061         /*
3062          * Cheat and hash the whiteout while the old dentry is still in
3063          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3064          *
3065          * d_lookup() will consistently find one of them at this point,
3066          * not sure which one, but that isn't even important.
3067          */
3068         d_rehash(whiteout);
3069         return 0;
3070 }
3071
3072 /*
3073  * The VFS layer already does all the dentry stuff for rename,
3074  * we just have to decrement the usage count for the target if
3075  * it exists so that the VFS layer correctly free's it when it
3076  * gets overwritten.
3077  */
3078 static int shmem_rename2(struct user_namespace *mnt_userns,
3079                          struct inode *old_dir, struct dentry *old_dentry,
3080                          struct inode *new_dir, struct dentry *new_dentry,
3081                          unsigned int flags)
3082 {
3083         struct inode *inode = d_inode(old_dentry);
3084         int they_are_dirs = S_ISDIR(inode->i_mode);
3085
3086         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3087                 return -EINVAL;
3088
3089         if (flags & RENAME_EXCHANGE)
3090                 return simple_rename_exchange(old_dir, old_dentry, new_dir, new_dentry);
3091
3092         if (!simple_empty(new_dentry))
3093                 return -ENOTEMPTY;
3094
3095         if (flags & RENAME_WHITEOUT) {
3096                 int error;
3097
3098                 error = shmem_whiteout(&init_user_ns, old_dir, old_dentry);
3099                 if (error)
3100                         return error;
3101         }
3102
3103         if (d_really_is_positive(new_dentry)) {
3104                 (void) shmem_unlink(new_dir, new_dentry);
3105                 if (they_are_dirs) {
3106                         drop_nlink(d_inode(new_dentry));
3107                         drop_nlink(old_dir);
3108                 }
3109         } else if (they_are_dirs) {
3110                 drop_nlink(old_dir);
3111                 inc_nlink(new_dir);
3112         }
3113
3114         old_dir->i_size -= BOGO_DIRENT_SIZE;
3115         new_dir->i_size += BOGO_DIRENT_SIZE;
3116         old_dir->i_ctime = old_dir->i_mtime =
3117         new_dir->i_ctime = new_dir->i_mtime =
3118         inode->i_ctime = current_time(old_dir);
3119         inode_inc_iversion(old_dir);
3120         inode_inc_iversion(new_dir);
3121         return 0;
3122 }
3123
3124 static int shmem_symlink(struct user_namespace *mnt_userns, struct inode *dir,
3125                          struct dentry *dentry, const char *symname)
3126 {
3127         int error;
3128         int len;
3129         struct inode *inode;
3130         struct folio *folio;
3131
3132         len = strlen(symname) + 1;
3133         if (len > PAGE_SIZE)
3134                 return -ENAMETOOLONG;
3135
3136         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3137                                 VM_NORESERVE);
3138         if (!inode)
3139                 return -ENOSPC;
3140
3141         error = security_inode_init_security(inode, dir, &dentry->d_name,
3142                                              shmem_initxattrs, NULL);
3143         if (error && error != -EOPNOTSUPP) {
3144                 iput(inode);
3145                 return error;
3146         }
3147
3148         inode->i_size = len-1;
3149         if (len <= SHORT_SYMLINK_LEN) {
3150                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3151                 if (!inode->i_link) {
3152                         iput(inode);
3153                         return -ENOMEM;
3154                 }
3155                 inode->i_op = &shmem_short_symlink_operations;
3156         } else {
3157                 inode_nohighmem(inode);
3158                 error = shmem_get_folio(inode, 0, &folio, SGP_WRITE);
3159                 if (error) {
3160                         iput(inode);
3161                         return error;
3162                 }
3163                 inode->i_mapping->a_ops = &shmem_aops;
3164                 inode->i_op = &shmem_symlink_inode_operations;
3165                 memcpy(folio_address(folio), symname, len);
3166                 folio_mark_uptodate(folio);
3167                 folio_mark_dirty(folio);
3168                 folio_unlock(folio);
3169                 folio_put(folio);
3170         }
3171         dir->i_size += BOGO_DIRENT_SIZE;
3172         dir->i_ctime = dir->i_mtime = current_time(dir);
3173         inode_inc_iversion(dir);
3174         d_instantiate(dentry, inode);
3175         dget(dentry);
3176         return 0;
3177 }
3178
3179 static void shmem_put_link(void *arg)
3180 {
3181         folio_mark_accessed(arg);
3182         folio_put(arg);
3183 }
3184
3185 static const char *shmem_get_link(struct dentry *dentry,
3186                                   struct inode *inode,
3187                                   struct delayed_call *done)
3188 {
3189         struct folio *folio = NULL;
3190         int error;
3191
3192         if (!dentry) {
3193                 folio = filemap_get_folio(inode->i_mapping, 0);
3194                 if (!folio)
3195                         return ERR_PTR(-ECHILD);
3196                 if (PageHWPoison(folio_page(folio, 0)) ||
3197                     !folio_test_uptodate(folio)) {
3198                         folio_put(folio);
3199                         return ERR_PTR(-ECHILD);
3200                 }
3201         } else {
3202                 error = shmem_get_folio(inode, 0, &folio, SGP_READ);
3203                 if (error)
3204                         return ERR_PTR(error);
3205                 if (!folio)
3206                         return ERR_PTR(-ECHILD);
3207                 if (PageHWPoison(folio_page(folio, 0))) {
3208                         folio_unlock(folio);
3209                         folio_put(folio);
3210                         return ERR_PTR(-ECHILD);
3211                 }
3212                 folio_unlock(folio);
3213         }
3214         set_delayed_call(done, shmem_put_link, folio);
3215         return folio_address(folio);
3216 }
3217
3218 #ifdef CONFIG_TMPFS_XATTR
3219
3220 static int shmem_fileattr_get(struct dentry *dentry, struct fileattr *fa)
3221 {
3222         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3223
3224         fileattr_fill_flags(fa, info->fsflags & SHMEM_FL_USER_VISIBLE);
3225
3226         return 0;
3227 }
3228
3229 static int shmem_fileattr_set(struct user_namespace *mnt_userns,
3230                               struct dentry *dentry, struct fileattr *fa)
3231 {
3232         struct inode *inode = d_inode(dentry);
3233         struct shmem_inode_info *info = SHMEM_I(inode);
3234
3235         if (fileattr_has_fsx(fa))
3236                 return -EOPNOTSUPP;
3237         if (fa->flags & ~SHMEM_FL_USER_MODIFIABLE)
3238                 return -EOPNOTSUPP;
3239
3240         info->fsflags = (info->fsflags & ~SHMEM_FL_USER_MODIFIABLE) |
3241                 (fa->flags & SHMEM_FL_USER_MODIFIABLE);
3242
3243         shmem_set_inode_flags(inode, info->fsflags);
3244         inode->i_ctime = current_time(inode);
3245         inode_inc_iversion(inode);
3246         return 0;
3247 }
3248
3249 /*
3250  * Superblocks without xattr inode operations may get some security.* xattr
3251  * support from the LSM "for free". As soon as we have any other xattrs
3252  * like ACLs, we also need to implement the security.* handlers at
3253  * filesystem level, though.
3254  */
3255
3256 /*
3257  * Callback for security_inode_init_security() for acquiring xattrs.
3258  */
3259 static int shmem_initxattrs(struct inode *inode,
3260                             const struct xattr *xattr_array,
3261                             void *fs_info)
3262 {
3263         struct shmem_inode_info *info = SHMEM_I(inode);
3264         const struct xattr *xattr;
3265         struct simple_xattr *new_xattr;
3266         size_t len;
3267
3268         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3269                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3270                 if (!new_xattr)
3271                         return -ENOMEM;
3272
3273                 len = strlen(xattr->name) + 1;
3274                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3275                                           GFP_KERNEL);
3276                 if (!new_xattr->name) {
3277                         kvfree(new_xattr);
3278                         return -ENOMEM;
3279                 }
3280
3281                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3282                        XATTR_SECURITY_PREFIX_LEN);
3283                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3284                        xattr->name, len);
3285
3286                 simple_xattr_list_add(&info->xattrs, new_xattr);
3287         }
3288
3289         return 0;
3290 }
3291
3292 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3293                                    struct dentry *unused, struct inode *inode,
3294                                    const char *name, void *buffer, size_t size)
3295 {
3296         struct shmem_inode_info *info = SHMEM_I(inode);
3297
3298         name = xattr_full_name(handler, name);
3299         return simple_xattr_get(&info->xattrs, name, buffer, size);
3300 }
3301
3302 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3303                                    struct user_namespace *mnt_userns,
3304                                    struct dentry *unused, struct inode *inode,
3305                                    const char *name, const void *value,
3306                                    size_t size, int flags)
3307 {
3308         struct shmem_inode_info *info = SHMEM_I(inode);
3309         int err;
3310
3311         name = xattr_full_name(handler, name);
3312         err = simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3313         if (!err) {
3314                 inode->i_ctime = current_time(inode);
3315                 inode_inc_iversion(inode);
3316         }
3317         return err;
3318 }
3319
3320 static const struct xattr_handler shmem_security_xattr_handler = {
3321         .prefix = XATTR_SECURITY_PREFIX,
3322         .get = shmem_xattr_handler_get,
3323         .set = shmem_xattr_handler_set,
3324 };
3325
3326 static const struct xattr_handler shmem_trusted_xattr_handler = {
3327         .prefix = XATTR_TRUSTED_PREFIX,
3328         .get = shmem_xattr_handler_get,
3329         .set = shmem_xattr_handler_set,
3330 };
3331
3332 static const struct xattr_handler *shmem_xattr_handlers[] = {
3333 #ifdef CONFIG_TMPFS_POSIX_ACL
3334         &posix_acl_access_xattr_handler,
3335         &posix_acl_default_xattr_handler,
3336 #endif
3337         &shmem_security_xattr_handler,
3338         &shmem_trusted_xattr_handler,
3339         NULL
3340 };
3341
3342 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3343 {
3344         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3345         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3346 }
3347 #endif /* CONFIG_TMPFS_XATTR */
3348
3349 static const struct inode_operations shmem_short_symlink_operations = {
3350         .getattr        = shmem_getattr,
3351         .get_link       = simple_get_link,
3352 #ifdef CONFIG_TMPFS_XATTR
3353         .listxattr      = shmem_listxattr,
3354 #endif
3355 };
3356
3357 static const struct inode_operations shmem_symlink_inode_operations = {
3358         .getattr        = shmem_getattr,
3359         .get_link       = shmem_get_link,
3360 #ifdef CONFIG_TMPFS_XATTR
3361         .listxattr      = shmem_listxattr,
3362 #endif
3363 };
3364
3365 static struct dentry *shmem_get_parent(struct dentry *child)
3366 {
3367         return ERR_PTR(-ESTALE);
3368 }
3369
3370 static int shmem_match(struct inode *ino, void *vfh)
3371 {
3372         __u32 *fh = vfh;
3373         __u64 inum = fh[2];
3374         inum = (inum << 32) | fh[1];
3375         return ino->i_ino == inum && fh[0] == ino->i_generation;
3376 }
3377
3378 /* Find any alias of inode, but prefer a hashed alias */
3379 static struct dentry *shmem_find_alias(struct inode *inode)
3380 {
3381         struct dentry *alias = d_find_alias(inode);
3382
3383         return alias ?: d_find_any_alias(inode);
3384 }
3385
3386
3387 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3388                 struct fid *fid, int fh_len, int fh_type)
3389 {
3390         struct inode *inode;
3391         struct dentry *dentry = NULL;
3392         u64 inum;
3393
3394         if (fh_len < 3)
3395                 return NULL;
3396
3397         inum = fid->raw[2];
3398         inum = (inum << 32) | fid->raw[1];
3399
3400         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3401                         shmem_match, fid->raw);
3402         if (inode) {
3403                 dentry = shmem_find_alias(inode);
3404                 iput(inode);
3405         }
3406
3407         return dentry;
3408 }
3409
3410 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3411                                 struct inode *parent)
3412 {
3413         if (*len < 3) {
3414                 *len = 3;
3415                 return FILEID_INVALID;
3416         }
3417
3418         if (inode_unhashed(inode)) {
3419                 /* Unfortunately insert_inode_hash is not idempotent,
3420                  * so as we hash inodes here rather than at creation
3421                  * time, we need a lock to ensure we only try
3422                  * to do it once
3423                  */
3424                 static DEFINE_SPINLOCK(lock);
3425                 spin_lock(&lock);
3426                 if (inode_unhashed(inode))
3427                         __insert_inode_hash(inode,
3428                                             inode->i_ino + inode->i_generation);
3429                 spin_unlock(&lock);
3430         }
3431
3432         fh[0] = inode->i_generation;
3433         fh[1] = inode->i_ino;
3434         fh[2] = ((__u64)inode->i_ino) >> 32;
3435
3436         *len = 3;
3437         return 1;
3438 }
3439
3440 static const struct export_operations shmem_export_ops = {
3441         .get_parent     = shmem_get_parent,
3442         .encode_fh      = shmem_encode_fh,
3443         .fh_to_dentry   = shmem_fh_to_dentry,
3444 };
3445
3446 enum shmem_param {
3447         Opt_gid,
3448         Opt_huge,
3449         Opt_mode,
3450         Opt_mpol,
3451         Opt_nr_blocks,
3452         Opt_nr_inodes,
3453         Opt_size,
3454         Opt_uid,
3455         Opt_inode32,
3456         Opt_inode64,
3457 };
3458
3459 static const struct constant_table shmem_param_enums_huge[] = {
3460         {"never",       SHMEM_HUGE_NEVER },
3461         {"always",      SHMEM_HUGE_ALWAYS },
3462         {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3463         {"advise",      SHMEM_HUGE_ADVISE },
3464         {}
3465 };
3466
3467 const struct fs_parameter_spec shmem_fs_parameters[] = {
3468         fsparam_u32   ("gid",           Opt_gid),
3469         fsparam_enum  ("huge",          Opt_huge,  shmem_param_enums_huge),
3470         fsparam_u32oct("mode",          Opt_mode),
3471         fsparam_string("mpol",          Opt_mpol),
3472         fsparam_string("nr_blocks",     Opt_nr_blocks),
3473         fsparam_string("nr_inodes",     Opt_nr_inodes),
3474         fsparam_string("size",          Opt_size),
3475         fsparam_u32   ("uid",           Opt_uid),
3476         fsparam_flag  ("inode32",       Opt_inode32),
3477         fsparam_flag  ("inode64",       Opt_inode64),
3478         {}
3479 };
3480
3481 static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3482 {
3483         struct shmem_options *ctx = fc->fs_private;
3484         struct fs_parse_result result;
3485         unsigned long long size;
3486         char *rest;
3487         int opt;
3488         kuid_t kuid;
3489         kgid_t kgid;
3490
3491         opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3492         if (opt < 0)
3493                 return opt;
3494
3495         switch (opt) {
3496         case Opt_size:
3497                 size = memparse(param->string, &rest);
3498                 if (*rest == '%') {
3499                         size <<= PAGE_SHIFT;
3500                         size *= totalram_pages();
3501                         do_div(size, 100);
3502                         rest++;
3503                 }
3504                 if (*rest)
3505                         goto bad_value;
3506                 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3507                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3508                 break;
3509         case Opt_nr_blocks:
3510                 ctx->blocks = memparse(param->string, &rest);
3511                 if (*rest || ctx->blocks > S64_MAX)
3512                         goto bad_value;
3513                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3514                 break;
3515         case Opt_nr_inodes:
3516                 ctx->inodes = memparse(param->string, &rest);
3517                 if (*rest)
3518                         goto bad_value;
3519                 ctx->seen |= SHMEM_SEEN_INODES;
3520                 break;
3521         case Opt_mode:
3522                 ctx->mode = result.uint_32 & 07777;
3523                 break;
3524         case Opt_uid:
3525                 kuid = make_kuid(current_user_ns(), result.uint_32);
3526                 if (!uid_valid(kuid))
3527                         goto bad_value;
3528
3529                 /*
3530                  * The requested uid must be representable in the
3531                  * filesystem's idmapping.
3532                  */
3533                 if (!kuid_has_mapping(fc->user_ns, kuid))
3534                         goto bad_value;
3535
3536                 ctx->uid = kuid;
3537                 break;
3538         case Opt_gid:
3539                 kgid = make_kgid(current_user_ns(), result.uint_32);
3540                 if (!gid_valid(kgid))
3541                         goto bad_value;
3542
3543                 /*
3544                  * The requested gid must be representable in the
3545                  * filesystem's idmapping.
3546                  */
3547                 if (!kgid_has_mapping(fc->user_ns, kgid))
3548                         goto bad_value;
3549
3550                 ctx->gid = kgid;
3551                 break;
3552         case Opt_huge:
3553                 ctx->huge = result.uint_32;
3554                 if (ctx->huge != SHMEM_HUGE_NEVER &&
3555                     !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3556                       has_transparent_hugepage()))
3557                         goto unsupported_parameter;
3558                 ctx->seen |= SHMEM_SEEN_HUGE;
3559                 break;
3560         case Opt_mpol:
3561                 if (IS_ENABLED(CONFIG_NUMA)) {
3562                         mpol_put(ctx->mpol);
3563                         ctx->mpol = NULL;
3564                         if (mpol_parse_str(param->string, &ctx->mpol))
3565                                 goto bad_value;
3566                         break;
3567                 }
3568                 goto unsupported_parameter;
3569         case Opt_inode32:
3570                 ctx->full_inums = false;
3571                 ctx->seen |= SHMEM_SEEN_INUMS;
3572                 break;
3573         case Opt_inode64:
3574                 if (sizeof(ino_t) < 8) {
3575                         return invalfc(fc,
3576                                        "Cannot use inode64 with <64bit inums in kernel\n");
3577                 }
3578                 ctx->full_inums = true;
3579                 ctx->seen |= SHMEM_SEEN_INUMS;
3580                 break;
3581         }
3582         return 0;
3583
3584 unsupported_parameter:
3585         return invalfc(fc, "Unsupported parameter '%s'", param->key);
3586 bad_value:
3587         return invalfc(fc, "Bad value for '%s'", param->key);
3588 }
3589
3590 static int shmem_parse_options(struct fs_context *fc, void *data)
3591 {
3592         char *options = data;
3593
3594         if (options) {
3595                 int err = security_sb_eat_lsm_opts(options, &fc->security);
3596                 if (err)
3597                         return err;
3598         }
3599
3600         while (options != NULL) {
3601                 char *this_char = options;
3602                 for (;;) {
3603                         /*
3604                          * NUL-terminate this option: unfortunately,
3605                          * mount options form a comma-separated list,
3606                          * but mpol's nodelist may also contain commas.
3607                          */
3608                         options = strchr(options, ',');
3609                         if (options == NULL)
3610                                 break;
3611                         options++;
3612                         if (!isdigit(*options)) {
3613                                 options[-1] = '\0';
3614                                 break;
3615                         }
3616                 }
3617                 if (*this_char) {
3618                         char *value = strchr(this_char, '=');
3619                         size_t len = 0;
3620                         int err;
3621
3622                         if (value) {
3623                                 *value++ = '\0';
3624                                 len = strlen(value);
3625                         }
3626                         err = vfs_parse_fs_string(fc, this_char, value, len);
3627                         if (err < 0)
3628                                 return err;
3629                 }
3630         }
3631         return 0;
3632 }
3633
3634 /*
3635  * Reconfigure a shmem filesystem.
3636  *
3637  * Note that we disallow change from limited->unlimited blocks/inodes while any
3638  * are in use; but we must separately disallow unlimited->limited, because in
3639  * that case we have no record of how much is already in use.
3640  */
3641 static int shmem_reconfigure(struct fs_context *fc)
3642 {
3643         struct shmem_options *ctx = fc->fs_private;
3644         struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3645         unsigned long inodes;
3646         struct mempolicy *mpol = NULL;
3647         const char *err;
3648
3649         raw_spin_lock(&sbinfo->stat_lock);
3650         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3651
3652         if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3653                 if (!sbinfo->max_blocks) {
3654                         err = "Cannot retroactively limit size";
3655                         goto out;
3656                 }
3657                 if (percpu_counter_compare(&sbinfo->used_blocks,
3658                                            ctx->blocks) > 0) {
3659                         err = "Too small a size for current use";
3660                         goto out;
3661                 }
3662         }
3663         if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3664                 if (!sbinfo->max_inodes) {
3665                         err = "Cannot retroactively limit inodes";
3666                         goto out;
3667                 }
3668                 if (ctx->inodes < inodes) {
3669                         err = "Too few inodes for current use";
3670                         goto out;
3671                 }
3672         }
3673
3674         if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3675             sbinfo->next_ino > UINT_MAX) {
3676                 err = "Current inum too high to switch to 32-bit inums";
3677                 goto out;
3678         }
3679
3680         if (ctx->seen & SHMEM_SEEN_HUGE)
3681                 sbinfo->huge = ctx->huge;
3682         if (ctx->seen & SHMEM_SEEN_INUMS)
3683                 sbinfo->full_inums = ctx->full_inums;
3684         if (ctx->seen & SHMEM_SEEN_BLOCKS)
3685                 sbinfo->max_blocks  = ctx->blocks;
3686         if (ctx->seen & SHMEM_SEEN_INODES) {
3687                 sbinfo->max_inodes  = ctx->inodes;
3688                 sbinfo->free_inodes = ctx->inodes - inodes;
3689         }
3690
3691         /*
3692          * Preserve previous mempolicy unless mpol remount option was specified.
3693          */
3694         if (ctx->mpol) {
3695                 mpol = sbinfo->mpol;
3696                 sbinfo->mpol = ctx->mpol;       /* transfers initial ref */
3697                 ctx->mpol = NULL;
3698         }
3699         raw_spin_unlock(&sbinfo->stat_lock);
3700         mpol_put(mpol);
3701         return 0;
3702 out:
3703         raw_spin_unlock(&sbinfo->stat_lock);
3704         return invalfc(fc, "%s", err);
3705 }
3706
3707 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3708 {
3709         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3710
3711         if (sbinfo->max_blocks != shmem_default_max_blocks())
3712                 seq_printf(seq, ",size=%luk",
3713                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3714         if (sbinfo->max_inodes != shmem_default_max_inodes())
3715                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3716         if (sbinfo->mode != (0777 | S_ISVTX))
3717                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3718         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3719                 seq_printf(seq, ",uid=%u",
3720                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3721         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3722                 seq_printf(seq, ",gid=%u",
3723                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3724
3725         /*
3726          * Showing inode{64,32} might be useful even if it's the system default,
3727          * since then people don't have to resort to checking both here and
3728          * /proc/config.gz to confirm 64-bit inums were successfully applied
3729          * (which may not even exist if IKCONFIG_PROC isn't enabled).
3730          *
3731          * We hide it when inode64 isn't the default and we are using 32-bit
3732          * inodes, since that probably just means the feature isn't even under
3733          * consideration.
3734          *
3735          * As such:
3736          *
3737          *                     +-----------------+-----------------+
3738          *                     | TMPFS_INODE64=y | TMPFS_INODE64=n |
3739          *  +------------------+-----------------+-----------------+
3740          *  | full_inums=true  | show            | show            |
3741          *  | full_inums=false | show            | hide            |
3742          *  +------------------+-----------------+-----------------+
3743          *
3744          */
3745         if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3746                 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3747 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3748         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3749         if (sbinfo->huge)
3750                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3751 #endif
3752         shmem_show_mpol(seq, sbinfo->mpol);
3753         return 0;
3754 }
3755
3756 #endif /* CONFIG_TMPFS */
3757
3758 static void shmem_put_super(struct super_block *sb)
3759 {
3760         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3761
3762         free_percpu(sbinfo->ino_batch);
3763         percpu_counter_destroy(&sbinfo->used_blocks);
3764         mpol_put(sbinfo->mpol);
3765         kfree(sbinfo);
3766         sb->s_fs_info = NULL;
3767 }
3768
3769 static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3770 {
3771         struct shmem_options *ctx = fc->fs_private;
3772         struct inode *inode;
3773         struct shmem_sb_info *sbinfo;
3774
3775         /* Round up to L1_CACHE_BYTES to resist false sharing */
3776         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3777                                 L1_CACHE_BYTES), GFP_KERNEL);
3778         if (!sbinfo)
3779                 return -ENOMEM;
3780
3781         sb->s_fs_info = sbinfo;
3782
3783 #ifdef CONFIG_TMPFS
3784         /*
3785          * Per default we only allow half of the physical ram per
3786          * tmpfs instance, limiting inodes to one per page of lowmem;
3787          * but the internal instance is left unlimited.
3788          */
3789         if (!(sb->s_flags & SB_KERNMOUNT)) {
3790                 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3791                         ctx->blocks = shmem_default_max_blocks();
3792                 if (!(ctx->seen & SHMEM_SEEN_INODES))
3793                         ctx->inodes = shmem_default_max_inodes();
3794                 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3795                         ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3796         } else {
3797                 sb->s_flags |= SB_NOUSER;
3798         }
3799         sb->s_export_op = &shmem_export_ops;
3800         sb->s_flags |= SB_NOSEC | SB_I_VERSION;
3801 #else
3802         sb->s_flags |= SB_NOUSER;
3803 #endif
3804         sbinfo->max_blocks = ctx->blocks;
3805         sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3806         if (sb->s_flags & SB_KERNMOUNT) {
3807                 sbinfo->ino_batch = alloc_percpu(ino_t);
3808                 if (!sbinfo->ino_batch)
3809                         goto failed;
3810         }
3811         sbinfo->uid = ctx->uid;
3812         sbinfo->gid = ctx->gid;
3813         sbinfo->full_inums = ctx->full_inums;
3814         sbinfo->mode = ctx->mode;
3815         sbinfo->huge = ctx->huge;
3816         sbinfo->mpol = ctx->mpol;
3817         ctx->mpol = NULL;
3818
3819         raw_spin_lock_init(&sbinfo->stat_lock);
3820         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3821                 goto failed;
3822         spin_lock_init(&sbinfo->shrinklist_lock);
3823         INIT_LIST_HEAD(&sbinfo->shrinklist);
3824
3825         sb->s_maxbytes = MAX_LFS_FILESIZE;
3826         sb->s_blocksize = PAGE_SIZE;
3827         sb->s_blocksize_bits = PAGE_SHIFT;
3828         sb->s_magic = TMPFS_MAGIC;
3829         sb->s_op = &shmem_ops;
3830         sb->s_time_gran = 1;
3831 #ifdef CONFIG_TMPFS_XATTR
3832         sb->s_xattr = shmem_xattr_handlers;
3833 #endif
3834 #ifdef CONFIG_TMPFS_POSIX_ACL
3835         sb->s_flags |= SB_POSIXACL;
3836 #endif
3837         uuid_gen(&sb->s_uuid);
3838
3839         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3840         if (!inode)
3841                 goto failed;
3842         inode->i_uid = sbinfo->uid;
3843         inode->i_gid = sbinfo->gid;
3844         sb->s_root = d_make_root(inode);
3845         if (!sb->s_root)
3846                 goto failed;
3847         return 0;
3848
3849 failed:
3850         shmem_put_super(sb);
3851         return -ENOMEM;
3852 }
3853
3854 static int shmem_get_tree(struct fs_context *fc)
3855 {
3856         return get_tree_nodev(fc, shmem_fill_super);
3857 }
3858
3859 static void shmem_free_fc(struct fs_context *fc)
3860 {
3861         struct shmem_options *ctx = fc->fs_private;
3862
3863         if (ctx) {
3864                 mpol_put(ctx->mpol);
3865                 kfree(ctx);
3866         }
3867 }
3868
3869 static const struct fs_context_operations shmem_fs_context_ops = {
3870         .free                   = shmem_free_fc,
3871         .get_tree               = shmem_get_tree,
3872 #ifdef CONFIG_TMPFS
3873         .parse_monolithic       = shmem_parse_options,
3874         .parse_param            = shmem_parse_one,
3875         .reconfigure            = shmem_reconfigure,
3876 #endif
3877 };
3878
3879 static struct kmem_cache *shmem_inode_cachep;
3880
3881 static struct inode *shmem_alloc_inode(struct super_block *sb)
3882 {
3883         struct shmem_inode_info *info;
3884         info = alloc_inode_sb(sb, shmem_inode_cachep, GFP_KERNEL);
3885         if (!info)
3886                 return NULL;
3887         return &info->vfs_inode;
3888 }
3889
3890 static void shmem_free_in_core_inode(struct inode *inode)
3891 {
3892         if (S_ISLNK(inode->i_mode))
3893                 kfree(inode->i_link);
3894         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3895 }
3896
3897 static void shmem_destroy_inode(struct inode *inode)
3898 {
3899         if (S_ISREG(inode->i_mode))
3900                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3901 }
3902
3903 static void shmem_init_inode(void *foo)
3904 {
3905         struct shmem_inode_info *info = foo;
3906         inode_init_once(&info->vfs_inode);
3907 }
3908
3909 static void shmem_init_inodecache(void)
3910 {
3911         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3912                                 sizeof(struct shmem_inode_info),
3913                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3914 }
3915
3916 static void shmem_destroy_inodecache(void)
3917 {
3918         kmem_cache_destroy(shmem_inode_cachep);
3919 }
3920
3921 /* Keep the page in page cache instead of truncating it */
3922 static int shmem_error_remove_page(struct address_space *mapping,
3923                                    struct page *page)
3924 {
3925         return 0;
3926 }
3927
3928 const struct address_space_operations shmem_aops = {
3929         .writepage      = shmem_writepage,
3930         .dirty_folio    = noop_dirty_folio,
3931 #ifdef CONFIG_TMPFS
3932         .write_begin    = shmem_write_begin,
3933         .write_end      = shmem_write_end,
3934 #endif
3935 #ifdef CONFIG_MIGRATION
3936         .migrate_folio  = migrate_folio,
3937 #endif
3938         .error_remove_page = shmem_error_remove_page,
3939 };
3940 EXPORT_SYMBOL(shmem_aops);
3941
3942 static const struct file_operations shmem_file_operations = {
3943         .mmap           = shmem_mmap,
3944         .get_unmapped_area = shmem_get_unmapped_area,
3945 #ifdef CONFIG_TMPFS
3946         .llseek         = shmem_file_llseek,
3947         .read_iter      = shmem_file_read_iter,
3948         .write_iter     = generic_file_write_iter,
3949         .fsync          = noop_fsync,
3950         .splice_read    = generic_file_splice_read,
3951         .splice_write   = iter_file_splice_write,
3952         .fallocate      = shmem_fallocate,
3953 #endif
3954 };
3955
3956 static const struct inode_operations shmem_inode_operations = {
3957         .getattr        = shmem_getattr,
3958         .setattr        = shmem_setattr,
3959 #ifdef CONFIG_TMPFS_XATTR
3960         .listxattr      = shmem_listxattr,
3961         .set_acl        = simple_set_acl,
3962         .fileattr_get   = shmem_fileattr_get,
3963         .fileattr_set   = shmem_fileattr_set,
3964 #endif
3965 };
3966
3967 static const struct inode_operations shmem_dir_inode_operations = {
3968 #ifdef CONFIG_TMPFS
3969         .getattr        = shmem_getattr,
3970         .create         = shmem_create,
3971         .lookup         = simple_lookup,
3972         .link           = shmem_link,
3973         .unlink         = shmem_unlink,
3974         .symlink        = shmem_symlink,
3975         .mkdir          = shmem_mkdir,
3976         .rmdir          = shmem_rmdir,
3977         .mknod          = shmem_mknod,
3978         .rename         = shmem_rename2,
3979         .tmpfile        = shmem_tmpfile,
3980 #endif
3981 #ifdef CONFIG_TMPFS_XATTR
3982         .listxattr      = shmem_listxattr,
3983         .fileattr_get   = shmem_fileattr_get,
3984         .fileattr_set   = shmem_fileattr_set,
3985 #endif
3986 #ifdef CONFIG_TMPFS_POSIX_ACL
3987         .setattr        = shmem_setattr,
3988         .set_acl        = simple_set_acl,
3989 #endif
3990 };
3991
3992 static const struct inode_operations shmem_special_inode_operations = {
3993         .getattr        = shmem_getattr,
3994 #ifdef CONFIG_TMPFS_XATTR
3995         .listxattr      = shmem_listxattr,
3996 #endif
3997 #ifdef CONFIG_TMPFS_POSIX_ACL
3998         .setattr        = shmem_setattr,
3999         .set_acl        = simple_set_acl,
4000 #endif
4001 };
4002
4003 static const struct super_operations shmem_ops = {
4004         .alloc_inode    = shmem_alloc_inode,
4005         .free_inode     = shmem_free_in_core_inode,
4006         .destroy_inode  = shmem_destroy_inode,
4007 #ifdef CONFIG_TMPFS
4008         .statfs         = shmem_statfs,
4009         .show_options   = shmem_show_options,
4010 #endif
4011         .evict_inode    = shmem_evict_inode,
4012         .drop_inode     = generic_delete_inode,
4013         .put_super      = shmem_put_super,
4014 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4015         .nr_cached_objects      = shmem_unused_huge_count,
4016         .free_cached_objects    = shmem_unused_huge_scan,
4017 #endif
4018 };
4019
4020 static const struct vm_operations_struct shmem_vm_ops = {
4021         .fault          = shmem_fault,
4022         .map_pages      = filemap_map_pages,
4023 #ifdef CONFIG_NUMA
4024         .set_policy     = shmem_set_policy,
4025         .get_policy     = shmem_get_policy,
4026 #endif
4027 };
4028
4029 int shmem_init_fs_context(struct fs_context *fc)
4030 {
4031         struct shmem_options *ctx;
4032
4033         ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
4034         if (!ctx)
4035                 return -ENOMEM;
4036
4037         ctx->mode = 0777 | S_ISVTX;
4038         ctx->uid = current_fsuid();
4039         ctx->gid = current_fsgid();
4040
4041         fc->fs_private = ctx;
4042         fc->ops = &shmem_fs_context_ops;
4043         return 0;
4044 }
4045
4046 static struct file_system_type shmem_fs_type = {
4047         .owner          = THIS_MODULE,
4048         .name           = "tmpfs",
4049         .init_fs_context = shmem_init_fs_context,
4050 #ifdef CONFIG_TMPFS
4051         .parameters     = shmem_fs_parameters,
4052 #endif
4053         .kill_sb        = kill_litter_super,
4054         .fs_flags       = FS_USERNS_MOUNT,
4055 };
4056
4057 void __init shmem_init(void)
4058 {
4059         int error;
4060
4061         shmem_init_inodecache();
4062
4063         error = register_filesystem(&shmem_fs_type);
4064         if (error) {
4065                 pr_err("Could not register tmpfs\n");
4066                 goto out2;
4067         }
4068
4069         shm_mnt = kern_mount(&shmem_fs_type);
4070         if (IS_ERR(shm_mnt)) {
4071                 error = PTR_ERR(shm_mnt);
4072                 pr_err("Could not kern_mount tmpfs\n");
4073                 goto out1;
4074         }
4075
4076 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4077         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4078                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4079         else
4080                 shmem_huge = SHMEM_HUGE_NEVER; /* just in case it was patched */
4081 #endif
4082         return;
4083
4084 out1:
4085         unregister_filesystem(&shmem_fs_type);
4086 out2:
4087         shmem_destroy_inodecache();
4088         shm_mnt = ERR_PTR(error);
4089 }
4090
4091 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4092 static ssize_t shmem_enabled_show(struct kobject *kobj,
4093                                   struct kobj_attribute *attr, char *buf)
4094 {
4095         static const int values[] = {
4096                 SHMEM_HUGE_ALWAYS,
4097                 SHMEM_HUGE_WITHIN_SIZE,
4098                 SHMEM_HUGE_ADVISE,
4099                 SHMEM_HUGE_NEVER,
4100                 SHMEM_HUGE_DENY,
4101                 SHMEM_HUGE_FORCE,
4102         };
4103         int len = 0;
4104         int i;
4105
4106         for (i = 0; i < ARRAY_SIZE(values); i++) {
4107                 len += sysfs_emit_at(buf, len,
4108                                      shmem_huge == values[i] ? "%s[%s]" : "%s%s",
4109                                      i ? " " : "",
4110                                      shmem_format_huge(values[i]));
4111         }
4112
4113         len += sysfs_emit_at(buf, len, "\n");
4114
4115         return len;
4116 }
4117
4118 static ssize_t shmem_enabled_store(struct kobject *kobj,
4119                 struct kobj_attribute *attr, const char *buf, size_t count)
4120 {
4121         char tmp[16];
4122         int huge;
4123
4124         if (count + 1 > sizeof(tmp))
4125                 return -EINVAL;
4126         memcpy(tmp, buf, count);
4127         tmp[count] = '\0';
4128         if (count && tmp[count - 1] == '\n')
4129                 tmp[count - 1] = '\0';
4130
4131         huge = shmem_parse_huge(tmp);
4132         if (huge == -EINVAL)
4133                 return -EINVAL;
4134         if (!has_transparent_hugepage() &&
4135                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4136                 return -EINVAL;
4137
4138         shmem_huge = huge;
4139         if (shmem_huge > SHMEM_HUGE_DENY)
4140                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4141         return count;
4142 }
4143
4144 struct kobj_attribute shmem_enabled_attr = __ATTR_RW(shmem_enabled);
4145 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4146
4147 #else /* !CONFIG_SHMEM */
4148
4149 /*
4150  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4151  *
4152  * This is intended for small system where the benefits of the full
4153  * shmem code (swap-backed and resource-limited) are outweighed by
4154  * their complexity. On systems without swap this code should be
4155  * effectively equivalent, but much lighter weight.
4156  */
4157
4158 static struct file_system_type shmem_fs_type = {
4159         .name           = "tmpfs",
4160         .init_fs_context = ramfs_init_fs_context,
4161         .parameters     = ramfs_fs_parameters,
4162         .kill_sb        = ramfs_kill_sb,
4163         .fs_flags       = FS_USERNS_MOUNT,
4164 };
4165
4166 void __init shmem_init(void)
4167 {
4168         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4169
4170         shm_mnt = kern_mount(&shmem_fs_type);
4171         BUG_ON(IS_ERR(shm_mnt));
4172 }
4173
4174 int shmem_unuse(unsigned int type)
4175 {
4176         return 0;
4177 }
4178
4179 int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
4180 {
4181         return 0;
4182 }
4183
4184 void shmem_unlock_mapping(struct address_space *mapping)
4185 {
4186 }
4187
4188 #ifdef CONFIG_MMU
4189 unsigned long shmem_get_unmapped_area(struct file *file,
4190                                       unsigned long addr, unsigned long len,
4191                                       unsigned long pgoff, unsigned long flags)
4192 {
4193         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4194 }
4195 #endif
4196
4197 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4198 {
4199         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4200 }
4201 EXPORT_SYMBOL_GPL(shmem_truncate_range);
4202
4203 #define shmem_vm_ops                            generic_file_vm_ops
4204 #define shmem_file_operations                   ramfs_file_operations
4205 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
4206 #define shmem_acct_size(flags, size)            0
4207 #define shmem_unacct_size(flags, size)          do {} while (0)
4208
4209 #endif /* CONFIG_SHMEM */
4210
4211 /* common code */
4212
4213 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4214                                        unsigned long flags, unsigned int i_flags)
4215 {
4216         struct inode *inode;
4217         struct file *res;
4218
4219         if (IS_ERR(mnt))
4220                 return ERR_CAST(mnt);
4221
4222         if (size < 0 || size > MAX_LFS_FILESIZE)
4223                 return ERR_PTR(-EINVAL);
4224
4225         if (shmem_acct_size(flags, size))
4226                 return ERR_PTR(-ENOMEM);
4227
4228         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4229                                 flags);
4230         if (unlikely(!inode)) {
4231                 shmem_unacct_size(flags, size);
4232                 return ERR_PTR(-ENOSPC);
4233         }
4234         inode->i_flags |= i_flags;
4235         inode->i_size = size;
4236         clear_nlink(inode);     /* It is unlinked */
4237         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4238         if (!IS_ERR(res))
4239                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4240                                 &shmem_file_operations);
4241         if (IS_ERR(res))
4242                 iput(inode);
4243         return res;
4244 }
4245
4246 /**
4247  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4248  *      kernel internal.  There will be NO LSM permission checks against the
4249  *      underlying inode.  So users of this interface must do LSM checks at a
4250  *      higher layer.  The users are the big_key and shm implementations.  LSM
4251  *      checks are provided at the key or shm level rather than the inode.
4252  * @name: name for dentry (to be seen in /proc/<pid>/maps
4253  * @size: size to be set for the file
4254  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4255  */
4256 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4257 {
4258         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4259 }
4260
4261 /**
4262  * shmem_file_setup - get an unlinked file living in tmpfs
4263  * @name: name for dentry (to be seen in /proc/<pid>/maps
4264  * @size: size to be set for the file
4265  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4266  */
4267 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4268 {
4269         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4270 }
4271 EXPORT_SYMBOL_GPL(shmem_file_setup);
4272
4273 /**
4274  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4275  * @mnt: the tmpfs mount where the file will be created
4276  * @name: name for dentry (to be seen in /proc/<pid>/maps
4277  * @size: size to be set for the file
4278  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4279  */
4280 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4281                                        loff_t size, unsigned long flags)
4282 {
4283         return __shmem_file_setup(mnt, name, size, flags, 0);
4284 }
4285 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4286
4287 /**
4288  * shmem_zero_setup - setup a shared anonymous mapping
4289  * @vma: the vma to be mmapped is prepared by do_mmap
4290  */
4291 int shmem_zero_setup(struct vm_area_struct *vma)
4292 {
4293         struct file *file;
4294         loff_t size = vma->vm_end - vma->vm_start;
4295
4296         /*
4297          * Cloning a new file under mmap_lock leads to a lock ordering conflict
4298          * between XFS directory reading and selinux: since this file is only
4299          * accessible to the user through its mapping, use S_PRIVATE flag to
4300          * bypass file security, in the same way as shmem_kernel_file_setup().
4301          */
4302         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4303         if (IS_ERR(file))
4304                 return PTR_ERR(file);
4305
4306         if (vma->vm_file)
4307                 fput(vma->vm_file);
4308         vma->vm_file = file;
4309         vma->vm_ops = &shmem_vm_ops;
4310
4311         return 0;
4312 }
4313
4314 /**
4315  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4316  * @mapping:    the page's address_space
4317  * @index:      the page index
4318  * @gfp:        the page allocator flags to use if allocating
4319  *
4320  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4321  * with any new page allocations done using the specified allocation flags.
4322  * But read_cache_page_gfp() uses the ->read_folio() method: which does not
4323  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4324  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4325  *
4326  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4327  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4328  */
4329 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4330                                          pgoff_t index, gfp_t gfp)
4331 {
4332 #ifdef CONFIG_SHMEM
4333         struct inode *inode = mapping->host;
4334         struct folio *folio;
4335         struct page *page;
4336         int error;
4337
4338         BUG_ON(!shmem_mapping(mapping));
4339         error = shmem_get_folio_gfp(inode, index, &folio, SGP_CACHE,
4340                                   gfp, NULL, NULL, NULL);
4341         if (error)
4342                 return ERR_PTR(error);
4343
4344         folio_unlock(folio);
4345         page = folio_file_page(folio, index);
4346         if (PageHWPoison(page)) {
4347                 folio_put(folio);
4348                 return ERR_PTR(-EIO);
4349         }
4350
4351         return page;
4352 #else
4353         /*
4354          * The tiny !SHMEM case uses ramfs without swap
4355          */
4356         return read_cache_page_gfp(mapping, index, gfp);
4357 #endif
4358 }
4359 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);