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
 804         rcu_read_lock();
 805         xas_for_each(&xas, page, end - 1) {
 806                 if (xas_retry(&xas, page))
 807                         continue;
 808                 if (xa_is_value(page))
 809                         swapped++;
 810
 811                 if (need_resched()) {
 812                         xas_pause(&xas);
 813                         cond_resched_rcu();
 814                 }
 815         }
 816
 817         rcu_read_unlock();
 818
 819         return swapped << PAGE_SHIFT;
 820 }
 821
 822 /*
 823  * Determine (in bytes) how many of the shmem object's pages mapped by the
 824  * given vma is swapped out.
 825  *
 826  * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
 827  * as long as the inode doesn't go away and racy results are not a problem.
 828  */
 829 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
 830 {
 831         struct inode *inode = file_inode(vma->vm_file);
 832         struct shmem_inode_info *info = SHMEM_I(inode);
 833         struct address_space *mapping = inode->i_mapping;
 834         unsigned long swapped;
 835
 836         /* Be careful as we don't hold info->lock */
 837         swapped = READ_ONCE(info->swapped);
 838
 839         /*
 840          * The easier cases are when the shmem object has nothing in swap, or
 841          * the vma maps it whole. Then we can simply use the stats that we
 842          * already track.
 843          */
 844         if (!swapped)
 845                 return 0;
 846
 847         if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
 848                 return swapped << PAGE_SHIFT;
 849
 850         /* Here comes the more involved part */
 851         return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
 852                                         vma->vm_pgoff + vma_pages(vma));
 853 }
 854
 855 /*
 856  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 857  */
 858 void shmem_unlock_mapping(struct address_space *mapping)
 859 {
 860         struct folio_batch fbatch;
 861         pgoff_t index = 0;
 862
 863         folio_batch_init(&fbatch);
 864         /*
 865          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 866          */
 867         while (!mapping_unevictable(mapping) &&
 868                filemap_get_folios(mapping, &index, ~0UL, &fbatch)) {
 869                 check_move_unevictable_folios(&fbatch);
 870                 folio_batch_release(&fbatch);
 871                 cond_resched();
 872         }
 873 }
 874
 875 static struct folio *shmem_get_partial_folio(struct inode *inode, pgoff_t index)
 876 {
 877         struct folio *folio;
 878
 879         /*
 880          * At first avoid shmem_get_folio(,,,SGP_READ): that fails
 881          * beyond i_size, and reports fallocated pages as holes.
 882          */
 883         folio = __filemap_get_folio(inode->i_mapping, index,
 884                                         FGP_ENTRY | FGP_LOCK, 0);
 885         if (!xa_is_value(folio))
 886                 return folio;
 887         /*
 888          * But read a page back from swap if any of it is within i_size
 889          * (although in some cases this is just a waste of time).
 890          */
 891         folio = NULL;
 892         shmem_get_folio(inode, index, &folio, SGP_READ);
 893         return folio;
 894 }
 895
 896 /*
 897  * Remove range of pages and swap entries from page cache, and free them.
 898  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 899  */
 900 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 901                                                                  bool unfalloc)
 902 {
 903         struct address_space *mapping = inode->i_mapping;
 904         struct shmem_inode_info *info = SHMEM_I(inode);
 905         pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
 906         pgoff_t end = (lend + 1) >> PAGE_SHIFT;
 907         struct folio_batch fbatch;
 908         pgoff_t indices[PAGEVEC_SIZE];
 909         struct folio *folio;
 910         bool same_folio;
 911         long nr_swaps_freed = 0;
 912         pgoff_t index;
 913         int i;
 914
 915         if (lend == -1)
 916                 end = -1;       /* unsigned, so actually very big */
 917
 918         if (info->fallocend > start && info->fallocend <= end && !unfalloc)
 919                 info->fallocend = start;
 920
 921         folio_batch_init(&fbatch);
 922         index = start;
 923         while (index < end && find_lock_entries(mapping, index, end - 1,
 924                         &fbatch, indices)) {
 925                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 926                         folio = fbatch.folios[i];
 927
 928                         index = indices[i];
 929
 930                         if (xa_is_value(folio)) {
 931                                 if (unfalloc)
 932                                         continue;
 933                                 nr_swaps_freed += !shmem_free_swap(mapping,
 934                                                                 index, folio);
 935                                 continue;
 936                         }
 937                         index += folio_nr_pages(folio) - 1;
 938
 939                         if (!unfalloc || !folio_test_uptodate(folio))
 940                                 truncate_inode_folio(mapping, folio);
 941                         folio_unlock(folio);
 942                 }
 943                 folio_batch_remove_exceptionals(&fbatch);
 944                 folio_batch_release(&fbatch);
 945                 cond_resched();
 946                 index++;
 947         }
 948
 949         /*
 950          * When undoing a failed fallocate, we want none of the partial folio
 951          * zeroing and splitting below, but shall want to truncate the whole
 952          * folio when !uptodate indicates that it was added by this fallocate,
 953          * even when [lstart, lend] covers only a part of the folio.
 954          */
 955         if (unfalloc)
 956                 goto whole_folios;
 957
 958         same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
 959         folio = shmem_get_partial_folio(inode, lstart >> PAGE_SHIFT);
 960         if (folio) {
 961                 same_folio = lend < folio_pos(folio) + folio_size(folio);
 962                 folio_mark_dirty(folio);
 963                 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
 964                         start = folio->index + folio_nr_pages(folio);
 965                         if (same_folio)
 966                                 end = folio->index;
 967                 }
 968                 folio_unlock(folio);
 969                 folio_put(folio);
 970                 folio = NULL;
 971         }
 972
 973         if (!same_folio)
 974                 folio = shmem_get_partial_folio(inode, lend >> PAGE_SHIFT);
 975         if (folio) {
 976                 folio_mark_dirty(folio);
 977                 if (!truncate_inode_partial_folio(folio, lstart, lend))
 978                         end = folio->index;
 979                 folio_unlock(folio);
 980                 folio_put(folio);
 981         }
 982
 983 whole_folios:
 984
 985         index = start;
 986         while (index < end) {
 987                 cond_resched();
 988
 989                 if (!find_get_entries(mapping, index, end - 1, &fbatch,
 990                                 indices)) {
 991                         /* If all gone or hole-punch or unfalloc, we're done */
 992                         if (index == start || end != -1)
 993                                 break;
 994                         /* But if truncating, restart to make sure all gone */
 995                         index = start;
 996                         continue;
 997                 }
 998                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 999                         folio = fbatch.folios[i];
1000
1001                         index = indices[i];
1002                         if (xa_is_value(folio)) {
1003                                 if (unfalloc)
1004                                         continue;
1005                                 if (shmem_free_swap(mapping, index, folio)) {
1006                                         /* Swap was replaced by page: retry */
1007                                         index--;
1008                                         break;
1009                                 }
1010                                 nr_swaps_freed++;
1011                                 continue;
1012                         }
1013
1014                         folio_lock(folio);
1015
1016                         if (!unfalloc || !folio_test_uptodate(folio)) {
1017                                 if (folio_mapping(folio) != mapping) {
1018                                         /* Page was replaced by swap: retry */
1019                                         folio_unlock(folio);
1020                                         index--;
1021                                         break;
1022                                 }
1023                                 VM_BUG_ON_FOLIO(folio_test_writeback(folio),
1024                                                 folio);
1025                                 truncate_inode_folio(mapping, folio);
1026                         }
1027                         index = folio->index + folio_nr_pages(folio) - 1;
1028                         folio_unlock(folio);
1029                 }
1030                 folio_batch_remove_exceptionals(&fbatch);
1031                 folio_batch_release(&fbatch);
1032                 index++;
1033         }
1034
1035         spin_lock_irq(&info->lock);
1036         info->swapped -= nr_swaps_freed;
1037         shmem_recalc_inode(inode);
1038         spin_unlock_irq(&info->lock);
1039 }
1040
1041 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1042 {
1043         shmem_undo_range(inode, lstart, lend, false);
1044         inode->i_ctime = inode->i_mtime = current_time(inode);
1045         inode_inc_iversion(inode);
1046 }
1047 EXPORT_SYMBOL_GPL(shmem_truncate_range);
1048
1049 static int shmem_getattr(struct user_namespace *mnt_userns,
1050                          const struct path *path, struct kstat *stat,
1051                          u32 request_mask, unsigned int query_flags)
1052 {
1053         struct inode *inode = path->dentry->d_inode;
1054         struct shmem_inode_info *info = SHMEM_I(inode);
1055
1056         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1057                 spin_lock_irq(&info->lock);
1058                 shmem_recalc_inode(inode);
1059                 spin_unlock_irq(&info->lock);
1060         }
1061         if (info->fsflags & FS_APPEND_FL)
1062                 stat->attributes |= STATX_ATTR_APPEND;
1063         if (info->fsflags & FS_IMMUTABLE_FL)
1064                 stat->attributes |= STATX_ATTR_IMMUTABLE;
1065         if (info->fsflags & FS_NODUMP_FL)
1066                 stat->attributes |= STATX_ATTR_NODUMP;
1067         stat->attributes_mask |= (STATX_ATTR_APPEND |
1068                         STATX_ATTR_IMMUTABLE |
1069                         STATX_ATTR_NODUMP);
1070         generic_fillattr(&init_user_ns, inode, stat);
1071
1072         if (shmem_is_huge(NULL, inode, 0, false))
1073                 stat->blksize = HPAGE_PMD_SIZE;
1074
1075         if (request_mask & STATX_BTIME) {
1076                 stat->result_mask |= STATX_BTIME;
1077                 stat->btime.tv_sec = info->i_crtime.tv_sec;
1078                 stat->btime.tv_nsec = info->i_crtime.tv_nsec;
1079         }
1080
1081         return 0;
1082 }
1083
1084 static int shmem_setattr(struct user_namespace *mnt_userns,
1085                          struct dentry *dentry, struct iattr *attr)
1086 {
1087         struct inode *inode = d_inode(dentry);
1088         struct shmem_inode_info *info = SHMEM_I(inode);
1089         int error;
1090         bool update_mtime = false;
1091         bool update_ctime = true;
1092
1093         error = setattr_prepare(&init_user_ns, dentry, attr);
1094         if (error)
1095                 return error;
1096
1097         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1098                 loff_t oldsize = inode->i_size;
1099                 loff_t newsize = attr->ia_size;
1100
1101                 /* protected by i_rwsem */
1102                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1103                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1104                         return -EPERM;
1105
1106                 if (newsize != oldsize) {
1107                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1108                                         oldsize, newsize);
1109                         if (error)
1110                                 return error;
1111                         i_size_write(inode, newsize);
1112                         update_mtime = true;
1113                 } else {
1114                         update_ctime = false;
1115                 }
1116                 if (newsize <= oldsize) {
1117                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1118                         if (oldsize > holebegin)
1119                                 unmap_mapping_range(inode->i_mapping,
1120                                                         holebegin, 0, 1);
1121                         if (info->alloced)
1122                                 shmem_truncate_range(inode,
1123                                                         newsize, (loff_t)-1);
1124                         /* unmap again to remove racily COWed private pages */
1125                         if (oldsize > holebegin)
1126                                 unmap_mapping_range(inode->i_mapping,
1127                                                         holebegin, 0, 1);
1128                 }
1129         }
1130
1131         setattr_copy(&init_user_ns, inode, attr);
1132         if (attr->ia_valid & ATTR_MODE)
1133                 error = posix_acl_chmod(&init_user_ns, inode, inode->i_mode);
1134         if (!error && update_ctime) {
1135                 inode->i_ctime = current_time(inode);
1136                 if (update_mtime)
1137                         inode->i_mtime = inode->i_ctime;
1138                 inode_inc_iversion(inode);
1139         }
1140         return error;
1141 }
1142
1143 static void shmem_evict_inode(struct inode *inode)
1144 {
1145         struct shmem_inode_info *info = SHMEM_I(inode);
1146         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1147
1148         if (shmem_mapping(inode->i_mapping)) {
1149                 shmem_unacct_size(info->flags, inode->i_size);
1150                 inode->i_size = 0;
1151                 mapping_set_exiting(inode->i_mapping);
1152                 shmem_truncate_range(inode, 0, (loff_t)-1);
1153                 if (!list_empty(&info->shrinklist)) {
1154                         spin_lock(&sbinfo->shrinklist_lock);
1155                         if (!list_empty(&info->shrinklist)) {
1156                                 list_del_init(&info->shrinklist);
1157                                 sbinfo->shrinklist_len--;
1158                         }
1159                         spin_unlock(&sbinfo->shrinklist_lock);
1160                 }
1161                 while (!list_empty(&info->swaplist)) {
1162                         /* Wait while shmem_unuse() is scanning this inode... */
1163                         wait_var_event(&info->stop_eviction,
1164                                        !atomic_read(&info->stop_eviction));
1165                         mutex_lock(&shmem_swaplist_mutex);
1166                         /* ...but beware of the race if we peeked too early */
1167                         if (!atomic_read(&info->stop_eviction))
1168                                 list_del_init(&info->swaplist);
1169                         mutex_unlock(&shmem_swaplist_mutex);
1170                 }
1171         }
1172
1173         simple_xattrs_free(&info->xattrs);
1174         WARN_ON(inode->i_blocks);
1175         shmem_free_inode(inode->i_sb);
1176         clear_inode(inode);
1177 }
1178
1179 static int shmem_find_swap_entries(struct address_space *mapping,
1180                                    pgoff_t start, struct folio_batch *fbatch,
1181                                    pgoff_t *indices, unsigned int type)
1182 {
1183         XA_STATE(xas, &mapping->i_pages, start);
1184         struct folio *folio;
1185         swp_entry_t entry;
1186
1187         rcu_read_lock();
1188         xas_for_each(&xas, folio, ULONG_MAX) {
1189                 if (xas_retry(&xas, folio))
1190                         continue;
1191
1192                 if (!xa_is_value(folio))
1193                         continue;
1194
1195                 entry = radix_to_swp_entry(folio);
1196                 /*
1197                  * swapin error entries can be found in the mapping. But they're
1198                  * deliberately ignored here as we've done everything we can do.
1199                  */
1200                 if (swp_type(entry) != type)
1201                         continue;
1202
1203                 indices[folio_batch_count(fbatch)] = xas.xa_index;
1204                 if (!folio_batch_add(fbatch, folio))
1205                         break;
1206
1207                 if (need_resched()) {
1208                         xas_pause(&xas);
1209                         cond_resched_rcu();
1210                 }
1211         }
1212         rcu_read_unlock();
1213
1214         return xas.xa_index;
1215 }
1216
1217 /*
1218  * Move the swapped pages for an inode to page cache. Returns the count
1219  * of pages swapped in, or the error in case of failure.
1220  */
1221 static int shmem_unuse_swap_entries(struct inode *inode,
1222                 struct folio_batch *fbatch, pgoff_t *indices)
1223 {
1224         int i = 0;
1225         int ret = 0;
1226         int error = 0;
1227         struct address_space *mapping = inode->i_mapping;
1228
1229         for (i = 0; i < folio_batch_count(fbatch); i++) {
1230                 struct folio *folio = fbatch->folios[i];
1231
1232                 if (!xa_is_value(folio))
1233                         continue;
1234                 error = shmem_swapin_folio(inode, indices[i],
1235                                           &folio, SGP_CACHE,
1236                                           mapping_gfp_mask(mapping),
1237                                           NULL, NULL);
1238                 if (error == 0) {
1239                         folio_unlock(folio);
1240                         folio_put(folio);
1241                         ret++;
1242                 }
1243                 if (error == -ENOMEM)
1244                         break;
1245                 error = 0;
1246         }
1247         return error ? error : ret;
1248 }
1249
1250 /*
1251  * If swap found in inode, free it and move page from swapcache to filecache.
1252  */
1253 static int shmem_unuse_inode(struct inode *inode, unsigned int type)
1254 {
1255         struct address_space *mapping = inode->i_mapping;
1256         pgoff_t start = 0;
1257         struct folio_batch fbatch;
1258         pgoff_t indices[PAGEVEC_SIZE];
1259         int ret = 0;
1260
1261         do {
1262                 folio_batch_init(&fbatch);
1263                 shmem_find_swap_entries(mapping, start, &fbatch, indices, type);
1264                 if (folio_batch_count(&fbatch) == 0) {
1265                         ret = 0;
1266                         break;
1267                 }
1268
1269                 ret = shmem_unuse_swap_entries(inode, &fbatch, indices);
1270                 if (ret < 0)
1271                         break;
1272
1273                 start = indices[folio_batch_count(&fbatch) - 1];
1274         } while (true);
1275
1276         return ret;
1277 }
1278
1279 /*
1280  * Read all the shared memory data that resides in the swap
1281  * device 'type' back into memory, so the swap device can be
1282  * unused.
1283  */
1284 int shmem_unuse(unsigned int type)
1285 {
1286         struct shmem_inode_info *info, *next;
1287         int error = 0;
1288
1289         if (list_empty(&shmem_swaplist))
1290                 return 0;
1291
1292         mutex_lock(&shmem_swaplist_mutex);
1293         list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1294                 if (!info->swapped) {
1295                         list_del_init(&info->swaplist);
1296                         continue;
1297                 }
1298                 /*
1299                  * Drop the swaplist mutex while searching the inode for swap;
1300                  * but before doing so, make sure shmem_evict_inode() will not
1301                  * remove placeholder inode from swaplist, nor let it be freed
1302                  * (igrab() would protect from unlink, but not from unmount).
1303                  */
1304                 atomic_inc(&info->stop_eviction);
1305                 mutex_unlock(&shmem_swaplist_mutex);
1306
1307                 error = shmem_unuse_inode(&info->vfs_inode, type);
1308                 cond_resched();
1309
1310                 mutex_lock(&shmem_swaplist_mutex);
1311                 next = list_next_entry(info, swaplist);
1312                 if (!info->swapped)
1313                         list_del_init(&info->swaplist);
1314                 if (atomic_dec_and_test(&info->stop_eviction))
1315                         wake_up_var(&info->stop_eviction);
1316                 if (error)
1317                         break;
1318         }
1319         mutex_unlock(&shmem_swaplist_mutex);
1320
1321         return error;
1322 }
1323
1324 /*
1325  * Move the page from the page cache to the swap cache.
1326  */
1327 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1328 {
1329         struct folio *folio = page_folio(page);
1330         struct shmem_inode_info *info;
1331         struct address_space *mapping;
1332         struct inode *inode;
1333         swp_entry_t swap;
1334         pgoff_t index;
1335
1336         /*
1337          * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "always" or
1338          * "force", drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
1339          * and its shmem_writeback() needs them to be split when swapping.
1340          */
1341         if (folio_test_large(folio)) {
1342                 /* Ensure the subpages are still dirty */
1343                 folio_test_set_dirty(folio);
1344                 if (split_huge_page(page) < 0)
1345                         goto redirty;
1346                 folio = page_folio(page);
1347                 folio_clear_dirty(folio);
1348         }
1349
1350         BUG_ON(!folio_test_locked(folio));
1351         mapping = folio->mapping;
1352         index = folio->index;
1353         inode = mapping->host;
1354         info = SHMEM_I(inode);
1355         if (info->flags & VM_LOCKED)
1356                 goto redirty;
1357         if (!total_swap_pages)
1358                 goto redirty;
1359
1360         /*
1361          * Our capabilities prevent regular writeback or sync from ever calling
1362          * shmem_writepage; but a stacking filesystem might use ->writepage of
1363          * its underlying filesystem, in which case tmpfs should write out to
1364          * swap only in response to memory pressure, and not for the writeback
1365          * threads or sync.
1366          */
1367         if (!wbc->for_reclaim) {
1368                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1369                 goto redirty;
1370         }
1371
1372         /*
1373          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1374          * value into swapfile.c, the only way we can correctly account for a
1375          * fallocated folio arriving here is now to initialize it and write it.
1376          *
1377          * That's okay for a folio already fallocated earlier, but if we have
1378          * not yet completed the fallocation, then (a) we want to keep track
1379          * of this folio in case we have to undo it, and (b) it may not be a
1380          * good idea to continue anyway, once we're pushing into swap.  So
1381          * reactivate the folio, and let shmem_fallocate() quit when too many.
1382          */
1383         if (!folio_test_uptodate(folio)) {
1384                 if (inode->i_private) {
1385                         struct shmem_falloc *shmem_falloc;
1386                         spin_lock(&inode->i_lock);
1387                         shmem_falloc = inode->i_private;
1388                         if (shmem_falloc &&
1389                             !shmem_falloc->waitq &&
1390                             index >= shmem_falloc->start &&
1391                             index < shmem_falloc->next)
1392                                 shmem_falloc->nr_unswapped++;
1393                         else
1394                                 shmem_falloc = NULL;
1395                         spin_unlock(&inode->i_lock);
1396                         if (shmem_falloc)
1397                                 goto redirty;
1398                 }
1399                 folio_zero_range(folio, 0, folio_size(folio));
1400                 flush_dcache_folio(folio);
1401                 folio_mark_uptodate(folio);
1402         }
1403
1404         swap = folio_alloc_swap(folio);
1405         if (!swap.val)
1406                 goto redirty;
1407
1408         /*
1409          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1410          * if it's not already there.  Do it now before the folio is
1411          * moved to swap cache, when its pagelock no longer protects
1412          * the inode from eviction.  But don't unlock the mutex until
1413          * we've incremented swapped, because shmem_unuse_inode() will
1414          * prune a !swapped inode from the swaplist under this mutex.
1415          */
1416         mutex_lock(&shmem_swaplist_mutex);
1417         if (list_empty(&info->swaplist))
1418                 list_add(&info->swaplist, &shmem_swaplist);
1419
1420         if (add_to_swap_cache(folio, swap,
1421                         __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1422                         NULL) == 0) {
1423                 spin_lock_irq(&info->lock);
1424                 shmem_recalc_inode(inode);
1425                 info->swapped++;
1426                 spin_unlock_irq(&info->lock);
1427
1428                 swap_shmem_alloc(swap);
1429                 shmem_delete_from_page_cache(folio, swp_to_radix_entry(swap));
1430
1431                 mutex_unlock(&shmem_swaplist_mutex);
1432                 BUG_ON(folio_mapped(folio));
1433                 swap_writepage(&folio->page, wbc);
1434                 return 0;
1435         }
1436
1437         mutex_unlock(&shmem_swaplist_mutex);
1438         put_swap_folio(folio, swap);
1439 redirty:
1440         folio_mark_dirty(folio);
1441         if (wbc->for_reclaim)
1442                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with folio locked */
1443         folio_unlock(folio);
1444         return 0;
1445 }
1446
1447 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1448 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1449 {
1450         char buffer[64];
1451
1452         if (!mpol || mpol->mode == MPOL_DEFAULT)
1453                 return;         /* show nothing */
1454
1455         mpol_to_str(buffer, sizeof(buffer), mpol);
1456
1457         seq_printf(seq, ",mpol=%s", buffer);
1458 }
1459
1460 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1461 {
1462         struct mempolicy *mpol = NULL;
1463         if (sbinfo->mpol) {
1464                 raw_spin_lock(&sbinfo->stat_lock);      /* prevent replace/use races */
1465                 mpol = sbinfo->mpol;
1466                 mpol_get(mpol);
1467                 raw_spin_unlock(&sbinfo->stat_lock);
1468         }
1469         return mpol;
1470 }
1471 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1472 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1473 {
1474 }
1475 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1476 {
1477         return NULL;
1478 }
1479 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1480 #ifndef CONFIG_NUMA
1481 #define vm_policy vm_private_data
1482 #endif
1483
1484 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1485                 struct shmem_inode_info *info, pgoff_t index)
1486 {
1487         /* Create a pseudo vma that just contains the policy */
1488         vma_init(vma, NULL);
1489         /* Bias interleave by inode number to distribute better across nodes */
1490         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1491         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1492 }
1493
1494 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1495 {
1496         /* Drop reference taken by mpol_shared_policy_lookup() */
1497         mpol_cond_put(vma->vm_policy);
1498 }
1499
1500 static struct folio *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1501                         struct shmem_inode_info *info, pgoff_t index)
1502 {
1503         struct vm_area_struct pvma;
1504         struct page *page;
1505         struct vm_fault vmf = {
1506                 .vma = &pvma,
1507         };
1508
1509         shmem_pseudo_vma_init(&pvma, info, index);
1510         page = swap_cluster_readahead(swap, gfp, &vmf);
1511         shmem_pseudo_vma_destroy(&pvma);
1512
1513         if (!page)
1514                 return NULL;
1515         return page_folio(page);
1516 }
1517
1518 /*
1519  * Make sure huge_gfp is always more limited than limit_gfp.
1520  * Some of the flags set permissions, while others set limitations.
1521  */
1522 static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t limit_gfp)
1523 {
1524         gfp_t allowflags = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
1525         gfp_t denyflags = __GFP_NOWARN | __GFP_NORETRY;
1526         gfp_t zoneflags = limit_gfp & GFP_ZONEMASK;
1527         gfp_t result = huge_gfp & ~(allowflags | GFP_ZONEMASK);
1528
1529         /* Allow allocations only from the originally specified zones. */
1530         result |= zoneflags;
1531
1532         /*
1533          * Minimize the result gfp by taking the union with the deny flags,
1534          * and the intersection of the allow flags.
1535          */
1536         result |= (limit_gfp & denyflags);
1537         result |= (huge_gfp & limit_gfp) & allowflags;
1538
1539         return result;
1540 }
1541
1542 static struct folio *shmem_alloc_hugefolio(gfp_t gfp,
1543                 struct shmem_inode_info *info, pgoff_t index)
1544 {
1545         struct vm_area_struct pvma;
1546         struct address_space *mapping = info->vfs_inode.i_mapping;
1547         pgoff_t hindex;
1548         struct folio *folio;
1549
1550         hindex = round_down(index, HPAGE_PMD_NR);
1551         if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1552                                                                 XA_PRESENT))
1553                 return NULL;
1554
1555         shmem_pseudo_vma_init(&pvma, info, hindex);
1556         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, &pvma, 0, true);
1557         shmem_pseudo_vma_destroy(&pvma);
1558         if (!folio)
1559                 count_vm_event(THP_FILE_FALLBACK);
1560         return folio;
1561 }
1562
1563 static struct folio *shmem_alloc_folio(gfp_t gfp,
1564                         struct shmem_inode_info *info, pgoff_t index)
1565 {
1566         struct vm_area_struct pvma;
1567         struct folio *folio;
1568
1569         shmem_pseudo_vma_init(&pvma, info, index);
1570         folio = vma_alloc_folio(gfp, 0, &pvma, 0, false);
1571         shmem_pseudo_vma_destroy(&pvma);
1572
1573         return folio;
1574 }
1575
1576 static struct folio *shmem_alloc_and_acct_folio(gfp_t gfp, struct inode *inode,
1577                 pgoff_t index, bool huge)
1578 {
1579         struct shmem_inode_info *info = SHMEM_I(inode);
1580         struct folio *folio;
1581         int nr;
1582         int err = -ENOSPC;
1583
1584         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1585                 huge = false;
1586         nr = huge ? HPAGE_PMD_NR : 1;
1587
1588         if (!shmem_inode_acct_block(inode, nr))
1589                 goto failed;
1590
1591         if (huge)
1592                 folio = shmem_alloc_hugefolio(gfp, info, index);
1593         else
1594                 folio = shmem_alloc_folio(gfp, info, index);
1595         if (folio) {
1596                 __folio_set_locked(folio);
1597                 __folio_set_swapbacked(folio);
1598                 return folio;
1599         }
1600
1601         err = -ENOMEM;
1602         shmem_inode_unacct_blocks(inode, nr);
1603 failed:
1604         return ERR_PTR(err);
1605 }
1606
1607 /*
1608  * When a page is moved from swapcache to shmem filecache (either by the
1609  * usual swapin of shmem_get_folio_gfp(), or by the less common swapoff of
1610  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1611  * ignorance of the mapping it belongs to.  If that mapping has special
1612  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1613  * we may need to copy to a suitable page before moving to filecache.
1614  *
1615  * In a future release, this may well be extended to respect cpuset and
1616  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1617  * but for now it is a simple matter of zone.
1618  */
1619 static bool shmem_should_replace_folio(struct folio *folio, gfp_t gfp)
1620 {
1621         return folio_zonenum(folio) > gfp_zone(gfp);
1622 }
1623
1624 static int shmem_replace_folio(struct folio **foliop, gfp_t gfp,
1625                                 struct shmem_inode_info *info, pgoff_t index)
1626 {
1627         struct folio *old, *new;
1628         struct address_space *swap_mapping;
1629         swp_entry_t entry;
1630         pgoff_t swap_index;
1631         int error;
1632
1633         old = *foliop;
1634         entry = folio_swap_entry(old);
1635         swap_index = swp_offset(entry);
1636         swap_mapping = swap_address_space(entry);
1637
1638         /*
1639          * We have arrived here because our zones are constrained, so don't
1640          * limit chance of success by further cpuset and node constraints.
1641          */
1642         gfp &= ~GFP_CONSTRAINT_MASK;
1643         VM_BUG_ON_FOLIO(folio_test_large(old), old);
1644         new = shmem_alloc_folio(gfp, info, index);
1645         if (!new)
1646                 return -ENOMEM;
1647
1648         folio_get(new);
1649         folio_copy(new, old);
1650         flush_dcache_folio(new);
1651
1652         __folio_set_locked(new);
1653         __folio_set_swapbacked(new);
1654         folio_mark_uptodate(new);
1655         folio_set_swap_entry(new, entry);
1656         folio_set_swapcache(new);
1657
1658         /*
1659          * Our caller will very soon move newpage out of swapcache, but it's
1660          * a nice clean interface for us to replace oldpage by newpage there.
1661          */
1662         xa_lock_irq(&swap_mapping->i_pages);
1663         error = shmem_replace_entry(swap_mapping, swap_index, old, new);
1664         if (!error) {
1665                 mem_cgroup_migrate(old, new);
1666                 __lruvec_stat_mod_folio(new, NR_FILE_PAGES, 1);
1667                 __lruvec_stat_mod_folio(new, NR_SHMEM, 1);
1668                 __lruvec_stat_mod_folio(old, NR_FILE_PAGES, -1);
1669                 __lruvec_stat_mod_folio(old, NR_SHMEM, -1);
1670         }
1671         xa_unlock_irq(&swap_mapping->i_pages);
1672
1673         if (unlikely(error)) {
1674                 /*
1675                  * Is this possible?  I think not, now that our callers check
1676                  * both PageSwapCache and page_private after getting page lock;
1677                  * but be defensive.  Reverse old to newpage for clear and free.
1678                  */
1679                 old = new;
1680         } else {
1681                 folio_add_lru(new);
1682                 *foliop = new;
1683         }
1684
1685         folio_clear_swapcache(old);
1686         old->private = NULL;
1687
1688         folio_unlock(old);
1689         folio_put_refs(old, 2);
1690         return error;
1691 }
1692
1693 static void shmem_set_folio_swapin_error(struct inode *inode, pgoff_t index,
1694                                          struct folio *folio, swp_entry_t swap)
1695 {
1696         struct address_space *mapping = inode->i_mapping;
1697         struct shmem_inode_info *info = SHMEM_I(inode);
1698         swp_entry_t swapin_error;
1699         void *old;
1700
1701         swapin_error = make_swapin_error_entry(&folio->page);
1702         old = xa_cmpxchg_irq(&mapping->i_pages, index,
1703                              swp_to_radix_entry(swap),
1704                              swp_to_radix_entry(swapin_error), 0);
1705         if (old != swp_to_radix_entry(swap))
1706                 return;
1707
1708         folio_wait_writeback(folio);
1709         delete_from_swap_cache(folio);
1710         spin_lock_irq(&info->lock);
1711         /*
1712          * Don't treat swapin error folio as alloced. Otherwise inode->i_blocks won't
1713          * be 0 when inode is released and thus trigger WARN_ON(inode->i_blocks) in
1714          * shmem_evict_inode.
1715          */
1716         info->alloced--;
1717         info->swapped--;
1718         shmem_recalc_inode(inode);
1719         spin_unlock_irq(&info->lock);
1720         swap_free(swap);
1721 }
1722
1723 /*
1724  * Swap in the folio pointed to by *foliop.
1725  * Caller has to make sure that *foliop contains a valid swapped folio.
1726  * Returns 0 and the folio in foliop if success. On failure, returns the
1727  * error code and NULL in *foliop.
1728  */
1729 static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
1730                              struct folio **foliop, enum sgp_type sgp,
1731                              gfp_t gfp, struct vm_area_struct *vma,
1732                              vm_fault_t *fault_type)
1733 {
1734         struct address_space *mapping = inode->i_mapping;
1735         struct shmem_inode_info *info = SHMEM_I(inode);
1736         struct mm_struct *charge_mm = vma ? vma->vm_mm : NULL;
1737         struct folio *folio = NULL;
1738         swp_entry_t swap;
1739         int error;
1740
1741         VM_BUG_ON(!*foliop || !xa_is_value(*foliop));
1742         swap = radix_to_swp_entry(*foliop);
1743         *foliop = NULL;
1744
1745         if (is_swapin_error_entry(swap))
1746                 return -EIO;
1747
1748         /* Look it up and read it in.. */
1749         folio = swap_cache_get_folio(swap, NULL, 0);
1750         if (!folio) {
1751                 /* Or update major stats only when swapin succeeds?? */
1752                 if (fault_type) {
1753                         *fault_type |= VM_FAULT_MAJOR;
1754                         count_vm_event(PGMAJFAULT);
1755                         count_memcg_event_mm(charge_mm, PGMAJFAULT);
1756                 }
1757                 /* Here we actually start the io */
1758                 folio = shmem_swapin(swap, gfp, info, index);
1759                 if (!folio) {
1760                         error = -ENOMEM;
1761                         goto failed;
1762                 }
1763         }
1764
1765         /* We have to do this with folio locked to prevent races */
1766         folio_lock(folio);
1767         if (!folio_test_swapcache(folio) ||
1768             folio_swap_entry(folio).val != swap.val ||
1769             !shmem_confirm_swap(mapping, index, swap)) {
1770                 error = -EEXIST;
1771                 goto unlock;
1772         }
1773         if (!folio_test_uptodate(folio)) {
1774                 error = -EIO;
1775                 goto failed;
1776         }
1777         folio_wait_writeback(folio);
1778
1779         /*
1780          * Some architectures may have to restore extra metadata to the
1781          * folio after reading from swap.
1782          */
1783         arch_swap_restore(swap, folio);
1784
1785         if (shmem_should_replace_folio(folio, gfp)) {
1786                 error = shmem_replace_folio(&folio, gfp, info, index);
1787                 if (error)
1788                         goto failed;
1789         }
1790
1791         error = shmem_add_to_page_cache(folio, mapping, index,
1792                                         swp_to_radix_entry(swap), gfp,
1793                                         charge_mm);
1794         if (error)
1795                 goto failed;
1796
1797         spin_lock_irq(&info->lock);
1798         info->swapped--;
1799         shmem_recalc_inode(inode);
1800         spin_unlock_irq(&info->lock);
1801
1802         if (sgp == SGP_WRITE)
1803                 folio_mark_accessed(folio);
1804
1805         delete_from_swap_cache(folio);
1806         folio_mark_dirty(folio);
1807         swap_free(swap);
1808
1809         *foliop = folio;
1810         return 0;
1811 failed:
1812         if (!shmem_confirm_swap(mapping, index, swap))
1813                 error = -EEXIST;
1814         if (error == -EIO)
1815                 shmem_set_folio_swapin_error(inode, index, folio, swap);
1816 unlock:
1817         if (folio) {
1818                 folio_unlock(folio);
1819                 folio_put(folio);
1820         }
1821
1822         return error;
1823 }
1824
1825 /*
1826  * shmem_get_folio_gfp - find page in cache, or get from swap, or allocate
1827  *
1828  * If we allocate a new one we do not mark it dirty. That's up to the
1829  * vm. If we swap it in we mark it dirty since we also free the swap
1830  * entry since a page cannot live in both the swap and page cache.
1831  *
1832  * vma, vmf, and fault_type are only supplied by shmem_fault:
1833  * otherwise they are NULL.
1834  */
1835 static int shmem_get_folio_gfp(struct inode *inode, pgoff_t index,
1836                 struct folio **foliop, enum sgp_type sgp, gfp_t gfp,
1837                 struct vm_area_struct *vma, struct vm_fault *vmf,
1838                 vm_fault_t *fault_type)
1839 {
1840         struct address_space *mapping = inode->i_mapping;
1841         struct shmem_inode_info *info = SHMEM_I(inode);
1842         struct shmem_sb_info *sbinfo;
1843         struct mm_struct *charge_mm;
1844         struct folio *folio;
1845         pgoff_t hindex = index;
1846         gfp_t huge_gfp;
1847         int error;
1848         int once = 0;
1849         int alloced = 0;
1850
1851         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1852                 return -EFBIG;
1853 repeat:
1854         if (sgp <= SGP_CACHE &&
1855             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1856                 return -EINVAL;
1857         }
1858
1859         sbinfo = SHMEM_SB(inode->i_sb);
1860         charge_mm = vma ? vma->vm_mm : NULL;
1861
1862         folio = __filemap_get_folio(mapping, index, FGP_ENTRY | FGP_LOCK, 0);
1863         if (folio && vma && userfaultfd_minor(vma)) {
1864                 if (!xa_is_value(folio)) {
1865                         folio_unlock(folio);
1866                         folio_put(folio);
1867                 }
1868                 *fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
1869                 return 0;
1870         }
1871
1872         if (xa_is_value(folio)) {
1873                 error = shmem_swapin_folio(inode, index, &folio,
1874                                           sgp, gfp, vma, fault_type);
1875                 if (error == -EEXIST)
1876                         goto repeat;
1877
1878                 *foliop = folio;
1879                 return error;
1880         }
1881
1882         if (folio) {
1883                 hindex = folio->index;
1884                 if (sgp == SGP_WRITE)
1885                         folio_mark_accessed(folio);
1886                 if (folio_test_uptodate(folio))
1887                         goto out;
1888                 /* fallocated folio */
1889                 if (sgp != SGP_READ)
1890                         goto clear;
1891                 folio_unlock(folio);
1892                 folio_put(folio);
1893         }
1894
1895         /*
1896          * SGP_READ: succeed on hole, with NULL folio, letting caller zero.
1897          * SGP_NOALLOC: fail on hole, with NULL folio, letting caller fail.
1898          */
1899         *foliop = NULL;
1900         if (sgp == SGP_READ)
1901                 return 0;
1902         if (sgp == SGP_NOALLOC)
1903                 return -ENOENT;
1904
1905         /*
1906          * Fast cache lookup and swap lookup did not find it: allocate.
1907          */
1908
1909         if (vma && userfaultfd_missing(vma)) {
1910                 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1911                 return 0;
1912         }
1913
1914         if (!shmem_is_huge(vma, inode, index, false))
1915                 goto alloc_nohuge;
1916
1917         huge_gfp = vma_thp_gfp_mask(vma);
1918         huge_gfp = limit_gfp_mask(huge_gfp, gfp);
1919         folio = shmem_alloc_and_acct_folio(huge_gfp, inode, index, true);
1920         if (IS_ERR(folio)) {
1921 alloc_nohuge:
1922                 folio = shmem_alloc_and_acct_folio(gfp, inode, index, false);
1923         }
1924         if (IS_ERR(folio)) {
1925                 int retry = 5;
1926
1927                 error = PTR_ERR(folio);
1928                 folio = NULL;
1929                 if (error != -ENOSPC)
1930                         goto unlock;
1931                 /*
1932                  * Try to reclaim some space by splitting a large folio
1933                  * beyond i_size on the filesystem.
1934                  */
1935                 while (retry--) {
1936                         int ret;
1937
1938                         ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1939                         if (ret == SHRINK_STOP)
1940                                 break;
1941                         if (ret)
1942                                 goto alloc_nohuge;
1943                 }
1944                 goto unlock;
1945         }
1946
1947         hindex = round_down(index, folio_nr_pages(folio));
1948
1949         if (sgp == SGP_WRITE)
1950                 __folio_set_referenced(folio);
1951
1952         error = shmem_add_to_page_cache(folio, mapping, hindex,
1953                                         NULL, gfp & GFP_RECLAIM_MASK,
1954                                         charge_mm);
1955         if (error)
1956                 goto unacct;
1957         folio_add_lru(folio);
1958
1959         spin_lock_irq(&info->lock);
1960         info->alloced += folio_nr_pages(folio);
1961         inode->i_blocks += (blkcnt_t)BLOCKS_PER_PAGE << folio_order(folio);
1962         shmem_recalc_inode(inode);
1963         spin_unlock_irq(&info->lock);
1964         alloced = true;
1965
1966         if (folio_test_pmd_mappable(folio) &&
1967             DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1968                                         folio_next_index(folio) - 1) {
1969                 /*
1970                  * Part of the large folio is beyond i_size: subject
1971                  * to shrink under memory pressure.
1972                  */
1973                 spin_lock(&sbinfo->shrinklist_lock);
1974                 /*
1975                  * _careful to defend against unlocked access to
1976                  * ->shrink_list in shmem_unused_huge_shrink()
1977                  */
1978                 if (list_empty_careful(&info->shrinklist)) {
1979                         list_add_tail(&info->shrinklist,
1980                                       &sbinfo->shrinklist);
1981                         sbinfo->shrinklist_len++;
1982                 }
1983                 spin_unlock(&sbinfo->shrinklist_lock);
1984         }
1985
1986         /*
1987          * Let SGP_FALLOC use the SGP_WRITE optimization on a new folio.
1988          */
1989         if (sgp == SGP_FALLOC)
1990                 sgp = SGP_WRITE;
1991 clear:
1992         /*
1993          * Let SGP_WRITE caller clear ends if write does not fill folio;
1994          * but SGP_FALLOC on a folio fallocated earlier must initialize
1995          * it now, lest undo on failure cancel our earlier guarantee.
1996          */
1997         if (sgp != SGP_WRITE && !folio_test_uptodate(folio)) {
1998                 long i, n = folio_nr_pages(folio);
1999
2000                 for (i = 0; i < n; i++)
2001                         clear_highpage(folio_page(folio, i));
2002                 flush_dcache_folio(folio);
2003                 folio_mark_uptodate(folio);
2004         }
2005
2006         /* Perhaps the file has been truncated since we checked */
2007         if (sgp <= SGP_CACHE &&
2008             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
2009                 if (alloced) {
2010                         folio_clear_dirty(folio);
2011                         filemap_remove_folio(folio);
2012                         spin_lock_irq(&info->lock);
2013                         shmem_recalc_inode(inode);
2014                         spin_unlock_irq(&info->lock);
2015                 }
2016                 error = -EINVAL;
2017                 goto unlock;
2018         }
2019 out:
2020         *foliop = folio;
2021         return 0;
2022
2023         /*
2024          * Error recovery.
2025          */
2026 unacct:
2027         shmem_inode_unacct_blocks(inode, folio_nr_pages(folio));
2028
2029         if (folio_test_large(folio)) {
2030                 folio_unlock(folio);
2031                 folio_put(folio);
2032                 goto alloc_nohuge;
2033         }
2034 unlock:
2035         if (folio) {
2036                 folio_unlock(folio);
2037                 folio_put(folio);
2038         }
2039         if (error == -ENOSPC && !once++) {
2040                 spin_lock_irq(&info->lock);
2041                 shmem_recalc_inode(inode);
2042                 spin_unlock_irq(&info->lock);
2043                 goto repeat;
2044         }
2045         if (error == -EEXIST)
2046                 goto repeat;
2047         return error;
2048 }
2049
2050 int shmem_get_folio(struct inode *inode, pgoff_t index, struct folio **foliop,
2051                 enum sgp_type sgp)
2052 {
2053         return shmem_get_folio_gfp(inode, index, foliop, sgp,
2054                         mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
2055 }
2056
2057 /*
2058  * This is like autoremove_wake_function, but it removes the wait queue
2059  * entry unconditionally - even if something else had already woken the
2060  * target.
2061  */
2062 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2063 {
2064         int ret = default_wake_function(wait, mode, sync, key);
2065         list_del_init(&wait->entry);
2066         return ret;
2067 }
2068
2069 static vm_fault_t shmem_fault(struct vm_fault *vmf)
2070 {
2071         struct vm_area_struct *vma = vmf->vma;
2072         struct inode *inode = file_inode(vma->vm_file);
2073         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2074         struct folio *folio = NULL;
2075         int err;
2076         vm_fault_t ret = VM_FAULT_LOCKED;
2077
2078         /*
2079          * Trinity finds that probing a hole which tmpfs is punching can
2080          * prevent the hole-punch from ever completing: which in turn
2081          * locks writers out with its hold on i_rwsem.  So refrain from
2082          * faulting pages into the hole while it's being punched.  Although
2083          * shmem_undo_range() does remove the additions, it may be unable to
2084          * keep up, as each new page needs its own unmap_mapping_range() call,
2085          * and the i_mmap tree grows ever slower to scan if new vmas are added.
2086          *
2087          * It does not matter if we sometimes reach this check just before the
2088          * hole-punch begins, so that one fault then races with the punch:
2089          * we just need to make racing faults a rare case.
2090          *
2091          * The implementation below would be much simpler if we just used a
2092          * standard mutex or completion: but we cannot take i_rwsem in fault,
2093          * and bloating every shmem inode for this unlikely case would be sad.
2094          */
2095         if (unlikely(inode->i_private)) {
2096                 struct shmem_falloc *shmem_falloc;
2097
2098                 spin_lock(&inode->i_lock);
2099                 shmem_falloc = inode->i_private;
2100                 if (shmem_falloc &&
2101                     shmem_falloc->waitq &&
2102                     vmf->pgoff >= shmem_falloc->start &&
2103                     vmf->pgoff < shmem_falloc->next) {
2104                         struct file *fpin;
2105                         wait_queue_head_t *shmem_falloc_waitq;
2106                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2107
2108                         ret = VM_FAULT_NOPAGE;
2109                         fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2110                         if (fpin)
2111                                 ret = VM_FAULT_RETRY;
2112
2113                         shmem_falloc_waitq = shmem_falloc->waitq;
2114                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2115                                         TASK_UNINTERRUPTIBLE);
2116                         spin_unlock(&inode->i_lock);
2117                         schedule();
2118
2119                         /*
2120                          * shmem_falloc_waitq points into the shmem_fallocate()
2121                          * stack of the hole-punching task: shmem_falloc_waitq
2122                          * is usually invalid by the time we reach here, but
2123                          * finish_wait() does not dereference it in that case;
2124                          * though i_lock needed lest racing with wake_up_all().
2125                          */
2126                         spin_lock(&inode->i_lock);
2127                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2128                         spin_unlock(&inode->i_lock);
2129
2130                         if (fpin)
2131                                 fput(fpin);
2132                         return ret;
2133                 }
2134                 spin_unlock(&inode->i_lock);
2135         }
2136
2137         err = shmem_get_folio_gfp(inode, vmf->pgoff, &folio, SGP_CACHE,
2138                                   gfp, vma, vmf, &ret);
2139         if (err)
2140                 return vmf_error(err);
2141         if (folio)
2142                 vmf->page = folio_file_page(folio, vmf->pgoff);
2143         return ret;
2144 }
2145
2146 unsigned long shmem_get_unmapped_area(struct file *file,
2147                                       unsigned long uaddr, unsigned long len,
2148                                       unsigned long pgoff, unsigned long flags)
2149 {
2150         unsigned long (*get_area)(struct file *,
2151                 unsigned long, unsigned long, unsigned long, unsigned long);
2152         unsigned long addr;
2153         unsigned long offset;
2154         unsigned long inflated_len;
2155         unsigned long inflated_addr;
2156         unsigned long inflated_offset;
2157
2158         if (len > TASK_SIZE)
2159                 return -ENOMEM;
2160
2161         get_area = current->mm->get_unmapped_area;
2162         addr = get_area(file, uaddr, len, pgoff, flags);
2163
2164         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2165                 return addr;
2166         if (IS_ERR_VALUE(addr))
2167                 return addr;
2168         if (addr & ~PAGE_MASK)
2169                 return addr;
2170         if (addr > TASK_SIZE - len)
2171                 return addr;
2172
2173         if (shmem_huge == SHMEM_HUGE_DENY)
2174                 return addr;
2175         if (len < HPAGE_PMD_SIZE)
2176                 return addr;
2177         if (flags & MAP_FIXED)
2178                 return addr;
2179         /*
2180          * Our priority is to support MAP_SHARED mapped hugely;
2181          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2182          * But if caller specified an address hint and we allocated area there
2183          * successfully, respect that as before.
2184          */
2185         if (uaddr == addr)
2186                 return addr;
2187
2188         if (shmem_huge != SHMEM_HUGE_FORCE) {
2189                 struct super_block *sb;
2190
2191                 if (file) {
2192                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2193                         sb = file_inode(file)->i_sb;
2194                 } else {
2195                         /*
2196                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2197                          * for "/dev/zero", to create a shared anonymous object.
2198                          */
2199                         if (IS_ERR(shm_mnt))
2200                                 return addr;
2201                         sb = shm_mnt->mnt_sb;
2202                 }
2203                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2204                         return addr;
2205         }
2206
2207         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2208         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2209                 return addr;
2210         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2211                 return addr;
2212
2213         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2214         if (inflated_len > TASK_SIZE)
2215                 return addr;
2216         if (inflated_len < len)
2217                 return addr;
2218
2219         inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2220         if (IS_ERR_VALUE(inflated_addr))
2221                 return addr;
2222         if (inflated_addr & ~PAGE_MASK)
2223                 return addr;
2224
2225         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2226         inflated_addr += offset - inflated_offset;
2227         if (inflated_offset > offset)
2228                 inflated_addr += HPAGE_PMD_SIZE;
2229
2230         if (inflated_addr > TASK_SIZE - len)
2231                 return addr;
2232         return inflated_addr;
2233 }
2234
2235 #ifdef CONFIG_NUMA
2236 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2237 {
2238         struct inode *inode = file_inode(vma->vm_file);
2239         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2240 }
2241
2242 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2243                                           unsigned long addr)
2244 {
2245         struct inode *inode = file_inode(vma->vm_file);
2246         pgoff_t index;
2247
2248         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2249         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2250 }
2251 #endif
2252
2253 int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
2254 {
2255         struct inode *inode = file_inode(file);
2256         struct shmem_inode_info *info = SHMEM_I(inode);
2257         int retval = -ENOMEM;
2258
2259         /*
2260          * What serializes the accesses to info->flags?
2261          * ipc_lock_object() when called from shmctl_do_lock(),
2262          * no serialization needed when called from shm_destroy().
2263          */
2264         if (lock && !(info->flags & VM_LOCKED)) {
2265                 if (!user_shm_lock(inode->i_size, ucounts))
2266                         goto out_nomem;
2267                 info->flags |= VM_LOCKED;
2268                 mapping_set_unevictable(file->f_mapping);
2269         }
2270         if (!lock && (info->flags & VM_LOCKED) && ucounts) {
2271                 user_shm_unlock(inode->i_size, ucounts);
2272                 info->flags &= ~VM_LOCKED;
2273                 mapping_clear_unevictable(file->f_mapping);
2274         }
2275         retval = 0;
2276
2277 out_nomem:
2278         return retval;
2279 }
2280
2281 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2282 {
2283         struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2284         int ret;
2285
2286         ret = seal_check_future_write(info->seals, vma);
2287         if (ret)
2288                 return ret;
2289
2290         /* arm64 - allow memory tagging on RAM-based files */
2291         vma->vm_flags |= VM_MTE_ALLOWED;
2292
2293         file_accessed(file);
2294         vma->vm_ops = &shmem_vm_ops;
2295         return 0;
2296 }
2297
2298 #ifdef CONFIG_TMPFS_XATTR
2299 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2300
2301 /*
2302  * chattr's fsflags are unrelated to extended attributes,
2303  * but tmpfs has chosen to enable them under the same config option.
2304  */
2305 static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2306 {
2307         unsigned int i_flags = 0;
2308
2309         if (fsflags & FS_NOATIME_FL)
2310                 i_flags |= S_NOATIME;
2311         if (fsflags & FS_APPEND_FL)
2312                 i_flags |= S_APPEND;
2313         if (fsflags & FS_IMMUTABLE_FL)
2314                 i_flags |= S_IMMUTABLE;
2315         /*
2316          * But FS_NODUMP_FL does not require any action in i_flags.
2317          */
2318         inode_set_flags(inode, i_flags, S_NOATIME | S_APPEND | S_IMMUTABLE);
2319 }
2320 #else
2321 static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2322 {
2323 }
2324 #define shmem_initxattrs NULL
2325 #endif
2326
2327 static struct inode *shmem_get_inode(struct super_block *sb, struct inode *dir,
2328                                      umode_t mode, dev_t dev, unsigned long flags)
2329 {
2330         struct inode *inode;
2331         struct shmem_inode_info *info;
2332         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2333         ino_t ino;
2334
2335         if (shmem_reserve_inode(sb, &ino))
2336                 return NULL;
2337
2338         inode = new_inode(sb);
2339         if (inode) {
2340                 inode->i_ino = ino;
2341                 inode_init_owner(&init_user_ns, inode, dir, mode);
2342                 inode->i_blocks = 0;
2343                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2344                 inode->i_generation = get_random_u32();
2345                 info = SHMEM_I(inode);
2346                 memset(info, 0, (char *)inode - (char *)info);
2347                 spin_lock_init(&info->lock);
2348                 atomic_set(&info->stop_eviction, 0);
2349                 info->seals = F_SEAL_SEAL;
2350                 info->flags = flags & VM_NORESERVE;
2351                 info->i_crtime = inode->i_mtime;
2352                 info->fsflags = (dir == NULL) ? 0 :
2353                         SHMEM_I(dir)->fsflags & SHMEM_FL_INHERITED;
2354                 if (info->fsflags)
2355                         shmem_set_inode_flags(inode, info->fsflags);
2356                 INIT_LIST_HEAD(&info->shrinklist);
2357                 INIT_LIST_HEAD(&info->swaplist);
2358                 simple_xattrs_init(&info->xattrs);
2359                 cache_no_acl(inode);
2360                 mapping_set_large_folios(inode->i_mapping);
2361
2362                 switch (mode & S_IFMT) {
2363                 default:
2364                         inode->i_op = &shmem_special_inode_operations;
2365                         init_special_inode(inode, mode, dev);
2366                         break;
2367                 case S_IFREG:
2368                         inode->i_mapping->a_ops = &shmem_aops;
2369                         inode->i_op = &shmem_inode_operations;
2370                         inode->i_fop = &shmem_file_operations;
2371                         mpol_shared_policy_init(&info->policy,
2372                                                  shmem_get_sbmpol(sbinfo));
2373                         break;
2374                 case S_IFDIR:
2375                         inc_nlink(inode);
2376                         /* Some things misbehave if size == 0 on a directory */
2377                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2378                         inode->i_op = &shmem_dir_inode_operations;
2379                         inode->i_fop = &simple_dir_operations;
2380                         break;
2381                 case S_IFLNK:
2382                         /*
2383                          * Must not load anything in the rbtree,
2384                          * mpol_free_shared_policy will not be called.
2385                          */
2386                         mpol_shared_policy_init(&info->policy, NULL);
2387                         break;
2388                 }
2389
2390                 lockdep_annotate_inode_mutex_key(inode);
2391         } else
2392                 shmem_free_inode(sb);
2393         return inode;
2394 }
2395
2396 #ifdef CONFIG_USERFAULTFD
2397 int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2398                            pmd_t *dst_pmd,
2399                            struct vm_area_struct *dst_vma,
2400                            unsigned long dst_addr,
2401                            unsigned long src_addr,
2402                            bool zeropage, bool wp_copy,
2403                            struct page **pagep)
2404 {
2405         struct inode *inode = file_inode(dst_vma->vm_file);
2406         struct shmem_inode_info *info = SHMEM_I(inode);
2407         struct address_space *mapping = inode->i_mapping;
2408         gfp_t gfp = mapping_gfp_mask(mapping);
2409         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2410         void *page_kaddr;
2411         struct folio *folio;
2412         int ret;
2413         pgoff_t max_off;
2414
2415         if (!shmem_inode_acct_block(inode, 1)) {
2416                 /*
2417                  * We may have got a page, returned -ENOENT triggering a retry,
2418                  * and now we find ourselves with -ENOMEM. Release the page, to
2419                  * avoid a BUG_ON in our caller.
2420                  */
2421                 if (unlikely(*pagep)) {
2422                         put_page(*pagep);
2423                         *pagep = NULL;
2424                 }
2425                 return -ENOMEM;
2426         }
2427
2428         if (!*pagep) {
2429                 ret = -ENOMEM;
2430                 folio = shmem_alloc_folio(gfp, info, pgoff);
2431                 if (!folio)
2432                         goto out_unacct_blocks;
2433
2434                 if (!zeropage) {        /* COPY */
2435                         page_kaddr = kmap_local_folio(folio, 0);
2436                         /*
2437                          * The read mmap_lock is held here.  Despite the
2438                          * mmap_lock being read recursive a deadlock is still
2439                          * possible if a writer has taken a lock.  For example:
2440                          *
2441                          * process A thread 1 takes read lock on own mmap_lock
2442                          * process A thread 2 calls mmap, blocks taking write lock
2443                          * process B thread 1 takes page fault, read lock on own mmap lock
2444                          * process B thread 2 calls mmap, blocks taking write lock
2445                          * process A thread 1 blocks taking read lock on process B
2446                          * process B thread 1 blocks taking read lock on process A
2447                          *
2448                          * Disable page faults to prevent potential deadlock
2449                          * and retry the copy outside the mmap_lock.
2450                          */
2451                         pagefault_disable();
2452                         ret = copy_from_user(page_kaddr,
2453                                              (const void __user *)src_addr,
2454                                              PAGE_SIZE);
2455                         pagefault_enable();
2456                         kunmap_local(page_kaddr);
2457
2458                         /* fallback to copy_from_user outside mmap_lock */
2459                         if (unlikely(ret)) {
2460                                 *pagep = &folio->page;
2461                                 ret = -ENOENT;
2462                                 /* don't free the page */
2463                                 goto out_unacct_blocks;
2464                         }
2465
2466                         flush_dcache_folio(folio);
2467                 } else {                /* ZEROPAGE */
2468                         clear_user_highpage(&folio->page, dst_addr);
2469                 }
2470         } else {
2471                 folio = page_folio(*pagep);
2472                 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2473                 *pagep = NULL;
2474         }
2475
2476         VM_BUG_ON(folio_test_locked(folio));
2477         VM_BUG_ON(folio_test_swapbacked(folio));
2478         __folio_set_locked(folio);
2479         __folio_set_swapbacked(folio);
2480         __folio_mark_uptodate(folio);
2481
2482         ret = -EFAULT;
2483         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2484         if (unlikely(pgoff >= max_off))
2485                 goto out_release;
2486
2487         ret = shmem_add_to_page_cache(folio, mapping, pgoff, NULL,
2488                                       gfp & GFP_RECLAIM_MASK, dst_mm);
2489         if (ret)
2490                 goto out_release;
2491
2492         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
2493                                        &folio->page, true, wp_copy);
2494         if (ret)
2495                 goto out_delete_from_cache;
2496
2497         spin_lock_irq(&info->lock);
2498         info->alloced++;
2499         inode->i_blocks += BLOCKS_PER_PAGE;
2500         shmem_recalc_inode(inode);
2501         spin_unlock_irq(&info->lock);
2502
2503         folio_unlock(folio);
2504         return 0;
2505 out_delete_from_cache:
2506         filemap_remove_folio(folio);
2507 out_release:
2508         folio_unlock(folio);
2509         folio_put(folio);
2510 out_unacct_blocks:
2511         shmem_inode_unacct_blocks(inode, 1);
2512         return ret;
2513 }
2514 #endif /* CONFIG_USERFAULTFD */
2515
2516 #ifdef CONFIG_TMPFS
2517 static const struct inode_operations shmem_symlink_inode_operations;
2518 static const struct inode_operations shmem_short_symlink_operations;
2519
2520 static int
2521 shmem_write_begin(struct file *file, struct address_space *mapping,
2522                         loff_t pos, unsigned len,
2523                         struct page **pagep, void **fsdata)
2524 {
2525         struct inode *inode = mapping->host;
2526         struct shmem_inode_info *info = SHMEM_I(inode);
2527         pgoff_t index = pos >> PAGE_SHIFT;
2528         struct folio *folio;
2529         int ret = 0;
2530
2531         /* i_rwsem is held by caller */
2532         if (unlikely(info->seals & (F_SEAL_GROW |
2533                                    F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2534                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2535                         return -EPERM;
2536                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2537                         return -EPERM;
2538         }
2539
2540         ret = shmem_get_folio(inode, index, &folio, SGP_WRITE);
2541
2542         if (ret)
2543                 return ret;
2544
2545         *pagep = folio_file_page(folio, index);
2546         if (PageHWPoison(*pagep)) {
2547                 folio_unlock(folio);
2548                 folio_put(folio);
2549                 *pagep = NULL;
2550                 return -EIO;
2551         }
2552
2553         return 0;
2554 }
2555
2556 static int
2557 shmem_write_end(struct file *file, struct address_space *mapping,
2558                         loff_t pos, unsigned len, unsigned copied,
2559                         struct page *page, void *fsdata)
2560 {
2561         struct inode *inode = mapping->host;
2562
2563         if (pos + copied > inode->i_size)
2564                 i_size_write(inode, pos + copied);
2565
2566         if (!PageUptodate(page)) {
2567                 struct page *head = compound_head(page);
2568                 if (PageTransCompound(page)) {
2569                         int i;
2570
2571                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2572                                 if (head + i == page)
2573                                         continue;
2574                                 clear_highpage(head + i);
2575                                 flush_dcache_page(head + i);
2576                         }
2577                 }
2578                 if (copied < PAGE_SIZE) {
2579                         unsigned from = pos & (PAGE_SIZE - 1);
2580                         zero_user_segments(page, 0, from,
2581                                         from + copied, PAGE_SIZE);
2582                 }
2583                 SetPageUptodate(head);
2584         }
2585         set_page_dirty(page);
2586         unlock_page(page);
2587         put_page(page);
2588
2589         return copied;
2590 }
2591
2592 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2593 {
2594         struct file *file = iocb->ki_filp;
2595         struct inode *inode = file_inode(file);
2596         struct address_space *mapping = inode->i_mapping;
2597         pgoff_t index;
2598         unsigned long offset;
2599         int error = 0;
2600         ssize_t retval = 0;
2601         loff_t *ppos = &iocb->ki_pos;
2602
2603         index = *ppos >> PAGE_SHIFT;
2604         offset = *ppos & ~PAGE_MASK;
2605
2606         for (;;) {
2607                 struct folio *folio = NULL;
2608                 struct page *page = NULL;
2609                 pgoff_t end_index;
2610                 unsigned long nr, ret;
2611                 loff_t i_size = i_size_read(inode);
2612
2613                 end_index = i_size >> PAGE_SHIFT;
2614                 if (index > end_index)
2615                         break;
2616                 if (index == end_index) {
2617                         nr = i_size & ~PAGE_MASK;
2618                         if (nr <= offset)
2619                                 break;
2620                 }
2621
2622                 error = shmem_get_folio(inode, index, &folio, SGP_READ);
2623                 if (error) {
2624                         if (error == -EINVAL)
2625                                 error = 0;
2626                         break;
2627                 }
2628                 if (folio) {
2629                         folio_unlock(folio);
2630
2631                         page = folio_file_page(folio, index);
2632                         if (PageHWPoison(page)) {
2633                                 folio_put(folio);
2634                                 error = -EIO;
2635                                 break;
2636                         }
2637                 }
2638
2639                 /*
2640                  * We must evaluate after, since reads (unlike writes)
2641                  * are called without i_rwsem protection against truncate
2642                  */
2643                 nr = PAGE_SIZE;
2644                 i_size = i_size_read(inode);
2645                 end_index = i_size >> PAGE_SHIFT;
2646                 if (index == end_index) {
2647                         nr = i_size & ~PAGE_MASK;
2648                         if (nr <= offset) {
2649                                 if (folio)
2650                                         folio_put(folio);
2651                                 break;
2652                         }
2653                 }
2654                 nr -= offset;
2655
2656                 if (folio) {
2657                         /*
2658                          * If users can be writing to this page using arbitrary
2659                          * virtual addresses, take care about potential aliasing
2660                          * before reading the page on the kernel side.
2661                          */
2662                         if (mapping_writably_mapped(mapping))
2663                                 flush_dcache_page(page);
2664                         /*
2665                          * Mark the page accessed if we read the beginning.
2666                          */
2667                         if (!offset)
2668                                 folio_mark_accessed(folio);
2669                         /*
2670                          * Ok, we have the page, and it's up-to-date, so
2671                          * now we can copy it to user space...
2672                          */
2673                         ret = copy_page_to_iter(page, offset, nr, to);
2674                         folio_put(folio);
2675
2676                 } else if (user_backed_iter(to)) {
2677                         /*
2678                          * Copy to user tends to be so well optimized, but
2679                          * clear_user() not so much, that it is noticeably
2680                          * faster to copy the zero page instead of clearing.
2681                          */
2682                         ret = copy_page_to_iter(ZERO_PAGE(0), offset, nr, to);
2683                 } else {
2684                         /*
2685                          * But submitting the same page twice in a row to
2686                          * splice() - or others? - can result in confusion:
2687                          * so don't attempt that optimization on pipes etc.
2688                          */
2689                         ret = iov_iter_zero(nr, to);
2690                 }
2691
2692                 retval += ret;
2693                 offset += ret;
2694                 index += offset >> PAGE_SHIFT;
2695                 offset &= ~PAGE_MASK;
2696
2697                 if (!iov_iter_count(to))
2698                         break;
2699                 if (ret < nr) {
2700                         error = -EFAULT;
2701                         break;
2702                 }
2703                 cond_resched();
2704         }
2705
2706         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2707         file_accessed(file);
2708         return retval ? retval : error;
2709 }
2710
2711 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2712 {
2713         struct address_space *mapping = file->f_mapping;
2714         struct inode *inode = mapping->host;
2715
2716         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2717                 return generic_file_llseek_size(file, offset, whence,
2718                                         MAX_LFS_FILESIZE, i_size_read(inode));
2719         if (offset < 0)
2720                 return -ENXIO;
2721
2722         inode_lock(inode);
2723         /* We're holding i_rwsem so we can access i_size directly */
2724         offset = mapping_seek_hole_data(mapping, offset, inode->i_size, whence);
2725         if (offset >= 0)
2726                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2727         inode_unlock(inode);
2728         return offset;
2729 }
2730
2731 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2732                                                          loff_t len)
2733 {
2734         struct inode *inode = file_inode(file);
2735         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2736         struct shmem_inode_info *info = SHMEM_I(inode);
2737         struct shmem_falloc shmem_falloc;
2738         pgoff_t start, index, end, undo_fallocend;
2739         int error;
2740
2741         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2742                 return -EOPNOTSUPP;
2743
2744         inode_lock(inode);
2745
2746         if (mode & FALLOC_FL_PUNCH_HOLE) {
2747                 struct address_space *mapping = file->f_mapping;
2748                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2749                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2750                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2751
2752                 /* protected by i_rwsem */
2753                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2754                         error = -EPERM;
2755                         goto out;
2756                 }
2757
2758                 shmem_falloc.waitq = &shmem_falloc_waitq;
2759                 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2760                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2761                 spin_lock(&inode->i_lock);
2762                 inode->i_private = &shmem_falloc;
2763                 spin_unlock(&inode->i_lock);
2764
2765                 if ((u64)unmap_end > (u64)unmap_start)
2766                         unmap_mapping_range(mapping, unmap_start,
2767                                             1 + unmap_end - unmap_start, 0);
2768                 shmem_truncate_range(inode, offset, offset + len - 1);
2769                 /* No need to unmap again: hole-punching leaves COWed pages */
2770
2771                 spin_lock(&inode->i_lock);
2772                 inode->i_private = NULL;
2773                 wake_up_all(&shmem_falloc_waitq);
2774                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2775                 spin_unlock(&inode->i_lock);
2776                 error = 0;
2777                 goto out;
2778         }
2779
2780         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2781         error = inode_newsize_ok(inode, offset + len);
2782         if (error)
2783                 goto out;
2784
2785         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2786                 error = -EPERM;
2787                 goto out;
2788         }
2789
2790         start = offset >> PAGE_SHIFT;
2791         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2792         /* Try to avoid a swapstorm if len is impossible to satisfy */
2793         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2794                 error = -ENOSPC;
2795                 goto out;
2796         }
2797
2798         shmem_falloc.waitq = NULL;
2799         shmem_falloc.start = start;
2800         shmem_falloc.next  = start;
2801         shmem_falloc.nr_falloced = 0;
2802         shmem_falloc.nr_unswapped = 0;
2803         spin_lock(&inode->i_lock);
2804         inode->i_private = &shmem_falloc;
2805         spin_unlock(&inode->i_lock);
2806
2807         /*
2808          * info->fallocend is only relevant when huge pages might be
2809          * involved: to prevent split_huge_page() freeing fallocated
2810          * pages when FALLOC_FL_KEEP_SIZE committed beyond i_size.
2811          */
2812         undo_fallocend = info->fallocend;
2813         if (info->fallocend < end)
2814                 info->fallocend = end;
2815
2816         for (index = start; index < end; ) {
2817                 struct folio *folio;
2818
2819                 /*
2820                  * Good, the fallocate(2) manpage permits EINTR: we may have
2821                  * been interrupted because we are using up too much memory.
2822                  */
2823                 if (signal_pending(current))
2824                         error = -EINTR;
2825                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2826                         error = -ENOMEM;
2827                 else
2828                         error = shmem_get_folio(inode, index, &folio,
2829                                                 SGP_FALLOC);
2830                 if (error) {
2831                         info->fallocend = undo_fallocend;
2832                         /* Remove the !uptodate folios we added */
2833                         if (index > start) {
2834                                 shmem_undo_range(inode,
2835                                     (loff_t)start << PAGE_SHIFT,
2836                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2837                         }
2838                         goto undone;
2839                 }
2840
2841                 /*
2842                  * Here is a more important optimization than it appears:
2843                  * a second SGP_FALLOC on the same large folio will clear it,
2844                  * making it uptodate and un-undoable if we fail later.
2845                  */
2846                 index = folio_next_index(folio);
2847                 /* Beware 32-bit wraparound */
2848                 if (!index)
2849                         index--;
2850
2851                 /*
2852                  * Inform shmem_writepage() how far we have reached.
2853                  * No need for lock or barrier: we have the page lock.
2854                  */
2855                 if (!folio_test_uptodate(folio))
2856                         shmem_falloc.nr_falloced += index - shmem_falloc.next;
2857                 shmem_falloc.next = index;
2858
2859                 /*
2860                  * If !uptodate, leave it that way so that freeable folios
2861                  * can be recognized if we need to rollback on error later.
2862                  * But mark it dirty so that memory pressure will swap rather
2863                  * than free the folios we are allocating (and SGP_CACHE folios
2864                  * might still be clean: we now need to mark those dirty too).
2865                  */
2866                 folio_mark_dirty(folio);
2867                 folio_unlock(folio);
2868                 folio_put(folio);
2869                 cond_resched();
2870         }
2871
2872         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2873                 i_size_write(inode, offset + len);
2874 undone:
2875         spin_lock(&inode->i_lock);
2876         inode->i_private = NULL;
2877         spin_unlock(&inode->i_lock);
2878 out:
2879         if (!error)
2880                 file_modified(file);
2881         inode_unlock(inode);
2882         return error;
2883 }
2884
2885 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2886 {
2887         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2888
2889         buf->f_type = TMPFS_MAGIC;
2890         buf->f_bsize = PAGE_SIZE;
2891         buf->f_namelen = NAME_MAX;
2892         if (sbinfo->max_blocks) {
2893                 buf->f_blocks = sbinfo->max_blocks;
2894                 buf->f_bavail =
2895                 buf->f_bfree  = sbinfo->max_blocks -
2896                                 percpu_counter_sum(&sbinfo->used_blocks);
2897         }
2898         if (sbinfo->max_inodes) {
2899                 buf->f_files = sbinfo->max_inodes;
2900                 buf->f_ffree = sbinfo->free_inodes;
2901         }
2902         /* else leave those fields 0 like simple_statfs */
2903
2904         buf->f_fsid = uuid_to_fsid(dentry->d_sb->s_uuid.b);
2905
2906         return 0;
2907 }
2908
2909 /*
2910  * File creation. Allocate an inode, and we're done..
2911  */
2912 static int
2913 shmem_mknod(struct user_namespace *mnt_userns, struct inode *dir,
2914             struct dentry *dentry, umode_t mode, dev_t dev)
2915 {
2916         struct inode *inode;
2917         int error = -ENOSPC;
2918
2919         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2920         if (inode) {
2921                 error = simple_acl_create(dir, inode);
2922                 if (error)
2923                         goto out_iput;
2924                 error = security_inode_init_security(inode, dir,
2925                                                      &dentry->d_name,
2926                                                      shmem_initxattrs, NULL);
2927                 if (error && error != -EOPNOTSUPP)
2928                         goto out_iput;
2929
2930                 error = 0;
2931                 dir->i_size += BOGO_DIRENT_SIZE;
2932                 dir->i_ctime = dir->i_mtime = current_time(dir);
2933                 inode_inc_iversion(dir);
2934                 d_instantiate(dentry, inode);
2935                 dget(dentry); /* Extra count - pin the dentry in core */
2936         }
2937         return error;
2938 out_iput:
2939         iput(inode);
2940         return error;
2941 }
2942
2943 static int
2944 shmem_tmpfile(struct user_namespace *mnt_userns, struct inode *dir,
2945               struct file *file, umode_t mode)
2946 {
2947         struct inode *inode;
2948         int error = -ENOSPC;
2949
2950         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2951         if (inode) {
2952                 error = security_inode_init_security(inode, dir,
2953                                                      NULL,
2954                                                      shmem_initxattrs, NULL);
2955                 if (error && error != -EOPNOTSUPP)
2956                         goto out_iput;
2957                 error = simple_acl_create(dir, inode);
2958                 if (error)
2959                         goto out_iput;
2960                 d_tmpfile(file, inode);
2961         }
2962         return finish_open_simple(file, error);
2963 out_iput:
2964         iput(inode);
2965         return error;
2966 }
2967
2968 static int shmem_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
2969                        struct dentry *dentry, umode_t mode)
2970 {
2971         int error;
2972
2973         if ((error = shmem_mknod(&init_user_ns, dir, dentry,
2974                                  mode | S_IFDIR, 0)))
2975                 return error;
2976         inc_nlink(dir);
2977         return 0;
2978 }
2979
2980 static int shmem_create(struct user_namespace *mnt_userns, struct inode *dir,
2981                         struct dentry *dentry, umode_t mode, bool excl)
2982 {
2983         return shmem_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
2984 }
2985
2986 /*
2987  * Link a file..
2988  */
2989 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2990 {
2991         struct inode *inode = d_inode(old_dentry);
2992         int ret = 0;
2993
2994         /*
2995          * No ordinary (disk based) filesystem counts links as inodes;
2996          * but each new link needs a new dentry, pinning lowmem, and
2997          * tmpfs dentries cannot be pruned until they are unlinked.
2998          * But if an O_TMPFILE file is linked into the tmpfs, the
2999          * first link must skip that, to get the accounting right.
3000          */
3001         if (inode->i_nlink) {
3002                 ret = shmem_reserve_inode(inode->i_sb, NULL);
3003                 if (ret)
3004                         goto out;
3005         }
3006
3007         dir->i_size += BOGO_DIRENT_SIZE;
3008         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3009         inode_inc_iversion(dir);
3010         inc_nlink(inode);
3011         ihold(inode);   /* New dentry reference */
3012         dget(dentry);           /* Extra pinning count for the created dentry */
3013         d_instantiate(dentry, inode);
3014 out:
3015         return ret;
3016 }
3017
3018 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3019 {
3020         struct inode *inode = d_inode(dentry);
3021
3022         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3023                 shmem_free_inode(inode->i_sb);
3024
3025         dir->i_size -= BOGO_DIRENT_SIZE;
3026         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3027         inode_inc_iversion(dir);
3028         drop_nlink(inode);
3029         dput(dentry);   /* Undo the count from "create" - this does all the work */
3030         return 0;
3031 }
3032
3033 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3034 {
3035         if (!simple_empty(dentry))
3036                 return -ENOTEMPTY;
3037
3038         drop_nlink(d_inode(dentry));
3039         drop_nlink(dir);
3040         return shmem_unlink(dir, dentry);
3041 }
3042
3043 static int shmem_whiteout(struct user_namespace *mnt_userns,
3044                           struct inode *old_dir, struct dentry *old_dentry)
3045 {
3046         struct dentry *whiteout;
3047         int error;
3048
3049         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3050         if (!whiteout)
3051                 return -ENOMEM;
3052
3053         error = shmem_mknod(&init_user_ns, old_dir, whiteout,
3054                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3055         dput(whiteout);
3056         if (error)
3057                 return error;
3058
3059         /*
3060          * Cheat and hash the whiteout while the old dentry is still in
3061          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3062          *
3063          * d_lookup() will consistently find one of them at this point,
3064          * not sure which one, but that isn't even important.
3065          */
3066         d_rehash(whiteout);
3067         return 0;
3068 }
3069
3070 /*
3071  * The VFS layer already does all the dentry stuff for rename,
3072  * we just have to decrement the usage count for the target if
3073  * it exists so that the VFS layer correctly free's it when it
3074  * gets overwritten.
3075  */
3076 static int shmem_rename2(struct user_namespace *mnt_userns,
3077                          struct inode *old_dir, struct dentry *old_dentry,
3078                          struct inode *new_dir, struct dentry *new_dentry,
3079                          unsigned int flags)
3080 {
3081         struct inode *inode = d_inode(old_dentry);
3082         int they_are_dirs = S_ISDIR(inode->i_mode);
3083
3084         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3085                 return -EINVAL;
3086
3087         if (flags & RENAME_EXCHANGE)
3088                 return simple_rename_exchange(old_dir, old_dentry, new_dir, new_dentry);
3089
3090         if (!simple_empty(new_dentry))
3091                 return -ENOTEMPTY;
3092
3093         if (flags & RENAME_WHITEOUT) {
3094                 int error;
3095
3096                 error = shmem_whiteout(&init_user_ns, old_dir, old_dentry);
3097                 if (error)
3098                         return error;
3099         }
3100
3101         if (d_really_is_positive(new_dentry)) {
3102                 (void) shmem_unlink(new_dir, new_dentry);
3103                 if (they_are_dirs) {
3104                         drop_nlink(d_inode(new_dentry));
3105                         drop_nlink(old_dir);
3106                 }
3107         } else if (they_are_dirs) {
3108                 drop_nlink(old_dir);
3109                 inc_nlink(new_dir);
3110         }
3111
3112         old_dir->i_size -= BOGO_DIRENT_SIZE;
3113         new_dir->i_size += BOGO_DIRENT_SIZE;
3114         old_dir->i_ctime = old_dir->i_mtime =
3115         new_dir->i_ctime = new_dir->i_mtime =
3116         inode->i_ctime = current_time(old_dir);
3117         inode_inc_iversion(old_dir);
3118         inode_inc_iversion(new_dir);
3119         return 0;
3120 }
3121
3122 static int shmem_symlink(struct user_namespace *mnt_userns, struct inode *dir,
3123                          struct dentry *dentry, const char *symname)
3124 {
3125         int error;
3126         int len;
3127         struct inode *inode;
3128         struct folio *folio;
3129
3130         len = strlen(symname) + 1;
3131         if (len > PAGE_SIZE)
3132                 return -ENAMETOOLONG;
3133
3134         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3135                                 VM_NORESERVE);
3136         if (!inode)
3137                 return -ENOSPC;
3138
3139         error = security_inode_init_security(inode, dir, &dentry->d_name,
3140                                              shmem_initxattrs, NULL);
3141         if (error && error != -EOPNOTSUPP) {
3142                 iput(inode);
3143                 return error;
3144         }
3145
3146         inode->i_size = len-1;
3147         if (len <= SHORT_SYMLINK_LEN) {
3148                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3149                 if (!inode->i_link) {
3150                         iput(inode);
3151                         return -ENOMEM;
3152                 }
3153                 inode->i_op = &shmem_short_symlink_operations;
3154         } else {
3155                 inode_nohighmem(inode);
3156                 error = shmem_get_folio(inode, 0, &folio, SGP_WRITE);
3157                 if (error) {
3158                         iput(inode);
3159                         return error;
3160                 }
3161                 inode->i_mapping->a_ops = &shmem_aops;
3162                 inode->i_op = &shmem_symlink_inode_operations;
3163                 memcpy(folio_address(folio), symname, len);
3164                 folio_mark_uptodate(folio);
3165                 folio_mark_dirty(folio);
3166                 folio_unlock(folio);
3167                 folio_put(folio);
3168         }
3169         dir->i_size += BOGO_DIRENT_SIZE;
3170         dir->i_ctime = dir->i_mtime = current_time(dir);
3171         inode_inc_iversion(dir);
3172         d_instantiate(dentry, inode);
3173         dget(dentry);
3174         return 0;
3175 }
3176
3177 static void shmem_put_link(void *arg)
3178 {
3179         folio_mark_accessed(arg);
3180         folio_put(arg);
3181 }
3182
3183 static const char *shmem_get_link(struct dentry *dentry,
3184                                   struct inode *inode,
3185                                   struct delayed_call *done)
3186 {
3187         struct folio *folio = NULL;
3188         int error;
3189
3190         if (!dentry) {
3191                 folio = filemap_get_folio(inode->i_mapping, 0);
3192                 if (!folio)
3193                         return ERR_PTR(-ECHILD);
3194                 if (PageHWPoison(folio_page(folio, 0)) ||
3195                     !folio_test_uptodate(folio)) {
3196                         folio_put(folio);
3197                         return ERR_PTR(-ECHILD);
3198                 }
3199         } else {
3200                 error = shmem_get_folio(inode, 0, &folio, SGP_READ);
3201                 if (error)
3202                         return ERR_PTR(error);
3203                 if (!folio)
3204                         return ERR_PTR(-ECHILD);
3205                 if (PageHWPoison(folio_page(folio, 0))) {
3206                         folio_unlock(folio);
3207                         folio_put(folio);
3208                         return ERR_PTR(-ECHILD);
3209                 }
3210                 folio_unlock(folio);
3211         }
3212         set_delayed_call(done, shmem_put_link, folio);
3213         return folio_address(folio);
3214 }
3215
3216 #ifdef CONFIG_TMPFS_XATTR
3217
3218 static int shmem_fileattr_get(struct dentry *dentry, struct fileattr *fa)
3219 {
3220         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3221
3222         fileattr_fill_flags(fa, info->fsflags & SHMEM_FL_USER_VISIBLE);
3223
3224         return 0;
3225 }
3226
3227 static int shmem_fileattr_set(struct user_namespace *mnt_userns,
3228                               struct dentry *dentry, struct fileattr *fa)
3229 {
3230         struct inode *inode = d_inode(dentry);
3231         struct shmem_inode_info *info = SHMEM_I(inode);
3232
3233         if (fileattr_has_fsx(fa))
3234                 return -EOPNOTSUPP;
3235         if (fa->flags & ~SHMEM_FL_USER_MODIFIABLE)
3236                 return -EOPNOTSUPP;
3237
3238         info->fsflags = (info->fsflags & ~SHMEM_FL_USER_MODIFIABLE) |
3239                 (fa->flags & SHMEM_FL_USER_MODIFIABLE);
3240
3241         shmem_set_inode_flags(inode, info->fsflags);
3242         inode->i_ctime = current_time(inode);
3243         inode_inc_iversion(inode);
3244         return 0;
3245 }
3246
3247 /*
3248  * Superblocks without xattr inode operations may get some security.* xattr
3249  * support from the LSM "for free". As soon as we have any other xattrs
3250  * like ACLs, we also need to implement the security.* handlers at
3251  * filesystem level, though.
3252  */
3253
3254 /*
3255  * Callback for security_inode_init_security() for acquiring xattrs.
3256  */
3257 static int shmem_initxattrs(struct inode *inode,
3258                             const struct xattr *xattr_array,
3259                             void *fs_info)
3260 {
3261         struct shmem_inode_info *info = SHMEM_I(inode);
3262         const struct xattr *xattr;
3263         struct simple_xattr *new_xattr;
3264         size_t len;
3265
3266         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3267                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3268                 if (!new_xattr)
3269                         return -ENOMEM;
3270
3271                 len = strlen(xattr->name) + 1;
3272                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3273                                           GFP_KERNEL);
3274                 if (!new_xattr->name) {
3275                         kvfree(new_xattr);
3276                         return -ENOMEM;
3277                 }
3278
3279                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3280                        XATTR_SECURITY_PREFIX_LEN);
3281                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3282                        xattr->name, len);
3283
3284                 simple_xattr_list_add(&info->xattrs, new_xattr);
3285         }
3286
3287         return 0;
3288 }
3289
3290 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3291                                    struct dentry *unused, struct inode *inode,
3292                                    const char *name, void *buffer, size_t size)
3293 {
3294         struct shmem_inode_info *info = SHMEM_I(inode);
3295
3296         name = xattr_full_name(handler, name);
3297         return simple_xattr_get(&info->xattrs, name, buffer, size);
3298 }
3299
3300 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3301                                    struct user_namespace *mnt_userns,
3302                                    struct dentry *unused, struct inode *inode,
3303                                    const char *name, const void *value,
3304                                    size_t size, int flags)
3305 {
3306         struct shmem_inode_info *info = SHMEM_I(inode);
3307         int err;
3308
3309         name = xattr_full_name(handler, name);
3310         err = simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3311         if (!err) {
3312                 inode->i_ctime = current_time(inode);
3313                 inode_inc_iversion(inode);
3314         }
3315         return err;
3316 }
3317
3318 static const struct xattr_handler shmem_security_xattr_handler = {
3319         .prefix = XATTR_SECURITY_PREFIX,
3320         .get = shmem_xattr_handler_get,
3321         .set = shmem_xattr_handler_set,
3322 };
3323
3324 static const struct xattr_handler shmem_trusted_xattr_handler = {
3325         .prefix = XATTR_TRUSTED_PREFIX,
3326         .get = shmem_xattr_handler_get,
3327         .set = shmem_xattr_handler_set,
3328 };
3329
3330 static const struct xattr_handler *shmem_xattr_handlers[] = {
3331 #ifdef CONFIG_TMPFS_POSIX_ACL
3332         &posix_acl_access_xattr_handler,
3333         &posix_acl_default_xattr_handler,
3334 #endif
3335         &shmem_security_xattr_handler,
3336         &shmem_trusted_xattr_handler,
3337         NULL
3338 };
3339
3340 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3341 {
3342         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3343         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3344 }
3345 #endif /* CONFIG_TMPFS_XATTR */
3346
3347 static const struct inode_operations shmem_short_symlink_operations = {
3348         .getattr        = shmem_getattr,
3349         .get_link       = simple_get_link,
3350 #ifdef CONFIG_TMPFS_XATTR
3351         .listxattr      = shmem_listxattr,
3352 #endif
3353 };
3354
3355 static const struct inode_operations shmem_symlink_inode_operations = {
3356         .getattr        = shmem_getattr,
3357         .get_link       = shmem_get_link,
3358 #ifdef CONFIG_TMPFS_XATTR
3359         .listxattr      = shmem_listxattr,
3360 #endif
3361 };
3362
3363 static struct dentry *shmem_get_parent(struct dentry *child)
3364 {
3365         return ERR_PTR(-ESTALE);
3366 }
3367
3368 static int shmem_match(struct inode *ino, void *vfh)
3369 {
3370         __u32 *fh = vfh;
3371         __u64 inum = fh[2];
3372         inum = (inum << 32) | fh[1];
3373         return ino->i_ino == inum && fh[0] == ino->i_generation;
3374 }
3375
3376 /* Find any alias of inode, but prefer a hashed alias */
3377 static struct dentry *shmem_find_alias(struct inode *inode)
3378 {
3379         struct dentry *alias = d_find_alias(inode);
3380
3381         return alias ?: d_find_any_alias(inode);
3382 }
3383
3384
3385 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3386                 struct fid *fid, int fh_len, int fh_type)
3387 {
3388         struct inode *inode;
3389         struct dentry *dentry = NULL;
3390         u64 inum;
3391
3392         if (fh_len < 3)
3393                 return NULL;
3394
3395         inum = fid->raw[2];
3396         inum = (inum << 32) | fid->raw[1];
3397
3398         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3399                         shmem_match, fid->raw);
3400         if (inode) {
3401                 dentry = shmem_find_alias(inode);
3402                 iput(inode);
3403         }
3404
3405         return dentry;
3406 }
3407
3408 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3409                                 struct inode *parent)
3410 {
3411         if (*len < 3) {
3412                 *len = 3;
3413                 return FILEID_INVALID;
3414         }
3415
3416         if (inode_unhashed(inode)) {
3417                 /* Unfortunately insert_inode_hash is not idempotent,
3418                  * so as we hash inodes here rather than at creation
3419                  * time, we need a lock to ensure we only try
3420                  * to do it once
3421                  */
3422                 static DEFINE_SPINLOCK(lock);
3423                 spin_lock(&lock);
3424                 if (inode_unhashed(inode))
3425                         __insert_inode_hash(inode,
3426                                             inode->i_ino + inode->i_generation);
3427                 spin_unlock(&lock);
3428         }
3429
3430         fh[0] = inode->i_generation;
3431         fh[1] = inode->i_ino;
3432         fh[2] = ((__u64)inode->i_ino) >> 32;
3433
3434         *len = 3;
3435         return 1;
3436 }
3437
3438 static const struct export_operations shmem_export_ops = {
3439         .get_parent     = shmem_get_parent,
3440         .encode_fh      = shmem_encode_fh,
3441         .fh_to_dentry   = shmem_fh_to_dentry,
3442 };
3443
3444 enum shmem_param {
3445         Opt_gid,
3446         Opt_huge,
3447         Opt_mode,
3448         Opt_mpol,
3449         Opt_nr_blocks,
3450         Opt_nr_inodes,
3451         Opt_size,
3452         Opt_uid,
3453         Opt_inode32,
3454         Opt_inode64,
3455 };
3456
3457 static const struct constant_table shmem_param_enums_huge[] = {
3458         {"never",       SHMEM_HUGE_NEVER },
3459         {"always",      SHMEM_HUGE_ALWAYS },
3460         {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3461         {"advise",      SHMEM_HUGE_ADVISE },
3462         {}
3463 };
3464
3465 const struct fs_parameter_spec shmem_fs_parameters[] = {
3466         fsparam_u32   ("gid",           Opt_gid),
3467         fsparam_enum  ("huge",          Opt_huge,  shmem_param_enums_huge),
3468         fsparam_u32oct("mode",          Opt_mode),
3469         fsparam_string("mpol",          Opt_mpol),
3470         fsparam_string("nr_blocks",     Opt_nr_blocks),
3471         fsparam_string("nr_inodes",     Opt_nr_inodes),
3472         fsparam_string("size",          Opt_size),
3473         fsparam_u32   ("uid",           Opt_uid),
3474         fsparam_flag  ("inode32",       Opt_inode32),
3475         fsparam_flag  ("inode64",       Opt_inode64),
3476         {}
3477 };
3478
3479 static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3480 {
3481         struct shmem_options *ctx = fc->fs_private;
3482         struct fs_parse_result result;
3483         unsigned long long size;
3484         char *rest;
3485         int opt;
3486
3487         opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3488         if (opt < 0)
3489                 return opt;
3490
3491         switch (opt) {
3492         case Opt_size:
3493                 size = memparse(param->string, &rest);
3494                 if (*rest == '%') {
3495                         size <<= PAGE_SHIFT;
3496                         size *= totalram_pages();
3497                         do_div(size, 100);
3498                         rest++;
3499                 }
3500                 if (*rest)
3501                         goto bad_value;
3502                 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3503                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3504                 break;
3505         case Opt_nr_blocks:
3506                 ctx->blocks = memparse(param->string, &rest);
3507                 if (*rest || ctx->blocks > S64_MAX)
3508                         goto bad_value;
3509                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3510                 break;
3511         case Opt_nr_inodes:
3512                 ctx->inodes = memparse(param->string, &rest);
3513                 if (*rest)
3514                         goto bad_value;
3515                 ctx->seen |= SHMEM_SEEN_INODES;
3516                 break;
3517         case Opt_mode:
3518                 ctx->mode = result.uint_32 & 07777;
3519                 break;
3520         case Opt_uid:
3521                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
3522                 if (!uid_valid(ctx->uid))
3523                         goto bad_value;
3524                 break;
3525         case Opt_gid:
3526                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
3527                 if (!gid_valid(ctx->gid))
3528                         goto bad_value;
3529                 break;
3530         case Opt_huge:
3531                 ctx->huge = result.uint_32;
3532                 if (ctx->huge != SHMEM_HUGE_NEVER &&
3533                     !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3534                       has_transparent_hugepage()))
3535                         goto unsupported_parameter;
3536                 ctx->seen |= SHMEM_SEEN_HUGE;
3537                 break;
3538         case Opt_mpol:
3539                 if (IS_ENABLED(CONFIG_NUMA)) {
3540                         mpol_put(ctx->mpol);
3541                         ctx->mpol = NULL;
3542                         if (mpol_parse_str(param->string, &ctx->mpol))
3543                                 goto bad_value;
3544                         break;
3545                 }
3546                 goto unsupported_parameter;
3547         case Opt_inode32:
3548                 ctx->full_inums = false;
3549                 ctx->seen |= SHMEM_SEEN_INUMS;
3550                 break;
3551         case Opt_inode64:
3552                 if (sizeof(ino_t) < 8) {
3553                         return invalfc(fc,
3554                                        "Cannot use inode64 with <64bit inums in kernel\n");
3555                 }
3556                 ctx->full_inums = true;
3557                 ctx->seen |= SHMEM_SEEN_INUMS;
3558                 break;
3559         }
3560         return 0;
3561
3562 unsupported_parameter:
3563         return invalfc(fc, "Unsupported parameter '%s'", param->key);
3564 bad_value:
3565         return invalfc(fc, "Bad value for '%s'", param->key);
3566 }
3567
3568 static int shmem_parse_options(struct fs_context *fc, void *data)
3569 {
3570         char *options = data;
3571
3572         if (options) {
3573                 int err = security_sb_eat_lsm_opts(options, &fc->security);
3574                 if (err)
3575                         return err;
3576         }
3577
3578         while (options != NULL) {
3579                 char *this_char = options;
3580                 for (;;) {
3581                         /*
3582                          * NUL-terminate this option: unfortunately,
3583                          * mount options form a comma-separated list,
3584                          * but mpol's nodelist may also contain commas.
3585                          */
3586                         options = strchr(options, ',');
3587                         if (options == NULL)
3588                                 break;
3589                         options++;
3590                         if (!isdigit(*options)) {
3591                                 options[-1] = '\0';
3592                                 break;
3593                         }
3594                 }
3595                 if (*this_char) {
3596                         char *value = strchr(this_char, '=');
3597                         size_t len = 0;
3598                         int err;
3599
3600                         if (value) {
3601                                 *value++ = '\0';
3602                                 len = strlen(value);
3603                         }
3604                         err = vfs_parse_fs_string(fc, this_char, value, len);
3605                         if (err < 0)
3606                                 return err;
3607                 }
3608         }
3609         return 0;
3610 }
3611
3612 /*
3613  * Reconfigure a shmem filesystem.
3614  *
3615  * Note that we disallow change from limited->unlimited blocks/inodes while any
3616  * are in use; but we must separately disallow unlimited->limited, because in
3617  * that case we have no record of how much is already in use.
3618  */
3619 static int shmem_reconfigure(struct fs_context *fc)
3620 {
3621         struct shmem_options *ctx = fc->fs_private;
3622         struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3623         unsigned long inodes;
3624         struct mempolicy *mpol = NULL;
3625         const char *err;
3626
3627         raw_spin_lock(&sbinfo->stat_lock);
3628         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3629
3630         if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3631                 if (!sbinfo->max_blocks) {
3632                         err = "Cannot retroactively limit size";
3633                         goto out;
3634                 }
3635                 if (percpu_counter_compare(&sbinfo->used_blocks,
3636                                            ctx->blocks) > 0) {
3637                         err = "Too small a size for current use";
3638                         goto out;
3639                 }
3640         }
3641         if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3642                 if (!sbinfo->max_inodes) {
3643                         err = "Cannot retroactively limit inodes";
3644                         goto out;
3645                 }
3646                 if (ctx->inodes < inodes) {
3647                         err = "Too few inodes for current use";
3648                         goto out;
3649                 }
3650         }
3651
3652         if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3653             sbinfo->next_ino > UINT_MAX) {
3654                 err = "Current inum too high to switch to 32-bit inums";
3655                 goto out;
3656         }
3657
3658         if (ctx->seen & SHMEM_SEEN_HUGE)
3659                 sbinfo->huge = ctx->huge;
3660         if (ctx->seen & SHMEM_SEEN_INUMS)
3661                 sbinfo->full_inums = ctx->full_inums;
3662         if (ctx->seen & SHMEM_SEEN_BLOCKS)
3663                 sbinfo->max_blocks  = ctx->blocks;
3664         if (ctx->seen & SHMEM_SEEN_INODES) {
3665                 sbinfo->max_inodes  = ctx->inodes;
3666                 sbinfo->free_inodes = ctx->inodes - inodes;
3667         }
3668
3669         /*
3670          * Preserve previous mempolicy unless mpol remount option was specified.
3671          */
3672         if (ctx->mpol) {
3673                 mpol = sbinfo->mpol;
3674                 sbinfo->mpol = ctx->mpol;       /* transfers initial ref */
3675                 ctx->mpol = NULL;
3676         }
3677         raw_spin_unlock(&sbinfo->stat_lock);
3678         mpol_put(mpol);
3679         return 0;
3680 out:
3681         raw_spin_unlock(&sbinfo->stat_lock);
3682         return invalfc(fc, "%s", err);
3683 }
3684
3685 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3686 {
3687         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3688
3689         if (sbinfo->max_blocks != shmem_default_max_blocks())
3690                 seq_printf(seq, ",size=%luk",
3691                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3692         if (sbinfo->max_inodes != shmem_default_max_inodes())
3693                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3694         if (sbinfo->mode != (0777 | S_ISVTX))
3695                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3696         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3697                 seq_printf(seq, ",uid=%u",
3698                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3699         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3700                 seq_printf(seq, ",gid=%u",
3701                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3702
3703         /*
3704          * Showing inode{64,32} might be useful even if it's the system default,
3705          * since then people don't have to resort to checking both here and
3706          * /proc/config.gz to confirm 64-bit inums were successfully applied
3707          * (which may not even exist if IKCONFIG_PROC isn't enabled).
3708          *
3709          * We hide it when inode64 isn't the default and we are using 32-bit
3710          * inodes, since that probably just means the feature isn't even under
3711          * consideration.
3712          *
3713          * As such:
3714          *
3715          *                     +-----------------+-----------------+
3716          *                     | TMPFS_INODE64=y | TMPFS_INODE64=n |
3717          *  +------------------+-----------------+-----------------+
3718          *  | full_inums=true  | show            | show            |
3719          *  | full_inums=false | show            | hide            |
3720          *  +------------------+-----------------+-----------------+
3721          *
3722          */
3723         if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3724                 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3725 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3726         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3727         if (sbinfo->huge)
3728                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3729 #endif
3730         shmem_show_mpol(seq, sbinfo->mpol);
3731         return 0;
3732 }
3733
3734 #endif /* CONFIG_TMPFS */
3735
3736 static void shmem_put_super(struct super_block *sb)
3737 {
3738         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3739
3740         free_percpu(sbinfo->ino_batch);
3741         percpu_counter_destroy(&sbinfo->used_blocks);
3742         mpol_put(sbinfo->mpol);
3743         kfree(sbinfo);
3744         sb->s_fs_info = NULL;
3745 }
3746
3747 static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3748 {
3749         struct shmem_options *ctx = fc->fs_private;
3750         struct inode *inode;
3751         struct shmem_sb_info *sbinfo;
3752
3753         /* Round up to L1_CACHE_BYTES to resist false sharing */
3754         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3755                                 L1_CACHE_BYTES), GFP_KERNEL);
3756         if (!sbinfo)
3757                 return -ENOMEM;
3758
3759         sb->s_fs_info = sbinfo;
3760
3761 #ifdef CONFIG_TMPFS
3762         /*
3763          * Per default we only allow half of the physical ram per
3764          * tmpfs instance, limiting inodes to one per page of lowmem;
3765          * but the internal instance is left unlimited.
3766          */
3767         if (!(sb->s_flags & SB_KERNMOUNT)) {
3768                 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3769                         ctx->blocks = shmem_default_max_blocks();
3770                 if (!(ctx->seen & SHMEM_SEEN_INODES))
3771                         ctx->inodes = shmem_default_max_inodes();
3772                 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3773                         ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3774         } else {
3775                 sb->s_flags |= SB_NOUSER;
3776         }
3777         sb->s_export_op = &shmem_export_ops;
3778         sb->s_flags |= SB_NOSEC | SB_I_VERSION;
3779 #else
3780         sb->s_flags |= SB_NOUSER;
3781 #endif
3782         sbinfo->max_blocks = ctx->blocks;
3783         sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3784         if (sb->s_flags & SB_KERNMOUNT) {
3785                 sbinfo->ino_batch = alloc_percpu(ino_t);
3786                 if (!sbinfo->ino_batch)
3787                         goto failed;
3788         }
3789         sbinfo->uid = ctx->uid;
3790         sbinfo->gid = ctx->gid;
3791         sbinfo->full_inums = ctx->full_inums;
3792         sbinfo->mode = ctx->mode;
3793         sbinfo->huge = ctx->huge;
3794         sbinfo->mpol = ctx->mpol;
3795         ctx->mpol = NULL;
3796
3797         raw_spin_lock_init(&sbinfo->stat_lock);
3798         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3799                 goto failed;
3800         spin_lock_init(&sbinfo->shrinklist_lock);
3801         INIT_LIST_HEAD(&sbinfo->shrinklist);
3802
3803         sb->s_maxbytes = MAX_LFS_FILESIZE;
3804         sb->s_blocksize = PAGE_SIZE;
3805         sb->s_blocksize_bits = PAGE_SHIFT;
3806         sb->s_magic = TMPFS_MAGIC;
3807         sb->s_op = &shmem_ops;
3808         sb->s_time_gran = 1;
3809 #ifdef CONFIG_TMPFS_XATTR
3810         sb->s_xattr = shmem_xattr_handlers;
3811 #endif
3812 #ifdef CONFIG_TMPFS_POSIX_ACL
3813         sb->s_flags |= SB_POSIXACL;
3814 #endif
3815         uuid_gen(&sb->s_uuid);
3816
3817         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3818         if (!inode)
3819                 goto failed;
3820         inode->i_uid = sbinfo->uid;
3821         inode->i_gid = sbinfo->gid;
3822         sb->s_root = d_make_root(inode);
3823         if (!sb->s_root)
3824                 goto failed;
3825         return 0;
3826
3827 failed:
3828         shmem_put_super(sb);
3829         return -ENOMEM;
3830 }
3831
3832 static int shmem_get_tree(struct fs_context *fc)
3833 {
3834         return get_tree_nodev(fc, shmem_fill_super);
3835 }
3836
3837 static void shmem_free_fc(struct fs_context *fc)
3838 {
3839         struct shmem_options *ctx = fc->fs_private;
3840
3841         if (ctx) {
3842                 mpol_put(ctx->mpol);
3843                 kfree(ctx);
3844         }
3845 }
3846
3847 static const struct fs_context_operations shmem_fs_context_ops = {
3848         .free                   = shmem_free_fc,
3849         .get_tree               = shmem_get_tree,
3850 #ifdef CONFIG_TMPFS
3851         .parse_monolithic       = shmem_parse_options,
3852         .parse_param            = shmem_parse_one,
3853         .reconfigure            = shmem_reconfigure,
3854 #endif
3855 };
3856
3857 static struct kmem_cache *shmem_inode_cachep;
3858
3859 static struct inode *shmem_alloc_inode(struct super_block *sb)
3860 {
3861         struct shmem_inode_info *info;
3862         info = alloc_inode_sb(sb, shmem_inode_cachep, GFP_KERNEL);
3863         if (!info)
3864                 return NULL;
3865         return &info->vfs_inode;
3866 }
3867
3868 static void shmem_free_in_core_inode(struct inode *inode)
3869 {
3870         if (S_ISLNK(inode->i_mode))
3871                 kfree(inode->i_link);
3872         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3873 }
3874
3875 static void shmem_destroy_inode(struct inode *inode)
3876 {
3877         if (S_ISREG(inode->i_mode))
3878                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3879 }
3880
3881 static void shmem_init_inode(void *foo)
3882 {
3883         struct shmem_inode_info *info = foo;
3884         inode_init_once(&info->vfs_inode);
3885 }
3886
3887 static void shmem_init_inodecache(void)
3888 {
3889         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3890                                 sizeof(struct shmem_inode_info),
3891                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3892 }
3893
3894 static void shmem_destroy_inodecache(void)
3895 {
3896         kmem_cache_destroy(shmem_inode_cachep);
3897 }
3898
3899 /* Keep the page in page cache instead of truncating it */
3900 static int shmem_error_remove_page(struct address_space *mapping,
3901                                    struct page *page)
3902 {
3903         return 0;
3904 }
3905
3906 const struct address_space_operations shmem_aops = {
3907         .writepage      = shmem_writepage,
3908         .dirty_folio    = noop_dirty_folio,
3909 #ifdef CONFIG_TMPFS
3910         .write_begin    = shmem_write_begin,
3911         .write_end      = shmem_write_end,
3912 #endif
3913 #ifdef CONFIG_MIGRATION
3914         .migrate_folio  = migrate_folio,
3915 #endif
3916         .error_remove_page = shmem_error_remove_page,
3917 };
3918 EXPORT_SYMBOL(shmem_aops);
3919
3920 static const struct file_operations shmem_file_operations = {
3921         .mmap           = shmem_mmap,
3922         .get_unmapped_area = shmem_get_unmapped_area,
3923 #ifdef CONFIG_TMPFS
3924         .llseek         = shmem_file_llseek,
3925         .read_iter      = shmem_file_read_iter,
3926         .write_iter     = generic_file_write_iter,
3927         .fsync          = noop_fsync,
3928         .splice_read    = generic_file_splice_read,
3929         .splice_write   = iter_file_splice_write,
3930         .fallocate      = shmem_fallocate,
3931 #endif
3932 };
3933
3934 static const struct inode_operations shmem_inode_operations = {
3935         .getattr        = shmem_getattr,
3936         .setattr        = shmem_setattr,
3937 #ifdef CONFIG_TMPFS_XATTR
3938         .listxattr      = shmem_listxattr,
3939         .set_acl        = simple_set_acl,
3940         .fileattr_get   = shmem_fileattr_get,
3941         .fileattr_set   = shmem_fileattr_set,
3942 #endif
3943 };
3944
3945 static const struct inode_operations shmem_dir_inode_operations = {
3946 #ifdef CONFIG_TMPFS
3947         .getattr        = shmem_getattr,
3948         .create         = shmem_create,
3949         .lookup         = simple_lookup,
3950         .link           = shmem_link,
3951         .unlink         = shmem_unlink,
3952         .symlink        = shmem_symlink,
3953         .mkdir          = shmem_mkdir,
3954         .rmdir          = shmem_rmdir,
3955         .mknod          = shmem_mknod,
3956         .rename         = shmem_rename2,
3957         .tmpfile        = shmem_tmpfile,
3958 #endif
3959 #ifdef CONFIG_TMPFS_XATTR
3960         .listxattr      = shmem_listxattr,
3961         .fileattr_get   = shmem_fileattr_get,
3962         .fileattr_set   = shmem_fileattr_set,
3963 #endif
3964 #ifdef CONFIG_TMPFS_POSIX_ACL
3965         .setattr        = shmem_setattr,
3966         .set_acl        = simple_set_acl,
3967 #endif
3968 };
3969
3970 static const struct inode_operations shmem_special_inode_operations = {
3971         .getattr        = shmem_getattr,
3972 #ifdef CONFIG_TMPFS_XATTR
3973         .listxattr      = shmem_listxattr,
3974 #endif
3975 #ifdef CONFIG_TMPFS_POSIX_ACL
3976         .setattr        = shmem_setattr,
3977         .set_acl        = simple_set_acl,
3978 #endif
3979 };
3980
3981 static const struct super_operations shmem_ops = {
3982         .alloc_inode    = shmem_alloc_inode,
3983         .free_inode     = shmem_free_in_core_inode,
3984         .destroy_inode  = shmem_destroy_inode,
3985 #ifdef CONFIG_TMPFS
3986         .statfs         = shmem_statfs,
3987         .show_options   = shmem_show_options,
3988 #endif
3989         .evict_inode    = shmem_evict_inode,
3990         .drop_inode     = generic_delete_inode,
3991         .put_super      = shmem_put_super,
3992 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3993         .nr_cached_objects      = shmem_unused_huge_count,
3994         .free_cached_objects    = shmem_unused_huge_scan,
3995 #endif
3996 };
3997
3998 static const struct vm_operations_struct shmem_vm_ops = {
3999         .fault          = shmem_fault,
4000         .map_pages      = filemap_map_pages,
4001 #ifdef CONFIG_NUMA
4002         .set_policy     = shmem_set_policy,
4003         .get_policy     = shmem_get_policy,
4004 #endif
4005 };
4006
4007 int shmem_init_fs_context(struct fs_context *fc)
4008 {
4009         struct shmem_options *ctx;
4010
4011         ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
4012         if (!ctx)
4013                 return -ENOMEM;
4014
4015         ctx->mode = 0777 | S_ISVTX;
4016         ctx->uid = current_fsuid();
4017         ctx->gid = current_fsgid();
4018
4019         fc->fs_private = ctx;
4020         fc->ops = &shmem_fs_context_ops;
4021         return 0;
4022 }
4023
4024 static struct file_system_type shmem_fs_type = {
4025         .owner          = THIS_MODULE,
4026         .name           = "tmpfs",
4027         .init_fs_context = shmem_init_fs_context,
4028 #ifdef CONFIG_TMPFS
4029         .parameters     = shmem_fs_parameters,
4030 #endif
4031         .kill_sb        = kill_litter_super,
4032         .fs_flags       = FS_USERNS_MOUNT,
4033 };
4034
4035 void __init shmem_init(void)
4036 {
4037         int error;
4038
4039         shmem_init_inodecache();
4040
4041         error = register_filesystem(&shmem_fs_type);
4042         if (error) {
4043                 pr_err("Could not register tmpfs\n");
4044                 goto out2;
4045         }
4046
4047         shm_mnt = kern_mount(&shmem_fs_type);
4048         if (IS_ERR(shm_mnt)) {
4049                 error = PTR_ERR(shm_mnt);
4050                 pr_err("Could not kern_mount tmpfs\n");
4051                 goto out1;
4052         }
4053
4054 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4055         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4056                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4057         else
4058                 shmem_huge = SHMEM_HUGE_NEVER; /* just in case it was patched */
4059 #endif
4060         return;
4061
4062 out1:
4063         unregister_filesystem(&shmem_fs_type);
4064 out2:
4065         shmem_destroy_inodecache();
4066         shm_mnt = ERR_PTR(error);
4067 }
4068
4069 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4070 static ssize_t shmem_enabled_show(struct kobject *kobj,
4071                                   struct kobj_attribute *attr, char *buf)
4072 {
4073         static const int values[] = {
4074                 SHMEM_HUGE_ALWAYS,
4075                 SHMEM_HUGE_WITHIN_SIZE,
4076                 SHMEM_HUGE_ADVISE,
4077                 SHMEM_HUGE_NEVER,
4078                 SHMEM_HUGE_DENY,
4079                 SHMEM_HUGE_FORCE,
4080         };
4081         int len = 0;
4082         int i;
4083
4084         for (i = 0; i < ARRAY_SIZE(values); i++) {
4085                 len += sysfs_emit_at(buf, len,
4086                                      shmem_huge == values[i] ? "%s[%s]" : "%s%s",
4087                                      i ? " " : "",
4088                                      shmem_format_huge(values[i]));
4089         }
4090
4091         len += sysfs_emit_at(buf, len, "\n");
4092
4093         return len;
4094 }
4095
4096 static ssize_t shmem_enabled_store(struct kobject *kobj,
4097                 struct kobj_attribute *attr, const char *buf, size_t count)
4098 {
4099         char tmp[16];
4100         int huge;
4101
4102         if (count + 1 > sizeof(tmp))
4103                 return -EINVAL;
4104         memcpy(tmp, buf, count);
4105         tmp[count] = '\0';
4106         if (count && tmp[count - 1] == '\n')
4107                 tmp[count - 1] = '\0';
4108
4109         huge = shmem_parse_huge(tmp);
4110         if (huge == -EINVAL)
4111                 return -EINVAL;
4112         if (!has_transparent_hugepage() &&
4113                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4114                 return -EINVAL;
4115
4116         shmem_huge = huge;
4117         if (shmem_huge > SHMEM_HUGE_DENY)
4118                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4119         return count;
4120 }
4121
4122 struct kobj_attribute shmem_enabled_attr = __ATTR_RW(shmem_enabled);
4123 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4124
4125 #else /* !CONFIG_SHMEM */
4126
4127 /*
4128  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4129  *
4130  * This is intended for small system where the benefits of the full
4131  * shmem code (swap-backed and resource-limited) are outweighed by
4132  * their complexity. On systems without swap this code should be
4133  * effectively equivalent, but much lighter weight.
4134  */
4135
4136 static struct file_system_type shmem_fs_type = {
4137         .name           = "tmpfs",
4138         .init_fs_context = ramfs_init_fs_context,
4139         .parameters     = ramfs_fs_parameters,
4140         .kill_sb        = kill_litter_super,
4141         .fs_flags       = FS_USERNS_MOUNT,
4142 };
4143
4144 void __init shmem_init(void)
4145 {
4146         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4147
4148         shm_mnt = kern_mount(&shmem_fs_type);
4149         BUG_ON(IS_ERR(shm_mnt));
4150 }
4151
4152 int shmem_unuse(unsigned int type)
4153 {
4154         return 0;
4155 }
4156
4157 int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
4158 {
4159         return 0;
4160 }
4161
4162 void shmem_unlock_mapping(struct address_space *mapping)
4163 {
4164 }
4165
4166 #ifdef CONFIG_MMU
4167 unsigned long shmem_get_unmapped_area(struct file *file,
4168                                       unsigned long addr, unsigned long len,
4169                                       unsigned long pgoff, unsigned long flags)
4170 {
4171         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4172 }
4173 #endif
4174
4175 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4176 {
4177         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4178 }
4179 EXPORT_SYMBOL_GPL(shmem_truncate_range);
4180
4181 #define shmem_vm_ops                            generic_file_vm_ops
4182 #define shmem_file_operations                   ramfs_file_operations
4183 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
4184 #define shmem_acct_size(flags, size)            0
4185 #define shmem_unacct_size(flags, size)          do {} while (0)
4186
4187 #endif /* CONFIG_SHMEM */
4188
4189 /* common code */
4190
4191 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4192                                        unsigned long flags, unsigned int i_flags)
4193 {
4194         struct inode *inode;
4195         struct file *res;
4196
4197         if (IS_ERR(mnt))
4198                 return ERR_CAST(mnt);
4199
4200         if (size < 0 || size > MAX_LFS_FILESIZE)
4201                 return ERR_PTR(-EINVAL);
4202
4203         if (shmem_acct_size(flags, size))
4204                 return ERR_PTR(-ENOMEM);
4205
4206         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4207                                 flags);
4208         if (unlikely(!inode)) {
4209                 shmem_unacct_size(flags, size);
4210                 return ERR_PTR(-ENOSPC);
4211         }
4212         inode->i_flags |= i_flags;
4213         inode->i_size = size;
4214         clear_nlink(inode);     /* It is unlinked */
4215         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4216         if (!IS_ERR(res))
4217                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4218                                 &shmem_file_operations);
4219         if (IS_ERR(res))
4220                 iput(inode);
4221         return res;
4222 }
4223
4224 /**
4225  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4226  *      kernel internal.  There will be NO LSM permission checks against the
4227  *      underlying inode.  So users of this interface must do LSM checks at a
4228  *      higher layer.  The users are the big_key and shm implementations.  LSM
4229  *      checks are provided at the key or shm level rather than the inode.
4230  * @name: name for dentry (to be seen in /proc/<pid>/maps
4231  * @size: size to be set for the file
4232  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4233  */
4234 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4235 {
4236         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4237 }
4238
4239 /**
4240  * shmem_file_setup - get an unlinked file living in tmpfs
4241  * @name: name for dentry (to be seen in /proc/<pid>/maps
4242  * @size: size to be set for the file
4243  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4244  */
4245 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4246 {
4247         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4248 }
4249 EXPORT_SYMBOL_GPL(shmem_file_setup);
4250
4251 /**
4252  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4253  * @mnt: the tmpfs mount where the file will be created
4254  * @name: name for dentry (to be seen in /proc/<pid>/maps
4255  * @size: size to be set for the file
4256  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4257  */
4258 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4259                                        loff_t size, unsigned long flags)
4260 {
4261         return __shmem_file_setup(mnt, name, size, flags, 0);
4262 }
4263 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4264
4265 /**
4266  * shmem_zero_setup - setup a shared anonymous mapping
4267  * @vma: the vma to be mmapped is prepared by do_mmap
4268  */
4269 int shmem_zero_setup(struct vm_area_struct *vma)
4270 {
4271         struct file *file;
4272         loff_t size = vma->vm_end - vma->vm_start;
4273
4274         /*
4275          * Cloning a new file under mmap_lock leads to a lock ordering conflict
4276          * between XFS directory reading and selinux: since this file is only
4277          * accessible to the user through its mapping, use S_PRIVATE flag to
4278          * bypass file security, in the same way as shmem_kernel_file_setup().
4279          */
4280         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4281         if (IS_ERR(file))
4282                 return PTR_ERR(file);
4283
4284         if (vma->vm_file)
4285                 fput(vma->vm_file);
4286         vma->vm_file = file;
4287         vma->vm_ops = &shmem_vm_ops;
4288
4289         return 0;
4290 }
4291
4292 /**
4293  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4294  * @mapping:    the page's address_space
4295  * @index:      the page index
4296  * @gfp:        the page allocator flags to use if allocating
4297  *
4298  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4299  * with any new page allocations done using the specified allocation flags.
4300  * But read_cache_page_gfp() uses the ->read_folio() method: which does not
4301  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4302  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4303  *
4304  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4305  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4306  */
4307 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4308                                          pgoff_t index, gfp_t gfp)
4309 {
4310 #ifdef CONFIG_SHMEM
4311         struct inode *inode = mapping->host;
4312         struct folio *folio;
4313         struct page *page;
4314         int error;
4315
4316         BUG_ON(!shmem_mapping(mapping));
4317         error = shmem_get_folio_gfp(inode, index, &folio, SGP_CACHE,
4318                                   gfp, NULL, NULL, NULL);
4319         if (error)
4320                 return ERR_PTR(error);
4321
4322         folio_unlock(folio);
4323         page = folio_file_page(folio, index);
4324         if (PageHWPoison(page)) {
4325                 folio_put(folio);
4326                 return ERR_PTR(-EIO);
4327         }
4328
4329         return page;
4330 #else
4331         /*
4332          * The tiny !SHMEM case uses ramfs without swap
4333          */
4334         return read_cache_page_gfp(mapping, index, gfp);
4335 #endif
4336 }
4337 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);