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 "swap.h"
  42
  43 static struct vfsmount *shm_mnt;
  44
  45 #ifdef CONFIG_SHMEM
  46 /*
  47  * This virtual memory filesystem is heavily based on the ramfs. It
  48  * extends ramfs by the ability to use swap and honor resource limits
  49  * which makes it a completely usable filesystem.
  50  */
  51
  52 #include <linux/xattr.h>
  53 #include <linux/exportfs.h>
  54 #include <linux/posix_acl.h>
  55 #include <linux/posix_acl_xattr.h>
  56 #include <linux/mman.h>
  57 #include <linux/string.h>
  58 #include <linux/slab.h>
  59 #include <linux/backing-dev.h>
  60 #include <linux/shmem_fs.h>
  61 #include <linux/writeback.h>
  62 #include <linux/pagevec.h>
  63 #include <linux/percpu_counter.h>
  64 #include <linux/falloc.h>
  65 #include <linux/splice.h>
  66 #include <linux/security.h>
  67 #include <linux/swapops.h>
  68 #include <linux/mempolicy.h>
  69 #include <linux/namei.h>
  70 #include <linux/ctype.h>
  71 #include <linux/migrate.h>
  72 #include <linux/highmem.h>
  73 #include <linux/seq_file.h>
  74 #include <linux/magic.h>
  75 #include <linux/syscalls.h>
  76 #include <linux/fcntl.h>
  77 #include <uapi/linux/memfd.h>
  78 #include <linux/userfaultfd_k.h>
  79 #include <linux/rmap.h>
  80 #include <linux/uuid.h>
  81
  82 #include <linux/uaccess.h>
  83
  84 #include "internal.h"
  85
  86 #define BLOCKS_PER_PAGE  (PAGE_SIZE/512)
  87 #define VM_ACCT(size)    (PAGE_ALIGN(size) >> PAGE_SHIFT)
  88
  89 /* Pretend that each entry is of this size in directory's i_size */
  90 #define BOGO_DIRENT_SIZE 20
  91
  92 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  93 #define SHORT_SYMLINK_LEN 128
  94
  95 /*
  96  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
  97  * inode->i_private (with i_rwsem making sure that it has only one user at
  98  * a time): we would prefer not to enlarge the shmem inode just for that.
  99  */
 100 struct shmem_falloc {
 101         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
 102         pgoff_t start;          /* start of range currently being fallocated */
 103         pgoff_t next;           /* the next page offset to be fallocated */
 104         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
 105         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
 106 };
 107
 108 struct shmem_options {
 109         unsigned long long blocks;
 110         unsigned long long inodes;
 111         struct mempolicy *mpol;
 112         kuid_t uid;
 113         kgid_t gid;
 114         umode_t mode;
 115         bool full_inums;
 116         int huge;
 117         int seen;
 118 #define SHMEM_SEEN_BLOCKS 1
 119 #define SHMEM_SEEN_INODES 2
 120 #define SHMEM_SEEN_HUGE 4
 121 #define SHMEM_SEEN_INUMS 8
 122 };
 123
 124 #ifdef CONFIG_TMPFS
 125 static unsigned long shmem_default_max_blocks(void)
 126 {
 127         return totalram_pages() / 2;
 128 }
 129
 130 static unsigned long shmem_default_max_inodes(void)
 131 {
 132         unsigned long nr_pages = totalram_pages();
 133
 134         return min(nr_pages - totalhigh_pages(), nr_pages / 2);
 135 }
 136 #endif
 137
 138 static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
 139                              struct folio **foliop, enum sgp_type sgp,
 140                              gfp_t gfp, struct vm_area_struct *vma,
 141                              vm_fault_t *fault_type);
 142 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 143                 struct page **pagep, enum sgp_type sgp,
 144                 gfp_t gfp, struct vm_area_struct *vma,
 145                 struct vm_fault *vmf, vm_fault_t *fault_type);
 146
 147 int shmem_getpage(struct inode *inode, pgoff_t index,
 148                 struct page **pagep, enum sgp_type sgp)
 149 {
 150         return shmem_getpage_gfp(inode, index, pagep, sgp,
 151                 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
 152 }
 153
 154 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 155 {
 156         return sb->s_fs_info;
 157 }
 158
 159 /*
 160  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 161  * for shared memory and for shared anonymous (/dev/zero) mappings
 162  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 163  * consistent with the pre-accounting of private mappings ...
 164  */
 165 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 166 {
 167         return (flags & VM_NORESERVE) ?
 168                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 169 }
 170
 171 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 172 {
 173         if (!(flags & VM_NORESERVE))
 174                 vm_unacct_memory(VM_ACCT(size));
 175 }
 176
 177 static inline int shmem_reacct_size(unsigned long flags,
 178                 loff_t oldsize, loff_t newsize)
 179 {
 180         if (!(flags & VM_NORESERVE)) {
 181                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
 182                         return security_vm_enough_memory_mm(current->mm,
 183                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
 184                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
 185                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
 186         }
 187         return 0;
 188 }
 189
 190 /*
 191  * ... whereas tmpfs objects are accounted incrementally as
 192  * pages are allocated, in order to allow large sparse files.
 193  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 194  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 195  */
 196 static inline int shmem_acct_block(unsigned long flags, long pages)
 197 {
 198         if (!(flags & VM_NORESERVE))
 199                 return 0;
 200
 201         return security_vm_enough_memory_mm(current->mm,
 202                         pages * VM_ACCT(PAGE_SIZE));
 203 }
 204
 205 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 206 {
 207         if (flags & VM_NORESERVE)
 208                 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
 209 }
 210
 211 static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
 212 {
 213         struct shmem_inode_info *info = SHMEM_I(inode);
 214         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 215
 216         if (shmem_acct_block(info->flags, pages))
 217                 return false;
 218
 219         if (sbinfo->max_blocks) {
 220                 if (percpu_counter_compare(&sbinfo->used_blocks,
 221                                            sbinfo->max_blocks - pages) > 0)
 222                         goto unacct;
 223                 percpu_counter_add(&sbinfo->used_blocks, pages);
 224         }
 225
 226         return true;
 227
 228 unacct:
 229         shmem_unacct_blocks(info->flags, pages);
 230         return false;
 231 }
 232
 233 static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
 234 {
 235         struct shmem_inode_info *info = SHMEM_I(inode);
 236         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 237
 238         if (sbinfo->max_blocks)
 239                 percpu_counter_sub(&sbinfo->used_blocks, pages);
 240         shmem_unacct_blocks(info->flags, pages);
 241 }
 242
 243 static const struct super_operations shmem_ops;
 244 const struct address_space_operations shmem_aops;
 245 static const struct file_operations shmem_file_operations;
 246 static const struct inode_operations shmem_inode_operations;
 247 static const struct inode_operations shmem_dir_inode_operations;
 248 static const struct inode_operations shmem_special_inode_operations;
 249 static const struct vm_operations_struct shmem_vm_ops;
 250 static struct file_system_type shmem_fs_type;
 251
 252 bool vma_is_shmem(struct vm_area_struct *vma)
 253 {
 254         return vma->vm_ops == &shmem_vm_ops;
 255 }
 256
 257 static LIST_HEAD(shmem_swaplist);
 258 static DEFINE_MUTEX(shmem_swaplist_mutex);
 259
 260 /*
 261  * shmem_reserve_inode() performs bookkeeping to reserve a shmem inode, and
 262  * produces a novel ino for the newly allocated inode.
 263  *
 264  * It may also be called when making a hard link to permit the space needed by
 265  * each dentry. However, in that case, no new inode number is needed since that
 266  * internally draws from another pool of inode numbers (currently global
 267  * get_next_ino()). This case is indicated by passing NULL as inop.
 268  */
 269 #define SHMEM_INO_BATCH 1024
 270 static int shmem_reserve_inode(struct super_block *sb, ino_t *inop)
 271 {
 272         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 273         ino_t ino;
 274
 275         if (!(sb->s_flags & SB_KERNMOUNT)) {
 276                 raw_spin_lock(&sbinfo->stat_lock);
 277                 if (sbinfo->max_inodes) {
 278                         if (!sbinfo->free_inodes) {
 279                                 raw_spin_unlock(&sbinfo->stat_lock);
 280                                 return -ENOSPC;
 281                         }
 282                         sbinfo->free_inodes--;
 283                 }
 284                 if (inop) {
 285                         ino = sbinfo->next_ino++;
 286                         if (unlikely(is_zero_ino(ino)))
 287                                 ino = sbinfo->next_ino++;
 288                         if (unlikely(!sbinfo->full_inums &&
 289                                      ino > UINT_MAX)) {
 290                                 /*
 291                                  * Emulate get_next_ino uint wraparound for
 292                                  * compatibility
 293                                  */
 294                                 if (IS_ENABLED(CONFIG_64BIT))
 295                                         pr_warn("%s: inode number overflow on device %d, consider using inode64 mount option\n",
 296                                                 __func__, MINOR(sb->s_dev));
 297                                 sbinfo->next_ino = 1;
 298                                 ino = sbinfo->next_ino++;
 299                         }
 300                         *inop = ino;
 301                 }
 302                 raw_spin_unlock(&sbinfo->stat_lock);
 303         } else if (inop) {
 304                 /*
 305                  * __shmem_file_setup, one of our callers, is lock-free: it
 306                  * doesn't hold stat_lock in shmem_reserve_inode since
 307                  * max_inodes is always 0, and is called from potentially
 308                  * unknown contexts. As such, use a per-cpu batched allocator
 309                  * which doesn't require the per-sb stat_lock unless we are at
 310                  * the batch boundary.
 311                  *
 312                  * We don't need to worry about inode{32,64} since SB_KERNMOUNT
 313                  * shmem mounts are not exposed to userspace, so we don't need
 314                  * to worry about things like glibc compatibility.
 315                  */
 316                 ino_t *next_ino;
 317
 318                 next_ino = per_cpu_ptr(sbinfo->ino_batch, get_cpu());
 319                 ino = *next_ino;
 320                 if (unlikely(ino % SHMEM_INO_BATCH == 0)) {
 321                         raw_spin_lock(&sbinfo->stat_lock);
 322                         ino = sbinfo->next_ino;
 323                         sbinfo->next_ino += SHMEM_INO_BATCH;
 324                         raw_spin_unlock(&sbinfo->stat_lock);
 325                         if (unlikely(is_zero_ino(ino)))
 326                                 ino++;
 327                 }
 328                 *inop = ino;
 329                 *next_ino = ++ino;
 330                 put_cpu();
 331         }
 332
 333         return 0;
 334 }
 335
 336 static void shmem_free_inode(struct super_block *sb)
 337 {
 338         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 339         if (sbinfo->max_inodes) {
 340                 raw_spin_lock(&sbinfo->stat_lock);
 341                 sbinfo->free_inodes++;
 342                 raw_spin_unlock(&sbinfo->stat_lock);
 343         }
 344 }
 345
 346 /**
 347  * shmem_recalc_inode - recalculate the block usage of an inode
 348  * @inode: inode to recalc
 349  *
 350  * We have to calculate the free blocks since the mm can drop
 351  * undirtied hole pages behind our back.
 352  *
 353  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 354  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 355  *
 356  * It has to be called with the spinlock held.
 357  */
 358 static void shmem_recalc_inode(struct inode *inode)
 359 {
 360         struct shmem_inode_info *info = SHMEM_I(inode);
 361         long freed;
 362
 363         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 364         if (freed > 0) {
 365                 info->alloced -= freed;
 366                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 367                 shmem_inode_unacct_blocks(inode, freed);
 368         }
 369 }
 370
 371 bool shmem_charge(struct inode *inode, long pages)
 372 {
 373         struct shmem_inode_info *info = SHMEM_I(inode);
 374         unsigned long flags;
 375
 376         if (!shmem_inode_acct_block(inode, pages))
 377                 return false;
 378
 379         /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
 380         inode->i_mapping->nrpages += pages;
 381
 382         spin_lock_irqsave(&info->lock, flags);
 383         info->alloced += pages;
 384         inode->i_blocks += pages * BLOCKS_PER_PAGE;
 385         shmem_recalc_inode(inode);
 386         spin_unlock_irqrestore(&info->lock, flags);
 387
 388         return true;
 389 }
 390
 391 void shmem_uncharge(struct inode *inode, long pages)
 392 {
 393         struct shmem_inode_info *info = SHMEM_I(inode);
 394         unsigned long flags;
 395
 396         /* nrpages adjustment done by __filemap_remove_folio() or caller */
 397
 398         spin_lock_irqsave(&info->lock, flags);
 399         info->alloced -= pages;
 400         inode->i_blocks -= pages * BLOCKS_PER_PAGE;
 401         shmem_recalc_inode(inode);
 402         spin_unlock_irqrestore(&info->lock, flags);
 403
 404         shmem_inode_unacct_blocks(inode, pages);
 405 }
 406
 407 /*
 408  * Replace item expected in xarray by a new item, while holding xa_lock.
 409  */
 410 static int shmem_replace_entry(struct address_space *mapping,
 411                         pgoff_t index, void *expected, void *replacement)
 412 {
 413         XA_STATE(xas, &mapping->i_pages, index);
 414         void *item;
 415
 416         VM_BUG_ON(!expected);
 417         VM_BUG_ON(!replacement);
 418         item = xas_load(&xas);
 419         if (item != expected)
 420                 return -ENOENT;
 421         xas_store(&xas, replacement);
 422         return 0;
 423 }
 424
 425 /*
 426  * Sometimes, before we decide whether to proceed or to fail, we must check
 427  * that an entry was not already brought back from swap by a racing thread.
 428  *
 429  * Checking page is not enough: by the time a SwapCache page is locked, it
 430  * might be reused, and again be SwapCache, using the same swap as before.
 431  */
 432 static bool shmem_confirm_swap(struct address_space *mapping,
 433                                pgoff_t index, swp_entry_t swap)
 434 {
 435         return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
 436 }
 437
 438 /*
 439  * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
 440  *
 441  * SHMEM_HUGE_NEVER:
 442  *      disables huge pages for the mount;
 443  * SHMEM_HUGE_ALWAYS:
 444  *      enables huge pages for the mount;
 445  * SHMEM_HUGE_WITHIN_SIZE:
 446  *      only allocate huge pages if the page will be fully within i_size,
 447  *      also respect fadvise()/madvise() hints;
 448  * SHMEM_HUGE_ADVISE:
 449  *      only allocate huge pages if requested with fadvise()/madvise();
 450  */
 451
 452 #define SHMEM_HUGE_NEVER        0
 453 #define SHMEM_HUGE_ALWAYS       1
 454 #define SHMEM_HUGE_WITHIN_SIZE  2
 455 #define SHMEM_HUGE_ADVISE       3
 456
 457 /*
 458  * Special values.
 459  * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
 460  *
 461  * SHMEM_HUGE_DENY:
 462  *      disables huge on shm_mnt and all mounts, for emergency use;
 463  * SHMEM_HUGE_FORCE:
 464  *      enables huge on shm_mnt and all mounts, w/o needing option, for testing;
 465  *
 466  */
 467 #define SHMEM_HUGE_DENY         (-1)
 468 #define SHMEM_HUGE_FORCE        (-2)
 469
 470 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 471 /* ifdef here to avoid bloating shmem.o when not necessary */
 472
 473 static int shmem_huge __read_mostly = SHMEM_HUGE_NEVER;
 474
 475 bool shmem_is_huge(struct vm_area_struct *vma,
 476                    struct inode *inode, pgoff_t index)
 477 {
 478         loff_t i_size;
 479
 480         if (!S_ISREG(inode->i_mode))
 481                 return false;
 482         if (shmem_huge == SHMEM_HUGE_DENY)
 483                 return false;
 484         if (vma && ((vma->vm_flags & VM_NOHUGEPAGE) ||
 485             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)))
 486                 return false;
 487         if (shmem_huge == SHMEM_HUGE_FORCE)
 488                 return true;
 489
 490         switch (SHMEM_SB(inode->i_sb)->huge) {
 491         case SHMEM_HUGE_ALWAYS:
 492                 return true;
 493         case SHMEM_HUGE_WITHIN_SIZE:
 494                 index = round_up(index + 1, HPAGE_PMD_NR);
 495                 i_size = round_up(i_size_read(inode), PAGE_SIZE);
 496                 if (i_size >> PAGE_SHIFT >= index)
 497                         return true;
 498                 fallthrough;
 499         case SHMEM_HUGE_ADVISE:
 500                 if (vma && (vma->vm_flags & VM_HUGEPAGE))
 501                         return true;
 502                 fallthrough;
 503         default:
 504                 return false;
 505         }
 506 }
 507
 508 #if defined(CONFIG_SYSFS)
 509 static int shmem_parse_huge(const char *str)
 510 {
 511         if (!strcmp(str, "never"))
 512                 return SHMEM_HUGE_NEVER;
 513         if (!strcmp(str, "always"))
 514                 return SHMEM_HUGE_ALWAYS;
 515         if (!strcmp(str, "within_size"))
 516                 return SHMEM_HUGE_WITHIN_SIZE;
 517         if (!strcmp(str, "advise"))
 518                 return SHMEM_HUGE_ADVISE;
 519         if (!strcmp(str, "deny"))
 520                 return SHMEM_HUGE_DENY;
 521         if (!strcmp(str, "force"))
 522                 return SHMEM_HUGE_FORCE;
 523         return -EINVAL;
 524 }
 525 #endif
 526
 527 #if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
 528 static const char *shmem_format_huge(int huge)
 529 {
 530         switch (huge) {
 531         case SHMEM_HUGE_NEVER:
 532                 return "never";
 533         case SHMEM_HUGE_ALWAYS:
 534                 return "always";
 535         case SHMEM_HUGE_WITHIN_SIZE:
 536                 return "within_size";
 537         case SHMEM_HUGE_ADVISE:
 538                 return "advise";
 539         case SHMEM_HUGE_DENY:
 540                 return "deny";
 541         case SHMEM_HUGE_FORCE:
 542                 return "force";
 543         default:
 544                 VM_BUG_ON(1);
 545                 return "bad_val";
 546         }
 547 }
 548 #endif
 549
 550 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 551                 struct shrink_control *sc, unsigned long nr_to_split)
 552 {
 553         LIST_HEAD(list), *pos, *next;
 554         LIST_HEAD(to_remove);
 555         struct inode *inode;
 556         struct shmem_inode_info *info;
 557         struct folio *folio;
 558         unsigned long batch = sc ? sc->nr_to_scan : 128;
 559         int split = 0;
 560
 561         if (list_empty(&sbinfo->shrinklist))
 562                 return SHRINK_STOP;
 563
 564         spin_lock(&sbinfo->shrinklist_lock);
 565         list_for_each_safe(pos, next, &sbinfo->shrinklist) {
 566                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 567
 568                 /* pin the inode */
 569                 inode = igrab(&info->vfs_inode);
 570
 571                 /* inode is about to be evicted */
 572                 if (!inode) {
 573                         list_del_init(&info->shrinklist);
 574                         goto next;
 575                 }
 576
 577                 /* Check if there's anything to gain */
 578                 if (round_up(inode->i_size, PAGE_SIZE) ==
 579                                 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
 580                         list_move(&info->shrinklist, &to_remove);
 581                         goto next;
 582                 }
 583
 584                 list_move(&info->shrinklist, &list);
 585 next:
 586                 sbinfo->shrinklist_len--;
 587                 if (!--batch)
 588                         break;
 589         }
 590         spin_unlock(&sbinfo->shrinklist_lock);
 591
 592         list_for_each_safe(pos, next, &to_remove) {
 593                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 594                 inode = &info->vfs_inode;
 595                 list_del_init(&info->shrinklist);
 596                 iput(inode);
 597         }
 598
 599         list_for_each_safe(pos, next, &list) {
 600                 int ret;
 601                 pgoff_t index;
 602
 603                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 604                 inode = &info->vfs_inode;
 605
 606                 if (nr_to_split && split >= nr_to_split)
 607                         goto move_back;
 608
 609                 index = (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT;
 610                 folio = filemap_get_folio(inode->i_mapping, index);
 611                 if (!folio)
 612                         goto drop;
 613
 614                 /* No huge page at the end of the file: nothing to split */
 615                 if (!folio_test_large(folio)) {
 616                         folio_put(folio);
 617                         goto drop;
 618                 }
 619
 620                 /*
 621                  * Move the inode on the list back to shrinklist if we failed
 622                  * to lock the page at this time.
 623                  *
 624                  * Waiting for the lock may lead to deadlock in the
 625                  * reclaim path.
 626                  */
 627                 if (!folio_trylock(folio)) {
 628                         folio_put(folio);
 629                         goto move_back;
 630                 }
 631
 632                 ret = split_huge_page(&folio->page);
 633                 folio_unlock(folio);
 634                 folio_put(folio);
 635
 636                 /* If split failed move the inode on the list back to shrinklist */
 637                 if (ret)
 638                         goto move_back;
 639
 640                 split++;
 641 drop:
 642                 list_del_init(&info->shrinklist);
 643                 goto put;
 644 move_back:
 645                 /*
 646                  * Make sure the inode is either on the global list or deleted
 647                  * from any local list before iput() since it could be deleted
 648                  * in another thread once we put the inode (then the local list
 649                  * is corrupted).
 650                  */
 651                 spin_lock(&sbinfo->shrinklist_lock);
 652                 list_move(&info->shrinklist, &sbinfo->shrinklist);
 653                 sbinfo->shrinklist_len++;
 654                 spin_unlock(&sbinfo->shrinklist_lock);
 655 put:
 656                 iput(inode);
 657         }
 658
 659         return split;
 660 }
 661
 662 static long shmem_unused_huge_scan(struct super_block *sb,
 663                 struct shrink_control *sc)
 664 {
 665         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 666
 667         if (!READ_ONCE(sbinfo->shrinklist_len))
 668                 return SHRINK_STOP;
 669
 670         return shmem_unused_huge_shrink(sbinfo, sc, 0);
 671 }
 672
 673 static long shmem_unused_huge_count(struct super_block *sb,
 674                 struct shrink_control *sc)
 675 {
 676         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 677         return READ_ONCE(sbinfo->shrinklist_len);
 678 }
 679 #else /* !CONFIG_TRANSPARENT_HUGEPAGE */
 680
 681 #define shmem_huge SHMEM_HUGE_DENY
 682
 683 bool shmem_is_huge(struct vm_area_struct *vma,
 684                    struct inode *inode, pgoff_t index)
 685 {
 686         return false;
 687 }
 688
 689 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 690                 struct shrink_control *sc, unsigned long nr_to_split)
 691 {
 692         return 0;
 693 }
 694 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 695
 696 /*
 697  * Like filemap_add_folio, but error if expected item has gone.
 698  */
 699 static int shmem_add_to_page_cache(struct folio *folio,
 700                                    struct address_space *mapping,
 701                                    pgoff_t index, void *expected, gfp_t gfp,
 702                                    struct mm_struct *charge_mm)
 703 {
 704         XA_STATE_ORDER(xas, &mapping->i_pages, index, folio_order(folio));
 705         long nr = folio_nr_pages(folio);
 706         int error;
 707
 708         VM_BUG_ON_FOLIO(index != round_down(index, nr), folio);
 709         VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 710         VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio);
 711         VM_BUG_ON(expected && folio_test_large(folio));
 712
 713         folio_ref_add(folio, nr);
 714         folio->mapping = mapping;
 715         folio->index = index;
 716
 717         if (!folio_test_swapcache(folio)) {
 718                 error = mem_cgroup_charge(folio, charge_mm, gfp);
 719                 if (error) {
 720                         if (folio_test_pmd_mappable(folio)) {
 721                                 count_vm_event(THP_FILE_FALLBACK);
 722                                 count_vm_event(THP_FILE_FALLBACK_CHARGE);
 723                         }
 724                         goto error;
 725                 }
 726         }
 727         folio_throttle_swaprate(folio, gfp);
 728
 729         do {
 730                 xas_lock_irq(&xas);
 731                 if (expected != xas_find_conflict(&xas)) {
 732                         xas_set_err(&xas, -EEXIST);
 733                         goto unlock;
 734                 }
 735                 if (expected && xas_find_conflict(&xas)) {
 736                         xas_set_err(&xas, -EEXIST);
 737                         goto unlock;
 738                 }
 739                 xas_store(&xas, folio);
 740                 if (xas_error(&xas))
 741                         goto unlock;
 742                 if (folio_test_pmd_mappable(folio)) {
 743                         count_vm_event(THP_FILE_ALLOC);
 744                         __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr);
 745                 }
 746                 mapping->nrpages += nr;
 747                 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
 748                 __lruvec_stat_mod_folio(folio, NR_SHMEM, nr);
 749 unlock:
 750                 xas_unlock_irq(&xas);
 751         } while (xas_nomem(&xas, gfp));
 752
 753         if (xas_error(&xas)) {
 754                 error = xas_error(&xas);
 755                 goto error;
 756         }
 757
 758         return 0;
 759 error:
 760         folio->mapping = NULL;
 761         folio_ref_sub(folio, nr);
 762         return error;
 763 }
 764
 765 /*
 766  * Like delete_from_page_cache, but substitutes swap for page.
 767  */
 768 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 769 {
 770         struct address_space *mapping = page->mapping;
 771         int error;
 772
 773         VM_BUG_ON_PAGE(PageCompound(page), page);
 774
 775         xa_lock_irq(&mapping->i_pages);
 776         error = shmem_replace_entry(mapping, page->index, page, radswap);
 777         page->mapping = NULL;
 778         mapping->nrpages--;
 779         __dec_lruvec_page_state(page, NR_FILE_PAGES);
 780         __dec_lruvec_page_state(page, NR_SHMEM);
 781         xa_unlock_irq(&mapping->i_pages);
 782         put_page(page);
 783         BUG_ON(error);
 784 }
 785
 786 /*
 787  * Remove swap entry from page cache, free the swap and its page cache.
 788  */
 789 static int shmem_free_swap(struct address_space *mapping,
 790                            pgoff_t index, void *radswap)
 791 {
 792         void *old;
 793
 794         old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
 795         if (old != radswap)
 796                 return -ENOENT;
 797         free_swap_and_cache(radix_to_swp_entry(radswap));
 798         return 0;
 799 }
 800
 801 /*
 802  * Determine (in bytes) how many of the shmem object's pages mapped by the
 803  * given offsets are swapped out.
 804  *
 805  * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
 806  * as long as the inode doesn't go away and racy results are not a problem.
 807  */
 808 unsigned long shmem_partial_swap_usage(struct address_space *mapping,
 809                                                 pgoff_t start, pgoff_t end)
 810 {
 811         XA_STATE(xas, &mapping->i_pages, start);
 812         struct page *page;
 813         unsigned long swapped = 0;
 814
 815         rcu_read_lock();
 816         xas_for_each(&xas, page, end - 1) {
 817                 if (xas_retry(&xas, page))
 818                         continue;
 819                 if (xa_is_value(page))
 820                         swapped++;
 821
 822                 if (need_resched()) {
 823                         xas_pause(&xas);
 824                         cond_resched_rcu();
 825                 }
 826         }
 827
 828         rcu_read_unlock();
 829
 830         return swapped << PAGE_SHIFT;
 831 }
 832
 833 /*
 834  * Determine (in bytes) how many of the shmem object's pages mapped by the
 835  * given vma is swapped out.
 836  *
 837  * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
 838  * as long as the inode doesn't go away and racy results are not a problem.
 839  */
 840 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
 841 {
 842         struct inode *inode = file_inode(vma->vm_file);
 843         struct shmem_inode_info *info = SHMEM_I(inode);
 844         struct address_space *mapping = inode->i_mapping;
 845         unsigned long swapped;
 846
 847         /* Be careful as we don't hold info->lock */
 848         swapped = READ_ONCE(info->swapped);
 849
 850         /*
 851          * The easier cases are when the shmem object has nothing in swap, or
 852          * the vma maps it whole. Then we can simply use the stats that we
 853          * already track.
 854          */
 855         if (!swapped)
 856                 return 0;
 857
 858         if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
 859                 return swapped << PAGE_SHIFT;
 860
 861         /* Here comes the more involved part */
 862         return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
 863                                         vma->vm_pgoff + vma_pages(vma));
 864 }
 865
 866 /*
 867  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 868  */
 869 void shmem_unlock_mapping(struct address_space *mapping)
 870 {
 871         struct folio_batch fbatch;
 872         pgoff_t index = 0;
 873
 874         folio_batch_init(&fbatch);
 875         /*
 876          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 877          */
 878         while (!mapping_unevictable(mapping) &&
 879                filemap_get_folios(mapping, &index, ~0UL, &fbatch)) {
 880                 check_move_unevictable_folios(&fbatch);
 881                 folio_batch_release(&fbatch);
 882                 cond_resched();
 883         }
 884 }
 885
 886 static struct folio *shmem_get_partial_folio(struct inode *inode, pgoff_t index)
 887 {
 888         struct folio *folio;
 889         struct page *page;
 890
 891         /*
 892          * At first avoid shmem_getpage(,,,SGP_READ): that fails
 893          * beyond i_size, and reports fallocated pages as holes.
 894          */
 895         folio = __filemap_get_folio(inode->i_mapping, index,
 896                                         FGP_ENTRY | FGP_LOCK, 0);
 897         if (!xa_is_value(folio))
 898                 return folio;
 899         /*
 900          * But read a page back from swap if any of it is within i_size
 901          * (although in some cases this is just a waste of time).
 902          */
 903         page = NULL;
 904         shmem_getpage(inode, index, &page, SGP_READ);
 905         return page ? page_folio(page) : NULL;
 906 }
 907
 908 /*
 909  * Remove range of pages and swap entries from page cache, and free them.
 910  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 911  */
 912 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 913                                                                  bool unfalloc)
 914 {
 915         struct address_space *mapping = inode->i_mapping;
 916         struct shmem_inode_info *info = SHMEM_I(inode);
 917         pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
 918         pgoff_t end = (lend + 1) >> PAGE_SHIFT;
 919         struct folio_batch fbatch;
 920         pgoff_t indices[PAGEVEC_SIZE];
 921         struct folio *folio;
 922         bool same_folio;
 923         long nr_swaps_freed = 0;
 924         pgoff_t index;
 925         int i;
 926
 927         if (lend == -1)
 928                 end = -1;       /* unsigned, so actually very big */
 929
 930         if (info->fallocend > start && info->fallocend <= end && !unfalloc)
 931                 info->fallocend = start;
 932
 933         folio_batch_init(&fbatch);
 934         index = start;
 935         while (index < end && find_lock_entries(mapping, index, end - 1,
 936                         &fbatch, indices)) {
 937                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 938                         folio = fbatch.folios[i];
 939
 940                         index = indices[i];
 941
 942                         if (xa_is_value(folio)) {
 943                                 if (unfalloc)
 944                                         continue;
 945                                 nr_swaps_freed += !shmem_free_swap(mapping,
 946                                                                 index, folio);
 947                                 continue;
 948                         }
 949                         index += folio_nr_pages(folio) - 1;
 950
 951                         if (!unfalloc || !folio_test_uptodate(folio))
 952                                 truncate_inode_folio(mapping, folio);
 953                         folio_unlock(folio);
 954                 }
 955                 folio_batch_remove_exceptionals(&fbatch);
 956                 folio_batch_release(&fbatch);
 957                 cond_resched();
 958                 index++;
 959         }
 960
 961         same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
 962         folio = shmem_get_partial_folio(inode, lstart >> PAGE_SHIFT);
 963         if (folio) {
 964                 same_folio = lend < folio_pos(folio) + folio_size(folio);
 965                 folio_mark_dirty(folio);
 966                 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
 967                         start = folio->index + folio_nr_pages(folio);
 968                         if (same_folio)
 969                                 end = folio->index;
 970                 }
 971                 folio_unlock(folio);
 972                 folio_put(folio);
 973                 folio = NULL;
 974         }
 975
 976         if (!same_folio)
 977                 folio = shmem_get_partial_folio(inode, lend >> PAGE_SHIFT);
 978         if (folio) {
 979                 folio_mark_dirty(folio);
 980                 if (!truncate_inode_partial_folio(folio, lstart, lend))
 981                         end = folio->index;
 982                 folio_unlock(folio);
 983                 folio_put(folio);
 984         }
 985
 986         index = start;
 987         while (index < end) {
 988                 cond_resched();
 989
 990                 if (!find_get_entries(mapping, index, end - 1, &fbatch,
 991                                 indices)) {
 992                         /* If all gone or hole-punch or unfalloc, we're done */
 993                         if (index == start || end != -1)
 994                                 break;
 995                         /* But if truncating, restart to make sure all gone */
 996                         index = start;
 997                         continue;
 998                 }
 999                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
1000                         folio = fbatch.folios[i];
1001
1002                         index = indices[i];
1003                         if (xa_is_value(folio)) {
1004                                 if (unfalloc)
1005                                         continue;
1006                                 if (shmem_free_swap(mapping, index, folio)) {
1007                                         /* Swap was replaced by page: retry */
1008                                         index--;
1009                                         break;
1010                                 }
1011                                 nr_swaps_freed++;
1012                                 continue;
1013                         }
1014
1015                         folio_lock(folio);
1016
1017                         if (!unfalloc || !folio_test_uptodate(folio)) {
1018                                 if (folio_mapping(folio) != mapping) {
1019                                         /* Page was replaced by swap: retry */
1020                                         folio_unlock(folio);
1021                                         index--;
1022                                         break;
1023                                 }
1024                                 VM_BUG_ON_FOLIO(folio_test_writeback(folio),
1025                                                 folio);
1026                                 truncate_inode_folio(mapping, folio);
1027                         }
1028                         index = folio->index + folio_nr_pages(folio) - 1;
1029                         folio_unlock(folio);
1030                 }
1031                 folio_batch_remove_exceptionals(&fbatch);
1032                 folio_batch_release(&fbatch);
1033                 index++;
1034         }
1035
1036         spin_lock_irq(&info->lock);
1037         info->swapped -= nr_swaps_freed;
1038         shmem_recalc_inode(inode);
1039         spin_unlock_irq(&info->lock);
1040 }
1041
1042 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1043 {
1044         shmem_undo_range(inode, lstart, lend, false);
1045         inode->i_ctime = inode->i_mtime = current_time(inode);
1046 }
1047 EXPORT_SYMBOL_GPL(shmem_truncate_range);
1048
1049 static int shmem_getattr(struct user_namespace *mnt_userns,
1050                          const struct path *path, struct kstat *stat,
1051                          u32 request_mask, unsigned int query_flags)
1052 {
1053         struct inode *inode = path->dentry->d_inode;
1054         struct shmem_inode_info *info = SHMEM_I(inode);
1055
1056         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1057                 spin_lock_irq(&info->lock);
1058                 shmem_recalc_inode(inode);
1059                 spin_unlock_irq(&info->lock);
1060         }
1061         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))
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
1091         error = setattr_prepare(&init_user_ns, dentry, attr);
1092         if (error)
1093                 return error;
1094
1095         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1096                 loff_t oldsize = inode->i_size;
1097                 loff_t newsize = attr->ia_size;
1098
1099                 /* protected by i_rwsem */
1100                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1101                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1102                         return -EPERM;
1103
1104                 if (newsize != oldsize) {
1105                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1106                                         oldsize, newsize);
1107                         if (error)
1108                                 return error;
1109                         i_size_write(inode, newsize);
1110                         inode->i_ctime = inode->i_mtime = current_time(inode);
1111                 }
1112                 if (newsize <= oldsize) {
1113                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1114                         if (oldsize > holebegin)
1115                                 unmap_mapping_range(inode->i_mapping,
1116                                                         holebegin, 0, 1);
1117                         if (info->alloced)
1118                                 shmem_truncate_range(inode,
1119                                                         newsize, (loff_t)-1);
1120                         /* unmap again to remove racily COWed private pages */
1121                         if (oldsize > holebegin)
1122                                 unmap_mapping_range(inode->i_mapping,
1123                                                         holebegin, 0, 1);
1124                 }
1125         }
1126
1127         setattr_copy(&init_user_ns, inode, attr);
1128         if (attr->ia_valid & ATTR_MODE)
1129                 error = posix_acl_chmod(&init_user_ns, inode, inode->i_mode);
1130         return error;
1131 }
1132
1133 static void shmem_evict_inode(struct inode *inode)
1134 {
1135         struct shmem_inode_info *info = SHMEM_I(inode);
1136         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1137
1138         if (shmem_mapping(inode->i_mapping)) {
1139                 shmem_unacct_size(info->flags, inode->i_size);
1140                 inode->i_size = 0;
1141                 mapping_set_exiting(inode->i_mapping);
1142                 shmem_truncate_range(inode, 0, (loff_t)-1);
1143                 if (!list_empty(&info->shrinklist)) {
1144                         spin_lock(&sbinfo->shrinklist_lock);
1145                         if (!list_empty(&info->shrinklist)) {
1146                                 list_del_init(&info->shrinklist);
1147                                 sbinfo->shrinklist_len--;
1148                         }
1149                         spin_unlock(&sbinfo->shrinklist_lock);
1150                 }
1151                 while (!list_empty(&info->swaplist)) {
1152                         /* Wait while shmem_unuse() is scanning this inode... */
1153                         wait_var_event(&info->stop_eviction,
1154                                        !atomic_read(&info->stop_eviction));
1155                         mutex_lock(&shmem_swaplist_mutex);
1156                         /* ...but beware of the race if we peeked too early */
1157                         if (!atomic_read(&info->stop_eviction))
1158                                 list_del_init(&info->swaplist);
1159                         mutex_unlock(&shmem_swaplist_mutex);
1160                 }
1161         }
1162
1163         simple_xattrs_free(&info->xattrs);
1164         WARN_ON(inode->i_blocks);
1165         shmem_free_inode(inode->i_sb);
1166         clear_inode(inode);
1167 }
1168
1169 static int shmem_find_swap_entries(struct address_space *mapping,
1170                                    pgoff_t start, struct folio_batch *fbatch,
1171                                    pgoff_t *indices, unsigned int type)
1172 {
1173         XA_STATE(xas, &mapping->i_pages, start);
1174         struct folio *folio;
1175         swp_entry_t entry;
1176
1177         rcu_read_lock();
1178         xas_for_each(&xas, folio, ULONG_MAX) {
1179                 if (xas_retry(&xas, folio))
1180                         continue;
1181
1182                 if (!xa_is_value(folio))
1183                         continue;
1184
1185                 entry = radix_to_swp_entry(folio);
1186                 /*
1187                  * swapin error entries can be found in the mapping. But they're
1188                  * deliberately ignored here as we've done everything we can do.
1189                  */
1190                 if (swp_type(entry) != type)
1191                         continue;
1192
1193                 indices[folio_batch_count(fbatch)] = xas.xa_index;
1194                 if (!folio_batch_add(fbatch, folio))
1195                         break;
1196
1197                 if (need_resched()) {
1198                         xas_pause(&xas);
1199                         cond_resched_rcu();
1200                 }
1201         }
1202         rcu_read_unlock();
1203
1204         return xas.xa_index;
1205 }
1206
1207 /*
1208  * Move the swapped pages for an inode to page cache. Returns the count
1209  * of pages swapped in, or the error in case of failure.
1210  */
1211 static int shmem_unuse_swap_entries(struct inode *inode,
1212                 struct folio_batch *fbatch, pgoff_t *indices)
1213 {
1214         int i = 0;
1215         int ret = 0;
1216         int error = 0;
1217         struct address_space *mapping = inode->i_mapping;
1218
1219         for (i = 0; i < folio_batch_count(fbatch); i++) {
1220                 struct folio *folio = fbatch->folios[i];
1221
1222                 if (!xa_is_value(folio))
1223                         continue;
1224                 error = shmem_swapin_folio(inode, indices[i],
1225                                           &folio, SGP_CACHE,
1226                                           mapping_gfp_mask(mapping),
1227                                           NULL, NULL);
1228                 if (error == 0) {
1229                         folio_unlock(folio);
1230                         folio_put(folio);
1231                         ret++;
1232                 }
1233                 if (error == -ENOMEM)
1234                         break;
1235                 error = 0;
1236         }
1237         return error ? error : ret;
1238 }
1239
1240 /*
1241  * If swap found in inode, free it and move page from swapcache to filecache.
1242  */
1243 static int shmem_unuse_inode(struct inode *inode, unsigned int type)
1244 {
1245         struct address_space *mapping = inode->i_mapping;
1246         pgoff_t start = 0;
1247         struct folio_batch fbatch;
1248         pgoff_t indices[PAGEVEC_SIZE];
1249         int ret = 0;
1250
1251         do {
1252                 folio_batch_init(&fbatch);
1253                 shmem_find_swap_entries(mapping, start, &fbatch, indices, type);
1254                 if (folio_batch_count(&fbatch) == 0) {
1255                         ret = 0;
1256                         break;
1257                 }
1258
1259                 ret = shmem_unuse_swap_entries(inode, &fbatch, indices);
1260                 if (ret < 0)
1261                         break;
1262
1263                 start = indices[folio_batch_count(&fbatch) - 1];
1264         } while (true);
1265
1266         return ret;
1267 }
1268
1269 /*
1270  * Read all the shared memory data that resides in the swap
1271  * device 'type' back into memory, so the swap device can be
1272  * unused.
1273  */
1274 int shmem_unuse(unsigned int type)
1275 {
1276         struct shmem_inode_info *info, *next;
1277         int error = 0;
1278
1279         if (list_empty(&shmem_swaplist))
1280                 return 0;
1281
1282         mutex_lock(&shmem_swaplist_mutex);
1283         list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1284                 if (!info->swapped) {
1285                         list_del_init(&info->swaplist);
1286                         continue;
1287                 }
1288                 /*
1289                  * Drop the swaplist mutex while searching the inode for swap;
1290                  * but before doing so, make sure shmem_evict_inode() will not
1291                  * remove placeholder inode from swaplist, nor let it be freed
1292                  * (igrab() would protect from unlink, but not from unmount).
1293                  */
1294                 atomic_inc(&info->stop_eviction);
1295                 mutex_unlock(&shmem_swaplist_mutex);
1296
1297                 error = shmem_unuse_inode(&info->vfs_inode, type);
1298                 cond_resched();
1299
1300                 mutex_lock(&shmem_swaplist_mutex);
1301                 next = list_next_entry(info, swaplist);
1302                 if (!info->swapped)
1303                         list_del_init(&info->swaplist);
1304                 if (atomic_dec_and_test(&info->stop_eviction))
1305                         wake_up_var(&info->stop_eviction);
1306                 if (error)
1307                         break;
1308         }
1309         mutex_unlock(&shmem_swaplist_mutex);
1310
1311         return error;
1312 }
1313
1314 /*
1315  * Move the page from the page cache to the swap cache.
1316  */
1317 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1318 {
1319         struct folio *folio = page_folio(page);
1320         struct shmem_inode_info *info;
1321         struct address_space *mapping;
1322         struct inode *inode;
1323         swp_entry_t swap;
1324         pgoff_t index;
1325
1326         /*
1327          * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "always" or
1328          * "force", drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
1329          * and its shmem_writeback() needs them to be split when swapping.
1330          */
1331         if (PageTransCompound(page)) {
1332                 /* Ensure the subpages are still dirty */
1333                 SetPageDirty(page);
1334                 if (split_huge_page(page) < 0)
1335                         goto redirty;
1336                 ClearPageDirty(page);
1337         }
1338
1339         BUG_ON(!PageLocked(page));
1340         mapping = page->mapping;
1341         index = page->index;
1342         inode = mapping->host;
1343         info = SHMEM_I(inode);
1344         if (info->flags & VM_LOCKED)
1345                 goto redirty;
1346         if (!total_swap_pages)
1347                 goto redirty;
1348
1349         /*
1350          * Our capabilities prevent regular writeback or sync from ever calling
1351          * shmem_writepage; but a stacking filesystem might use ->writepage of
1352          * its underlying filesystem, in which case tmpfs should write out to
1353          * swap only in response to memory pressure, and not for the writeback
1354          * threads or sync.
1355          */
1356         if (!wbc->for_reclaim) {
1357                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1358                 goto redirty;
1359         }
1360
1361         /*
1362          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1363          * value into swapfile.c, the only way we can correctly account for a
1364          * fallocated page arriving here is now to initialize it and write it.
1365          *
1366          * That's okay for a page already fallocated earlier, but if we have
1367          * not yet completed the fallocation, then (a) we want to keep track
1368          * of this page in case we have to undo it, and (b) it may not be a
1369          * good idea to continue anyway, once we're pushing into swap.  So
1370          * reactivate the page, and let shmem_fallocate() quit when too many.
1371          */
1372         if (!PageUptodate(page)) {
1373                 if (inode->i_private) {
1374                         struct shmem_falloc *shmem_falloc;
1375                         spin_lock(&inode->i_lock);
1376                         shmem_falloc = inode->i_private;
1377                         if (shmem_falloc &&
1378                             !shmem_falloc->waitq &&
1379                             index >= shmem_falloc->start &&
1380                             index < shmem_falloc->next)
1381                                 shmem_falloc->nr_unswapped++;
1382                         else
1383                                 shmem_falloc = NULL;
1384                         spin_unlock(&inode->i_lock);
1385                         if (shmem_falloc)
1386                                 goto redirty;
1387                 }
1388                 clear_highpage(page);
1389                 flush_dcache_page(page);
1390                 SetPageUptodate(page);
1391         }
1392
1393         swap = folio_alloc_swap(folio);
1394         if (!swap.val)
1395                 goto redirty;
1396
1397         /*
1398          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1399          * if it's not already there.  Do it now before the page is
1400          * moved to swap cache, when its pagelock no longer protects
1401          * the inode from eviction.  But don't unlock the mutex until
1402          * we've incremented swapped, because shmem_unuse_inode() will
1403          * prune a !swapped inode from the swaplist under this mutex.
1404          */
1405         mutex_lock(&shmem_swaplist_mutex);
1406         if (list_empty(&info->swaplist))
1407                 list_add(&info->swaplist, &shmem_swaplist);
1408
1409         if (add_to_swap_cache(page, swap,
1410                         __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1411                         NULL) == 0) {
1412                 spin_lock_irq(&info->lock);
1413                 shmem_recalc_inode(inode);
1414                 info->swapped++;
1415                 spin_unlock_irq(&info->lock);
1416
1417                 swap_shmem_alloc(swap);
1418                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1419
1420                 mutex_unlock(&shmem_swaplist_mutex);
1421                 BUG_ON(page_mapped(page));
1422                 swap_writepage(page, wbc);
1423                 return 0;
1424         }
1425
1426         mutex_unlock(&shmem_swaplist_mutex);
1427         put_swap_page(page, swap);
1428 redirty:
1429         set_page_dirty(page);
1430         if (wbc->for_reclaim)
1431                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1432         unlock_page(page);
1433         return 0;
1434 }
1435
1436 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1437 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1438 {
1439         char buffer[64];
1440
1441         if (!mpol || mpol->mode == MPOL_DEFAULT)
1442                 return;         /* show nothing */
1443
1444         mpol_to_str(buffer, sizeof(buffer), mpol);
1445
1446         seq_printf(seq, ",mpol=%s", buffer);
1447 }
1448
1449 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1450 {
1451         struct mempolicy *mpol = NULL;
1452         if (sbinfo->mpol) {
1453                 raw_spin_lock(&sbinfo->stat_lock);      /* prevent replace/use races */
1454                 mpol = sbinfo->mpol;
1455                 mpol_get(mpol);
1456                 raw_spin_unlock(&sbinfo->stat_lock);
1457         }
1458         return mpol;
1459 }
1460 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1461 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1462 {
1463 }
1464 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1465 {
1466         return NULL;
1467 }
1468 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1469 #ifndef CONFIG_NUMA
1470 #define vm_policy vm_private_data
1471 #endif
1472
1473 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1474                 struct shmem_inode_info *info, pgoff_t index)
1475 {
1476         /* Create a pseudo vma that just contains the policy */
1477         vma_init(vma, NULL);
1478         /* Bias interleave by inode number to distribute better across nodes */
1479         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1480         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1481 }
1482
1483 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1484 {
1485         /* Drop reference taken by mpol_shared_policy_lookup() */
1486         mpol_cond_put(vma->vm_policy);
1487 }
1488
1489 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1490                         struct shmem_inode_info *info, pgoff_t index)
1491 {
1492         struct vm_area_struct pvma;
1493         struct page *page;
1494         struct vm_fault vmf = {
1495                 .vma = &pvma,
1496         };
1497
1498         shmem_pseudo_vma_init(&pvma, info, index);
1499         page = swap_cluster_readahead(swap, gfp, &vmf);
1500         shmem_pseudo_vma_destroy(&pvma);
1501
1502         return page;
1503 }
1504
1505 /*
1506  * Make sure huge_gfp is always more limited than limit_gfp.
1507  * Some of the flags set permissions, while others set limitations.
1508  */
1509 static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t limit_gfp)
1510 {
1511         gfp_t allowflags = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
1512         gfp_t denyflags = __GFP_NOWARN | __GFP_NORETRY;
1513         gfp_t zoneflags = limit_gfp & GFP_ZONEMASK;
1514         gfp_t result = huge_gfp & ~(allowflags | GFP_ZONEMASK);
1515
1516         /* Allow allocations only from the originally specified zones. */
1517         result |= zoneflags;
1518
1519         /*
1520          * Minimize the result gfp by taking the union with the deny flags,
1521          * and the intersection of the allow flags.
1522          */
1523         result |= (limit_gfp & denyflags);
1524         result |= (huge_gfp & limit_gfp) & allowflags;
1525
1526         return result;
1527 }
1528
1529 static struct folio *shmem_alloc_hugefolio(gfp_t gfp,
1530                 struct shmem_inode_info *info, pgoff_t index)
1531 {
1532         struct vm_area_struct pvma;
1533         struct address_space *mapping = info->vfs_inode.i_mapping;
1534         pgoff_t hindex;
1535         struct folio *folio;
1536
1537         hindex = round_down(index, HPAGE_PMD_NR);
1538         if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1539                                                                 XA_PRESENT))
1540                 return NULL;
1541
1542         shmem_pseudo_vma_init(&pvma, info, hindex);
1543         folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, &pvma, 0, true);
1544         shmem_pseudo_vma_destroy(&pvma);
1545         if (!folio)
1546                 count_vm_event(THP_FILE_FALLBACK);
1547         return folio;
1548 }
1549
1550 static struct folio *shmem_alloc_folio(gfp_t gfp,
1551                         struct shmem_inode_info *info, pgoff_t index)
1552 {
1553         struct vm_area_struct pvma;
1554         struct folio *folio;
1555
1556         shmem_pseudo_vma_init(&pvma, info, index);
1557         folio = vma_alloc_folio(gfp, 0, &pvma, 0, false);
1558         shmem_pseudo_vma_destroy(&pvma);
1559
1560         return folio;
1561 }
1562
1563 static struct page *shmem_alloc_page(gfp_t gfp,
1564                         struct shmem_inode_info *info, pgoff_t index)
1565 {
1566         return &shmem_alloc_folio(gfp, info, index)->page;
1567 }
1568
1569 static struct folio *shmem_alloc_and_acct_folio(gfp_t gfp, struct inode *inode,
1570                 pgoff_t index, bool huge)
1571 {
1572         struct shmem_inode_info *info = SHMEM_I(inode);
1573         struct folio *folio;
1574         int nr;
1575         int err = -ENOSPC;
1576
1577         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1578                 huge = false;
1579         nr = huge ? HPAGE_PMD_NR : 1;
1580
1581         if (!shmem_inode_acct_block(inode, nr))
1582                 goto failed;
1583
1584         if (huge)
1585                 folio = shmem_alloc_hugefolio(gfp, info, index);
1586         else
1587                 folio = shmem_alloc_folio(gfp, info, index);
1588         if (folio) {
1589                 __folio_set_locked(folio);
1590                 __folio_set_swapbacked(folio);
1591                 return folio;
1592         }
1593
1594         err = -ENOMEM;
1595         shmem_inode_unacct_blocks(inode, nr);
1596 failed:
1597         return ERR_PTR(err);
1598 }
1599
1600 /*
1601  * When a page is moved from swapcache to shmem filecache (either by the
1602  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1603  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1604  * ignorance of the mapping it belongs to.  If that mapping has special
1605  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1606  * we may need to copy to a suitable page before moving to filecache.
1607  *
1608  * In a future release, this may well be extended to respect cpuset and
1609  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1610  * but for now it is a simple matter of zone.
1611  */
1612 static bool shmem_should_replace_folio(struct folio *folio, gfp_t gfp)
1613 {
1614         return folio_zonenum(folio) > gfp_zone(gfp);
1615 }
1616
1617 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1618                                 struct shmem_inode_info *info, pgoff_t index)
1619 {
1620         struct page *oldpage, *newpage;
1621         struct folio *old, *new;
1622         struct address_space *swap_mapping;
1623         swp_entry_t entry;
1624         pgoff_t swap_index;
1625         int error;
1626
1627         oldpage = *pagep;
1628         entry.val = page_private(oldpage);
1629         swap_index = swp_offset(entry);
1630         swap_mapping = page_mapping(oldpage);
1631
1632         /*
1633          * We have arrived here because our zones are constrained, so don't
1634          * limit chance of success by further cpuset and node constraints.
1635          */
1636         gfp &= ~GFP_CONSTRAINT_MASK;
1637         newpage = shmem_alloc_page(gfp, info, index);
1638         if (!newpage)
1639                 return -ENOMEM;
1640
1641         get_page(newpage);
1642         copy_highpage(newpage, oldpage);
1643         flush_dcache_page(newpage);
1644
1645         __SetPageLocked(newpage);
1646         __SetPageSwapBacked(newpage);
1647         SetPageUptodate(newpage);
1648         set_page_private(newpage, entry.val);
1649         SetPageSwapCache(newpage);
1650
1651         /*
1652          * Our caller will very soon move newpage out of swapcache, but it's
1653          * a nice clean interface for us to replace oldpage by newpage there.
1654          */
1655         xa_lock_irq(&swap_mapping->i_pages);
1656         error = shmem_replace_entry(swap_mapping, swap_index, oldpage, newpage);
1657         if (!error) {
1658                 old = page_folio(oldpage);
1659                 new = page_folio(newpage);
1660                 mem_cgroup_migrate(old, new);
1661                 __inc_lruvec_page_state(newpage, NR_FILE_PAGES);
1662                 __dec_lruvec_page_state(oldpage, NR_FILE_PAGES);
1663         }
1664         xa_unlock_irq(&swap_mapping->i_pages);
1665
1666         if (unlikely(error)) {
1667                 /*
1668                  * Is this possible?  I think not, now that our callers check
1669                  * both PageSwapCache and page_private after getting page lock;
1670                  * but be defensive.  Reverse old to newpage for clear and free.
1671                  */
1672                 oldpage = newpage;
1673         } else {
1674                 lru_cache_add(newpage);
1675                 *pagep = newpage;
1676         }
1677
1678         ClearPageSwapCache(oldpage);
1679         set_page_private(oldpage, 0);
1680
1681         unlock_page(oldpage);
1682         put_page(oldpage);
1683         put_page(oldpage);
1684         return error;
1685 }
1686
1687 static void shmem_set_folio_swapin_error(struct inode *inode, pgoff_t index,
1688                                          struct folio *folio, swp_entry_t swap)
1689 {
1690         struct address_space *mapping = inode->i_mapping;
1691         struct shmem_inode_info *info = SHMEM_I(inode);
1692         swp_entry_t swapin_error;
1693         void *old;
1694
1695         swapin_error = make_swapin_error_entry(&folio->page);
1696         old = xa_cmpxchg_irq(&mapping->i_pages, index,
1697                              swp_to_radix_entry(swap),
1698                              swp_to_radix_entry(swapin_error), 0);
1699         if (old != swp_to_radix_entry(swap))
1700                 return;
1701
1702         folio_wait_writeback(folio);
1703         delete_from_swap_cache(folio);
1704         spin_lock_irq(&info->lock);
1705         /*
1706          * Don't treat swapin error folio as alloced. Otherwise inode->i_blocks won't
1707          * be 0 when inode is released and thus trigger WARN_ON(inode->i_blocks) in
1708          * shmem_evict_inode.
1709          */
1710         info->alloced--;
1711         info->swapped--;
1712         shmem_recalc_inode(inode);
1713         spin_unlock_irq(&info->lock);
1714         swap_free(swap);
1715 }
1716
1717 /*
1718  * Swap in the folio pointed to by *foliop.
1719  * Caller has to make sure that *foliop contains a valid swapped folio.
1720  * Returns 0 and the folio in foliop if success. On failure, returns the
1721  * error code and NULL in *foliop.
1722  */
1723 static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
1724                              struct folio **foliop, enum sgp_type sgp,
1725                              gfp_t gfp, struct vm_area_struct *vma,
1726                              vm_fault_t *fault_type)
1727 {
1728         struct address_space *mapping = inode->i_mapping;
1729         struct shmem_inode_info *info = SHMEM_I(inode);
1730         struct mm_struct *charge_mm = vma ? vma->vm_mm : NULL;
1731         struct page *page;
1732         struct folio *folio = NULL;
1733         swp_entry_t swap;
1734         int error;
1735
1736         VM_BUG_ON(!*foliop || !xa_is_value(*foliop));
1737         swap = radix_to_swp_entry(*foliop);
1738         *foliop = NULL;
1739
1740         if (is_swapin_error_entry(swap))
1741                 return -EIO;
1742
1743         /* Look it up and read it in.. */
1744         page = lookup_swap_cache(swap, NULL, 0);
1745         if (!page) {
1746                 /* Or update major stats only when swapin succeeds?? */
1747                 if (fault_type) {
1748                         *fault_type |= VM_FAULT_MAJOR;
1749                         count_vm_event(PGMAJFAULT);
1750                         count_memcg_event_mm(charge_mm, PGMAJFAULT);
1751                 }
1752                 /* Here we actually start the io */
1753                 page = shmem_swapin(swap, gfp, info, index);
1754                 if (!page) {
1755                         error = -ENOMEM;
1756                         goto failed;
1757                 }
1758         }
1759         folio = page_folio(page);
1760
1761         /* We have to do this with folio locked to prevent races */
1762         folio_lock(folio);
1763         if (!folio_test_swapcache(folio) ||
1764             folio_swap_entry(folio).val != swap.val ||
1765             !shmem_confirm_swap(mapping, index, swap)) {
1766                 error = -EEXIST;
1767                 goto unlock;
1768         }
1769         if (!folio_test_uptodate(folio)) {
1770                 error = -EIO;
1771                 goto failed;
1772         }
1773         folio_wait_writeback(folio);
1774
1775         /*
1776          * Some architectures may have to restore extra metadata to the
1777          * folio after reading from swap.
1778          */
1779         arch_swap_restore(swap, folio);
1780
1781         if (shmem_should_replace_folio(folio, gfp)) {
1782                 error = shmem_replace_page(&page, gfp, info, index);
1783                 if (error)
1784                         goto failed;
1785         }
1786
1787         error = shmem_add_to_page_cache(folio, mapping, index,
1788                                         swp_to_radix_entry(swap), gfp,
1789                                         charge_mm);
1790         if (error)
1791                 goto failed;
1792
1793         spin_lock_irq(&info->lock);
1794         info->swapped--;
1795         shmem_recalc_inode(inode);
1796         spin_unlock_irq(&info->lock);
1797
1798         if (sgp == SGP_WRITE)
1799                 folio_mark_accessed(folio);
1800
1801         delete_from_swap_cache(folio);
1802         folio_mark_dirty(folio);
1803         swap_free(swap);
1804
1805         *foliop = folio;
1806         return 0;
1807 failed:
1808         if (!shmem_confirm_swap(mapping, index, swap))
1809                 error = -EEXIST;
1810         if (error == -EIO)
1811                 shmem_set_folio_swapin_error(inode, index, folio, swap);
1812 unlock:
1813         if (folio) {
1814                 folio_unlock(folio);
1815                 folio_put(folio);
1816         }
1817
1818         return error;
1819 }
1820
1821 /*
1822  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1823  *
1824  * If we allocate a new one we do not mark it dirty. That's up to the
1825  * vm. If we swap it in we mark it dirty since we also free the swap
1826  * entry since a page cannot live in both the swap and page cache.
1827  *
1828  * vma, vmf, and fault_type are only supplied by shmem_fault:
1829  * otherwise they are NULL.
1830  */
1831 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1832         struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1833         struct vm_area_struct *vma, struct vm_fault *vmf,
1834                         vm_fault_t *fault_type)
1835 {
1836         struct address_space *mapping = inode->i_mapping;
1837         struct shmem_inode_info *info = SHMEM_I(inode);
1838         struct shmem_sb_info *sbinfo;
1839         struct mm_struct *charge_mm;
1840         struct folio *folio;
1841         pgoff_t hindex = index;
1842         gfp_t huge_gfp;
1843         int error;
1844         int once = 0;
1845         int alloced = 0;
1846
1847         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1848                 return -EFBIG;
1849 repeat:
1850         if (sgp <= SGP_CACHE &&
1851             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1852                 return -EINVAL;
1853         }
1854
1855         sbinfo = SHMEM_SB(inode->i_sb);
1856         charge_mm = vma ? vma->vm_mm : NULL;
1857
1858         folio = __filemap_get_folio(mapping, index, FGP_ENTRY | FGP_LOCK, 0);
1859         if (folio && vma && userfaultfd_minor(vma)) {
1860                 if (!xa_is_value(folio)) {
1861                         folio_unlock(folio);
1862                         folio_put(folio);
1863                 }
1864                 *fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
1865                 return 0;
1866         }
1867
1868         if (xa_is_value(folio)) {
1869                 error = shmem_swapin_folio(inode, index, &folio,
1870                                           sgp, gfp, vma, fault_type);
1871                 if (error == -EEXIST)
1872                         goto repeat;
1873
1874                 *pagep = &folio->page;
1875                 return error;
1876         }
1877
1878         if (folio) {
1879                 hindex = folio->index;
1880                 if (sgp == SGP_WRITE)
1881                         folio_mark_accessed(folio);
1882                 if (folio_test_uptodate(folio))
1883                         goto out;
1884                 /* fallocated page */
1885                 if (sgp != SGP_READ)
1886                         goto clear;
1887                 folio_unlock(folio);
1888                 folio_put(folio);
1889         }
1890
1891         /*
1892          * SGP_READ: succeed on hole, with NULL page, letting caller zero.
1893          * SGP_NOALLOC: fail on hole, with NULL page, letting caller fail.
1894          */
1895         *pagep = NULL;
1896         if (sgp == SGP_READ)
1897                 return 0;
1898         if (sgp == SGP_NOALLOC)
1899                 return -ENOENT;
1900
1901         /*
1902          * Fast cache lookup and swap lookup did not find it: allocate.
1903          */
1904
1905         if (vma && userfaultfd_missing(vma)) {
1906                 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1907                 return 0;
1908         }
1909
1910         if (!shmem_is_huge(vma, inode, index))
1911                 goto alloc_nohuge;
1912
1913         huge_gfp = vma_thp_gfp_mask(vma);
1914         huge_gfp = limit_gfp_mask(huge_gfp, gfp);
1915         folio = shmem_alloc_and_acct_folio(huge_gfp, inode, index, true);
1916         if (IS_ERR(folio)) {
1917 alloc_nohuge:
1918                 folio = shmem_alloc_and_acct_folio(gfp, inode, index, false);
1919         }
1920         if (IS_ERR(folio)) {
1921                 int retry = 5;
1922
1923                 error = PTR_ERR(folio);
1924                 folio = NULL;
1925                 if (error != -ENOSPC)
1926                         goto unlock;
1927                 /*
1928                  * Try to reclaim some space by splitting a huge page
1929                  * beyond i_size on the filesystem.
1930                  */
1931                 while (retry--) {
1932                         int ret;
1933
1934                         ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1935                         if (ret == SHRINK_STOP)
1936                                 break;
1937                         if (ret)
1938                                 goto alloc_nohuge;
1939                 }
1940                 goto unlock;
1941         }
1942
1943         hindex = round_down(index, folio_nr_pages(folio));
1944
1945         if (sgp == SGP_WRITE)
1946                 __folio_set_referenced(folio);
1947
1948         error = shmem_add_to_page_cache(folio, mapping, hindex,
1949                                         NULL, gfp & GFP_RECLAIM_MASK,
1950                                         charge_mm);
1951         if (error)
1952                 goto unacct;
1953         folio_add_lru(folio);
1954
1955         spin_lock_irq(&info->lock);
1956         info->alloced += folio_nr_pages(folio);
1957         inode->i_blocks += (blkcnt_t)BLOCKS_PER_PAGE << folio_order(folio);
1958         shmem_recalc_inode(inode);
1959         spin_unlock_irq(&info->lock);
1960         alloced = true;
1961
1962         if (folio_test_pmd_mappable(folio) &&
1963             DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1964                         hindex + HPAGE_PMD_NR - 1) {
1965                 /*
1966                  * Part of the huge page is beyond i_size: subject
1967                  * to shrink under memory pressure.
1968                  */
1969                 spin_lock(&sbinfo->shrinklist_lock);
1970                 /*
1971                  * _careful to defend against unlocked access to
1972                  * ->shrink_list in shmem_unused_huge_shrink()
1973                  */
1974                 if (list_empty_careful(&info->shrinklist)) {
1975                         list_add_tail(&info->shrinklist,
1976                                       &sbinfo->shrinklist);
1977                         sbinfo->shrinklist_len++;
1978                 }
1979                 spin_unlock(&sbinfo->shrinklist_lock);
1980         }
1981
1982         /*
1983          * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1984          */
1985         if (sgp == SGP_FALLOC)
1986                 sgp = SGP_WRITE;
1987 clear:
1988         /*
1989          * Let SGP_WRITE caller clear ends if write does not fill page;
1990          * but SGP_FALLOC on a page fallocated earlier must initialize
1991          * it now, lest undo on failure cancel our earlier guarantee.
1992          */
1993         if (sgp != SGP_WRITE && !folio_test_uptodate(folio)) {
1994                 long i, n = folio_nr_pages(folio);
1995
1996                 for (i = 0; i < n; i++)
1997                         clear_highpage(folio_page(folio, i));
1998                 flush_dcache_folio(folio);
1999                 folio_mark_uptodate(folio);
2000         }
2001
2002         /* Perhaps the file has been truncated since we checked */
2003         if (sgp <= SGP_CACHE &&
2004             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
2005                 if (alloced) {
2006                         folio_clear_dirty(folio);
2007                         filemap_remove_folio(folio);
2008                         spin_lock_irq(&info->lock);
2009                         shmem_recalc_inode(inode);
2010                         spin_unlock_irq(&info->lock);
2011                 }
2012                 error = -EINVAL;
2013                 goto unlock;
2014         }
2015 out:
2016         *pagep = folio_page(folio, index - hindex);
2017         return 0;
2018
2019         /*
2020          * Error recovery.
2021          */
2022 unacct:
2023         shmem_inode_unacct_blocks(inode, folio_nr_pages(folio));
2024
2025         if (folio_test_large(folio)) {
2026                 folio_unlock(folio);
2027                 folio_put(folio);
2028                 goto alloc_nohuge;
2029         }
2030 unlock:
2031         if (folio) {
2032                 folio_unlock(folio);
2033                 folio_put(folio);
2034         }
2035         if (error == -ENOSPC && !once++) {
2036                 spin_lock_irq(&info->lock);
2037                 shmem_recalc_inode(inode);
2038                 spin_unlock_irq(&info->lock);
2039                 goto repeat;
2040         }
2041         if (error == -EEXIST)
2042                 goto repeat;
2043         return error;
2044 }
2045
2046 /*
2047  * This is like autoremove_wake_function, but it removes the wait queue
2048  * entry unconditionally - even if something else had already woken the
2049  * target.
2050  */
2051 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2052 {
2053         int ret = default_wake_function(wait, mode, sync, key);
2054         list_del_init(&wait->entry);
2055         return ret;
2056 }
2057
2058 static vm_fault_t shmem_fault(struct vm_fault *vmf)
2059 {
2060         struct vm_area_struct *vma = vmf->vma;
2061         struct inode *inode = file_inode(vma->vm_file);
2062         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2063         int err;
2064         vm_fault_t ret = VM_FAULT_LOCKED;
2065
2066         /*
2067          * Trinity finds that probing a hole which tmpfs is punching can
2068          * prevent the hole-punch from ever completing: which in turn
2069          * locks writers out with its hold on i_rwsem.  So refrain from
2070          * faulting pages into the hole while it's being punched.  Although
2071          * shmem_undo_range() does remove the additions, it may be unable to
2072          * keep up, as each new page needs its own unmap_mapping_range() call,
2073          * and the i_mmap tree grows ever slower to scan if new vmas are added.
2074          *
2075          * It does not matter if we sometimes reach this check just before the
2076          * hole-punch begins, so that one fault then races with the punch:
2077          * we just need to make racing faults a rare case.
2078          *
2079          * The implementation below would be much simpler if we just used a
2080          * standard mutex or completion: but we cannot take i_rwsem in fault,
2081          * and bloating every shmem inode for this unlikely case would be sad.
2082          */
2083         if (unlikely(inode->i_private)) {
2084                 struct shmem_falloc *shmem_falloc;
2085
2086                 spin_lock(&inode->i_lock);
2087                 shmem_falloc = inode->i_private;
2088                 if (shmem_falloc &&
2089                     shmem_falloc->waitq &&
2090                     vmf->pgoff >= shmem_falloc->start &&
2091                     vmf->pgoff < shmem_falloc->next) {
2092                         struct file *fpin;
2093                         wait_queue_head_t *shmem_falloc_waitq;
2094                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2095
2096                         ret = VM_FAULT_NOPAGE;
2097                         fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2098                         if (fpin)
2099                                 ret = VM_FAULT_RETRY;
2100
2101                         shmem_falloc_waitq = shmem_falloc->waitq;
2102                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2103                                         TASK_UNINTERRUPTIBLE);
2104                         spin_unlock(&inode->i_lock);
2105                         schedule();
2106
2107                         /*
2108                          * shmem_falloc_waitq points into the shmem_fallocate()
2109                          * stack of the hole-punching task: shmem_falloc_waitq
2110                          * is usually invalid by the time we reach here, but
2111                          * finish_wait() does not dereference it in that case;
2112                          * though i_lock needed lest racing with wake_up_all().
2113                          */
2114                         spin_lock(&inode->i_lock);
2115                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2116                         spin_unlock(&inode->i_lock);
2117
2118                         if (fpin)
2119                                 fput(fpin);
2120                         return ret;
2121                 }
2122                 spin_unlock(&inode->i_lock);
2123         }
2124
2125         err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, SGP_CACHE,
2126                                   gfp, vma, vmf, &ret);
2127         if (err)
2128                 return vmf_error(err);
2129         return ret;
2130 }
2131
2132 unsigned long shmem_get_unmapped_area(struct file *file,
2133                                       unsigned long uaddr, unsigned long len,
2134                                       unsigned long pgoff, unsigned long flags)
2135 {
2136         unsigned long (*get_area)(struct file *,
2137                 unsigned long, unsigned long, unsigned long, unsigned long);
2138         unsigned long addr;
2139         unsigned long offset;
2140         unsigned long inflated_len;
2141         unsigned long inflated_addr;
2142         unsigned long inflated_offset;
2143
2144         if (len > TASK_SIZE)
2145                 return -ENOMEM;
2146
2147         get_area = current->mm->get_unmapped_area;
2148         addr = get_area(file, uaddr, len, pgoff, flags);
2149
2150         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2151                 return addr;
2152         if (IS_ERR_VALUE(addr))
2153                 return addr;
2154         if (addr & ~PAGE_MASK)
2155                 return addr;
2156         if (addr > TASK_SIZE - len)
2157                 return addr;
2158
2159         if (shmem_huge == SHMEM_HUGE_DENY)
2160                 return addr;
2161         if (len < HPAGE_PMD_SIZE)
2162                 return addr;
2163         if (flags & MAP_FIXED)
2164                 return addr;
2165         /*
2166          * Our priority is to support MAP_SHARED mapped hugely;
2167          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2168          * But if caller specified an address hint and we allocated area there
2169          * successfully, respect that as before.
2170          */
2171         if (uaddr == addr)
2172                 return addr;
2173
2174         if (shmem_huge != SHMEM_HUGE_FORCE) {
2175                 struct super_block *sb;
2176
2177                 if (file) {
2178                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2179                         sb = file_inode(file)->i_sb;
2180                 } else {
2181                         /*
2182                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2183                          * for "/dev/zero", to create a shared anonymous object.
2184                          */
2185                         if (IS_ERR(shm_mnt))
2186                                 return addr;
2187                         sb = shm_mnt->mnt_sb;
2188                 }
2189                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2190                         return addr;
2191         }
2192
2193         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2194         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2195                 return addr;
2196         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2197                 return addr;
2198
2199         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2200         if (inflated_len > TASK_SIZE)
2201                 return addr;
2202         if (inflated_len < len)
2203                 return addr;
2204
2205         inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2206         if (IS_ERR_VALUE(inflated_addr))
2207                 return addr;
2208         if (inflated_addr & ~PAGE_MASK)
2209                 return addr;
2210
2211         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2212         inflated_addr += offset - inflated_offset;
2213         if (inflated_offset > offset)
2214                 inflated_addr += HPAGE_PMD_SIZE;
2215
2216         if (inflated_addr > TASK_SIZE - len)
2217                 return addr;
2218         return inflated_addr;
2219 }
2220
2221 #ifdef CONFIG_NUMA
2222 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2223 {
2224         struct inode *inode = file_inode(vma->vm_file);
2225         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2226 }
2227
2228 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2229                                           unsigned long addr)
2230 {
2231         struct inode *inode = file_inode(vma->vm_file);
2232         pgoff_t index;
2233
2234         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2235         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2236 }
2237 #endif
2238
2239 int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
2240 {
2241         struct inode *inode = file_inode(file);
2242         struct shmem_inode_info *info = SHMEM_I(inode);
2243         int retval = -ENOMEM;
2244
2245         /*
2246          * What serializes the accesses to info->flags?
2247          * ipc_lock_object() when called from shmctl_do_lock(),
2248          * no serialization needed when called from shm_destroy().
2249          */
2250         if (lock && !(info->flags & VM_LOCKED)) {
2251                 if (!user_shm_lock(inode->i_size, ucounts))
2252                         goto out_nomem;
2253                 info->flags |= VM_LOCKED;
2254                 mapping_set_unevictable(file->f_mapping);
2255         }
2256         if (!lock && (info->flags & VM_LOCKED) && ucounts) {
2257                 user_shm_unlock(inode->i_size, ucounts);
2258                 info->flags &= ~VM_LOCKED;
2259                 mapping_clear_unevictable(file->f_mapping);
2260         }
2261         retval = 0;
2262
2263 out_nomem:
2264         return retval;
2265 }
2266
2267 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2268 {
2269         struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2270         int ret;
2271
2272         ret = seal_check_future_write(info->seals, vma);
2273         if (ret)
2274                 return ret;
2275
2276         /* arm64 - allow memory tagging on RAM-based files */
2277         vma->vm_flags |= VM_MTE_ALLOWED;
2278
2279         file_accessed(file);
2280         vma->vm_ops = &shmem_vm_ops;
2281         return 0;
2282 }
2283
2284 /* Mask out flags that are inappropriate for the given type of inode. */
2285 static unsigned shmem_mask_flags(umode_t mode, __u32 flags)
2286 {
2287         if (S_ISDIR(mode))
2288                 return flags;
2289         else if (S_ISREG(mode))
2290                 return flags & SHMEM_REG_FLMASK;
2291         else
2292                 return flags & SHMEM_OTHER_FLMASK;
2293 }
2294
2295 static struct inode *shmem_get_inode(struct super_block *sb, struct inode *dir,
2296                                      umode_t mode, dev_t dev, unsigned long flags)
2297 {
2298         struct inode *inode;
2299         struct shmem_inode_info *info;
2300         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2301         ino_t ino;
2302
2303         if (shmem_reserve_inode(sb, &ino))
2304                 return NULL;
2305
2306         inode = new_inode(sb);
2307         if (inode) {
2308                 inode->i_ino = ino;
2309                 inode_init_owner(&init_user_ns, inode, dir, mode);
2310                 inode->i_blocks = 0;
2311                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2312                 inode->i_generation = prandom_u32();
2313                 info = SHMEM_I(inode);
2314                 memset(info, 0, (char *)inode - (char *)info);
2315                 spin_lock_init(&info->lock);
2316                 atomic_set(&info->stop_eviction, 0);
2317                 info->seals = F_SEAL_SEAL;
2318                 info->flags = flags & VM_NORESERVE;
2319                 info->i_crtime = inode->i_mtime;
2320                 info->fsflags = (dir == NULL) ? 0 :
2321                         SHMEM_I(dir)->fsflags & SHMEM_FL_INHERITED;
2322                 info->fsflags = shmem_mask_flags(mode, info->fsflags);
2323                 INIT_LIST_HEAD(&info->shrinklist);
2324                 INIT_LIST_HEAD(&info->swaplist);
2325                 simple_xattrs_init(&info->xattrs);
2326                 cache_no_acl(inode);
2327                 mapping_set_large_folios(inode->i_mapping);
2328
2329                 switch (mode & S_IFMT) {
2330                 default:
2331                         inode->i_op = &shmem_special_inode_operations;
2332                         init_special_inode(inode, mode, dev);
2333                         break;
2334                 case S_IFREG:
2335                         inode->i_mapping->a_ops = &shmem_aops;
2336                         inode->i_op = &shmem_inode_operations;
2337                         inode->i_fop = &shmem_file_operations;
2338                         mpol_shared_policy_init(&info->policy,
2339                                                  shmem_get_sbmpol(sbinfo));
2340                         break;
2341                 case S_IFDIR:
2342                         inc_nlink(inode);
2343                         /* Some things misbehave if size == 0 on a directory */
2344                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2345                         inode->i_op = &shmem_dir_inode_operations;
2346                         inode->i_fop = &simple_dir_operations;
2347                         break;
2348                 case S_IFLNK:
2349                         /*
2350                          * Must not load anything in the rbtree,
2351                          * mpol_free_shared_policy will not be called.
2352                          */
2353                         mpol_shared_policy_init(&info->policy, NULL);
2354                         break;
2355                 }
2356
2357                 lockdep_annotate_inode_mutex_key(inode);
2358         } else
2359                 shmem_free_inode(sb);
2360         return inode;
2361 }
2362
2363 #ifdef CONFIG_USERFAULTFD
2364 int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2365                            pmd_t *dst_pmd,
2366                            struct vm_area_struct *dst_vma,
2367                            unsigned long dst_addr,
2368                            unsigned long src_addr,
2369                            bool zeropage, bool wp_copy,
2370                            struct page **pagep)
2371 {
2372         struct inode *inode = file_inode(dst_vma->vm_file);
2373         struct shmem_inode_info *info = SHMEM_I(inode);
2374         struct address_space *mapping = inode->i_mapping;
2375         gfp_t gfp = mapping_gfp_mask(mapping);
2376         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2377         void *page_kaddr;
2378         struct folio *folio;
2379         struct page *page;
2380         int ret;
2381         pgoff_t max_off;
2382
2383         if (!shmem_inode_acct_block(inode, 1)) {
2384                 /*
2385                  * We may have got a page, returned -ENOENT triggering a retry,
2386                  * and now we find ourselves with -ENOMEM. Release the page, to
2387                  * avoid a BUG_ON in our caller.
2388                  */
2389                 if (unlikely(*pagep)) {
2390                         put_page(*pagep);
2391                         *pagep = NULL;
2392                 }
2393                 return -ENOMEM;
2394         }
2395
2396         if (!*pagep) {
2397                 ret = -ENOMEM;
2398                 page = shmem_alloc_page(gfp, info, pgoff);
2399                 if (!page)
2400                         goto out_unacct_blocks;
2401
2402                 if (!zeropage) {        /* COPY */
2403                         page_kaddr = kmap_atomic(page);
2404                         ret = copy_from_user(page_kaddr,
2405                                              (const void __user *)src_addr,
2406                                              PAGE_SIZE);
2407                         kunmap_atomic(page_kaddr);
2408
2409                         /* fallback to copy_from_user outside mmap_lock */
2410                         if (unlikely(ret)) {
2411                                 *pagep = page;
2412                                 ret = -ENOENT;
2413                                 /* don't free the page */
2414                                 goto out_unacct_blocks;
2415                         }
2416
2417                         flush_dcache_page(page);
2418                 } else {                /* ZEROPAGE */
2419                         clear_user_highpage(page, dst_addr);
2420                 }
2421         } else {
2422                 page = *pagep;
2423                 *pagep = NULL;
2424         }
2425
2426         VM_BUG_ON(PageLocked(page));
2427         VM_BUG_ON(PageSwapBacked(page));
2428         __SetPageLocked(page);
2429         __SetPageSwapBacked(page);
2430         __SetPageUptodate(page);
2431
2432         ret = -EFAULT;
2433         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2434         if (unlikely(pgoff >= max_off))
2435                 goto out_release;
2436
2437         folio = page_folio(page);
2438         ret = shmem_add_to_page_cache(folio, mapping, pgoff, NULL,
2439                                       gfp & GFP_RECLAIM_MASK, dst_mm);
2440         if (ret)
2441                 goto out_release;
2442
2443         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
2444                                        page, true, wp_copy);
2445         if (ret)
2446                 goto out_delete_from_cache;
2447
2448         spin_lock_irq(&info->lock);
2449         info->alloced++;
2450         inode->i_blocks += BLOCKS_PER_PAGE;
2451         shmem_recalc_inode(inode);
2452         spin_unlock_irq(&info->lock);
2453
2454         unlock_page(page);
2455         return 0;
2456 out_delete_from_cache:
2457         delete_from_page_cache(page);
2458 out_release:
2459         unlock_page(page);
2460         put_page(page);
2461 out_unacct_blocks:
2462         shmem_inode_unacct_blocks(inode, 1);
2463         return ret;
2464 }
2465 #endif /* CONFIG_USERFAULTFD */
2466
2467 #ifdef CONFIG_TMPFS
2468 static const struct inode_operations shmem_symlink_inode_operations;
2469 static const struct inode_operations shmem_short_symlink_operations;
2470
2471 #ifdef CONFIG_TMPFS_XATTR
2472 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2473 #else
2474 #define shmem_initxattrs NULL
2475 #endif
2476
2477 static int
2478 shmem_write_begin(struct file *file, struct address_space *mapping,
2479                         loff_t pos, unsigned len,
2480                         struct page **pagep, void **fsdata)
2481 {
2482         struct inode *inode = mapping->host;
2483         struct shmem_inode_info *info = SHMEM_I(inode);
2484         pgoff_t index = pos >> PAGE_SHIFT;
2485         int ret = 0;
2486
2487         /* i_rwsem is held by caller */
2488         if (unlikely(info->seals & (F_SEAL_GROW |
2489                                    F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2490                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2491                         return -EPERM;
2492                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2493                         return -EPERM;
2494         }
2495
2496         ret = shmem_getpage(inode, index, pagep, SGP_WRITE);
2497
2498         if (ret)
2499                 return ret;
2500
2501         if (PageHWPoison(*pagep)) {
2502                 unlock_page(*pagep);
2503                 put_page(*pagep);
2504                 *pagep = NULL;
2505                 return -EIO;
2506         }
2507
2508         return 0;
2509 }
2510
2511 static int
2512 shmem_write_end(struct file *file, struct address_space *mapping,
2513                         loff_t pos, unsigned len, unsigned copied,
2514                         struct page *page, void *fsdata)
2515 {
2516         struct inode *inode = mapping->host;
2517
2518         if (pos + copied > inode->i_size)
2519                 i_size_write(inode, pos + copied);
2520
2521         if (!PageUptodate(page)) {
2522                 struct page *head = compound_head(page);
2523                 if (PageTransCompound(page)) {
2524                         int i;
2525
2526                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2527                                 if (head + i == page)
2528                                         continue;
2529                                 clear_highpage(head + i);
2530                                 flush_dcache_page(head + i);
2531                         }
2532                 }
2533                 if (copied < PAGE_SIZE) {
2534                         unsigned from = pos & (PAGE_SIZE - 1);
2535                         zero_user_segments(page, 0, from,
2536                                         from + copied, PAGE_SIZE);
2537                 }
2538                 SetPageUptodate(head);
2539         }
2540         set_page_dirty(page);
2541         unlock_page(page);
2542         put_page(page);
2543
2544         return copied;
2545 }
2546
2547 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2548 {
2549         struct file *file = iocb->ki_filp;
2550         struct inode *inode = file_inode(file);
2551         struct address_space *mapping = inode->i_mapping;
2552         pgoff_t index;
2553         unsigned long offset;
2554         int error = 0;
2555         ssize_t retval = 0;
2556         loff_t *ppos = &iocb->ki_pos;
2557
2558         index = *ppos >> PAGE_SHIFT;
2559         offset = *ppos & ~PAGE_MASK;
2560
2561         for (;;) {
2562                 struct page *page = NULL;
2563                 pgoff_t end_index;
2564                 unsigned long nr, ret;
2565                 loff_t i_size = i_size_read(inode);
2566
2567                 end_index = i_size >> PAGE_SHIFT;
2568                 if (index > end_index)
2569                         break;
2570                 if (index == end_index) {
2571                         nr = i_size & ~PAGE_MASK;
2572                         if (nr <= offset)
2573                                 break;
2574                 }
2575
2576                 error = shmem_getpage(inode, index, &page, SGP_READ);
2577                 if (error) {
2578                         if (error == -EINVAL)
2579                                 error = 0;
2580                         break;
2581                 }
2582                 if (page) {
2583                         unlock_page(page);
2584
2585                         if (PageHWPoison(page)) {
2586                                 put_page(page);
2587                                 error = -EIO;
2588                                 break;
2589                         }
2590                 }
2591
2592                 /*
2593                  * We must evaluate after, since reads (unlike writes)
2594                  * are called without i_rwsem protection against truncate
2595                  */
2596                 nr = PAGE_SIZE;
2597                 i_size = i_size_read(inode);
2598                 end_index = i_size >> PAGE_SHIFT;
2599                 if (index == end_index) {
2600                         nr = i_size & ~PAGE_MASK;
2601                         if (nr <= offset) {
2602                                 if (page)
2603                                         put_page(page);
2604                                 break;
2605                         }
2606                 }
2607                 nr -= offset;
2608
2609                 if (page) {
2610                         /*
2611                          * If users can be writing to this page using arbitrary
2612                          * virtual addresses, take care about potential aliasing
2613                          * before reading the page on the kernel side.
2614                          */
2615                         if (mapping_writably_mapped(mapping))
2616                                 flush_dcache_page(page);
2617                         /*
2618                          * Mark the page accessed if we read the beginning.
2619                          */
2620                         if (!offset)
2621                                 mark_page_accessed(page);
2622                         /*
2623                          * Ok, we have the page, and it's up-to-date, so
2624                          * now we can copy it to user space...
2625                          */
2626                         ret = copy_page_to_iter(page, offset, nr, to);
2627                         put_page(page);
2628
2629                 } else if (user_backed_iter(to)) {
2630                         /*
2631                          * Copy to user tends to be so well optimized, but
2632                          * clear_user() not so much, that it is noticeably
2633                          * faster to copy the zero page instead of clearing.
2634                          */
2635                         ret = copy_page_to_iter(ZERO_PAGE(0), offset, nr, to);
2636                 } else {
2637                         /*
2638                          * But submitting the same page twice in a row to
2639                          * splice() - or others? - can result in confusion:
2640                          * so don't attempt that optimization on pipes etc.
2641                          */
2642                         ret = iov_iter_zero(nr, to);
2643                 }
2644
2645                 retval += ret;
2646                 offset += ret;
2647                 index += offset >> PAGE_SHIFT;
2648                 offset &= ~PAGE_MASK;
2649
2650                 if (!iov_iter_count(to))
2651                         break;
2652                 if (ret < nr) {
2653                         error = -EFAULT;
2654                         break;
2655                 }
2656                 cond_resched();
2657         }
2658
2659         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2660         file_accessed(file);
2661         return retval ? retval : error;
2662 }
2663
2664 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2665 {
2666         struct address_space *mapping = file->f_mapping;
2667         struct inode *inode = mapping->host;
2668
2669         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2670                 return generic_file_llseek_size(file, offset, whence,
2671                                         MAX_LFS_FILESIZE, i_size_read(inode));
2672         if (offset < 0)
2673                 return -ENXIO;
2674
2675         inode_lock(inode);
2676         /* We're holding i_rwsem so we can access i_size directly */
2677         offset = mapping_seek_hole_data(mapping, offset, inode->i_size, whence);
2678         if (offset >= 0)
2679                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2680         inode_unlock(inode);
2681         return offset;
2682 }
2683
2684 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2685                                                          loff_t len)
2686 {
2687         struct inode *inode = file_inode(file);
2688         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2689         struct shmem_inode_info *info = SHMEM_I(inode);
2690         struct shmem_falloc shmem_falloc;
2691         pgoff_t start, index, end, undo_fallocend;
2692         int error;
2693
2694         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2695                 return -EOPNOTSUPP;
2696
2697         inode_lock(inode);
2698
2699         if (mode & FALLOC_FL_PUNCH_HOLE) {
2700                 struct address_space *mapping = file->f_mapping;
2701                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2702                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2703                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2704
2705                 /* protected by i_rwsem */
2706                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2707                         error = -EPERM;
2708                         goto out;
2709                 }
2710
2711                 shmem_falloc.waitq = &shmem_falloc_waitq;
2712                 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2713                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2714                 spin_lock(&inode->i_lock);
2715                 inode->i_private = &shmem_falloc;
2716                 spin_unlock(&inode->i_lock);
2717
2718                 if ((u64)unmap_end > (u64)unmap_start)
2719                         unmap_mapping_range(mapping, unmap_start,
2720                                             1 + unmap_end - unmap_start, 0);
2721                 shmem_truncate_range(inode, offset, offset + len - 1);
2722                 /* No need to unmap again: hole-punching leaves COWed pages */
2723
2724                 spin_lock(&inode->i_lock);
2725                 inode->i_private = NULL;
2726                 wake_up_all(&shmem_falloc_waitq);
2727                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2728                 spin_unlock(&inode->i_lock);
2729                 error = 0;
2730                 goto out;
2731         }
2732
2733         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2734         error = inode_newsize_ok(inode, offset + len);
2735         if (error)
2736                 goto out;
2737
2738         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2739                 error = -EPERM;
2740                 goto out;
2741         }
2742
2743         start = offset >> PAGE_SHIFT;
2744         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2745         /* Try to avoid a swapstorm if len is impossible to satisfy */
2746         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2747                 error = -ENOSPC;
2748                 goto out;
2749         }
2750
2751         shmem_falloc.waitq = NULL;
2752         shmem_falloc.start = start;
2753         shmem_falloc.next  = start;
2754         shmem_falloc.nr_falloced = 0;
2755         shmem_falloc.nr_unswapped = 0;
2756         spin_lock(&inode->i_lock);
2757         inode->i_private = &shmem_falloc;
2758         spin_unlock(&inode->i_lock);
2759
2760         /*
2761          * info->fallocend is only relevant when huge pages might be
2762          * involved: to prevent split_huge_page() freeing fallocated
2763          * pages when FALLOC_FL_KEEP_SIZE committed beyond i_size.
2764          */
2765         undo_fallocend = info->fallocend;
2766         if (info->fallocend < end)
2767                 info->fallocend = end;
2768
2769         for (index = start; index < end; ) {
2770                 struct page *page;
2771
2772                 /*
2773                  * Good, the fallocate(2) manpage permits EINTR: we may have
2774                  * been interrupted because we are using up too much memory.
2775                  */
2776                 if (signal_pending(current))
2777                         error = -EINTR;
2778                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2779                         error = -ENOMEM;
2780                 else
2781                         error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2782                 if (error) {
2783                         info->fallocend = undo_fallocend;
2784                         /* Remove the !PageUptodate pages we added */
2785                         if (index > start) {
2786                                 shmem_undo_range(inode,
2787                                     (loff_t)start << PAGE_SHIFT,
2788                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2789                         }
2790                         goto undone;
2791                 }
2792
2793                 index++;
2794                 /*
2795                  * Here is a more important optimization than it appears:
2796                  * a second SGP_FALLOC on the same huge page will clear it,
2797                  * making it PageUptodate and un-undoable if we fail later.
2798                  */
2799                 if (PageTransCompound(page)) {
2800                         index = round_up(index, HPAGE_PMD_NR);
2801                         /* Beware 32-bit wraparound */
2802                         if (!index)
2803                                 index--;
2804                 }
2805
2806                 /*
2807                  * Inform shmem_writepage() how far we have reached.
2808                  * No need for lock or barrier: we have the page lock.
2809                  */
2810                 if (!PageUptodate(page))
2811                         shmem_falloc.nr_falloced += index - shmem_falloc.next;
2812                 shmem_falloc.next = index;
2813
2814                 /*
2815                  * If !PageUptodate, leave it that way so that freeable pages
2816                  * can be recognized if we need to rollback on error later.
2817                  * But set_page_dirty so that memory pressure will swap rather
2818                  * than free the pages we are allocating (and SGP_CACHE pages
2819                  * might still be clean: we now need to mark those dirty too).
2820                  */
2821                 set_page_dirty(page);
2822                 unlock_page(page);
2823                 put_page(page);
2824                 cond_resched();
2825         }
2826
2827         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2828                 i_size_write(inode, offset + len);
2829         inode->i_ctime = current_time(inode);
2830 undone:
2831         spin_lock(&inode->i_lock);
2832         inode->i_private = NULL;
2833         spin_unlock(&inode->i_lock);
2834 out:
2835         inode_unlock(inode);
2836         return error;
2837 }
2838
2839 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2840 {
2841         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2842
2843         buf->f_type = TMPFS_MAGIC;
2844         buf->f_bsize = PAGE_SIZE;
2845         buf->f_namelen = NAME_MAX;
2846         if (sbinfo->max_blocks) {
2847                 buf->f_blocks = sbinfo->max_blocks;
2848                 buf->f_bavail =
2849                 buf->f_bfree  = sbinfo->max_blocks -
2850                                 percpu_counter_sum(&sbinfo->used_blocks);
2851         }
2852         if (sbinfo->max_inodes) {
2853                 buf->f_files = sbinfo->max_inodes;
2854                 buf->f_ffree = sbinfo->free_inodes;
2855         }
2856         /* else leave those fields 0 like simple_statfs */
2857
2858         buf->f_fsid = uuid_to_fsid(dentry->d_sb->s_uuid.b);
2859
2860         return 0;
2861 }
2862
2863 /*
2864  * File creation. Allocate an inode, and we're done..
2865  */
2866 static int
2867 shmem_mknod(struct user_namespace *mnt_userns, struct inode *dir,
2868             struct dentry *dentry, umode_t mode, dev_t dev)
2869 {
2870         struct inode *inode;
2871         int error = -ENOSPC;
2872
2873         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2874         if (inode) {
2875                 error = simple_acl_create(dir, inode);
2876                 if (error)
2877                         goto out_iput;
2878                 error = security_inode_init_security(inode, dir,
2879                                                      &dentry->d_name,
2880                                                      shmem_initxattrs, NULL);
2881                 if (error && error != -EOPNOTSUPP)
2882                         goto out_iput;
2883
2884                 error = 0;
2885                 dir->i_size += BOGO_DIRENT_SIZE;
2886                 dir->i_ctime = dir->i_mtime = current_time(dir);
2887                 d_instantiate(dentry, inode);
2888                 dget(dentry); /* Extra count - pin the dentry in core */
2889         }
2890         return error;
2891 out_iput:
2892         iput(inode);
2893         return error;
2894 }
2895
2896 static int
2897 shmem_tmpfile(struct user_namespace *mnt_userns, struct inode *dir,
2898               struct dentry *dentry, umode_t mode)
2899 {
2900         struct inode *inode;
2901         int error = -ENOSPC;
2902
2903         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2904         if (inode) {
2905                 error = security_inode_init_security(inode, dir,
2906                                                      NULL,
2907                                                      shmem_initxattrs, NULL);
2908                 if (error && error != -EOPNOTSUPP)
2909                         goto out_iput;
2910                 error = simple_acl_create(dir, inode);
2911                 if (error)
2912                         goto out_iput;
2913                 d_tmpfile(dentry, inode);
2914         }
2915         return error;
2916 out_iput:
2917         iput(inode);
2918         return error;
2919 }
2920
2921 static int shmem_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
2922                        struct dentry *dentry, umode_t mode)
2923 {
2924         int error;
2925
2926         if ((error = shmem_mknod(&init_user_ns, dir, dentry,
2927                                  mode | S_IFDIR, 0)))
2928                 return error;
2929         inc_nlink(dir);
2930         return 0;
2931 }
2932
2933 static int shmem_create(struct user_namespace *mnt_userns, struct inode *dir,
2934                         struct dentry *dentry, umode_t mode, bool excl)
2935 {
2936         return shmem_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
2937 }
2938
2939 /*
2940  * Link a file..
2941  */
2942 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2943 {
2944         struct inode *inode = d_inode(old_dentry);
2945         int ret = 0;
2946
2947         /*
2948          * No ordinary (disk based) filesystem counts links as inodes;
2949          * but each new link needs a new dentry, pinning lowmem, and
2950          * tmpfs dentries cannot be pruned until they are unlinked.
2951          * But if an O_TMPFILE file is linked into the tmpfs, the
2952          * first link must skip that, to get the accounting right.
2953          */
2954         if (inode->i_nlink) {
2955                 ret = shmem_reserve_inode(inode->i_sb, NULL);
2956                 if (ret)
2957                         goto out;
2958         }
2959
2960         dir->i_size += BOGO_DIRENT_SIZE;
2961         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2962         inc_nlink(inode);
2963         ihold(inode);   /* New dentry reference */
2964         dget(dentry);           /* Extra pinning count for the created dentry */
2965         d_instantiate(dentry, inode);
2966 out:
2967         return ret;
2968 }
2969
2970 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2971 {
2972         struct inode *inode = d_inode(dentry);
2973
2974         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2975                 shmem_free_inode(inode->i_sb);
2976
2977         dir->i_size -= BOGO_DIRENT_SIZE;
2978         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2979         drop_nlink(inode);
2980         dput(dentry);   /* Undo the count from "create" - this does all the work */
2981         return 0;
2982 }
2983
2984 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2985 {
2986         if (!simple_empty(dentry))
2987                 return -ENOTEMPTY;
2988
2989         drop_nlink(d_inode(dentry));
2990         drop_nlink(dir);
2991         return shmem_unlink(dir, dentry);
2992 }
2993
2994 static int shmem_whiteout(struct user_namespace *mnt_userns,
2995                           struct inode *old_dir, struct dentry *old_dentry)
2996 {
2997         struct dentry *whiteout;
2998         int error;
2999
3000         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3001         if (!whiteout)
3002                 return -ENOMEM;
3003
3004         error = shmem_mknod(&init_user_ns, old_dir, whiteout,
3005                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3006         dput(whiteout);
3007         if (error)
3008                 return error;
3009
3010         /*
3011          * Cheat and hash the whiteout while the old dentry is still in
3012          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3013          *
3014          * d_lookup() will consistently find one of them at this point,
3015          * not sure which one, but that isn't even important.
3016          */
3017         d_rehash(whiteout);
3018         return 0;
3019 }
3020
3021 /*
3022  * The VFS layer already does all the dentry stuff for rename,
3023  * we just have to decrement the usage count for the target if
3024  * it exists so that the VFS layer correctly free's it when it
3025  * gets overwritten.
3026  */
3027 static int shmem_rename2(struct user_namespace *mnt_userns,
3028                          struct inode *old_dir, struct dentry *old_dentry,
3029                          struct inode *new_dir, struct dentry *new_dentry,
3030                          unsigned int flags)
3031 {
3032         struct inode *inode = d_inode(old_dentry);
3033         int they_are_dirs = S_ISDIR(inode->i_mode);
3034
3035         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3036                 return -EINVAL;
3037
3038         if (flags & RENAME_EXCHANGE)
3039                 return simple_rename_exchange(old_dir, old_dentry, new_dir, new_dentry);
3040
3041         if (!simple_empty(new_dentry))
3042                 return -ENOTEMPTY;
3043
3044         if (flags & RENAME_WHITEOUT) {
3045                 int error;
3046
3047                 error = shmem_whiteout(&init_user_ns, old_dir, old_dentry);
3048                 if (error)
3049                         return error;
3050         }
3051
3052         if (d_really_is_positive(new_dentry)) {
3053                 (void) shmem_unlink(new_dir, new_dentry);
3054                 if (they_are_dirs) {
3055                         drop_nlink(d_inode(new_dentry));
3056                         drop_nlink(old_dir);
3057                 }
3058         } else if (they_are_dirs) {
3059                 drop_nlink(old_dir);
3060                 inc_nlink(new_dir);
3061         }
3062
3063         old_dir->i_size -= BOGO_DIRENT_SIZE;
3064         new_dir->i_size += BOGO_DIRENT_SIZE;
3065         old_dir->i_ctime = old_dir->i_mtime =
3066         new_dir->i_ctime = new_dir->i_mtime =
3067         inode->i_ctime = current_time(old_dir);
3068         return 0;
3069 }
3070
3071 static int shmem_symlink(struct user_namespace *mnt_userns, struct inode *dir,
3072                          struct dentry *dentry, const char *symname)
3073 {
3074         int error;
3075         int len;
3076         struct inode *inode;
3077         struct page *page;
3078
3079         len = strlen(symname) + 1;
3080         if (len > PAGE_SIZE)
3081                 return -ENAMETOOLONG;
3082
3083         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3084                                 VM_NORESERVE);
3085         if (!inode)
3086                 return -ENOSPC;
3087
3088         error = security_inode_init_security(inode, dir, &dentry->d_name,
3089                                              shmem_initxattrs, NULL);
3090         if (error && error != -EOPNOTSUPP) {
3091                 iput(inode);
3092                 return error;
3093         }
3094
3095         inode->i_size = len-1;
3096         if (len <= SHORT_SYMLINK_LEN) {
3097                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3098                 if (!inode->i_link) {
3099                         iput(inode);
3100                         return -ENOMEM;
3101                 }
3102                 inode->i_op = &shmem_short_symlink_operations;
3103         } else {
3104                 inode_nohighmem(inode);
3105                 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3106                 if (error) {
3107                         iput(inode);
3108                         return error;
3109                 }
3110                 inode->i_mapping->a_ops = &shmem_aops;
3111                 inode->i_op = &shmem_symlink_inode_operations;
3112                 memcpy(page_address(page), symname, len);
3113                 SetPageUptodate(page);
3114                 set_page_dirty(page);
3115                 unlock_page(page);
3116                 put_page(page);
3117         }
3118         dir->i_size += BOGO_DIRENT_SIZE;
3119         dir->i_ctime = dir->i_mtime = current_time(dir);
3120         d_instantiate(dentry, inode);
3121         dget(dentry);
3122         return 0;
3123 }
3124
3125 static void shmem_put_link(void *arg)
3126 {
3127         mark_page_accessed(arg);
3128         put_page(arg);
3129 }
3130
3131 static const char *shmem_get_link(struct dentry *dentry,
3132                                   struct inode *inode,
3133                                   struct delayed_call *done)
3134 {
3135         struct page *page = NULL;
3136         int error;
3137         if (!dentry) {
3138                 page = find_get_page(inode->i_mapping, 0);
3139                 if (!page)
3140                         return ERR_PTR(-ECHILD);
3141                 if (PageHWPoison(page) ||
3142                     !PageUptodate(page)) {
3143                         put_page(page);
3144                         return ERR_PTR(-ECHILD);
3145                 }
3146         } else {
3147                 error = shmem_getpage(inode, 0, &page, SGP_READ);
3148                 if (error)
3149                         return ERR_PTR(error);
3150                 if (!page)
3151                         return ERR_PTR(-ECHILD);
3152                 if (PageHWPoison(page)) {
3153                         unlock_page(page);
3154                         put_page(page);
3155                         return ERR_PTR(-ECHILD);
3156                 }
3157                 unlock_page(page);
3158         }
3159         set_delayed_call(done, shmem_put_link, page);
3160         return page_address(page);
3161 }
3162
3163 #ifdef CONFIG_TMPFS_XATTR
3164
3165 static int shmem_fileattr_get(struct dentry *dentry, struct fileattr *fa)
3166 {
3167         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3168
3169         fileattr_fill_flags(fa, info->fsflags & SHMEM_FL_USER_VISIBLE);
3170
3171         return 0;
3172 }
3173
3174 static int shmem_fileattr_set(struct user_namespace *mnt_userns,
3175                               struct dentry *dentry, struct fileattr *fa)
3176 {
3177         struct inode *inode = d_inode(dentry);
3178         struct shmem_inode_info *info = SHMEM_I(inode);
3179
3180         if (fileattr_has_fsx(fa))
3181                 return -EOPNOTSUPP;
3182
3183         info->fsflags = (info->fsflags & ~SHMEM_FL_USER_MODIFIABLE) |
3184                 (fa->flags & SHMEM_FL_USER_MODIFIABLE);
3185
3186         inode->i_flags &= ~(S_APPEND | S_IMMUTABLE | S_NOATIME);
3187         if (info->fsflags & FS_APPEND_FL)
3188                 inode->i_flags |= S_APPEND;
3189         if (info->fsflags & FS_IMMUTABLE_FL)
3190                 inode->i_flags |= S_IMMUTABLE;
3191         if (info->fsflags & FS_NOATIME_FL)
3192                 inode->i_flags |= S_NOATIME;
3193
3194         inode->i_ctime = current_time(inode);
3195         return 0;
3196 }
3197
3198 /*
3199  * Superblocks without xattr inode operations may get some security.* xattr
3200  * support from the LSM "for free". As soon as we have any other xattrs
3201  * like ACLs, we also need to implement the security.* handlers at
3202  * filesystem level, though.
3203  */
3204
3205 /*
3206  * Callback for security_inode_init_security() for acquiring xattrs.
3207  */
3208 static int shmem_initxattrs(struct inode *inode,
3209                             const struct xattr *xattr_array,
3210                             void *fs_info)
3211 {
3212         struct shmem_inode_info *info = SHMEM_I(inode);
3213         const struct xattr *xattr;
3214         struct simple_xattr *new_xattr;
3215         size_t len;
3216
3217         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3218                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3219                 if (!new_xattr)
3220                         return -ENOMEM;
3221
3222                 len = strlen(xattr->name) + 1;
3223                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3224                                           GFP_KERNEL);
3225                 if (!new_xattr->name) {
3226                         kvfree(new_xattr);
3227                         return -ENOMEM;
3228                 }
3229
3230                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3231                        XATTR_SECURITY_PREFIX_LEN);
3232                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3233                        xattr->name, len);
3234
3235                 simple_xattr_list_add(&info->xattrs, new_xattr);
3236         }
3237
3238         return 0;
3239 }
3240
3241 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3242                                    struct dentry *unused, struct inode *inode,
3243                                    const char *name, void *buffer, size_t size)
3244 {
3245         struct shmem_inode_info *info = SHMEM_I(inode);
3246
3247         name = xattr_full_name(handler, name);
3248         return simple_xattr_get(&info->xattrs, name, buffer, size);
3249 }
3250
3251 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3252                                    struct user_namespace *mnt_userns,
3253                                    struct dentry *unused, struct inode *inode,
3254                                    const char *name, const void *value,
3255                                    size_t size, int flags)
3256 {
3257         struct shmem_inode_info *info = SHMEM_I(inode);
3258
3259         name = xattr_full_name(handler, name);
3260         return simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3261 }
3262
3263 static const struct xattr_handler shmem_security_xattr_handler = {
3264         .prefix = XATTR_SECURITY_PREFIX,
3265         .get = shmem_xattr_handler_get,
3266         .set = shmem_xattr_handler_set,
3267 };
3268
3269 static const struct xattr_handler shmem_trusted_xattr_handler = {
3270         .prefix = XATTR_TRUSTED_PREFIX,
3271         .get = shmem_xattr_handler_get,
3272         .set = shmem_xattr_handler_set,
3273 };
3274
3275 static const struct xattr_handler *shmem_xattr_handlers[] = {
3276 #ifdef CONFIG_TMPFS_POSIX_ACL
3277         &posix_acl_access_xattr_handler,
3278         &posix_acl_default_xattr_handler,
3279 #endif
3280         &shmem_security_xattr_handler,
3281         &shmem_trusted_xattr_handler,
3282         NULL
3283 };
3284
3285 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3286 {
3287         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3288         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3289 }
3290 #endif /* CONFIG_TMPFS_XATTR */
3291
3292 static const struct inode_operations shmem_short_symlink_operations = {
3293         .getattr        = shmem_getattr,
3294         .get_link       = simple_get_link,
3295 #ifdef CONFIG_TMPFS_XATTR
3296         .listxattr      = shmem_listxattr,
3297 #endif
3298 };
3299
3300 static const struct inode_operations shmem_symlink_inode_operations = {
3301         .getattr        = shmem_getattr,
3302         .get_link       = shmem_get_link,
3303 #ifdef CONFIG_TMPFS_XATTR
3304         .listxattr      = shmem_listxattr,
3305 #endif
3306 };
3307
3308 static struct dentry *shmem_get_parent(struct dentry *child)
3309 {
3310         return ERR_PTR(-ESTALE);
3311 }
3312
3313 static int shmem_match(struct inode *ino, void *vfh)
3314 {
3315         __u32 *fh = vfh;
3316         __u64 inum = fh[2];
3317         inum = (inum << 32) | fh[1];
3318         return ino->i_ino == inum && fh[0] == ino->i_generation;
3319 }
3320
3321 /* Find any alias of inode, but prefer a hashed alias */
3322 static struct dentry *shmem_find_alias(struct inode *inode)
3323 {
3324         struct dentry *alias = d_find_alias(inode);
3325
3326         return alias ?: d_find_any_alias(inode);
3327 }
3328
3329
3330 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3331                 struct fid *fid, int fh_len, int fh_type)
3332 {
3333         struct inode *inode;
3334         struct dentry *dentry = NULL;
3335         u64 inum;
3336
3337         if (fh_len < 3)
3338                 return NULL;
3339
3340         inum = fid->raw[2];
3341         inum = (inum << 32) | fid->raw[1];
3342
3343         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3344                         shmem_match, fid->raw);
3345         if (inode) {
3346                 dentry = shmem_find_alias(inode);
3347                 iput(inode);
3348         }
3349
3350         return dentry;
3351 }
3352
3353 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3354                                 struct inode *parent)
3355 {
3356         if (*len < 3) {
3357                 *len = 3;
3358                 return FILEID_INVALID;
3359         }
3360
3361         if (inode_unhashed(inode)) {
3362                 /* Unfortunately insert_inode_hash is not idempotent,
3363                  * so as we hash inodes here rather than at creation
3364                  * time, we need a lock to ensure we only try
3365                  * to do it once
3366                  */
3367                 static DEFINE_SPINLOCK(lock);
3368                 spin_lock(&lock);
3369                 if (inode_unhashed(inode))
3370                         __insert_inode_hash(inode,
3371                                             inode->i_ino + inode->i_generation);
3372                 spin_unlock(&lock);
3373         }
3374
3375         fh[0] = inode->i_generation;
3376         fh[1] = inode->i_ino;
3377         fh[2] = ((__u64)inode->i_ino) >> 32;
3378
3379         *len = 3;
3380         return 1;
3381 }
3382
3383 static const struct export_operations shmem_export_ops = {
3384         .get_parent     = shmem_get_parent,
3385         .encode_fh      = shmem_encode_fh,
3386         .fh_to_dentry   = shmem_fh_to_dentry,
3387 };
3388
3389 enum shmem_param {
3390         Opt_gid,
3391         Opt_huge,
3392         Opt_mode,
3393         Opt_mpol,
3394         Opt_nr_blocks,
3395         Opt_nr_inodes,
3396         Opt_size,
3397         Opt_uid,
3398         Opt_inode32,
3399         Opt_inode64,
3400 };
3401
3402 static const struct constant_table shmem_param_enums_huge[] = {
3403         {"never",       SHMEM_HUGE_NEVER },
3404         {"always",      SHMEM_HUGE_ALWAYS },
3405         {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3406         {"advise",      SHMEM_HUGE_ADVISE },
3407         {}
3408 };
3409
3410 const struct fs_parameter_spec shmem_fs_parameters[] = {
3411         fsparam_u32   ("gid",           Opt_gid),
3412         fsparam_enum  ("huge",          Opt_huge,  shmem_param_enums_huge),
3413         fsparam_u32oct("mode",          Opt_mode),
3414         fsparam_string("mpol",          Opt_mpol),
3415         fsparam_string("nr_blocks",     Opt_nr_blocks),
3416         fsparam_string("nr_inodes",     Opt_nr_inodes),
3417         fsparam_string("size",          Opt_size),
3418         fsparam_u32   ("uid",           Opt_uid),
3419         fsparam_flag  ("inode32",       Opt_inode32),
3420         fsparam_flag  ("inode64",       Opt_inode64),
3421         {}
3422 };
3423
3424 static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3425 {
3426         struct shmem_options *ctx = fc->fs_private;
3427         struct fs_parse_result result;
3428         unsigned long long size;
3429         char *rest;
3430         int opt;
3431
3432         opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3433         if (opt < 0)
3434                 return opt;
3435
3436         switch (opt) {
3437         case Opt_size:
3438                 size = memparse(param->string, &rest);
3439                 if (*rest == '%') {
3440                         size <<= PAGE_SHIFT;
3441                         size *= totalram_pages();
3442                         do_div(size, 100);
3443                         rest++;
3444                 }
3445                 if (*rest)
3446                         goto bad_value;
3447                 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3448                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3449                 break;
3450         case Opt_nr_blocks:
3451                 ctx->blocks = memparse(param->string, &rest);
3452                 if (*rest || ctx->blocks > S64_MAX)
3453                         goto bad_value;
3454                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3455                 break;
3456         case Opt_nr_inodes:
3457                 ctx->inodes = memparse(param->string, &rest);
3458                 if (*rest)
3459                         goto bad_value;
3460                 ctx->seen |= SHMEM_SEEN_INODES;
3461                 break;
3462         case Opt_mode:
3463                 ctx->mode = result.uint_32 & 07777;
3464                 break;
3465         case Opt_uid:
3466                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
3467                 if (!uid_valid(ctx->uid))
3468                         goto bad_value;
3469                 break;
3470         case Opt_gid:
3471                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
3472                 if (!gid_valid(ctx->gid))
3473                         goto bad_value;
3474                 break;
3475         case Opt_huge:
3476                 ctx->huge = result.uint_32;
3477                 if (ctx->huge != SHMEM_HUGE_NEVER &&
3478                     !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3479                       has_transparent_hugepage()))
3480                         goto unsupported_parameter;
3481                 ctx->seen |= SHMEM_SEEN_HUGE;
3482                 break;
3483         case Opt_mpol:
3484                 if (IS_ENABLED(CONFIG_NUMA)) {
3485                         mpol_put(ctx->mpol);
3486                         ctx->mpol = NULL;
3487                         if (mpol_parse_str(param->string, &ctx->mpol))
3488                                 goto bad_value;
3489                         break;
3490                 }
3491                 goto unsupported_parameter;
3492         case Opt_inode32:
3493                 ctx->full_inums = false;
3494                 ctx->seen |= SHMEM_SEEN_INUMS;
3495                 break;
3496         case Opt_inode64:
3497                 if (sizeof(ino_t) < 8) {
3498                         return invalfc(fc,
3499                                        "Cannot use inode64 with <64bit inums in kernel\n");
3500                 }
3501                 ctx->full_inums = true;
3502                 ctx->seen |= SHMEM_SEEN_INUMS;
3503                 break;
3504         }
3505         return 0;
3506
3507 unsupported_parameter:
3508         return invalfc(fc, "Unsupported parameter '%s'", param->key);
3509 bad_value:
3510         return invalfc(fc, "Bad value for '%s'", param->key);
3511 }
3512
3513 static int shmem_parse_options(struct fs_context *fc, void *data)
3514 {
3515         char *options = data;
3516
3517         if (options) {
3518                 int err = security_sb_eat_lsm_opts(options, &fc->security);
3519                 if (err)
3520                         return err;
3521         }
3522
3523         while (options != NULL) {
3524                 char *this_char = options;
3525                 for (;;) {
3526                         /*
3527                          * NUL-terminate this option: unfortunately,
3528                          * mount options form a comma-separated list,
3529                          * but mpol's nodelist may also contain commas.
3530                          */
3531                         options = strchr(options, ',');
3532                         if (options == NULL)
3533                                 break;
3534                         options++;
3535                         if (!isdigit(*options)) {
3536                                 options[-1] = '\0';
3537                                 break;
3538                         }
3539                 }
3540                 if (*this_char) {
3541                         char *value = strchr(this_char, '=');
3542                         size_t len = 0;
3543                         int err;
3544
3545                         if (value) {
3546                                 *value++ = '\0';
3547                                 len = strlen(value);
3548                         }
3549                         err = vfs_parse_fs_string(fc, this_char, value, len);
3550                         if (err < 0)
3551                                 return err;
3552                 }
3553         }
3554         return 0;
3555 }
3556
3557 /*
3558  * Reconfigure a shmem filesystem.
3559  *
3560  * Note that we disallow change from limited->unlimited blocks/inodes while any
3561  * are in use; but we must separately disallow unlimited->limited, because in
3562  * that case we have no record of how much is already in use.
3563  */
3564 static int shmem_reconfigure(struct fs_context *fc)
3565 {
3566         struct shmem_options *ctx = fc->fs_private;
3567         struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3568         unsigned long inodes;
3569         struct mempolicy *mpol = NULL;
3570         const char *err;
3571
3572         raw_spin_lock(&sbinfo->stat_lock);
3573         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3574
3575         if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3576                 if (!sbinfo->max_blocks) {
3577                         err = "Cannot retroactively limit size";
3578                         goto out;
3579                 }
3580                 if (percpu_counter_compare(&sbinfo->used_blocks,
3581                                            ctx->blocks) > 0) {
3582                         err = "Too small a size for current use";
3583                         goto out;
3584                 }
3585         }
3586         if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3587                 if (!sbinfo->max_inodes) {
3588                         err = "Cannot retroactively limit inodes";
3589                         goto out;
3590                 }
3591                 if (ctx->inodes < inodes) {
3592                         err = "Too few inodes for current use";
3593                         goto out;
3594                 }
3595         }
3596
3597         if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3598             sbinfo->next_ino > UINT_MAX) {
3599                 err = "Current inum too high to switch to 32-bit inums";
3600                 goto out;
3601         }
3602
3603         if (ctx->seen & SHMEM_SEEN_HUGE)
3604                 sbinfo->huge = ctx->huge;
3605         if (ctx->seen & SHMEM_SEEN_INUMS)
3606                 sbinfo->full_inums = ctx->full_inums;
3607         if (ctx->seen & SHMEM_SEEN_BLOCKS)
3608                 sbinfo->max_blocks  = ctx->blocks;
3609         if (ctx->seen & SHMEM_SEEN_INODES) {
3610                 sbinfo->max_inodes  = ctx->inodes;
3611                 sbinfo->free_inodes = ctx->inodes - inodes;
3612         }
3613
3614         /*
3615          * Preserve previous mempolicy unless mpol remount option was specified.
3616          */
3617         if (ctx->mpol) {
3618                 mpol = sbinfo->mpol;
3619                 sbinfo->mpol = ctx->mpol;       /* transfers initial ref */
3620                 ctx->mpol = NULL;
3621         }
3622         raw_spin_unlock(&sbinfo->stat_lock);
3623         mpol_put(mpol);
3624         return 0;
3625 out:
3626         raw_spin_unlock(&sbinfo->stat_lock);
3627         return invalfc(fc, "%s", err);
3628 }
3629
3630 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3631 {
3632         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3633
3634         if (sbinfo->max_blocks != shmem_default_max_blocks())
3635                 seq_printf(seq, ",size=%luk",
3636                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3637         if (sbinfo->max_inodes != shmem_default_max_inodes())
3638                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3639         if (sbinfo->mode != (0777 | S_ISVTX))
3640                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3641         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3642                 seq_printf(seq, ",uid=%u",
3643                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3644         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3645                 seq_printf(seq, ",gid=%u",
3646                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3647
3648         /*
3649          * Showing inode{64,32} might be useful even if it's the system default,
3650          * since then people don't have to resort to checking both here and
3651          * /proc/config.gz to confirm 64-bit inums were successfully applied
3652          * (which may not even exist if IKCONFIG_PROC isn't enabled).
3653          *
3654          * We hide it when inode64 isn't the default and we are using 32-bit
3655          * inodes, since that probably just means the feature isn't even under
3656          * consideration.
3657          *
3658          * As such:
3659          *
3660          *                     +-----------------+-----------------+
3661          *                     | TMPFS_INODE64=y | TMPFS_INODE64=n |
3662          *  +------------------+-----------------+-----------------+
3663          *  | full_inums=true  | show            | show            |
3664          *  | full_inums=false | show            | hide            |
3665          *  +------------------+-----------------+-----------------+
3666          *
3667          */
3668         if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3669                 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3670 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3671         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3672         if (sbinfo->huge)
3673                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3674 #endif
3675         shmem_show_mpol(seq, sbinfo->mpol);
3676         return 0;
3677 }
3678
3679 #endif /* CONFIG_TMPFS */
3680
3681 static void shmem_put_super(struct super_block *sb)
3682 {
3683         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3684
3685         free_percpu(sbinfo->ino_batch);
3686         percpu_counter_destroy(&sbinfo->used_blocks);
3687         mpol_put(sbinfo->mpol);
3688         kfree(sbinfo);
3689         sb->s_fs_info = NULL;
3690 }
3691
3692 static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3693 {
3694         struct shmem_options *ctx = fc->fs_private;
3695         struct inode *inode;
3696         struct shmem_sb_info *sbinfo;
3697
3698         /* Round up to L1_CACHE_BYTES to resist false sharing */
3699         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3700                                 L1_CACHE_BYTES), GFP_KERNEL);
3701         if (!sbinfo)
3702                 return -ENOMEM;
3703
3704         sb->s_fs_info = sbinfo;
3705
3706 #ifdef CONFIG_TMPFS
3707         /*
3708          * Per default we only allow half of the physical ram per
3709          * tmpfs instance, limiting inodes to one per page of lowmem;
3710          * but the internal instance is left unlimited.
3711          */
3712         if (!(sb->s_flags & SB_KERNMOUNT)) {
3713                 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3714                         ctx->blocks = shmem_default_max_blocks();
3715                 if (!(ctx->seen & SHMEM_SEEN_INODES))
3716                         ctx->inodes = shmem_default_max_inodes();
3717                 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3718                         ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3719         } else {
3720                 sb->s_flags |= SB_NOUSER;
3721         }
3722         sb->s_export_op = &shmem_export_ops;
3723         sb->s_flags |= SB_NOSEC;
3724 #else
3725         sb->s_flags |= SB_NOUSER;
3726 #endif
3727         sbinfo->max_blocks = ctx->blocks;
3728         sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3729         if (sb->s_flags & SB_KERNMOUNT) {
3730                 sbinfo->ino_batch = alloc_percpu(ino_t);
3731                 if (!sbinfo->ino_batch)
3732                         goto failed;
3733         }
3734         sbinfo->uid = ctx->uid;
3735         sbinfo->gid = ctx->gid;
3736         sbinfo->full_inums = ctx->full_inums;
3737         sbinfo->mode = ctx->mode;
3738         sbinfo->huge = ctx->huge;
3739         sbinfo->mpol = ctx->mpol;
3740         ctx->mpol = NULL;
3741
3742         raw_spin_lock_init(&sbinfo->stat_lock);
3743         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3744                 goto failed;
3745         spin_lock_init(&sbinfo->shrinklist_lock);
3746         INIT_LIST_HEAD(&sbinfo->shrinklist);
3747
3748         sb->s_maxbytes = MAX_LFS_FILESIZE;
3749         sb->s_blocksize = PAGE_SIZE;
3750         sb->s_blocksize_bits = PAGE_SHIFT;
3751         sb->s_magic = TMPFS_MAGIC;
3752         sb->s_op = &shmem_ops;
3753         sb->s_time_gran = 1;
3754 #ifdef CONFIG_TMPFS_XATTR
3755         sb->s_xattr = shmem_xattr_handlers;
3756 #endif
3757 #ifdef CONFIG_TMPFS_POSIX_ACL
3758         sb->s_flags |= SB_POSIXACL;
3759 #endif
3760         uuid_gen(&sb->s_uuid);
3761
3762         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3763         if (!inode)
3764                 goto failed;
3765         inode->i_uid = sbinfo->uid;
3766         inode->i_gid = sbinfo->gid;
3767         sb->s_root = d_make_root(inode);
3768         if (!sb->s_root)
3769                 goto failed;
3770         return 0;
3771
3772 failed:
3773         shmem_put_super(sb);
3774         return -ENOMEM;
3775 }
3776
3777 static int shmem_get_tree(struct fs_context *fc)
3778 {
3779         return get_tree_nodev(fc, shmem_fill_super);
3780 }
3781
3782 static void shmem_free_fc(struct fs_context *fc)
3783 {
3784         struct shmem_options *ctx = fc->fs_private;
3785
3786         if (ctx) {
3787                 mpol_put(ctx->mpol);
3788                 kfree(ctx);
3789         }
3790 }
3791
3792 static const struct fs_context_operations shmem_fs_context_ops = {
3793         .free                   = shmem_free_fc,
3794         .get_tree               = shmem_get_tree,
3795 #ifdef CONFIG_TMPFS
3796         .parse_monolithic       = shmem_parse_options,
3797         .parse_param            = shmem_parse_one,
3798         .reconfigure            = shmem_reconfigure,
3799 #endif
3800 };
3801
3802 static struct kmem_cache *shmem_inode_cachep;
3803
3804 static struct inode *shmem_alloc_inode(struct super_block *sb)
3805 {
3806         struct shmem_inode_info *info;
3807         info = alloc_inode_sb(sb, shmem_inode_cachep, GFP_KERNEL);
3808         if (!info)
3809                 return NULL;
3810         return &info->vfs_inode;
3811 }
3812
3813 static void shmem_free_in_core_inode(struct inode *inode)
3814 {
3815         if (S_ISLNK(inode->i_mode))
3816                 kfree(inode->i_link);
3817         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3818 }
3819
3820 static void shmem_destroy_inode(struct inode *inode)
3821 {
3822         if (S_ISREG(inode->i_mode))
3823                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3824 }
3825
3826 static void shmem_init_inode(void *foo)
3827 {
3828         struct shmem_inode_info *info = foo;
3829         inode_init_once(&info->vfs_inode);
3830 }
3831
3832 static void shmem_init_inodecache(void)
3833 {
3834         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3835                                 sizeof(struct shmem_inode_info),
3836                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3837 }
3838
3839 static void shmem_destroy_inodecache(void)
3840 {
3841         kmem_cache_destroy(shmem_inode_cachep);
3842 }
3843
3844 /* Keep the page in page cache instead of truncating it */
3845 static int shmem_error_remove_page(struct address_space *mapping,
3846                                    struct page *page)
3847 {
3848         return 0;
3849 }
3850
3851 const struct address_space_operations shmem_aops = {
3852         .writepage      = shmem_writepage,
3853         .dirty_folio    = noop_dirty_folio,
3854 #ifdef CONFIG_TMPFS
3855         .write_begin    = shmem_write_begin,
3856         .write_end      = shmem_write_end,
3857 #endif
3858 #ifdef CONFIG_MIGRATION
3859         .migrate_folio  = migrate_folio,
3860 #endif
3861         .error_remove_page = shmem_error_remove_page,
3862 };
3863 EXPORT_SYMBOL(shmem_aops);
3864
3865 static const struct file_operations shmem_file_operations = {
3866         .mmap           = shmem_mmap,
3867         .get_unmapped_area = shmem_get_unmapped_area,
3868 #ifdef CONFIG_TMPFS
3869         .llseek         = shmem_file_llseek,
3870         .read_iter      = shmem_file_read_iter,
3871         .write_iter     = generic_file_write_iter,
3872         .fsync          = noop_fsync,
3873         .splice_read    = generic_file_splice_read,
3874         .splice_write   = iter_file_splice_write,
3875         .fallocate      = shmem_fallocate,
3876 #endif
3877 };
3878
3879 static const struct inode_operations shmem_inode_operations = {
3880         .getattr        = shmem_getattr,
3881         .setattr        = shmem_setattr,
3882 #ifdef CONFIG_TMPFS_XATTR
3883         .listxattr      = shmem_listxattr,
3884         .set_acl        = simple_set_acl,
3885         .fileattr_get   = shmem_fileattr_get,
3886         .fileattr_set   = shmem_fileattr_set,
3887 #endif
3888 };
3889
3890 static const struct inode_operations shmem_dir_inode_operations = {
3891 #ifdef CONFIG_TMPFS
3892         .getattr        = shmem_getattr,
3893         .create         = shmem_create,
3894         .lookup         = simple_lookup,
3895         .link           = shmem_link,
3896         .unlink         = shmem_unlink,
3897         .symlink        = shmem_symlink,
3898         .mkdir          = shmem_mkdir,
3899         .rmdir          = shmem_rmdir,
3900         .mknod          = shmem_mknod,
3901         .rename         = shmem_rename2,
3902         .tmpfile        = shmem_tmpfile,
3903 #endif
3904 #ifdef CONFIG_TMPFS_XATTR
3905         .listxattr      = shmem_listxattr,
3906         .fileattr_get   = shmem_fileattr_get,
3907         .fileattr_set   = shmem_fileattr_set,
3908 #endif
3909 #ifdef CONFIG_TMPFS_POSIX_ACL
3910         .setattr        = shmem_setattr,
3911         .set_acl        = simple_set_acl,
3912 #endif
3913 };
3914
3915 static const struct inode_operations shmem_special_inode_operations = {
3916         .getattr        = shmem_getattr,
3917 #ifdef CONFIG_TMPFS_XATTR
3918         .listxattr      = shmem_listxattr,
3919 #endif
3920 #ifdef CONFIG_TMPFS_POSIX_ACL
3921         .setattr        = shmem_setattr,
3922         .set_acl        = simple_set_acl,
3923 #endif
3924 };
3925
3926 static const struct super_operations shmem_ops = {
3927         .alloc_inode    = shmem_alloc_inode,
3928         .free_inode     = shmem_free_in_core_inode,
3929         .destroy_inode  = shmem_destroy_inode,
3930 #ifdef CONFIG_TMPFS
3931         .statfs         = shmem_statfs,
3932         .show_options   = shmem_show_options,
3933 #endif
3934         .evict_inode    = shmem_evict_inode,
3935         .drop_inode     = generic_delete_inode,
3936         .put_super      = shmem_put_super,
3937 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3938         .nr_cached_objects      = shmem_unused_huge_count,
3939         .free_cached_objects    = shmem_unused_huge_scan,
3940 #endif
3941 };
3942
3943 static const struct vm_operations_struct shmem_vm_ops = {
3944         .fault          = shmem_fault,
3945         .map_pages      = filemap_map_pages,
3946 #ifdef CONFIG_NUMA
3947         .set_policy     = shmem_set_policy,
3948         .get_policy     = shmem_get_policy,
3949 #endif
3950 };
3951
3952 int shmem_init_fs_context(struct fs_context *fc)
3953 {
3954         struct shmem_options *ctx;
3955
3956         ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
3957         if (!ctx)
3958                 return -ENOMEM;
3959
3960         ctx->mode = 0777 | S_ISVTX;
3961         ctx->uid = current_fsuid();
3962         ctx->gid = current_fsgid();
3963
3964         fc->fs_private = ctx;
3965         fc->ops = &shmem_fs_context_ops;
3966         return 0;
3967 }
3968
3969 static struct file_system_type shmem_fs_type = {
3970         .owner          = THIS_MODULE,
3971         .name           = "tmpfs",
3972         .init_fs_context = shmem_init_fs_context,
3973 #ifdef CONFIG_TMPFS
3974         .parameters     = shmem_fs_parameters,
3975 #endif
3976         .kill_sb        = kill_litter_super,
3977         .fs_flags       = FS_USERNS_MOUNT,
3978 };
3979
3980 void __init shmem_init(void)
3981 {
3982         int error;
3983
3984         shmem_init_inodecache();
3985
3986         error = register_filesystem(&shmem_fs_type);
3987         if (error) {
3988                 pr_err("Could not register tmpfs\n");
3989                 goto out2;
3990         }
3991
3992         shm_mnt = kern_mount(&shmem_fs_type);
3993         if (IS_ERR(shm_mnt)) {
3994                 error = PTR_ERR(shm_mnt);
3995                 pr_err("Could not kern_mount tmpfs\n");
3996                 goto out1;
3997         }
3998
3999 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4000         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4001                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4002         else
4003                 shmem_huge = SHMEM_HUGE_NEVER; /* just in case it was patched */
4004 #endif
4005         return;
4006
4007 out1:
4008         unregister_filesystem(&shmem_fs_type);
4009 out2:
4010         shmem_destroy_inodecache();
4011         shm_mnt = ERR_PTR(error);
4012 }
4013
4014 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4015 static ssize_t shmem_enabled_show(struct kobject *kobj,
4016                                   struct kobj_attribute *attr, char *buf)
4017 {
4018         static const int values[] = {
4019                 SHMEM_HUGE_ALWAYS,
4020                 SHMEM_HUGE_WITHIN_SIZE,
4021                 SHMEM_HUGE_ADVISE,
4022                 SHMEM_HUGE_NEVER,
4023                 SHMEM_HUGE_DENY,
4024                 SHMEM_HUGE_FORCE,
4025         };
4026         int len = 0;
4027         int i;
4028
4029         for (i = 0; i < ARRAY_SIZE(values); i++) {
4030                 len += sysfs_emit_at(buf, len,
4031                                      shmem_huge == values[i] ? "%s[%s]" : "%s%s",
4032                                      i ? " " : "",
4033                                      shmem_format_huge(values[i]));
4034         }
4035
4036         len += sysfs_emit_at(buf, len, "\n");
4037
4038         return len;
4039 }
4040
4041 static ssize_t shmem_enabled_store(struct kobject *kobj,
4042                 struct kobj_attribute *attr, const char *buf, size_t count)
4043 {
4044         char tmp[16];
4045         int huge;
4046
4047         if (count + 1 > sizeof(tmp))
4048                 return -EINVAL;
4049         memcpy(tmp, buf, count);
4050         tmp[count] = '\0';
4051         if (count && tmp[count - 1] == '\n')
4052                 tmp[count - 1] = '\0';
4053
4054         huge = shmem_parse_huge(tmp);
4055         if (huge == -EINVAL)
4056                 return -EINVAL;
4057         if (!has_transparent_hugepage() &&
4058                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4059                 return -EINVAL;
4060
4061         shmem_huge = huge;
4062         if (shmem_huge > SHMEM_HUGE_DENY)
4063                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4064         return count;
4065 }
4066
4067 struct kobj_attribute shmem_enabled_attr = __ATTR_RW(shmem_enabled);
4068 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4069
4070 #else /* !CONFIG_SHMEM */
4071
4072 /*
4073  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4074  *
4075  * This is intended for small system where the benefits of the full
4076  * shmem code (swap-backed and resource-limited) are outweighed by
4077  * their complexity. On systems without swap this code should be
4078  * effectively equivalent, but much lighter weight.
4079  */
4080
4081 static struct file_system_type shmem_fs_type = {
4082         .name           = "tmpfs",
4083         .init_fs_context = ramfs_init_fs_context,
4084         .parameters     = ramfs_fs_parameters,
4085         .kill_sb        = kill_litter_super,
4086         .fs_flags       = FS_USERNS_MOUNT,
4087 };
4088
4089 void __init shmem_init(void)
4090 {
4091         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4092
4093         shm_mnt = kern_mount(&shmem_fs_type);
4094         BUG_ON(IS_ERR(shm_mnt));
4095 }
4096
4097 int shmem_unuse(unsigned int type)
4098 {
4099         return 0;
4100 }
4101
4102 int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
4103 {
4104         return 0;
4105 }
4106
4107 void shmem_unlock_mapping(struct address_space *mapping)
4108 {
4109 }
4110
4111 #ifdef CONFIG_MMU
4112 unsigned long shmem_get_unmapped_area(struct file *file,
4113                                       unsigned long addr, unsigned long len,
4114                                       unsigned long pgoff, unsigned long flags)
4115 {
4116         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4117 }
4118 #endif
4119
4120 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4121 {
4122         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4123 }
4124 EXPORT_SYMBOL_GPL(shmem_truncate_range);
4125
4126 #define shmem_vm_ops                            generic_file_vm_ops
4127 #define shmem_file_operations                   ramfs_file_operations
4128 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
4129 #define shmem_acct_size(flags, size)            0
4130 #define shmem_unacct_size(flags, size)          do {} while (0)
4131
4132 #endif /* CONFIG_SHMEM */
4133
4134 /* common code */
4135
4136 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4137                                        unsigned long flags, unsigned int i_flags)
4138 {
4139         struct inode *inode;
4140         struct file *res;
4141
4142         if (IS_ERR(mnt))
4143                 return ERR_CAST(mnt);
4144
4145         if (size < 0 || size > MAX_LFS_FILESIZE)
4146                 return ERR_PTR(-EINVAL);
4147
4148         if (shmem_acct_size(flags, size))
4149                 return ERR_PTR(-ENOMEM);
4150
4151         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4152                                 flags);
4153         if (unlikely(!inode)) {
4154                 shmem_unacct_size(flags, size);
4155                 return ERR_PTR(-ENOSPC);
4156         }
4157         inode->i_flags |= i_flags;
4158         inode->i_size = size;
4159         clear_nlink(inode);     /* It is unlinked */
4160         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4161         if (!IS_ERR(res))
4162                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4163                                 &shmem_file_operations);
4164         if (IS_ERR(res))
4165                 iput(inode);
4166         return res;
4167 }
4168
4169 /**
4170  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4171  *      kernel internal.  There will be NO LSM permission checks against the
4172  *      underlying inode.  So users of this interface must do LSM checks at a
4173  *      higher layer.  The users are the big_key and shm implementations.  LSM
4174  *      checks are provided at the key or shm level rather than the inode.
4175  * @name: name for dentry (to be seen in /proc/<pid>/maps
4176  * @size: size to be set for the file
4177  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4178  */
4179 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4180 {
4181         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4182 }
4183
4184 /**
4185  * shmem_file_setup - get an unlinked file living in tmpfs
4186  * @name: name for dentry (to be seen in /proc/<pid>/maps
4187  * @size: size to be set for the file
4188  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4189  */
4190 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4191 {
4192         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4193 }
4194 EXPORT_SYMBOL_GPL(shmem_file_setup);
4195
4196 /**
4197  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4198  * @mnt: the tmpfs mount where the file will be created
4199  * @name: name for dentry (to be seen in /proc/<pid>/maps
4200  * @size: size to be set for the file
4201  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4202  */
4203 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4204                                        loff_t size, unsigned long flags)
4205 {
4206         return __shmem_file_setup(mnt, name, size, flags, 0);
4207 }
4208 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4209
4210 /**
4211  * shmem_zero_setup - setup a shared anonymous mapping
4212  * @vma: the vma to be mmapped is prepared by do_mmap
4213  */
4214 int shmem_zero_setup(struct vm_area_struct *vma)
4215 {
4216         struct file *file;
4217         loff_t size = vma->vm_end - vma->vm_start;
4218
4219         /*
4220          * Cloning a new file under mmap_lock leads to a lock ordering conflict
4221          * between XFS directory reading and selinux: since this file is only
4222          * accessible to the user through its mapping, use S_PRIVATE flag to
4223          * bypass file security, in the same way as shmem_kernel_file_setup().
4224          */
4225         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4226         if (IS_ERR(file))
4227                 return PTR_ERR(file);
4228
4229         if (vma->vm_file)
4230                 fput(vma->vm_file);
4231         vma->vm_file = file;
4232         vma->vm_ops = &shmem_vm_ops;
4233
4234         return 0;
4235 }
4236
4237 /**
4238  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4239  * @mapping:    the page's address_space
4240  * @index:      the page index
4241  * @gfp:        the page allocator flags to use if allocating
4242  *
4243  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4244  * with any new page allocations done using the specified allocation flags.
4245  * But read_cache_page_gfp() uses the ->read_folio() method: which does not
4246  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4247  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4248  *
4249  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4250  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4251  */
4252 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4253                                          pgoff_t index, gfp_t gfp)
4254 {
4255 #ifdef CONFIG_SHMEM
4256         struct inode *inode = mapping->host;
4257         struct page *page;
4258         int error;
4259
4260         BUG_ON(!shmem_mapping(mapping));
4261         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4262                                   gfp, NULL, NULL, NULL);
4263         if (error)
4264                 return ERR_PTR(error);
4265
4266         unlock_page(page);
4267         if (PageHWPoison(page)) {
4268                 put_page(page);
4269                 return ERR_PTR(-EIO);
4270         }
4271
4272         return page;
4273 #else
4274         /*
4275          * The tiny !SHMEM case uses ramfs without swap
4276          */
4277         return read_cache_page_gfp(mapping, index, gfp);
4278 #endif
4279 }
4280 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);