mm/shmem.c

   1 /*
   2  * Resizable virtual memory filesystem for Linux.
   3  *
   4  * Copyright (C) 2000 Linus Torvalds.
   5  *               2000 Transmeta Corp.
   6  *               2000-2001 Christoph Rohland
   7  *               2000-2001 SAP AG
   8  *               2002 Red Hat Inc.
   9  * Copyright (C) 2002-2011 Hugh Dickins.
  10  * Copyright (C) 2011 Google Inc.
  11  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  12  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  13  *
  14  * Extended attribute support for tmpfs:
  15  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
  16  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
  17  *
  18  * tiny-shmem:
  19  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
  20  *
  21  * This file is released under the GPL.
  22  */
  23
  24 #include <linux/fs.h>
  25 #include <linux/init.h>
  26 #include <linux/vfs.h>
  27 #include <linux/mount.h>
  28 #include <linux/ramfs.h>
  29 #include <linux/pagemap.h>
  30 #include <linux/file.h>
  31 #include <linux/mm.h>
  32 #include <linux/random.h>
  33 #include <linux/sched/signal.h>
  34 #include <linux/export.h>
  35 #include <linux/swap.h>
  36 #include <linux/uio.h>
  37 #include <linux/khugepaged.h>
  38 #include <linux/hugetlb.h>
  39 #include <linux/frontswap.h>
  40 #include <linux/fs_parser.h>
  41
  42 #include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
  43
  44 static struct vfsmount *shm_mnt;
  45
  46 #ifdef CONFIG_SHMEM
  47 /*
  48  * This virtual memory filesystem is heavily based on the ramfs. It
  49  * extends ramfs by the ability to use swap and honor resource limits
  50  * which makes it a completely usable filesystem.
  51  */
  52
  53 #include <linux/xattr.h>
  54 #include <linux/exportfs.h>
  55 #include <linux/posix_acl.h>
  56 #include <linux/posix_acl_xattr.h>
  57 #include <linux/mman.h>
  58 #include <linux/string.h>
  59 #include <linux/slab.h>
  60 #include <linux/backing-dev.h>
  61 #include <linux/shmem_fs.h>
  62 #include <linux/writeback.h>
  63 #include <linux/blkdev.h>
  64 #include <linux/pagevec.h>
  65 #include <linux/percpu_counter.h>
  66 #include <linux/falloc.h>
  67 #include <linux/splice.h>
  68 #include <linux/security.h>
  69 #include <linux/swapops.h>
  70 #include <linux/mempolicy.h>
  71 #include <linux/namei.h>
  72 #include <linux/ctype.h>
  73 #include <linux/migrate.h>
  74 #include <linux/highmem.h>
  75 #include <linux/seq_file.h>
  76 #include <linux/magic.h>
  77 #include <linux/syscalls.h>
  78 #include <linux/fcntl.h>
  79 #include <uapi/linux/memfd.h>
  80 #include <linux/userfaultfd_k.h>
  81 #include <linux/rmap.h>
  82 #include <linux/uuid.h>
  83
  84 #include <linux/uaccess.h>
  85
  86 #include "internal.h"
  87
  88 #define BLOCKS_PER_PAGE  (PAGE_SIZE/512)
  89 #define VM_ACCT(size)    (PAGE_ALIGN(size) >> PAGE_SHIFT)
  90
  91 /* Pretend that each entry is of this size in directory's i_size */
  92 #define BOGO_DIRENT_SIZE 20
  93
  94 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  95 #define SHORT_SYMLINK_LEN 128
  96
  97 /*
  98  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
  99  * inode->i_private (with i_mutex making sure that it has only one user at
 100  * a time): we would prefer not to enlarge the shmem inode just for that.
 101  */
 102 struct shmem_falloc {
 103         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
 104         pgoff_t start;          /* start of range currently being fallocated */
 105         pgoff_t next;           /* the next page offset to be fallocated */
 106         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
 107         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
 108 };
 109
 110 struct shmem_options {
 111         unsigned long long blocks;
 112         unsigned long long inodes;
 113         struct mempolicy *mpol;
 114         kuid_t uid;
 115         kgid_t gid;
 116         umode_t mode;
 117         bool full_inums;
 118         int huge;
 119         int seen;
 120 #define SHMEM_SEEN_BLOCKS 1
 121 #define SHMEM_SEEN_INODES 2
 122 #define SHMEM_SEEN_HUGE 4
 123 #define SHMEM_SEEN_INUMS 8
 124 };
 125
 126 #ifdef CONFIG_TMPFS
 127 static unsigned long shmem_default_max_blocks(void)
 128 {
 129         return totalram_pages() / 2;
 130 }
 131
 132 static unsigned long shmem_default_max_inodes(void)
 133 {
 134         unsigned long nr_pages = totalram_pages();
 135
 136         return min(nr_pages - totalhigh_pages(), nr_pages / 2);
 137 }
 138 #endif
 139
 140 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
 141 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 142                                 struct shmem_inode_info *info, pgoff_t index);
 143 static int shmem_swapin_page(struct inode *inode, pgoff_t index,
 144                              struct page **pagep, enum sgp_type sgp,
 145                              gfp_t gfp, struct vm_area_struct *vma,
 146                              vm_fault_t *fault_type);
 147 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 148                 struct page **pagep, enum sgp_type sgp,
 149                 gfp_t gfp, struct vm_area_struct *vma,
 150                 struct vm_fault *vmf, vm_fault_t *fault_type);
 151
 152 int shmem_getpage(struct inode *inode, pgoff_t index,
 153                 struct page **pagep, enum sgp_type sgp)
 154 {
 155         return shmem_getpage_gfp(inode, index, pagep, sgp,
 156                 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
 157 }
 158
 159 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 160 {
 161         return sb->s_fs_info;
 162 }
 163
 164 /*
 165  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 166  * for shared memory and for shared anonymous (/dev/zero) mappings
 167  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 168  * consistent with the pre-accounting of private mappings ...
 169  */
 170 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 171 {
 172         return (flags & VM_NORESERVE) ?
 173                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 174 }
 175
 176 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 177 {
 178         if (!(flags & VM_NORESERVE))
 179                 vm_unacct_memory(VM_ACCT(size));
 180 }
 181
 182 static inline int shmem_reacct_size(unsigned long flags,
 183                 loff_t oldsize, loff_t newsize)
 184 {
 185         if (!(flags & VM_NORESERVE)) {
 186                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
 187                         return security_vm_enough_memory_mm(current->mm,
 188                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
 189                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
 190                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
 191         }
 192         return 0;
 193 }
 194
 195 /*
 196  * ... whereas tmpfs objects are accounted incrementally as
 197  * pages are allocated, in order to allow large sparse files.
 198  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 199  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 200  */
 201 static inline int shmem_acct_block(unsigned long flags, long pages)
 202 {
 203         if (!(flags & VM_NORESERVE))
 204                 return 0;
 205
 206         return security_vm_enough_memory_mm(current->mm,
 207                         pages * VM_ACCT(PAGE_SIZE));
 208 }
 209
 210 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 211 {
 212         if (flags & VM_NORESERVE)
 213                 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
 214 }
 215
 216 static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
 217 {
 218         struct shmem_inode_info *info = SHMEM_I(inode);
 219         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 220
 221         if (shmem_acct_block(info->flags, pages))
 222                 return false;
 223
 224         if (sbinfo->max_blocks) {
 225                 if (percpu_counter_compare(&sbinfo->used_blocks,
 226                                            sbinfo->max_blocks - pages) > 0)
 227                         goto unacct;
 228                 percpu_counter_add(&sbinfo->used_blocks, pages);
 229         }
 230
 231         return true;
 232
 233 unacct:
 234         shmem_unacct_blocks(info->flags, pages);
 235         return false;
 236 }
 237
 238 static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
 239 {
 240         struct shmem_inode_info *info = SHMEM_I(inode);
 241         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 242
 243         if (sbinfo->max_blocks)
 244                 percpu_counter_sub(&sbinfo->used_blocks, pages);
 245         shmem_unacct_blocks(info->flags, pages);
 246 }
 247
 248 static const struct super_operations shmem_ops;
 249 static const struct address_space_operations shmem_aops;
 250 static const struct file_operations shmem_file_operations;
 251 static const struct inode_operations shmem_inode_operations;
 252 static const struct inode_operations shmem_dir_inode_operations;
 253 static const struct inode_operations shmem_special_inode_operations;
 254 static const struct vm_operations_struct shmem_vm_ops;
 255 static struct file_system_type shmem_fs_type;
 256
 257 bool vma_is_shmem(struct vm_area_struct *vma)
 258 {
 259         return vma->vm_ops == &shmem_vm_ops;
 260 }
 261
 262 static LIST_HEAD(shmem_swaplist);
 263 static DEFINE_MUTEX(shmem_swaplist_mutex);
 264
 265 /*
 266  * shmem_reserve_inode() performs bookkeeping to reserve a shmem inode, and
 267  * produces a novel ino for the newly allocated inode.
 268  *
 269  * It may also be called when making a hard link to permit the space needed by
 270  * each dentry. However, in that case, no new inode number is needed since that
 271  * internally draws from another pool of inode numbers (currently global
 272  * get_next_ino()). This case is indicated by passing NULL as inop.
 273  */
 274 #define SHMEM_INO_BATCH 1024
 275 static int shmem_reserve_inode(struct super_block *sb, ino_t *inop)
 276 {
 277         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 278         ino_t ino;
 279
 280         if (!(sb->s_flags & SB_KERNMOUNT)) {
 281                 spin_lock(&sbinfo->stat_lock);
 282                 if (sbinfo->max_inodes) {
 283                         if (!sbinfo->free_inodes) {
 284                                 spin_unlock(&sbinfo->stat_lock);
 285                                 return -ENOSPC;
 286                         }
 287                         sbinfo->free_inodes--;
 288                 }
 289                 if (inop) {
 290                         ino = sbinfo->next_ino++;
 291                         if (unlikely(is_zero_ino(ino)))
 292                                 ino = sbinfo->next_ino++;
 293                         if (unlikely(!sbinfo->full_inums &&
 294                                      ino > UINT_MAX)) {
 295                                 /*
 296                                  * Emulate get_next_ino uint wraparound for
 297                                  * compatibility
 298                                  */
 299                                 if (IS_ENABLED(CONFIG_64BIT))
 300                                         pr_warn("%s: inode number overflow on device %d, consider using inode64 mount option\n",
 301                                                 __func__, MINOR(sb->s_dev));
 302                                 sbinfo->next_ino = 1;
 303                                 ino = sbinfo->next_ino++;
 304                         }
 305                         *inop = ino;
 306                 }
 307                 spin_unlock(&sbinfo->stat_lock);
 308         } else if (inop) {
 309                 /*
 310                  * __shmem_file_setup, one of our callers, is lock-free: it
 311                  * doesn't hold stat_lock in shmem_reserve_inode since
 312                  * max_inodes is always 0, and is called from potentially
 313                  * unknown contexts. As such, use a per-cpu batched allocator
 314                  * which doesn't require the per-sb stat_lock unless we are at
 315                  * the batch boundary.
 316                  *
 317                  * We don't need to worry about inode{32,64} since SB_KERNMOUNT
 318                  * shmem mounts are not exposed to userspace, so we don't need
 319                  * to worry about things like glibc compatibility.
 320                  */
 321                 ino_t *next_ino;
 322                 next_ino = per_cpu_ptr(sbinfo->ino_batch, get_cpu());
 323                 ino = *next_ino;
 324                 if (unlikely(ino % SHMEM_INO_BATCH == 0)) {
 325                         spin_lock(&sbinfo->stat_lock);
 326                         ino = sbinfo->next_ino;
 327                         sbinfo->next_ino += SHMEM_INO_BATCH;
 328                         spin_unlock(&sbinfo->stat_lock);
 329                         if (unlikely(is_zero_ino(ino)))
 330                                 ino++;
 331                 }
 332                 *inop = ino;
 333                 *next_ino = ++ino;
 334                 put_cpu();
 335         }
 336
 337         return 0;
 338 }
 339
 340 static void shmem_free_inode(struct super_block *sb)
 341 {
 342         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 343         if (sbinfo->max_inodes) {
 344                 spin_lock(&sbinfo->stat_lock);
 345                 sbinfo->free_inodes++;
 346                 spin_unlock(&sbinfo->stat_lock);
 347         }
 348 }
 349
 350 /**
 351  * shmem_recalc_inode - recalculate the block usage of an inode
 352  * @inode: inode to recalc
 353  *
 354  * We have to calculate the free blocks since the mm can drop
 355  * undirtied hole pages behind our back.
 356  *
 357  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 358  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 359  *
 360  * It has to be called with the spinlock held.
 361  */
 362 static void shmem_recalc_inode(struct inode *inode)
 363 {
 364         struct shmem_inode_info *info = SHMEM_I(inode);
 365         long freed;
 366
 367         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 368         if (freed > 0) {
 369                 info->alloced -= freed;
 370                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 371                 shmem_inode_unacct_blocks(inode, freed);
 372         }
 373 }
 374
 375 bool shmem_charge(struct inode *inode, long pages)
 376 {
 377         struct shmem_inode_info *info = SHMEM_I(inode);
 378         unsigned long flags;
 379
 380         if (!shmem_inode_acct_block(inode, pages))
 381                 return false;
 382
 383         /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
 384         inode->i_mapping->nrpages += pages;
 385
 386         spin_lock_irqsave(&info->lock, flags);
 387         info->alloced += pages;
 388         inode->i_blocks += pages * BLOCKS_PER_PAGE;
 389         shmem_recalc_inode(inode);
 390         spin_unlock_irqrestore(&info->lock, flags);
 391
 392         return true;
 393 }
 394
 395 void shmem_uncharge(struct inode *inode, long pages)
 396 {
 397         struct shmem_inode_info *info = SHMEM_I(inode);
 398         unsigned long flags;
 399
 400         /* nrpages adjustment done by __delete_from_page_cache() or caller */
 401
 402         spin_lock_irqsave(&info->lock, flags);
 403         info->alloced -= pages;
 404         inode->i_blocks -= pages * BLOCKS_PER_PAGE;
 405         shmem_recalc_inode(inode);
 406         spin_unlock_irqrestore(&info->lock, flags);
 407
 408         shmem_inode_unacct_blocks(inode, pages);
 409 }
 410
 411 /*
 412  * Replace item expected in xarray by a new item, while holding xa_lock.
 413  */
 414 static int shmem_replace_entry(struct address_space *mapping,
 415                         pgoff_t index, void *expected, void *replacement)
 416 {
 417         XA_STATE(xas, &mapping->i_pages, index);
 418         void *item;
 419
 420         VM_BUG_ON(!expected);
 421         VM_BUG_ON(!replacement);
 422         item = xas_load(&xas);
 423         if (item != expected)
 424                 return -ENOENT;
 425         xas_store(&xas, replacement);
 426         return 0;
 427 }
 428
 429 /*
 430  * Sometimes, before we decide whether to proceed or to fail, we must check
 431  * that an entry was not already brought back from swap by a racing thread.
 432  *
 433  * Checking page is not enough: by the time a SwapCache page is locked, it
 434  * might be reused, and again be SwapCache, using the same swap as before.
 435  */
 436 static bool shmem_confirm_swap(struct address_space *mapping,
 437                                pgoff_t index, swp_entry_t swap)
 438 {
 439         return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
 440 }
 441
 442 /*
 443  * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
 444  *
 445  * SHMEM_HUGE_NEVER:
 446  *      disables huge pages for the mount;
 447  * SHMEM_HUGE_ALWAYS:
 448  *      enables huge pages for the mount;
 449  * SHMEM_HUGE_WITHIN_SIZE:
 450  *      only allocate huge pages if the page will be fully within i_size,
 451  *      also respect fadvise()/madvise() hints;
 452  * SHMEM_HUGE_ADVISE:
 453  *      only allocate huge pages if requested with fadvise()/madvise();
 454  */
 455
 456 #define SHMEM_HUGE_NEVER        0
 457 #define SHMEM_HUGE_ALWAYS       1
 458 #define SHMEM_HUGE_WITHIN_SIZE  2
 459 #define SHMEM_HUGE_ADVISE       3
 460
 461 /*
 462  * Special values.
 463  * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
 464  *
 465  * SHMEM_HUGE_DENY:
 466  *      disables huge on shm_mnt and all mounts, for emergency use;
 467  * SHMEM_HUGE_FORCE:
 468  *      enables huge on shm_mnt and all mounts, w/o needing option, for testing;
 469  *
 470  */
 471 #define SHMEM_HUGE_DENY         (-1)
 472 #define SHMEM_HUGE_FORCE        (-2)
 473
 474 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 475 /* ifdef here to avoid bloating shmem.o when not necessary */
 476
 477 static int shmem_huge __read_mostly;
 478
 479 #if defined(CONFIG_SYSFS)
 480 static int shmem_parse_huge(const char *str)
 481 {
 482         if (!strcmp(str, "never"))
 483                 return SHMEM_HUGE_NEVER;
 484         if (!strcmp(str, "always"))
 485                 return SHMEM_HUGE_ALWAYS;
 486         if (!strcmp(str, "within_size"))
 487                 return SHMEM_HUGE_WITHIN_SIZE;
 488         if (!strcmp(str, "advise"))
 489                 return SHMEM_HUGE_ADVISE;
 490         if (!strcmp(str, "deny"))
 491                 return SHMEM_HUGE_DENY;
 492         if (!strcmp(str, "force"))
 493                 return SHMEM_HUGE_FORCE;
 494         return -EINVAL;
 495 }
 496 #endif
 497
 498 #if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
 499 static const char *shmem_format_huge(int huge)
 500 {
 501         switch (huge) {
 502         case SHMEM_HUGE_NEVER:
 503                 return "never";
 504         case SHMEM_HUGE_ALWAYS:
 505                 return "always";
 506         case SHMEM_HUGE_WITHIN_SIZE:
 507                 return "within_size";
 508         case SHMEM_HUGE_ADVISE:
 509                 return "advise";
 510         case SHMEM_HUGE_DENY:
 511                 return "deny";
 512         case SHMEM_HUGE_FORCE:
 513                 return "force";
 514         default:
 515                 VM_BUG_ON(1);
 516                 return "bad_val";
 517         }
 518 }
 519 #endif
 520
 521 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 522                 struct shrink_control *sc, unsigned long nr_to_split)
 523 {
 524         LIST_HEAD(list), *pos, *next;
 525         LIST_HEAD(to_remove);
 526         struct inode *inode;
 527         struct shmem_inode_info *info;
 528         struct page *page;
 529         unsigned long batch = sc ? sc->nr_to_scan : 128;
 530         int split = 0;
 531
 532         if (list_empty(&sbinfo->shrinklist))
 533                 return SHRINK_STOP;
 534
 535         spin_lock(&sbinfo->shrinklist_lock);
 536         list_for_each_safe(pos, next, &sbinfo->shrinklist) {
 537                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 538
 539                 /* pin the inode */
 540                 inode = igrab(&info->vfs_inode);
 541
 542                 /* inode is about to be evicted */
 543                 if (!inode) {
 544                         list_del_init(&info->shrinklist);
 545                         goto next;
 546                 }
 547
 548                 /* Check if there's anything to gain */
 549                 if (round_up(inode->i_size, PAGE_SIZE) ==
 550                                 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
 551                         list_move(&info->shrinklist, &to_remove);
 552                         goto next;
 553                 }
 554
 555                 list_move(&info->shrinklist, &list);
 556 next:
 557                 sbinfo->shrinklist_len--;
 558                 if (!--batch)
 559                         break;
 560         }
 561         spin_unlock(&sbinfo->shrinklist_lock);
 562
 563         list_for_each_safe(pos, next, &to_remove) {
 564                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 565                 inode = &info->vfs_inode;
 566                 list_del_init(&info->shrinklist);
 567                 iput(inode);
 568         }
 569
 570         list_for_each_safe(pos, next, &list) {
 571                 int ret;
 572
 573                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
 574                 inode = &info->vfs_inode;
 575
 576                 if (nr_to_split && split >= nr_to_split)
 577                         goto move_back;
 578
 579                 page = find_get_page(inode->i_mapping,
 580                                 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
 581                 if (!page)
 582                         goto drop;
 583
 584                 /* No huge page at the end of the file: nothing to split */
 585                 if (!PageTransHuge(page)) {
 586                         put_page(page);
 587                         goto drop;
 588                 }
 589
 590                 /*
 591                  * Move the inode on the list back to shrinklist if we failed
 592                  * to lock the page at this time.
 593                  *
 594                  * Waiting for the lock may lead to deadlock in the
 595                  * reclaim path.
 596                  */
 597                 if (!trylock_page(page)) {
 598                         put_page(page);
 599                         goto move_back;
 600                 }
 601
 602                 ret = split_huge_page(page);
 603                 unlock_page(page);
 604                 put_page(page);
 605
 606                 /* If split failed move the inode on the list back to shrinklist */
 607                 if (ret)
 608                         goto move_back;
 609
 610                 split++;
 611 drop:
 612                 list_del_init(&info->shrinklist);
 613                 goto put;
 614 move_back:
 615                 /*
 616                  * Make sure the inode is either on the global list or deleted
 617                  * from any local list before iput() since it could be deleted
 618                  * in another thread once we put the inode (then the local list
 619                  * is corrupted).
 620                  */
 621                 spin_lock(&sbinfo->shrinklist_lock);
 622                 list_move(&info->shrinklist, &sbinfo->shrinklist);
 623                 sbinfo->shrinklist_len++;
 624                 spin_unlock(&sbinfo->shrinklist_lock);
 625 put:
 626                 iput(inode);
 627         }
 628
 629         return split;
 630 }
 631
 632 static long shmem_unused_huge_scan(struct super_block *sb,
 633                 struct shrink_control *sc)
 634 {
 635         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 636
 637         if (!READ_ONCE(sbinfo->shrinklist_len))
 638                 return SHRINK_STOP;
 639
 640         return shmem_unused_huge_shrink(sbinfo, sc, 0);
 641 }
 642
 643 static long shmem_unused_huge_count(struct super_block *sb,
 644                 struct shrink_control *sc)
 645 {
 646         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 647         return READ_ONCE(sbinfo->shrinklist_len);
 648 }
 649 #else /* !CONFIG_TRANSPARENT_HUGEPAGE */
 650
 651 #define shmem_huge SHMEM_HUGE_DENY
 652
 653 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 654                 struct shrink_control *sc, unsigned long nr_to_split)
 655 {
 656         return 0;
 657 }
 658 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 659
 660 static inline bool is_huge_enabled(struct shmem_sb_info *sbinfo)
 661 {
 662         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 663             (shmem_huge == SHMEM_HUGE_FORCE || sbinfo->huge) &&
 664             shmem_huge != SHMEM_HUGE_DENY)
 665                 return true;
 666         return false;
 667 }
 668
 669 /*
 670  * Like add_to_page_cache_locked, but error if expected item has gone.
 671  */
 672 static int shmem_add_to_page_cache(struct page *page,
 673                                    struct address_space *mapping,
 674                                    pgoff_t index, void *expected, gfp_t gfp,
 675                                    struct mm_struct *charge_mm)
 676 {
 677         XA_STATE_ORDER(xas, &mapping->i_pages, index, compound_order(page));
 678         unsigned long i = 0;
 679         unsigned long nr = compound_nr(page);
 680         int error;
 681
 682         VM_BUG_ON_PAGE(PageTail(page), page);
 683         VM_BUG_ON_PAGE(index != round_down(index, nr), page);
 684         VM_BUG_ON_PAGE(!PageLocked(page), page);
 685         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 686         VM_BUG_ON(expected && PageTransHuge(page));
 687
 688         page_ref_add(page, nr);
 689         page->mapping = mapping;
 690         page->index = index;
 691
 692         if (!PageSwapCache(page)) {
 693                 error = mem_cgroup_charge(page, charge_mm, gfp);
 694                 if (error) {
 695                         if (PageTransHuge(page)) {
 696                                 count_vm_event(THP_FILE_FALLBACK);
 697                                 count_vm_event(THP_FILE_FALLBACK_CHARGE);
 698                         }
 699                         goto error;
 700                 }
 701         }
 702         cgroup_throttle_swaprate(page, gfp);
 703
 704         do {
 705                 void *entry;
 706                 xas_lock_irq(&xas);
 707                 entry = xas_find_conflict(&xas);
 708                 if (entry != expected)
 709                         xas_set_err(&xas, -EEXIST);
 710                 xas_create_range(&xas);
 711                 if (xas_error(&xas))
 712                         goto unlock;
 713 next:
 714                 xas_store(&xas, page);
 715                 if (++i < nr) {
 716                         xas_next(&xas);
 717                         goto next;
 718                 }
 719                 if (PageTransHuge(page)) {
 720                         count_vm_event(THP_FILE_ALLOC);
 721                         __inc_node_page_state(page, NR_SHMEM_THPS);
 722                 }
 723                 mapping->nrpages += nr;
 724                 __mod_lruvec_page_state(page, NR_FILE_PAGES, nr);
 725                 __mod_lruvec_page_state(page, NR_SHMEM, nr);
 726 unlock:
 727                 xas_unlock_irq(&xas);
 728         } while (xas_nomem(&xas, gfp));
 729
 730         if (xas_error(&xas)) {
 731                 error = xas_error(&xas);
 732                 goto error;
 733         }
 734
 735         return 0;
 736 error:
 737         page->mapping = NULL;
 738         page_ref_sub(page, nr);
 739         return error;
 740 }
 741
 742 /*
 743  * Like delete_from_page_cache, but substitutes swap for page.
 744  */
 745 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 746 {
 747         struct address_space *mapping = page->mapping;
 748         int error;
 749
 750         VM_BUG_ON_PAGE(PageCompound(page), page);
 751
 752         xa_lock_irq(&mapping->i_pages);
 753         error = shmem_replace_entry(mapping, page->index, page, radswap);
 754         page->mapping = NULL;
 755         mapping->nrpages--;
 756         __dec_lruvec_page_state(page, NR_FILE_PAGES);
 757         __dec_lruvec_page_state(page, NR_SHMEM);
 758         xa_unlock_irq(&mapping->i_pages);
 759         put_page(page);
 760         BUG_ON(error);
 761 }
 762
 763 /*
 764  * Remove swap entry from page cache, free the swap and its page cache.
 765  */
 766 static int shmem_free_swap(struct address_space *mapping,
 767                            pgoff_t index, void *radswap)
 768 {
 769         void *old;
 770
 771         old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
 772         if (old != radswap)
 773                 return -ENOENT;
 774         free_swap_and_cache(radix_to_swp_entry(radswap));
 775         return 0;
 776 }
 777
 778 /*
 779  * Determine (in bytes) how many of the shmem object's pages mapped by the
 780  * given offsets are swapped out.
 781  *
 782  * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
 783  * as long as the inode doesn't go away and racy results are not a problem.
 784  */
 785 unsigned long shmem_partial_swap_usage(struct address_space *mapping,
 786                                                 pgoff_t start, pgoff_t end)
 787 {
 788         XA_STATE(xas, &mapping->i_pages, start);
 789         struct page *page;
 790         unsigned long swapped = 0;
 791
 792         rcu_read_lock();
 793         xas_for_each(&xas, page, end - 1) {
 794                 if (xas_retry(&xas, page))
 795                         continue;
 796                 if (xa_is_value(page))
 797                         swapped++;
 798
 799                 if (need_resched()) {
 800                         xas_pause(&xas);
 801                         cond_resched_rcu();
 802                 }
 803         }
 804
 805         rcu_read_unlock();
 806
 807         return swapped << PAGE_SHIFT;
 808 }
 809
 810 /*
 811  * Determine (in bytes) how many of the shmem object's pages mapped by the
 812  * given vma is swapped out.
 813  *
 814  * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
 815  * as long as the inode doesn't go away and racy results are not a problem.
 816  */
 817 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
 818 {
 819         struct inode *inode = file_inode(vma->vm_file);
 820         struct shmem_inode_info *info = SHMEM_I(inode);
 821         struct address_space *mapping = inode->i_mapping;
 822         unsigned long swapped;
 823
 824         /* Be careful as we don't hold info->lock */
 825         swapped = READ_ONCE(info->swapped);
 826
 827         /*
 828          * The easier cases are when the shmem object has nothing in swap, or
 829          * the vma maps it whole. Then we can simply use the stats that we
 830          * already track.
 831          */
 832         if (!swapped)
 833                 return 0;
 834
 835         if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
 836                 return swapped << PAGE_SHIFT;
 837
 838         /* Here comes the more involved part */
 839         return shmem_partial_swap_usage(mapping,
 840                         linear_page_index(vma, vma->vm_start),
 841                         linear_page_index(vma, vma->vm_end));
 842 }
 843
 844 /*
 845  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 846  */
 847 void shmem_unlock_mapping(struct address_space *mapping)
 848 {
 849         struct pagevec pvec;
 850         pgoff_t indices[PAGEVEC_SIZE];
 851         pgoff_t index = 0;
 852
 853         pagevec_init(&pvec);
 854         /*
 855          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 856          */
 857         while (!mapping_unevictable(mapping)) {
 858                 /*
 859                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
 860                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
 861                  */
 862                 pvec.nr = find_get_entries(mapping, index,
 863                                            PAGEVEC_SIZE, pvec.pages, indices);
 864                 if (!pvec.nr)
 865                         break;
 866                 index = indices[pvec.nr - 1] + 1;
 867                 pagevec_remove_exceptionals(&pvec);
 868                 check_move_unevictable_pages(&pvec);
 869                 pagevec_release(&pvec);
 870                 cond_resched();
 871         }
 872 }
 873
 874 /*
 875  * Check whether a hole-punch or truncation needs to split a huge page,
 876  * returning true if no split was required, or the split has been successful.
 877  *
 878  * Eviction (or truncation to 0 size) should never need to split a huge page;
 879  * but in rare cases might do so, if shmem_undo_range() failed to trylock on
 880  * head, and then succeeded to trylock on tail.
 881  *
 882  * A split can only succeed when there are no additional references on the
 883  * huge page: so the split below relies upon find_get_entries() having stopped
 884  * when it found a subpage of the huge page, without getting further references.
 885  */
 886 static bool shmem_punch_compound(struct page *page, pgoff_t start, pgoff_t end)
 887 {
 888         if (!PageTransCompound(page))
 889                 return true;
 890
 891         /* Just proceed to delete a huge page wholly within the range punched */
 892         if (PageHead(page) &&
 893             page->index >= start && page->index + HPAGE_PMD_NR <= end)
 894                 return true;
 895
 896         /* Try to split huge page, so we can truly punch the hole or truncate */
 897         return split_huge_page(page) >= 0;
 898 }
 899
 900 /*
 901  * Remove range of pages and swap entries from page cache, and free them.
 902  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 903  */
 904 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 905                                                                  bool unfalloc)
 906 {
 907         struct address_space *mapping = inode->i_mapping;
 908         struct shmem_inode_info *info = SHMEM_I(inode);
 909         pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
 910         pgoff_t end = (lend + 1) >> PAGE_SHIFT;
 911         unsigned int partial_start = lstart & (PAGE_SIZE - 1);
 912         unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
 913         struct pagevec pvec;
 914         pgoff_t indices[PAGEVEC_SIZE];
 915         long nr_swaps_freed = 0;
 916         pgoff_t index;
 917         int i;
 918
 919         if (lend == -1)
 920                 end = -1;       /* unsigned, so actually very big */
 921
 922         pagevec_init(&pvec);
 923         index = start;
 924         while (index < end) {
 925                 pvec.nr = find_get_entries(mapping, index,
 926                         min(end - index, (pgoff_t)PAGEVEC_SIZE),
 927                         pvec.pages, indices);
 928                 if (!pvec.nr)
 929                         break;
 930                 for (i = 0; i < pagevec_count(&pvec); i++) {
 931                         struct page *page = pvec.pages[i];
 932
 933                         index = indices[i];
 934                         if (index >= end)
 935                                 break;
 936
 937                         if (xa_is_value(page)) {
 938                                 if (unfalloc)
 939                                         continue;
 940                                 nr_swaps_freed += !shmem_free_swap(mapping,
 941                                                                 index, page);
 942                                 continue;
 943                         }
 944
 945                         VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
 946
 947                         if (!trylock_page(page))
 948                                 continue;
 949
 950                         if ((!unfalloc || !PageUptodate(page)) &&
 951                             page_mapping(page) == mapping) {
 952                                 VM_BUG_ON_PAGE(PageWriteback(page), page);
 953                                 if (shmem_punch_compound(page, start, end))
 954                                         truncate_inode_page(mapping, page);
 955                         }
 956                         unlock_page(page);
 957                 }
 958                 pagevec_remove_exceptionals(&pvec);
 959                 pagevec_release(&pvec);
 960                 cond_resched();
 961                 index++;
 962         }
 963
 964         if (partial_start) {
 965                 struct page *page = NULL;
 966                 shmem_getpage(inode, start - 1, &page, SGP_READ);
 967                 if (page) {
 968                         unsigned int top = PAGE_SIZE;
 969                         if (start > end) {
 970                                 top = partial_end;
 971                                 partial_end = 0;
 972                         }
 973                         zero_user_segment(page, partial_start, top);
 974                         set_page_dirty(page);
 975                         unlock_page(page);
 976                         put_page(page);
 977                 }
 978         }
 979         if (partial_end) {
 980                 struct page *page = NULL;
 981                 shmem_getpage(inode, end, &page, SGP_READ);
 982                 if (page) {
 983                         zero_user_segment(page, 0, partial_end);
 984                         set_page_dirty(page);
 985                         unlock_page(page);
 986                         put_page(page);
 987                 }
 988         }
 989         if (start >= end)
 990                 return;
 991
 992         index = start;
 993         while (index < end) {
 994                 cond_resched();
 995
 996                 pvec.nr = find_get_entries(mapping, index,
 997                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 998                                 pvec.pages, indices);
 999                 if (!pvec.nr) {
1000                         /* If all gone or hole-punch or unfalloc, we're done */
1001                         if (index == start || end != -1)
1002                                 break;
1003                         /* But if truncating, restart to make sure all gone */
1004                         index = start;
1005                         continue;
1006                 }
1007                 for (i = 0; i < pagevec_count(&pvec); i++) {
1008                         struct page *page = pvec.pages[i];
1009
1010                         index = indices[i];
1011                         if (index >= end)
1012                                 break;
1013
1014                         if (xa_is_value(page)) {
1015                                 if (unfalloc)
1016                                         continue;
1017                                 if (shmem_free_swap(mapping, index, page)) {
1018                                         /* Swap was replaced by page: retry */
1019                                         index--;
1020                                         break;
1021                                 }
1022                                 nr_swaps_freed++;
1023                                 continue;
1024                         }
1025
1026                         lock_page(page);
1027
1028                         if (!unfalloc || !PageUptodate(page)) {
1029                                 if (page_mapping(page) != mapping) {
1030                                         /* Page was replaced by swap: retry */
1031                                         unlock_page(page);
1032                                         index--;
1033                                         break;
1034                                 }
1035                                 VM_BUG_ON_PAGE(PageWriteback(page), page);
1036                                 if (shmem_punch_compound(page, start, end))
1037                                         truncate_inode_page(mapping, page);
1038                                 else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1039                                         /* Wipe the page and don't get stuck */
1040                                         clear_highpage(page);
1041                                         flush_dcache_page(page);
1042                                         set_page_dirty(page);
1043                                         if (index <
1044                                             round_up(start, HPAGE_PMD_NR))
1045                                                 start = index + 1;
1046                                 }
1047                         }
1048                         unlock_page(page);
1049                 }
1050                 pagevec_remove_exceptionals(&pvec);
1051                 pagevec_release(&pvec);
1052                 index++;
1053         }
1054
1055         spin_lock_irq(&info->lock);
1056         info->swapped -= nr_swaps_freed;
1057         shmem_recalc_inode(inode);
1058         spin_unlock_irq(&info->lock);
1059 }
1060
1061 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1062 {
1063         shmem_undo_range(inode, lstart, lend, false);
1064         inode->i_ctime = inode->i_mtime = current_time(inode);
1065 }
1066 EXPORT_SYMBOL_GPL(shmem_truncate_range);
1067
1068 static int shmem_getattr(const struct path *path, struct kstat *stat,
1069                          u32 request_mask, unsigned int query_flags)
1070 {
1071         struct inode *inode = path->dentry->d_inode;
1072         struct shmem_inode_info *info = SHMEM_I(inode);
1073         struct shmem_sb_info *sb_info = SHMEM_SB(inode->i_sb);
1074
1075         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1076                 spin_lock_irq(&info->lock);
1077                 shmem_recalc_inode(inode);
1078                 spin_unlock_irq(&info->lock);
1079         }
1080         generic_fillattr(inode, stat);
1081
1082         if (is_huge_enabled(sb_info))
1083                 stat->blksize = HPAGE_PMD_SIZE;
1084
1085         return 0;
1086 }
1087
1088 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1089 {
1090         struct inode *inode = d_inode(dentry);
1091         struct shmem_inode_info *info = SHMEM_I(inode);
1092         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1093         int error;
1094
1095         error = setattr_prepare(dentry, attr);
1096         if (error)
1097                 return error;
1098
1099         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1100                 loff_t oldsize = inode->i_size;
1101                 loff_t newsize = attr->ia_size;
1102
1103                 /* protected by i_mutex */
1104                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1105                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1106                         return -EPERM;
1107
1108                 if (newsize != oldsize) {
1109                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1110                                         oldsize, newsize);
1111                         if (error)
1112                                 return error;
1113                         i_size_write(inode, newsize);
1114                         inode->i_ctime = inode->i_mtime = current_time(inode);
1115                 }
1116                 if (newsize <= oldsize) {
1117                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1118                         if (oldsize > holebegin)
1119                                 unmap_mapping_range(inode->i_mapping,
1120                                                         holebegin, 0, 1);
1121                         if (info->alloced)
1122                                 shmem_truncate_range(inode,
1123                                                         newsize, (loff_t)-1);
1124                         /* unmap again to remove racily COWed private pages */
1125                         if (oldsize > holebegin)
1126                                 unmap_mapping_range(inode->i_mapping,
1127                                                         holebegin, 0, 1);
1128
1129                         /*
1130                          * Part of the huge page can be beyond i_size: subject
1131                          * to shrink under memory pressure.
1132                          */
1133                         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1134                                 spin_lock(&sbinfo->shrinklist_lock);
1135                                 /*
1136                                  * _careful to defend against unlocked access to
1137                                  * ->shrink_list in shmem_unused_huge_shrink()
1138                                  */
1139                                 if (list_empty_careful(&info->shrinklist)) {
1140                                         list_add_tail(&info->shrinklist,
1141                                                         &sbinfo->shrinklist);
1142                                         sbinfo->shrinklist_len++;
1143                                 }
1144                                 spin_unlock(&sbinfo->shrinklist_lock);
1145                         }
1146                 }
1147         }
1148
1149         setattr_copy(inode, attr);
1150         if (attr->ia_valid & ATTR_MODE)
1151                 error = posix_acl_chmod(inode, inode->i_mode);
1152         return error;
1153 }
1154
1155 static void shmem_evict_inode(struct inode *inode)
1156 {
1157         struct shmem_inode_info *info = SHMEM_I(inode);
1158         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1159
1160         if (inode->i_mapping->a_ops == &shmem_aops) {
1161                 shmem_unacct_size(info->flags, inode->i_size);
1162                 inode->i_size = 0;
1163                 shmem_truncate_range(inode, 0, (loff_t)-1);
1164                 if (!list_empty(&info->shrinklist)) {
1165                         spin_lock(&sbinfo->shrinklist_lock);
1166                         if (!list_empty(&info->shrinklist)) {
1167                                 list_del_init(&info->shrinklist);
1168                                 sbinfo->shrinklist_len--;
1169                         }
1170                         spin_unlock(&sbinfo->shrinklist_lock);
1171                 }
1172                 while (!list_empty(&info->swaplist)) {
1173                         /* Wait while shmem_unuse() is scanning this inode... */
1174                         wait_var_event(&info->stop_eviction,
1175                                        !atomic_read(&info->stop_eviction));
1176                         mutex_lock(&shmem_swaplist_mutex);
1177                         /* ...but beware of the race if we peeked too early */
1178                         if (!atomic_read(&info->stop_eviction))
1179                                 list_del_init(&info->swaplist);
1180                         mutex_unlock(&shmem_swaplist_mutex);
1181                 }
1182         }
1183
1184         simple_xattrs_free(&info->xattrs);
1185         WARN_ON(inode->i_blocks);
1186         shmem_free_inode(inode->i_sb);
1187         clear_inode(inode);
1188 }
1189
1190 extern struct swap_info_struct *swap_info[];
1191
1192 static int shmem_find_swap_entries(struct address_space *mapping,
1193                                    pgoff_t start, unsigned int nr_entries,
1194                                    struct page **entries, pgoff_t *indices,
1195                                    unsigned int type, bool frontswap)
1196 {
1197         XA_STATE(xas, &mapping->i_pages, start);
1198         struct page *page;
1199         swp_entry_t entry;
1200         unsigned int ret = 0;
1201
1202         if (!nr_entries)
1203                 return 0;
1204
1205         rcu_read_lock();
1206         xas_for_each(&xas, page, ULONG_MAX) {
1207                 if (xas_retry(&xas, page))
1208                         continue;
1209
1210                 if (!xa_is_value(page))
1211                         continue;
1212
1213                 entry = radix_to_swp_entry(page);
1214                 if (swp_type(entry) != type)
1215                         continue;
1216                 if (frontswap &&
1217                     !frontswap_test(swap_info[type], swp_offset(entry)))
1218                         continue;
1219
1220                 indices[ret] = xas.xa_index;
1221                 entries[ret] = page;
1222
1223                 if (need_resched()) {
1224                         xas_pause(&xas);
1225                         cond_resched_rcu();
1226                 }
1227                 if (++ret == nr_entries)
1228                         break;
1229         }
1230         rcu_read_unlock();
1231
1232         return ret;
1233 }
1234
1235 /*
1236  * Move the swapped pages for an inode to page cache. Returns the count
1237  * of pages swapped in, or the error in case of failure.
1238  */
1239 static int shmem_unuse_swap_entries(struct inode *inode, struct pagevec pvec,
1240                                     pgoff_t *indices)
1241 {
1242         int i = 0;
1243         int ret = 0;
1244         int error = 0;
1245         struct address_space *mapping = inode->i_mapping;
1246
1247         for (i = 0; i < pvec.nr; i++) {
1248                 struct page *page = pvec.pages[i];
1249
1250                 if (!xa_is_value(page))
1251                         continue;
1252                 error = shmem_swapin_page(inode, indices[i],
1253                                           &page, SGP_CACHE,
1254                                           mapping_gfp_mask(mapping),
1255                                           NULL, NULL);
1256                 if (error == 0) {
1257                         unlock_page(page);
1258                         put_page(page);
1259                         ret++;
1260                 }
1261                 if (error == -ENOMEM)
1262                         break;
1263                 error = 0;
1264         }
1265         return error ? error : ret;
1266 }
1267
1268 /*
1269  * If swap found in inode, free it and move page from swapcache to filecache.
1270  */
1271 static int shmem_unuse_inode(struct inode *inode, unsigned int type,
1272                              bool frontswap, unsigned long *fs_pages_to_unuse)
1273 {
1274         struct address_space *mapping = inode->i_mapping;
1275         pgoff_t start = 0;
1276         struct pagevec pvec;
1277         pgoff_t indices[PAGEVEC_SIZE];
1278         bool frontswap_partial = (frontswap && *fs_pages_to_unuse > 0);
1279         int ret = 0;
1280
1281         pagevec_init(&pvec);
1282         do {
1283                 unsigned int nr_entries = PAGEVEC_SIZE;
1284
1285                 if (frontswap_partial && *fs_pages_to_unuse < PAGEVEC_SIZE)
1286                         nr_entries = *fs_pages_to_unuse;
1287
1288                 pvec.nr = shmem_find_swap_entries(mapping, start, nr_entries,
1289                                                   pvec.pages, indices,
1290                                                   type, frontswap);
1291                 if (pvec.nr == 0) {
1292                         ret = 0;
1293                         break;
1294                 }
1295
1296                 ret = shmem_unuse_swap_entries(inode, pvec, indices);
1297                 if (ret < 0)
1298                         break;
1299
1300                 if (frontswap_partial) {
1301                         *fs_pages_to_unuse -= ret;
1302                         if (*fs_pages_to_unuse == 0) {
1303                                 ret = FRONTSWAP_PAGES_UNUSED;
1304                                 break;
1305                         }
1306                 }
1307
1308                 start = indices[pvec.nr - 1];
1309         } while (true);
1310
1311         return ret;
1312 }
1313
1314 /*
1315  * Read all the shared memory data that resides in the swap
1316  * device 'type' back into memory, so the swap device can be
1317  * unused.
1318  */
1319 int shmem_unuse(unsigned int type, bool frontswap,
1320                 unsigned long *fs_pages_to_unuse)
1321 {
1322         struct shmem_inode_info *info, *next;
1323         int error = 0;
1324
1325         if (list_empty(&shmem_swaplist))
1326                 return 0;
1327
1328         mutex_lock(&shmem_swaplist_mutex);
1329         list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1330                 if (!info->swapped) {
1331                         list_del_init(&info->swaplist);
1332                         continue;
1333                 }
1334                 /*
1335                  * Drop the swaplist mutex while searching the inode for swap;
1336                  * but before doing so, make sure shmem_evict_inode() will not
1337                  * remove placeholder inode from swaplist, nor let it be freed
1338                  * (igrab() would protect from unlink, but not from unmount).
1339                  */
1340                 atomic_inc(&info->stop_eviction);
1341                 mutex_unlock(&shmem_swaplist_mutex);
1342
1343                 error = shmem_unuse_inode(&info->vfs_inode, type, frontswap,
1344                                           fs_pages_to_unuse);
1345                 cond_resched();
1346
1347                 mutex_lock(&shmem_swaplist_mutex);
1348                 next = list_next_entry(info, swaplist);
1349                 if (!info->swapped)
1350                         list_del_init(&info->swaplist);
1351                 if (atomic_dec_and_test(&info->stop_eviction))
1352                         wake_up_var(&info->stop_eviction);
1353                 if (error)
1354                         break;
1355         }
1356         mutex_unlock(&shmem_swaplist_mutex);
1357
1358         return error;
1359 }
1360
1361 /*
1362  * Move the page from the page cache to the swap cache.
1363  */
1364 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1365 {
1366         struct shmem_inode_info *info;
1367         struct address_space *mapping;
1368         struct inode *inode;
1369         swp_entry_t swap;
1370         pgoff_t index;
1371
1372         VM_BUG_ON_PAGE(PageCompound(page), page);
1373         BUG_ON(!PageLocked(page));
1374         mapping = page->mapping;
1375         index = page->index;
1376         inode = mapping->host;
1377         info = SHMEM_I(inode);
1378         if (info->flags & VM_LOCKED)
1379                 goto redirty;
1380         if (!total_swap_pages)
1381                 goto redirty;
1382
1383         /*
1384          * Our capabilities prevent regular writeback or sync from ever calling
1385          * shmem_writepage; but a stacking filesystem might use ->writepage of
1386          * its underlying filesystem, in which case tmpfs should write out to
1387          * swap only in response to memory pressure, and not for the writeback
1388          * threads or sync.
1389          */
1390         if (!wbc->for_reclaim) {
1391                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1392                 goto redirty;
1393         }
1394
1395         /*
1396          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1397          * value into swapfile.c, the only way we can correctly account for a
1398          * fallocated page arriving here is now to initialize it and write it.
1399          *
1400          * That's okay for a page already fallocated earlier, but if we have
1401          * not yet completed the fallocation, then (a) we want to keep track
1402          * of this page in case we have to undo it, and (b) it may not be a
1403          * good idea to continue anyway, once we're pushing into swap.  So
1404          * reactivate the page, and let shmem_fallocate() quit when too many.
1405          */
1406         if (!PageUptodate(page)) {
1407                 if (inode->i_private) {
1408                         struct shmem_falloc *shmem_falloc;
1409                         spin_lock(&inode->i_lock);
1410                         shmem_falloc = inode->i_private;
1411                         if (shmem_falloc &&
1412                             !shmem_falloc->waitq &&
1413                             index >= shmem_falloc->start &&
1414                             index < shmem_falloc->next)
1415                                 shmem_falloc->nr_unswapped++;
1416                         else
1417                                 shmem_falloc = NULL;
1418                         spin_unlock(&inode->i_lock);
1419                         if (shmem_falloc)
1420                                 goto redirty;
1421                 }
1422                 clear_highpage(page);
1423                 flush_dcache_page(page);
1424                 SetPageUptodate(page);
1425         }
1426
1427         swap = get_swap_page(page);
1428         if (!swap.val)
1429                 goto redirty;
1430
1431         /*
1432          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1433          * if it's not already there.  Do it now before the page is
1434          * moved to swap cache, when its pagelock no longer protects
1435          * the inode from eviction.  But don't unlock the mutex until
1436          * we've incremented swapped, because shmem_unuse_inode() will
1437          * prune a !swapped inode from the swaplist under this mutex.
1438          */
1439         mutex_lock(&shmem_swaplist_mutex);
1440         if (list_empty(&info->swaplist))
1441                 list_add(&info->swaplist, &shmem_swaplist);
1442
1443         if (add_to_swap_cache(page, swap,
1444                         __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1445                         NULL) == 0) {
1446                 spin_lock_irq(&info->lock);
1447                 shmem_recalc_inode(inode);
1448                 info->swapped++;
1449                 spin_unlock_irq(&info->lock);
1450
1451                 swap_shmem_alloc(swap);
1452                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1453
1454                 mutex_unlock(&shmem_swaplist_mutex);
1455                 BUG_ON(page_mapped(page));
1456                 swap_writepage(page, wbc);
1457                 return 0;
1458         }
1459
1460         mutex_unlock(&shmem_swaplist_mutex);
1461         put_swap_page(page, swap);
1462 redirty:
1463         set_page_dirty(page);
1464         if (wbc->for_reclaim)
1465                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1466         unlock_page(page);
1467         return 0;
1468 }
1469
1470 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1471 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1472 {
1473         char buffer[64];
1474
1475         if (!mpol || mpol->mode == MPOL_DEFAULT)
1476                 return;         /* show nothing */
1477
1478         mpol_to_str(buffer, sizeof(buffer), mpol);
1479
1480         seq_printf(seq, ",mpol=%s", buffer);
1481 }
1482
1483 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1484 {
1485         struct mempolicy *mpol = NULL;
1486         if (sbinfo->mpol) {
1487                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1488                 mpol = sbinfo->mpol;
1489                 mpol_get(mpol);
1490                 spin_unlock(&sbinfo->stat_lock);
1491         }
1492         return mpol;
1493 }
1494 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1495 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1496 {
1497 }
1498 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1499 {
1500         return NULL;
1501 }
1502 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1503 #ifndef CONFIG_NUMA
1504 #define vm_policy vm_private_data
1505 #endif
1506
1507 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1508                 struct shmem_inode_info *info, pgoff_t index)
1509 {
1510         /* Create a pseudo vma that just contains the policy */
1511         vma_init(vma, NULL);
1512         /* Bias interleave by inode number to distribute better across nodes */
1513         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1514         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1515 }
1516
1517 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1518 {
1519         /* Drop reference taken by mpol_shared_policy_lookup() */
1520         mpol_cond_put(vma->vm_policy);
1521 }
1522
1523 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1524                         struct shmem_inode_info *info, pgoff_t index)
1525 {
1526         struct vm_area_struct pvma;
1527         struct page *page;
1528         struct vm_fault vmf;
1529
1530         shmem_pseudo_vma_init(&pvma, info, index);
1531         vmf.vma = &pvma;
1532         vmf.address = 0;
1533         page = swap_cluster_readahead(swap, gfp, &vmf);
1534         shmem_pseudo_vma_destroy(&pvma);
1535
1536         return page;
1537 }
1538
1539 static struct page *shmem_alloc_hugepage(gfp_t gfp,
1540                 struct shmem_inode_info *info, pgoff_t index)
1541 {
1542         struct vm_area_struct pvma;
1543         struct address_space *mapping = info->vfs_inode.i_mapping;
1544         pgoff_t hindex;
1545         struct page *page;
1546
1547         hindex = round_down(index, HPAGE_PMD_NR);
1548         if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1549                                                                 XA_PRESENT))
1550                 return NULL;
1551
1552         shmem_pseudo_vma_init(&pvma, info, hindex);
1553         page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1554                         HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1555         shmem_pseudo_vma_destroy(&pvma);
1556         if (page)
1557                 prep_transhuge_page(page);
1558         else
1559                 count_vm_event(THP_FILE_FALLBACK);
1560         return page;
1561 }
1562
1563 static struct page *shmem_alloc_page(gfp_t gfp,
1564                         struct shmem_inode_info *info, pgoff_t index)
1565 {
1566         struct vm_area_struct pvma;
1567         struct page *page;
1568
1569         shmem_pseudo_vma_init(&pvma, info, index);
1570         page = alloc_page_vma(gfp, &pvma, 0);
1571         shmem_pseudo_vma_destroy(&pvma);
1572
1573         return page;
1574 }
1575
1576 static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1577                 struct inode *inode,
1578                 pgoff_t index, bool huge)
1579 {
1580         struct shmem_inode_info *info = SHMEM_I(inode);
1581         struct page *page;
1582         int nr;
1583         int err = -ENOSPC;
1584
1585         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1586                 huge = false;
1587         nr = huge ? HPAGE_PMD_NR : 1;
1588
1589         if (!shmem_inode_acct_block(inode, nr))
1590                 goto failed;
1591
1592         if (huge)
1593                 page = shmem_alloc_hugepage(gfp, info, index);
1594         else
1595                 page = shmem_alloc_page(gfp, info, index);
1596         if (page) {
1597                 __SetPageLocked(page);
1598                 __SetPageSwapBacked(page);
1599                 return page;
1600         }
1601
1602         err = -ENOMEM;
1603         shmem_inode_unacct_blocks(inode, nr);
1604 failed:
1605         return ERR_PTR(err);
1606 }
1607
1608 /*
1609  * When a page is moved from swapcache to shmem filecache (either by the
1610  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1611  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1612  * ignorance of the mapping it belongs to.  If that mapping has special
1613  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1614  * we may need to copy to a suitable page before moving to filecache.
1615  *
1616  * In a future release, this may well be extended to respect cpuset and
1617  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1618  * but for now it is a simple matter of zone.
1619  */
1620 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1621 {
1622         return page_zonenum(page) > gfp_zone(gfp);
1623 }
1624
1625 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1626                                 struct shmem_inode_info *info, pgoff_t index)
1627 {
1628         struct page *oldpage, *newpage;
1629         struct address_space *swap_mapping;
1630         swp_entry_t entry;
1631         pgoff_t swap_index;
1632         int error;
1633
1634         oldpage = *pagep;
1635         entry.val = page_private(oldpage);
1636         swap_index = swp_offset(entry);
1637         swap_mapping = page_mapping(oldpage);
1638
1639         /*
1640          * We have arrived here because our zones are constrained, so don't
1641          * limit chance of success by further cpuset and node constraints.
1642          */
1643         gfp &= ~GFP_CONSTRAINT_MASK;
1644         newpage = shmem_alloc_page(gfp, info, index);
1645         if (!newpage)
1646                 return -ENOMEM;
1647
1648         get_page(newpage);
1649         copy_highpage(newpage, oldpage);
1650         flush_dcache_page(newpage);
1651
1652         __SetPageLocked(newpage);
1653         __SetPageSwapBacked(newpage);
1654         SetPageUptodate(newpage);
1655         set_page_private(newpage, entry.val);
1656         SetPageSwapCache(newpage);
1657
1658         /*
1659          * Our caller will very soon move newpage out of swapcache, but it's
1660          * a nice clean interface for us to replace oldpage by newpage there.
1661          */
1662         xa_lock_irq(&swap_mapping->i_pages);
1663         error = shmem_replace_entry(swap_mapping, swap_index, oldpage, newpage);
1664         if (!error) {
1665                 mem_cgroup_migrate(oldpage, newpage);
1666                 __inc_lruvec_page_state(newpage, NR_FILE_PAGES);
1667                 __dec_lruvec_page_state(oldpage, NR_FILE_PAGES);
1668         }
1669         xa_unlock_irq(&swap_mapping->i_pages);
1670
1671         if (unlikely(error)) {
1672                 /*
1673                  * Is this possible?  I think not, now that our callers check
1674                  * both PageSwapCache and page_private after getting page lock;
1675                  * but be defensive.  Reverse old to newpage for clear and free.
1676                  */
1677                 oldpage = newpage;
1678         } else {
1679                 lru_cache_add(newpage);
1680                 *pagep = newpage;
1681         }
1682
1683         ClearPageSwapCache(oldpage);
1684         set_page_private(oldpage, 0);
1685
1686         unlock_page(oldpage);
1687         put_page(oldpage);
1688         put_page(oldpage);
1689         return error;
1690 }
1691
1692 /*
1693  * Swap in the page pointed to by *pagep.
1694  * Caller has to make sure that *pagep contains a valid swapped page.
1695  * Returns 0 and the page in pagep if success. On failure, returns the
1696  * error code and NULL in *pagep.
1697  */
1698 static int shmem_swapin_page(struct inode *inode, pgoff_t index,
1699                              struct page **pagep, enum sgp_type sgp,
1700                              gfp_t gfp, struct vm_area_struct *vma,
1701                              vm_fault_t *fault_type)
1702 {
1703         struct address_space *mapping = inode->i_mapping;
1704         struct shmem_inode_info *info = SHMEM_I(inode);
1705         struct mm_struct *charge_mm = vma ? vma->vm_mm : current->mm;
1706         struct page *page;
1707         swp_entry_t swap;
1708         int error;
1709
1710         VM_BUG_ON(!*pagep || !xa_is_value(*pagep));
1711         swap = radix_to_swp_entry(*pagep);
1712         *pagep = NULL;
1713
1714         /* Look it up and read it in.. */
1715         page = lookup_swap_cache(swap, NULL, 0);
1716         if (!page) {
1717                 /* Or update major stats only when swapin succeeds?? */
1718                 if (fault_type) {
1719                         *fault_type |= VM_FAULT_MAJOR;
1720                         count_vm_event(PGMAJFAULT);
1721                         count_memcg_event_mm(charge_mm, PGMAJFAULT);
1722                 }
1723                 /* Here we actually start the io */
1724                 page = shmem_swapin(swap, gfp, info, index);
1725                 if (!page) {
1726                         error = -ENOMEM;
1727                         goto failed;
1728                 }
1729         }
1730
1731         /* We have to do this with page locked to prevent races */
1732         lock_page(page);
1733         if (!PageSwapCache(page) || page_private(page) != swap.val ||
1734             !shmem_confirm_swap(mapping, index, swap)) {
1735                 error = -EEXIST;
1736                 goto unlock;
1737         }
1738         if (!PageUptodate(page)) {
1739                 error = -EIO;
1740                 goto failed;
1741         }
1742         wait_on_page_writeback(page);
1743
1744         /*
1745          * Some architectures may have to restore extra metadata to the
1746          * physical page after reading from swap.
1747          */
1748         arch_swap_restore(swap, page);
1749
1750         if (shmem_should_replace_page(page, gfp)) {
1751                 error = shmem_replace_page(&page, gfp, info, index);
1752                 if (error)
1753                         goto failed;
1754         }
1755
1756         error = shmem_add_to_page_cache(page, mapping, index,
1757                                         swp_to_radix_entry(swap), gfp,
1758                                         charge_mm);
1759         if (error)
1760                 goto failed;
1761
1762         spin_lock_irq(&info->lock);
1763         info->swapped--;
1764         shmem_recalc_inode(inode);
1765         spin_unlock_irq(&info->lock);
1766
1767         if (sgp == SGP_WRITE)
1768                 mark_page_accessed(page);
1769
1770         delete_from_swap_cache(page);
1771         set_page_dirty(page);
1772         swap_free(swap);
1773
1774         *pagep = page;
1775         return 0;
1776 failed:
1777         if (!shmem_confirm_swap(mapping, index, swap))
1778                 error = -EEXIST;
1779 unlock:
1780         if (page) {
1781                 unlock_page(page);
1782                 put_page(page);
1783         }
1784
1785         return error;
1786 }
1787
1788 /*
1789  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1790  *
1791  * If we allocate a new one we do not mark it dirty. That's up to the
1792  * vm. If we swap it in we mark it dirty since we also free the swap
1793  * entry since a page cannot live in both the swap and page cache.
1794  *
1795  * vmf and fault_type are only supplied by shmem_fault:
1796  * otherwise they are NULL.
1797  */
1798 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1799         struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1800         struct vm_area_struct *vma, struct vm_fault *vmf,
1801                         vm_fault_t *fault_type)
1802 {
1803         struct address_space *mapping = inode->i_mapping;
1804         struct shmem_inode_info *info = SHMEM_I(inode);
1805         struct shmem_sb_info *sbinfo;
1806         struct mm_struct *charge_mm;
1807         struct page *page;
1808         enum sgp_type sgp_huge = sgp;
1809         pgoff_t hindex = index;
1810         int error;
1811         int once = 0;
1812         int alloced = 0;
1813
1814         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1815                 return -EFBIG;
1816         if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1817                 sgp = SGP_CACHE;
1818 repeat:
1819         if (sgp <= SGP_CACHE &&
1820             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1821                 return -EINVAL;
1822         }
1823
1824         sbinfo = SHMEM_SB(inode->i_sb);
1825         charge_mm = vma ? vma->vm_mm : current->mm;
1826
1827         page = find_lock_entry(mapping, index);
1828         if (xa_is_value(page)) {
1829                 error = shmem_swapin_page(inode, index, &page,
1830                                           sgp, gfp, vma, fault_type);
1831                 if (error == -EEXIST)
1832                         goto repeat;
1833
1834                 *pagep = page;
1835                 return error;
1836         }
1837
1838         if (page)
1839                 hindex = page->index;
1840         if (page && sgp == SGP_WRITE)
1841                 mark_page_accessed(page);
1842
1843         /* fallocated page? */
1844         if (page && !PageUptodate(page)) {
1845                 if (sgp != SGP_READ)
1846                         goto clear;
1847                 unlock_page(page);
1848                 put_page(page);
1849                 page = NULL;
1850                 hindex = index;
1851         }
1852         if (page || sgp == SGP_READ)
1853                 goto out;
1854
1855         /*
1856          * Fast cache lookup did not find it:
1857          * bring it back from swap or allocate.
1858          */
1859
1860         if (vma && userfaultfd_missing(vma)) {
1861                 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1862                 return 0;
1863         }
1864
1865         /* shmem_symlink() */
1866         if (mapping->a_ops != &shmem_aops)
1867                 goto alloc_nohuge;
1868         if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1869                 goto alloc_nohuge;
1870         if (shmem_huge == SHMEM_HUGE_FORCE)
1871                 goto alloc_huge;
1872         switch (sbinfo->huge) {
1873         case SHMEM_HUGE_NEVER:
1874                 goto alloc_nohuge;
1875         case SHMEM_HUGE_WITHIN_SIZE: {
1876                 loff_t i_size;
1877                 pgoff_t off;
1878
1879                 off = round_up(index, HPAGE_PMD_NR);
1880                 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1881                 if (i_size >= HPAGE_PMD_SIZE &&
1882                     i_size >> PAGE_SHIFT >= off)
1883                         goto alloc_huge;
1884
1885                 fallthrough;
1886         }
1887         case SHMEM_HUGE_ADVISE:
1888                 if (sgp_huge == SGP_HUGE)
1889                         goto alloc_huge;
1890                 /* TODO: implement fadvise() hints */
1891                 goto alloc_nohuge;
1892         }
1893
1894 alloc_huge:
1895         page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1896         if (IS_ERR(page)) {
1897 alloc_nohuge:
1898                 page = shmem_alloc_and_acct_page(gfp, inode,
1899                                                  index, false);
1900         }
1901         if (IS_ERR(page)) {
1902                 int retry = 5;
1903
1904                 error = PTR_ERR(page);
1905                 page = NULL;
1906                 if (error != -ENOSPC)
1907                         goto unlock;
1908                 /*
1909                  * Try to reclaim some space by splitting a huge page
1910                  * beyond i_size on the filesystem.
1911                  */
1912                 while (retry--) {
1913                         int ret;
1914
1915                         ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1916                         if (ret == SHRINK_STOP)
1917                                 break;
1918                         if (ret)
1919                                 goto alloc_nohuge;
1920                 }
1921                 goto unlock;
1922         }
1923
1924         if (PageTransHuge(page))
1925                 hindex = round_down(index, HPAGE_PMD_NR);
1926         else
1927                 hindex = index;
1928
1929         if (sgp == SGP_WRITE)
1930                 __SetPageReferenced(page);
1931
1932         error = shmem_add_to_page_cache(page, mapping, hindex,
1933                                         NULL, gfp & GFP_RECLAIM_MASK,
1934                                         charge_mm);
1935         if (error)
1936                 goto unacct;
1937         lru_cache_add(page);
1938
1939         spin_lock_irq(&info->lock);
1940         info->alloced += compound_nr(page);
1941         inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1942         shmem_recalc_inode(inode);
1943         spin_unlock_irq(&info->lock);
1944         alloced = true;
1945
1946         if (PageTransHuge(page) &&
1947             DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1948                         hindex + HPAGE_PMD_NR - 1) {
1949                 /*
1950                  * Part of the huge page is beyond i_size: subject
1951                  * to shrink under memory pressure.
1952                  */
1953                 spin_lock(&sbinfo->shrinklist_lock);
1954                 /*
1955                  * _careful to defend against unlocked access to
1956                  * ->shrink_list in shmem_unused_huge_shrink()
1957                  */
1958                 if (list_empty_careful(&info->shrinklist)) {
1959                         list_add_tail(&info->shrinklist,
1960                                       &sbinfo->shrinklist);
1961                         sbinfo->shrinklist_len++;
1962                 }
1963                 spin_unlock(&sbinfo->shrinklist_lock);
1964         }
1965
1966         /*
1967          * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1968          */
1969         if (sgp == SGP_FALLOC)
1970                 sgp = SGP_WRITE;
1971 clear:
1972         /*
1973          * Let SGP_WRITE caller clear ends if write does not fill page;
1974          * but SGP_FALLOC on a page fallocated earlier must initialize
1975          * it now, lest undo on failure cancel our earlier guarantee.
1976          */
1977         if (sgp != SGP_WRITE && !PageUptodate(page)) {
1978                 int i;
1979
1980                 for (i = 0; i < compound_nr(page); i++) {
1981                         clear_highpage(page + i);
1982                         flush_dcache_page(page + i);
1983                 }
1984                 SetPageUptodate(page);
1985         }
1986
1987         /* Perhaps the file has been truncated since we checked */
1988         if (sgp <= SGP_CACHE &&
1989             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1990                 if (alloced) {
1991                         ClearPageDirty(page);
1992                         delete_from_page_cache(page);
1993                         spin_lock_irq(&info->lock);
1994                         shmem_recalc_inode(inode);
1995                         spin_unlock_irq(&info->lock);
1996                 }
1997                 error = -EINVAL;
1998                 goto unlock;
1999         }
2000 out:
2001         *pagep = page + index - hindex;
2002         return 0;
2003
2004         /*
2005          * Error recovery.
2006          */
2007 unacct:
2008         shmem_inode_unacct_blocks(inode, compound_nr(page));
2009
2010         if (PageTransHuge(page)) {
2011                 unlock_page(page);
2012                 put_page(page);
2013                 goto alloc_nohuge;
2014         }
2015 unlock:
2016         if (page) {
2017                 unlock_page(page);
2018                 put_page(page);
2019         }
2020         if (error == -ENOSPC && !once++) {
2021                 spin_lock_irq(&info->lock);
2022                 shmem_recalc_inode(inode);
2023                 spin_unlock_irq(&info->lock);
2024                 goto repeat;
2025         }
2026         if (error == -EEXIST)
2027                 goto repeat;
2028         return error;
2029 }
2030
2031 /*
2032  * This is like autoremove_wake_function, but it removes the wait queue
2033  * entry unconditionally - even if something else had already woken the
2034  * target.
2035  */
2036 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2037 {
2038         int ret = default_wake_function(wait, mode, sync, key);
2039         list_del_init(&wait->entry);
2040         return ret;
2041 }
2042
2043 static vm_fault_t shmem_fault(struct vm_fault *vmf)
2044 {
2045         struct vm_area_struct *vma = vmf->vma;
2046         struct inode *inode = file_inode(vma->vm_file);
2047         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2048         enum sgp_type sgp;
2049         int err;
2050         vm_fault_t ret = VM_FAULT_LOCKED;
2051
2052         /*
2053          * Trinity finds that probing a hole which tmpfs is punching can
2054          * prevent the hole-punch from ever completing: which in turn
2055          * locks writers out with its hold on i_mutex.  So refrain from
2056          * faulting pages into the hole while it's being punched.  Although
2057          * shmem_undo_range() does remove the additions, it may be unable to
2058          * keep up, as each new page needs its own unmap_mapping_range() call,
2059          * and the i_mmap tree grows ever slower to scan if new vmas are added.
2060          *
2061          * It does not matter if we sometimes reach this check just before the
2062          * hole-punch begins, so that one fault then races with the punch:
2063          * we just need to make racing faults a rare case.
2064          *
2065          * The implementation below would be much simpler if we just used a
2066          * standard mutex or completion: but we cannot take i_mutex in fault,
2067          * and bloating every shmem inode for this unlikely case would be sad.
2068          */
2069         if (unlikely(inode->i_private)) {
2070                 struct shmem_falloc *shmem_falloc;
2071
2072                 spin_lock(&inode->i_lock);
2073                 shmem_falloc = inode->i_private;
2074                 if (shmem_falloc &&
2075                     shmem_falloc->waitq &&
2076                     vmf->pgoff >= shmem_falloc->start &&
2077                     vmf->pgoff < shmem_falloc->next) {
2078                         struct file *fpin;
2079                         wait_queue_head_t *shmem_falloc_waitq;
2080                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2081
2082                         ret = VM_FAULT_NOPAGE;
2083                         fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2084                         if (fpin)
2085                                 ret = VM_FAULT_RETRY;
2086
2087                         shmem_falloc_waitq = shmem_falloc->waitq;
2088                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2089                                         TASK_UNINTERRUPTIBLE);
2090                         spin_unlock(&inode->i_lock);
2091                         schedule();
2092
2093                         /*
2094                          * shmem_falloc_waitq points into the shmem_fallocate()
2095                          * stack of the hole-punching task: shmem_falloc_waitq
2096                          * is usually invalid by the time we reach here, but
2097                          * finish_wait() does not dereference it in that case;
2098                          * though i_lock needed lest racing with wake_up_all().
2099                          */
2100                         spin_lock(&inode->i_lock);
2101                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2102                         spin_unlock(&inode->i_lock);
2103
2104                         if (fpin)
2105                                 fput(fpin);
2106                         return ret;
2107                 }
2108                 spin_unlock(&inode->i_lock);
2109         }
2110
2111         sgp = SGP_CACHE;
2112
2113         if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2114             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2115                 sgp = SGP_NOHUGE;
2116         else if (vma->vm_flags & VM_HUGEPAGE)
2117                 sgp = SGP_HUGE;
2118
2119         err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2120                                   gfp, vma, vmf, &ret);
2121         if (err)
2122                 return vmf_error(err);
2123         return ret;
2124 }
2125
2126 unsigned long shmem_get_unmapped_area(struct file *file,
2127                                       unsigned long uaddr, unsigned long len,
2128                                       unsigned long pgoff, unsigned long flags)
2129 {
2130         unsigned long (*get_area)(struct file *,
2131                 unsigned long, unsigned long, unsigned long, unsigned long);
2132         unsigned long addr;
2133         unsigned long offset;
2134         unsigned long inflated_len;
2135         unsigned long inflated_addr;
2136         unsigned long inflated_offset;
2137
2138         if (len > TASK_SIZE)
2139                 return -ENOMEM;
2140
2141         get_area = current->mm->get_unmapped_area;
2142         addr = get_area(file, uaddr, len, pgoff, flags);
2143
2144         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2145                 return addr;
2146         if (IS_ERR_VALUE(addr))
2147                 return addr;
2148         if (addr & ~PAGE_MASK)
2149                 return addr;
2150         if (addr > TASK_SIZE - len)
2151                 return addr;
2152
2153         if (shmem_huge == SHMEM_HUGE_DENY)
2154                 return addr;
2155         if (len < HPAGE_PMD_SIZE)
2156                 return addr;
2157         if (flags & MAP_FIXED)
2158                 return addr;
2159         /*
2160          * Our priority is to support MAP_SHARED mapped hugely;
2161          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2162          * But if caller specified an address hint and we allocated area there
2163          * successfully, respect that as before.
2164          */
2165         if (uaddr == addr)
2166                 return addr;
2167
2168         if (shmem_huge != SHMEM_HUGE_FORCE) {
2169                 struct super_block *sb;
2170
2171                 if (file) {
2172                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2173                         sb = file_inode(file)->i_sb;
2174                 } else {
2175                         /*
2176                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2177                          * for "/dev/zero", to create a shared anonymous object.
2178                          */
2179                         if (IS_ERR(shm_mnt))
2180                                 return addr;
2181                         sb = shm_mnt->mnt_sb;
2182                 }
2183                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2184                         return addr;
2185         }
2186
2187         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2188         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2189                 return addr;
2190         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2191                 return addr;
2192
2193         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2194         if (inflated_len > TASK_SIZE)
2195                 return addr;
2196         if (inflated_len < len)
2197                 return addr;
2198
2199         inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2200         if (IS_ERR_VALUE(inflated_addr))
2201                 return addr;
2202         if (inflated_addr & ~PAGE_MASK)
2203                 return addr;
2204
2205         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2206         inflated_addr += offset - inflated_offset;
2207         if (inflated_offset > offset)
2208                 inflated_addr += HPAGE_PMD_SIZE;
2209
2210         if (inflated_addr > TASK_SIZE - len)
2211                 return addr;
2212         return inflated_addr;
2213 }
2214
2215 #ifdef CONFIG_NUMA
2216 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2217 {
2218         struct inode *inode = file_inode(vma->vm_file);
2219         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2220 }
2221
2222 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2223                                           unsigned long addr)
2224 {
2225         struct inode *inode = file_inode(vma->vm_file);
2226         pgoff_t index;
2227
2228         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2229         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2230 }
2231 #endif
2232
2233 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2234 {
2235         struct inode *inode = file_inode(file);
2236         struct shmem_inode_info *info = SHMEM_I(inode);
2237         int retval = -ENOMEM;
2238
2239         /*
2240          * What serializes the accesses to info->flags?
2241          * ipc_lock_object() when called from shmctl_do_lock(),
2242          * no serialization needed when called from shm_destroy().
2243          */
2244         if (lock && !(info->flags & VM_LOCKED)) {
2245                 if (!user_shm_lock(inode->i_size, user))
2246                         goto out_nomem;
2247                 info->flags |= VM_LOCKED;
2248                 mapping_set_unevictable(file->f_mapping);
2249         }
2250         if (!lock && (info->flags & VM_LOCKED) && user) {
2251                 user_shm_unlock(inode->i_size, user);
2252                 info->flags &= ~VM_LOCKED;
2253                 mapping_clear_unevictable(file->f_mapping);
2254         }
2255         retval = 0;
2256
2257 out_nomem:
2258         return retval;
2259 }
2260
2261 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2262 {
2263         struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2264         int ret;
2265
2266         ret = seal_check_future_write(info->seals, vma);
2267         if (ret)
2268                 return ret;
2269
2270         /* arm64 - allow memory tagging on RAM-based files */
2271         vma->vm_flags |= VM_MTE_ALLOWED;
2272
2273         file_accessed(file);
2274         vma->vm_ops = &shmem_vm_ops;
2275         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
2276                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2277                         (vma->vm_end & HPAGE_PMD_MASK)) {
2278                 khugepaged_enter(vma, vma->vm_flags);
2279         }
2280         return 0;
2281 }
2282
2283 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2284                                      umode_t mode, dev_t dev, unsigned long flags)
2285 {
2286         struct inode *inode;
2287         struct shmem_inode_info *info;
2288         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2289         ino_t ino;
2290
2291         if (shmem_reserve_inode(sb, &ino))
2292                 return NULL;
2293
2294         inode = new_inode(sb);
2295         if (inode) {
2296                 inode->i_ino = ino;
2297                 inode_init_owner(inode, dir, mode);
2298                 inode->i_blocks = 0;
2299                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2300                 inode->i_generation = prandom_u32();
2301                 info = SHMEM_I(inode);
2302                 memset(info, 0, (char *)inode - (char *)info);
2303                 spin_lock_init(&info->lock);
2304                 atomic_set(&info->stop_eviction, 0);
2305                 info->seals = F_SEAL_SEAL;
2306                 info->flags = flags & VM_NORESERVE;
2307                 INIT_LIST_HEAD(&info->shrinklist);
2308                 INIT_LIST_HEAD(&info->swaplist);
2309                 simple_xattrs_init(&info->xattrs);
2310                 cache_no_acl(inode);
2311
2312                 switch (mode & S_IFMT) {
2313                 default:
2314                         inode->i_op = &shmem_special_inode_operations;
2315                         init_special_inode(inode, mode, dev);
2316                         break;
2317                 case S_IFREG:
2318                         inode->i_mapping->a_ops = &shmem_aops;
2319                         inode->i_op = &shmem_inode_operations;
2320                         inode->i_fop = &shmem_file_operations;
2321                         mpol_shared_policy_init(&info->policy,
2322                                                  shmem_get_sbmpol(sbinfo));
2323                         break;
2324                 case S_IFDIR:
2325                         inc_nlink(inode);
2326                         /* Some things misbehave if size == 0 on a directory */
2327                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2328                         inode->i_op = &shmem_dir_inode_operations;
2329                         inode->i_fop = &simple_dir_operations;
2330                         break;
2331                 case S_IFLNK:
2332                         /*
2333                          * Must not load anything in the rbtree,
2334                          * mpol_free_shared_policy will not be called.
2335                          */
2336                         mpol_shared_policy_init(&info->policy, NULL);
2337                         break;
2338                 }
2339
2340                 lockdep_annotate_inode_mutex_key(inode);
2341         } else
2342                 shmem_free_inode(sb);
2343         return inode;
2344 }
2345
2346 bool shmem_mapping(struct address_space *mapping)
2347 {
2348         return mapping->a_ops == &shmem_aops;
2349 }
2350
2351 static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2352                                   pmd_t *dst_pmd,
2353                                   struct vm_area_struct *dst_vma,
2354                                   unsigned long dst_addr,
2355                                   unsigned long src_addr,
2356                                   bool zeropage,
2357                                   struct page **pagep)
2358 {
2359         struct inode *inode = file_inode(dst_vma->vm_file);
2360         struct shmem_inode_info *info = SHMEM_I(inode);
2361         struct address_space *mapping = inode->i_mapping;
2362         gfp_t gfp = mapping_gfp_mask(mapping);
2363         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2364         spinlock_t *ptl;
2365         void *page_kaddr;
2366         struct page *page;
2367         pte_t _dst_pte, *dst_pte;
2368         int ret;
2369         pgoff_t offset, max_off;
2370
2371         ret = -ENOMEM;
2372         if (!shmem_inode_acct_block(inode, 1)) {
2373                 /*
2374                  * We may have got a page, returned -ENOENT triggering a retry,
2375                  * and now we find ourselves with -ENOMEM. Release the page, to
2376                  * avoid a BUG_ON in our caller.
2377                  */
2378                 if (unlikely(*pagep)) {
2379                         put_page(*pagep);
2380                         *pagep = NULL;
2381                 }
2382                 goto out;
2383         }
2384
2385         if (!*pagep) {
2386                 page = shmem_alloc_page(gfp, info, pgoff);
2387                 if (!page)
2388                         goto out_unacct_blocks;
2389
2390                 if (!zeropage) {        /* mcopy_atomic */
2391                         page_kaddr = kmap_atomic(page);
2392                         ret = copy_from_user(page_kaddr,
2393                                              (const void __user *)src_addr,
2394                                              PAGE_SIZE);
2395                         kunmap_atomic(page_kaddr);
2396
2397                         /* fallback to copy_from_user outside mmap_lock */
2398                         if (unlikely(ret)) {
2399                                 *pagep = page;
2400                                 shmem_inode_unacct_blocks(inode, 1);
2401                                 /* don't free the page */
2402                                 return -ENOENT;
2403                         }
2404                 } else {                /* mfill_zeropage_atomic */
2405                         clear_highpage(page);
2406                 }
2407         } else {
2408                 page = *pagep;
2409                 *pagep = NULL;
2410         }
2411
2412         VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2413         __SetPageLocked(page);
2414         __SetPageSwapBacked(page);
2415         __SetPageUptodate(page);
2416
2417         ret = -EFAULT;
2418         offset = linear_page_index(dst_vma, dst_addr);
2419         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2420         if (unlikely(offset >= max_off))
2421                 goto out_release;
2422
2423         ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL,
2424                                       gfp & GFP_RECLAIM_MASK, dst_mm);
2425         if (ret)
2426                 goto out_release;
2427
2428         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2429         if (dst_vma->vm_flags & VM_WRITE)
2430                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2431         else {
2432                 /*
2433                  * We don't set the pte dirty if the vma has no
2434                  * VM_WRITE permission, so mark the page dirty or it
2435                  * could be freed from under us. We could do it
2436                  * unconditionally before unlock_page(), but doing it
2437                  * only if VM_WRITE is not set is faster.
2438                  */
2439                 set_page_dirty(page);
2440         }
2441
2442         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2443
2444         ret = -EFAULT;
2445         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2446         if (unlikely(offset >= max_off))
2447                 goto out_release_unlock;
2448
2449         ret = -EEXIST;
2450         if (!pte_none(*dst_pte))
2451                 goto out_release_unlock;
2452
2453         lru_cache_add(page);
2454
2455         spin_lock_irq(&info->lock);
2456         info->alloced++;
2457         inode->i_blocks += BLOCKS_PER_PAGE;
2458         shmem_recalc_inode(inode);
2459         spin_unlock_irq(&info->lock);
2460
2461         inc_mm_counter(dst_mm, mm_counter_file(page));
2462         page_add_file_rmap(page, false);
2463         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2464
2465         /* No need to invalidate - it was non-present before */
2466         update_mmu_cache(dst_vma, dst_addr, dst_pte);
2467         pte_unmap_unlock(dst_pte, ptl);
2468         unlock_page(page);
2469         ret = 0;
2470 out:
2471         return ret;
2472 out_release_unlock:
2473         pte_unmap_unlock(dst_pte, ptl);
2474         ClearPageDirty(page);
2475         delete_from_page_cache(page);
2476 out_release:
2477         unlock_page(page);
2478         put_page(page);
2479 out_unacct_blocks:
2480         shmem_inode_unacct_blocks(inode, 1);
2481         goto out;
2482 }
2483
2484 int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2485                            pmd_t *dst_pmd,
2486                            struct vm_area_struct *dst_vma,
2487                            unsigned long dst_addr,
2488                            unsigned long src_addr,
2489                            struct page **pagep)
2490 {
2491         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2492                                       dst_addr, src_addr, false, pagep);
2493 }
2494
2495 int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2496                              pmd_t *dst_pmd,
2497                              struct vm_area_struct *dst_vma,
2498                              unsigned long dst_addr)
2499 {
2500         struct page *page = NULL;
2501
2502         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2503                                       dst_addr, 0, true, &page);
2504 }
2505
2506 #ifdef CONFIG_TMPFS
2507 static const struct inode_operations shmem_symlink_inode_operations;
2508 static const struct inode_operations shmem_short_symlink_operations;
2509
2510 #ifdef CONFIG_TMPFS_XATTR
2511 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2512 #else
2513 #define shmem_initxattrs NULL
2514 #endif
2515
2516 static int
2517 shmem_write_begin(struct file *file, struct address_space *mapping,
2518                         loff_t pos, unsigned len, unsigned flags,
2519                         struct page **pagep, void **fsdata)
2520 {
2521         struct inode *inode = mapping->host;
2522         struct shmem_inode_info *info = SHMEM_I(inode);
2523         pgoff_t index = pos >> PAGE_SHIFT;
2524
2525         /* i_mutex is held by caller */
2526         if (unlikely(info->seals & (F_SEAL_GROW |
2527                                    F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2528                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2529                         return -EPERM;
2530                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2531                         return -EPERM;
2532         }
2533
2534         return shmem_getpage(inode, index, pagep, SGP_WRITE);
2535 }
2536
2537 static int
2538 shmem_write_end(struct file *file, struct address_space *mapping,
2539                         loff_t pos, unsigned len, unsigned copied,
2540                         struct page *page, void *fsdata)
2541 {
2542         struct inode *inode = mapping->host;
2543
2544         if (pos + copied > inode->i_size)
2545                 i_size_write(inode, pos + copied);
2546
2547         if (!PageUptodate(page)) {
2548                 struct page *head = compound_head(page);
2549                 if (PageTransCompound(page)) {
2550                         int i;
2551
2552                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2553                                 if (head + i == page)
2554                                         continue;
2555                                 clear_highpage(head + i);
2556                                 flush_dcache_page(head + i);
2557                         }
2558                 }
2559                 if (copied < PAGE_SIZE) {
2560                         unsigned from = pos & (PAGE_SIZE - 1);
2561                         zero_user_segments(page, 0, from,
2562                                         from + copied, PAGE_SIZE);
2563                 }
2564                 SetPageUptodate(head);
2565         }
2566         set_page_dirty(page);
2567         unlock_page(page);
2568         put_page(page);
2569
2570         return copied;
2571 }
2572
2573 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2574 {
2575         struct file *file = iocb->ki_filp;
2576         struct inode *inode = file_inode(file);
2577         struct address_space *mapping = inode->i_mapping;
2578         pgoff_t index;
2579         unsigned long offset;
2580         enum sgp_type sgp = SGP_READ;
2581         int error = 0;
2582         ssize_t retval = 0;
2583         loff_t *ppos = &iocb->ki_pos;
2584
2585         /*
2586          * Might this read be for a stacking filesystem?  Then when reading
2587          * holes of a sparse file, we actually need to allocate those pages,
2588          * and even mark them dirty, so it cannot exceed the max_blocks limit.
2589          */
2590         if (!iter_is_iovec(to))
2591                 sgp = SGP_CACHE;
2592
2593         index = *ppos >> PAGE_SHIFT;
2594         offset = *ppos & ~PAGE_MASK;
2595
2596         for (;;) {
2597                 struct page *page = NULL;
2598                 pgoff_t end_index;
2599                 unsigned long nr, ret;
2600                 loff_t i_size = i_size_read(inode);
2601
2602                 end_index = i_size >> PAGE_SHIFT;
2603                 if (index > end_index)
2604                         break;
2605                 if (index == end_index) {
2606                         nr = i_size & ~PAGE_MASK;
2607                         if (nr <= offset)
2608                                 break;
2609                 }
2610
2611                 error = shmem_getpage(inode, index, &page, sgp);
2612                 if (error) {
2613                         if (error == -EINVAL)
2614                                 error = 0;
2615                         break;
2616                 }
2617                 if (page) {
2618                         if (sgp == SGP_CACHE)
2619                                 set_page_dirty(page);
2620                         unlock_page(page);
2621                 }
2622
2623                 /*
2624                  * We must evaluate after, since reads (unlike writes)
2625                  * are called without i_mutex protection against truncate
2626                  */
2627                 nr = PAGE_SIZE;
2628                 i_size = i_size_read(inode);
2629                 end_index = i_size >> PAGE_SHIFT;
2630                 if (index == end_index) {
2631                         nr = i_size & ~PAGE_MASK;
2632                         if (nr <= offset) {
2633                                 if (page)
2634                                         put_page(page);
2635                                 break;
2636                         }
2637                 }
2638                 nr -= offset;
2639
2640                 if (page) {
2641                         /*
2642                          * If users can be writing to this page using arbitrary
2643                          * virtual addresses, take care about potential aliasing
2644                          * before reading the page on the kernel side.
2645                          */
2646                         if (mapping_writably_mapped(mapping))
2647                                 flush_dcache_page(page);
2648                         /*
2649                          * Mark the page accessed if we read the beginning.
2650                          */
2651                         if (!offset)
2652                                 mark_page_accessed(page);
2653                 } else {
2654                         page = ZERO_PAGE(0);
2655                         get_page(page);
2656                 }
2657
2658                 /*
2659                  * Ok, we have the page, and it's up-to-date, so
2660                  * now we can copy it to user space...
2661                  */
2662                 ret = copy_page_to_iter(page, offset, nr, to);
2663                 retval += ret;
2664                 offset += ret;
2665                 index += offset >> PAGE_SHIFT;
2666                 offset &= ~PAGE_MASK;
2667
2668                 put_page(page);
2669                 if (!iov_iter_count(to))
2670                         break;
2671                 if (ret < nr) {
2672                         error = -EFAULT;
2673                         break;
2674                 }
2675                 cond_resched();
2676         }
2677
2678         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2679         file_accessed(file);
2680         return retval ? retval : error;
2681 }
2682
2683 /*
2684  * llseek SEEK_DATA or SEEK_HOLE through the page cache.
2685  */
2686 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2687                                     pgoff_t index, pgoff_t end, int whence)
2688 {
2689         struct page *page;
2690         struct pagevec pvec;
2691         pgoff_t indices[PAGEVEC_SIZE];
2692         bool done = false;
2693         int i;
2694
2695         pagevec_init(&pvec);
2696         pvec.nr = 1;            /* start small: we may be there already */
2697         while (!done) {
2698                 pvec.nr = find_get_entries(mapping, index,
2699                                         pvec.nr, pvec.pages, indices);
2700                 if (!pvec.nr) {
2701                         if (whence == SEEK_DATA)
2702                                 index = end;
2703                         break;
2704                 }
2705                 for (i = 0; i < pvec.nr; i++, index++) {
2706                         if (index < indices[i]) {
2707                                 if (whence == SEEK_HOLE) {
2708                                         done = true;
2709                                         break;
2710                                 }
2711                                 index = indices[i];
2712                         }
2713                         page = pvec.pages[i];
2714                         if (page && !xa_is_value(page)) {
2715                                 if (!PageUptodate(page))
2716                                         page = NULL;
2717                         }
2718                         if (index >= end ||
2719                             (page && whence == SEEK_DATA) ||
2720                             (!page && whence == SEEK_HOLE)) {
2721                                 done = true;
2722                                 break;
2723                         }
2724                 }
2725                 pagevec_remove_exceptionals(&pvec);
2726                 pagevec_release(&pvec);
2727                 pvec.nr = PAGEVEC_SIZE;
2728                 cond_resched();
2729         }
2730         return index;
2731 }
2732
2733 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2734 {
2735         struct address_space *mapping = file->f_mapping;
2736         struct inode *inode = mapping->host;
2737         pgoff_t start, end;
2738         loff_t new_offset;
2739
2740         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2741                 return generic_file_llseek_size(file, offset, whence,
2742                                         MAX_LFS_FILESIZE, i_size_read(inode));
2743         inode_lock(inode);
2744         /* We're holding i_mutex so we can access i_size directly */
2745
2746         if (offset < 0 || offset >= inode->i_size)
2747                 offset = -ENXIO;
2748         else {
2749                 start = offset >> PAGE_SHIFT;
2750                 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2751                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2752                 new_offset <<= PAGE_SHIFT;
2753                 if (new_offset > offset) {
2754                         if (new_offset < inode->i_size)
2755                                 offset = new_offset;
2756                         else if (whence == SEEK_DATA)
2757                                 offset = -ENXIO;
2758                         else
2759                                 offset = inode->i_size;
2760                 }
2761         }
2762
2763         if (offset >= 0)
2764                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2765         inode_unlock(inode);
2766         return offset;
2767 }
2768
2769 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2770                                                          loff_t len)
2771 {
2772         struct inode *inode = file_inode(file);
2773         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2774         struct shmem_inode_info *info = SHMEM_I(inode);
2775         struct shmem_falloc shmem_falloc;
2776         pgoff_t start, index, end;
2777         int error;
2778
2779         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2780                 return -EOPNOTSUPP;
2781
2782         inode_lock(inode);
2783
2784         if (mode & FALLOC_FL_PUNCH_HOLE) {
2785                 struct address_space *mapping = file->f_mapping;
2786                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2787                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2788                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2789
2790                 /* protected by i_mutex */
2791                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2792                         error = -EPERM;
2793                         goto out;
2794                 }
2795
2796                 shmem_falloc.waitq = &shmem_falloc_waitq;
2797                 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2798                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2799                 spin_lock(&inode->i_lock);
2800                 inode->i_private = &shmem_falloc;
2801                 spin_unlock(&inode->i_lock);
2802
2803                 if ((u64)unmap_end > (u64)unmap_start)
2804                         unmap_mapping_range(mapping, unmap_start,
2805                                             1 + unmap_end - unmap_start, 0);
2806                 shmem_truncate_range(inode, offset, offset + len - 1);
2807                 /* No need to unmap again: hole-punching leaves COWed pages */
2808
2809                 spin_lock(&inode->i_lock);
2810                 inode->i_private = NULL;
2811                 wake_up_all(&shmem_falloc_waitq);
2812                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2813                 spin_unlock(&inode->i_lock);
2814                 error = 0;
2815                 goto out;
2816         }
2817
2818         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2819         error = inode_newsize_ok(inode, offset + len);
2820         if (error)
2821                 goto out;
2822
2823         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2824                 error = -EPERM;
2825                 goto out;
2826         }
2827
2828         start = offset >> PAGE_SHIFT;
2829         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2830         /* Try to avoid a swapstorm if len is impossible to satisfy */
2831         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2832                 error = -ENOSPC;
2833                 goto out;
2834         }
2835
2836         shmem_falloc.waitq = NULL;
2837         shmem_falloc.start = start;
2838         shmem_falloc.next  = start;
2839         shmem_falloc.nr_falloced = 0;
2840         shmem_falloc.nr_unswapped = 0;
2841         spin_lock(&inode->i_lock);
2842         inode->i_private = &shmem_falloc;
2843         spin_unlock(&inode->i_lock);
2844
2845         for (index = start; index < end; index++) {
2846                 struct page *page;
2847
2848                 /*
2849                  * Good, the fallocate(2) manpage permits EINTR: we may have
2850                  * been interrupted because we are using up too much memory.
2851                  */
2852                 if (signal_pending(current))
2853                         error = -EINTR;
2854                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2855                         error = -ENOMEM;
2856                 else
2857                         error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2858                 if (error) {
2859                         /* Remove the !PageUptodate pages we added */
2860                         if (index > start) {
2861                                 shmem_undo_range(inode,
2862                                     (loff_t)start << PAGE_SHIFT,
2863                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2864                         }
2865                         goto undone;
2866                 }
2867
2868                 /*
2869                  * Inform shmem_writepage() how far we have reached.
2870                  * No need for lock or barrier: we have the page lock.
2871                  */
2872                 shmem_falloc.next++;
2873                 if (!PageUptodate(page))
2874                         shmem_falloc.nr_falloced++;
2875
2876                 /*
2877                  * If !PageUptodate, leave it that way so that freeable pages
2878                  * can be recognized if we need to rollback on error later.
2879                  * But set_page_dirty so that memory pressure will swap rather
2880                  * than free the pages we are allocating (and SGP_CACHE pages
2881                  * might still be clean: we now need to mark those dirty too).
2882                  */
2883                 set_page_dirty(page);
2884                 unlock_page(page);
2885                 put_page(page);
2886                 cond_resched();
2887         }
2888
2889         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2890                 i_size_write(inode, offset + len);
2891         inode->i_ctime = current_time(inode);
2892 undone:
2893         spin_lock(&inode->i_lock);
2894         inode->i_private = NULL;
2895         spin_unlock(&inode->i_lock);
2896 out:
2897         inode_unlock(inode);
2898         return error;
2899 }
2900
2901 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2902 {
2903         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2904
2905         buf->f_type = TMPFS_MAGIC;
2906         buf->f_bsize = PAGE_SIZE;
2907         buf->f_namelen = NAME_MAX;
2908         if (sbinfo->max_blocks) {
2909                 buf->f_blocks = sbinfo->max_blocks;
2910                 buf->f_bavail =
2911                 buf->f_bfree  = sbinfo->max_blocks -
2912                                 percpu_counter_sum(&sbinfo->used_blocks);
2913         }
2914         if (sbinfo->max_inodes) {
2915                 buf->f_files = sbinfo->max_inodes;
2916                 buf->f_ffree = sbinfo->free_inodes;
2917         }
2918         /* else leave those fields 0 like simple_statfs */
2919         return 0;
2920 }
2921
2922 /*
2923  * File creation. Allocate an inode, and we're done..
2924  */
2925 static int
2926 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2927 {
2928         struct inode *inode;
2929         int error = -ENOSPC;
2930
2931         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2932         if (inode) {
2933                 error = simple_acl_create(dir, inode);
2934                 if (error)
2935                         goto out_iput;
2936                 error = security_inode_init_security(inode, dir,
2937                                                      &dentry->d_name,
2938                                                      shmem_initxattrs, NULL);
2939                 if (error && error != -EOPNOTSUPP)
2940                         goto out_iput;
2941
2942                 error = 0;
2943                 dir->i_size += BOGO_DIRENT_SIZE;
2944                 dir->i_ctime = dir->i_mtime = current_time(dir);
2945                 d_instantiate(dentry, inode);
2946                 dget(dentry); /* Extra count - pin the dentry in core */
2947         }
2948         return error;
2949 out_iput:
2950         iput(inode);
2951         return error;
2952 }
2953
2954 static int
2955 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2956 {
2957         struct inode *inode;
2958         int error = -ENOSPC;
2959
2960         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2961         if (inode) {
2962                 error = security_inode_init_security(inode, dir,
2963                                                      NULL,
2964                                                      shmem_initxattrs, NULL);
2965                 if (error && error != -EOPNOTSUPP)
2966                         goto out_iput;
2967                 error = simple_acl_create(dir, inode);
2968                 if (error)
2969                         goto out_iput;
2970                 d_tmpfile(dentry, inode);
2971         }
2972         return error;
2973 out_iput:
2974         iput(inode);
2975         return error;
2976 }
2977
2978 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2979 {
2980         int error;
2981
2982         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2983                 return error;
2984         inc_nlink(dir);
2985         return 0;
2986 }
2987
2988 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2989                 bool excl)
2990 {
2991         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2992 }
2993
2994 /*
2995  * Link a file..
2996  */
2997 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2998 {
2999         struct inode *inode = d_inode(old_dentry);
3000         int ret = 0;
3001
3002         /*
3003          * No ordinary (disk based) filesystem counts links as inodes;
3004          * but each new link needs a new dentry, pinning lowmem, and
3005          * tmpfs dentries cannot be pruned until they are unlinked.
3006          * But if an O_TMPFILE file is linked into the tmpfs, the
3007          * first link must skip that, to get the accounting right.
3008          */
3009         if (inode->i_nlink) {
3010                 ret = shmem_reserve_inode(inode->i_sb, NULL);
3011                 if (ret)
3012                         goto out;
3013         }
3014
3015         dir->i_size += BOGO_DIRENT_SIZE;
3016         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3017         inc_nlink(inode);
3018         ihold(inode);   /* New dentry reference */
3019         dget(dentry);           /* Extra pinning count for the created dentry */
3020         d_instantiate(dentry, inode);
3021 out:
3022         return ret;
3023 }
3024
3025 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3026 {
3027         struct inode *inode = d_inode(dentry);
3028
3029         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3030                 shmem_free_inode(inode->i_sb);
3031
3032         dir->i_size -= BOGO_DIRENT_SIZE;
3033         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3034         drop_nlink(inode);
3035         dput(dentry);   /* Undo the count from "create" - this does all the work */
3036         return 0;
3037 }
3038
3039 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3040 {
3041         if (!simple_empty(dentry))
3042                 return -ENOTEMPTY;
3043
3044         drop_nlink(d_inode(dentry));
3045         drop_nlink(dir);
3046         return shmem_unlink(dir, dentry);
3047 }
3048
3049 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
3050 {
3051         bool old_is_dir = d_is_dir(old_dentry);
3052         bool new_is_dir = d_is_dir(new_dentry);
3053
3054         if (old_dir != new_dir && old_is_dir != new_is_dir) {
3055                 if (old_is_dir) {
3056                         drop_nlink(old_dir);
3057                         inc_nlink(new_dir);
3058                 } else {
3059                         drop_nlink(new_dir);
3060                         inc_nlink(old_dir);
3061                 }
3062         }
3063         old_dir->i_ctime = old_dir->i_mtime =
3064         new_dir->i_ctime = new_dir->i_mtime =
3065         d_inode(old_dentry)->i_ctime =
3066         d_inode(new_dentry)->i_ctime = current_time(old_dir);
3067
3068         return 0;
3069 }
3070
3071 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3072 {
3073         struct dentry *whiteout;
3074         int error;
3075
3076         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3077         if (!whiteout)
3078                 return -ENOMEM;
3079
3080         error = shmem_mknod(old_dir, whiteout,
3081                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3082         dput(whiteout);
3083         if (error)
3084                 return error;
3085
3086         /*
3087          * Cheat and hash the whiteout while the old dentry is still in
3088          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3089          *
3090          * d_lookup() will consistently find one of them at this point,
3091          * not sure which one, but that isn't even important.
3092          */
3093         d_rehash(whiteout);
3094         return 0;
3095 }
3096
3097 /*
3098  * The VFS layer already does all the dentry stuff for rename,
3099  * we just have to decrement the usage count for the target if
3100  * it exists so that the VFS layer correctly free's it when it
3101  * gets overwritten.
3102  */
3103 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3104 {
3105         struct inode *inode = d_inode(old_dentry);
3106         int they_are_dirs = S_ISDIR(inode->i_mode);
3107
3108         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3109                 return -EINVAL;
3110
3111         if (flags & RENAME_EXCHANGE)
3112                 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3113
3114         if (!simple_empty(new_dentry))
3115                 return -ENOTEMPTY;
3116
3117         if (flags & RENAME_WHITEOUT) {
3118                 int error;
3119
3120                 error = shmem_whiteout(old_dir, old_dentry);
3121                 if (error)
3122                         return error;
3123         }
3124
3125         if (d_really_is_positive(new_dentry)) {
3126                 (void) shmem_unlink(new_dir, new_dentry);
3127                 if (they_are_dirs) {
3128                         drop_nlink(d_inode(new_dentry));
3129                         drop_nlink(old_dir);
3130                 }
3131         } else if (they_are_dirs) {
3132                 drop_nlink(old_dir);
3133                 inc_nlink(new_dir);
3134         }
3135
3136         old_dir->i_size -= BOGO_DIRENT_SIZE;
3137         new_dir->i_size += BOGO_DIRENT_SIZE;
3138         old_dir->i_ctime = old_dir->i_mtime =
3139         new_dir->i_ctime = new_dir->i_mtime =
3140         inode->i_ctime = current_time(old_dir);
3141         return 0;
3142 }
3143
3144 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3145 {
3146         int error;
3147         int len;
3148         struct inode *inode;
3149         struct page *page;
3150
3151         len = strlen(symname) + 1;
3152         if (len > PAGE_SIZE)
3153                 return -ENAMETOOLONG;
3154
3155         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3156                                 VM_NORESERVE);
3157         if (!inode)
3158                 return -ENOSPC;
3159
3160         error = security_inode_init_security(inode, dir, &dentry->d_name,
3161                                              shmem_initxattrs, NULL);
3162         if (error && error != -EOPNOTSUPP) {
3163                 iput(inode);
3164                 return error;
3165         }
3166
3167         inode->i_size = len-1;
3168         if (len <= SHORT_SYMLINK_LEN) {
3169                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3170                 if (!inode->i_link) {
3171                         iput(inode);
3172                         return -ENOMEM;
3173                 }
3174                 inode->i_op = &shmem_short_symlink_operations;
3175         } else {
3176                 inode_nohighmem(inode);
3177                 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3178                 if (error) {
3179                         iput(inode);
3180                         return error;
3181                 }
3182                 inode->i_mapping->a_ops = &shmem_aops;
3183                 inode->i_op = &shmem_symlink_inode_operations;
3184                 memcpy(page_address(page), symname, len);
3185                 SetPageUptodate(page);
3186                 set_page_dirty(page);
3187                 unlock_page(page);
3188                 put_page(page);
3189         }
3190         dir->i_size += BOGO_DIRENT_SIZE;
3191         dir->i_ctime = dir->i_mtime = current_time(dir);
3192         d_instantiate(dentry, inode);
3193         dget(dentry);
3194         return 0;
3195 }
3196
3197 static void shmem_put_link(void *arg)
3198 {
3199         mark_page_accessed(arg);
3200         put_page(arg);
3201 }
3202
3203 static const char *shmem_get_link(struct dentry *dentry,
3204                                   struct inode *inode,
3205                                   struct delayed_call *done)
3206 {
3207         struct page *page = NULL;
3208         int error;
3209         if (!dentry) {
3210                 page = find_get_page(inode->i_mapping, 0);
3211                 if (!page)
3212                         return ERR_PTR(-ECHILD);
3213                 if (!PageUptodate(page)) {
3214                         put_page(page);
3215                         return ERR_PTR(-ECHILD);
3216                 }
3217         } else {
3218                 error = shmem_getpage(inode, 0, &page, SGP_READ);
3219                 if (error)
3220                         return ERR_PTR(error);
3221                 unlock_page(page);
3222         }
3223         set_delayed_call(done, shmem_put_link, page);
3224         return page_address(page);
3225 }
3226
3227 #ifdef CONFIG_TMPFS_XATTR
3228 /*
3229  * Superblocks without xattr inode operations may get some security.* xattr
3230  * support from the LSM "for free". As soon as we have any other xattrs
3231  * like ACLs, we also need to implement the security.* handlers at
3232  * filesystem level, though.
3233  */
3234
3235 /*
3236  * Callback for security_inode_init_security() for acquiring xattrs.
3237  */
3238 static int shmem_initxattrs(struct inode *inode,
3239                             const struct xattr *xattr_array,
3240                             void *fs_info)
3241 {
3242         struct shmem_inode_info *info = SHMEM_I(inode);
3243         const struct xattr *xattr;
3244         struct simple_xattr *new_xattr;
3245         size_t len;
3246
3247         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3248                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3249                 if (!new_xattr)
3250                         return -ENOMEM;
3251
3252                 len = strlen(xattr->name) + 1;
3253                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3254                                           GFP_KERNEL);
3255                 if (!new_xattr->name) {
3256                         kvfree(new_xattr);
3257                         return -ENOMEM;
3258                 }
3259
3260                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3261                        XATTR_SECURITY_PREFIX_LEN);
3262                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3263                        xattr->name, len);
3264
3265                 simple_xattr_list_add(&info->xattrs, new_xattr);
3266         }
3267
3268         return 0;
3269 }
3270
3271 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3272                                    struct dentry *unused, struct inode *inode,
3273                                    const char *name, void *buffer, size_t size)
3274 {
3275         struct shmem_inode_info *info = SHMEM_I(inode);
3276
3277         name = xattr_full_name(handler, name);
3278         return simple_xattr_get(&info->xattrs, name, buffer, size);
3279 }
3280
3281 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3282                                    struct dentry *unused, struct inode *inode,
3283                                    const char *name, const void *value,
3284                                    size_t size, int flags)
3285 {
3286         struct shmem_inode_info *info = SHMEM_I(inode);
3287
3288         name = xattr_full_name(handler, name);
3289         return simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3290 }
3291
3292 static const struct xattr_handler shmem_security_xattr_handler = {
3293         .prefix = XATTR_SECURITY_PREFIX,
3294         .get = shmem_xattr_handler_get,
3295         .set = shmem_xattr_handler_set,
3296 };
3297
3298 static const struct xattr_handler shmem_trusted_xattr_handler = {
3299         .prefix = XATTR_TRUSTED_PREFIX,
3300         .get = shmem_xattr_handler_get,
3301         .set = shmem_xattr_handler_set,
3302 };
3303
3304 static const struct xattr_handler *shmem_xattr_handlers[] = {
3305 #ifdef CONFIG_TMPFS_POSIX_ACL
3306         &posix_acl_access_xattr_handler,
3307         &posix_acl_default_xattr_handler,
3308 #endif
3309         &shmem_security_xattr_handler,
3310         &shmem_trusted_xattr_handler,
3311         NULL
3312 };
3313
3314 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3315 {
3316         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3317         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3318 }
3319 #endif /* CONFIG_TMPFS_XATTR */
3320
3321 static const struct inode_operations shmem_short_symlink_operations = {
3322         .get_link       = simple_get_link,
3323 #ifdef CONFIG_TMPFS_XATTR
3324         .listxattr      = shmem_listxattr,
3325 #endif
3326 };
3327
3328 static const struct inode_operations shmem_symlink_inode_operations = {
3329         .get_link       = shmem_get_link,
3330 #ifdef CONFIG_TMPFS_XATTR
3331         .listxattr      = shmem_listxattr,
3332 #endif
3333 };
3334
3335 static struct dentry *shmem_get_parent(struct dentry *child)
3336 {
3337         return ERR_PTR(-ESTALE);
3338 }
3339
3340 static int shmem_match(struct inode *ino, void *vfh)
3341 {
3342         __u32 *fh = vfh;
3343         __u64 inum = fh[2];
3344         inum = (inum << 32) | fh[1];
3345         return ino->i_ino == inum && fh[0] == ino->i_generation;
3346 }
3347
3348 /* Find any alias of inode, but prefer a hashed alias */
3349 static struct dentry *shmem_find_alias(struct inode *inode)
3350 {
3351         struct dentry *alias = d_find_alias(inode);
3352
3353         return alias ?: d_find_any_alias(inode);
3354 }
3355
3356
3357 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3358                 struct fid *fid, int fh_len, int fh_type)
3359 {
3360         struct inode *inode;
3361         struct dentry *dentry = NULL;
3362         u64 inum;
3363
3364         if (fh_len < 3)
3365                 return NULL;
3366
3367         inum = fid->raw[2];
3368         inum = (inum << 32) | fid->raw[1];
3369
3370         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3371                         shmem_match, fid->raw);
3372         if (inode) {
3373                 dentry = shmem_find_alias(inode);
3374                 iput(inode);
3375         }
3376
3377         return dentry;
3378 }
3379
3380 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3381                                 struct inode *parent)
3382 {
3383         if (*len < 3) {
3384                 *len = 3;
3385                 return FILEID_INVALID;
3386         }
3387
3388         if (inode_unhashed(inode)) {
3389                 /* Unfortunately insert_inode_hash is not idempotent,
3390                  * so as we hash inodes here rather than at creation
3391                  * time, we need a lock to ensure we only try
3392                  * to do it once
3393                  */
3394                 static DEFINE_SPINLOCK(lock);
3395                 spin_lock(&lock);
3396                 if (inode_unhashed(inode))
3397                         __insert_inode_hash(inode,
3398                                             inode->i_ino + inode->i_generation);
3399                 spin_unlock(&lock);
3400         }
3401
3402         fh[0] = inode->i_generation;
3403         fh[1] = inode->i_ino;
3404         fh[2] = ((__u64)inode->i_ino) >> 32;
3405
3406         *len = 3;
3407         return 1;
3408 }
3409
3410 static const struct export_operations shmem_export_ops = {
3411         .get_parent     = shmem_get_parent,
3412         .encode_fh      = shmem_encode_fh,
3413         .fh_to_dentry   = shmem_fh_to_dentry,
3414 };
3415
3416 enum shmem_param {
3417         Opt_gid,
3418         Opt_huge,
3419         Opt_mode,
3420         Opt_mpol,
3421         Opt_nr_blocks,
3422         Opt_nr_inodes,
3423         Opt_size,
3424         Opt_uid,
3425         Opt_inode32,
3426         Opt_inode64,
3427 };
3428
3429 static const struct constant_table shmem_param_enums_huge[] = {
3430         {"never",       SHMEM_HUGE_NEVER },
3431         {"always",      SHMEM_HUGE_ALWAYS },
3432         {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3433         {"advise",      SHMEM_HUGE_ADVISE },
3434         {}
3435 };
3436
3437 const struct fs_parameter_spec shmem_fs_parameters[] = {
3438         fsparam_u32   ("gid",           Opt_gid),
3439         fsparam_enum  ("huge",          Opt_huge,  shmem_param_enums_huge),
3440         fsparam_u32oct("mode",          Opt_mode),
3441         fsparam_string("mpol",          Opt_mpol),
3442         fsparam_string("nr_blocks",     Opt_nr_blocks),
3443         fsparam_string("nr_inodes",     Opt_nr_inodes),
3444         fsparam_string("size",          Opt_size),
3445         fsparam_u32   ("uid",           Opt_uid),
3446         fsparam_flag  ("inode32",       Opt_inode32),
3447         fsparam_flag  ("inode64",       Opt_inode64),
3448         {}
3449 };
3450
3451 static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3452 {
3453         struct shmem_options *ctx = fc->fs_private;
3454         struct fs_parse_result result;
3455         unsigned long long size;
3456         char *rest;
3457         int opt;
3458         kuid_t kuid;
3459         kgid_t kgid;
3460
3461         opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3462         if (opt < 0)
3463                 return opt;
3464
3465         switch (opt) {
3466         case Opt_size:
3467                 size = memparse(param->string, &rest);
3468                 if (*rest == '%') {
3469                         size <<= PAGE_SHIFT;
3470                         size *= totalram_pages();
3471                         do_div(size, 100);
3472                         rest++;
3473                 }
3474                 if (*rest)
3475                         goto bad_value;
3476                 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3477                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3478                 break;
3479         case Opt_nr_blocks:
3480                 ctx->blocks = memparse(param->string, &rest);
3481                 if (*rest)
3482                         goto bad_value;
3483                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3484                 break;
3485         case Opt_nr_inodes:
3486                 ctx->inodes = memparse(param->string, &rest);
3487                 if (*rest)
3488                         goto bad_value;
3489                 ctx->seen |= SHMEM_SEEN_INODES;
3490                 break;
3491         case Opt_mode:
3492                 ctx->mode = result.uint_32 & 07777;
3493                 break;
3494         case Opt_uid:
3495                 kuid = make_kuid(current_user_ns(), result.uint_32);
3496                 if (!uid_valid(kuid))
3497                         goto bad_value;
3498
3499                 /*
3500                  * The requested uid must be representable in the
3501                  * filesystem's idmapping.
3502                  */
3503                 if (!kuid_has_mapping(fc->user_ns, kuid))
3504                         goto bad_value;
3505
3506                 ctx->uid = kuid;
3507                 break;
3508         case Opt_gid:
3509                 kgid = make_kgid(current_user_ns(), result.uint_32);
3510                 if (!gid_valid(kgid))
3511                         goto bad_value;
3512
3513                 /*
3514                  * The requested gid must be representable in the
3515                  * filesystem's idmapping.
3516                  */
3517                 if (!kgid_has_mapping(fc->user_ns, kgid))
3518                         goto bad_value;
3519
3520                 ctx->gid = kgid;
3521                 break;
3522         case Opt_huge:
3523                 ctx->huge = result.uint_32;
3524                 if (ctx->huge != SHMEM_HUGE_NEVER &&
3525                     !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3526                       has_transparent_hugepage()))
3527                         goto unsupported_parameter;
3528                 ctx->seen |= SHMEM_SEEN_HUGE;
3529                 break;
3530         case Opt_mpol:
3531                 if (IS_ENABLED(CONFIG_NUMA)) {
3532                         mpol_put(ctx->mpol);
3533                         ctx->mpol = NULL;
3534                         if (mpol_parse_str(param->string, &ctx->mpol))
3535                                 goto bad_value;
3536                         break;
3537                 }
3538                 goto unsupported_parameter;
3539         case Opt_inode32:
3540                 ctx->full_inums = false;
3541                 ctx->seen |= SHMEM_SEEN_INUMS;
3542                 break;
3543         case Opt_inode64:
3544                 if (sizeof(ino_t) < 8) {
3545                         return invalfc(fc,
3546                                        "Cannot use inode64 with <64bit inums in kernel\n");
3547                 }
3548                 ctx->full_inums = true;
3549                 ctx->seen |= SHMEM_SEEN_INUMS;
3550                 break;
3551         }
3552         return 0;
3553
3554 unsupported_parameter:
3555         return invalfc(fc, "Unsupported parameter '%s'", param->key);
3556 bad_value:
3557         return invalfc(fc, "Bad value for '%s'", param->key);
3558 }
3559
3560 static int shmem_parse_options(struct fs_context *fc, void *data)
3561 {
3562         char *options = data;
3563
3564         if (options) {
3565                 int err = security_sb_eat_lsm_opts(options, &fc->security);
3566                 if (err)
3567                         return err;
3568         }
3569
3570         while (options != NULL) {
3571                 char *this_char = options;
3572                 for (;;) {
3573                         /*
3574                          * NUL-terminate this option: unfortunately,
3575                          * mount options form a comma-separated list,
3576                          * but mpol's nodelist may also contain commas.
3577                          */
3578                         options = strchr(options, ',');
3579                         if (options == NULL)
3580                                 break;
3581                         options++;
3582                         if (!isdigit(*options)) {
3583                                 options[-1] = '\0';
3584                                 break;
3585                         }
3586                 }
3587                 if (*this_char) {
3588                         char *value = strchr(this_char,'=');
3589                         size_t len = 0;
3590                         int err;
3591
3592                         if (value) {
3593                                 *value++ = '\0';
3594                                 len = strlen(value);
3595                         }
3596                         err = vfs_parse_fs_string(fc, this_char, value, len);
3597                         if (err < 0)
3598                                 return err;
3599                 }
3600         }
3601         return 0;
3602 }
3603
3604 /*
3605  * Reconfigure a shmem filesystem.
3606  *
3607  * Note that we disallow change from limited->unlimited blocks/inodes while any
3608  * are in use; but we must separately disallow unlimited->limited, because in
3609  * that case we have no record of how much is already in use.
3610  */
3611 static int shmem_reconfigure(struct fs_context *fc)
3612 {
3613         struct shmem_options *ctx = fc->fs_private;
3614         struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3615         unsigned long inodes;
3616         const char *err;
3617
3618         spin_lock(&sbinfo->stat_lock);
3619         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3620         if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3621                 if (!sbinfo->max_blocks) {
3622                         err = "Cannot retroactively limit size";
3623                         goto out;
3624                 }
3625                 if (percpu_counter_compare(&sbinfo->used_blocks,
3626                                            ctx->blocks) > 0) {
3627                         err = "Too small a size for current use";
3628                         goto out;
3629                 }
3630         }
3631         if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3632                 if (!sbinfo->max_inodes) {
3633                         err = "Cannot retroactively limit inodes";
3634                         goto out;
3635                 }
3636                 if (ctx->inodes < inodes) {
3637                         err = "Too few inodes for current use";
3638                         goto out;
3639                 }
3640         }
3641
3642         if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3643             sbinfo->next_ino > UINT_MAX) {
3644                 err = "Current inum too high to switch to 32-bit inums";
3645                 goto out;
3646         }
3647
3648         if (ctx->seen & SHMEM_SEEN_HUGE)
3649                 sbinfo->huge = ctx->huge;
3650         if (ctx->seen & SHMEM_SEEN_INUMS)
3651                 sbinfo->full_inums = ctx->full_inums;
3652         if (ctx->seen & SHMEM_SEEN_BLOCKS)
3653                 sbinfo->max_blocks  = ctx->blocks;
3654         if (ctx->seen & SHMEM_SEEN_INODES) {
3655                 sbinfo->max_inodes  = ctx->inodes;
3656                 sbinfo->free_inodes = ctx->inodes - inodes;
3657         }
3658
3659         /*
3660          * Preserve previous mempolicy unless mpol remount option was specified.
3661          */
3662         if (ctx->mpol) {
3663                 mpol_put(sbinfo->mpol);
3664                 sbinfo->mpol = ctx->mpol;       /* transfers initial ref */
3665                 ctx->mpol = NULL;
3666         }
3667         spin_unlock(&sbinfo->stat_lock);
3668         return 0;
3669 out:
3670         spin_unlock(&sbinfo->stat_lock);
3671         return invalfc(fc, "%s", err);
3672 }
3673
3674 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3675 {
3676         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3677
3678         if (sbinfo->max_blocks != shmem_default_max_blocks())
3679                 seq_printf(seq, ",size=%luk",
3680                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3681         if (sbinfo->max_inodes != shmem_default_max_inodes())
3682                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3683         if (sbinfo->mode != (0777 | S_ISVTX))
3684                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3685         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3686                 seq_printf(seq, ",uid=%u",
3687                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3688         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3689                 seq_printf(seq, ",gid=%u",
3690                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3691
3692         /*
3693          * Showing inode{64,32} might be useful even if it's the system default,
3694          * since then people don't have to resort to checking both here and
3695          * /proc/config.gz to confirm 64-bit inums were successfully applied
3696          * (which may not even exist if IKCONFIG_PROC isn't enabled).
3697          *
3698          * We hide it when inode64 isn't the default and we are using 32-bit
3699          * inodes, since that probably just means the feature isn't even under
3700          * consideration.
3701          *
3702          * As such:
3703          *
3704          *                     +-----------------+-----------------+
3705          *                     | TMPFS_INODE64=y | TMPFS_INODE64=n |
3706          *  +------------------+-----------------+-----------------+
3707          *  | full_inums=true  | show            | show            |
3708          *  | full_inums=false | show            | hide            |
3709          *  +------------------+-----------------+-----------------+
3710          *
3711          */
3712         if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3713                 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3714 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3715         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3716         if (sbinfo->huge)
3717                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3718 #endif
3719         shmem_show_mpol(seq, sbinfo->mpol);
3720         return 0;
3721 }
3722
3723 #endif /* CONFIG_TMPFS */
3724
3725 static void shmem_put_super(struct super_block *sb)
3726 {
3727         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3728
3729         free_percpu(sbinfo->ino_batch);
3730         percpu_counter_destroy(&sbinfo->used_blocks);
3731         mpol_put(sbinfo->mpol);
3732         kfree(sbinfo);
3733         sb->s_fs_info = NULL;
3734 }
3735
3736 static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3737 {
3738         struct shmem_options *ctx = fc->fs_private;
3739         struct inode *inode;
3740         struct shmem_sb_info *sbinfo;
3741         int err = -ENOMEM;
3742
3743         /* Round up to L1_CACHE_BYTES to resist false sharing */
3744         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3745                                 L1_CACHE_BYTES), GFP_KERNEL);
3746         if (!sbinfo)
3747                 return -ENOMEM;
3748
3749         sb->s_fs_info = sbinfo;
3750
3751 #ifdef CONFIG_TMPFS
3752         /*
3753          * Per default we only allow half of the physical ram per
3754          * tmpfs instance, limiting inodes to one per page of lowmem;
3755          * but the internal instance is left unlimited.
3756          */
3757         if (!(sb->s_flags & SB_KERNMOUNT)) {
3758                 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3759                         ctx->blocks = shmem_default_max_blocks();
3760                 if (!(ctx->seen & SHMEM_SEEN_INODES))
3761                         ctx->inodes = shmem_default_max_inodes();
3762                 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3763                         ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3764         } else {
3765                 sb->s_flags |= SB_NOUSER;
3766         }
3767         sb->s_export_op = &shmem_export_ops;
3768         sb->s_flags |= SB_NOSEC;
3769 #else
3770         sb->s_flags |= SB_NOUSER;
3771 #endif
3772         sbinfo->max_blocks = ctx->blocks;
3773         sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3774         if (sb->s_flags & SB_KERNMOUNT) {
3775                 sbinfo->ino_batch = alloc_percpu(ino_t);
3776                 if (!sbinfo->ino_batch)
3777                         goto failed;
3778         }
3779         sbinfo->uid = ctx->uid;
3780         sbinfo->gid = ctx->gid;
3781         sbinfo->full_inums = ctx->full_inums;
3782         sbinfo->mode = ctx->mode;
3783         sbinfo->huge = ctx->huge;
3784         sbinfo->mpol = ctx->mpol;
3785         ctx->mpol = NULL;
3786
3787         spin_lock_init(&sbinfo->stat_lock);
3788         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3789                 goto failed;
3790         spin_lock_init(&sbinfo->shrinklist_lock);
3791         INIT_LIST_HEAD(&sbinfo->shrinklist);
3792
3793         sb->s_maxbytes = MAX_LFS_FILESIZE;
3794         sb->s_blocksize = PAGE_SIZE;
3795         sb->s_blocksize_bits = PAGE_SHIFT;
3796         sb->s_magic = TMPFS_MAGIC;
3797         sb->s_op = &shmem_ops;
3798         sb->s_time_gran = 1;
3799 #ifdef CONFIG_TMPFS_XATTR
3800         sb->s_xattr = shmem_xattr_handlers;
3801 #endif
3802 #ifdef CONFIG_TMPFS_POSIX_ACL
3803         sb->s_flags |= SB_POSIXACL;
3804 #endif
3805         uuid_gen(&sb->s_uuid);
3806
3807         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3808         if (!inode)
3809                 goto failed;
3810         inode->i_uid = sbinfo->uid;
3811         inode->i_gid = sbinfo->gid;
3812         sb->s_root = d_make_root(inode);
3813         if (!sb->s_root)
3814                 goto failed;
3815         return 0;
3816
3817 failed:
3818         shmem_put_super(sb);
3819         return err;
3820 }
3821
3822 static int shmem_get_tree(struct fs_context *fc)
3823 {
3824         return get_tree_nodev(fc, shmem_fill_super);
3825 }
3826
3827 static void shmem_free_fc(struct fs_context *fc)
3828 {
3829         struct shmem_options *ctx = fc->fs_private;
3830
3831         if (ctx) {
3832                 mpol_put(ctx->mpol);
3833                 kfree(ctx);
3834         }
3835 }
3836
3837 static const struct fs_context_operations shmem_fs_context_ops = {
3838         .free                   = shmem_free_fc,
3839         .get_tree               = shmem_get_tree,
3840 #ifdef CONFIG_TMPFS
3841         .parse_monolithic       = shmem_parse_options,
3842         .parse_param            = shmem_parse_one,
3843         .reconfigure            = shmem_reconfigure,
3844 #endif
3845 };
3846
3847 static struct kmem_cache *shmem_inode_cachep;
3848
3849 static struct inode *shmem_alloc_inode(struct super_block *sb)
3850 {
3851         struct shmem_inode_info *info;
3852         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3853         if (!info)
3854                 return NULL;
3855         return &info->vfs_inode;
3856 }
3857
3858 static void shmem_free_in_core_inode(struct inode *inode)
3859 {
3860         if (S_ISLNK(inode->i_mode))
3861                 kfree(inode->i_link);
3862         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3863 }
3864
3865 static void shmem_destroy_inode(struct inode *inode)
3866 {
3867         if (S_ISREG(inode->i_mode))
3868                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3869 }
3870
3871 static void shmem_init_inode(void *foo)
3872 {
3873         struct shmem_inode_info *info = foo;
3874         inode_init_once(&info->vfs_inode);
3875 }
3876
3877 static void shmem_init_inodecache(void)
3878 {
3879         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3880                                 sizeof(struct shmem_inode_info),
3881                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3882 }
3883
3884 static void shmem_destroy_inodecache(void)
3885 {
3886         kmem_cache_destroy(shmem_inode_cachep);
3887 }
3888
3889 static const struct address_space_operations shmem_aops = {
3890         .writepage      = shmem_writepage,
3891         .set_page_dirty = __set_page_dirty_no_writeback,
3892 #ifdef CONFIG_TMPFS
3893         .write_begin    = shmem_write_begin,
3894         .write_end      = shmem_write_end,
3895 #endif
3896 #ifdef CONFIG_MIGRATION
3897         .migratepage    = migrate_page,
3898 #endif
3899         .error_remove_page = generic_error_remove_page,
3900 };
3901
3902 static const struct file_operations shmem_file_operations = {
3903         .mmap           = shmem_mmap,
3904         .get_unmapped_area = shmem_get_unmapped_area,
3905 #ifdef CONFIG_TMPFS
3906         .llseek         = shmem_file_llseek,
3907         .read_iter      = shmem_file_read_iter,
3908         .write_iter     = generic_file_write_iter,
3909         .fsync          = noop_fsync,
3910         .splice_read    = generic_file_splice_read,
3911         .splice_write   = iter_file_splice_write,
3912         .fallocate      = shmem_fallocate,
3913 #endif
3914 };
3915
3916 static const struct inode_operations shmem_inode_operations = {
3917         .getattr        = shmem_getattr,
3918         .setattr        = shmem_setattr,
3919 #ifdef CONFIG_TMPFS_XATTR
3920         .listxattr      = shmem_listxattr,
3921         .set_acl        = simple_set_acl,
3922 #endif
3923 };
3924
3925 static const struct inode_operations shmem_dir_inode_operations = {
3926 #ifdef CONFIG_TMPFS
3927         .create         = shmem_create,
3928         .lookup         = simple_lookup,
3929         .link           = shmem_link,
3930         .unlink         = shmem_unlink,
3931         .symlink        = shmem_symlink,
3932         .mkdir          = shmem_mkdir,
3933         .rmdir          = shmem_rmdir,
3934         .mknod          = shmem_mknod,
3935         .rename         = shmem_rename2,
3936         .tmpfile        = shmem_tmpfile,
3937 #endif
3938 #ifdef CONFIG_TMPFS_XATTR
3939         .listxattr      = shmem_listxattr,
3940 #endif
3941 #ifdef CONFIG_TMPFS_POSIX_ACL
3942         .setattr        = shmem_setattr,
3943         .set_acl        = simple_set_acl,
3944 #endif
3945 };
3946
3947 static const struct inode_operations shmem_special_inode_operations = {
3948 #ifdef CONFIG_TMPFS_XATTR
3949         .listxattr      = shmem_listxattr,
3950 #endif
3951 #ifdef CONFIG_TMPFS_POSIX_ACL
3952         .setattr        = shmem_setattr,
3953         .set_acl        = simple_set_acl,
3954 #endif
3955 };
3956
3957 static const struct super_operations shmem_ops = {
3958         .alloc_inode    = shmem_alloc_inode,
3959         .free_inode     = shmem_free_in_core_inode,
3960         .destroy_inode  = shmem_destroy_inode,
3961 #ifdef CONFIG_TMPFS
3962         .statfs         = shmem_statfs,
3963         .show_options   = shmem_show_options,
3964 #endif
3965         .evict_inode    = shmem_evict_inode,
3966         .drop_inode     = generic_delete_inode,
3967         .put_super      = shmem_put_super,
3968 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3969         .nr_cached_objects      = shmem_unused_huge_count,
3970         .free_cached_objects    = shmem_unused_huge_scan,
3971 #endif
3972 };
3973
3974 static const struct vm_operations_struct shmem_vm_ops = {
3975         .fault          = shmem_fault,
3976         .map_pages      = filemap_map_pages,
3977 #ifdef CONFIG_NUMA
3978         .set_policy     = shmem_set_policy,
3979         .get_policy     = shmem_get_policy,
3980 #endif
3981 };
3982
3983 int shmem_init_fs_context(struct fs_context *fc)
3984 {
3985         struct shmem_options *ctx;
3986
3987         ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
3988         if (!ctx)
3989                 return -ENOMEM;
3990
3991         ctx->mode = 0777 | S_ISVTX;
3992         ctx->uid = current_fsuid();
3993         ctx->gid = current_fsgid();
3994
3995         fc->fs_private = ctx;
3996         fc->ops = &shmem_fs_context_ops;
3997         return 0;
3998 }
3999
4000 static struct file_system_type shmem_fs_type = {
4001         .owner          = THIS_MODULE,
4002         .name           = "tmpfs",
4003         .init_fs_context = shmem_init_fs_context,
4004 #ifdef CONFIG_TMPFS
4005         .parameters     = shmem_fs_parameters,
4006 #endif
4007         .kill_sb        = kill_litter_super,
4008         .fs_flags       = FS_USERNS_MOUNT | FS_THP_SUPPORT,
4009 };
4010
4011 int __init shmem_init(void)
4012 {
4013         int error;
4014
4015         shmem_init_inodecache();
4016
4017         error = register_filesystem(&shmem_fs_type);
4018         if (error) {
4019                 pr_err("Could not register tmpfs\n");
4020                 goto out2;
4021         }
4022
4023         shm_mnt = kern_mount(&shmem_fs_type);
4024         if (IS_ERR(shm_mnt)) {
4025                 error = PTR_ERR(shm_mnt);
4026                 pr_err("Could not kern_mount tmpfs\n");
4027                 goto out1;
4028         }
4029
4030 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4031         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4032                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4033         else
4034                 shmem_huge = 0; /* just in case it was patched */
4035 #endif
4036         return 0;
4037
4038 out1:
4039         unregister_filesystem(&shmem_fs_type);
4040 out2:
4041         shmem_destroy_inodecache();
4042         shm_mnt = ERR_PTR(error);
4043         return error;
4044 }
4045
4046 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4047 static ssize_t shmem_enabled_show(struct kobject *kobj,
4048                 struct kobj_attribute *attr, char *buf)
4049 {
4050         static const int values[] = {
4051                 SHMEM_HUGE_ALWAYS,
4052                 SHMEM_HUGE_WITHIN_SIZE,
4053                 SHMEM_HUGE_ADVISE,
4054                 SHMEM_HUGE_NEVER,
4055                 SHMEM_HUGE_DENY,
4056                 SHMEM_HUGE_FORCE,
4057         };
4058         int i, count;
4059
4060         for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
4061                 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
4062
4063                 count += sprintf(buf + count, fmt,
4064                                 shmem_format_huge(values[i]));
4065         }
4066         buf[count - 1] = '\n';
4067         return count;
4068 }
4069
4070 static ssize_t shmem_enabled_store(struct kobject *kobj,
4071                 struct kobj_attribute *attr, const char *buf, size_t count)
4072 {
4073         char tmp[16];
4074         int huge;
4075
4076         if (count + 1 > sizeof(tmp))
4077                 return -EINVAL;
4078         memcpy(tmp, buf, count);
4079         tmp[count] = '\0';
4080         if (count && tmp[count - 1] == '\n')
4081                 tmp[count - 1] = '\0';
4082
4083         huge = shmem_parse_huge(tmp);
4084         if (huge == -EINVAL)
4085                 return -EINVAL;
4086         if (!has_transparent_hugepage() &&
4087                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4088                 return -EINVAL;
4089
4090         shmem_huge = huge;
4091         if (shmem_huge > SHMEM_HUGE_DENY)
4092                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4093         return count;
4094 }
4095
4096 struct kobj_attribute shmem_enabled_attr =
4097         __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
4098 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4099
4100 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4101 bool shmem_huge_enabled(struct vm_area_struct *vma)
4102 {
4103         struct inode *inode = file_inode(vma->vm_file);
4104         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
4105         loff_t i_size;
4106         pgoff_t off;
4107
4108         if (!transhuge_vma_enabled(vma, vma->vm_flags))
4109                 return false;
4110         if (shmem_huge == SHMEM_HUGE_FORCE)
4111                 return true;
4112         if (shmem_huge == SHMEM_HUGE_DENY)
4113                 return false;
4114         switch (sbinfo->huge) {
4115                 case SHMEM_HUGE_NEVER:
4116                         return false;
4117                 case SHMEM_HUGE_ALWAYS:
4118                         return true;
4119                 case SHMEM_HUGE_WITHIN_SIZE:
4120                         off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
4121                         i_size = round_up(i_size_read(inode), PAGE_SIZE);
4122                         if (i_size >= HPAGE_PMD_SIZE &&
4123                                         i_size >> PAGE_SHIFT >= off)
4124                                 return true;
4125                         fallthrough;
4126                 case SHMEM_HUGE_ADVISE:
4127                         /* TODO: implement fadvise() hints */
4128                         return (vma->vm_flags & VM_HUGEPAGE);
4129                 default:
4130                         VM_BUG_ON(1);
4131                         return false;
4132         }
4133 }
4134 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4135
4136 #else /* !CONFIG_SHMEM */
4137
4138 /*
4139  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4140  *
4141  * This is intended for small system where the benefits of the full
4142  * shmem code (swap-backed and resource-limited) are outweighed by
4143  * their complexity. On systems without swap this code should be
4144  * effectively equivalent, but much lighter weight.
4145  */
4146
4147 static struct file_system_type shmem_fs_type = {
4148         .name           = "tmpfs",
4149         .init_fs_context = ramfs_init_fs_context,
4150         .parameters     = ramfs_fs_parameters,
4151         .kill_sb        = ramfs_kill_sb,
4152         .fs_flags       = FS_USERNS_MOUNT,
4153 };
4154
4155 int __init shmem_init(void)
4156 {
4157         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4158
4159         shm_mnt = kern_mount(&shmem_fs_type);
4160         BUG_ON(IS_ERR(shm_mnt));
4161
4162         return 0;
4163 }
4164
4165 int shmem_unuse(unsigned int type, bool frontswap,
4166                 unsigned long *fs_pages_to_unuse)
4167 {
4168         return 0;
4169 }
4170
4171 int shmem_lock(struct file *file, int lock, struct user_struct *user)
4172 {
4173         return 0;
4174 }
4175
4176 void shmem_unlock_mapping(struct address_space *mapping)
4177 {
4178 }
4179
4180 #ifdef CONFIG_MMU
4181 unsigned long shmem_get_unmapped_area(struct file *file,
4182                                       unsigned long addr, unsigned long len,
4183                                       unsigned long pgoff, unsigned long flags)
4184 {
4185         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4186 }
4187 #endif
4188
4189 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4190 {
4191         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4192 }
4193 EXPORT_SYMBOL_GPL(shmem_truncate_range);
4194
4195 #define shmem_vm_ops                            generic_file_vm_ops
4196 #define shmem_file_operations                   ramfs_file_operations
4197 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
4198 #define shmem_acct_size(flags, size)            0
4199 #define shmem_unacct_size(flags, size)          do {} while (0)
4200
4201 #endif /* CONFIG_SHMEM */
4202
4203 /* common code */
4204
4205 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4206                                        unsigned long flags, unsigned int i_flags)
4207 {
4208         struct inode *inode;
4209         struct file *res;
4210
4211         if (IS_ERR(mnt))
4212                 return ERR_CAST(mnt);
4213
4214         if (size < 0 || size > MAX_LFS_FILESIZE)
4215                 return ERR_PTR(-EINVAL);
4216
4217         if (shmem_acct_size(flags, size))
4218                 return ERR_PTR(-ENOMEM);
4219
4220         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4221                                 flags);
4222         if (unlikely(!inode)) {
4223                 shmem_unacct_size(flags, size);
4224                 return ERR_PTR(-ENOSPC);
4225         }
4226         inode->i_flags |= i_flags;
4227         inode->i_size = size;
4228         clear_nlink(inode);     /* It is unlinked */
4229         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4230         if (!IS_ERR(res))
4231                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4232                                 &shmem_file_operations);
4233         if (IS_ERR(res))
4234                 iput(inode);
4235         return res;
4236 }
4237
4238 /**
4239  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4240  *      kernel internal.  There will be NO LSM permission checks against the
4241  *      underlying inode.  So users of this interface must do LSM checks at a
4242  *      higher layer.  The users are the big_key and shm implementations.  LSM
4243  *      checks are provided at the key or shm level rather than the inode.
4244  * @name: name for dentry (to be seen in /proc/<pid>/maps
4245  * @size: size to be set for the file
4246  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4247  */
4248 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4249 {
4250         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4251 }
4252
4253 /**
4254  * shmem_file_setup - get an unlinked file living in tmpfs
4255  * @name: name for dentry (to be seen in /proc/<pid>/maps
4256  * @size: size to be set for the file
4257  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4258  */
4259 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4260 {
4261         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4262 }
4263 EXPORT_SYMBOL_GPL(shmem_file_setup);
4264
4265 /**
4266  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4267  * @mnt: the tmpfs mount where the file will be created
4268  * @name: name for dentry (to be seen in /proc/<pid>/maps
4269  * @size: size to be set for the file
4270  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4271  */
4272 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4273                                        loff_t size, unsigned long flags)
4274 {
4275         return __shmem_file_setup(mnt, name, size, flags, 0);
4276 }
4277 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4278
4279 /**
4280  * shmem_zero_setup - setup a shared anonymous mapping
4281  * @vma: the vma to be mmapped is prepared by do_mmap
4282  */
4283 int shmem_zero_setup(struct vm_area_struct *vma)
4284 {
4285         struct file *file;
4286         loff_t size = vma->vm_end - vma->vm_start;
4287
4288         /*
4289          * Cloning a new file under mmap_lock leads to a lock ordering conflict
4290          * between XFS directory reading and selinux: since this file is only
4291          * accessible to the user through its mapping, use S_PRIVATE flag to
4292          * bypass file security, in the same way as shmem_kernel_file_setup().
4293          */
4294         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4295         if (IS_ERR(file))
4296                 return PTR_ERR(file);
4297
4298         if (vma->vm_file)
4299                 fput(vma->vm_file);
4300         vma->vm_file = file;
4301         vma->vm_ops = &shmem_vm_ops;
4302
4303         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
4304                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4305                         (vma->vm_end & HPAGE_PMD_MASK)) {
4306                 khugepaged_enter(vma, vma->vm_flags);
4307         }
4308
4309         return 0;
4310 }
4311
4312 /**
4313  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4314  * @mapping:    the page's address_space
4315  * @index:      the page index
4316  * @gfp:        the page allocator flags to use if allocating
4317  *
4318  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4319  * with any new page allocations done using the specified allocation flags.
4320  * But read_cache_page_gfp() uses the ->readpage() method: which does not
4321  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4322  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4323  *
4324  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4325  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4326  */
4327 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4328                                          pgoff_t index, gfp_t gfp)
4329 {
4330 #ifdef CONFIG_SHMEM
4331         struct inode *inode = mapping->host;
4332         struct page *page;
4333         int error;
4334
4335         BUG_ON(mapping->a_ops != &shmem_aops);
4336         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4337                                   gfp, NULL, NULL, NULL);
4338         if (error)
4339                 page = ERR_PTR(error);
4340         else
4341                 unlock_page(page);
4342         return page;
4343 #else
4344         /*
4345          * The tiny !SHMEM case uses ramfs without swap
4346          */
4347         return read_cache_page_gfp(mapping, index, gfp);
4348 #endif
4349 }
4350 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);