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