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(const struct path *path, struct kstat *stat,
1064                          u32 request_mask, unsigned int query_flags)
1065 {
1066         struct inode *inode = path->dentry->d_inode;
1067         struct shmem_inode_info *info = SHMEM_I(inode);
1068         struct shmem_sb_info *sb_info = SHMEM_SB(inode->i_sb);
1069
1070         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1071                 spin_lock_irq(&info->lock);
1072                 shmem_recalc_inode(inode);
1073                 spin_unlock_irq(&info->lock);
1074         }
1075         generic_fillattr(inode, stat);
1076
1077         if (is_huge_enabled(sb_info))
1078                 stat->blksize = HPAGE_PMD_SIZE;
1079
1080         return 0;
1081 }
1082
1083 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1084 {
1085         struct inode *inode = d_inode(dentry);
1086         struct shmem_inode_info *info = SHMEM_I(inode);
1087         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1088         int error;
1089
1090         error = setattr_prepare(dentry, attr);
1091         if (error)
1092                 return error;
1093
1094         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1095                 loff_t oldsize = inode->i_size;
1096                 loff_t newsize = attr->ia_size;
1097
1098                 /* protected by i_mutex */
1099                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1100                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1101                         return -EPERM;
1102
1103                 if (newsize != oldsize) {
1104                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1105                                         oldsize, newsize);
1106                         if (error)
1107                                 return error;
1108                         i_size_write(inode, newsize);
1109                         inode->i_ctime = inode->i_mtime = current_time(inode);
1110                 }
1111                 if (newsize <= oldsize) {
1112                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1113                         if (oldsize > holebegin)
1114                                 unmap_mapping_range(inode->i_mapping,
1115                                                         holebegin, 0, 1);
1116                         if (info->alloced)
1117                                 shmem_truncate_range(inode,
1118                                                         newsize, (loff_t)-1);
1119                         /* unmap again to remove racily COWed private pages */
1120                         if (oldsize > holebegin)
1121                                 unmap_mapping_range(inode->i_mapping,
1122                                                         holebegin, 0, 1);
1123
1124                         /*
1125                          * Part of the huge page can be beyond i_size: subject
1126                          * to shrink under memory pressure.
1127                          */
1128                         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1129                                 spin_lock(&sbinfo->shrinklist_lock);
1130                                 /*
1131                                  * _careful to defend against unlocked access to
1132                                  * ->shrink_list in shmem_unused_huge_shrink()
1133                                  */
1134                                 if (list_empty_careful(&info->shrinklist)) {
1135                                         list_add_tail(&info->shrinklist,
1136                                                         &sbinfo->shrinklist);
1137                                         sbinfo->shrinklist_len++;
1138                                 }
1139                                 spin_unlock(&sbinfo->shrinklist_lock);
1140                         }
1141                 }
1142         }
1143
1144         setattr_copy(inode, attr);
1145         if (attr->ia_valid & ATTR_MODE)
1146                 error = posix_acl_chmod(inode, inode->i_mode);
1147         return error;
1148 }
1149
1150 static void shmem_evict_inode(struct inode *inode)
1151 {
1152         struct shmem_inode_info *info = SHMEM_I(inode);
1153         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1154
1155         if (shmem_mapping(inode->i_mapping)) {
1156                 shmem_unacct_size(info->flags, inode->i_size);
1157                 inode->i_size = 0;
1158                 shmem_truncate_range(inode, 0, (loff_t)-1);
1159                 if (!list_empty(&info->shrinklist)) {
1160                         spin_lock(&sbinfo->shrinklist_lock);
1161                         if (!list_empty(&info->shrinklist)) {
1162                                 list_del_init(&info->shrinklist);
1163                                 sbinfo->shrinklist_len--;
1164                         }
1165                         spin_unlock(&sbinfo->shrinklist_lock);
1166                 }
1167                 while (!list_empty(&info->swaplist)) {
1168                         /* Wait while shmem_unuse() is scanning this inode... */
1169                         wait_var_event(&info->stop_eviction,
1170                                        !atomic_read(&info->stop_eviction));
1171                         mutex_lock(&shmem_swaplist_mutex);
1172                         /* ...but beware of the race if we peeked too early */
1173                         if (!atomic_read(&info->stop_eviction))
1174                                 list_del_init(&info->swaplist);
1175                         mutex_unlock(&shmem_swaplist_mutex);
1176                 }
1177         }
1178
1179         simple_xattrs_free(&info->xattrs);
1180         WARN_ON(inode->i_blocks);
1181         shmem_free_inode(inode->i_sb);
1182         clear_inode(inode);
1183 }
1184
1185 extern struct swap_info_struct *swap_info[];
1186
1187 static int shmem_find_swap_entries(struct address_space *mapping,
1188                                    pgoff_t start, unsigned int nr_entries,
1189                                    struct page **entries, pgoff_t *indices,
1190                                    unsigned int type, bool frontswap)
1191 {
1192         XA_STATE(xas, &mapping->i_pages, start);
1193         struct page *page;
1194         swp_entry_t entry;
1195         unsigned int ret = 0;
1196
1197         if (!nr_entries)
1198                 return 0;
1199
1200         rcu_read_lock();
1201         xas_for_each(&xas, page, ULONG_MAX) {
1202                 if (xas_retry(&xas, page))
1203                         continue;
1204
1205                 if (!xa_is_value(page))
1206                         continue;
1207
1208                 entry = radix_to_swp_entry(page);
1209                 if (swp_type(entry) != type)
1210                         continue;
1211                 if (frontswap &&
1212                     !frontswap_test(swap_info[type], swp_offset(entry)))
1213                         continue;
1214
1215                 indices[ret] = xas.xa_index;
1216                 entries[ret] = page;
1217
1218                 if (need_resched()) {
1219                         xas_pause(&xas);
1220                         cond_resched_rcu();
1221                 }
1222                 if (++ret == nr_entries)
1223                         break;
1224         }
1225         rcu_read_unlock();
1226
1227         return ret;
1228 }
1229
1230 /*
1231  * Move the swapped pages for an inode to page cache. Returns the count
1232  * of pages swapped in, or the error in case of failure.
1233  */
1234 static int shmem_unuse_swap_entries(struct inode *inode, struct pagevec pvec,
1235                                     pgoff_t *indices)
1236 {
1237         int i = 0;
1238         int ret = 0;
1239         int error = 0;
1240         struct address_space *mapping = inode->i_mapping;
1241
1242         for (i = 0; i < pvec.nr; i++) {
1243                 struct page *page = pvec.pages[i];
1244
1245                 if (!xa_is_value(page))
1246                         continue;
1247                 error = shmem_swapin_page(inode, indices[i],
1248                                           &page, SGP_CACHE,
1249                                           mapping_gfp_mask(mapping),
1250                                           NULL, NULL);
1251                 if (error == 0) {
1252                         unlock_page(page);
1253                         put_page(page);
1254                         ret++;
1255                 }
1256                 if (error == -ENOMEM)
1257                         break;
1258                 error = 0;
1259         }
1260         return error ? error : ret;
1261 }
1262
1263 /*
1264  * If swap found in inode, free it and move page from swapcache to filecache.
1265  */
1266 static int shmem_unuse_inode(struct inode *inode, unsigned int type,
1267                              bool frontswap, unsigned long *fs_pages_to_unuse)
1268 {
1269         struct address_space *mapping = inode->i_mapping;
1270         pgoff_t start = 0;
1271         struct pagevec pvec;
1272         pgoff_t indices[PAGEVEC_SIZE];
1273         bool frontswap_partial = (frontswap && *fs_pages_to_unuse > 0);
1274         int ret = 0;
1275
1276         pagevec_init(&pvec);
1277         do {
1278                 unsigned int nr_entries = PAGEVEC_SIZE;
1279
1280                 if (frontswap_partial && *fs_pages_to_unuse < PAGEVEC_SIZE)
1281                         nr_entries = *fs_pages_to_unuse;
1282
1283                 pvec.nr = shmem_find_swap_entries(mapping, start, nr_entries,
1284                                                   pvec.pages, indices,
1285                                                   type, frontswap);
1286                 if (pvec.nr == 0) {
1287                         ret = 0;
1288                         break;
1289                 }
1290
1291                 ret = shmem_unuse_swap_entries(inode, pvec, indices);
1292                 if (ret < 0)
1293                         break;
1294
1295                 if (frontswap_partial) {
1296                         *fs_pages_to_unuse -= ret;
1297                         if (*fs_pages_to_unuse == 0) {
1298                                 ret = FRONTSWAP_PAGES_UNUSED;
1299                                 break;
1300                         }
1301                 }
1302
1303                 start = indices[pvec.nr - 1];
1304         } while (true);
1305
1306         return ret;
1307 }
1308
1309 /*
1310  * Read all the shared memory data that resides in the swap
1311  * device 'type' back into memory, so the swap device can be
1312  * unused.
1313  */
1314 int shmem_unuse(unsigned int type, bool frontswap,
1315                 unsigned long *fs_pages_to_unuse)
1316 {
1317         struct shmem_inode_info *info, *next;
1318         int error = 0;
1319
1320         if (list_empty(&shmem_swaplist))
1321                 return 0;
1322
1323         mutex_lock(&shmem_swaplist_mutex);
1324         list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1325                 if (!info->swapped) {
1326                         list_del_init(&info->swaplist);
1327                         continue;
1328                 }
1329                 /*
1330                  * Drop the swaplist mutex while searching the inode for swap;
1331                  * but before doing so, make sure shmem_evict_inode() will not
1332                  * remove placeholder inode from swaplist, nor let it be freed
1333                  * (igrab() would protect from unlink, but not from unmount).
1334                  */
1335                 atomic_inc(&info->stop_eviction);
1336                 mutex_unlock(&shmem_swaplist_mutex);
1337
1338                 error = shmem_unuse_inode(&info->vfs_inode, type, frontswap,
1339                                           fs_pages_to_unuse);
1340                 cond_resched();
1341
1342                 mutex_lock(&shmem_swaplist_mutex);
1343                 next = list_next_entry(info, swaplist);
1344                 if (!info->swapped)
1345                         list_del_init(&info->swaplist);
1346                 if (atomic_dec_and_test(&info->stop_eviction))
1347                         wake_up_var(&info->stop_eviction);
1348                 if (error)
1349                         break;
1350         }
1351         mutex_unlock(&shmem_swaplist_mutex);
1352
1353         return error;
1354 }
1355
1356 /*
1357  * Move the page from the page cache to the swap cache.
1358  */
1359 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1360 {
1361         struct shmem_inode_info *info;
1362         struct address_space *mapping;
1363         struct inode *inode;
1364         swp_entry_t swap;
1365         pgoff_t index;
1366
1367         VM_BUG_ON_PAGE(PageCompound(page), page);
1368         BUG_ON(!PageLocked(page));
1369         mapping = page->mapping;
1370         index = page->index;
1371         inode = mapping->host;
1372         info = SHMEM_I(inode);
1373         if (info->flags & VM_LOCKED)
1374                 goto redirty;
1375         if (!total_swap_pages)
1376                 goto redirty;
1377
1378         /*
1379          * Our capabilities prevent regular writeback or sync from ever calling
1380          * shmem_writepage; but a stacking filesystem might use ->writepage of
1381          * its underlying filesystem, in which case tmpfs should write out to
1382          * swap only in response to memory pressure, and not for the writeback
1383          * threads or sync.
1384          */
1385         if (!wbc->for_reclaim) {
1386                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1387                 goto redirty;
1388         }
1389
1390         /*
1391          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1392          * value into swapfile.c, the only way we can correctly account for a
1393          * fallocated page arriving here is now to initialize it and write it.
1394          *
1395          * That's okay for a page already fallocated earlier, but if we have
1396          * not yet completed the fallocation, then (a) we want to keep track
1397          * of this page in case we have to undo it, and (b) it may not be a
1398          * good idea to continue anyway, once we're pushing into swap.  So
1399          * reactivate the page, and let shmem_fallocate() quit when too many.
1400          */
1401         if (!PageUptodate(page)) {
1402                 if (inode->i_private) {
1403                         struct shmem_falloc *shmem_falloc;
1404                         spin_lock(&inode->i_lock);
1405                         shmem_falloc = inode->i_private;
1406                         if (shmem_falloc &&
1407                             !shmem_falloc->waitq &&
1408                             index >= shmem_falloc->start &&
1409                             index < shmem_falloc->next)
1410                                 shmem_falloc->nr_unswapped++;
1411                         else
1412                                 shmem_falloc = NULL;
1413                         spin_unlock(&inode->i_lock);
1414                         if (shmem_falloc)
1415                                 goto redirty;
1416                 }
1417                 clear_highpage(page);
1418                 flush_dcache_page(page);
1419                 SetPageUptodate(page);
1420         }
1421
1422         swap = get_swap_page(page);
1423         if (!swap.val)
1424                 goto redirty;
1425
1426         /*
1427          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1428          * if it's not already there.  Do it now before the page is
1429          * moved to swap cache, when its pagelock no longer protects
1430          * the inode from eviction.  But don't unlock the mutex until
1431          * we've incremented swapped, because shmem_unuse_inode() will
1432          * prune a !swapped inode from the swaplist under this mutex.
1433          */
1434         mutex_lock(&shmem_swaplist_mutex);
1435         if (list_empty(&info->swaplist))
1436                 list_add(&info->swaplist, &shmem_swaplist);
1437
1438         if (add_to_swap_cache(page, swap,
1439                         __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1440                         NULL) == 0) {
1441                 spin_lock_irq(&info->lock);
1442                 shmem_recalc_inode(inode);
1443                 info->swapped++;
1444                 spin_unlock_irq(&info->lock);
1445
1446                 swap_shmem_alloc(swap);
1447                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1448
1449                 mutex_unlock(&shmem_swaplist_mutex);
1450                 BUG_ON(page_mapped(page));
1451                 swap_writepage(page, wbc);
1452                 return 0;
1453         }
1454
1455         mutex_unlock(&shmem_swaplist_mutex);
1456         put_swap_page(page, swap);
1457 redirty:
1458         set_page_dirty(page);
1459         if (wbc->for_reclaim)
1460                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1461         unlock_page(page);
1462         return 0;
1463 }
1464
1465 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1466 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1467 {
1468         char buffer[64];
1469
1470         if (!mpol || mpol->mode == MPOL_DEFAULT)
1471                 return;         /* show nothing */
1472
1473         mpol_to_str(buffer, sizeof(buffer), mpol);
1474
1475         seq_printf(seq, ",mpol=%s", buffer);
1476 }
1477
1478 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1479 {
1480         struct mempolicy *mpol = NULL;
1481         if (sbinfo->mpol) {
1482                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1483                 mpol = sbinfo->mpol;
1484                 mpol_get(mpol);
1485                 spin_unlock(&sbinfo->stat_lock);
1486         }
1487         return mpol;
1488 }
1489 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1490 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1491 {
1492 }
1493 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1494 {
1495         return NULL;
1496 }
1497 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1498 #ifndef CONFIG_NUMA
1499 #define vm_policy vm_private_data
1500 #endif
1501
1502 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1503                 struct shmem_inode_info *info, pgoff_t index)
1504 {
1505         /* Create a pseudo vma that just contains the policy */
1506         vma_init(vma, NULL);
1507         /* Bias interleave by inode number to distribute better across nodes */
1508         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1509         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1510 }
1511
1512 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1513 {
1514         /* Drop reference taken by mpol_shared_policy_lookup() */
1515         mpol_cond_put(vma->vm_policy);
1516 }
1517
1518 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1519                         struct shmem_inode_info *info, pgoff_t index)
1520 {
1521         struct vm_area_struct pvma;
1522         struct page *page;
1523         struct vm_fault vmf = {
1524                 .vma = &pvma,
1525         };
1526
1527         shmem_pseudo_vma_init(&pvma, info, index);
1528         page = swap_cluster_readahead(swap, gfp, &vmf);
1529         shmem_pseudo_vma_destroy(&pvma);
1530
1531         return page;
1532 }
1533
1534 static struct page *shmem_alloc_hugepage(gfp_t gfp,
1535                 struct shmem_inode_info *info, pgoff_t index)
1536 {
1537         struct vm_area_struct pvma;
1538         struct address_space *mapping = info->vfs_inode.i_mapping;
1539         pgoff_t hindex;
1540         struct page *page;
1541
1542         hindex = round_down(index, HPAGE_PMD_NR);
1543         if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1544                                                                 XA_PRESENT))
1545                 return NULL;
1546
1547         shmem_pseudo_vma_init(&pvma, info, hindex);
1548         page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1549                         HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1550         shmem_pseudo_vma_destroy(&pvma);
1551         if (page)
1552                 prep_transhuge_page(page);
1553         else
1554                 count_vm_event(THP_FILE_FALLBACK);
1555         return page;
1556 }
1557
1558 static struct page *shmem_alloc_page(gfp_t gfp,
1559                         struct shmem_inode_info *info, pgoff_t index)
1560 {
1561         struct vm_area_struct pvma;
1562         struct page *page;
1563
1564         shmem_pseudo_vma_init(&pvma, info, index);
1565         page = alloc_page_vma(gfp, &pvma, 0);
1566         shmem_pseudo_vma_destroy(&pvma);
1567
1568         return page;
1569 }
1570
1571 static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1572                 struct inode *inode,
1573                 pgoff_t index, bool huge)
1574 {
1575         struct shmem_inode_info *info = SHMEM_I(inode);
1576         struct page *page;
1577         int nr;
1578         int err = -ENOSPC;
1579
1580         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1581                 huge = false;
1582         nr = huge ? HPAGE_PMD_NR : 1;
1583
1584         if (!shmem_inode_acct_block(inode, nr))
1585                 goto failed;
1586
1587         if (huge)
1588                 page = shmem_alloc_hugepage(gfp, info, index);
1589         else
1590                 page = shmem_alloc_page(gfp, info, index);
1591         if (page) {
1592                 __SetPageLocked(page);
1593                 __SetPageSwapBacked(page);
1594                 return page;
1595         }
1596
1597         err = -ENOMEM;
1598         shmem_inode_unacct_blocks(inode, nr);
1599 failed:
1600         return ERR_PTR(err);
1601 }
1602
1603 /*
1604  * When a page is moved from swapcache to shmem filecache (either by the
1605  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1606  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1607  * ignorance of the mapping it belongs to.  If that mapping has special
1608  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1609  * we may need to copy to a suitable page before moving to filecache.
1610  *
1611  * In a future release, this may well be extended to respect cpuset and
1612  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1613  * but for now it is a simple matter of zone.
1614  */
1615 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1616 {
1617         return page_zonenum(page) > gfp_zone(gfp);
1618 }
1619
1620 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1621                                 struct shmem_inode_info *info, pgoff_t index)
1622 {
1623         struct page *oldpage, *newpage;
1624         struct address_space *swap_mapping;
1625         swp_entry_t entry;
1626         pgoff_t swap_index;
1627         int error;
1628
1629         oldpage = *pagep;
1630         entry.val = page_private(oldpage);
1631         swap_index = swp_offset(entry);
1632         swap_mapping = page_mapping(oldpage);
1633
1634         /*
1635          * We have arrived here because our zones are constrained, so don't
1636          * limit chance of success by further cpuset and node constraints.
1637          */
1638         gfp &= ~GFP_CONSTRAINT_MASK;
1639         newpage = shmem_alloc_page(gfp, info, index);
1640         if (!newpage)
1641                 return -ENOMEM;
1642
1643         get_page(newpage);
1644         copy_highpage(newpage, oldpage);
1645         flush_dcache_page(newpage);
1646
1647         __SetPageLocked(newpage);
1648         __SetPageSwapBacked(newpage);
1649         SetPageUptodate(newpage);
1650         set_page_private(newpage, entry.val);
1651         SetPageSwapCache(newpage);
1652
1653         /*
1654          * Our caller will very soon move newpage out of swapcache, but it's
1655          * a nice clean interface for us to replace oldpage by newpage there.
1656          */
1657         xa_lock_irq(&swap_mapping->i_pages);
1658         error = shmem_replace_entry(swap_mapping, swap_index, oldpage, newpage);
1659         if (!error) {
1660                 mem_cgroup_migrate(oldpage, newpage);
1661                 __inc_lruvec_page_state(newpage, NR_FILE_PAGES);
1662                 __dec_lruvec_page_state(oldpage, NR_FILE_PAGES);
1663         }
1664         xa_unlock_irq(&swap_mapping->i_pages);
1665
1666         if (unlikely(error)) {
1667                 /*
1668                  * Is this possible?  I think not, now that our callers check
1669                  * both PageSwapCache and page_private after getting page lock;
1670                  * but be defensive.  Reverse old to newpage for clear and free.
1671                  */
1672                 oldpage = newpage;
1673         } else {
1674                 lru_cache_add(newpage);
1675                 *pagep = newpage;
1676         }
1677
1678         ClearPageSwapCache(oldpage);
1679         set_page_private(oldpage, 0);
1680
1681         unlock_page(oldpage);
1682         put_page(oldpage);
1683         put_page(oldpage);
1684         return error;
1685 }
1686
1687 /*
1688  * Swap in the page pointed to by *pagep.
1689  * Caller has to make sure that *pagep contains a valid swapped page.
1690  * Returns 0 and the page in pagep if success. On failure, returns the
1691  * error code and NULL in *pagep.
1692  */
1693 static int shmem_swapin_page(struct inode *inode, pgoff_t index,
1694                              struct page **pagep, enum sgp_type sgp,
1695                              gfp_t gfp, struct vm_area_struct *vma,
1696                              vm_fault_t *fault_type)
1697 {
1698         struct address_space *mapping = inode->i_mapping;
1699         struct shmem_inode_info *info = SHMEM_I(inode);
1700         struct mm_struct *charge_mm = vma ? vma->vm_mm : current->mm;
1701         struct page *page;
1702         swp_entry_t swap;
1703         int error;
1704
1705         VM_BUG_ON(!*pagep || !xa_is_value(*pagep));
1706         swap = radix_to_swp_entry(*pagep);
1707         *pagep = NULL;
1708
1709         /* Look it up and read it in.. */
1710         page = lookup_swap_cache(swap, NULL, 0);
1711         if (!page) {
1712                 /* Or update major stats only when swapin succeeds?? */
1713                 if (fault_type) {
1714                         *fault_type |= VM_FAULT_MAJOR;
1715                         count_vm_event(PGMAJFAULT);
1716                         count_memcg_event_mm(charge_mm, PGMAJFAULT);
1717                 }
1718                 /* Here we actually start the io */
1719                 page = shmem_swapin(swap, gfp, info, index);
1720                 if (!page) {
1721                         error = -ENOMEM;
1722                         goto failed;
1723                 }
1724         }
1725
1726         /* We have to do this with page locked to prevent races */
1727         lock_page(page);
1728         if (!PageSwapCache(page) || page_private(page) != swap.val ||
1729             !shmem_confirm_swap(mapping, index, swap)) {
1730                 error = -EEXIST;
1731                 goto unlock;
1732         }
1733         if (!PageUptodate(page)) {
1734                 error = -EIO;
1735                 goto failed;
1736         }
1737         wait_on_page_writeback(page);
1738
1739         /*
1740          * Some architectures may have to restore extra metadata to the
1741          * physical page after reading from swap.
1742          */
1743         arch_swap_restore(swap, page);
1744
1745         if (shmem_should_replace_page(page, gfp)) {
1746                 error = shmem_replace_page(&page, gfp, info, index);
1747                 if (error)
1748                         goto failed;
1749         }
1750
1751         error = shmem_add_to_page_cache(page, mapping, index,
1752                                         swp_to_radix_entry(swap), gfp,
1753                                         charge_mm);
1754         if (error)
1755                 goto failed;
1756
1757         spin_lock_irq(&info->lock);
1758         info->swapped--;
1759         shmem_recalc_inode(inode);
1760         spin_unlock_irq(&info->lock);
1761
1762         if (sgp == SGP_WRITE)
1763                 mark_page_accessed(page);
1764
1765         delete_from_swap_cache(page);
1766         set_page_dirty(page);
1767         swap_free(swap);
1768
1769         *pagep = page;
1770         return 0;
1771 failed:
1772         if (!shmem_confirm_swap(mapping, index, swap))
1773                 error = -EEXIST;
1774 unlock:
1775         if (page) {
1776                 unlock_page(page);
1777                 put_page(page);
1778         }
1779
1780         return error;
1781 }
1782
1783 /*
1784  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1785  *
1786  * If we allocate a new one we do not mark it dirty. That's up to the
1787  * vm. If we swap it in we mark it dirty since we also free the swap
1788  * entry since a page cannot live in both the swap and page cache.
1789  *
1790  * vmf and fault_type are only supplied by shmem_fault:
1791  * otherwise they are NULL.
1792  */
1793 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1794         struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1795         struct vm_area_struct *vma, struct vm_fault *vmf,
1796                         vm_fault_t *fault_type)
1797 {
1798         struct address_space *mapping = inode->i_mapping;
1799         struct shmem_inode_info *info = SHMEM_I(inode);
1800         struct shmem_sb_info *sbinfo;
1801         struct mm_struct *charge_mm;
1802         struct page *page;
1803         enum sgp_type sgp_huge = sgp;
1804         pgoff_t hindex = index;
1805         int error;
1806         int once = 0;
1807         int alloced = 0;
1808
1809         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1810                 return -EFBIG;
1811         if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1812                 sgp = SGP_CACHE;
1813 repeat:
1814         if (sgp <= SGP_CACHE &&
1815             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1816                 return -EINVAL;
1817         }
1818
1819         sbinfo = SHMEM_SB(inode->i_sb);
1820         charge_mm = vma ? vma->vm_mm : current->mm;
1821
1822         page = find_lock_entry(mapping, index);
1823         if (xa_is_value(page)) {
1824                 error = shmem_swapin_page(inode, index, &page,
1825                                           sgp, gfp, vma, fault_type);
1826                 if (error == -EEXIST)
1827                         goto repeat;
1828
1829                 *pagep = page;
1830                 return error;
1831         }
1832
1833         if (page)
1834                 hindex = page->index;
1835         if (page && sgp == SGP_WRITE)
1836                 mark_page_accessed(page);
1837
1838         /* fallocated page? */
1839         if (page && !PageUptodate(page)) {
1840                 if (sgp != SGP_READ)
1841                         goto clear;
1842                 unlock_page(page);
1843                 put_page(page);
1844                 page = NULL;
1845                 hindex = index;
1846         }
1847         if (page || sgp == SGP_READ)
1848                 goto out;
1849
1850         /*
1851          * Fast cache lookup did not find it:
1852          * bring it back from swap or allocate.
1853          */
1854
1855         if (vma && userfaultfd_missing(vma)) {
1856                 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1857                 return 0;
1858         }
1859
1860         /* shmem_symlink() */
1861         if (!shmem_mapping(mapping))
1862                 goto alloc_nohuge;
1863         if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1864                 goto alloc_nohuge;
1865         if (shmem_huge == SHMEM_HUGE_FORCE)
1866                 goto alloc_huge;
1867         switch (sbinfo->huge) {
1868         case SHMEM_HUGE_NEVER:
1869                 goto alloc_nohuge;
1870         case SHMEM_HUGE_WITHIN_SIZE: {
1871                 loff_t i_size;
1872                 pgoff_t off;
1873
1874                 off = round_up(index, HPAGE_PMD_NR);
1875                 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1876                 if (i_size >= HPAGE_PMD_SIZE &&
1877                     i_size >> PAGE_SHIFT >= off)
1878                         goto alloc_huge;
1879
1880                 fallthrough;
1881         }
1882         case SHMEM_HUGE_ADVISE:
1883                 if (sgp_huge == SGP_HUGE)
1884                         goto alloc_huge;
1885                 /* TODO: implement fadvise() hints */
1886                 goto alloc_nohuge;
1887         }
1888
1889 alloc_huge:
1890         page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1891         if (IS_ERR(page)) {
1892 alloc_nohuge:
1893                 page = shmem_alloc_and_acct_page(gfp, inode,
1894                                                  index, false);
1895         }
1896         if (IS_ERR(page)) {
1897                 int retry = 5;
1898
1899                 error = PTR_ERR(page);
1900                 page = NULL;
1901                 if (error != -ENOSPC)
1902                         goto unlock;
1903                 /*
1904                  * Try to reclaim some space by splitting a huge page
1905                  * beyond i_size on the filesystem.
1906                  */
1907                 while (retry--) {
1908                         int ret;
1909
1910                         ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1911                         if (ret == SHRINK_STOP)
1912                                 break;
1913                         if (ret)
1914                                 goto alloc_nohuge;
1915                 }
1916                 goto unlock;
1917         }
1918
1919         if (PageTransHuge(page))
1920                 hindex = round_down(index, HPAGE_PMD_NR);
1921         else
1922                 hindex = index;
1923
1924         if (sgp == SGP_WRITE)
1925                 __SetPageReferenced(page);
1926
1927         error = shmem_add_to_page_cache(page, mapping, hindex,
1928                                         NULL, gfp & GFP_RECLAIM_MASK,
1929                                         charge_mm);
1930         if (error)
1931                 goto unacct;
1932         lru_cache_add(page);
1933
1934         spin_lock_irq(&info->lock);
1935         info->alloced += compound_nr(page);
1936         inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1937         shmem_recalc_inode(inode);
1938         spin_unlock_irq(&info->lock);
1939         alloced = true;
1940
1941         if (PageTransHuge(page) &&
1942             DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1943                         hindex + HPAGE_PMD_NR - 1) {
1944                 /*
1945                  * Part of the huge page is beyond i_size: subject
1946                  * to shrink under memory pressure.
1947                  */
1948                 spin_lock(&sbinfo->shrinklist_lock);
1949                 /*
1950                  * _careful to defend against unlocked access to
1951                  * ->shrink_list in shmem_unused_huge_shrink()
1952                  */
1953                 if (list_empty_careful(&info->shrinklist)) {
1954                         list_add_tail(&info->shrinklist,
1955                                       &sbinfo->shrinklist);
1956                         sbinfo->shrinklist_len++;
1957                 }
1958                 spin_unlock(&sbinfo->shrinklist_lock);
1959         }
1960
1961         /*
1962          * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1963          */
1964         if (sgp == SGP_FALLOC)
1965                 sgp = SGP_WRITE;
1966 clear:
1967         /*
1968          * Let SGP_WRITE caller clear ends if write does not fill page;
1969          * but SGP_FALLOC on a page fallocated earlier must initialize
1970          * it now, lest undo on failure cancel our earlier guarantee.
1971          */
1972         if (sgp != SGP_WRITE && !PageUptodate(page)) {
1973                 int i;
1974
1975                 for (i = 0; i < compound_nr(page); i++) {
1976                         clear_highpage(page + i);
1977                         flush_dcache_page(page + i);
1978                 }
1979                 SetPageUptodate(page);
1980         }
1981
1982         /* Perhaps the file has been truncated since we checked */
1983         if (sgp <= SGP_CACHE &&
1984             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1985                 if (alloced) {
1986                         ClearPageDirty(page);
1987                         delete_from_page_cache(page);
1988                         spin_lock_irq(&info->lock);
1989                         shmem_recalc_inode(inode);
1990                         spin_unlock_irq(&info->lock);
1991                 }
1992                 error = -EINVAL;
1993                 goto unlock;
1994         }
1995 out:
1996         *pagep = page + index - hindex;
1997         return 0;
1998
1999         /*
2000          * Error recovery.
2001          */
2002 unacct:
2003         shmem_inode_unacct_blocks(inode, compound_nr(page));
2004
2005         if (PageTransHuge(page)) {
2006                 unlock_page(page);
2007                 put_page(page);
2008                 goto alloc_nohuge;
2009         }
2010 unlock:
2011         if (page) {
2012                 unlock_page(page);
2013                 put_page(page);
2014         }
2015         if (error == -ENOSPC && !once++) {
2016                 spin_lock_irq(&info->lock);
2017                 shmem_recalc_inode(inode);
2018                 spin_unlock_irq(&info->lock);
2019                 goto repeat;
2020         }
2021         if (error == -EEXIST)
2022                 goto repeat;
2023         return error;
2024 }
2025
2026 /*
2027  * This is like autoremove_wake_function, but it removes the wait queue
2028  * entry unconditionally - even if something else had already woken the
2029  * target.
2030  */
2031 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2032 {
2033         int ret = default_wake_function(wait, mode, sync, key);
2034         list_del_init(&wait->entry);
2035         return ret;
2036 }
2037
2038 static vm_fault_t shmem_fault(struct vm_fault *vmf)
2039 {
2040         struct vm_area_struct *vma = vmf->vma;
2041         struct inode *inode = file_inode(vma->vm_file);
2042         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2043         enum sgp_type sgp;
2044         int err;
2045         vm_fault_t ret = VM_FAULT_LOCKED;
2046
2047         /*
2048          * Trinity finds that probing a hole which tmpfs is punching can
2049          * prevent the hole-punch from ever completing: which in turn
2050          * locks writers out with its hold on i_mutex.  So refrain from
2051          * faulting pages into the hole while it's being punched.  Although
2052          * shmem_undo_range() does remove the additions, it may be unable to
2053          * keep up, as each new page needs its own unmap_mapping_range() call,
2054          * and the i_mmap tree grows ever slower to scan if new vmas are added.
2055          *
2056          * It does not matter if we sometimes reach this check just before the
2057          * hole-punch begins, so that one fault then races with the punch:
2058          * we just need to make racing faults a rare case.
2059          *
2060          * The implementation below would be much simpler if we just used a
2061          * standard mutex or completion: but we cannot take i_mutex in fault,
2062          * and bloating every shmem inode for this unlikely case would be sad.
2063          */
2064         if (unlikely(inode->i_private)) {
2065                 struct shmem_falloc *shmem_falloc;
2066
2067                 spin_lock(&inode->i_lock);
2068                 shmem_falloc = inode->i_private;
2069                 if (shmem_falloc &&
2070                     shmem_falloc->waitq &&
2071                     vmf->pgoff >= shmem_falloc->start &&
2072                     vmf->pgoff < shmem_falloc->next) {
2073                         struct file *fpin;
2074                         wait_queue_head_t *shmem_falloc_waitq;
2075                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2076
2077                         ret = VM_FAULT_NOPAGE;
2078                         fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2079                         if (fpin)
2080                                 ret = VM_FAULT_RETRY;
2081
2082                         shmem_falloc_waitq = shmem_falloc->waitq;
2083                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2084                                         TASK_UNINTERRUPTIBLE);
2085                         spin_unlock(&inode->i_lock);
2086                         schedule();
2087
2088                         /*
2089                          * shmem_falloc_waitq points into the shmem_fallocate()
2090                          * stack of the hole-punching task: shmem_falloc_waitq
2091                          * is usually invalid by the time we reach here, but
2092                          * finish_wait() does not dereference it in that case;
2093                          * though i_lock needed lest racing with wake_up_all().
2094                          */
2095                         spin_lock(&inode->i_lock);
2096                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2097                         spin_unlock(&inode->i_lock);
2098
2099                         if (fpin)
2100                                 fput(fpin);
2101                         return ret;
2102                 }
2103                 spin_unlock(&inode->i_lock);
2104         }
2105
2106         sgp = SGP_CACHE;
2107
2108         if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2109             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2110                 sgp = SGP_NOHUGE;
2111         else if (vma->vm_flags & VM_HUGEPAGE)
2112                 sgp = SGP_HUGE;
2113
2114         err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2115                                   gfp, vma, vmf, &ret);
2116         if (err)
2117                 return vmf_error(err);
2118         return ret;
2119 }
2120
2121 unsigned long shmem_get_unmapped_area(struct file *file,
2122                                       unsigned long uaddr, unsigned long len,
2123                                       unsigned long pgoff, unsigned long flags)
2124 {
2125         unsigned long (*get_area)(struct file *,
2126                 unsigned long, unsigned long, unsigned long, unsigned long);
2127         unsigned long addr;
2128         unsigned long offset;
2129         unsigned long inflated_len;
2130         unsigned long inflated_addr;
2131         unsigned long inflated_offset;
2132
2133         if (len > TASK_SIZE)
2134                 return -ENOMEM;
2135
2136         get_area = current->mm->get_unmapped_area;
2137         addr = get_area(file, uaddr, len, pgoff, flags);
2138
2139         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2140                 return addr;
2141         if (IS_ERR_VALUE(addr))
2142                 return addr;
2143         if (addr & ~PAGE_MASK)
2144                 return addr;
2145         if (addr > TASK_SIZE - len)
2146                 return addr;
2147
2148         if (shmem_huge == SHMEM_HUGE_DENY)
2149                 return addr;
2150         if (len < HPAGE_PMD_SIZE)
2151                 return addr;
2152         if (flags & MAP_FIXED)
2153                 return addr;
2154         /*
2155          * Our priority is to support MAP_SHARED mapped hugely;
2156          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2157          * But if caller specified an address hint and we allocated area there
2158          * successfully, respect that as before.
2159          */
2160         if (uaddr == addr)
2161                 return addr;
2162
2163         if (shmem_huge != SHMEM_HUGE_FORCE) {
2164                 struct super_block *sb;
2165
2166                 if (file) {
2167                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2168                         sb = file_inode(file)->i_sb;
2169                 } else {
2170                         /*
2171                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2172                          * for "/dev/zero", to create a shared anonymous object.
2173                          */
2174                         if (IS_ERR(shm_mnt))
2175                                 return addr;
2176                         sb = shm_mnt->mnt_sb;
2177                 }
2178                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2179                         return addr;
2180         }
2181
2182         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2183         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2184                 return addr;
2185         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2186                 return addr;
2187
2188         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2189         if (inflated_len > TASK_SIZE)
2190                 return addr;
2191         if (inflated_len < len)
2192                 return addr;
2193
2194         inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2195         if (IS_ERR_VALUE(inflated_addr))
2196                 return addr;
2197         if (inflated_addr & ~PAGE_MASK)
2198                 return addr;
2199
2200         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2201         inflated_addr += offset - inflated_offset;
2202         if (inflated_offset > offset)
2203                 inflated_addr += HPAGE_PMD_SIZE;
2204
2205         if (inflated_addr > TASK_SIZE - len)
2206                 return addr;
2207         return inflated_addr;
2208 }
2209
2210 #ifdef CONFIG_NUMA
2211 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2212 {
2213         struct inode *inode = file_inode(vma->vm_file);
2214         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2215 }
2216
2217 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2218                                           unsigned long addr)
2219 {
2220         struct inode *inode = file_inode(vma->vm_file);
2221         pgoff_t index;
2222
2223         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2224         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2225 }
2226 #endif
2227
2228 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2229 {
2230         struct inode *inode = file_inode(file);
2231         struct shmem_inode_info *info = SHMEM_I(inode);
2232         int retval = -ENOMEM;
2233
2234         /*
2235          * What serializes the accesses to info->flags?
2236          * ipc_lock_object() when called from shmctl_do_lock(),
2237          * no serialization needed when called from shm_destroy().
2238          */
2239         if (lock && !(info->flags & VM_LOCKED)) {
2240                 if (!user_shm_lock(inode->i_size, user))
2241                         goto out_nomem;
2242                 info->flags |= VM_LOCKED;
2243                 mapping_set_unevictable(file->f_mapping);
2244         }
2245         if (!lock && (info->flags & VM_LOCKED) && user) {
2246                 user_shm_unlock(inode->i_size, user);
2247                 info->flags &= ~VM_LOCKED;
2248                 mapping_clear_unevictable(file->f_mapping);
2249         }
2250         retval = 0;
2251
2252 out_nomem:
2253         return retval;
2254 }
2255
2256 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2257 {
2258         struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2259
2260         if (info->seals & F_SEAL_FUTURE_WRITE) {
2261                 /*
2262                  * New PROT_WRITE and MAP_SHARED mmaps are not allowed when
2263                  * "future write" seal active.
2264                  */
2265                 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
2266                         return -EPERM;
2267
2268                 /*
2269                  * Since an F_SEAL_FUTURE_WRITE sealed memfd can be mapped as
2270                  * MAP_SHARED and read-only, take care to not allow mprotect to
2271                  * revert protections on such mappings. Do this only for shared
2272                  * mappings. For private mappings, don't need to mask
2273                  * VM_MAYWRITE as we still want them to be COW-writable.
2274                  */
2275                 if (vma->vm_flags & VM_SHARED)
2276                         vma->vm_flags &= ~(VM_MAYWRITE);
2277         }
2278
2279         /* arm64 - allow memory tagging on RAM-based files */
2280         vma->vm_flags |= VM_MTE_ALLOWED;
2281
2282         file_accessed(file);
2283         vma->vm_ops = &shmem_vm_ops;
2284         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
2285                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2286                         (vma->vm_end & HPAGE_PMD_MASK)) {
2287                 khugepaged_enter(vma, vma->vm_flags);
2288         }
2289         return 0;
2290 }
2291
2292 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2293                                      umode_t mode, dev_t dev, unsigned long flags)
2294 {
2295         struct inode *inode;
2296         struct shmem_inode_info *info;
2297         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2298         ino_t ino;
2299
2300         if (shmem_reserve_inode(sb, &ino))
2301                 return NULL;
2302
2303         inode = new_inode(sb);
2304         if (inode) {
2305                 inode->i_ino = ino;
2306                 inode_init_owner(inode, dir, mode);
2307                 inode->i_blocks = 0;
2308                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2309                 inode->i_generation = prandom_u32();
2310                 info = SHMEM_I(inode);
2311                 memset(info, 0, (char *)inode - (char *)info);
2312                 spin_lock_init(&info->lock);
2313                 atomic_set(&info->stop_eviction, 0);
2314                 info->seals = F_SEAL_SEAL;
2315                 info->flags = flags & VM_NORESERVE;
2316                 INIT_LIST_HEAD(&info->shrinklist);
2317                 INIT_LIST_HEAD(&info->swaplist);
2318                 simple_xattrs_init(&info->xattrs);
2319                 cache_no_acl(inode);
2320
2321                 switch (mode & S_IFMT) {
2322                 default:
2323                         inode->i_op = &shmem_special_inode_operations;
2324                         init_special_inode(inode, mode, dev);
2325                         break;
2326                 case S_IFREG:
2327                         inode->i_mapping->a_ops = &shmem_aops;
2328                         inode->i_op = &shmem_inode_operations;
2329                         inode->i_fop = &shmem_file_operations;
2330                         mpol_shared_policy_init(&info->policy,
2331                                                  shmem_get_sbmpol(sbinfo));
2332                         break;
2333                 case S_IFDIR:
2334                         inc_nlink(inode);
2335                         /* Some things misbehave if size == 0 on a directory */
2336                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2337                         inode->i_op = &shmem_dir_inode_operations;
2338                         inode->i_fop = &simple_dir_operations;
2339                         break;
2340                 case S_IFLNK:
2341                         /*
2342                          * Must not load anything in the rbtree,
2343                          * mpol_free_shared_policy will not be called.
2344                          */
2345                         mpol_shared_policy_init(&info->policy, NULL);
2346                         break;
2347                 }
2348
2349                 lockdep_annotate_inode_mutex_key(inode);
2350         } else
2351                 shmem_free_inode(sb);
2352         return inode;
2353 }
2354
2355 static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2356                                   pmd_t *dst_pmd,
2357                                   struct vm_area_struct *dst_vma,
2358                                   unsigned long dst_addr,
2359                                   unsigned long src_addr,
2360                                   bool zeropage,
2361                                   struct page **pagep)
2362 {
2363         struct inode *inode = file_inode(dst_vma->vm_file);
2364         struct shmem_inode_info *info = SHMEM_I(inode);
2365         struct address_space *mapping = inode->i_mapping;
2366         gfp_t gfp = mapping_gfp_mask(mapping);
2367         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2368         spinlock_t *ptl;
2369         void *page_kaddr;
2370         struct page *page;
2371         pte_t _dst_pte, *dst_pte;
2372         int ret;
2373         pgoff_t offset, max_off;
2374
2375         ret = -ENOMEM;
2376         if (!shmem_inode_acct_block(inode, 1))
2377                 goto out;
2378
2379         if (!*pagep) {
2380                 page = shmem_alloc_page(gfp, info, pgoff);
2381                 if (!page)
2382                         goto out_unacct_blocks;
2383
2384                 if (!zeropage) {        /* mcopy_atomic */
2385                         page_kaddr = kmap_atomic(page);
2386                         ret = copy_from_user(page_kaddr,
2387                                              (const void __user *)src_addr,
2388                                              PAGE_SIZE);
2389                         kunmap_atomic(page_kaddr);
2390
2391                         /* fallback to copy_from_user outside mmap_lock */
2392                         if (unlikely(ret)) {
2393                                 *pagep = page;
2394                                 shmem_inode_unacct_blocks(inode, 1);
2395                                 /* don't free the page */
2396                                 return -ENOENT;
2397                         }
2398                 } else {                /* mfill_zeropage_atomic */
2399                         clear_highpage(page);
2400                 }
2401         } else {
2402                 page = *pagep;
2403                 *pagep = NULL;
2404         }
2405
2406         VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2407         __SetPageLocked(page);
2408         __SetPageSwapBacked(page);
2409         __SetPageUptodate(page);
2410
2411         ret = -EFAULT;
2412         offset = linear_page_index(dst_vma, dst_addr);
2413         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2414         if (unlikely(offset >= max_off))
2415                 goto out_release;
2416
2417         ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL,
2418                                       gfp & GFP_RECLAIM_MASK, dst_mm);
2419         if (ret)
2420                 goto out_release;
2421
2422         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2423         if (dst_vma->vm_flags & VM_WRITE)
2424                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2425         else {
2426                 /*
2427                  * We don't set the pte dirty if the vma has no
2428                  * VM_WRITE permission, so mark the page dirty or it
2429                  * could be freed from under us. We could do it
2430                  * unconditionally before unlock_page(), but doing it
2431                  * only if VM_WRITE is not set is faster.
2432                  */
2433                 set_page_dirty(page);
2434         }
2435
2436         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2437
2438         ret = -EFAULT;
2439         max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2440         if (unlikely(offset >= max_off))
2441                 goto out_release_unlock;
2442
2443         ret = -EEXIST;
2444         if (!pte_none(*dst_pte))
2445                 goto out_release_unlock;
2446
2447         lru_cache_add(page);
2448
2449         spin_lock_irq(&info->lock);
2450         info->alloced++;
2451         inode->i_blocks += BLOCKS_PER_PAGE;
2452         shmem_recalc_inode(inode);
2453         spin_unlock_irq(&info->lock);
2454
2455         inc_mm_counter(dst_mm, mm_counter_file(page));
2456         page_add_file_rmap(page, false);
2457         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2458
2459         /* No need to invalidate - it was non-present before */
2460         update_mmu_cache(dst_vma, dst_addr, dst_pte);
2461         pte_unmap_unlock(dst_pte, ptl);
2462         unlock_page(page);
2463         ret = 0;
2464 out:
2465         return ret;
2466 out_release_unlock:
2467         pte_unmap_unlock(dst_pte, ptl);
2468         ClearPageDirty(page);
2469         delete_from_page_cache(page);
2470 out_release:
2471         unlock_page(page);
2472         put_page(page);
2473 out_unacct_blocks:
2474         shmem_inode_unacct_blocks(inode, 1);
2475         goto out;
2476 }
2477
2478 int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2479                            pmd_t *dst_pmd,
2480                            struct vm_area_struct *dst_vma,
2481                            unsigned long dst_addr,
2482                            unsigned long src_addr,
2483                            struct page **pagep)
2484 {
2485         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2486                                       dst_addr, src_addr, false, pagep);
2487 }
2488
2489 int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2490                              pmd_t *dst_pmd,
2491                              struct vm_area_struct *dst_vma,
2492                              unsigned long dst_addr)
2493 {
2494         struct page *page = NULL;
2495
2496         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2497                                       dst_addr, 0, true, &page);
2498 }
2499
2500 #ifdef CONFIG_TMPFS
2501 static const struct inode_operations shmem_symlink_inode_operations;
2502 static const struct inode_operations shmem_short_symlink_operations;
2503
2504 #ifdef CONFIG_TMPFS_XATTR
2505 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2506 #else
2507 #define shmem_initxattrs NULL
2508 #endif
2509
2510 static int
2511 shmem_write_begin(struct file *file, struct address_space *mapping,
2512                         loff_t pos, unsigned len, unsigned flags,
2513                         struct page **pagep, void **fsdata)
2514 {
2515         struct inode *inode = mapping->host;
2516         struct shmem_inode_info *info = SHMEM_I(inode);
2517         pgoff_t index = pos >> PAGE_SHIFT;
2518
2519         /* i_mutex is held by caller */
2520         if (unlikely(info->seals & (F_SEAL_GROW |
2521                                    F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2522                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2523                         return -EPERM;
2524                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2525                         return -EPERM;
2526         }
2527
2528         return shmem_getpage(inode, index, pagep, SGP_WRITE);
2529 }
2530
2531 static int
2532 shmem_write_end(struct file *file, struct address_space *mapping,
2533                         loff_t pos, unsigned len, unsigned copied,
2534                         struct page *page, void *fsdata)
2535 {
2536         struct inode *inode = mapping->host;
2537
2538         if (pos + copied > inode->i_size)
2539                 i_size_write(inode, pos + copied);
2540
2541         if (!PageUptodate(page)) {
2542                 struct page *head = compound_head(page);
2543                 if (PageTransCompound(page)) {
2544                         int i;
2545
2546                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2547                                 if (head + i == page)
2548                                         continue;
2549                                 clear_highpage(head + i);
2550                                 flush_dcache_page(head + i);
2551                         }
2552                 }
2553                 if (copied < PAGE_SIZE) {
2554                         unsigned from = pos & (PAGE_SIZE - 1);
2555                         zero_user_segments(page, 0, from,
2556                                         from + copied, PAGE_SIZE);
2557                 }
2558                 SetPageUptodate(head);
2559         }
2560         set_page_dirty(page);
2561         unlock_page(page);
2562         put_page(page);
2563
2564         return copied;
2565 }
2566
2567 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2568 {
2569         struct file *file = iocb->ki_filp;
2570         struct inode *inode = file_inode(file);
2571         struct address_space *mapping = inode->i_mapping;
2572         pgoff_t index;
2573         unsigned long offset;
2574         enum sgp_type sgp = SGP_READ;
2575         int error = 0;
2576         ssize_t retval = 0;
2577         loff_t *ppos = &iocb->ki_pos;
2578
2579         /*
2580          * Might this read be for a stacking filesystem?  Then when reading
2581          * holes of a sparse file, we actually need to allocate those pages,
2582          * and even mark them dirty, so it cannot exceed the max_blocks limit.
2583          */
2584         if (!iter_is_iovec(to))
2585                 sgp = SGP_CACHE;
2586
2587         index = *ppos >> PAGE_SHIFT;
2588         offset = *ppos & ~PAGE_MASK;
2589
2590         for (;;) {
2591                 struct page *page = NULL;
2592                 pgoff_t end_index;
2593                 unsigned long nr, ret;
2594                 loff_t i_size = i_size_read(inode);
2595
2596                 end_index = i_size >> PAGE_SHIFT;
2597                 if (index > end_index)
2598                         break;
2599                 if (index == end_index) {
2600                         nr = i_size & ~PAGE_MASK;
2601                         if (nr <= offset)
2602                                 break;
2603                 }
2604
2605                 error = shmem_getpage(inode, index, &page, sgp);
2606                 if (error) {
2607                         if (error == -EINVAL)
2608                                 error = 0;
2609                         break;
2610                 }
2611                 if (page) {
2612                         if (sgp == SGP_CACHE)
2613                                 set_page_dirty(page);
2614                         unlock_page(page);
2615                 }
2616
2617                 /*
2618                  * We must evaluate after, since reads (unlike writes)
2619                  * are called without i_mutex protection against truncate
2620                  */
2621                 nr = PAGE_SIZE;
2622                 i_size = i_size_read(inode);
2623                 end_index = i_size >> PAGE_SHIFT;
2624                 if (index == end_index) {
2625                         nr = i_size & ~PAGE_MASK;
2626                         if (nr <= offset) {
2627                                 if (page)
2628                                         put_page(page);
2629                                 break;
2630                         }
2631                 }
2632                 nr -= offset;
2633
2634                 if (page) {
2635                         /*
2636                          * If users can be writing to this page using arbitrary
2637                          * virtual addresses, take care about potential aliasing
2638                          * before reading the page on the kernel side.
2639                          */
2640                         if (mapping_writably_mapped(mapping))
2641                                 flush_dcache_page(page);
2642                         /*
2643                          * Mark the page accessed if we read the beginning.
2644                          */
2645                         if (!offset)
2646                                 mark_page_accessed(page);
2647                 } else {
2648                         page = ZERO_PAGE(0);
2649                         get_page(page);
2650                 }
2651
2652                 /*
2653                  * Ok, we have the page, and it's up-to-date, so
2654                  * now we can copy it to user space...
2655                  */
2656                 ret = copy_page_to_iter(page, offset, nr, to);
2657                 retval += ret;
2658                 offset += ret;
2659                 index += offset >> PAGE_SHIFT;
2660                 offset &= ~PAGE_MASK;
2661
2662                 put_page(page);
2663                 if (!iov_iter_count(to))
2664                         break;
2665                 if (ret < nr) {
2666                         error = -EFAULT;
2667                         break;
2668                 }
2669                 cond_resched();
2670         }
2671
2672         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2673         file_accessed(file);
2674         return retval ? retval : error;
2675 }
2676
2677 /*
2678  * llseek SEEK_DATA or SEEK_HOLE through the page cache.
2679  */
2680 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2681                                     pgoff_t index, pgoff_t end, int whence)
2682 {
2683         struct page *page;
2684         struct pagevec pvec;
2685         pgoff_t indices[PAGEVEC_SIZE];
2686         bool done = false;
2687         int i;
2688
2689         pagevec_init(&pvec);
2690         pvec.nr = 1;            /* start small: we may be there already */
2691         while (!done) {
2692                 pvec.nr = find_get_entries(mapping, index,
2693                                         pvec.nr, pvec.pages, indices);
2694                 if (!pvec.nr) {
2695                         if (whence == SEEK_DATA)
2696                                 index = end;
2697                         break;
2698                 }
2699                 for (i = 0; i < pvec.nr; i++, index++) {
2700                         if (index < indices[i]) {
2701                                 if (whence == SEEK_HOLE) {
2702                                         done = true;
2703                                         break;
2704                                 }
2705                                 index = indices[i];
2706                         }
2707                         page = pvec.pages[i];
2708                         if (page && !xa_is_value(page)) {
2709                                 if (!PageUptodate(page))
2710                                         page = NULL;
2711                         }
2712                         if (index >= end ||
2713                             (page && whence == SEEK_DATA) ||
2714                             (!page && whence == SEEK_HOLE)) {
2715                                 done = true;
2716                                 break;
2717                         }
2718                 }
2719                 pagevec_remove_exceptionals(&pvec);
2720                 pagevec_release(&pvec);
2721                 pvec.nr = PAGEVEC_SIZE;
2722                 cond_resched();
2723         }
2724         return index;
2725 }
2726
2727 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2728 {
2729         struct address_space *mapping = file->f_mapping;
2730         struct inode *inode = mapping->host;
2731         pgoff_t start, end;
2732         loff_t new_offset;
2733
2734         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2735                 return generic_file_llseek_size(file, offset, whence,
2736                                         MAX_LFS_FILESIZE, i_size_read(inode));
2737         inode_lock(inode);
2738         /* We're holding i_mutex so we can access i_size directly */
2739
2740         if (offset < 0 || offset >= inode->i_size)
2741                 offset = -ENXIO;
2742         else {
2743                 start = offset >> PAGE_SHIFT;
2744                 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2745                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2746                 new_offset <<= PAGE_SHIFT;
2747                 if (new_offset > offset) {
2748                         if (new_offset < inode->i_size)
2749                                 offset = new_offset;
2750                         else if (whence == SEEK_DATA)
2751                                 offset = -ENXIO;
2752                         else
2753                                 offset = inode->i_size;
2754                 }
2755         }
2756
2757         if (offset >= 0)
2758                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2759         inode_unlock(inode);
2760         return offset;
2761 }
2762
2763 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2764                                                          loff_t len)
2765 {
2766         struct inode *inode = file_inode(file);
2767         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2768         struct shmem_inode_info *info = SHMEM_I(inode);
2769         struct shmem_falloc shmem_falloc;
2770         pgoff_t start, index, end;
2771         int error;
2772
2773         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2774                 return -EOPNOTSUPP;
2775
2776         inode_lock(inode);
2777
2778         if (mode & FALLOC_FL_PUNCH_HOLE) {
2779                 struct address_space *mapping = file->f_mapping;
2780                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2781                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2782                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2783
2784                 /* protected by i_mutex */
2785                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2786                         error = -EPERM;
2787                         goto out;
2788                 }
2789
2790                 shmem_falloc.waitq = &shmem_falloc_waitq;
2791                 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2792                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2793                 spin_lock(&inode->i_lock);
2794                 inode->i_private = &shmem_falloc;
2795                 spin_unlock(&inode->i_lock);
2796
2797                 if ((u64)unmap_end > (u64)unmap_start)
2798                         unmap_mapping_range(mapping, unmap_start,
2799                                             1 + unmap_end - unmap_start, 0);
2800                 shmem_truncate_range(inode, offset, offset + len - 1);
2801                 /* No need to unmap again: hole-punching leaves COWed pages */
2802
2803                 spin_lock(&inode->i_lock);
2804                 inode->i_private = NULL;
2805                 wake_up_all(&shmem_falloc_waitq);
2806                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2807                 spin_unlock(&inode->i_lock);
2808                 error = 0;
2809                 goto out;
2810         }
2811
2812         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2813         error = inode_newsize_ok(inode, offset + len);
2814         if (error)
2815                 goto out;
2816
2817         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2818                 error = -EPERM;
2819                 goto out;
2820         }
2821
2822         start = offset >> PAGE_SHIFT;
2823         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2824         /* Try to avoid a swapstorm if len is impossible to satisfy */
2825         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2826                 error = -ENOSPC;
2827                 goto out;
2828         }
2829
2830         shmem_falloc.waitq = NULL;
2831         shmem_falloc.start = start;
2832         shmem_falloc.next  = start;
2833         shmem_falloc.nr_falloced = 0;
2834         shmem_falloc.nr_unswapped = 0;
2835         spin_lock(&inode->i_lock);
2836         inode->i_private = &shmem_falloc;
2837         spin_unlock(&inode->i_lock);
2838
2839         for (index = start; index < end; index++) {
2840                 struct page *page;
2841
2842                 /*
2843                  * Good, the fallocate(2) manpage permits EINTR: we may have
2844                  * been interrupted because we are using up too much memory.
2845                  */
2846                 if (signal_pending(current))
2847                         error = -EINTR;
2848                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2849                         error = -ENOMEM;
2850                 else
2851                         error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2852                 if (error) {
2853                         /* Remove the !PageUptodate pages we added */
2854                         if (index > start) {
2855                                 shmem_undo_range(inode,
2856                                     (loff_t)start << PAGE_SHIFT,
2857                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2858                         }
2859                         goto undone;
2860                 }
2861
2862                 /*
2863                  * Inform shmem_writepage() how far we have reached.
2864                  * No need for lock or barrier: we have the page lock.
2865                  */
2866                 shmem_falloc.next++;
2867                 if (!PageUptodate(page))
2868                         shmem_falloc.nr_falloced++;
2869
2870                 /*
2871                  * If !PageUptodate, leave it that way so that freeable pages
2872                  * can be recognized if we need to rollback on error later.
2873                  * But set_page_dirty so that memory pressure will swap rather
2874                  * than free the pages we are allocating (and SGP_CACHE pages
2875                  * might still be clean: we now need to mark those dirty too).
2876                  */
2877                 set_page_dirty(page);
2878                 unlock_page(page);
2879                 put_page(page);
2880                 cond_resched();
2881         }
2882
2883         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2884                 i_size_write(inode, offset + len);
2885         inode->i_ctime = current_time(inode);
2886 undone:
2887         spin_lock(&inode->i_lock);
2888         inode->i_private = NULL;
2889         spin_unlock(&inode->i_lock);
2890 out:
2891         inode_unlock(inode);
2892         return error;
2893 }
2894
2895 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2896 {
2897         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2898
2899         buf->f_type = TMPFS_MAGIC;
2900         buf->f_bsize = PAGE_SIZE;
2901         buf->f_namelen = NAME_MAX;
2902         if (sbinfo->max_blocks) {
2903                 buf->f_blocks = sbinfo->max_blocks;
2904                 buf->f_bavail =
2905                 buf->f_bfree  = sbinfo->max_blocks -
2906                                 percpu_counter_sum(&sbinfo->used_blocks);
2907         }
2908         if (sbinfo->max_inodes) {
2909                 buf->f_files = sbinfo->max_inodes;
2910                 buf->f_ffree = sbinfo->free_inodes;
2911         }
2912         /* else leave those fields 0 like simple_statfs */
2913         return 0;
2914 }
2915
2916 /*
2917  * File creation. Allocate an inode, and we're done..
2918  */
2919 static int
2920 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2921 {
2922         struct inode *inode;
2923         int error = -ENOSPC;
2924
2925         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2926         if (inode) {
2927                 error = simple_acl_create(dir, inode);
2928                 if (error)
2929                         goto out_iput;
2930                 error = security_inode_init_security(inode, dir,
2931                                                      &dentry->d_name,
2932                                                      shmem_initxattrs, NULL);
2933                 if (error && error != -EOPNOTSUPP)
2934                         goto out_iput;
2935
2936                 error = 0;
2937                 dir->i_size += BOGO_DIRENT_SIZE;
2938                 dir->i_ctime = dir->i_mtime = current_time(dir);
2939                 d_instantiate(dentry, inode);
2940                 dget(dentry); /* Extra count - pin the dentry in core */
2941         }
2942         return error;
2943 out_iput:
2944         iput(inode);
2945         return error;
2946 }
2947
2948 static int
2949 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2950 {
2951         struct inode *inode;
2952         int error = -ENOSPC;
2953
2954         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2955         if (inode) {
2956                 error = security_inode_init_security(inode, dir,
2957                                                      NULL,
2958                                                      shmem_initxattrs, NULL);
2959                 if (error && error != -EOPNOTSUPP)
2960                         goto out_iput;
2961                 error = simple_acl_create(dir, inode);
2962                 if (error)
2963                         goto out_iput;
2964                 d_tmpfile(dentry, inode);
2965         }
2966         return error;
2967 out_iput:
2968         iput(inode);
2969         return error;
2970 }
2971
2972 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2973 {
2974         int error;
2975
2976         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2977                 return error;
2978         inc_nlink(dir);
2979         return 0;
2980 }
2981
2982 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2983                 bool excl)
2984 {
2985         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2986 }
2987
2988 /*
2989  * Link a file..
2990  */
2991 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2992 {
2993         struct inode *inode = d_inode(old_dentry);
2994         int ret = 0;
2995
2996         /*
2997          * No ordinary (disk based) filesystem counts links as inodes;
2998          * but each new link needs a new dentry, pinning lowmem, and
2999          * tmpfs dentries cannot be pruned until they are unlinked.
3000          * But if an O_TMPFILE file is linked into the tmpfs, the
3001          * first link must skip that, to get the accounting right.
3002          */
3003         if (inode->i_nlink) {
3004                 ret = shmem_reserve_inode(inode->i_sb, NULL);
3005                 if (ret)
3006                         goto out;
3007         }
3008
3009         dir->i_size += BOGO_DIRENT_SIZE;
3010         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3011         inc_nlink(inode);
3012         ihold(inode);   /* New dentry reference */
3013         dget(dentry);           /* Extra pinning count for the created dentry */
3014         d_instantiate(dentry, inode);
3015 out:
3016         return ret;
3017 }
3018
3019 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3020 {
3021         struct inode *inode = d_inode(dentry);
3022
3023         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3024                 shmem_free_inode(inode->i_sb);
3025
3026         dir->i_size -= BOGO_DIRENT_SIZE;
3027         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3028         drop_nlink(inode);
3029         dput(dentry);   /* Undo the count from "create" - this does all the work */
3030         return 0;
3031 }
3032
3033 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3034 {
3035         if (!simple_empty(dentry))
3036                 return -ENOTEMPTY;
3037
3038         drop_nlink(d_inode(dentry));
3039         drop_nlink(dir);
3040         return shmem_unlink(dir, dentry);
3041 }
3042
3043 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
3044 {
3045         bool old_is_dir = d_is_dir(old_dentry);
3046         bool new_is_dir = d_is_dir(new_dentry);
3047
3048         if (old_dir != new_dir && old_is_dir != new_is_dir) {
3049                 if (old_is_dir) {
3050                         drop_nlink(old_dir);
3051                         inc_nlink(new_dir);
3052                 } else {
3053                         drop_nlink(new_dir);
3054                         inc_nlink(old_dir);
3055                 }
3056         }
3057         old_dir->i_ctime = old_dir->i_mtime =
3058         new_dir->i_ctime = new_dir->i_mtime =
3059         d_inode(old_dentry)->i_ctime =
3060         d_inode(new_dentry)->i_ctime = current_time(old_dir);
3061
3062         return 0;
3063 }
3064
3065 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3066 {
3067         struct dentry *whiteout;
3068         int error;
3069
3070         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3071         if (!whiteout)
3072                 return -ENOMEM;
3073
3074         error = shmem_mknod(old_dir, whiteout,
3075                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3076         dput(whiteout);
3077         if (error)
3078                 return error;
3079
3080         /*
3081          * Cheat and hash the whiteout while the old dentry is still in
3082          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3083          *
3084          * d_lookup() will consistently find one of them at this point,
3085          * not sure which one, but that isn't even important.
3086          */
3087         d_rehash(whiteout);
3088         return 0;
3089 }
3090
3091 /*
3092  * The VFS layer already does all the dentry stuff for rename,
3093  * we just have to decrement the usage count for the target if
3094  * it exists so that the VFS layer correctly free's it when it
3095  * gets overwritten.
3096  */
3097 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3098 {
3099         struct inode *inode = d_inode(old_dentry);
3100         int they_are_dirs = S_ISDIR(inode->i_mode);
3101
3102         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3103                 return -EINVAL;
3104
3105         if (flags & RENAME_EXCHANGE)
3106                 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3107
3108         if (!simple_empty(new_dentry))
3109                 return -ENOTEMPTY;
3110
3111         if (flags & RENAME_WHITEOUT) {
3112                 int error;
3113
3114                 error = shmem_whiteout(old_dir, old_dentry);
3115                 if (error)
3116                         return error;
3117         }
3118
3119         if (d_really_is_positive(new_dentry)) {
3120                 (void) shmem_unlink(new_dir, new_dentry);
3121                 if (they_are_dirs) {
3122                         drop_nlink(d_inode(new_dentry));
3123                         drop_nlink(old_dir);
3124                 }
3125         } else if (they_are_dirs) {
3126                 drop_nlink(old_dir);
3127                 inc_nlink(new_dir);
3128         }
3129
3130         old_dir->i_size -= BOGO_DIRENT_SIZE;
3131         new_dir->i_size += BOGO_DIRENT_SIZE;
3132         old_dir->i_ctime = old_dir->i_mtime =
3133         new_dir->i_ctime = new_dir->i_mtime =
3134         inode->i_ctime = current_time(old_dir);
3135         return 0;
3136 }
3137
3138 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3139 {
3140         int error;
3141         int len;
3142         struct inode *inode;
3143         struct page *page;
3144
3145         len = strlen(symname) + 1;
3146         if (len > PAGE_SIZE)
3147                 return -ENAMETOOLONG;
3148
3149         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3150                                 VM_NORESERVE);
3151         if (!inode)
3152                 return -ENOSPC;
3153
3154         error = security_inode_init_security(inode, dir, &dentry->d_name,
3155                                              shmem_initxattrs, NULL);
3156         if (error && error != -EOPNOTSUPP) {
3157                 iput(inode);
3158                 return error;
3159         }
3160
3161         inode->i_size = len-1;
3162         if (len <= SHORT_SYMLINK_LEN) {
3163                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3164                 if (!inode->i_link) {
3165                         iput(inode);
3166                         return -ENOMEM;
3167                 }
3168                 inode->i_op = &shmem_short_symlink_operations;
3169         } else {
3170                 inode_nohighmem(inode);
3171                 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3172                 if (error) {
3173                         iput(inode);
3174                         return error;
3175                 }
3176                 inode->i_mapping->a_ops = &shmem_aops;
3177                 inode->i_op = &shmem_symlink_inode_operations;
3178                 memcpy(page_address(page), symname, len);
3179                 SetPageUptodate(page);
3180                 set_page_dirty(page);
3181                 unlock_page(page);
3182                 put_page(page);
3183         }
3184         dir->i_size += BOGO_DIRENT_SIZE;
3185         dir->i_ctime = dir->i_mtime = current_time(dir);
3186         d_instantiate(dentry, inode);
3187         dget(dentry);
3188         return 0;
3189 }
3190
3191 static void shmem_put_link(void *arg)
3192 {
3193         mark_page_accessed(arg);
3194         put_page(arg);
3195 }
3196
3197 static const char *shmem_get_link(struct dentry *dentry,
3198                                   struct inode *inode,
3199                                   struct delayed_call *done)
3200 {
3201         struct page *page = NULL;
3202         int error;
3203         if (!dentry) {
3204                 page = find_get_page(inode->i_mapping, 0);
3205                 if (!page)
3206                         return ERR_PTR(-ECHILD);
3207                 if (!PageUptodate(page)) {
3208                         put_page(page);
3209                         return ERR_PTR(-ECHILD);
3210                 }
3211         } else {
3212                 error = shmem_getpage(inode, 0, &page, SGP_READ);
3213                 if (error)
3214                         return ERR_PTR(error);
3215                 unlock_page(page);
3216         }
3217         set_delayed_call(done, shmem_put_link, page);
3218         return page_address(page);
3219 }
3220
3221 #ifdef CONFIG_TMPFS_XATTR
3222 /*
3223  * Superblocks without xattr inode operations may get some security.* xattr
3224  * support from the LSM "for free". As soon as we have any other xattrs
3225  * like ACLs, we also need to implement the security.* handlers at
3226  * filesystem level, though.
3227  */
3228
3229 /*
3230  * Callback for security_inode_init_security() for acquiring xattrs.
3231  */
3232 static int shmem_initxattrs(struct inode *inode,
3233                             const struct xattr *xattr_array,
3234                             void *fs_info)
3235 {
3236         struct shmem_inode_info *info = SHMEM_I(inode);
3237         const struct xattr *xattr;
3238         struct simple_xattr *new_xattr;
3239         size_t len;
3240
3241         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3242                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3243                 if (!new_xattr)
3244                         return -ENOMEM;
3245
3246                 len = strlen(xattr->name) + 1;
3247                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3248                                           GFP_KERNEL);
3249                 if (!new_xattr->name) {
3250                         kvfree(new_xattr);
3251                         return -ENOMEM;
3252                 }
3253
3254                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3255                        XATTR_SECURITY_PREFIX_LEN);
3256                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3257                        xattr->name, len);
3258
3259                 simple_xattr_list_add(&info->xattrs, new_xattr);
3260         }
3261
3262         return 0;
3263 }
3264
3265 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3266                                    struct dentry *unused, struct inode *inode,
3267                                    const char *name, void *buffer, size_t size)
3268 {
3269         struct shmem_inode_info *info = SHMEM_I(inode);
3270
3271         name = xattr_full_name(handler, name);
3272         return simple_xattr_get(&info->xattrs, name, buffer, size);
3273 }
3274
3275 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3276                                    struct dentry *unused, struct inode *inode,
3277                                    const char *name, const void *value,
3278                                    size_t size, int flags)
3279 {
3280         struct shmem_inode_info *info = SHMEM_I(inode);
3281
3282         name = xattr_full_name(handler, name);
3283         return simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3284 }
3285
3286 static const struct xattr_handler shmem_security_xattr_handler = {
3287         .prefix = XATTR_SECURITY_PREFIX,
3288         .get = shmem_xattr_handler_get,
3289         .set = shmem_xattr_handler_set,
3290 };
3291
3292 static const struct xattr_handler shmem_trusted_xattr_handler = {
3293         .prefix = XATTR_TRUSTED_PREFIX,
3294         .get = shmem_xattr_handler_get,
3295         .set = shmem_xattr_handler_set,
3296 };
3297
3298 static const struct xattr_handler *shmem_xattr_handlers[] = {
3299 #ifdef CONFIG_TMPFS_POSIX_ACL
3300         &posix_acl_access_xattr_handler,
3301         &posix_acl_default_xattr_handler,
3302 #endif
3303         &shmem_security_xattr_handler,
3304         &shmem_trusted_xattr_handler,
3305         NULL
3306 };
3307
3308 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3309 {
3310         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3311         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3312 }
3313 #endif /* CONFIG_TMPFS_XATTR */
3314
3315 static const struct inode_operations shmem_short_symlink_operations = {
3316         .get_link       = simple_get_link,
3317 #ifdef CONFIG_TMPFS_XATTR
3318         .listxattr      = shmem_listxattr,
3319 #endif
3320 };
3321
3322 static const struct inode_operations shmem_symlink_inode_operations = {
3323         .get_link       = shmem_get_link,
3324 #ifdef CONFIG_TMPFS_XATTR
3325         .listxattr      = shmem_listxattr,
3326 #endif
3327 };
3328
3329 static struct dentry *shmem_get_parent(struct dentry *child)
3330 {
3331         return ERR_PTR(-ESTALE);
3332 }
3333
3334 static int shmem_match(struct inode *ino, void *vfh)
3335 {
3336         __u32 *fh = vfh;
3337         __u64 inum = fh[2];
3338         inum = (inum << 32) | fh[1];
3339         return ino->i_ino == inum && fh[0] == ino->i_generation;
3340 }
3341
3342 /* Find any alias of inode, but prefer a hashed alias */
3343 static struct dentry *shmem_find_alias(struct inode *inode)
3344 {
3345         struct dentry *alias = d_find_alias(inode);
3346
3347         return alias ?: d_find_any_alias(inode);
3348 }
3349
3350
3351 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3352                 struct fid *fid, int fh_len, int fh_type)
3353 {
3354         struct inode *inode;
3355         struct dentry *dentry = NULL;
3356         u64 inum;
3357
3358         if (fh_len < 3)
3359                 return NULL;
3360
3361         inum = fid->raw[2];
3362         inum = (inum << 32) | fid->raw[1];
3363
3364         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3365                         shmem_match, fid->raw);
3366         if (inode) {
3367                 dentry = shmem_find_alias(inode);
3368                 iput(inode);
3369         }
3370
3371         return dentry;
3372 }
3373
3374 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3375                                 struct inode *parent)
3376 {
3377         if (*len < 3) {
3378                 *len = 3;
3379                 return FILEID_INVALID;
3380         }
3381
3382         if (inode_unhashed(inode)) {
3383                 /* Unfortunately insert_inode_hash is not idempotent,
3384                  * so as we hash inodes here rather than at creation
3385                  * time, we need a lock to ensure we only try
3386                  * to do it once
3387                  */
3388                 static DEFINE_SPINLOCK(lock);
3389                 spin_lock(&lock);
3390                 if (inode_unhashed(inode))
3391                         __insert_inode_hash(inode,
3392                                             inode->i_ino + inode->i_generation);
3393                 spin_unlock(&lock);
3394         }
3395
3396         fh[0] = inode->i_generation;
3397         fh[1] = inode->i_ino;
3398         fh[2] = ((__u64)inode->i_ino) >> 32;
3399
3400         *len = 3;
3401         return 1;
3402 }
3403
3404 static const struct export_operations shmem_export_ops = {
3405         .get_parent     = shmem_get_parent,
3406         .encode_fh      = shmem_encode_fh,
3407         .fh_to_dentry   = shmem_fh_to_dentry,
3408 };
3409
3410 enum shmem_param {
3411         Opt_gid,
3412         Opt_huge,
3413         Opt_mode,
3414         Opt_mpol,
3415         Opt_nr_blocks,
3416         Opt_nr_inodes,
3417         Opt_size,
3418         Opt_uid,
3419         Opt_inode32,
3420         Opt_inode64,
3421 };
3422
3423 static const struct constant_table shmem_param_enums_huge[] = {
3424         {"never",       SHMEM_HUGE_NEVER },
3425         {"always",      SHMEM_HUGE_ALWAYS },
3426         {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3427         {"advise",      SHMEM_HUGE_ADVISE },
3428         {}
3429 };
3430
3431 const struct fs_parameter_spec shmem_fs_parameters[] = {
3432         fsparam_u32   ("gid",           Opt_gid),
3433         fsparam_enum  ("huge",          Opt_huge,  shmem_param_enums_huge),
3434         fsparam_u32oct("mode",          Opt_mode),
3435         fsparam_string("mpol",          Opt_mpol),
3436         fsparam_string("nr_blocks",     Opt_nr_blocks),
3437         fsparam_string("nr_inodes",     Opt_nr_inodes),
3438         fsparam_string("size",          Opt_size),
3439         fsparam_u32   ("uid",           Opt_uid),
3440         fsparam_flag  ("inode32",       Opt_inode32),
3441         fsparam_flag  ("inode64",       Opt_inode64),
3442         {}
3443 };
3444
3445 static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3446 {
3447         struct shmem_options *ctx = fc->fs_private;
3448         struct fs_parse_result result;
3449         unsigned long long size;
3450         char *rest;
3451         int opt;
3452
3453         opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3454         if (opt < 0)
3455                 return opt;
3456
3457         switch (opt) {
3458         case Opt_size:
3459                 size = memparse(param->string, &rest);
3460                 if (*rest == '%') {
3461                         size <<= PAGE_SHIFT;
3462                         size *= totalram_pages();
3463                         do_div(size, 100);
3464                         rest++;
3465                 }
3466                 if (*rest)
3467                         goto bad_value;
3468                 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3469                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3470                 break;
3471         case Opt_nr_blocks:
3472                 ctx->blocks = memparse(param->string, &rest);
3473                 if (*rest)
3474                         goto bad_value;
3475                 ctx->seen |= SHMEM_SEEN_BLOCKS;
3476                 break;
3477         case Opt_nr_inodes:
3478                 ctx->inodes = memparse(param->string, &rest);
3479                 if (*rest)
3480                         goto bad_value;
3481                 ctx->seen |= SHMEM_SEEN_INODES;
3482                 break;
3483         case Opt_mode:
3484                 ctx->mode = result.uint_32 & 07777;
3485                 break;
3486         case Opt_uid:
3487                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
3488                 if (!uid_valid(ctx->uid))
3489                         goto bad_value;
3490                 break;
3491         case Opt_gid:
3492                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
3493                 if (!gid_valid(ctx->gid))
3494                         goto bad_value;
3495                 break;
3496         case Opt_huge:
3497                 ctx->huge = result.uint_32;
3498                 if (ctx->huge != SHMEM_HUGE_NEVER &&
3499                     !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3500                       has_transparent_hugepage()))
3501                         goto unsupported_parameter;
3502                 ctx->seen |= SHMEM_SEEN_HUGE;
3503                 break;
3504         case Opt_mpol:
3505                 if (IS_ENABLED(CONFIG_NUMA)) {
3506                         mpol_put(ctx->mpol);
3507                         ctx->mpol = NULL;
3508                         if (mpol_parse_str(param->string, &ctx->mpol))
3509                                 goto bad_value;
3510                         break;
3511                 }
3512                 goto unsupported_parameter;
3513         case Opt_inode32:
3514                 ctx->full_inums = false;
3515                 ctx->seen |= SHMEM_SEEN_INUMS;
3516                 break;
3517         case Opt_inode64:
3518                 if (sizeof(ino_t) < 8) {
3519                         return invalfc(fc,
3520                                        "Cannot use inode64 with <64bit inums in kernel\n");
3521                 }
3522                 ctx->full_inums = true;
3523                 ctx->seen |= SHMEM_SEEN_INUMS;
3524                 break;
3525         }
3526         return 0;
3527
3528 unsupported_parameter:
3529         return invalfc(fc, "Unsupported parameter '%s'", param->key);
3530 bad_value:
3531         return invalfc(fc, "Bad value for '%s'", param->key);
3532 }
3533
3534 static int shmem_parse_options(struct fs_context *fc, void *data)
3535 {
3536         char *options = data;
3537
3538         if (options) {
3539                 int err = security_sb_eat_lsm_opts(options, &fc->security);
3540                 if (err)
3541                         return err;
3542         }
3543
3544         while (options != NULL) {
3545                 char *this_char = options;
3546                 for (;;) {
3547                         /*
3548                          * NUL-terminate this option: unfortunately,
3549                          * mount options form a comma-separated list,
3550                          * but mpol's nodelist may also contain commas.
3551                          */
3552                         options = strchr(options, ',');
3553                         if (options == NULL)
3554                                 break;
3555                         options++;
3556                         if (!isdigit(*options)) {
3557                                 options[-1] = '\0';
3558                                 break;
3559                         }
3560                 }
3561                 if (*this_char) {
3562                         char *value = strchr(this_char,'=');
3563                         size_t len = 0;
3564                         int err;
3565
3566                         if (value) {
3567                                 *value++ = '\0';
3568                                 len = strlen(value);
3569                         }
3570                         err = vfs_parse_fs_string(fc, this_char, value, len);
3571                         if (err < 0)
3572                                 return err;
3573                 }
3574         }
3575         return 0;
3576 }
3577
3578 /*
3579  * Reconfigure a shmem filesystem.
3580  *
3581  * Note that we disallow change from limited->unlimited blocks/inodes while any
3582  * are in use; but we must separately disallow unlimited->limited, because in
3583  * that case we have no record of how much is already in use.
3584  */
3585 static int shmem_reconfigure(struct fs_context *fc)
3586 {
3587         struct shmem_options *ctx = fc->fs_private;
3588         struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3589         unsigned long inodes;
3590         const char *err;
3591
3592         spin_lock(&sbinfo->stat_lock);
3593         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3594         if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3595                 if (!sbinfo->max_blocks) {
3596                         err = "Cannot retroactively limit size";
3597                         goto out;
3598                 }
3599                 if (percpu_counter_compare(&sbinfo->used_blocks,
3600                                            ctx->blocks) > 0) {
3601                         err = "Too small a size for current use";
3602                         goto out;
3603                 }
3604         }
3605         if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3606                 if (!sbinfo->max_inodes) {
3607                         err = "Cannot retroactively limit inodes";
3608                         goto out;
3609                 }
3610                 if (ctx->inodes < inodes) {
3611                         err = "Too few inodes for current use";
3612                         goto out;
3613                 }
3614         }
3615
3616         if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3617             sbinfo->next_ino > UINT_MAX) {
3618                 err = "Current inum too high to switch to 32-bit inums";
3619                 goto out;
3620         }
3621
3622         if (ctx->seen & SHMEM_SEEN_HUGE)
3623                 sbinfo->huge = ctx->huge;
3624         if (ctx->seen & SHMEM_SEEN_INUMS)
3625                 sbinfo->full_inums = ctx->full_inums;
3626         if (ctx->seen & SHMEM_SEEN_BLOCKS)
3627                 sbinfo->max_blocks  = ctx->blocks;
3628         if (ctx->seen & SHMEM_SEEN_INODES) {
3629                 sbinfo->max_inodes  = ctx->inodes;
3630                 sbinfo->free_inodes = ctx->inodes - inodes;
3631         }
3632
3633         /*
3634          * Preserve previous mempolicy unless mpol remount option was specified.
3635          */
3636         if (ctx->mpol) {
3637                 mpol_put(sbinfo->mpol);
3638                 sbinfo->mpol = ctx->mpol;       /* transfers initial ref */
3639                 ctx->mpol = NULL;
3640         }
3641         spin_unlock(&sbinfo->stat_lock);
3642         return 0;
3643 out:
3644         spin_unlock(&sbinfo->stat_lock);
3645         return invalfc(fc, "%s", err);
3646 }
3647
3648 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3649 {
3650         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3651
3652         if (sbinfo->max_blocks != shmem_default_max_blocks())
3653                 seq_printf(seq, ",size=%luk",
3654                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3655         if (sbinfo->max_inodes != shmem_default_max_inodes())
3656                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3657         if (sbinfo->mode != (0777 | S_ISVTX))
3658                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3659         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3660                 seq_printf(seq, ",uid=%u",
3661                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3662         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3663                 seq_printf(seq, ",gid=%u",
3664                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3665
3666         /*
3667          * Showing inode{64,32} might be useful even if it's the system default,
3668          * since then people don't have to resort to checking both here and
3669          * /proc/config.gz to confirm 64-bit inums were successfully applied
3670          * (which may not even exist if IKCONFIG_PROC isn't enabled).
3671          *
3672          * We hide it when inode64 isn't the default and we are using 32-bit
3673          * inodes, since that probably just means the feature isn't even under
3674          * consideration.
3675          *
3676          * As such:
3677          *
3678          *                     +-----------------+-----------------+
3679          *                     | TMPFS_INODE64=y | TMPFS_INODE64=n |
3680          *  +------------------+-----------------+-----------------+
3681          *  | full_inums=true  | show            | show            |
3682          *  | full_inums=false | show            | hide            |
3683          *  +------------------+-----------------+-----------------+
3684          *
3685          */
3686         if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3687                 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3688 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3689         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3690         if (sbinfo->huge)
3691                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3692 #endif
3693         shmem_show_mpol(seq, sbinfo->mpol);
3694         return 0;
3695 }
3696
3697 #endif /* CONFIG_TMPFS */
3698
3699 static void shmem_put_super(struct super_block *sb)
3700 {
3701         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3702
3703         free_percpu(sbinfo->ino_batch);
3704         percpu_counter_destroy(&sbinfo->used_blocks);
3705         mpol_put(sbinfo->mpol);
3706         kfree(sbinfo);
3707         sb->s_fs_info = NULL;
3708 }
3709
3710 static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3711 {
3712         struct shmem_options *ctx = fc->fs_private;
3713         struct inode *inode;
3714         struct shmem_sb_info *sbinfo;
3715         int err = -ENOMEM;
3716
3717         /* Round up to L1_CACHE_BYTES to resist false sharing */
3718         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3719                                 L1_CACHE_BYTES), GFP_KERNEL);
3720         if (!sbinfo)
3721                 return -ENOMEM;
3722
3723         sb->s_fs_info = sbinfo;
3724
3725 #ifdef CONFIG_TMPFS
3726         /*
3727          * Per default we only allow half of the physical ram per
3728          * tmpfs instance, limiting inodes to one per page of lowmem;
3729          * but the internal instance is left unlimited.
3730          */
3731         if (!(sb->s_flags & SB_KERNMOUNT)) {
3732                 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3733                         ctx->blocks = shmem_default_max_blocks();
3734                 if (!(ctx->seen & SHMEM_SEEN_INODES))
3735                         ctx->inodes = shmem_default_max_inodes();
3736                 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3737                         ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3738         } else {
3739                 sb->s_flags |= SB_NOUSER;
3740         }
3741         sb->s_export_op = &shmem_export_ops;
3742         sb->s_flags |= SB_NOSEC;
3743 #else
3744         sb->s_flags |= SB_NOUSER;
3745 #endif
3746         sbinfo->max_blocks = ctx->blocks;
3747         sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3748         if (sb->s_flags & SB_KERNMOUNT) {
3749                 sbinfo->ino_batch = alloc_percpu(ino_t);
3750                 if (!sbinfo->ino_batch)
3751                         goto failed;
3752         }
3753         sbinfo->uid = ctx->uid;
3754         sbinfo->gid = ctx->gid;
3755         sbinfo->full_inums = ctx->full_inums;
3756         sbinfo->mode = ctx->mode;
3757         sbinfo->huge = ctx->huge;
3758         sbinfo->mpol = ctx->mpol;
3759         ctx->mpol = NULL;
3760
3761         spin_lock_init(&sbinfo->stat_lock);
3762         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3763                 goto failed;
3764         spin_lock_init(&sbinfo->shrinklist_lock);
3765         INIT_LIST_HEAD(&sbinfo->shrinklist);
3766
3767         sb->s_maxbytes = MAX_LFS_FILESIZE;
3768         sb->s_blocksize = PAGE_SIZE;
3769         sb->s_blocksize_bits = PAGE_SHIFT;
3770         sb->s_magic = TMPFS_MAGIC;
3771         sb->s_op = &shmem_ops;
3772         sb->s_time_gran = 1;
3773 #ifdef CONFIG_TMPFS_XATTR
3774         sb->s_xattr = shmem_xattr_handlers;
3775 #endif
3776 #ifdef CONFIG_TMPFS_POSIX_ACL
3777         sb->s_flags |= SB_POSIXACL;
3778 #endif
3779         uuid_gen(&sb->s_uuid);
3780
3781         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3782         if (!inode)
3783                 goto failed;
3784         inode->i_uid = sbinfo->uid;
3785         inode->i_gid = sbinfo->gid;
3786         sb->s_root = d_make_root(inode);
3787         if (!sb->s_root)
3788                 goto failed;
3789         return 0;
3790
3791 failed:
3792         shmem_put_super(sb);
3793         return err;
3794 }
3795
3796 static int shmem_get_tree(struct fs_context *fc)
3797 {
3798         return get_tree_nodev(fc, shmem_fill_super);
3799 }
3800
3801 static void shmem_free_fc(struct fs_context *fc)
3802 {
3803         struct shmem_options *ctx = fc->fs_private;
3804
3805         if (ctx) {
3806                 mpol_put(ctx->mpol);
3807                 kfree(ctx);
3808         }
3809 }
3810
3811 static const struct fs_context_operations shmem_fs_context_ops = {
3812         .free                   = shmem_free_fc,
3813         .get_tree               = shmem_get_tree,
3814 #ifdef CONFIG_TMPFS
3815         .parse_monolithic       = shmem_parse_options,
3816         .parse_param            = shmem_parse_one,
3817         .reconfigure            = shmem_reconfigure,
3818 #endif
3819 };
3820
3821 static struct kmem_cache *shmem_inode_cachep;
3822
3823 static struct inode *shmem_alloc_inode(struct super_block *sb)
3824 {
3825         struct shmem_inode_info *info;
3826         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3827         if (!info)
3828                 return NULL;
3829         return &info->vfs_inode;
3830 }
3831
3832 static void shmem_free_in_core_inode(struct inode *inode)
3833 {
3834         if (S_ISLNK(inode->i_mode))
3835                 kfree(inode->i_link);
3836         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3837 }
3838
3839 static void shmem_destroy_inode(struct inode *inode)
3840 {
3841         if (S_ISREG(inode->i_mode))
3842                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3843 }
3844
3845 static void shmem_init_inode(void *foo)
3846 {
3847         struct shmem_inode_info *info = foo;
3848         inode_init_once(&info->vfs_inode);
3849 }
3850
3851 static void shmem_init_inodecache(void)
3852 {
3853         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3854                                 sizeof(struct shmem_inode_info),
3855                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3856 }
3857
3858 static void shmem_destroy_inodecache(void)
3859 {
3860         kmem_cache_destroy(shmem_inode_cachep);
3861 }
3862
3863 const struct address_space_operations shmem_aops = {
3864         .writepage      = shmem_writepage,
3865         .set_page_dirty = __set_page_dirty_no_writeback,
3866 #ifdef CONFIG_TMPFS
3867         .write_begin    = shmem_write_begin,
3868         .write_end      = shmem_write_end,
3869 #endif
3870 #ifdef CONFIG_MIGRATION
3871         .migratepage    = migrate_page,
3872 #endif
3873         .error_remove_page = generic_error_remove_page,
3874 };
3875 EXPORT_SYMBOL(shmem_aops);
3876
3877 static const struct file_operations shmem_file_operations = {
3878         .mmap           = shmem_mmap,
3879         .get_unmapped_area = shmem_get_unmapped_area,
3880 #ifdef CONFIG_TMPFS
3881         .llseek         = shmem_file_llseek,
3882         .read_iter      = shmem_file_read_iter,
3883         .write_iter     = generic_file_write_iter,
3884         .fsync          = noop_fsync,
3885         .splice_read    = generic_file_splice_read,
3886         .splice_write   = iter_file_splice_write,
3887         .fallocate      = shmem_fallocate,
3888 #endif
3889 };
3890
3891 static const struct inode_operations shmem_inode_operations = {
3892         .getattr        = shmem_getattr,
3893         .setattr        = shmem_setattr,
3894 #ifdef CONFIG_TMPFS_XATTR
3895         .listxattr      = shmem_listxattr,
3896         .set_acl        = simple_set_acl,
3897 #endif
3898 };
3899
3900 static const struct inode_operations shmem_dir_inode_operations = {
3901 #ifdef CONFIG_TMPFS
3902         .create         = shmem_create,
3903         .lookup         = simple_lookup,
3904         .link           = shmem_link,
3905         .unlink         = shmem_unlink,
3906         .symlink        = shmem_symlink,
3907         .mkdir          = shmem_mkdir,
3908         .rmdir          = shmem_rmdir,
3909         .mknod          = shmem_mknod,
3910         .rename         = shmem_rename2,
3911         .tmpfile        = shmem_tmpfile,
3912 #endif
3913 #ifdef CONFIG_TMPFS_XATTR
3914         .listxattr      = shmem_listxattr,
3915 #endif
3916 #ifdef CONFIG_TMPFS_POSIX_ACL
3917         .setattr        = shmem_setattr,
3918         .set_acl        = simple_set_acl,
3919 #endif
3920 };
3921
3922 static const struct inode_operations shmem_special_inode_operations = {
3923 #ifdef CONFIG_TMPFS_XATTR
3924         .listxattr      = shmem_listxattr,
3925 #endif
3926 #ifdef CONFIG_TMPFS_POSIX_ACL
3927         .setattr        = shmem_setattr,
3928         .set_acl        = simple_set_acl,
3929 #endif
3930 };
3931
3932 static const struct super_operations shmem_ops = {
3933         .alloc_inode    = shmem_alloc_inode,
3934         .free_inode     = shmem_free_in_core_inode,
3935         .destroy_inode  = shmem_destroy_inode,
3936 #ifdef CONFIG_TMPFS
3937         .statfs         = shmem_statfs,
3938         .show_options   = shmem_show_options,
3939 #endif
3940         .evict_inode    = shmem_evict_inode,
3941         .drop_inode     = generic_delete_inode,
3942         .put_super      = shmem_put_super,
3943 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3944         .nr_cached_objects      = shmem_unused_huge_count,
3945         .free_cached_objects    = shmem_unused_huge_scan,
3946 #endif
3947 };
3948
3949 static const struct vm_operations_struct shmem_vm_ops = {
3950         .fault          = shmem_fault,
3951         .map_pages      = filemap_map_pages,
3952 #ifdef CONFIG_NUMA
3953         .set_policy     = shmem_set_policy,
3954         .get_policy     = shmem_get_policy,
3955 #endif
3956 };
3957
3958 int shmem_init_fs_context(struct fs_context *fc)
3959 {
3960         struct shmem_options *ctx;
3961
3962         ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
3963         if (!ctx)
3964                 return -ENOMEM;
3965
3966         ctx->mode = 0777 | S_ISVTX;
3967         ctx->uid = current_fsuid();
3968         ctx->gid = current_fsgid();
3969
3970         fc->fs_private = ctx;
3971         fc->ops = &shmem_fs_context_ops;
3972         return 0;
3973 }
3974
3975 static struct file_system_type shmem_fs_type = {
3976         .owner          = THIS_MODULE,
3977         .name           = "tmpfs",
3978         .init_fs_context = shmem_init_fs_context,
3979 #ifdef CONFIG_TMPFS
3980         .parameters     = shmem_fs_parameters,
3981 #endif
3982         .kill_sb        = kill_litter_super,
3983         .fs_flags       = FS_USERNS_MOUNT | FS_THP_SUPPORT,
3984 };
3985
3986 int __init shmem_init(void)
3987 {
3988         int error;
3989
3990         shmem_init_inodecache();
3991
3992         error = register_filesystem(&shmem_fs_type);
3993         if (error) {
3994                 pr_err("Could not register tmpfs\n");
3995                 goto out2;
3996         }
3997
3998         shm_mnt = kern_mount(&shmem_fs_type);
3999         if (IS_ERR(shm_mnt)) {
4000                 error = PTR_ERR(shm_mnt);
4001                 pr_err("Could not kern_mount tmpfs\n");
4002                 goto out1;
4003         }
4004
4005 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4006         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4007                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4008         else
4009                 shmem_huge = 0; /* just in case it was patched */
4010 #endif
4011         return 0;
4012
4013 out1:
4014         unregister_filesystem(&shmem_fs_type);
4015 out2:
4016         shmem_destroy_inodecache();
4017         shm_mnt = ERR_PTR(error);
4018         return error;
4019 }
4020
4021 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4022 static ssize_t shmem_enabled_show(struct kobject *kobj,
4023                                   struct kobj_attribute *attr, char *buf)
4024 {
4025         static const int values[] = {
4026                 SHMEM_HUGE_ALWAYS,
4027                 SHMEM_HUGE_WITHIN_SIZE,
4028                 SHMEM_HUGE_ADVISE,
4029                 SHMEM_HUGE_NEVER,
4030                 SHMEM_HUGE_DENY,
4031                 SHMEM_HUGE_FORCE,
4032         };
4033         int len = 0;
4034         int i;
4035
4036         for (i = 0; i < ARRAY_SIZE(values); i++) {
4037                 len += sysfs_emit_at(buf, len,
4038                                      shmem_huge == values[i] ? "%s[%s]" : "%s%s",
4039                                      i ? " " : "",
4040                                      shmem_format_huge(values[i]));
4041         }
4042
4043         len += sysfs_emit_at(buf, len, "\n");
4044
4045         return len;
4046 }
4047
4048 static ssize_t shmem_enabled_store(struct kobject *kobj,
4049                 struct kobj_attribute *attr, const char *buf, size_t count)
4050 {
4051         char tmp[16];
4052         int huge;
4053
4054         if (count + 1 > sizeof(tmp))
4055                 return -EINVAL;
4056         memcpy(tmp, buf, count);
4057         tmp[count] = '\0';
4058         if (count && tmp[count - 1] == '\n')
4059                 tmp[count - 1] = '\0';
4060
4061         huge = shmem_parse_huge(tmp);
4062         if (huge == -EINVAL)
4063                 return -EINVAL;
4064         if (!has_transparent_hugepage() &&
4065                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4066                 return -EINVAL;
4067
4068         shmem_huge = huge;
4069         if (shmem_huge > SHMEM_HUGE_DENY)
4070                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4071         return count;
4072 }
4073
4074 struct kobj_attribute shmem_enabled_attr =
4075         __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
4076 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4077
4078 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4079 bool shmem_huge_enabled(struct vm_area_struct *vma)
4080 {
4081         struct inode *inode = file_inode(vma->vm_file);
4082         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
4083         loff_t i_size;
4084         pgoff_t off;
4085
4086         if ((vma->vm_flags & VM_NOHUGEPAGE) ||
4087             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
4088                 return false;
4089         if (shmem_huge == SHMEM_HUGE_FORCE)
4090                 return true;
4091         if (shmem_huge == SHMEM_HUGE_DENY)
4092                 return false;
4093         switch (sbinfo->huge) {
4094                 case SHMEM_HUGE_NEVER:
4095                         return false;
4096                 case SHMEM_HUGE_ALWAYS:
4097                         return true;
4098                 case SHMEM_HUGE_WITHIN_SIZE:
4099                         off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
4100                         i_size = round_up(i_size_read(inode), PAGE_SIZE);
4101                         if (i_size >= HPAGE_PMD_SIZE &&
4102                                         i_size >> PAGE_SHIFT >= off)
4103                                 return true;
4104                         fallthrough;
4105                 case SHMEM_HUGE_ADVISE:
4106                         /* TODO: implement fadvise() hints */
4107                         return (vma->vm_flags & VM_HUGEPAGE);
4108                 default:
4109                         VM_BUG_ON(1);
4110                         return false;
4111         }
4112 }
4113 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4114
4115 #else /* !CONFIG_SHMEM */
4116
4117 /*
4118  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4119  *
4120  * This is intended for small system where the benefits of the full
4121  * shmem code (swap-backed and resource-limited) are outweighed by
4122  * their complexity. On systems without swap this code should be
4123  * effectively equivalent, but much lighter weight.
4124  */
4125
4126 static struct file_system_type shmem_fs_type = {
4127         .name           = "tmpfs",
4128         .init_fs_context = ramfs_init_fs_context,
4129         .parameters     = ramfs_fs_parameters,
4130         .kill_sb        = kill_litter_super,
4131         .fs_flags       = FS_USERNS_MOUNT,
4132 };
4133
4134 int __init shmem_init(void)
4135 {
4136         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4137
4138         shm_mnt = kern_mount(&shmem_fs_type);
4139         BUG_ON(IS_ERR(shm_mnt));
4140
4141         return 0;
4142 }
4143
4144 int shmem_unuse(unsigned int type, bool frontswap,
4145                 unsigned long *fs_pages_to_unuse)
4146 {
4147         return 0;
4148 }
4149
4150 int shmem_lock(struct file *file, int lock, struct user_struct *user)
4151 {
4152         return 0;
4153 }
4154
4155 void shmem_unlock_mapping(struct address_space *mapping)
4156 {
4157 }
4158
4159 #ifdef CONFIG_MMU
4160 unsigned long shmem_get_unmapped_area(struct file *file,
4161                                       unsigned long addr, unsigned long len,
4162                                       unsigned long pgoff, unsigned long flags)
4163 {
4164         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4165 }
4166 #endif
4167
4168 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4169 {
4170         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4171 }
4172 EXPORT_SYMBOL_GPL(shmem_truncate_range);
4173
4174 #define shmem_vm_ops                            generic_file_vm_ops
4175 #define shmem_file_operations                   ramfs_file_operations
4176 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
4177 #define shmem_acct_size(flags, size)            0
4178 #define shmem_unacct_size(flags, size)          do {} while (0)
4179
4180 #endif /* CONFIG_SHMEM */
4181
4182 /* common code */
4183
4184 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4185                                        unsigned long flags, unsigned int i_flags)
4186 {
4187         struct inode *inode;
4188         struct file *res;
4189
4190         if (IS_ERR(mnt))
4191                 return ERR_CAST(mnt);
4192
4193         if (size < 0 || size > MAX_LFS_FILESIZE)
4194                 return ERR_PTR(-EINVAL);
4195
4196         if (shmem_acct_size(flags, size))
4197                 return ERR_PTR(-ENOMEM);
4198
4199         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4200                                 flags);
4201         if (unlikely(!inode)) {
4202                 shmem_unacct_size(flags, size);
4203                 return ERR_PTR(-ENOSPC);
4204         }
4205         inode->i_flags |= i_flags;
4206         inode->i_size = size;
4207         clear_nlink(inode);     /* It is unlinked */
4208         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4209         if (!IS_ERR(res))
4210                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4211                                 &shmem_file_operations);
4212         if (IS_ERR(res))
4213                 iput(inode);
4214         return res;
4215 }
4216
4217 /**
4218  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4219  *      kernel internal.  There will be NO LSM permission checks against the
4220  *      underlying inode.  So users of this interface must do LSM checks at a
4221  *      higher layer.  The users are the big_key and shm implementations.  LSM
4222  *      checks are provided at the key or shm level rather than the inode.
4223  * @name: name for dentry (to be seen in /proc/<pid>/maps
4224  * @size: size to be set for the file
4225  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4226  */
4227 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4228 {
4229         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4230 }
4231
4232 /**
4233  * shmem_file_setup - get an unlinked file living in tmpfs
4234  * @name: name for dentry (to be seen in /proc/<pid>/maps
4235  * @size: size to be set for the file
4236  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4237  */
4238 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4239 {
4240         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4241 }
4242 EXPORT_SYMBOL_GPL(shmem_file_setup);
4243
4244 /**
4245  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4246  * @mnt: the tmpfs mount where the file will be created
4247  * @name: name for dentry (to be seen in /proc/<pid>/maps
4248  * @size: size to be set for the file
4249  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4250  */
4251 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4252                                        loff_t size, unsigned long flags)
4253 {
4254         return __shmem_file_setup(mnt, name, size, flags, 0);
4255 }
4256 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4257
4258 /**
4259  * shmem_zero_setup - setup a shared anonymous mapping
4260  * @vma: the vma to be mmapped is prepared by do_mmap
4261  */
4262 int shmem_zero_setup(struct vm_area_struct *vma)
4263 {
4264         struct file *file;
4265         loff_t size = vma->vm_end - vma->vm_start;
4266
4267         /*
4268          * Cloning a new file under mmap_lock leads to a lock ordering conflict
4269          * between XFS directory reading and selinux: since this file is only
4270          * accessible to the user through its mapping, use S_PRIVATE flag to
4271          * bypass file security, in the same way as shmem_kernel_file_setup().
4272          */
4273         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4274         if (IS_ERR(file))
4275                 return PTR_ERR(file);
4276
4277         if (vma->vm_file)
4278                 fput(vma->vm_file);
4279         vma->vm_file = file;
4280         vma->vm_ops = &shmem_vm_ops;
4281
4282         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
4283                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4284                         (vma->vm_end & HPAGE_PMD_MASK)) {
4285                 khugepaged_enter(vma, vma->vm_flags);
4286         }
4287
4288         return 0;
4289 }
4290
4291 /**
4292  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4293  * @mapping:    the page's address_space
4294  * @index:      the page index
4295  * @gfp:        the page allocator flags to use if allocating
4296  *
4297  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4298  * with any new page allocations done using the specified allocation flags.
4299  * But read_cache_page_gfp() uses the ->readpage() method: which does not
4300  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4301  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4302  *
4303  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4304  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4305  */
4306 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4307                                          pgoff_t index, gfp_t gfp)
4308 {
4309 #ifdef CONFIG_SHMEM
4310         struct inode *inode = mapping->host;
4311         struct page *page;
4312         int error;
4313
4314         BUG_ON(!shmem_mapping(mapping));
4315         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4316                                   gfp, NULL, NULL, NULL);
4317         if (error)
4318                 page = ERR_PTR(error);
4319         else
4320                 unlock_page(page);
4321         return page;
4322 #else
4323         /*
4324          * The tiny !SHMEM case uses ramfs without swap
4325          */
4326         return read_cache_page_gfp(mapping, index, gfp);
4327 #endif
4328 }
4329 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);