fs/ext4/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/fs/ext4/file.c
   4  *
   5  * Copyright (C) 1992, 1993, 1994, 1995
   6  * Remy Card ([email protected])
   7  * Laboratoire MASI - Institut Blaise Pascal
   8  * Universite Pierre et Marie Curie (Paris VI)
   9  *
  10  *  from
  11  *
  12  *  linux/fs/minix/file.c
  13  *
  14  *  Copyright (C) 1991, 1992  Linus Torvalds
  15  *
  16  *  ext4 fs regular file handling primitives
  17  *
  18  *  64-bit file support on 64-bit platforms by Jakub Jelinek
  19  *      ([email protected])
  20  */
  21
  22 #include <linux/time.h>
  23 #include <linux/fs.h>
  24 #include <linux/iomap.h>
  25 #include <linux/mount.h>
  26 #include <linux/path.h>
  27 #include <linux/dax.h>
  28 #include <linux/quotaops.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/uio.h>
  31 #include <linux/mman.h>
  32 #include <linux/backing-dev.h>
  33 #include "ext4.h"
  34 #include "ext4_jbd2.h"
  35 #include "xattr.h"
  36 #include "acl.h"
  37 #include "truncate.h"
  38
  39 static bool ext4_dio_supported(struct inode *inode)
  40 {
  41         if (IS_ENABLED(CONFIG_FS_ENCRYPTION) && IS_ENCRYPTED(inode))
  42                 return false;
  43         if (fsverity_active(inode))
  44                 return false;
  45         if (ext4_should_journal_data(inode))
  46                 return false;
  47         if (ext4_has_inline_data(inode))
  48                 return false;
  49         return true;
  50 }
  51
  52 static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
  53 {
  54         ssize_t ret;
  55         struct inode *inode = file_inode(iocb->ki_filp);
  56
  57         if (iocb->ki_flags & IOCB_NOWAIT) {
  58                 if (!inode_trylock_shared(inode))
  59                         return -EAGAIN;
  60         } else {
  61                 inode_lock_shared(inode);
  62         }
  63
  64         if (!ext4_dio_supported(inode)) {
  65                 inode_unlock_shared(inode);
  66                 /*
  67                  * Fallback to buffered I/O if the operation being performed on
  68                  * the inode is not supported by direct I/O. The IOCB_DIRECT
  69                  * flag needs to be cleared here in order to ensure that the
  70                  * direct I/O path within generic_file_read_iter() is not
  71                  * taken.
  72                  */
  73                 iocb->ki_flags &= ~IOCB_DIRECT;
  74                 return generic_file_read_iter(iocb, to);
  75         }
  76
  77         ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL, 0);
  78         inode_unlock_shared(inode);
  79
  80         file_accessed(iocb->ki_filp);
  81         return ret;
  82 }
  83
  84 #ifdef CONFIG_FS_DAX
  85 static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
  86 {
  87         struct inode *inode = file_inode(iocb->ki_filp);
  88         ssize_t ret;
  89
  90         if (iocb->ki_flags & IOCB_NOWAIT) {
  91                 if (!inode_trylock_shared(inode))
  92                         return -EAGAIN;
  93         } else {
  94                 inode_lock_shared(inode);
  95         }
  96         /*
  97          * Recheck under inode lock - at this point we are sure it cannot
  98          * change anymore
  99          */
 100         if (!IS_DAX(inode)) {
 101                 inode_unlock_shared(inode);
 102                 /* Fallback to buffered IO in case we cannot support DAX */
 103                 return generic_file_read_iter(iocb, to);
 104         }
 105         ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
 106         inode_unlock_shared(inode);
 107
 108         file_accessed(iocb->ki_filp);
 109         return ret;
 110 }
 111 #endif
 112
 113 static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 114 {
 115         struct inode *inode = file_inode(iocb->ki_filp);
 116
 117         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 118                 return -EIO;
 119
 120         if (!iov_iter_count(to))
 121                 return 0; /* skip atime */
 122
 123 #ifdef CONFIG_FS_DAX
 124         if (IS_DAX(inode))
 125                 return ext4_dax_read_iter(iocb, to);
 126 #endif
 127         if (iocb->ki_flags & IOCB_DIRECT)
 128                 return ext4_dio_read_iter(iocb, to);
 129
 130         return generic_file_read_iter(iocb, to);
 131 }
 132
 133 /*
 134  * Called when an inode is released. Note that this is different
 135  * from ext4_file_open: open gets called at every open, but release
 136  * gets called only when /all/ the files are closed.
 137  */
 138 static int ext4_release_file(struct inode *inode, struct file *filp)
 139 {
 140         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
 141                 ext4_alloc_da_blocks(inode);
 142                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
 143         }
 144         /* if we are the last writer on the inode, drop the block reservation */
 145         if ((filp->f_mode & FMODE_WRITE) &&
 146                         (atomic_read(&inode->i_writecount) == 1) &&
 147                         !EXT4_I(inode)->i_reserved_data_blocks) {
 148                 down_write(&EXT4_I(inode)->i_data_sem);
 149                 ext4_discard_preallocations(inode, 0);
 150                 up_write(&EXT4_I(inode)->i_data_sem);
 151         }
 152         if (is_dx(inode) && filp->private_data)
 153                 ext4_htree_free_dir_info(filp->private_data);
 154
 155         return 0;
 156 }
 157
 158 /*
 159  * This tests whether the IO in question is block-aligned or not.
 160  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
 161  * are converted to written only after the IO is complete.  Until they are
 162  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
 163  * it needs to zero out portions of the start and/or end block.  If 2 AIO
 164  * threads are at work on the same unwritten block, they must be synchronized
 165  * or one thread will zero the other's data, causing corruption.
 166  */
 167 static bool
 168 ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
 169 {
 170         struct super_block *sb = inode->i_sb;
 171         unsigned long blockmask = sb->s_blocksize - 1;
 172
 173         if ((pos | iov_iter_alignment(from)) & blockmask)
 174                 return true;
 175
 176         return false;
 177 }
 178
 179 static bool
 180 ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
 181 {
 182         if (offset + len > i_size_read(inode) ||
 183             offset + len > EXT4_I(inode)->i_disksize)
 184                 return true;
 185         return false;
 186 }
 187
 188 /* Is IO overwriting allocated and initialized blocks? */
 189 static bool ext4_overwrite_io(struct inode *inode, loff_t pos, loff_t len)
 190 {
 191         struct ext4_map_blocks map;
 192         unsigned int blkbits = inode->i_blkbits;
 193         int err, blklen;
 194
 195         if (pos + len > i_size_read(inode))
 196                 return false;
 197
 198         map.m_lblk = pos >> blkbits;
 199         map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
 200         blklen = map.m_len;
 201
 202         err = ext4_map_blocks(NULL, inode, &map, 0);
 203         /*
 204          * 'err==len' means that all of the blocks have been preallocated,
 205          * regardless of whether they have been initialized or not. To exclude
 206          * unwritten extents, we need to check m_flags.
 207          */
 208         return err == blklen && (map.m_flags & EXT4_MAP_MAPPED);
 209 }
 210
 211 static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
 212                                          struct iov_iter *from)
 213 {
 214         struct inode *inode = file_inode(iocb->ki_filp);
 215         ssize_t ret;
 216
 217         if (unlikely(IS_IMMUTABLE(inode)))
 218                 return -EPERM;
 219
 220         ret = generic_write_checks(iocb, from);
 221         if (ret <= 0)
 222                 return ret;
 223
 224         /*
 225          * If we have encountered a bitmap-format file, the size limit
 226          * is smaller than s_maxbytes, which is for extent-mapped files.
 227          */
 228         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 229                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 230
 231                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
 232                         return -EFBIG;
 233                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 234         }
 235
 236         return iov_iter_count(from);
 237 }
 238
 239 static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
 240 {
 241         ssize_t ret, count;
 242
 243         count = ext4_generic_write_checks(iocb, from);
 244         if (count <= 0)
 245                 return count;
 246
 247         ret = file_modified(iocb->ki_filp);
 248         if (ret)
 249                 return ret;
 250         return count;
 251 }
 252
 253 static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
 254                                         struct iov_iter *from)
 255 {
 256         ssize_t ret;
 257         struct inode *inode = file_inode(iocb->ki_filp);
 258
 259         if (iocb->ki_flags & IOCB_NOWAIT)
 260                 return -EOPNOTSUPP;
 261
 262         ext4_fc_start_update(inode);
 263         inode_lock(inode);
 264         ret = ext4_write_checks(iocb, from);
 265         if (ret <= 0)
 266                 goto out;
 267
 268         current->backing_dev_info = inode_to_bdi(inode);
 269         ret = generic_perform_write(iocb->ki_filp, from, iocb->ki_pos);
 270         current->backing_dev_info = NULL;
 271
 272 out:
 273         inode_unlock(inode);
 274         ext4_fc_stop_update(inode);
 275         if (likely(ret > 0)) {
 276                 iocb->ki_pos += ret;
 277                 ret = generic_write_sync(iocb, ret);
 278         }
 279
 280         return ret;
 281 }
 282
 283 static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
 284                                            ssize_t written, size_t count)
 285 {
 286         handle_t *handle;
 287         bool truncate = false;
 288         u8 blkbits = inode->i_blkbits;
 289         ext4_lblk_t written_blk, end_blk;
 290         int ret;
 291
 292         /*
 293          * Note that EXT4_I(inode)->i_disksize can get extended up to
 294          * inode->i_size while the I/O was running due to writeback of delalloc
 295          * blocks. But, the code in ext4_iomap_alloc() is careful to use
 296          * zeroed/unwritten extents if this is possible; thus we won't leave
 297          * uninitialized blocks in a file even if we didn't succeed in writing
 298          * as much as we intended.
 299          */
 300         WARN_ON_ONCE(i_size_read(inode) < EXT4_I(inode)->i_disksize);
 301         if (offset + count <= EXT4_I(inode)->i_disksize) {
 302                 /*
 303                  * We need to ensure that the inode is removed from the orphan
 304                  * list if it has been added prematurely, due to writeback of
 305                  * delalloc blocks.
 306                  */
 307                 if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
 308                         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 309
 310                         if (IS_ERR(handle)) {
 311                                 ext4_orphan_del(NULL, inode);
 312                                 return PTR_ERR(handle);
 313                         }
 314
 315                         ext4_orphan_del(handle, inode);
 316                         ext4_journal_stop(handle);
 317                 }
 318
 319                 return written;
 320         }
 321
 322         if (written < 0)
 323                 goto truncate;
 324
 325         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 326         if (IS_ERR(handle)) {
 327                 written = PTR_ERR(handle);
 328                 goto truncate;
 329         }
 330
 331         if (ext4_update_inode_size(inode, offset + written)) {
 332                 ret = ext4_mark_inode_dirty(handle, inode);
 333                 if (unlikely(ret)) {
 334                         written = ret;
 335                         ext4_journal_stop(handle);
 336                         goto truncate;
 337                 }
 338         }
 339
 340         /*
 341          * We may need to truncate allocated but not written blocks beyond EOF.
 342          */
 343         written_blk = ALIGN(offset + written, 1 << blkbits);
 344         end_blk = ALIGN(offset + count, 1 << blkbits);
 345         if (written_blk < end_blk && ext4_can_truncate(inode))
 346                 truncate = true;
 347
 348         /*
 349          * Remove the inode from the orphan list if it has been extended and
 350          * everything went OK.
 351          */
 352         if (!truncate && inode->i_nlink)
 353                 ext4_orphan_del(handle, inode);
 354         ext4_journal_stop(handle);
 355
 356         if (truncate) {
 357 truncate:
 358                 ext4_truncate_failed_write(inode);
 359                 /*
 360                  * If the truncate operation failed early, then the inode may
 361                  * still be on the orphan list. In that case, we need to try
 362                  * remove the inode from the in-memory linked list.
 363                  */
 364                 if (inode->i_nlink)
 365                         ext4_orphan_del(NULL, inode);
 366         }
 367
 368         return written;
 369 }
 370
 371 static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
 372                                  int error, unsigned int flags)
 373 {
 374         loff_t offset = iocb->ki_pos;
 375         struct inode *inode = file_inode(iocb->ki_filp);
 376
 377         if (error)
 378                 return error;
 379
 380         if (size && flags & IOMAP_DIO_UNWRITTEN)
 381                 return ext4_convert_unwritten_extents(NULL, inode,
 382                                                       offset, size);
 383
 384         return 0;
 385 }
 386
 387 static const struct iomap_dio_ops ext4_dio_write_ops = {
 388         .end_io = ext4_dio_write_end_io,
 389 };
 390
 391 /*
 392  * The intention here is to start with shared lock acquired then see if any
 393  * condition requires an exclusive inode lock. If yes, then we restart the
 394  * whole operation by releasing the shared lock and acquiring exclusive lock.
 395  *
 396  * - For unaligned_io we never take shared lock as it may cause data corruption
 397  *   when two unaligned IO tries to modify the same block e.g. while zeroing.
 398  *
 399  * - For extending writes case we don't take the shared lock, since it requires
 400  *   updating inode i_disksize and/or orphan handling with exclusive lock.
 401  *
 402  * - shared locking will only be true mostly with overwrites. Otherwise we will
 403  *   switch to exclusive i_rwsem lock.
 404  */
 405 static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
 406                                      bool *ilock_shared, bool *extend)
 407 {
 408         struct file *file = iocb->ki_filp;
 409         struct inode *inode = file_inode(file);
 410         loff_t offset;
 411         size_t count;
 412         ssize_t ret;
 413
 414 restart:
 415         ret = ext4_generic_write_checks(iocb, from);
 416         if (ret <= 0)
 417                 goto out;
 418
 419         offset = iocb->ki_pos;
 420         count = ret;
 421         if (ext4_extending_io(inode, offset, count))
 422                 *extend = true;
 423         /*
 424          * Determine whether the IO operation will overwrite allocated
 425          * and initialized blocks.
 426          * We need exclusive i_rwsem for changing security info
 427          * in file_modified().
 428          */
 429         if (*ilock_shared && (!IS_NOSEC(inode) || *extend ||
 430              !ext4_overwrite_io(inode, offset, count))) {
 431                 if (iocb->ki_flags & IOCB_NOWAIT) {
 432                         ret = -EAGAIN;
 433                         goto out;
 434                 }
 435                 inode_unlock_shared(inode);
 436                 *ilock_shared = false;
 437                 inode_lock(inode);
 438                 goto restart;
 439         }
 440
 441         ret = file_modified(file);
 442         if (ret < 0)
 443                 goto out;
 444
 445         return count;
 446 out:
 447         if (*ilock_shared)
 448                 inode_unlock_shared(inode);
 449         else
 450                 inode_unlock(inode);
 451         return ret;
 452 }
 453
 454 static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 455 {
 456         ssize_t ret;
 457         handle_t *handle;
 458         struct inode *inode = file_inode(iocb->ki_filp);
 459         loff_t offset = iocb->ki_pos;
 460         size_t count = iov_iter_count(from);
 461         const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
 462         bool extend = false, unaligned_io = false;
 463         bool ilock_shared = true;
 464
 465         /*
 466          * We initially start with shared inode lock unless it is
 467          * unaligned IO which needs exclusive lock anyways.
 468          */
 469         if (ext4_unaligned_io(inode, from, offset)) {
 470                 unaligned_io = true;
 471                 ilock_shared = false;
 472         }
 473         /*
 474          * Quick check here without any i_rwsem lock to see if it is extending
 475          * IO. A more reliable check is done in ext4_dio_write_checks() with
 476          * proper locking in place.
 477          */
 478         if (offset + count > i_size_read(inode))
 479                 ilock_shared = false;
 480
 481         if (iocb->ki_flags & IOCB_NOWAIT) {
 482                 if (ilock_shared) {
 483                         if (!inode_trylock_shared(inode))
 484                                 return -EAGAIN;
 485                 } else {
 486                         if (!inode_trylock(inode))
 487                                 return -EAGAIN;
 488                 }
 489         } else {
 490                 if (ilock_shared)
 491                         inode_lock_shared(inode);
 492                 else
 493                         inode_lock(inode);
 494         }
 495
 496         /* Fallback to buffered I/O if the inode does not support direct I/O. */
 497         if (!ext4_dio_supported(inode)) {
 498                 if (ilock_shared)
 499                         inode_unlock_shared(inode);
 500                 else
 501                         inode_unlock(inode);
 502                 return ext4_buffered_write_iter(iocb, from);
 503         }
 504
 505         ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend);
 506         if (ret <= 0)
 507                 return ret;
 508
 509         /* if we're going to block and IOCB_NOWAIT is set, return -EAGAIN */
 510         if ((iocb->ki_flags & IOCB_NOWAIT) && (unaligned_io || extend)) {
 511                 ret = -EAGAIN;
 512                 goto out;
 513         }
 514
 515         offset = iocb->ki_pos;
 516         count = ret;
 517
 518         /*
 519          * Unaligned direct IO must be serialized among each other as zeroing
 520          * of partial blocks of two competing unaligned IOs can result in data
 521          * corruption.
 522          *
 523          * So we make sure we don't allow any unaligned IO in flight.
 524          * For IOs where we need not wait (like unaligned non-AIO DIO),
 525          * below inode_dio_wait() may anyway become a no-op, since we start
 526          * with exclusive lock.
 527          */
 528         if (unaligned_io)
 529                 inode_dio_wait(inode);
 530
 531         if (extend) {
 532                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 533                 if (IS_ERR(handle)) {
 534                         ret = PTR_ERR(handle);
 535                         goto out;
 536                 }
 537
 538                 ext4_fc_start_update(inode);
 539                 ret = ext4_orphan_add(handle, inode);
 540                 ext4_fc_stop_update(inode);
 541                 if (ret) {
 542                         ext4_journal_stop(handle);
 543                         goto out;
 544                 }
 545
 546                 ext4_journal_stop(handle);
 547         }
 548
 549         if (ilock_shared)
 550                 iomap_ops = &ext4_iomap_overwrite_ops;
 551         ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 552                            (unaligned_io || extend) ? IOMAP_DIO_FORCE_WAIT : 0);
 553         if (ret == -ENOTBLK)
 554                 ret = 0;
 555
 556         if (extend)
 557                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 558
 559 out:
 560         if (ilock_shared)
 561                 inode_unlock_shared(inode);
 562         else
 563                 inode_unlock(inode);
 564
 565         if (ret >= 0 && iov_iter_count(from)) {
 566                 ssize_t err;
 567                 loff_t endbyte;
 568
 569                 offset = iocb->ki_pos;
 570                 err = ext4_buffered_write_iter(iocb, from);
 571                 if (err < 0)
 572                         return err;
 573
 574                 /*
 575                  * We need to ensure that the pages within the page cache for
 576                  * the range covered by this I/O are written to disk and
 577                  * invalidated. This is in attempt to preserve the expected
 578                  * direct I/O semantics in the case we fallback to buffered I/O
 579                  * to complete off the I/O request.
 580                  */
 581                 ret += err;
 582                 endbyte = offset + err - 1;
 583                 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
 584                                                    offset, endbyte);
 585                 if (!err)
 586                         invalidate_mapping_pages(iocb->ki_filp->f_mapping,
 587                                                  offset >> PAGE_SHIFT,
 588                                                  endbyte >> PAGE_SHIFT);
 589         }
 590
 591         return ret;
 592 }
 593
 594 #ifdef CONFIG_FS_DAX
 595 static ssize_t
 596 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 597 {
 598         ssize_t ret;
 599         size_t count;
 600         loff_t offset;
 601         handle_t *handle;
 602         bool extend = false;
 603         struct inode *inode = file_inode(iocb->ki_filp);
 604
 605         if (iocb->ki_flags & IOCB_NOWAIT) {
 606                 if (!inode_trylock(inode))
 607                         return -EAGAIN;
 608         } else {
 609                 inode_lock(inode);
 610         }
 611
 612         ret = ext4_write_checks(iocb, from);
 613         if (ret <= 0)
 614                 goto out;
 615
 616         offset = iocb->ki_pos;
 617         count = iov_iter_count(from);
 618
 619         if (offset + count > EXT4_I(inode)->i_disksize) {
 620                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 621                 if (IS_ERR(handle)) {
 622                         ret = PTR_ERR(handle);
 623                         goto out;
 624                 }
 625
 626                 ret = ext4_orphan_add(handle, inode);
 627                 if (ret) {
 628                         ext4_journal_stop(handle);
 629                         goto out;
 630                 }
 631
 632                 extend = true;
 633                 ext4_journal_stop(handle);
 634         }
 635
 636         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 637
 638         if (extend)
 639                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 640 out:
 641         inode_unlock(inode);
 642         if (ret > 0)
 643                 ret = generic_write_sync(iocb, ret);
 644         return ret;
 645 }
 646 #endif
 647
 648 static ssize_t
 649 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 650 {
 651         struct inode *inode = file_inode(iocb->ki_filp);
 652
 653         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 654                 return -EIO;
 655
 656 #ifdef CONFIG_FS_DAX
 657         if (IS_DAX(inode))
 658                 return ext4_dax_write_iter(iocb, from);
 659 #endif
 660         if (iocb->ki_flags & IOCB_DIRECT)
 661                 return ext4_dio_write_iter(iocb, from);
 662         else
 663                 return ext4_buffered_write_iter(iocb, from);
 664 }
 665
 666 #ifdef CONFIG_FS_DAX
 667 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
 668                 enum page_entry_size pe_size)
 669 {
 670         int error = 0;
 671         vm_fault_t result;
 672         int retries = 0;
 673         handle_t *handle = NULL;
 674         struct inode *inode = file_inode(vmf->vma->vm_file);
 675         struct super_block *sb = inode->i_sb;
 676
 677         /*
 678          * We have to distinguish real writes from writes which will result in a
 679          * COW page; COW writes should *not* poke the journal (the file will not
 680          * be changed). Doing so would cause unintended failures when mounted
 681          * read-only.
 682          *
 683          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 684          * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
 685          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 686          * we eventually come back with a COW page.
 687          */
 688         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 689                 (vmf->vma->vm_flags & VM_SHARED);
 690         pfn_t pfn;
 691
 692         if (write) {
 693                 sb_start_pagefault(sb);
 694                 file_update_time(vmf->vma->vm_file);
 695                 down_read(&EXT4_I(inode)->i_mmap_sem);
 696 retry:
 697                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 698                                                EXT4_DATA_TRANS_BLOCKS(sb));
 699                 if (IS_ERR(handle)) {
 700                         up_read(&EXT4_I(inode)->i_mmap_sem);
 701                         sb_end_pagefault(sb);
 702                         return VM_FAULT_SIGBUS;
 703                 }
 704         } else {
 705                 down_read(&EXT4_I(inode)->i_mmap_sem);
 706         }
 707         result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
 708         if (write) {
 709                 ext4_journal_stop(handle);
 710
 711                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 712                     ext4_should_retry_alloc(sb, &retries))
 713                         goto retry;
 714                 /* Handling synchronous page fault? */
 715                 if (result & VM_FAULT_NEEDDSYNC)
 716                         result = dax_finish_sync_fault(vmf, pe_size, pfn);
 717                 up_read(&EXT4_I(inode)->i_mmap_sem);
 718                 sb_end_pagefault(sb);
 719         } else {
 720                 up_read(&EXT4_I(inode)->i_mmap_sem);
 721         }
 722
 723         return result;
 724 }
 725
 726 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 727 {
 728         return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
 729 }
 730
 731 static const struct vm_operations_struct ext4_dax_vm_ops = {
 732         .fault          = ext4_dax_fault,
 733         .huge_fault     = ext4_dax_huge_fault,
 734         .page_mkwrite   = ext4_dax_fault,
 735         .pfn_mkwrite    = ext4_dax_fault,
 736 };
 737 #else
 738 #define ext4_dax_vm_ops ext4_file_vm_ops
 739 #endif
 740
 741 static const struct vm_operations_struct ext4_file_vm_ops = {
 742         .fault          = ext4_filemap_fault,
 743         .map_pages      = filemap_map_pages,
 744         .page_mkwrite   = ext4_page_mkwrite,
 745 };
 746
 747 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 748 {
 749         struct inode *inode = file->f_mapping->host;
 750         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 751         struct dax_device *dax_dev = sbi->s_daxdev;
 752
 753         if (unlikely(ext4_forced_shutdown(sbi)))
 754                 return -EIO;
 755
 756         /*
 757          * We don't support synchronous mappings for non-DAX files and
 758          * for DAX files if underneath dax_device is not synchronous.
 759          */
 760         if (!daxdev_mapping_supported(vma, dax_dev))
 761                 return -EOPNOTSUPP;
 762
 763         file_accessed(file);
 764         if (IS_DAX(file_inode(file))) {
 765                 vma->vm_ops = &ext4_dax_vm_ops;
 766                 vma->vm_flags |= VM_HUGEPAGE;
 767         } else {
 768                 vma->vm_ops = &ext4_file_vm_ops;
 769         }
 770         return 0;
 771 }
 772
 773 static int ext4_sample_last_mounted(struct super_block *sb,
 774                                     struct vfsmount *mnt)
 775 {
 776         struct ext4_sb_info *sbi = EXT4_SB(sb);
 777         struct path path;
 778         char buf[64], *cp;
 779         handle_t *handle;
 780         int err;
 781
 782         if (likely(ext4_test_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED)))
 783                 return 0;
 784
 785         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 786                 return 0;
 787
 788         ext4_set_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED);
 789         /*
 790          * Sample where the filesystem has been mounted and
 791          * store it in the superblock for sysadmin convenience
 792          * when trying to sort through large numbers of block
 793          * devices or filesystem images.
 794          */
 795         memset(buf, 0, sizeof(buf));
 796         path.mnt = mnt;
 797         path.dentry = mnt->mnt_root;
 798         cp = d_path(&path, buf, sizeof(buf));
 799         err = 0;
 800         if (IS_ERR(cp))
 801                 goto out;
 802
 803         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 804         err = PTR_ERR(handle);
 805         if (IS_ERR(handle))
 806                 goto out;
 807         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 808         err = ext4_journal_get_write_access(handle, sbi->s_sbh);
 809         if (err)
 810                 goto out_journal;
 811         lock_buffer(sbi->s_sbh);
 812         strncpy(sbi->s_es->s_last_mounted, cp,
 813                 sizeof(sbi->s_es->s_last_mounted));
 814         ext4_superblock_csum_set(sb);
 815         unlock_buffer(sbi->s_sbh);
 816         ext4_handle_dirty_metadata(handle, NULL, sbi->s_sbh);
 817 out_journal:
 818         ext4_journal_stop(handle);
 819 out:
 820         sb_end_intwrite(sb);
 821         return err;
 822 }
 823
 824 static int ext4_file_open(struct inode *inode, struct file *filp)
 825 {
 826         int ret;
 827
 828         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 829                 return -EIO;
 830
 831         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 832         if (ret)
 833                 return ret;
 834
 835         ret = fscrypt_file_open(inode, filp);
 836         if (ret)
 837                 return ret;
 838
 839         ret = fsverity_file_open(inode, filp);
 840         if (ret)
 841                 return ret;
 842
 843         /*
 844          * Set up the jbd2_inode if we are opening the inode for
 845          * writing and the journal is present
 846          */
 847         if (filp->f_mode & FMODE_WRITE) {
 848                 ret = ext4_inode_attach_jinode(inode);
 849                 if (ret < 0)
 850                         return ret;
 851         }
 852
 853         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
 854         return dquot_file_open(inode, filp);
 855 }
 856
 857 /*
 858  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 859  * by calling generic_file_llseek_size() with the appropriate maxbytes
 860  * value for each.
 861  */
 862 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 863 {
 864         struct inode *inode = file->f_mapping->host;
 865         loff_t maxbytes;
 866
 867         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 868                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 869         else
 870                 maxbytes = inode->i_sb->s_maxbytes;
 871
 872         switch (whence) {
 873         default:
 874                 return generic_file_llseek_size(file, offset, whence,
 875                                                 maxbytes, i_size_read(inode));
 876         case SEEK_HOLE:
 877                 inode_lock_shared(inode);
 878                 offset = iomap_seek_hole(inode, offset,
 879                                          &ext4_iomap_report_ops);
 880                 inode_unlock_shared(inode);
 881                 break;
 882         case SEEK_DATA:
 883                 inode_lock_shared(inode);
 884                 offset = iomap_seek_data(inode, offset,
 885                                          &ext4_iomap_report_ops);
 886                 inode_unlock_shared(inode);
 887                 break;
 888         }
 889
 890         if (offset < 0)
 891                 return offset;
 892         return vfs_setpos(file, offset, maxbytes);
 893 }
 894
 895 const struct file_operations ext4_file_operations = {
 896         .llseek         = ext4_llseek,
 897         .read_iter      = ext4_file_read_iter,
 898         .write_iter     = ext4_file_write_iter,
 899         .iopoll         = iomap_dio_iopoll,
 900         .unlocked_ioctl = ext4_ioctl,
 901 #ifdef CONFIG_COMPAT
 902         .compat_ioctl   = ext4_compat_ioctl,
 903 #endif
 904         .mmap           = ext4_file_mmap,
 905         .mmap_supported_flags = MAP_SYNC,
 906         .open           = ext4_file_open,
 907         .release        = ext4_release_file,
 908         .fsync          = ext4_sync_file,
 909         .get_unmapped_area = thp_get_unmapped_area,
 910         .splice_read    = generic_file_splice_read,
 911         .splice_write   = iter_file_splice_write,
 912         .fallocate      = ext4_fallocate,
 913 };
 914
 915 const struct inode_operations ext4_file_inode_operations = {
 916         .setattr        = ext4_setattr,
 917         .getattr        = ext4_file_getattr,
 918         .listxattr      = ext4_listxattr,
 919         .get_acl        = ext4_get_acl,
 920         .set_acl        = ext4_set_acl,
 921         .fiemap         = ext4_fiemap,
 922 };
 923