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 /*
  40  * Returns %true if the given DIO request should be attempted with DIO, or
  41  * %false if it should fall back to buffered I/O.
  42  *
  43  * DIO isn't well specified; when it's unsupported (either due to the request
  44  * being misaligned, or due to the file not supporting DIO at all), filesystems
  45  * either fall back to buffered I/O or return EINVAL.  For files that don't use
  46  * any special features like encryption or verity, ext4 has traditionally
  47  * returned EINVAL for misaligned DIO.  iomap_dio_rw() uses this convention too.
  48  * In this case, we should attempt the DIO, *not* fall back to buffered I/O.
  49  *
  50  * In contrast, in cases where DIO is unsupported due to ext4 features, ext4
  51  * traditionally falls back to buffered I/O.
  52  *
  53  * This function implements the traditional ext4 behavior in all these cases.
  54  */
  55 static bool ext4_should_use_dio(struct kiocb *iocb, struct iov_iter *iter)
  56 {
  57         struct inode *inode = file_inode(iocb->ki_filp);
  58         u32 dio_align = ext4_dio_alignment(inode);
  59
  60         if (dio_align == 0)
  61                 return false;
  62
  63         if (dio_align == 1)
  64                 return true;
  65
  66         return IS_ALIGNED(iocb->ki_pos | iov_iter_alignment(iter), dio_align);
  67 }
  68
  69 static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
  70 {
  71         ssize_t ret;
  72         struct inode *inode = file_inode(iocb->ki_filp);
  73
  74         if (iocb->ki_flags & IOCB_NOWAIT) {
  75                 if (!inode_trylock_shared(inode))
  76                         return -EAGAIN;
  77         } else {
  78                 inode_lock_shared(inode);
  79         }
  80
  81         if (!ext4_should_use_dio(iocb, to)) {
  82                 inode_unlock_shared(inode);
  83                 /*
  84                  * Fallback to buffered I/O if the operation being performed on
  85                  * the inode is not supported by direct I/O. The IOCB_DIRECT
  86                  * flag needs to be cleared here in order to ensure that the
  87                  * direct I/O path within generic_file_read_iter() is not
  88                  * taken.
  89                  */
  90                 iocb->ki_flags &= ~IOCB_DIRECT;
  91                 return generic_file_read_iter(iocb, to);
  92         }
  93
  94         ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL, 0, NULL, 0);
  95         inode_unlock_shared(inode);
  96
  97         file_accessed(iocb->ki_filp);
  98         return ret;
  99 }
 100
 101 #ifdef CONFIG_FS_DAX
 102 static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
 103 {
 104         struct inode *inode = file_inode(iocb->ki_filp);
 105         ssize_t ret;
 106
 107         if (iocb->ki_flags & IOCB_NOWAIT) {
 108                 if (!inode_trylock_shared(inode))
 109                         return -EAGAIN;
 110         } else {
 111                 inode_lock_shared(inode);
 112         }
 113         /*
 114          * Recheck under inode lock - at this point we are sure it cannot
 115          * change anymore
 116          */
 117         if (!IS_DAX(inode)) {
 118                 inode_unlock_shared(inode);
 119                 /* Fallback to buffered IO in case we cannot support DAX */
 120                 return generic_file_read_iter(iocb, to);
 121         }
 122         ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
 123         inode_unlock_shared(inode);
 124
 125         file_accessed(iocb->ki_filp);
 126         return ret;
 127 }
 128 #endif
 129
 130 static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 131 {
 132         struct inode *inode = file_inode(iocb->ki_filp);
 133
 134         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 135                 return -EIO;
 136
 137         if (!iov_iter_count(to))
 138                 return 0; /* skip atime */
 139
 140 #ifdef CONFIG_FS_DAX
 141         if (IS_DAX(inode))
 142                 return ext4_dax_read_iter(iocb, to);
 143 #endif
 144         if (iocb->ki_flags & IOCB_DIRECT)
 145                 return ext4_dio_read_iter(iocb, to);
 146
 147         return generic_file_read_iter(iocb, to);
 148 }
 149
 150 static ssize_t ext4_file_splice_read(struct file *in, loff_t *ppos,
 151                                      struct pipe_inode_info *pipe,
 152                                      size_t len, unsigned int flags)
 153 {
 154         struct inode *inode = file_inode(in);
 155
 156         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 157                 return -EIO;
 158         return filemap_splice_read(in, ppos, pipe, len, flags);
 159 }
 160
 161 /*
 162  * Called when an inode is released. Note that this is different
 163  * from ext4_file_open: open gets called at every open, but release
 164  * gets called only when /all/ the files are closed.
 165  */
 166 static int ext4_release_file(struct inode *inode, struct file *filp)
 167 {
 168         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
 169                 ext4_alloc_da_blocks(inode);
 170                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
 171         }
 172         /* if we are the last writer on the inode, drop the block reservation */
 173         if ((filp->f_mode & FMODE_WRITE) &&
 174                         (atomic_read(&inode->i_writecount) == 1) &&
 175                         !EXT4_I(inode)->i_reserved_data_blocks) {
 176                 down_write(&EXT4_I(inode)->i_data_sem);
 177                 ext4_discard_preallocations(inode, 0);
 178                 up_write(&EXT4_I(inode)->i_data_sem);
 179         }
 180         if (is_dx(inode) && filp->private_data)
 181                 ext4_htree_free_dir_info(filp->private_data);
 182
 183         return 0;
 184 }
 185
 186 /*
 187  * This tests whether the IO in question is block-aligned or not.
 188  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
 189  * are converted to written only after the IO is complete.  Until they are
 190  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
 191  * it needs to zero out portions of the start and/or end block.  If 2 AIO
 192  * threads are at work on the same unwritten block, they must be synchronized
 193  * or one thread will zero the other's data, causing corruption.
 194  */
 195 static bool
 196 ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
 197 {
 198         struct super_block *sb = inode->i_sb;
 199         unsigned long blockmask = sb->s_blocksize - 1;
 200
 201         if ((pos | iov_iter_alignment(from)) & blockmask)
 202                 return true;
 203
 204         return false;
 205 }
 206
 207 static bool
 208 ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
 209 {
 210         if (offset + len > i_size_read(inode) ||
 211             offset + len > EXT4_I(inode)->i_disksize)
 212                 return true;
 213         return false;
 214 }
 215
 216 /* Is IO overwriting allocated or initialized blocks? */
 217 static bool ext4_overwrite_io(struct inode *inode,
 218                               loff_t pos, loff_t len, bool *unwritten)
 219 {
 220         struct ext4_map_blocks map;
 221         unsigned int blkbits = inode->i_blkbits;
 222         int err, blklen;
 223
 224         if (pos + len > i_size_read(inode))
 225                 return false;
 226
 227         map.m_lblk = pos >> blkbits;
 228         map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
 229         blklen = map.m_len;
 230
 231         err = ext4_map_blocks(NULL, inode, &map, 0);
 232         if (err != blklen)
 233                 return false;
 234         /*
 235          * 'err==len' means that all of the blocks have been preallocated,
 236          * regardless of whether they have been initialized or not. We need to
 237          * check m_flags to distinguish the unwritten extents.
 238          */
 239         *unwritten = !(map.m_flags & EXT4_MAP_MAPPED);
 240         return true;
 241 }
 242
 243 static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
 244                                          struct iov_iter *from)
 245 {
 246         struct inode *inode = file_inode(iocb->ki_filp);
 247         ssize_t ret;
 248
 249         if (unlikely(IS_IMMUTABLE(inode)))
 250                 return -EPERM;
 251
 252         ret = generic_write_checks(iocb, from);
 253         if (ret <= 0)
 254                 return ret;
 255
 256         /*
 257          * If we have encountered a bitmap-format file, the size limit
 258          * is smaller than s_maxbytes, which is for extent-mapped files.
 259          */
 260         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 261                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 262
 263                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
 264                         return -EFBIG;
 265                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 266         }
 267
 268         return iov_iter_count(from);
 269 }
 270
 271 static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
 272 {
 273         ssize_t ret, count;
 274
 275         count = ext4_generic_write_checks(iocb, from);
 276         if (count <= 0)
 277                 return count;
 278
 279         ret = file_modified(iocb->ki_filp);
 280         if (ret)
 281                 return ret;
 282         return count;
 283 }
 284
 285 static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
 286                                         struct iov_iter *from)
 287 {
 288         ssize_t ret;
 289         struct inode *inode = file_inode(iocb->ki_filp);
 290
 291         if (iocb->ki_flags & IOCB_NOWAIT)
 292                 return -EOPNOTSUPP;
 293
 294         inode_lock(inode);
 295         ret = ext4_write_checks(iocb, from);
 296         if (ret <= 0)
 297                 goto out;
 298
 299         ret = generic_perform_write(iocb, from);
 300
 301 out:
 302         inode_unlock(inode);
 303         if (unlikely(ret <= 0))
 304                 return ret;
 305         return generic_write_sync(iocb, ret);
 306 }
 307
 308 static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
 309                                            ssize_t written, size_t count)
 310 {
 311         handle_t *handle;
 312         bool truncate = false;
 313         u8 blkbits = inode->i_blkbits;
 314         ext4_lblk_t written_blk, end_blk;
 315         int ret;
 316
 317         /*
 318          * Note that EXT4_I(inode)->i_disksize can get extended up to
 319          * inode->i_size while the I/O was running due to writeback of delalloc
 320          * blocks. But, the code in ext4_iomap_alloc() is careful to use
 321          * zeroed/unwritten extents if this is possible; thus we won't leave
 322          * uninitialized blocks in a file even if we didn't succeed in writing
 323          * as much as we intended.
 324          */
 325         WARN_ON_ONCE(i_size_read(inode) < EXT4_I(inode)->i_disksize);
 326         if (offset + count <= EXT4_I(inode)->i_disksize) {
 327                 /*
 328                  * We need to ensure that the inode is removed from the orphan
 329                  * list if it has been added prematurely, due to writeback of
 330                  * delalloc blocks.
 331                  */
 332                 if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
 333                         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 334
 335                         if (IS_ERR(handle)) {
 336                                 ext4_orphan_del(NULL, inode);
 337                                 return PTR_ERR(handle);
 338                         }
 339
 340                         ext4_orphan_del(handle, inode);
 341                         ext4_journal_stop(handle);
 342                 }
 343
 344                 return written;
 345         }
 346
 347         if (written < 0)
 348                 goto truncate;
 349
 350         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 351         if (IS_ERR(handle)) {
 352                 written = PTR_ERR(handle);
 353                 goto truncate;
 354         }
 355
 356         if (ext4_update_inode_size(inode, offset + written)) {
 357                 ret = ext4_mark_inode_dirty(handle, inode);
 358                 if (unlikely(ret)) {
 359                         written = ret;
 360                         ext4_journal_stop(handle);
 361                         goto truncate;
 362                 }
 363         }
 364
 365         /*
 366          * We may need to truncate allocated but not written blocks beyond EOF.
 367          */
 368         written_blk = ALIGN(offset + written, 1 << blkbits);
 369         end_blk = ALIGN(offset + count, 1 << blkbits);
 370         if (written_blk < end_blk && ext4_can_truncate(inode))
 371                 truncate = true;
 372
 373         /*
 374          * Remove the inode from the orphan list if it has been extended and
 375          * everything went OK.
 376          */
 377         if (!truncate && inode->i_nlink)
 378                 ext4_orphan_del(handle, inode);
 379         ext4_journal_stop(handle);
 380
 381         if (truncate) {
 382 truncate:
 383                 ext4_truncate_failed_write(inode);
 384                 /*
 385                  * If the truncate operation failed early, then the inode may
 386                  * still be on the orphan list. In that case, we need to try
 387                  * remove the inode from the in-memory linked list.
 388                  */
 389                 if (inode->i_nlink)
 390                         ext4_orphan_del(NULL, inode);
 391         }
 392
 393         return written;
 394 }
 395
 396 static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
 397                                  int error, unsigned int flags)
 398 {
 399         loff_t pos = iocb->ki_pos;
 400         struct inode *inode = file_inode(iocb->ki_filp);
 401
 402         if (error)
 403                 return error;
 404
 405         if (size && flags & IOMAP_DIO_UNWRITTEN) {
 406                 error = ext4_convert_unwritten_extents(NULL, inode, pos, size);
 407                 if (error < 0)
 408                         return error;
 409         }
 410         /*
 411          * If we are extending the file, we have to update i_size here before
 412          * page cache gets invalidated in iomap_dio_rw(). Otherwise racing
 413          * buffered reads could zero out too much from page cache pages. Update
 414          * of on-disk size will happen later in ext4_dio_write_iter() where
 415          * we have enough information to also perform orphan list handling etc.
 416          * Note that we perform all extending writes synchronously under
 417          * i_rwsem held exclusively so i_size update is safe here in that case.
 418          * If the write was not extending, we cannot see pos > i_size here
 419          * because operations reducing i_size like truncate wait for all
 420          * outstanding DIO before updating i_size.
 421          */
 422         pos += size;
 423         if (pos > i_size_read(inode))
 424                 i_size_write(inode, pos);
 425
 426         return 0;
 427 }
 428
 429 static const struct iomap_dio_ops ext4_dio_write_ops = {
 430         .end_io = ext4_dio_write_end_io,
 431 };
 432
 433 /*
 434  * The intention here is to start with shared lock acquired then see if any
 435  * condition requires an exclusive inode lock. If yes, then we restart the
 436  * whole operation by releasing the shared lock and acquiring exclusive lock.
 437  *
 438  * - For unaligned_io we never take shared lock as it may cause data corruption
 439  *   when two unaligned IO tries to modify the same block e.g. while zeroing.
 440  *
 441  * - For extending writes case we don't take the shared lock, since it requires
 442  *   updating inode i_disksize and/or orphan handling with exclusive lock.
 443  *
 444  * - shared locking will only be true mostly with overwrites, including
 445  *   initialized blocks and unwritten blocks. For overwrite unwritten blocks
 446  *   we protect splitting extents by i_data_sem in ext4_inode_info, so we can
 447  *   also release exclusive i_rwsem lock.
 448  *
 449  * - Otherwise we will switch to exclusive i_rwsem lock.
 450  */
 451 static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
 452                                      bool *ilock_shared, bool *extend,
 453                                      bool *unwritten, int *dio_flags)
 454 {
 455         struct file *file = iocb->ki_filp;
 456         struct inode *inode = file_inode(file);
 457         loff_t offset;
 458         size_t count;
 459         ssize_t ret;
 460         bool overwrite, unaligned_io;
 461
 462 restart:
 463         ret = ext4_generic_write_checks(iocb, from);
 464         if (ret <= 0)
 465                 goto out;
 466
 467         offset = iocb->ki_pos;
 468         count = ret;
 469
 470         unaligned_io = ext4_unaligned_io(inode, from, offset);
 471         *extend = ext4_extending_io(inode, offset, count);
 472         overwrite = ext4_overwrite_io(inode, offset, count, unwritten);
 473
 474         /*
 475          * Determine whether we need to upgrade to an exclusive lock. This is
 476          * required to change security info in file_modified(), for extending
 477          * I/O, any form of non-overwrite I/O, and unaligned I/O to unwritten
 478          * extents (as partial block zeroing may be required).
 479          */
 480         if (*ilock_shared &&
 481             ((!IS_NOSEC(inode) || *extend || !overwrite ||
 482              (unaligned_io && *unwritten)))) {
 483                 if (iocb->ki_flags & IOCB_NOWAIT) {
 484                         ret = -EAGAIN;
 485                         goto out;
 486                 }
 487                 inode_unlock_shared(inode);
 488                 *ilock_shared = false;
 489                 inode_lock(inode);
 490                 goto restart;
 491         }
 492
 493         /*
 494          * Now that locking is settled, determine dio flags and exclusivity
 495          * requirements. Unaligned writes are allowed under shared lock so long
 496          * as they are pure overwrites. Set the iomap overwrite only flag as an
 497          * added precaution in this case. Even though this is unnecessary, we
 498          * can detect and warn on unexpected -EAGAIN if an unsafe unaligned
 499          * write is ever submitted.
 500          *
 501          * Otherwise, concurrent unaligned writes risk data corruption due to
 502          * partial block zeroing in the dio layer, and so the I/O must occur
 503          * exclusively. The inode lock is already held exclusive if the write is
 504          * non-overwrite or extending, so drain all outstanding dio and set the
 505          * force wait dio flag.
 506          */
 507         if (*ilock_shared && unaligned_io) {
 508                 *dio_flags = IOMAP_DIO_OVERWRITE_ONLY;
 509         } else if (!*ilock_shared && (unaligned_io || *extend)) {
 510                 if (iocb->ki_flags & IOCB_NOWAIT) {
 511                         ret = -EAGAIN;
 512                         goto out;
 513                 }
 514                 if (unaligned_io && (!overwrite || *unwritten))
 515                         inode_dio_wait(inode);
 516                 *dio_flags = IOMAP_DIO_FORCE_WAIT;
 517         }
 518
 519         ret = file_modified(file);
 520         if (ret < 0)
 521                 goto out;
 522
 523         return count;
 524 out:
 525         if (*ilock_shared)
 526                 inode_unlock_shared(inode);
 527         else
 528                 inode_unlock(inode);
 529         return ret;
 530 }
 531
 532 static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 533 {
 534         ssize_t ret;
 535         handle_t *handle;
 536         struct inode *inode = file_inode(iocb->ki_filp);
 537         loff_t offset = iocb->ki_pos;
 538         size_t count = iov_iter_count(from);
 539         const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
 540         bool extend = false, unwritten = false;
 541         bool ilock_shared = true;
 542         int dio_flags = 0;
 543
 544         /*
 545          * Quick check here without any i_rwsem lock to see if it is extending
 546          * IO. A more reliable check is done in ext4_dio_write_checks() with
 547          * proper locking in place.
 548          */
 549         if (offset + count > i_size_read(inode))
 550                 ilock_shared = false;
 551
 552         if (iocb->ki_flags & IOCB_NOWAIT) {
 553                 if (ilock_shared) {
 554                         if (!inode_trylock_shared(inode))
 555                                 return -EAGAIN;
 556                 } else {
 557                         if (!inode_trylock(inode))
 558                                 return -EAGAIN;
 559                 }
 560         } else {
 561                 if (ilock_shared)
 562                         inode_lock_shared(inode);
 563                 else
 564                         inode_lock(inode);
 565         }
 566
 567         /* Fallback to buffered I/O if the inode does not support direct I/O. */
 568         if (!ext4_should_use_dio(iocb, from)) {
 569                 if (ilock_shared)
 570                         inode_unlock_shared(inode);
 571                 else
 572                         inode_unlock(inode);
 573                 return ext4_buffered_write_iter(iocb, from);
 574         }
 575
 576         ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend,
 577                                     &unwritten, &dio_flags);
 578         if (ret <= 0)
 579                 return ret;
 580
 581         /*
 582          * Make sure inline data cannot be created anymore since we are going
 583          * to allocate blocks for DIO. We know the inode does not have any
 584          * inline data now because ext4_dio_supported() checked for that.
 585          */
 586         ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
 587
 588         offset = iocb->ki_pos;
 589         count = ret;
 590
 591         if (extend) {
 592                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 593                 if (IS_ERR(handle)) {
 594                         ret = PTR_ERR(handle);
 595                         goto out;
 596                 }
 597
 598                 ret = ext4_orphan_add(handle, inode);
 599                 if (ret) {
 600                         ext4_journal_stop(handle);
 601                         goto out;
 602                 }
 603
 604                 ext4_journal_stop(handle);
 605         }
 606
 607         if (ilock_shared && !unwritten)
 608                 iomap_ops = &ext4_iomap_overwrite_ops;
 609         ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 610                            dio_flags, NULL, 0);
 611         WARN_ON_ONCE(ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT));
 612         if (ret == -ENOTBLK)
 613                 ret = 0;
 614
 615         if (extend)
 616                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 617
 618 out:
 619         if (ilock_shared)
 620                 inode_unlock_shared(inode);
 621         else
 622                 inode_unlock(inode);
 623
 624         if (ret >= 0 && iov_iter_count(from)) {
 625                 ssize_t err;
 626                 loff_t endbyte;
 627
 628                 offset = iocb->ki_pos;
 629                 err = ext4_buffered_write_iter(iocb, from);
 630                 if (err < 0)
 631                         return err;
 632
 633                 /*
 634                  * We need to ensure that the pages within the page cache for
 635                  * the range covered by this I/O are written to disk and
 636                  * invalidated. This is in attempt to preserve the expected
 637                  * direct I/O semantics in the case we fallback to buffered I/O
 638                  * to complete off the I/O request.
 639                  */
 640                 ret += err;
 641                 endbyte = offset + err - 1;
 642                 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
 643                                                    offset, endbyte);
 644                 if (!err)
 645                         invalidate_mapping_pages(iocb->ki_filp->f_mapping,
 646                                                  offset >> PAGE_SHIFT,
 647                                                  endbyte >> PAGE_SHIFT);
 648         }
 649
 650         return ret;
 651 }
 652
 653 #ifdef CONFIG_FS_DAX
 654 static ssize_t
 655 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 656 {
 657         ssize_t ret;
 658         size_t count;
 659         loff_t offset;
 660         handle_t *handle;
 661         bool extend = false;
 662         struct inode *inode = file_inode(iocb->ki_filp);
 663
 664         if (iocb->ki_flags & IOCB_NOWAIT) {
 665                 if (!inode_trylock(inode))
 666                         return -EAGAIN;
 667         } else {
 668                 inode_lock(inode);
 669         }
 670
 671         ret = ext4_write_checks(iocb, from);
 672         if (ret <= 0)
 673                 goto out;
 674
 675         offset = iocb->ki_pos;
 676         count = iov_iter_count(from);
 677
 678         if (offset + count > EXT4_I(inode)->i_disksize) {
 679                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 680                 if (IS_ERR(handle)) {
 681                         ret = PTR_ERR(handle);
 682                         goto out;
 683                 }
 684
 685                 ret = ext4_orphan_add(handle, inode);
 686                 if (ret) {
 687                         ext4_journal_stop(handle);
 688                         goto out;
 689                 }
 690
 691                 extend = true;
 692                 ext4_journal_stop(handle);
 693         }
 694
 695         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 696
 697         if (extend)
 698                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 699 out:
 700         inode_unlock(inode);
 701         if (ret > 0)
 702                 ret = generic_write_sync(iocb, ret);
 703         return ret;
 704 }
 705 #endif
 706
 707 static ssize_t
 708 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 709 {
 710         struct inode *inode = file_inode(iocb->ki_filp);
 711
 712         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 713                 return -EIO;
 714
 715 #ifdef CONFIG_FS_DAX
 716         if (IS_DAX(inode))
 717                 return ext4_dax_write_iter(iocb, from);
 718 #endif
 719         if (iocb->ki_flags & IOCB_DIRECT)
 720                 return ext4_dio_write_iter(iocb, from);
 721         else
 722                 return ext4_buffered_write_iter(iocb, from);
 723 }
 724
 725 #ifdef CONFIG_FS_DAX
 726 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
 727                 enum page_entry_size pe_size)
 728 {
 729         int error = 0;
 730         vm_fault_t result;
 731         int retries = 0;
 732         handle_t *handle = NULL;
 733         struct inode *inode = file_inode(vmf->vma->vm_file);
 734         struct super_block *sb = inode->i_sb;
 735
 736         /*
 737          * We have to distinguish real writes from writes which will result in a
 738          * COW page; COW writes should *not* poke the journal (the file will not
 739          * be changed). Doing so would cause unintended failures when mounted
 740          * read-only.
 741          *
 742          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 743          * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
 744          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 745          * we eventually come back with a COW page.
 746          */
 747         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 748                 (vmf->vma->vm_flags & VM_SHARED);
 749         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
 750         pfn_t pfn;
 751
 752         if (write) {
 753                 sb_start_pagefault(sb);
 754                 file_update_time(vmf->vma->vm_file);
 755                 filemap_invalidate_lock_shared(mapping);
 756 retry:
 757                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 758                                                EXT4_DATA_TRANS_BLOCKS(sb));
 759                 if (IS_ERR(handle)) {
 760                         filemap_invalidate_unlock_shared(mapping);
 761                         sb_end_pagefault(sb);
 762                         return VM_FAULT_SIGBUS;
 763                 }
 764         } else {
 765                 filemap_invalidate_lock_shared(mapping);
 766         }
 767         result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
 768         if (write) {
 769                 ext4_journal_stop(handle);
 770
 771                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 772                     ext4_should_retry_alloc(sb, &retries))
 773                         goto retry;
 774                 /* Handling synchronous page fault? */
 775                 if (result & VM_FAULT_NEEDDSYNC)
 776                         result = dax_finish_sync_fault(vmf, pe_size, pfn);
 777                 filemap_invalidate_unlock_shared(mapping);
 778                 sb_end_pagefault(sb);
 779         } else {
 780                 filemap_invalidate_unlock_shared(mapping);
 781         }
 782
 783         return result;
 784 }
 785
 786 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 787 {
 788         return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
 789 }
 790
 791 static const struct vm_operations_struct ext4_dax_vm_ops = {
 792         .fault          = ext4_dax_fault,
 793         .huge_fault     = ext4_dax_huge_fault,
 794         .page_mkwrite   = ext4_dax_fault,
 795         .pfn_mkwrite    = ext4_dax_fault,
 796 };
 797 #else
 798 #define ext4_dax_vm_ops ext4_file_vm_ops
 799 #endif
 800
 801 static const struct vm_operations_struct ext4_file_vm_ops = {
 802         .fault          = filemap_fault,
 803         .map_pages      = filemap_map_pages,
 804         .page_mkwrite   = ext4_page_mkwrite,
 805 };
 806
 807 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 808 {
 809         struct inode *inode = file->f_mapping->host;
 810         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 811         struct dax_device *dax_dev = sbi->s_daxdev;
 812
 813         if (unlikely(ext4_forced_shutdown(sbi)))
 814                 return -EIO;
 815
 816         /*
 817          * We don't support synchronous mappings for non-DAX files and
 818          * for DAX files if underneath dax_device is not synchronous.
 819          */
 820         if (!daxdev_mapping_supported(vma, dax_dev))
 821                 return -EOPNOTSUPP;
 822
 823         file_accessed(file);
 824         if (IS_DAX(file_inode(file))) {
 825                 vma->vm_ops = &ext4_dax_vm_ops;
 826                 vm_flags_set(vma, VM_HUGEPAGE);
 827         } else {
 828                 vma->vm_ops = &ext4_file_vm_ops;
 829         }
 830         return 0;
 831 }
 832
 833 static int ext4_sample_last_mounted(struct super_block *sb,
 834                                     struct vfsmount *mnt)
 835 {
 836         struct ext4_sb_info *sbi = EXT4_SB(sb);
 837         struct path path;
 838         char buf[64], *cp;
 839         handle_t *handle;
 840         int err;
 841
 842         if (likely(ext4_test_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED)))
 843                 return 0;
 844
 845         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 846                 return 0;
 847
 848         ext4_set_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED);
 849         /*
 850          * Sample where the filesystem has been mounted and
 851          * store it in the superblock for sysadmin convenience
 852          * when trying to sort through large numbers of block
 853          * devices or filesystem images.
 854          */
 855         memset(buf, 0, sizeof(buf));
 856         path.mnt = mnt;
 857         path.dentry = mnt->mnt_root;
 858         cp = d_path(&path, buf, sizeof(buf));
 859         err = 0;
 860         if (IS_ERR(cp))
 861                 goto out;
 862
 863         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 864         err = PTR_ERR(handle);
 865         if (IS_ERR(handle))
 866                 goto out;
 867         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 868         err = ext4_journal_get_write_access(handle, sb, sbi->s_sbh,
 869                                             EXT4_JTR_NONE);
 870         if (err)
 871                 goto out_journal;
 872         lock_buffer(sbi->s_sbh);
 873         strncpy(sbi->s_es->s_last_mounted, cp,
 874                 sizeof(sbi->s_es->s_last_mounted));
 875         ext4_superblock_csum_set(sb);
 876         unlock_buffer(sbi->s_sbh);
 877         ext4_handle_dirty_metadata(handle, NULL, sbi->s_sbh);
 878 out_journal:
 879         ext4_journal_stop(handle);
 880 out:
 881         sb_end_intwrite(sb);
 882         return err;
 883 }
 884
 885 static int ext4_file_open(struct inode *inode, struct file *filp)
 886 {
 887         int ret;
 888
 889         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 890                 return -EIO;
 891
 892         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 893         if (ret)
 894                 return ret;
 895
 896         ret = fscrypt_file_open(inode, filp);
 897         if (ret)
 898                 return ret;
 899
 900         ret = fsverity_file_open(inode, filp);
 901         if (ret)
 902                 return ret;
 903
 904         /*
 905          * Set up the jbd2_inode if we are opening the inode for
 906          * writing and the journal is present
 907          */
 908         if (filp->f_mode & FMODE_WRITE) {
 909                 ret = ext4_inode_attach_jinode(inode);
 910                 if (ret < 0)
 911                         return ret;
 912         }
 913
 914         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC |
 915                         FMODE_DIO_PARALLEL_WRITE;
 916         return dquot_file_open(inode, filp);
 917 }
 918
 919 /*
 920  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 921  * by calling generic_file_llseek_size() with the appropriate maxbytes
 922  * value for each.
 923  */
 924 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 925 {
 926         struct inode *inode = file->f_mapping->host;
 927         loff_t maxbytes;
 928
 929         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 930                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 931         else
 932                 maxbytes = inode->i_sb->s_maxbytes;
 933
 934         switch (whence) {
 935         default:
 936                 return generic_file_llseek_size(file, offset, whence,
 937                                                 maxbytes, i_size_read(inode));
 938         case SEEK_HOLE:
 939                 inode_lock_shared(inode);
 940                 offset = iomap_seek_hole(inode, offset,
 941                                          &ext4_iomap_report_ops);
 942                 inode_unlock_shared(inode);
 943                 break;
 944         case SEEK_DATA:
 945                 inode_lock_shared(inode);
 946                 offset = iomap_seek_data(inode, offset,
 947                                          &ext4_iomap_report_ops);
 948                 inode_unlock_shared(inode);
 949                 break;
 950         }
 951
 952         if (offset < 0)
 953                 return offset;
 954         return vfs_setpos(file, offset, maxbytes);
 955 }
 956
 957 const struct file_operations ext4_file_operations = {
 958         .llseek         = ext4_llseek,
 959         .read_iter      = ext4_file_read_iter,
 960         .write_iter     = ext4_file_write_iter,
 961         .iopoll         = iocb_bio_iopoll,
 962         .unlocked_ioctl = ext4_ioctl,
 963 #ifdef CONFIG_COMPAT
 964         .compat_ioctl   = ext4_compat_ioctl,
 965 #endif
 966         .mmap           = ext4_file_mmap,
 967         .mmap_supported_flags = MAP_SYNC,
 968         .open           = ext4_file_open,
 969         .release        = ext4_release_file,
 970         .fsync          = ext4_sync_file,
 971         .get_unmapped_area = thp_get_unmapped_area,
 972         .splice_read    = ext4_file_splice_read,
 973         .splice_write   = iter_file_splice_write,
 974         .fallocate      = ext4_fallocate,
 975 };
 976
 977 const struct inode_operations ext4_file_inode_operations = {
 978         .setattr        = ext4_setattr,
 979         .getattr        = ext4_file_getattr,
 980         .listxattr      = ext4_listxattr,
 981         .get_inode_acl  = ext4_get_acl,
 982         .set_acl        = ext4_set_acl,
 983         .fiemap         = ext4_fiemap,
 984         .fileattr_get   = ext4_fileattr_get,
 985         .fileattr_set   = ext4_fileattr_set,
 986 };
 987