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 (card@masi.ibp.fr)
   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  *      (jj@sunsite.ms.mff.cuni.cz)
  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(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(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 count)
 310 {
 311         handle_t *handle;
 312
 313         lockdep_assert_held_write(&inode->i_rwsem);
 314         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 315         if (IS_ERR(handle))
 316                 return PTR_ERR(handle);
 317
 318         if (ext4_update_inode_size(inode, offset + count)) {
 319                 int ret = ext4_mark_inode_dirty(handle, inode);
 320                 if (unlikely(ret)) {
 321                         ext4_journal_stop(handle);
 322                         return ret;
 323                 }
 324         }
 325
 326         if (inode->i_nlink)
 327                 ext4_orphan_del(handle, inode);
 328         ext4_journal_stop(handle);
 329
 330         return count;
 331 }
 332
 333 /*
 334  * Clean up the inode after DIO or DAX extending write has completed and the
 335  * inode size has been updated using ext4_handle_inode_extension().
 336  */
 337 static void ext4_inode_extension_cleanup(struct inode *inode, ssize_t count)
 338 {
 339         lockdep_assert_held_write(&inode->i_rwsem);
 340         if (count < 0) {
 341                 ext4_truncate_failed_write(inode);
 342                 /*
 343                  * If the truncate operation failed early, then the inode may
 344                  * still be on the orphan list. In that case, we need to try
 345                  * remove the inode from the in-memory linked list.
 346                  */
 347                 if (inode->i_nlink)
 348                         ext4_orphan_del(NULL, inode);
 349                 return;
 350         }
 351         /*
 352          * If i_disksize got extended due to writeback of delalloc blocks while
 353          * the DIO was running we could fail to cleanup the orphan list in
 354          * ext4_handle_inode_extension(). Do it now.
 355          */
 356         if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
 357                 handle_t *handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 358
 359                 if (IS_ERR(handle)) {
 360                         /*
 361                          * The write has successfully completed. Not much to
 362                          * do with the error here so just cleanup the orphan
 363                          * list and hope for the best.
 364                          */
 365                         ext4_orphan_del(NULL, inode);
 366                         return;
 367                 }
 368                 ext4_orphan_del(handle, inode);
 369                 ext4_journal_stop(handle);
 370         }
 371 }
 372
 373 static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
 374                                  int error, unsigned int flags)
 375 {
 376         loff_t pos = iocb->ki_pos;
 377         struct inode *inode = file_inode(iocb->ki_filp);
 378
 379         if (!error && size && flags & IOMAP_DIO_UNWRITTEN)
 380                 error = ext4_convert_unwritten_extents(NULL, inode, pos, size);
 381         if (error)
 382                 return error;
 383         /*
 384          * Note that EXT4_I(inode)->i_disksize can get extended up to
 385          * inode->i_size while the I/O was running due to writeback of delalloc
 386          * blocks. But the code in ext4_iomap_alloc() is careful to use
 387          * zeroed/unwritten extents if this is possible; thus we won't leave
 388          * uninitialized blocks in a file even if we didn't succeed in writing
 389          * as much as we intended.
 390          */
 391         WARN_ON_ONCE(i_size_read(inode) < READ_ONCE(EXT4_I(inode)->i_disksize));
 392         if (pos + size <= READ_ONCE(EXT4_I(inode)->i_disksize))
 393                 return size;
 394         return ext4_handle_inode_extension(inode, pos, size);
 395 }
 396
 397 static const struct iomap_dio_ops ext4_dio_write_ops = {
 398         .end_io = ext4_dio_write_end_io,
 399 };
 400
 401 /*
 402  * The intention here is to start with shared lock acquired then see if any
 403  * condition requires an exclusive inode lock. If yes, then we restart the
 404  * whole operation by releasing the shared lock and acquiring exclusive lock.
 405  *
 406  * - For unaligned_io we never take shared lock as it may cause data corruption
 407  *   when two unaligned IO tries to modify the same block e.g. while zeroing.
 408  *
 409  * - For extending writes case we don't take the shared lock, since it requires
 410  *   updating inode i_disksize and/or orphan handling with exclusive lock.
 411  *
 412  * - shared locking will only be true mostly with overwrites, including
 413  *   initialized blocks and unwritten blocks. For overwrite unwritten blocks
 414  *   we protect splitting extents by i_data_sem in ext4_inode_info, so we can
 415  *   also release exclusive i_rwsem lock.
 416  *
 417  * - Otherwise we will switch to exclusive i_rwsem lock.
 418  */
 419 static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
 420                                      bool *ilock_shared, bool *extend,
 421                                      bool *unwritten, int *dio_flags)
 422 {
 423         struct file *file = iocb->ki_filp;
 424         struct inode *inode = file_inode(file);
 425         loff_t offset;
 426         size_t count;
 427         ssize_t ret;
 428         bool overwrite, unaligned_io;
 429
 430 restart:
 431         ret = ext4_generic_write_checks(iocb, from);
 432         if (ret <= 0)
 433                 goto out;
 434
 435         offset = iocb->ki_pos;
 436         count = ret;
 437
 438         unaligned_io = ext4_unaligned_io(inode, from, offset);
 439         *extend = ext4_extending_io(inode, offset, count);
 440         overwrite = ext4_overwrite_io(inode, offset, count, unwritten);
 441
 442         /*
 443          * Determine whether we need to upgrade to an exclusive lock. This is
 444          * required to change security info in file_modified(), for extending
 445          * I/O, any form of non-overwrite I/O, and unaligned I/O to unwritten
 446          * extents (as partial block zeroing may be required).
 447          *
 448          * Note that unaligned writes are allowed under shared lock so long as
 449          * they are pure overwrites. Otherwise, concurrent unaligned writes risk
 450          * data corruption due to partial block zeroing in the dio layer, and so
 451          * the I/O must occur exclusively.
 452          */
 453         if (*ilock_shared &&
 454             ((!IS_NOSEC(inode) || *extend || !overwrite ||
 455              (unaligned_io && *unwritten)))) {
 456                 if (iocb->ki_flags & IOCB_NOWAIT) {
 457                         ret = -EAGAIN;
 458                         goto out;
 459                 }
 460                 inode_unlock_shared(inode);
 461                 *ilock_shared = false;
 462                 inode_lock(inode);
 463                 goto restart;
 464         }
 465
 466         /*
 467          * Now that locking is settled, determine dio flags and exclusivity
 468          * requirements. We don't use DIO_OVERWRITE_ONLY because we enforce
 469          * behavior already. The inode lock is already held exclusive if the
 470          * write is non-overwrite or extending, so drain all outstanding dio and
 471          * set the force wait dio flag.
 472          */
 473         if (!*ilock_shared && (unaligned_io || *extend)) {
 474                 if (iocb->ki_flags & IOCB_NOWAIT) {
 475                         ret = -EAGAIN;
 476                         goto out;
 477                 }
 478                 if (unaligned_io && (!overwrite || *unwritten))
 479                         inode_dio_wait(inode);
 480                 *dio_flags = IOMAP_DIO_FORCE_WAIT;
 481         }
 482
 483         ret = file_modified(file);
 484         if (ret < 0)
 485                 goto out;
 486
 487         return count;
 488 out:
 489         if (*ilock_shared)
 490                 inode_unlock_shared(inode);
 491         else
 492                 inode_unlock(inode);
 493         return ret;
 494 }
 495
 496 static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 497 {
 498         ssize_t ret;
 499         handle_t *handle;
 500         struct inode *inode = file_inode(iocb->ki_filp);
 501         loff_t offset = iocb->ki_pos;
 502         size_t count = iov_iter_count(from);
 503         const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
 504         bool extend = false, unwritten = false;
 505         bool ilock_shared = true;
 506         int dio_flags = 0;
 507
 508         /*
 509          * Quick check here without any i_rwsem lock to see if it is extending
 510          * IO. A more reliable check is done in ext4_dio_write_checks() with
 511          * proper locking in place.
 512          */
 513         if (offset + count > i_size_read(inode))
 514                 ilock_shared = false;
 515
 516         if (iocb->ki_flags & IOCB_NOWAIT) {
 517                 if (ilock_shared) {
 518                         if (!inode_trylock_shared(inode))
 519                                 return -EAGAIN;
 520                 } else {
 521                         if (!inode_trylock(inode))
 522                                 return -EAGAIN;
 523                 }
 524         } else {
 525                 if (ilock_shared)
 526                         inode_lock_shared(inode);
 527                 else
 528                         inode_lock(inode);
 529         }
 530
 531         /* Fallback to buffered I/O if the inode does not support direct I/O. */
 532         if (!ext4_should_use_dio(iocb, from)) {
 533                 if (ilock_shared)
 534                         inode_unlock_shared(inode);
 535                 else
 536                         inode_unlock(inode);
 537                 return ext4_buffered_write_iter(iocb, from);
 538         }
 539
 540         /*
 541          * Prevent inline data from being created since we are going to allocate
 542          * blocks for DIO. We know the inode does not currently have inline data
 543          * because ext4_should_use_dio() checked for it, but we have to clear
 544          * the state flag before the write checks because a lock cycle could
 545          * introduce races with other writers.
 546          */
 547         ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
 548
 549         ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend,
 550                                     &unwritten, &dio_flags);
 551         if (ret <= 0)
 552                 return ret;
 553
 554         offset = iocb->ki_pos;
 555         count = ret;
 556
 557         if (extend) {
 558                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 559                 if (IS_ERR(handle)) {
 560                         ret = PTR_ERR(handle);
 561                         goto out;
 562                 }
 563
 564                 ret = ext4_orphan_add(handle, inode);
 565                 if (ret) {
 566                         ext4_journal_stop(handle);
 567                         goto out;
 568                 }
 569
 570                 ext4_journal_stop(handle);
 571         }
 572
 573         if (ilock_shared && !unwritten)
 574                 iomap_ops = &ext4_iomap_overwrite_ops;
 575         ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 576                            dio_flags, NULL, 0);
 577         if (ret == -ENOTBLK)
 578                 ret = 0;
 579         if (extend) {
 580                 /*
 581                  * We always perform extending DIO write synchronously so by
 582                  * now the IO is completed and ext4_handle_inode_extension()
 583                  * was called. Cleanup the inode in case of error or race with
 584                  * writeback of delalloc blocks.
 585                  */
 586                 WARN_ON_ONCE(ret == -EIOCBQUEUED);
 587                 ext4_inode_extension_cleanup(inode, ret);
 588         }
 589
 590 out:
 591         if (ilock_shared)
 592                 inode_unlock_shared(inode);
 593         else
 594                 inode_unlock(inode);
 595
 596         if (ret >= 0 && iov_iter_count(from)) {
 597                 ssize_t err;
 598                 loff_t endbyte;
 599
 600                 offset = iocb->ki_pos;
 601                 err = ext4_buffered_write_iter(iocb, from);
 602                 if (err < 0)
 603                         return err;
 604
 605                 /*
 606                  * We need to ensure that the pages within the page cache for
 607                  * the range covered by this I/O are written to disk and
 608                  * invalidated. This is in attempt to preserve the expected
 609                  * direct I/O semantics in the case we fallback to buffered I/O
 610                  * to complete off the I/O request.
 611                  */
 612                 ret += err;
 613                 endbyte = offset + err - 1;
 614                 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
 615                                                    offset, endbyte);
 616                 if (!err)
 617                         invalidate_mapping_pages(iocb->ki_filp->f_mapping,
 618                                                  offset >> PAGE_SHIFT,
 619                                                  endbyte >> PAGE_SHIFT);
 620         }
 621
 622         return ret;
 623 }
 624
 625 #ifdef CONFIG_FS_DAX
 626 static ssize_t
 627 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 628 {
 629         ssize_t ret;
 630         size_t count;
 631         loff_t offset;
 632         handle_t *handle;
 633         bool extend = false;
 634         struct inode *inode = file_inode(iocb->ki_filp);
 635
 636         if (iocb->ki_flags & IOCB_NOWAIT) {
 637                 if (!inode_trylock(inode))
 638                         return -EAGAIN;
 639         } else {
 640                 inode_lock(inode);
 641         }
 642
 643         ret = ext4_write_checks(iocb, from);
 644         if (ret <= 0)
 645                 goto out;
 646
 647         offset = iocb->ki_pos;
 648         count = iov_iter_count(from);
 649
 650         if (offset + count > EXT4_I(inode)->i_disksize) {
 651                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 652                 if (IS_ERR(handle)) {
 653                         ret = PTR_ERR(handle);
 654                         goto out;
 655                 }
 656
 657                 ret = ext4_orphan_add(handle, inode);
 658                 if (ret) {
 659                         ext4_journal_stop(handle);
 660                         goto out;
 661                 }
 662
 663                 extend = true;
 664                 ext4_journal_stop(handle);
 665         }
 666
 667         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 668
 669         if (extend) {
 670                 ret = ext4_handle_inode_extension(inode, offset, ret);
 671                 ext4_inode_extension_cleanup(inode, ret);
 672         }
 673 out:
 674         inode_unlock(inode);
 675         if (ret > 0)
 676                 ret = generic_write_sync(iocb, ret);
 677         return ret;
 678 }
 679 #endif
 680
 681 static ssize_t
 682 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 683 {
 684         struct inode *inode = file_inode(iocb->ki_filp);
 685
 686         if (unlikely(ext4_forced_shutdown(inode->i_sb)))
 687                 return -EIO;
 688
 689 #ifdef CONFIG_FS_DAX
 690         if (IS_DAX(inode))
 691                 return ext4_dax_write_iter(iocb, from);
 692 #endif
 693         if (iocb->ki_flags & IOCB_DIRECT)
 694                 return ext4_dio_write_iter(iocb, from);
 695         else
 696                 return ext4_buffered_write_iter(iocb, from);
 697 }
 698
 699 #ifdef CONFIG_FS_DAX
 700 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf, unsigned int order)
 701 {
 702         int error = 0;
 703         vm_fault_t result;
 704         int retries = 0;
 705         handle_t *handle = NULL;
 706         struct inode *inode = file_inode(vmf->vma->vm_file);
 707         struct super_block *sb = inode->i_sb;
 708
 709         /*
 710          * We have to distinguish real writes from writes which will result in a
 711          * COW page; COW writes should *not* poke the journal (the file will not
 712          * be changed). Doing so would cause unintended failures when mounted
 713          * read-only.
 714          *
 715          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 716          * unset for order != 0 (i.e. only in do_cow_fault); for
 717          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 718          * we eventually come back with a COW page.
 719          */
 720         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 721                 (vmf->vma->vm_flags & VM_SHARED);
 722         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
 723         pfn_t pfn;
 724
 725         if (write) {
 726                 sb_start_pagefault(sb);
 727                 file_update_time(vmf->vma->vm_file);
 728                 filemap_invalidate_lock_shared(mapping);
 729 retry:
 730                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 731                                                EXT4_DATA_TRANS_BLOCKS(sb));
 732                 if (IS_ERR(handle)) {
 733                         filemap_invalidate_unlock_shared(mapping);
 734                         sb_end_pagefault(sb);
 735                         return VM_FAULT_SIGBUS;
 736                 }
 737         } else {
 738                 filemap_invalidate_lock_shared(mapping);
 739         }
 740         result = dax_iomap_fault(vmf, order, &pfn, &error, &ext4_iomap_ops);
 741         if (write) {
 742                 ext4_journal_stop(handle);
 743
 744                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 745                     ext4_should_retry_alloc(sb, &retries))
 746                         goto retry;
 747                 /* Handling synchronous page fault? */
 748                 if (result & VM_FAULT_NEEDDSYNC)
 749                         result = dax_finish_sync_fault(vmf, order, pfn);
 750                 filemap_invalidate_unlock_shared(mapping);
 751                 sb_end_pagefault(sb);
 752         } else {
 753                 filemap_invalidate_unlock_shared(mapping);
 754         }
 755
 756         return result;
 757 }
 758
 759 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 760 {
 761         return ext4_dax_huge_fault(vmf, 0);
 762 }
 763
 764 static const struct vm_operations_struct ext4_dax_vm_ops = {
 765         .fault          = ext4_dax_fault,
 766         .huge_fault     = ext4_dax_huge_fault,
 767         .page_mkwrite   = ext4_dax_fault,
 768         .pfn_mkwrite    = ext4_dax_fault,
 769 };
 770 #else
 771 #define ext4_dax_vm_ops ext4_file_vm_ops
 772 #endif
 773
 774 static const struct vm_operations_struct ext4_file_vm_ops = {
 775         .fault          = filemap_fault,
 776         .map_pages      = filemap_map_pages,
 777         .page_mkwrite   = ext4_page_mkwrite,
 778 };
 779
 780 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 781 {
 782         struct inode *inode = file->f_mapping->host;
 783         struct dax_device *dax_dev = EXT4_SB(inode->i_sb)->s_daxdev;
 784
 785         if (unlikely(ext4_forced_shutdown(inode->i_sb)))
 786                 return -EIO;
 787
 788         /*
 789          * We don't support synchronous mappings for non-DAX files and
 790          * for DAX files if underneath dax_device is not synchronous.
 791          */
 792         if (!daxdev_mapping_supported(vma, dax_dev))
 793                 return -EOPNOTSUPP;
 794
 795         file_accessed(file);
 796         if (IS_DAX(file_inode(file))) {
 797                 vma->vm_ops = &ext4_dax_vm_ops;
 798                 vm_flags_set(vma, VM_HUGEPAGE);
 799         } else {
 800                 vma->vm_ops = &ext4_file_vm_ops;
 801         }
 802         return 0;
 803 }
 804
 805 static int ext4_sample_last_mounted(struct super_block *sb,
 806                                     struct vfsmount *mnt)
 807 {
 808         struct ext4_sb_info *sbi = EXT4_SB(sb);
 809         struct path path;
 810         char buf[64], *cp;
 811         handle_t *handle;
 812         int err;
 813
 814         if (likely(ext4_test_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED)))
 815                 return 0;
 816
 817         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 818                 return 0;
 819
 820         ext4_set_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED);
 821         /*
 822          * Sample where the filesystem has been mounted and
 823          * store it in the superblock for sysadmin convenience
 824          * when trying to sort through large numbers of block
 825          * devices or filesystem images.
 826          */
 827         memset(buf, 0, sizeof(buf));
 828         path.mnt = mnt;
 829         path.dentry = mnt->mnt_root;
 830         cp = d_path(&path, buf, sizeof(buf));
 831         err = 0;
 832         if (IS_ERR(cp))
 833                 goto out;
 834
 835         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 836         err = PTR_ERR(handle);
 837         if (IS_ERR(handle))
 838                 goto out;
 839         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 840         err = ext4_journal_get_write_access(handle, sb, sbi->s_sbh,
 841                                             EXT4_JTR_NONE);
 842         if (err)
 843                 goto out_journal;
 844         lock_buffer(sbi->s_sbh);
 845         strncpy(sbi->s_es->s_last_mounted, cp,
 846                 sizeof(sbi->s_es->s_last_mounted));
 847         ext4_superblock_csum_set(sb);
 848         unlock_buffer(sbi->s_sbh);
 849         ext4_handle_dirty_metadata(handle, NULL, sbi->s_sbh);
 850 out_journal:
 851         ext4_journal_stop(handle);
 852 out:
 853         sb_end_intwrite(sb);
 854         return err;
 855 }
 856
 857 static int ext4_file_open(struct inode *inode, struct file *filp)
 858 {
 859         int ret;
 860
 861         if (unlikely(ext4_forced_shutdown(inode->i_sb)))
 862                 return -EIO;
 863
 864         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 865         if (ret)
 866                 return ret;
 867
 868         ret = fscrypt_file_open(inode, filp);
 869         if (ret)
 870                 return ret;
 871
 872         ret = fsverity_file_open(inode, filp);
 873         if (ret)
 874                 return ret;
 875
 876         /*
 877          * Set up the jbd2_inode if we are opening the inode for
 878          * writing and the journal is present
 879          */
 880         if (filp->f_mode & FMODE_WRITE) {
 881                 ret = ext4_inode_attach_jinode(inode);
 882                 if (ret < 0)
 883                         return ret;
 884         }
 885
 886         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC |
 887                         FMODE_DIO_PARALLEL_WRITE;
 888         return dquot_file_open(inode, filp);
 889 }
 890
 891 /*
 892  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 893  * by calling generic_file_llseek_size() with the appropriate maxbytes
 894  * value for each.
 895  */
 896 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 897 {
 898         struct inode *inode = file->f_mapping->host;
 899         loff_t maxbytes;
 900
 901         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 902                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 903         else
 904                 maxbytes = inode->i_sb->s_maxbytes;
 905
 906         switch (whence) {
 907         default:
 908                 return generic_file_llseek_size(file, offset, whence,
 909                                                 maxbytes, i_size_read(inode));
 910         case SEEK_HOLE:
 911                 inode_lock_shared(inode);
 912                 offset = iomap_seek_hole(inode, offset,
 913                                          &ext4_iomap_report_ops);
 914                 inode_unlock_shared(inode);
 915                 break;
 916         case SEEK_DATA:
 917                 inode_lock_shared(inode);
 918                 offset = iomap_seek_data(inode, offset,
 919                                          &ext4_iomap_report_ops);
 920                 inode_unlock_shared(inode);
 921                 break;
 922         }
 923
 924         if (offset < 0)
 925                 return offset;
 926         return vfs_setpos(file, offset, maxbytes);
 927 }
 928
 929 const struct file_operations ext4_file_operations = {
 930         .llseek         = ext4_llseek,
 931         .read_iter      = ext4_file_read_iter,
 932         .write_iter     = ext4_file_write_iter,
 933         .iopoll         = iocb_bio_iopoll,
 934         .unlocked_ioctl = ext4_ioctl,
 935 #ifdef CONFIG_COMPAT
 936         .compat_ioctl   = ext4_compat_ioctl,
 937 #endif
 938         .mmap           = ext4_file_mmap,
 939         .mmap_supported_flags = MAP_SYNC,
 940         .open           = ext4_file_open,
 941         .release        = ext4_release_file,
 942         .fsync          = ext4_sync_file,
 943         .get_unmapped_area = thp_get_unmapped_area,
 944         .splice_read    = ext4_file_splice_read,
 945         .splice_write   = iter_file_splice_write,
 946         .fallocate      = ext4_fallocate,
 947 };
 948
 949 const struct inode_operations ext4_file_inode_operations = {
 950         .setattr        = ext4_setattr,
 951         .getattr        = ext4_file_getattr,
 952         .listxattr      = ext4_listxattr,
 953         .get_inode_acl  = ext4_get_acl,
 954         .set_acl        = ext4_set_acl,
 955         .fiemap         = ext4_fiemap,
 956         .fileattr_get   = ext4_fileattr_get,
 957         .fileattr_set   = ext4_fileattr_set,
 958 };
 959