fs/ext4/file.c

   1 /*
   2  *  linux/fs/ext4/file.c
   3  *
   4  * Copyright (C) 1992, 1993, 1994, 1995
   5  * Remy Card (card@masi.ibp.fr)
   6  * Laboratoire MASI - Institut Blaise Pascal
   7  * Universite Pierre et Marie Curie (Paris VI)
   8  *
   9  *  from
  10  *
  11  *  linux/fs/minix/file.c
  12  *
  13  *  Copyright (C) 1991, 1992  Linus Torvalds
  14  *
  15  *  ext4 fs regular file handling primitives
  16  *
  17  *  64-bit file support on 64-bit platforms by Jakub Jelinek
  18  *      (jj@sunsite.ms.mff.cuni.cz)
  19  */
  20
  21 #include <linux/time.h>
  22 #include <linux/fs.h>
  23 #include <linux/jbd2.h>
  24 #include <linux/mount.h>
  25 #include <linux/path.h>
  26 #include <linux/quotaops.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/uio.h>
  29 #include "ext4.h"
  30 #include "ext4_jbd2.h"
  31 #include "xattr.h"
  32 #include "acl.h"
  33
  34 /*
  35  * Called when an inode is released. Note that this is different
  36  * from ext4_file_open: open gets called at every open, but release
  37  * gets called only when /all/ the files are closed.
  38  */
  39 static int ext4_release_file(struct inode *inode, struct file *filp)
  40 {
  41         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
  42                 ext4_alloc_da_blocks(inode);
  43                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
  44         }
  45         /* if we are the last writer on the inode, drop the block reservation */
  46         if ((filp->f_mode & FMODE_WRITE) &&
  47                         (atomic_read(&inode->i_writecount) == 1) &&
  48                         !EXT4_I(inode)->i_reserved_data_blocks)
  49         {
  50                 down_write(&EXT4_I(inode)->i_data_sem);
  51                 ext4_discard_preallocations(inode);
  52                 up_write(&EXT4_I(inode)->i_data_sem);
  53         }
  54         if (is_dx(inode) && filp->private_data)
  55                 ext4_htree_free_dir_info(filp->private_data);
  56
  57         return 0;
  58 }
  59
  60 static void ext4_unwritten_wait(struct inode *inode)
  61 {
  62         wait_queue_head_t *wq = ext4_ioend_wq(inode);
  63
  64         wait_event(*wq, (atomic_read(&EXT4_I(inode)->i_unwritten) == 0));
  65 }
  66
  67 /*
  68  * This tests whether the IO in question is block-aligned or not.
  69  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
  70  * are converted to written only after the IO is complete.  Until they are
  71  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
  72  * it needs to zero out portions of the start and/or end block.  If 2 AIO
  73  * threads are at work on the same unwritten block, they must be synchronized
  74  * or one thread will zero the other's data, causing corruption.
  75  */
  76 static int
  77 ext4_unaligned_aio(struct inode *inode, struct iov_iter *from, loff_t pos)
  78 {
  79         struct super_block *sb = inode->i_sb;
  80         int blockmask = sb->s_blocksize - 1;
  81
  82         if (pos >= i_size_read(inode))
  83                 return 0;
  84
  85         if ((pos | iov_iter_alignment(from)) & blockmask)
  86                 return 1;
  87
  88         return 0;
  89 }
  90
  91 static ssize_t
  92 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
  93 {
  94         struct file *file = iocb->ki_filp;
  95         struct inode *inode = file_inode(iocb->ki_filp);
  96         struct mutex *aio_mutex = NULL;
  97         struct blk_plug plug;
  98         int o_direct = iocb->ki_flags & IOCB_DIRECT;
  99         int overwrite = 0;
 100         ssize_t ret;
 101
 102         /*
 103          * Unaligned direct AIO must be serialized; see comment above
 104          * In the case of O_APPEND, assume that we must always serialize
 105          */
 106         if (o_direct &&
 107             ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) &&
 108             !is_sync_kiocb(iocb) &&
 109             (iocb->ki_flags & IOCB_APPEND ||
 110              ext4_unaligned_aio(inode, from, iocb->ki_pos))) {
 111                 aio_mutex = ext4_aio_mutex(inode);
 112                 mutex_lock(aio_mutex);
 113                 ext4_unwritten_wait(inode);
 114         }
 115
 116         mutex_lock(&inode->i_mutex);
 117         ret = generic_write_checks(iocb, from);
 118         if (ret <= 0)
 119                 goto out;
 120
 121         /*
 122          * If we have encountered a bitmap-format file, the size limit
 123          * is smaller than s_maxbytes, which is for extent-mapped files.
 124          */
 125         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 126                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 127
 128                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes) {
 129                         ret = -EFBIG;
 130                         goto out;
 131                 }
 132                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 133         }
 134
 135         iocb->private = &overwrite;
 136         if (o_direct) {
 137                 size_t length = iov_iter_count(from);
 138                 loff_t pos = iocb->ki_pos;
 139                 blk_start_plug(&plug);
 140
 141                 /* check whether we do a DIO overwrite or not */
 142                 if (ext4_should_dioread_nolock(inode) && !aio_mutex &&
 143                     !file->f_mapping->nrpages && pos + length <= i_size_read(inode)) {
 144                         struct ext4_map_blocks map;
 145                         unsigned int blkbits = inode->i_blkbits;
 146                         int err, len;
 147
 148                         map.m_lblk = pos >> blkbits;
 149                         map.m_len = (EXT4_BLOCK_ALIGN(pos + length, blkbits) >> blkbits)
 150                                 - map.m_lblk;
 151                         len = map.m_len;
 152
 153                         err = ext4_map_blocks(NULL, inode, &map, 0);
 154                         /*
 155                          * 'err==len' means that all of blocks has
 156                          * been preallocated no matter they are
 157                          * initialized or not.  For excluding
 158                          * unwritten extents, we need to check
 159                          * m_flags.  There are two conditions that
 160                          * indicate for initialized extents.  1) If we
 161                          * hit extent cache, EXT4_MAP_MAPPED flag is
 162                          * returned; 2) If we do a real lookup,
 163                          * non-flags are returned.  So we should check
 164                          * these two conditions.
 165                          */
 166                         if (err == len && (map.m_flags & EXT4_MAP_MAPPED))
 167                                 overwrite = 1;
 168                 }
 169         }
 170
 171         ret = __generic_file_write_iter(iocb, from);
 172         mutex_unlock(&inode->i_mutex);
 173
 174         if (ret > 0) {
 175                 ssize_t err;
 176
 177                 err = generic_write_sync(file, iocb->ki_pos - ret, ret);
 178                 if (err < 0)
 179                         ret = err;
 180         }
 181         if (o_direct)
 182                 blk_finish_plug(&plug);
 183
 184         if (aio_mutex)
 185                 mutex_unlock(aio_mutex);
 186         return ret;
 187
 188 out:
 189         mutex_unlock(&inode->i_mutex);
 190         if (aio_mutex)
 191                 mutex_unlock(aio_mutex);
 192         return ret;
 193 }
 194
 195 #ifdef CONFIG_FS_DAX
 196 static int ext4_dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 197 {
 198         return dax_fault(vma, vmf, ext4_get_block);
 199                                         /* Is this the right get_block? */
 200 }
 201
 202 static int ext4_dax_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 203 {
 204         return dax_mkwrite(vma, vmf, ext4_get_block);
 205 }
 206
 207 static const struct vm_operations_struct ext4_dax_vm_ops = {
 208         .fault          = ext4_dax_fault,
 209         .page_mkwrite   = ext4_dax_mkwrite,
 210 };
 211 #else
 212 #define ext4_dax_vm_ops ext4_file_vm_ops
 213 #endif
 214
 215 static const struct vm_operations_struct ext4_file_vm_ops = {
 216         .fault          = filemap_fault,
 217         .map_pages      = filemap_map_pages,
 218         .page_mkwrite   = ext4_page_mkwrite,
 219 };
 220
 221 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 222 {
 223         file_accessed(file);
 224         if (IS_DAX(file_inode(file))) {
 225                 vma->vm_ops = &ext4_dax_vm_ops;
 226                 vma->vm_flags |= VM_MIXEDMAP;
 227         } else {
 228                 vma->vm_ops = &ext4_file_vm_ops;
 229         }
 230         return 0;
 231 }
 232
 233 static int ext4_file_open(struct inode * inode, struct file * filp)
 234 {
 235         struct super_block *sb = inode->i_sb;
 236         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 237         struct vfsmount *mnt = filp->f_path.mnt;
 238         struct path path;
 239         char buf[64], *cp;
 240
 241         if (unlikely(!(sbi->s_mount_flags & EXT4_MF_MNTDIR_SAMPLED) &&
 242                      !(sb->s_flags & MS_RDONLY))) {
 243                 sbi->s_mount_flags |= EXT4_MF_MNTDIR_SAMPLED;
 244                 /*
 245                  * Sample where the filesystem has been mounted and
 246                  * store it in the superblock for sysadmin convenience
 247                  * when trying to sort through large numbers of block
 248                  * devices or filesystem images.
 249                  */
 250                 memset(buf, 0, sizeof(buf));
 251                 path.mnt = mnt;
 252                 path.dentry = mnt->mnt_root;
 253                 cp = d_path(&path, buf, sizeof(buf));
 254                 if (!IS_ERR(cp)) {
 255                         handle_t *handle;
 256                         int err;
 257
 258                         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 259                         if (IS_ERR(handle))
 260                                 return PTR_ERR(handle);
 261                         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 262                         err = ext4_journal_get_write_access(handle, sbi->s_sbh);
 263                         if (err) {
 264                                 ext4_journal_stop(handle);
 265                                 return err;
 266                         }
 267                         strlcpy(sbi->s_es->s_last_mounted, cp,
 268                                 sizeof(sbi->s_es->s_last_mounted));
 269                         ext4_handle_dirty_super(handle, sb);
 270                         ext4_journal_stop(handle);
 271                 }
 272         }
 273         /*
 274          * Set up the jbd2_inode if we are opening the inode for
 275          * writing and the journal is present
 276          */
 277         if (filp->f_mode & FMODE_WRITE) {
 278                 int ret = ext4_inode_attach_jinode(inode);
 279                 if (ret < 0)
 280                         return ret;
 281         }
 282         return dquot_file_open(inode, filp);
 283 }
 284
 285 /*
 286  * Here we use ext4_map_blocks() to get a block mapping for a extent-based
 287  * file rather than ext4_ext_walk_space() because we can introduce
 288  * SEEK_DATA/SEEK_HOLE for block-mapped and extent-mapped file at the same
 289  * function.  When extent status tree has been fully implemented, it will
 290  * track all extent status for a file and we can directly use it to
 291  * retrieve the offset for SEEK_DATA/SEEK_HOLE.
 292  */
 293
 294 /*
 295  * When we retrieve the offset for SEEK_DATA/SEEK_HOLE, we would need to
 296  * lookup page cache to check whether or not there has some data between
 297  * [startoff, endoff] because, if this range contains an unwritten extent,
 298  * we determine this extent as a data or a hole according to whether the
 299  * page cache has data or not.
 300  */
 301 static int ext4_find_unwritten_pgoff(struct inode *inode,
 302                                      int whence,
 303                                      struct ext4_map_blocks *map,
 304                                      loff_t *offset)
 305 {
 306         struct pagevec pvec;
 307         unsigned int blkbits;
 308         pgoff_t index;
 309         pgoff_t end;
 310         loff_t endoff;
 311         loff_t startoff;
 312         loff_t lastoff;
 313         int found = 0;
 314
 315         blkbits = inode->i_sb->s_blocksize_bits;
 316         startoff = *offset;
 317         lastoff = startoff;
 318         endoff = (loff_t)(map->m_lblk + map->m_len) << blkbits;
 319
 320         index = startoff >> PAGE_CACHE_SHIFT;
 321         end = endoff >> PAGE_CACHE_SHIFT;
 322
 323         pagevec_init(&pvec, 0);
 324         do {
 325                 int i, num;
 326                 unsigned long nr_pages;
 327
 328                 num = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
 329                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
 330                                           (pgoff_t)num);
 331                 if (nr_pages == 0) {
 332                         if (whence == SEEK_DATA)
 333                                 break;
 334
 335                         BUG_ON(whence != SEEK_HOLE);
 336                         /*
 337                          * If this is the first time to go into the loop and
 338                          * offset is not beyond the end offset, it will be a
 339                          * hole at this offset
 340                          */
 341                         if (lastoff == startoff || lastoff < endoff)
 342                                 found = 1;
 343                         break;
 344                 }
 345
 346                 /*
 347                  * If this is the first time to go into the loop and
 348                  * offset is smaller than the first page offset, it will be a
 349                  * hole at this offset.
 350                  */
 351                 if (lastoff == startoff && whence == SEEK_HOLE &&
 352                     lastoff < page_offset(pvec.pages[0])) {
 353                         found = 1;
 354                         break;
 355                 }
 356
 357                 for (i = 0; i < nr_pages; i++) {
 358                         struct page *page = pvec.pages[i];
 359                         struct buffer_head *bh, *head;
 360
 361                         /*
 362                          * If the current offset is not beyond the end of given
 363                          * range, it will be a hole.
 364                          */
 365                         if (lastoff < endoff && whence == SEEK_HOLE &&
 366                             page->index > end) {
 367                                 found = 1;
 368                                 *offset = lastoff;
 369                                 goto out;
 370                         }
 371
 372                         lock_page(page);
 373
 374                         if (unlikely(page->mapping != inode->i_mapping)) {
 375                                 unlock_page(page);
 376                                 continue;
 377                         }
 378
 379                         if (!page_has_buffers(page)) {
 380                                 unlock_page(page);
 381                                 continue;
 382                         }
 383
 384                         if (page_has_buffers(page)) {
 385                                 lastoff = page_offset(page);
 386                                 bh = head = page_buffers(page);
 387                                 do {
 388                                         if (buffer_uptodate(bh) ||
 389                                             buffer_unwritten(bh)) {
 390                                                 if (whence == SEEK_DATA)
 391                                                         found = 1;
 392                                         } else {
 393                                                 if (whence == SEEK_HOLE)
 394                                                         found = 1;
 395                                         }
 396                                         if (found) {
 397                                                 *offset = max_t(loff_t,
 398                                                         startoff, lastoff);
 399                                                 unlock_page(page);
 400                                                 goto out;
 401                                         }
 402                                         lastoff += bh->b_size;
 403                                         bh = bh->b_this_page;
 404                                 } while (bh != head);
 405                         }
 406
 407                         lastoff = page_offset(page) + PAGE_SIZE;
 408                         unlock_page(page);
 409                 }
 410
 411                 /*
 412                  * The no. of pages is less than our desired, that would be a
 413                  * hole in there.
 414                  */
 415                 if (nr_pages < num && whence == SEEK_HOLE) {
 416                         found = 1;
 417                         *offset = lastoff;
 418                         break;
 419                 }
 420
 421                 index = pvec.pages[i - 1]->index + 1;
 422                 pagevec_release(&pvec);
 423         } while (index <= end);
 424
 425 out:
 426         pagevec_release(&pvec);
 427         return found;
 428 }
 429
 430 /*
 431  * ext4_seek_data() retrieves the offset for SEEK_DATA.
 432  */
 433 static loff_t ext4_seek_data(struct file *file, loff_t offset, loff_t maxsize)
 434 {
 435         struct inode *inode = file->f_mapping->host;
 436         struct ext4_map_blocks map;
 437         struct extent_status es;
 438         ext4_lblk_t start, last, end;
 439         loff_t dataoff, isize;
 440         int blkbits;
 441         int ret = 0;
 442
 443         mutex_lock(&inode->i_mutex);
 444
 445         isize = i_size_read(inode);
 446         if (offset >= isize) {
 447                 mutex_unlock(&inode->i_mutex);
 448                 return -ENXIO;
 449         }
 450
 451         blkbits = inode->i_sb->s_blocksize_bits;
 452         start = offset >> blkbits;
 453         last = start;
 454         end = isize >> blkbits;
 455         dataoff = offset;
 456
 457         do {
 458                 map.m_lblk = last;
 459                 map.m_len = end - last + 1;
 460                 ret = ext4_map_blocks(NULL, inode, &map, 0);
 461                 if (ret > 0 && !(map.m_flags & EXT4_MAP_UNWRITTEN)) {
 462                         if (last != start)
 463                                 dataoff = (loff_t)last << blkbits;
 464                         break;
 465                 }
 466
 467                 /*
 468                  * If there is a delay extent at this offset,
 469                  * it will be as a data.
 470                  */
 471                 ext4_es_find_delayed_extent_range(inode, last, last, &es);
 472                 if (es.es_len != 0 && in_range(last, es.es_lblk, es.es_len)) {
 473                         if (last != start)
 474                                 dataoff = (loff_t)last << blkbits;
 475                         break;
 476                 }
 477
 478                 /*
 479                  * If there is a unwritten extent at this offset,
 480                  * it will be as a data or a hole according to page
 481                  * cache that has data or not.
 482                  */
 483                 if (map.m_flags & EXT4_MAP_UNWRITTEN) {
 484                         int unwritten;
 485                         unwritten = ext4_find_unwritten_pgoff(inode, SEEK_DATA,
 486                                                               &map, &dataoff);
 487                         if (unwritten)
 488                                 break;
 489                 }
 490
 491                 last++;
 492                 dataoff = (loff_t)last << blkbits;
 493         } while (last <= end);
 494
 495         mutex_unlock(&inode->i_mutex);
 496
 497         if (dataoff > isize)
 498                 return -ENXIO;
 499
 500         return vfs_setpos(file, dataoff, maxsize);
 501 }
 502
 503 /*
 504  * ext4_seek_hole() retrieves the offset for SEEK_HOLE.
 505  */
 506 static loff_t ext4_seek_hole(struct file *file, loff_t offset, loff_t maxsize)
 507 {
 508         struct inode *inode = file->f_mapping->host;
 509         struct ext4_map_blocks map;
 510         struct extent_status es;
 511         ext4_lblk_t start, last, end;
 512         loff_t holeoff, isize;
 513         int blkbits;
 514         int ret = 0;
 515
 516         mutex_lock(&inode->i_mutex);
 517
 518         isize = i_size_read(inode);
 519         if (offset >= isize) {
 520                 mutex_unlock(&inode->i_mutex);
 521                 return -ENXIO;
 522         }
 523
 524         blkbits = inode->i_sb->s_blocksize_bits;
 525         start = offset >> blkbits;
 526         last = start;
 527         end = isize >> blkbits;
 528         holeoff = offset;
 529
 530         do {
 531                 map.m_lblk = last;
 532                 map.m_len = end - last + 1;
 533                 ret = ext4_map_blocks(NULL, inode, &map, 0);
 534                 if (ret > 0 && !(map.m_flags & EXT4_MAP_UNWRITTEN)) {
 535                         last += ret;
 536                         holeoff = (loff_t)last << blkbits;
 537                         continue;
 538                 }
 539
 540                 /*
 541                  * If there is a delay extent at this offset,
 542                  * we will skip this extent.
 543                  */
 544                 ext4_es_find_delayed_extent_range(inode, last, last, &es);
 545                 if (es.es_len != 0 && in_range(last, es.es_lblk, es.es_len)) {
 546                         last = es.es_lblk + es.es_len;
 547                         holeoff = (loff_t)last << blkbits;
 548                         continue;
 549                 }
 550
 551                 /*
 552                  * If there is a unwritten extent at this offset,
 553                  * it will be as a data or a hole according to page
 554                  * cache that has data or not.
 555                  */
 556                 if (map.m_flags & EXT4_MAP_UNWRITTEN) {
 557                         int unwritten;
 558                         unwritten = ext4_find_unwritten_pgoff(inode, SEEK_HOLE,
 559                                                               &map, &holeoff);
 560                         if (!unwritten) {
 561                                 last += ret;
 562                                 holeoff = (loff_t)last << blkbits;
 563                                 continue;
 564                         }
 565                 }
 566
 567                 /* find a hole */
 568                 break;
 569         } while (last <= end);
 570
 571         mutex_unlock(&inode->i_mutex);
 572
 573         if (holeoff > isize)
 574                 holeoff = isize;
 575
 576         return vfs_setpos(file, holeoff, maxsize);
 577 }
 578
 579 /*
 580  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 581  * by calling generic_file_llseek_size() with the appropriate maxbytes
 582  * value for each.
 583  */
 584 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 585 {
 586         struct inode *inode = file->f_mapping->host;
 587         loff_t maxbytes;
 588
 589         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 590                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 591         else
 592                 maxbytes = inode->i_sb->s_maxbytes;
 593
 594         switch (whence) {
 595         case SEEK_SET:
 596         case SEEK_CUR:
 597         case SEEK_END:
 598                 return generic_file_llseek_size(file, offset, whence,
 599                                                 maxbytes, i_size_read(inode));
 600         case SEEK_DATA:
 601                 return ext4_seek_data(file, offset, maxbytes);
 602         case SEEK_HOLE:
 603                 return ext4_seek_hole(file, offset, maxbytes);
 604         }
 605
 606         return -EINVAL;
 607 }
 608
 609 const struct file_operations ext4_file_operations = {
 610         .llseek         = ext4_llseek,
 611         .read_iter      = generic_file_read_iter,
 612         .write_iter     = ext4_file_write_iter,
 613         .unlocked_ioctl = ext4_ioctl,
 614 #ifdef CONFIG_COMPAT
 615         .compat_ioctl   = ext4_compat_ioctl,
 616 #endif
 617         .mmap           = ext4_file_mmap,
 618         .open           = ext4_file_open,
 619         .release        = ext4_release_file,
 620         .fsync          = ext4_sync_file,
 621         .splice_read    = generic_file_splice_read,
 622         .splice_write   = iter_file_splice_write,
 623         .fallocate      = ext4_fallocate,
 624 };
 625
 626 #ifdef CONFIG_FS_DAX
 627 const struct file_operations ext4_dax_file_operations = {
 628         .llseek         = ext4_llseek,
 629         .read_iter      = generic_file_read_iter,
 630         .write_iter     = ext4_file_write_iter,
 631         .unlocked_ioctl = ext4_ioctl,
 632 #ifdef CONFIG_COMPAT
 633         .compat_ioctl   = ext4_compat_ioctl,
 634 #endif
 635         .mmap           = ext4_file_mmap,
 636         .open           = ext4_file_open,
 637         .release        = ext4_release_file,
 638         .fsync          = ext4_sync_file,
 639         /* Splice not yet supported with DAX */
 640         .fallocate      = ext4_fallocate,
 641 };
 642 #endif
 643
 644 const struct inode_operations ext4_file_inode_operations = {
 645         .setattr        = ext4_setattr,
 646         .getattr        = ext4_getattr,
 647         .setxattr       = generic_setxattr,
 648         .getxattr       = generic_getxattr,
 649         .listxattr      = ext4_listxattr,
 650         .removexattr    = generic_removexattr,
 651         .get_acl        = ext4_get_acl,
 652         .set_acl        = ext4_set_acl,
 653         .fiemap         = ext4_fiemap,
 654 };
 655