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 "ext4.h"
  33 #include "ext4_jbd2.h"
  34 #include "xattr.h"
  35 #include "acl.h"
  36
  37 #ifdef CONFIG_FS_DAX
  38 static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
  39 {
  40         struct inode *inode = file_inode(iocb->ki_filp);
  41         ssize_t ret;
  42
  43         if (!inode_trylock_shared(inode)) {
  44                 if (iocb->ki_flags & IOCB_NOWAIT)
  45                         return -EAGAIN;
  46                 inode_lock_shared(inode);
  47         }
  48         /*
  49          * Recheck under inode lock - at this point we are sure it cannot
  50          * change anymore
  51          */
  52         if (!IS_DAX(inode)) {
  53                 inode_unlock_shared(inode);
  54                 /* Fallback to buffered IO in case we cannot support DAX */
  55                 return generic_file_read_iter(iocb, to);
  56         }
  57         ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
  58         inode_unlock_shared(inode);
  59
  60         file_accessed(iocb->ki_filp);
  61         return ret;
  62 }
  63 #endif
  64
  65 static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
  66 {
  67         if (unlikely(ext4_forced_shutdown(EXT4_SB(file_inode(iocb->ki_filp)->i_sb))))
  68                 return -EIO;
  69
  70         if (!iov_iter_count(to))
  71                 return 0; /* skip atime */
  72
  73 #ifdef CONFIG_FS_DAX
  74         if (IS_DAX(file_inode(iocb->ki_filp)))
  75                 return ext4_dax_read_iter(iocb, to);
  76 #endif
  77         return generic_file_read_iter(iocb, to);
  78 }
  79
  80 /*
  81  * Called when an inode is released. Note that this is different
  82  * from ext4_file_open: open gets called at every open, but release
  83  * gets called only when /all/ the files are closed.
  84  */
  85 static int ext4_release_file(struct inode *inode, struct file *filp)
  86 {
  87         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
  88                 ext4_alloc_da_blocks(inode);
  89                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
  90         }
  91         /* if we are the last writer on the inode, drop the block reservation */
  92         if ((filp->f_mode & FMODE_WRITE) &&
  93                         (atomic_read(&inode->i_writecount) == 1) &&
  94                         !EXT4_I(inode)->i_reserved_data_blocks)
  95         {
  96                 down_write(&EXT4_I(inode)->i_data_sem);
  97                 ext4_discard_preallocations(inode);
  98                 up_write(&EXT4_I(inode)->i_data_sem);
  99         }
 100         if (is_dx(inode) && filp->private_data)
 101                 ext4_htree_free_dir_info(filp->private_data);
 102
 103         return 0;
 104 }
 105
 106 static void ext4_unwritten_wait(struct inode *inode)
 107 {
 108         wait_queue_head_t *wq = ext4_ioend_wq(inode);
 109
 110         wait_event(*wq, (atomic_read(&EXT4_I(inode)->i_unwritten) == 0));
 111 }
 112
 113 /*
 114  * This tests whether the IO in question is block-aligned or not.
 115  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
 116  * are converted to written only after the IO is complete.  Until they are
 117  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
 118  * it needs to zero out portions of the start and/or end block.  If 2 AIO
 119  * threads are at work on the same unwritten block, they must be synchronized
 120  * or one thread will zero the other's data, causing corruption.
 121  */
 122 static int
 123 ext4_unaligned_aio(struct inode *inode, struct iov_iter *from, loff_t pos)
 124 {
 125         struct super_block *sb = inode->i_sb;
 126         int blockmask = sb->s_blocksize - 1;
 127
 128         if (pos >= ALIGN(i_size_read(inode), sb->s_blocksize))
 129                 return 0;
 130
 131         if ((pos | iov_iter_alignment(from)) & blockmask)
 132                 return 1;
 133
 134         return 0;
 135 }
 136
 137 /* Is IO overwriting allocated and initialized blocks? */
 138 static bool ext4_overwrite_io(struct inode *inode, loff_t pos, loff_t len)
 139 {
 140         struct ext4_map_blocks map;
 141         unsigned int blkbits = inode->i_blkbits;
 142         int err, blklen;
 143
 144         if (pos + len > i_size_read(inode))
 145                 return false;
 146
 147         map.m_lblk = pos >> blkbits;
 148         map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
 149         blklen = map.m_len;
 150
 151         err = ext4_map_blocks(NULL, inode, &map, 0);
 152         /*
 153          * 'err==len' means that all of the blocks have been preallocated,
 154          * regardless of whether they have been initialized or not. To exclude
 155          * unwritten extents, we need to check m_flags.
 156          */
 157         return err == blklen && (map.m_flags & EXT4_MAP_MAPPED);
 158 }
 159
 160 static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
 161 {
 162         struct inode *inode = file_inode(iocb->ki_filp);
 163         ssize_t ret;
 164
 165         ret = generic_write_checks(iocb, from);
 166         if (ret <= 0)
 167                 return ret;
 168
 169         if (unlikely(IS_IMMUTABLE(inode)))
 170                 return -EPERM;
 171
 172         /*
 173          * If we have encountered a bitmap-format file, the size limit
 174          * is smaller than s_maxbytes, which is for extent-mapped files.
 175          */
 176         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 177                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 178
 179                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
 180                         return -EFBIG;
 181                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 182         }
 183         return iov_iter_count(from);
 184 }
 185
 186 #ifdef CONFIG_FS_DAX
 187 static ssize_t
 188 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 189 {
 190         struct inode *inode = file_inode(iocb->ki_filp);
 191         ssize_t ret;
 192
 193         if (!inode_trylock(inode)) {
 194                 if (iocb->ki_flags & IOCB_NOWAIT)
 195                         return -EAGAIN;
 196                 inode_lock(inode);
 197         }
 198         ret = ext4_write_checks(iocb, from);
 199         if (ret <= 0)
 200                 goto out;
 201         ret = file_remove_privs(iocb->ki_filp);
 202         if (ret)
 203                 goto out;
 204         ret = file_update_time(iocb->ki_filp);
 205         if (ret)
 206                 goto out;
 207
 208         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 209 out:
 210         inode_unlock(inode);
 211         if (ret > 0)
 212                 ret = generic_write_sync(iocb, ret);
 213         return ret;
 214 }
 215 #endif
 216
 217 static ssize_t
 218 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 219 {
 220         struct inode *inode = file_inode(iocb->ki_filp);
 221         int o_direct = iocb->ki_flags & IOCB_DIRECT;
 222         int unaligned_aio = 0;
 223         int overwrite = 0;
 224         ssize_t ret;
 225
 226         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 227                 return -EIO;
 228
 229 #ifdef CONFIG_FS_DAX
 230         if (IS_DAX(inode))
 231                 return ext4_dax_write_iter(iocb, from);
 232 #endif
 233
 234         if (!inode_trylock(inode)) {
 235                 if (iocb->ki_flags & IOCB_NOWAIT)
 236                         return -EAGAIN;
 237                 inode_lock(inode);
 238         }
 239
 240         ret = ext4_write_checks(iocb, from);
 241         if (ret <= 0)
 242                 goto out;
 243
 244         /*
 245          * Unaligned direct AIO must be serialized among each other as zeroing
 246          * of partial blocks of two competing unaligned AIOs can result in data
 247          * corruption.
 248          */
 249         if (o_direct && ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) &&
 250             !is_sync_kiocb(iocb) &&
 251             ext4_unaligned_aio(inode, from, iocb->ki_pos)) {
 252                 unaligned_aio = 1;
 253                 ext4_unwritten_wait(inode);
 254         }
 255
 256         iocb->private = &overwrite;
 257         /* Check whether we do a DIO overwrite or not */
 258         if (o_direct && !unaligned_aio) {
 259                 if (ext4_overwrite_io(inode, iocb->ki_pos, iov_iter_count(from))) {
 260                         if (ext4_should_dioread_nolock(inode))
 261                                 overwrite = 1;
 262                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
 263                         ret = -EAGAIN;
 264                         goto out;
 265                 }
 266         }
 267
 268         ret = __generic_file_write_iter(iocb, from);
 269         /*
 270          * Unaligned direct AIO must be the only IO in flight. Otherwise
 271          * overlapping aligned IO after unaligned might result in data
 272          * corruption.
 273          */
 274         if (ret == -EIOCBQUEUED && unaligned_aio)
 275                 ext4_unwritten_wait(inode);
 276         inode_unlock(inode);
 277
 278         if (ret > 0)
 279                 ret = generic_write_sync(iocb, ret);
 280
 281         return ret;
 282
 283 out:
 284         inode_unlock(inode);
 285         return ret;
 286 }
 287
 288 #ifdef CONFIG_FS_DAX
 289 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
 290                 enum page_entry_size pe_size)
 291 {
 292         int error = 0;
 293         vm_fault_t result;
 294         int retries = 0;
 295         handle_t *handle = NULL;
 296         struct inode *inode = file_inode(vmf->vma->vm_file);
 297         struct super_block *sb = inode->i_sb;
 298
 299         /*
 300          * We have to distinguish real writes from writes which will result in a
 301          * COW page; COW writes should *not* poke the journal (the file will not
 302          * be changed). Doing so would cause unintended failures when mounted
 303          * read-only.
 304          *
 305          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 306          * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
 307          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 308          * we eventually come back with a COW page.
 309          */
 310         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 311                 (vmf->vma->vm_flags & VM_SHARED);
 312         pfn_t pfn;
 313
 314         if (write) {
 315                 sb_start_pagefault(sb);
 316                 file_update_time(vmf->vma->vm_file);
 317                 down_read(&EXT4_I(inode)->i_mmap_sem);
 318 retry:
 319                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 320                                                EXT4_DATA_TRANS_BLOCKS(sb));
 321                 if (IS_ERR(handle)) {
 322                         up_read(&EXT4_I(inode)->i_mmap_sem);
 323                         sb_end_pagefault(sb);
 324                         return VM_FAULT_SIGBUS;
 325                 }
 326         } else {
 327                 down_read(&EXT4_I(inode)->i_mmap_sem);
 328         }
 329         result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
 330         if (write) {
 331                 ext4_journal_stop(handle);
 332
 333                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 334                     ext4_should_retry_alloc(sb, &retries))
 335                         goto retry;
 336                 /* Handling synchronous page fault? */
 337                 if (result & VM_FAULT_NEEDDSYNC)
 338                         result = dax_finish_sync_fault(vmf, pe_size, pfn);
 339                 up_read(&EXT4_I(inode)->i_mmap_sem);
 340                 sb_end_pagefault(sb);
 341         } else {
 342                 up_read(&EXT4_I(inode)->i_mmap_sem);
 343         }
 344
 345         return result;
 346 }
 347
 348 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 349 {
 350         return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
 351 }
 352
 353 static const struct vm_operations_struct ext4_dax_vm_ops = {
 354         .fault          = ext4_dax_fault,
 355         .huge_fault     = ext4_dax_huge_fault,
 356         .page_mkwrite   = ext4_dax_fault,
 357         .pfn_mkwrite    = ext4_dax_fault,
 358 };
 359 #else
 360 #define ext4_dax_vm_ops ext4_file_vm_ops
 361 #endif
 362
 363 static const struct vm_operations_struct ext4_file_vm_ops = {
 364         .fault          = ext4_filemap_fault,
 365         .map_pages      = filemap_map_pages,
 366         .page_mkwrite   = ext4_page_mkwrite,
 367 };
 368
 369 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 370 {
 371         struct inode *inode = file->f_mapping->host;
 372         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 373         struct dax_device *dax_dev = sbi->s_daxdev;
 374
 375         if (unlikely(ext4_forced_shutdown(sbi)))
 376                 return -EIO;
 377
 378         /*
 379          * We don't support synchronous mappings for non-DAX files and
 380          * for DAX files if underneath dax_device is not synchronous.
 381          */
 382         if (!daxdev_mapping_supported(vma, dax_dev))
 383                 return -EOPNOTSUPP;
 384
 385         file_accessed(file);
 386         if (IS_DAX(file_inode(file))) {
 387                 vma->vm_ops = &ext4_dax_vm_ops;
 388                 vma->vm_flags |= VM_HUGEPAGE;
 389         } else {
 390                 vma->vm_ops = &ext4_file_vm_ops;
 391         }
 392         return 0;
 393 }
 394
 395 static int ext4_sample_last_mounted(struct super_block *sb,
 396                                     struct vfsmount *mnt)
 397 {
 398         struct ext4_sb_info *sbi = EXT4_SB(sb);
 399         struct path path;
 400         char buf[64], *cp;
 401         handle_t *handle;
 402         int err;
 403
 404         if (likely(sbi->s_mount_flags & EXT4_MF_MNTDIR_SAMPLED))
 405                 return 0;
 406
 407         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 408                 return 0;
 409
 410         sbi->s_mount_flags |= EXT4_MF_MNTDIR_SAMPLED;
 411         /*
 412          * Sample where the filesystem has been mounted and
 413          * store it in the superblock for sysadmin convenience
 414          * when trying to sort through large numbers of block
 415          * devices or filesystem images.
 416          */
 417         memset(buf, 0, sizeof(buf));
 418         path.mnt = mnt;
 419         path.dentry = mnt->mnt_root;
 420         cp = d_path(&path, buf, sizeof(buf));
 421         err = 0;
 422         if (IS_ERR(cp))
 423                 goto out;
 424
 425         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 426         err = PTR_ERR(handle);
 427         if (IS_ERR(handle))
 428                 goto out;
 429         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 430         err = ext4_journal_get_write_access(handle, sbi->s_sbh);
 431         if (err)
 432                 goto out_journal;
 433         strlcpy(sbi->s_es->s_last_mounted, cp,
 434                 sizeof(sbi->s_es->s_last_mounted));
 435         ext4_handle_dirty_super(handle, sb);
 436 out_journal:
 437         ext4_journal_stop(handle);
 438 out:
 439         sb_end_intwrite(sb);
 440         return err;
 441 }
 442
 443 static int ext4_file_open(struct inode * inode, struct file * filp)
 444 {
 445         int ret;
 446
 447         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 448                 return -EIO;
 449
 450         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 451         if (ret)
 452                 return ret;
 453
 454         ret = fscrypt_file_open(inode, filp);
 455         if (ret)
 456                 return ret;
 457
 458         ret = fsverity_file_open(inode, filp);
 459         if (ret)
 460                 return ret;
 461
 462         /*
 463          * Set up the jbd2_inode if we are opening the inode for
 464          * writing and the journal is present
 465          */
 466         if (filp->f_mode & FMODE_WRITE) {
 467                 ret = ext4_inode_attach_jinode(inode);
 468                 if (ret < 0)
 469                         return ret;
 470         }
 471
 472         filp->f_mode |= FMODE_NOWAIT;
 473         return dquot_file_open(inode, filp);
 474 }
 475
 476 /*
 477  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 478  * by calling generic_file_llseek_size() with the appropriate maxbytes
 479  * value for each.
 480  */
 481 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 482 {
 483         struct inode *inode = file->f_mapping->host;
 484         loff_t maxbytes;
 485
 486         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 487                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 488         else
 489                 maxbytes = inode->i_sb->s_maxbytes;
 490
 491         switch (whence) {
 492         default:
 493                 return generic_file_llseek_size(file, offset, whence,
 494                                                 maxbytes, i_size_read(inode));
 495         case SEEK_HOLE:
 496                 inode_lock_shared(inode);
 497                 offset = iomap_seek_hole(inode, offset, &ext4_iomap_ops);
 498                 inode_unlock_shared(inode);
 499                 break;
 500         case SEEK_DATA:
 501                 inode_lock_shared(inode);
 502                 offset = iomap_seek_data(inode, offset, &ext4_iomap_ops);
 503                 inode_unlock_shared(inode);
 504                 break;
 505         }
 506
 507         if (offset < 0)
 508                 return offset;
 509         return vfs_setpos(file, offset, maxbytes);
 510 }
 511
 512 const struct file_operations ext4_file_operations = {
 513         .llseek         = ext4_llseek,
 514         .read_iter      = ext4_file_read_iter,
 515         .write_iter     = ext4_file_write_iter,
 516         .unlocked_ioctl = ext4_ioctl,
 517 #ifdef CONFIG_COMPAT
 518         .compat_ioctl   = ext4_compat_ioctl,
 519 #endif
 520         .mmap           = ext4_file_mmap,
 521         .mmap_supported_flags = MAP_SYNC,
 522         .open           = ext4_file_open,
 523         .release        = ext4_release_file,
 524         .fsync          = ext4_sync_file,
 525         .get_unmapped_area = thp_get_unmapped_area,
 526         .splice_read    = generic_file_splice_read,
 527         .splice_write   = iter_file_splice_write,
 528         .fallocate      = ext4_fallocate,
 529 };
 530
 531 const struct inode_operations ext4_file_inode_operations = {
 532         .setattr        = ext4_setattr,
 533         .getattr        = ext4_file_getattr,
 534         .listxattr      = ext4_listxattr,
 535         .get_acl        = ext4_get_acl,
 536         .set_acl        = ext4_set_acl,
 537         .fiemap         = ext4_fiemap,
 538 };
 539