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 "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 we have encountered a bitmap-format file, the size limit
 170          * is smaller than s_maxbytes, which is for extent-mapped files.
 171          */
 172         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 173                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 174
 175                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
 176                         return -EFBIG;
 177                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 178         }
 179         return iov_iter_count(from);
 180 }
 181
 182 #ifdef CONFIG_FS_DAX
 183 static ssize_t
 184 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 185 {
 186         struct inode *inode = file_inode(iocb->ki_filp);
 187         ssize_t ret;
 188
 189         if (!inode_trylock(inode)) {
 190                 if (iocb->ki_flags & IOCB_NOWAIT)
 191                         return -EAGAIN;
 192                 inode_lock(inode);
 193         }
 194         ret = ext4_write_checks(iocb, from);
 195         if (ret <= 0)
 196                 goto out;
 197         ret = file_remove_privs(iocb->ki_filp);
 198         if (ret)
 199                 goto out;
 200         ret = file_update_time(iocb->ki_filp);
 201         if (ret)
 202                 goto out;
 203
 204         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 205 out:
 206         inode_unlock(inode);
 207         if (ret > 0)
 208                 ret = generic_write_sync(iocb, ret);
 209         return ret;
 210 }
 211 #endif
 212
 213 static ssize_t
 214 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 215 {
 216         struct inode *inode = file_inode(iocb->ki_filp);
 217         int o_direct = iocb->ki_flags & IOCB_DIRECT;
 218         int unaligned_aio = 0;
 219         int overwrite = 0;
 220         ssize_t ret;
 221
 222         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 223                 return -EIO;
 224
 225 #ifdef CONFIG_FS_DAX
 226         if (IS_DAX(inode))
 227                 return ext4_dax_write_iter(iocb, from);
 228 #endif
 229         if (!o_direct && (iocb->ki_flags & IOCB_NOWAIT))
 230                 return -EOPNOTSUPP;
 231
 232         if (!inode_trylock(inode)) {
 233                 if (iocb->ki_flags & IOCB_NOWAIT)
 234                         return -EAGAIN;
 235                 inode_lock(inode);
 236         }
 237
 238         ret = ext4_write_checks(iocb, from);
 239         if (ret <= 0)
 240                 goto out;
 241
 242         /*
 243          * Unaligned direct AIO must be serialized among each other as zeroing
 244          * of partial blocks of two competing unaligned AIOs can result in data
 245          * corruption.
 246          */
 247         if (o_direct && ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) &&
 248             !is_sync_kiocb(iocb) &&
 249             ext4_unaligned_aio(inode, from, iocb->ki_pos)) {
 250                 unaligned_aio = 1;
 251                 ext4_unwritten_wait(inode);
 252         }
 253
 254         iocb->private = &overwrite;
 255         /* Check whether we do a DIO overwrite or not */
 256         if (o_direct && !unaligned_aio) {
 257                 if (ext4_overwrite_io(inode, iocb->ki_pos, iov_iter_count(from))) {
 258                         if (ext4_should_dioread_nolock(inode))
 259                                 overwrite = 1;
 260                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
 261                         ret = -EAGAIN;
 262                         goto out;
 263                 }
 264         }
 265
 266         ret = __generic_file_write_iter(iocb, from);
 267         /*
 268          * Unaligned direct AIO must be the only IO in flight. Otherwise
 269          * overlapping aligned IO after unaligned might result in data
 270          * corruption.
 271          */
 272         if (ret == -EIOCBQUEUED && unaligned_aio)
 273                 ext4_unwritten_wait(inode);
 274         inode_unlock(inode);
 275
 276         if (ret > 0)
 277                 ret = generic_write_sync(iocb, ret);
 278
 279         return ret;
 280
 281 out:
 282         inode_unlock(inode);
 283         return ret;
 284 }
 285
 286 #ifdef CONFIG_FS_DAX
 287 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
 288                 enum page_entry_size pe_size)
 289 {
 290         int error = 0;
 291         vm_fault_t result;
 292         int retries = 0;
 293         handle_t *handle = NULL;
 294         struct inode *inode = file_inode(vmf->vma->vm_file);
 295         struct super_block *sb = inode->i_sb;
 296
 297         /*
 298          * We have to distinguish real writes from writes which will result in a
 299          * COW page; COW writes should *not* poke the journal (the file will not
 300          * be changed). Doing so would cause unintended failures when mounted
 301          * read-only.
 302          *
 303          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 304          * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
 305          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 306          * we eventually come back with a COW page.
 307          */
 308         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 309                 (vmf->vma->vm_flags & VM_SHARED);
 310         pfn_t pfn;
 311
 312         if (write) {
 313                 sb_start_pagefault(sb);
 314                 file_update_time(vmf->vma->vm_file);
 315                 down_read(&EXT4_I(inode)->i_mmap_sem);
 316 retry:
 317                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 318                                                EXT4_DATA_TRANS_BLOCKS(sb));
 319                 if (IS_ERR(handle)) {
 320                         up_read(&EXT4_I(inode)->i_mmap_sem);
 321                         sb_end_pagefault(sb);
 322                         return VM_FAULT_SIGBUS;
 323                 }
 324         } else {
 325                 down_read(&EXT4_I(inode)->i_mmap_sem);
 326         }
 327         result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
 328         if (write) {
 329                 ext4_journal_stop(handle);
 330
 331                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 332                     ext4_should_retry_alloc(sb, &retries))
 333                         goto retry;
 334                 /* Handling synchronous page fault? */
 335                 if (result & VM_FAULT_NEEDDSYNC)
 336                         result = dax_finish_sync_fault(vmf, pe_size, pfn);
 337                 up_read(&EXT4_I(inode)->i_mmap_sem);
 338                 sb_end_pagefault(sb);
 339         } else {
 340                 up_read(&EXT4_I(inode)->i_mmap_sem);
 341         }
 342
 343         return result;
 344 }
 345
 346 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 347 {
 348         return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
 349 }
 350
 351 static const struct vm_operations_struct ext4_dax_vm_ops = {
 352         .fault          = ext4_dax_fault,
 353         .huge_fault     = ext4_dax_huge_fault,
 354         .page_mkwrite   = ext4_dax_fault,
 355         .pfn_mkwrite    = ext4_dax_fault,
 356 };
 357 #else
 358 #define ext4_dax_vm_ops ext4_file_vm_ops
 359 #endif
 360
 361 static const struct vm_operations_struct ext4_file_vm_ops = {
 362         .fault          = ext4_filemap_fault,
 363         .map_pages      = filemap_map_pages,
 364         .page_mkwrite   = ext4_page_mkwrite,
 365 };
 366
 367 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 368 {
 369         struct inode *inode = file->f_mapping->host;
 370
 371         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 372                 return -EIO;
 373
 374         /*
 375          * We don't support synchronous mappings for non-DAX files. At least
 376          * until someone comes with a sensible use case.
 377          */
 378         if (!IS_DAX(file_inode(file)) && (vma->vm_flags & VM_SYNC))
 379                 return -EOPNOTSUPP;
 380
 381         file_accessed(file);
 382         if (IS_DAX(file_inode(file))) {
 383                 vma->vm_ops = &ext4_dax_vm_ops;
 384                 vma->vm_flags |= VM_HUGEPAGE;
 385         } else {
 386                 vma->vm_ops = &ext4_file_vm_ops;
 387         }
 388         return 0;
 389 }
 390
 391 static int ext4_sample_last_mounted(struct super_block *sb,
 392                                     struct vfsmount *mnt)
 393 {
 394         struct ext4_sb_info *sbi = EXT4_SB(sb);
 395         struct path path;
 396         char buf[64], *cp;
 397         handle_t *handle;
 398         int err;
 399
 400         if (likely(sbi->s_mount_flags & EXT4_MF_MNTDIR_SAMPLED))
 401                 return 0;
 402
 403         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 404                 return 0;
 405
 406         sbi->s_mount_flags |= EXT4_MF_MNTDIR_SAMPLED;
 407         /*
 408          * Sample where the filesystem has been mounted and
 409          * store it in the superblock for sysadmin convenience
 410          * when trying to sort through large numbers of block
 411          * devices or filesystem images.
 412          */
 413         memset(buf, 0, sizeof(buf));
 414         path.mnt = mnt;
 415         path.dentry = mnt->mnt_root;
 416         cp = d_path(&path, buf, sizeof(buf));
 417         err = 0;
 418         if (IS_ERR(cp))
 419                 goto out;
 420
 421         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 422         err = PTR_ERR(handle);
 423         if (IS_ERR(handle))
 424                 goto out;
 425         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 426         err = ext4_journal_get_write_access(handle, sbi->s_sbh);
 427         if (err)
 428                 goto out_journal;
 429         strlcpy(sbi->s_es->s_last_mounted, cp,
 430                 sizeof(sbi->s_es->s_last_mounted));
 431         ext4_handle_dirty_super(handle, sb);
 432 out_journal:
 433         ext4_journal_stop(handle);
 434 out:
 435         sb_end_intwrite(sb);
 436         return err;
 437 }
 438
 439 static int ext4_file_open(struct inode * inode, struct file * filp)
 440 {
 441         int ret;
 442
 443         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 444                 return -EIO;
 445
 446         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 447         if (ret)
 448                 return ret;
 449
 450         ret = fscrypt_file_open(inode, filp);
 451         if (ret)
 452                 return ret;
 453
 454         /*
 455          * Set up the jbd2_inode if we are opening the inode for
 456          * writing and the journal is present
 457          */
 458         if (filp->f_mode & FMODE_WRITE) {
 459                 ret = ext4_inode_attach_jinode(inode);
 460                 if (ret < 0)
 461                         return ret;
 462         }
 463
 464         filp->f_mode |= FMODE_NOWAIT;
 465         return dquot_file_open(inode, filp);
 466 }
 467
 468 /*
 469  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 470  * by calling generic_file_llseek_size() with the appropriate maxbytes
 471  * value for each.
 472  */
 473 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 474 {
 475         struct inode *inode = file->f_mapping->host;
 476         loff_t maxbytes;
 477
 478         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 479                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 480         else
 481                 maxbytes = inode->i_sb->s_maxbytes;
 482
 483         switch (whence) {
 484         default:
 485                 return generic_file_llseek_size(file, offset, whence,
 486                                                 maxbytes, i_size_read(inode));
 487         case SEEK_HOLE:
 488                 inode_lock_shared(inode);
 489                 offset = iomap_seek_hole(inode, offset, &ext4_iomap_ops);
 490                 inode_unlock_shared(inode);
 491                 break;
 492         case SEEK_DATA:
 493                 inode_lock_shared(inode);
 494                 offset = iomap_seek_data(inode, offset, &ext4_iomap_ops);
 495                 inode_unlock_shared(inode);
 496                 break;
 497         }
 498
 499         if (offset < 0)
 500                 return offset;
 501         return vfs_setpos(file, offset, maxbytes);
 502 }
 503
 504 const struct file_operations ext4_file_operations = {
 505         .llseek         = ext4_llseek,
 506         .read_iter      = ext4_file_read_iter,
 507         .write_iter     = ext4_file_write_iter,
 508         .unlocked_ioctl = ext4_ioctl,
 509 #ifdef CONFIG_COMPAT
 510         .compat_ioctl   = ext4_compat_ioctl,
 511 #endif
 512         .mmap           = ext4_file_mmap,
 513         .mmap_supported_flags = MAP_SYNC,
 514         .open           = ext4_file_open,
 515         .release        = ext4_release_file,
 516         .fsync          = ext4_sync_file,
 517         .get_unmapped_area = thp_get_unmapped_area,
 518         .splice_read    = generic_file_splice_read,
 519         .splice_write   = iter_file_splice_write,
 520         .fallocate      = ext4_fallocate,
 521 };
 522
 523 const struct inode_operations ext4_file_inode_operations = {
 524         .setattr        = ext4_setattr,
 525         .getattr        = ext4_file_getattr,
 526         .listxattr      = ext4_listxattr,
 527         .get_acl        = ext4_get_acl,
 528         .set_acl        = ext4_set_acl,
 529         .fiemap         = ext4_fiemap,
 530 };
 531