fs/overlayfs/file.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2017 Red Hat, Inc.
   4  */
   5
   6 #include <linux/cred.h>
   7 #include <linux/file.h>
   8 #include <linux/mount.h>
   9 #include <linux/xattr.h>
  10 #include <linux/uio.h>
  11 #include <linux/uaccess.h>
  12 #include <linux/splice.h>
  13 #include <linux/security.h>
  14 #include <linux/mm.h>
  15 #include <linux/fs.h>
  16 #include "overlayfs.h"
  17
  18 struct ovl_aio_req {
  19         struct kiocb iocb;
  20         struct kiocb *orig_iocb;
  21         struct fd fd;
  22 };
  23
  24 static struct kmem_cache *ovl_aio_request_cachep;
  25
  26 static char ovl_whatisit(struct inode *inode, struct inode *realinode)
  27 {
  28         if (realinode != ovl_inode_upper(inode))
  29                 return 'l';
  30         if (ovl_has_upperdata(inode))
  31                 return 'u';
  32         else
  33                 return 'm';
  34 }
  35
  36 /* No atime modificaton nor notify on underlying */
  37 #define OVL_OPEN_FLAGS (O_NOATIME | FMODE_NONOTIFY)
  38
  39 static struct file *ovl_open_realfile(const struct file *file,
  40                                       struct inode *realinode)
  41 {
  42         struct inode *inode = file_inode(file);
  43         struct file *realfile;
  44         const struct cred *old_cred;
  45         int flags = file->f_flags | OVL_OPEN_FLAGS;
  46         int acc_mode = ACC_MODE(flags);
  47         int err;
  48
  49         if (flags & O_APPEND)
  50                 acc_mode |= MAY_APPEND;
  51
  52         old_cred = ovl_override_creds(inode->i_sb);
  53         err = inode_permission(realinode, MAY_OPEN | acc_mode);
  54         if (err) {
  55                 realfile = ERR_PTR(err);
  56         } else if (!inode_owner_or_capable(realinode)) {
  57                 realfile = ERR_PTR(-EPERM);
  58         } else {
  59                 realfile = open_with_fake_path(&file->f_path, flags, realinode,
  60                                                current_cred());
  61         }
  62         revert_creds(old_cred);
  63
  64         pr_debug("open(%p[%pD2/%c], 0%o) -> (%p, 0%o)\n",
  65                  file, file, ovl_whatisit(inode, realinode), file->f_flags,
  66                  realfile, IS_ERR(realfile) ? 0 : realfile->f_flags);
  67
  68         return realfile;
  69 }
  70
  71 #define OVL_SETFL_MASK (O_APPEND | O_NONBLOCK | O_NDELAY | O_DIRECT)
  72
  73 static int ovl_change_flags(struct file *file, unsigned int flags)
  74 {
  75         struct inode *inode = file_inode(file);
  76         int err;
  77
  78         flags |= OVL_OPEN_FLAGS;
  79
  80         /* If some flag changed that cannot be changed then something's amiss */
  81         if (WARN_ON((file->f_flags ^ flags) & ~OVL_SETFL_MASK))
  82                 return -EIO;
  83
  84         flags &= OVL_SETFL_MASK;
  85
  86         if (((flags ^ file->f_flags) & O_APPEND) && IS_APPEND(inode))
  87                 return -EPERM;
  88
  89         if (flags & O_DIRECT) {
  90                 if (!file->f_mapping->a_ops ||
  91                     !file->f_mapping->a_ops->direct_IO)
  92                         return -EINVAL;
  93         }
  94
  95         if (file->f_op->check_flags) {
  96                 err = file->f_op->check_flags(flags);
  97                 if (err)
  98                         return err;
  99         }
 100
 101         spin_lock(&file->f_lock);
 102         file->f_flags = (file->f_flags & ~OVL_SETFL_MASK) | flags;
 103         spin_unlock(&file->f_lock);
 104
 105         return 0;
 106 }
 107
 108 static int ovl_real_fdget_meta(const struct file *file, struct fd *real,
 109                                bool allow_meta)
 110 {
 111         struct inode *inode = file_inode(file);
 112         struct inode *realinode;
 113
 114         real->flags = 0;
 115         real->file = file->private_data;
 116
 117         if (allow_meta)
 118                 realinode = ovl_inode_real(inode);
 119         else
 120                 realinode = ovl_inode_realdata(inode);
 121
 122         /* Has it been copied up since we'd opened it? */
 123         if (unlikely(file_inode(real->file) != realinode)) {
 124                 real->flags = FDPUT_FPUT;
 125                 real->file = ovl_open_realfile(file, realinode);
 126
 127                 return PTR_ERR_OR_ZERO(real->file);
 128         }
 129
 130         /* Did the flags change since open? */
 131         if (unlikely((file->f_flags ^ real->file->f_flags) & ~OVL_OPEN_FLAGS))
 132                 return ovl_change_flags(real->file, file->f_flags);
 133
 134         return 0;
 135 }
 136
 137 static int ovl_real_fdget(const struct file *file, struct fd *real)
 138 {
 139         if (d_is_dir(file_dentry(file))) {
 140                 real->flags = 0;
 141                 real->file = ovl_dir_real_file(file, false);
 142
 143                 return PTR_ERR_OR_ZERO(real->file);
 144         }
 145
 146         return ovl_real_fdget_meta(file, real, false);
 147 }
 148
 149 static int ovl_open(struct inode *inode, struct file *file)
 150 {
 151         struct file *realfile;
 152         int err;
 153
 154         err = ovl_maybe_copy_up(file_dentry(file), file->f_flags);
 155         if (err)
 156                 return err;
 157
 158         /* No longer need these flags, so don't pass them on to underlying fs */
 159         file->f_flags &= ~(O_CREAT | O_EXCL | O_NOCTTY | O_TRUNC);
 160
 161         realfile = ovl_open_realfile(file, ovl_inode_realdata(inode));
 162         if (IS_ERR(realfile))
 163                 return PTR_ERR(realfile);
 164
 165         file->private_data = realfile;
 166
 167         return 0;
 168 }
 169
 170 static int ovl_release(struct inode *inode, struct file *file)
 171 {
 172         fput(file->private_data);
 173
 174         return 0;
 175 }
 176
 177 static loff_t ovl_llseek(struct file *file, loff_t offset, int whence)
 178 {
 179         struct inode *inode = file_inode(file);
 180         struct fd real;
 181         const struct cred *old_cred;
 182         loff_t ret;
 183
 184         /*
 185          * The two special cases below do not need to involve real fs,
 186          * so we can optimizing concurrent callers.
 187          */
 188         if (offset == 0) {
 189                 if (whence == SEEK_CUR)
 190                         return file->f_pos;
 191
 192                 if (whence == SEEK_SET)
 193                         return vfs_setpos(file, 0, 0);
 194         }
 195
 196         ret = ovl_real_fdget(file, &real);
 197         if (ret)
 198                 return ret;
 199
 200         /*
 201          * Overlay file f_pos is the master copy that is preserved
 202          * through copy up and modified on read/write, but only real
 203          * fs knows how to SEEK_HOLE/SEEK_DATA and real fs may impose
 204          * limitations that are more strict than ->s_maxbytes for specific
 205          * files, so we use the real file to perform seeks.
 206          */
 207         ovl_inode_lock(inode);
 208         real.file->f_pos = file->f_pos;
 209
 210         old_cred = ovl_override_creds(inode->i_sb);
 211         ret = vfs_llseek(real.file, offset, whence);
 212         revert_creds(old_cred);
 213
 214         file->f_pos = real.file->f_pos;
 215         ovl_inode_unlock(inode);
 216
 217         fdput(real);
 218
 219         return ret;
 220 }
 221
 222 static void ovl_file_accessed(struct file *file)
 223 {
 224         struct inode *inode, *upperinode;
 225
 226         if (file->f_flags & O_NOATIME)
 227                 return;
 228
 229         inode = file_inode(file);
 230         upperinode = ovl_inode_upper(inode);
 231
 232         if (!upperinode)
 233                 return;
 234
 235         if ((!timespec64_equal(&inode->i_mtime, &upperinode->i_mtime) ||
 236              !timespec64_equal(&inode->i_ctime, &upperinode->i_ctime))) {
 237                 inode->i_mtime = upperinode->i_mtime;
 238                 inode->i_ctime = upperinode->i_ctime;
 239         }
 240
 241         touch_atime(&file->f_path);
 242 }
 243
 244 static rwf_t ovl_iocb_to_rwf(int ifl)
 245 {
 246         rwf_t flags = 0;
 247
 248         if (ifl & IOCB_NOWAIT)
 249                 flags |= RWF_NOWAIT;
 250         if (ifl & IOCB_HIPRI)
 251                 flags |= RWF_HIPRI;
 252         if (ifl & IOCB_DSYNC)
 253                 flags |= RWF_DSYNC;
 254         if (ifl & IOCB_SYNC)
 255                 flags |= RWF_SYNC;
 256
 257         return flags;
 258 }
 259
 260 static void ovl_aio_cleanup_handler(struct ovl_aio_req *aio_req)
 261 {
 262         struct kiocb *iocb = &aio_req->iocb;
 263         struct kiocb *orig_iocb = aio_req->orig_iocb;
 264
 265         if (iocb->ki_flags & IOCB_WRITE) {
 266                 struct inode *inode = file_inode(orig_iocb->ki_filp);
 267
 268                 /* Actually acquired in ovl_write_iter() */
 269                 __sb_writers_acquired(file_inode(iocb->ki_filp)->i_sb,
 270                                       SB_FREEZE_WRITE);
 271                 file_end_write(iocb->ki_filp);
 272                 ovl_copyattr(ovl_inode_real(inode), inode);
 273         }
 274
 275         orig_iocb->ki_pos = iocb->ki_pos;
 276         fdput(aio_req->fd);
 277         kmem_cache_free(ovl_aio_request_cachep, aio_req);
 278 }
 279
 280 static void ovl_aio_rw_complete(struct kiocb *iocb, long res, long res2)
 281 {
 282         struct ovl_aio_req *aio_req = container_of(iocb,
 283                                                    struct ovl_aio_req, iocb);
 284         struct kiocb *orig_iocb = aio_req->orig_iocb;
 285
 286         ovl_aio_cleanup_handler(aio_req);
 287         orig_iocb->ki_complete(orig_iocb, res, res2);
 288 }
 289
 290 static ssize_t ovl_read_iter(struct kiocb *iocb, struct iov_iter *iter)
 291 {
 292         struct file *file = iocb->ki_filp;
 293         struct fd real;
 294         const struct cred *old_cred;
 295         ssize_t ret;
 296
 297         if (!iov_iter_count(iter))
 298                 return 0;
 299
 300         ret = ovl_real_fdget(file, &real);
 301         if (ret)
 302                 return ret;
 303
 304         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 305         if (is_sync_kiocb(iocb)) {
 306                 ret = vfs_iter_read(real.file, iter, &iocb->ki_pos,
 307                                     ovl_iocb_to_rwf(iocb->ki_flags));
 308         } else {
 309                 struct ovl_aio_req *aio_req;
 310
 311                 ret = -ENOMEM;
 312                 aio_req = kmem_cache_zalloc(ovl_aio_request_cachep, GFP_KERNEL);
 313                 if (!aio_req)
 314                         goto out;
 315
 316                 aio_req->fd = real;
 317                 real.flags = 0;
 318                 aio_req->orig_iocb = iocb;
 319                 kiocb_clone(&aio_req->iocb, iocb, real.file);
 320                 aio_req->iocb.ki_complete = ovl_aio_rw_complete;
 321                 ret = vfs_iocb_iter_read(real.file, &aio_req->iocb, iter);
 322                 if (ret != -EIOCBQUEUED)
 323                         ovl_aio_cleanup_handler(aio_req);
 324         }
 325 out:
 326         revert_creds(old_cred);
 327         ovl_file_accessed(file);
 328
 329         fdput(real);
 330
 331         return ret;
 332 }
 333
 334 static ssize_t ovl_write_iter(struct kiocb *iocb, struct iov_iter *iter)
 335 {
 336         struct file *file = iocb->ki_filp;
 337         struct inode *inode = file_inode(file);
 338         struct fd real;
 339         const struct cred *old_cred;
 340         ssize_t ret;
 341         int ifl = iocb->ki_flags;
 342
 343         if (!iov_iter_count(iter))
 344                 return 0;
 345
 346         inode_lock(inode);
 347         /* Update mode */
 348         ovl_copyattr(ovl_inode_real(inode), inode);
 349         ret = file_remove_privs(file);
 350         if (ret)
 351                 goto out_unlock;
 352
 353         ret = ovl_real_fdget(file, &real);
 354         if (ret)
 355                 goto out_unlock;
 356
 357         if (!ovl_should_sync(OVL_FS(inode->i_sb)))
 358                 ifl &= ~(IOCB_DSYNC | IOCB_SYNC);
 359
 360         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 361         if (is_sync_kiocb(iocb)) {
 362                 file_start_write(real.file);
 363                 ret = vfs_iter_write(real.file, iter, &iocb->ki_pos,
 364                                      ovl_iocb_to_rwf(ifl));
 365                 file_end_write(real.file);
 366                 /* Update size */
 367                 ovl_copyattr(ovl_inode_real(inode), inode);
 368         } else {
 369                 struct ovl_aio_req *aio_req;
 370
 371                 ret = -ENOMEM;
 372                 aio_req = kmem_cache_zalloc(ovl_aio_request_cachep, GFP_KERNEL);
 373                 if (!aio_req)
 374                         goto out;
 375
 376                 file_start_write(real.file);
 377                 /* Pacify lockdep, same trick as done in aio_write() */
 378                 __sb_writers_release(file_inode(real.file)->i_sb,
 379                                      SB_FREEZE_WRITE);
 380                 aio_req->fd = real;
 381                 real.flags = 0;
 382                 aio_req->orig_iocb = iocb;
 383                 kiocb_clone(&aio_req->iocb, iocb, real.file);
 384                 aio_req->iocb.ki_flags = ifl;
 385                 aio_req->iocb.ki_complete = ovl_aio_rw_complete;
 386                 ret = vfs_iocb_iter_write(real.file, &aio_req->iocb, iter);
 387                 if (ret != -EIOCBQUEUED)
 388                         ovl_aio_cleanup_handler(aio_req);
 389         }
 390 out:
 391         revert_creds(old_cred);
 392         fdput(real);
 393
 394 out_unlock:
 395         inode_unlock(inode);
 396
 397         return ret;
 398 }
 399
 400 static ssize_t ovl_splice_read(struct file *in, loff_t *ppos,
 401                          struct pipe_inode_info *pipe, size_t len,
 402                          unsigned int flags)
 403 {
 404         ssize_t ret;
 405         struct fd real;
 406         const struct cred *old_cred;
 407
 408         ret = ovl_real_fdget(in, &real);
 409         if (ret)
 410                 return ret;
 411
 412         old_cred = ovl_override_creds(file_inode(in)->i_sb);
 413         ret = generic_file_splice_read(real.file, ppos, pipe, len, flags);
 414         revert_creds(old_cred);
 415
 416         ovl_file_accessed(in);
 417         fdput(real);
 418         return ret;
 419 }
 420
 421 static ssize_t
 422 ovl_splice_write(struct pipe_inode_info *pipe, struct file *out,
 423                           loff_t *ppos, size_t len, unsigned int flags)
 424 {
 425         struct fd real;
 426         const struct cred *old_cred;
 427         ssize_t ret;
 428
 429         ret = ovl_real_fdget(out, &real);
 430         if (ret)
 431                 return ret;
 432
 433         old_cred = ovl_override_creds(file_inode(out)->i_sb);
 434         ret = iter_file_splice_write(pipe, real.file, ppos, len, flags);
 435         revert_creds(old_cred);
 436
 437         ovl_file_accessed(out);
 438         fdput(real);
 439         return ret;
 440 }
 441
 442 static int ovl_fsync(struct file *file, loff_t start, loff_t end, int datasync)
 443 {
 444         struct fd real;
 445         const struct cred *old_cred;
 446         int ret;
 447
 448         if (!ovl_should_sync(OVL_FS(file_inode(file)->i_sb)))
 449                 return 0;
 450
 451         ret = ovl_real_fdget_meta(file, &real, !datasync);
 452         if (ret)
 453                 return ret;
 454
 455         /* Don't sync lower file for fear of receiving EROFS error */
 456         if (file_inode(real.file) == ovl_inode_upper(file_inode(file))) {
 457                 old_cred = ovl_override_creds(file_inode(file)->i_sb);
 458                 ret = vfs_fsync_range(real.file, start, end, datasync);
 459                 revert_creds(old_cred);
 460         }
 461
 462         fdput(real);
 463
 464         return ret;
 465 }
 466
 467 static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
 468 {
 469         struct file *realfile = file->private_data;
 470         const struct cred *old_cred;
 471         int ret;
 472
 473         if (!realfile->f_op->mmap)
 474                 return -ENODEV;
 475
 476         if (WARN_ON(file != vma->vm_file))
 477                 return -EIO;
 478
 479         vma->vm_file = get_file(realfile);
 480
 481         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 482         ret = call_mmap(vma->vm_file, vma);
 483         revert_creds(old_cred);
 484
 485         if (ret) {
 486                 /* Drop reference count from new vm_file value */
 487                 fput(realfile);
 488         } else {
 489                 /* Drop reference count from previous vm_file value */
 490                 fput(file);
 491         }
 492
 493         ovl_file_accessed(file);
 494
 495         return ret;
 496 }
 497
 498 static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
 499 {
 500         struct inode *inode = file_inode(file);
 501         struct fd real;
 502         const struct cred *old_cred;
 503         int ret;
 504
 505         ret = ovl_real_fdget(file, &real);
 506         if (ret)
 507                 return ret;
 508
 509         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 510         ret = vfs_fallocate(real.file, mode, offset, len);
 511         revert_creds(old_cred);
 512
 513         /* Update size */
 514         ovl_copyattr(ovl_inode_real(inode), inode);
 515
 516         fdput(real);
 517
 518         return ret;
 519 }
 520
 521 static int ovl_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
 522 {
 523         struct fd real;
 524         const struct cred *old_cred;
 525         int ret;
 526
 527         ret = ovl_real_fdget(file, &real);
 528         if (ret)
 529                 return ret;
 530
 531         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 532         ret = vfs_fadvise(real.file, offset, len, advice);
 533         revert_creds(old_cred);
 534
 535         fdput(real);
 536
 537         return ret;
 538 }
 539
 540 static long ovl_real_ioctl(struct file *file, unsigned int cmd,
 541                            unsigned long arg)
 542 {
 543         struct fd real;
 544         const struct cred *old_cred;
 545         long ret;
 546
 547         ret = ovl_real_fdget(file, &real);
 548         if (ret)
 549                 return ret;
 550
 551         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 552         ret = security_file_ioctl(real.file, cmd, arg);
 553         if (!ret)
 554                 ret = vfs_ioctl(real.file, cmd, arg);
 555         revert_creds(old_cred);
 556
 557         fdput(real);
 558
 559         return ret;
 560 }
 561
 562 static long ovl_ioctl_set_flags(struct file *file, unsigned int cmd,
 563                                 unsigned long arg, unsigned int iflags)
 564 {
 565         long ret;
 566         struct inode *inode = file_inode(file);
 567         unsigned int old_iflags;
 568
 569         if (!inode_owner_or_capable(inode))
 570                 return -EACCES;
 571
 572         ret = mnt_want_write_file(file);
 573         if (ret)
 574                 return ret;
 575
 576         inode_lock(inode);
 577
 578         /* Check the capability before cred override */
 579         ret = -EPERM;
 580         old_iflags = READ_ONCE(inode->i_flags);
 581         if (((iflags ^ old_iflags) & (S_APPEND | S_IMMUTABLE)) &&
 582             !capable(CAP_LINUX_IMMUTABLE))
 583                 goto unlock;
 584
 585         ret = ovl_maybe_copy_up(file_dentry(file), O_WRONLY);
 586         if (ret)
 587                 goto unlock;
 588
 589         ret = ovl_real_ioctl(file, cmd, arg);
 590
 591         ovl_copyflags(ovl_inode_real(inode), inode);
 592 unlock:
 593         inode_unlock(inode);
 594
 595         mnt_drop_write_file(file);
 596
 597         return ret;
 598
 599 }
 600
 601 static unsigned int ovl_fsflags_to_iflags(unsigned int flags)
 602 {
 603         unsigned int iflags = 0;
 604
 605         if (flags & FS_SYNC_FL)
 606                 iflags |= S_SYNC;
 607         if (flags & FS_APPEND_FL)
 608                 iflags |= S_APPEND;
 609         if (flags & FS_IMMUTABLE_FL)
 610                 iflags |= S_IMMUTABLE;
 611         if (flags & FS_NOATIME_FL)
 612                 iflags |= S_NOATIME;
 613
 614         return iflags;
 615 }
 616
 617 static long ovl_ioctl_set_fsflags(struct file *file, unsigned int cmd,
 618                                   unsigned long arg)
 619 {
 620         unsigned int flags;
 621
 622         if (get_user(flags, (int __user *) arg))
 623                 return -EFAULT;
 624
 625         return ovl_ioctl_set_flags(file, cmd, arg,
 626                                    ovl_fsflags_to_iflags(flags));
 627 }
 628
 629 static unsigned int ovl_fsxflags_to_iflags(unsigned int xflags)
 630 {
 631         unsigned int iflags = 0;
 632
 633         if (xflags & FS_XFLAG_SYNC)
 634                 iflags |= S_SYNC;
 635         if (xflags & FS_XFLAG_APPEND)
 636                 iflags |= S_APPEND;
 637         if (xflags & FS_XFLAG_IMMUTABLE)
 638                 iflags |= S_IMMUTABLE;
 639         if (xflags & FS_XFLAG_NOATIME)
 640                 iflags |= S_NOATIME;
 641
 642         return iflags;
 643 }
 644
 645 static long ovl_ioctl_set_fsxflags(struct file *file, unsigned int cmd,
 646                                    unsigned long arg)
 647 {
 648         struct fsxattr fa;
 649
 650         memset(&fa, 0, sizeof(fa));
 651         if (copy_from_user(&fa, (void __user *) arg, sizeof(fa)))
 652                 return -EFAULT;
 653
 654         return ovl_ioctl_set_flags(file, cmd, arg,
 655                                    ovl_fsxflags_to_iflags(fa.fsx_xflags));
 656 }
 657
 658 long ovl_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 659 {
 660         long ret;
 661
 662         switch (cmd) {
 663         case FS_IOC_GETFLAGS:
 664         case FS_IOC_FSGETXATTR:
 665                 ret = ovl_real_ioctl(file, cmd, arg);
 666                 break;
 667
 668         case FS_IOC_SETFLAGS:
 669                 ret = ovl_ioctl_set_fsflags(file, cmd, arg);
 670                 break;
 671
 672         case FS_IOC_FSSETXATTR:
 673                 ret = ovl_ioctl_set_fsxflags(file, cmd, arg);
 674                 break;
 675
 676         default:
 677                 ret = -ENOTTY;
 678         }
 679
 680         return ret;
 681 }
 682
 683 #ifdef CONFIG_COMPAT
 684 long ovl_compat_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 685 {
 686         switch (cmd) {
 687         case FS_IOC32_GETFLAGS:
 688                 cmd = FS_IOC_GETFLAGS;
 689                 break;
 690
 691         case FS_IOC32_SETFLAGS:
 692                 cmd = FS_IOC_SETFLAGS;
 693                 break;
 694
 695         default:
 696                 return -ENOIOCTLCMD;
 697         }
 698
 699         return ovl_ioctl(file, cmd, arg);
 700 }
 701 #endif
 702
 703 enum ovl_copyop {
 704         OVL_COPY,
 705         OVL_CLONE,
 706         OVL_DEDUPE,
 707 };
 708
 709 static loff_t ovl_copyfile(struct file *file_in, loff_t pos_in,
 710                             struct file *file_out, loff_t pos_out,
 711                             loff_t len, unsigned int flags, enum ovl_copyop op)
 712 {
 713         struct inode *inode_out = file_inode(file_out);
 714         struct fd real_in, real_out;
 715         const struct cred *old_cred;
 716         loff_t ret;
 717
 718         ret = ovl_real_fdget(file_out, &real_out);
 719         if (ret)
 720                 return ret;
 721
 722         ret = ovl_real_fdget(file_in, &real_in);
 723         if (ret) {
 724                 fdput(real_out);
 725                 return ret;
 726         }
 727
 728         old_cred = ovl_override_creds(file_inode(file_out)->i_sb);
 729         switch (op) {
 730         case OVL_COPY:
 731                 ret = vfs_copy_file_range(real_in.file, pos_in,
 732                                           real_out.file, pos_out, len, flags);
 733                 break;
 734
 735         case OVL_CLONE:
 736                 ret = vfs_clone_file_range(real_in.file, pos_in,
 737                                            real_out.file, pos_out, len, flags);
 738                 break;
 739
 740         case OVL_DEDUPE:
 741                 ret = vfs_dedupe_file_range_one(real_in.file, pos_in,
 742                                                 real_out.file, pos_out, len,
 743                                                 flags);
 744                 break;
 745         }
 746         revert_creds(old_cred);
 747
 748         /* Update size */
 749         ovl_copyattr(ovl_inode_real(inode_out), inode_out);
 750
 751         fdput(real_in);
 752         fdput(real_out);
 753
 754         return ret;
 755 }
 756
 757 static ssize_t ovl_copy_file_range(struct file *file_in, loff_t pos_in,
 758                                    struct file *file_out, loff_t pos_out,
 759                                    size_t len, unsigned int flags)
 760 {
 761         return ovl_copyfile(file_in, pos_in, file_out, pos_out, len, flags,
 762                             OVL_COPY);
 763 }
 764
 765 static loff_t ovl_remap_file_range(struct file *file_in, loff_t pos_in,
 766                                    struct file *file_out, loff_t pos_out,
 767                                    loff_t len, unsigned int remap_flags)
 768 {
 769         enum ovl_copyop op;
 770
 771         if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
 772                 return -EINVAL;
 773
 774         if (remap_flags & REMAP_FILE_DEDUP)
 775                 op = OVL_DEDUPE;
 776         else
 777                 op = OVL_CLONE;
 778
 779         /*
 780          * Don't copy up because of a dedupe request, this wouldn't make sense
 781          * most of the time (data would be duplicated instead of deduplicated).
 782          */
 783         if (op == OVL_DEDUPE &&
 784             (!ovl_inode_upper(file_inode(file_in)) ||
 785              !ovl_inode_upper(file_inode(file_out))))
 786                 return -EPERM;
 787
 788         return ovl_copyfile(file_in, pos_in, file_out, pos_out, len,
 789                             remap_flags, op);
 790 }
 791
 792 const struct file_operations ovl_file_operations = {
 793         .open           = ovl_open,
 794         .release        = ovl_release,
 795         .llseek         = ovl_llseek,
 796         .read_iter      = ovl_read_iter,
 797         .write_iter     = ovl_write_iter,
 798         .fsync          = ovl_fsync,
 799         .mmap           = ovl_mmap,
 800         .fallocate      = ovl_fallocate,
 801         .fadvise        = ovl_fadvise,
 802         .unlocked_ioctl = ovl_ioctl,
 803 #ifdef CONFIG_COMPAT
 804         .compat_ioctl   = ovl_compat_ioctl,
 805 #endif
 806         .splice_read    = ovl_splice_read,
 807         .splice_write   = ovl_splice_write,
 808
 809         .copy_file_range        = ovl_copy_file_range,
 810         .remap_file_range       = ovl_remap_file_range,
 811 };
 812
 813 int __init ovl_aio_request_cache_init(void)
 814 {
 815         ovl_aio_request_cachep = kmem_cache_create("ovl_aio_req",
 816                                                    sizeof(struct ovl_aio_req),
 817                                                    0, SLAB_HWCACHE_ALIGN, NULL);
 818         if (!ovl_aio_request_cachep)
 819                 return -ENOMEM;
 820
 821         return 0;
 822 }
 823
 824 void ovl_aio_request_cache_destroy(void)
 825 {
 826         kmem_cache_destroy(ovl_aio_request_cachep);
 827 }