fs/overlayfs/file.c

   1 /*
   2  * Copyright (C) 2017 Red Hat, Inc.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 as published by
   6  * the Free Software Foundation.
   7  */
   8
   9 #include <linux/cred.h>
  10 #include <linux/file.h>
  11 #include <linux/mount.h>
  12 #include <linux/xattr.h>
  13 #include <linux/uio.h>
  14 #include "overlayfs.h"
  15
  16 static char ovl_whatisit(struct inode *inode, struct inode *realinode)
  17 {
  18         if (realinode != ovl_inode_upper(inode))
  19                 return 'l';
  20         if (ovl_has_upperdata(inode))
  21                 return 'u';
  22         else
  23                 return 'm';
  24 }
  25
  26 static struct file *ovl_open_realfile(const struct file *file,
  27                                       struct inode *realinode)
  28 {
  29         struct inode *inode = file_inode(file);
  30         struct file *realfile;
  31         const struct cred *old_cred;
  32
  33         old_cred = ovl_override_creds(inode->i_sb);
  34         realfile = open_with_fake_path(&file->f_path, file->f_flags | O_NOATIME,
  35                                        realinode, current_cred());
  36         revert_creds(old_cred);
  37
  38         pr_debug("open(%p[%pD2/%c], 0%o) -> (%p, 0%o)\n",
  39                  file, file, ovl_whatisit(inode, realinode), file->f_flags,
  40                  realfile, IS_ERR(realfile) ? 0 : realfile->f_flags);
  41
  42         return realfile;
  43 }
  44
  45 #define OVL_SETFL_MASK (O_APPEND | O_NONBLOCK | O_NDELAY | O_DIRECT)
  46
  47 static int ovl_change_flags(struct file *file, unsigned int flags)
  48 {
  49         struct inode *inode = file_inode(file);
  50         int err;
  51
  52         /* No atime modificaton on underlying */
  53         flags |= O_NOATIME;
  54
  55         /* If some flag changed that cannot be changed then something's amiss */
  56         if (WARN_ON((file->f_flags ^ flags) & ~OVL_SETFL_MASK))
  57                 return -EIO;
  58
  59         flags &= OVL_SETFL_MASK;
  60
  61         if (((flags ^ file->f_flags) & O_APPEND) && IS_APPEND(inode))
  62                 return -EPERM;
  63
  64         if (flags & O_DIRECT) {
  65                 if (!file->f_mapping->a_ops ||
  66                     !file->f_mapping->a_ops->direct_IO)
  67                         return -EINVAL;
  68         }
  69
  70         if (file->f_op->check_flags) {
  71                 err = file->f_op->check_flags(flags);
  72                 if (err)
  73                         return err;
  74         }
  75
  76         spin_lock(&file->f_lock);
  77         file->f_flags = (file->f_flags & ~OVL_SETFL_MASK) | flags;
  78         spin_unlock(&file->f_lock);
  79
  80         return 0;
  81 }
  82
  83 static int ovl_real_fdget_meta(const struct file *file, struct fd *real,
  84                                bool allow_meta)
  85 {
  86         struct inode *inode = file_inode(file);
  87         struct inode *realinode;
  88
  89         real->flags = 0;
  90         real->file = file->private_data;
  91
  92         if (allow_meta)
  93                 realinode = ovl_inode_real(inode);
  94         else
  95                 realinode = ovl_inode_realdata(inode);
  96
  97         /* Has it been copied up since we'd opened it? */
  98         if (unlikely(file_inode(real->file) != realinode)) {
  99                 real->flags = FDPUT_FPUT;
 100                 real->file = ovl_open_realfile(file, realinode);
 101
 102                 return PTR_ERR_OR_ZERO(real->file);
 103         }
 104
 105         /* Did the flags change since open? */
 106         if (unlikely((file->f_flags ^ real->file->f_flags) & ~O_NOATIME))
 107                 return ovl_change_flags(real->file, file->f_flags);
 108
 109         return 0;
 110 }
 111
 112 static int ovl_real_fdget(const struct file *file, struct fd *real)
 113 {
 114         return ovl_real_fdget_meta(file, real, false);
 115 }
 116
 117 static int ovl_open(struct inode *inode, struct file *file)
 118 {
 119         struct dentry *dentry = file_dentry(file);
 120         struct file *realfile;
 121         int err;
 122
 123         err = ovl_open_maybe_copy_up(dentry, file->f_flags);
 124         if (err)
 125                 return err;
 126
 127         /* No longer need these flags, so don't pass them on to underlying fs */
 128         file->f_flags &= ~(O_CREAT | O_EXCL | O_NOCTTY | O_TRUNC);
 129
 130         realfile = ovl_open_realfile(file, ovl_inode_realdata(inode));
 131         if (IS_ERR(realfile))
 132                 return PTR_ERR(realfile);
 133
 134         file->private_data = realfile;
 135
 136         return 0;
 137 }
 138
 139 static int ovl_release(struct inode *inode, struct file *file)
 140 {
 141         fput(file->private_data);
 142
 143         return 0;
 144 }
 145
 146 static loff_t ovl_llseek(struct file *file, loff_t offset, int whence)
 147 {
 148         struct inode *realinode = ovl_inode_real(file_inode(file));
 149
 150         return generic_file_llseek_size(file, offset, whence,
 151                                         realinode->i_sb->s_maxbytes,
 152                                         i_size_read(realinode));
 153 }
 154
 155 static void ovl_file_accessed(struct file *file)
 156 {
 157         struct inode *inode, *upperinode;
 158
 159         if (file->f_flags & O_NOATIME)
 160                 return;
 161
 162         inode = file_inode(file);
 163         upperinode = ovl_inode_upper(inode);
 164
 165         if (!upperinode)
 166                 return;
 167
 168         if ((!timespec64_equal(&inode->i_mtime, &upperinode->i_mtime) ||
 169              !timespec64_equal(&inode->i_ctime, &upperinode->i_ctime))) {
 170                 inode->i_mtime = upperinode->i_mtime;
 171                 inode->i_ctime = upperinode->i_ctime;
 172         }
 173
 174         touch_atime(&file->f_path);
 175 }
 176
 177 static rwf_t ovl_iocb_to_rwf(struct kiocb *iocb)
 178 {
 179         int ifl = iocb->ki_flags;
 180         rwf_t flags = 0;
 181
 182         if (ifl & IOCB_NOWAIT)
 183                 flags |= RWF_NOWAIT;
 184         if (ifl & IOCB_HIPRI)
 185                 flags |= RWF_HIPRI;
 186         if (ifl & IOCB_DSYNC)
 187                 flags |= RWF_DSYNC;
 188         if (ifl & IOCB_SYNC)
 189                 flags |= RWF_SYNC;
 190
 191         return flags;
 192 }
 193
 194 static ssize_t ovl_read_iter(struct kiocb *iocb, struct iov_iter *iter)
 195 {
 196         struct file *file = iocb->ki_filp;
 197         struct fd real;
 198         const struct cred *old_cred;
 199         ssize_t ret;
 200
 201         if (!iov_iter_count(iter))
 202                 return 0;
 203
 204         ret = ovl_real_fdget(file, &real);
 205         if (ret)
 206                 return ret;
 207
 208         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 209         ret = vfs_iter_read(real.file, iter, &iocb->ki_pos,
 210                             ovl_iocb_to_rwf(iocb));
 211         revert_creds(old_cred);
 212
 213         ovl_file_accessed(file);
 214
 215         fdput(real);
 216
 217         return ret;
 218 }
 219
 220 static ssize_t ovl_write_iter(struct kiocb *iocb, struct iov_iter *iter)
 221 {
 222         struct file *file = iocb->ki_filp;
 223         struct inode *inode = file_inode(file);
 224         struct fd real;
 225         const struct cred *old_cred;
 226         ssize_t ret;
 227
 228         if (!iov_iter_count(iter))
 229                 return 0;
 230
 231         inode_lock(inode);
 232         /* Update mode */
 233         ovl_copyattr(ovl_inode_real(inode), inode);
 234         ret = file_remove_privs(file);
 235         if (ret)
 236                 goto out_unlock;
 237
 238         ret = ovl_real_fdget(file, &real);
 239         if (ret)
 240                 goto out_unlock;
 241
 242         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 243         ret = vfs_iter_write(real.file, iter, &iocb->ki_pos,
 244                              ovl_iocb_to_rwf(iocb));
 245         revert_creds(old_cred);
 246
 247         /* Update size */
 248         ovl_copyattr(ovl_inode_real(inode), inode);
 249
 250         fdput(real);
 251
 252 out_unlock:
 253         inode_unlock(inode);
 254
 255         return ret;
 256 }
 257
 258 static int ovl_fsync(struct file *file, loff_t start, loff_t end, int datasync)
 259 {
 260         struct fd real;
 261         const struct cred *old_cred;
 262         int ret;
 263
 264         ret = ovl_real_fdget_meta(file, &real, !datasync);
 265         if (ret)
 266                 return ret;
 267
 268         /* Don't sync lower file for fear of receiving EROFS error */
 269         if (file_inode(real.file) == ovl_inode_upper(file_inode(file))) {
 270                 old_cred = ovl_override_creds(file_inode(file)->i_sb);
 271                 ret = vfs_fsync_range(real.file, start, end, datasync);
 272                 revert_creds(old_cred);
 273         }
 274
 275         fdput(real);
 276
 277         return ret;
 278 }
 279
 280 static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
 281 {
 282         struct file *realfile = file->private_data;
 283         const struct cred *old_cred;
 284         int ret;
 285
 286         if (!realfile->f_op->mmap)
 287                 return -ENODEV;
 288
 289         if (WARN_ON(file != vma->vm_file))
 290                 return -EIO;
 291
 292         vma->vm_file = get_file(realfile);
 293
 294         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 295         ret = call_mmap(vma->vm_file, vma);
 296         revert_creds(old_cred);
 297
 298         if (ret) {
 299                 /* Drop reference count from new vm_file value */
 300                 fput(realfile);
 301         } else {
 302                 /* Drop reference count from previous vm_file value */
 303                 fput(file);
 304         }
 305
 306         ovl_file_accessed(file);
 307
 308         return ret;
 309 }
 310
 311 static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
 312 {
 313         struct inode *inode = file_inode(file);
 314         struct fd real;
 315         const struct cred *old_cred;
 316         int ret;
 317
 318         ret = ovl_real_fdget(file, &real);
 319         if (ret)
 320                 return ret;
 321
 322         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 323         ret = vfs_fallocate(real.file, mode, offset, len);
 324         revert_creds(old_cred);
 325
 326         /* Update size */
 327         ovl_copyattr(ovl_inode_real(inode), inode);
 328
 329         fdput(real);
 330
 331         return ret;
 332 }
 333
 334 static int ovl_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
 335 {
 336         struct fd real;
 337         const struct cred *old_cred;
 338         int ret;
 339
 340         ret = ovl_real_fdget(file, &real);
 341         if (ret)
 342                 return ret;
 343
 344         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 345         ret = vfs_fadvise(real.file, offset, len, advice);
 346         revert_creds(old_cred);
 347
 348         fdput(real);
 349
 350         return ret;
 351 }
 352
 353 static long ovl_real_ioctl(struct file *file, unsigned int cmd,
 354                            unsigned long arg)
 355 {
 356         struct fd real;
 357         const struct cred *old_cred;
 358         long ret;
 359
 360         ret = ovl_real_fdget(file, &real);
 361         if (ret)
 362                 return ret;
 363
 364         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 365         ret = vfs_ioctl(real.file, cmd, arg);
 366         revert_creds(old_cred);
 367
 368         fdput(real);
 369
 370         return ret;
 371 }
 372
 373 static long ovl_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 374 {
 375         long ret;
 376         struct inode *inode = file_inode(file);
 377
 378         switch (cmd) {
 379         case FS_IOC_GETFLAGS:
 380                 ret = ovl_real_ioctl(file, cmd, arg);
 381                 break;
 382
 383         case FS_IOC_SETFLAGS:
 384                 if (!inode_owner_or_capable(inode))
 385                         return -EACCES;
 386
 387                 ret = mnt_want_write_file(file);
 388                 if (ret)
 389                         return ret;
 390
 391                 ret = ovl_copy_up_with_data(file_dentry(file));
 392                 if (!ret) {
 393                         ret = ovl_real_ioctl(file, cmd, arg);
 394
 395                         inode_lock(inode);
 396                         ovl_copyflags(ovl_inode_real(inode), inode);
 397                         inode_unlock(inode);
 398                 }
 399
 400                 mnt_drop_write_file(file);
 401                 break;
 402
 403         default:
 404                 ret = -ENOTTY;
 405         }
 406
 407         return ret;
 408 }
 409
 410 static long ovl_compat_ioctl(struct file *file, unsigned int cmd,
 411                              unsigned long arg)
 412 {
 413         switch (cmd) {
 414         case FS_IOC32_GETFLAGS:
 415                 cmd = FS_IOC_GETFLAGS;
 416                 break;
 417
 418         case FS_IOC32_SETFLAGS:
 419                 cmd = FS_IOC_SETFLAGS;
 420                 break;
 421
 422         default:
 423                 return -ENOIOCTLCMD;
 424         }
 425
 426         return ovl_ioctl(file, cmd, arg);
 427 }
 428
 429 enum ovl_copyop {
 430         OVL_COPY,
 431         OVL_CLONE,
 432         OVL_DEDUPE,
 433 };
 434
 435 static ssize_t ovl_copyfile(struct file *file_in, loff_t pos_in,
 436                             struct file *file_out, loff_t pos_out,
 437                             u64 len, unsigned int flags, enum ovl_copyop op)
 438 {
 439         struct inode *inode_out = file_inode(file_out);
 440         struct fd real_in, real_out;
 441         const struct cred *old_cred;
 442         ssize_t ret;
 443
 444         ret = ovl_real_fdget(file_out, &real_out);
 445         if (ret)
 446                 return ret;
 447
 448         ret = ovl_real_fdget(file_in, &real_in);
 449         if (ret) {
 450                 fdput(real_out);
 451                 return ret;
 452         }
 453
 454         old_cred = ovl_override_creds(file_inode(file_out)->i_sb);
 455         switch (op) {
 456         case OVL_COPY:
 457                 ret = vfs_copy_file_range(real_in.file, pos_in,
 458                                           real_out.file, pos_out, len, flags);
 459                 break;
 460
 461         case OVL_CLONE:
 462                 ret = vfs_clone_file_range(real_in.file, pos_in,
 463                                            real_out.file, pos_out, len);
 464                 break;
 465
 466         case OVL_DEDUPE:
 467                 ret = vfs_dedupe_file_range_one(real_in.file, pos_in,
 468                                                 real_out.file, pos_out, len);
 469                 break;
 470         }
 471         revert_creds(old_cred);
 472
 473         /* Update size */
 474         ovl_copyattr(ovl_inode_real(inode_out), inode_out);
 475
 476         fdput(real_in);
 477         fdput(real_out);
 478
 479         return ret;
 480 }
 481
 482 static ssize_t ovl_copy_file_range(struct file *file_in, loff_t pos_in,
 483                                    struct file *file_out, loff_t pos_out,
 484                                    size_t len, unsigned int flags)
 485 {
 486         return ovl_copyfile(file_in, pos_in, file_out, pos_out, len, flags,
 487                             OVL_COPY);
 488 }
 489
 490 static int ovl_clone_file_range(struct file *file_in, loff_t pos_in,
 491                                 struct file *file_out, loff_t pos_out, u64 len)
 492 {
 493         return ovl_copyfile(file_in, pos_in, file_out, pos_out, len, 0,
 494                             OVL_CLONE);
 495 }
 496
 497 static int ovl_dedupe_file_range(struct file *file_in, loff_t pos_in,
 498                                  struct file *file_out, loff_t pos_out, u64 len)
 499 {
 500         /*
 501          * Don't copy up because of a dedupe request, this wouldn't make sense
 502          * most of the time (data would be duplicated instead of deduplicated).
 503          */
 504         if (!ovl_inode_upper(file_inode(file_in)) ||
 505             !ovl_inode_upper(file_inode(file_out)))
 506                 return -EPERM;
 507
 508         return ovl_copyfile(file_in, pos_in, file_out, pos_out, len, 0,
 509                             OVL_DEDUPE);
 510 }
 511
 512 const struct file_operations ovl_file_operations = {
 513         .open           = ovl_open,
 514         .release        = ovl_release,
 515         .llseek         = ovl_llseek,
 516         .read_iter      = ovl_read_iter,
 517         .write_iter     = ovl_write_iter,
 518         .fsync          = ovl_fsync,
 519         .mmap           = ovl_mmap,
 520         .fallocate      = ovl_fallocate,
 521         .fadvise        = ovl_fadvise,
 522         .unlocked_ioctl = ovl_ioctl,
 523         .compat_ioctl   = ovl_compat_ioctl,
 524
 525         .copy_file_range        = ovl_copy_file_range,
 526         .clone_file_range       = ovl_clone_file_range,
 527         .dedupe_file_range      = ovl_dedupe_file_range,
 528 };