fs/exportfs/expfs.c

   1 /*
   2  * Copyright (C) Neil Brown 2002
   3  * Copyright (C) Christoph Hellwig 2007
   4  *
   5  * This file contains the code mapping from inodes to NFS file handles,
   6  * and for mapping back from file handles to dentries.
   7  *
   8  * For details on why we do all the strange and hairy things in here
   9  * take a look at Documentation/filesystems/nfs/Exporting.
  10  */
  11 #include <linux/exportfs.h>
  12 #include <linux/fs.h>
  13 #include <linux/file.h>
  14 #include <linux/module.h>
  15 #include <linux/mount.h>
  16 #include <linux/namei.h>
  17 #include <linux/sched.h>
  18 #include <linux/cred.h>
  19
  20 #define dprintk(fmt, args...) do{}while(0)
  21
  22
  23 static int get_name(const struct path *path, char *name, struct dentry *child);
  24
  25
  26 static int exportfs_get_name(struct vfsmount *mnt, struct dentry *dir,
  27                 char *name, struct dentry *child)
  28 {
  29         const struct export_operations *nop = dir->d_sb->s_export_op;
  30         struct path path = {.mnt = mnt, .dentry = dir};
  31
  32         if (nop->get_name)
  33                 return nop->get_name(dir, name, child);
  34         else
  35                 return get_name(&path, name, child);
  36 }
  37
  38 /*
  39  * Check if the dentry or any of it's aliases is acceptable.
  40  */
  41 static struct dentry *
  42 find_acceptable_alias(struct dentry *result,
  43                 int (*acceptable)(void *context, struct dentry *dentry),
  44                 void *context)
  45 {
  46         struct dentry *dentry, *toput = NULL;
  47         struct inode *inode;
  48
  49         if (acceptable(context, result))
  50                 return result;
  51
  52         inode = result->d_inode;
  53         spin_lock(&inode->i_lock);
  54         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
  55                 dget(dentry);
  56                 spin_unlock(&inode->i_lock);
  57                 if (toput)
  58                         dput(toput);
  59                 if (dentry != result && acceptable(context, dentry)) {
  60                         dput(result);
  61                         return dentry;
  62                 }
  63                 spin_lock(&inode->i_lock);
  64                 toput = dentry;
  65         }
  66         spin_unlock(&inode->i_lock);
  67
  68         if (toput)
  69                 dput(toput);
  70         return NULL;
  71 }
  72
  73 static bool dentry_connected(struct dentry *dentry)
  74 {
  75         dget(dentry);
  76         while (dentry->d_flags & DCACHE_DISCONNECTED) {
  77                 struct dentry *parent = dget_parent(dentry);
  78
  79                 dput(dentry);
  80                 if (dentry == parent) {
  81                         dput(parent);
  82                         return false;
  83                 }
  84                 dentry = parent;
  85         }
  86         dput(dentry);
  87         return true;
  88 }
  89
  90 static void clear_disconnected(struct dentry *dentry)
  91 {
  92         dget(dentry);
  93         while (dentry->d_flags & DCACHE_DISCONNECTED) {
  94                 struct dentry *parent = dget_parent(dentry);
  95
  96                 WARN_ON_ONCE(IS_ROOT(dentry));
  97
  98                 spin_lock(&dentry->d_lock);
  99                 dentry->d_flags &= ~DCACHE_DISCONNECTED;
 100                 spin_unlock(&dentry->d_lock);
 101
 102                 dput(dentry);
 103                 dentry = parent;
 104         }
 105         dput(dentry);
 106 }
 107
 108 /*
 109  * Reconnect a directory dentry with its parent.
 110  *
 111  * This can return a dentry, or NULL, or an error.
 112  *
 113  * In the first case the returned dentry is the parent of the given
 114  * dentry, and may itself need to be reconnected to its parent.
 115  *
 116  * In the NULL case, a concurrent VFS operation has either renamed or
 117  * removed this directory.  The concurrent operation has reconnected our
 118  * dentry, so we no longer need to.
 119  */
 120 static struct dentry *reconnect_one(struct vfsmount *mnt,
 121                 struct dentry *dentry, char *nbuf)
 122 {
 123         struct dentry *parent;
 124         struct dentry *tmp;
 125         int err;
 126
 127         parent = ERR_PTR(-EACCES);
 128         inode_lock(dentry->d_inode);
 129         if (mnt->mnt_sb->s_export_op->get_parent)
 130                 parent = mnt->mnt_sb->s_export_op->get_parent(dentry);
 131         inode_unlock(dentry->d_inode);
 132
 133         if (IS_ERR(parent)) {
 134                 dprintk("%s: get_parent of %ld failed, err %d\n",
 135                         __func__, dentry->d_inode->i_ino, PTR_ERR(parent));
 136                 return parent;
 137         }
 138
 139         dprintk("%s: find name of %lu in %lu\n", __func__,
 140                 dentry->d_inode->i_ino, parent->d_inode->i_ino);
 141         err = exportfs_get_name(mnt, parent, nbuf, dentry);
 142         if (err == -ENOENT)
 143                 goto out_reconnected;
 144         if (err)
 145                 goto out_err;
 146         dprintk("%s: found name: %s\n", __func__, nbuf);
 147         tmp = lookup_one_len_unlocked(nbuf, parent, strlen(nbuf));
 148         if (IS_ERR(tmp)) {
 149                 dprintk("%s: lookup failed: %d\n", __func__, PTR_ERR(tmp));
 150                 err = PTR_ERR(tmp);
 151                 goto out_err;
 152         }
 153         if (tmp != dentry) {
 154                 /*
 155                  * Somebody has renamed it since exportfs_get_name();
 156                  * great, since it could've only been renamed if it
 157                  * got looked up and thus connected, and it would
 158                  * remain connected afterwards.  We are done.
 159                  */
 160                 dput(tmp);
 161                 goto out_reconnected;
 162         }
 163         dput(tmp);
 164         if (IS_ROOT(dentry)) {
 165                 err = -ESTALE;
 166                 goto out_err;
 167         }
 168         return parent;
 169
 170 out_err:
 171         dput(parent);
 172         return ERR_PTR(err);
 173 out_reconnected:
 174         dput(parent);
 175         /*
 176          * Someone must have renamed our entry into another parent, in
 177          * which case it has been reconnected by the rename.
 178          *
 179          * Or someone removed it entirely, in which case filehandle
 180          * lookup will succeed but the directory is now IS_DEAD and
 181          * subsequent operations on it will fail.
 182          *
 183          * Alternatively, maybe there was no race at all, and the
 184          * filesystem is just corrupt and gave us a parent that doesn't
 185          * actually contain any entry pointing to this inode.  So,
 186          * double check that this worked and return -ESTALE if not:
 187          */
 188         if (!dentry_connected(dentry))
 189                 return ERR_PTR(-ESTALE);
 190         return NULL;
 191 }
 192
 193 /*
 194  * Make sure target_dir is fully connected to the dentry tree.
 195  *
 196  * On successful return, DCACHE_DISCONNECTED will be cleared on
 197  * target_dir, and target_dir->d_parent->...->d_parent will reach the
 198  * root of the filesystem.
 199  *
 200  * Whenever DCACHE_DISCONNECTED is unset, target_dir is fully connected.
 201  * But the converse is not true: target_dir may have DCACHE_DISCONNECTED
 202  * set but already be connected.  In that case we'll verify the
 203  * connection to root and then clear the flag.
 204  *
 205  * Note that target_dir could be removed by a concurrent operation.  In
 206  * that case reconnect_path may still succeed with target_dir fully
 207  * connected, but further operations using the filehandle will fail when
 208  * necessary (due to S_DEAD being set on the directory).
 209  */
 210 static int
 211 reconnect_path(struct vfsmount *mnt, struct dentry *target_dir, char *nbuf)
 212 {
 213         struct dentry *dentry, *parent;
 214
 215         dentry = dget(target_dir);
 216
 217         while (dentry->d_flags & DCACHE_DISCONNECTED) {
 218                 BUG_ON(dentry == mnt->mnt_sb->s_root);
 219
 220                 if (IS_ROOT(dentry))
 221                         parent = reconnect_one(mnt, dentry, nbuf);
 222                 else
 223                         parent = dget_parent(dentry);
 224
 225                 if (!parent)
 226                         break;
 227                 dput(dentry);
 228                 if (IS_ERR(parent))
 229                         return PTR_ERR(parent);
 230                 dentry = parent;
 231         }
 232         dput(dentry);
 233         clear_disconnected(target_dir);
 234         return 0;
 235 }
 236
 237 struct getdents_callback {
 238         struct dir_context ctx;
 239         char *name;             /* name that was found. It already points to a
 240                                    buffer NAME_MAX+1 is size */
 241         u64 ino;                /* the inum we are looking for */
 242         int found;              /* inode matched? */
 243         int sequence;           /* sequence counter */
 244 };
 245
 246 /*
 247  * A rather strange filldir function to capture
 248  * the name matching the specified inode number.
 249  */
 250 static int filldir_one(struct dir_context *ctx, const char *name, int len,
 251                         loff_t pos, u64 ino, unsigned int d_type)
 252 {
 253         struct getdents_callback *buf =
 254                 container_of(ctx, struct getdents_callback, ctx);
 255         int result = 0;
 256
 257         buf->sequence++;
 258         if (buf->ino == ino && len <= NAME_MAX) {
 259                 memcpy(buf->name, name, len);
 260                 buf->name[len] = '\0';
 261                 buf->found = 1;
 262                 result = -1;
 263         }
 264         return result;
 265 }
 266
 267 /**
 268  * get_name - default export_operations->get_name function
 269  * @path:   the directory in which to find a name
 270  * @name:   a pointer to a %NAME_MAX+1 char buffer to store the name
 271  * @child:  the dentry for the child directory.
 272  *
 273  * calls readdir on the parent until it finds an entry with
 274  * the same inode number as the child, and returns that.
 275  */
 276 static int get_name(const struct path *path, char *name, struct dentry *child)
 277 {
 278         const struct cred *cred = current_cred();
 279         struct inode *dir = path->dentry->d_inode;
 280         int error;
 281         struct file *file;
 282         struct kstat stat;
 283         struct path child_path = {
 284                 .mnt = path->mnt,
 285                 .dentry = child,
 286         };
 287         struct getdents_callback buffer = {
 288                 .ctx.actor = filldir_one,
 289                 .name = name,
 290         };
 291
 292         error = -ENOTDIR;
 293         if (!dir || !S_ISDIR(dir->i_mode))
 294                 goto out;
 295         error = -EINVAL;
 296         if (!dir->i_fop)
 297                 goto out;
 298         /*
 299          * inode->i_ino is unsigned long, kstat->ino is u64, so the
 300          * former would be insufficient on 32-bit hosts when the
 301          * filesystem supports 64-bit inode numbers.  So we need to
 302          * actually call ->getattr, not just read i_ino:
 303          */
 304         error = vfs_getattr_nosec(&child_path, &stat,
 305                                   STATX_INO, AT_STATX_SYNC_AS_STAT);
 306         if (error)
 307                 return error;
 308         buffer.ino = stat.ino;
 309         /*
 310          * Open the directory ...
 311          */
 312         file = dentry_open(path, O_RDONLY, cred);
 313         error = PTR_ERR(file);
 314         if (IS_ERR(file))
 315                 goto out;
 316
 317         error = -EINVAL;
 318         if (!file->f_op->iterate && !file->f_op->iterate_shared)
 319                 goto out_close;
 320
 321         buffer.sequence = 0;
 322         while (1) {
 323                 int old_seq = buffer.sequence;
 324
 325                 error = iterate_dir(file, &buffer.ctx);
 326                 if (buffer.found) {
 327                         error = 0;
 328                         break;
 329                 }
 330
 331                 if (error < 0)
 332                         break;
 333
 334                 error = -ENOENT;
 335                 if (old_seq == buffer.sequence)
 336                         break;
 337         }
 338
 339 out_close:
 340         fput(file);
 341 out:
 342         return error;
 343 }
 344
 345 /**
 346  * export_encode_fh - default export_operations->encode_fh function
 347  * @inode:   the object to encode
 348  * @fid:     where to store the file handle fragment
 349  * @max_len: maximum length to store there
 350  * @parent:  parent directory inode, if wanted
 351  *
 352  * This default encode_fh function assumes that the 32 inode number
 353  * is suitable for locating an inode, and that the generation number
 354  * can be used to check that it is still valid.  It places them in the
 355  * filehandle fragment where export_decode_fh expects to find them.
 356  */
 357 static int export_encode_fh(struct inode *inode, struct fid *fid,
 358                 int *max_len, struct inode *parent)
 359 {
 360         int len = *max_len;
 361         int type = FILEID_INO32_GEN;
 362
 363         if (parent && (len < 4)) {
 364                 *max_len = 4;
 365                 return FILEID_INVALID;
 366         } else if (len < 2) {
 367                 *max_len = 2;
 368                 return FILEID_INVALID;
 369         }
 370
 371         len = 2;
 372         fid->i32.ino = inode->i_ino;
 373         fid->i32.gen = inode->i_generation;
 374         if (parent) {
 375                 fid->i32.parent_ino = parent->i_ino;
 376                 fid->i32.parent_gen = parent->i_generation;
 377                 len = 4;
 378                 type = FILEID_INO32_GEN_PARENT;
 379         }
 380         *max_len = len;
 381         return type;
 382 }
 383
 384 int exportfs_encode_inode_fh(struct inode *inode, struct fid *fid,
 385                              int *max_len, struct inode *parent)
 386 {
 387         const struct export_operations *nop = inode->i_sb->s_export_op;
 388
 389         if (nop && nop->encode_fh)
 390                 return nop->encode_fh(inode, fid->raw, max_len, parent);
 391
 392         return export_encode_fh(inode, fid, max_len, parent);
 393 }
 394 EXPORT_SYMBOL_GPL(exportfs_encode_inode_fh);
 395
 396 int exportfs_encode_fh(struct dentry *dentry, struct fid *fid, int *max_len,
 397                 int connectable)
 398 {
 399         int error;
 400         struct dentry *p = NULL;
 401         struct inode *inode = dentry->d_inode, *parent = NULL;
 402
 403         if (connectable && !S_ISDIR(inode->i_mode)) {
 404                 p = dget_parent(dentry);
 405                 /*
 406                  * note that while p might've ceased to be our parent already,
 407                  * it's still pinned by and still positive.
 408                  */
 409                 parent = p->d_inode;
 410         }
 411
 412         error = exportfs_encode_inode_fh(inode, fid, max_len, parent);
 413         dput(p);
 414
 415         return error;
 416 }
 417 EXPORT_SYMBOL_GPL(exportfs_encode_fh);
 418
 419 struct dentry *exportfs_decode_fh(struct vfsmount *mnt, struct fid *fid,
 420                 int fh_len, int fileid_type,
 421                 int (*acceptable)(void *, struct dentry *), void *context)
 422 {
 423         const struct export_operations *nop = mnt->mnt_sb->s_export_op;
 424         struct dentry *result, *alias;
 425         char nbuf[NAME_MAX+1];
 426         int err;
 427
 428         /*
 429          * Try to get any dentry for the given file handle from the filesystem.
 430          */
 431         if (!nop || !nop->fh_to_dentry)
 432                 return ERR_PTR(-ESTALE);
 433         result = nop->fh_to_dentry(mnt->mnt_sb, fid, fh_len, fileid_type);
 434         if (PTR_ERR(result) == -ENOMEM)
 435                 return ERR_CAST(result);
 436         if (IS_ERR_OR_NULL(result))
 437                 return ERR_PTR(-ESTALE);
 438
 439         /*
 440          * If no acceptance criteria was specified by caller, a disconnected
 441          * dentry is also accepatable. Callers may use this mode to query if
 442          * file handle is stale or to get a reference to an inode without
 443          * risking the high overhead caused by directory reconnect.
 444          */
 445         if (!acceptable)
 446                 return result;
 447
 448         if (d_is_dir(result)) {
 449                 /*
 450                  * This request is for a directory.
 451                  *
 452                  * On the positive side there is only one dentry for each
 453                  * directory inode.  On the negative side this implies that we
 454                  * to ensure our dentry is connected all the way up to the
 455                  * filesystem root.
 456                  */
 457                 if (result->d_flags & DCACHE_DISCONNECTED) {
 458                         err = reconnect_path(mnt, result, nbuf);
 459                         if (err)
 460                                 goto err_result;
 461                 }
 462
 463                 if (!acceptable(context, result)) {
 464                         err = -EACCES;
 465                         goto err_result;
 466                 }
 467
 468                 return result;
 469         } else {
 470                 /*
 471                  * It's not a directory.  Life is a little more complicated.
 472                  */
 473                 struct dentry *target_dir, *nresult;
 474
 475                 /*
 476                  * See if either the dentry we just got from the filesystem
 477                  * or any alias for it is acceptable.  This is always true
 478                  * if this filesystem is exported without the subtreecheck
 479                  * option.  If the filesystem is exported with the subtree
 480                  * check option there's a fair chance we need to look at
 481                  * the parent directory in the file handle and make sure
 482                  * it's connected to the filesystem root.
 483                  */
 484                 alias = find_acceptable_alias(result, acceptable, context);
 485                 if (alias)
 486                         return alias;
 487
 488                 /*
 489                  * Try to extract a dentry for the parent directory from the
 490                  * file handle.  If this fails we'll have to give up.
 491                  */
 492                 err = -ESTALE;
 493                 if (!nop->fh_to_parent)
 494                         goto err_result;
 495
 496                 target_dir = nop->fh_to_parent(mnt->mnt_sb, fid,
 497                                 fh_len, fileid_type);
 498                 if (!target_dir)
 499                         goto err_result;
 500                 err = PTR_ERR(target_dir);
 501                 if (IS_ERR(target_dir))
 502                         goto err_result;
 503
 504                 /*
 505                  * And as usual we need to make sure the parent directory is
 506                  * connected to the filesystem root.  The VFS really doesn't
 507                  * like disconnected directories..
 508                  */
 509                 err = reconnect_path(mnt, target_dir, nbuf);
 510                 if (err) {
 511                         dput(target_dir);
 512                         goto err_result;
 513                 }
 514
 515                 /*
 516                  * Now that we've got both a well-connected parent and a
 517                  * dentry for the inode we're after, make sure that our
 518                  * inode is actually connected to the parent.
 519                  */
 520                 err = exportfs_get_name(mnt, target_dir, nbuf, result);
 521                 if (!err) {
 522                         inode_lock(target_dir->d_inode);
 523                         nresult = lookup_one_len(nbuf, target_dir,
 524                                                  strlen(nbuf));
 525                         inode_unlock(target_dir->d_inode);
 526                         if (!IS_ERR(nresult)) {
 527                                 if (nresult->d_inode) {
 528                                         dput(result);
 529                                         result = nresult;
 530                                 } else
 531                                         dput(nresult);
 532                         }
 533                 }
 534
 535                 /*
 536                  * At this point we are done with the parent, but it's pinned
 537                  * by the child dentry anyway.
 538                  */
 539                 dput(target_dir);
 540
 541                 /*
 542                  * And finally make sure the dentry is actually acceptable
 543                  * to NFSD.
 544                  */
 545                 alias = find_acceptable_alias(result, acceptable, context);
 546                 if (!alias) {
 547                         err = -EACCES;
 548                         goto err_result;
 549                 }
 550
 551                 return alias;
 552         }
 553
 554  err_result:
 555         dput(result);
 556         if (err != -ENOMEM)
 557                 err = -ESTALE;
 558         return ERR_PTR(err);
 559 }
 560 EXPORT_SYMBOL_GPL(exportfs_decode_fh);
 561
 562 MODULE_LICENSE("GPL");