Whamcloud - gitweb
Land b_release_1_4_6 onto HEAD (20060223_1455)
[fs/lustre-release.git] / libsysio / src / namei.c
1 /*
2  *    This Cplant(TM) source code is the property of Sandia National
3  *    Laboratories.
4  *
5  *    This Cplant(TM) source code is copyrighted by Sandia National
6  *    Laboratories.
7  *
8  *    The redistribution of this Cplant(TM) source code is subject to the
9  *    terms of the GNU Lesser General Public License
10  *    (see cit/LGPL or http://www.gnu.org/licenses/lgpl.html)
11  *
12  *    Cplant(TM) Copyright 1998-2003 Sandia Corporation. 
13  *    Under the terms of Contract DE-AC04-94AL85000, there is a non-exclusive
14  *    license for use of this work by or on behalf of the US Government.
15  *    Export of this program may require a license from the United States
16  *    Government.
17  */
18
19 /*
20  * This library is free software; you can redistribute it and/or
21  * modify it under the terms of the GNU Lesser General Public
22  * License as published by the Free Software Foundation; either
23  * version 2.1 of the License, or (at your option) any later version.
24  * 
25  * This library is distributed in the hope that it will be useful,
26  * but WITHOUT ANY WARRANTY; without even the implied warranty of
27  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
28  * Lesser General Public License for more details.
29  * 
30  * You should have received a copy of the GNU Lesser General Public
31  * License along with this library; if not, write to the Free Software
32  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
33  *
34  * Questions or comments about this library should be sent to:
35  *
36  * Lee Ward
37  * Sandia National Laboratories, New Mexico
38  * P.O. Box 5800
39  * Albuquerque, NM 87185-1110
40  *
41  * lee@sandia.gov
42  */
43
44 #if defined(AUTOMOUNT_FILE_NAME) && defined(__linux__)
45 #define _BSD_SOURCE
46 #endif
47
48 #include <stdlib.h>
49 #include <string.h>
50 #include <errno.h>
51 #include <assert.h>
52 #include <unistd.h>
53 #include <sys/param.h>
54 #include <sys/types.h>
55 #include <sys/stat.h>
56 #include <sys/queue.h>
57
58 #include "sysio.h"
59 #include "mount.h"
60 #include "inode.h"
61
62 /*
63  * Parse next component in path.
64  */
65 #ifndef AUTOMOUNT_FILE_NAME
66 static
67 #endif
68 void
69 _sysio_next_component(const char *path, struct qstr *name)
70 {
71         while (*path == PATH_SEPARATOR)
72                 path++;
73         name->name = path;
74         name->len = 0;
75         name->hashval = 0;
76         while (*path && *path != PATH_SEPARATOR) {
77                 name->hashval =
78                     37 * name->hashval + *path++;
79                 name->len++;
80         }
81 }
82
83 /*
84  * Given parent, look up component.
85  */
86 static int
87 lookup(struct pnode *parent,
88        struct qstr *name,
89        struct pnode **pnop,
90        struct intent *intnt,
91        const char *path)
92 {
93         int     err;
94         struct pnode *pno;
95
96         if (!parent->p_base->pb_ino)
97                 return -ENOTDIR;
98
99         err = _sysio_permitted(parent->p_base->pb_ino, X_OK);
100         if (err)
101                 return err;
102
103         /*
104          * Short-circuit `.' and `..'; We don't cache those.
105          */
106         pno = NULL;
107         if (name->len == 1 && name->name[0] == '.')
108                 pno = parent;
109         else if (name->len == 2 && name->name[0] == '.' && name->name[1] == '.')
110                 pno = parent->p_parent;
111         if (pno)
112                 P_REF(pno);
113         else {
114                 /*
115                  * Get cache entry then.
116                  */
117                 err = _sysio_p_find_alias(parent, name, &pno);
118                 if (err)
119                         return err;
120         }
121
122         /*
123          * While covered, move to the covering node.
124          */
125         while (pno->p_cover && pno->p_cover != pno) {
126                 struct pnode *cover;
127
128                 cover = pno->p_cover;
129                 P_REF(cover);
130                 P_RELE(pno);
131                 pno = cover;
132         }
133
134         *pnop = pno;
135
136         /*
137          * (Re)validate the pnode.
138          */
139         err = _sysio_p_validate(pno, intnt, path);
140         if (err)
141                 return err;
142
143         return 0;
144 }
145
146 /*
147  * The meat. Walk an absolute or relative path, looking up each
148  * component. Various flags in the nameidata argument govern actions
149  * and return values/state. They are:
150  *
151  * ND_NOFOLLOW          symbolic links are not followed
152  * ND_NEGOK             if terminal/leaf does not exist, return
153  *                       path node (alias) anyway.
154  */
155 int
156 _sysio_path_walk(struct pnode *parent, struct nameidata *nd)
157 {
158         int     err;
159         const char *path;
160         struct qstr this, next;
161         struct inode *ino;
162
163         /*
164          * NULL path?
165          */
166         if (!nd->nd_path)
167                 return -EFAULT;
168
169         /*
170          * Empty path?
171          */
172         if (!*nd->nd_path)
173                 return -ENOENT;
174
175         /*
176          * Leading slash?
177          */
178         if (*nd->nd_path == PATH_SEPARATOR) {
179                 /*
180                  * Make parent the root of the name space.
181                  */
182                 parent = nd->nd_root;
183         }
184
185 #if DEFER_INIT_CWD
186         if (!parent) {
187                 const char *icwd;
188
189                 if (!_sysio_init_cwd && !nd->nd_root)
190                         abort();
191
192                 /*
193                  * Finally have to set the current working directory. We can
194                  * not tolerate errors here or else risk leaving the process
195                  * in a very unexpected location. We abort then unless all goes
196                  * well.
197                  */
198                 icwd = _sysio_init_cwd;
199                 _sysio_init_cwd = NULL;
200                 parent = nd->nd_root;
201                 if (!parent)
202                         abort();
203                 (void )_sysio_namei(nd->nd_root, icwd, 0, NULL, &parent);
204                 if (_sysio_p_chdir(parent) != 0)
205                         abort();
206         }
207 #endif
208
209         /*
210          * (Re)Validate the parent.
211          */
212         err = _sysio_p_validate(parent, NULL, NULL);
213         if (err)
214                 return err;
215
216         /*
217          * Prime everything for the loop. Will need another reference to the
218          * initial directory. It'll be dropped later.
219          */
220         nd->nd_pno = parent;
221         P_REF(nd->nd_pno);
222         _sysio_next_component(nd->nd_path, &next);
223         path = next.name;
224         parent = NULL;
225         err = 0;
226
227         /*
228          * Derecurse the path tree-walk.
229          */
230         for (;;) {
231                 ino = nd->nd_pno->p_base->pb_ino;
232                 if (S_ISLNK(ino->i_stbuf.st_mode) &&
233                     (next.len || !(nd->nd_flags & ND_NOFOLLOW))) {
234                         char    *lpath;
235                         ssize_t cc;
236                         struct nameidata nameidata;
237
238                         if (nd->nd_slicnt >= MAX_SYMLINK) {
239                                 err = -ELOOP;
240                                 break;
241                         }
242
243                         /*
244                          * Follow symbolic link.
245                          */
246                         lpath = malloc(MAXPATHLEN + 1);
247                         if (!lpath) {
248                                 err = -ENOMEM;
249                                 break;
250                         }
251                         cc =
252                             ino->i_ops.inop_readlink(nd->nd_pno,
253                                                      lpath,
254                                                      MAXPATHLEN);
255                         if (cc < 0) {
256                                 free(lpath);
257                                 err = (int )cc;
258                                 break;
259                         }
260                         lpath[cc] = '\0';                       /* NUL term */
261                         /*
262                          * Handle symbolic links with recursion. Yuck!
263                          */
264                         ND_INIT(&nameidata,
265                                 (nd->nd_flags | ND_NEGOK),
266                                 lpath,
267                                 nd->nd_root,
268                                 !next.len ? nd->nd_intent : NULL);
269                         nameidata.nd_slicnt = nd->nd_slicnt + 1;
270                         err =
271                             _sysio_path_walk(nd->nd_pno->p_parent, &nameidata);
272                         free(lpath);
273                         if (err)
274                                 break;
275                         P_RELE(nd->nd_pno);
276                         nd->nd_pno = nameidata.nd_pno;
277                         ino = nd->nd_pno->p_base->pb_ino;
278                 }
279 #ifdef AUTOMOUNT_FILE_NAME
280                 else if (ino &&
281                          S_ISDIR(ino->i_stbuf.st_mode) &&
282                          (nd->nd_pno->p_mount->mnt_flags & MOUNT_F_AUTO) &&
283                          nd->nd_amcnt < MAX_MOUNT_DEPTH &&
284                          ino->i_stbuf.st_mode & S_ISUID) {
285                         struct pnode *pno;
286
287                         /*
288                          * We're committed to a lookup. It's time to see if
289                          * we're going to do it in an automount-point and
290                          * arrange the mount if so.
291                          */
292                         assert(!nd->nd_pno->p_cover);
293                         err =
294                             lookup(nd->nd_pno,
295                                    &_sysio_mount_file_name,
296                                    &pno,
297                                    NULL,
298                                    NULL);
299                         if (pno)
300                                 P_RELE(pno);
301                         if (!err && _sysio_automount(pno) == 0) {
302                                 struct pnode *root;
303
304                                 /*
305                                  * All went well. Need to switch
306                                  * parent pno and ino to the
307                                  * root of the newly mounted sub-tree.
308                                  *
309                                  * NB:
310                                  * We don't recurseively retry these
311                                  * things. It's OK to have the new root
312                                  * be an automount-point but it's going
313                                  * to take another lookup to accomplish it.
314                                  * The alternative could get us into an
315                                  * infinite loop.
316                                  */
317                                 root = nd->nd_pno->p_cover;
318                                 assert(root);
319                                 P_RELE(nd->nd_pno);
320                                 nd->nd_pno = root;
321 #if 0
322                                 P_REF(nd->nd_pno);
323 #endif
324                                 ino = nd->nd_pno->p_base->pb_ino;
325                                 assert(ino);
326
327                                 /*
328                                  * Must send the intent-path again.
329                                  */
330                                 path = nd->nd_path;
331                                 nd->nd_amcnt++;
332
333                                 /*
334                                  * Must go back top and retry with this
335                                  * new pnode as parent.
336                                  */
337                                 continue;
338                         }
339                         err = 0;                        /* it never happened */
340                 }
341 #endif
342
343                 /*
344                  * Set up for next component.
345                  */
346                 this = next;
347                 if (path)
348                         path = this.name;
349                 if (!this.len)
350                         break;
351                 if (!ino) {
352                         /*
353                          * Should only be here if final component was
354                          * target of a symlink.
355                          */
356                         nd->nd_path = this.name + this.len;
357                         err = -ENOENT;
358                         break;
359                 }
360                 nd->nd_path = this.name + this.len;
361                 _sysio_next_component(nd->nd_path, &next);
362                 parent = nd->nd_pno;
363                 nd->nd_pno = NULL;
364
365                 /*
366                  * Parent must be a directory.
367                  */
368                 if (ino && !S_ISDIR(ino->i_stbuf.st_mode)) {
369                         err = -ENOTDIR;
370                         break;
371                 }
372
373                 /*
374                  * The extra path arg is passed only on the first lookup in the
375                  * walk as we cross into each file system, anew. The intent is
376                  * passed both on the first lookup and when trying to look up
377                  * the final component -- Of the original path, not on the
378                  * file system.
379                  *
380                  * Confused? Me too and I came up with this weirdness. It's
381                  * hints to the file system drivers. Read on.
382                  *
383                  * The first lookup will give everything one needs to ready
384                  * everything for the entire operation before the path is
385                  * walked. The file system driver knows it's the first lookup
386                  * in the walk because it has both the path and the intent.
387                  *
388                  * Alternatively, one could split the duties; The first lookup
389                  * can be used to prime the file system inode cache with the
390                  * interior nodes we'll want in the path-walk. Then, when
391                  * looking up the last component, ready everything for the
392                  * operations(s) to come. The file system driver knows it's
393                  * the last lookup in the walk because it has the intent,
394                  * again, but without the path.
395                  *
396                  * One special case; If we were asked to look up a single
397                  * component, we treat it as the last component. The file
398                  * system driver never sees the extra path argument. It should
399                  * be noted that the driver always has the fully qualified
400                  * path, on the target file system, available to it for any
401                  * node it is looking up, including the last, via the base
402                  * path node and it's ancestor chain.
403                  */
404                 err =
405                     lookup(parent,
406                            &this,
407                            &nd->nd_pno,
408                            (path || !next.len)
409                              ? nd->nd_intent
410                              : NULL,
411                            (path && next.len) ? path : NULL);
412                 if (err) {
413                         if (err == -ENOENT &&
414                             !next.len &&
415                             (nd->nd_flags & ND_NEGOK))
416                                 err = 0;
417                         break;
418                 }
419                 path = NULL;                            /* Stop that! */
420                 if ((parent->p_mount->mnt_fs !=
421                      nd->nd_pno->p_mount->mnt_fs)) {
422                         /*
423                          * Crossed into a new fs. We'll want the next lookup
424                          * to include the path again.
425                          */
426                         path = nd->nd_path;
427                 }
428
429                 /*
430                  * Release the parent.
431                  */
432                 P_RELE(parent);
433                 parent = NULL;
434         }
435
436         /*
437          * Trailing separators cause us to break from the loop with
438          * a parent set but no pnode. Check for that.
439          */
440         if (!nd->nd_pno) {
441                 nd->nd_pno = parent;
442                 parent = NULL;
443                 /*
444                  * Make sure the last processed component was a directory. The
445                  * trailing slashes are illegal behind anything else.
446                  */
447                 if (!(err ||
448                       S_ISDIR(nd->nd_pno->p_base->pb_ino->i_stbuf.st_mode)))
449                         err = -ENOTDIR;
450         }
451
452         /*
453          * Drop reference to parent if set. Either we have a dup of the original
454          * parent or an intermediate reference.
455          */
456         if (parent)
457                 P_RELE(parent);
458
459         /*
460          * On error, we will want to drop our reference to the current
461          * path node if at end.
462          */
463         if (err && nd->nd_pno) {
464                 P_RELE(nd->nd_pno);
465                 nd->nd_pno = NULL;
466         }
467
468         return err;
469 }
470
471 #ifdef CPLANT_YOD
472 /* 
473  * for backward compatibility w/protocol switch
474  * remove everything up to the first ':'
475  * fortran libs prepend cwd to path, so not much choice
476  */
477 #define STRIP_PREFIX(p) strchr(p,':') ? strchr(p,':')+1 : p
478 #else
479 #define STRIP_PREFIX(p) p
480 #endif
481
482 /*
483  * Expanded form of the path-walk routine, with the common arguments, builds
484  * the nameidata bundle and calls path-walk.
485  */
486 int
487 _sysio_namei(struct pnode *parent,
488              const char *path,
489              unsigned flags,
490              struct intent *intnt,
491              struct pnode **pnop)
492 {
493         struct nameidata nameidata;
494         int     err;
495
496         ND_INIT(&nameidata, flags, STRIP_PREFIX(path), _sysio_root, intnt);
497         err = _sysio_path_walk(parent, &nameidata);
498         if (!err)
499                 *pnop = nameidata.nd_pno;
500         return err;
501 }