xref: /illumos-gate/usr/src/cmd/fs.d/ufs/fsck/pass3.c (revision 8119dad84d6416f13557b0ba8e2aaf9064cbcfd3)
1 /*
2  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
3  * Use is subject to license terms.
4  */
5 
6 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
7 /*	  All Rights Reserved  	*/
8 
9 /*
10  * Copyright (c) 1980, 1986, 1990 The Regents of the University of California.
11  * All rights reserved.
12  *
13  * Redistribution and use in source and binary forms are permitted
14  * provided that: (1) source distributions retain this entire copyright
15  * notice and comment, and (2) distributions including binaries display
16  * the following acknowledgement:  ``This product includes software
17  * developed by the University of California, Berkeley and its contributors''
18  * in the documentation or other materials provided with the distribution
19  * and in all advertising materials mentioning features or use of this
20  * software. Neither the name of the University nor the names of its
21  * contributors may be used to endorse or promote products derived
22  * from this software without specific prior written permission.
23  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
25  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
26  */
27 
28 #include <stdio.h>
29 #include <stdlib.h>
30 #include <unistd.h>
31 #include <string.h>
32 #include <sys/param.h>
33 #include <sys/types.h>
34 #include <sys/mntent.h>
35 #include <sys/fs/ufs_fs.h>
36 #include <sys/vnode.h>
37 #include <sys/fs/ufs_inode.h>
38 #define	_KERNEL
39 #include <sys/fs/ufs_fsdir.h>
40 #undef _KERNEL
41 #include "fsck.h"
42 
43 static int pass3acheck(struct inodesc *);
44 static void setcurino(struct inodesc *, struct dinode *, struct inoinfo *);
45 
46 void
47 pass3a(void)
48 {
49 	caddr_t flow;
50 	struct inoinfo **inpp, *inp;
51 	fsck_ino_t orphan;
52 	int loopcnt;
53 	int state;
54 	struct shadowclientinfo *sci, *sci_victim, *sci_prev, **sci_rootp;
55 	struct inodesc curino;
56 	struct dinode *dp;
57 	struct inodesc idesc;
58 	char namebuf[MAXNAMLEN + 1];
59 
60 	for (inpp = &inpsort[inplast - 1]; inpp >= inpsort; inpp--) {
61 		inp = *inpp;
62 		state = statemap[inp->i_number];
63 		if (inp->i_number == UFSROOTINO ||
64 		    (inp->i_parent != 0 && !S_IS_DUNFOUND(state)))
65 			continue;
66 		if (state == DCLEAR || state == USTATE || (state & INORPHAN))
67 			continue;
68 		/*
69 		 * If we are running with logging and we come
70 		 * across unreferenced directories, we just leave
71 		 * them in DSTATE which will cause them to be pitched
72 		 * in pass 4.
73 		 */
74 		if (preen && !iscorrupt && islog && S_IS_DUNFOUND(state)) {
75 			if (inp->i_dotdot >= UFSROOTINO) {
76 				LINK_RANGE(flow, lncntp[inp->i_dotdot], 1);
77 				if (flow != NULL) {
78 					dp = ginode(inp->i_dotdot);
79 					LINK_CLEAR(flow, inp->i_dotdot,
80 					    dp->di_mode, &idesc);
81 					if (statemap[inp->i_dotdot] == USTATE)
82 						continue;
83 				}
84 				TRACK_LNCNTP(inp->i_dotdot,
85 				    lncntp[inp->i_dotdot]++);
86 			}
87 			continue;
88 		}
89 
90 		for (loopcnt = 0; ; loopcnt++) {
91 			orphan = inp->i_number;
92 			/*
93 			 * Skip out if we aren't connected to the name
94 			 * space, or our parent is connected, or we've
95 			 * looked at too many directories.  Our parent
96 			 * being connected means that orphan is the
97 			 * first ancestor of *inpp with questionable
98 			 * antecedents.
99 			 */
100 			if (inp->i_parent == 0 ||
101 			    !INO_IS_DUNFOUND(inp->i_parent) ||
102 			    loopcnt > numdirs)
103 				break;
104 			inp = getinoinfo(inp->i_parent);
105 			/*
106 			 * Can't happen, because a non-zero parent's already
107 			 * been seen and therefore cached.
108 			 */
109 			if (inp == NULL)
110 				errexit("pass3 could not find cached "
111 					"inode I=%d\n",
112 					inp->i_parent);
113 		}
114 
115 		/*
116 		 * Already did this one.  Don't bother the user
117 		 * with redundant questions.
118 		 */
119 		if (statemap[orphan] & INORPHAN)
120 			continue;
121 
122 		/*
123 		 * A link count of 0 with parent and .. inodes of 0
124 		 * indicates a partly deleted directory.
125 		 * Clear it.
126 		 */
127 		dp = ginode(orphan);
128 		if (dp->di_nlink == 0 && inp->i_dotdot == 0 &&
129 		    inp->i_parent == 0) {
130 			/*
131 			 * clri() just uses curino.id_number; in other
132 			 * words, it won't use the callback that setcurino()
133 			 * puts in.
134 			 */
135 			setcurino(&curino, dp, inp);
136 			clri(&curino, "UNREF", CLRI_VERBOSE, CLRI_NOP_OK);
137 
138 			/*
139 			 * If we didn't clear it, at least mark it so
140 			 * we don't waste time on it again.
141 			 */
142 			if (statemap[orphan] != USTATE) {
143 				statemap[orphan] |= INORPHAN;
144 			}
145 			continue;
146 		}
147 
148 		/*
149 		 * We can call linkup() multiple times on the same directory
150 		 * inode, if we were told not to reconnect it the first time.
151 		 * This is because we find it as a disconnected parent of
152 		 * of its children (and mark it found), and then finally get
153 		 * to it in the inpsort array.  This is better than in the
154 		 * past, where we'd call it every time we found it as a
155 		 * child's parent.  Ideally, we'd suppress even the second
156 		 * query, but that confuses pass 4's interpretation of
157 		 * the state flags.
158 		 */
159 		if (loopcnt <= countdirs) {
160 			if (linkup(orphan, inp->i_dotdot, NULL)) {
161 				/*
162 				 * Bookkeeping for any sort of relinked
163 				 * directory.
164 				 */
165 				inp->i_dotdot = lfdir;
166 				inp->i_parent = inp->i_dotdot;
167 				statemap[orphan] &= ~(INORPHAN);
168 			} else {
169 				statemap[orphan] |= INORPHAN;
170 			}
171 			propagate();
172 			continue;
173 		}
174 
175 		/*
176 		 * We visited more directories than exist in the
177 		 * filesystem.  The only way to do that is if there's
178 		 * a loop.
179 		 */
180 		pfatal("ORPHANED DIRECTORY LOOP DETECTED I=%d\n", orphan);
181 
182 		/*
183 		 * Can never get here with inp->i_parent zero, because
184 		 * of the interactions between the for() and the
185 		 * if (loopcnt <= countdirs) above.
186 		 */
187 		init_inodesc(&idesc);
188 		idesc.id_type = DATA;
189 		idesc.id_number = inp->i_parent;
190 		idesc.id_parent = orphan;
191 		idesc.id_func = findname;
192 		idesc.id_name = namebuf;
193 		namebuf[0] = '\0';
194 
195 		/*
196 		 * Theoretically, this lookup via ckinode can't fail
197 		 * (if orphan doesn't exist in i_parent, then i_parent
198 		 * would not have been filled in by pass2check()).
199 		 * However, if we're interactive, we want to at least
200 		 * attempt to continue.  The worst case is that it
201 		 * gets reconnected as #nnn into lost+found instead of
202 		 * to its old parent with its old name.
203 		 */
204 		if ((ckinode(ginode(inp->i_parent),
205 		    &idesc, CKI_TRAVERSE) & FOUND) == 0)
206 			pfatal("COULD NOT FIND NAME IN PARENT DIRECTORY");
207 
208 		if (linkup(orphan, inp->i_parent, namebuf)) {
209 			if (cleardirentry(inp->i_parent, orphan) & FOUND) {
210 				LFDIR_LINK_RANGE_NORVAL(flow, lncntp[lfdir], 1,
211 				    &idesc);
212 				TRACK_LNCNTP(orphan, lncntp[orphan]++);
213 			}
214 			inp->i_parent = inp->i_dotdot = lfdir;
215 			LFDIR_LINK_RANGE_NORVAL(flow, lncntp[lfdir], -1,
216 			    &idesc);
217 			TRACK_LNCNTP(lfdir, lncntp[lfdir]--);
218 			statemap[orphan] = DFOUND;
219 		} else {
220 			/*
221 			 * Represents a on-disk leak, not an inconsistency,
222 			 * so don't set iscorrupt.  Such leaks are harmless
223 			 * in the context of discrepancies that the kernel
224 			 * will panic over.
225 			 *
226 			 * We don't care if tsearch() returns non-NULL
227 			 * != orphan, since there's no dynamic memory
228 			 * to free here.
229 			 */
230 			if (tsearch((void *)orphan, &limbo_dirs,
231 				    ino_t_cmp) == NULL)
232 				errexit("out of memory");
233 			statemap[orphan] |= INORPHAN;
234 			continue;
235 		}
236 		propagate();
237 	}
238 
239 	/*
240 	 * The essence of the inner loop is to update the inode of
241 	 * every shadow or attribute inode's lncntp[] by the number of
242 	 * links we've found to them in pass 2 and above.  Logically,
243 	 * all that is needed is just the one line:
244 	 *
245 	 * lncntp[sci->shadow] -= sci->totalclients;
246 	 *
247 	 * However, there's the possibility of wrapping the link count
248 	 * (this is especially true for shadows, which are expected to
249 	 * be shared amongst many files).  This means that we have to
250 	 * range-check before changing anything, and if the check
251 	 * fails, offer to clear the shadow or attribute.  If we do
252 	 * clear it, then we have to remove it from the linked list of
253 	 * all of the type of inodes that we're going through.
254 	 *
255 	 * Just to make things a little more complicated, these are
256 	 * singly-linked lists, so we have to do all the extra
257 	 * bookkeeping that goes along with that as well.
258 	 *
259 	 * The only connection between the shadowclientinfo and
260 	 * attrclientinfo lists is that they use the same underlying
261 	 * struct.  Both need this scan, so the outer loop is just to
262 	 * pick which one we're working on at the moment.  There is no
263 	 * requirement as to which of these lists is scanned first.
264 	 */
265 	for (loopcnt = 0; loopcnt < 2; loopcnt++) {
266 		if (loopcnt == 0)
267 			sci_rootp = &shadowclientinfo;
268 		else
269 			sci_rootp = &attrclientinfo;
270 
271 		sci = *sci_rootp;
272 		sci_prev = NULL;
273 		while (sci != NULL) {
274 			sci_victim = NULL;
275 			LINK_RANGE(flow, lncntp[sci->shadow],
276 			    -(sci->totalClients));
277 			if (flow != NULL) {
278 				/*
279 				 * Overflowed the link count.
280 				 */
281 				dp = ginode(sci->shadow);
282 				LINK_CLEAR(flow, sci->shadow, dp->di_mode,
283 				    &idesc);
284 				if (statemap[sci->shadow] == USTATE) {
285 					/*
286 					 * It's been cleared, fix the
287 					 * lists.
288 					 */
289 					if (sci_prev == NULL) {
290 						*sci_rootp = sci->next;
291 					} else {
292 						sci_prev->next = sci->next;
293 					}
294 					sci_victim = sci;
295 				}
296 			}
297 
298 			/*
299 			 * If we did not clear the shadow, then we
300 			 * need to update the count and advance the
301 			 * previous pointer.  Otherwise, finish the
302 			 * clean up once we're done with the struct.
303 			 */
304 			if (sci_victim == NULL) {
305 				TRACK_LNCNTP(sci->shadow,
306 				    lncntp[sci->shadow] -= sci->totalClients);
307 				sci_prev = sci;
308 			}
309 			sci = sci->next;
310 			if (sci_victim != NULL)
311 				deshadow(sci_victim, NULL);
312 		}
313 	}
314 }
315 
316 
317 /*
318  * This is used to verify the cflags of files
319  * under a directory that used to be an attrdir.
320  */
321 
322 static int
323 pass3acheck(struct inodesc *idesc)
324 {
325 	struct direct *dirp = idesc->id_dirp;
326 	int n = 0, ret = 0;
327 	struct dinode *dp, *pdirp;
328 	int isattr;
329 	int dirtype;
330 	int inotype;
331 
332 	if (dirp->d_ino == 0)
333 		return (KEEPON);
334 
335 	idesc->id_entryno++;
336 	if ((strcmp(dirp->d_name, ".") == 0) ||
337 	    (strcmp(dirp->d_name, "..") == 0)) {
338 		return (KEEPON);
339 	}
340 
341 	switch (statemap[dirp->d_ino] & ~(INDELAYD)) {
342 	case DSTATE:
343 	case DFOUND:
344 	case FSTATE:
345 		/*
346 		 * Accept DSTATE and DFOUND so we can handle normal
347 		 * directories as well as xattr directories.
348 		 *
349 		 * For extended attribute directories .. may point
350 		 * to a file.  In this situation we don't want
351 		 * to decrement link count as it was already
352 		 * decremented when the entry was seen and decremented
353 		 * in the directory it actually lives in.
354 		 */
355 		dp = ginode(dirp->d_ino);
356 		isattr = (dp->di_cflags & IXATTR);
357 		inotype = (dp->di_mode & IFMT);
358 		pdirp = ginode(idesc->id_number);
359 		dirtype = (pdirp->di_mode & IFMT);
360 		/*
361 		 * IXATTR indicates that an object is itself an extended
362 		 * attribute.  An IFMT of IFATTRDIR means we are looking
363 		 * at a directory which contains files which should all
364 		 * have IXATTR set.  The IFATTRDIR case was handled in
365 		 * pass 2b.
366 		 *
367 		 * Note that the following code actually handles
368 		 * anything that's marked as an extended attribute but
369 		 * in a regular directory, not just files.
370 		 */
371 		if ((dirtype == IFDIR) && isattr) {
372 			fileerror(idesc->id_number, dirp->d_ino,
373 		    "%s I=%d should NOT be marked as extended attribute\n",
374 			    (inotype == IFDIR) ? "Directory" : "File",
375 			    dirp->d_ino);
376 			dp = ginode(dirp->d_ino);
377 			dp->di_cflags &= ~IXATTR;
378 			if ((n = reply("FIX")) == 1) {
379 				inodirty();
380 			} else {
381 				iscorrupt = 1;
382 			}
383 			if (n != 0)
384 				return (KEEPON | ALTERED);
385 		}
386 		break;
387 	default:
388 		errexit("PASS3: BAD STATE %d FOR INODE I=%d",
389 		    statemap[dirp->d_ino], dirp->d_ino);
390 		/* NOTREACHED */
391 	}
392 	if (n == 0)
393 		return (ret|KEEPON);
394 	return (ret|KEEPON|ALTERED);
395 }
396 
397 static void
398 setcurino(struct inodesc *idesc, struct dinode *dp, struct inoinfo *inp)
399 {
400 	(void) memmove((void *)&dp->di_db[0], (void *)&inp->i_blks[0],
401 		inp->i_blkssize);
402 
403 	init_inodesc(idesc);
404 	idesc->id_number = inp->i_number;
405 	idesc->id_parent = inp->i_parent;
406 	idesc->id_fix = DONTKNOW;
407 	idesc->id_type = DATA;
408 	idesc->id_func = pass3acheck;
409 }
410 
411 void
412 maybe_convert_attrdir_to_dir(fsck_ino_t orphan)
413 {
414 	struct dinode *dp = ginode(orphan);
415 	struct inoinfo *inp = getinoinfo(orphan);
416 	struct inodesc idesc;
417 
418 	if (dp->di_cflags & IXATTR) {
419 		dp->di_cflags &= ~IXATTR;
420 		inodirty();
421 	}
422 
423 	if ((dp->di_mode & IFMT) == IFATTRDIR) {
424 		dp->di_mode &= ~IFATTRDIR;
425 		dp->di_mode |= IFDIR;
426 		inodirty();
427 
428 		setcurino(&idesc, dp, inp);
429 		idesc.id_fix = FIX;
430 		idesc.id_filesize = dp->di_size;
431 		(void) ckinode(dp, &idesc, CKI_TRAVERSE);
432 	}
433 }
434