xref: /openbmc/linux/fs/xfs/scrub/parent.c (revision d9fd5a71)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Copyright (C) 2017 Oracle.  All Rights Reserved.
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_trans_resv.h"
11 #include "xfs_mount.h"
12 #include "xfs_log_format.h"
13 #include "xfs_inode.h"
14 #include "xfs_icache.h"
15 #include "xfs_dir2.h"
16 #include "xfs_dir2_priv.h"
17 #include "scrub/scrub.h"
18 #include "scrub/common.h"
19 
20 /* Set us up to scrub parents. */
21 int
22 xchk_setup_parent(
23 	struct xfs_scrub	*sc,
24 	struct xfs_inode	*ip)
25 {
26 	return xchk_setup_inode_contents(sc, ip, 0);
27 }
28 
29 /* Parent pointers */
30 
31 /* Look for an entry in a parent pointing to this inode. */
32 
33 struct xchk_parent_ctx {
34 	struct dir_context	dc;
35 	struct xfs_scrub	*sc;
36 	xfs_ino_t		ino;
37 	xfs_nlink_t		nlink;
38 	bool			cancelled;
39 };
40 
41 /* Look for a single entry in a directory pointing to an inode. */
42 STATIC int
43 xchk_parent_actor(
44 	struct dir_context	*dc,
45 	const char		*name,
46 	int			namelen,
47 	loff_t			pos,
48 	u64			ino,
49 	unsigned		type)
50 {
51 	struct xchk_parent_ctx	*spc;
52 	int			error = 0;
53 
54 	spc = container_of(dc, struct xchk_parent_ctx, dc);
55 	if (spc->ino == ino)
56 		spc->nlink++;
57 
58 	/*
59 	 * If we're facing a fatal signal, bail out.  Store the cancellation
60 	 * status separately because the VFS readdir code squashes error codes
61 	 * into short directory reads.
62 	 */
63 	if (xchk_should_terminate(spc->sc, &error))
64 		spc->cancelled = true;
65 
66 	return error;
67 }
68 
69 /* Count the number of dentries in the parent dir that point to this inode. */
70 STATIC int
71 xchk_parent_count_parent_dentries(
72 	struct xfs_scrub	*sc,
73 	struct xfs_inode	*parent,
74 	xfs_nlink_t		*nlink)
75 {
76 	struct xchk_parent_ctx	spc = {
77 		.dc.actor	= xchk_parent_actor,
78 		.ino		= sc->ip->i_ino,
79 		.sc		= sc,
80 	};
81 	size_t			bufsize;
82 	loff_t			oldpos;
83 	uint			lock_mode;
84 	int			error = 0;
85 
86 	/*
87 	 * If there are any blocks, read-ahead block 0 as we're almost
88 	 * certain to have the next operation be a read there.  This is
89 	 * how we guarantee that the parent's extent map has been loaded,
90 	 * if there is one.
91 	 */
92 	lock_mode = xfs_ilock_data_map_shared(parent);
93 	if (parent->i_df.if_nextents > 0)
94 		error = xfs_dir3_data_readahead(parent, 0, 0);
95 	xfs_iunlock(parent, lock_mode);
96 	if (error)
97 		return error;
98 
99 	/*
100 	 * Iterate the parent dir to confirm that there is
101 	 * exactly one entry pointing back to the inode being
102 	 * scanned.
103 	 */
104 	bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE,
105 			parent->i_d.di_size);
106 	oldpos = 0;
107 	while (true) {
108 		error = xfs_readdir(sc->tp, parent, &spc.dc, bufsize);
109 		if (error)
110 			goto out;
111 		if (spc.cancelled) {
112 			error = -EAGAIN;
113 			goto out;
114 		}
115 		if (oldpos == spc.dc.pos)
116 			break;
117 		oldpos = spc.dc.pos;
118 	}
119 	*nlink = spc.nlink;
120 out:
121 	return error;
122 }
123 
124 /*
125  * Given the inode number of the alleged parent of the inode being
126  * scrubbed, try to validate that the parent has exactly one directory
127  * entry pointing back to the inode being scrubbed.
128  */
129 STATIC int
130 xchk_parent_validate(
131 	struct xfs_scrub	*sc,
132 	xfs_ino_t		dnum,
133 	bool			*try_again)
134 {
135 	struct xfs_mount	*mp = sc->mp;
136 	struct xfs_inode	*dp = NULL;
137 	xfs_nlink_t		expected_nlink;
138 	xfs_nlink_t		nlink;
139 	int			error = 0;
140 
141 	*try_again = false;
142 
143 	if (sc->sm->sm_flags & XFS_SCRUB_OFLAG_CORRUPT)
144 		goto out;
145 
146 	/* '..' must not point to ourselves. */
147 	if (sc->ip->i_ino == dnum) {
148 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
149 		goto out;
150 	}
151 
152 	/*
153 	 * If we're an unlinked directory, the parent /won't/ have a link
154 	 * to us.  Otherwise, it should have one link.
155 	 */
156 	expected_nlink = VFS_I(sc->ip)->i_nlink == 0 ? 0 : 1;
157 
158 	/*
159 	 * Grab this parent inode.  We release the inode before we
160 	 * cancel the scrub transaction.  Since we're don't know a
161 	 * priori that releasing the inode won't trigger eofblocks
162 	 * cleanup (which allocates what would be a nested transaction)
163 	 * if the parent pointer erroneously points to a file, we
164 	 * can't use DONTCACHE here because DONTCACHE inodes can trigger
165 	 * immediate inactive cleanup of the inode.
166 	 *
167 	 * If _iget returns -EINVAL or -ENOENT then the parent inode number is
168 	 * garbage and the directory is corrupt.  If the _iget returns
169 	 * -EFSCORRUPTED or -EFSBADCRC then the parent is corrupt which is a
170 	 *  cross referencing error.  Any other error is an operational error.
171 	 */
172 	error = xfs_iget(mp, sc->tp, dnum, XFS_IGET_UNTRUSTED, 0, &dp);
173 	if (error == -EINVAL || error == -ENOENT) {
174 		error = -EFSCORRUPTED;
175 		xchk_fblock_process_error(sc, XFS_DATA_FORK, 0, &error);
176 		goto out;
177 	}
178 	if (!xchk_fblock_xref_process_error(sc, XFS_DATA_FORK, 0, &error))
179 		goto out;
180 	if (dp == sc->ip || !S_ISDIR(VFS_I(dp)->i_mode)) {
181 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
182 		goto out_rele;
183 	}
184 
185 	/*
186 	 * We prefer to keep the inode locked while we lock and search
187 	 * its alleged parent for a forward reference.  If we can grab
188 	 * the iolock, validate the pointers and we're done.  We must
189 	 * use nowait here to avoid an ABBA deadlock on the parent and
190 	 * the child inodes.
191 	 */
192 	if (xfs_ilock_nowait(dp, XFS_IOLOCK_SHARED)) {
193 		error = xchk_parent_count_parent_dentries(sc, dp, &nlink);
194 		if (!xchk_fblock_xref_process_error(sc, XFS_DATA_FORK, 0,
195 				&error))
196 			goto out_unlock;
197 		if (nlink != expected_nlink)
198 			xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
199 		goto out_unlock;
200 	}
201 
202 	/*
203 	 * The game changes if we get here.  We failed to lock the parent,
204 	 * so we're going to try to verify both pointers while only holding
205 	 * one lock so as to avoid deadlocking with something that's actually
206 	 * trying to traverse down the directory tree.
207 	 */
208 	xfs_iunlock(sc->ip, sc->ilock_flags);
209 	sc->ilock_flags = 0;
210 	error = xchk_ilock_inverted(dp, XFS_IOLOCK_SHARED);
211 	if (error)
212 		goto out_rele;
213 
214 	/* Go looking for our dentry. */
215 	error = xchk_parent_count_parent_dentries(sc, dp, &nlink);
216 	if (!xchk_fblock_xref_process_error(sc, XFS_DATA_FORK, 0, &error))
217 		goto out_unlock;
218 
219 	/* Drop the parent lock, relock this inode. */
220 	xfs_iunlock(dp, XFS_IOLOCK_SHARED);
221 	error = xchk_ilock_inverted(sc->ip, XFS_IOLOCK_EXCL);
222 	if (error)
223 		goto out_rele;
224 	sc->ilock_flags = XFS_IOLOCK_EXCL;
225 
226 	/*
227 	 * If we're an unlinked directory, the parent /won't/ have a link
228 	 * to us.  Otherwise, it should have one link.  We have to re-set
229 	 * it here because we dropped the lock on sc->ip.
230 	 */
231 	expected_nlink = VFS_I(sc->ip)->i_nlink == 0 ? 0 : 1;
232 
233 	/* Look up '..' to see if the inode changed. */
234 	error = xfs_dir_lookup(sc->tp, sc->ip, &xfs_name_dotdot, &dnum, NULL);
235 	if (!xchk_fblock_process_error(sc, XFS_DATA_FORK, 0, &error))
236 		goto out_rele;
237 
238 	/* Drat, parent changed.  Try again! */
239 	if (dnum != dp->i_ino) {
240 		xfs_irele(dp);
241 		*try_again = true;
242 		return 0;
243 	}
244 	xfs_irele(dp);
245 
246 	/*
247 	 * '..' didn't change, so check that there was only one entry
248 	 * for us in the parent.
249 	 */
250 	if (nlink != expected_nlink)
251 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
252 	return error;
253 
254 out_unlock:
255 	xfs_iunlock(dp, XFS_IOLOCK_SHARED);
256 out_rele:
257 	xfs_irele(dp);
258 out:
259 	return error;
260 }
261 
262 /* Scrub a parent pointer. */
263 int
264 xchk_parent(
265 	struct xfs_scrub	*sc)
266 {
267 	struct xfs_mount	*mp = sc->mp;
268 	xfs_ino_t		dnum;
269 	bool			try_again;
270 	int			tries = 0;
271 	int			error = 0;
272 
273 	/*
274 	 * If we're a directory, check that the '..' link points up to
275 	 * a directory that has one entry pointing to us.
276 	 */
277 	if (!S_ISDIR(VFS_I(sc->ip)->i_mode))
278 		return -ENOENT;
279 
280 	/* We're not a special inode, are we? */
281 	if (!xfs_verify_dir_ino(mp, sc->ip->i_ino)) {
282 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
283 		goto out;
284 	}
285 
286 	/*
287 	 * The VFS grabs a read or write lock via i_rwsem before it reads
288 	 * or writes to a directory.  If we've gotten this far we've
289 	 * already obtained IOLOCK_EXCL, which (since 4.10) is the same as
290 	 * getting a write lock on i_rwsem.  Therefore, it is safe for us
291 	 * to drop the ILOCK here in order to do directory lookups.
292 	 */
293 	sc->ilock_flags &= ~(XFS_ILOCK_EXCL | XFS_MMAPLOCK_EXCL);
294 	xfs_iunlock(sc->ip, XFS_ILOCK_EXCL | XFS_MMAPLOCK_EXCL);
295 
296 	/* Look up '..' */
297 	error = xfs_dir_lookup(sc->tp, sc->ip, &xfs_name_dotdot, &dnum, NULL);
298 	if (!xchk_fblock_process_error(sc, XFS_DATA_FORK, 0, &error))
299 		goto out;
300 	if (!xfs_verify_dir_ino(mp, dnum)) {
301 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
302 		goto out;
303 	}
304 
305 	/* Is this the root dir?  Then '..' must point to itself. */
306 	if (sc->ip == mp->m_rootip) {
307 		if (sc->ip->i_ino != mp->m_sb.sb_rootino ||
308 		    sc->ip->i_ino != dnum)
309 			xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
310 		goto out;
311 	}
312 
313 	do {
314 		error = xchk_parent_validate(sc, dnum, &try_again);
315 		if (error)
316 			goto out;
317 	} while (try_again && ++tries < 20);
318 
319 	/*
320 	 * We gave it our best shot but failed, so mark this scrub
321 	 * incomplete.  Userspace can decide if it wants to try again.
322 	 */
323 	if (try_again && tries == 20)
324 		xchk_set_incomplete(sc);
325 out:
326 	/*
327 	 * If we failed to lock the parent inode even after a retry, just mark
328 	 * this scrub incomplete and return.
329 	 */
330 	if ((sc->flags & XCHK_TRY_HARDER) && error == -EDEADLOCK) {
331 		error = 0;
332 		xchk_set_incomplete(sc);
333 	}
334 	return error;
335 }
336