xref: /openbmc/linux/fs/xfs/scrub/parent.c (revision 5e012745)
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 	xfs_ino_t		ino;
36 	xfs_nlink_t		nlink;
37 };
38 
39 /* Look for a single entry in a directory pointing to an inode. */
40 STATIC int
41 xchk_parent_actor(
42 	struct dir_context	*dc,
43 	const char		*name,
44 	int			namelen,
45 	loff_t			pos,
46 	u64			ino,
47 	unsigned		type)
48 {
49 	struct xchk_parent_ctx	*spc;
50 
51 	spc = container_of(dc, struct xchk_parent_ctx, dc);
52 	if (spc->ino == ino)
53 		spc->nlink++;
54 	return 0;
55 }
56 
57 /* Count the number of dentries in the parent dir that point to this inode. */
58 STATIC int
59 xchk_parent_count_parent_dentries(
60 	struct xfs_scrub	*sc,
61 	struct xfs_inode	*parent,
62 	xfs_nlink_t		*nlink)
63 {
64 	struct xchk_parent_ctx	spc = {
65 		.dc.actor = xchk_parent_actor,
66 		.dc.pos = 0,
67 		.ino = sc->ip->i_ino,
68 		.nlink = 0,
69 	};
70 	size_t			bufsize;
71 	loff_t			oldpos;
72 	uint			lock_mode;
73 	int			error = 0;
74 
75 	/*
76 	 * If there are any blocks, read-ahead block 0 as we're almost
77 	 * certain to have the next operation be a read there.  This is
78 	 * how we guarantee that the parent's extent map has been loaded,
79 	 * if there is one.
80 	 */
81 	lock_mode = xfs_ilock_data_map_shared(parent);
82 	if (parent->i_d.di_nextents > 0)
83 		error = xfs_dir3_data_readahead(parent, 0, -1);
84 	xfs_iunlock(parent, lock_mode);
85 	if (error)
86 		return error;
87 
88 	/*
89 	 * Iterate the parent dir to confirm that there is
90 	 * exactly one entry pointing back to the inode being
91 	 * scanned.
92 	 */
93 	bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE,
94 			parent->i_d.di_size);
95 	oldpos = 0;
96 	while (true) {
97 		error = xfs_readdir(sc->tp, parent, &spc.dc, bufsize);
98 		if (error)
99 			goto out;
100 		if (oldpos == spc.dc.pos)
101 			break;
102 		oldpos = spc.dc.pos;
103 	}
104 	*nlink = spc.nlink;
105 out:
106 	return error;
107 }
108 
109 /*
110  * Given the inode number of the alleged parent of the inode being
111  * scrubbed, try to validate that the parent has exactly one directory
112  * entry pointing back to the inode being scrubbed.
113  */
114 STATIC int
115 xchk_parent_validate(
116 	struct xfs_scrub	*sc,
117 	xfs_ino_t		dnum,
118 	bool			*try_again)
119 {
120 	struct xfs_mount	*mp = sc->mp;
121 	struct xfs_inode	*dp = NULL;
122 	xfs_nlink_t		expected_nlink;
123 	xfs_nlink_t		nlink;
124 	int			error = 0;
125 
126 	*try_again = false;
127 
128 	if (sc->sm->sm_flags & XFS_SCRUB_OFLAG_CORRUPT)
129 		goto out;
130 
131 	/* '..' must not point to ourselves. */
132 	if (sc->ip->i_ino == dnum) {
133 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
134 		goto out;
135 	}
136 
137 	/*
138 	 * If we're an unlinked directory, the parent /won't/ have a link
139 	 * to us.  Otherwise, it should have one link.
140 	 */
141 	expected_nlink = VFS_I(sc->ip)->i_nlink == 0 ? 0 : 1;
142 
143 	/*
144 	 * Grab this parent inode.  We release the inode before we
145 	 * cancel the scrub transaction.  Since we're don't know a
146 	 * priori that releasing the inode won't trigger eofblocks
147 	 * cleanup (which allocates what would be a nested transaction)
148 	 * if the parent pointer erroneously points to a file, we
149 	 * can't use DONTCACHE here because DONTCACHE inodes can trigger
150 	 * immediate inactive cleanup of the inode.
151 	 *
152 	 * If _iget returns -EINVAL then the parent inode number is garbage
153 	 * and the directory is corrupt.  If the _iget returns -EFSCORRUPTED
154 	 * or -EFSBADCRC then the parent is corrupt which is a cross
155 	 * referencing error.  Any other error is an operational error.
156 	 */
157 	error = xfs_iget(mp, sc->tp, dnum, XFS_IGET_UNTRUSTED, 0, &dp);
158 	if (error == -EINVAL) {
159 		error = -EFSCORRUPTED;
160 		xchk_fblock_process_error(sc, XFS_DATA_FORK, 0, &error);
161 		goto out;
162 	}
163 	if (!xchk_fblock_xref_process_error(sc, XFS_DATA_FORK, 0, &error))
164 		goto out;
165 	if (dp == sc->ip || !S_ISDIR(VFS_I(dp)->i_mode)) {
166 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
167 		goto out_rele;
168 	}
169 
170 	/*
171 	 * We prefer to keep the inode locked while we lock and search
172 	 * its alleged parent for a forward reference.  If we can grab
173 	 * the iolock, validate the pointers and we're done.  We must
174 	 * use nowait here to avoid an ABBA deadlock on the parent and
175 	 * the child inodes.
176 	 */
177 	if (xfs_ilock_nowait(dp, XFS_IOLOCK_SHARED)) {
178 		error = xchk_parent_count_parent_dentries(sc, dp, &nlink);
179 		if (!xchk_fblock_xref_process_error(sc, XFS_DATA_FORK, 0,
180 				&error))
181 			goto out_unlock;
182 		if (nlink != expected_nlink)
183 			xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
184 		goto out_unlock;
185 	}
186 
187 	/*
188 	 * The game changes if we get here.  We failed to lock the parent,
189 	 * so we're going to try to verify both pointers while only holding
190 	 * one lock so as to avoid deadlocking with something that's actually
191 	 * trying to traverse down the directory tree.
192 	 */
193 	xfs_iunlock(sc->ip, sc->ilock_flags);
194 	sc->ilock_flags = 0;
195 	error = xchk_ilock_inverted(dp, XFS_IOLOCK_SHARED);
196 	if (error)
197 		goto out_rele;
198 
199 	/* Go looking for our dentry. */
200 	error = xchk_parent_count_parent_dentries(sc, dp, &nlink);
201 	if (!xchk_fblock_xref_process_error(sc, XFS_DATA_FORK, 0, &error))
202 		goto out_unlock;
203 
204 	/* Drop the parent lock, relock this inode. */
205 	xfs_iunlock(dp, XFS_IOLOCK_SHARED);
206 	error = xchk_ilock_inverted(sc->ip, XFS_IOLOCK_EXCL);
207 	if (error)
208 		goto out_rele;
209 	sc->ilock_flags = XFS_IOLOCK_EXCL;
210 
211 	/*
212 	 * If we're an unlinked directory, the parent /won't/ have a link
213 	 * to us.  Otherwise, it should have one link.  We have to re-set
214 	 * it here because we dropped the lock on sc->ip.
215 	 */
216 	expected_nlink = VFS_I(sc->ip)->i_nlink == 0 ? 0 : 1;
217 
218 	/* Look up '..' to see if the inode changed. */
219 	error = xfs_dir_lookup(sc->tp, sc->ip, &xfs_name_dotdot, &dnum, NULL);
220 	if (!xchk_fblock_process_error(sc, XFS_DATA_FORK, 0, &error))
221 		goto out_rele;
222 
223 	/* Drat, parent changed.  Try again! */
224 	if (dnum != dp->i_ino) {
225 		xfs_irele(dp);
226 		*try_again = true;
227 		return 0;
228 	}
229 	xfs_irele(dp);
230 
231 	/*
232 	 * '..' didn't change, so check that there was only one entry
233 	 * for us in the parent.
234 	 */
235 	if (nlink != expected_nlink)
236 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
237 	return error;
238 
239 out_unlock:
240 	xfs_iunlock(dp, XFS_IOLOCK_SHARED);
241 out_rele:
242 	xfs_irele(dp);
243 out:
244 	return error;
245 }
246 
247 /* Scrub a parent pointer. */
248 int
249 xchk_parent(
250 	struct xfs_scrub	*sc)
251 {
252 	struct xfs_mount	*mp = sc->mp;
253 	xfs_ino_t		dnum;
254 	bool			try_again;
255 	int			tries = 0;
256 	int			error = 0;
257 
258 	/*
259 	 * If we're a directory, check that the '..' link points up to
260 	 * a directory that has one entry pointing to us.
261 	 */
262 	if (!S_ISDIR(VFS_I(sc->ip)->i_mode))
263 		return -ENOENT;
264 
265 	/* We're not a special inode, are we? */
266 	if (!xfs_verify_dir_ino(mp, sc->ip->i_ino)) {
267 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
268 		goto out;
269 	}
270 
271 	/*
272 	 * The VFS grabs a read or write lock via i_rwsem before it reads
273 	 * or writes to a directory.  If we've gotten this far we've
274 	 * already obtained IOLOCK_EXCL, which (since 4.10) is the same as
275 	 * getting a write lock on i_rwsem.  Therefore, it is safe for us
276 	 * to drop the ILOCK here in order to do directory lookups.
277 	 */
278 	sc->ilock_flags &= ~(XFS_ILOCK_EXCL | XFS_MMAPLOCK_EXCL);
279 	xfs_iunlock(sc->ip, XFS_ILOCK_EXCL | XFS_MMAPLOCK_EXCL);
280 
281 	/* Look up '..' */
282 	error = xfs_dir_lookup(sc->tp, sc->ip, &xfs_name_dotdot, &dnum, NULL);
283 	if (!xchk_fblock_process_error(sc, XFS_DATA_FORK, 0, &error))
284 		goto out;
285 	if (!xfs_verify_dir_ino(mp, dnum)) {
286 		xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
287 		goto out;
288 	}
289 
290 	/* Is this the root dir?  Then '..' must point to itself. */
291 	if (sc->ip == mp->m_rootip) {
292 		if (sc->ip->i_ino != mp->m_sb.sb_rootino ||
293 		    sc->ip->i_ino != dnum)
294 			xchk_fblock_set_corrupt(sc, XFS_DATA_FORK, 0);
295 		goto out;
296 	}
297 
298 	do {
299 		error = xchk_parent_validate(sc, dnum, &try_again);
300 		if (error)
301 			goto out;
302 	} while (try_again && ++tries < 20);
303 
304 	/*
305 	 * We gave it our best shot but failed, so mark this scrub
306 	 * incomplete.  Userspace can decide if it wants to try again.
307 	 */
308 	if (try_again && tries == 20)
309 		xchk_set_incomplete(sc);
310 out:
311 	/*
312 	 * If we failed to lock the parent inode even after a retry, just mark
313 	 * this scrub incomplete and return.
314 	 */
315 	if ((sc->flags & XCHK_TRY_HARDER) && error == -EDEADLOCK) {
316 		error = 0;
317 		xchk_set_incomplete(sc);
318 	}
319 	return error;
320 }
321