1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * dcache.c
4  *
5  * dentry cache handling code
6  *
7  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
8  */
9 
10 #include <linux/fs.h>
11 #include <linux/types.h>
12 #include <linux/slab.h>
13 #include <linux/namei.h>
14 
15 #include <cluster/masklog.h>
16 
17 #include "ocfs2.h"
18 
19 #include "alloc.h"
20 #include "dcache.h"
21 #include "dlmglue.h"
22 #include "file.h"
23 #include "inode.h"
24 #include "ocfs2_trace.h"
25 
ocfs2_dentry_attach_gen(struct dentry * dentry)26 void ocfs2_dentry_attach_gen(struct dentry *dentry)
27 {
28 	unsigned long gen =
29 		OCFS2_I(d_inode(dentry->d_parent))->ip_dir_lock_gen;
30 	BUG_ON(d_inode(dentry));
31 	dentry->d_fsdata = (void *)gen;
32 }
33 
34 
ocfs2_dentry_revalidate(struct dentry * dentry,unsigned int flags)35 static int ocfs2_dentry_revalidate(struct dentry *dentry, unsigned int flags)
36 {
37 	struct inode *inode;
38 	int ret = 0;    /* if all else fails, just return false */
39 	struct ocfs2_super *osb;
40 
41 	if (flags & LOOKUP_RCU)
42 		return -ECHILD;
43 
44 	inode = d_inode(dentry);
45 	osb = OCFS2_SB(dentry->d_sb);
46 
47 	trace_ocfs2_dentry_revalidate(dentry, dentry->d_name.len,
48 				      dentry->d_name.name);
49 
50 	/* For a negative dentry -
51 	 * check the generation number of the parent and compare with the
52 	 * one stored in the inode.
53 	 */
54 	if (inode == NULL) {
55 		unsigned long gen = (unsigned long) dentry->d_fsdata;
56 		unsigned long pgen;
57 		spin_lock(&dentry->d_lock);
58 		pgen = OCFS2_I(d_inode(dentry->d_parent))->ip_dir_lock_gen;
59 		spin_unlock(&dentry->d_lock);
60 		trace_ocfs2_dentry_revalidate_negative(dentry->d_name.len,
61 						       dentry->d_name.name,
62 						       pgen, gen);
63 		if (gen != pgen)
64 			goto bail;
65 		goto valid;
66 	}
67 
68 	BUG_ON(!osb);
69 
70 	if (inode == osb->root_inode || is_bad_inode(inode))
71 		goto bail;
72 
73 	spin_lock(&OCFS2_I(inode)->ip_lock);
74 	/* did we or someone else delete this inode? */
75 	if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
76 		spin_unlock(&OCFS2_I(inode)->ip_lock);
77 		trace_ocfs2_dentry_revalidate_delete(
78 				(unsigned long long)OCFS2_I(inode)->ip_blkno);
79 		goto bail;
80 	}
81 	spin_unlock(&OCFS2_I(inode)->ip_lock);
82 
83 	/*
84 	 * We don't need a cluster lock to test this because once an
85 	 * inode nlink hits zero, it never goes back.
86 	 */
87 	if (inode->i_nlink == 0) {
88 		trace_ocfs2_dentry_revalidate_orphaned(
89 			(unsigned long long)OCFS2_I(inode)->ip_blkno,
90 			S_ISDIR(inode->i_mode));
91 		goto bail;
92 	}
93 
94 	/*
95 	 * If the last lookup failed to create dentry lock, let us
96 	 * redo it.
97 	 */
98 	if (!dentry->d_fsdata) {
99 		trace_ocfs2_dentry_revalidate_nofsdata(
100 				(unsigned long long)OCFS2_I(inode)->ip_blkno);
101 		goto bail;
102 	}
103 
104 valid:
105 	ret = 1;
106 
107 bail:
108 	trace_ocfs2_dentry_revalidate_ret(ret);
109 	return ret;
110 }
111 
ocfs2_match_dentry(struct dentry * dentry,u64 parent_blkno,int skip_unhashed)112 static int ocfs2_match_dentry(struct dentry *dentry,
113 			      u64 parent_blkno,
114 			      int skip_unhashed)
115 {
116 	struct inode *parent;
117 
118 	/*
119 	 * ocfs2_lookup() does a d_splice_alias() _before_ attaching
120 	 * to the lock data, so we skip those here, otherwise
121 	 * ocfs2_dentry_attach_lock() will get its original dentry
122 	 * back.
123 	 */
124 	if (!dentry->d_fsdata)
125 		return 0;
126 
127 	if (!dentry->d_parent)
128 		return 0;
129 
130 	if (skip_unhashed && d_unhashed(dentry))
131 		return 0;
132 
133 	parent = d_inode(dentry->d_parent);
134 	/* Negative parent dentry? */
135 	if (!parent)
136 		return 0;
137 
138 	/* Name is in a different directory. */
139 	if (OCFS2_I(parent)->ip_blkno != parent_blkno)
140 		return 0;
141 
142 	return 1;
143 }
144 
145 /*
146  * Walk the inode alias list, and find a dentry which has a given
147  * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it
148  * is looking for a dentry_lock reference. The downconvert thread is
149  * looking to unhash aliases, so we allow it to skip any that already
150  * have that property.
151  */
ocfs2_find_local_alias(struct inode * inode,u64 parent_blkno,int skip_unhashed)152 struct dentry *ocfs2_find_local_alias(struct inode *inode,
153 				      u64 parent_blkno,
154 				      int skip_unhashed)
155 {
156 	struct dentry *dentry;
157 
158 	spin_lock(&inode->i_lock);
159 	hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
160 		spin_lock(&dentry->d_lock);
161 		if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) {
162 			trace_ocfs2_find_local_alias(dentry->d_name.len,
163 						     dentry->d_name.name);
164 
165 			dget_dlock(dentry);
166 			spin_unlock(&dentry->d_lock);
167 			spin_unlock(&inode->i_lock);
168 			return dentry;
169 		}
170 		spin_unlock(&dentry->d_lock);
171 	}
172 	spin_unlock(&inode->i_lock);
173 	return NULL;
174 }
175 
176 DEFINE_SPINLOCK(dentry_attach_lock);
177 
178 /*
179  * Attach this dentry to a cluster lock.
180  *
181  * Dentry locks cover all links in a given directory to a particular
182  * inode. We do this so that ocfs2 can build a lock name which all
183  * nodes in the cluster can agree on at all times. Shoving full names
184  * in the cluster lock won't work due to size restrictions. Covering
185  * links inside of a directory is a good compromise because it still
186  * allows us to use the parent directory lock to synchronize
187  * operations.
188  *
189  * Call this function with the parent dir semaphore and the parent dir
190  * cluster lock held.
191  *
192  * The dir semaphore will protect us from having to worry about
193  * concurrent processes on our node trying to attach a lock at the
194  * same time.
195  *
196  * The dir cluster lock (held at either PR or EX mode) protects us
197  * from unlink and rename on other nodes.
198  *
199  * A dput() can happen asynchronously due to pruning, so we cover
200  * attaching and detaching the dentry lock with a
201  * dentry_attach_lock.
202  *
203  * A node which has done lookup on a name retains a protected read
204  * lock until final dput. If the user requests and unlink or rename,
205  * the protected read is upgraded to an exclusive lock. Other nodes
206  * who have seen the dentry will then be informed that they need to
207  * downgrade their lock, which will involve d_delete on the
208  * dentry. This happens in ocfs2_dentry_convert_worker().
209  */
ocfs2_dentry_attach_lock(struct dentry * dentry,struct inode * inode,u64 parent_blkno)210 int ocfs2_dentry_attach_lock(struct dentry *dentry,
211 			     struct inode *inode,
212 			     u64 parent_blkno)
213 {
214 	int ret;
215 	struct dentry *alias;
216 	struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
217 
218 	trace_ocfs2_dentry_attach_lock(dentry->d_name.len, dentry->d_name.name,
219 				       (unsigned long long)parent_blkno, dl);
220 
221 	/*
222 	 * Negative dentry. We ignore these for now.
223 	 *
224 	 * XXX: Could we can improve ocfs2_dentry_revalidate() by
225 	 * tracking these?
226 	 */
227 	if (!inode)
228 		return 0;
229 
230 	if (d_really_is_negative(dentry) && dentry->d_fsdata) {
231 		/* Converting a negative dentry to positive
232 		   Clear dentry->d_fsdata */
233 		dentry->d_fsdata = dl = NULL;
234 	}
235 
236 	if (dl) {
237 		mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
238 				" \"%pd\": old parent: %llu, new: %llu\n",
239 				dentry,
240 				(unsigned long long)parent_blkno,
241 				(unsigned long long)dl->dl_parent_blkno);
242 		return 0;
243 	}
244 
245 	alias = ocfs2_find_local_alias(inode, parent_blkno, 0);
246 	if (alias) {
247 		/*
248 		 * Great, an alias exists, which means we must have a
249 		 * dentry lock already. We can just grab the lock off
250 		 * the alias and add it to the list.
251 		 *
252 		 * We're depending here on the fact that this dentry
253 		 * was found and exists in the dcache and so must have
254 		 * a reference to the dentry_lock because we can't
255 		 * race creates. Final dput() cannot happen on it
256 		 * since we have it pinned, so our reference is safe.
257 		 */
258 		dl = alias->d_fsdata;
259 		mlog_bug_on_msg(!dl, "parent %llu, ino %llu\n",
260 				(unsigned long long)parent_blkno,
261 				(unsigned long long)OCFS2_I(inode)->ip_blkno);
262 
263 		mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
264 				" \"%pd\": old parent: %llu, new: %llu\n",
265 				dentry,
266 				(unsigned long long)parent_blkno,
267 				(unsigned long long)dl->dl_parent_blkno);
268 
269 		trace_ocfs2_dentry_attach_lock_found(dl->dl_lockres.l_name,
270 				(unsigned long long)parent_blkno,
271 				(unsigned long long)OCFS2_I(inode)->ip_blkno);
272 
273 		goto out_attach;
274 	}
275 
276 	/*
277 	 * There are no other aliases
278 	 */
279 	dl = kmalloc(sizeof(*dl), GFP_NOFS);
280 	if (!dl) {
281 		ret = -ENOMEM;
282 		mlog_errno(ret);
283 		return ret;
284 	}
285 
286 	dl->dl_count = 0;
287 	/*
288 	 * Does this have to happen below, for all attaches, in case
289 	 * the struct inode gets blown away by the downconvert thread?
290 	 */
291 	dl->dl_inode = igrab(inode);
292 	dl->dl_parent_blkno = parent_blkno;
293 	ocfs2_dentry_lock_res_init(dl, parent_blkno, inode);
294 
295 out_attach:
296 	spin_lock(&dentry_attach_lock);
297 	if (unlikely(dentry->d_fsdata && !alias)) {
298 		/* d_fsdata is set by a racing thread which is doing
299 		 * the same thing as this thread is doing. Leave the racing
300 		 * thread going ahead and we return here.
301 		 */
302 		spin_unlock(&dentry_attach_lock);
303 		iput(dl->dl_inode);
304 		ocfs2_lock_res_free(&dl->dl_lockres);
305 		kfree(dl);
306 		return 0;
307 	}
308 
309 	dentry->d_fsdata = dl;
310 	dl->dl_count++;
311 	spin_unlock(&dentry_attach_lock);
312 
313 	/*
314 	 * This actually gets us our PRMODE level lock. From now on,
315 	 * we'll have a notification if one of these names is
316 	 * destroyed on another node.
317 	 */
318 	ret = ocfs2_dentry_lock(dentry, 0);
319 	if (!ret)
320 		ocfs2_dentry_unlock(dentry, 0);
321 	else
322 		mlog_errno(ret);
323 
324 	/*
325 	 * In case of error, manually free the allocation and do the iput().
326 	 * We need to do this because error here means no d_instantiate(),
327 	 * which means iput() will not be called during dput(dentry).
328 	 */
329 	if (ret < 0 && !alias) {
330 		ocfs2_lock_res_free(&dl->dl_lockres);
331 		BUG_ON(dl->dl_count != 1);
332 		spin_lock(&dentry_attach_lock);
333 		dentry->d_fsdata = NULL;
334 		spin_unlock(&dentry_attach_lock);
335 		kfree(dl);
336 		iput(inode);
337 	}
338 
339 	dput(alias);
340 
341 	return ret;
342 }
343 
344 /*
345  * ocfs2_dentry_iput() and friends.
346  *
347  * At this point, our particular dentry is detached from the inodes
348  * alias list, so there's no way that the locking code can find it.
349  *
350  * The interesting stuff happens when we determine that our lock needs
351  * to go away because this is the last subdir alias in the
352  * system. This function needs to handle a couple things:
353  *
354  * 1) Synchronizing lock shutdown with the downconvert threads. This
355  *    is already handled for us via the lockres release drop function
356  *    called in ocfs2_release_dentry_lock()
357  *
358  * 2) A race may occur when we're doing our lock shutdown and
359  *    another process wants to create a new dentry lock. Right now we
360  *    let them race, which means that for a very short while, this
361  *    node might have two locks on a lock resource. This should be a
362  *    problem though because one of them is in the process of being
363  *    thrown out.
364  */
ocfs2_drop_dentry_lock(struct ocfs2_super * osb,struct ocfs2_dentry_lock * dl)365 static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb,
366 				   struct ocfs2_dentry_lock *dl)
367 {
368 	iput(dl->dl_inode);
369 	ocfs2_simple_drop_lockres(osb, &dl->dl_lockres);
370 	ocfs2_lock_res_free(&dl->dl_lockres);
371 	kfree(dl);
372 }
373 
ocfs2_dentry_lock_put(struct ocfs2_super * osb,struct ocfs2_dentry_lock * dl)374 void ocfs2_dentry_lock_put(struct ocfs2_super *osb,
375 			   struct ocfs2_dentry_lock *dl)
376 {
377 	int unlock = 0;
378 
379 	BUG_ON(dl->dl_count == 0);
380 
381 	spin_lock(&dentry_attach_lock);
382 	dl->dl_count--;
383 	unlock = !dl->dl_count;
384 	spin_unlock(&dentry_attach_lock);
385 
386 	if (unlock)
387 		ocfs2_drop_dentry_lock(osb, dl);
388 }
389 
ocfs2_dentry_iput(struct dentry * dentry,struct inode * inode)390 static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode)
391 {
392 	struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
393 
394 	if (!dl) {
395 		/*
396 		 * No dentry lock is ok if we're disconnected or
397 		 * unhashed.
398 		 */
399 		if (!(dentry->d_flags & DCACHE_DISCONNECTED) &&
400 		    !d_unhashed(dentry)) {
401 			unsigned long long ino = 0ULL;
402 			if (inode)
403 				ino = (unsigned long long)OCFS2_I(inode)->ip_blkno;
404 			mlog(ML_ERROR, "Dentry is missing cluster lock. "
405 			     "inode: %llu, d_flags: 0x%x, d_name: %pd\n",
406 			     ino, dentry->d_flags, dentry);
407 		}
408 
409 		goto out;
410 	}
411 
412 	mlog_bug_on_msg(dl->dl_count == 0, "dentry: %pd, count: %u\n",
413 			dentry, dl->dl_count);
414 
415 	ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl);
416 
417 out:
418 	iput(inode);
419 }
420 
421 /*
422  * d_move(), but keep the locks in sync.
423  *
424  * When we are done, "dentry" will have the parent dir and name of
425  * "target", which will be thrown away.
426  *
427  * We manually update the lock of "dentry" if need be.
428  *
429  * "target" doesn't have it's dentry lock touched - we allow the later
430  * dput() to handle this for us.
431  *
432  * This is called during ocfs2_rename(), while holding parent
433  * directory locks. The dentries have already been deleted on other
434  * nodes via ocfs2_remote_dentry_delete().
435  *
436  * Normally, the VFS handles the d_move() for the file system, after
437  * the ->rename() callback. OCFS2 wants to handle this internally, so
438  * the new lock can be created atomically with respect to the cluster.
439  */
ocfs2_dentry_move(struct dentry * dentry,struct dentry * target,struct inode * old_dir,struct inode * new_dir)440 void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target,
441 		       struct inode *old_dir, struct inode *new_dir)
442 {
443 	int ret;
444 	struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb);
445 	struct inode *inode = d_inode(dentry);
446 
447 	/*
448 	 * Move within the same directory, so the actual lock info won't
449 	 * change.
450 	 *
451 	 * XXX: Is there any advantage to dropping the lock here?
452 	 */
453 	if (old_dir == new_dir)
454 		goto out_move;
455 
456 	ocfs2_dentry_lock_put(osb, dentry->d_fsdata);
457 
458 	dentry->d_fsdata = NULL;
459 	ret = ocfs2_dentry_attach_lock(dentry, inode, OCFS2_I(new_dir)->ip_blkno);
460 	if (ret)
461 		mlog_errno(ret);
462 
463 out_move:
464 	d_move(dentry, target);
465 }
466 
467 const struct dentry_operations ocfs2_dentry_ops = {
468 	.d_revalidate		= ocfs2_dentry_revalidate,
469 	.d_iput			= ocfs2_dentry_iput,
470 };
471