1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dcache.c
5 *
6 * dentry cache handling code
7 *
8 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26#include <linux/fs.h>
27#include <linux/types.h>
28#include <linux/slab.h>
29#include <linux/namei.h>
30
31#include <cluster/masklog.h>
32
33#include "ocfs2.h"
34
35#include "alloc.h"
36#include "dcache.h"
37#include "dlmglue.h"
38#include "file.h"
39#include "inode.h"
40#include "ocfs2_trace.h"
41
42void ocfs2_dentry_attach_gen(struct dentry *dentry)
43{
44	unsigned long gen =
45		OCFS2_I(d_inode(dentry->d_parent))->ip_dir_lock_gen;
46	BUG_ON(d_inode(dentry));
47	dentry->d_fsdata = (void *)gen;
48}
49
50
51static int ocfs2_dentry_revalidate(struct dentry *dentry, unsigned int flags)
52{
53	struct inode *inode;
54	int ret = 0;    /* if all else fails, just return false */
55	struct ocfs2_super *osb;
56
57	if (flags & LOOKUP_RCU)
58		return -ECHILD;
59
60	inode = d_inode(dentry);
61	osb = OCFS2_SB(dentry->d_sb);
62
63	trace_ocfs2_dentry_revalidate(dentry, dentry->d_name.len,
64				      dentry->d_name.name);
65
66	/* For a negative dentry -
67	 * check the generation number of the parent and compare with the
68	 * one stored in the inode.
69	 */
70	if (inode == NULL) {
71		unsigned long gen = (unsigned long) dentry->d_fsdata;
72		unsigned long pgen;
73		spin_lock(&dentry->d_lock);
74		pgen = OCFS2_I(d_inode(dentry->d_parent))->ip_dir_lock_gen;
75		spin_unlock(&dentry->d_lock);
76		trace_ocfs2_dentry_revalidate_negative(dentry->d_name.len,
77						       dentry->d_name.name,
78						       pgen, gen);
79		if (gen != pgen)
80			goto bail;
81		goto valid;
82	}
83
84	BUG_ON(!osb);
85
86	if (inode == osb->root_inode || is_bad_inode(inode))
87		goto bail;
88
89	spin_lock(&OCFS2_I(inode)->ip_lock);
90	/* did we or someone else delete this inode? */
91	if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
92		spin_unlock(&OCFS2_I(inode)->ip_lock);
93		trace_ocfs2_dentry_revalidate_delete(
94				(unsigned long long)OCFS2_I(inode)->ip_blkno);
95		goto bail;
96	}
97	spin_unlock(&OCFS2_I(inode)->ip_lock);
98
99	/*
100	 * We don't need a cluster lock to test this because once an
101	 * inode nlink hits zero, it never goes back.
102	 */
103	if (inode->i_nlink == 0) {
104		trace_ocfs2_dentry_revalidate_orphaned(
105			(unsigned long long)OCFS2_I(inode)->ip_blkno,
106			S_ISDIR(inode->i_mode));
107		goto bail;
108	}
109
110	/*
111	 * If the last lookup failed to create dentry lock, let us
112	 * redo it.
113	 */
114	if (!dentry->d_fsdata) {
115		trace_ocfs2_dentry_revalidate_nofsdata(
116				(unsigned long long)OCFS2_I(inode)->ip_blkno);
117		goto bail;
118	}
119
120valid:
121	ret = 1;
122
123bail:
124	trace_ocfs2_dentry_revalidate_ret(ret);
125	return ret;
126}
127
128static int ocfs2_match_dentry(struct dentry *dentry,
129			      u64 parent_blkno,
130			      int skip_unhashed)
131{
132	struct inode *parent;
133
134	/*
135	 * ocfs2_lookup() does a d_splice_alias() _before_ attaching
136	 * to the lock data, so we skip those here, otherwise
137	 * ocfs2_dentry_attach_lock() will get its original dentry
138	 * back.
139	 */
140	if (!dentry->d_fsdata)
141		return 0;
142
143	if (!dentry->d_parent)
144		return 0;
145
146	if (skip_unhashed && d_unhashed(dentry))
147		return 0;
148
149	parent = d_inode(dentry->d_parent);
150	/* Negative parent dentry? */
151	if (!parent)
152		return 0;
153
154	/* Name is in a different directory. */
155	if (OCFS2_I(parent)->ip_blkno != parent_blkno)
156		return 0;
157
158	return 1;
159}
160
161/*
162 * Walk the inode alias list, and find a dentry which has a given
163 * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it
164 * is looking for a dentry_lock reference. The downconvert thread is
165 * looking to unhash aliases, so we allow it to skip any that already
166 * have that property.
167 */
168struct dentry *ocfs2_find_local_alias(struct inode *inode,
169				      u64 parent_blkno,
170				      int skip_unhashed)
171{
172	struct dentry *dentry;
173
174	spin_lock(&inode->i_lock);
175	hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
176		spin_lock(&dentry->d_lock);
177		if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) {
178			trace_ocfs2_find_local_alias(dentry->d_name.len,
179						     dentry->d_name.name);
180
181			dget_dlock(dentry);
182			spin_unlock(&dentry->d_lock);
183			spin_unlock(&inode->i_lock);
184			return dentry;
185		}
186		spin_unlock(&dentry->d_lock);
187	}
188	spin_unlock(&inode->i_lock);
189	return NULL;
190}
191
192DEFINE_SPINLOCK(dentry_attach_lock);
193
194/*
195 * Attach this dentry to a cluster lock.
196 *
197 * Dentry locks cover all links in a given directory to a particular
198 * inode. We do this so that ocfs2 can build a lock name which all
199 * nodes in the cluster can agree on at all times. Shoving full names
200 * in the cluster lock won't work due to size restrictions. Covering
201 * links inside of a directory is a good compromise because it still
202 * allows us to use the parent directory lock to synchronize
203 * operations.
204 *
205 * Call this function with the parent dir semaphore and the parent dir
206 * cluster lock held.
207 *
208 * The dir semaphore will protect us from having to worry about
209 * concurrent processes on our node trying to attach a lock at the
210 * same time.
211 *
212 * The dir cluster lock (held at either PR or EX mode) protects us
213 * from unlink and rename on other nodes.
214 *
215 * A dput() can happen asynchronously due to pruning, so we cover
216 * attaching and detaching the dentry lock with a
217 * dentry_attach_lock.
218 *
219 * A node which has done lookup on a name retains a protected read
220 * lock until final dput. If the user requests and unlink or rename,
221 * the protected read is upgraded to an exclusive lock. Other nodes
222 * who have seen the dentry will then be informed that they need to
223 * downgrade their lock, which will involve d_delete on the
224 * dentry. This happens in ocfs2_dentry_convert_worker().
225 */
226int ocfs2_dentry_attach_lock(struct dentry *dentry,
227			     struct inode *inode,
228			     u64 parent_blkno)
229{
230	int ret;
231	struct dentry *alias;
232	struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
233
234	trace_ocfs2_dentry_attach_lock(dentry->d_name.len, dentry->d_name.name,
235				       (unsigned long long)parent_blkno, dl);
236
237	/*
238	 * Negative dentry. We ignore these for now.
239	 *
240	 * XXX: Could we can improve ocfs2_dentry_revalidate() by
241	 * tracking these?
242	 */
243	if (!inode)
244		return 0;
245
246	if (d_really_is_negative(dentry) && dentry->d_fsdata) {
247		/* Converting a negative dentry to positive
248		   Clear dentry->d_fsdata */
249		dentry->d_fsdata = dl = NULL;
250	}
251
252	if (dl) {
253		mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
254				" \"%pd\": old parent: %llu, new: %llu\n",
255				dentry,
256				(unsigned long long)parent_blkno,
257				(unsigned long long)dl->dl_parent_blkno);
258		return 0;
259	}
260
261	alias = ocfs2_find_local_alias(inode, parent_blkno, 0);
262	if (alias) {
263		/*
264		 * Great, an alias exists, which means we must have a
265		 * dentry lock already. We can just grab the lock off
266		 * the alias and add it to the list.
267		 *
268		 * We're depending here on the fact that this dentry
269		 * was found and exists in the dcache and so must have
270		 * a reference to the dentry_lock because we can't
271		 * race creates. Final dput() cannot happen on it
272		 * since we have it pinned, so our reference is safe.
273		 */
274		dl = alias->d_fsdata;
275		mlog_bug_on_msg(!dl, "parent %llu, ino %llu\n",
276				(unsigned long long)parent_blkno,
277				(unsigned long long)OCFS2_I(inode)->ip_blkno);
278
279		mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
280				" \"%pd\": old parent: %llu, new: %llu\n",
281				dentry,
282				(unsigned long long)parent_blkno,
283				(unsigned long long)dl->dl_parent_blkno);
284
285		trace_ocfs2_dentry_attach_lock_found(dl->dl_lockres.l_name,
286				(unsigned long long)parent_blkno,
287				(unsigned long long)OCFS2_I(inode)->ip_blkno);
288
289		goto out_attach;
290	}
291
292	/*
293	 * There are no other aliases
294	 */
295	dl = kmalloc(sizeof(*dl), GFP_NOFS);
296	if (!dl) {
297		ret = -ENOMEM;
298		mlog_errno(ret);
299		return ret;
300	}
301
302	dl->dl_count = 0;
303	/*
304	 * Does this have to happen below, for all attaches, in case
305	 * the struct inode gets blown away by the downconvert thread?
306	 */
307	dl->dl_inode = igrab(inode);
308	dl->dl_parent_blkno = parent_blkno;
309	ocfs2_dentry_lock_res_init(dl, parent_blkno, inode);
310
311out_attach:
312	spin_lock(&dentry_attach_lock);
313	dentry->d_fsdata = dl;
314	dl->dl_count++;
315	spin_unlock(&dentry_attach_lock);
316
317	/*
318	 * This actually gets us our PRMODE level lock. From now on,
319	 * we'll have a notification if one of these names is
320	 * destroyed on another node.
321	 */
322	ret = ocfs2_dentry_lock(dentry, 0);
323	if (!ret)
324		ocfs2_dentry_unlock(dentry, 0);
325	else
326		mlog_errno(ret);
327
328	/*
329	 * In case of error, manually free the allocation and do the iput().
330	 * We need to do this because error here means no d_instantiate(),
331	 * which means iput() will not be called during dput(dentry).
332	 */
333	if (ret < 0 && !alias) {
334		ocfs2_lock_res_free(&dl->dl_lockres);
335		BUG_ON(dl->dl_count != 1);
336		spin_lock(&dentry_attach_lock);
337		dentry->d_fsdata = NULL;
338		spin_unlock(&dentry_attach_lock);
339		kfree(dl);
340		iput(inode);
341	}
342
343	dput(alias);
344
345	return ret;
346}
347
348/*
349 * ocfs2_dentry_iput() and friends.
350 *
351 * At this point, our particular dentry is detached from the inodes
352 * alias list, so there's no way that the locking code can find it.
353 *
354 * The interesting stuff happens when we determine that our lock needs
355 * to go away because this is the last subdir alias in the
356 * system. This function needs to handle a couple things:
357 *
358 * 1) Synchronizing lock shutdown with the downconvert threads. This
359 *    is already handled for us via the lockres release drop function
360 *    called in ocfs2_release_dentry_lock()
361 *
362 * 2) A race may occur when we're doing our lock shutdown and
363 *    another process wants to create a new dentry lock. Right now we
364 *    let them race, which means that for a very short while, this
365 *    node might have two locks on a lock resource. This should be a
366 *    problem though because one of them is in the process of being
367 *    thrown out.
368 */
369static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb,
370				   struct ocfs2_dentry_lock *dl)
371{
372	iput(dl->dl_inode);
373	ocfs2_simple_drop_lockres(osb, &dl->dl_lockres);
374	ocfs2_lock_res_free(&dl->dl_lockres);
375	kfree(dl);
376}
377
378void ocfs2_dentry_lock_put(struct ocfs2_super *osb,
379			   struct ocfs2_dentry_lock *dl)
380{
381	int unlock = 0;
382
383	BUG_ON(dl->dl_count == 0);
384
385	spin_lock(&dentry_attach_lock);
386	dl->dl_count--;
387	unlock = !dl->dl_count;
388	spin_unlock(&dentry_attach_lock);
389
390	if (unlock)
391		ocfs2_drop_dentry_lock(osb, dl);
392}
393
394static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode)
395{
396	struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
397
398	if (!dl) {
399		/*
400		 * No dentry lock is ok if we're disconnected or
401		 * unhashed.
402		 */
403		if (!(dentry->d_flags & DCACHE_DISCONNECTED) &&
404		    !d_unhashed(dentry)) {
405			unsigned long long ino = 0ULL;
406			if (inode)
407				ino = (unsigned long long)OCFS2_I(inode)->ip_blkno;
408			mlog(ML_ERROR, "Dentry is missing cluster lock. "
409			     "inode: %llu, d_flags: 0x%x, d_name: %pd\n",
410			     ino, dentry->d_flags, dentry);
411		}
412
413		goto out;
414	}
415
416	mlog_bug_on_msg(dl->dl_count == 0, "dentry: %pd, count: %u\n",
417			dentry, dl->dl_count);
418
419	ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl);
420
421out:
422	iput(inode);
423}
424
425/*
426 * d_move(), but keep the locks in sync.
427 *
428 * When we are done, "dentry" will have the parent dir and name of
429 * "target", which will be thrown away.
430 *
431 * We manually update the lock of "dentry" if need be.
432 *
433 * "target" doesn't have it's dentry lock touched - we allow the later
434 * dput() to handle this for us.
435 *
436 * This is called during ocfs2_rename(), while holding parent
437 * directory locks. The dentries have already been deleted on other
438 * nodes via ocfs2_remote_dentry_delete().
439 *
440 * Normally, the VFS handles the d_move() for the file system, after
441 * the ->rename() callback. OCFS2 wants to handle this internally, so
442 * the new lock can be created atomically with respect to the cluster.
443 */
444void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target,
445		       struct inode *old_dir, struct inode *new_dir)
446{
447	int ret;
448	struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb);
449	struct inode *inode = d_inode(dentry);
450
451	/*
452	 * Move within the same directory, so the actual lock info won't
453	 * change.
454	 *
455	 * XXX: Is there any advantage to dropping the lock here?
456	 */
457	if (old_dir == new_dir)
458		goto out_move;
459
460	ocfs2_dentry_lock_put(osb, dentry->d_fsdata);
461
462	dentry->d_fsdata = NULL;
463	ret = ocfs2_dentry_attach_lock(dentry, inode, OCFS2_I(new_dir)->ip_blkno);
464	if (ret)
465		mlog_errno(ret);
466
467out_move:
468	d_move(dentry, target);
469}
470
471const struct dentry_operations ocfs2_dentry_ops = {
472	.d_revalidate		= ocfs2_dentry_revalidate,
473	.d_iput			= ocfs2_dentry_iput,
474};
475