1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2006, 2010, Oracle and/or its affiliates. All rights reserved.
23 */
24
25/*
26 * Routines to manage the on-disk persistent error log.
27 *
28 * Each pool stores a log of all logical data errors seen during normal
29 * operation.  This is actually the union of two distinct logs: the last log,
30 * and the current log.  All errors seen are logged to the current log.  When a
31 * scrub completes, the current log becomes the last log, the last log is thrown
32 * out, and the current log is reinitialized.  This way, if an error is somehow
33 * corrected, a new scrub will show that that it no longer exists, and will be
34 * deleted from the log when the scrub completes.
35 *
36 * The log is stored using a ZAP object whose key is a string form of the
37 * zbookmark tuple (objset, object, level, blkid), and whose contents is an
38 * optional 'objset:object' human-readable string describing the data.  When an
39 * error is first logged, this string will be empty, indicating that no name is
40 * known.  This prevents us from having to issue a potentially large amount of
41 * I/O to discover the object name during an error path.  Instead, we do the
42 * calculation when the data is requested, storing the result so future queries
43 * will be faster.
44 *
45 * This log is then shipped into an nvlist where the key is the dataset name and
46 * the value is the object name.  Userland is then responsible for uniquifying
47 * this list and displaying it to the user.
48 */
49
50#include <sys/dmu_tx.h>
51#include <sys/spa.h>
52#include <sys/spa_impl.h>
53#include <sys/zap.h>
54#include <sys/zio.h>
55
56
57/*
58 * Convert a bookmark to a string.
59 */
60static void
61bookmark_to_name(zbookmark_t *zb, char *buf, size_t len)
62{
63	(void) snprintf(buf, len, "%llx:%llx:%llx:%llx",
64	    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
65	    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid);
66}
67
68/*
69 * Convert a string to a bookmark
70 */
71#ifdef _KERNEL
72static void
73name_to_bookmark(char *buf, zbookmark_t *zb)
74{
75	zb->zb_objset = strtonum(buf, &buf);
76	ASSERT(*buf == ':');
77	zb->zb_object = strtonum(buf + 1, &buf);
78	ASSERT(*buf == ':');
79	zb->zb_level = (int)strtonum(buf + 1, &buf);
80	ASSERT(*buf == ':');
81	zb->zb_blkid = strtonum(buf + 1, &buf);
82	ASSERT(*buf == '\0');
83}
84#endif
85
86/*
87 * Log an uncorrectable error to the persistent error log.  We add it to the
88 * spa's list of pending errors.  The changes are actually synced out to disk
89 * during spa_errlog_sync().
90 */
91void
92spa_log_error(spa_t *spa, zio_t *zio)
93{
94	zbookmark_t *zb = &zio->io_logical->io_bookmark;
95	spa_error_entry_t search;
96	spa_error_entry_t *new;
97	avl_tree_t *tree;
98	avl_index_t where;
99
100	/*
101	 * If we are trying to import a pool, ignore any errors, as we won't be
102	 * writing to the pool any time soon.
103	 */
104	if (spa_load_state(spa) == SPA_LOAD_TRYIMPORT)
105		return;
106
107	mutex_enter(&spa->spa_errlist_lock);
108
109	/*
110	 * If we have had a request to rotate the log, log it to the next list
111	 * instead of the current one.
112	 */
113	if (spa->spa_scrub_active || spa->spa_scrub_finished)
114		tree = &spa->spa_errlist_scrub;
115	else
116		tree = &spa->spa_errlist_last;
117
118	search.se_bookmark = *zb;
119	if (avl_find(tree, &search, &where) != NULL) {
120		mutex_exit(&spa->spa_errlist_lock);
121		return;
122	}
123
124	new = kmem_zalloc(sizeof (spa_error_entry_t), KM_SLEEP);
125	new->se_bookmark = *zb;
126	avl_insert(tree, new, where);
127
128	mutex_exit(&spa->spa_errlist_lock);
129}
130
131/*
132 * Return the number of errors currently in the error log.  This is actually the
133 * sum of both the last log and the current log, since we don't know the union
134 * of these logs until we reach userland.
135 */
136uint64_t
137spa_get_errlog_size(spa_t *spa)
138{
139	uint64_t total = 0, count;
140
141	mutex_enter(&spa->spa_errlog_lock);
142	if (spa->spa_errlog_scrub != 0 &&
143	    zap_count(spa->spa_meta_objset, spa->spa_errlog_scrub,
144	    &count) == 0)
145		total += count;
146
147	if (spa->spa_errlog_last != 0 && !spa->spa_scrub_finished &&
148	    zap_count(spa->spa_meta_objset, spa->spa_errlog_last,
149	    &count) == 0)
150		total += count;
151	mutex_exit(&spa->spa_errlog_lock);
152
153	mutex_enter(&spa->spa_errlist_lock);
154	total += avl_numnodes(&spa->spa_errlist_last);
155	total += avl_numnodes(&spa->spa_errlist_scrub);
156	mutex_exit(&spa->spa_errlist_lock);
157
158	return (total);
159}
160
161#ifdef _KERNEL
162static int
163process_error_log(spa_t *spa, uint64_t obj, void *addr, size_t *count)
164{
165	zap_cursor_t zc;
166	zap_attribute_t za;
167	zbookmark_t zb;
168
169	if (obj == 0)
170		return (0);
171
172	for (zap_cursor_init(&zc, spa->spa_meta_objset, obj);
173	    zap_cursor_retrieve(&zc, &za) == 0;
174	    zap_cursor_advance(&zc)) {
175
176		if (*count == 0) {
177			zap_cursor_fini(&zc);
178			return (ENOMEM);
179		}
180
181		name_to_bookmark(za.za_name, &zb);
182
183		if (copyout(&zb, (char *)addr +
184		    (*count - 1) * sizeof (zbookmark_t),
185		    sizeof (zbookmark_t)) != 0)
186			return (EFAULT);
187
188		*count -= 1;
189	}
190
191	zap_cursor_fini(&zc);
192
193	return (0);
194}
195
196static int
197process_error_list(avl_tree_t *list, void *addr, size_t *count)
198{
199	spa_error_entry_t *se;
200
201	for (se = avl_first(list); se != NULL; se = AVL_NEXT(list, se)) {
202
203		if (*count == 0)
204			return (ENOMEM);
205
206		if (copyout(&se->se_bookmark, (char *)addr +
207		    (*count - 1) * sizeof (zbookmark_t),
208		    sizeof (zbookmark_t)) != 0)
209			return (EFAULT);
210
211		*count -= 1;
212	}
213
214	return (0);
215}
216#endif
217
218/*
219 * Copy all known errors to userland as an array of bookmarks.  This is
220 * actually a union of the on-disk last log and current log, as well as any
221 * pending error requests.
222 *
223 * Because the act of reading the on-disk log could cause errors to be
224 * generated, we have two separate locks: one for the error log and one for the
225 * in-core error lists.  We only need the error list lock to log and error, so
226 * we grab the error log lock while we read the on-disk logs, and only pick up
227 * the error list lock when we are finished.
228 */
229int
230spa_get_errlog(spa_t *spa, void *uaddr, size_t *count)
231{
232	int ret = 0;
233
234#ifdef _KERNEL
235	mutex_enter(&spa->spa_errlog_lock);
236
237	ret = process_error_log(spa, spa->spa_errlog_scrub, uaddr, count);
238
239	if (!ret && !spa->spa_scrub_finished)
240		ret = process_error_log(spa, spa->spa_errlog_last, uaddr,
241		    count);
242
243	mutex_enter(&spa->spa_errlist_lock);
244	if (!ret)
245		ret = process_error_list(&spa->spa_errlist_scrub, uaddr,
246		    count);
247	if (!ret)
248		ret = process_error_list(&spa->spa_errlist_last, uaddr,
249		    count);
250	mutex_exit(&spa->spa_errlist_lock);
251
252	mutex_exit(&spa->spa_errlog_lock);
253#endif
254
255	return (ret);
256}
257
258/*
259 * Called when a scrub completes.  This simply set a bit which tells which AVL
260 * tree to add new errors.  spa_errlog_sync() is responsible for actually
261 * syncing the changes to the underlying objects.
262 */
263void
264spa_errlog_rotate(spa_t *spa)
265{
266	mutex_enter(&spa->spa_errlist_lock);
267	spa->spa_scrub_finished = B_TRUE;
268	mutex_exit(&spa->spa_errlist_lock);
269}
270
271/*
272 * Discard any pending errors from the spa_t.  Called when unloading a faulted
273 * pool, as the errors encountered during the open cannot be synced to disk.
274 */
275void
276spa_errlog_drain(spa_t *spa)
277{
278	spa_error_entry_t *se;
279	void *cookie;
280
281	mutex_enter(&spa->spa_errlist_lock);
282
283	cookie = NULL;
284	while ((se = avl_destroy_nodes(&spa->spa_errlist_last,
285	    &cookie)) != NULL)
286		kmem_free(se, sizeof (spa_error_entry_t));
287	cookie = NULL;
288	while ((se = avl_destroy_nodes(&spa->spa_errlist_scrub,
289	    &cookie)) != NULL)
290		kmem_free(se, sizeof (spa_error_entry_t));
291
292	mutex_exit(&spa->spa_errlist_lock);
293}
294
295/*
296 * Process a list of errors into the current on-disk log.
297 */
298static void
299sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj, dmu_tx_t *tx)
300{
301	spa_error_entry_t *se;
302	char buf[64];
303	void *cookie;
304
305	if (avl_numnodes(t) != 0) {
306		/* create log if necessary */
307		if (*obj == 0)
308			*obj = zap_create(spa->spa_meta_objset,
309			    DMU_OT_ERROR_LOG, DMU_OT_NONE,
310			    0, tx);
311
312		/* add errors to the current log */
313		for (se = avl_first(t); se != NULL; se = AVL_NEXT(t, se)) {
314			char *name = se->se_name ? se->se_name : "";
315
316			bookmark_to_name(&se->se_bookmark, buf, sizeof (buf));
317
318			(void) zap_update(spa->spa_meta_objset,
319			    *obj, buf, 1, strlen(name) + 1, name, tx);
320		}
321
322		/* purge the error list */
323		cookie = NULL;
324		while ((se = avl_destroy_nodes(t, &cookie)) != NULL)
325			kmem_free(se, sizeof (spa_error_entry_t));
326	}
327}
328
329/*
330 * Sync the error log out to disk.  This is a little tricky because the act of
331 * writing the error log requires the spa_errlist_lock.  So, we need to lock the
332 * error lists, take a copy of the lists, and then reinitialize them.  Then, we
333 * drop the error list lock and take the error log lock, at which point we
334 * do the errlog processing.  Then, if we encounter an I/O error during this
335 * process, we can successfully add the error to the list.  Note that this will
336 * result in the perpetual recycling of errors, but it is an unlikely situation
337 * and not a performance critical operation.
338 */
339void
340spa_errlog_sync(spa_t *spa, uint64_t txg)
341{
342	dmu_tx_t *tx;
343	avl_tree_t scrub, last;
344	int scrub_finished;
345
346	mutex_enter(&spa->spa_errlist_lock);
347
348	/*
349	 * Bail out early under normal circumstances.
350	 */
351	if (avl_numnodes(&spa->spa_errlist_scrub) == 0 &&
352	    avl_numnodes(&spa->spa_errlist_last) == 0 &&
353	    !spa->spa_scrub_finished) {
354		mutex_exit(&spa->spa_errlist_lock);
355		return;
356	}
357
358	spa_get_errlists(spa, &last, &scrub);
359	scrub_finished = spa->spa_scrub_finished;
360	spa->spa_scrub_finished = B_FALSE;
361
362	mutex_exit(&spa->spa_errlist_lock);
363	mutex_enter(&spa->spa_errlog_lock);
364
365	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
366
367	/*
368	 * Sync out the current list of errors.
369	 */
370	sync_error_list(spa, &last, &spa->spa_errlog_last, tx);
371
372	/*
373	 * Rotate the log if necessary.
374	 */
375	if (scrub_finished) {
376		if (spa->spa_errlog_last != 0)
377			VERIFY(dmu_object_free(spa->spa_meta_objset,
378			    spa->spa_errlog_last, tx) == 0);
379		spa->spa_errlog_last = spa->spa_errlog_scrub;
380		spa->spa_errlog_scrub = 0;
381
382		sync_error_list(spa, &scrub, &spa->spa_errlog_last, tx);
383	}
384
385	/*
386	 * Sync out any pending scrub errors.
387	 */
388	sync_error_list(spa, &scrub, &spa->spa_errlog_scrub, tx);
389
390	/*
391	 * Update the MOS to reflect the new values.
392	 */
393	(void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
394	    DMU_POOL_ERRLOG_LAST, sizeof (uint64_t), 1,
395	    &spa->spa_errlog_last, tx);
396	(void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
397	    DMU_POOL_ERRLOG_SCRUB, sizeof (uint64_t), 1,
398	    &spa->spa_errlog_scrub, tx);
399
400	dmu_tx_commit(tx);
401
402	mutex_exit(&spa->spa_errlog_lock);
403}
404