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 * Copyright (c) 2013, 2014 by Delphix. All rights reserved.
24 */
25
26/*
27 * Routines to manage the on-disk persistent error log.
28 *
29 * Each pool stores a log of all logical data errors seen during normal
30 * operation.  This is actually the union of two distinct logs: the last log,
31 * and the current log.  All errors seen are logged to the current log.  When a
32 * scrub completes, the current log becomes the last log, the last log is thrown
33 * out, and the current log is reinitialized.  This way, if an error is somehow
34 * corrected, a new scrub will show that that it no longer exists, and will be
35 * deleted from the log when the scrub completes.
36 *
37 * The log is stored using a ZAP object whose key is a string form of the
38 * zbookmark_phys tuple (objset, object, level, blkid), and whose contents is an
39 * optional 'objset:object' human-readable string describing the data.  When an
40 * error is first logged, this string will be empty, indicating that no name is
41 * known.  This prevents us from having to issue a potentially large amount of
42 * I/O to discover the object name during an error path.  Instead, we do the
43 * calculation when the data is requested, storing the result so future queries
44 * will be faster.
45 *
46 * This log is then shipped into an nvlist where the key is the dataset name and
47 * the value is the object name.  Userland is then responsible for uniquifying
48 * this list and displaying it to the user.
49 */
50
51#include <sys/dmu_tx.h>
52#include <sys/spa.h>
53#include <sys/spa_impl.h>
54#include <sys/zap.h>
55#include <sys/zio.h>
56
57
58/*
59 * Convert a bookmark to a string.
60 */
61static void
62bookmark_to_name(zbookmark_phys_t *zb, char *buf, size_t len)
63{
64	(void) snprintf(buf, len, "%llx:%llx:%llx:%llx",
65	    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
66	    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid);
67}
68
69/*
70 * Convert a string to a bookmark
71 */
72#ifdef _KERNEL
73static void
74name_to_bookmark(char *buf, zbookmark_phys_t *zb)
75{
76	zb->zb_objset = zfs_strtonum(buf, &buf);
77	ASSERT(*buf == ':');
78	zb->zb_object = zfs_strtonum(buf + 1, &buf);
79	ASSERT(*buf == ':');
80	zb->zb_level = (int)zfs_strtonum(buf + 1, &buf);
81	ASSERT(*buf == ':');
82	zb->zb_blkid = zfs_strtonum(buf + 1, &buf);
83	ASSERT(*buf == '\0');
84}
85#endif
86
87/*
88 * Log an uncorrectable error to the persistent error log.  We add it to the
89 * spa's list of pending errors.  The changes are actually synced out to disk
90 * during spa_errlog_sync().
91 */
92void
93spa_log_error(spa_t *spa, const zbookmark_phys_t *zb)
94{
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_phys_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 (SET_ERROR(ENOMEM));
179		}
180
181		name_to_bookmark(za.za_name, &zb);
182
183		if (copyout(&zb, (char *)addr +
184		    (*count - 1) * sizeof (zbookmark_phys_t),
185		    sizeof (zbookmark_phys_t)) != 0) {
186			zap_cursor_fini(&zc);
187			return (SET_ERROR(EFAULT));
188		}
189
190		*count -= 1;
191	}
192
193	zap_cursor_fini(&zc);
194
195	return (0);
196}
197
198static int
199process_error_list(avl_tree_t *list, void *addr, size_t *count)
200{
201	spa_error_entry_t *se;
202
203	for (se = avl_first(list); se != NULL; se = AVL_NEXT(list, se)) {
204
205		if (*count == 0)
206			return (SET_ERROR(ENOMEM));
207
208		if (copyout(&se->se_bookmark, (char *)addr +
209		    (*count - 1) * sizeof (zbookmark_phys_t),
210		    sizeof (zbookmark_phys_t)) != 0)
211			return (SET_ERROR(EFAULT));
212
213		*count -= 1;
214	}
215
216	return (0);
217}
218#endif
219
220/*
221 * Copy all known errors to userland as an array of bookmarks.  This is
222 * actually a union of the on-disk last log and current log, as well as any
223 * pending error requests.
224 *
225 * Because the act of reading the on-disk log could cause errors to be
226 * generated, we have two separate locks: one for the error log and one for the
227 * in-core error lists.  We only need the error list lock to log and error, so
228 * we grab the error log lock while we read the on-disk logs, and only pick up
229 * the error list lock when we are finished.
230 */
231int
232spa_get_errlog(spa_t *spa, void *uaddr, size_t *count)
233{
234	int ret = 0;
235
236#ifdef _KERNEL
237	mutex_enter(&spa->spa_errlog_lock);
238
239	ret = process_error_log(spa, spa->spa_errlog_scrub, uaddr, count);
240
241	if (!ret && !spa->spa_scrub_finished)
242		ret = process_error_log(spa, spa->spa_errlog_last, uaddr,
243		    count);
244
245	mutex_enter(&spa->spa_errlist_lock);
246	if (!ret)
247		ret = process_error_list(&spa->spa_errlist_scrub, uaddr,
248		    count);
249	if (!ret)
250		ret = process_error_list(&spa->spa_errlist_last, uaddr,
251		    count);
252	mutex_exit(&spa->spa_errlist_lock);
253
254	mutex_exit(&spa->spa_errlog_lock);
255#endif
256
257	return (ret);
258}
259
260/*
261 * Called when a scrub completes.  This simply set a bit which tells which AVL
262 * tree to add new errors.  spa_errlog_sync() is responsible for actually
263 * syncing the changes to the underlying objects.
264 */
265void
266spa_errlog_rotate(spa_t *spa)
267{
268	mutex_enter(&spa->spa_errlist_lock);
269	spa->spa_scrub_finished = B_TRUE;
270	mutex_exit(&spa->spa_errlist_lock);
271}
272
273/*
274 * Discard any pending errors from the spa_t.  Called when unloading a faulted
275 * pool, as the errors encountered during the open cannot be synced to disk.
276 */
277void
278spa_errlog_drain(spa_t *spa)
279{
280	spa_error_entry_t *se;
281	void *cookie;
282
283	mutex_enter(&spa->spa_errlist_lock);
284
285	cookie = NULL;
286	while ((se = avl_destroy_nodes(&spa->spa_errlist_last,
287	    &cookie)) != NULL)
288		kmem_free(se, sizeof (spa_error_entry_t));
289	cookie = NULL;
290	while ((se = avl_destroy_nodes(&spa->spa_errlist_scrub,
291	    &cookie)) != NULL)
292		kmem_free(se, sizeof (spa_error_entry_t));
293
294	mutex_exit(&spa->spa_errlist_lock);
295}
296
297/*
298 * Process a list of errors into the current on-disk log.
299 */
300static void
301sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj, dmu_tx_t *tx)
302{
303	spa_error_entry_t *se;
304	char buf[64];
305	void *cookie;
306
307	if (avl_numnodes(t) != 0) {
308		/* create log if necessary */
309		if (*obj == 0)
310			*obj = zap_create(spa->spa_meta_objset,
311			    DMU_OT_ERROR_LOG, DMU_OT_NONE,
312			    0, tx);
313
314		/* add errors to the current log */
315		for (se = avl_first(t); se != NULL; se = AVL_NEXT(t, se)) {
316			char *name = se->se_name ? se->se_name : "";
317
318			bookmark_to_name(&se->se_bookmark, buf, sizeof (buf));
319
320			(void) zap_update(spa->spa_meta_objset,
321			    *obj, buf, 1, strlen(name) + 1, name, tx);
322		}
323
324		/* purge the error list */
325		cookie = NULL;
326		while ((se = avl_destroy_nodes(t, &cookie)) != NULL)
327			kmem_free(se, sizeof (spa_error_entry_t));
328	}
329}
330
331/*
332 * Sync the error log out to disk.  This is a little tricky because the act of
333 * writing the error log requires the spa_errlist_lock.  So, we need to lock the
334 * error lists, take a copy of the lists, and then reinitialize them.  Then, we
335 * drop the error list lock and take the error log lock, at which point we
336 * do the errlog processing.  Then, if we encounter an I/O error during this
337 * process, we can successfully add the error to the list.  Note that this will
338 * result in the perpetual recycling of errors, but it is an unlikely situation
339 * and not a performance critical operation.
340 */
341void
342spa_errlog_sync(spa_t *spa, uint64_t txg)
343{
344	dmu_tx_t *tx;
345	avl_tree_t scrub, last;
346	int scrub_finished;
347
348	mutex_enter(&spa->spa_errlist_lock);
349
350	/*
351	 * Bail out early under normal circumstances.
352	 */
353	if (avl_numnodes(&spa->spa_errlist_scrub) == 0 &&
354	    avl_numnodes(&spa->spa_errlist_last) == 0 &&
355	    !spa->spa_scrub_finished) {
356		mutex_exit(&spa->spa_errlist_lock);
357		return;
358	}
359
360	spa_get_errlists(spa, &last, &scrub);
361	scrub_finished = spa->spa_scrub_finished;
362	spa->spa_scrub_finished = B_FALSE;
363
364	mutex_exit(&spa->spa_errlist_lock);
365	mutex_enter(&spa->spa_errlog_lock);
366
367	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
368
369	/*
370	 * Sync out the current list of errors.
371	 */
372	sync_error_list(spa, &last, &spa->spa_errlog_last, tx);
373
374	/*
375	 * Rotate the log if necessary.
376	 */
377	if (scrub_finished) {
378		if (spa->spa_errlog_last != 0)
379			VERIFY(dmu_object_free(spa->spa_meta_objset,
380			    spa->spa_errlog_last, tx) == 0);
381		spa->spa_errlog_last = spa->spa_errlog_scrub;
382		spa->spa_errlog_scrub = 0;
383
384		sync_error_list(spa, &scrub, &spa->spa_errlog_last, tx);
385	}
386
387	/*
388	 * Sync out any pending scrub errors.
389	 */
390	sync_error_list(spa, &scrub, &spa->spa_errlog_scrub, tx);
391
392	/*
393	 * Update the MOS to reflect the new values.
394	 */
395	(void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
396	    DMU_POOL_ERRLOG_LAST, sizeof (uint64_t), 1,
397	    &spa->spa_errlog_last, tx);
398	(void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
399	    DMU_POOL_ERRLOG_SCRUB, sizeof (uint64_t), 1,
400	    &spa->spa_errlog_scrub, tx);
401
402	dmu_tx_commit(tx);
403
404	mutex_exit(&spa->spa_errlog_lock);
405}
406