xref: /illumos-gate/usr/src/uts/common/fs/zfs/spa_errlog.c (revision 088f389458728c464569a5506b58070254fa4f7d)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * Routines to manage the on-disk persistent error log.
30  *
31  * Each pool stores a log of all logical data errors seen during normal
32  * operation.  This is actually the union of two distinct logs: the last log,
33  * and the current log.  All errors seen are logged to the current log.  When a
34  * scrub completes, the current log becomes the last log, the last log is thrown
35  * out, and the current log is reinitialized.  This way, if an error is somehow
36  * corrected, a new scrub will show that that it no longer exists, and will be
37  * deleted from the log when the scrub completes.
38  *
39  * The log is stored using a ZAP object whose key is a string form of the
40  * zbookmark tuple (objset, object, level, blkid), and whose contents is an
41  * optional 'objset:object' human-readable string describing the data.  When an
42  * error is first logged, this string will be empty, indicating that no name is
43  * known.  This prevents us from having to issue a potentially large amount of
44  * I/O to discover the object name during an error path.  Instead, we do the
45  * calculation when the data is requested, storing the result so future queries
46  * will be faster.
47  *
48  * This log is then shipped into an nvlist where the key is the dataset name and
49  * the value is the object name.  Userland is then responsible for uniquifying
50  * this list and displaying it to the user.
51  */
52 
53 #include <sys/dmu_tx.h>
54 #include <sys/spa.h>
55 #include <sys/spa_impl.h>
56 #include <sys/zap.h>
57 #include <sys/zio.h>
58 
59 /*
60  * This is a stripped-down version of strtoull, suitable only for converting
61  * lowercase hexidecimal numbers that don't overflow.
62  */
63 #ifdef _KERNEL
64 static uint64_t
65 strtonum(char *str, char **nptr)
66 {
67 	uint64_t val = 0;
68 	char c;
69 	int digit;
70 
71 	while ((c = *str) != '\0') {
72 		if (c >= '0' && c <= '9')
73 			digit = c - '0';
74 		else if (c >= 'a' && c <= 'f')
75 			digit = 10 + c - 'a';
76 		else
77 			break;
78 
79 		val *= 16;
80 		val += digit;
81 
82 		str++;
83 	}
84 
85 	*nptr = str;
86 
87 	return (val);
88 }
89 #endif
90 
91 /*
92  * Convert a bookmark to a string.
93  */
94 static void
95 bookmark_to_name(zbookmark_t *zb, char *buf, size_t len)
96 {
97 	(void) snprintf(buf, len, "%llx:%llx:%llx:%llx",
98 	    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
99 	    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid);
100 }
101 
102 /*
103  * Convert a string to a bookmark
104  */
105 #ifdef _KERNEL
106 static void
107 name_to_bookmark(char *buf, zbookmark_t *zb)
108 {
109 	zb->zb_objset = strtonum(buf, &buf);
110 	ASSERT(*buf == ':');
111 	zb->zb_object = strtonum(buf + 1, &buf);
112 	ASSERT(*buf == ':');
113 	zb->zb_level = (int)strtonum(buf + 1, &buf);
114 	ASSERT(*buf == ':');
115 	zb->zb_blkid = strtonum(buf + 1, &buf);
116 	ASSERT(*buf == '\0');
117 }
118 #endif
119 
120 /*
121  * Log an uncorrectable error to the persistent error log.  We add it to the
122  * spa's list of pending errors.  The changes are actually synced out to disk
123  * during spa_errlog_sync().
124  */
125 void
126 spa_log_error(spa_t *spa, zio_t *zio)
127 {
128 	zbookmark_t *zb = &zio->io_logical->io_bookmark;
129 	spa_error_entry_t search;
130 	spa_error_entry_t *new;
131 	avl_tree_t *tree;
132 	avl_index_t where;
133 
134 	/*
135 	 * If we are trying to import a pool, ignore any errors, as we won't be
136 	 * writing to the pool any time soon.
137 	 */
138 	if (spa->spa_load_state == SPA_LOAD_TRYIMPORT)
139 		return;
140 
141 	mutex_enter(&spa->spa_errlist_lock);
142 
143 	/*
144 	 * If we have had a request to rotate the log, log it to the next list
145 	 * instead of the current one.
146 	 */
147 	if (spa->spa_scrub_active || spa->spa_scrub_finished)
148 		tree = &spa->spa_errlist_scrub;
149 	else
150 		tree = &spa->spa_errlist_last;
151 
152 	search.se_bookmark = *zb;
153 	if (avl_find(tree, &search, &where) != NULL) {
154 		mutex_exit(&spa->spa_errlist_lock);
155 		return;
156 	}
157 
158 	new = kmem_zalloc(sizeof (spa_error_entry_t), KM_SLEEP);
159 	new->se_bookmark = *zb;
160 	avl_insert(tree, new, where);
161 
162 	mutex_exit(&spa->spa_errlist_lock);
163 }
164 
165 /*
166  * Return the number of errors currently in the error log.  This is actually the
167  * sum of both the last log and the current log, since we don't know the union
168  * of these logs until we reach userland.
169  */
170 uint64_t
171 spa_get_errlog_size(spa_t *spa)
172 {
173 	uint64_t total = 0, count;
174 
175 	mutex_enter(&spa->spa_errlog_lock);
176 	if (spa->spa_errlog_scrub != 0 &&
177 	    zap_count(spa->spa_meta_objset, spa->spa_errlog_scrub,
178 	    &count) == 0)
179 		total += count;
180 
181 	if (spa->spa_errlog_last != 0 && !spa->spa_scrub_finished &&
182 	    zap_count(spa->spa_meta_objset, spa->spa_errlog_last,
183 	    &count) == 0)
184 		total += count;
185 	mutex_exit(&spa->spa_errlog_lock);
186 
187 	mutex_enter(&spa->spa_errlist_lock);
188 	total += avl_numnodes(&spa->spa_errlist_last);
189 	total += avl_numnodes(&spa->spa_errlist_scrub);
190 	mutex_exit(&spa->spa_errlist_lock);
191 
192 	return (total);
193 }
194 
195 #ifdef _KERNEL
196 static int
197 process_error_log(spa_t *spa, uint64_t obj, void *addr, size_t *count)
198 {
199 	zap_cursor_t zc;
200 	zap_attribute_t za;
201 	zbookmark_t zb;
202 
203 	if (obj == 0)
204 		return (0);
205 
206 	for (zap_cursor_init(&zc, spa->spa_meta_objset, obj);
207 	    zap_cursor_retrieve(&zc, &za) == 0;
208 	    zap_cursor_advance(&zc)) {
209 
210 		if (*count == 0) {
211 			zap_cursor_fini(&zc);
212 			return (ENOMEM);
213 		}
214 
215 		name_to_bookmark(za.za_name, &zb);
216 
217 		if (copyout(&zb, (char *)addr +
218 		    (*count - 1) * sizeof (zbookmark_t),
219 		    sizeof (zbookmark_t)) != 0)
220 			return (EFAULT);
221 
222 		*count -= 1;
223 	}
224 
225 	zap_cursor_fini(&zc);
226 
227 	return (0);
228 }
229 
230 static int
231 process_error_list(avl_tree_t *list, void *addr, size_t *count)
232 {
233 	spa_error_entry_t *se;
234 
235 	for (se = avl_first(list); se != NULL; se = AVL_NEXT(list, se)) {
236 
237 		if (*count == 0)
238 			return (ENOMEM);
239 
240 		if (copyout(&se->se_bookmark, (char *)addr +
241 		    (*count - 1) * sizeof (zbookmark_t),
242 		    sizeof (zbookmark_t)) != 0)
243 			return (EFAULT);
244 
245 		*count -= 1;
246 	}
247 
248 	return (0);
249 }
250 #endif
251 
252 /*
253  * Copy all known errors to userland as an array of bookmarks.  This is
254  * actually a union of the on-disk last log and current log, as well as any
255  * pending error requests.
256  *
257  * Because the act of reading the on-disk log could cause errors to be
258  * generated, we have two separate locks: one for the error log and one for the
259  * in-core error lists.  We only need the error list lock to log and error, so
260  * we grab the error log lock while we read the on-disk logs, and only pick up
261  * the error list lock when we are finished.
262  */
263 int
264 spa_get_errlog(spa_t *spa, void *uaddr, size_t *count)
265 {
266 	int ret = 0;
267 
268 #ifdef _KERNEL
269 	mutex_enter(&spa->spa_errlog_lock);
270 
271 	ret = process_error_log(spa, spa->spa_errlog_scrub, uaddr, count);
272 
273 	if (!ret && !spa->spa_scrub_finished)
274 		ret = process_error_log(spa, spa->spa_errlog_last, uaddr,
275 		    count);
276 
277 	mutex_enter(&spa->spa_errlist_lock);
278 	if (!ret)
279 		ret = process_error_list(&spa->spa_errlist_scrub, uaddr,
280 		    count);
281 	if (!ret)
282 		ret = process_error_list(&spa->spa_errlist_last, uaddr,
283 		    count);
284 	mutex_exit(&spa->spa_errlist_lock);
285 
286 	mutex_exit(&spa->spa_errlog_lock);
287 #endif
288 
289 	return (ret);
290 }
291 
292 /*
293  * Called when a scrub completes.  This simply set a bit which tells which AVL
294  * tree to add new errors.  spa_errlog_sync() is responsible for actually
295  * syncing the changes to the underlying objects.
296  */
297 void
298 spa_errlog_rotate(spa_t *spa)
299 {
300 	mutex_enter(&spa->spa_errlist_lock);
301 	spa->spa_scrub_finished = B_TRUE;
302 	mutex_exit(&spa->spa_errlist_lock);
303 }
304 
305 /*
306  * Discard any pending errors from the spa_t.  Called when unloading a faulted
307  * pool, as the errors encountered during the open cannot be synced to disk.
308  */
309 void
310 spa_errlog_drain(spa_t *spa)
311 {
312 	spa_error_entry_t *se;
313 	void *cookie;
314 
315 	mutex_enter(&spa->spa_errlist_lock);
316 
317 	cookie = NULL;
318 	while ((se = avl_destroy_nodes(&spa->spa_errlist_last,
319 	    &cookie)) != NULL)
320 		kmem_free(se, sizeof (spa_error_entry_t));
321 	cookie = NULL;
322 	while ((se = avl_destroy_nodes(&spa->spa_errlist_scrub,
323 	    &cookie)) != NULL)
324 		kmem_free(se, sizeof (spa_error_entry_t));
325 
326 	mutex_exit(&spa->spa_errlist_lock);
327 }
328 
329 /*
330  * Process a list of errors into the current on-disk log.
331  */
332 static void
333 sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj, dmu_tx_t *tx)
334 {
335 	spa_error_entry_t *se;
336 	char buf[64];
337 	void *cookie;
338 
339 	if (avl_numnodes(t) != 0) {
340 		/* create log if necessary */
341 		if (*obj == 0)
342 			*obj = zap_create(spa->spa_meta_objset,
343 			    DMU_OT_ERROR_LOG, DMU_OT_NONE,
344 			    0, tx);
345 
346 		/* add errors to the current log */
347 		for (se = avl_first(t); se != NULL; se = AVL_NEXT(t, se)) {
348 			char *name = se->se_name ? se->se_name : "";
349 
350 			bookmark_to_name(&se->se_bookmark, buf, sizeof (buf));
351 
352 			(void) zap_update(spa->spa_meta_objset,
353 			    *obj, buf, 1, strlen(name) + 1, name, tx);
354 		}
355 
356 		/* purge the error list */
357 		cookie = NULL;
358 		while ((se = avl_destroy_nodes(t, &cookie)) != NULL)
359 			kmem_free(se, sizeof (spa_error_entry_t));
360 	}
361 }
362 
363 /*
364  * Sync the error log out to disk.  This is a little tricky because the act of
365  * writing the error log requires the spa_errlist_lock.  So, we need to lock the
366  * error lists, take a copy of the lists, and then reinitialize them.  Then, we
367  * drop the error list lock and take the error log lock, at which point we
368  * do the errlog processing.  Then, if we encounter an I/O error during this
369  * process, we can successfully add the error to the list.  Note that this will
370  * result in the perpetual recycling of errors, but it is an unlikely situation
371  * and not a performance critical operation.
372  */
373 void
374 spa_errlog_sync(spa_t *spa, uint64_t txg)
375 {
376 	dmu_tx_t *tx;
377 	avl_tree_t scrub, last;
378 	int scrub_finished;
379 
380 	mutex_enter(&spa->spa_errlist_lock);
381 
382 	/*
383 	 * Bail out early under normal circumstances.
384 	 */
385 	if (avl_numnodes(&spa->spa_errlist_scrub) == 0 &&
386 	    avl_numnodes(&spa->spa_errlist_last) == 0 &&
387 	    !spa->spa_scrub_finished) {
388 		mutex_exit(&spa->spa_errlist_lock);
389 		return;
390 	}
391 
392 	spa_get_errlists(spa, &last, &scrub);
393 	scrub_finished = spa->spa_scrub_finished;
394 	spa->spa_scrub_finished = B_FALSE;
395 
396 	mutex_exit(&spa->spa_errlist_lock);
397 	mutex_enter(&spa->spa_errlog_lock);
398 
399 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
400 
401 	/*
402 	 * Sync out the current list of errors.
403 	 */
404 	sync_error_list(spa, &last, &spa->spa_errlog_last, tx);
405 
406 	/*
407 	 * Rotate the log if necessary.
408 	 */
409 	if (scrub_finished) {
410 		if (spa->spa_errlog_last != 0)
411 			VERIFY(dmu_object_free(spa->spa_meta_objset,
412 			    spa->spa_errlog_last, tx) == 0);
413 		spa->spa_errlog_last = spa->spa_errlog_scrub;
414 		spa->spa_errlog_scrub = 0;
415 
416 		sync_error_list(spa, &scrub, &spa->spa_errlog_last, tx);
417 	}
418 
419 	/*
420 	 * Sync out any pending scrub errors.
421 	 */
422 	sync_error_list(spa, &scrub, &spa->spa_errlog_scrub, tx);
423 
424 	/*
425 	 * Update the MOS to reflect the new values.
426 	 */
427 	(void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
428 	    DMU_POOL_ERRLOG_LAST, sizeof (uint64_t), 1,
429 	    &spa->spa_errlog_last, tx);
430 	(void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
431 	    DMU_POOL_ERRLOG_SCRUB, sizeof (uint64_t), 1,
432 	    &spa->spa_errlog_scrub, tx);
433 
434 	dmu_tx_commit(tx);
435 
436 	mutex_exit(&spa->spa_errlog_lock);
437 }
438