xref: /titanic_41/usr/src/uts/common/fs/zfs/spa_errlog.c (revision 9a0e82381c494d655823a54fc713a8c2a571131c)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
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 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  * This is a stripped-down version of strtoull, suitable only for converting
59  * lowercase hexidecimal numbers that don't overflow.
60  */
61 #ifdef _KERNEL
62 uint64_t
63 strtonum(const char *str, char **nptr)
64 {
65 	uint64_t val = 0;
66 	char c;
67 	int digit;
68 
69 	while ((c = *str) != '\0') {
70 		if (c >= '0' && c <= '9')
71 			digit = c - '0';
72 		else if (c >= 'a' && c <= 'f')
73 			digit = 10 + c - 'a';
74 		else
75 			break;
76 
77 		val *= 16;
78 		val += digit;
79 
80 		str++;
81 	}
82 
83 	if (nptr)
84 		*nptr = (char *)str;
85 
86 	return (val);
87 }
88 #endif
89 
90 /*
91  * Convert a bookmark to a string.
92  */
93 static void
94 bookmark_to_name(zbookmark_t *zb, char *buf, size_t len)
95 {
96 	(void) snprintf(buf, len, "%llx:%llx:%llx:%llx",
97 	    (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
98 	    (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid);
99 }
100 
101 /*
102  * Convert a string to a bookmark
103  */
104 #ifdef _KERNEL
105 static void
106 name_to_bookmark(char *buf, zbookmark_t *zb)
107 {
108 	zb->zb_objset = strtonum(buf, &buf);
109 	ASSERT(*buf == ':');
110 	zb->zb_object = strtonum(buf + 1, &buf);
111 	ASSERT(*buf == ':');
112 	zb->zb_level = (int)strtonum(buf + 1, &buf);
113 	ASSERT(*buf == ':');
114 	zb->zb_blkid = strtonum(buf + 1, &buf);
115 	ASSERT(*buf == '\0');
116 }
117 #endif
118 
119 /*
120  * Log an uncorrectable error to the persistent error log.  We add it to the
121  * spa's list of pending errors.  The changes are actually synced out to disk
122  * during spa_errlog_sync().
123  */
124 void
125 spa_log_error(spa_t *spa, zio_t *zio)
126 {
127 	zbookmark_t *zb = &zio->io_logical->io_bookmark;
128 	spa_error_entry_t search;
129 	spa_error_entry_t *new;
130 	avl_tree_t *tree;
131 	avl_index_t where;
132 
133 	/*
134 	 * If we are trying to import a pool, ignore any errors, as we won't be
135 	 * writing to the pool any time soon.
136 	 */
137 	if (spa->spa_load_state == SPA_LOAD_TRYIMPORT)
138 		return;
139 
140 	mutex_enter(&spa->spa_errlist_lock);
141 
142 	/*
143 	 * If we have had a request to rotate the log, log it to the next list
144 	 * instead of the current one.
145 	 */
146 	if (spa->spa_scrub_active || spa->spa_scrub_finished)
147 		tree = &spa->spa_errlist_scrub;
148 	else
149 		tree = &spa->spa_errlist_last;
150 
151 	search.se_bookmark = *zb;
152 	if (avl_find(tree, &search, &where) != NULL) {
153 		mutex_exit(&spa->spa_errlist_lock);
154 		return;
155 	}
156 
157 	new = kmem_zalloc(sizeof (spa_error_entry_t), KM_SLEEP);
158 	new->se_bookmark = *zb;
159 	avl_insert(tree, new, where);
160 
161 	mutex_exit(&spa->spa_errlist_lock);
162 }
163 
164 /*
165  * Return the number of errors currently in the error log.  This is actually the
166  * sum of both the last log and the current log, since we don't know the union
167  * of these logs until we reach userland.
168  */
169 uint64_t
170 spa_get_errlog_size(spa_t *spa)
171 {
172 	uint64_t total = 0, count;
173 
174 	mutex_enter(&spa->spa_errlog_lock);
175 	if (spa->spa_errlog_scrub != 0 &&
176 	    zap_count(spa->spa_meta_objset, spa->spa_errlog_scrub,
177 	    &count) == 0)
178 		total += count;
179 
180 	if (spa->spa_errlog_last != 0 && !spa->spa_scrub_finished &&
181 	    zap_count(spa->spa_meta_objset, spa->spa_errlog_last,
182 	    &count) == 0)
183 		total += count;
184 	mutex_exit(&spa->spa_errlog_lock);
185 
186 	mutex_enter(&spa->spa_errlist_lock);
187 	total += avl_numnodes(&spa->spa_errlist_last);
188 	total += avl_numnodes(&spa->spa_errlist_scrub);
189 	mutex_exit(&spa->spa_errlist_lock);
190 
191 	return (total);
192 }
193 
194 #ifdef _KERNEL
195 static int
196 process_error_log(spa_t *spa, uint64_t obj, void *addr, size_t *count)
197 {
198 	zap_cursor_t zc;
199 	zap_attribute_t za;
200 	zbookmark_t zb;
201 
202 	if (obj == 0)
203 		return (0);
204 
205 	for (zap_cursor_init(&zc, spa->spa_meta_objset, obj);
206 	    zap_cursor_retrieve(&zc, &za) == 0;
207 	    zap_cursor_advance(&zc)) {
208 
209 		if (*count == 0) {
210 			zap_cursor_fini(&zc);
211 			return (ENOMEM);
212 		}
213 
214 		name_to_bookmark(za.za_name, &zb);
215 
216 		if (copyout(&zb, (char *)addr +
217 		    (*count - 1) * sizeof (zbookmark_t),
218 		    sizeof (zbookmark_t)) != 0)
219 			return (EFAULT);
220 
221 		*count -= 1;
222 	}
223 
224 	zap_cursor_fini(&zc);
225 
226 	return (0);
227 }
228 
229 static int
230 process_error_list(avl_tree_t *list, void *addr, size_t *count)
231 {
232 	spa_error_entry_t *se;
233 
234 	for (se = avl_first(list); se != NULL; se = AVL_NEXT(list, se)) {
235 
236 		if (*count == 0)
237 			return (ENOMEM);
238 
239 		if (copyout(&se->se_bookmark, (char *)addr +
240 		    (*count - 1) * sizeof (zbookmark_t),
241 		    sizeof (zbookmark_t)) != 0)
242 			return (EFAULT);
243 
244 		*count -= 1;
245 	}
246 
247 	return (0);
248 }
249 #endif
250 
251 /*
252  * Copy all known errors to userland as an array of bookmarks.  This is
253  * actually a union of the on-disk last log and current log, as well as any
254  * pending error requests.
255  *
256  * Because the act of reading the on-disk log could cause errors to be
257  * generated, we have two separate locks: one for the error log and one for the
258  * in-core error lists.  We only need the error list lock to log and error, so
259  * we grab the error log lock while we read the on-disk logs, and only pick up
260  * the error list lock when we are finished.
261  */
262 int
263 spa_get_errlog(spa_t *spa, void *uaddr, size_t *count)
264 {
265 	int ret = 0;
266 
267 #ifdef _KERNEL
268 	mutex_enter(&spa->spa_errlog_lock);
269 
270 	ret = process_error_log(spa, spa->spa_errlog_scrub, uaddr, count);
271 
272 	if (!ret && !spa->spa_scrub_finished)
273 		ret = process_error_log(spa, spa->spa_errlog_last, uaddr,
274 		    count);
275 
276 	mutex_enter(&spa->spa_errlist_lock);
277 	if (!ret)
278 		ret = process_error_list(&spa->spa_errlist_scrub, uaddr,
279 		    count);
280 	if (!ret)
281 		ret = process_error_list(&spa->spa_errlist_last, uaddr,
282 		    count);
283 	mutex_exit(&spa->spa_errlist_lock);
284 
285 	mutex_exit(&spa->spa_errlog_lock);
286 #endif
287 
288 	return (ret);
289 }
290 
291 /*
292  * Called when a scrub completes.  This simply set a bit which tells which AVL
293  * tree to add new errors.  spa_errlog_sync() is responsible for actually
294  * syncing the changes to the underlying objects.
295  */
296 void
297 spa_errlog_rotate(spa_t *spa)
298 {
299 	mutex_enter(&spa->spa_errlist_lock);
300 	spa->spa_scrub_finished = B_TRUE;
301 	mutex_exit(&spa->spa_errlist_lock);
302 }
303 
304 /*
305  * Discard any pending errors from the spa_t.  Called when unloading a faulted
306  * pool, as the errors encountered during the open cannot be synced to disk.
307  */
308 void
309 spa_errlog_drain(spa_t *spa)
310 {
311 	spa_error_entry_t *se;
312 	void *cookie;
313 
314 	mutex_enter(&spa->spa_errlist_lock);
315 
316 	cookie = NULL;
317 	while ((se = avl_destroy_nodes(&spa->spa_errlist_last,
318 	    &cookie)) != NULL)
319 		kmem_free(se, sizeof (spa_error_entry_t));
320 	cookie = NULL;
321 	while ((se = avl_destroy_nodes(&spa->spa_errlist_scrub,
322 	    &cookie)) != NULL)
323 		kmem_free(se, sizeof (spa_error_entry_t));
324 
325 	mutex_exit(&spa->spa_errlist_lock);
326 }
327 
328 /*
329  * Process a list of errors into the current on-disk log.
330  */
331 static void
332 sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj, dmu_tx_t *tx)
333 {
334 	spa_error_entry_t *se;
335 	char buf[64];
336 	void *cookie;
337 
338 	if (avl_numnodes(t) != 0) {
339 		/* create log if necessary */
340 		if (*obj == 0)
341 			*obj = zap_create(spa->spa_meta_objset,
342 			    DMU_OT_ERROR_LOG, DMU_OT_NONE,
343 			    0, tx);
344 
345 		/* add errors to the current log */
346 		for (se = avl_first(t); se != NULL; se = AVL_NEXT(t, se)) {
347 			char *name = se->se_name ? se->se_name : "";
348 
349 			bookmark_to_name(&se->se_bookmark, buf, sizeof (buf));
350 
351 			(void) zap_update(spa->spa_meta_objset,
352 			    *obj, buf, 1, strlen(name) + 1, name, tx);
353 		}
354 
355 		/* purge the error list */
356 		cookie = NULL;
357 		while ((se = avl_destroy_nodes(t, &cookie)) != NULL)
358 			kmem_free(se, sizeof (spa_error_entry_t));
359 	}
360 }
361 
362 /*
363  * Sync the error log out to disk.  This is a little tricky because the act of
364  * writing the error log requires the spa_errlist_lock.  So, we need to lock the
365  * error lists, take a copy of the lists, and then reinitialize them.  Then, we
366  * drop the error list lock and take the error log lock, at which point we
367  * do the errlog processing.  Then, if we encounter an I/O error during this
368  * process, we can successfully add the error to the list.  Note that this will
369  * result in the perpetual recycling of errors, but it is an unlikely situation
370  * and not a performance critical operation.
371  */
372 void
373 spa_errlog_sync(spa_t *spa, uint64_t txg)
374 {
375 	dmu_tx_t *tx;
376 	avl_tree_t scrub, last;
377 	int scrub_finished;
378 
379 	mutex_enter(&spa->spa_errlist_lock);
380 
381 	/*
382 	 * Bail out early under normal circumstances.
383 	 */
384 	if (avl_numnodes(&spa->spa_errlist_scrub) == 0 &&
385 	    avl_numnodes(&spa->spa_errlist_last) == 0 &&
386 	    !spa->spa_scrub_finished) {
387 		mutex_exit(&spa->spa_errlist_lock);
388 		return;
389 	}
390 
391 	spa_get_errlists(spa, &last, &scrub);
392 	scrub_finished = spa->spa_scrub_finished;
393 	spa->spa_scrub_finished = B_FALSE;
394 
395 	mutex_exit(&spa->spa_errlist_lock);
396 	mutex_enter(&spa->spa_errlog_lock);
397 
398 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
399 
400 	/*
401 	 * Sync out the current list of errors.
402 	 */
403 	sync_error_list(spa, &last, &spa->spa_errlog_last, tx);
404 
405 	/*
406 	 * Rotate the log if necessary.
407 	 */
408 	if (scrub_finished) {
409 		if (spa->spa_errlog_last != 0)
410 			VERIFY(dmu_object_free(spa->spa_meta_objset,
411 			    spa->spa_errlog_last, tx) == 0);
412 		spa->spa_errlog_last = spa->spa_errlog_scrub;
413 		spa->spa_errlog_scrub = 0;
414 
415 		sync_error_list(spa, &scrub, &spa->spa_errlog_last, tx);
416 	}
417 
418 	/*
419 	 * Sync out any pending scrub errors.
420 	 */
421 	sync_error_list(spa, &scrub, &spa->spa_errlog_scrub, tx);
422 
423 	/*
424 	 * Update the MOS to reflect the new values.
425 	 */
426 	(void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
427 	    DMU_POOL_ERRLOG_LAST, sizeof (uint64_t), 1,
428 	    &spa->spa_errlog_last, tx);
429 	(void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
430 	    DMU_POOL_ERRLOG_SCRUB, sizeof (uint64_t), 1,
431 	    &spa->spa_errlog_scrub, tx);
432 
433 	dmu_tx_commit(tx);
434 
435 	mutex_exit(&spa->spa_errlog_lock);
436 }
437