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