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