1168404Spjd/* 2168404Spjd * CDDL HEADER START 3168404Spjd * 4168404Spjd * The contents of this file are subject to the terms of the 5168404Spjd * Common Development and Distribution License (the "License"). 6168404Spjd * You may not use this file except in compliance with the License. 7168404Spjd * 8168404Spjd * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9168404Spjd * or http://www.opensolaris.org/os/licensing. 10168404Spjd * See the License for the specific language governing permissions 11168404Spjd * and limitations under the License. 12168404Spjd * 13168404Spjd * When distributing Covered Code, include this CDDL HEADER in each 14168404Spjd * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15168404Spjd * If applicable, add the following below this CDDL HEADER, with the 16168404Spjd * fields enclosed by brackets "[]" replaced with your own identifying 17168404Spjd * information: Portions Copyright [yyyy] [name of copyright owner] 18168404Spjd * 19168404Spjd * CDDL HEADER END 20168404Spjd */ 21168404Spjd/* 22219089Spjd * Copyright (c) 2006, 2010, Oracle and/or its affiliates. All rights reserved. 23268123Sdelphij * Copyright (c) 2013, 2014 by Delphix. All rights reserved. 24168404Spjd */ 25168404Spjd 26168404Spjd/* 27168404Spjd * Routines to manage the on-disk persistent error log. 28168404Spjd * 29168404Spjd * Each pool stores a log of all logical data errors seen during normal 30168404Spjd * operation. This is actually the union of two distinct logs: the last log, 31168404Spjd * and the current log. All errors seen are logged to the current log. When a 32168404Spjd * scrub completes, the current log becomes the last log, the last log is thrown 33168404Spjd * out, and the current log is reinitialized. This way, if an error is somehow 34168404Spjd * corrected, a new scrub will show that that it no longer exists, and will be 35168404Spjd * deleted from the log when the scrub completes. 36168404Spjd * 37168404Spjd * The log is stored using a ZAP object whose key is a string form of the 38268123Sdelphij * zbookmark_phys tuple (objset, object, level, blkid), and whose contents is an 39168404Spjd * optional 'objset:object' human-readable string describing the data. When an 40168404Spjd * error is first logged, this string will be empty, indicating that no name is 41168404Spjd * known. This prevents us from having to issue a potentially large amount of 42168404Spjd * I/O to discover the object name during an error path. Instead, we do the 43168404Spjd * calculation when the data is requested, storing the result so future queries 44168404Spjd * will be faster. 45168404Spjd * 46168404Spjd * This log is then shipped into an nvlist where the key is the dataset name and 47168404Spjd * the value is the object name. Userland is then responsible for uniquifying 48168404Spjd * this list and displaying it to the user. 49168404Spjd */ 50168404Spjd 51168404Spjd#include <sys/dmu_tx.h> 52168404Spjd#include <sys/spa.h> 53168404Spjd#include <sys/spa_impl.h> 54168404Spjd#include <sys/zap.h> 55168404Spjd#include <sys/zio.h> 56168404Spjd 57168404Spjd 58168404Spjd/* 59168404Spjd * Convert a bookmark to a string. 60168404Spjd */ 61168404Spjdstatic void 62268123Sdelphijbookmark_to_name(zbookmark_phys_t *zb, char *buf, size_t len) 63168404Spjd{ 64168404Spjd (void) snprintf(buf, len, "%llx:%llx:%llx:%llx", 65168404Spjd (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object, 66168404Spjd (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid); 67168404Spjd} 68168404Spjd 69168404Spjd/* 70168404Spjd * Convert a string to a bookmark 71168404Spjd */ 72168404Spjd#ifdef _KERNEL 73168404Spjdstatic void 74268123Sdelphijname_to_bookmark(char *buf, zbookmark_phys_t *zb) 75168404Spjd{ 76321578Smav zb->zb_objset = zfs_strtonum(buf, &buf); 77168404Spjd ASSERT(*buf == ':'); 78321578Smav zb->zb_object = zfs_strtonum(buf + 1, &buf); 79168404Spjd ASSERT(*buf == ':'); 80321578Smav zb->zb_level = (int)zfs_strtonum(buf + 1, &buf); 81168404Spjd ASSERT(*buf == ':'); 82321578Smav zb->zb_blkid = zfs_strtonum(buf + 1, &buf); 83168404Spjd ASSERT(*buf == '\0'); 84168404Spjd} 85168404Spjd#endif 86168404Spjd 87168404Spjd/* 88168404Spjd * Log an uncorrectable error to the persistent error log. We add it to the 89168404Spjd * spa's list of pending errors. The changes are actually synced out to disk 90168404Spjd * during spa_errlog_sync(). 91168404Spjd */ 92168404Spjdvoid 93168404Spjdspa_log_error(spa_t *spa, zio_t *zio) 94168404Spjd{ 95268123Sdelphij zbookmark_phys_t *zb = &zio->io_logical->io_bookmark; 96168404Spjd spa_error_entry_t search; 97168404Spjd spa_error_entry_t *new; 98168404Spjd avl_tree_t *tree; 99168404Spjd avl_index_t where; 100168404Spjd 101168404Spjd /* 102168404Spjd * If we are trying to import a pool, ignore any errors, as we won't be 103168404Spjd * writing to the pool any time soon. 104168404Spjd */ 105219089Spjd if (spa_load_state(spa) == SPA_LOAD_TRYIMPORT) 106168404Spjd return; 107168404Spjd 108168404Spjd mutex_enter(&spa->spa_errlist_lock); 109168404Spjd 110168404Spjd /* 111168404Spjd * If we have had a request to rotate the log, log it to the next list 112168404Spjd * instead of the current one. 113168404Spjd */ 114168404Spjd if (spa->spa_scrub_active || spa->spa_scrub_finished) 115168404Spjd tree = &spa->spa_errlist_scrub; 116168404Spjd else 117168404Spjd tree = &spa->spa_errlist_last; 118168404Spjd 119168404Spjd search.se_bookmark = *zb; 120168404Spjd if (avl_find(tree, &search, &where) != NULL) { 121168404Spjd mutex_exit(&spa->spa_errlist_lock); 122168404Spjd return; 123168404Spjd } 124168404Spjd 125168404Spjd new = kmem_zalloc(sizeof (spa_error_entry_t), KM_SLEEP); 126168404Spjd new->se_bookmark = *zb; 127168404Spjd avl_insert(tree, new, where); 128168404Spjd 129168404Spjd mutex_exit(&spa->spa_errlist_lock); 130168404Spjd} 131168404Spjd 132168404Spjd/* 133168404Spjd * Return the number of errors currently in the error log. This is actually the 134168404Spjd * sum of both the last log and the current log, since we don't know the union 135168404Spjd * of these logs until we reach userland. 136168404Spjd */ 137168404Spjduint64_t 138168404Spjdspa_get_errlog_size(spa_t *spa) 139168404Spjd{ 140168404Spjd uint64_t total = 0, count; 141168404Spjd 142168404Spjd mutex_enter(&spa->spa_errlog_lock); 143168404Spjd if (spa->spa_errlog_scrub != 0 && 144168404Spjd zap_count(spa->spa_meta_objset, spa->spa_errlog_scrub, 145168404Spjd &count) == 0) 146168404Spjd total += count; 147168404Spjd 148168404Spjd if (spa->spa_errlog_last != 0 && !spa->spa_scrub_finished && 149168404Spjd zap_count(spa->spa_meta_objset, spa->spa_errlog_last, 150168404Spjd &count) == 0) 151168404Spjd total += count; 152168404Spjd mutex_exit(&spa->spa_errlog_lock); 153168404Spjd 154168404Spjd mutex_enter(&spa->spa_errlist_lock); 155168404Spjd total += avl_numnodes(&spa->spa_errlist_last); 156168404Spjd total += avl_numnodes(&spa->spa_errlist_scrub); 157168404Spjd mutex_exit(&spa->spa_errlist_lock); 158168404Spjd 159168404Spjd return (total); 160168404Spjd} 161168404Spjd 162168404Spjd#ifdef _KERNEL 163168404Spjdstatic int 164168404Spjdprocess_error_log(spa_t *spa, uint64_t obj, void *addr, size_t *count) 165168404Spjd{ 166168404Spjd zap_cursor_t zc; 167168404Spjd zap_attribute_t za; 168268123Sdelphij zbookmark_phys_t zb; 169168404Spjd 170168404Spjd if (obj == 0) 171168404Spjd return (0); 172168404Spjd 173168404Spjd for (zap_cursor_init(&zc, spa->spa_meta_objset, obj); 174168404Spjd zap_cursor_retrieve(&zc, &za) == 0; 175168404Spjd zap_cursor_advance(&zc)) { 176168404Spjd 177168404Spjd if (*count == 0) { 178168404Spjd zap_cursor_fini(&zc); 179249195Smm return (SET_ERROR(ENOMEM)); 180168404Spjd } 181168404Spjd 182168404Spjd name_to_bookmark(za.za_name, &zb); 183168404Spjd 184168404Spjd if (copyout(&zb, (char *)addr + 185268123Sdelphij (*count - 1) * sizeof (zbookmark_phys_t), 186268123Sdelphij sizeof (zbookmark_phys_t)) != 0) { 187251632Sdelphij zap_cursor_fini(&zc); 188249195Smm return (SET_ERROR(EFAULT)); 189251632Sdelphij } 190168404Spjd 191168404Spjd *count -= 1; 192168404Spjd } 193168404Spjd 194168404Spjd zap_cursor_fini(&zc); 195168404Spjd 196168404Spjd return (0); 197168404Spjd} 198168404Spjd 199168404Spjdstatic int 200168404Spjdprocess_error_list(avl_tree_t *list, void *addr, size_t *count) 201168404Spjd{ 202168404Spjd spa_error_entry_t *se; 203168404Spjd 204168404Spjd for (se = avl_first(list); se != NULL; se = AVL_NEXT(list, se)) { 205168404Spjd 206168404Spjd if (*count == 0) 207249195Smm return (SET_ERROR(ENOMEM)); 208168404Spjd 209168404Spjd if (copyout(&se->se_bookmark, (char *)addr + 210268123Sdelphij (*count - 1) * sizeof (zbookmark_phys_t), 211268123Sdelphij sizeof (zbookmark_phys_t)) != 0) 212249195Smm return (SET_ERROR(EFAULT)); 213168404Spjd 214168404Spjd *count -= 1; 215168404Spjd } 216168404Spjd 217168404Spjd return (0); 218168404Spjd} 219168404Spjd#endif 220168404Spjd 221168404Spjd/* 222168404Spjd * Copy all known errors to userland as an array of bookmarks. This is 223168404Spjd * actually a union of the on-disk last log and current log, as well as any 224168404Spjd * pending error requests. 225168404Spjd * 226168404Spjd * Because the act of reading the on-disk log could cause errors to be 227168404Spjd * generated, we have two separate locks: one for the error log and one for the 228168404Spjd * in-core error lists. We only need the error list lock to log and error, so 229168404Spjd * we grab the error log lock while we read the on-disk logs, and only pick up 230168404Spjd * the error list lock when we are finished. 231168404Spjd */ 232168404Spjdint 233168404Spjdspa_get_errlog(spa_t *spa, void *uaddr, size_t *count) 234168404Spjd{ 235168404Spjd int ret = 0; 236168404Spjd 237168404Spjd#ifdef _KERNEL 238168404Spjd mutex_enter(&spa->spa_errlog_lock); 239168404Spjd 240168404Spjd ret = process_error_log(spa, spa->spa_errlog_scrub, uaddr, count); 241168404Spjd 242168404Spjd if (!ret && !spa->spa_scrub_finished) 243168404Spjd ret = process_error_log(spa, spa->spa_errlog_last, uaddr, 244168404Spjd count); 245168404Spjd 246168404Spjd mutex_enter(&spa->spa_errlist_lock); 247168404Spjd if (!ret) 248168404Spjd ret = process_error_list(&spa->spa_errlist_scrub, uaddr, 249168404Spjd count); 250168404Spjd if (!ret) 251168404Spjd ret = process_error_list(&spa->spa_errlist_last, uaddr, 252168404Spjd count); 253168404Spjd mutex_exit(&spa->spa_errlist_lock); 254168404Spjd 255168404Spjd mutex_exit(&spa->spa_errlog_lock); 256168404Spjd#endif 257168404Spjd 258168404Spjd return (ret); 259168404Spjd} 260168404Spjd 261168404Spjd/* 262168404Spjd * Called when a scrub completes. This simply set a bit which tells which AVL 263168404Spjd * tree to add new errors. spa_errlog_sync() is responsible for actually 264168404Spjd * syncing the changes to the underlying objects. 265168404Spjd */ 266168404Spjdvoid 267168404Spjdspa_errlog_rotate(spa_t *spa) 268168404Spjd{ 269168404Spjd mutex_enter(&spa->spa_errlist_lock); 270168404Spjd spa->spa_scrub_finished = B_TRUE; 271168404Spjd mutex_exit(&spa->spa_errlist_lock); 272168404Spjd} 273168404Spjd 274168404Spjd/* 275168404Spjd * Discard any pending errors from the spa_t. Called when unloading a faulted 276168404Spjd * pool, as the errors encountered during the open cannot be synced to disk. 277168404Spjd */ 278168404Spjdvoid 279168404Spjdspa_errlog_drain(spa_t *spa) 280168404Spjd{ 281168404Spjd spa_error_entry_t *se; 282168404Spjd void *cookie; 283168404Spjd 284168404Spjd mutex_enter(&spa->spa_errlist_lock); 285168404Spjd 286168404Spjd cookie = NULL; 287168404Spjd while ((se = avl_destroy_nodes(&spa->spa_errlist_last, 288168404Spjd &cookie)) != NULL) 289168404Spjd kmem_free(se, sizeof (spa_error_entry_t)); 290168404Spjd cookie = NULL; 291168404Spjd while ((se = avl_destroy_nodes(&spa->spa_errlist_scrub, 292168404Spjd &cookie)) != NULL) 293168404Spjd kmem_free(se, sizeof (spa_error_entry_t)); 294168404Spjd 295168404Spjd mutex_exit(&spa->spa_errlist_lock); 296168404Spjd} 297168404Spjd 298168404Spjd/* 299168404Spjd * Process a list of errors into the current on-disk log. 300168404Spjd */ 301168404Spjdstatic void 302168404Spjdsync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj, dmu_tx_t *tx) 303168404Spjd{ 304168404Spjd spa_error_entry_t *se; 305168404Spjd char buf[64]; 306168404Spjd void *cookie; 307168404Spjd 308168404Spjd if (avl_numnodes(t) != 0) { 309168404Spjd /* create log if necessary */ 310168404Spjd if (*obj == 0) 311168404Spjd *obj = zap_create(spa->spa_meta_objset, 312168404Spjd DMU_OT_ERROR_LOG, DMU_OT_NONE, 313168404Spjd 0, tx); 314168404Spjd 315168404Spjd /* add errors to the current log */ 316168404Spjd for (se = avl_first(t); se != NULL; se = AVL_NEXT(t, se)) { 317168404Spjd char *name = se->se_name ? se->se_name : ""; 318168404Spjd 319168404Spjd bookmark_to_name(&se->se_bookmark, buf, sizeof (buf)); 320168404Spjd 321168404Spjd (void) zap_update(spa->spa_meta_objset, 322168404Spjd *obj, buf, 1, strlen(name) + 1, name, tx); 323168404Spjd } 324168404Spjd 325168404Spjd /* purge the error list */ 326168404Spjd cookie = NULL; 327168404Spjd while ((se = avl_destroy_nodes(t, &cookie)) != NULL) 328168404Spjd kmem_free(se, sizeof (spa_error_entry_t)); 329168404Spjd } 330168404Spjd} 331168404Spjd 332168404Spjd/* 333168404Spjd * Sync the error log out to disk. This is a little tricky because the act of 334168404Spjd * writing the error log requires the spa_errlist_lock. So, we need to lock the 335168404Spjd * error lists, take a copy of the lists, and then reinitialize them. Then, we 336168404Spjd * drop the error list lock and take the error log lock, at which point we 337168404Spjd * do the errlog processing. Then, if we encounter an I/O error during this 338168404Spjd * process, we can successfully add the error to the list. Note that this will 339168404Spjd * result in the perpetual recycling of errors, but it is an unlikely situation 340168404Spjd * and not a performance critical operation. 341168404Spjd */ 342168404Spjdvoid 343168404Spjdspa_errlog_sync(spa_t *spa, uint64_t txg) 344168404Spjd{ 345168404Spjd dmu_tx_t *tx; 346168404Spjd avl_tree_t scrub, last; 347168404Spjd int scrub_finished; 348168404Spjd 349168404Spjd mutex_enter(&spa->spa_errlist_lock); 350168404Spjd 351168404Spjd /* 352168404Spjd * Bail out early under normal circumstances. 353168404Spjd */ 354168404Spjd if (avl_numnodes(&spa->spa_errlist_scrub) == 0 && 355168404Spjd avl_numnodes(&spa->spa_errlist_last) == 0 && 356168404Spjd !spa->spa_scrub_finished) { 357168404Spjd mutex_exit(&spa->spa_errlist_lock); 358168404Spjd return; 359168404Spjd } 360168404Spjd 361168404Spjd spa_get_errlists(spa, &last, &scrub); 362168404Spjd scrub_finished = spa->spa_scrub_finished; 363168404Spjd spa->spa_scrub_finished = B_FALSE; 364168404Spjd 365168404Spjd mutex_exit(&spa->spa_errlist_lock); 366168404Spjd mutex_enter(&spa->spa_errlog_lock); 367168404Spjd 368168404Spjd tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 369168404Spjd 370168404Spjd /* 371168404Spjd * Sync out the current list of errors. 372168404Spjd */ 373168404Spjd sync_error_list(spa, &last, &spa->spa_errlog_last, tx); 374168404Spjd 375168404Spjd /* 376168404Spjd * Rotate the log if necessary. 377168404Spjd */ 378168404Spjd if (scrub_finished) { 379168404Spjd if (spa->spa_errlog_last != 0) 380168404Spjd VERIFY(dmu_object_free(spa->spa_meta_objset, 381168404Spjd spa->spa_errlog_last, tx) == 0); 382168404Spjd spa->spa_errlog_last = spa->spa_errlog_scrub; 383168404Spjd spa->spa_errlog_scrub = 0; 384168404Spjd 385168404Spjd sync_error_list(spa, &scrub, &spa->spa_errlog_last, tx); 386168404Spjd } 387168404Spjd 388168404Spjd /* 389168404Spjd * Sync out any pending scrub errors. 390168404Spjd */ 391168404Spjd sync_error_list(spa, &scrub, &spa->spa_errlog_scrub, tx); 392168404Spjd 393168404Spjd /* 394168404Spjd * Update the MOS to reflect the new values. 395168404Spjd */ 396168404Spjd (void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 397168404Spjd DMU_POOL_ERRLOG_LAST, sizeof (uint64_t), 1, 398168404Spjd &spa->spa_errlog_last, tx); 399168404Spjd (void) zap_update(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 400168404Spjd DMU_POOL_ERRLOG_SCRUB, sizeof (uint64_t), 1, 401168404Spjd &spa->spa_errlog_scrub, tx); 402168404Spjd 403168404Spjd dmu_tx_commit(tx); 404168404Spjd 405168404Spjd mutex_exit(&spa->spa_errlog_lock); 406168404Spjd} 407