lustre/obdclass/llog.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *   Author: Andreas Dilger <adilger@clusterfs.com>
   6  *
   7  *   This file is part of the Lustre file system, http://www.lustre.org
   8  *   Lustre is a trademark of Cluster File Systems, Inc.
   9  *
  10  *   You may have signed or agreed to another license before downloading
  11  *   this software.  If so, you are bound by the terms and conditions
  12  *   of that agreement, and the following does not apply to you.  See the
  13  *   LICENSE file included with this distribution for more information.
  14  *
  15  *   If you did not agree to a different license, then this copy of Lustre
  16  *   is open source software; you can redistribute it and/or modify it
  17  *   under the terms of version 2 of the GNU General Public License as
  18  *   published by the Free Software Foundation.
  19  *
  20  *   In either case, Lustre is distributed in the hope that it will be
  21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  23  *   license text for more details.
  24  *
  25  * OST<->MDS recovery logging infrastructure.
  26  *
  27  * Invariants in implementation:
  28  * - we do not share logs among different OST<->MDS connections, so that
  29  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_LOG
  33
  34 #ifndef EXPORT_SYMTAB
  35 #define EXPORT_SYMTAB
  36 #endif
  37
  38 #ifdef __KERNEL__
  39 #include <linux/fs.h>
  40 #else
  41 #include <liblustre.h>
  42 #endif
  43
  44 #include <linux/obd_class.h>
  45 #include <linux/lustre_log.h>
  46 #include <libcfs/list.h>
  47
  48 /* Allocate a new log or catalog handle */
  49 struct llog_handle *llog_alloc_handle(void)
  50 {
  51         struct llog_handle *loghandle;
  52         ENTRY;
  53
  54         OBD_ALLOC(loghandle, sizeof(*loghandle));
  55         if (loghandle == NULL)
  56                 RETURN(ERR_PTR(-ENOMEM));
  57
  58         init_rwsem(&loghandle->lgh_lock);
  59
  60         RETURN(loghandle);
  61 }
  62 EXPORT_SYMBOL(llog_alloc_handle);
  63
  64
  65 void llog_free_handle(struct llog_handle *loghandle)
  66 {
  67         if (!loghandle)
  68                 return;
  69
  70         if (!loghandle->lgh_hdr)
  71                 goto out;
  72         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  73                 list_del_init(&loghandle->u.phd.phd_entry);
  74         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  75                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  76         OBD_FREE(loghandle->lgh_hdr, LLOG_CHUNK_SIZE);
  77
  78  out:
  79         OBD_FREE(loghandle, sizeof(*loghandle));
  80 }
  81 EXPORT_SYMBOL(llog_free_handle);
  82
  83 /* returns negative on error; 0 if success; 1 if success & log destroyed */
  84 int llog_cancel_rec(struct llog_handle *loghandle, int index)
  85 {
  86         struct llog_log_hdr *llh = loghandle->lgh_hdr;
  87         int rc = 0;
  88         ENTRY;
  89
  90         CDEBUG(D_HA, "canceling %d in log "LPX64"\n",
  91                index, loghandle->lgh_id.lgl_oid);
  92
  93         if (index == 0) {
  94                 CERROR("cannot cancel index 0 (which is header)\n");
  95                 RETURN(-EINVAL);
  96         }
  97
  98         if (!ext2_clear_bit(index, llh->llh_bitmap)) {
  99                 CDEBUG(D_HA, "catalog index %u already clear?\n", index);
 100                 RETURN(-EINVAL);
 101         }
 102
 103         llh->llh_count--;
 104
 105         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 106             (llh->llh_count == 1) &&
 107             (loghandle->lgh_last_idx == (LLOG_BITMAP_BYTES * 8) - 1)) {
 108                 rc = llog_destroy(loghandle);
 109                 if (rc) {
 110                         CERROR("failure destroying log after last cancel: %d\n",
 111                                rc);
 112                         ext2_set_bit(index, llh->llh_bitmap);
 113                         llh->llh_count++;
 114                 } else {
 115                         rc = 1;
 116                 }
 117                 RETURN(rc);
 118         }
 119
 120         rc = llog_write_rec(loghandle, &llh->llh_hdr, NULL, 0, NULL, 0);
 121         if (rc) {
 122                 CERROR("failure re-writing header %d\n", rc);
 123                 ext2_set_bit(index, llh->llh_bitmap);
 124                 llh->llh_count++;
 125         }
 126         RETURN(rc);
 127 }
 128 EXPORT_SYMBOL(llog_cancel_rec);
 129
 130 int llog_init_handle(struct llog_handle *handle, int flags,
 131                      struct obd_uuid *uuid)
 132 {
 133         int rc;
 134         struct llog_log_hdr *llh;
 135         ENTRY;
 136         LASSERT(handle->lgh_hdr == NULL);
 137
 138         OBD_ALLOC(llh, sizeof(*llh));
 139         if (llh == NULL)
 140                 RETURN(-ENOMEM);
 141         handle->lgh_hdr = llh;
 142         /* first assign flags to use llog_client_ops */
 143         llh->llh_flags = flags;
 144         rc = llog_read_header(handle);
 145         if (rc == 0) {
 146                 flags = llh->llh_flags;
 147                 if (uuid)
 148                         LASSERT(obd_uuid_equals(uuid, &llh->llh_tgtuuid));
 149                 GOTO(out, rc);
 150         } else if (rc != LLOG_EEMPTY || !flags) {
 151                 /* set a pesudo flag for initialization */
 152                 flags = LLOG_F_IS_CAT;
 153                 GOTO(out, rc);
 154         }
 155         rc = 0;
 156
 157         handle->lgh_last_idx = 0; /* header is record with index 0 */
 158         llh->llh_count = 1;         /* for the header record */
 159         llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 160         llh->llh_hdr.lrh_len = llh->llh_tail.lrt_len = LLOG_CHUNK_SIZE;
 161         llh->llh_hdr.lrh_index = llh->llh_tail.lrt_index = 0;
 162         llh->llh_timestamp = CURRENT_SECONDS;
 163         if (uuid)
 164                 memcpy(&llh->llh_tgtuuid, uuid, sizeof(llh->llh_tgtuuid));
 165         llh->llh_bitmap_offset = offsetof(typeof(*llh),llh_bitmap);
 166         ext2_set_bit(0, llh->llh_bitmap);
 167
 168 out:
 169         if (flags & LLOG_F_IS_CAT) {
 170                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 171                 llh->llh_size = sizeof(struct llog_logid_rec);
 172         } else if (flags & LLOG_F_IS_PLAIN) {
 173                 INIT_LIST_HEAD(&handle->u.phd.phd_entry);
 174         } else {
 175                 CERROR("Unknown flags: %#x (Expected %#x or %#x\n",
 176                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 177                 LBUG();
 178         }
 179
 180         if (rc) {
 181                 OBD_FREE(llh, sizeof(*llh));
 182                 handle->lgh_hdr = NULL;
 183         }
 184         RETURN(rc);
 185 }
 186 EXPORT_SYMBOL(llog_init_handle);
 187
 188 int llog_close(struct llog_handle *loghandle)
 189 {
 190         struct llog_operations *lop;
 191         int rc;
 192         ENTRY;
 193
 194         rc = llog_handle2ops(loghandle, &lop);
 195         if (rc)
 196                 GOTO(out, rc);
 197         if (lop->lop_close == NULL)
 198                 GOTO(out, -EOPNOTSUPP);
 199         rc = lop->lop_close(loghandle);
 200  out:
 201         llog_free_handle(loghandle);
 202         RETURN(rc);
 203 }
 204 EXPORT_SYMBOL(llog_close);
 205
 206 int llog_process(struct llog_handle *loghandle, llog_cb_t cb,
 207                  void *data, void *catdata)
 208 {
 209         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 210         struct llog_process_cat_data *cd = catdata;
 211         char *buf;
 212         __u64 cur_offset = LLOG_CHUNK_SIZE;
 213         int rc = 0, index = 1, last_index;
 214         int saved_index = 0, last_called_index = 0;
 215         ENTRY;
 216
 217         LASSERT(llh);
 218
 219         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 220         if (!buf)
 221                 RETURN(-ENOMEM);
 222
 223         if (cd != NULL) {
 224                 last_called_index = cd->first_idx;
 225                 index = cd->first_idx + 1;
 226         }
 227         if (cd != NULL && cd->last_idx)
 228                 last_index = cd->last_idx;
 229         else
 230                 last_index = LLOG_BITMAP_BYTES * 8 - 1;
 231
 232         while (rc == 0) {
 233                 struct llog_rec_hdr *rec;
 234
 235                 /* skip records not set in bitmap */
 236                 while (index <= last_index &&
 237                        !ext2_test_bit(index, llh->llh_bitmap))
 238                         ++index;
 239
 240                 LASSERT(index <= last_index + 1);
 241                 if (index == last_index + 1)
 242                         break;
 243
 244                 CDEBUG(D_OTHER, "index: %d last_index %d\n",
 245                        index, last_index);
 246
 247                 /* get the buf with our target record; avoid old garbage */
 248                 memset(buf, 0, LLOG_CHUNK_SIZE);
 249                 rc = llog_next_block(loghandle, &saved_index, index,
 250                                      &cur_offset, buf, LLOG_CHUNK_SIZE);
 251                 if (rc)
 252                         GOTO(out, rc);
 253
 254                 /* NB: when rec->lrh_len is accessed it is already swabbed
 255                  * since it is used at the "end" of the loop and the rec
 256                  * swabbing is done at the beginning of the loop. */
 257                 for (rec = (struct llog_rec_hdr *)buf;
 258                      (char *)rec < buf + LLOG_CHUNK_SIZE;
 259                      rec = (struct llog_rec_hdr *)((char *)rec + rec->lrh_len)){
 260
 261                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 262                                rec, rec->lrh_type);
 263
 264                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 265                                 lustre_swab_llog_rec(rec, NULL);
 266
 267                         CDEBUG(D_OTHER, "after swabbing, type: %#x\n",
 268                                rec->lrh_type);
 269
 270                         if (rec->lrh_index == 0)
 271                                 GOTO(out, 0); /* no more records */
 272
 273                         if (rec->lrh_len == 0 || rec->lrh_len >LLOG_CHUNK_SIZE){
 274                                 CWARN("invalid length %d in llog record for "
 275                                       "index %d\n", rec->lrh_len,
 276                                 rec->lrh_index);
 277                                 GOTO(out, 0);
 278                         }
 279
 280                         if (rec->lrh_index < index) {
 281                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 282                                        rec->lrh_index);
 283                                 continue;
 284                         }
 285
 286                         CDEBUG(D_OTHER,
 287                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 288                                rec->lrh_index, rec->lrh_len,
 289                                (int)(buf + LLOG_CHUNK_SIZE - (char *)rec));
 290
 291                         /* if set, process the callback on this record */
 292                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 293                                 rc = cb(loghandle, rec, data);
 294                                 last_called_index = index;
 295                                 if (rc == LLOG_PROC_BREAK) {
 296                                         CWARN("recovery from log: "LPX64":%x"
 297                                               " stopped\n",
 298                                               loghandle->lgh_id.lgl_oid,
 299                                               loghandle->lgh_id.lgl_ogen);
 300                                         GOTO(out, rc);
 301                                 } else if (rc == LLOG_DEL_RECORD) {
 302                                         llog_cancel_rec(loghandle, rec->lrh_index);
 303                                         rc = 0;
 304                                 }
 305                                 if (rc)
 306                                         GOTO(out, rc);
 307                         } else {
 308                                 CDEBUG(D_OTHER, "Skipped index %d\n", index);
 309                         }
 310
 311                         /* next record, still in buffer? */
 312                         ++index;
 313                         if (index > last_index)
 314                                 GOTO(out, rc = 0);
 315                 }
 316         }
 317
 318  out:
 319         if (cd != NULL)
 320                 cd->last_idx = last_called_index;
 321         if (buf)
 322                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 323         RETURN(rc);
 324 }
 325 EXPORT_SYMBOL(llog_process);
 326
 327 int llog_get_size(struct llog_handle *loghandle)
 328 {
 329         if (loghandle && loghandle->lgh_hdr)
 330                 return loghandle->lgh_hdr->llh_count;
 331         return 0;
 332 }
 333 EXPORT_SYMBOL(llog_get_size);
 334
 335 int llog_reverse_process(struct llog_handle *loghandle, llog_cb_t cb,
 336                          void *data, void *catdata)
 337 {
 338         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 339         struct llog_process_cat_data *cd = catdata;
 340         void *buf;
 341         int rc = 0, first_index = 1, index, idx;
 342         struct llog_rec_tail *tail;
 343         ENTRY;
 344
 345         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 346         if (!buf)
 347                 RETURN(-ENOMEM);
 348
 349         if (cd != NULL)
 350                 first_index = cd->first_idx + 1;
 351         if (cd != NULL && cd->last_idx)
 352                 index = cd->last_idx;
 353         else
 354                 index = LLOG_BITMAP_BYTES * 8 - 1;
 355
 356         while (rc == 0) {
 357                 struct llog_rec_hdr *rec;
 358
 359                 /* skip records not set in bitmap */
 360                 while (index >= first_index &&
 361                        !ext2_test_bit(index, llh->llh_bitmap))
 362                         --index;
 363
 364                 LASSERT(index >= first_index - 1);
 365                 if (index == first_index - 1)
 366                         break;
 367
 368                 /* get the buf with our target record; avoid old garbage */
 369                 memset(buf, 0, LLOG_CHUNK_SIZE);
 370                 rc = llog_prev_block(loghandle, index, buf, LLOG_CHUNK_SIZE);
 371                 if (rc)
 372                         GOTO(out, rc);
 373
 374                 rec = buf;
 375                 idx = le32_to_cpu(rec->lrh_index);
 376                 if (idx < index)
 377                         CDEBUG(D_HA, "index %u : idx %u\n", index, idx);
 378                 while (idx < index) {
 379                         rec = ((void *)rec + le32_to_cpu(rec->lrh_len));
 380                         idx ++;
 381                 }
 382
 383                 /* process records in buffer, starting where we found one */
 384                 while ((void *)rec >= buf) {
 385                         if (rec->lrh_index == 0)
 386                                 GOTO(out, 0); /* no more records */
 387
 388                         /* if set, process the callback on this record */
 389                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 390                                 rc = cb(loghandle, rec, data);
 391                                 if (rc == LLOG_PROC_BREAK) {
 392                                         CWARN("recovery from log: "LPX64":%x"
 393                                               " stopped\n",
 394                                               loghandle->lgh_id.lgl_oid,
 395                                               loghandle->lgh_id.lgl_ogen);
 396                                         GOTO(out, rc);
 397                                 }
 398                                 if (rc)
 399                                         GOTO(out, rc);
 400                         }
 401
 402                         /* previous record, still in buffer? */
 403                         --index;
 404                         if (index < first_index)
 405                                 GOTO(out, rc = 0);
 406                         tail = (void *)rec - sizeof(struct llog_rec_tail);
 407                         rec = ((void *)rec - le32_to_cpu(tail->lrt_len));
 408                 }
 409         }
 410
 411 out:
 412         if (buf)
 413                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 414         RETURN(rc);
 415 }
 416 EXPORT_SYMBOL(llog_reverse_process);