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 #ifndef __KERNEL__
  39 #include <liblustre.h>
  40 #endif
  41
  42 #include <obd_class.h>
  43 #include <lustre_log.h>
  44 #include <libcfs/list.h>
  45
  46 /* Allocate a new log or catalog handle */
  47 struct llog_handle *llog_alloc_handle(void)
  48 {
  49         struct llog_handle *loghandle;
  50         ENTRY;
  51
  52         OBD_ALLOC(loghandle, sizeof(*loghandle));
  53         if (loghandle == NULL)
  54                 RETURN(ERR_PTR(-ENOMEM));
  55
  56         init_rwsem(&loghandle->lgh_lock);
  57
  58         RETURN(loghandle);
  59 }
  60 EXPORT_SYMBOL(llog_alloc_handle);
  61
  62
  63 void llog_free_handle(struct llog_handle *loghandle)
  64 {
  65         if (!loghandle)
  66                 return;
  67
  68         if (!loghandle->lgh_hdr)
  69                 goto out;
  70         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  71                 list_del_init(&loghandle->u.phd.phd_entry);
  72         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  73                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  74         OBD_FREE(loghandle->lgh_hdr, LLOG_CHUNK_SIZE);
  75
  76  out:
  77         OBD_FREE(loghandle, sizeof(*loghandle));
  78 }
  79 EXPORT_SYMBOL(llog_free_handle);
  80
  81 /* returns negative on error; 0 if success; 1 if success & log destroyed */
  82 int llog_cancel_rec(struct llog_handle *loghandle, int index)
  83 {
  84         struct llog_log_hdr *llh = loghandle->lgh_hdr;
  85         int rc = 0;
  86         ENTRY;
  87
  88         CDEBUG(D_RPCTRACE, "canceling %d in log "LPX64"\n",
  89                index, loghandle->lgh_id.lgl_oid);
  90
  91         if (index == 0) {
  92                 CERROR("cannot cancel index 0 (which is header)\n");
  93                 RETURN(-EINVAL);
  94         }
  95
  96         if (!ext2_clear_bit(index, llh->llh_bitmap)) {
  97                 CDEBUG(D_RPCTRACE, "catalog index %u already clear?\n", index);
  98                 RETURN(-EINVAL);
  99         }
 100
 101         llh->llh_count--;
 102
 103         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 104             (llh->llh_count == 1) &&
 105             (loghandle->lgh_last_idx == (LLOG_BITMAP_BYTES * 8) - 1)) {
 106                 rc = llog_destroy(loghandle);
 107                 if (rc) {
 108                         CERROR("failure destroying log after last cancel: %d\n",
 109                                rc);
 110                         ext2_set_bit(index, llh->llh_bitmap);
 111                         llh->llh_count++;
 112                 } else {
 113                         rc = 1;
 114                 }
 115                 RETURN(rc);
 116         }
 117
 118         rc = llog_write_rec(loghandle, &llh->llh_hdr, NULL, 0, NULL, 0);
 119         if (rc) {
 120                 CERROR("failure re-writing header %d\n", rc);
 121                 ext2_set_bit(index, llh->llh_bitmap);
 122                 llh->llh_count++;
 123         }
 124         RETURN(rc);
 125 }
 126 EXPORT_SYMBOL(llog_cancel_rec);
 127
 128 int llog_init_handle(struct llog_handle *handle, int flags,
 129                      struct obd_uuid *uuid)
 130 {
 131         int rc;
 132         struct llog_log_hdr *llh;
 133         ENTRY;
 134         LASSERT(handle->lgh_hdr == NULL);
 135
 136         OBD_ALLOC(llh, sizeof(*llh));
 137         if (llh == NULL)
 138                 RETURN(-ENOMEM);
 139         handle->lgh_hdr = llh;
 140         /* first assign flags to use llog_client_ops */
 141         llh->llh_flags = flags;
 142         rc = llog_read_header(handle);
 143         if (rc == 0) {
 144                 flags = llh->llh_flags;
 145                 if (uuid && !obd_uuid_equals(uuid, &llh->llh_tgtuuid)) {
 146                         CERROR("uuid mismatch: %s/%s\n", (char *)uuid->uuid,
 147                                (char *)llh->llh_tgtuuid.uuid);
 148                         rc = -EEXIST;
 149                 }
 150                 GOTO(out, rc);
 151         } else if (rc != LLOG_EEMPTY || !flags) {
 152                 /* set a pesudo flag for initialization */
 153                 flags = LLOG_F_IS_CAT;
 154                 GOTO(out, rc);
 155         }
 156         rc = 0;
 157
 158         handle->lgh_last_idx = 0; /* header is record with index 0 */
 159         llh->llh_count = 1;         /* for the header record */
 160         llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 161         llh->llh_hdr.lrh_len = llh->llh_tail.lrt_len = LLOG_CHUNK_SIZE;
 162         llh->llh_hdr.lrh_index = llh->llh_tail.lrt_index = 0;
 163         llh->llh_timestamp = CURRENT_SECONDS;
 164         if (uuid)
 165                 memcpy(&llh->llh_tgtuuid, uuid, sizeof(llh->llh_tgtuuid));
 166         llh->llh_bitmap_offset = offsetof(typeof(*llh),llh_bitmap);
 167         ext2_set_bit(0, llh->llh_bitmap);
 168
 169 out:
 170         if (flags & LLOG_F_IS_CAT) {
 171                 CFS_INIT_LIST_HEAD(&handle->u.chd.chd_head);
 172                 llh->llh_size = sizeof(struct llog_logid_rec);
 173         } else if (flags & LLOG_F_IS_PLAIN) {
 174                 CFS_INIT_LIST_HEAD(&handle->u.phd.phd_entry);
 175         } else {
 176                 CERROR("Unknown flags: %#x (Expected %#x or %#x\n",
 177                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 178                 LBUG();
 179         }
 180
 181         if (rc) {
 182                 OBD_FREE(llh, sizeof(*llh));
 183                 handle->lgh_hdr = NULL;
 184         }
 185         RETURN(rc);
 186 }
 187 EXPORT_SYMBOL(llog_init_handle);
 188
 189 int llog_close(struct llog_handle *loghandle)
 190 {
 191         struct llog_operations *lop;
 192         int rc;
 193         ENTRY;
 194
 195         rc = llog_handle2ops(loghandle, &lop);
 196         if (rc)
 197                 GOTO(out, rc);
 198         if (lop->lop_close == NULL)
 199                 GOTO(out, -EOPNOTSUPP);
 200         rc = lop->lop_close(loghandle);
 201  out:
 202         llog_free_handle(loghandle);
 203         RETURN(rc);
 204 }
 205 EXPORT_SYMBOL(llog_close);
 206
 207 int llog_process(struct llog_handle *loghandle, llog_cb_t cb,
 208                  void *data, void *catdata)
 209 {
 210         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 211         struct llog_process_cat_data *cd = catdata;
 212         char *buf;
 213         __u64 cur_offset = LLOG_CHUNK_SIZE, last_offset;
 214         int rc = 0, index = 1, last_index;
 215         int saved_index = 0, last_called_index = 0;
 216         ENTRY;
 217
 218         LASSERT(llh);
 219
 220         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 221         if (!buf)
 222                 RETURN(-ENOMEM);
 223
 224         if (cd != NULL) {
 225                 last_called_index = cd->first_idx;
 226                 index = cd->first_idx + 1;
 227         }
 228         if (cd != NULL && cd->last_idx)
 229                 last_index = cd->last_idx;
 230         else
 231                 last_index = LLOG_BITMAP_BYTES * 8 - 1;
 232
 233         while (rc == 0) {
 234                 struct llog_rec_hdr *rec;
 235
 236                 /* skip records not set in bitmap */
 237                 while (index <= last_index &&
 238                        !ext2_test_bit(index, llh->llh_bitmap))
 239                         ++index;
 240
 241                 LASSERT(index <= last_index + 1);
 242                 if (index == last_index + 1)
 243                         break;
 244
 245                 CDEBUG(D_OTHER, "index: %d last_index %d\n",
 246                        index, last_index);
 247
 248                 /* get the buf with our target record; avoid old garbage */
 249                 memset(buf, 0, LLOG_CHUNK_SIZE);
 250                 last_offset = cur_offset;
 251                 rc = llog_next_block(loghandle, &saved_index, index,
 252                                      &cur_offset, buf, LLOG_CHUNK_SIZE);
 253                 if (rc)
 254                         GOTO(out, rc);
 255
 256                 /* NB: when rec->lrh_len is accessed it is already swabbed
 257                  * since it is used at the "end" of the loop and the rec
 258                  * swabbing is done at the beginning of the loop. */
 259                 for (rec = (struct llog_rec_hdr *)buf;
 260                      (char *)rec < buf + LLOG_CHUNK_SIZE;
 261                      rec = (struct llog_rec_hdr *)((char *)rec + rec->lrh_len)){
 262
 263                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 264                                rec, rec->lrh_type);
 265
 266                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 267                                 lustre_swab_llog_rec(rec, NULL);
 268
 269                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 270                                rec->lrh_type, rec->lrh_index);
 271
 272                         if (rec->lrh_index == 0)
 273                                 GOTO(out, 0); /* no more records */
 274
 275                         if (rec->lrh_len == 0 || rec->lrh_len >LLOG_CHUNK_SIZE){
 276                                 CWARN("invalid length %d in llog record for "
 277                                       "index %d/%d\n", rec->lrh_len,
 278                                       rec->lrh_index, index);
 279                                 GOTO(out, rc = -EINVAL);
 280                         }
 281
 282                         if (rec->lrh_index < index) {
 283                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 284                                        rec->lrh_index);
 285                                 continue;
 286                         }
 287
 288                         CDEBUG(D_OTHER,
 289                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 290                                rec->lrh_index, rec->lrh_len,
 291                                (int)(buf + LLOG_CHUNK_SIZE - (char *)rec));
 292
 293                         loghandle->lgh_cur_idx = rec->lrh_index;
 294                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 295                                 last_offset;
 296
 297                         /* if set, process the callback on this record */
 298                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 299                                 rc = cb(loghandle, rec, data);
 300                                 last_called_index = index;
 301                                 if (rc == LLOG_PROC_BREAK) {
 302                                         CDEBUG(D_HA, "recovery from log: "LPX64
 303                                                ":%x stopped\n",
 304                                                loghandle->lgh_id.lgl_oid,
 305                                                loghandle->lgh_id.lgl_ogen);
 306                                         GOTO(out, rc);
 307                                 } else if (rc == LLOG_DEL_RECORD) {
 308                                         llog_cancel_rec(loghandle, rec->lrh_index);
 309                                         rc = 0;
 310                                 }
 311                                 if (rc)
 312                                         GOTO(out, rc);
 313                         } else {
 314                                 CDEBUG(D_OTHER, "Skipped index %d\n", index);
 315                         }
 316
 317                         /* next record, still in buffer? */
 318                         ++index;
 319                         if (index > last_index)
 320                                 GOTO(out, rc = 0);
 321                 }
 322         }
 323
 324  out:
 325         if (cd != NULL)
 326                 cd->last_idx = last_called_index;
 327         if (buf)
 328                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 329         RETURN(rc);
 330 }
 331 EXPORT_SYMBOL(llog_process);
 332
 333 inline int llog_get_size(struct llog_handle *loghandle)
 334 {
 335         if (loghandle && loghandle->lgh_hdr)
 336                 return loghandle->lgh_hdr->llh_count;
 337         return 0;
 338 }
 339 EXPORT_SYMBOL(llog_get_size);
 340
 341 int llog_reverse_process(struct llog_handle *loghandle, llog_cb_t cb,
 342                          void *data, void *catdata)
 343 {
 344         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 345         struct llog_process_cat_data *cd = catdata;
 346         void *buf;
 347         int rc = 0, first_index = 1, index, idx;
 348         ENTRY;
 349
 350         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 351         if (!buf)
 352                 RETURN(-ENOMEM);
 353
 354         if (cd != NULL)
 355                 first_index = cd->first_idx + 1;
 356         if (cd != NULL && cd->last_idx)
 357                 index = cd->last_idx;
 358         else
 359                 index = LLOG_BITMAP_BYTES * 8 - 1;
 360
 361         while (rc == 0) {
 362                 struct llog_rec_hdr *rec;
 363                 struct llog_rec_tail *tail;
 364
 365                 /* skip records not set in bitmap */
 366                 while (index >= first_index &&
 367                        !ext2_test_bit(index, llh->llh_bitmap))
 368                         --index;
 369
 370                 LASSERT(index >= first_index - 1);
 371                 if (index == first_index - 1)
 372                         break;
 373
 374                 /* get the buf with our target record; avoid old garbage */
 375                 memset(buf, 0, LLOG_CHUNK_SIZE);
 376                 rc = llog_prev_block(loghandle, index, buf, LLOG_CHUNK_SIZE);
 377                 if (rc)
 378                         GOTO(out, rc);
 379
 380                 rec = buf;
 381                 idx = le32_to_cpu(rec->lrh_index);
 382                 if (idx < index)
 383                         CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 384                 while (idx < index) {
 385                         rec = ((void *)rec + le32_to_cpu(rec->lrh_len));
 386                         idx ++;
 387                 }
 388                 tail = (void *)rec + le32_to_cpu(rec->lrh_len) - sizeof(*tail);
 389
 390                 /* process records in buffer, starting where we found one */
 391                 while ((void *)tail > buf) {
 392                         rec = (void *)tail - le32_to_cpu(tail->lrt_len) +
 393                                 sizeof(*tail);
 394
 395                         if (rec->lrh_index == 0)
 396                                 GOTO(out, 0); /* no more records */
 397
 398                         /* if set, process the callback on this record */
 399                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 400                                 rc = cb(loghandle, rec, data);
 401                                 if (rc == LLOG_PROC_BREAK) {
 402                                         CWARN("recovery from log: "LPX64":%x"
 403                                               " stopped\n",
 404                                               loghandle->lgh_id.lgl_oid,
 405                                               loghandle->lgh_id.lgl_ogen);
 406                                         GOTO(out, rc);
 407                                 }
 408                                 if (rc)
 409                                         GOTO(out, rc);
 410                         }
 411
 412                         /* previous record, still in buffer? */
 413                         --index;
 414                         if (index < first_index)
 415                                 GOTO(out, rc = 0);
 416                         tail = (void *)rec - sizeof(*tail);
 417                 }
 418         }
 419
 420 out:
 421         if (buf)
 422                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 423         RETURN(rc);
 424 }
 425 EXPORT_SYMBOL(llog_reverse_process);