lustre/lvfs/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 Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  * OST<->MDS recovery logging infrastructure.
  23  *
  24  * Invariants in implementation:
  25  * - we do not share logs among different OST<->MDS connections, so that
  26  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  27  */
  28
  29 #define DEBUG_SUBSYSTEM S_LOG
  30
  31 #ifndef EXPORT_SYMTAB
  32 #define EXPORT_SYMTAB
  33 #endif
  34
  35 #ifdef __KERNEL__
  36 #include <linux/fs.h>
  37 #else
  38 #include <liblustre.h>
  39 #endif
  40
  41 #include <linux/lustre_log.h>
  42
  43 /* Allocate a new log or catalog handle */
  44 struct llog_handle *llog_alloc_handle(void)
  45 {
  46         struct llog_handle *loghandle;
  47         ENTRY;
  48
  49         OBD_ALLOC(loghandle, sizeof(*loghandle));
  50         if (loghandle == NULL)
  51                 RETURN(ERR_PTR(-ENOMEM));
  52
  53         init_rwsem(&loghandle->lgh_lock);
  54
  55         RETURN(loghandle);
  56 }
  57 EXPORT_SYMBOL(llog_alloc_handle);
  58
  59 void llog_free_handle(struct llog_handle *loghandle)
  60 {
  61         if (!loghandle)
  62                 return;
  63
  64         if (!loghandle->lgh_hdr)
  65                 goto out;
  66         if (le32_to_cpu(loghandle->lgh_hdr->llh_flags) & LLOG_F_IS_PLAIN)
  67                 list_del_init(&loghandle->u.phd.phd_entry);
  68         if (le32_to_cpu(loghandle->lgh_hdr->llh_flags) & LLOG_F_IS_CAT)
  69                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  70         OBD_FREE(loghandle->lgh_hdr, LLOG_CHUNK_SIZE);
  71
  72  out:
  73         OBD_FREE(loghandle, sizeof(*loghandle));
  74 }
  75 EXPORT_SYMBOL(llog_free_handle);
  76
  77 /* returns negative on error; 0 if success; 1 if success & log destroyed */
  78 int llog_cancel_rec(struct llog_handle *loghandle, int index)
  79 {
  80         struct llog_log_hdr *llh = loghandle->lgh_hdr;
  81         int rc = 0;
  82         ENTRY;
  83
  84         CDEBUG(D_HA, "canceling %d in log "LPX64"\n",
  85                index, loghandle->lgh_id.lgl_oid);
  86
  87         if (index == 0) {
  88                 CERROR("cannot cancel index 0 (which is header)\n");
  89                 RETURN(-EINVAL);
  90         }
  91
  92         if (!ext2_clear_bit(index, llh->llh_bitmap)) {
  93                 CERROR("catalog index %u already clear?\n", index);
  94                 RETURN(-EINVAL);
  95         }
  96
  97         llh->llh_count = cpu_to_le32(le32_to_cpu(llh->llh_count) - 1);
  98
  99         if ((le32_to_cpu(llh->llh_flags) & LLOG_F_ZAP_WHEN_EMPTY) &&
 100             (le32_to_cpu(llh->llh_count) == 1) &&
 101             (loghandle->lgh_last_idx == (LLOG_BITMAP_BYTES * 8) - 1)) {
 102                 rc = llog_destroy(loghandle);
 103                 if (rc) {
 104                         CERROR("failure destroying log after last cancel: %d\n",
 105                                rc);
 106                         ext2_set_bit(index, llh->llh_bitmap);
 107                         llh->llh_count++;
 108                 } else {
 109                         rc = 1;
 110                 }
 111                 RETURN(rc);
 112         }
 113
 114         rc = llog_write_rec(loghandle, &llh->llh_hdr, NULL, 0, NULL, 0);
 115         if (rc) {
 116                 CERROR("failure re-writing header %d\n", rc);
 117                 ext2_set_bit(index, llh->llh_bitmap);
 118                 llh->llh_count++;
 119         }
 120         RETURN(rc);
 121 }
 122 EXPORT_SYMBOL(llog_cancel_rec);
 123
 124 int llog_init_handle(struct llog_handle *handle, int flags,
 125                      struct obd_uuid *uuid)
 126 {
 127         int rc;
 128         struct llog_log_hdr *llh;
 129         ENTRY;
 130         LASSERT(handle->lgh_hdr == NULL);
 131
 132         OBD_ALLOC(llh, sizeof(*llh));
 133         if (llh == NULL)
 134                 RETURN(-ENOMEM);
 135         handle->lgh_hdr = llh;
 136         /* first assign flags to use llog_client_ops */
 137         llh->llh_flags = cpu_to_le32(flags);
 138         rc = llog_read_header(handle);
 139         if (rc == 0) {
 140                 flags = le32_to_cpu(llh->llh_flags);
 141                 if (uuid)
 142                         LASSERT(obd_uuid_equals(uuid, &llh->llh_tgtuuid));
 143                 GOTO(out, rc);
 144         } else if (rc != LLOG_EEMPTY || !flags) {
 145                 /* set a pesudo flag for initialization */
 146                 flags = LLOG_F_IS_CAT;
 147                 GOTO(out, rc);
 148         }
 149         rc = 0;
 150
 151         handle->lgh_last_idx = 0; /* header is record with index 0 */
 152         llh->llh_count = cpu_to_le32(1);         /* for the header record */
 153         llh->llh_hdr.lrh_type = cpu_to_le32(LLOG_HDR_MAGIC);
 154         llh->llh_hdr.lrh_len = llh->llh_tail.lrt_len =
 155                 cpu_to_le32(LLOG_CHUNK_SIZE);
 156         llh->llh_hdr.lrh_index = llh->llh_tail.lrt_index = 0;
 157         llh->llh_timestamp = cpu_to_le64(LTIME_S(CURRENT_TIME));
 158         if (uuid)
 159                 memcpy(&llh->llh_tgtuuid, uuid, sizeof(llh->llh_tgtuuid));
 160         llh->llh_bitmap_offset = cpu_to_le32(offsetof(typeof(*llh),llh_bitmap));
 161         ext2_set_bit(0, llh->llh_bitmap);
 162
 163 out:
 164         if (flags & LLOG_F_IS_CAT) {
 165                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 166                 llh->llh_size = cpu_to_le32(sizeof(struct llog_logid_rec));
 167         }
 168         else if (flags & LLOG_F_IS_PLAIN)
 169                 INIT_LIST_HEAD(&handle->u.phd.phd_entry);
 170         else
 171                 LBUG();
 172
 173         if (rc) {
 174                 OBD_FREE(llh, sizeof(*llh));
 175                 handle->lgh_hdr = NULL;
 176         }
 177         RETURN(rc);
 178 }
 179 EXPORT_SYMBOL(llog_init_handle);
 180
 181 int llog_close(struct llog_handle *loghandle)
 182 {
 183         struct llog_operations *lop;
 184         int rc;
 185         ENTRY;
 186
 187         rc = llog_handle2ops(loghandle, &lop);
 188         if (rc)
 189                 GOTO(out, rc);
 190         if (lop->lop_close == NULL)
 191                 GOTO(out, rc = -EOPNOTSUPP);
 192         rc = lop->lop_close(loghandle);
 193  out:
 194         llog_free_handle(loghandle);
 195         RETURN(rc);
 196 }
 197 EXPORT_SYMBOL(llog_close);
 198
 199 int llog_process(struct llog_handle *loghandle, llog_cb_t cb,
 200                  void *data, void *catdata)
 201 {
 202         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 203         struct llog_process_cat_data *cd = catdata;
 204         void *buf;
 205         __u64 cur_offset = LLOG_CHUNK_SIZE;
 206         int rc = 0, index = 1, last_index, idx;
 207         int saved_index = 0;
 208         ENTRY;
 209
 210         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 211         if (!buf)
 212                 RETURN(-ENOMEM);
 213
 214         if (cd != NULL)
 215                 index = cd->first_idx + 1;
 216         if (cd != NULL && cd->last_idx)
 217                 last_index = cd->last_idx;
 218         else
 219                 last_index = LLOG_BITMAP_BYTES * 8 - 1;
 220
 221         while (rc == 0) {
 222                 struct llog_rec_hdr *rec;
 223
 224                 /* skip records not set in bitmap */
 225                 while (index <= last_index &&
 226                        !ext2_test_bit(index, llh->llh_bitmap))
 227                         ++index;
 228
 229                 LASSERT(index <= last_index + 1);
 230                 if (index == last_index + 1)
 231                         break;
 232
 233                 /* get the buf with our target record; avoid old garbage */
 234                 memset(buf, 0, LLOG_CHUNK_SIZE);
 235                 rc = llog_next_block(loghandle, &saved_index, index,
 236                                      &cur_offset, buf, LLOG_CHUNK_SIZE);
 237                 if (rc)
 238                         GOTO(out, rc);
 239
 240                 rec = buf;
 241                 idx = le32_to_cpu(rec->lrh_index);
 242                 if (idx < index)
 243                         CDEBUG(D_HA, "index %u : idx %u\n", index, idx);
 244                 while (idx < index) {
 245                         rec = ((void *)rec + le32_to_cpu(rec->lrh_len));
 246                         idx ++;
 247                 }
 248
 249                 /* process records in buffer, starting where we found one */
 250                 while ((void *)rec < buf + LLOG_CHUNK_SIZE) {
 251                         if (rec->lrh_index == 0)
 252                                 GOTO(out, rc = 0); /* no more records */
 253
 254                         /* if set, process the callback on this record */
 255                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 256                                 rc = cb(loghandle, rec, data);
 257                                 if (rc == LLOG_PROC_BREAK) {
 258                                         CDEBUG(D_HA, "recovery from log: "LPX64":%x"
 259                                               " stopped\n",
 260                                               loghandle->lgh_id.lgl_oid,
 261                                               loghandle->lgh_id.lgl_ogen);
 262                                         GOTO(out, rc);
 263                                 } else if (rc == LLOG_DEL_RECORD) {
 264                                         llog_cancel_rec(loghandle, rec->lrh_index);
 265                                         rc = 0;
 266                                 }
 267                                 if (rc)
 268                                         GOTO(out, rc);
 269                         }
 270                         /* next record, still in buffer? */
 271                         ++index;
 272                         if (index > last_index)
 273                                 GOTO(out, rc = 0);
 274                         rec = ((void *)rec + le32_to_cpu(rec->lrh_len));
 275                 }
 276         }
 277
 278  out:
 279         if (buf)
 280                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 281         RETURN(rc);
 282 }
 283 EXPORT_SYMBOL(llog_process);
 284
 285 int llog_reverse_process(struct llog_handle *loghandle, llog_cb_t cb,
 286                          void *data, void *catdata)
 287 {
 288         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 289         struct llog_process_cat_data *cd = catdata;
 290         void *buf;
 291         int rc = 0, first_index = 1, index, idx;
 292         struct llog_rec_tail *tail;
 293         ENTRY;
 294
 295         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 296         if (!buf)
 297                 RETURN(-ENOMEM);
 298
 299         if (cd != NULL)
 300                 first_index = cd->first_idx + 1;
 301         if (cd != NULL && cd->last_idx)
 302                 index = cd->last_idx;
 303         else
 304                 index = LLOG_BITMAP_BYTES * 8 - 1;
 305
 306         while (rc == 0) {
 307                 struct llog_rec_hdr *rec;
 308
 309                 /* skip records not set in bitmap */
 310                 while (index >= first_index &&
 311                        !ext2_test_bit(index, llh->llh_bitmap))
 312                         --index;
 313
 314                 LASSERT(index >= first_index - 1);
 315                 if (index == first_index - 1)
 316                         break;
 317
 318                 /* get the buf with our target record; avoid old garbage */
 319                 memset(buf, 0, LLOG_CHUNK_SIZE);
 320                 rc = llog_prev_block(loghandle, index, buf, LLOG_CHUNK_SIZE);
 321                 if (rc)
 322                         GOTO(out, rc);
 323
 324                 rec = buf;
 325                 idx = le32_to_cpu(rec->lrh_index);
 326                 if (idx < index)
 327                         CDEBUG(D_HA, "index %u : idx %u\n", index, idx);
 328                 while (idx < index) {
 329                         rec = ((void *)rec + le32_to_cpu(rec->lrh_len));
 330                         idx ++;
 331                 }
 332
 333                 /* process records in buffer, starting where we found one */
 334                 while ((void *)rec >= buf) {
 335                         if (rec->lrh_index == 0)
 336                                 GOTO(out, 0); /* no more records */
 337
 338                         /* if set, process the callback on this record */
 339                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 340                                 rc = cb(loghandle, rec, data);
 341                                 if (rc == LLOG_PROC_BREAK) {
 342                                         CWARN("recovery from log: "LPX64":%x"
 343                                               " stopped\n",
 344                                               loghandle->lgh_id.lgl_oid,
 345                                               loghandle->lgh_id.lgl_ogen);
 346                                         GOTO(out, rc);
 347                                 }
 348                                 if (rc)
 349                                         GOTO(out, rc);
 350                         }
 351
 352                         /* previous record, still in buffer? */
 353                         --index;
 354                         if (index < first_index)
 355                                 GOTO(out, rc = 0);
 356
 357                         if ((void *)rec == buf)
 358                                 break;
 359
 360                         tail = (void *)rec - sizeof(struct llog_rec_tail);
 361                         rec = ((void *)rec - le32_to_cpu(tail->lrt_len));
 362                 }
 363         }
 364
 365  out:
 366         if (buf)
 367                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 368         RETURN(rc);
 369 }
 370 EXPORT_SYMBOL(llog_reverse_process);