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 #include "llog_internal.h"
  46
  47 /* Allocate a new log or catalog handle */
  48 struct llog_handle *llog_alloc_handle(void)
  49 {
  50         struct llog_handle *loghandle;
  51         ENTRY;
  52
  53         OBD_ALLOC(loghandle, sizeof(*loghandle));
  54         if (loghandle == NULL)
  55                 RETURN(ERR_PTR(-ENOMEM));
  56
  57         init_rwsem(&loghandle->lgh_lock);
  58
  59         RETURN(loghandle);
  60 }
  61 EXPORT_SYMBOL(llog_alloc_handle);
  62
  63
  64 void llog_free_handle(struct llog_handle *loghandle)
  65 {
  66         if (!loghandle)
  67                 return;
  68
  69         if (!loghandle->lgh_hdr)
  70                 goto out;
  71         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  72                 list_del_init(&loghandle->u.phd.phd_entry);
  73         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  74                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  75         OBD_FREE(loghandle->lgh_hdr, LLOG_CHUNK_SIZE);
  76
  77  out:
  78         OBD_FREE(loghandle, sizeof(*loghandle));
  79 }
  80 EXPORT_SYMBOL(llog_free_handle);
  81
  82 /* returns negative on error; 0 if success; 1 if success & log destroyed */
  83 int llog_cancel_rec(struct llog_handle *loghandle, int index)
  84 {
  85         struct llog_log_hdr *llh = loghandle->lgh_hdr;
  86         int rc = 0;
  87         ENTRY;
  88
  89         CDEBUG(D_RPCTRACE, "canceling %d in log "LPX64"\n",
  90                index, loghandle->lgh_id.lgl_oid);
  91
  92         if (index == 0) {
  93                 CERROR("cannot cancel index 0 (which is header)\n");
  94                 RETURN(-EINVAL);
  95         }
  96
  97         if (!ext2_clear_bit(index, llh->llh_bitmap)) {
  98                 CDEBUG(D_RPCTRACE, "catalog index %u already clear?\n", index);
  99                 RETURN(-EINVAL);
 100         }
 101
 102         llh->llh_count--;
 103
 104         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 105             (llh->llh_count == 1) &&
 106             (loghandle->lgh_last_idx == (LLOG_BITMAP_BYTES * 8) - 1)) {
 107                 rc = llog_destroy(loghandle);
 108                 if (rc) {
 109                         CERROR("failure destroying log after last cancel: %d\n",
 110                                rc);
 111                         ext2_set_bit(index, llh->llh_bitmap);
 112                         llh->llh_count++;
 113                 } else {
 114                         rc = 1;
 115                 }
 116                 RETURN(rc);
 117         }
 118
 119         rc = llog_write_rec(loghandle, &llh->llh_hdr, NULL, 0, NULL, 0);
 120         if (rc) {
 121                 CERROR("failure re-writing header %d\n", rc);
 122                 ext2_set_bit(index, llh->llh_bitmap);
 123                 llh->llh_count++;
 124         }
 125         RETURN(rc);
 126 }
 127 EXPORT_SYMBOL(llog_cancel_rec);
 128
 129 int llog_init_handle(struct llog_handle *handle, int flags,
 130                      struct obd_uuid *uuid)
 131 {
 132         int rc;
 133         struct llog_log_hdr *llh;
 134         ENTRY;
 135         LASSERT(handle->lgh_hdr == NULL);
 136
 137         OBD_ALLOC(llh, sizeof(*llh));
 138         if (llh == NULL)
 139                 RETURN(-ENOMEM);
 140         handle->lgh_hdr = llh;
 141         /* first assign flags to use llog_client_ops */
 142         llh->llh_flags = flags;
 143         rc = llog_read_header(handle);
 144         if (rc == 0) {
 145                 flags = llh->llh_flags;
 146                 if (uuid && !obd_uuid_equals(uuid, &llh->llh_tgtuuid)) {
 147                         CERROR("uuid mismatch: %s/%s\n", (char *)uuid->uuid,
 148                                (char *)llh->llh_tgtuuid.uuid);
 149                         rc = -EEXIST;
 150                 }
 151                 GOTO(out, rc);
 152         } else if (rc != LLOG_EEMPTY || !flags) {
 153                 /* set a pesudo flag for initialization */
 154                 flags = LLOG_F_IS_CAT;
 155                 GOTO(out, rc);
 156         }
 157         rc = 0;
 158
 159         handle->lgh_last_idx = 0; /* header is record with index 0 */
 160         llh->llh_count = 1;         /* for the header record */
 161         llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 162         llh->llh_hdr.lrh_len = llh->llh_tail.lrt_len = LLOG_CHUNK_SIZE;
 163         llh->llh_hdr.lrh_index = llh->llh_tail.lrt_index = 0;
 164         llh->llh_timestamp = cfs_time_current_sec();
 165         if (uuid)
 166                 memcpy(&llh->llh_tgtuuid, uuid, sizeof(llh->llh_tgtuuid));
 167         llh->llh_bitmap_offset = offsetof(typeof(*llh),llh_bitmap);
 168         ext2_set_bit(0, llh->llh_bitmap);
 169
 170 out:
 171         if (flags & LLOG_F_IS_CAT) {
 172                 CFS_INIT_LIST_HEAD(&handle->u.chd.chd_head);
 173                 llh->llh_size = sizeof(struct llog_logid_rec);
 174         } else if (flags & LLOG_F_IS_PLAIN) {
 175                 CFS_INIT_LIST_HEAD(&handle->u.phd.phd_entry);
 176         } else {
 177                 CERROR("Unknown flags: %#x (Expected %#x or %#x\n",
 178                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 179                 LBUG();
 180         }
 181
 182         if (rc) {
 183                 OBD_FREE(llh, sizeof(*llh));
 184                 handle->lgh_hdr = NULL;
 185         }
 186         RETURN(rc);
 187 }
 188 EXPORT_SYMBOL(llog_init_handle);
 189
 190 int llog_close(struct llog_handle *loghandle)
 191 {
 192         struct llog_operations *lop;
 193         int rc;
 194         ENTRY;
 195
 196         rc = llog_handle2ops(loghandle, &lop);
 197         if (rc)
 198                 GOTO(out, rc);
 199         if (lop->lop_close == NULL)
 200                 GOTO(out, -EOPNOTSUPP);
 201         rc = lop->lop_close(loghandle);
 202  out:
 203         llog_free_handle(loghandle);
 204         RETURN(rc);
 205 }
 206 EXPORT_SYMBOL(llog_close);
 207
 208 static int llog_process_thread(void *arg)
 209 {
 210         struct llog_process_info     *lpi = (struct llog_process_info *)arg;
 211         struct llog_handle           *loghandle = lpi->lpi_loghandle;
 212         struct llog_log_hdr          *llh = loghandle->lgh_hdr;
 213         struct llog_process_cat_data *cd  = lpi->lpi_catdata;
 214         char                         *buf;
 215         __u64                         cur_offset = LLOG_CHUNK_SIZE;
 216         __u64                         last_offset;
 217         int                           rc = 0, index = 1, last_index;
 218         int                           saved_index = 0, last_called_index = 0;
 219
 220         LASSERT(llh);
 221
 222         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 223         if (!buf) {
 224                 lpi->lpi_rc = -ENOMEM;
 225 #ifdef __KERNEL__
 226                 complete(&lpi->lpi_completion);
 227 #endif
 228                 return 0;
 229         }
 230
 231         cfs_daemonize("llog_process");
 232
 233         if (cd != NULL) {
 234                 last_called_index = cd->first_idx;
 235                 index = cd->first_idx + 1;
 236         }
 237         if (cd != NULL && cd->last_idx)
 238                 last_index = cd->last_idx;
 239         else
 240                 last_index = LLOG_BITMAP_BYTES * 8 - 1;
 241
 242         while (rc == 0) {
 243                 struct llog_rec_hdr *rec;
 244
 245                 /* skip records not set in bitmap */
 246                 while (index <= last_index &&
 247                        !ext2_test_bit(index, llh->llh_bitmap))
 248                         ++index;
 249
 250                 LASSERT(index <= last_index + 1);
 251                 if (index == last_index + 1)
 252                         break;
 253
 254                 CDEBUG(D_OTHER, "index: %d last_index %d\n",
 255                        index, last_index);
 256
 257                 /* get the buf with our target record; avoid old garbage */
 258                 memset(buf, 0, LLOG_CHUNK_SIZE);
 259                 last_offset = cur_offset;
 260                 rc = llog_next_block(loghandle, &saved_index, index,
 261                                      &cur_offset, buf, LLOG_CHUNK_SIZE);
 262                 if (rc)
 263                         GOTO(out, rc);
 264
 265                 /* NB: when rec->lrh_len is accessed it is already swabbed
 266                  * since it is used at the "end" of the loop and the rec
 267                  * swabbing is done at the beginning of the loop. */
 268                 for (rec = (struct llog_rec_hdr *)buf;
 269                      (char *)rec < buf + LLOG_CHUNK_SIZE;
 270                      rec = (struct llog_rec_hdr *)((char *)rec + rec->lrh_len)){
 271
 272                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 273                                rec, rec->lrh_type);
 274
 275                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 276                                 lustre_swab_llog_rec(rec, NULL);
 277
 278                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 279                                rec->lrh_type, rec->lrh_index);
 280
 281                         if (rec->lrh_index == 0)
 282                                 GOTO(out, 0); /* no more records */
 283
 284                         if (rec->lrh_len == 0 || rec->lrh_len >LLOG_CHUNK_SIZE){
 285                                 CWARN("invalid length %d in llog record for "
 286                                       "index %d/%d\n", rec->lrh_len,
 287                                       rec->lrh_index, index);
 288                                 GOTO(out, rc = -EINVAL);
 289                         }
 290
 291                         if (rec->lrh_index < index) {
 292                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 293                                        rec->lrh_index);
 294                                 continue;
 295                         }
 296
 297                         CDEBUG(D_OTHER,
 298                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 299                                rec->lrh_index, rec->lrh_len,
 300                                (int)(buf + LLOG_CHUNK_SIZE - (char *)rec));
 301
 302                         loghandle->lgh_cur_idx = rec->lrh_index;
 303                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 304                                                     last_offset;
 305
 306                         /* if set, process the callback on this record */
 307                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 308                                 rc = lpi->lpi_cb(loghandle, rec, lpi->lpi_cbdata);
 309                                 last_called_index = index;
 310                                 if (rc == LLOG_PROC_BREAK) {
 311                                         CDEBUG(D_HA, "recovery from log: "LPX64
 312                                                ":%x stopped\n",
 313                                                loghandle->lgh_id.lgl_oid,
 314                                                loghandle->lgh_id.lgl_ogen);
 315                                         GOTO(out, rc);
 316                                 } else if (rc == LLOG_DEL_RECORD) {
 317                                         llog_cancel_rec(loghandle,
 318                                                         rec->lrh_index);
 319                                         rc = 0;
 320                                 }
 321                                 if (rc)
 322                                         GOTO(out, rc);
 323                         } else {
 324                                 CDEBUG(D_OTHER, "Skipped index %d\n", index);
 325                         }
 326
 327                         /* next record, still in buffer? */
 328                         ++index;
 329                         if (index > last_index)
 330                                 GOTO(out, rc = 0);
 331                 }
 332         }
 333
 334  out:
 335         if (cd != NULL)
 336                 cd->last_idx = last_called_index;
 337         if (buf)
 338                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 339         lpi->lpi_rc = rc;
 340 #ifdef __KERNEL__
 341         complete(&lpi->lpi_completion);
 342 #endif
 343         return 0;
 344 }
 345
 346 int llog_process(struct llog_handle *loghandle, llog_cb_t cb,
 347                  void *data, void *catdata)
 348 {
 349         struct llog_process_info *lpi;
 350         int                      rc;
 351         ENTRY;
 352
 353         OBD_ALLOC_PTR(lpi);
 354         if (lpi == NULL) {
 355                 CERROR("cannot alloc pointer\n");
 356                 RETURN(-ENOMEM);
 357         }
 358         lpi->lpi_loghandle = loghandle;
 359         lpi->lpi_cb        = cb;
 360         lpi->lpi_cbdata    = data;
 361         lpi->lpi_catdata   = catdata;
 362
 363 #ifdef __KERNEL__
 364         init_completion(&lpi->lpi_completion);
 365         rc = cfs_kernel_thread(llog_process_thread, lpi, CLONE_VM | CLONE_FILES);
 366         if (rc < 0) {
 367                 CERROR("cannot start thread: %d\n", rc);
 368                 OBD_FREE_PTR(lpi);
 369                 RETURN(rc);
 370         }
 371         wait_for_completion(&lpi->lpi_completion);
 372 #else
 373         llog_process_thread(lpi);
 374 #endif
 375         rc = lpi->lpi_rc;
 376         OBD_FREE_PTR(lpi);
 377         RETURN(rc);
 378 }
 379 EXPORT_SYMBOL(llog_process);
 380
 381 inline int llog_get_size(struct llog_handle *loghandle)
 382 {
 383         if (loghandle && loghandle->lgh_hdr)
 384                 return loghandle->lgh_hdr->llh_count;
 385         return 0;
 386 }
 387 EXPORT_SYMBOL(llog_get_size);
 388
 389 int llog_reverse_process(struct llog_handle *loghandle, llog_cb_t cb,
 390                          void *data, void *catdata)
 391 {
 392         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 393         struct llog_process_cat_data *cd = catdata;
 394         void *buf;
 395         int rc = 0, first_index = 1, index, idx;
 396         ENTRY;
 397
 398         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 399         if (!buf)
 400                 RETURN(-ENOMEM);
 401
 402         if (cd != NULL)
 403                 first_index = cd->first_idx + 1;
 404         if (cd != NULL && cd->last_idx)
 405                 index = cd->last_idx;
 406         else
 407                 index = LLOG_BITMAP_BYTES * 8 - 1;
 408
 409         while (rc == 0) {
 410                 struct llog_rec_hdr *rec;
 411                 struct llog_rec_tail *tail;
 412
 413                 /* skip records not set in bitmap */
 414                 while (index >= first_index &&
 415                        !ext2_test_bit(index, llh->llh_bitmap))
 416                         --index;
 417
 418                 LASSERT(index >= first_index - 1);
 419                 if (index == first_index - 1)
 420                         break;
 421
 422                 /* get the buf with our target record; avoid old garbage */
 423                 memset(buf, 0, LLOG_CHUNK_SIZE);
 424                 rc = llog_prev_block(loghandle, index, buf, LLOG_CHUNK_SIZE);
 425                 if (rc)
 426                         GOTO(out, rc);
 427
 428                 rec = buf;
 429                 idx = le32_to_cpu(rec->lrh_index);
 430                 if (idx < index)
 431                         CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 432                 while (idx < index) {
 433                         rec = ((void *)rec + le32_to_cpu(rec->lrh_len));
 434                         idx ++;
 435                 }
 436                 tail = (void *)rec + le32_to_cpu(rec->lrh_len) - sizeof(*tail);
 437
 438                 /* process records in buffer, starting where we found one */
 439                 while ((void *)tail > buf) {
 440                         rec = (void *)tail - le32_to_cpu(tail->lrt_len) +
 441                                 sizeof(*tail);
 442
 443                         if (rec->lrh_index == 0)
 444                                 GOTO(out, 0); /* no more records */
 445
 446                         /* if set, process the callback on this record */
 447                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 448                                 rc = cb(loghandle, rec, data);
 449                                 if (rc == LLOG_PROC_BREAK) {
 450                                         CWARN("recovery from log: "LPX64":%x"
 451                                               " stopped\n",
 452                                               loghandle->lgh_id.lgl_oid,
 453                                               loghandle->lgh_id.lgl_ogen);
 454                                         GOTO(out, rc);
 455                                 }
 456                                 if (rc)
 457                                         GOTO(out, rc);
 458                         }
 459
 460                         /* previous record, still in buffer? */
 461                         --index;
 462                         if (index < first_index)
 463                                 GOTO(out, rc = 0);
 464                         tail = (void *)rec - sizeof(*tail);
 465                 }
 466         }
 467
 468 out:
 469         if (buf)
 470                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 471         RETURN(rc);
 472 }
 473 EXPORT_SYMBOL(llog_reverse_process);