lustre/obdclass/llog.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/llog.c
  37  *
  38  * OST<->MDS recovery logging infrastructure.
  39  * Invariants in implementation:
  40  * - we do not share logs among different OST<->MDS connections, so that
  41  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  42  *
  43  * Author: Andreas Dilger <adilger@clusterfs.com>
  44  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  45  * Author: Mikhail Pershin <tappro@whamcloud.com>
  46  */
  47
  48 #define DEBUG_SUBSYSTEM S_LOG
  49
  50 #ifndef __KERNEL__
  51 #include <liblustre.h>
  52 #endif
  53
  54 #include <obd_class.h>
  55 #include <lustre_log.h>
  56 #include "llog_internal.h"
  57
  58 /*
  59  * Allocate a new log or catalog handle
  60  * Used inside llog_open().
  61  */
  62 struct llog_handle *llog_alloc_handle(void)
  63 {
  64         struct llog_handle *loghandle;
  65
  66         OBD_ALLOC_PTR(loghandle);
  67         if (loghandle == NULL)
  68                 return ERR_PTR(-ENOMEM);
  69
  70         init_rwsem(&loghandle->lgh_lock);
  71         spin_lock_init(&loghandle->lgh_hdr_lock);
  72         CFS_INIT_LIST_HEAD(&loghandle->u.phd.phd_entry);
  73         cfs_atomic_set(&loghandle->lgh_refcount, 1);
  74
  75         return loghandle;
  76 }
  77
  78 /*
  79  * Free llog handle and header data if exists. Used in llog_close() only
  80  */
  81 void llog_free_handle(struct llog_handle *loghandle)
  82 {
  83         LASSERT(loghandle != NULL);
  84
  85         /* failed llog_init_handle */
  86         if (!loghandle->lgh_hdr)
  87                 goto out;
  88
  89         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  90                 LASSERT(cfs_list_empty(&loghandle->u.phd.phd_entry));
  91         else if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  92                 LASSERT(cfs_list_empty(&loghandle->u.chd.chd_head));
  93         LASSERT(sizeof(*(loghandle->lgh_hdr)) == LLOG_CHUNK_SIZE);
  94         OBD_FREE(loghandle->lgh_hdr, LLOG_CHUNK_SIZE);
  95 out:
  96         OBD_FREE_PTR(loghandle);
  97 }
  98
  99 void llog_handle_get(struct llog_handle *loghandle)
 100 {
 101         cfs_atomic_inc(&loghandle->lgh_refcount);
 102 }
 103
 104 void llog_handle_put(struct llog_handle *loghandle)
 105 {
 106         LASSERT(cfs_atomic_read(&loghandle->lgh_refcount) > 0);
 107         if (cfs_atomic_dec_and_test(&loghandle->lgh_refcount))
 108                 llog_free_handle(loghandle);
 109 }
 110
 111 /* returns negative on error; 0 if success; 1 if success & log destroyed */
 112 int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
 113                     int index)
 114 {
 115         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 116         int rc = 0;
 117         ENTRY;
 118
 119         CDEBUG(D_RPCTRACE, "Canceling %d in log "DOSTID"\n",
 120                index, POSTID(&loghandle->lgh_id.lgl_oi));
 121
 122         if (index == 0) {
 123                 CERROR("Can't cancel index 0 which is header\n");
 124                 RETURN(-EINVAL);
 125         }
 126
 127         spin_lock(&loghandle->lgh_hdr_lock);
 128         if (!ext2_clear_bit(index, llh->llh_bitmap)) {
 129                 spin_unlock(&loghandle->lgh_hdr_lock);
 130                 CDEBUG(D_RPCTRACE, "Catalog index %u already clear?\n", index);
 131                 RETURN(-ENOENT);
 132         }
 133
 134         llh->llh_count--;
 135
 136         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 137             (llh->llh_count == 1) &&
 138             (loghandle->lgh_last_idx == (LLOG_BITMAP_BYTES * 8) - 1)) {
 139                 spin_unlock(&loghandle->lgh_hdr_lock);
 140                 rc = llog_destroy(env, loghandle);
 141                 if (rc < 0) {
 142                         CERROR("%s: can't destroy empty llog #"DOSTID
 143                                "#%08x: rc = %d\n",
 144                                loghandle->lgh_ctxt->loc_obd->obd_name,
 145                                POSTID(&loghandle->lgh_id.lgl_oi),
 146                                loghandle->lgh_id.lgl_ogen, rc);
 147                         GOTO(out_err, rc);
 148                 }
 149                 RETURN(1);
 150         }
 151         spin_unlock(&loghandle->lgh_hdr_lock);
 152
 153         rc = llog_write(env, loghandle, &llh->llh_hdr, NULL, 0, NULL, 0);
 154         if (rc < 0) {
 155                 CERROR("%s: fail to write header for llog #"DOSTID
 156                        "#%08x: rc = %d\n",
 157                        loghandle->lgh_ctxt->loc_obd->obd_name,
 158                        POSTID(&loghandle->lgh_id.lgl_oi),
 159                        loghandle->lgh_id.lgl_ogen, rc);
 160                 GOTO(out_err, rc);
 161         }
 162         RETURN(0);
 163 out_err:
 164         spin_lock(&loghandle->lgh_hdr_lock);
 165         ext2_set_bit(index, llh->llh_bitmap);
 166         llh->llh_count++;
 167         spin_unlock(&loghandle->lgh_hdr_lock);
 168         return rc;
 169 }
 170 EXPORT_SYMBOL(llog_cancel_rec);
 171
 172 static int llog_read_header(const struct lu_env *env,
 173                             struct llog_handle *handle,
 174                             struct obd_uuid *uuid)
 175 {
 176         struct llog_operations *lop;
 177         int rc;
 178
 179         rc = llog_handle2ops(handle, &lop);
 180         if (rc)
 181                 RETURN(rc);
 182
 183         if (lop->lop_read_header == NULL)
 184                 RETURN(-EOPNOTSUPP);
 185
 186         rc = lop->lop_read_header(env, handle);
 187         if (rc == LLOG_EEMPTY) {
 188                 struct llog_log_hdr *llh = handle->lgh_hdr;
 189
 190                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 191                 llh->llh_count = 1;         /* for the header record */
 192                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 193                 llh->llh_hdr.lrh_len = llh->llh_tail.lrt_len = LLOG_CHUNK_SIZE;
 194                 llh->llh_hdr.lrh_index = llh->llh_tail.lrt_index = 0;
 195                 llh->llh_timestamp = cfs_time_current_sec();
 196                 if (uuid)
 197                         memcpy(&llh->llh_tgtuuid, uuid,
 198                                sizeof(llh->llh_tgtuuid));
 199                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 200                 ext2_set_bit(0, llh->llh_bitmap);
 201                 rc = 0;
 202         }
 203         return rc;
 204 }
 205
 206 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 207                      int flags, struct obd_uuid *uuid)
 208 {
 209         struct llog_log_hdr     *llh;
 210         int                      rc;
 211
 212         ENTRY;
 213         LASSERT(handle->lgh_hdr == NULL);
 214
 215         OBD_ALLOC_PTR(llh);
 216         if (llh == NULL)
 217                 RETURN(-ENOMEM);
 218         handle->lgh_hdr = llh;
 219         /* first assign flags to use llog_client_ops */
 220         llh->llh_flags = flags;
 221         rc = llog_read_header(env, handle, uuid);
 222         if (rc == 0) {
 223                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 224                               flags & LLOG_F_IS_CAT) ||
 225                              (llh->llh_flags & LLOG_F_IS_CAT &&
 226                               flags & LLOG_F_IS_PLAIN))) {
 227                         CERROR("%s: llog type is %s but initializing %s\n",
 228                                handle->lgh_ctxt->loc_obd->obd_name,
 229                                llh->llh_flags & LLOG_F_IS_CAT ?
 230                                "catalog" : "plain",
 231                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 232                         GOTO(out, rc = -EINVAL);
 233                 } else if (llh->llh_flags &
 234                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 235                         /*
 236                          * it is possible to open llog without specifying llog
 237                          * type so it is taken from llh_flags
 238                          */
 239                         flags = llh->llh_flags;
 240                 } else {
 241                         /* for some reason the llh_flags has no type set */
 242                         CERROR("llog type is not specified!\n");
 243                         GOTO(out, rc = -EINVAL);
 244                 }
 245                 if (unlikely(uuid &&
 246                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 247                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 248                                handle->lgh_ctxt->loc_obd->obd_name,
 249                                (char *)uuid->uuid,
 250                                (char *)llh->llh_tgtuuid.uuid);
 251                         GOTO(out, rc = -EEXIST);
 252                 }
 253         }
 254         if (flags & LLOG_F_IS_CAT) {
 255                 LASSERT(cfs_list_empty(&handle->u.chd.chd_head));
 256                 CFS_INIT_LIST_HEAD(&handle->u.chd.chd_head);
 257                 llh->llh_size = sizeof(struct llog_logid_rec);
 258         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 259                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 260                        handle->lgh_ctxt->loc_obd->obd_name,
 261                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 262                 rc = -EINVAL;
 263         }
 264 out:
 265         if (rc) {
 266                 OBD_FREE_PTR(llh);
 267                 handle->lgh_hdr = NULL;
 268         }
 269         RETURN(rc);
 270 }
 271 EXPORT_SYMBOL(llog_init_handle);
 272
 273 int llog_copy_handler(const struct lu_env *env,
 274                       struct llog_handle *llh,
 275                       struct llog_rec_hdr *rec,
 276                       void *data)
 277 {
 278         struct llog_rec_hdr local_rec = *rec;
 279         struct llog_handle *local_llh = (struct llog_handle *)data;
 280         char *cfg_buf = (char*) (rec + 1);
 281         struct lustre_cfg *lcfg;
 282         int rc = 0;
 283         ENTRY;
 284
 285         /* Append all records */
 286         local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
 287         rc = llog_write(env, local_llh, &local_rec, NULL, 0,
 288                         (void *)cfg_buf, -1);
 289
 290         lcfg = (struct lustre_cfg *)cfg_buf;
 291         CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
 292                rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
 293                lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
 294
 295         RETURN(rc);
 296 }
 297 EXPORT_SYMBOL(llog_copy_handler);
 298
 299 static int llog_process_thread(void *arg)
 300 {
 301         struct llog_process_info        *lpi = arg;
 302         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 303         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 304         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 305         char                            *buf;
 306         __u64                            cur_offset = LLOG_CHUNK_SIZE;
 307         __u64                            last_offset;
 308         int                              rc = 0, index = 1, last_index;
 309         int                              saved_index = 0;
 310         int                              last_called_index = 0;
 311
 312         ENTRY;
 313
 314         LASSERT(llh);
 315
 316         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 317         if (!buf) {
 318                 lpi->lpi_rc = -ENOMEM;
 319                 RETURN(0);
 320         }
 321
 322         if (cd != NULL) {
 323                 last_called_index = cd->lpcd_first_idx;
 324                 index = cd->lpcd_first_idx + 1;
 325         }
 326         if (cd != NULL && cd->lpcd_last_idx)
 327                 last_index = cd->lpcd_last_idx;
 328         else
 329                 last_index = LLOG_BITMAP_BYTES * 8 - 1;
 330
 331         while (rc == 0) {
 332                 struct llog_rec_hdr *rec;
 333
 334                 /* skip records not set in bitmap */
 335                 while (index <= last_index &&
 336                        !ext2_test_bit(index, llh->llh_bitmap))
 337                         ++index;
 338
 339                 LASSERT(index <= last_index + 1);
 340                 if (index == last_index + 1)
 341                         break;
 342 repeat:
 343                 CDEBUG(D_OTHER, "index: %d last_index %d\n",
 344                        index, last_index);
 345
 346                 /* get the buf with our target record; avoid old garbage */
 347                 memset(buf, 0, LLOG_CHUNK_SIZE);
 348                 last_offset = cur_offset;
 349                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 350                                      index, &cur_offset, buf, LLOG_CHUNK_SIZE);
 351                 if (rc)
 352                         GOTO(out, rc);
 353
 354                 /* NB: when rec->lrh_len is accessed it is already swabbed
 355                  * since it is used at the "end" of the loop and the rec
 356                  * swabbing is done at the beginning of the loop. */
 357                 for (rec = (struct llog_rec_hdr *)buf;
 358                      (char *)rec < buf + LLOG_CHUNK_SIZE;
 359                      rec = (struct llog_rec_hdr *)((char *)rec + rec->lrh_len)){
 360
 361                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 362                                rec, rec->lrh_type);
 363
 364                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 365                                 lustre_swab_llog_rec(rec);
 366
 367                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 368                                rec->lrh_type, rec->lrh_index);
 369
 370                         if (rec->lrh_index == 0) {
 371                                 /* probably another rec just got added? */
 372                                 if (index <= loghandle->lgh_last_idx)
 373                                         GOTO(repeat, rc = 0);
 374                                 GOTO(out, rc = 0); /* no more records */
 375                         }
 376                         if (rec->lrh_len == 0 ||
 377                             rec->lrh_len > LLOG_CHUNK_SIZE) {
 378                                 CWARN("invalid length %d in llog record for "
 379                                       "index %d/%d\n", rec->lrh_len,
 380                                       rec->lrh_index, index);
 381                                 GOTO(out, rc = -EINVAL);
 382                         }
 383
 384                         if (rec->lrh_index < index) {
 385                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 386                                        rec->lrh_index);
 387                                 continue;
 388                         }
 389
 390                         CDEBUG(D_OTHER,
 391                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 392                                rec->lrh_index, rec->lrh_len,
 393                                (int)(buf + LLOG_CHUNK_SIZE - (char *)rec));
 394
 395                         loghandle->lgh_cur_idx = rec->lrh_index;
 396                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 397                                                     last_offset;
 398
 399                         /* if set, process the callback on this record */
 400                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 401                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 402                                                  lpi->lpi_cbdata);
 403                                 last_called_index = index;
 404                                 if (rc == LLOG_PROC_BREAK) {
 405                                         GOTO(out, rc);
 406                                 } else if (rc == LLOG_DEL_RECORD) {
 407                                         llog_cancel_rec(lpi->lpi_env,
 408                                                         loghandle,
 409                                                         rec->lrh_index);
 410                                         rc = 0;
 411                                 }
 412                                 if (rc)
 413                                         GOTO(out, rc);
 414                         } else {
 415                                 CDEBUG(D_OTHER, "Skipped index %d\n", index);
 416                         }
 417
 418                         /* next record, still in buffer? */
 419                         ++index;
 420                         if (index > last_index)
 421                                 GOTO(out, rc = 0);
 422                 }
 423         }
 424
 425 out:
 426         if (cd != NULL)
 427                 cd->lpcd_last_idx = last_called_index;
 428
 429         OBD_FREE(buf, LLOG_CHUNK_SIZE);
 430         lpi->lpi_rc = rc;
 431         return 0;
 432 }
 433
 434 #ifdef __KERNEL__
 435 static int llog_process_thread_daemonize(void *arg)
 436 {
 437         struct llog_process_info        *lpi = arg;
 438         struct lu_env                    env;
 439         int                              rc;
 440
 441         cfs_daemonize_ctxt("llog_process_thread");
 442
 443         /* client env has no keys, tags is just 0 */
 444         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 445         if (rc)
 446                 goto out;
 447         lpi->lpi_env = &env;
 448
 449         rc = llog_process_thread(arg);
 450
 451         lu_env_fini(&env);
 452 out:
 453         complete(&lpi->lpi_completion);
 454         return rc;
 455 }
 456 #endif
 457
 458 int llog_process_or_fork(const struct lu_env *env,
 459                          struct llog_handle *loghandle,
 460                          llog_cb_t cb, void *data, void *catdata, bool fork)
 461 {
 462         struct llog_process_info *lpi;
 463         int                      rc;
 464
 465         ENTRY;
 466
 467         OBD_ALLOC_PTR(lpi);
 468         if (lpi == NULL) {
 469                 CERROR("cannot alloc pointer\n");
 470                 RETURN(-ENOMEM);
 471         }
 472         lpi->lpi_loghandle = loghandle;
 473         lpi->lpi_cb        = cb;
 474         lpi->lpi_cbdata    = data;
 475         lpi->lpi_catdata   = catdata;
 476
 477 #ifdef __KERNEL__
 478         if (fork) {
 479                 /* The new thread can't use parent env,
 480                  * init the new one in llog_process_thread_daemonize. */
 481                 lpi->lpi_env = NULL;
 482                 init_completion(&lpi->lpi_completion);
 483                 rc = cfs_create_thread(llog_process_thread_daemonize, lpi,
 484                                        CFS_DAEMON_FLAGS);
 485                 if (rc < 0) {
 486                         CERROR("%s: cannot start thread: rc = %d\n",
 487                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 488                         OBD_FREE_PTR(lpi);
 489                         RETURN(rc);
 490                 }
 491                 wait_for_completion(&lpi->lpi_completion);
 492         } else {
 493                 lpi->lpi_env = env;
 494                 llog_process_thread(lpi);
 495         }
 496 #else
 497         lpi->lpi_env = env;
 498         llog_process_thread(lpi);
 499 #endif
 500         rc = lpi->lpi_rc;
 501         OBD_FREE_PTR(lpi);
 502         RETURN(rc);
 503 }
 504 EXPORT_SYMBOL(llog_process_or_fork);
 505
 506 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 507                  llog_cb_t cb, void *data, void *catdata)
 508 {
 509         return llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 510 }
 511 EXPORT_SYMBOL(llog_process);
 512
 513 inline int llog_get_size(struct llog_handle *loghandle)
 514 {
 515         if (loghandle && loghandle->lgh_hdr)
 516                 return loghandle->lgh_hdr->llh_count;
 517         return 0;
 518 }
 519 EXPORT_SYMBOL(llog_get_size);
 520
 521 int llog_reverse_process(const struct lu_env *env,
 522                          struct llog_handle *loghandle, llog_cb_t cb,
 523                          void *data, void *catdata)
 524 {
 525         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 526         struct llog_process_cat_data *cd = catdata;
 527         void *buf;
 528         int rc = 0, first_index = 1, index, idx;
 529         ENTRY;
 530
 531         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 532         if (!buf)
 533                 RETURN(-ENOMEM);
 534
 535         if (cd != NULL)
 536                 first_index = cd->lpcd_first_idx + 1;
 537         if (cd != NULL && cd->lpcd_last_idx)
 538                 index = cd->lpcd_last_idx;
 539         else
 540                 index = LLOG_BITMAP_BYTES * 8 - 1;
 541
 542         while (rc == 0) {
 543                 struct llog_rec_hdr *rec;
 544                 struct llog_rec_tail *tail;
 545
 546                 /* skip records not set in bitmap */
 547                 while (index >= first_index &&
 548                        !ext2_test_bit(index, llh->llh_bitmap))
 549                         --index;
 550
 551                 LASSERT(index >= first_index - 1);
 552                 if (index == first_index - 1)
 553                         break;
 554
 555                 /* get the buf with our target record; avoid old garbage */
 556                 memset(buf, 0, LLOG_CHUNK_SIZE);
 557                 rc = llog_prev_block(env, loghandle, index, buf,
 558                                      LLOG_CHUNK_SIZE);
 559                 if (rc)
 560                         GOTO(out, rc);
 561
 562                 rec = buf;
 563                 idx = rec->lrh_index;
 564                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 565                 while (idx < index) {
 566                         rec = (void *)rec + rec->lrh_len;
 567                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 568                                 lustre_swab_llog_rec(rec);
 569                         idx ++;
 570                 }
 571                 LASSERT(idx == index);
 572                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 573
 574                 /* process records in buffer, starting where we found one */
 575                 while ((void *)tail > buf) {
 576                         if (tail->lrt_index == 0)
 577                                 GOTO(out, rc = 0); /* no more records */
 578
 579                         /* if set, process the callback on this record */
 580                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 581                                 rec = (void *)tail - tail->lrt_len +
 582                                       sizeof(*tail);
 583
 584                                 rc = cb(env, loghandle, rec, data);
 585                                 if (rc == LLOG_PROC_BREAK) {
 586                                         GOTO(out, rc);
 587                                 } else if (rc == LLOG_DEL_RECORD) {
 588                                         llog_cancel_rec(env, loghandle,
 589                                                         tail->lrt_index);
 590                                         rc = 0;
 591                                 }
 592                                 if (rc)
 593                                         GOTO(out, rc);
 594                         }
 595
 596                         /* previous record, still in buffer? */
 597                         --index;
 598                         if (index < first_index)
 599                                 GOTO(out, rc = 0);
 600                         tail = (void *)tail - tail->lrt_len;
 601                 }
 602         }
 603
 604 out:
 605         if (buf)
 606                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 607         RETURN(rc);
 608 }
 609 EXPORT_SYMBOL(llog_reverse_process);
 610
 611 /**
 612  * new llog API
 613  *
 614  * API functions:
 615  *      llog_open - open llog, may not exist
 616  *      llog_exist - check if llog exists
 617  *      llog_close - close opened llog, pair for open, frees llog_handle
 618  *      llog_declare_create - declare llog creation
 619  *      llog_create - create new llog on disk, need transaction handle
 620  *      llog_declare_write_rec - declaration of llog write
 621  *      llog_write_rec - write llog record on disk, need transaction handle
 622  *      llog_declare_add - declare llog catalog record addition
 623  *      llog_add - add llog record in catalog, need transaction handle
 624  */
 625 int llog_exist(struct llog_handle *loghandle)
 626 {
 627         struct llog_operations  *lop;
 628         int                      rc;
 629
 630         ENTRY;
 631
 632         rc = llog_handle2ops(loghandle, &lop);
 633         if (rc)
 634                 RETURN(rc);
 635         if (lop->lop_exist == NULL)
 636                 RETURN(-EOPNOTSUPP);
 637
 638         rc = lop->lop_exist(loghandle);
 639         RETURN(rc);
 640 }
 641 EXPORT_SYMBOL(llog_exist);
 642
 643 int llog_declare_create(const struct lu_env *env,
 644                         struct llog_handle *loghandle, struct thandle *th)
 645 {
 646         struct llog_operations  *lop;
 647         int                      raised, rc;
 648
 649         ENTRY;
 650
 651         rc = llog_handle2ops(loghandle, &lop);
 652         if (rc)
 653                 RETURN(rc);
 654         if (lop->lop_declare_create == NULL)
 655                 RETURN(-EOPNOTSUPP);
 656
 657         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 658         if (!raised)
 659                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 660         rc = lop->lop_declare_create(env, loghandle, th);
 661         if (!raised)
 662                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 663         RETURN(rc);
 664 }
 665 EXPORT_SYMBOL(llog_declare_create);
 666
 667 int llog_create(const struct lu_env *env, struct llog_handle *handle,
 668                 struct thandle *th)
 669 {
 670         struct llog_operations  *lop;
 671         int                      raised, rc;
 672
 673         ENTRY;
 674
 675         rc = llog_handle2ops(handle, &lop);
 676         if (rc)
 677                 RETURN(rc);
 678         if (lop->lop_create == NULL)
 679                 RETURN(-EOPNOTSUPP);
 680
 681         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 682         if (!raised)
 683                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 684         rc = lop->lop_create(env, handle, th);
 685         if (!raised)
 686                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 687         RETURN(rc);
 688 }
 689 EXPORT_SYMBOL(llog_create);
 690
 691 int llog_declare_write_rec(const struct lu_env *env,
 692                            struct llog_handle *handle,
 693                            struct llog_rec_hdr *rec, int idx,
 694                            struct thandle *th)
 695 {
 696         struct llog_operations  *lop;
 697         int                      raised, rc;
 698
 699         ENTRY;
 700
 701         rc = llog_handle2ops(handle, &lop);
 702         if (rc)
 703                 RETURN(rc);
 704         LASSERT(lop);
 705         if (lop->lop_declare_write_rec == NULL)
 706                 RETURN(-EOPNOTSUPP);
 707
 708         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 709         if (!raised)
 710                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 711         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
 712         if (!raised)
 713                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 714         RETURN(rc);
 715 }
 716 EXPORT_SYMBOL(llog_declare_write_rec);
 717
 718 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
 719                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 720                    int numcookies, void *buf, int idx, struct thandle *th)
 721 {
 722         struct llog_operations  *lop;
 723         int                      raised, rc, buflen;
 724
 725         ENTRY;
 726
 727         rc = llog_handle2ops(handle, &lop);
 728         if (rc)
 729                 RETURN(rc);
 730
 731         LASSERT(lop);
 732         if (lop->lop_write_rec == NULL)
 733                 RETURN(-EOPNOTSUPP);
 734
 735         if (buf)
 736                 buflen = rec->lrh_len + sizeof(struct llog_rec_hdr) +
 737                          sizeof(struct llog_rec_tail);
 738         else
 739                 buflen = rec->lrh_len;
 740         LASSERT(cfs_size_round(buflen) == buflen);
 741
 742         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 743         if (!raised)
 744                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 745         rc = lop->lop_write_rec(env, handle, rec, logcookies, numcookies,
 746                                 buf, idx, th);
 747         if (!raised)
 748                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 749         RETURN(rc);
 750 }
 751 EXPORT_SYMBOL(llog_write_rec);
 752
 753 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
 754              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 755              void *buf, struct thandle *th)
 756 {
 757         int raised, rc;
 758
 759         ENTRY;
 760
 761         if (lgh->lgh_logops->lop_add == NULL)
 762                 RETURN(-EOPNOTSUPP);
 763
 764         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 765         if (!raised)
 766                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 767         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, buf, th);
 768         if (!raised)
 769                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 770         RETURN(rc);
 771 }
 772 EXPORT_SYMBOL(llog_add);
 773
 774 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
 775                      struct llog_rec_hdr *rec, struct thandle *th)
 776 {
 777         int raised, rc;
 778
 779         ENTRY;
 780
 781         if (lgh->lgh_logops->lop_declare_add == NULL)
 782                 RETURN(-EOPNOTSUPP);
 783
 784         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 785         if (!raised)
 786                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 787         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
 788         if (!raised)
 789                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 790         RETURN(rc);
 791 }
 792 EXPORT_SYMBOL(llog_declare_add);
 793
 794 /**
 795  * Helper function to open llog or create it if doesn't exist.
 796  * It hides all transaction handling from caller.
 797  */
 798 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
 799                      struct llog_handle **res, struct llog_logid *logid,
 800                      char *name)
 801 {
 802         struct thandle  *th;
 803         int              rc;
 804
 805         ENTRY;
 806
 807         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
 808         if (rc)
 809                 RETURN(rc);
 810
 811         if (llog_exist(*res))
 812                 RETURN(0);
 813
 814         if ((*res)->lgh_obj != NULL) {
 815                 struct dt_device *d;
 816
 817                 d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
 818
 819                 th = dt_trans_create(env, d);
 820                 if (IS_ERR(th))
 821                         GOTO(out, rc = PTR_ERR(th));
 822
 823                 rc = llog_declare_create(env, *res, th);
 824                 if (rc == 0) {
 825                         rc = dt_trans_start_local(env, d, th);
 826                         if (rc == 0)
 827                                 rc = llog_create(env, *res, th);
 828                 }
 829                 dt_trans_stop(env, d, th);
 830         } else {
 831                 /* lvfs compat code */
 832                 LASSERT((*res)->lgh_file == NULL);
 833                 rc = llog_create(env, *res, NULL);
 834         }
 835 out:
 836         if (rc)
 837                 llog_close(env, *res);
 838         RETURN(rc);
 839 }
 840 EXPORT_SYMBOL(llog_open_create);
 841
 842 /**
 843  * Helper function to delete existent llog.
 844  */
 845 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
 846                struct llog_logid *logid, char *name)
 847 {
 848         struct llog_handle      *handle;
 849         int                      rc = 0, rc2;
 850
 851         ENTRY;
 852
 853         /* nothing to erase */
 854         if (name == NULL && logid == NULL)
 855                 RETURN(0);
 856
 857         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
 858         if (rc < 0)
 859                 RETURN(rc);
 860
 861         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
 862         if (rc == 0)
 863                 rc = llog_destroy(env, handle);
 864
 865         rc2 = llog_close(env, handle);
 866         if (rc == 0)
 867                 rc = rc2;
 868         RETURN(rc);
 869 }
 870 EXPORT_SYMBOL(llog_erase);
 871
 872 /*
 873  * Helper function for write record in llog.
 874  * It hides all transaction handling from caller.
 875  * Valid only with local llog.
 876  */
 877 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
 878                struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 879                int cookiecount, void *buf, int idx)
 880 {
 881         int rc;
 882
 883         ENTRY;
 884
 885         LASSERT(loghandle);
 886         LASSERT(loghandle->lgh_ctxt);
 887
 888         if (loghandle->lgh_obj != NULL) {
 889                 struct dt_device        *dt;
 890                 struct thandle          *th;
 891
 892                 dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 893
 894                 th = dt_trans_create(env, dt);
 895                 if (IS_ERR(th))
 896                         RETURN(PTR_ERR(th));
 897
 898                 rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
 899                 if (rc)
 900                         GOTO(out_trans, rc);
 901
 902                 rc = dt_trans_start_local(env, dt, th);
 903                 if (rc)
 904                         GOTO(out_trans, rc);
 905
 906                 down_write(&loghandle->lgh_lock);
 907                 rc = llog_write_rec(env, loghandle, rec, reccookie,
 908                                     cookiecount, buf, idx, th);
 909                 up_write(&loghandle->lgh_lock);
 910 out_trans:
 911                 dt_trans_stop(env, dt, th);
 912         } else { /* lvfs compatibility */
 913                 down_write(&loghandle->lgh_lock);
 914                 rc = llog_write_rec(env, loghandle, rec, reccookie,
 915                                     cookiecount, buf, idx, NULL);
 916                 up_write(&loghandle->lgh_lock);
 917         }
 918         RETURN(rc);
 919 }
 920 EXPORT_SYMBOL(llog_write);
 921
 922 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
 923               struct llog_handle **lgh, struct llog_logid *logid,
 924               char *name, enum llog_open_param open_param)
 925 {
 926         int      raised;
 927         int      rc;
 928
 929         ENTRY;
 930
 931         LASSERT(ctxt);
 932         LASSERT(ctxt->loc_logops);
 933
 934         if (ctxt->loc_logops->lop_open == NULL) {
 935                 *lgh = NULL;
 936                 RETURN(-EOPNOTSUPP);
 937         }
 938
 939         *lgh = llog_alloc_handle();
 940         if (*lgh == NULL)
 941                 RETURN(-ENOMEM);
 942         (*lgh)->lgh_ctxt = ctxt;
 943         (*lgh)->lgh_logops = ctxt->loc_logops;
 944
 945         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 946         if (!raised)
 947                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 948         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
 949         if (!raised)
 950                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 951         if (rc) {
 952                 llog_free_handle(*lgh);
 953                 *lgh = NULL;
 954         }
 955         RETURN(rc);
 956 }
 957 EXPORT_SYMBOL(llog_open);
 958
 959 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
 960 {
 961         struct llog_operations  *lop;
 962         int                      rc;
 963
 964         ENTRY;
 965
 966         rc = llog_handle2ops(loghandle, &lop);
 967         if (rc)
 968                 GOTO(out, rc);
 969         if (lop->lop_close == NULL)
 970                 GOTO(out, rc = -EOPNOTSUPP);
 971         rc = lop->lop_close(env, loghandle);
 972 out:
 973         llog_handle_put(loghandle);
 974         RETURN(rc);
 975 }
 976 EXPORT_SYMBOL(llog_close);