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 "LPX64"\n",
 120                index, loghandle->lgh_id.lgl_oid);
 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 #"LPX64"#"LPX64
 143                                "#%08x: rc = %d\n",
 144                                loghandle->lgh_ctxt->loc_obd->obd_name,
 145                                loghandle->lgh_id.lgl_oid,
 146                                loghandle->lgh_id.lgl_oseq,
 147                                loghandle->lgh_id.lgl_ogen, rc);
 148                         GOTO(out_err, rc);
 149                 }
 150                 RETURN(1);
 151         }
 152         spin_unlock(&loghandle->lgh_hdr_lock);
 153
 154         rc = llog_write(env, loghandle, &llh->llh_hdr, NULL, 0, NULL, 0);
 155         if (rc < 0) {
 156                 CERROR("%s: fail to write header for llog #"LPX64"#"LPX64
 157                        "#%08x: rc = %d\n",
 158                        loghandle->lgh_ctxt->loc_obd->obd_name,
 159                        loghandle->lgh_id.lgl_oid,
 160                        loghandle->lgh_id.lgl_oseq,
 161                        loghandle->lgh_id.lgl_ogen, rc);
 162                 GOTO(out_err, rc);
 163         }
 164         RETURN(0);
 165 out_err:
 166         spin_lock(&loghandle->lgh_hdr_lock);
 167         ext2_set_bit(index, llh->llh_bitmap);
 168         llh->llh_count++;
 169         spin_unlock(&loghandle->lgh_hdr_lock);
 170         return rc;
 171 }
 172 EXPORT_SYMBOL(llog_cancel_rec);
 173
 174 static int llog_read_header(const struct lu_env *env,
 175                             struct llog_handle *handle,
 176                             struct obd_uuid *uuid)
 177 {
 178         struct llog_operations *lop;
 179         int rc;
 180
 181         rc = llog_handle2ops(handle, &lop);
 182         if (rc)
 183                 RETURN(rc);
 184
 185         if (lop->lop_read_header == NULL)
 186                 RETURN(-EOPNOTSUPP);
 187
 188         rc = lop->lop_read_header(env, handle);
 189         if (rc == LLOG_EEMPTY) {
 190                 struct llog_log_hdr *llh = handle->lgh_hdr;
 191
 192                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 193                 llh->llh_count = 1;         /* for the header record */
 194                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 195                 llh->llh_hdr.lrh_len = llh->llh_tail.lrt_len = LLOG_CHUNK_SIZE;
 196                 llh->llh_hdr.lrh_index = llh->llh_tail.lrt_index = 0;
 197                 llh->llh_timestamp = cfs_time_current_sec();
 198                 if (uuid)
 199                         memcpy(&llh->llh_tgtuuid, uuid,
 200                                sizeof(llh->llh_tgtuuid));
 201                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 202                 ext2_set_bit(0, llh->llh_bitmap);
 203                 rc = 0;
 204         }
 205         return rc;
 206 }
 207
 208 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 209                      int flags, struct obd_uuid *uuid)
 210 {
 211         struct llog_log_hdr     *llh;
 212         int                      rc;
 213
 214         ENTRY;
 215         LASSERT(handle->lgh_hdr == NULL);
 216
 217         OBD_ALLOC_PTR(llh);
 218         if (llh == NULL)
 219                 RETURN(-ENOMEM);
 220         handle->lgh_hdr = llh;
 221         /* first assign flags to use llog_client_ops */
 222         llh->llh_flags = flags;
 223         rc = llog_read_header(env, handle, uuid);
 224         if (rc == 0) {
 225                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 226                               flags & LLOG_F_IS_CAT) ||
 227                              (llh->llh_flags & LLOG_F_IS_CAT &&
 228                               flags & LLOG_F_IS_PLAIN))) {
 229                         CERROR("%s: llog type is %s but initializing %s\n",
 230                                handle->lgh_ctxt->loc_obd->obd_name,
 231                                llh->llh_flags & LLOG_F_IS_CAT ?
 232                                "catalog" : "plain",
 233                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 234                         GOTO(out, rc = -EINVAL);
 235                 } else if (llh->llh_flags &
 236                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 237                         /*
 238                          * it is possible to open llog without specifying llog
 239                          * type so it is taken from llh_flags
 240                          */
 241                         flags = llh->llh_flags;
 242                 } else {
 243                         /* for some reason the llh_flags has no type set */
 244                         CERROR("llog type is not specified!\n");
 245                         GOTO(out, rc = -EINVAL);
 246                 }
 247                 if (unlikely(uuid &&
 248                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 249                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 250                                handle->lgh_ctxt->loc_obd->obd_name,
 251                                (char *)uuid->uuid,
 252                                (char *)llh->llh_tgtuuid.uuid);
 253                         GOTO(out, rc = -EEXIST);
 254                 }
 255         }
 256         if (flags & LLOG_F_IS_CAT) {
 257                 LASSERT(cfs_list_empty(&handle->u.chd.chd_head));
 258                 CFS_INIT_LIST_HEAD(&handle->u.chd.chd_head);
 259                 llh->llh_size = sizeof(struct llog_logid_rec);
 260         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 261                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 262                        handle->lgh_ctxt->loc_obd->obd_name,
 263                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 264                 rc = -EINVAL;
 265         }
 266 out:
 267         if (rc) {
 268                 OBD_FREE_PTR(llh);
 269                 handle->lgh_hdr = NULL;
 270         }
 271         RETURN(rc);
 272 }
 273 EXPORT_SYMBOL(llog_init_handle);
 274
 275 int llog_copy_handler(const struct lu_env *env,
 276                       struct llog_handle *llh,
 277                       struct llog_rec_hdr *rec,
 278                       void *data)
 279 {
 280         struct llog_rec_hdr local_rec = *rec;
 281         struct llog_handle *local_llh = (struct llog_handle *)data;
 282         char *cfg_buf = (char*) (rec + 1);
 283         struct lustre_cfg *lcfg;
 284         int rc = 0;
 285         ENTRY;
 286
 287         /* Append all records */
 288         local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
 289         rc = llog_write(env, local_llh, &local_rec, NULL, 0,
 290                         (void *)cfg_buf, -1);
 291
 292         lcfg = (struct lustre_cfg *)cfg_buf;
 293         CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
 294                rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
 295                lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
 296
 297         RETURN(rc);
 298 }
 299 EXPORT_SYMBOL(llog_copy_handler);
 300
 301 static int llog_process_thread(void *arg)
 302 {
 303         struct llog_process_info        *lpi = arg;
 304         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 305         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 306         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 307         char                            *buf;
 308         __u64                            cur_offset = LLOG_CHUNK_SIZE;
 309         __u64                            last_offset;
 310         int                              rc = 0, index = 1, last_index;
 311         int                              saved_index = 0;
 312         int                              last_called_index = 0;
 313
 314         ENTRY;
 315
 316         LASSERT(llh);
 317
 318         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 319         if (!buf) {
 320                 lpi->lpi_rc = -ENOMEM;
 321                 RETURN(0);
 322         }
 323
 324         if (cd != NULL) {
 325                 last_called_index = cd->lpcd_first_idx;
 326                 index = cd->lpcd_first_idx + 1;
 327         }
 328         if (cd != NULL && cd->lpcd_last_idx)
 329                 last_index = cd->lpcd_last_idx;
 330         else
 331                 last_index = LLOG_BITMAP_BYTES * 8 - 1;
 332
 333         while (rc == 0) {
 334                 struct llog_rec_hdr *rec;
 335
 336                 /* skip records not set in bitmap */
 337                 while (index <= last_index &&
 338                        !ext2_test_bit(index, llh->llh_bitmap))
 339                         ++index;
 340
 341                 LASSERT(index <= last_index + 1);
 342                 if (index == last_index + 1)
 343                         break;
 344 repeat:
 345                 CDEBUG(D_OTHER, "index: %d last_index %d\n",
 346                        index, last_index);
 347
 348                 /* get the buf with our target record; avoid old garbage */
 349                 memset(buf, 0, LLOG_CHUNK_SIZE);
 350                 last_offset = cur_offset;
 351                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 352                                      index, &cur_offset, buf, LLOG_CHUNK_SIZE);
 353                 if (rc)
 354                         GOTO(out, rc);
 355
 356                 /* NB: when rec->lrh_len is accessed it is already swabbed
 357                  * since it is used at the "end" of the loop and the rec
 358                  * swabbing is done at the beginning of the loop. */
 359                 for (rec = (struct llog_rec_hdr *)buf;
 360                      (char *)rec < buf + LLOG_CHUNK_SIZE;
 361                      rec = (struct llog_rec_hdr *)((char *)rec + rec->lrh_len)){
 362
 363                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 364                                rec, rec->lrh_type);
 365
 366                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 367                                 lustre_swab_llog_rec(rec);
 368
 369                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 370                                rec->lrh_type, rec->lrh_index);
 371
 372                         if (rec->lrh_index == 0) {
 373                                 /* probably another rec just got added? */
 374                                 if (index <= loghandle->lgh_last_idx)
 375                                         GOTO(repeat, rc = 0);
 376                                 GOTO(out, rc = 0); /* no more records */
 377                         }
 378                         if (rec->lrh_len == 0 ||
 379                             rec->lrh_len > LLOG_CHUNK_SIZE) {
 380                                 CWARN("invalid length %d in llog record for "
 381                                       "index %d/%d\n", rec->lrh_len,
 382                                       rec->lrh_index, index);
 383                                 GOTO(out, rc = -EINVAL);
 384                         }
 385
 386                         if (rec->lrh_index < index) {
 387                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 388                                        rec->lrh_index);
 389                                 continue;
 390                         }
 391
 392                         CDEBUG(D_OTHER,
 393                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 394                                rec->lrh_index, rec->lrh_len,
 395                                (int)(buf + LLOG_CHUNK_SIZE - (char *)rec));
 396
 397                         loghandle->lgh_cur_idx = rec->lrh_index;
 398                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 399                                                     last_offset;
 400
 401                         /* if set, process the callback on this record */
 402                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 403                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 404                                                  lpi->lpi_cbdata);
 405                                 last_called_index = index;
 406                                 if (rc == LLOG_PROC_BREAK) {
 407                                         GOTO(out, rc);
 408                                 } else if (rc == LLOG_DEL_RECORD) {
 409                                         llog_cancel_rec(lpi->lpi_env,
 410                                                         loghandle,
 411                                                         rec->lrh_index);
 412                                         rc = 0;
 413                                 }
 414                                 if (rc)
 415                                         GOTO(out, rc);
 416                         } else {
 417                                 CDEBUG(D_OTHER, "Skipped index %d\n", index);
 418                         }
 419
 420                         /* next record, still in buffer? */
 421                         ++index;
 422                         if (index > last_index)
 423                                 GOTO(out, rc = 0);
 424                 }
 425         }
 426
 427 out:
 428         if (cd != NULL)
 429                 cd->lpcd_last_idx = last_called_index;
 430
 431         OBD_FREE(buf, LLOG_CHUNK_SIZE);
 432         lpi->lpi_rc = rc;
 433         return 0;
 434 }
 435
 436 #ifdef __KERNEL__
 437 static int llog_process_thread_daemonize(void *arg)
 438 {
 439         struct llog_process_info        *lpi = arg;
 440         struct lu_env                    env;
 441         int                              rc;
 442
 443         cfs_daemonize_ctxt("llog_process_thread");
 444
 445         /* client env has no keys, tags is just 0 */
 446         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 447         if (rc)
 448                 goto out;
 449         lpi->lpi_env = &env;
 450
 451         rc = llog_process_thread(arg);
 452
 453         lu_env_fini(&env);
 454 out:
 455         complete(&lpi->lpi_completion);
 456         return rc;
 457 }
 458 #endif
 459
 460 int llog_process_or_fork(const struct lu_env *env,
 461                          struct llog_handle *loghandle,
 462                          llog_cb_t cb, void *data, void *catdata, bool fork)
 463 {
 464         struct llog_process_info *lpi;
 465         int                      rc;
 466
 467         ENTRY;
 468
 469         OBD_ALLOC_PTR(lpi);
 470         if (lpi == NULL) {
 471                 CERROR("cannot alloc pointer\n");
 472                 RETURN(-ENOMEM);
 473         }
 474         lpi->lpi_loghandle = loghandle;
 475         lpi->lpi_cb        = cb;
 476         lpi->lpi_cbdata    = data;
 477         lpi->lpi_catdata   = catdata;
 478
 479 #ifdef __KERNEL__
 480         if (fork) {
 481                 /* The new thread can't use parent env,
 482                  * init the new one in llog_process_thread_daemonize. */
 483                 lpi->lpi_env = NULL;
 484                 init_completion(&lpi->lpi_completion);
 485                 rc = cfs_create_thread(llog_process_thread_daemonize, lpi,
 486                                        CFS_DAEMON_FLAGS);
 487                 if (rc < 0) {
 488                         CERROR("%s: cannot start thread: rc = %d\n",
 489                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 490                         OBD_FREE_PTR(lpi);
 491                         RETURN(rc);
 492                 }
 493                 wait_for_completion(&lpi->lpi_completion);
 494         } else {
 495                 lpi->lpi_env = env;
 496                 llog_process_thread(lpi);
 497         }
 498 #else
 499         lpi->lpi_env = env;
 500         llog_process_thread(lpi);
 501 #endif
 502         rc = lpi->lpi_rc;
 503         OBD_FREE_PTR(lpi);
 504         RETURN(rc);
 505 }
 506 EXPORT_SYMBOL(llog_process_or_fork);
 507
 508 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 509                  llog_cb_t cb, void *data, void *catdata)
 510 {
 511         return llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 512 }
 513 EXPORT_SYMBOL(llog_process);
 514
 515 inline int llog_get_size(struct llog_handle *loghandle)
 516 {
 517         if (loghandle && loghandle->lgh_hdr)
 518                 return loghandle->lgh_hdr->llh_count;
 519         return 0;
 520 }
 521 EXPORT_SYMBOL(llog_get_size);
 522
 523 int llog_reverse_process(const struct lu_env *env,
 524                          struct llog_handle *loghandle, llog_cb_t cb,
 525                          void *data, void *catdata)
 526 {
 527         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 528         struct llog_process_cat_data *cd = catdata;
 529         void *buf;
 530         int rc = 0, first_index = 1, index, idx;
 531         ENTRY;
 532
 533         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 534         if (!buf)
 535                 RETURN(-ENOMEM);
 536
 537         if (cd != NULL)
 538                 first_index = cd->lpcd_first_idx + 1;
 539         if (cd != NULL && cd->lpcd_last_idx)
 540                 index = cd->lpcd_last_idx;
 541         else
 542                 index = LLOG_BITMAP_BYTES * 8 - 1;
 543
 544         while (rc == 0) {
 545                 struct llog_rec_hdr *rec;
 546                 struct llog_rec_tail *tail;
 547
 548                 /* skip records not set in bitmap */
 549                 while (index >= first_index &&
 550                        !ext2_test_bit(index, llh->llh_bitmap))
 551                         --index;
 552
 553                 LASSERT(index >= first_index - 1);
 554                 if (index == first_index - 1)
 555                         break;
 556
 557                 /* get the buf with our target record; avoid old garbage */
 558                 memset(buf, 0, LLOG_CHUNK_SIZE);
 559                 rc = llog_prev_block(env, loghandle, index, buf,
 560                                      LLOG_CHUNK_SIZE);
 561                 if (rc)
 562                         GOTO(out, rc);
 563
 564                 rec = buf;
 565                 idx = rec->lrh_index;
 566                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 567                 while (idx < index) {
 568                         rec = (void *)rec + rec->lrh_len;
 569                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 570                                 lustre_swab_llog_rec(rec);
 571                         idx ++;
 572                 }
 573                 LASSERT(idx == index);
 574                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 575
 576                 /* process records in buffer, starting where we found one */
 577                 while ((void *)tail > buf) {
 578                         if (tail->lrt_index == 0)
 579                                 GOTO(out, rc = 0); /* no more records */
 580
 581                         /* if set, process the callback on this record */
 582                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 583                                 rec = (void *)tail - tail->lrt_len +
 584                                       sizeof(*tail);
 585
 586                                 rc = cb(env, loghandle, rec, data);
 587                                 if (rc == LLOG_PROC_BREAK) {
 588                                         GOTO(out, rc);
 589                                 } else if (rc == LLOG_DEL_RECORD) {
 590                                         llog_cancel_rec(env, loghandle,
 591                                                         tail->lrt_index);
 592                                         rc = 0;
 593                                 }
 594                                 if (rc)
 595                                         GOTO(out, rc);
 596                         }
 597
 598                         /* previous record, still in buffer? */
 599                         --index;
 600                         if (index < first_index)
 601                                 GOTO(out, rc = 0);
 602                         tail = (void *)tail - tail->lrt_len;
 603                 }
 604         }
 605
 606 out:
 607         if (buf)
 608                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 609         RETURN(rc);
 610 }
 611 EXPORT_SYMBOL(llog_reverse_process);
 612
 613 /**
 614  * new llog API
 615  *
 616  * API functions:
 617  *      llog_open - open llog, may not exist
 618  *      llog_exist - check if llog exists
 619  *      llog_close - close opened llog, pair for open, frees llog_handle
 620  *      llog_declare_create - declare llog creation
 621  *      llog_create - create new llog on disk, need transaction handle
 622  *      llog_declare_write_rec - declaration of llog write
 623  *      llog_write_rec - write llog record on disk, need transaction handle
 624  *      llog_declare_add - declare llog catalog record addition
 625  *      llog_add - add llog record in catalog, need transaction handle
 626  */
 627 int llog_exist(struct llog_handle *loghandle)
 628 {
 629         struct llog_operations  *lop;
 630         int                      rc;
 631
 632         ENTRY;
 633
 634         rc = llog_handle2ops(loghandle, &lop);
 635         if (rc)
 636                 RETURN(rc);
 637         if (lop->lop_exist == NULL)
 638                 RETURN(-EOPNOTSUPP);
 639
 640         rc = lop->lop_exist(loghandle);
 641         RETURN(rc);
 642 }
 643 EXPORT_SYMBOL(llog_exist);
 644
 645 int llog_declare_create(const struct lu_env *env,
 646                         struct llog_handle *loghandle, struct thandle *th)
 647 {
 648         struct llog_operations  *lop;
 649         int                      raised, rc;
 650
 651         ENTRY;
 652
 653         rc = llog_handle2ops(loghandle, &lop);
 654         if (rc)
 655                 RETURN(rc);
 656         if (lop->lop_declare_create == NULL)
 657                 RETURN(-EOPNOTSUPP);
 658
 659         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 660         if (!raised)
 661                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 662         rc = lop->lop_declare_create(env, loghandle, th);
 663         if (!raised)
 664                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 665         RETURN(rc);
 666 }
 667 EXPORT_SYMBOL(llog_declare_create);
 668
 669 int llog_create(const struct lu_env *env, struct llog_handle *handle,
 670                 struct thandle *th)
 671 {
 672         struct llog_operations  *lop;
 673         int                      raised, rc;
 674
 675         ENTRY;
 676
 677         rc = llog_handle2ops(handle, &lop);
 678         if (rc)
 679                 RETURN(rc);
 680         if (lop->lop_create == NULL)
 681                 RETURN(-EOPNOTSUPP);
 682
 683         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 684         if (!raised)
 685                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 686         rc = lop->lop_create(env, handle, th);
 687         if (!raised)
 688                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 689         RETURN(rc);
 690 }
 691 EXPORT_SYMBOL(llog_create);
 692
 693 int llog_declare_write_rec(const struct lu_env *env,
 694                            struct llog_handle *handle,
 695                            struct llog_rec_hdr *rec, int idx,
 696                            struct thandle *th)
 697 {
 698         struct llog_operations  *lop;
 699         int                      raised, rc;
 700
 701         ENTRY;
 702
 703         rc = llog_handle2ops(handle, &lop);
 704         if (rc)
 705                 RETURN(rc);
 706         LASSERT(lop);
 707         if (lop->lop_declare_write_rec == NULL)
 708                 RETURN(-EOPNOTSUPP);
 709
 710         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 711         if (!raised)
 712                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 713         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
 714         if (!raised)
 715                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 716         RETURN(rc);
 717 }
 718 EXPORT_SYMBOL(llog_declare_write_rec);
 719
 720 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
 721                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 722                    int numcookies, void *buf, int idx, struct thandle *th)
 723 {
 724         struct llog_operations  *lop;
 725         int                      raised, rc, buflen;
 726
 727         ENTRY;
 728
 729         rc = llog_handle2ops(handle, &lop);
 730         if (rc)
 731                 RETURN(rc);
 732
 733         LASSERT(lop);
 734         if (lop->lop_write_rec == NULL)
 735                 RETURN(-EOPNOTSUPP);
 736
 737         if (buf)
 738                 buflen = rec->lrh_len + sizeof(struct llog_rec_hdr) +
 739                          sizeof(struct llog_rec_tail);
 740         else
 741                 buflen = rec->lrh_len;
 742         LASSERT(cfs_size_round(buflen) == buflen);
 743
 744         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 745         if (!raised)
 746                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 747         rc = lop->lop_write_rec(env, handle, rec, logcookies, numcookies,
 748                                 buf, idx, th);
 749         if (!raised)
 750                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 751         RETURN(rc);
 752 }
 753 EXPORT_SYMBOL(llog_write_rec);
 754
 755 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
 756              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 757              void *buf, struct thandle *th)
 758 {
 759         int raised, rc;
 760
 761         ENTRY;
 762
 763         if (lgh->lgh_logops->lop_add == NULL)
 764                 RETURN(-EOPNOTSUPP);
 765
 766         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 767         if (!raised)
 768                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 769         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, buf, th);
 770         if (!raised)
 771                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 772         RETURN(rc);
 773 }
 774 EXPORT_SYMBOL(llog_add);
 775
 776 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
 777                      struct llog_rec_hdr *rec, struct thandle *th)
 778 {
 779         int raised, rc;
 780
 781         ENTRY;
 782
 783         if (lgh->lgh_logops->lop_declare_add == NULL)
 784                 RETURN(-EOPNOTSUPP);
 785
 786         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 787         if (!raised)
 788                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 789         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
 790         if (!raised)
 791                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 792         RETURN(rc);
 793 }
 794 EXPORT_SYMBOL(llog_declare_add);
 795
 796 /**
 797  * Helper function to open llog or create it if doesn't exist.
 798  * It hides all transaction handling from caller.
 799  */
 800 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
 801                      struct llog_handle **res, struct llog_logid *logid,
 802                      char *name)
 803 {
 804         struct thandle  *th;
 805         int              rc;
 806
 807         ENTRY;
 808
 809         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
 810         if (rc)
 811                 RETURN(rc);
 812
 813         if (llog_exist(*res))
 814                 RETURN(0);
 815
 816         if ((*res)->lgh_obj != NULL) {
 817                 struct dt_device *d;
 818
 819                 d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
 820
 821                 th = dt_trans_create(env, d);
 822                 if (IS_ERR(th))
 823                         GOTO(out, rc = PTR_ERR(th));
 824
 825                 rc = llog_declare_create(env, *res, th);
 826                 if (rc == 0) {
 827                         rc = dt_trans_start_local(env, d, th);
 828                         if (rc == 0)
 829                                 rc = llog_create(env, *res, th);
 830                 }
 831                 dt_trans_stop(env, d, th);
 832         } else {
 833                 /* lvfs compat code */
 834                 LASSERT((*res)->lgh_file == NULL);
 835                 rc = llog_create(env, *res, NULL);
 836         }
 837 out:
 838         if (rc)
 839                 llog_close(env, *res);
 840         RETURN(rc);
 841 }
 842 EXPORT_SYMBOL(llog_open_create);
 843
 844 /**
 845  * Helper function to delete existent llog.
 846  */
 847 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
 848                struct llog_logid *logid, char *name)
 849 {
 850         struct llog_handle      *handle;
 851         int                      rc = 0, rc2;
 852
 853         ENTRY;
 854
 855         /* nothing to erase */
 856         if (name == NULL && logid == NULL)
 857                 RETURN(0);
 858
 859         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
 860         if (rc < 0)
 861                 RETURN(rc);
 862
 863         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
 864         if (rc == 0)
 865                 rc = llog_destroy(env, handle);
 866
 867         rc2 = llog_close(env, handle);
 868         if (rc == 0)
 869                 rc = rc2;
 870         RETURN(rc);
 871 }
 872 EXPORT_SYMBOL(llog_erase);
 873
 874 /*
 875  * Helper function for write record in llog.
 876  * It hides all transaction handling from caller.
 877  * Valid only with local llog.
 878  */
 879 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
 880                struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 881                int cookiecount, void *buf, int idx)
 882 {
 883         int rc;
 884
 885         ENTRY;
 886
 887         LASSERT(loghandle);
 888         LASSERT(loghandle->lgh_ctxt);
 889
 890         if (loghandle->lgh_obj != NULL) {
 891                 struct dt_device        *dt;
 892                 struct thandle          *th;
 893
 894                 dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 895
 896                 th = dt_trans_create(env, dt);
 897                 if (IS_ERR(th))
 898                         RETURN(PTR_ERR(th));
 899
 900                 rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
 901                 if (rc)
 902                         GOTO(out_trans, rc);
 903
 904                 rc = dt_trans_start_local(env, dt, th);
 905                 if (rc)
 906                         GOTO(out_trans, rc);
 907
 908                 down_write(&loghandle->lgh_lock);
 909                 rc = llog_write_rec(env, loghandle, rec, reccookie,
 910                                     cookiecount, buf, idx, th);
 911                 up_write(&loghandle->lgh_lock);
 912 out_trans:
 913                 dt_trans_stop(env, dt, th);
 914         } else { /* lvfs compatibility */
 915                 down_write(&loghandle->lgh_lock);
 916                 rc = llog_write_rec(env, loghandle, rec, reccookie,
 917                                     cookiecount, buf, idx, NULL);
 918                 up_write(&loghandle->lgh_lock);
 919         }
 920         RETURN(rc);
 921 }
 922 EXPORT_SYMBOL(llog_write);
 923
 924 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
 925               struct llog_handle **lgh, struct llog_logid *logid,
 926               char *name, enum llog_open_param open_param)
 927 {
 928         int      raised;
 929         int      rc;
 930
 931         ENTRY;
 932
 933         LASSERT(ctxt);
 934         LASSERT(ctxt->loc_logops);
 935
 936         if (ctxt->loc_logops->lop_open == NULL) {
 937                 *lgh = NULL;
 938                 RETURN(-EOPNOTSUPP);
 939         }
 940
 941         *lgh = llog_alloc_handle();
 942         if (*lgh == NULL)
 943                 RETURN(-ENOMEM);
 944         (*lgh)->lgh_ctxt = ctxt;
 945         (*lgh)->lgh_logops = ctxt->loc_logops;
 946
 947         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 948         if (!raised)
 949                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 950         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
 951         if (!raised)
 952                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 953         if (rc) {
 954                 llog_free_handle(*lgh);
 955                 *lgh = NULL;
 956         }
 957         RETURN(rc);
 958 }
 959 EXPORT_SYMBOL(llog_open);
 960
 961 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
 962 {
 963         struct llog_operations  *lop;
 964         int                      rc;
 965
 966         ENTRY;
 967
 968         rc = llog_handle2ops(loghandle, &lop);
 969         if (rc)
 970                 GOTO(out, rc);
 971         if (lop->lop_close == NULL)
 972                 GOTO(out, rc = -EOPNOTSUPP);
 973         rc = lop->lop_close(env, loghandle);
 974 out:
 975         llog_handle_put(loghandle);
 976         RETURN(rc);
 977 }
 978 EXPORT_SYMBOL(llog_close);