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, 2013, 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 NULL;
  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 static int llog_process_thread(void *arg)
 274 {
 275         struct llog_process_info        *lpi = arg;
 276         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 277         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 278         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 279         char                            *buf;
 280         __u64                            cur_offset = LLOG_CHUNK_SIZE;
 281         __u64                            last_offset;
 282         int                              rc = 0, index = 1, last_index;
 283         int                              saved_index = 0;
 284         int                              last_called_index = 0;
 285
 286         ENTRY;
 287
 288         LASSERT(llh);
 289
 290         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 291         if (!buf) {
 292                 lpi->lpi_rc = -ENOMEM;
 293                 RETURN(0);
 294         }
 295
 296         if (cd != NULL) {
 297                 last_called_index = cd->lpcd_first_idx;
 298                 index = cd->lpcd_first_idx + 1;
 299         }
 300         if (cd != NULL && cd->lpcd_last_idx)
 301                 last_index = cd->lpcd_last_idx;
 302         else
 303                 last_index = LLOG_BITMAP_BYTES * 8 - 1;
 304
 305         while (rc == 0) {
 306                 struct llog_rec_hdr *rec;
 307
 308                 /* skip records not set in bitmap */
 309                 while (index <= last_index &&
 310                        !ext2_test_bit(index, llh->llh_bitmap))
 311                         ++index;
 312
 313                 LASSERT(index <= last_index + 1);
 314                 if (index == last_index + 1)
 315                         break;
 316 repeat:
 317                 CDEBUG(D_OTHER, "index: %d last_index %d\n",
 318                        index, last_index);
 319
 320                 /* get the buf with our target record; avoid old garbage */
 321                 memset(buf, 0, LLOG_CHUNK_SIZE);
 322                 last_offset = cur_offset;
 323                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 324                                      index, &cur_offset, buf, LLOG_CHUNK_SIZE);
 325                 if (rc)
 326                         GOTO(out, rc);
 327
 328                 /* NB: when rec->lrh_len is accessed it is already swabbed
 329                  * since it is used at the "end" of the loop and the rec
 330                  * swabbing is done at the beginning of the loop. */
 331                 for (rec = (struct llog_rec_hdr *)buf;
 332                      (char *)rec < buf + LLOG_CHUNK_SIZE;
 333                      rec = (struct llog_rec_hdr *)((char *)rec + rec->lrh_len)){
 334
 335                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 336                                rec, rec->lrh_type);
 337
 338                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 339                                 lustre_swab_llog_rec(rec);
 340
 341                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 342                                rec->lrh_type, rec->lrh_index);
 343
 344                         if (rec->lrh_index == 0) {
 345                                 /* probably another rec just got added? */
 346                                 if (index <= loghandle->lgh_last_idx)
 347                                         GOTO(repeat, rc = 0);
 348                                 GOTO(out, rc = 0); /* no more records */
 349                         }
 350                         if (rec->lrh_len == 0 ||
 351                             rec->lrh_len > LLOG_CHUNK_SIZE) {
 352                                 CWARN("invalid length %d in llog record for "
 353                                       "index %d/%d\n", rec->lrh_len,
 354                                       rec->lrh_index, index);
 355                                 GOTO(out, rc = -EINVAL);
 356                         }
 357
 358                         if (rec->lrh_index < index) {
 359                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 360                                        rec->lrh_index);
 361                                 continue;
 362                         }
 363
 364                         CDEBUG(D_OTHER,
 365                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 366                                rec->lrh_index, rec->lrh_len,
 367                                (int)(buf + LLOG_CHUNK_SIZE - (char *)rec));
 368
 369                         loghandle->lgh_cur_idx = rec->lrh_index;
 370                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 371                                                     last_offset;
 372
 373                         /* if set, process the callback on this record */
 374                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 375                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 376                                                  lpi->lpi_cbdata);
 377                                 last_called_index = index;
 378                                 if (rc == LLOG_PROC_BREAK) {
 379                                         GOTO(out, rc);
 380                                 } else if (rc == LLOG_DEL_RECORD) {
 381                                         llog_cancel_rec(lpi->lpi_env,
 382                                                         loghandle,
 383                                                         rec->lrh_index);
 384                                         rc = 0;
 385                                 }
 386                                 if (rc)
 387                                         GOTO(out, rc);
 388                         } else {
 389                                 CDEBUG(D_OTHER, "Skipped index %d\n", index);
 390                         }
 391
 392                         /* next record, still in buffer? */
 393                         ++index;
 394                         if (index > last_index)
 395                                 GOTO(out, rc = 0);
 396                 }
 397         }
 398
 399 out:
 400         if (cd != NULL)
 401                 cd->lpcd_last_idx = last_called_index;
 402
 403         OBD_FREE(buf, LLOG_CHUNK_SIZE);
 404         lpi->lpi_rc = rc;
 405         return 0;
 406 }
 407
 408 #ifdef __KERNEL__
 409 static int llog_process_thread_daemonize(void *arg)
 410 {
 411         struct llog_process_info        *lpi = arg;
 412         struct lu_env                    env;
 413         int                              rc;
 414
 415         unshare_fs_struct();
 416
 417         /* client env has no keys, tags is just 0 */
 418         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 419         if (rc)
 420                 goto out;
 421         lpi->lpi_env = &env;
 422
 423         rc = llog_process_thread(arg);
 424
 425         lu_env_fini(&env);
 426 out:
 427         complete(&lpi->lpi_completion);
 428         return rc;
 429 }
 430 #endif
 431
 432 int llog_process_or_fork(const struct lu_env *env,
 433                          struct llog_handle *loghandle,
 434                          llog_cb_t cb, void *data, void *catdata, bool fork)
 435 {
 436         struct llog_process_info *lpi;
 437         int                      rc;
 438
 439         ENTRY;
 440
 441         OBD_ALLOC_PTR(lpi);
 442         if (lpi == NULL) {
 443                 CERROR("cannot alloc pointer\n");
 444                 RETURN(-ENOMEM);
 445         }
 446         lpi->lpi_loghandle = loghandle;
 447         lpi->lpi_cb        = cb;
 448         lpi->lpi_cbdata    = data;
 449         lpi->lpi_catdata   = catdata;
 450
 451 #ifdef __KERNEL__
 452         if (fork) {
 453                 /* The new thread can't use parent env,
 454                  * init the new one in llog_process_thread_daemonize. */
 455                 lpi->lpi_env = NULL;
 456                 init_completion(&lpi->lpi_completion);
 457                 rc = PTR_ERR(kthread_run(llog_process_thread_daemonize, lpi,
 458                                              "llog_process_thread"));
 459                 if (IS_ERR_VALUE(rc)) {
 460                         CERROR("%s: cannot start thread: rc = %d\n",
 461                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 462                         OBD_FREE_PTR(lpi);
 463                         RETURN(rc);
 464                 }
 465                 wait_for_completion(&lpi->lpi_completion);
 466         } else {
 467                 lpi->lpi_env = env;
 468                 llog_process_thread(lpi);
 469         }
 470 #else
 471         lpi->lpi_env = env;
 472         llog_process_thread(lpi);
 473 #endif
 474         rc = lpi->lpi_rc;
 475         OBD_FREE_PTR(lpi);
 476         RETURN(rc);
 477 }
 478 EXPORT_SYMBOL(llog_process_or_fork);
 479
 480 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 481                  llog_cb_t cb, void *data, void *catdata)
 482 {
 483         return llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 484 }
 485 EXPORT_SYMBOL(llog_process);
 486
 487 int llog_reverse_process(const struct lu_env *env,
 488                          struct llog_handle *loghandle, llog_cb_t cb,
 489                          void *data, void *catdata)
 490 {
 491         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 492         struct llog_process_cat_data *cd = catdata;
 493         void *buf;
 494         int rc = 0, first_index = 1, index, idx;
 495         ENTRY;
 496
 497         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 498         if (!buf)
 499                 RETURN(-ENOMEM);
 500
 501         if (cd != NULL)
 502                 first_index = cd->lpcd_first_idx + 1;
 503         if (cd != NULL && cd->lpcd_last_idx)
 504                 index = cd->lpcd_last_idx;
 505         else
 506                 index = LLOG_BITMAP_BYTES * 8 - 1;
 507
 508         while (rc == 0) {
 509                 struct llog_rec_hdr *rec;
 510                 struct llog_rec_tail *tail;
 511
 512                 /* skip records not set in bitmap */
 513                 while (index >= first_index &&
 514                        !ext2_test_bit(index, llh->llh_bitmap))
 515                         --index;
 516
 517                 LASSERT(index >= first_index - 1);
 518                 if (index == first_index - 1)
 519                         break;
 520
 521                 /* get the buf with our target record; avoid old garbage */
 522                 memset(buf, 0, LLOG_CHUNK_SIZE);
 523                 rc = llog_prev_block(env, loghandle, index, buf,
 524                                      LLOG_CHUNK_SIZE);
 525                 if (rc)
 526                         GOTO(out, rc);
 527
 528                 rec = buf;
 529                 idx = rec->lrh_index;
 530                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 531                 while (idx < index) {
 532                         rec = (void *)rec + rec->lrh_len;
 533                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 534                                 lustre_swab_llog_rec(rec);
 535                         idx ++;
 536                 }
 537                 LASSERT(idx == index);
 538                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 539
 540                 /* process records in buffer, starting where we found one */
 541                 while ((void *)tail > buf) {
 542                         if (tail->lrt_index == 0)
 543                                 GOTO(out, rc = 0); /* no more records */
 544
 545                         /* if set, process the callback on this record */
 546                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 547                                 rec = (void *)tail - tail->lrt_len +
 548                                       sizeof(*tail);
 549
 550                                 rc = cb(env, loghandle, rec, data);
 551                                 if (rc == LLOG_PROC_BREAK) {
 552                                         GOTO(out, rc);
 553                                 } else if (rc == LLOG_DEL_RECORD) {
 554                                         llog_cancel_rec(env, loghandle,
 555                                                         tail->lrt_index);
 556                                         rc = 0;
 557                                 }
 558                                 if (rc)
 559                                         GOTO(out, rc);
 560                         }
 561
 562                         /* previous record, still in buffer? */
 563                         --index;
 564                         if (index < first_index)
 565                                 GOTO(out, rc = 0);
 566                         tail = (void *)tail - tail->lrt_len;
 567                 }
 568         }
 569
 570 out:
 571         if (buf)
 572                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 573         RETURN(rc);
 574 }
 575 EXPORT_SYMBOL(llog_reverse_process);
 576
 577 /**
 578  * new llog API
 579  *
 580  * API functions:
 581  *      llog_open - open llog, may not exist
 582  *      llog_exist - check if llog exists
 583  *      llog_close - close opened llog, pair for open, frees llog_handle
 584  *      llog_declare_create - declare llog creation
 585  *      llog_create - create new llog on disk, need transaction handle
 586  *      llog_declare_write_rec - declaration of llog write
 587  *      llog_write_rec - write llog record on disk, need transaction handle
 588  *      llog_declare_add - declare llog catalog record addition
 589  *      llog_add - add llog record in catalog, need transaction handle
 590  */
 591 int llog_exist(struct llog_handle *loghandle)
 592 {
 593         struct llog_operations  *lop;
 594         int                      rc;
 595
 596         ENTRY;
 597
 598         rc = llog_handle2ops(loghandle, &lop);
 599         if (rc)
 600                 RETURN(rc);
 601         if (lop->lop_exist == NULL)
 602                 RETURN(-EOPNOTSUPP);
 603
 604         rc = lop->lop_exist(loghandle);
 605         RETURN(rc);
 606 }
 607 EXPORT_SYMBOL(llog_exist);
 608
 609 int llog_declare_create(const struct lu_env *env,
 610                         struct llog_handle *loghandle, struct thandle *th)
 611 {
 612         struct llog_operations  *lop;
 613         int                      raised, rc;
 614
 615         ENTRY;
 616
 617         rc = llog_handle2ops(loghandle, &lop);
 618         if (rc)
 619                 RETURN(rc);
 620         if (lop->lop_declare_create == NULL)
 621                 RETURN(-EOPNOTSUPP);
 622
 623         raised = cfs_cap_raised(CAP_SYS_RESOURCE);
 624         if (!raised)
 625                 cfs_cap_raise(CAP_SYS_RESOURCE);
 626         rc = lop->lop_declare_create(env, loghandle, th);
 627         if (!raised)
 628                 cfs_cap_lower(CAP_SYS_RESOURCE);
 629         RETURN(rc);
 630 }
 631 EXPORT_SYMBOL(llog_declare_create);
 632
 633 int llog_create(const struct lu_env *env, struct llog_handle *handle,
 634                 struct thandle *th)
 635 {
 636         struct llog_operations  *lop;
 637         int                      raised, rc;
 638
 639         ENTRY;
 640
 641         rc = llog_handle2ops(handle, &lop);
 642         if (rc)
 643                 RETURN(rc);
 644         if (lop->lop_create == NULL)
 645                 RETURN(-EOPNOTSUPP);
 646
 647         raised = cfs_cap_raised(CAP_SYS_RESOURCE);
 648         if (!raised)
 649                 cfs_cap_raise(CAP_SYS_RESOURCE);
 650         rc = lop->lop_create(env, handle, th);
 651         if (!raised)
 652                 cfs_cap_lower(CAP_SYS_RESOURCE);
 653         RETURN(rc);
 654 }
 655 EXPORT_SYMBOL(llog_create);
 656
 657 int llog_declare_write_rec(const struct lu_env *env,
 658                            struct llog_handle *handle,
 659                            struct llog_rec_hdr *rec, int idx,
 660                            struct thandle *th)
 661 {
 662         struct llog_operations  *lop;
 663         int                      raised, rc;
 664
 665         ENTRY;
 666
 667         rc = llog_handle2ops(handle, &lop);
 668         if (rc)
 669                 RETURN(rc);
 670         LASSERT(lop);
 671         if (lop->lop_declare_write_rec == NULL)
 672                 RETURN(-EOPNOTSUPP);
 673
 674         raised = cfs_cap_raised(CAP_SYS_RESOURCE);
 675         if (!raised)
 676                 cfs_cap_raise(CAP_SYS_RESOURCE);
 677         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
 678         if (!raised)
 679                 cfs_cap_lower(CAP_SYS_RESOURCE);
 680         RETURN(rc);
 681 }
 682 EXPORT_SYMBOL(llog_declare_write_rec);
 683
 684 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
 685                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 686                    int numcookies, void *buf, int idx, struct thandle *th)
 687 {
 688         struct llog_operations  *lop;
 689         int                      raised, rc, buflen;
 690
 691         ENTRY;
 692
 693         rc = llog_handle2ops(handle, &lop);
 694         if (rc)
 695                 RETURN(rc);
 696
 697         LASSERT(lop);
 698         if (lop->lop_write_rec == NULL)
 699                 RETURN(-EOPNOTSUPP);
 700
 701         if (buf)
 702                 buflen = rec->lrh_len + sizeof(struct llog_rec_hdr) +
 703                          sizeof(struct llog_rec_tail);
 704         else
 705                 buflen = rec->lrh_len;
 706         LASSERT(cfs_size_round(buflen) == buflen);
 707
 708         raised = cfs_cap_raised(CAP_SYS_RESOURCE);
 709         if (!raised)
 710                 cfs_cap_raise(CAP_SYS_RESOURCE);
 711         rc = lop->lop_write_rec(env, handle, rec, logcookies, numcookies,
 712                                 buf, idx, th);
 713         if (!raised)
 714                 cfs_cap_lower(CAP_SYS_RESOURCE);
 715         RETURN(rc);
 716 }
 717 EXPORT_SYMBOL(llog_write_rec);
 718
 719 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
 720              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 721              void *buf, struct thandle *th)
 722 {
 723         int raised, rc;
 724
 725         ENTRY;
 726
 727         if (lgh->lgh_logops->lop_add == NULL)
 728                 RETURN(-EOPNOTSUPP);
 729
 730         raised = cfs_cap_raised(CAP_SYS_RESOURCE);
 731         if (!raised)
 732                 cfs_cap_raise(CAP_SYS_RESOURCE);
 733         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, buf, th);
 734         if (!raised)
 735                 cfs_cap_lower(CAP_SYS_RESOURCE);
 736         RETURN(rc);
 737 }
 738 EXPORT_SYMBOL(llog_add);
 739
 740 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
 741                      struct llog_rec_hdr *rec, struct thandle *th)
 742 {
 743         int raised, rc;
 744
 745         ENTRY;
 746
 747         if (lgh->lgh_logops->lop_declare_add == NULL)
 748                 RETURN(-EOPNOTSUPP);
 749
 750         raised = cfs_cap_raised(CAP_SYS_RESOURCE);
 751         if (!raised)
 752                 cfs_cap_raise(CAP_SYS_RESOURCE);
 753         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
 754         if (!raised)
 755                 cfs_cap_lower(CAP_SYS_RESOURCE);
 756         RETURN(rc);
 757 }
 758 EXPORT_SYMBOL(llog_declare_add);
 759
 760 /**
 761  * Helper function to open llog or create it if doesn't exist.
 762  * It hides all transaction handling from caller.
 763  */
 764 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
 765                      struct llog_handle **res, struct llog_logid *logid,
 766                      char *name)
 767 {
 768         struct dt_device        *d;
 769         struct thandle          *th;
 770         int                      rc;
 771
 772         ENTRY;
 773
 774         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
 775         if (rc)
 776                 RETURN(rc);
 777
 778         if (llog_exist(*res))
 779                 RETURN(0);
 780
 781         LASSERT((*res)->lgh_obj != NULL);
 782
 783         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
 784
 785         th = dt_trans_create(env, d);
 786         if (IS_ERR(th))
 787                 GOTO(out, rc = PTR_ERR(th));
 788
 789         rc = llog_declare_create(env, *res, th);
 790         if (rc == 0) {
 791                 rc = dt_trans_start_local(env, d, th);
 792                 if (rc == 0)
 793                         rc = llog_create(env, *res, th);
 794         }
 795         dt_trans_stop(env, d, th);
 796 out:
 797         if (rc)
 798                 llog_close(env, *res);
 799         RETURN(rc);
 800 }
 801 EXPORT_SYMBOL(llog_open_create);
 802
 803 /**
 804  * Helper function to delete existent llog.
 805  */
 806 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
 807                struct llog_logid *logid, char *name)
 808 {
 809         struct llog_handle      *handle;
 810         int                      rc = 0, rc2;
 811
 812         ENTRY;
 813
 814         /* nothing to erase */
 815         if (name == NULL && logid == NULL)
 816                 RETURN(0);
 817
 818         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
 819         if (rc < 0)
 820                 RETURN(rc);
 821
 822         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
 823         if (rc == 0)
 824                 rc = llog_destroy(env, handle);
 825
 826         rc2 = llog_close(env, handle);
 827         if (rc == 0)
 828                 rc = rc2;
 829         RETURN(rc);
 830 }
 831 EXPORT_SYMBOL(llog_erase);
 832
 833 /*
 834  * Helper function for write record in llog.
 835  * It hides all transaction handling from caller.
 836  * Valid only with local llog.
 837  */
 838 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
 839                struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 840                int cookiecount, void *buf, int idx)
 841 {
 842         struct dt_device        *dt;
 843         struct thandle          *th;
 844         int                      rc;
 845
 846         ENTRY;
 847
 848         LASSERT(loghandle);
 849         LASSERT(loghandle->lgh_ctxt);
 850         LASSERT(loghandle->lgh_obj != NULL);
 851
 852         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 853
 854         th = dt_trans_create(env, dt);
 855         if (IS_ERR(th))
 856                 RETURN(PTR_ERR(th));
 857
 858         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
 859         if (rc)
 860                 GOTO(out_trans, rc);
 861
 862         rc = dt_trans_start_local(env, dt, th);
 863         if (rc)
 864                 GOTO(out_trans, rc);
 865
 866         down_write(&loghandle->lgh_lock);
 867         rc = llog_write_rec(env, loghandle, rec, reccookie,
 868                             cookiecount, buf, idx, th);
 869         up_write(&loghandle->lgh_lock);
 870 out_trans:
 871         dt_trans_stop(env, dt, th);
 872         RETURN(rc);
 873 }
 874 EXPORT_SYMBOL(llog_write);
 875
 876 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
 877               struct llog_handle **lgh, struct llog_logid *logid,
 878               char *name, enum llog_open_param open_param)
 879 {
 880         int      raised;
 881         int      rc;
 882
 883         ENTRY;
 884
 885         LASSERT(ctxt);
 886         LASSERT(ctxt->loc_logops);
 887
 888         if (ctxt->loc_logops->lop_open == NULL) {
 889                 *lgh = NULL;
 890                 RETURN(-EOPNOTSUPP);
 891         }
 892
 893         *lgh = llog_alloc_handle();
 894         if (*lgh == NULL)
 895                 RETURN(-ENOMEM);
 896         (*lgh)->lgh_ctxt = ctxt;
 897         (*lgh)->lgh_logops = ctxt->loc_logops;
 898
 899         raised = cfs_cap_raised(CAP_SYS_RESOURCE);
 900         if (!raised)
 901                 cfs_cap_raise(CAP_SYS_RESOURCE);
 902         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
 903         if (!raised)
 904                 cfs_cap_lower(CAP_SYS_RESOURCE);
 905         if (rc) {
 906                 llog_free_handle(*lgh);
 907                 *lgh = NULL;
 908         }
 909         RETURN(rc);
 910 }
 911 EXPORT_SYMBOL(llog_open);
 912
 913 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
 914 {
 915         struct llog_operations  *lop;
 916         int                      rc;
 917
 918         ENTRY;
 919
 920         rc = llog_handle2ops(loghandle, &lop);
 921         if (rc)
 922                 GOTO(out, rc);
 923         if (lop->lop_close == NULL)
 924                 GOTO(out, rc = -EOPNOTSUPP);
 925         rc = lop->lop_close(env, loghandle);
 926 out:
 927         llog_handle_put(loghandle);
 928         RETURN(rc);
 929 }
 930 EXPORT_SYMBOL(llog_close);
 931
 932 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
 933                   char *name)
 934 {
 935         struct llog_handle      *llh;
 936         int                      rc = 0;
 937
 938         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
 939         if (rc < 0) {
 940                 if (likely(rc == -ENOENT))
 941                         rc = 0;
 942                 GOTO(out, rc);
 943         }
 944
 945         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
 946         if (rc)
 947                 GOTO(out_close, rc);
 948         rc = llog_get_size(llh);
 949
 950 out_close:
 951         llog_close(env, llh);
 952 out:
 953         /* header is record 1 */
 954         return (rc <= 1);
 955 }
 956 EXPORT_SYMBOL(llog_is_empty);
 957
 958 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
 959                       struct llog_rec_hdr *rec, void *data)
 960 {
 961         struct llog_handle      *copy_llh = data;
 962
 963         /* Append all records */
 964         return llog_write(env, copy_llh, rec, NULL, 0, NULL, -1);
 965 }
 966 EXPORT_SYMBOL(llog_copy_handler);
 967
 968 /* backup plain llog */
 969 int llog_backup(const struct lu_env *env, struct obd_device *obd,
 970                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
 971                 char *name, char *backup)
 972 {
 973         struct llog_handle      *llh, *bllh;
 974         int                      rc;
 975
 976         ENTRY;
 977
 978         /* open original log */
 979         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
 980         if (rc < 0) {
 981                 /* the -ENOENT case is also reported to the caller
 982                  * but silently so it should handle that if needed.
 983                  */
 984                 if (rc != -ENOENT)
 985                         CERROR("%s: failed to open log %s: rc = %d\n",
 986                                obd->obd_name, name, rc);
 987                 RETURN(rc);
 988         }
 989
 990         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
 991         if (rc)
 992                 GOTO(out_close, rc);
 993
 994         /* Make sure there's no old backup log */
 995         rc = llog_erase(env, bctxt, NULL, backup);
 996         if (rc < 0 && rc != -ENOENT)
 997                 GOTO(out_close, rc);
 998
 999         /* open backup log */
1000         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1001         if (rc) {
1002                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1003                        obd->obd_name, backup, rc);
1004                 GOTO(out_close, rc);
1005         }
1006
1007         /* check that backup llog is not the same object as original one */
1008         if (llh->lgh_obj == bllh->lgh_obj) {
1009                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1010                        obd->obd_name, name, backup, llh->lgh_obj,
1011                        bllh->lgh_obj);
1012                 GOTO(out_backup, rc = -EEXIST);
1013         }
1014
1015         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1016         if (rc)
1017                 GOTO(out_backup, rc);
1018
1019         /* Copy log record by record */
1020         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1021                                   NULL, false);
1022         if (rc)
1023                 CERROR("%s: failed to backup log %s: rc = %d\n",
1024                        obd->obd_name, name, rc);
1025 out_backup:
1026         llog_close(env, bllh);
1027 out_close:
1028         llog_close(env, llh);
1029         RETURN(rc);
1030 }
1031 EXPORT_SYMBOL(llog_backup);