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         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         atomic_inc(&loghandle->lgh_refcount);
 102 }
 103
 104 void llog_handle_put(struct llog_handle *loghandle)
 105 {
 106         LASSERT(atomic_read(&loghandle->lgh_refcount) > 0);
 107         if (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         if (unlikely(rc == -EIO)) {
 404                 /* something bad happened to the processing, probably I/O
 405                  * error or the log got corrupted..
 406                  * to be able to finally release the log we discard any
 407                  * remaining bits in the header */
 408                 CERROR("llog found corrupted\n");
 409                 while (index <= last_index) {
 410                         if (ext2_test_bit(index, llh->llh_bitmap) != 0)
 411                                 llog_cancel_rec(lpi->lpi_env, loghandle, index);
 412                         index++;
 413                 }
 414                 rc = 0;
 415         }
 416
 417         OBD_FREE(buf, LLOG_CHUNK_SIZE);
 418         lpi->lpi_rc = rc;
 419         return 0;
 420 }
 421
 422 #ifdef __KERNEL__
 423 static int llog_process_thread_daemonize(void *arg)
 424 {
 425         struct llog_process_info        *lpi = arg;
 426         struct lu_env                    env;
 427         int                              rc;
 428
 429         unshare_fs_struct();
 430
 431         /* client env has no keys, tags is just 0 */
 432         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 433         if (rc)
 434                 goto out;
 435         lpi->lpi_env = &env;
 436
 437         rc = llog_process_thread(arg);
 438
 439         lu_env_fini(&env);
 440 out:
 441         complete(&lpi->lpi_completion);
 442         return rc;
 443 }
 444 #endif
 445
 446 int llog_process_or_fork(const struct lu_env *env,
 447                          struct llog_handle *loghandle,
 448                          llog_cb_t cb, void *data, void *catdata, bool fork)
 449 {
 450         struct llog_process_info *lpi;
 451         int                      rc;
 452
 453         ENTRY;
 454
 455         OBD_ALLOC_PTR(lpi);
 456         if (lpi == NULL) {
 457                 CERROR("cannot alloc pointer\n");
 458                 RETURN(-ENOMEM);
 459         }
 460         lpi->lpi_loghandle = loghandle;
 461         lpi->lpi_cb        = cb;
 462         lpi->lpi_cbdata    = data;
 463         lpi->lpi_catdata   = catdata;
 464
 465 #ifdef __KERNEL__
 466         if (fork) {
 467                 /* The new thread can't use parent env,
 468                  * init the new one in llog_process_thread_daemonize. */
 469                 lpi->lpi_env = NULL;
 470                 init_completion(&lpi->lpi_completion);
 471                 rc = PTR_ERR(kthread_run(llog_process_thread_daemonize, lpi,
 472                                              "llog_process_thread"));
 473                 if (IS_ERR_VALUE(rc)) {
 474                         CERROR("%s: cannot start thread: rc = %d\n",
 475                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 476                         OBD_FREE_PTR(lpi);
 477                         RETURN(rc);
 478                 }
 479                 wait_for_completion(&lpi->lpi_completion);
 480         } else {
 481                 lpi->lpi_env = env;
 482                 llog_process_thread(lpi);
 483         }
 484 #else
 485         lpi->lpi_env = env;
 486         llog_process_thread(lpi);
 487 #endif
 488         rc = lpi->lpi_rc;
 489         OBD_FREE_PTR(lpi);
 490         RETURN(rc);
 491 }
 492 EXPORT_SYMBOL(llog_process_or_fork);
 493
 494 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 495                  llog_cb_t cb, void *data, void *catdata)
 496 {
 497         return llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 498 }
 499 EXPORT_SYMBOL(llog_process);
 500
 501 int llog_reverse_process(const struct lu_env *env,
 502                          struct llog_handle *loghandle, llog_cb_t cb,
 503                          void *data, void *catdata)
 504 {
 505         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 506         struct llog_process_cat_data *cd = catdata;
 507         void *buf;
 508         int rc = 0, first_index = 1, index, idx;
 509         ENTRY;
 510
 511         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 512         if (!buf)
 513                 RETURN(-ENOMEM);
 514
 515         if (cd != NULL)
 516                 first_index = cd->lpcd_first_idx + 1;
 517         if (cd != NULL && cd->lpcd_last_idx)
 518                 index = cd->lpcd_last_idx;
 519         else
 520                 index = LLOG_BITMAP_BYTES * 8 - 1;
 521
 522         while (rc == 0) {
 523                 struct llog_rec_hdr *rec;
 524                 struct llog_rec_tail *tail;
 525
 526                 /* skip records not set in bitmap */
 527                 while (index >= first_index &&
 528                        !ext2_test_bit(index, llh->llh_bitmap))
 529                         --index;
 530
 531                 LASSERT(index >= first_index - 1);
 532                 if (index == first_index - 1)
 533                         break;
 534
 535                 /* get the buf with our target record; avoid old garbage */
 536                 memset(buf, 0, LLOG_CHUNK_SIZE);
 537                 rc = llog_prev_block(env, loghandle, index, buf,
 538                                      LLOG_CHUNK_SIZE);
 539                 if (rc)
 540                         GOTO(out, rc);
 541
 542                 rec = buf;
 543                 idx = rec->lrh_index;
 544                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 545                 while (idx < index) {
 546                         rec = (void *)rec + rec->lrh_len;
 547                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 548                                 lustre_swab_llog_rec(rec);
 549                         idx ++;
 550                 }
 551                 LASSERT(idx == index);
 552                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 553
 554                 /* process records in buffer, starting where we found one */
 555                 while ((void *)tail > buf) {
 556                         if (tail->lrt_index == 0)
 557                                 GOTO(out, rc = 0); /* no more records */
 558
 559                         /* if set, process the callback on this record */
 560                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 561                                 rec = (void *)tail - tail->lrt_len +
 562                                       sizeof(*tail);
 563
 564                                 rc = cb(env, loghandle, rec, data);
 565                                 if (rc == LLOG_PROC_BREAK) {
 566                                         GOTO(out, rc);
 567                                 } else if (rc == LLOG_DEL_RECORD) {
 568                                         llog_cancel_rec(env, loghandle,
 569                                                         tail->lrt_index);
 570                                         rc = 0;
 571                                 }
 572                                 if (rc)
 573                                         GOTO(out, rc);
 574                         }
 575
 576                         /* previous record, still in buffer? */
 577                         --index;
 578                         if (index < first_index)
 579                                 GOTO(out, rc = 0);
 580                         tail = (void *)tail - tail->lrt_len;
 581                 }
 582         }
 583
 584 out:
 585         if (buf)
 586                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 587         RETURN(rc);
 588 }
 589 EXPORT_SYMBOL(llog_reverse_process);
 590
 591 /**
 592  * new llog API
 593  *
 594  * API functions:
 595  *      llog_open - open llog, may not exist
 596  *      llog_exist - check if llog exists
 597  *      llog_close - close opened llog, pair for open, frees llog_handle
 598  *      llog_declare_create - declare llog creation
 599  *      llog_create - create new llog on disk, need transaction handle
 600  *      llog_declare_write_rec - declaration of llog write
 601  *      llog_write_rec - write llog record on disk, need transaction handle
 602  *      llog_declare_add - declare llog catalog record addition
 603  *      llog_add - add llog record in catalog, need transaction handle
 604  */
 605 int llog_exist(struct llog_handle *loghandle)
 606 {
 607         struct llog_operations  *lop;
 608         int                      rc;
 609
 610         ENTRY;
 611
 612         rc = llog_handle2ops(loghandle, &lop);
 613         if (rc)
 614                 RETURN(rc);
 615         if (lop->lop_exist == NULL)
 616                 RETURN(-EOPNOTSUPP);
 617
 618         rc = lop->lop_exist(loghandle);
 619         RETURN(rc);
 620 }
 621 EXPORT_SYMBOL(llog_exist);
 622
 623 int llog_declare_create(const struct lu_env *env,
 624                         struct llog_handle *loghandle, struct thandle *th)
 625 {
 626         struct llog_operations  *lop;
 627         int                      raised, rc;
 628
 629         ENTRY;
 630
 631         rc = llog_handle2ops(loghandle, &lop);
 632         if (rc)
 633                 RETURN(rc);
 634         if (lop->lop_declare_create == NULL)
 635                 RETURN(-EOPNOTSUPP);
 636
 637         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 638         if (!raised)
 639                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 640         rc = lop->lop_declare_create(env, loghandle, th);
 641         if (!raised)
 642                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 643         RETURN(rc);
 644 }
 645 EXPORT_SYMBOL(llog_declare_create);
 646
 647 int llog_create(const struct lu_env *env, struct llog_handle *handle,
 648                 struct thandle *th)
 649 {
 650         struct llog_operations  *lop;
 651         int                      raised, rc;
 652
 653         ENTRY;
 654
 655         rc = llog_handle2ops(handle, &lop);
 656         if (rc)
 657                 RETURN(rc);
 658         if (lop->lop_create == NULL)
 659                 RETURN(-EOPNOTSUPP);
 660
 661         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 662         if (!raised)
 663                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 664         rc = lop->lop_create(env, handle, th);
 665         if (!raised)
 666                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 667         RETURN(rc);
 668 }
 669 EXPORT_SYMBOL(llog_create);
 670
 671 int llog_declare_write_rec(const struct lu_env *env,
 672                            struct llog_handle *handle,
 673                            struct llog_rec_hdr *rec, int idx,
 674                            struct thandle *th)
 675 {
 676         struct llog_operations  *lop;
 677         int                      raised, rc;
 678
 679         ENTRY;
 680
 681         rc = llog_handle2ops(handle, &lop);
 682         if (rc)
 683                 RETURN(rc);
 684         LASSERT(lop);
 685         if (lop->lop_declare_write_rec == NULL)
 686                 RETURN(-EOPNOTSUPP);
 687
 688         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 689         if (!raised)
 690                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 691         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
 692         if (!raised)
 693                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 694         RETURN(rc);
 695 }
 696 EXPORT_SYMBOL(llog_declare_write_rec);
 697
 698 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
 699                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 700                    int numcookies, void *buf, int idx, struct thandle *th)
 701 {
 702         struct llog_operations  *lop;
 703         int                      raised, rc, buflen;
 704
 705         ENTRY;
 706
 707         rc = llog_handle2ops(handle, &lop);
 708         if (rc)
 709                 RETURN(rc);
 710
 711         LASSERT(lop);
 712         if (lop->lop_write_rec == NULL)
 713                 RETURN(-EOPNOTSUPP);
 714
 715         if (buf)
 716                 buflen = rec->lrh_len + sizeof(struct llog_rec_hdr) +
 717                          sizeof(struct llog_rec_tail);
 718         else
 719                 buflen = rec->lrh_len;
 720         LASSERT(cfs_size_round(buflen) == buflen);
 721
 722         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 723         if (!raised)
 724                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 725         rc = lop->lop_write_rec(env, handle, rec, logcookies, numcookies,
 726                                 buf, idx, th);
 727         if (!raised)
 728                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 729         RETURN(rc);
 730 }
 731 EXPORT_SYMBOL(llog_write_rec);
 732
 733 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
 734              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 735              void *buf, struct thandle *th)
 736 {
 737         int raised, rc;
 738
 739         ENTRY;
 740
 741         if (lgh->lgh_logops->lop_add == NULL)
 742                 RETURN(-EOPNOTSUPP);
 743
 744         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 745         if (!raised)
 746                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 747         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, buf, th);
 748         if (!raised)
 749                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 750         RETURN(rc);
 751 }
 752 EXPORT_SYMBOL(llog_add);
 753
 754 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
 755                      struct llog_rec_hdr *rec, struct thandle *th)
 756 {
 757         int raised, rc;
 758
 759         ENTRY;
 760
 761         if (lgh->lgh_logops->lop_declare_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_declare_add(env, lgh, rec, th);
 768         if (!raised)
 769                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 770         RETURN(rc);
 771 }
 772 EXPORT_SYMBOL(llog_declare_add);
 773
 774 /**
 775  * Helper function to open llog or create it if doesn't exist.
 776  * It hides all transaction handling from caller.
 777  */
 778 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
 779                      struct llog_handle **res, struct llog_logid *logid,
 780                      char *name)
 781 {
 782         struct dt_device        *d;
 783         struct thandle          *th;
 784         int                      rc;
 785
 786         ENTRY;
 787
 788         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
 789         if (rc)
 790                 RETURN(rc);
 791
 792         if (llog_exist(*res))
 793                 RETURN(0);
 794
 795         LASSERT((*res)->lgh_obj != NULL);
 796
 797         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
 798
 799         th = dt_trans_create(env, d);
 800         if (IS_ERR(th))
 801                 GOTO(out, rc = PTR_ERR(th));
 802
 803         rc = llog_declare_create(env, *res, th);
 804         if (rc == 0) {
 805                 rc = dt_trans_start_local(env, d, th);
 806                 if (rc == 0)
 807                         rc = llog_create(env, *res, th);
 808         }
 809         dt_trans_stop(env, d, th);
 810 out:
 811         if (rc)
 812                 llog_close(env, *res);
 813         RETURN(rc);
 814 }
 815 EXPORT_SYMBOL(llog_open_create);
 816
 817 /**
 818  * Helper function to delete existent llog.
 819  */
 820 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
 821                struct llog_logid *logid, char *name)
 822 {
 823         struct llog_handle      *handle;
 824         int                      rc = 0, rc2;
 825
 826         ENTRY;
 827
 828         /* nothing to erase */
 829         if (name == NULL && logid == NULL)
 830                 RETURN(0);
 831
 832         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
 833         if (rc < 0)
 834                 RETURN(rc);
 835
 836         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
 837         if (rc == 0)
 838                 rc = llog_destroy(env, handle);
 839
 840         rc2 = llog_close(env, handle);
 841         if (rc == 0)
 842                 rc = rc2;
 843         RETURN(rc);
 844 }
 845 EXPORT_SYMBOL(llog_erase);
 846
 847 /*
 848  * Helper function for write record in llog.
 849  * It hides all transaction handling from caller.
 850  * Valid only with local llog.
 851  */
 852 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
 853                struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 854                int cookiecount, void *buf, int idx)
 855 {
 856         struct dt_device        *dt;
 857         struct thandle          *th;
 858         int                      rc;
 859
 860         ENTRY;
 861
 862         LASSERT(loghandle);
 863         LASSERT(loghandle->lgh_ctxt);
 864         LASSERT(loghandle->lgh_obj != NULL);
 865
 866         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 867
 868         th = dt_trans_create(env, dt);
 869         if (IS_ERR(th))
 870                 RETURN(PTR_ERR(th));
 871
 872         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
 873         if (rc)
 874                 GOTO(out_trans, rc);
 875
 876         rc = dt_trans_start_local(env, dt, th);
 877         if (rc)
 878                 GOTO(out_trans, rc);
 879
 880         down_write(&loghandle->lgh_lock);
 881         rc = llog_write_rec(env, loghandle, rec, reccookie,
 882                             cookiecount, buf, idx, th);
 883         up_write(&loghandle->lgh_lock);
 884 out_trans:
 885         dt_trans_stop(env, dt, th);
 886         RETURN(rc);
 887 }
 888 EXPORT_SYMBOL(llog_write);
 889
 890 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
 891               struct llog_handle **lgh, struct llog_logid *logid,
 892               char *name, enum llog_open_param open_param)
 893 {
 894         int      raised;
 895         int      rc;
 896
 897         ENTRY;
 898
 899         LASSERT(ctxt);
 900         LASSERT(ctxt->loc_logops);
 901
 902         if (ctxt->loc_logops->lop_open == NULL) {
 903                 *lgh = NULL;
 904                 RETURN(-EOPNOTSUPP);
 905         }
 906
 907         *lgh = llog_alloc_handle();
 908         if (*lgh == NULL)
 909                 RETURN(-ENOMEM);
 910         (*lgh)->lgh_ctxt = ctxt;
 911         (*lgh)->lgh_logops = ctxt->loc_logops;
 912
 913         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 914         if (!raised)
 915                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 916         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
 917         if (!raised)
 918                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 919         if (rc) {
 920                 llog_free_handle(*lgh);
 921                 *lgh = NULL;
 922         }
 923         RETURN(rc);
 924 }
 925 EXPORT_SYMBOL(llog_open);
 926
 927 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
 928 {
 929         struct llog_operations  *lop;
 930         int                      rc;
 931
 932         ENTRY;
 933
 934         rc = llog_handle2ops(loghandle, &lop);
 935         if (rc)
 936                 GOTO(out, rc);
 937         if (lop->lop_close == NULL)
 938                 GOTO(out, rc = -EOPNOTSUPP);
 939         rc = lop->lop_close(env, loghandle);
 940 out:
 941         llog_handle_put(loghandle);
 942         RETURN(rc);
 943 }
 944 EXPORT_SYMBOL(llog_close);
 945
 946 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
 947                   char *name)
 948 {
 949         struct llog_handle      *llh;
 950         int                      rc = 0;
 951
 952         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
 953         if (rc < 0) {
 954                 if (likely(rc == -ENOENT))
 955                         rc = 0;
 956                 GOTO(out, rc);
 957         }
 958
 959         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
 960         if (rc)
 961                 GOTO(out_close, rc);
 962         rc = llog_get_size(llh);
 963
 964 out_close:
 965         llog_close(env, llh);
 966 out:
 967         /* header is record 1 */
 968         return (rc <= 1);
 969 }
 970 EXPORT_SYMBOL(llog_is_empty);
 971
 972 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
 973                       struct llog_rec_hdr *rec, void *data)
 974 {
 975         struct llog_handle      *copy_llh = data;
 976
 977         /* Append all records */
 978         return llog_write(env, copy_llh, rec, NULL, 0, NULL, -1);
 979 }
 980 EXPORT_SYMBOL(llog_copy_handler);
 981
 982 /* backup plain llog */
 983 int llog_backup(const struct lu_env *env, struct obd_device *obd,
 984                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
 985                 char *name, char *backup)
 986 {
 987         struct llog_handle      *llh, *bllh;
 988         int                      rc;
 989
 990         ENTRY;
 991
 992         /* open original log */
 993         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
 994         if (rc < 0) {
 995                 /* the -ENOENT case is also reported to the caller
 996                  * but silently so it should handle that if needed.
 997                  */
 998                 if (rc != -ENOENT)
 999                         CERROR("%s: failed to open log %s: rc = %d\n",
1000                                obd->obd_name, name, rc);
1001                 RETURN(rc);
1002         }
1003
1004         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1005         if (rc)
1006                 GOTO(out_close, rc);
1007
1008         /* Make sure there's no old backup log */
1009         rc = llog_erase(env, bctxt, NULL, backup);
1010         if (rc < 0 && rc != -ENOENT)
1011                 GOTO(out_close, rc);
1012
1013         /* open backup log */
1014         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1015         if (rc) {
1016                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1017                        obd->obd_name, backup, rc);
1018                 GOTO(out_close, rc);
1019         }
1020
1021         /* check that backup llog is not the same object as original one */
1022         if (llh->lgh_obj == bllh->lgh_obj) {
1023                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1024                        obd->obd_name, name, backup, llh->lgh_obj,
1025                        bllh->lgh_obj);
1026                 GOTO(out_backup, rc = -EEXIST);
1027         }
1028
1029         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1030         if (rc)
1031                 GOTO(out_backup, rc);
1032
1033         /* Copy log record by record */
1034         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1035                                   NULL, false);
1036         if (rc)
1037                 CERROR("%s: failed to backup log %s: rc = %d\n",
1038                        obd->obd_name, name, rc);
1039 out_backup:
1040         llog_close(env, bllh);
1041 out_close:
1042         llog_close(env, llh);
1043         RETURN(rc);
1044 }
1045 EXPORT_SYMBOL(llog_backup);