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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/obdclass/llog.c
  33  *
  34  * OST<->MDS recovery logging infrastructure.
  35  * Invariants in implementation:
  36  * - we do not share logs among different OST<->MDS connections, so that
  37  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  38  *
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  41  * Author: Mikhail Pershin <tappro@whamcloud.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LOG
  45
  46 #include <linux/pid_namespace.h>
  47 #include <linux/kthread.h>
  48 #include <llog_swab.h>
  49 #include <lustre_log.h>
  50 #include <obd_support.h>
  51 #include <obd_class.h>
  52 #include "llog_internal.h"
  53 /*
  54  * Allocate a new log or catalog handle
  55  * Used inside llog_open().
  56  */
  57 static struct llog_handle *llog_alloc_handle(void)
  58 {
  59         struct llog_handle *loghandle;
  60
  61         OBD_ALLOC_PTR(loghandle);
  62         if (loghandle == NULL)
  63                 return NULL;
  64
  65         init_rwsem(&loghandle->lgh_lock);
  66         mutex_init(&loghandle->lgh_hdr_mutex);
  67         init_rwsem(&loghandle->lgh_last_sem);
  68         INIT_LIST_HEAD(&loghandle->u.phd.phd_entry);
  69         refcount_set(&loghandle->lgh_refcount, 1);
  70
  71         return loghandle;
  72 }
  73
  74 /*
  75  * Free llog handle and header data if exists. Used in llog_close() only
  76  */
  77 static void llog_free_handle(struct llog_handle *loghandle)
  78 {
  79         LASSERT(loghandle != NULL);
  80
  81         /* failed llog_init_handle */
  82         if (loghandle->lgh_hdr == NULL)
  83                 goto out;
  84
  85         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  86                 LASSERT(list_empty(&loghandle->u.phd.phd_entry));
  87         else if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  88                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  89         OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
  90 out:
  91         OBD_FREE_PTR(loghandle);
  92 }
  93
  94 struct llog_handle *llog_handle_get(struct llog_handle *loghandle)
  95 {
  96         if (refcount_inc_not_zero(&loghandle->lgh_refcount))
  97                 return loghandle;
  98         return NULL;
  99 }
 100
 101 int llog_handle_put(const struct lu_env *env, struct llog_handle *loghandle)
 102 {
 103         int rc = 0;
 104
 105         if (refcount_dec_and_test(&loghandle->lgh_refcount)) {
 106                 struct llog_operations *lop;
 107
 108                 rc = llog_handle2ops(loghandle, &lop);
 109                 if (!rc) {
 110                         if (lop->lop_close)
 111                                 rc = lop->lop_close(env, loghandle);
 112                         else
 113                                 rc = -EOPNOTSUPP;
 114                 }
 115                 llog_free_handle(loghandle);
 116         }
 117         return rc;
 118 }
 119
 120 static int llog_declare_destroy(const struct lu_env *env,
 121                                 struct llog_handle *handle,
 122                                 struct thandle *th)
 123 {
 124         struct llog_operations *lop;
 125         int rc;
 126
 127         ENTRY;
 128
 129         rc = llog_handle2ops(handle, &lop);
 130         if (rc)
 131                 RETURN(rc);
 132         if (lop->lop_declare_destroy == NULL)
 133                 RETURN(-EOPNOTSUPP);
 134
 135         rc = lop->lop_declare_destroy(env, handle, th);
 136
 137         RETURN(rc);
 138 }
 139
 140 int llog_trans_destroy(const struct lu_env *env, struct llog_handle *handle,
 141                        struct thandle *th)
 142 {
 143         struct llog_operations  *lop;
 144         int rc;
 145         ENTRY;
 146
 147         rc = llog_handle2ops(handle, &lop);
 148         if (rc < 0)
 149                 RETURN(rc);
 150         if (lop->lop_destroy == NULL)
 151                 RETURN(-EOPNOTSUPP);
 152
 153         LASSERT(handle->lgh_obj != NULL);
 154         if (!llog_exist(handle))
 155                 RETURN(0);
 156
 157         rc = lop->lop_destroy(env, handle, th);
 158
 159         RETURN(rc);
 160 }
 161
 162 int llog_destroy(const struct lu_env *env, struct llog_handle *handle)
 163 {
 164         struct llog_operations  *lop;
 165         struct dt_device        *dt;
 166         struct thandle          *th;
 167         int rc;
 168
 169         ENTRY;
 170
 171         rc = llog_handle2ops(handle, &lop);
 172         if (rc < 0)
 173                 RETURN(rc);
 174         if (lop->lop_destroy == NULL)
 175                 RETURN(-EOPNOTSUPP);
 176
 177         if (handle->lgh_obj == NULL) {
 178                 /* if lgh_obj == NULL, then it is from client side destroy */
 179                 rc = lop->lop_destroy(env, handle, NULL);
 180                 RETURN(rc);
 181         }
 182
 183         if (!llog_exist(handle))
 184                 RETURN(0);
 185
 186         dt = lu2dt_dev(handle->lgh_obj->do_lu.lo_dev);
 187
 188         if (unlikely(unlikely(dt->dd_rdonly)))
 189                 RETURN(-EROFS);
 190
 191         th = dt_trans_create(env, dt);
 192         if (IS_ERR(th))
 193                 RETURN(PTR_ERR(th));
 194
 195         rc = llog_declare_destroy(env, handle, th);
 196         if (rc != 0)
 197                 GOTO(out_trans, rc);
 198
 199         rc = dt_trans_start_local(env, dt, th);
 200         if (rc < 0)
 201                 GOTO(out_trans, rc);
 202
 203         rc = lop->lop_destroy(env, handle, th);
 204
 205 out_trans:
 206         dt_trans_stop(env, dt, th);
 207
 208         RETURN(rc);
 209 }
 210 EXPORT_SYMBOL(llog_destroy);
 211
 212 /* returns negative on error; 0 if success; 1 if success & log destroyed */
 213 int llog_cancel_arr_rec(const struct lu_env *env, struct llog_handle *loghandle,
 214                         int num, int *index)
 215 {
 216         struct llog_thread_info *lgi = llog_info(env);
 217         struct dt_device        *dt;
 218         struct llog_log_hdr     *llh;
 219         struct thandle          *th;
 220         __u32                    tmp_lgc_index;
 221         int                      rc, i = 0;
 222         int rc1;
 223         bool subtract_count = false;
 224
 225         ENTRY;
 226
 227         LASSERT(loghandle != NULL);
 228         LASSERT(loghandle->lgh_ctxt != NULL);
 229         LASSERT(loghandle->lgh_obj != NULL);
 230
 231         llh = loghandle->lgh_hdr;
 232
 233         CDEBUG(D_RPCTRACE, "Canceling %d records, first %d in log "DFID"\n",
 234                num, index[0], PFID(&loghandle->lgh_id.lgl_oi.oi_fid));
 235
 236         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 237
 238         if (unlikely(unlikely(dt->dd_rdonly)))
 239                 RETURN(0);
 240
 241         th = dt_trans_create(env, dt);
 242         if (IS_ERR(th))
 243                 RETURN(PTR_ERR(th));
 244
 245         rc = llog_declare_write_rec(env, loghandle, &llh->llh_hdr, 0, th);
 246         if (rc < 0)
 247                 GOTO(out_trans, rc);
 248
 249         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY)) {
 250                 rc = llog_declare_destroy(env, loghandle, th);
 251                 if (rc < 0)
 252                         GOTO(out_trans, rc);
 253         }
 254
 255         th->th_wait_submit = 1;
 256         rc = dt_trans_start_local(env, dt, th);
 257         if (rc < 0)
 258                 GOTO(out_trans, rc);
 259
 260         down_write(&loghandle->lgh_lock);
 261         /* clear bitmap */
 262         mutex_lock(&loghandle->lgh_hdr_mutex);
 263         for (i = 0; i < num; ++i) {
 264                 if (index[i] == 0) {
 265                         CERROR("Can't cancel index 0 which is header\n");
 266                         GOTO(out_unlock, rc = -EINVAL);
 267                 }
 268                 if (!__test_and_clear_bit_le(index[i], LLOG_HDR_BITMAP(llh))) {
 269                         CDEBUG(D_RPCTRACE, "Catalog index %u already clear?\n",
 270                                index[i]);
 271                         GOTO(out_unlock, rc = -ENOENT);
 272                 }
 273         }
 274         loghandle->lgh_hdr->llh_count -= num;
 275         subtract_count = true;
 276
 277         /* Since llog_process_thread use lgi_cookie, it`s better to save them
 278          * and restore after using
 279          */
 280         tmp_lgc_index = lgi->lgi_cookie.lgc_index;
 281         /* Pass this index to llog_osd_write_rec(), which will use the index
 282          * to only update the necesary bitmap. */
 283         lgi->lgi_cookie.lgc_index = index[0];
 284         /* update header */
 285         rc = llog_write_rec(env, loghandle, &llh->llh_hdr, (num != 1 ? NULL :
 286                             &lgi->lgi_cookie), LLOG_HEADER_IDX, th);
 287         lgi->lgi_cookie.lgc_index = tmp_lgc_index;
 288
 289         if (rc != 0)
 290                 GOTO(out_unlock, rc);
 291
 292         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 293             (llh->llh_count == 1) &&
 294             ((loghandle->lgh_last_idx == LLOG_HDR_BITMAP_SIZE(llh) - 1) ||
 295              (loghandle->u.phd.phd_cat_handle != NULL &&
 296               loghandle->u.phd.phd_cat_handle->u.chd.chd_current_log !=
 297                 loghandle))) {
 298                 /* never try to destroy it again */
 299                 llh->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 300                 rc = llog_trans_destroy(env, loghandle, th);
 301                 if (rc < 0) {
 302                         /* Sigh, can not destroy the final plain llog, but
 303                          * the bitmap has been clearly, so the record can not
 304                          * be accessed anymore, let's return 0 for now, and
 305                          * the orphan will be handled by LFSCK. */
 306                         CERROR("%s: can't destroy empty llog "DFID": rc = %d\n",
 307                                loghandle->lgh_ctxt->loc_obd->obd_name,
 308                                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rc);
 309                         GOTO(out_unlock, rc = 0);
 310                 }
 311                 rc = LLOG_DEL_PLAIN;
 312         }
 313
 314 out_unlock:
 315         mutex_unlock(&loghandle->lgh_hdr_mutex);
 316         up_write(&loghandle->lgh_lock);
 317 out_trans:
 318         rc1 = dt_trans_stop(env, dt, th);
 319         if (rc == 0)
 320                 rc = rc1;
 321         if (rc < 0) {
 322                 mutex_lock(&loghandle->lgh_hdr_mutex);
 323                 if (subtract_count)
 324                         loghandle->lgh_hdr->llh_count += num;
 325                 for (i = i - 1; i >= 0; i--)
 326                         set_bit_le(index[i], LLOG_HDR_BITMAP(llh));
 327                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 328         }
 329         RETURN(rc);
 330 }
 331
 332 int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
 333                     int index)
 334 {
 335         return llog_cancel_arr_rec(env, loghandle, 1, &index);
 336 }
 337
 338 int llog_read_header(const struct lu_env *env, struct llog_handle *handle,
 339                      const struct obd_uuid *uuid)
 340 {
 341         struct llog_operations *lop;
 342         int rc;
 343         ENTRY;
 344
 345         rc = llog_handle2ops(handle, &lop);
 346         if (rc)
 347                 RETURN(rc);
 348
 349         if (lop->lop_read_header == NULL)
 350                 RETURN(-EOPNOTSUPP);
 351
 352         rc = lop->lop_read_header(env, handle);
 353         if (rc == LLOG_EEMPTY) {
 354                 struct llog_log_hdr *llh = handle->lgh_hdr;
 355
 356                 /* lrh_len should be initialized in llog_init_handle */
 357                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 358                 llh->llh_count = 1;         /* for the header record */
 359                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 360                 LASSERT(handle->lgh_ctxt->loc_chunk_size >=
 361                                                 LLOG_MIN_CHUNK_SIZE);
 362                 llh->llh_hdr.lrh_len = handle->lgh_ctxt->loc_chunk_size;
 363                 llh->llh_hdr.lrh_index = 0;
 364                 llh->llh_timestamp = ktime_get_real_seconds();
 365                 if (uuid)
 366                         memcpy(&llh->llh_tgtuuid, uuid,
 367                                sizeof(llh->llh_tgtuuid));
 368                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 369                 /* Since update llog header might also call this function,
 370                  * let's reset the bitmap to 0 here */
 371                 memset(LLOG_HDR_BITMAP(llh), 0, llh->llh_hdr.lrh_len -
 372                                                 llh->llh_bitmap_offset -
 373                                                 sizeof(llh->llh_tail));
 374                 set_bit_le(0, LLOG_HDR_BITMAP(llh));
 375                 LLOG_HDR_TAIL(llh)->lrt_len = llh->llh_hdr.lrh_len;
 376                 LLOG_HDR_TAIL(llh)->lrt_index = llh->llh_hdr.lrh_index;
 377                 rc = 0;
 378         }
 379         RETURN(rc);
 380 }
 381 EXPORT_SYMBOL(llog_read_header);
 382
 383 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 384                      int flags, struct obd_uuid *uuid)
 385 {
 386         struct llog_log_hdr     *llh;
 387         enum llog_flag           fmt = flags & LLOG_F_EXT_MASK;
 388         int                      rc;
 389         int                     chunk_size = handle->lgh_ctxt->loc_chunk_size;
 390         ENTRY;
 391
 392         LASSERT(handle->lgh_hdr == NULL);
 393
 394         LASSERT(chunk_size >= LLOG_MIN_CHUNK_SIZE);
 395         OBD_ALLOC_LARGE(llh, chunk_size);
 396         if (llh == NULL)
 397                 RETURN(-ENOMEM);
 398
 399         handle->lgh_hdr = llh;
 400         handle->lgh_hdr_size = chunk_size;
 401         /* first assign flags to use llog_client_ops */
 402         llh->llh_flags = flags;
 403         rc = llog_read_header(env, handle, uuid);
 404         if (rc == 0) {
 405                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 406                               flags & LLOG_F_IS_CAT) ||
 407                              (llh->llh_flags & LLOG_F_IS_CAT &&
 408                               flags & LLOG_F_IS_PLAIN))) {
 409                         CERROR("%s: llog type is %s but initializing %s\n",
 410                                handle->lgh_ctxt->loc_obd->obd_name,
 411                                llh->llh_flags & LLOG_F_IS_CAT ?
 412                                "catalog" : "plain",
 413                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 414                         GOTO(out, rc = -EINVAL);
 415                 } else if (llh->llh_flags &
 416                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 417                         /*
 418                          * it is possible to open llog without specifying llog
 419                          * type so it is taken from llh_flags
 420                          */
 421                         flags = llh->llh_flags;
 422                 } else {
 423                         /* for some reason the llh_flags has no type set */
 424                         CERROR("llog type is not specified!\n");
 425                         GOTO(out, rc = -EINVAL);
 426                 }
 427                 if (unlikely(uuid &&
 428                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 429                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 430                                handle->lgh_ctxt->loc_obd->obd_name,
 431                                (char *)uuid->uuid,
 432                                (char *)llh->llh_tgtuuid.uuid);
 433                         GOTO(out, rc = -EEXIST);
 434                 }
 435         }
 436         if (flags & LLOG_F_IS_CAT) {
 437                 LASSERT(list_empty(&handle->u.chd.chd_head));
 438                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 439                 llh->llh_size = sizeof(struct llog_logid_rec);
 440                 llh->llh_flags |= LLOG_F_IS_FIXSIZE;
 441         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 442                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 443                        handle->lgh_ctxt->loc_obd->obd_name,
 444                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 445                 rc = -EINVAL;
 446         }
 447         llh->llh_flags |= fmt;
 448 out:
 449         if (rc) {
 450                 OBD_FREE_LARGE(llh, chunk_size);
 451                 handle->lgh_hdr = NULL;
 452         }
 453         RETURN(rc);
 454 }
 455 EXPORT_SYMBOL(llog_init_handle);
 456
 457 static int llog_process_thread(void *arg)
 458 {
 459         struct llog_process_info        *lpi = arg;
 460         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 461         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 462         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 463         struct llog_thread_info         *lti;
 464         char                            *buf;
 465         size_t                           chunk_size;
 466         __u64                            cur_offset;
 467         int                              rc = 0, index = 1, last_index;
 468         int                              saved_index = 0;
 469         int                              last_called_index = 0;
 470         bool                             repeated = false;
 471
 472         ENTRY;
 473
 474         if (llh == NULL)
 475                 RETURN(-EINVAL);
 476
 477         lti = lpi->lpi_env == NULL ? NULL : llog_info(lpi->lpi_env);
 478
 479         cur_offset = chunk_size = llh->llh_hdr.lrh_len;
 480         /* expect chunk_size to be power of two */
 481         LASSERT(is_power_of_2(chunk_size));
 482
 483         OBD_ALLOC_LARGE(buf, chunk_size);
 484         if (buf == NULL) {
 485                 lpi->lpi_rc = -ENOMEM;
 486                 RETURN(0);
 487         }
 488
 489         if (cd != NULL) {
 490                 last_called_index = cd->lpcd_first_idx;
 491                 index = cd->lpcd_first_idx + 1;
 492         }
 493         if (cd != NULL && cd->lpcd_last_idx)
 494                 last_index = cd->lpcd_last_idx;
 495         else
 496                 last_index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 497
 498         while (rc == 0) {
 499                 struct llog_rec_hdr *rec;
 500                 off_t chunk_offset = 0;
 501                 unsigned int buf_offset = 0;
 502                 int     lh_last_idx;
 503                 int     synced_idx = 0;
 504
 505                 /* skip records not set in bitmap */
 506                 while (index <= last_index &&
 507                        !test_bit_le(index, LLOG_HDR_BITMAP(llh)))
 508                         ++index;
 509
 510                 /* There are no indices prior the last_index */
 511                 if (index > last_index)
 512                         break;
 513
 514                 CDEBUG(D_OTHER, "index: %d last_index %d\n", index,
 515                        last_index);
 516
 517 repeat:
 518                 /* get the buf with our target record; avoid old garbage */
 519                 memset(buf, 0, chunk_size);
 520                 /* the record index for outdated chunk data */
 521                 /* it is safe to process buffer until saved lgh_last_idx */
 522                 lh_last_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 523                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 524                                      index, &cur_offset, buf, chunk_size);
 525                 if (repeated && rc)
 526                         CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
 527                                " buf_offset %u, rc = %d\n", cur_offset,
 528                                (__u64)chunk_offset, buf_offset, rc);
 529                 /* we`ve tried to reread the chunk, but there is no
 530                  * new records */
 531                 if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
 532                     cur_offset)
 533                         GOTO(out, rc = 0);
 534                 if (rc != 0)
 535                         GOTO(out, rc);
 536
 537                 /* NB: after llog_next_block() call the cur_offset is the
 538                  * offset of the next block after read one.
 539                  * The absolute offset of the current chunk is calculated
 540                  * from cur_offset value and stored in chunk_offset variable.
 541                  */
 542                 if ((cur_offset & (chunk_size - 1)) != 0)
 543                         chunk_offset = cur_offset & ~(chunk_size - 1);
 544                 else
 545                         chunk_offset = cur_offset - chunk_size;
 546
 547                 /* NB: when rec->lrh_len is accessed it is already swabbed
 548                  * since it is used at the "end" of the loop and the rec
 549                  * swabbing is done at the beginning of the loop. */
 550                 for (rec = (struct llog_rec_hdr *)(buf + buf_offset);
 551                      (char *)rec < buf + chunk_size;
 552                      rec = llog_rec_hdr_next(rec)) {
 553
 554                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 555                                rec, rec->lrh_type);
 556
 557                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 558                                 lustre_swab_llog_rec(rec);
 559
 560                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 561                                rec->lrh_type, rec->lrh_index);
 562
 563                         if (index == (synced_idx + 1) &&
 564                             synced_idx == LLOG_HDR_TAIL(llh)->lrt_index)
 565                                 GOTO(out, rc = 0);
 566
 567                         if (OBD_FAIL_PRECHECK(OBD_FAIL_LLOG_PROCESS_TIMEOUT) &&
 568                                 cfs_fail_val == (unsigned int)
 569                                         (loghandle->lgh_id.lgl_oi.oi.oi_id &
 570                                          0xFFFFFFFF)) {
 571                                 OBD_RACE(OBD_FAIL_LLOG_PROCESS_TIMEOUT);
 572                         }
 573
 574                         /* the bitmap could be changed during processing
 575                          * records from the chunk. For wrapped catalog
 576                          * it means we can read deleted record and try to
 577                          * process it. Check this case and reread the chunk.
 578                          * It is safe to process to lh_last_idx, including
 579                          * lh_last_idx if it was synced. We can not do <=
 580                          * comparison, cause for wrapped catalog lgh_last_idx
 581                          * could be less than index. So we detect last index
 582                          * for processing as index == lh_last_idx+1. But when
 583                          * catalog is wrapped and full lgh_last_idx=llh_cat_idx,
 584                          * the first processing index is llh_cat_idx+1.The
 585                          * exception is !(lgh_last_idx == llh_cat_idx &&
 586                          * index == llh_cat_idx + 1), and after simplification
 587                          * it turns to
 588                          * lh_last_idx != LLOG_HDR_TAIL(llh)->lrt_index
 589                          * This exception is working for catalog only.
 590                          */
 591
 592                         if ((index == lh_last_idx && synced_idx != index) ||
 593                             (index == (lh_last_idx + 1) &&
 594                              lh_last_idx != LLOG_HDR_TAIL(llh)->lrt_index) ||
 595                             (rec->lrh_index == 0 && !repeated)) {
 596
 597                                 /* save offset inside buffer for the re-read */
 598                                 buf_offset = (char *)rec - (char *)buf;
 599                                 cur_offset = chunk_offset;
 600                                 repeated = true;
 601                                 /* We need to be sure lgh_last_idx
 602                                  * record was saved to disk
 603                                  */
 604                                 down_read(&loghandle->lgh_last_sem);
 605                                 synced_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 606                                 up_read(&loghandle->lgh_last_sem);
 607                                 CDEBUG(D_OTHER, "synced_idx: %d\n", synced_idx);
 608                                 goto repeat;
 609
 610                         }
 611
 612                         repeated = false;
 613
 614                         if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
 615                                 CWARN("%s: invalid length %d in llog "DFID
 616                                       "record for index %d/%d\n",
 617                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 618                                        rec->lrh_len,
 619                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 620                                        rec->lrh_index, index);
 621
 622                                 GOTO(out, rc = -EINVAL);
 623                         }
 624
 625                         if (rec->lrh_index < index) {
 626                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 627                                        rec->lrh_index);
 628                                 continue;
 629                         }
 630
 631                         if (rec->lrh_index != index) {
 632                                 CERROR("%s: "DFID" Invalid record: index %u"
 633                                        " but expected %u\n",
 634                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 635                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 636                                        rec->lrh_index, index);
 637                                 GOTO(out, rc = -ERANGE);
 638                         }
 639
 640                         CDEBUG(D_OTHER,
 641                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 642                                rec->lrh_index, rec->lrh_len,
 643                                (int)(buf + chunk_size - (char *)rec));
 644
 645                         /* lgh_cur_offset is used only at llog_test_3 */
 646                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 647                                                     chunk_offset;
 648
 649                         /* if set, process the callback on this record */
 650                         if (test_bit_le(index, LLOG_HDR_BITMAP(llh))) {
 651                                 struct llog_cookie *lgc;
 652                                 __u64   tmp_off;
 653                                 int     tmp_idx;
 654
 655                                 CDEBUG((llh->llh_flags & LLOG_F_IS_CAT ?
 656                                         D_HA : D_OTHER),
 657                                        "index: %d, lh_last_idx: %d "
 658                                        "synced_idx: %d lgh_last_idx: %d\n",
 659                                        index, lh_last_idx, synced_idx,
 660                                        loghandle->lgh_last_idx);
 661
 662                                 if (lti != NULL) {
 663                                         lgc = &lti->lgi_cookie;
 664                                         /* store lu_env for recursive calls */
 665                                         tmp_off = lgc->lgc_offset;
 666                                         tmp_idx = lgc->lgc_index;
 667
 668                                         lgc->lgc_offset = (char *)rec -
 669                                                 (char *)buf + chunk_offset;
 670                                         lgc->lgc_index = rec->lrh_index;
 671                                 }
 672                                 /* using lu_env for passing record offset to
 673                                  * llog_write through various callbacks */
 674                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 675                                                  lpi->lpi_cbdata);
 676                                 last_called_index = index;
 677
 678                                 if (lti != NULL) {
 679                                         lgc->lgc_offset = tmp_off;
 680                                         lgc->lgc_index = tmp_idx;
 681                                 }
 682
 683                                 if (rc == LLOG_PROC_BREAK) {
 684                                         GOTO(out, rc);
 685                                 } else if (rc == LLOG_DEL_RECORD) {
 686                                         rc = llog_cancel_rec(lpi->lpi_env,
 687                                                              loghandle,
 688                                                              rec->lrh_index);
 689                                 }
 690                                 if (rc)
 691                                         GOTO(out, rc);
 692                                 /* some stupid callbacks directly cancel records
 693                                  * and delete llog. Check it and stop
 694                                  * processing. */
 695                                 if (loghandle->lgh_hdr == NULL ||
 696                                     loghandle->lgh_hdr->llh_count == 1)
 697                                         GOTO(out, rc = 0);
 698                         }
 699                         /* exit if the last index is reached */
 700                         if (index >= last_index)
 701                                 GOTO(out, rc = 0);
 702                         ++index;
 703                 }
 704         }
 705
 706 out:
 707         CDEBUG(D_HA, "stop processing %s "DOSTID":%x index %d count %d\n",
 708                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" : "plain"),
 709                POSTID(&loghandle->lgh_id.lgl_oi), loghandle->lgh_id.lgl_ogen,
 710                index, llh->llh_count);
 711
 712         if (cd != NULL)
 713                 cd->lpcd_last_idx = last_called_index;
 714
 715         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 716                 if (dt_object_remote(loghandle->lgh_obj)) {
 717                         /* If it is remote object, then -EIO might means
 718                          * disconnection or eviction, let's return -EAGAIN,
 719                          * so for update recovery log processing, it will
 720                          * retry until the umount or abort recovery, see
 721                          * lod_sub_recovery_thread() */
 722                         CERROR("%s retry remote llog process\n",
 723                                loghandle->lgh_ctxt->loc_obd->obd_name);
 724                         rc = -EAGAIN;
 725                 } else {
 726                         /* something bad happened to the processing of a local
 727                          * llog file, probably I/O error or the log got
 728                          * corrupted to be able to finally release the log we
 729                          * discard any remaining bits in the header */
 730                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 731                                " %s index %d count %d\n",
 732                                loghandle->lgh_ctxt->loc_obd->obd_name,
 733                                POSTID(&loghandle->lgh_id.lgl_oi),
 734                                loghandle->lgh_id.lgl_ogen,
 735                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 736                                 "plain"), index, llh->llh_count);
 737
 738                         while (index <= last_index) {
 739                                 if (test_bit_le(index,
 740                                                   LLOG_HDR_BITMAP(llh)) != 0)
 741                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 742                                                         index);
 743                                 index++;
 744                         }
 745                         rc = 0;
 746                 }
 747         }
 748
 749         OBD_FREE_LARGE(buf, chunk_size);
 750         lpi->lpi_rc = rc;
 751         return 0;
 752 }
 753
 754 static int llog_process_thread_daemonize(void *arg)
 755 {
 756         struct llog_process_info        *lpi = arg;
 757         struct lu_env                    env;
 758         int                              rc;
 759         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 760
 761         task_lock(lpi->lpi_reftask);
 762         new_ns = lpi->lpi_reftask->nsproxy;
 763         if (curr_ns != new_ns) {
 764                 get_nsproxy(new_ns);
 765
 766                 current->nsproxy = new_ns;
 767                 /* XXX: we should call put_nsproxy() instead of
 768                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 769                  * used outside of the kernel itself, because it calls
 770                  * free_nsproxy() which is not exported by the kernel
 771                  * (defined in kernel/nsproxy.c) */
 772                 if (curr_ns)
 773                         atomic_dec(&curr_ns->count);
 774         }
 775         task_unlock(lpi->lpi_reftask);
 776
 777         /* client env has no keys, tags is just 0 */
 778         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 779         if (rc)
 780                 goto out;
 781         lpi->lpi_env = &env;
 782
 783         rc = llog_process_thread(arg);
 784
 785         lu_env_fini(&env);
 786 out:
 787         complete(&lpi->lpi_completion);
 788         return rc;
 789 }
 790
 791 int llog_process_or_fork(const struct lu_env *env,
 792                          struct llog_handle *loghandle,
 793                          llog_cb_t cb, void *data, void *catdata, bool fork)
 794 {
 795         struct llog_process_info *lpi;
 796         struct llog_process_data *d = data;
 797         struct llog_process_cat_data *cd = catdata;
 798         __u32 flags = loghandle->lgh_hdr->llh_flags;
 799         int rc;
 800
 801         ENTRY;
 802
 803         OBD_ALLOC_PTR(lpi);
 804         if (lpi == NULL) {
 805                 CERROR("cannot alloc pointer\n");
 806                 RETURN(-ENOMEM);
 807         }
 808         lpi->lpi_loghandle = loghandle;
 809         lpi->lpi_cb        = cb;
 810         lpi->lpi_cbdata    = data;
 811         lpi->lpi_catdata   = catdata;
 812
 813         CDEBUG(D_OTHER, "Processing "DFID" flags 0x%03x startcat %d startidx %d first_idx %d last_idx %d\n",
 814                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), flags,
 815                (flags & LLOG_F_IS_CAT) && d ? d->lpd_startcat : -1,
 816                (flags & LLOG_F_IS_CAT) && d ? d->lpd_startidx : -1,
 817                cd ? cd->lpcd_first_idx : -1, cd ? cd->lpcd_last_idx : -1);
 818         if (fork) {
 819                 struct task_struct *task;
 820
 821                 /* The new thread can't use parent env,
 822                  * init the new one in llog_process_thread_daemonize. */
 823                 lpi->lpi_env = NULL;
 824                 init_completion(&lpi->lpi_completion);
 825                 /* take reference to current, so that
 826                  * llog_process_thread_daemonize() can use it to switch to
 827                  * namespace associated with current  */
 828                 lpi->lpi_reftask = current;
 829                 task = kthread_run(llog_process_thread_daemonize, lpi,
 830                                    "llog_process_thread");
 831                 if (IS_ERR(task)) {
 832                         rc = PTR_ERR(task);
 833                         CERROR("%s: cannot start thread: rc = %d\n",
 834                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 835                         GOTO(out_lpi, rc);
 836                 }
 837                 wait_for_completion(&lpi->lpi_completion);
 838         } else {
 839                 lpi->lpi_env = env;
 840                 llog_process_thread(lpi);
 841         }
 842         rc = lpi->lpi_rc;
 843
 844 out_lpi:
 845         OBD_FREE_PTR(lpi);
 846         RETURN(rc);
 847 }
 848 EXPORT_SYMBOL(llog_process_or_fork);
 849
 850 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 851                  llog_cb_t cb, void *data, void *catdata)
 852 {
 853         int rc;
 854         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 855         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 856 }
 857 EXPORT_SYMBOL(llog_process);
 858
 859 static inline const struct cred *llog_raise_resource(void)
 860 {
 861         struct cred *cred = NULL;
 862
 863         if (cap_raised(current_cap(), CAP_SYS_RESOURCE))
 864                 return cred;
 865
 866         cred = prepare_creds();
 867         if (!cred)
 868                 return cred;
 869
 870         cap_raise(cred->cap_effective, CAP_SYS_RESOURCE);
 871         return override_creds(cred);
 872 }
 873
 874 static inline void llog_restore_resource(const struct cred *old_cred)
 875 {
 876         if (old_cred)
 877                 revert_creds(old_cred);
 878 }
 879
 880 int llog_reverse_process(const struct lu_env *env,
 881                          struct llog_handle *loghandle, llog_cb_t cb,
 882                          void *data, void *catdata)
 883 {
 884         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 885         struct llog_process_cat_data *cd = catdata;
 886         void *buf;
 887         int rc = 0, first_index = 1, index, idx;
 888         __u32   chunk_size = llh->llh_hdr.lrh_len;
 889         ENTRY;
 890
 891         OBD_ALLOC_LARGE(buf, chunk_size);
 892         if (buf == NULL)
 893                 RETURN(-ENOMEM);
 894
 895         if (cd != NULL)
 896                 first_index = cd->lpcd_first_idx + 1;
 897         if (cd != NULL && cd->lpcd_last_idx)
 898                 index = cd->lpcd_last_idx;
 899         else
 900                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 901
 902         while (rc == 0) {
 903                 struct llog_rec_hdr *rec;
 904                 struct llog_rec_tail *tail;
 905
 906                 /* skip records not set in bitmap */
 907                 while (index >= first_index &&
 908                        !test_bit_le(index, LLOG_HDR_BITMAP(llh)))
 909                         --index;
 910
 911                 LASSERT(index >= first_index - 1);
 912                 if (index == first_index - 1)
 913                         break;
 914
 915                 /* get the buf with our target record; avoid old garbage */
 916                 memset(buf, 0, chunk_size);
 917                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 918                 if (rc)
 919                         GOTO(out, rc);
 920
 921                 rec = buf;
 922                 idx = rec->lrh_index;
 923                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 924                 while (idx < index) {
 925                         rec = (void *)rec + rec->lrh_len;
 926                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 927                                 lustre_swab_llog_rec(rec);
 928                         idx ++;
 929                 }
 930                 LASSERT(idx == index);
 931                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 932
 933                 /* process records in buffer, starting where we found one */
 934                 while ((void *)tail > buf) {
 935                         if (tail->lrt_index == 0)
 936                                 GOTO(out, rc = 0); /* no more records */
 937
 938                         /* if set, process the callback on this record */
 939                         if (test_bit_le(index, LLOG_HDR_BITMAP(llh))) {
 940                                 rec = (void *)tail - tail->lrt_len +
 941                                       sizeof(*tail);
 942
 943                                 rc = cb(env, loghandle, rec, data);
 944                                 if (rc == LLOG_PROC_BREAK) {
 945                                         GOTO(out, rc);
 946                                 } else if (rc == LLOG_DEL_RECORD) {
 947                                         rc = llog_cancel_rec(env, loghandle,
 948                                                              tail->lrt_index);
 949                                 }
 950                                 if (rc)
 951                                         GOTO(out, rc);
 952                         }
 953
 954                         /* previous record, still in buffer? */
 955                         --index;
 956                         if (index < first_index)
 957                                 GOTO(out, rc = 0);
 958                         tail = (void *)tail - tail->lrt_len;
 959                 }
 960         }
 961
 962 out:
 963         if (buf != NULL)
 964                 OBD_FREE_LARGE(buf, chunk_size);
 965         RETURN(rc);
 966 }
 967 EXPORT_SYMBOL(llog_reverse_process);
 968
 969 /**
 970  * new llog API
 971  *
 972  * API functions:
 973  *      llog_open - open llog, may not exist
 974  *      llog_exist - check if llog exists
 975  *      llog_close - close opened llog, pair for open, frees llog_handle
 976  *      llog_declare_create - declare llog creation
 977  *      llog_create - create new llog on disk, need transaction handle
 978  *      llog_declare_write_rec - declaration of llog write
 979  *      llog_write_rec - write llog record on disk, need transaction handle
 980  *      llog_declare_add - declare llog catalog record addition
 981  *      llog_add - add llog record in catalog, need transaction handle
 982  */
 983 int llog_exist(struct llog_handle *loghandle)
 984 {
 985         struct llog_operations  *lop;
 986         int                      rc;
 987
 988         ENTRY;
 989
 990         rc = llog_handle2ops(loghandle, &lop);
 991         if (rc)
 992                 RETURN(rc);
 993         if (lop->lop_exist == NULL)
 994                 RETURN(-EOPNOTSUPP);
 995
 996         rc = lop->lop_exist(loghandle);
 997         RETURN(rc);
 998 }
 999 EXPORT_SYMBOL(llog_exist);
1000
1001 int llog_declare_create(const struct lu_env *env,
1002                         struct llog_handle *loghandle, struct thandle *th)
1003 {
1004         const struct cred *old_cred;
1005         struct llog_operations  *lop;
1006         int rc;
1007
1008         ENTRY;
1009
1010         rc = llog_handle2ops(loghandle, &lop);
1011         if (rc)
1012                 RETURN(rc);
1013         if (lop->lop_declare_create == NULL)
1014                 RETURN(-EOPNOTSUPP);
1015
1016         old_cred = llog_raise_resource();
1017         rc = lop->lop_declare_create(env, loghandle, th);
1018         llog_restore_resource(old_cred);
1019         RETURN(rc);
1020 }
1021
1022 int llog_create(const struct lu_env *env, struct llog_handle *handle,
1023                 struct thandle *th)
1024 {
1025         const struct cred *old_cred;
1026         struct llog_operations  *lop;
1027         int rc;
1028
1029         ENTRY;
1030
1031         rc = llog_handle2ops(handle, &lop);
1032         if (rc)
1033                 RETURN(rc);
1034         if (lop->lop_create == NULL)
1035                 RETURN(-EOPNOTSUPP);
1036
1037         old_cred = llog_raise_resource();
1038         rc = lop->lop_create(env, handle, th);
1039         llog_restore_resource(old_cred);
1040         RETURN(rc);
1041 }
1042
1043 int llog_declare_write_rec(const struct lu_env *env,
1044                            struct llog_handle *handle,
1045                            struct llog_rec_hdr *rec, int idx,
1046                            struct thandle *th)
1047 {
1048         const struct cred *old_cred;
1049         struct llog_operations  *lop;
1050         int rc;
1051
1052         ENTRY;
1053
1054         rc = llog_handle2ops(handle, &lop);
1055         if (rc)
1056                 RETURN(rc);
1057         LASSERT(lop);
1058         if (lop->lop_declare_write_rec == NULL)
1059                 RETURN(-EOPNOTSUPP);
1060
1061         old_cred = llog_raise_resource();
1062         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
1063         llog_restore_resource(old_cred);
1064         RETURN(rc);
1065 }
1066
1067 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
1068                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1069                    int idx, struct thandle *th)
1070 {
1071         const struct cred *old_cred;
1072         struct llog_operations  *lop;
1073         int rc, buflen;
1074
1075         ENTRY;
1076
1077         /* API sanity checks */
1078         if (handle == NULL) {
1079                 CERROR("loghandle is missed\n");
1080                 RETURN(-EPROTO);
1081         } else if (handle->lgh_obj == NULL) {
1082                 CERROR("loghandle %p with NULL object\n",
1083                         handle);
1084                 RETURN(-EPROTO);
1085         } else if (th == NULL) {
1086                 CERROR("%s: missed transaction handle\n",
1087                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name);
1088                 RETURN(-EPROTO);
1089         } else if (handle->lgh_hdr == NULL) {
1090                 CERROR("%s: loghandle %p with no header\n",
1091                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name,
1092                         handle);
1093                 RETURN(-EPROTO);
1094         }
1095
1096         rc = llog_handle2ops(handle, &lop);
1097         if (rc)
1098                 RETURN(rc);
1099
1100         if (lop->lop_write_rec == NULL)
1101                 RETURN(-EOPNOTSUPP);
1102
1103         buflen = rec->lrh_len;
1104         LASSERT(cfs_size_round(buflen) == buflen);
1105
1106         old_cred = llog_raise_resource();
1107         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1108         llog_restore_resource(old_cred);
1109         RETURN(rc);
1110 }
1111
1112 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1113              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1114              struct thandle *th)
1115 {
1116         const struct cred *old_cred;
1117         int rc;
1118
1119         ENTRY;
1120
1121         if (lgh->lgh_logops->lop_add == NULL)
1122                 RETURN(-EOPNOTSUPP);
1123
1124         old_cred = llog_raise_resource();
1125         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1126         llog_restore_resource(old_cred);
1127         RETURN(rc);
1128 }
1129 EXPORT_SYMBOL(llog_add);
1130
1131 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1132                      struct llog_rec_hdr *rec, struct thandle *th)
1133 {
1134         const struct cred *old_cred;
1135         int rc;
1136
1137         ENTRY;
1138
1139         if (lgh->lgh_logops->lop_declare_add == NULL)
1140                 RETURN(-EOPNOTSUPP);
1141
1142         old_cred = llog_raise_resource();
1143         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1144         llog_restore_resource(old_cred);
1145         RETURN(rc);
1146 }
1147 EXPORT_SYMBOL(llog_declare_add);
1148
1149 /**
1150  * Helper function to open llog or create it if doesn't exist.
1151  * It hides all transaction handling from caller.
1152  */
1153 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1154                      struct llog_handle **res, struct llog_logid *logid,
1155                      char *name)
1156 {
1157         struct dt_device        *d;
1158         struct thandle          *th;
1159         int                      rc;
1160
1161         ENTRY;
1162
1163         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1164         if (rc)
1165                 RETURN(rc);
1166
1167         if (llog_exist(*res))
1168                 RETURN(0);
1169
1170         LASSERT((*res)->lgh_obj != NULL);
1171
1172         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1173
1174         if (unlikely(unlikely(d->dd_rdonly)))
1175                 RETURN(-EROFS);
1176
1177         th = dt_trans_create(env, d);
1178         if (IS_ERR(th))
1179                 GOTO(out, rc = PTR_ERR(th));
1180
1181         /* Create update llog object synchronously, which
1182          * happens during inialization process see
1183          * lod_sub_prep_llog(), to make sure the update
1184          * llog object is created before corss-MDT writing
1185          * updates into the llog object */
1186         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1187                 th->th_sync = 1;
1188
1189         th->th_wait_submit = 1;
1190         rc = llog_declare_create(env, *res, th);
1191         if (rc == 0) {
1192                 rc = dt_trans_start_local(env, d, th);
1193                 if (rc == 0)
1194                         rc = llog_create(env, *res, th);
1195         }
1196         dt_trans_stop(env, d, th);
1197 out:
1198         if (rc)
1199                 llog_close(env, *res);
1200         RETURN(rc);
1201 }
1202 EXPORT_SYMBOL(llog_open_create);
1203
1204 /**
1205  * Helper function to delete existent llog.
1206  */
1207 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1208                struct llog_logid *logid, char *name)
1209 {
1210         struct llog_handle      *handle;
1211         int                      rc = 0, rc2;
1212
1213         ENTRY;
1214
1215         /* nothing to erase */
1216         if (name == NULL && logid == NULL)
1217                 RETURN(0);
1218
1219         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1220         if (rc < 0)
1221                 RETURN(rc);
1222
1223         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1224         if (rc == 0)
1225                 rc = llog_destroy(env, handle);
1226
1227         rc2 = llog_close(env, handle);
1228         if (rc == 0)
1229                 rc = rc2;
1230         RETURN(rc);
1231 }
1232 EXPORT_SYMBOL(llog_erase);
1233
1234 /*
1235  * Helper function for write record in llog.
1236  * It hides all transaction handling from caller.
1237  * Valid only with local llog.
1238  */
1239 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1240                struct llog_rec_hdr *rec, int idx)
1241 {
1242         struct dt_device        *dt;
1243         struct thandle          *th;
1244         bool                    need_cookie;
1245         int                     rc;
1246
1247         ENTRY;
1248
1249         LASSERT(loghandle);
1250         LASSERT(loghandle->lgh_ctxt);
1251         LASSERT(loghandle->lgh_obj != NULL);
1252
1253         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1254
1255         if (unlikely(unlikely(dt->dd_rdonly)))
1256                 RETURN(-EROFS);
1257
1258         th = dt_trans_create(env, dt);
1259         if (IS_ERR(th))
1260                 RETURN(PTR_ERR(th));
1261
1262         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1263         if (rc)
1264                 GOTO(out_trans, rc);
1265
1266         th->th_wait_submit = 1;
1267         rc = dt_trans_start_local(env, dt, th);
1268         if (rc)
1269                 GOTO(out_trans, rc);
1270
1271         need_cookie = !(idx == LLOG_HEADER_IDX || idx == LLOG_NEXT_IDX);
1272
1273         down_write(&loghandle->lgh_lock);
1274         if (need_cookie) {
1275                 struct llog_thread_info *lti = llog_info(env);
1276
1277                 /* cookie comes from llog_process_thread */
1278                 rc = llog_write_rec(env, loghandle, rec, &lti->lgi_cookie,
1279                                     rec->lrh_index, th);
1280                 /* upper layer didn`t pass cookie so change rc */
1281                 rc = (rc == 1 ? 0 : rc);
1282         } else {
1283                 rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1284         }
1285
1286         up_write(&loghandle->lgh_lock);
1287 out_trans:
1288         dt_trans_stop(env, dt, th);
1289         RETURN(rc);
1290 }
1291 EXPORT_SYMBOL(llog_write);
1292
1293 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1294               struct llog_handle **lgh, struct llog_logid *logid,
1295               char *name, enum llog_open_param open_param)
1296 {
1297         const struct cred *old_cred;
1298         int      rc;
1299
1300         ENTRY;
1301
1302         LASSERT(ctxt);
1303         LASSERT(ctxt->loc_logops);
1304
1305         if (ctxt->loc_logops->lop_open == NULL) {
1306                 *lgh = NULL;
1307                 RETURN(-EOPNOTSUPP);
1308         }
1309
1310         *lgh = llog_alloc_handle();
1311         if (*lgh == NULL)
1312                 RETURN(-ENOMEM);
1313         (*lgh)->lgh_ctxt = ctxt;
1314         (*lgh)->lgh_logops = ctxt->loc_logops;
1315
1316         old_cred = llog_raise_resource();
1317         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1318         llog_restore_resource(old_cred);
1319         if (rc) {
1320                 llog_free_handle(*lgh);
1321                 *lgh = NULL;
1322         }
1323         RETURN(rc);
1324 }
1325 EXPORT_SYMBOL(llog_open);
1326
1327 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1328 {
1329         return llog_handle_put(env, loghandle);
1330 }
1331 EXPORT_SYMBOL(llog_close);
1332
1333 /**
1334  * Helper function to get the llog size in records. It is used by MGS
1335  * mostly to check that config llog exists and contains data.
1336  *
1337  * \param[in] env       execution environment
1338  * \param[in] ctxt      llog context
1339  * \param[in] name      llog name
1340  *
1341  * \retval              true if there are records in llog besides a header
1342  * \retval              false on error or llog without records
1343  */
1344 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1345                   char *name)
1346 {
1347         struct llog_handle      *llh;
1348         int                      rc = 0;
1349
1350         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1351         if (rc < 0) {
1352                 if (likely(rc == -ENOENT))
1353                         rc = 0;
1354                 GOTO(out, rc);
1355         }
1356
1357         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1358         if (rc)
1359                 GOTO(out_close, rc);
1360         rc = llog_get_size(llh);
1361
1362 out_close:
1363         llog_close(env, llh);
1364 out:
1365         /* The header is record 1, the llog is still considered as empty
1366          * if there is only header */
1367         return (rc <= 1);
1368 }
1369 EXPORT_SYMBOL(llog_is_empty);
1370
1371 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1372                       struct llog_rec_hdr *rec, void *data)
1373 {
1374         struct llog_handle      *copy_llh = data;
1375
1376         /* Append all records */
1377         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1378 }
1379
1380 /* backup plain llog */
1381 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1382                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1383                 char *name, char *backup)
1384 {
1385         struct llog_handle      *llh, *bllh;
1386         int                      rc;
1387
1388         ENTRY;
1389
1390         /* open original log */
1391         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1392         if (rc < 0) {
1393                 /* the -ENOENT case is also reported to the caller
1394                  * but silently so it should handle that if needed.
1395                  */
1396                 if (rc != -ENOENT)
1397                         CERROR("%s: failed to open log %s: rc = %d\n",
1398                                obd->obd_name, name, rc);
1399                 RETURN(rc);
1400         }
1401
1402         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1403         if (rc)
1404                 GOTO(out_close, rc);
1405
1406         /* Make sure there's no old backup log */
1407         rc = llog_erase(env, bctxt, NULL, backup);
1408         if (rc < 0 && rc != -ENOENT)
1409                 GOTO(out_close, rc);
1410
1411         /* open backup log */
1412         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1413         if (rc) {
1414                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1415                        obd->obd_name, backup, rc);
1416                 GOTO(out_close, rc);
1417         }
1418
1419         /* check that backup llog is not the same object as original one */
1420         if (llh->lgh_obj == bllh->lgh_obj) {
1421                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1422                        obd->obd_name, name, backup, llh->lgh_obj,
1423                        bllh->lgh_obj);
1424                 GOTO(out_backup, rc = -EEXIST);
1425         }
1426
1427         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1428         if (rc)
1429                 GOTO(out_backup, rc);
1430
1431         /* Copy log record by record */
1432         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1433                                   NULL, false);
1434         if (rc)
1435                 CERROR("%s: failed to backup log %s: rc = %d\n",
1436                        obd->obd_name, name, rc);
1437 out_backup:
1438         llog_close(env, bllh);
1439 out_close:
1440         llog_close(env, llh);
1441         RETURN(rc);
1442 }
1443 EXPORT_SYMBOL(llog_backup);
1444
1445 /* Get size of llog */
1446 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1447 {
1448         int rc;
1449         struct lu_attr la;
1450
1451         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1452         if (rc) {
1453                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1454                        llh->lgh_ctxt->loc_obd->obd_name,
1455                        PFID(&llh->lgh_id.lgl_oi.oi_fid), rc);
1456                 return 0;
1457         }
1458
1459         return la.la_size;
1460 }
1461 EXPORT_SYMBOL(llog_size);
1462