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 (!ext2_clear_bit(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                         ext2_set_bit(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                 ext2_set_bit(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                 bool partial_chunk;
 503                 int     lh_last_idx;
 504                 int     synced_idx = 0;
 505
 506                 /* skip records not set in bitmap */
 507                 while (index <= last_index &&
 508                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 509                         ++index;
 510
 511                 /* There are no indices prior the last_index */
 512                 if (index > last_index)
 513                         break;
 514
 515                 CDEBUG(D_OTHER, "index: %d last_index %d\n", index,
 516                        last_index);
 517
 518 repeat:
 519                 /* get the buf with our target record; avoid old garbage */
 520                 memset(buf, 0, chunk_size);
 521                 /* the record index for outdated chunk data */
 522                 /* it is safe to process buffer until saved lgh_last_idx */
 523                 lh_last_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 524                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 525                                      index, &cur_offset, buf, chunk_size);
 526                 if (repeated && rc)
 527                         CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
 528                                " buf_offset %u, rc = %d\n", cur_offset,
 529                                (__u64)chunk_offset, buf_offset, rc);
 530                 /* we`ve tried to reread the chunk, but there is no
 531                  * new records */
 532                 if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
 533                     cur_offset)
 534                         GOTO(out, rc = 0);
 535                 if (rc != 0)
 536                         GOTO(out, rc);
 537
 538                 /* NB: after llog_next_block() call the cur_offset is the
 539                  * offset of the next block after read one.
 540                  * The absolute offset of the current chunk is calculated
 541                  * from cur_offset value and stored in chunk_offset variable.
 542                  */
 543                 if ((cur_offset & (chunk_size - 1)) != 0) {
 544                         partial_chunk = true;
 545                         chunk_offset = cur_offset & ~(chunk_size - 1);
 546                 } else {
 547                         partial_chunk = false;
 548                         chunk_offset = cur_offset - chunk_size;
 549                 }
 550
 551                 /* NB: when rec->lrh_len is accessed it is already swabbed
 552                  * since it is used at the "end" of the loop and the rec
 553                  * swabbing is done at the beginning of the loop. */
 554                 for (rec = (struct llog_rec_hdr *)(buf + buf_offset);
 555                      (char *)rec < buf + chunk_size;
 556                      rec = llog_rec_hdr_next(rec)) {
 557
 558                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 559                                rec, rec->lrh_type);
 560
 561                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 562                                 lustre_swab_llog_rec(rec);
 563
 564                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 565                                rec->lrh_type, rec->lrh_index);
 566
 567                         if (index == (synced_idx + 1) &&
 568                             synced_idx == LLOG_HDR_TAIL(llh)->lrt_index)
 569                                 GOTO(out, rc = 0);
 570
 571                         if (OBD_FAIL_PRECHECK(OBD_FAIL_LLOG_PROCESS_TIMEOUT) &&
 572                                 cfs_fail_val == (unsigned int)
 573                                         (loghandle->lgh_id.lgl_oi.oi.oi_id &
 574                                          0xFFFFFFFF)) {
 575                                 OBD_RACE(OBD_FAIL_LLOG_PROCESS_TIMEOUT);
 576                         }
 577
 578                         /* the bitmap could be changed during processing
 579                          * records from the chunk. For wrapped catalog
 580                          * it means we can read deleted record and try to
 581                          * process it. Check this case and reread the chunk.
 582                          * It is safe to process to lh_last_idx, including
 583                          * lh_last_idx if it was synced. We can not do <=
 584                          * comparison, cause for wrapped catalog lgh_last_idx
 585                          * could be less than index. So we detect last index
 586                          * for processing as index == lh_last_idx+1. But when
 587                          * catalog is wrapped and full lgh_last_idx=llh_cat_idx,
 588                          * the first processing index is llh_cat_idx+1.
 589                          */
 590
 591                         if ((index == lh_last_idx && synced_idx != index) ||
 592                             (index == (lh_last_idx + 1) &&
 593                              !(index == (llh->llh_cat_idx + 1) &&
 594                                (llh->llh_flags & LLOG_F_IS_CAT))) ||
 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 (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 651                                 struct llog_cookie *lgc;
 652                                 __u64   tmp_off;
 653                                 int     tmp_idx;
 654
 655                                 CDEBUG(D_OTHER, "index: %d, lh_last_idx: %d "
 656                                        "synced_idx: %d lgh_last_idx: %d\n",
 657                                        index, lh_last_idx, synced_idx,
 658                                        loghandle->lgh_last_idx);
 659
 660                                 if (lti != NULL) {
 661                                         lgc = &lti->lgi_cookie;
 662                                         /* store lu_env for recursive calls */
 663                                         tmp_off = lgc->lgc_offset;
 664                                         tmp_idx = lgc->lgc_index;
 665
 666                                         lgc->lgc_offset = (char *)rec -
 667                                                 (char *)buf + chunk_offset;
 668                                         lgc->lgc_index = rec->lrh_index;
 669                                 }
 670                                 /* using lu_env for passing record offset to
 671                                  * llog_write through various callbacks */
 672                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 673                                                  lpi->lpi_cbdata);
 674                                 last_called_index = index;
 675
 676                                 if (lti != NULL) {
 677                                         lgc->lgc_offset = tmp_off;
 678                                         lgc->lgc_index = tmp_idx;
 679                                 }
 680
 681                                 if (rc == LLOG_PROC_BREAK) {
 682                                         GOTO(out, rc);
 683                                 } else if (rc == LLOG_DEL_RECORD) {
 684                                         rc = llog_cancel_rec(lpi->lpi_env,
 685                                                              loghandle,
 686                                                              rec->lrh_index);
 687                                 }
 688                                 if (rc)
 689                                         GOTO(out, rc);
 690                                 /* some stupid callbacks directly cancel records
 691                                  * and delete llog. Check it and stop
 692                                  * processing. */
 693                                 if (loghandle->lgh_hdr == NULL ||
 694                                     loghandle->lgh_hdr->llh_count == 1)
 695                                         GOTO(out, rc = 0);
 696                         }
 697                         /* exit if the last index is reached */
 698                         if (index >= last_index)
 699                                 GOTO(out, rc = 0);
 700                         ++index;
 701                 }
 702         }
 703
 704 out:
 705         if (cd != NULL)
 706                 cd->lpcd_last_idx = last_called_index;
 707
 708         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 709                 if (dt_object_remote(loghandle->lgh_obj)) {
 710                         /* If it is remote object, then -EIO might means
 711                          * disconnection or eviction, let's return -EAGAIN,
 712                          * so for update recovery log processing, it will
 713                          * retry until the umount or abort recovery, see
 714                          * lod_sub_recovery_thread() */
 715                         CERROR("%s retry remote llog process\n",
 716                                loghandle->lgh_ctxt->loc_obd->obd_name);
 717                         rc = -EAGAIN;
 718                 } else {
 719                         /* something bad happened to the processing of a local
 720                          * llog file, probably I/O error or the log got
 721                          * corrupted to be able to finally release the log we
 722                          * discard any remaining bits in the header */
 723                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 724                                " %s index %d count %d\n",
 725                                loghandle->lgh_ctxt->loc_obd->obd_name,
 726                                POSTID(&loghandle->lgh_id.lgl_oi),
 727                                loghandle->lgh_id.lgl_ogen,
 728                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 729                                 "plain"), index, llh->llh_count);
 730
 731                         while (index <= last_index) {
 732                                 if (ext2_test_bit(index,
 733                                                   LLOG_HDR_BITMAP(llh)) != 0)
 734                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 735                                                         index);
 736                                 index++;
 737                         }
 738                         rc = 0;
 739                 }
 740         }
 741
 742         OBD_FREE_LARGE(buf, chunk_size);
 743         lpi->lpi_rc = rc;
 744         return 0;
 745 }
 746
 747 static int llog_process_thread_daemonize(void *arg)
 748 {
 749         struct llog_process_info        *lpi = arg;
 750         struct lu_env                    env;
 751         int                              rc;
 752         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 753
 754         task_lock(lpi->lpi_reftask);
 755         new_ns = lpi->lpi_reftask->nsproxy;
 756         if (curr_ns != new_ns) {
 757                 get_nsproxy(new_ns);
 758
 759                 current->nsproxy = new_ns;
 760                 /* XXX: we should call put_nsproxy() instead of
 761                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 762                  * used outside of the kernel itself, because it calls
 763                  * free_nsproxy() which is not exported by the kernel
 764                  * (defined in kernel/nsproxy.c) */
 765                 if (curr_ns)
 766                         atomic_dec(&curr_ns->count);
 767         }
 768         task_unlock(lpi->lpi_reftask);
 769
 770         unshare_fs_struct();
 771
 772         /* client env has no keys, tags is just 0 */
 773         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 774         if (rc)
 775                 goto out;
 776         lpi->lpi_env = &env;
 777
 778         rc = llog_process_thread(arg);
 779
 780         lu_env_fini(&env);
 781 out:
 782         complete(&lpi->lpi_completion);
 783         return rc;
 784 }
 785
 786 int llog_process_or_fork(const struct lu_env *env,
 787                          struct llog_handle *loghandle,
 788                          llog_cb_t cb, void *data, void *catdata, bool fork)
 789 {
 790         struct llog_process_info *lpi;
 791         struct llog_process_data *d = data;
 792         struct llog_process_cat_data *cd = catdata;
 793         int                      rc;
 794
 795         ENTRY;
 796
 797         OBD_ALLOC_PTR(lpi);
 798         if (lpi == NULL) {
 799                 CERROR("cannot alloc pointer\n");
 800                 RETURN(-ENOMEM);
 801         }
 802         lpi->lpi_loghandle = loghandle;
 803         lpi->lpi_cb        = cb;
 804         lpi->lpi_cbdata    = data;
 805         lpi->lpi_catdata   = catdata;
 806
 807         CDEBUG(D_OTHER, "Processing "DFID" flags 0x%03x startcat %d startidx %d first_idx %d last_idx %d\n",
 808                PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 809                loghandle->lgh_hdr->llh_flags, d ? d->lpd_startcat : -1,
 810                d ? d->lpd_startidx : -1, cd ? cd->lpcd_first_idx : -1,
 811                cd ? cd->lpcd_last_idx : -1);
 812         if (fork) {
 813                 struct task_struct *task;
 814
 815                 /* The new thread can't use parent env,
 816                  * init the new one in llog_process_thread_daemonize. */
 817                 lpi->lpi_env = NULL;
 818                 init_completion(&lpi->lpi_completion);
 819                 /* take reference to current, so that
 820                  * llog_process_thread_daemonize() can use it to switch to
 821                  * namespace associated with current  */
 822                 lpi->lpi_reftask = current;
 823                 task = kthread_run(llog_process_thread_daemonize, lpi,
 824                                    "llog_process_thread");
 825                 if (IS_ERR(task)) {
 826                         rc = PTR_ERR(task);
 827                         CERROR("%s: cannot start thread: rc = %d\n",
 828                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 829                         GOTO(out_lpi, rc);
 830                 }
 831                 wait_for_completion(&lpi->lpi_completion);
 832         } else {
 833                 lpi->lpi_env = env;
 834                 llog_process_thread(lpi);
 835         }
 836         rc = lpi->lpi_rc;
 837
 838 out_lpi:
 839         OBD_FREE_PTR(lpi);
 840         RETURN(rc);
 841 }
 842 EXPORT_SYMBOL(llog_process_or_fork);
 843
 844 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 845                  llog_cb_t cb, void *data, void *catdata)
 846 {
 847         int rc;
 848         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 849         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 850 }
 851 EXPORT_SYMBOL(llog_process);
 852
 853 static inline const struct cred *llog_raise_resource(void)
 854 {
 855         struct cred *cred = NULL;
 856
 857         if (cap_raised(current_cap(), CAP_SYS_RESOURCE))
 858                 return cred;
 859
 860         cred = prepare_creds();
 861         if (!cred)
 862                 return cred;
 863
 864         cap_raise(cred->cap_effective, CAP_SYS_RESOURCE);
 865         return override_creds(cred);
 866 }
 867
 868 static inline void llog_restore_resource(const struct cred *old_cred)
 869 {
 870         if (old_cred)
 871                 revert_creds(old_cred);
 872 }
 873
 874 int llog_reverse_process(const struct lu_env *env,
 875                          struct llog_handle *loghandle, llog_cb_t cb,
 876                          void *data, void *catdata)
 877 {
 878         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 879         struct llog_process_cat_data *cd = catdata;
 880         void *buf;
 881         int rc = 0, first_index = 1, index, idx;
 882         __u32   chunk_size = llh->llh_hdr.lrh_len;
 883         ENTRY;
 884
 885         OBD_ALLOC_LARGE(buf, chunk_size);
 886         if (buf == NULL)
 887                 RETURN(-ENOMEM);
 888
 889         if (cd != NULL)
 890                 first_index = cd->lpcd_first_idx + 1;
 891         if (cd != NULL && cd->lpcd_last_idx)
 892                 index = cd->lpcd_last_idx;
 893         else
 894                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 895
 896         while (rc == 0) {
 897                 struct llog_rec_hdr *rec;
 898                 struct llog_rec_tail *tail;
 899
 900                 /* skip records not set in bitmap */
 901                 while (index >= first_index &&
 902                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 903                         --index;
 904
 905                 LASSERT(index >= first_index - 1);
 906                 if (index == first_index - 1)
 907                         break;
 908
 909                 /* get the buf with our target record; avoid old garbage */
 910                 memset(buf, 0, chunk_size);
 911                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 912                 if (rc)
 913                         GOTO(out, rc);
 914
 915                 rec = buf;
 916                 idx = rec->lrh_index;
 917                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 918                 while (idx < index) {
 919                         rec = (void *)rec + rec->lrh_len;
 920                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 921                                 lustre_swab_llog_rec(rec);
 922                         idx ++;
 923                 }
 924                 LASSERT(idx == index);
 925                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 926
 927                 /* process records in buffer, starting where we found one */
 928                 while ((void *)tail > buf) {
 929                         if (tail->lrt_index == 0)
 930                                 GOTO(out, rc = 0); /* no more records */
 931
 932                         /* if set, process the callback on this record */
 933                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 934                                 rec = (void *)tail - tail->lrt_len +
 935                                       sizeof(*tail);
 936
 937                                 rc = cb(env, loghandle, rec, data);
 938                                 if (rc == LLOG_PROC_BREAK) {
 939                                         GOTO(out, rc);
 940                                 } else if (rc == LLOG_DEL_RECORD) {
 941                                         rc = llog_cancel_rec(env, loghandle,
 942                                                              tail->lrt_index);
 943                                 }
 944                                 if (rc)
 945                                         GOTO(out, rc);
 946                         }
 947
 948                         /* previous record, still in buffer? */
 949                         --index;
 950                         if (index < first_index)
 951                                 GOTO(out, rc = 0);
 952                         tail = (void *)tail - tail->lrt_len;
 953                 }
 954         }
 955
 956 out:
 957         if (buf != NULL)
 958                 OBD_FREE_LARGE(buf, chunk_size);
 959         RETURN(rc);
 960 }
 961 EXPORT_SYMBOL(llog_reverse_process);
 962
 963 /**
 964  * new llog API
 965  *
 966  * API functions:
 967  *      llog_open - open llog, may not exist
 968  *      llog_exist - check if llog exists
 969  *      llog_close - close opened llog, pair for open, frees llog_handle
 970  *      llog_declare_create - declare llog creation
 971  *      llog_create - create new llog on disk, need transaction handle
 972  *      llog_declare_write_rec - declaration of llog write
 973  *      llog_write_rec - write llog record on disk, need transaction handle
 974  *      llog_declare_add - declare llog catalog record addition
 975  *      llog_add - add llog record in catalog, need transaction handle
 976  */
 977 int llog_exist(struct llog_handle *loghandle)
 978 {
 979         struct llog_operations  *lop;
 980         int                      rc;
 981
 982         ENTRY;
 983
 984         rc = llog_handle2ops(loghandle, &lop);
 985         if (rc)
 986                 RETURN(rc);
 987         if (lop->lop_exist == NULL)
 988                 RETURN(-EOPNOTSUPP);
 989
 990         rc = lop->lop_exist(loghandle);
 991         RETURN(rc);
 992 }
 993 EXPORT_SYMBOL(llog_exist);
 994
 995 int llog_declare_create(const struct lu_env *env,
 996                         struct llog_handle *loghandle, struct thandle *th)
 997 {
 998         const struct cred *old_cred;
 999         struct llog_operations  *lop;
1000         int rc;
1001
1002         ENTRY;
1003
1004         rc = llog_handle2ops(loghandle, &lop);
1005         if (rc)
1006                 RETURN(rc);
1007         if (lop->lop_declare_create == NULL)
1008                 RETURN(-EOPNOTSUPP);
1009
1010         old_cred = llog_raise_resource();
1011         rc = lop->lop_declare_create(env, loghandle, th);
1012         llog_restore_resource(old_cred);
1013         RETURN(rc);
1014 }
1015
1016 int llog_create(const struct lu_env *env, struct llog_handle *handle,
1017                 struct thandle *th)
1018 {
1019         const struct cred *old_cred;
1020         struct llog_operations  *lop;
1021         int rc;
1022
1023         ENTRY;
1024
1025         rc = llog_handle2ops(handle, &lop);
1026         if (rc)
1027                 RETURN(rc);
1028         if (lop->lop_create == NULL)
1029                 RETURN(-EOPNOTSUPP);
1030
1031         old_cred = llog_raise_resource();
1032         rc = lop->lop_create(env, handle, th);
1033         llog_restore_resource(old_cred);
1034         RETURN(rc);
1035 }
1036
1037 int llog_declare_write_rec(const struct lu_env *env,
1038                            struct llog_handle *handle,
1039                            struct llog_rec_hdr *rec, int idx,
1040                            struct thandle *th)
1041 {
1042         const struct cred *old_cred;
1043         struct llog_operations  *lop;
1044         int rc;
1045
1046         ENTRY;
1047
1048         rc = llog_handle2ops(handle, &lop);
1049         if (rc)
1050                 RETURN(rc);
1051         LASSERT(lop);
1052         if (lop->lop_declare_write_rec == NULL)
1053                 RETURN(-EOPNOTSUPP);
1054
1055         old_cred = llog_raise_resource();
1056         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
1057         llog_restore_resource(old_cred);
1058         RETURN(rc);
1059 }
1060
1061 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
1062                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1063                    int idx, struct thandle *th)
1064 {
1065         const struct cred *old_cred;
1066         struct llog_operations  *lop;
1067         int rc, buflen;
1068
1069         ENTRY;
1070
1071         /* API sanity checks */
1072         if (handle == NULL) {
1073                 CERROR("loghandle is missed\n");
1074                 RETURN(-EPROTO);
1075         } else if (handle->lgh_obj == NULL) {
1076                 CERROR("loghandle %p with NULL object\n",
1077                         handle);
1078                 RETURN(-EPROTO);
1079         } else if (th == NULL) {
1080                 CERROR("%s: missed transaction handle\n",
1081                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name);
1082                 RETURN(-EPROTO);
1083         } else if (handle->lgh_hdr == NULL) {
1084                 CERROR("%s: loghandle %p with no header\n",
1085                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name,
1086                         handle);
1087                 RETURN(-EPROTO);
1088         }
1089
1090         rc = llog_handle2ops(handle, &lop);
1091         if (rc)
1092                 RETURN(rc);
1093
1094         if (lop->lop_write_rec == NULL)
1095                 RETURN(-EOPNOTSUPP);
1096
1097         buflen = rec->lrh_len;
1098         LASSERT(cfs_size_round(buflen) == buflen);
1099
1100         old_cred = llog_raise_resource();
1101         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1102         llog_restore_resource(old_cred);
1103         RETURN(rc);
1104 }
1105
1106 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1107              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1108              struct thandle *th)
1109 {
1110         const struct cred *old_cred;
1111         int rc;
1112
1113         ENTRY;
1114
1115         if (lgh->lgh_logops->lop_add == NULL)
1116                 RETURN(-EOPNOTSUPP);
1117
1118         old_cred = llog_raise_resource();
1119         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1120         llog_restore_resource(old_cred);
1121         RETURN(rc);
1122 }
1123 EXPORT_SYMBOL(llog_add);
1124
1125 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1126                      struct llog_rec_hdr *rec, struct thandle *th)
1127 {
1128         const struct cred *old_cred;
1129         int rc;
1130
1131         ENTRY;
1132
1133         if (lgh->lgh_logops->lop_declare_add == NULL)
1134                 RETURN(-EOPNOTSUPP);
1135
1136         old_cred = llog_raise_resource();
1137         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1138         llog_restore_resource(old_cred);
1139         RETURN(rc);
1140 }
1141 EXPORT_SYMBOL(llog_declare_add);
1142
1143 /**
1144  * Helper function to open llog or create it if doesn't exist.
1145  * It hides all transaction handling from caller.
1146  */
1147 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1148                      struct llog_handle **res, struct llog_logid *logid,
1149                      char *name)
1150 {
1151         struct dt_device        *d;
1152         struct thandle          *th;
1153         int                      rc;
1154
1155         ENTRY;
1156
1157         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1158         if (rc)
1159                 RETURN(rc);
1160
1161         if (llog_exist(*res))
1162                 RETURN(0);
1163
1164         LASSERT((*res)->lgh_obj != NULL);
1165
1166         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1167
1168         if (unlikely(unlikely(d->dd_rdonly)))
1169                 RETURN(-EROFS);
1170
1171         th = dt_trans_create(env, d);
1172         if (IS_ERR(th))
1173                 GOTO(out, rc = PTR_ERR(th));
1174
1175         /* Create update llog object synchronously, which
1176          * happens during inialization process see
1177          * lod_sub_prep_llog(), to make sure the update
1178          * llog object is created before corss-MDT writing
1179          * updates into the llog object */
1180         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1181                 th->th_sync = 1;
1182
1183         th->th_wait_submit = 1;
1184         rc = llog_declare_create(env, *res, th);
1185         if (rc == 0) {
1186                 rc = dt_trans_start_local(env, d, th);
1187                 if (rc == 0)
1188                         rc = llog_create(env, *res, th);
1189         }
1190         dt_trans_stop(env, d, th);
1191 out:
1192         if (rc)
1193                 llog_close(env, *res);
1194         RETURN(rc);
1195 }
1196 EXPORT_SYMBOL(llog_open_create);
1197
1198 /**
1199  * Helper function to delete existent llog.
1200  */
1201 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1202                struct llog_logid *logid, char *name)
1203 {
1204         struct llog_handle      *handle;
1205         int                      rc = 0, rc2;
1206
1207         ENTRY;
1208
1209         /* nothing to erase */
1210         if (name == NULL && logid == NULL)
1211                 RETURN(0);
1212
1213         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1214         if (rc < 0)
1215                 RETURN(rc);
1216
1217         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1218         if (rc == 0)
1219                 rc = llog_destroy(env, handle);
1220
1221         rc2 = llog_close(env, handle);
1222         if (rc == 0)
1223                 rc = rc2;
1224         RETURN(rc);
1225 }
1226 EXPORT_SYMBOL(llog_erase);
1227
1228 /*
1229  * Helper function for write record in llog.
1230  * It hides all transaction handling from caller.
1231  * Valid only with local llog.
1232  */
1233 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1234                struct llog_rec_hdr *rec, int idx)
1235 {
1236         struct dt_device        *dt;
1237         struct thandle          *th;
1238         bool                    need_cookie;
1239         int                     rc;
1240
1241         ENTRY;
1242
1243         LASSERT(loghandle);
1244         LASSERT(loghandle->lgh_ctxt);
1245         LASSERT(loghandle->lgh_obj != NULL);
1246
1247         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1248
1249         if (unlikely(unlikely(dt->dd_rdonly)))
1250                 RETURN(-EROFS);
1251
1252         th = dt_trans_create(env, dt);
1253         if (IS_ERR(th))
1254                 RETURN(PTR_ERR(th));
1255
1256         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1257         if (rc)
1258                 GOTO(out_trans, rc);
1259
1260         th->th_wait_submit = 1;
1261         rc = dt_trans_start_local(env, dt, th);
1262         if (rc)
1263                 GOTO(out_trans, rc);
1264
1265         need_cookie = !(idx == LLOG_HEADER_IDX || idx == LLOG_NEXT_IDX);
1266
1267         down_write(&loghandle->lgh_lock);
1268         if (need_cookie) {
1269                 struct llog_thread_info *lti = llog_info(env);
1270
1271                 /* cookie comes from llog_process_thread */
1272                 rc = llog_write_rec(env, loghandle, rec, &lti->lgi_cookie,
1273                                     rec->lrh_index, th);
1274                 /* upper layer didn`t pass cookie so change rc */
1275                 rc = (rc == 1 ? 0 : rc);
1276         } else {
1277                 rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1278         }
1279
1280         up_write(&loghandle->lgh_lock);
1281 out_trans:
1282         dt_trans_stop(env, dt, th);
1283         RETURN(rc);
1284 }
1285 EXPORT_SYMBOL(llog_write);
1286
1287 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1288               struct llog_handle **lgh, struct llog_logid *logid,
1289               char *name, enum llog_open_param open_param)
1290 {
1291         const struct cred *old_cred;
1292         int      rc;
1293
1294         ENTRY;
1295
1296         LASSERT(ctxt);
1297         LASSERT(ctxt->loc_logops);
1298
1299         if (ctxt->loc_logops->lop_open == NULL) {
1300                 *lgh = NULL;
1301                 RETURN(-EOPNOTSUPP);
1302         }
1303
1304         *lgh = llog_alloc_handle();
1305         if (*lgh == NULL)
1306                 RETURN(-ENOMEM);
1307         (*lgh)->lgh_ctxt = ctxt;
1308         (*lgh)->lgh_logops = ctxt->loc_logops;
1309
1310         old_cred = llog_raise_resource();
1311         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1312         llog_restore_resource(old_cred);
1313         if (rc) {
1314                 llog_free_handle(*lgh);
1315                 *lgh = NULL;
1316         }
1317         RETURN(rc);
1318 }
1319 EXPORT_SYMBOL(llog_open);
1320
1321 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1322 {
1323         return llog_handle_put(env, loghandle);
1324 }
1325 EXPORT_SYMBOL(llog_close);
1326
1327 /**
1328  * Helper function to get the llog size in records. It is used by MGS
1329  * mostly to check that config llog exists and contains data.
1330  *
1331  * \param[in] env       execution environment
1332  * \param[in] ctxt      llog context
1333  * \param[in] name      llog name
1334  *
1335  * \retval              true if there are records in llog besides a header
1336  * \retval              false on error or llog without records
1337  */
1338 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1339                   char *name)
1340 {
1341         struct llog_handle      *llh;
1342         int                      rc = 0;
1343
1344         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1345         if (rc < 0) {
1346                 if (likely(rc == -ENOENT))
1347                         rc = 0;
1348                 GOTO(out, rc);
1349         }
1350
1351         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1352         if (rc)
1353                 GOTO(out_close, rc);
1354         rc = llog_get_size(llh);
1355
1356 out_close:
1357         llog_close(env, llh);
1358 out:
1359         /* The header is record 1, the llog is still considered as empty
1360          * if there is only header */
1361         return (rc <= 1);
1362 }
1363 EXPORT_SYMBOL(llog_is_empty);
1364
1365 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1366                       struct llog_rec_hdr *rec, void *data)
1367 {
1368         struct llog_handle      *copy_llh = data;
1369
1370         /* Append all records */
1371         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1372 }
1373
1374 /* backup plain llog */
1375 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1376                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1377                 char *name, char *backup)
1378 {
1379         struct llog_handle      *llh, *bllh;
1380         int                      rc;
1381
1382         ENTRY;
1383
1384         /* open original log */
1385         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1386         if (rc < 0) {
1387                 /* the -ENOENT case is also reported to the caller
1388                  * but silently so it should handle that if needed.
1389                  */
1390                 if (rc != -ENOENT)
1391                         CERROR("%s: failed to open log %s: rc = %d\n",
1392                                obd->obd_name, name, rc);
1393                 RETURN(rc);
1394         }
1395
1396         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1397         if (rc)
1398                 GOTO(out_close, rc);
1399
1400         /* Make sure there's no old backup log */
1401         rc = llog_erase(env, bctxt, NULL, backup);
1402         if (rc < 0 && rc != -ENOENT)
1403                 GOTO(out_close, rc);
1404
1405         /* open backup log */
1406         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1407         if (rc) {
1408                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1409                        obd->obd_name, backup, rc);
1410                 GOTO(out_close, rc);
1411         }
1412
1413         /* check that backup llog is not the same object as original one */
1414         if (llh->lgh_obj == bllh->lgh_obj) {
1415                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1416                        obd->obd_name, name, backup, llh->lgh_obj,
1417                        bllh->lgh_obj);
1418                 GOTO(out_backup, rc = -EEXIST);
1419         }
1420
1421         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1422         if (rc)
1423                 GOTO(out_backup, rc);
1424
1425         /* Copy log record by record */
1426         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1427                                   NULL, false);
1428         if (rc)
1429                 CERROR("%s: failed to backup log %s: rc = %d\n",
1430                        obd->obd_name, name, rc);
1431 out_backup:
1432         llog_close(env, bllh);
1433 out_close:
1434         llog_close(env, llh);
1435         RETURN(rc);
1436 }
1437 EXPORT_SYMBOL(llog_backup);
1438
1439 /* Get size of llog */
1440 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1441 {
1442         int rc;
1443         struct lu_attr la;
1444
1445         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1446         if (rc) {
1447                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1448                        llh->lgh_ctxt->loc_obd->obd_name,
1449                        PFID(&llh->lgh_id.lgl_oi.oi_fid), rc);
1450                 return 0;
1451         }
1452
1453         return la.la_size;
1454 }
1455 EXPORT_SYMBOL(llog_size);
1456