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                 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                        !test_bit_le(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.The
 589                          * exception is !(lgh_last_idx == llh_cat_idx &&
 590                          * index == llh_cat_idx + 1), and after simplification
 591                          * it turns to
 592                          * lh_last_idx != LLOG_HDR_TAIL(llh)->lrt_index
 593                          * This exception is working for catalog only.
 594                          */
 595
 596                         if ((index == lh_last_idx && synced_idx != index) ||
 597                             (index == (lh_last_idx + 1) &&
 598                              lh_last_idx != LLOG_HDR_TAIL(llh)->lrt_index) ||
 599                             (rec->lrh_index == 0 && !repeated)) {
 600
 601                                 /* save offset inside buffer for the re-read */
 602                                 buf_offset = (char *)rec - (char *)buf;
 603                                 cur_offset = chunk_offset;
 604                                 repeated = true;
 605                                 /* We need to be sure lgh_last_idx
 606                                  * record was saved to disk
 607                                  */
 608                                 down_read(&loghandle->lgh_last_sem);
 609                                 synced_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 610                                 up_read(&loghandle->lgh_last_sem);
 611                                 CDEBUG(D_OTHER, "synced_idx: %d\n", synced_idx);
 612                                 goto repeat;
 613
 614                         }
 615
 616                         repeated = false;
 617
 618                         if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
 619                                 CWARN("%s: invalid length %d in llog "DFID
 620                                       "record for index %d/%d\n",
 621                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 622                                        rec->lrh_len,
 623                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 624                                        rec->lrh_index, index);
 625
 626                                 GOTO(out, rc = -EINVAL);
 627                         }
 628
 629                         if (rec->lrh_index < index) {
 630                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 631                                        rec->lrh_index);
 632                                 continue;
 633                         }
 634
 635                         if (rec->lrh_index != index) {
 636                                 CERROR("%s: "DFID" Invalid record: index %u"
 637                                        " but expected %u\n",
 638                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 639                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 640                                        rec->lrh_index, index);
 641                                 GOTO(out, rc = -ERANGE);
 642                         }
 643
 644                         CDEBUG(D_OTHER,
 645                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 646                                rec->lrh_index, rec->lrh_len,
 647                                (int)(buf + chunk_size - (char *)rec));
 648
 649                         /* lgh_cur_offset is used only at llog_test_3 */
 650                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 651                                                     chunk_offset;
 652
 653                         /* if set, process the callback on this record */
 654                         if (test_bit_le(index, LLOG_HDR_BITMAP(llh))) {
 655                                 struct llog_cookie *lgc;
 656                                 __u64   tmp_off;
 657                                 int     tmp_idx;
 658
 659                                 CDEBUG((llh->llh_flags & LLOG_F_IS_CAT ?
 660                                         D_HA : D_OTHER),
 661                                        "index: %d, lh_last_idx: %d "
 662                                        "synced_idx: %d lgh_last_idx: %d\n",
 663                                        index, lh_last_idx, synced_idx,
 664                                        loghandle->lgh_last_idx);
 665
 666                                 if (lti != NULL) {
 667                                         lgc = &lti->lgi_cookie;
 668                                         /* store lu_env for recursive calls */
 669                                         tmp_off = lgc->lgc_offset;
 670                                         tmp_idx = lgc->lgc_index;
 671
 672                                         lgc->lgc_offset = (char *)rec -
 673                                                 (char *)buf + chunk_offset;
 674                                         lgc->lgc_index = rec->lrh_index;
 675                                 }
 676                                 /* using lu_env for passing record offset to
 677                                  * llog_write through various callbacks */
 678                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 679                                                  lpi->lpi_cbdata);
 680                                 last_called_index = index;
 681
 682                                 if (lti != NULL) {
 683                                         lgc->lgc_offset = tmp_off;
 684                                         lgc->lgc_index = tmp_idx;
 685                                 }
 686
 687                                 if (rc == LLOG_PROC_BREAK) {
 688                                         GOTO(out, rc);
 689                                 } else if (rc == LLOG_DEL_RECORD) {
 690                                         rc = llog_cancel_rec(lpi->lpi_env,
 691                                                              loghandle,
 692                                                              rec->lrh_index);
 693                                 }
 694                                 if (rc)
 695                                         GOTO(out, rc);
 696                                 /* some stupid callbacks directly cancel records
 697                                  * and delete llog. Check it and stop
 698                                  * processing. */
 699                                 if (loghandle->lgh_hdr == NULL ||
 700                                     loghandle->lgh_hdr->llh_count == 1)
 701                                         GOTO(out, rc = 0);
 702                         }
 703                         /* exit if the last index is reached */
 704                         if (index >= last_index)
 705                                 GOTO(out, rc = 0);
 706                         ++index;
 707                 }
 708         }
 709
 710 out:
 711         CDEBUG(D_HA, "stop processing %s "DOSTID":%x index %d count %d\n",
 712                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" : "plain"),
 713                POSTID(&loghandle->lgh_id.lgl_oi), loghandle->lgh_id.lgl_ogen,
 714                index, llh->llh_count);
 715
 716         if (cd != NULL)
 717                 cd->lpcd_last_idx = last_called_index;
 718
 719         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 720                 if (dt_object_remote(loghandle->lgh_obj)) {
 721                         /* If it is remote object, then -EIO might means
 722                          * disconnection or eviction, let's return -EAGAIN,
 723                          * so for update recovery log processing, it will
 724                          * retry until the umount or abort recovery, see
 725                          * lod_sub_recovery_thread() */
 726                         CERROR("%s retry remote llog process\n",
 727                                loghandle->lgh_ctxt->loc_obd->obd_name);
 728                         rc = -EAGAIN;
 729                 } else {
 730                         /* something bad happened to the processing of a local
 731                          * llog file, probably I/O error or the log got
 732                          * corrupted to be able to finally release the log we
 733                          * discard any remaining bits in the header */
 734                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 735                                " %s index %d count %d\n",
 736                                loghandle->lgh_ctxt->loc_obd->obd_name,
 737                                POSTID(&loghandle->lgh_id.lgl_oi),
 738                                loghandle->lgh_id.lgl_ogen,
 739                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 740                                 "plain"), index, llh->llh_count);
 741
 742                         while (index <= last_index) {
 743                                 if (test_bit_le(index,
 744                                                   LLOG_HDR_BITMAP(llh)) != 0)
 745                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 746                                                         index);
 747                                 index++;
 748                         }
 749                         rc = 0;
 750                 }
 751         }
 752
 753         OBD_FREE_LARGE(buf, chunk_size);
 754         lpi->lpi_rc = rc;
 755         return 0;
 756 }
 757
 758 static int llog_process_thread_daemonize(void *arg)
 759 {
 760         struct llog_process_info        *lpi = arg;
 761         struct lu_env                    env;
 762         int                              rc;
 763         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 764
 765         task_lock(lpi->lpi_reftask);
 766         new_ns = lpi->lpi_reftask->nsproxy;
 767         if (curr_ns != new_ns) {
 768                 get_nsproxy(new_ns);
 769
 770                 current->nsproxy = new_ns;
 771                 /* XXX: we should call put_nsproxy() instead of
 772                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 773                  * used outside of the kernel itself, because it calls
 774                  * free_nsproxy() which is not exported by the kernel
 775                  * (defined in kernel/nsproxy.c) */
 776                 if (curr_ns)
 777                         atomic_dec(&curr_ns->count);
 778         }
 779         task_unlock(lpi->lpi_reftask);
 780
 781         unshare_fs_struct();
 782
 783         /* client env has no keys, tags is just 0 */
 784         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 785         if (rc)
 786                 goto out;
 787         lpi->lpi_env = &env;
 788
 789         rc = llog_process_thread(arg);
 790
 791         lu_env_fini(&env);
 792 out:
 793         complete(&lpi->lpi_completion);
 794         return rc;
 795 }
 796
 797 int llog_process_or_fork(const struct lu_env *env,
 798                          struct llog_handle *loghandle,
 799                          llog_cb_t cb, void *data, void *catdata, bool fork)
 800 {
 801         struct llog_process_info *lpi;
 802         struct llog_process_data *d = data;
 803         struct llog_process_cat_data *cd = catdata;
 804         int                      rc;
 805
 806         ENTRY;
 807
 808         OBD_ALLOC_PTR(lpi);
 809         if (lpi == NULL) {
 810                 CERROR("cannot alloc pointer\n");
 811                 RETURN(-ENOMEM);
 812         }
 813         lpi->lpi_loghandle = loghandle;
 814         lpi->lpi_cb        = cb;
 815         lpi->lpi_cbdata    = data;
 816         lpi->lpi_catdata   = catdata;
 817
 818         CDEBUG(D_OTHER, "Processing "DFID" flags 0x%03x startcat %d startidx %d first_idx %d last_idx %d\n",
 819                PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 820                loghandle->lgh_hdr->llh_flags, d ? d->lpd_startcat : -1,
 821                d ? d->lpd_startidx : -1, cd ? cd->lpcd_first_idx : -1,
 822                cd ? cd->lpcd_last_idx : -1);
 823         if (fork) {
 824                 struct task_struct *task;
 825
 826                 /* The new thread can't use parent env,
 827                  * init the new one in llog_process_thread_daemonize. */
 828                 lpi->lpi_env = NULL;
 829                 init_completion(&lpi->lpi_completion);
 830                 /* take reference to current, so that
 831                  * llog_process_thread_daemonize() can use it to switch to
 832                  * namespace associated with current  */
 833                 lpi->lpi_reftask = current;
 834                 task = kthread_run(llog_process_thread_daemonize, lpi,
 835                                    "llog_process_thread");
 836                 if (IS_ERR(task)) {
 837                         rc = PTR_ERR(task);
 838                         CERROR("%s: cannot start thread: rc = %d\n",
 839                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 840                         GOTO(out_lpi, rc);
 841                 }
 842                 wait_for_completion(&lpi->lpi_completion);
 843         } else {
 844                 lpi->lpi_env = env;
 845                 llog_process_thread(lpi);
 846         }
 847         rc = lpi->lpi_rc;
 848
 849 out_lpi:
 850         OBD_FREE_PTR(lpi);
 851         RETURN(rc);
 852 }
 853 EXPORT_SYMBOL(llog_process_or_fork);
 854
 855 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 856                  llog_cb_t cb, void *data, void *catdata)
 857 {
 858         int rc;
 859         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 860         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 861 }
 862 EXPORT_SYMBOL(llog_process);
 863
 864 static inline const struct cred *llog_raise_resource(void)
 865 {
 866         struct cred *cred = NULL;
 867
 868         if (cap_raised(current_cap(), CAP_SYS_RESOURCE))
 869                 return cred;
 870
 871         cred = prepare_creds();
 872         if (!cred)
 873                 return cred;
 874
 875         cap_raise(cred->cap_effective, CAP_SYS_RESOURCE);
 876         return override_creds(cred);
 877 }
 878
 879 static inline void llog_restore_resource(const struct cred *old_cred)
 880 {
 881         if (old_cred)
 882                 revert_creds(old_cred);
 883 }
 884
 885 int llog_reverse_process(const struct lu_env *env,
 886                          struct llog_handle *loghandle, llog_cb_t cb,
 887                          void *data, void *catdata)
 888 {
 889         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 890         struct llog_process_cat_data *cd = catdata;
 891         void *buf;
 892         int rc = 0, first_index = 1, index, idx;
 893         __u32   chunk_size = llh->llh_hdr.lrh_len;
 894         ENTRY;
 895
 896         OBD_ALLOC_LARGE(buf, chunk_size);
 897         if (buf == NULL)
 898                 RETURN(-ENOMEM);
 899
 900         if (cd != NULL)
 901                 first_index = cd->lpcd_first_idx + 1;
 902         if (cd != NULL && cd->lpcd_last_idx)
 903                 index = cd->lpcd_last_idx;
 904         else
 905                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 906
 907         while (rc == 0) {
 908                 struct llog_rec_hdr *rec;
 909                 struct llog_rec_tail *tail;
 910
 911                 /* skip records not set in bitmap */
 912                 while (index >= first_index &&
 913                        !test_bit_le(index, LLOG_HDR_BITMAP(llh)))
 914                         --index;
 915
 916                 LASSERT(index >= first_index - 1);
 917                 if (index == first_index - 1)
 918                         break;
 919
 920                 /* get the buf with our target record; avoid old garbage */
 921                 memset(buf, 0, chunk_size);
 922                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 923                 if (rc)
 924                         GOTO(out, rc);
 925
 926                 rec = buf;
 927                 idx = rec->lrh_index;
 928                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 929                 while (idx < index) {
 930                         rec = (void *)rec + rec->lrh_len;
 931                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 932                                 lustre_swab_llog_rec(rec);
 933                         idx ++;
 934                 }
 935                 LASSERT(idx == index);
 936                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 937
 938                 /* process records in buffer, starting where we found one */
 939                 while ((void *)tail > buf) {
 940                         if (tail->lrt_index == 0)
 941                                 GOTO(out, rc = 0); /* no more records */
 942
 943                         /* if set, process the callback on this record */
 944                         if (test_bit_le(index, LLOG_HDR_BITMAP(llh))) {
 945                                 rec = (void *)tail - tail->lrt_len +
 946                                       sizeof(*tail);
 947
 948                                 rc = cb(env, loghandle, rec, data);
 949                                 if (rc == LLOG_PROC_BREAK) {
 950                                         GOTO(out, rc);
 951                                 } else if (rc == LLOG_DEL_RECORD) {
 952                                         rc = llog_cancel_rec(env, loghandle,
 953                                                              tail->lrt_index);
 954                                 }
 955                                 if (rc)
 956                                         GOTO(out, rc);
 957                         }
 958
 959                         /* previous record, still in buffer? */
 960                         --index;
 961                         if (index < first_index)
 962                                 GOTO(out, rc = 0);
 963                         tail = (void *)tail - tail->lrt_len;
 964                 }
 965         }
 966
 967 out:
 968         if (buf != NULL)
 969                 OBD_FREE_LARGE(buf, chunk_size);
 970         RETURN(rc);
 971 }
 972 EXPORT_SYMBOL(llog_reverse_process);
 973
 974 /**
 975  * new llog API
 976  *
 977  * API functions:
 978  *      llog_open - open llog, may not exist
 979  *      llog_exist - check if llog exists
 980  *      llog_close - close opened llog, pair for open, frees llog_handle
 981  *      llog_declare_create - declare llog creation
 982  *      llog_create - create new llog on disk, need transaction handle
 983  *      llog_declare_write_rec - declaration of llog write
 984  *      llog_write_rec - write llog record on disk, need transaction handle
 985  *      llog_declare_add - declare llog catalog record addition
 986  *      llog_add - add llog record in catalog, need transaction handle
 987  */
 988 int llog_exist(struct llog_handle *loghandle)
 989 {
 990         struct llog_operations  *lop;
 991         int                      rc;
 992
 993         ENTRY;
 994
 995         rc = llog_handle2ops(loghandle, &lop);
 996         if (rc)
 997                 RETURN(rc);
 998         if (lop->lop_exist == NULL)
 999                 RETURN(-EOPNOTSUPP);
1000
1001         rc = lop->lop_exist(loghandle);
1002         RETURN(rc);
1003 }
1004 EXPORT_SYMBOL(llog_exist);
1005
1006 int llog_declare_create(const struct lu_env *env,
1007                         struct llog_handle *loghandle, struct thandle *th)
1008 {
1009         const struct cred *old_cred;
1010         struct llog_operations  *lop;
1011         int rc;
1012
1013         ENTRY;
1014
1015         rc = llog_handle2ops(loghandle, &lop);
1016         if (rc)
1017                 RETURN(rc);
1018         if (lop->lop_declare_create == NULL)
1019                 RETURN(-EOPNOTSUPP);
1020
1021         old_cred = llog_raise_resource();
1022         rc = lop->lop_declare_create(env, loghandle, th);
1023         llog_restore_resource(old_cred);
1024         RETURN(rc);
1025 }
1026
1027 int llog_create(const struct lu_env *env, struct llog_handle *handle,
1028                 struct thandle *th)
1029 {
1030         const struct cred *old_cred;
1031         struct llog_operations  *lop;
1032         int rc;
1033
1034         ENTRY;
1035
1036         rc = llog_handle2ops(handle, &lop);
1037         if (rc)
1038                 RETURN(rc);
1039         if (lop->lop_create == NULL)
1040                 RETURN(-EOPNOTSUPP);
1041
1042         old_cred = llog_raise_resource();
1043         rc = lop->lop_create(env, handle, th);
1044         llog_restore_resource(old_cred);
1045         RETURN(rc);
1046 }
1047
1048 int llog_declare_write_rec(const struct lu_env *env,
1049                            struct llog_handle *handle,
1050                            struct llog_rec_hdr *rec, int idx,
1051                            struct thandle *th)
1052 {
1053         const struct cred *old_cred;
1054         struct llog_operations  *lop;
1055         int rc;
1056
1057         ENTRY;
1058
1059         rc = llog_handle2ops(handle, &lop);
1060         if (rc)
1061                 RETURN(rc);
1062         LASSERT(lop);
1063         if (lop->lop_declare_write_rec == NULL)
1064                 RETURN(-EOPNOTSUPP);
1065
1066         old_cred = llog_raise_resource();
1067         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
1068         llog_restore_resource(old_cred);
1069         RETURN(rc);
1070 }
1071
1072 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
1073                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1074                    int idx, struct thandle *th)
1075 {
1076         const struct cred *old_cred;
1077         struct llog_operations  *lop;
1078         int rc, buflen;
1079
1080         ENTRY;
1081
1082         /* API sanity checks */
1083         if (handle == NULL) {
1084                 CERROR("loghandle is missed\n");
1085                 RETURN(-EPROTO);
1086         } else if (handle->lgh_obj == NULL) {
1087                 CERROR("loghandle %p with NULL object\n",
1088                         handle);
1089                 RETURN(-EPROTO);
1090         } else if (th == NULL) {
1091                 CERROR("%s: missed transaction handle\n",
1092                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name);
1093                 RETURN(-EPROTO);
1094         } else if (handle->lgh_hdr == NULL) {
1095                 CERROR("%s: loghandle %p with no header\n",
1096                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name,
1097                         handle);
1098                 RETURN(-EPROTO);
1099         }
1100
1101         rc = llog_handle2ops(handle, &lop);
1102         if (rc)
1103                 RETURN(rc);
1104
1105         if (lop->lop_write_rec == NULL)
1106                 RETURN(-EOPNOTSUPP);
1107
1108         buflen = rec->lrh_len;
1109         LASSERT(cfs_size_round(buflen) == buflen);
1110
1111         old_cred = llog_raise_resource();
1112         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1113         llog_restore_resource(old_cred);
1114         RETURN(rc);
1115 }
1116
1117 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1118              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1119              struct thandle *th)
1120 {
1121         const struct cred *old_cred;
1122         int rc;
1123
1124         ENTRY;
1125
1126         if (lgh->lgh_logops->lop_add == NULL)
1127                 RETURN(-EOPNOTSUPP);
1128
1129         old_cred = llog_raise_resource();
1130         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1131         llog_restore_resource(old_cred);
1132         RETURN(rc);
1133 }
1134 EXPORT_SYMBOL(llog_add);
1135
1136 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1137                      struct llog_rec_hdr *rec, struct thandle *th)
1138 {
1139         const struct cred *old_cred;
1140         int rc;
1141
1142         ENTRY;
1143
1144         if (lgh->lgh_logops->lop_declare_add == NULL)
1145                 RETURN(-EOPNOTSUPP);
1146
1147         old_cred = llog_raise_resource();
1148         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1149         llog_restore_resource(old_cred);
1150         RETURN(rc);
1151 }
1152 EXPORT_SYMBOL(llog_declare_add);
1153
1154 /**
1155  * Helper function to open llog or create it if doesn't exist.
1156  * It hides all transaction handling from caller.
1157  */
1158 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1159                      struct llog_handle **res, struct llog_logid *logid,
1160                      char *name)
1161 {
1162         struct dt_device        *d;
1163         struct thandle          *th;
1164         int                      rc;
1165
1166         ENTRY;
1167
1168         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1169         if (rc)
1170                 RETURN(rc);
1171
1172         if (llog_exist(*res))
1173                 RETURN(0);
1174
1175         LASSERT((*res)->lgh_obj != NULL);
1176
1177         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1178
1179         if (unlikely(unlikely(d->dd_rdonly)))
1180                 RETURN(-EROFS);
1181
1182         th = dt_trans_create(env, d);
1183         if (IS_ERR(th))
1184                 GOTO(out, rc = PTR_ERR(th));
1185
1186         /* Create update llog object synchronously, which
1187          * happens during inialization process see
1188          * lod_sub_prep_llog(), to make sure the update
1189          * llog object is created before corss-MDT writing
1190          * updates into the llog object */
1191         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1192                 th->th_sync = 1;
1193
1194         th->th_wait_submit = 1;
1195         rc = llog_declare_create(env, *res, th);
1196         if (rc == 0) {
1197                 rc = dt_trans_start_local(env, d, th);
1198                 if (rc == 0)
1199                         rc = llog_create(env, *res, th);
1200         }
1201         dt_trans_stop(env, d, th);
1202 out:
1203         if (rc)
1204                 llog_close(env, *res);
1205         RETURN(rc);
1206 }
1207 EXPORT_SYMBOL(llog_open_create);
1208
1209 /**
1210  * Helper function to delete existent llog.
1211  */
1212 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1213                struct llog_logid *logid, char *name)
1214 {
1215         struct llog_handle      *handle;
1216         int                      rc = 0, rc2;
1217
1218         ENTRY;
1219
1220         /* nothing to erase */
1221         if (name == NULL && logid == NULL)
1222                 RETURN(0);
1223
1224         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1225         if (rc < 0)
1226                 RETURN(rc);
1227
1228         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1229         if (rc == 0)
1230                 rc = llog_destroy(env, handle);
1231
1232         rc2 = llog_close(env, handle);
1233         if (rc == 0)
1234                 rc = rc2;
1235         RETURN(rc);
1236 }
1237 EXPORT_SYMBOL(llog_erase);
1238
1239 /*
1240  * Helper function for write record in llog.
1241  * It hides all transaction handling from caller.
1242  * Valid only with local llog.
1243  */
1244 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1245                struct llog_rec_hdr *rec, int idx)
1246 {
1247         struct dt_device        *dt;
1248         struct thandle          *th;
1249         bool                    need_cookie;
1250         int                     rc;
1251
1252         ENTRY;
1253
1254         LASSERT(loghandle);
1255         LASSERT(loghandle->lgh_ctxt);
1256         LASSERT(loghandle->lgh_obj != NULL);
1257
1258         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1259
1260         if (unlikely(unlikely(dt->dd_rdonly)))
1261                 RETURN(-EROFS);
1262
1263         th = dt_trans_create(env, dt);
1264         if (IS_ERR(th))
1265                 RETURN(PTR_ERR(th));
1266
1267         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1268         if (rc)
1269                 GOTO(out_trans, rc);
1270
1271         th->th_wait_submit = 1;
1272         rc = dt_trans_start_local(env, dt, th);
1273         if (rc)
1274                 GOTO(out_trans, rc);
1275
1276         need_cookie = !(idx == LLOG_HEADER_IDX || idx == LLOG_NEXT_IDX);
1277
1278         down_write(&loghandle->lgh_lock);
1279         if (need_cookie) {
1280                 struct llog_thread_info *lti = llog_info(env);
1281
1282                 /* cookie comes from llog_process_thread */
1283                 rc = llog_write_rec(env, loghandle, rec, &lti->lgi_cookie,
1284                                     rec->lrh_index, th);
1285                 /* upper layer didn`t pass cookie so change rc */
1286                 rc = (rc == 1 ? 0 : rc);
1287         } else {
1288                 rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1289         }
1290
1291         up_write(&loghandle->lgh_lock);
1292 out_trans:
1293         dt_trans_stop(env, dt, th);
1294         RETURN(rc);
1295 }
1296 EXPORT_SYMBOL(llog_write);
1297
1298 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1299               struct llog_handle **lgh, struct llog_logid *logid,
1300               char *name, enum llog_open_param open_param)
1301 {
1302         const struct cred *old_cred;
1303         int      rc;
1304
1305         ENTRY;
1306
1307         LASSERT(ctxt);
1308         LASSERT(ctxt->loc_logops);
1309
1310         if (ctxt->loc_logops->lop_open == NULL) {
1311                 *lgh = NULL;
1312                 RETURN(-EOPNOTSUPP);
1313         }
1314
1315         *lgh = llog_alloc_handle();
1316         if (*lgh == NULL)
1317                 RETURN(-ENOMEM);
1318         (*lgh)->lgh_ctxt = ctxt;
1319         (*lgh)->lgh_logops = ctxt->loc_logops;
1320
1321         old_cred = llog_raise_resource();
1322         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1323         llog_restore_resource(old_cred);
1324         if (rc) {
1325                 llog_free_handle(*lgh);
1326                 *lgh = NULL;
1327         }
1328         RETURN(rc);
1329 }
1330 EXPORT_SYMBOL(llog_open);
1331
1332 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1333 {
1334         return llog_handle_put(env, loghandle);
1335 }
1336 EXPORT_SYMBOL(llog_close);
1337
1338 /**
1339  * Helper function to get the llog size in records. It is used by MGS
1340  * mostly to check that config llog exists and contains data.
1341  *
1342  * \param[in] env       execution environment
1343  * \param[in] ctxt      llog context
1344  * \param[in] name      llog name
1345  *
1346  * \retval              true if there are records in llog besides a header
1347  * \retval              false on error or llog without records
1348  */
1349 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1350                   char *name)
1351 {
1352         struct llog_handle      *llh;
1353         int                      rc = 0;
1354
1355         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1356         if (rc < 0) {
1357                 if (likely(rc == -ENOENT))
1358                         rc = 0;
1359                 GOTO(out, rc);
1360         }
1361
1362         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1363         if (rc)
1364                 GOTO(out_close, rc);
1365         rc = llog_get_size(llh);
1366
1367 out_close:
1368         llog_close(env, llh);
1369 out:
1370         /* The header is record 1, the llog is still considered as empty
1371          * if there is only header */
1372         return (rc <= 1);
1373 }
1374 EXPORT_SYMBOL(llog_is_empty);
1375
1376 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1377                       struct llog_rec_hdr *rec, void *data)
1378 {
1379         struct llog_handle      *copy_llh = data;
1380
1381         /* Append all records */
1382         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1383 }
1384
1385 /* backup plain llog */
1386 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1387                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1388                 char *name, char *backup)
1389 {
1390         struct llog_handle      *llh, *bllh;
1391         int                      rc;
1392
1393         ENTRY;
1394
1395         /* open original log */
1396         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1397         if (rc < 0) {
1398                 /* the -ENOENT case is also reported to the caller
1399                  * but silently so it should handle that if needed.
1400                  */
1401                 if (rc != -ENOENT)
1402                         CERROR("%s: failed to open log %s: rc = %d\n",
1403                                obd->obd_name, name, rc);
1404                 RETURN(rc);
1405         }
1406
1407         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1408         if (rc)
1409                 GOTO(out_close, rc);
1410
1411         /* Make sure there's no old backup log */
1412         rc = llog_erase(env, bctxt, NULL, backup);
1413         if (rc < 0 && rc != -ENOENT)
1414                 GOTO(out_close, rc);
1415
1416         /* open backup log */
1417         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1418         if (rc) {
1419                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1420                        obd->obd_name, backup, rc);
1421                 GOTO(out_close, rc);
1422         }
1423
1424         /* check that backup llog is not the same object as original one */
1425         if (llh->lgh_obj == bllh->lgh_obj) {
1426                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1427                        obd->obd_name, name, backup, llh->lgh_obj,
1428                        bllh->lgh_obj);
1429                 GOTO(out_backup, rc = -EEXIST);
1430         }
1431
1432         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1433         if (rc)
1434                 GOTO(out_backup, rc);
1435
1436         /* Copy log record by record */
1437         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1438                                   NULL, false);
1439         if (rc)
1440                 CERROR("%s: failed to backup log %s: rc = %d\n",
1441                        obd->obd_name, name, rc);
1442 out_backup:
1443         llog_close(env, bllh);
1444 out_close:
1445         llog_close(env, llh);
1446         RETURN(rc);
1447 }
1448 EXPORT_SYMBOL(llog_backup);
1449
1450 /* Get size of llog */
1451 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1452 {
1453         int rc;
1454         struct lu_attr la;
1455
1456         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1457         if (rc) {
1458                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1459                        llh->lgh_ctxt->loc_obd->obd_name,
1460                        PFID(&llh->lgh_id.lgl_oi.oi_fid), rc);
1461                 return 0;
1462         }
1463
1464         return la.la_size;
1465 }
1466 EXPORT_SYMBOL(llog_size);
1467