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, 2016, 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_class.h>
  51 #include "llog_internal.h"
  52 /*
  53  * Allocate a new log or catalog handle
  54  * Used inside llog_open().
  55  */
  56 static struct llog_handle *llog_alloc_handle(void)
  57 {
  58         struct llog_handle *loghandle;
  59
  60         OBD_ALLOC_PTR(loghandle);
  61         if (loghandle == NULL)
  62                 return NULL;
  63
  64         init_rwsem(&loghandle->lgh_lock);
  65         mutex_init(&loghandle->lgh_hdr_mutex);
  66         INIT_LIST_HEAD(&loghandle->u.phd.phd_entry);
  67         atomic_set(&loghandle->lgh_refcount, 1);
  68
  69         return loghandle;
  70 }
  71
  72 /*
  73  * Free llog handle and header data if exists. Used in llog_close() only
  74  */
  75 static void llog_free_handle(struct llog_handle *loghandle)
  76 {
  77         LASSERT(loghandle != NULL);
  78
  79         /* failed llog_init_handle */
  80         if (loghandle->lgh_hdr == NULL)
  81                 goto out;
  82
  83         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  84                 LASSERT(list_empty(&loghandle->u.phd.phd_entry));
  85         else if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  86                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  87         OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
  88 out:
  89         OBD_FREE_PTR(loghandle);
  90 }
  91
  92 void llog_handle_get(struct llog_handle *loghandle)
  93 {
  94         atomic_inc(&loghandle->lgh_refcount);
  95 }
  96
  97 void llog_handle_put(struct llog_handle *loghandle)
  98 {
  99         LASSERT(atomic_read(&loghandle->lgh_refcount) > 0);
 100         if (atomic_dec_and_test(&loghandle->lgh_refcount))
 101                 llog_free_handle(loghandle);
 102 }
 103
 104 static int llog_declare_destroy(const struct lu_env *env,
 105                                 struct llog_handle *handle,
 106                                 struct thandle *th)
 107 {
 108         struct llog_operations *lop;
 109         int rc;
 110
 111         ENTRY;
 112
 113         rc = llog_handle2ops(handle, &lop);
 114         if (rc)
 115                 RETURN(rc);
 116         if (lop->lop_declare_destroy == NULL)
 117                 RETURN(-EOPNOTSUPP);
 118
 119         rc = lop->lop_declare_destroy(env, handle, th);
 120
 121         RETURN(rc);
 122 }
 123
 124 int llog_trans_destroy(const struct lu_env *env, struct llog_handle *handle,
 125                        struct thandle *th)
 126 {
 127         struct llog_operations  *lop;
 128         int rc;
 129         ENTRY;
 130
 131         rc = llog_handle2ops(handle, &lop);
 132         if (rc < 0)
 133                 RETURN(rc);
 134         if (lop->lop_destroy == NULL)
 135                 RETURN(-EOPNOTSUPP);
 136
 137         LASSERT(handle->lgh_obj != NULL);
 138         if (!dt_object_exists(handle->lgh_obj))
 139                 RETURN(0);
 140
 141         rc = lop->lop_destroy(env, handle, th);
 142
 143         RETURN(rc);
 144 }
 145
 146 int llog_destroy(const struct lu_env *env, struct llog_handle *handle)
 147 {
 148         struct llog_operations  *lop;
 149         struct dt_device        *dt;
 150         struct thandle          *th;
 151         int rc;
 152
 153         ENTRY;
 154
 155         rc = llog_handle2ops(handle, &lop);
 156         if (rc < 0)
 157                 RETURN(rc);
 158         if (lop->lop_destroy == NULL)
 159                 RETURN(-EOPNOTSUPP);
 160
 161         if (handle->lgh_obj == NULL) {
 162                 /* if lgh_obj == NULL, then it is from client side destroy */
 163                 rc = lop->lop_destroy(env, handle, NULL);
 164                 RETURN(rc);
 165         }
 166
 167         if (!dt_object_exists(handle->lgh_obj))
 168                 RETURN(0);
 169
 170         dt = lu2dt_dev(handle->lgh_obj->do_lu.lo_dev);
 171
 172         th = dt_trans_create(env, dt);
 173         if (IS_ERR(th))
 174                 RETURN(PTR_ERR(th));
 175
 176         rc = llog_declare_destroy(env, handle, th);
 177         if (rc != 0)
 178                 GOTO(out_trans, rc);
 179
 180         rc = dt_trans_start_local(env, dt, th);
 181         if (rc < 0)
 182                 GOTO(out_trans, rc);
 183
 184         rc = lop->lop_destroy(env, handle, th);
 185
 186 out_trans:
 187         dt_trans_stop(env, dt, th);
 188
 189         RETURN(rc);
 190 }
 191 EXPORT_SYMBOL(llog_destroy);
 192
 193 /* returns negative on error; 0 if success; 1 if success & log destroyed */
 194 int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
 195                     int index)
 196 {
 197         struct llog_thread_info *lgi = llog_info(env);
 198         struct dt_device        *dt;
 199         struct llog_log_hdr     *llh = loghandle->lgh_hdr;
 200         struct thandle          *th;
 201         int                      rc;
 202         int rc1;
 203         bool subtract_count = false;
 204
 205         ENTRY;
 206
 207         CDEBUG(D_RPCTRACE, "Canceling %d in log "DFID"\n", index,
 208                PFID(&loghandle->lgh_id.lgl_oi.oi_fid));
 209
 210         if (index == 0) {
 211                 CERROR("Can't cancel index 0 which is header\n");
 212                 RETURN(-EINVAL);
 213         }
 214
 215         LASSERT(loghandle != NULL);
 216         LASSERT(loghandle->lgh_ctxt != NULL);
 217         LASSERT(loghandle->lgh_obj != NULL);
 218
 219         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 220
 221         th = dt_trans_create(env, dt);
 222         if (IS_ERR(th))
 223                 RETURN(PTR_ERR(th));
 224
 225         rc = llog_declare_write_rec(env, loghandle, &llh->llh_hdr, index, th);
 226         if (rc < 0)
 227                 GOTO(out_trans, rc);
 228
 229         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY)) {
 230                 rc = llog_declare_destroy(env, loghandle, th);
 231                 if (rc < 0)
 232                         GOTO(out_trans, rc);
 233         }
 234
 235         th->th_wait_submit = 1;
 236         rc = dt_trans_start_local(env, dt, th);
 237         if (rc < 0)
 238                 GOTO(out_trans, rc);
 239
 240         down_write(&loghandle->lgh_lock);
 241         /* clear bitmap */
 242         mutex_lock(&loghandle->lgh_hdr_mutex);
 243         if (!ext2_clear_bit(index, LLOG_HDR_BITMAP(llh))) {
 244                 CDEBUG(D_RPCTRACE, "Catalog index %u already clear?\n", index);
 245                 GOTO(out_unlock, rc);
 246         }
 247
 248         loghandle->lgh_hdr->llh_count--;
 249         subtract_count = true;
 250         /* Pass this index to llog_osd_write_rec(), which will use the index
 251          * to only update the necesary bitmap. */
 252         lgi->lgi_cookie.lgc_index = index;
 253         /* update header */
 254         rc = llog_write_rec(env, loghandle, &llh->llh_hdr, &lgi->lgi_cookie,
 255                             LLOG_HEADER_IDX, th);
 256         if (rc != 0)
 257                 GOTO(out_unlock, rc);
 258
 259         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 260             (llh->llh_count == 1) &&
 261             ((loghandle->lgh_last_idx == LLOG_HDR_BITMAP_SIZE(llh) - 1) ||
 262              (loghandle->u.phd.phd_cat_handle != NULL &&
 263               loghandle->u.phd.phd_cat_handle->u.chd.chd_current_log !=
 264                 loghandle))) {
 265                 /* never try to destroy it again */
 266                 llh->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 267                 rc = llog_trans_destroy(env, loghandle, th);
 268                 if (rc < 0) {
 269                         /* Sigh, can not destroy the final plain llog, but
 270                          * the bitmap has been clearly, so the record can not
 271                          * be accessed anymore, let's return 0 for now, and
 272                          * the orphan will be handled by LFSCK. */
 273                         CERROR("%s: can't destroy empty llog "DFID": rc = %d\n",
 274                                loghandle->lgh_ctxt->loc_obd->obd_name,
 275                                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rc);
 276                         GOTO(out_unlock, rc = 0);
 277                 }
 278                 rc = LLOG_DEL_PLAIN;
 279         }
 280
 281 out_unlock:
 282         mutex_unlock(&loghandle->lgh_hdr_mutex);
 283         up_write(&loghandle->lgh_lock);
 284 out_trans:
 285         rc1 = dt_trans_stop(env, dt, th);
 286         if (rc == 0)
 287                 rc = rc1;
 288         if (rc < 0 && subtract_count) {
 289                 mutex_lock(&loghandle->lgh_hdr_mutex);
 290                 loghandle->lgh_hdr->llh_count++;
 291                 ext2_set_bit(index, LLOG_HDR_BITMAP(llh));
 292                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 293         }
 294         RETURN(rc);
 295 }
 296
 297 int llog_read_header(const struct lu_env *env, struct llog_handle *handle,
 298                      const struct obd_uuid *uuid)
 299 {
 300         struct llog_operations *lop;
 301         int rc;
 302         ENTRY;
 303
 304         rc = llog_handle2ops(handle, &lop);
 305         if (rc)
 306                 RETURN(rc);
 307
 308         if (lop->lop_read_header == NULL)
 309                 RETURN(-EOPNOTSUPP);
 310
 311         rc = lop->lop_read_header(env, handle);
 312         if (rc == LLOG_EEMPTY) {
 313                 struct llog_log_hdr *llh = handle->lgh_hdr;
 314
 315                 /* lrh_len should be initialized in llog_init_handle */
 316                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 317                 llh->llh_count = 1;         /* for the header record */
 318                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 319                 LASSERT(handle->lgh_ctxt->loc_chunk_size >=
 320                                                 LLOG_MIN_CHUNK_SIZE);
 321                 llh->llh_hdr.lrh_len = handle->lgh_ctxt->loc_chunk_size;
 322                 llh->llh_hdr.lrh_index = 0;
 323                 llh->llh_timestamp = ktime_get_real_seconds();
 324                 if (uuid)
 325                         memcpy(&llh->llh_tgtuuid, uuid,
 326                                sizeof(llh->llh_tgtuuid));
 327                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 328                 /* Since update llog header might also call this function,
 329                  * let's reset the bitmap to 0 here */
 330                 memset(LLOG_HDR_BITMAP(llh), 0, llh->llh_hdr.lrh_len -
 331                                                 llh->llh_bitmap_offset -
 332                                                 sizeof(llh->llh_tail));
 333                 ext2_set_bit(0, LLOG_HDR_BITMAP(llh));
 334                 LLOG_HDR_TAIL(llh)->lrt_len = llh->llh_hdr.lrh_len;
 335                 LLOG_HDR_TAIL(llh)->lrt_index = llh->llh_hdr.lrh_index;
 336                 rc = 0;
 337         }
 338         RETURN(rc);
 339 }
 340 EXPORT_SYMBOL(llog_read_header);
 341
 342 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 343                      int flags, struct obd_uuid *uuid)
 344 {
 345         struct llog_log_hdr     *llh;
 346         enum llog_flag           fmt = flags & LLOG_F_EXT_MASK;
 347         int                      rc;
 348         int                     chunk_size = handle->lgh_ctxt->loc_chunk_size;
 349         ENTRY;
 350
 351         LASSERT(handle->lgh_hdr == NULL);
 352
 353         LASSERT(chunk_size >= LLOG_MIN_CHUNK_SIZE);
 354         OBD_ALLOC_LARGE(llh, chunk_size);
 355         if (llh == NULL)
 356                 RETURN(-ENOMEM);
 357
 358         handle->lgh_hdr = llh;
 359         handle->lgh_hdr_size = chunk_size;
 360         /* first assign flags to use llog_client_ops */
 361         llh->llh_flags = flags;
 362         rc = llog_read_header(env, handle, uuid);
 363         if (rc == 0) {
 364                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 365                               flags & LLOG_F_IS_CAT) ||
 366                              (llh->llh_flags & LLOG_F_IS_CAT &&
 367                               flags & LLOG_F_IS_PLAIN))) {
 368                         CERROR("%s: llog type is %s but initializing %s\n",
 369                                handle->lgh_ctxt->loc_obd->obd_name,
 370                                llh->llh_flags & LLOG_F_IS_CAT ?
 371                                "catalog" : "plain",
 372                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 373                         GOTO(out, rc = -EINVAL);
 374                 } else if (llh->llh_flags &
 375                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 376                         /*
 377                          * it is possible to open llog without specifying llog
 378                          * type so it is taken from llh_flags
 379                          */
 380                         flags = llh->llh_flags;
 381                 } else {
 382                         /* for some reason the llh_flags has no type set */
 383                         CERROR("llog type is not specified!\n");
 384                         GOTO(out, rc = -EINVAL);
 385                 }
 386                 if (unlikely(uuid &&
 387                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 388                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 389                                handle->lgh_ctxt->loc_obd->obd_name,
 390                                (char *)uuid->uuid,
 391                                (char *)llh->llh_tgtuuid.uuid);
 392                         GOTO(out, rc = -EEXIST);
 393                 }
 394         }
 395         if (flags & LLOG_F_IS_CAT) {
 396                 LASSERT(list_empty(&handle->u.chd.chd_head));
 397                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 398                 llh->llh_size = sizeof(struct llog_logid_rec);
 399                 llh->llh_flags |= LLOG_F_IS_FIXSIZE;
 400         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 401                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 402                        handle->lgh_ctxt->loc_obd->obd_name,
 403                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 404                 rc = -EINVAL;
 405         }
 406         llh->llh_flags |= fmt;
 407 out:
 408         if (rc) {
 409                 OBD_FREE_LARGE(llh, chunk_size);
 410                 handle->lgh_hdr = NULL;
 411         }
 412         RETURN(rc);
 413 }
 414 EXPORT_SYMBOL(llog_init_handle);
 415
 416 static int llog_process_thread(void *arg)
 417 {
 418         struct llog_process_info        *lpi = arg;
 419         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 420         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 421         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 422         char                            *buf;
 423         size_t                           chunk_size;
 424         __u64                            cur_offset;
 425         int                              rc = 0, index = 1, last_index;
 426         int                              saved_index = 0;
 427         int                              last_called_index = 0;
 428         bool                             repeated = false;
 429
 430         ENTRY;
 431
 432         if (llh == NULL)
 433                 RETURN(-EINVAL);
 434
 435         cur_offset = chunk_size = llh->llh_hdr.lrh_len;
 436         /* expect chunk_size to be power of two */
 437         LASSERT(is_power_of_2(chunk_size));
 438
 439         OBD_ALLOC_LARGE(buf, chunk_size);
 440         if (buf == NULL) {
 441                 lpi->lpi_rc = -ENOMEM;
 442                 RETURN(0);
 443         }
 444
 445         if (cd != NULL) {
 446                 last_called_index = cd->lpcd_first_idx;
 447                 index = cd->lpcd_first_idx + 1;
 448         }
 449         if (cd != NULL && cd->lpcd_last_idx)
 450                 last_index = cd->lpcd_last_idx;
 451         else
 452                 last_index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 453
 454         while (rc == 0) {
 455                 struct llog_rec_hdr *rec;
 456                 off_t chunk_offset = 0;
 457                 unsigned int buf_offset = 0;
 458                 bool partial_chunk;
 459                 int     lh_last_idx;
 460
 461                 /* skip records not set in bitmap */
 462                 while (index <= last_index &&
 463                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 464                         ++index;
 465
 466                 /* There are no indices prior the last_index */
 467                 if (index > last_index)
 468                         break;
 469
 470                 CDEBUG(D_OTHER, "index: %d last_index %d\n", index,
 471                        last_index);
 472
 473 repeat:
 474                 /* get the buf with our target record; avoid old garbage */
 475                 memset(buf, 0, chunk_size);
 476                 /* the record index for outdated chunk data */
 477                 lh_last_idx = loghandle->lgh_last_idx + 1;
 478                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 479                                      index, &cur_offset, buf, chunk_size);
 480                 if (repeated && rc)
 481                         CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
 482                                " buf_offset %u, rc = %d\n", cur_offset,
 483                                (__u64)chunk_offset, buf_offset, rc);
 484                 /* we`ve tried to reread the chunk, but there is no
 485                  * new records */
 486                 if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
 487                     cur_offset)
 488                         GOTO(out, rc = 0);
 489                 if (rc != 0)
 490                         GOTO(out, rc);
 491
 492                 /* NB: after llog_next_block() call the cur_offset is the
 493                  * offset of the next block after read one.
 494                  * The absolute offset of the current chunk is calculated
 495                  * from cur_offset value and stored in chunk_offset variable.
 496                  */
 497                 if ((cur_offset & (chunk_size - 1)) != 0) {
 498                         partial_chunk = true;
 499                         chunk_offset = cur_offset & ~(chunk_size - 1);
 500                 } else {
 501                         partial_chunk = false;
 502                         chunk_offset = cur_offset - chunk_size;
 503                 }
 504
 505                 /* NB: when rec->lrh_len is accessed it is already swabbed
 506                  * since it is used at the "end" of the loop and the rec
 507                  * swabbing is done at the beginning of the loop. */
 508                 for (rec = (struct llog_rec_hdr *)(buf + buf_offset);
 509                      (char *)rec < buf + chunk_size;
 510                      rec = llog_rec_hdr_next(rec)) {
 511
 512                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 513                                rec, rec->lrh_type);
 514
 515                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 516                                 lustre_swab_llog_rec(rec);
 517
 518                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 519                                rec->lrh_type, rec->lrh_index);
 520
 521                         /* the bitmap could be changed during processing
 522                          * records from the chunk. For wrapped catalog
 523                          * it means we can read deleted record and try to
 524                          * process it. Check this case and reread the chunk. */
 525
 526                         /* for partial chunk the end of it is zeroed, check
 527                          * for index 0 to distinguish it. */
 528                         if ((partial_chunk && rec->lrh_index == 0) ||
 529                              (index == lh_last_idx &&
 530                               lh_last_idx != (loghandle->lgh_last_idx + 1))) {
 531                                 /* concurrent llog_add() might add new records
 532                                  * while llog_processing, check this is not
 533                                  * the case and re-read the current chunk
 534                                  * otherwise. */
 535                                 int records;
 536                                 /* lgh_last_idx could be less then index
 537                                  * for catalog, if catalog is wrapped */
 538                                 if ((index > loghandle->lgh_last_idx &&
 539                                     !(loghandle->lgh_hdr->llh_flags &
 540                                       LLOG_F_IS_CAT)) || repeated ||
 541                                     (loghandle->lgh_obj != NULL &&
 542                                      dt_object_remote(loghandle->lgh_obj)))
 543                                         GOTO(out, rc = 0);
 544                                 /* <2 records means no more records
 545                                  * if the last record we processed was
 546                                  * the final one, then the underlying
 547                                  * object might have been destroyed yet.
 548                                  * we better don't access that.. */
 549                                 mutex_lock(&loghandle->lgh_hdr_mutex);
 550                                 records = loghandle->lgh_hdr->llh_count;
 551                                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 552                                 if (records <= 1)
 553                                         GOTO(out, rc = 0);
 554                                 CDEBUG(D_OTHER, "Re-read last llog buffer for "
 555                                        "new records, index %u, last %u\n",
 556                                        index, loghandle->lgh_last_idx);
 557                                 /* save offset inside buffer for the re-read */
 558                                 buf_offset = (char *)rec - (char *)buf;
 559                                 cur_offset = chunk_offset;
 560                                 repeated = true;
 561                                 goto repeat;
 562                         }
 563
 564                         repeated = false;
 565
 566                         if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
 567                                 CWARN("%s: invalid length %d in llog "DFID
 568                                       "record for index %d/%d\n",
 569                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 570                                        rec->lrh_len,
 571                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 572                                        rec->lrh_index, index);
 573
 574                                 GOTO(out, rc = -EINVAL);
 575                         }
 576
 577                         if (rec->lrh_index < index) {
 578                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 579                                        rec->lrh_index);
 580                                 continue;
 581                         }
 582
 583                         if (rec->lrh_index != index) {
 584                                 CERROR("%s: "DFID" Invalid record: index %u"
 585                                        " but expected %u\n",
 586                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 587                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 588                                        rec->lrh_index, index);
 589                                 GOTO(out, rc = -ERANGE);
 590                         }
 591
 592                         CDEBUG(D_OTHER,
 593                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 594                                rec->lrh_index, rec->lrh_len,
 595                                (int)(buf + chunk_size - (char *)rec));
 596
 597                         loghandle->lgh_cur_idx = rec->lrh_index;
 598                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 599                                                     chunk_offset;
 600
 601                         /* if set, process the callback on this record */
 602                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 603                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 604                                                  lpi->lpi_cbdata);
 605                                 last_called_index = index;
 606                                 if (rc == LLOG_PROC_BREAK) {
 607                                         GOTO(out, rc);
 608                                 } else if (rc == LLOG_DEL_RECORD) {
 609                                         rc = llog_cancel_rec(lpi->lpi_env,
 610                                                              loghandle,
 611                                                              rec->lrh_index);
 612                                 }
 613                                 if (rc)
 614                                         GOTO(out, rc);
 615                                 /* some stupid callbacks directly cancel records
 616                                  * and delete llog. Check it and stop
 617                                  * processing. */
 618                                 if (loghandle->lgh_hdr == NULL ||
 619                                     loghandle->lgh_hdr->llh_count == 1)
 620                                         GOTO(out, rc = 0);
 621                         }
 622                         /* exit if the last index is reached */
 623                         if (index >= last_index)
 624                                 GOTO(out, rc = 0);
 625                         ++index;
 626                 }
 627         }
 628
 629 out:
 630         if (cd != NULL)
 631                 cd->lpcd_last_idx = last_called_index;
 632
 633         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 634                 if (dt_object_remote(loghandle->lgh_obj)) {
 635                         /* If it is remote object, then -EIO might means
 636                          * disconnection or eviction, let's return -EAGAIN,
 637                          * so for update recovery log processing, it will
 638                          * retry until the umount or abort recovery, see
 639                          * lod_sub_recovery_thread() */
 640                         CERROR("%s retry remote llog process\n",
 641                                loghandle->lgh_ctxt->loc_obd->obd_name);
 642                         rc = -EAGAIN;
 643                 } else {
 644                         /* something bad happened to the processing of a local
 645                          * llog file, probably I/O error or the log got
 646                          * corrupted to be able to finally release the log we
 647                          * discard any remaining bits in the header */
 648                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 649                                " %s index %d count %d\n",
 650                                loghandle->lgh_ctxt->loc_obd->obd_name,
 651                                POSTID(&loghandle->lgh_id.lgl_oi),
 652                                loghandle->lgh_id.lgl_ogen,
 653                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 654                                 "plain"), index, llh->llh_count);
 655
 656                         while (index <= last_index) {
 657                                 if (ext2_test_bit(index,
 658                                                   LLOG_HDR_BITMAP(llh)) != 0)
 659                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 660                                                         index);
 661                                 index++;
 662                         }
 663                         rc = 0;
 664                 }
 665         }
 666
 667         OBD_FREE_LARGE(buf, chunk_size);
 668         lpi->lpi_rc = rc;
 669         return 0;
 670 }
 671
 672 static int llog_process_thread_daemonize(void *arg)
 673 {
 674         struct llog_process_info        *lpi = arg;
 675         struct lu_env                    env;
 676         int                              rc;
 677         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 678
 679         task_lock(lpi->lpi_reftask);
 680         new_ns = lpi->lpi_reftask->nsproxy;
 681         if (curr_ns != new_ns) {
 682                 get_nsproxy(new_ns);
 683
 684                 current->nsproxy = new_ns;
 685                 /* XXX: we should call put_nsproxy() instead of
 686                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 687                  * used outside of the kernel itself, because it calls
 688                  * free_nsproxy() which is not exported by the kernel
 689                  * (defined in kernel/nsproxy.c) */
 690                 atomic_dec(&curr_ns->count);
 691         }
 692         task_unlock(lpi->lpi_reftask);
 693
 694         unshare_fs_struct();
 695
 696         /* client env has no keys, tags is just 0 */
 697         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 698         if (rc)
 699                 goto out;
 700         lpi->lpi_env = &env;
 701
 702         rc = llog_process_thread(arg);
 703
 704         lu_env_fini(&env);
 705 out:
 706         complete(&lpi->lpi_completion);
 707         return rc;
 708 }
 709
 710 int llog_process_or_fork(const struct lu_env *env,
 711                          struct llog_handle *loghandle,
 712                          llog_cb_t cb, void *data, void *catdata, bool fork)
 713 {
 714         struct llog_process_info *lpi;
 715         int                      rc;
 716
 717         ENTRY;
 718
 719         OBD_ALLOC_PTR(lpi);
 720         if (lpi == NULL) {
 721                 CERROR("cannot alloc pointer\n");
 722                 RETURN(-ENOMEM);
 723         }
 724         lpi->lpi_loghandle = loghandle;
 725         lpi->lpi_cb        = cb;
 726         lpi->lpi_cbdata    = data;
 727         lpi->lpi_catdata   = catdata;
 728
 729         if (fork) {
 730                 struct task_struct *task;
 731
 732                 /* The new thread can't use parent env,
 733                  * init the new one in llog_process_thread_daemonize. */
 734                 lpi->lpi_env = NULL;
 735                 init_completion(&lpi->lpi_completion);
 736                 /* take reference to current, so that
 737                  * llog_process_thread_daemonize() can use it to switch to
 738                  * namespace associated with current  */
 739                 lpi->lpi_reftask = current;
 740                 task = kthread_run(llog_process_thread_daemonize, lpi,
 741                                    "llog_process_thread");
 742                 if (IS_ERR(task)) {
 743                         rc = PTR_ERR(task);
 744                         CERROR("%s: cannot start thread: rc = %d\n",
 745                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 746                         GOTO(out_lpi, rc);
 747                 }
 748                 wait_for_completion(&lpi->lpi_completion);
 749         } else {
 750                 lpi->lpi_env = env;
 751                 llog_process_thread(lpi);
 752         }
 753         rc = lpi->lpi_rc;
 754
 755 out_lpi:
 756         OBD_FREE_PTR(lpi);
 757         RETURN(rc);
 758 }
 759 EXPORT_SYMBOL(llog_process_or_fork);
 760
 761 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 762                  llog_cb_t cb, void *data, void *catdata)
 763 {
 764         int rc;
 765         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 766         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 767 }
 768 EXPORT_SYMBOL(llog_process);
 769
 770 int llog_reverse_process(const struct lu_env *env,
 771                          struct llog_handle *loghandle, llog_cb_t cb,
 772                          void *data, void *catdata)
 773 {
 774         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 775         struct llog_process_cat_data *cd = catdata;
 776         void *buf;
 777         int rc = 0, first_index = 1, index, idx;
 778         __u32   chunk_size = llh->llh_hdr.lrh_len;
 779         ENTRY;
 780
 781         OBD_ALLOC_LARGE(buf, chunk_size);
 782         if (buf == NULL)
 783                 RETURN(-ENOMEM);
 784
 785         if (cd != NULL)
 786                 first_index = cd->lpcd_first_idx + 1;
 787         if (cd != NULL && cd->lpcd_last_idx)
 788                 index = cd->lpcd_last_idx;
 789         else
 790                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 791
 792         while (rc == 0) {
 793                 struct llog_rec_hdr *rec;
 794                 struct llog_rec_tail *tail;
 795
 796                 /* skip records not set in bitmap */
 797                 while (index >= first_index &&
 798                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 799                         --index;
 800
 801                 LASSERT(index >= first_index - 1);
 802                 if (index == first_index - 1)
 803                         break;
 804
 805                 /* get the buf with our target record; avoid old garbage */
 806                 memset(buf, 0, chunk_size);
 807                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 808                 if (rc)
 809                         GOTO(out, rc);
 810
 811                 rec = buf;
 812                 idx = rec->lrh_index;
 813                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 814                 while (idx < index) {
 815                         rec = (void *)rec + rec->lrh_len;
 816                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 817                                 lustre_swab_llog_rec(rec);
 818                         idx ++;
 819                 }
 820                 LASSERT(idx == index);
 821                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 822
 823                 /* process records in buffer, starting where we found one */
 824                 while ((void *)tail > buf) {
 825                         if (tail->lrt_index == 0)
 826                                 GOTO(out, rc = 0); /* no more records */
 827
 828                         /* if set, process the callback on this record */
 829                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 830                                 rec = (void *)tail - tail->lrt_len +
 831                                       sizeof(*tail);
 832
 833                                 rc = cb(env, loghandle, rec, data);
 834                                 if (rc == LLOG_PROC_BREAK) {
 835                                         GOTO(out, rc);
 836                                 } else if (rc == LLOG_DEL_RECORD) {
 837                                         rc = llog_cancel_rec(env, loghandle,
 838                                                              tail->lrt_index);
 839                                 }
 840                                 if (rc)
 841                                         GOTO(out, rc);
 842                         }
 843
 844                         /* previous record, still in buffer? */
 845                         --index;
 846                         if (index < first_index)
 847                                 GOTO(out, rc = 0);
 848                         tail = (void *)tail - tail->lrt_len;
 849                 }
 850         }
 851
 852 out:
 853         if (buf != NULL)
 854                 OBD_FREE_LARGE(buf, chunk_size);
 855         RETURN(rc);
 856 }
 857 EXPORT_SYMBOL(llog_reverse_process);
 858
 859 /**
 860  * new llog API
 861  *
 862  * API functions:
 863  *      llog_open - open llog, may not exist
 864  *      llog_exist - check if llog exists
 865  *      llog_close - close opened llog, pair for open, frees llog_handle
 866  *      llog_declare_create - declare llog creation
 867  *      llog_create - create new llog on disk, need transaction handle
 868  *      llog_declare_write_rec - declaration of llog write
 869  *      llog_write_rec - write llog record on disk, need transaction handle
 870  *      llog_declare_add - declare llog catalog record addition
 871  *      llog_add - add llog record in catalog, need transaction handle
 872  */
 873 int llog_exist(struct llog_handle *loghandle)
 874 {
 875         struct llog_operations  *lop;
 876         int                      rc;
 877
 878         ENTRY;
 879
 880         rc = llog_handle2ops(loghandle, &lop);
 881         if (rc)
 882                 RETURN(rc);
 883         if (lop->lop_exist == NULL)
 884                 RETURN(-EOPNOTSUPP);
 885
 886         rc = lop->lop_exist(loghandle);
 887         RETURN(rc);
 888 }
 889 EXPORT_SYMBOL(llog_exist);
 890
 891 int llog_declare_create(const struct lu_env *env,
 892                         struct llog_handle *loghandle, struct thandle *th)
 893 {
 894         struct llog_operations  *lop;
 895         int                      raised, rc;
 896
 897         ENTRY;
 898
 899         rc = llog_handle2ops(loghandle, &lop);
 900         if (rc)
 901                 RETURN(rc);
 902         if (lop->lop_declare_create == NULL)
 903                 RETURN(-EOPNOTSUPP);
 904
 905         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 906         if (!raised)
 907                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 908         rc = lop->lop_declare_create(env, loghandle, th);
 909         if (!raised)
 910                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 911         RETURN(rc);
 912 }
 913
 914 int llog_create(const struct lu_env *env, struct llog_handle *handle,
 915                 struct thandle *th)
 916 {
 917         struct llog_operations  *lop;
 918         int                      raised, rc;
 919
 920         ENTRY;
 921
 922         rc = llog_handle2ops(handle, &lop);
 923         if (rc)
 924                 RETURN(rc);
 925         if (lop->lop_create == NULL)
 926                 RETURN(-EOPNOTSUPP);
 927
 928         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 929         if (!raised)
 930                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 931         rc = lop->lop_create(env, handle, th);
 932         if (!raised)
 933                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 934         RETURN(rc);
 935 }
 936
 937 int llog_declare_write_rec(const struct lu_env *env,
 938                            struct llog_handle *handle,
 939                            struct llog_rec_hdr *rec, int idx,
 940                            struct thandle *th)
 941 {
 942         struct llog_operations  *lop;
 943         int                      raised, rc;
 944
 945         ENTRY;
 946
 947         rc = llog_handle2ops(handle, &lop);
 948         if (rc)
 949                 RETURN(rc);
 950         LASSERT(lop);
 951         if (lop->lop_declare_write_rec == NULL)
 952                 RETURN(-EOPNOTSUPP);
 953
 954         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 955         if (!raised)
 956                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 957         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
 958         if (!raised)
 959                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 960         RETURN(rc);
 961 }
 962
 963 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
 964                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
 965                    int idx, struct thandle *th)
 966 {
 967         struct llog_operations  *lop;
 968         int                      raised, rc, buflen;
 969
 970         ENTRY;
 971
 972         /* API sanity checks */
 973         if (handle == NULL) {
 974                 CERROR("loghandle is missed\n");
 975                 RETURN(-EPROTO);
 976         } else if (handle->lgh_obj == NULL) {
 977                 CERROR("loghandle %p with NULL object\n",
 978                         handle);
 979                 RETURN(-EPROTO);
 980         } else if (th == NULL) {
 981                 CERROR("%s: missed transaction handle\n",
 982                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name);
 983                 RETURN(-EPROTO);
 984         } else if (handle->lgh_hdr == NULL) {
 985                 CERROR("%s: loghandle %p with no header\n",
 986                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name,
 987                         handle);
 988                 RETURN(-EPROTO);
 989         }
 990
 991         rc = llog_handle2ops(handle, &lop);
 992         if (rc)
 993                 RETURN(rc);
 994
 995         if (lop->lop_write_rec == NULL)
 996                 RETURN(-EOPNOTSUPP);
 997
 998         buflen = rec->lrh_len;
 999         LASSERT(cfs_size_round(buflen) == buflen);
1000
1001         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1002         if (!raised)
1003                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1004         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1005         if (!raised)
1006                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1007         RETURN(rc);
1008 }
1009
1010 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1011              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1012              struct thandle *th)
1013 {
1014         int raised, rc;
1015
1016         ENTRY;
1017
1018         if (lgh->lgh_logops->lop_add == NULL)
1019                 RETURN(-EOPNOTSUPP);
1020
1021         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1022         if (!raised)
1023                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1024         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1025         if (!raised)
1026                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1027         RETURN(rc);
1028 }
1029 EXPORT_SYMBOL(llog_add);
1030
1031 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1032                      struct llog_rec_hdr *rec, struct thandle *th)
1033 {
1034         int raised, rc;
1035
1036         ENTRY;
1037
1038         if (lgh->lgh_logops->lop_declare_add == NULL)
1039                 RETURN(-EOPNOTSUPP);
1040
1041         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1042         if (!raised)
1043                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1044         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1045         if (!raised)
1046                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1047         RETURN(rc);
1048 }
1049 EXPORT_SYMBOL(llog_declare_add);
1050
1051 /**
1052  * Helper function to open llog or create it if doesn't exist.
1053  * It hides all transaction handling from caller.
1054  */
1055 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1056                      struct llog_handle **res, struct llog_logid *logid,
1057                      char *name)
1058 {
1059         struct dt_device        *d;
1060         struct thandle          *th;
1061         int                      rc;
1062
1063         ENTRY;
1064
1065         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1066         if (rc)
1067                 RETURN(rc);
1068
1069         if (llog_exist(*res))
1070                 RETURN(0);
1071
1072         LASSERT((*res)->lgh_obj != NULL);
1073
1074         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1075
1076         th = dt_trans_create(env, d);
1077         if (IS_ERR(th))
1078                 GOTO(out, rc = PTR_ERR(th));
1079
1080         /* Create update llog object synchronously, which
1081          * happens during inialization process see
1082          * lod_sub_prep_llog(), to make sure the update
1083          * llog object is created before corss-MDT writing
1084          * updates into the llog object */
1085         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1086                 th->th_sync = 1;
1087
1088         th->th_wait_submit = 1;
1089         rc = llog_declare_create(env, *res, th);
1090         if (rc == 0) {
1091                 rc = dt_trans_start_local(env, d, th);
1092                 if (rc == 0)
1093                         rc = llog_create(env, *res, th);
1094         }
1095         dt_trans_stop(env, d, th);
1096 out:
1097         if (rc)
1098                 llog_close(env, *res);
1099         RETURN(rc);
1100 }
1101 EXPORT_SYMBOL(llog_open_create);
1102
1103 /**
1104  * Helper function to delete existent llog.
1105  */
1106 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1107                struct llog_logid *logid, char *name)
1108 {
1109         struct llog_handle      *handle;
1110         int                      rc = 0, rc2;
1111
1112         ENTRY;
1113
1114         /* nothing to erase */
1115         if (name == NULL && logid == NULL)
1116                 RETURN(0);
1117
1118         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1119         if (rc < 0)
1120                 RETURN(rc);
1121
1122         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1123         if (rc == 0)
1124                 rc = llog_destroy(env, handle);
1125
1126         rc2 = llog_close(env, handle);
1127         if (rc == 0)
1128                 rc = rc2;
1129         RETURN(rc);
1130 }
1131 EXPORT_SYMBOL(llog_erase);
1132
1133 /*
1134  * Helper function for write record in llog.
1135  * It hides all transaction handling from caller.
1136  * Valid only with local llog.
1137  */
1138 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1139                struct llog_rec_hdr *rec, int idx)
1140 {
1141         struct dt_device        *dt;
1142         struct thandle          *th;
1143         int                      rc;
1144
1145         ENTRY;
1146
1147         LASSERT(loghandle);
1148         LASSERT(loghandle->lgh_ctxt);
1149         LASSERT(loghandle->lgh_obj != NULL);
1150
1151         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1152
1153         th = dt_trans_create(env, dt);
1154         if (IS_ERR(th))
1155                 RETURN(PTR_ERR(th));
1156
1157         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1158         if (rc)
1159                 GOTO(out_trans, rc);
1160
1161         th->th_wait_submit = 1;
1162         rc = dt_trans_start_local(env, dt, th);
1163         if (rc)
1164                 GOTO(out_trans, rc);
1165
1166         down_write(&loghandle->lgh_lock);
1167         rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1168         up_write(&loghandle->lgh_lock);
1169 out_trans:
1170         dt_trans_stop(env, dt, th);
1171         RETURN(rc);
1172 }
1173 EXPORT_SYMBOL(llog_write);
1174
1175 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1176               struct llog_handle **lgh, struct llog_logid *logid,
1177               char *name, enum llog_open_param open_param)
1178 {
1179         int      raised;
1180         int      rc;
1181
1182         ENTRY;
1183
1184         LASSERT(ctxt);
1185         LASSERT(ctxt->loc_logops);
1186
1187         if (ctxt->loc_logops->lop_open == NULL) {
1188                 *lgh = NULL;
1189                 RETURN(-EOPNOTSUPP);
1190         }
1191
1192         *lgh = llog_alloc_handle();
1193         if (*lgh == NULL)
1194                 RETURN(-ENOMEM);
1195         (*lgh)->lgh_ctxt = ctxt;
1196         (*lgh)->lgh_logops = ctxt->loc_logops;
1197
1198         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1199         if (!raised)
1200                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1201         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1202         if (!raised)
1203                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1204         if (rc) {
1205                 llog_free_handle(*lgh);
1206                 *lgh = NULL;
1207         }
1208         RETURN(rc);
1209 }
1210 EXPORT_SYMBOL(llog_open);
1211
1212 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1213 {
1214         struct llog_operations  *lop;
1215         int                      rc;
1216
1217         ENTRY;
1218
1219         rc = llog_handle2ops(loghandle, &lop);
1220         if (rc)
1221                 GOTO(out, rc);
1222         if (lop->lop_close == NULL)
1223                 GOTO(out, rc = -EOPNOTSUPP);
1224         rc = lop->lop_close(env, loghandle);
1225 out:
1226         llog_handle_put(loghandle);
1227         RETURN(rc);
1228 }
1229 EXPORT_SYMBOL(llog_close);
1230
1231 /**
1232  * Helper function to get the llog size in records. It is used by MGS
1233  * mostly to check that config llog exists and contains data.
1234  *
1235  * \param[in] env       execution environment
1236  * \param[in] ctxt      llog context
1237  * \param[in] name      llog name
1238  *
1239  * \retval              true if there are records in llog besides a header
1240  * \retval              false on error or llog without records
1241  */
1242 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1243                   char *name)
1244 {
1245         struct llog_handle      *llh;
1246         int                      rc = 0;
1247
1248         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1249         if (rc < 0) {
1250                 if (likely(rc == -ENOENT))
1251                         rc = 0;
1252                 GOTO(out, rc);
1253         }
1254
1255         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1256         if (rc)
1257                 GOTO(out_close, rc);
1258         rc = llog_get_size(llh);
1259
1260 out_close:
1261         llog_close(env, llh);
1262 out:
1263         /* The header is record 1, the llog is still considered as empty
1264          * if there is only header */
1265         return (rc <= 1);
1266 }
1267 EXPORT_SYMBOL(llog_is_empty);
1268
1269 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1270                       struct llog_rec_hdr *rec, void *data)
1271 {
1272         struct llog_handle      *copy_llh = data;
1273
1274         /* Append all records */
1275         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1276 }
1277
1278 /* backup plain llog */
1279 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1280                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1281                 char *name, char *backup)
1282 {
1283         struct llog_handle      *llh, *bllh;
1284         int                      rc;
1285
1286         ENTRY;
1287
1288         /* open original log */
1289         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1290         if (rc < 0) {
1291                 /* the -ENOENT case is also reported to the caller
1292                  * but silently so it should handle that if needed.
1293                  */
1294                 if (rc != -ENOENT)
1295                         CERROR("%s: failed to open log %s: rc = %d\n",
1296                                obd->obd_name, name, rc);
1297                 RETURN(rc);
1298         }
1299
1300         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1301         if (rc)
1302                 GOTO(out_close, rc);
1303
1304         /* Make sure there's no old backup log */
1305         rc = llog_erase(env, bctxt, NULL, backup);
1306         if (rc < 0 && rc != -ENOENT)
1307                 GOTO(out_close, rc);
1308
1309         /* open backup log */
1310         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1311         if (rc) {
1312                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1313                        obd->obd_name, backup, rc);
1314                 GOTO(out_close, rc);
1315         }
1316
1317         /* check that backup llog is not the same object as original one */
1318         if (llh->lgh_obj == bllh->lgh_obj) {
1319                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1320                        obd->obd_name, name, backup, llh->lgh_obj,
1321                        bllh->lgh_obj);
1322                 GOTO(out_backup, rc = -EEXIST);
1323         }
1324
1325         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1326         if (rc)
1327                 GOTO(out_backup, rc);
1328
1329         /* Copy log record by record */
1330         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1331                                   NULL, false);
1332         if (rc)
1333                 CERROR("%s: failed to backup log %s: rc = %d\n",
1334                        obd->obd_name, name, rc);
1335 out_backup:
1336         llog_close(env, bllh);
1337 out_close:
1338         llog_close(env, llh);
1339         RETURN(rc);
1340 }
1341 EXPORT_SYMBOL(llog_backup);
1342
1343 /* Get size of llog */
1344 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1345 {
1346         int rc;
1347         struct lu_attr la;
1348
1349         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1350         if (rc) {
1351                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1352                        llh->lgh_ctxt->loc_obd->obd_name,
1353                        PFID(&llh->lgh_id.lgl_oi.oi_fid), rc);
1354                 return 0;
1355         }
1356
1357         return la.la_size;
1358 }
1359 EXPORT_SYMBOL(llog_size);
1360