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         atomic_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 void llog_handle_get(struct llog_handle *loghandle)
  95 {
  96         atomic_inc(&loghandle->lgh_refcount);
  97 }
  98
  99 void llog_handle_put(struct llog_handle *loghandle)
 100 {
 101         LASSERT(atomic_read(&loghandle->lgh_refcount) > 0);
 102         if (atomic_dec_and_test(&loghandle->lgh_refcount))
 103                 llog_free_handle(loghandle);
 104 }
 105
 106 static int llog_declare_destroy(const struct lu_env *env,
 107                                 struct llog_handle *handle,
 108                                 struct thandle *th)
 109 {
 110         struct llog_operations *lop;
 111         int rc;
 112
 113         ENTRY;
 114
 115         rc = llog_handle2ops(handle, &lop);
 116         if (rc)
 117                 RETURN(rc);
 118         if (lop->lop_declare_destroy == NULL)
 119                 RETURN(-EOPNOTSUPP);
 120
 121         rc = lop->lop_declare_destroy(env, handle, th);
 122
 123         RETURN(rc);
 124 }
 125
 126 int llog_trans_destroy(const struct lu_env *env, struct llog_handle *handle,
 127                        struct thandle *th)
 128 {
 129         struct llog_operations  *lop;
 130         int rc;
 131         ENTRY;
 132
 133         rc = llog_handle2ops(handle, &lop);
 134         if (rc < 0)
 135                 RETURN(rc);
 136         if (lop->lop_destroy == NULL)
 137                 RETURN(-EOPNOTSUPP);
 138
 139         LASSERT(handle->lgh_obj != NULL);
 140         if (!dt_object_exists(handle->lgh_obj))
 141                 RETURN(0);
 142
 143         rc = lop->lop_destroy(env, handle, th);
 144
 145         RETURN(rc);
 146 }
 147
 148 int llog_destroy(const struct lu_env *env, struct llog_handle *handle)
 149 {
 150         struct llog_operations  *lop;
 151         struct dt_device        *dt;
 152         struct thandle          *th;
 153         int rc;
 154
 155         ENTRY;
 156
 157         rc = llog_handle2ops(handle, &lop);
 158         if (rc < 0)
 159                 RETURN(rc);
 160         if (lop->lop_destroy == NULL)
 161                 RETURN(-EOPNOTSUPP);
 162
 163         if (handle->lgh_obj == NULL) {
 164                 /* if lgh_obj == NULL, then it is from client side destroy */
 165                 rc = lop->lop_destroy(env, handle, NULL);
 166                 RETURN(rc);
 167         }
 168
 169         if (!dt_object_exists(handle->lgh_obj))
 170                 RETURN(0);
 171
 172         dt = lu2dt_dev(handle->lgh_obj->do_lu.lo_dev);
 173
 174         if (unlikely(unlikely(dt->dd_rdonly)))
 175                 RETURN(-EROFS);
 176
 177         th = dt_trans_create(env, dt);
 178         if (IS_ERR(th))
 179                 RETURN(PTR_ERR(th));
 180
 181         rc = llog_declare_destroy(env, handle, th);
 182         if (rc != 0)
 183                 GOTO(out_trans, rc);
 184
 185         rc = dt_trans_start_local(env, dt, th);
 186         if (rc < 0)
 187                 GOTO(out_trans, rc);
 188
 189         rc = lop->lop_destroy(env, handle, th);
 190
 191 out_trans:
 192         dt_trans_stop(env, dt, th);
 193
 194         RETURN(rc);
 195 }
 196 EXPORT_SYMBOL(llog_destroy);
 197
 198 /* returns negative on error; 0 if success; 1 if success & log destroyed */
 199 int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
 200                     int index)
 201 {
 202         struct llog_thread_info *lgi = llog_info(env);
 203         struct dt_device        *dt;
 204         struct llog_log_hdr     *llh;
 205         struct thandle          *th;
 206         __u32                    tmp_lgc_index;
 207         int                      rc;
 208         int rc1;
 209         bool subtract_count = false;
 210
 211         ENTRY;
 212
 213         LASSERT(loghandle != NULL);
 214         LASSERT(loghandle->lgh_ctxt != NULL);
 215         LASSERT(loghandle->lgh_obj != NULL);
 216
 217         llh = loghandle->lgh_hdr;
 218
 219         CDEBUG(D_RPCTRACE, "Canceling %d in log "DFID"\n", index,
 220                PFID(&loghandle->lgh_id.lgl_oi.oi_fid));
 221
 222         if (index == 0) {
 223                 CERROR("Can't cancel index 0 which is header\n");
 224                 RETURN(-EINVAL);
 225         }
 226
 227         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 228
 229         if (unlikely(unlikely(dt->dd_rdonly)))
 230                 RETURN(0);
 231
 232         th = dt_trans_create(env, dt);
 233         if (IS_ERR(th))
 234                 RETURN(PTR_ERR(th));
 235
 236         rc = llog_declare_write_rec(env, loghandle, &llh->llh_hdr, index, th);
 237         if (rc < 0)
 238                 GOTO(out_trans, rc);
 239
 240         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY)) {
 241                 rc = llog_declare_destroy(env, loghandle, th);
 242                 if (rc < 0)
 243                         GOTO(out_trans, rc);
 244         }
 245
 246         th->th_wait_submit = 1;
 247         rc = dt_trans_start_local(env, dt, th);
 248         if (rc < 0)
 249                 GOTO(out_trans, rc);
 250
 251         down_write(&loghandle->lgh_lock);
 252         /* clear bitmap */
 253         mutex_lock(&loghandle->lgh_hdr_mutex);
 254         if (!ext2_clear_bit(index, LLOG_HDR_BITMAP(llh))) {
 255                 CDEBUG(D_RPCTRACE, "Catalog index %u already clear?\n", index);
 256                 GOTO(out_unlock, rc);
 257         }
 258
 259         loghandle->lgh_hdr->llh_count--;
 260         subtract_count = true;
 261
 262         /* Since llog_process_thread use lgi_cookie, it`s better to save them
 263          * and restore after using
 264          */
 265         tmp_lgc_index = lgi->lgi_cookie.lgc_index;
 266         /* Pass this index to llog_osd_write_rec(), which will use the index
 267          * to only update the necesary bitmap. */
 268         lgi->lgi_cookie.lgc_index = index;
 269         /* update header */
 270         rc = llog_write_rec(env, loghandle, &llh->llh_hdr, &lgi->lgi_cookie,
 271                             LLOG_HEADER_IDX, th);
 272         lgi->lgi_cookie.lgc_index = tmp_lgc_index;
 273
 274         if (rc != 0)
 275                 GOTO(out_unlock, rc);
 276
 277         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 278             (llh->llh_count == 1) &&
 279             ((loghandle->lgh_last_idx == LLOG_HDR_BITMAP_SIZE(llh) - 1) ||
 280              (loghandle->u.phd.phd_cat_handle != NULL &&
 281               loghandle->u.phd.phd_cat_handle->u.chd.chd_current_log !=
 282                 loghandle))) {
 283                 /* never try to destroy it again */
 284                 llh->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 285                 rc = llog_trans_destroy(env, loghandle, th);
 286                 if (rc < 0) {
 287                         /* Sigh, can not destroy the final plain llog, but
 288                          * the bitmap has been clearly, so the record can not
 289                          * be accessed anymore, let's return 0 for now, and
 290                          * the orphan will be handled by LFSCK. */
 291                         CERROR("%s: can't destroy empty llog "DFID": rc = %d\n",
 292                                loghandle->lgh_ctxt->loc_obd->obd_name,
 293                                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rc);
 294                         GOTO(out_unlock, rc = 0);
 295                 }
 296                 rc = LLOG_DEL_PLAIN;
 297         }
 298
 299 out_unlock:
 300         mutex_unlock(&loghandle->lgh_hdr_mutex);
 301         up_write(&loghandle->lgh_lock);
 302 out_trans:
 303         rc1 = dt_trans_stop(env, dt, th);
 304         if (rc == 0)
 305                 rc = rc1;
 306         if (rc < 0 && subtract_count) {
 307                 mutex_lock(&loghandle->lgh_hdr_mutex);
 308                 loghandle->lgh_hdr->llh_count++;
 309                 ext2_set_bit(index, LLOG_HDR_BITMAP(llh));
 310                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 311         }
 312         RETURN(rc);
 313 }
 314
 315 int llog_read_header(const struct lu_env *env, struct llog_handle *handle,
 316                      const struct obd_uuid *uuid)
 317 {
 318         struct llog_operations *lop;
 319         int rc;
 320         ENTRY;
 321
 322         rc = llog_handle2ops(handle, &lop);
 323         if (rc)
 324                 RETURN(rc);
 325
 326         if (lop->lop_read_header == NULL)
 327                 RETURN(-EOPNOTSUPP);
 328
 329         rc = lop->lop_read_header(env, handle);
 330         if (rc == LLOG_EEMPTY) {
 331                 struct llog_log_hdr *llh = handle->lgh_hdr;
 332
 333                 /* lrh_len should be initialized in llog_init_handle */
 334                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 335                 llh->llh_count = 1;         /* for the header record */
 336                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 337                 LASSERT(handle->lgh_ctxt->loc_chunk_size >=
 338                                                 LLOG_MIN_CHUNK_SIZE);
 339                 llh->llh_hdr.lrh_len = handle->lgh_ctxt->loc_chunk_size;
 340                 llh->llh_hdr.lrh_index = 0;
 341                 llh->llh_timestamp = ktime_get_real_seconds();
 342                 if (uuid)
 343                         memcpy(&llh->llh_tgtuuid, uuid,
 344                                sizeof(llh->llh_tgtuuid));
 345                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 346                 /* Since update llog header might also call this function,
 347                  * let's reset the bitmap to 0 here */
 348                 memset(LLOG_HDR_BITMAP(llh), 0, llh->llh_hdr.lrh_len -
 349                                                 llh->llh_bitmap_offset -
 350                                                 sizeof(llh->llh_tail));
 351                 ext2_set_bit(0, LLOG_HDR_BITMAP(llh));
 352                 LLOG_HDR_TAIL(llh)->lrt_len = llh->llh_hdr.lrh_len;
 353                 LLOG_HDR_TAIL(llh)->lrt_index = llh->llh_hdr.lrh_index;
 354                 rc = 0;
 355         }
 356         RETURN(rc);
 357 }
 358 EXPORT_SYMBOL(llog_read_header);
 359
 360 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 361                      int flags, struct obd_uuid *uuid)
 362 {
 363         struct llog_log_hdr     *llh;
 364         enum llog_flag           fmt = flags & LLOG_F_EXT_MASK;
 365         int                      rc;
 366         int                     chunk_size = handle->lgh_ctxt->loc_chunk_size;
 367         ENTRY;
 368
 369         LASSERT(handle->lgh_hdr == NULL);
 370
 371         LASSERT(chunk_size >= LLOG_MIN_CHUNK_SIZE);
 372         OBD_ALLOC_LARGE(llh, chunk_size);
 373         if (llh == NULL)
 374                 RETURN(-ENOMEM);
 375
 376         handle->lgh_hdr = llh;
 377         handle->lgh_hdr_size = chunk_size;
 378         /* first assign flags to use llog_client_ops */
 379         llh->llh_flags = flags;
 380         rc = llog_read_header(env, handle, uuid);
 381         if (rc == 0) {
 382                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 383                               flags & LLOG_F_IS_CAT) ||
 384                              (llh->llh_flags & LLOG_F_IS_CAT &&
 385                               flags & LLOG_F_IS_PLAIN))) {
 386                         CERROR("%s: llog type is %s but initializing %s\n",
 387                                handle->lgh_ctxt->loc_obd->obd_name,
 388                                llh->llh_flags & LLOG_F_IS_CAT ?
 389                                "catalog" : "plain",
 390                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 391                         GOTO(out, rc = -EINVAL);
 392                 } else if (llh->llh_flags &
 393                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 394                         /*
 395                          * it is possible to open llog without specifying llog
 396                          * type so it is taken from llh_flags
 397                          */
 398                         flags = llh->llh_flags;
 399                 } else {
 400                         /* for some reason the llh_flags has no type set */
 401                         CERROR("llog type is not specified!\n");
 402                         GOTO(out, rc = -EINVAL);
 403                 }
 404                 if (unlikely(uuid &&
 405                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 406                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 407                                handle->lgh_ctxt->loc_obd->obd_name,
 408                                (char *)uuid->uuid,
 409                                (char *)llh->llh_tgtuuid.uuid);
 410                         GOTO(out, rc = -EEXIST);
 411                 }
 412         }
 413         if (flags & LLOG_F_IS_CAT) {
 414                 LASSERT(list_empty(&handle->u.chd.chd_head));
 415                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 416                 llh->llh_size = sizeof(struct llog_logid_rec);
 417                 llh->llh_flags |= LLOG_F_IS_FIXSIZE;
 418         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 419                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 420                        handle->lgh_ctxt->loc_obd->obd_name,
 421                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 422                 rc = -EINVAL;
 423         }
 424         llh->llh_flags |= fmt;
 425 out:
 426         if (rc) {
 427                 OBD_FREE_LARGE(llh, chunk_size);
 428                 handle->lgh_hdr = NULL;
 429         }
 430         RETURN(rc);
 431 }
 432 EXPORT_SYMBOL(llog_init_handle);
 433
 434 static int llog_process_thread(void *arg)
 435 {
 436         struct llog_process_info        *lpi = arg;
 437         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 438         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 439         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 440         struct llog_thread_info         *lti;
 441         char                            *buf;
 442         size_t                           chunk_size;
 443         __u64                            cur_offset;
 444         int                              rc = 0, index = 1, last_index;
 445         int                              saved_index = 0;
 446         int                              last_called_index = 0;
 447         bool                             repeated = false;
 448
 449         ENTRY;
 450
 451         if (llh == NULL)
 452                 RETURN(-EINVAL);
 453
 454         lti = lpi->lpi_env == NULL ? NULL : llog_info(lpi->lpi_env);
 455
 456         cur_offset = chunk_size = llh->llh_hdr.lrh_len;
 457         /* expect chunk_size to be power of two */
 458         LASSERT(is_power_of_2(chunk_size));
 459
 460         OBD_ALLOC_LARGE(buf, chunk_size);
 461         if (buf == NULL) {
 462                 lpi->lpi_rc = -ENOMEM;
 463                 RETURN(0);
 464         }
 465
 466         if (cd != NULL) {
 467                 last_called_index = cd->lpcd_first_idx;
 468                 index = cd->lpcd_first_idx + 1;
 469         }
 470         if (cd != NULL && cd->lpcd_last_idx)
 471                 last_index = cd->lpcd_last_idx;
 472         else
 473                 last_index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 474
 475         while (rc == 0) {
 476                 struct llog_rec_hdr *rec;
 477                 off_t chunk_offset = 0;
 478                 unsigned int buf_offset = 0;
 479                 bool partial_chunk;
 480                 int     lh_last_idx;
 481                 int     synced_idx = 0;
 482
 483                 /* skip records not set in bitmap */
 484                 while (index <= last_index &&
 485                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 486                         ++index;
 487
 488                 /* There are no indices prior the last_index */
 489                 if (index > last_index)
 490                         break;
 491
 492                 CDEBUG(D_OTHER, "index: %d last_index %d\n", index,
 493                        last_index);
 494
 495 repeat:
 496                 /* get the buf with our target record; avoid old garbage */
 497                 memset(buf, 0, chunk_size);
 498                 /* the record index for outdated chunk data */
 499                 /* it is safe to process buffer until saved lgh_last_idx */
 500                 lh_last_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 501                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 502                                      index, &cur_offset, buf, chunk_size);
 503                 if (repeated && rc)
 504                         CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
 505                                " buf_offset %u, rc = %d\n", cur_offset,
 506                                (__u64)chunk_offset, buf_offset, rc);
 507                 /* we`ve tried to reread the chunk, but there is no
 508                  * new records */
 509                 if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
 510                     cur_offset)
 511                         GOTO(out, rc = 0);
 512                 if (rc != 0)
 513                         GOTO(out, rc);
 514
 515                 /* NB: after llog_next_block() call the cur_offset is the
 516                  * offset of the next block after read one.
 517                  * The absolute offset of the current chunk is calculated
 518                  * from cur_offset value and stored in chunk_offset variable.
 519                  */
 520                 if ((cur_offset & (chunk_size - 1)) != 0) {
 521                         partial_chunk = true;
 522                         chunk_offset = cur_offset & ~(chunk_size - 1);
 523                 } else {
 524                         partial_chunk = false;
 525                         chunk_offset = cur_offset - chunk_size;
 526                 }
 527
 528                 /* NB: when rec->lrh_len is accessed it is already swabbed
 529                  * since it is used at the "end" of the loop and the rec
 530                  * swabbing is done at the beginning of the loop. */
 531                 for (rec = (struct llog_rec_hdr *)(buf + buf_offset);
 532                      (char *)rec < buf + chunk_size;
 533                      rec = llog_rec_hdr_next(rec)) {
 534
 535                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 536                                rec, rec->lrh_type);
 537
 538                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 539                                 lustre_swab_llog_rec(rec);
 540
 541                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 542                                rec->lrh_type, rec->lrh_index);
 543
 544                         if (index == (synced_idx + 1) &&
 545                             synced_idx == LLOG_HDR_TAIL(llh)->lrt_index)
 546                                 GOTO(out, rc = 0);
 547
 548                         /* the bitmap could be changed during processing
 549                          * records from the chunk. For wrapped catalog
 550                          * it means we can read deleted record and try to
 551                          * process it. Check this case and reread the chunk.
 552                          * It is safe to process to lh_last_idx, including
 553                          * lh_last_idx if it was synced. We can not do <=
 554                          * comparison, cause for wrapped catalog lgh_last_idx
 555                          * could be less than index. So we detect last index
 556                          * for processing as index == lh_last_idx+1. But when
 557                          * catalog is wrapped and full lgh_last_idx=llh_cat_idx,
 558                          * the first processing index is llh_cat_idx+1.
 559                          */
 560
 561                         if ((index == lh_last_idx && synced_idx != index) ||
 562                             (index == (lh_last_idx + 1) &&
 563                              !(index == (llh->llh_cat_idx + 1) &&
 564                                (llh->llh_flags & LLOG_F_IS_CAT))) ||
 565                             (rec->lrh_index == 0 && !repeated)) {
 566
 567                                 /* save offset inside buffer for the re-read */
 568                                 buf_offset = (char *)rec - (char *)buf;
 569                                 cur_offset = chunk_offset;
 570                                 repeated = true;
 571                                 /* We need to be sure lgh_last_idx
 572                                  * record was saved to disk
 573                                  */
 574                                 down_read(&loghandle->lgh_last_sem);
 575                                 synced_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 576                                 up_read(&loghandle->lgh_last_sem);
 577                                 CDEBUG(D_OTHER, "synced_idx: %d\n", synced_idx);
 578                                 goto repeat;
 579
 580                         }
 581
 582                         repeated = false;
 583
 584                         if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
 585                                 CWARN("%s: invalid length %d in llog "DFID
 586                                       "record for index %d/%d\n",
 587                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 588                                        rec->lrh_len,
 589                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 590                                        rec->lrh_index, index);
 591
 592                                 GOTO(out, rc = -EINVAL);
 593                         }
 594
 595                         if (rec->lrh_index < index) {
 596                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 597                                        rec->lrh_index);
 598                                 continue;
 599                         }
 600
 601                         if (rec->lrh_index != index) {
 602                                 CERROR("%s: "DFID" Invalid record: index %u"
 603                                        " but expected %u\n",
 604                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 605                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 606                                        rec->lrh_index, index);
 607                                 GOTO(out, rc = -ERANGE);
 608                         }
 609
 610                         CDEBUG(D_OTHER,
 611                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 612                                rec->lrh_index, rec->lrh_len,
 613                                (int)(buf + chunk_size - (char *)rec));
 614
 615                         /* lgh_cur_offset is used only at llog_test_3 */
 616                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 617                                                     chunk_offset;
 618
 619                         if (OBD_FAIL_PRECHECK(OBD_FAIL_LLOG_PROCESS_TIMEOUT) &&
 620                                 index == lh_last_idx &&
 621                                 cfs_fail_val == (unsigned int)
 622                                         (loghandle->lgh_id.lgl_oi.oi.oi_id &
 623                                          0xFFFFFFFF)) {
 624                                 OBD_RACE(OBD_FAIL_LLOG_PROCESS_TIMEOUT);
 625                         }
 626                         /* if set, process the callback on this record */
 627                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 628                                 struct llog_cookie *lgc;
 629                                 __u64   tmp_off;
 630                                 int     tmp_idx;
 631
 632                                 CDEBUG(D_OTHER, "index: %d, lh_last_idx: %d "
 633                                        "synced_idx: %d lgh_last_idx: %d\n",
 634                                        index, lh_last_idx, synced_idx,
 635                                        loghandle->lgh_last_idx);
 636
 637                                 if (lti != NULL) {
 638                                         lgc = &lti->lgi_cookie;
 639                                         /* store lu_env for recursive calls */
 640                                         tmp_off = lgc->lgc_offset;
 641                                         tmp_idx = lgc->lgc_index;
 642
 643                                         lgc->lgc_offset = (char *)rec -
 644                                                 (char *)buf + chunk_offset;
 645                                         lgc->lgc_index = rec->lrh_index;
 646                                 }
 647                                 /* using lu_env for passing record offset to
 648                                  * llog_write through various callbacks */
 649                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 650                                                  lpi->lpi_cbdata);
 651                                 last_called_index = index;
 652
 653                                 if (lti != NULL) {
 654                                         lgc->lgc_offset = tmp_off;
 655                                         lgc->lgc_index = tmp_idx;
 656                                 }
 657
 658                                 if (rc == LLOG_PROC_BREAK) {
 659                                         GOTO(out, rc);
 660                                 } else if (rc == LLOG_DEL_RECORD) {
 661                                         rc = llog_cancel_rec(lpi->lpi_env,
 662                                                              loghandle,
 663                                                              rec->lrh_index);
 664                                 }
 665                                 if (rc)
 666                                         GOTO(out, rc);
 667                                 /* some stupid callbacks directly cancel records
 668                                  * and delete llog. Check it and stop
 669                                  * processing. */
 670                                 if (loghandle->lgh_hdr == NULL ||
 671                                     loghandle->lgh_hdr->llh_count == 1)
 672                                         GOTO(out, rc = 0);
 673                         }
 674                         /* exit if the last index is reached */
 675                         if (index >= last_index)
 676                                 GOTO(out, rc = 0);
 677                         ++index;
 678                 }
 679         }
 680
 681 out:
 682         if (cd != NULL)
 683                 cd->lpcd_last_idx = last_called_index;
 684
 685         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 686                 if (dt_object_remote(loghandle->lgh_obj)) {
 687                         /* If it is remote object, then -EIO might means
 688                          * disconnection or eviction, let's return -EAGAIN,
 689                          * so for update recovery log processing, it will
 690                          * retry until the umount or abort recovery, see
 691                          * lod_sub_recovery_thread() */
 692                         CERROR("%s retry remote llog process\n",
 693                                loghandle->lgh_ctxt->loc_obd->obd_name);
 694                         rc = -EAGAIN;
 695                 } else {
 696                         /* something bad happened to the processing of a local
 697                          * llog file, probably I/O error or the log got
 698                          * corrupted to be able to finally release the log we
 699                          * discard any remaining bits in the header */
 700                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 701                                " %s index %d count %d\n",
 702                                loghandle->lgh_ctxt->loc_obd->obd_name,
 703                                POSTID(&loghandle->lgh_id.lgl_oi),
 704                                loghandle->lgh_id.lgl_ogen,
 705                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 706                                 "plain"), index, llh->llh_count);
 707
 708                         while (index <= last_index) {
 709                                 if (ext2_test_bit(index,
 710                                                   LLOG_HDR_BITMAP(llh)) != 0)
 711                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 712                                                         index);
 713                                 index++;
 714                         }
 715                         rc = 0;
 716                 }
 717         }
 718
 719         OBD_FREE_LARGE(buf, chunk_size);
 720         lpi->lpi_rc = rc;
 721         return 0;
 722 }
 723
 724 static int llog_process_thread_daemonize(void *arg)
 725 {
 726         struct llog_process_info        *lpi = arg;
 727         struct lu_env                    env;
 728         int                              rc;
 729         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 730
 731         task_lock(lpi->lpi_reftask);
 732         new_ns = lpi->lpi_reftask->nsproxy;
 733         if (curr_ns != new_ns) {
 734                 get_nsproxy(new_ns);
 735
 736                 current->nsproxy = new_ns;
 737                 /* XXX: we should call put_nsproxy() instead of
 738                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 739                  * used outside of the kernel itself, because it calls
 740                  * free_nsproxy() which is not exported by the kernel
 741                  * (defined in kernel/nsproxy.c) */
 742                 atomic_dec(&curr_ns->count);
 743         }
 744         task_unlock(lpi->lpi_reftask);
 745
 746         unshare_fs_struct();
 747
 748         /* client env has no keys, tags is just 0 */
 749         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 750         if (rc)
 751                 goto out;
 752         lpi->lpi_env = &env;
 753
 754         rc = llog_process_thread(arg);
 755
 756         lu_env_fini(&env);
 757 out:
 758         complete(&lpi->lpi_completion);
 759         return rc;
 760 }
 761
 762 int llog_process_or_fork(const struct lu_env *env,
 763                          struct llog_handle *loghandle,
 764                          llog_cb_t cb, void *data, void *catdata, bool fork)
 765 {
 766         struct llog_process_info *lpi;
 767         int                      rc;
 768
 769         ENTRY;
 770
 771         OBD_ALLOC_PTR(lpi);
 772         if (lpi == NULL) {
 773                 CERROR("cannot alloc pointer\n");
 774                 RETURN(-ENOMEM);
 775         }
 776         lpi->lpi_loghandle = loghandle;
 777         lpi->lpi_cb        = cb;
 778         lpi->lpi_cbdata    = data;
 779         lpi->lpi_catdata   = catdata;
 780
 781         if (fork) {
 782                 struct task_struct *task;
 783
 784                 /* The new thread can't use parent env,
 785                  * init the new one in llog_process_thread_daemonize. */
 786                 lpi->lpi_env = NULL;
 787                 init_completion(&lpi->lpi_completion);
 788                 /* take reference to current, so that
 789                  * llog_process_thread_daemonize() can use it to switch to
 790                  * namespace associated with current  */
 791                 lpi->lpi_reftask = current;
 792                 task = kthread_run(llog_process_thread_daemonize, lpi,
 793                                    "llog_process_thread");
 794                 if (IS_ERR(task)) {
 795                         rc = PTR_ERR(task);
 796                         CERROR("%s: cannot start thread: rc = %d\n",
 797                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 798                         GOTO(out_lpi, rc);
 799                 }
 800                 wait_for_completion(&lpi->lpi_completion);
 801         } else {
 802                 lpi->lpi_env = env;
 803                 llog_process_thread(lpi);
 804         }
 805         rc = lpi->lpi_rc;
 806
 807 out_lpi:
 808         OBD_FREE_PTR(lpi);
 809         RETURN(rc);
 810 }
 811 EXPORT_SYMBOL(llog_process_or_fork);
 812
 813 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 814                  llog_cb_t cb, void *data, void *catdata)
 815 {
 816         int rc;
 817         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 818         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 819 }
 820 EXPORT_SYMBOL(llog_process);
 821
 822 int llog_reverse_process(const struct lu_env *env,
 823                          struct llog_handle *loghandle, llog_cb_t cb,
 824                          void *data, void *catdata)
 825 {
 826         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 827         struct llog_process_cat_data *cd = catdata;
 828         void *buf;
 829         int rc = 0, first_index = 1, index, idx;
 830         __u32   chunk_size = llh->llh_hdr.lrh_len;
 831         ENTRY;
 832
 833         OBD_ALLOC_LARGE(buf, chunk_size);
 834         if (buf == NULL)
 835                 RETURN(-ENOMEM);
 836
 837         if (cd != NULL)
 838                 first_index = cd->lpcd_first_idx + 1;
 839         if (cd != NULL && cd->lpcd_last_idx)
 840                 index = cd->lpcd_last_idx;
 841         else
 842                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 843
 844         while (rc == 0) {
 845                 struct llog_rec_hdr *rec;
 846                 struct llog_rec_tail *tail;
 847
 848                 /* skip records not set in bitmap */
 849                 while (index >= first_index &&
 850                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 851                         --index;
 852
 853                 LASSERT(index >= first_index - 1);
 854                 if (index == first_index - 1)
 855                         break;
 856
 857                 /* get the buf with our target record; avoid old garbage */
 858                 memset(buf, 0, chunk_size);
 859                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 860                 if (rc)
 861                         GOTO(out, rc);
 862
 863                 rec = buf;
 864                 idx = rec->lrh_index;
 865                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 866                 while (idx < index) {
 867                         rec = (void *)rec + rec->lrh_len;
 868                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 869                                 lustre_swab_llog_rec(rec);
 870                         idx ++;
 871                 }
 872                 LASSERT(idx == index);
 873                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 874
 875                 /* process records in buffer, starting where we found one */
 876                 while ((void *)tail > buf) {
 877                         if (tail->lrt_index == 0)
 878                                 GOTO(out, rc = 0); /* no more records */
 879
 880                         /* if set, process the callback on this record */
 881                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 882                                 rec = (void *)tail - tail->lrt_len +
 883                                       sizeof(*tail);
 884
 885                                 rc = cb(env, loghandle, rec, data);
 886                                 if (rc == LLOG_PROC_BREAK) {
 887                                         GOTO(out, rc);
 888                                 } else if (rc == LLOG_DEL_RECORD) {
 889                                         rc = llog_cancel_rec(env, loghandle,
 890                                                              tail->lrt_index);
 891                                 }
 892                                 if (rc)
 893                                         GOTO(out, rc);
 894                         }
 895
 896                         /* previous record, still in buffer? */
 897                         --index;
 898                         if (index < first_index)
 899                                 GOTO(out, rc = 0);
 900                         tail = (void *)tail - tail->lrt_len;
 901                 }
 902         }
 903
 904 out:
 905         if (buf != NULL)
 906                 OBD_FREE_LARGE(buf, chunk_size);
 907         RETURN(rc);
 908 }
 909 EXPORT_SYMBOL(llog_reverse_process);
 910
 911 /**
 912  * new llog API
 913  *
 914  * API functions:
 915  *      llog_open - open llog, may not exist
 916  *      llog_exist - check if llog exists
 917  *      llog_close - close opened llog, pair for open, frees llog_handle
 918  *      llog_declare_create - declare llog creation
 919  *      llog_create - create new llog on disk, need transaction handle
 920  *      llog_declare_write_rec - declaration of llog write
 921  *      llog_write_rec - write llog record on disk, need transaction handle
 922  *      llog_declare_add - declare llog catalog record addition
 923  *      llog_add - add llog record in catalog, need transaction handle
 924  */
 925 int llog_exist(struct llog_handle *loghandle)
 926 {
 927         struct llog_operations  *lop;
 928         int                      rc;
 929
 930         ENTRY;
 931
 932         rc = llog_handle2ops(loghandle, &lop);
 933         if (rc)
 934                 RETURN(rc);
 935         if (lop->lop_exist == NULL)
 936                 RETURN(-EOPNOTSUPP);
 937
 938         rc = lop->lop_exist(loghandle);
 939         RETURN(rc);
 940 }
 941 EXPORT_SYMBOL(llog_exist);
 942
 943 int llog_declare_create(const struct lu_env *env,
 944                         struct llog_handle *loghandle, struct thandle *th)
 945 {
 946         struct llog_operations  *lop;
 947         int                      raised, rc;
 948
 949         ENTRY;
 950
 951         rc = llog_handle2ops(loghandle, &lop);
 952         if (rc)
 953                 RETURN(rc);
 954         if (lop->lop_declare_create == NULL)
 955                 RETURN(-EOPNOTSUPP);
 956
 957         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 958         if (!raised)
 959                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 960         rc = lop->lop_declare_create(env, loghandle, th);
 961         if (!raised)
 962                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 963         RETURN(rc);
 964 }
 965
 966 int llog_create(const struct lu_env *env, struct llog_handle *handle,
 967                 struct thandle *th)
 968 {
 969         struct llog_operations  *lop;
 970         int                      raised, rc;
 971
 972         ENTRY;
 973
 974         rc = llog_handle2ops(handle, &lop);
 975         if (rc)
 976                 RETURN(rc);
 977         if (lop->lop_create == NULL)
 978                 RETURN(-EOPNOTSUPP);
 979
 980         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 981         if (!raised)
 982                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 983         rc = lop->lop_create(env, handle, th);
 984         if (!raised)
 985                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 986         RETURN(rc);
 987 }
 988
 989 int llog_declare_write_rec(const struct lu_env *env,
 990                            struct llog_handle *handle,
 991                            struct llog_rec_hdr *rec, int idx,
 992                            struct thandle *th)
 993 {
 994         struct llog_operations  *lop;
 995         int                      raised, rc;
 996
 997         ENTRY;
 998
 999         rc = llog_handle2ops(handle, &lop);
1000         if (rc)
1001                 RETURN(rc);
1002         LASSERT(lop);
1003         if (lop->lop_declare_write_rec == NULL)
1004                 RETURN(-EOPNOTSUPP);
1005
1006         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1007         if (!raised)
1008                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1009         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
1010         if (!raised)
1011                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1012         RETURN(rc);
1013 }
1014
1015 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
1016                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1017                    int idx, struct thandle *th)
1018 {
1019         struct llog_operations  *lop;
1020         int                      raised, rc, buflen;
1021
1022         ENTRY;
1023
1024         /* API sanity checks */
1025         if (handle == NULL) {
1026                 CERROR("loghandle is missed\n");
1027                 RETURN(-EPROTO);
1028         } else if (handle->lgh_obj == NULL) {
1029                 CERROR("loghandle %p with NULL object\n",
1030                         handle);
1031                 RETURN(-EPROTO);
1032         } else if (th == NULL) {
1033                 CERROR("%s: missed transaction handle\n",
1034                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name);
1035                 RETURN(-EPROTO);
1036         } else if (handle->lgh_hdr == NULL) {
1037                 CERROR("%s: loghandle %p with no header\n",
1038                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name,
1039                         handle);
1040                 RETURN(-EPROTO);
1041         }
1042
1043         rc = llog_handle2ops(handle, &lop);
1044         if (rc)
1045                 RETURN(rc);
1046
1047         if (lop->lop_write_rec == NULL)
1048                 RETURN(-EOPNOTSUPP);
1049
1050         buflen = rec->lrh_len;
1051         LASSERT(cfs_size_round(buflen) == buflen);
1052
1053         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1054         if (!raised)
1055                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1056         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1057         if (!raised)
1058                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1059         RETURN(rc);
1060 }
1061
1062 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1063              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1064              struct thandle *th)
1065 {
1066         int raised, rc;
1067
1068         ENTRY;
1069
1070         if (lgh->lgh_logops->lop_add == NULL)
1071                 RETURN(-EOPNOTSUPP);
1072
1073         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1074         if (!raised)
1075                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1076         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1077         if (!raised)
1078                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1079         RETURN(rc);
1080 }
1081 EXPORT_SYMBOL(llog_add);
1082
1083 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1084                      struct llog_rec_hdr *rec, struct thandle *th)
1085 {
1086         int raised, rc;
1087
1088         ENTRY;
1089
1090         if (lgh->lgh_logops->lop_declare_add == NULL)
1091                 RETURN(-EOPNOTSUPP);
1092
1093         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1094         if (!raised)
1095                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1096         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1097         if (!raised)
1098                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1099         RETURN(rc);
1100 }
1101 EXPORT_SYMBOL(llog_declare_add);
1102
1103 /**
1104  * Helper function to open llog or create it if doesn't exist.
1105  * It hides all transaction handling from caller.
1106  */
1107 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1108                      struct llog_handle **res, struct llog_logid *logid,
1109                      char *name)
1110 {
1111         struct dt_device        *d;
1112         struct thandle          *th;
1113         int                      rc;
1114
1115         ENTRY;
1116
1117         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1118         if (rc)
1119                 RETURN(rc);
1120
1121         if (llog_exist(*res))
1122                 RETURN(0);
1123
1124         LASSERT((*res)->lgh_obj != NULL);
1125
1126         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1127
1128         if (unlikely(unlikely(d->dd_rdonly)))
1129                 RETURN(-EROFS);
1130
1131         th = dt_trans_create(env, d);
1132         if (IS_ERR(th))
1133                 GOTO(out, rc = PTR_ERR(th));
1134
1135         /* Create update llog object synchronously, which
1136          * happens during inialization process see
1137          * lod_sub_prep_llog(), to make sure the update
1138          * llog object is created before corss-MDT writing
1139          * updates into the llog object */
1140         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1141                 th->th_sync = 1;
1142
1143         th->th_wait_submit = 1;
1144         rc = llog_declare_create(env, *res, th);
1145         if (rc == 0) {
1146                 rc = dt_trans_start_local(env, d, th);
1147                 if (rc == 0)
1148                         rc = llog_create(env, *res, th);
1149         }
1150         dt_trans_stop(env, d, th);
1151 out:
1152         if (rc)
1153                 llog_close(env, *res);
1154         RETURN(rc);
1155 }
1156 EXPORT_SYMBOL(llog_open_create);
1157
1158 /**
1159  * Helper function to delete existent llog.
1160  */
1161 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1162                struct llog_logid *logid, char *name)
1163 {
1164         struct llog_handle      *handle;
1165         int                      rc = 0, rc2;
1166
1167         ENTRY;
1168
1169         /* nothing to erase */
1170         if (name == NULL && logid == NULL)
1171                 RETURN(0);
1172
1173         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1174         if (rc < 0)
1175                 RETURN(rc);
1176
1177         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1178         if (rc == 0)
1179                 rc = llog_destroy(env, handle);
1180
1181         rc2 = llog_close(env, handle);
1182         if (rc == 0)
1183                 rc = rc2;
1184         RETURN(rc);
1185 }
1186 EXPORT_SYMBOL(llog_erase);
1187
1188 /*
1189  * Helper function for write record in llog.
1190  * It hides all transaction handling from caller.
1191  * Valid only with local llog.
1192  */
1193 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1194                struct llog_rec_hdr *rec, int idx)
1195 {
1196         struct dt_device        *dt;
1197         struct thandle          *th;
1198         bool                    need_cookie;
1199         int                     rc;
1200
1201         ENTRY;
1202
1203         LASSERT(loghandle);
1204         LASSERT(loghandle->lgh_ctxt);
1205         LASSERT(loghandle->lgh_obj != NULL);
1206
1207         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1208
1209         if (unlikely(unlikely(dt->dd_rdonly)))
1210                 RETURN(-EROFS);
1211
1212         th = dt_trans_create(env, dt);
1213         if (IS_ERR(th))
1214                 RETURN(PTR_ERR(th));
1215
1216         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1217         if (rc)
1218                 GOTO(out_trans, rc);
1219
1220         th->th_wait_submit = 1;
1221         rc = dt_trans_start_local(env, dt, th);
1222         if (rc)
1223                 GOTO(out_trans, rc);
1224
1225         need_cookie = !(idx == LLOG_HEADER_IDX || idx == LLOG_NEXT_IDX);
1226
1227         down_write(&loghandle->lgh_lock);
1228         if (need_cookie) {
1229                 struct llog_thread_info *lti = llog_info(env);
1230
1231                 /* cookie comes from llog_process_thread */
1232                 rc = llog_write_rec(env, loghandle, rec, &lti->lgi_cookie,
1233                                     rec->lrh_index, th);
1234                 /* upper layer didn`t pass cookie so change rc */
1235                 rc = (rc == 1 ? 0 : rc);
1236         } else {
1237                 rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1238         }
1239
1240         up_write(&loghandle->lgh_lock);
1241 out_trans:
1242         dt_trans_stop(env, dt, th);
1243         RETURN(rc);
1244 }
1245 EXPORT_SYMBOL(llog_write);
1246
1247 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1248               struct llog_handle **lgh, struct llog_logid *logid,
1249               char *name, enum llog_open_param open_param)
1250 {
1251         int      raised;
1252         int      rc;
1253
1254         ENTRY;
1255
1256         LASSERT(ctxt);
1257         LASSERT(ctxt->loc_logops);
1258
1259         if (ctxt->loc_logops->lop_open == NULL) {
1260                 *lgh = NULL;
1261                 RETURN(-EOPNOTSUPP);
1262         }
1263
1264         *lgh = llog_alloc_handle();
1265         if (*lgh == NULL)
1266                 RETURN(-ENOMEM);
1267         (*lgh)->lgh_ctxt = ctxt;
1268         (*lgh)->lgh_logops = ctxt->loc_logops;
1269
1270         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1271         if (!raised)
1272                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1273         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1274         if (!raised)
1275                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1276         if (rc) {
1277                 llog_free_handle(*lgh);
1278                 *lgh = NULL;
1279         }
1280         RETURN(rc);
1281 }
1282 EXPORT_SYMBOL(llog_open);
1283
1284 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1285 {
1286         struct llog_operations  *lop;
1287         int                      rc;
1288
1289         ENTRY;
1290
1291         rc = llog_handle2ops(loghandle, &lop);
1292         if (rc)
1293                 GOTO(out, rc);
1294         if (lop->lop_close == NULL)
1295                 GOTO(out, rc = -EOPNOTSUPP);
1296         rc = lop->lop_close(env, loghandle);
1297 out:
1298         llog_handle_put(loghandle);
1299         RETURN(rc);
1300 }
1301 EXPORT_SYMBOL(llog_close);
1302
1303 /**
1304  * Helper function to get the llog size in records. It is used by MGS
1305  * mostly to check that config llog exists and contains data.
1306  *
1307  * \param[in] env       execution environment
1308  * \param[in] ctxt      llog context
1309  * \param[in] name      llog name
1310  *
1311  * \retval              true if there are records in llog besides a header
1312  * \retval              false on error or llog without records
1313  */
1314 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1315                   char *name)
1316 {
1317         struct llog_handle      *llh;
1318         int                      rc = 0;
1319
1320         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1321         if (rc < 0) {
1322                 if (likely(rc == -ENOENT))
1323                         rc = 0;
1324                 GOTO(out, rc);
1325         }
1326
1327         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1328         if (rc)
1329                 GOTO(out_close, rc);
1330         rc = llog_get_size(llh);
1331
1332 out_close:
1333         llog_close(env, llh);
1334 out:
1335         /* The header is record 1, the llog is still considered as empty
1336          * if there is only header */
1337         return (rc <= 1);
1338 }
1339 EXPORT_SYMBOL(llog_is_empty);
1340
1341 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1342                       struct llog_rec_hdr *rec, void *data)
1343 {
1344         struct llog_handle      *copy_llh = data;
1345
1346         /* Append all records */
1347         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1348 }
1349
1350 /* backup plain llog */
1351 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1352                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1353                 char *name, char *backup)
1354 {
1355         struct llog_handle      *llh, *bllh;
1356         int                      rc;
1357
1358         ENTRY;
1359
1360         /* open original log */
1361         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1362         if (rc < 0) {
1363                 /* the -ENOENT case is also reported to the caller
1364                  * but silently so it should handle that if needed.
1365                  */
1366                 if (rc != -ENOENT)
1367                         CERROR("%s: failed to open log %s: rc = %d\n",
1368                                obd->obd_name, name, rc);
1369                 RETURN(rc);
1370         }
1371
1372         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1373         if (rc)
1374                 GOTO(out_close, rc);
1375
1376         /* Make sure there's no old backup log */
1377         rc = llog_erase(env, bctxt, NULL, backup);
1378         if (rc < 0 && rc != -ENOENT)
1379                 GOTO(out_close, rc);
1380
1381         /* open backup log */
1382         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1383         if (rc) {
1384                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1385                        obd->obd_name, backup, rc);
1386                 GOTO(out_close, rc);
1387         }
1388
1389         /* check that backup llog is not the same object as original one */
1390         if (llh->lgh_obj == bllh->lgh_obj) {
1391                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1392                        obd->obd_name, name, backup, llh->lgh_obj,
1393                        bllh->lgh_obj);
1394                 GOTO(out_backup, rc = -EEXIST);
1395         }
1396
1397         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1398         if (rc)
1399                 GOTO(out_backup, rc);
1400
1401         /* Copy log record by record */
1402         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1403                                   NULL, false);
1404         if (rc)
1405                 CERROR("%s: failed to backup log %s: rc = %d\n",
1406                        obd->obd_name, name, rc);
1407 out_backup:
1408         llog_close(env, bllh);
1409 out_close:
1410         llog_close(env, llh);
1411         RETURN(rc);
1412 }
1413 EXPORT_SYMBOL(llog_backup);
1414
1415 /* Get size of llog */
1416 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1417 {
1418         int rc;
1419         struct lu_attr la;
1420
1421         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1422         if (rc) {
1423                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1424                        llh->lgh_ctxt->loc_obd->obd_name,
1425                        PFID(&llh->lgh_id.lgl_oi.oi_fid), rc);
1426                 return 0;
1427         }
1428
1429         return la.la_size;
1430 }
1431 EXPORT_SYMBOL(llog_size);
1432