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_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;
 200         struct thandle          *th;
 201         __u32                    tmp_lgc_index;
 202         int                      rc;
 203         int rc1;
 204         bool subtract_count = false;
 205
 206         ENTRY;
 207
 208         LASSERT(loghandle != NULL);
 209         LASSERT(loghandle->lgh_ctxt != NULL);
 210         LASSERT(loghandle->lgh_obj != NULL);
 211
 212         llh = loghandle->lgh_hdr;
 213
 214         CDEBUG(D_RPCTRACE, "Canceling %d in log "DFID"\n", index,
 215                PFID(&loghandle->lgh_id.lgl_oi.oi_fid));
 216
 217         if (index == 0) {
 218                 CERROR("Can't cancel index 0 which is header\n");
 219                 RETURN(-EINVAL);
 220         }
 221
 222         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 223
 224         th = dt_trans_create(env, dt);
 225         if (IS_ERR(th))
 226                 RETURN(PTR_ERR(th));
 227
 228         rc = llog_declare_write_rec(env, loghandle, &llh->llh_hdr, index, th);
 229         if (rc < 0)
 230                 GOTO(out_trans, rc);
 231
 232         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY)) {
 233                 rc = llog_declare_destroy(env, loghandle, th);
 234                 if (rc < 0)
 235                         GOTO(out_trans, rc);
 236         }
 237
 238         th->th_wait_submit = 1;
 239         rc = dt_trans_start_local(env, dt, th);
 240         if (rc < 0)
 241                 GOTO(out_trans, rc);
 242
 243         down_write(&loghandle->lgh_lock);
 244         /* clear bitmap */
 245         mutex_lock(&loghandle->lgh_hdr_mutex);
 246         if (!ext2_clear_bit(index, LLOG_HDR_BITMAP(llh))) {
 247                 CDEBUG(D_RPCTRACE, "Catalog index %u already clear?\n", index);
 248                 GOTO(out_unlock, rc);
 249         }
 250
 251         loghandle->lgh_hdr->llh_count--;
 252         subtract_count = true;
 253
 254         /* Since llog_process_thread use lgi_cookie, it`s better to save them
 255          * and restore after using
 256          */
 257         tmp_lgc_index = lgi->lgi_cookie.lgc_index;
 258         /* Pass this index to llog_osd_write_rec(), which will use the index
 259          * to only update the necesary bitmap. */
 260         lgi->lgi_cookie.lgc_index = index;
 261         /* update header */
 262         rc = llog_write_rec(env, loghandle, &llh->llh_hdr, &lgi->lgi_cookie,
 263                             LLOG_HEADER_IDX, th);
 264         lgi->lgi_cookie.lgc_index = tmp_lgc_index;
 265
 266         if (rc != 0)
 267                 GOTO(out_unlock, rc);
 268
 269         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 270             (llh->llh_count == 1) &&
 271             ((loghandle->lgh_last_idx == LLOG_HDR_BITMAP_SIZE(llh) - 1) ||
 272              (loghandle->u.phd.phd_cat_handle != NULL &&
 273               loghandle->u.phd.phd_cat_handle->u.chd.chd_current_log !=
 274                 loghandle))) {
 275                 /* never try to destroy it again */
 276                 llh->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 277                 rc = llog_trans_destroy(env, loghandle, th);
 278                 if (rc < 0) {
 279                         /* Sigh, can not destroy the final plain llog, but
 280                          * the bitmap has been clearly, so the record can not
 281                          * be accessed anymore, let's return 0 for now, and
 282                          * the orphan will be handled by LFSCK. */
 283                         CERROR("%s: can't destroy empty llog "DFID": rc = %d\n",
 284                                loghandle->lgh_ctxt->loc_obd->obd_name,
 285                                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rc);
 286                         GOTO(out_unlock, rc = 0);
 287                 }
 288                 rc = LLOG_DEL_PLAIN;
 289         }
 290
 291 out_unlock:
 292         mutex_unlock(&loghandle->lgh_hdr_mutex);
 293         up_write(&loghandle->lgh_lock);
 294 out_trans:
 295         rc1 = dt_trans_stop(env, dt, th);
 296         if (rc == 0)
 297                 rc = rc1;
 298         if (rc < 0 && subtract_count) {
 299                 mutex_lock(&loghandle->lgh_hdr_mutex);
 300                 loghandle->lgh_hdr->llh_count++;
 301                 ext2_set_bit(index, LLOG_HDR_BITMAP(llh));
 302                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 303         }
 304         RETURN(rc);
 305 }
 306
 307 int llog_read_header(const struct lu_env *env, struct llog_handle *handle,
 308                      const struct obd_uuid *uuid)
 309 {
 310         struct llog_operations *lop;
 311         int rc;
 312         ENTRY;
 313
 314         rc = llog_handle2ops(handle, &lop);
 315         if (rc)
 316                 RETURN(rc);
 317
 318         if (lop->lop_read_header == NULL)
 319                 RETURN(-EOPNOTSUPP);
 320
 321         rc = lop->lop_read_header(env, handle);
 322         if (rc == LLOG_EEMPTY) {
 323                 struct llog_log_hdr *llh = handle->lgh_hdr;
 324
 325                 /* lrh_len should be initialized in llog_init_handle */
 326                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 327                 llh->llh_count = 1;         /* for the header record */
 328                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 329                 LASSERT(handle->lgh_ctxt->loc_chunk_size >=
 330                                                 LLOG_MIN_CHUNK_SIZE);
 331                 llh->llh_hdr.lrh_len = handle->lgh_ctxt->loc_chunk_size;
 332                 llh->llh_hdr.lrh_index = 0;
 333                 llh->llh_timestamp = ktime_get_real_seconds();
 334                 if (uuid)
 335                         memcpy(&llh->llh_tgtuuid, uuid,
 336                                sizeof(llh->llh_tgtuuid));
 337                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 338                 /* Since update llog header might also call this function,
 339                  * let's reset the bitmap to 0 here */
 340                 memset(LLOG_HDR_BITMAP(llh), 0, llh->llh_hdr.lrh_len -
 341                                                 llh->llh_bitmap_offset -
 342                                                 sizeof(llh->llh_tail));
 343                 ext2_set_bit(0, LLOG_HDR_BITMAP(llh));
 344                 LLOG_HDR_TAIL(llh)->lrt_len = llh->llh_hdr.lrh_len;
 345                 LLOG_HDR_TAIL(llh)->lrt_index = llh->llh_hdr.lrh_index;
 346                 rc = 0;
 347         }
 348         RETURN(rc);
 349 }
 350 EXPORT_SYMBOL(llog_read_header);
 351
 352 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 353                      int flags, struct obd_uuid *uuid)
 354 {
 355         struct llog_log_hdr     *llh;
 356         enum llog_flag           fmt = flags & LLOG_F_EXT_MASK;
 357         int                      rc;
 358         int                     chunk_size = handle->lgh_ctxt->loc_chunk_size;
 359         ENTRY;
 360
 361         LASSERT(handle->lgh_hdr == NULL);
 362
 363         LASSERT(chunk_size >= LLOG_MIN_CHUNK_SIZE);
 364         OBD_ALLOC_LARGE(llh, chunk_size);
 365         if (llh == NULL)
 366                 RETURN(-ENOMEM);
 367
 368         handle->lgh_hdr = llh;
 369         handle->lgh_hdr_size = chunk_size;
 370         /* first assign flags to use llog_client_ops */
 371         llh->llh_flags = flags;
 372         rc = llog_read_header(env, handle, uuid);
 373         if (rc == 0) {
 374                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 375                               flags & LLOG_F_IS_CAT) ||
 376                              (llh->llh_flags & LLOG_F_IS_CAT &&
 377                               flags & LLOG_F_IS_PLAIN))) {
 378                         CERROR("%s: llog type is %s but initializing %s\n",
 379                                handle->lgh_ctxt->loc_obd->obd_name,
 380                                llh->llh_flags & LLOG_F_IS_CAT ?
 381                                "catalog" : "plain",
 382                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 383                         GOTO(out, rc = -EINVAL);
 384                 } else if (llh->llh_flags &
 385                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 386                         /*
 387                          * it is possible to open llog without specifying llog
 388                          * type so it is taken from llh_flags
 389                          */
 390                         flags = llh->llh_flags;
 391                 } else {
 392                         /* for some reason the llh_flags has no type set */
 393                         CERROR("llog type is not specified!\n");
 394                         GOTO(out, rc = -EINVAL);
 395                 }
 396                 if (unlikely(uuid &&
 397                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 398                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 399                                handle->lgh_ctxt->loc_obd->obd_name,
 400                                (char *)uuid->uuid,
 401                                (char *)llh->llh_tgtuuid.uuid);
 402                         GOTO(out, rc = -EEXIST);
 403                 }
 404         }
 405         if (flags & LLOG_F_IS_CAT) {
 406                 LASSERT(list_empty(&handle->u.chd.chd_head));
 407                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 408                 llh->llh_size = sizeof(struct llog_logid_rec);
 409                 llh->llh_flags |= LLOG_F_IS_FIXSIZE;
 410         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 411                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 412                        handle->lgh_ctxt->loc_obd->obd_name,
 413                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 414                 rc = -EINVAL;
 415         }
 416         llh->llh_flags |= fmt;
 417 out:
 418         if (rc) {
 419                 OBD_FREE_LARGE(llh, chunk_size);
 420                 handle->lgh_hdr = NULL;
 421         }
 422         RETURN(rc);
 423 }
 424 EXPORT_SYMBOL(llog_init_handle);
 425
 426 static int llog_process_thread(void *arg)
 427 {
 428         struct llog_process_info        *lpi = arg;
 429         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 430         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 431         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 432         struct llog_thread_info         *lti;
 433         char                            *buf;
 434         size_t                           chunk_size;
 435         __u64                            cur_offset;
 436         int                              rc = 0, index = 1, last_index;
 437         int                              saved_index = 0;
 438         int                              last_called_index = 0;
 439         bool                             repeated = false;
 440
 441         ENTRY;
 442
 443         if (llh == NULL)
 444                 RETURN(-EINVAL);
 445
 446         lti = lpi->lpi_env == NULL ? NULL : llog_info(lpi->lpi_env);
 447
 448         cur_offset = chunk_size = llh->llh_hdr.lrh_len;
 449         /* expect chunk_size to be power of two */
 450         LASSERT(is_power_of_2(chunk_size));
 451
 452         OBD_ALLOC_LARGE(buf, chunk_size);
 453         if (buf == NULL) {
 454                 lpi->lpi_rc = -ENOMEM;
 455                 RETURN(0);
 456         }
 457
 458         if (cd != NULL) {
 459                 last_called_index = cd->lpcd_first_idx;
 460                 index = cd->lpcd_first_idx + 1;
 461         }
 462         if (cd != NULL && cd->lpcd_last_idx)
 463                 last_index = cd->lpcd_last_idx;
 464         else
 465                 last_index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 466
 467         while (rc == 0) {
 468                 struct llog_rec_hdr *rec;
 469                 off_t chunk_offset = 0;
 470                 unsigned int buf_offset = 0;
 471                 bool partial_chunk;
 472                 int     lh_last_idx;
 473
 474                 /* skip records not set in bitmap */
 475                 while (index <= last_index &&
 476                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 477                         ++index;
 478
 479                 /* There are no indices prior the last_index */
 480                 if (index > last_index)
 481                         break;
 482
 483                 CDEBUG(D_OTHER, "index: %d last_index %d\n", index,
 484                        last_index);
 485
 486 repeat:
 487                 /* get the buf with our target record; avoid old garbage */
 488                 memset(buf, 0, chunk_size);
 489                 /* the record index for outdated chunk data */
 490                 lh_last_idx = loghandle->lgh_last_idx + 1;
 491                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 492                                      index, &cur_offset, buf, chunk_size);
 493                 if (repeated && rc)
 494                         CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
 495                                " buf_offset %u, rc = %d\n", cur_offset,
 496                                (__u64)chunk_offset, buf_offset, rc);
 497                 /* we`ve tried to reread the chunk, but there is no
 498                  * new records */
 499                 if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
 500                     cur_offset)
 501                         GOTO(out, rc = 0);
 502                 if (rc != 0)
 503                         GOTO(out, rc);
 504
 505                 /* NB: after llog_next_block() call the cur_offset is the
 506                  * offset of the next block after read one.
 507                  * The absolute offset of the current chunk is calculated
 508                  * from cur_offset value and stored in chunk_offset variable.
 509                  */
 510                 if ((cur_offset & (chunk_size - 1)) != 0) {
 511                         partial_chunk = true;
 512                         chunk_offset = cur_offset & ~(chunk_size - 1);
 513                 } else {
 514                         partial_chunk = false;
 515                         chunk_offset = cur_offset - chunk_size;
 516                 }
 517
 518                 /* NB: when rec->lrh_len is accessed it is already swabbed
 519                  * since it is used at the "end" of the loop and the rec
 520                  * swabbing is done at the beginning of the loop. */
 521                 for (rec = (struct llog_rec_hdr *)(buf + buf_offset);
 522                      (char *)rec < buf + chunk_size;
 523                      rec = llog_rec_hdr_next(rec)) {
 524
 525                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 526                                rec, rec->lrh_type);
 527
 528                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 529                                 lustre_swab_llog_rec(rec);
 530
 531                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 532                                rec->lrh_type, rec->lrh_index);
 533
 534                         /* the bitmap could be changed during processing
 535                          * records from the chunk. For wrapped catalog
 536                          * it means we can read deleted record and try to
 537                          * process it. Check this case and reread the chunk. */
 538
 539                         /* for partial chunk the end of it is zeroed, check
 540                          * for index 0 to distinguish it. */
 541                         if ((partial_chunk && rec->lrh_index == 0) ||
 542                              (index == lh_last_idx &&
 543                               lh_last_idx != (loghandle->lgh_last_idx + 1))) {
 544                                 /* concurrent llog_add() might add new records
 545                                  * while llog_processing, check this is not
 546                                  * the case and re-read the current chunk
 547                                  * otherwise. */
 548                                 int records;
 549                                 /* lgh_last_idx could be less then index
 550                                  * for catalog, if catalog is wrapped */
 551                                 if ((index > loghandle->lgh_last_idx &&
 552                                     !(loghandle->lgh_hdr->llh_flags &
 553                                       LLOG_F_IS_CAT)) || repeated ||
 554                                     (loghandle->lgh_obj != NULL &&
 555                                      dt_object_remote(loghandle->lgh_obj)))
 556                                         GOTO(out, rc = 0);
 557                                 /* <2 records means no more records
 558                                  * if the last record we processed was
 559                                  * the final one, then the underlying
 560                                  * object might have been destroyed yet.
 561                                  * we better don't access that.. */
 562                                 mutex_lock(&loghandle->lgh_hdr_mutex);
 563                                 records = loghandle->lgh_hdr->llh_count;
 564                                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 565                                 if (records <= 1)
 566                                         GOTO(out, rc = 0);
 567                                 CDEBUG(D_OTHER, "Re-read last llog buffer for "
 568                                        "new records, index %u, last %u\n",
 569                                        index, loghandle->lgh_last_idx);
 570                                 /* save offset inside buffer for the re-read */
 571                                 buf_offset = (char *)rec - (char *)buf;
 572                                 cur_offset = chunk_offset;
 573                                 repeated = true;
 574                                 goto repeat;
 575                         }
 576
 577                         repeated = false;
 578
 579                         if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
 580                                 CWARN("%s: invalid length %d in llog "DFID
 581                                       "record for index %d/%d\n",
 582                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 583                                        rec->lrh_len,
 584                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 585                                        rec->lrh_index, index);
 586
 587                                 GOTO(out, rc = -EINVAL);
 588                         }
 589
 590                         if (rec->lrh_index < index) {
 591                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 592                                        rec->lrh_index);
 593                                 continue;
 594                         }
 595
 596                         if (rec->lrh_index != index) {
 597                                 CERROR("%s: "DFID" Invalid record: index %u"
 598                                        " but expected %u\n",
 599                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 600                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 601                                        rec->lrh_index, index);
 602                                 GOTO(out, rc = -ERANGE);
 603                         }
 604
 605                         CDEBUG(D_OTHER,
 606                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 607                                rec->lrh_index, rec->lrh_len,
 608                                (int)(buf + chunk_size - (char *)rec));
 609
 610                         /* lgh_cur_offset is used only at llog_test_3 */
 611                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 612                                                     chunk_offset;
 613
 614                         /* if set, process the callback on this record */
 615                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 616                                 struct llog_cookie *lgc;
 617                                 __u64   tmp_off;
 618                                 int     tmp_idx;
 619
 620                                 if (lti != NULL) {
 621                                         lgc = &lti->lgi_cookie;
 622                                         /* store lu_env for recursive calls */
 623                                         tmp_off = lgc->lgc_offset;
 624                                         tmp_idx = lgc->lgc_index;
 625
 626                                         lgc->lgc_offset = (char *)rec -
 627                                                 (char *)buf + chunk_offset;
 628                                         lgc->lgc_index = rec->lrh_index;
 629                                 }
 630                                 /* using lu_env for passing record offset to
 631                                  * llog_write through various callbacks */
 632                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 633                                                  lpi->lpi_cbdata);
 634                                 last_called_index = index;
 635
 636                                 if (lti != NULL) {
 637                                         lgc->lgc_offset = tmp_off;
 638                                         lgc->lgc_index = tmp_idx;
 639                                 }
 640
 641                                 if (rc == LLOG_PROC_BREAK) {
 642                                         GOTO(out, rc);
 643                                 } else if (rc == LLOG_DEL_RECORD) {
 644                                         rc = llog_cancel_rec(lpi->lpi_env,
 645                                                              loghandle,
 646                                                              rec->lrh_index);
 647                                 }
 648                                 if (rc)
 649                                         GOTO(out, rc);
 650                                 /* some stupid callbacks directly cancel records
 651                                  * and delete llog. Check it and stop
 652                                  * processing. */
 653                                 if (loghandle->lgh_hdr == NULL ||
 654                                     loghandle->lgh_hdr->llh_count == 1)
 655                                         GOTO(out, rc = 0);
 656                         }
 657                         /* exit if the last index is reached */
 658                         if (index >= last_index)
 659                                 GOTO(out, rc = 0);
 660                         ++index;
 661                 }
 662         }
 663
 664 out:
 665         if (cd != NULL)
 666                 cd->lpcd_last_idx = last_called_index;
 667
 668         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 669                 if (dt_object_remote(loghandle->lgh_obj)) {
 670                         /* If it is remote object, then -EIO might means
 671                          * disconnection or eviction, let's return -EAGAIN,
 672                          * so for update recovery log processing, it will
 673                          * retry until the umount or abort recovery, see
 674                          * lod_sub_recovery_thread() */
 675                         CERROR("%s retry remote llog process\n",
 676                                loghandle->lgh_ctxt->loc_obd->obd_name);
 677                         rc = -EAGAIN;
 678                 } else {
 679                         /* something bad happened to the processing of a local
 680                          * llog file, probably I/O error or the log got
 681                          * corrupted to be able to finally release the log we
 682                          * discard any remaining bits in the header */
 683                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 684                                " %s index %d count %d\n",
 685                                loghandle->lgh_ctxt->loc_obd->obd_name,
 686                                POSTID(&loghandle->lgh_id.lgl_oi),
 687                                loghandle->lgh_id.lgl_ogen,
 688                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 689                                 "plain"), index, llh->llh_count);
 690
 691                         while (index <= last_index) {
 692                                 if (ext2_test_bit(index,
 693                                                   LLOG_HDR_BITMAP(llh)) != 0)
 694                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 695                                                         index);
 696                                 index++;
 697                         }
 698                         rc = 0;
 699                 }
 700         }
 701
 702         OBD_FREE_LARGE(buf, chunk_size);
 703         lpi->lpi_rc = rc;
 704         return 0;
 705 }
 706
 707 static int llog_process_thread_daemonize(void *arg)
 708 {
 709         struct llog_process_info        *lpi = arg;
 710         struct lu_env                    env;
 711         int                              rc;
 712         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 713
 714         task_lock(lpi->lpi_reftask);
 715         new_ns = lpi->lpi_reftask->nsproxy;
 716         if (curr_ns != new_ns) {
 717                 get_nsproxy(new_ns);
 718
 719                 current->nsproxy = new_ns;
 720                 /* XXX: we should call put_nsproxy() instead of
 721                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 722                  * used outside of the kernel itself, because it calls
 723                  * free_nsproxy() which is not exported by the kernel
 724                  * (defined in kernel/nsproxy.c) */
 725                 atomic_dec(&curr_ns->count);
 726         }
 727         task_unlock(lpi->lpi_reftask);
 728
 729         unshare_fs_struct();
 730
 731         /* client env has no keys, tags is just 0 */
 732         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 733         if (rc)
 734                 goto out;
 735         lpi->lpi_env = &env;
 736
 737         rc = llog_process_thread(arg);
 738
 739         lu_env_fini(&env);
 740 out:
 741         complete(&lpi->lpi_completion);
 742         return rc;
 743 }
 744
 745 int llog_process_or_fork(const struct lu_env *env,
 746                          struct llog_handle *loghandle,
 747                          llog_cb_t cb, void *data, void *catdata, bool fork)
 748 {
 749         struct llog_process_info *lpi;
 750         int                      rc;
 751
 752         ENTRY;
 753
 754         OBD_ALLOC_PTR(lpi);
 755         if (lpi == NULL) {
 756                 CERROR("cannot alloc pointer\n");
 757                 RETURN(-ENOMEM);
 758         }
 759         lpi->lpi_loghandle = loghandle;
 760         lpi->lpi_cb        = cb;
 761         lpi->lpi_cbdata    = data;
 762         lpi->lpi_catdata   = catdata;
 763
 764         if (fork) {
 765                 struct task_struct *task;
 766
 767                 /* The new thread can't use parent env,
 768                  * init the new one in llog_process_thread_daemonize. */
 769                 lpi->lpi_env = NULL;
 770                 init_completion(&lpi->lpi_completion);
 771                 /* take reference to current, so that
 772                  * llog_process_thread_daemonize() can use it to switch to
 773                  * namespace associated with current  */
 774                 lpi->lpi_reftask = current;
 775                 task = kthread_run(llog_process_thread_daemonize, lpi,
 776                                    "llog_process_thread");
 777                 if (IS_ERR(task)) {
 778                         rc = PTR_ERR(task);
 779                         CERROR("%s: cannot start thread: rc = %d\n",
 780                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 781                         GOTO(out_lpi, rc);
 782                 }
 783                 wait_for_completion(&lpi->lpi_completion);
 784         } else {
 785                 lpi->lpi_env = env;
 786                 llog_process_thread(lpi);
 787         }
 788         rc = lpi->lpi_rc;
 789
 790 out_lpi:
 791         OBD_FREE_PTR(lpi);
 792         RETURN(rc);
 793 }
 794 EXPORT_SYMBOL(llog_process_or_fork);
 795
 796 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 797                  llog_cb_t cb, void *data, void *catdata)
 798 {
 799         int rc;
 800         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 801         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 802 }
 803 EXPORT_SYMBOL(llog_process);
 804
 805 int llog_reverse_process(const struct lu_env *env,
 806                          struct llog_handle *loghandle, llog_cb_t cb,
 807                          void *data, void *catdata)
 808 {
 809         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 810         struct llog_process_cat_data *cd = catdata;
 811         void *buf;
 812         int rc = 0, first_index = 1, index, idx;
 813         __u32   chunk_size = llh->llh_hdr.lrh_len;
 814         ENTRY;
 815
 816         OBD_ALLOC_LARGE(buf, chunk_size);
 817         if (buf == NULL)
 818                 RETURN(-ENOMEM);
 819
 820         if (cd != NULL)
 821                 first_index = cd->lpcd_first_idx + 1;
 822         if (cd != NULL && cd->lpcd_last_idx)
 823                 index = cd->lpcd_last_idx;
 824         else
 825                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 826
 827         while (rc == 0) {
 828                 struct llog_rec_hdr *rec;
 829                 struct llog_rec_tail *tail;
 830
 831                 /* skip records not set in bitmap */
 832                 while (index >= first_index &&
 833                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 834                         --index;
 835
 836                 LASSERT(index >= first_index - 1);
 837                 if (index == first_index - 1)
 838                         break;
 839
 840                 /* get the buf with our target record; avoid old garbage */
 841                 memset(buf, 0, chunk_size);
 842                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 843                 if (rc)
 844                         GOTO(out, rc);
 845
 846                 rec = buf;
 847                 idx = rec->lrh_index;
 848                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 849                 while (idx < index) {
 850                         rec = (void *)rec + rec->lrh_len;
 851                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 852                                 lustre_swab_llog_rec(rec);
 853                         idx ++;
 854                 }
 855                 LASSERT(idx == index);
 856                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 857
 858                 /* process records in buffer, starting where we found one */
 859                 while ((void *)tail > buf) {
 860                         if (tail->lrt_index == 0)
 861                                 GOTO(out, rc = 0); /* no more records */
 862
 863                         /* if set, process the callback on this record */
 864                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 865                                 rec = (void *)tail - tail->lrt_len +
 866                                       sizeof(*tail);
 867
 868                                 rc = cb(env, loghandle, rec, data);
 869                                 if (rc == LLOG_PROC_BREAK) {
 870                                         GOTO(out, rc);
 871                                 } else if (rc == LLOG_DEL_RECORD) {
 872                                         rc = llog_cancel_rec(env, loghandle,
 873                                                              tail->lrt_index);
 874                                 }
 875                                 if (rc)
 876                                         GOTO(out, rc);
 877                         }
 878
 879                         /* previous record, still in buffer? */
 880                         --index;
 881                         if (index < first_index)
 882                                 GOTO(out, rc = 0);
 883                         tail = (void *)tail - tail->lrt_len;
 884                 }
 885         }
 886
 887 out:
 888         if (buf != NULL)
 889                 OBD_FREE_LARGE(buf, chunk_size);
 890         RETURN(rc);
 891 }
 892 EXPORT_SYMBOL(llog_reverse_process);
 893
 894 /**
 895  * new llog API
 896  *
 897  * API functions:
 898  *      llog_open - open llog, may not exist
 899  *      llog_exist - check if llog exists
 900  *      llog_close - close opened llog, pair for open, frees llog_handle
 901  *      llog_declare_create - declare llog creation
 902  *      llog_create - create new llog on disk, need transaction handle
 903  *      llog_declare_write_rec - declaration of llog write
 904  *      llog_write_rec - write llog record on disk, need transaction handle
 905  *      llog_declare_add - declare llog catalog record addition
 906  *      llog_add - add llog record in catalog, need transaction handle
 907  */
 908 int llog_exist(struct llog_handle *loghandle)
 909 {
 910         struct llog_operations  *lop;
 911         int                      rc;
 912
 913         ENTRY;
 914
 915         rc = llog_handle2ops(loghandle, &lop);
 916         if (rc)
 917                 RETURN(rc);
 918         if (lop->lop_exist == NULL)
 919                 RETURN(-EOPNOTSUPP);
 920
 921         rc = lop->lop_exist(loghandle);
 922         RETURN(rc);
 923 }
 924 EXPORT_SYMBOL(llog_exist);
 925
 926 int llog_declare_create(const struct lu_env *env,
 927                         struct llog_handle *loghandle, struct thandle *th)
 928 {
 929         struct llog_operations  *lop;
 930         int                      raised, rc;
 931
 932         ENTRY;
 933
 934         rc = llog_handle2ops(loghandle, &lop);
 935         if (rc)
 936                 RETURN(rc);
 937         if (lop->lop_declare_create == NULL)
 938                 RETURN(-EOPNOTSUPP);
 939
 940         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 941         if (!raised)
 942                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 943         rc = lop->lop_declare_create(env, loghandle, th);
 944         if (!raised)
 945                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 946         RETURN(rc);
 947 }
 948
 949 int llog_create(const struct lu_env *env, struct llog_handle *handle,
 950                 struct thandle *th)
 951 {
 952         struct llog_operations  *lop;
 953         int                      raised, rc;
 954
 955         ENTRY;
 956
 957         rc = llog_handle2ops(handle, &lop);
 958         if (rc)
 959                 RETURN(rc);
 960         if (lop->lop_create == NULL)
 961                 RETURN(-EOPNOTSUPP);
 962
 963         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 964         if (!raised)
 965                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 966         rc = lop->lop_create(env, handle, th);
 967         if (!raised)
 968                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 969         RETURN(rc);
 970 }
 971
 972 int llog_declare_write_rec(const struct lu_env *env,
 973                            struct llog_handle *handle,
 974                            struct llog_rec_hdr *rec, int idx,
 975                            struct thandle *th)
 976 {
 977         struct llog_operations  *lop;
 978         int                      raised, rc;
 979
 980         ENTRY;
 981
 982         rc = llog_handle2ops(handle, &lop);
 983         if (rc)
 984                 RETURN(rc);
 985         LASSERT(lop);
 986         if (lop->lop_declare_write_rec == NULL)
 987                 RETURN(-EOPNOTSUPP);
 988
 989         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 990         if (!raised)
 991                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 992         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
 993         if (!raised)
 994                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 995         RETURN(rc);
 996 }
 997
 998 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
 999                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1000                    int idx, struct thandle *th)
1001 {
1002         struct llog_operations  *lop;
1003         int                      raised, rc, buflen;
1004
1005         ENTRY;
1006
1007         /* API sanity checks */
1008         if (handle == NULL) {
1009                 CERROR("loghandle is missed\n");
1010                 RETURN(-EPROTO);
1011         } else if (handle->lgh_obj == NULL) {
1012                 CERROR("loghandle %p with NULL object\n",
1013                         handle);
1014                 RETURN(-EPROTO);
1015         } else if (th == NULL) {
1016                 CERROR("%s: missed transaction handle\n",
1017                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name);
1018                 RETURN(-EPROTO);
1019         } else if (handle->lgh_hdr == NULL) {
1020                 CERROR("%s: loghandle %p with no header\n",
1021                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name,
1022                         handle);
1023                 RETURN(-EPROTO);
1024         }
1025
1026         rc = llog_handle2ops(handle, &lop);
1027         if (rc)
1028                 RETURN(rc);
1029
1030         if (lop->lop_write_rec == NULL)
1031                 RETURN(-EOPNOTSUPP);
1032
1033         buflen = rec->lrh_len;
1034         LASSERT(cfs_size_round(buflen) == buflen);
1035
1036         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1037         if (!raised)
1038                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1039         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1040         if (!raised)
1041                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1042         RETURN(rc);
1043 }
1044
1045 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1046              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1047              struct thandle *th)
1048 {
1049         int raised, rc;
1050
1051         ENTRY;
1052
1053         if (lgh->lgh_logops->lop_add == NULL)
1054                 RETURN(-EOPNOTSUPP);
1055
1056         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1057         if (!raised)
1058                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1059         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1060         if (!raised)
1061                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1062         RETURN(rc);
1063 }
1064 EXPORT_SYMBOL(llog_add);
1065
1066 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1067                      struct llog_rec_hdr *rec, struct thandle *th)
1068 {
1069         int raised, rc;
1070
1071         ENTRY;
1072
1073         if (lgh->lgh_logops->lop_declare_add == NULL)
1074                 RETURN(-EOPNOTSUPP);
1075
1076         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1077         if (!raised)
1078                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1079         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1080         if (!raised)
1081                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1082         RETURN(rc);
1083 }
1084 EXPORT_SYMBOL(llog_declare_add);
1085
1086 /**
1087  * Helper function to open llog or create it if doesn't exist.
1088  * It hides all transaction handling from caller.
1089  */
1090 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1091                      struct llog_handle **res, struct llog_logid *logid,
1092                      char *name)
1093 {
1094         struct dt_device        *d;
1095         struct thandle          *th;
1096         int                      rc;
1097
1098         ENTRY;
1099
1100         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1101         if (rc)
1102                 RETURN(rc);
1103
1104         if (llog_exist(*res))
1105                 RETURN(0);
1106
1107         LASSERT((*res)->lgh_obj != NULL);
1108
1109         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1110
1111         th = dt_trans_create(env, d);
1112         if (IS_ERR(th))
1113                 GOTO(out, rc = PTR_ERR(th));
1114
1115         /* Create update llog object synchronously, which
1116          * happens during inialization process see
1117          * lod_sub_prep_llog(), to make sure the update
1118          * llog object is created before corss-MDT writing
1119          * updates into the llog object */
1120         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1121                 th->th_sync = 1;
1122
1123         th->th_wait_submit = 1;
1124         rc = llog_declare_create(env, *res, th);
1125         if (rc == 0) {
1126                 rc = dt_trans_start_local(env, d, th);
1127                 if (rc == 0)
1128                         rc = llog_create(env, *res, th);
1129         }
1130         dt_trans_stop(env, d, th);
1131 out:
1132         if (rc)
1133                 llog_close(env, *res);
1134         RETURN(rc);
1135 }
1136 EXPORT_SYMBOL(llog_open_create);
1137
1138 /**
1139  * Helper function to delete existent llog.
1140  */
1141 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1142                struct llog_logid *logid, char *name)
1143 {
1144         struct llog_handle      *handle;
1145         int                      rc = 0, rc2;
1146
1147         ENTRY;
1148
1149         /* nothing to erase */
1150         if (name == NULL && logid == NULL)
1151                 RETURN(0);
1152
1153         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1154         if (rc < 0)
1155                 RETURN(rc);
1156
1157         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1158         if (rc == 0)
1159                 rc = llog_destroy(env, handle);
1160
1161         rc2 = llog_close(env, handle);
1162         if (rc == 0)
1163                 rc = rc2;
1164         RETURN(rc);
1165 }
1166 EXPORT_SYMBOL(llog_erase);
1167
1168 /*
1169  * Helper function for write record in llog.
1170  * It hides all transaction handling from caller.
1171  * Valid only with local llog.
1172  */
1173 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1174                struct llog_rec_hdr *rec, int idx)
1175 {
1176         struct dt_device        *dt;
1177         struct thandle          *th;
1178         bool                    need_cookie;
1179         int                     rc;
1180
1181         ENTRY;
1182
1183         LASSERT(loghandle);
1184         LASSERT(loghandle->lgh_ctxt);
1185         LASSERT(loghandle->lgh_obj != NULL);
1186
1187         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1188
1189         th = dt_trans_create(env, dt);
1190         if (IS_ERR(th))
1191                 RETURN(PTR_ERR(th));
1192
1193         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1194         if (rc)
1195                 GOTO(out_trans, rc);
1196
1197         th->th_wait_submit = 1;
1198         rc = dt_trans_start_local(env, dt, th);
1199         if (rc)
1200                 GOTO(out_trans, rc);
1201
1202         need_cookie = !(idx == LLOG_HEADER_IDX || idx == LLOG_NEXT_IDX);
1203
1204         down_write(&loghandle->lgh_lock);
1205         if (need_cookie) {
1206                 struct llog_thread_info *lti = llog_info(env);
1207
1208                 /* cookie comes from llog_process_thread */
1209                 rc = llog_write_rec(env, loghandle, rec, &lti->lgi_cookie,
1210                                     rec->lrh_index, th);
1211                 /* upper layer didn`t pass cookie so change rc */
1212                 rc = (rc == 1 ? 0 : rc);
1213         } else {
1214                 rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1215         }
1216
1217         up_write(&loghandle->lgh_lock);
1218 out_trans:
1219         dt_trans_stop(env, dt, th);
1220         RETURN(rc);
1221 }
1222 EXPORT_SYMBOL(llog_write);
1223
1224 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1225               struct llog_handle **lgh, struct llog_logid *logid,
1226               char *name, enum llog_open_param open_param)
1227 {
1228         int      raised;
1229         int      rc;
1230
1231         ENTRY;
1232
1233         LASSERT(ctxt);
1234         LASSERT(ctxt->loc_logops);
1235
1236         if (ctxt->loc_logops->lop_open == NULL) {
1237                 *lgh = NULL;
1238                 RETURN(-EOPNOTSUPP);
1239         }
1240
1241         *lgh = llog_alloc_handle();
1242         if (*lgh == NULL)
1243                 RETURN(-ENOMEM);
1244         (*lgh)->lgh_ctxt = ctxt;
1245         (*lgh)->lgh_logops = ctxt->loc_logops;
1246
1247         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1248         if (!raised)
1249                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1250         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1251         if (!raised)
1252                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1253         if (rc) {
1254                 llog_free_handle(*lgh);
1255                 *lgh = NULL;
1256         }
1257         RETURN(rc);
1258 }
1259 EXPORT_SYMBOL(llog_open);
1260
1261 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1262 {
1263         struct llog_operations  *lop;
1264         int                      rc;
1265
1266         ENTRY;
1267
1268         rc = llog_handle2ops(loghandle, &lop);
1269         if (rc)
1270                 GOTO(out, rc);
1271         if (lop->lop_close == NULL)
1272                 GOTO(out, rc = -EOPNOTSUPP);
1273         rc = lop->lop_close(env, loghandle);
1274 out:
1275         llog_handle_put(loghandle);
1276         RETURN(rc);
1277 }
1278 EXPORT_SYMBOL(llog_close);
1279
1280 /**
1281  * Helper function to get the llog size in records. It is used by MGS
1282  * mostly to check that config llog exists and contains data.
1283  *
1284  * \param[in] env       execution environment
1285  * \param[in] ctxt      llog context
1286  * \param[in] name      llog name
1287  *
1288  * \retval              true if there are records in llog besides a header
1289  * \retval              false on error or llog without records
1290  */
1291 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1292                   char *name)
1293 {
1294         struct llog_handle      *llh;
1295         int                      rc = 0;
1296
1297         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1298         if (rc < 0) {
1299                 if (likely(rc == -ENOENT))
1300                         rc = 0;
1301                 GOTO(out, rc);
1302         }
1303
1304         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1305         if (rc)
1306                 GOTO(out_close, rc);
1307         rc = llog_get_size(llh);
1308
1309 out_close:
1310         llog_close(env, llh);
1311 out:
1312         /* The header is record 1, the llog is still considered as empty
1313          * if there is only header */
1314         return (rc <= 1);
1315 }
1316 EXPORT_SYMBOL(llog_is_empty);
1317
1318 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1319                       struct llog_rec_hdr *rec, void *data)
1320 {
1321         struct llog_handle      *copy_llh = data;
1322
1323         /* Append all records */
1324         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1325 }
1326
1327 /* backup plain llog */
1328 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1329                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1330                 char *name, char *backup)
1331 {
1332         struct llog_handle      *llh, *bllh;
1333         int                      rc;
1334
1335         ENTRY;
1336
1337         /* open original log */
1338         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1339         if (rc < 0) {
1340                 /* the -ENOENT case is also reported to the caller
1341                  * but silently so it should handle that if needed.
1342                  */
1343                 if (rc != -ENOENT)
1344                         CERROR("%s: failed to open log %s: rc = %d\n",
1345                                obd->obd_name, name, rc);
1346                 RETURN(rc);
1347         }
1348
1349         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1350         if (rc)
1351                 GOTO(out_close, rc);
1352
1353         /* Make sure there's no old backup log */
1354         rc = llog_erase(env, bctxt, NULL, backup);
1355         if (rc < 0 && rc != -ENOENT)
1356                 GOTO(out_close, rc);
1357
1358         /* open backup log */
1359         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1360         if (rc) {
1361                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1362                        obd->obd_name, backup, rc);
1363                 GOTO(out_close, rc);
1364         }
1365
1366         /* check that backup llog is not the same object as original one */
1367         if (llh->lgh_obj == bllh->lgh_obj) {
1368                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1369                        obd->obd_name, name, backup, llh->lgh_obj,
1370                        bllh->lgh_obj);
1371                 GOTO(out_backup, rc = -EEXIST);
1372         }
1373
1374         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1375         if (rc)
1376                 GOTO(out_backup, rc);
1377
1378         /* Copy log record by record */
1379         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1380                                   NULL, false);
1381         if (rc)
1382                 CERROR("%s: failed to backup log %s: rc = %d\n",
1383                        obd->obd_name, name, rc);
1384 out_backup:
1385         llog_close(env, bllh);
1386 out_close:
1387         llog_close(env, llh);
1388         RETURN(rc);
1389 }
1390 EXPORT_SYMBOL(llog_backup);
1391
1392 /* Get size of llog */
1393 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1394 {
1395         int rc;
1396         struct lu_attr la;
1397
1398         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1399         if (rc) {
1400                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1401                        llh->lgh_ctxt->loc_obd->obd_name,
1402                        PFID(&llh->lgh_id.lgl_oi.oi_fid), rc);
1403                 return 0;
1404         }
1405
1406         return la.la_size;
1407 }
1408 EXPORT_SYMBOL(llog_size);
1409