lustre/obdclass/llog_cat.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/obdclass/llog_cat.c
  33  *
  34  * OST<->MDS recovery logging infrastructure.
  35  *
  36  * Invariants in implementation:
  37  * - we do not share logs among different OST<->MDS connections, so that
  38  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  39  *
  40  * Author: Andreas Dilger <adilger@clusterfs.com>
  41  * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
  42  * Author: Mikhail Pershin <mike.pershin@intel.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_LOG
  46
  47
  48 #include <obd_class.h>
  49
  50 #include "llog_internal.h"
  51
  52 /* Create a new log handle and add it to the open list.
  53  * This log handle will be closed when all of the records in it are removed.
  54  *
  55  * Assumes caller has already pushed us into the kernel context and is locking.
  56  */
  57 static int llog_cat_new_log(const struct lu_env *env,
  58                             struct llog_handle *cathandle,
  59                             struct llog_handle *loghandle,
  60                             struct thandle *th)
  61 {
  62         struct llog_thread_info *lgi = llog_info(env);
  63         struct llog_logid_rec   *rec = &lgi->lgi_logid;
  64         struct thandle *handle = NULL;
  65         struct dt_device *dt = NULL;
  66         struct llog_log_hdr     *llh = cathandle->lgh_hdr;
  67         int                      rc, index;
  68
  69         ENTRY;
  70
  71         index = (cathandle->lgh_last_idx + 1) %
  72                 (OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ? (cfs_fail_val + 1) :
  73                                                 LLOG_HDR_BITMAP_SIZE(llh));
  74
  75         /* check that new llog index will not overlap with the first one.
  76          * - llh_cat_idx is the index just before the first/oldest still in-use
  77          *      index in catalog
  78          * - lgh_last_idx is the last/newest used index in catalog
  79          *
  80          * When catalog is not wrapped yet then lgh_last_idx is always larger
  81          * than llh_cat_idx. After the wrap around lgh_last_idx re-starts
  82          * from 0 and llh_cat_idx becomes the upper limit for it
  83          *
  84          * Check if catalog has already wrapped around or not by comparing
  85          * last_idx and cat_idx */
  86         if ((index == llh->llh_cat_idx + 1 && llh->llh_count > 1) ||
  87             (index == 0 && llh->llh_cat_idx == 0)) {
  88                 CWARN("%s: there are no more free slots in catalog\n",
  89                       loghandle->lgh_ctxt->loc_obd->obd_name);
  90                 RETURN(-ENOSPC);
  91         }
  92
  93         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED))
  94                 RETURN(-ENOSPC);
  95
  96         if (loghandle->lgh_hdr != NULL) {
  97                 /* If llog object is remote and creation is failed, lgh_hdr
  98                  * might be left over here, free it first */
  99                 LASSERT(!llog_exist(loghandle));
 100                 OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
 101                 loghandle->lgh_hdr = NULL;
 102         }
 103
 104         if (th == NULL) {
 105                 dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 106
 107                 handle = dt_trans_create(env, dt);
 108                 if (IS_ERR(handle))
 109                         RETURN(PTR_ERR(handle));
 110
 111                 /* Create update llog object synchronously, which
 112                  * happens during inialization process see
 113                  * lod_sub_prep_llog(), to make sure the update
 114                  * llog object is created before corss-MDT writing
 115                  * updates into the llog object */
 116                 if (cathandle->lgh_ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
 117                         handle->th_sync = 1;
 118
 119                 handle->th_wait_submit = 1;
 120
 121                 rc = llog_declare_create(env, loghandle, handle);
 122                 if (rc != 0)
 123                         GOTO(out, rc);
 124
 125                 rec->lid_hdr.lrh_len = sizeof(*rec);
 126                 rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
 127                 rec->lid_id = loghandle->lgh_id;
 128                 rc = llog_declare_write_rec(env, cathandle, &rec->lid_hdr, -1,
 129                                             handle);
 130                 if (rc != 0)
 131                         GOTO(out, rc);
 132
 133                 rc = dt_trans_start_local(env, dt, handle);
 134                 if (rc != 0)
 135                         GOTO(out, rc);
 136
 137                 th = handle;
 138         }
 139
 140         rc = llog_create(env, loghandle, th);
 141         /* if llog is already created, no need to initialize it */
 142         if (rc == -EEXIST) {
 143                 GOTO(out, rc = 0);
 144         } else if (rc != 0) {
 145                 CERROR("%s: can't create new plain llog in catalog: rc = %d\n",
 146                        loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 147                 GOTO(out, rc);
 148         }
 149
 150         rc = llog_init_handle(env, loghandle,
 151                               LLOG_F_IS_PLAIN | LLOG_F_ZAP_WHEN_EMPTY,
 152                               &cathandle->lgh_hdr->llh_tgtuuid);
 153         if (rc < 0)
 154                 GOTO(out, rc);
 155
 156         /* build the record for this log in the catalog */
 157         rec->lid_hdr.lrh_len = sizeof(*rec);
 158         rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
 159         rec->lid_id = loghandle->lgh_id;
 160
 161         /* append the new record into catalog. The new index will be
 162          * assigned to the record and updated in rec header */
 163         rc = llog_write_rec(env, cathandle, &rec->lid_hdr,
 164                             &loghandle->u.phd.phd_cookie, LLOG_NEXT_IDX, th);
 165         if (rc < 0)
 166                 GOTO(out_destroy, rc);
 167
 168         CDEBUG(D_OTHER, "new plain log "DFID".%u of catalog "DFID"\n",
 169                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rec->lid_hdr.lrh_index,
 170                PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
 171
 172         loghandle->lgh_hdr->llh_cat_idx = rec->lid_hdr.lrh_index;
 173
 174         /* limit max size of plain llog so that space can be
 175          * released sooner, especially on small filesystems */
 176         /* 2MB for the cases when free space hasn't been learned yet */
 177         loghandle->lgh_max_size = 2 << 20;
 178         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 179         rc = dt_statfs(env, dt, &lgi->lgi_statfs);
 180         if (rc == 0 && lgi->lgi_statfs.os_bfree > 0) {
 181                 __u64 freespace = (lgi->lgi_statfs.os_bfree *
 182                                   lgi->lgi_statfs.os_bsize) >> 6;
 183                 if (freespace < loghandle->lgh_max_size)
 184                         loghandle->lgh_max_size = freespace;
 185                 /* shouldn't be > 128MB in any case?
 186                  * it's 256K records of 512 bytes each */
 187                 if (freespace > (128 << 20))
 188                         loghandle->lgh_max_size = 128 << 20;
 189         }
 190         rc = 0;
 191
 192 out:
 193         if (handle != NULL) {
 194                 handle->th_result = rc >= 0 ? 0 : rc;
 195                 dt_trans_stop(env, dt, handle);
 196         }
 197         RETURN(rc);
 198
 199 out_destroy:
 200         /* to signal llog_cat_close() it shouldn't try to destroy the llog,
 201          * we want to destroy it in this transaction, otherwise the object
 202          * becomes an orphan */
 203         loghandle->lgh_hdr->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 204         /* this is to mimic full log, so another llog_cat_current_log()
 205          * can skip it and ask for another onet */
 206         loghandle->lgh_last_idx = LLOG_HDR_BITMAP_SIZE(llh) + 1;
 207         llog_trans_destroy(env, loghandle, th);
 208         RETURN(rc);
 209 }
 210
 211 /* Open an existent log handle and add it to the open list.
 212  * This log handle will be closed when all of the records in it are removed.
 213  *
 214  * Assumes caller has already pushed us into the kernel context and is locking.
 215  * We return a lock on the handle to ensure nobody yanks it from us.
 216  *
 217  * This takes extra reference on llog_handle via llog_handle_get() and require
 218  * this reference to be put by caller using llog_handle_put()
 219  */
 220 int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
 221                        struct llog_handle **res, struct llog_logid *logid)
 222 {
 223         struct llog_handle      *loghandle;
 224         enum llog_flag           fmt;
 225         int                      rc = 0;
 226
 227         ENTRY;
 228
 229         if (cathandle == NULL)
 230                 RETURN(-EBADF);
 231
 232         fmt = cathandle->lgh_hdr->llh_flags & LLOG_F_EXT_MASK;
 233         down_write(&cathandle->lgh_lock);
 234         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 235                             u.phd.phd_entry) {
 236                 struct llog_logid *cgl = &loghandle->lgh_id;
 237
 238                 if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
 239                     ostid_seq(&cgl->lgl_oi) == ostid_seq(&logid->lgl_oi)) {
 240                         if (cgl->lgl_ogen != logid->lgl_ogen) {
 241                                 CWARN("%s: log "DFID" generation %x != %x\n",
 242                                       loghandle->lgh_ctxt->loc_obd->obd_name,
 243                                       PFID(&logid->lgl_oi.oi_fid),
 244                                       cgl->lgl_ogen, logid->lgl_ogen);
 245                                 continue;
 246                         }
 247                         loghandle->u.phd.phd_cat_handle = cathandle;
 248                         up_write(&cathandle->lgh_lock);
 249                         GOTO(out, rc = 0);
 250                 }
 251         }
 252         up_write(&cathandle->lgh_lock);
 253
 254         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle, logid, NULL,
 255                        LLOG_OPEN_EXISTS);
 256         if (rc < 0) {
 257                 CERROR("%s: error opening log id "DFID":%x: rc = %d\n",
 258                        cathandle->lgh_ctxt->loc_obd->obd_name,
 259                        PFID(&logid->lgl_oi.oi_fid), logid->lgl_ogen, rc);
 260                 RETURN(rc);
 261         }
 262
 263         rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN | fmt, NULL);
 264         if (rc < 0) {
 265                 llog_close(env, loghandle);
 266                 loghandle = NULL;
 267                 RETURN(rc);
 268         }
 269
 270         down_write(&cathandle->lgh_lock);
 271         list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
 272         up_write(&cathandle->lgh_lock);
 273
 274         loghandle->u.phd.phd_cat_handle = cathandle;
 275         loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
 276         loghandle->u.phd.phd_cookie.lgc_index =
 277                                 loghandle->lgh_hdr->llh_cat_idx;
 278         EXIT;
 279 out:
 280         llog_handle_get(loghandle);
 281         *res = loghandle;
 282         return 0;
 283 }
 284
 285 int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
 286 {
 287         struct llog_handle      *loghandle, *n;
 288         int                      rc;
 289
 290         ENTRY;
 291
 292         list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
 293                                  u.phd.phd_entry) {
 294                 struct llog_log_hdr     *llh = loghandle->lgh_hdr;
 295                 int                      index;
 296
 297                 /* unlink open-not-created llogs */
 298                 list_del_init(&loghandle->u.phd.phd_entry);
 299                 llh = loghandle->lgh_hdr;
 300                 if (loghandle->lgh_obj != NULL && llh != NULL &&
 301                     (llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 302                     (llh->llh_count == 1)) {
 303                         rc = llog_destroy(env, loghandle);
 304                         if (rc)
 305                                 CERROR("%s: failure destroying log during "
 306                                        "cleanup: rc = %d\n",
 307                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 308                                        rc);
 309
 310                         index = loghandle->u.phd.phd_cookie.lgc_index;
 311                         llog_cat_cleanup(env, cathandle, NULL, index);
 312                 }
 313                 llog_close(env, loghandle);
 314         }
 315         /* if handle was stored in ctxt, remove it too */
 316         if (cathandle->lgh_ctxt->loc_handle == cathandle)
 317                 cathandle->lgh_ctxt->loc_handle = NULL;
 318         rc = llog_close(env, cathandle);
 319         RETURN(rc);
 320 }
 321 EXPORT_SYMBOL(llog_cat_close);
 322
 323 /**
 324  * lockdep markers for nested struct llog_handle::lgh_lock locking.
 325  */
 326 enum {
 327         LLOGH_CAT,
 328         LLOGH_LOG
 329 };
 330
 331 /** Return the currently active log handle.  If the current log handle doesn't
 332  * have enough space left for the current record, start a new one.
 333  *
 334  * If reclen is 0, we only want to know what the currently active log is,
 335  * otherwise we get a lock on this log so nobody can steal our space.
 336  *
 337  * Assumes caller has already pushed us into the kernel context and is locking.
 338  *
 339  * NOTE: loghandle is write-locked upon successful return
 340  */
 341 static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
 342                                                 struct thandle *th)
 343 {
 344         struct llog_handle *loghandle = NULL;
 345         ENTRY;
 346
 347
 348         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2)) {
 349                 down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 350                 GOTO(next, loghandle);
 351         }
 352
 353         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 354         loghandle = cathandle->u.chd.chd_current_log;
 355         if (loghandle) {
 356                 struct llog_log_hdr *llh;
 357
 358                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 359                 llh = loghandle->lgh_hdr;
 360                 if (llh == NULL || !llog_is_full(loghandle)) {
 361                         up_read(&cathandle->lgh_lock);
 362                         RETURN(loghandle);
 363                 } else {
 364                         up_write(&loghandle->lgh_lock);
 365                 }
 366         }
 367         up_read(&cathandle->lgh_lock);
 368
 369         /* time to use next log */
 370
 371         /* first, we have to make sure the state hasn't changed */
 372         down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 373         loghandle = cathandle->u.chd.chd_current_log;
 374         if (loghandle) {
 375                 struct llog_log_hdr *llh;
 376
 377                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 378                 llh = loghandle->lgh_hdr;
 379                 LASSERT(llh);
 380                 if (!llog_is_full(loghandle))
 381                         GOTO(out_unlock, loghandle);
 382                 else
 383                         up_write(&loghandle->lgh_lock);
 384         }
 385
 386 next:
 387         /* Sigh, the chd_next_log and chd_current_log is initialized
 388          * in declare phase, and we do not serialize the catlog
 389          * accessing, so it might be possible the llog creation
 390          * thread (see llog_cat_declare_add_rec()) did not create
 391          * llog successfully, then the following thread might
 392          * meet this situation. */
 393         if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
 394                 CERROR("%s: next log does not exist!\n",
 395                        cathandle->lgh_ctxt->loc_obd->obd_name);
 396                 loghandle = ERR_PTR(-EIO);
 397                 if (cathandle->u.chd.chd_next_log == NULL) {
 398                         /* Store the error in chd_next_log, so
 399                          * the following process can get correct
 400                          * failure value */
 401                         cathandle->u.chd.chd_next_log = loghandle;
 402                 }
 403                 GOTO(out_unlock, loghandle);
 404         }
 405
 406         CDEBUG(D_INODE, "use next log\n");
 407
 408         loghandle = cathandle->u.chd.chd_next_log;
 409         cathandle->u.chd.chd_current_log = loghandle;
 410         cathandle->u.chd.chd_next_log = NULL;
 411         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 412
 413 out_unlock:
 414         up_write(&cathandle->lgh_lock);
 415         LASSERT(loghandle);
 416         RETURN(loghandle);
 417 }
 418
 419 static int llog_cat_update_header(const struct lu_env *env,
 420                            struct llog_handle *cathandle)
 421 {
 422         struct llog_handle *loghandle;
 423         int rc;
 424         ENTRY;
 425
 426         /* refresh llog */
 427         down_write(&cathandle->lgh_lock);
 428         if (!cathandle->lgh_stale) {
 429                 up_write(&cathandle->lgh_lock);
 430                 RETURN(0);
 431         }
 432         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 433                             u.phd.phd_entry) {
 434                 if (!llog_exist(loghandle))
 435                         continue;
 436
 437                 rc = llog_read_header(env, loghandle, NULL);
 438                 if (rc != 0) {
 439                         up_write(&cathandle->lgh_lock);
 440                         GOTO(out, rc);
 441                 }
 442         }
 443         rc = llog_read_header(env, cathandle, NULL);
 444         if (rc == 0)
 445                 cathandle->lgh_stale = 0;
 446         up_write(&cathandle->lgh_lock);
 447         if (rc != 0)
 448                 GOTO(out, rc);
 449 out:
 450         RETURN(rc);
 451 }
 452
 453 /* Add a single record to the recovery log(s) using a catalog
 454  * Returns as llog_write_record
 455  *
 456  * Assumes caller has already pushed us into the kernel context.
 457  */
 458 int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
 459                      struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 460                      struct thandle *th)
 461 {
 462         struct llog_handle *loghandle;
 463         int rc, retried = 0;
 464         ENTRY;
 465
 466         LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
 467
 468 retry:
 469         loghandle = llog_cat_current_log(cathandle, th);
 470         if (IS_ERR(loghandle))
 471                 RETURN(PTR_ERR(loghandle));
 472
 473         /* loghandle is already locked by llog_cat_current_log() for us */
 474         if (!llog_exist(loghandle)) {
 475                 rc = llog_cat_new_log(env, cathandle, loghandle, th);
 476                 if (rc < 0) {
 477                         up_write(&loghandle->lgh_lock);
 478                         /* nobody should be trying to use this llog */
 479                         down_write(&cathandle->lgh_lock);
 480                         if (cathandle->u.chd.chd_current_log == loghandle)
 481                                 cathandle->u.chd.chd_current_log = NULL;
 482                         up_write(&cathandle->lgh_lock);
 483                         RETURN(rc);
 484                 }
 485         }
 486         /* now let's try to add the record */
 487         rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
 488         if (rc < 0) {
 489                 CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
 490                              "llog_write_rec %d: lh=%p\n", rc, loghandle);
 491                 /* -ENOSPC is returned if no empty records left
 492                  * and when it's lack of space on the stogage.
 493                  * there is no point to try again if it's the second
 494                  * case. many callers (like llog test) expect ENOSPC,
 495                  * so we preserve this error code, but look for the
 496                  * actual cause here */
 497                 if (rc == -ENOSPC && llog_is_full(loghandle))
 498                         rc = -ENOBUFS;
 499         }
 500         up_write(&loghandle->lgh_lock);
 501
 502         if (rc == -ENOBUFS) {
 503                 if (retried++ == 0)
 504                         GOTO(retry, rc);
 505                 CERROR("%s: error on 2nd llog: rc = %d\n",
 506                        cathandle->lgh_ctxt->loc_obd->obd_name, rc);
 507         }
 508
 509         RETURN(rc);
 510 }
 511 EXPORT_SYMBOL(llog_cat_add_rec);
 512
 513 int llog_cat_declare_add_rec(const struct lu_env *env,
 514                              struct llog_handle *cathandle,
 515                              struct llog_rec_hdr *rec, struct thandle *th)
 516 {
 517         struct llog_thread_info *lgi = llog_info(env);
 518         struct llog_logid_rec   *lirec = &lgi->lgi_logid;
 519         struct llog_handle      *loghandle, *next;
 520         int                      rc = 0;
 521
 522         ENTRY;
 523
 524         if (cathandle->u.chd.chd_current_log == NULL) {
 525                 /* declare new plain llog */
 526                 down_write(&cathandle->lgh_lock);
 527                 if (cathandle->u.chd.chd_current_log == NULL) {
 528                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
 529                                        NULL, NULL, LLOG_OPEN_NEW);
 530                         if (rc == 0) {
 531                                 cathandle->u.chd.chd_current_log = loghandle;
 532                                 list_add_tail(&loghandle->u.phd.phd_entry,
 533                                               &cathandle->u.chd.chd_head);
 534                         }
 535                 }
 536                 up_write(&cathandle->lgh_lock);
 537         } else if (cathandle->u.chd.chd_next_log == NULL ||
 538                    IS_ERR(cathandle->u.chd.chd_next_log)) {
 539                 /* declare next plain llog */
 540                 down_write(&cathandle->lgh_lock);
 541                 if (cathandle->u.chd.chd_next_log == NULL ||
 542                     IS_ERR(cathandle->u.chd.chd_next_log)) {
 543                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
 544                                        NULL, NULL, LLOG_OPEN_NEW);
 545                         if (rc == 0) {
 546                                 cathandle->u.chd.chd_next_log = loghandle;
 547                                 list_add_tail(&loghandle->u.phd.phd_entry,
 548                                               &cathandle->u.chd.chd_head);
 549                         }
 550                 }
 551                 up_write(&cathandle->lgh_lock);
 552         }
 553         if (rc)
 554                 GOTO(out, rc);
 555
 556         lirec->lid_hdr.lrh_len = sizeof(*lirec);
 557
 558         if (!llog_exist(cathandle->u.chd.chd_current_log)) {
 559                 if (dt_object_remote(cathandle->lgh_obj)) {
 560                         /* For remote operation, if we put the llog object
 561                          * creation in the current transaction, then the
 562                          * llog object will not be created on the remote
 563                          * target until the transaction stop, if other
 564                          * operations start before the transaction stop,
 565                          * and use the same llog object, will be dependent
 566                          * on the success of this transaction. So let's
 567                          * create the llog object synchronously here to
 568                          * remove the dependency. */
 569 create_again:
 570                         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 571                         loghandle = cathandle->u.chd.chd_current_log;
 572                         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 573                         if (cathandle->lgh_stale) {
 574                                 up_write(&loghandle->lgh_lock);
 575                                 up_read(&cathandle->lgh_lock);
 576                                 GOTO(out, rc = -EIO);
 577                         }
 578                         if (!llog_exist(loghandle)) {
 579                                 rc = llog_cat_new_log(env, cathandle, loghandle,
 580                                                       NULL);
 581                                 if (rc == -ESTALE)
 582                                         cathandle->lgh_stale = 1;
 583                         }
 584                         up_write(&loghandle->lgh_lock);
 585                         up_read(&cathandle->lgh_lock);
 586                         if (rc == -ESTALE) {
 587                                 rc = llog_cat_update_header(env, cathandle);
 588                                 if (rc != 0)
 589                                         GOTO(out, rc);
 590                                 goto create_again;
 591                         } else if (rc < 0) {
 592                                 GOTO(out, rc);
 593                         }
 594                 } else {
 595                         rc = llog_declare_create(env,
 596                                         cathandle->u.chd.chd_current_log, th);
 597                         if (rc)
 598                                 GOTO(out, rc);
 599                         llog_declare_write_rec(env, cathandle,
 600                                                &lirec->lid_hdr, -1, th);
 601                 }
 602         }
 603
 604 write_again:
 605         /* declare records in the llogs */
 606         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
 607                                     rec, -1, th);
 608         if (rc == -ESTALE) {
 609                 down_write(&cathandle->lgh_lock);
 610                 if (cathandle->lgh_stale) {
 611                         up_write(&cathandle->lgh_lock);
 612                         GOTO(out, rc = -EIO);
 613                 }
 614
 615                 cathandle->lgh_stale = 1;
 616                 up_write(&cathandle->lgh_lock);
 617                 rc = llog_cat_update_header(env, cathandle);
 618                 if (rc != 0)
 619                         GOTO(out, rc);
 620                 goto write_again;
 621         } else if (rc < 0) {
 622                 GOTO(out, rc);
 623         }
 624
 625         next = cathandle->u.chd.chd_next_log;
 626         if (!IS_ERR_OR_NULL(next)) {
 627                 if (!llog_exist(next)) {
 628                         if (dt_object_remote(cathandle->lgh_obj)) {
 629                                 /* For remote operation, if we put the llog
 630                                  * object creation in the current transaction,
 631                                  * then the llog object will not be created on
 632                                  * the remote target until the transaction stop,
 633                                  * if other operations start before the
 634                                  * transaction stop, and use the same llog
 635                                  * object, will be dependent on the success of
 636                                  * this transaction. So let's create the llog
 637                                  * object synchronously here to remove the
 638                                  * dependency. */
 639                                 down_write_nested(&cathandle->lgh_lock,
 640                                                  LLOGH_CAT);
 641                                 next = cathandle->u.chd.chd_next_log;
 642                                 if (IS_ERR_OR_NULL(next)) {
 643                                         /* Sigh, another thread just tried,
 644                                          * let's fail as well */
 645                                         up_write(&cathandle->lgh_lock);
 646                                         if (next == NULL)
 647                                                 rc = -EIO;
 648                                         else
 649                                                 rc = PTR_ERR(next);
 650                                         GOTO(out, rc);
 651                                 }
 652
 653                                 down_write_nested(&next->lgh_lock, LLOGH_LOG);
 654                                 if (!llog_exist(next)) {
 655                                         rc = llog_cat_new_log(env, cathandle,
 656                                                               next, NULL);
 657                                         if (rc < 0)
 658                                                 cathandle->u.chd.chd_next_log =
 659                                                                 ERR_PTR(rc);
 660                                 }
 661                                 up_write(&next->lgh_lock);
 662                                 up_write(&cathandle->lgh_lock);
 663                                 if (rc < 0)
 664                                         GOTO(out, rc);
 665                         } else {
 666                                 rc = llog_declare_create(env, next, th);
 667                                 llog_declare_write_rec(env, cathandle,
 668                                                 &lirec->lid_hdr, -1, th);
 669                         }
 670                 }
 671                 /* XXX: we hope for declarations made for existing llog
 672                  *      this might be not correct with some backends
 673                  *      where declarations are expected against specific
 674                  *      object like ZFS with full debugging enabled */
 675                 /*llog_declare_write_rec(env, next, rec, -1, th);*/
 676         }
 677 out:
 678         RETURN(rc);
 679 }
 680 EXPORT_SYMBOL(llog_cat_declare_add_rec);
 681
 682 int llog_cat_add(const struct lu_env *env, struct llog_handle *cathandle,
 683                  struct llog_rec_hdr *rec, struct llog_cookie *reccookie)
 684 {
 685         struct llog_ctxt        *ctxt;
 686         struct dt_device        *dt;
 687         struct thandle          *th = NULL;
 688         int                      rc;
 689
 690         ctxt = cathandle->lgh_ctxt;
 691         LASSERT(ctxt);
 692         LASSERT(ctxt->loc_exp);
 693
 694         LASSERT(cathandle->lgh_obj != NULL);
 695         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 696
 697         th = dt_trans_create(env, dt);
 698         if (IS_ERR(th))
 699                 RETURN(PTR_ERR(th));
 700
 701         rc = llog_cat_declare_add_rec(env, cathandle, rec, th);
 702         if (rc)
 703                 GOTO(out_trans, rc);
 704
 705         rc = dt_trans_start_local(env, dt, th);
 706         if (rc)
 707                 GOTO(out_trans, rc);
 708         rc = llog_cat_add_rec(env, cathandle, rec, reccookie, th);
 709 out_trans:
 710         dt_trans_stop(env, dt, th);
 711         RETURN(rc);
 712 }
 713 EXPORT_SYMBOL(llog_cat_add);
 714
 715 /* For each cookie in the cookie array, we clear the log in-use bit and either:
 716  * - the log is empty, so mark it free in the catalog header and delete it
 717  * - the log is not empty, just write out the log header
 718  *
 719  * The cookies may be in different log files, so we need to get new logs
 720  * each time.
 721  *
 722  * Assumes caller has already pushed us into the kernel context.
 723  */
 724 int llog_cat_cancel_records(const struct lu_env *env,
 725                             struct llog_handle *cathandle, int count,
 726                             struct llog_cookie *cookies)
 727 {
 728         int i, index, rc = 0, failed = 0;
 729
 730         ENTRY;
 731
 732         for (i = 0; i < count; i++, cookies++) {
 733                 struct llog_handle *loghandle;
 734                 struct llog_logid *lgl = &cookies->lgc_lgl;
 735                 int  lrc;
 736
 737                 rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
 738                 if (rc) {
 739                         CDEBUG(D_HA, "%s: cannot find llog for handle "DFID":%x"
 740                                ": rc = %d\n",
 741                                cathandle->lgh_ctxt->loc_obd->obd_name,
 742                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, rc);
 743                         failed++;
 744                         continue;
 745                 }
 746
 747                 lrc = llog_cancel_rec(env, loghandle, cookies->lgc_index);
 748                 if (lrc == LLOG_DEL_PLAIN) { /* log has been destroyed */
 749                         index = loghandle->u.phd.phd_cookie.lgc_index;
 750                         rc = llog_cat_cleanup(env, cathandle, loghandle, index);
 751                 } else if (lrc == -ENOENT) {
 752                         if (rc == 0) /* ENOENT shouldn't rewrite any error */
 753                                 rc = lrc;
 754                 } else if (lrc < 0) {
 755                         failed++;
 756                         rc = lrc;
 757                 }
 758                 llog_handle_put(loghandle);
 759         }
 760         if (rc)
 761                 CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
 762                        cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
 763                        rc);
 764
 765         RETURN(rc);
 766 }
 767 EXPORT_SYMBOL(llog_cat_cancel_records);
 768
 769 static int llog_cat_process_cb(const struct lu_env *env,
 770                                struct llog_handle *cat_llh,
 771                                struct llog_rec_hdr *rec, void *data)
 772 {
 773         struct llog_process_data *d = data;
 774         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
 775         struct llog_handle *llh;
 776         struct llog_log_hdr *hdr;
 777         int rc;
 778
 779         ENTRY;
 780         if (rec->lrh_type != LLOG_LOGID_MAGIC) {
 781                 CERROR("invalid record in catalog\n");
 782                 RETURN(-EINVAL);
 783         }
 784         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "
 785                DFID"\n", PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
 786                rec->lrh_index, PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 787
 788         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
 789         if (rc) {
 790                 CERROR("%s: cannot find handle for llog "DFID": rc = %d\n",
 791                        cat_llh->lgh_ctxt->loc_obd->obd_name,
 792                        PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
 793                 if (rc == -ENOENT || rc == -ESTALE) {
 794                         /* After a server crash, a stub of index
 795                          * record in catlog could be kept, because
 796                          * plain log destroy + catlog index record
 797                          * deletion are not atomic. So we end up with
 798                          * an index but no actual record. Destroy the
 799                          * index and move on. */
 800                         rc = llog_cat_cleanup(env, cat_llh, NULL,
 801                                               rec->lrh_index);
 802                 }
 803
 804                 RETURN(rc);
 805         }
 806
 807         /* clean old empty llogs, do not consider current llog in use */
 808         /* ignore remote (lgh_obj=NULL) llogs */
 809         hdr = llh->lgh_hdr;
 810         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 811             hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
 812             llh != cat_llh->u.chd.chd_current_log) {
 813                 rc = llog_destroy(env, llh);
 814                 if (rc)
 815                         CERROR("%s: fail to destroy empty log: rc = %d\n",
 816                                llh->lgh_ctxt->loc_obd->obd_name, rc);
 817                 GOTO(out, rc = LLOG_DEL_PLAIN);
 818         }
 819
 820         if (rec->lrh_index < d->lpd_startcat) {
 821                 /* Skip processing of the logs until startcat */
 822                 rc = 0;
 823         } else if (d->lpd_startidx > 0) {
 824                 struct llog_process_cat_data cd;
 825
 826                 cd.lpcd_first_idx = d->lpd_startidx;
 827                 cd.lpcd_last_idx = 0;
 828                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 829                                           &cd, false);
 830                 /* Continue processing the next log from idx 0 */
 831                 d->lpd_startidx = 0;
 832         } else {
 833                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 834                                           NULL, false);
 835         }
 836
 837 out:
 838         /* The empty plain log was destroyed while processing */
 839         if (rc == LLOG_DEL_PLAIN)
 840                 rc = llog_cat_cleanup(env, cat_llh, llh,
 841                                       llh->u.phd.phd_cookie.lgc_index);
 842         llog_handle_put(llh);
 843
 844         RETURN(rc);
 845 }
 846
 847 int llog_cat_process_or_fork(const struct lu_env *env,
 848                              struct llog_handle *cat_llh, llog_cb_t cat_cb,
 849                              llog_cb_t cb, void *data, int startcat,
 850                              int startidx, bool fork)
 851 {
 852         struct llog_process_data d;
 853         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
 854         int rc;
 855         ENTRY;
 856
 857         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
 858         d.lpd_data = data;
 859         d.lpd_cb = cb;
 860         d.lpd_startcat = startcat;
 861         d.lpd_startidx = startidx;
 862
 863         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
 864             llh->llh_count > 1) {
 865                 struct llog_process_cat_data cd;
 866
 867                 CWARN("%s: catlog "DFID" crosses index zero\n",
 868                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 869                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 870
 871                 cd.lpcd_first_idx = llh->llh_cat_idx;
 872                 cd.lpcd_last_idx = 0;
 873                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 874                                           &d, &cd, fork);
 875                 if (rc != 0)
 876                         RETURN(rc);
 877
 878                 cd.lpcd_first_idx = 0;
 879                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
 880                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 881                                           &d, &cd, fork);
 882         } else {
 883                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 884                                           &d, NULL, fork);
 885         }
 886
 887         RETURN(rc);
 888 }
 889
 890 int llog_cat_process(const struct lu_env *env, struct llog_handle *cat_llh,
 891                      llog_cb_t cb, void *data, int startcat, int startidx)
 892 {
 893         return llog_cat_process_or_fork(env, cat_llh, llog_cat_process_cb,
 894                                         cb, data, startcat, startidx, false);
 895 }
 896 EXPORT_SYMBOL(llog_cat_process);
 897
 898 static int llog_cat_size_cb(const struct lu_env *env,
 899                              struct llog_handle *cat_llh,
 900                              struct llog_rec_hdr *rec, void *data)
 901 {
 902         struct llog_process_data *d = data;
 903         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
 904         struct llog_handle *llh;
 905         int rc;
 906         __u64 *cum_size = d->lpd_data;
 907         __u64 size;
 908
 909         ENTRY;
 910         if (rec->lrh_type != LLOG_LOGID_MAGIC) {
 911                 CERROR("%s: invalid record in catalog, rc = %d\n",
 912                        cat_llh->lgh_ctxt->loc_obd->obd_name, -EINVAL);
 913                 RETURN(-EINVAL);
 914         }
 915         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "
 916                DFID"\n", PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
 917                rec->lrh_index, PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 918
 919         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
 920         if (rc) {
 921                 CWARN("%s: cannot find handle for llog "DFID": rc = %d\n",
 922                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 923                       PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
 924                 RETURN(0);
 925         }
 926         size = llog_size(env, llh);
 927         *cum_size += size;
 928
 929         CDEBUG(D_INFO, "Add llog entry "DFID" size %llu\n",
 930                PFID(&llh->lgh_id.lgl_oi.oi_fid), size);
 931
 932         llog_handle_put(llh);
 933
 934         RETURN(0);
 935
 936 }
 937
 938 __u64 llog_cat_size(const struct lu_env *env, struct llog_handle *cat_llh)
 939 {
 940         __u64 size = llog_size(env, cat_llh);
 941
 942         llog_cat_process_or_fork(env, cat_llh, llog_cat_size_cb,
 943                                  NULL, &size, 0, 0, false);
 944
 945         return size;
 946 }
 947 EXPORT_SYMBOL(llog_cat_size);
 948
 949 static int llog_cat_reverse_process_cb(const struct lu_env *env,
 950                                        struct llog_handle *cat_llh,
 951                                        struct llog_rec_hdr *rec, void *data)
 952 {
 953         struct llog_process_data *d = data;
 954         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
 955         struct llog_handle *llh;
 956         struct llog_log_hdr *hdr;
 957         int rc;
 958
 959         if (le32_to_cpu(rec->lrh_type) != LLOG_LOGID_MAGIC) {
 960                 CERROR("invalid record in catalog\n");
 961                 RETURN(-EINVAL);
 962         }
 963         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "
 964                DFID"\n", PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
 965                le32_to_cpu(rec->lrh_index),
 966                PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 967
 968         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
 969         if (rc) {
 970                 CERROR("%s: cannot find handle for llog "DFID": rc = %d\n",
 971                        cat_llh->lgh_ctxt->loc_obd->obd_name,
 972                        PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
 973                 if (rc == -ENOENT || rc == -ESTALE) {
 974                         /* After a server crash, a stub of index
 975                          * record in catlog could be kept, because
 976                          * plain log destroy + catlog index record
 977                          * deletion are not atomic. So we end up with
 978                          * an index but no actual record. Destroy the
 979                          * index and move on. */
 980                         rc = llog_cat_cleanup(env, cat_llh, NULL,
 981                                               rec->lrh_index);
 982                 }
 983
 984                 RETURN(rc);
 985         }
 986
 987         /* clean old empty llogs, do not consider current llog in use */
 988         hdr = llh->lgh_hdr;
 989         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 990             hdr->llh_count == 1 &&
 991             llh != cat_llh->u.chd.chd_current_log) {
 992                 rc = llog_destroy(env, llh);
 993                 if (rc)
 994                         CERROR("%s: fail to destroy empty log: rc = %d\n",
 995                                llh->lgh_ctxt->loc_obd->obd_name, rc);
 996                 GOTO(out, rc = LLOG_DEL_PLAIN);
 997         }
 998
 999         rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
1000
1001 out:
1002         /* The empty plain was destroyed while processing */
1003         if (rc == LLOG_DEL_PLAIN)
1004                 rc = llog_cat_cleanup(env, cat_llh, llh,
1005                                       llh->u.phd.phd_cookie.lgc_index);
1006
1007         llog_handle_put(llh);
1008         RETURN(rc);
1009 }
1010
1011 int llog_cat_reverse_process(const struct lu_env *env,
1012                              struct llog_handle *cat_llh,
1013                              llog_cb_t cb, void *data)
1014 {
1015         struct llog_process_data d;
1016         struct llog_process_cat_data cd;
1017         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
1018         int rc;
1019         ENTRY;
1020
1021         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
1022         d.lpd_data = data;
1023         d.lpd_cb = cb;
1024
1025         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
1026             llh->llh_count > 1) {
1027                 CWARN("%s: catalog "DFID" crosses index zero\n",
1028                       cat_llh->lgh_ctxt->loc_obd->obd_name,
1029                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
1030
1031                 cd.lpcd_first_idx = 0;
1032                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
1033                 rc = llog_reverse_process(env, cat_llh,
1034                                           llog_cat_reverse_process_cb,
1035                                           &d, &cd);
1036                 if (rc != 0)
1037                         RETURN(rc);
1038
1039                 cd.lpcd_first_idx = le32_to_cpu(llh->llh_cat_idx);
1040                 cd.lpcd_last_idx = 0;
1041                 rc = llog_reverse_process(env, cat_llh,
1042                                           llog_cat_reverse_process_cb,
1043                                           &d, &cd);
1044         } else {
1045                 rc = llog_reverse_process(env, cat_llh,
1046                                           llog_cat_reverse_process_cb,
1047                                           &d, NULL);
1048         }
1049
1050         RETURN(rc);
1051 }
1052 EXPORT_SYMBOL(llog_cat_reverse_process);
1053
1054 static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
1055 {
1056         struct llog_log_hdr *llh = cathandle->lgh_hdr;
1057         int bitmap_size;
1058
1059         ENTRY;
1060
1061         bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
1062         /*
1063          * The llh_cat_idx equals to the first used index minus 1
1064          * so if we canceled the first index then llh_cat_idx
1065          * must be renewed.
1066          */
1067         if (llh->llh_cat_idx == (idx - 1)) {
1068                 llh->llh_cat_idx = idx;
1069
1070                 while (idx != cathandle->lgh_last_idx) {
1071                         idx = (idx + 1) % bitmap_size;
1072                         if (!ext2_test_bit(idx, LLOG_HDR_BITMAP(llh))) {
1073                                 /* update llh_cat_idx for each unset bit,
1074                                  * expecting the next one is set */
1075                                 llh->llh_cat_idx = idx;
1076                         } else if (idx == 0) {
1077                                 /* skip header bit */
1078                                 llh->llh_cat_idx = 0;
1079                                 continue;
1080                         } else {
1081                                 /* the first index is found */
1082                                 break;
1083                         }
1084                 }
1085
1086                 CDEBUG(D_RPCTRACE, "catlog "DFID" first idx %u, last_idx %u\n",
1087                        PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
1088                        llh->llh_cat_idx, cathandle->lgh_last_idx);
1089         }
1090
1091         RETURN(0);
1092 }
1093
1094 /* Cleanup deleted plain llog traces from catalog */
1095 int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
1096                      struct llog_handle *loghandle, int index)
1097 {
1098         int rc;
1099
1100         LASSERT(index);
1101         if (loghandle != NULL) {
1102                 /* remove destroyed llog from catalog list and
1103                  * chd_current_log variable */
1104                 down_write(&cathandle->lgh_lock);
1105                 if (cathandle->u.chd.chd_current_log == loghandle)
1106                         cathandle->u.chd.chd_current_log = NULL;
1107                 list_del_init(&loghandle->u.phd.phd_entry);
1108                 up_write(&cathandle->lgh_lock);
1109                 LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
1110                 /* llog was opened and keep in a list, close it now */
1111                 llog_close(env, loghandle);
1112         }
1113
1114         /* do not attempt to cleanup on-disk llog if on client side */
1115         if (cathandle->lgh_obj == NULL)
1116                 return 0;
1117
1118         /* remove plain llog entry from catalog by index */
1119         llog_cat_set_first_idx(cathandle, index);
1120         rc = llog_cancel_rec(env, cathandle, index);
1121         if (rc == 0)
1122                 CDEBUG(D_HA, "cancel plain log at index %u of catalog "DFID"\n",
1123                        index, PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
1124         return rc;
1125 }