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, 2015, 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 "DOSTID":%x for index %u of catalog"
 169                DOSTID"\n", POSTID(&loghandle->lgh_id.lgl_oi),
 170                loghandle->lgh_id.lgl_ogen, rec->lid_hdr.lrh_index,
 171                POSTID(&cathandle->lgh_id.lgl_oi));
 172
 173         loghandle->lgh_hdr->llh_cat_idx = rec->lid_hdr.lrh_index;
 174
 175         /* limit max size of plain llog so that space can be
 176          * released sooner, especially on small filesystems */
 177         /* 2MB for the cases when free space hasn't been learned yet */
 178         loghandle->lgh_max_size = 2 << 20;
 179         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 180         rc = dt_statfs(env, dt, &lgi->lgi_statfs);
 181         if (rc == 0 && lgi->lgi_statfs.os_bfree > 0) {
 182                 __u64 freespace = (lgi->lgi_statfs.os_bfree *
 183                                   lgi->lgi_statfs.os_bsize) >> 6;
 184                 if (freespace < loghandle->lgh_max_size)
 185                         loghandle->lgh_max_size = freespace;
 186                 /* shouldn't be > 128MB in any case?
 187                  * it's 256K records of 512 bytes each */
 188                 if (freespace > (128 << 20))
 189                         loghandle->lgh_max_size = 128 << 20;
 190         }
 191         rc = 0;
 192
 193 out:
 194         if (handle != NULL) {
 195                 handle->th_result = rc >= 0 ? 0 : rc;
 196                 dt_trans_stop(env, dt, handle);
 197         }
 198         RETURN(rc);
 199
 200 out_destroy:
 201         /* to signal llog_cat_close() it shouldn't try to destroy the llog,
 202          * we want to destroy it in this transaction, otherwise the object
 203          * becomes an orphan */
 204         loghandle->lgh_hdr->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 205         /* this is to mimic full log, so another llog_cat_current_log()
 206          * can skip it and ask for another onet */
 207         loghandle->lgh_last_idx = LLOG_HDR_BITMAP_SIZE(llh) + 1;
 208         llog_trans_destroy(env, loghandle, th);
 209         RETURN(rc);
 210 }
 211
 212 /* Open an existent log handle and add it to the open list.
 213  * This log handle will be closed when all of the records in it are removed.
 214  *
 215  * Assumes caller has already pushed us into the kernel context and is locking.
 216  * We return a lock on the handle to ensure nobody yanks it from us.
 217  *
 218  * This takes extra reference on llog_handle via llog_handle_get() and require
 219  * this reference to be put by caller using llog_handle_put()
 220  */
 221 int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
 222                        struct llog_handle **res, struct llog_logid *logid)
 223 {
 224         struct llog_handle      *loghandle;
 225         enum llog_flag           fmt;
 226         int                      rc = 0;
 227
 228         ENTRY;
 229
 230         if (cathandle == NULL)
 231                 RETURN(-EBADF);
 232
 233         fmt = cathandle->lgh_hdr->llh_flags & LLOG_F_EXT_MASK;
 234         down_write(&cathandle->lgh_lock);
 235         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 236                             u.phd.phd_entry) {
 237                 struct llog_logid *cgl = &loghandle->lgh_id;
 238
 239                 if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
 240                     ostid_seq(&cgl->lgl_oi) == ostid_seq(&logid->lgl_oi)) {
 241                         if (cgl->lgl_ogen != logid->lgl_ogen) {
 242                                 CERROR("%s: log "DOSTID" generation %x != %x\n",
 243                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 244                                        POSTID(&logid->lgl_oi), cgl->lgl_ogen,
 245                                        logid->lgl_ogen);
 246                                 continue;
 247                         }
 248                         loghandle->u.phd.phd_cat_handle = cathandle;
 249                         up_write(&cathandle->lgh_lock);
 250                         GOTO(out, rc = 0);
 251                 }
 252         }
 253         up_write(&cathandle->lgh_lock);
 254
 255         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle, logid, NULL,
 256                        LLOG_OPEN_EXISTS);
 257         if (rc < 0) {
 258                 CERROR("%s: error opening log id "DOSTID":%x: rc = %d\n",
 259                        cathandle->lgh_ctxt->loc_obd->obd_name,
 260                        POSTID(&logid->lgl_oi), logid->lgl_ogen, rc);
 261                 RETURN(rc);
 262         }
 263
 264         rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN | fmt, NULL);
 265         if (rc < 0) {
 266                 llog_close(env, loghandle);
 267                 loghandle = NULL;
 268                 RETURN(rc);
 269         }
 270
 271         down_write(&cathandle->lgh_lock);
 272         list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
 273         up_write(&cathandle->lgh_lock);
 274
 275         loghandle->u.phd.phd_cat_handle = cathandle;
 276         loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
 277         loghandle->u.phd.phd_cookie.lgc_index =
 278                                 loghandle->lgh_hdr->llh_cat_idx;
 279         EXIT;
 280 out:
 281         llog_handle_get(loghandle);
 282         *res = loghandle;
 283         return 0;
 284 }
 285
 286 int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
 287 {
 288         struct llog_handle      *loghandle, *n;
 289         int                      rc;
 290
 291         ENTRY;
 292
 293         list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
 294                                  u.phd.phd_entry) {
 295                 struct llog_log_hdr     *llh = loghandle->lgh_hdr;
 296                 int                      index;
 297
 298                 /* unlink open-not-created llogs */
 299                 list_del_init(&loghandle->u.phd.phd_entry);
 300                 llh = loghandle->lgh_hdr;
 301                 if (loghandle->lgh_obj != NULL && llh != NULL &&
 302                     (llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 303                     (llh->llh_count == 1)) {
 304                         rc = llog_destroy(env, loghandle);
 305                         if (rc)
 306                                 CERROR("%s: failure destroying log during "
 307                                        "cleanup: rc = %d\n",
 308                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 309                                        rc);
 310
 311                         index = loghandle->u.phd.phd_cookie.lgc_index;
 312                         llog_cat_cleanup(env, cathandle, NULL, index);
 313                 }
 314                 llog_close(env, loghandle);
 315         }
 316         /* if handle was stored in ctxt, remove it too */
 317         if (cathandle->lgh_ctxt->loc_handle == cathandle)
 318                 cathandle->lgh_ctxt->loc_handle = NULL;
 319         rc = llog_close(env, cathandle);
 320         RETURN(rc);
 321 }
 322 EXPORT_SYMBOL(llog_cat_close);
 323
 324 /**
 325  * lockdep markers for nested struct llog_handle::lgh_lock locking.
 326  */
 327 enum {
 328         LLOGH_CAT,
 329         LLOGH_LOG
 330 };
 331
 332 /** Return the currently active log handle.  If the current log handle doesn't
 333  * have enough space left for the current record, start a new one.
 334  *
 335  * If reclen is 0, we only want to know what the currently active log is,
 336  * otherwise we get a lock on this log so nobody can steal our space.
 337  *
 338  * Assumes caller has already pushed us into the kernel context and is locking.
 339  *
 340  * NOTE: loghandle is write-locked upon successful return
 341  */
 342 static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
 343                                                 struct thandle *th)
 344 {
 345         struct llog_handle *loghandle = NULL;
 346         ENTRY;
 347
 348
 349         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2)) {
 350                 down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 351                 GOTO(next, loghandle);
 352         }
 353
 354         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 355         loghandle = cathandle->u.chd.chd_current_log;
 356         if (loghandle) {
 357                 struct llog_log_hdr *llh;
 358
 359                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 360                 llh = loghandle->lgh_hdr;
 361                 if (llh == NULL || !llog_is_full(loghandle)) {
 362                         up_read(&cathandle->lgh_lock);
 363                         RETURN(loghandle);
 364                 } else {
 365                         up_write(&loghandle->lgh_lock);
 366                 }
 367         }
 368         up_read(&cathandle->lgh_lock);
 369
 370         /* time to use next log */
 371
 372         /* first, we have to make sure the state hasn't changed */
 373         down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 374         loghandle = cathandle->u.chd.chd_current_log;
 375         if (loghandle) {
 376                 struct llog_log_hdr *llh;
 377
 378                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 379                 llh = loghandle->lgh_hdr;
 380                 LASSERT(llh);
 381                 if (!llog_is_full(loghandle))
 382                         GOTO(out_unlock, loghandle);
 383                 else
 384                         up_write(&loghandle->lgh_lock);
 385         }
 386
 387 next:
 388         /* Sigh, the chd_next_log and chd_current_log is initialized
 389          * in declare phase, and we do not serialize the catlog
 390          * accessing, so it might be possible the llog creation
 391          * thread (see llog_cat_declare_add_rec()) did not create
 392          * llog successfully, then the following thread might
 393          * meet this situation. */
 394         if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
 395                 CERROR("%s: next log does not exist!\n",
 396                        cathandle->lgh_ctxt->loc_obd->obd_name);
 397                 loghandle = ERR_PTR(-EIO);
 398                 if (cathandle->u.chd.chd_next_log == NULL) {
 399                         /* Store the error in chd_next_log, so
 400                          * the following process can get correct
 401                          * failure value */
 402                         cathandle->u.chd.chd_next_log = loghandle;
 403                 }
 404                 GOTO(out_unlock, loghandle);
 405         }
 406
 407         CDEBUG(D_INODE, "use next log\n");
 408
 409         loghandle = cathandle->u.chd.chd_next_log;
 410         cathandle->u.chd.chd_current_log = loghandle;
 411         cathandle->u.chd.chd_next_log = NULL;
 412         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 413
 414 out_unlock:
 415         up_write(&cathandle->lgh_lock);
 416         LASSERT(loghandle);
 417         RETURN(loghandle);
 418 }
 419
 420 static int llog_cat_update_header(const struct lu_env *env,
 421                            struct llog_handle *cathandle)
 422 {
 423         struct llog_handle *loghandle;
 424         int rc;
 425         ENTRY;
 426
 427         /* refresh llog */
 428         down_write(&cathandle->lgh_lock);
 429         if (!cathandle->lgh_stale) {
 430                 up_write(&cathandle->lgh_lock);
 431                 RETURN(0);
 432         }
 433         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 434                             u.phd.phd_entry) {
 435                 if (!llog_exist(loghandle))
 436                         continue;
 437
 438                 rc = llog_read_header(env, loghandle, NULL);
 439                 if (rc != 0) {
 440                         up_write(&cathandle->lgh_lock);
 441                         GOTO(out, rc);
 442                 }
 443         }
 444         rc = llog_read_header(env, cathandle, NULL);
 445         if (rc == 0)
 446                 cathandle->lgh_stale = 0;
 447         up_write(&cathandle->lgh_lock);
 448         if (rc != 0)
 449                 GOTO(out, rc);
 450 out:
 451         RETURN(rc);
 452 }
 453
 454 /* Add a single record to the recovery log(s) using a catalog
 455  * Returns as llog_write_record
 456  *
 457  * Assumes caller has already pushed us into the kernel context.
 458  */
 459 int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
 460                      struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 461                      struct thandle *th)
 462 {
 463         struct llog_handle *loghandle;
 464         int rc, retried = 0;
 465         ENTRY;
 466
 467         LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
 468
 469 retry:
 470         loghandle = llog_cat_current_log(cathandle, th);
 471         if (IS_ERR(loghandle))
 472                 RETURN(PTR_ERR(loghandle));
 473
 474         /* loghandle is already locked by llog_cat_current_log() for us */
 475         if (!llog_exist(loghandle)) {
 476                 rc = llog_cat_new_log(env, cathandle, loghandle, th);
 477                 if (rc < 0) {
 478                         up_write(&loghandle->lgh_lock);
 479                         /* nobody should be trying to use this llog */
 480                         down_write(&cathandle->lgh_lock);
 481                         if (cathandle->u.chd.chd_current_log == loghandle)
 482                                 cathandle->u.chd.chd_current_log = NULL;
 483                         up_write(&cathandle->lgh_lock);
 484                         RETURN(rc);
 485                 }
 486         }
 487         /* now let's try to add the record */
 488         rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
 489         if (rc < 0) {
 490                 CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
 491                              "llog_write_rec %d: lh=%p\n", rc, loghandle);
 492                 /* -ENOSPC is returned if no empty records left
 493                  * and when it's lack of space on the stogage.
 494                  * there is no point to try again if it's the second
 495                  * case. many callers (like llog test) expect ENOSPC,
 496                  * so we preserve this error code, but look for the
 497                  * actual cause here */
 498                 if (rc == -ENOSPC && llog_is_full(loghandle))
 499                         rc = -ENOBUFS;
 500         }
 501         up_write(&loghandle->lgh_lock);
 502
 503         if (rc == -ENOBUFS) {
 504                 if (retried++ == 0)
 505                         GOTO(retry, rc);
 506                 CERROR("%s: error on 2nd llog: rc = %d\n",
 507                        cathandle->lgh_ctxt->loc_obd->obd_name, rc);
 508         }
 509
 510         RETURN(rc);
 511 }
 512 EXPORT_SYMBOL(llog_cat_add_rec);
 513
 514 int llog_cat_declare_add_rec(const struct lu_env *env,
 515                              struct llog_handle *cathandle,
 516                              struct llog_rec_hdr *rec, struct thandle *th)
 517 {
 518         struct llog_thread_info *lgi = llog_info(env);
 519         struct llog_logid_rec   *lirec = &lgi->lgi_logid;
 520         struct llog_handle      *loghandle, *next;
 521         int                      rc = 0;
 522
 523         ENTRY;
 524
 525         if (cathandle->u.chd.chd_current_log == NULL) {
 526                 /* declare new plain llog */
 527                 down_write(&cathandle->lgh_lock);
 528                 if (cathandle->u.chd.chd_current_log == NULL) {
 529                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
 530                                        NULL, NULL, LLOG_OPEN_NEW);
 531                         if (rc == 0) {
 532                                 cathandle->u.chd.chd_current_log = loghandle;
 533                                 list_add_tail(&loghandle->u.phd.phd_entry,
 534                                               &cathandle->u.chd.chd_head);
 535                         }
 536                 }
 537                 up_write(&cathandle->lgh_lock);
 538         } else if (cathandle->u.chd.chd_next_log == NULL ||
 539                    IS_ERR(cathandle->u.chd.chd_next_log)) {
 540                 /* declare next plain llog */
 541                 down_write(&cathandle->lgh_lock);
 542                 if (cathandle->u.chd.chd_next_log == NULL ||
 543                     IS_ERR(cathandle->u.chd.chd_next_log)) {
 544                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
 545                                        NULL, NULL, LLOG_OPEN_NEW);
 546                         if (rc == 0) {
 547                                 cathandle->u.chd.chd_next_log = loghandle;
 548                                 list_add_tail(&loghandle->u.phd.phd_entry,
 549                                               &cathandle->u.chd.chd_head);
 550                         }
 551                 }
 552                 up_write(&cathandle->lgh_lock);
 553         }
 554         if (rc)
 555                 GOTO(out, rc);
 556
 557         lirec->lid_hdr.lrh_len = sizeof(*lirec);
 558
 559         if (!llog_exist(cathandle->u.chd.chd_current_log)) {
 560                 if (dt_object_remote(cathandle->lgh_obj)) {
 561                         /* For remote operation, if we put the llog object
 562                          * creation in the current transaction, then the
 563                          * llog object will not be created on the remote
 564                          * target until the transaction stop, if other
 565                          * operations start before the transaction stop,
 566                          * and use the same llog object, will be dependent
 567                          * on the success of this transaction. So let's
 568                          * create the llog object synchronously here to
 569                          * remove the dependency. */
 570 create_again:
 571                         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 572                         loghandle = cathandle->u.chd.chd_current_log;
 573                         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 574                         if (cathandle->lgh_stale) {
 575                                 up_write(&loghandle->lgh_lock);
 576                                 up_read(&cathandle->lgh_lock);
 577                                 GOTO(out, rc = -EIO);
 578                         }
 579                         if (!llog_exist(loghandle)) {
 580                                 rc = llog_cat_new_log(env, cathandle, loghandle,
 581                                                       NULL);
 582                                 if (rc == -ESTALE)
 583                                         cathandle->lgh_stale = 1;
 584                         }
 585                         up_write(&loghandle->lgh_lock);
 586                         up_read(&cathandle->lgh_lock);
 587                         if (rc == -ESTALE) {
 588                                 rc = llog_cat_update_header(env, cathandle);
 589                                 if (rc != 0)
 590                                         GOTO(out, rc);
 591                                 goto create_again;
 592                         } else if (rc < 0) {
 593                                 GOTO(out, rc);
 594                         }
 595                 } else {
 596                         rc = llog_declare_create(env,
 597                                         cathandle->u.chd.chd_current_log, th);
 598                         if (rc)
 599                                 GOTO(out, rc);
 600                         llog_declare_write_rec(env, cathandle,
 601                                                &lirec->lid_hdr, -1, th);
 602                 }
 603         }
 604
 605 write_again:
 606         /* declare records in the llogs */
 607         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
 608                                     rec, -1, th);
 609         if (rc == -ESTALE) {
 610                 down_write(&cathandle->lgh_lock);
 611                 if (cathandle->lgh_stale) {
 612                         up_write(&cathandle->lgh_lock);
 613                         GOTO(out, rc = -EIO);
 614                 }
 615
 616                 cathandle->lgh_stale = 1;
 617                 up_write(&cathandle->lgh_lock);
 618                 rc = llog_cat_update_header(env, cathandle);
 619                 if (rc != 0)
 620                         GOTO(out, rc);
 621                 goto write_again;
 622         } else if (rc < 0) {
 623                 GOTO(out, rc);
 624         }
 625
 626         next = cathandle->u.chd.chd_next_log;
 627         if (!IS_ERR_OR_NULL(next)) {
 628                 if (!llog_exist(next)) {
 629                         if (dt_object_remote(cathandle->lgh_obj)) {
 630                                 /* For remote operation, if we put the llog
 631                                  * object creation in the current transaction,
 632                                  * then the llog object will not be created on
 633                                  * the remote target until the transaction stop,
 634                                  * if other operations start before the
 635                                  * transaction stop, and use the same llog
 636                                  * object, will be dependent on the success of
 637                                  * this transaction. So let's create the llog
 638                                  * object synchronously here to remove the
 639                                  * dependency. */
 640                                 down_write_nested(&cathandle->lgh_lock,
 641                                                  LLOGH_CAT);
 642                                 next = cathandle->u.chd.chd_next_log;
 643                                 if (IS_ERR_OR_NULL(next)) {
 644                                         /* Sigh, another thread just tried,
 645                                          * let's fail as well */
 646                                         up_write(&cathandle->lgh_lock);
 647                                         if (next == NULL)
 648                                                 rc = -EIO;
 649                                         else
 650                                                 rc = PTR_ERR(next);
 651                                         GOTO(out, rc);
 652                                 }
 653
 654                                 down_write_nested(&next->lgh_lock, LLOGH_LOG);
 655                                 if (!llog_exist(next)) {
 656                                         rc = llog_cat_new_log(env, cathandle,
 657                                                               next, NULL);
 658                                         if (rc < 0)
 659                                                 cathandle->u.chd.chd_next_log =
 660                                                                 ERR_PTR(rc);
 661                                 }
 662                                 up_write(&next->lgh_lock);
 663                                 up_write(&cathandle->lgh_lock);
 664                                 if (rc < 0)
 665                                         GOTO(out, rc);
 666                         } else {
 667                                 rc = llog_declare_create(env, next, th);
 668                                 llog_declare_write_rec(env, cathandle,
 669                                                 &lirec->lid_hdr, -1, th);
 670                         }
 671                 }
 672                 /* XXX: we hope for declarations made for existing llog
 673                  *      this might be not correct with some backends
 674                  *      where declarations are expected against specific
 675                  *      object like ZFS with full debugging enabled */
 676                 /*llog_declare_write_rec(env, next, rec, -1, th);*/
 677         }
 678 out:
 679         RETURN(rc);
 680 }
 681 EXPORT_SYMBOL(llog_cat_declare_add_rec);
 682
 683 int llog_cat_add(const struct lu_env *env, struct llog_handle *cathandle,
 684                  struct llog_rec_hdr *rec, struct llog_cookie *reccookie)
 685 {
 686         struct llog_ctxt        *ctxt;
 687         struct dt_device        *dt;
 688         struct thandle          *th = NULL;
 689         int                      rc;
 690
 691         ctxt = cathandle->lgh_ctxt;
 692         LASSERT(ctxt);
 693         LASSERT(ctxt->loc_exp);
 694
 695         LASSERT(cathandle->lgh_obj != NULL);
 696         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 697
 698         th = dt_trans_create(env, dt);
 699         if (IS_ERR(th))
 700                 RETURN(PTR_ERR(th));
 701
 702         rc = llog_cat_declare_add_rec(env, cathandle, rec, th);
 703         if (rc)
 704                 GOTO(out_trans, rc);
 705
 706         rc = dt_trans_start_local(env, dt, th);
 707         if (rc)
 708                 GOTO(out_trans, rc);
 709         rc = llog_cat_add_rec(env, cathandle, rec, reccookie, th);
 710 out_trans:
 711         dt_trans_stop(env, dt, th);
 712         RETURN(rc);
 713 }
 714 EXPORT_SYMBOL(llog_cat_add);
 715
 716 /* For each cookie in the cookie array, we clear the log in-use bit and either:
 717  * - the log is empty, so mark it free in the catalog header and delete it
 718  * - the log is not empty, just write out the log header
 719  *
 720  * The cookies may be in different log files, so we need to get new logs
 721  * each time.
 722  *
 723  * Assumes caller has already pushed us into the kernel context.
 724  */
 725 int llog_cat_cancel_records(const struct lu_env *env,
 726                             struct llog_handle *cathandle, int count,
 727                             struct llog_cookie *cookies)
 728 {
 729         int i, index, rc = 0, failed = 0;
 730
 731         ENTRY;
 732
 733         for (i = 0; i < count; i++, cookies++) {
 734                 struct llog_handle      *loghandle;
 735                 struct llog_logid       *lgl = &cookies->lgc_lgl;
 736                 int                      lrc;
 737
 738                 rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
 739                 if (rc) {
 740                         CERROR("%s: cannot find handle for llog "DOSTID": %d\n",
 741                                cathandle->lgh_ctxt->loc_obd->obd_name,
 742                                POSTID(&lgl->lgl_oi), 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,
 751                                               index);
 752                 } else if (lrc == -ENOENT) {
 753                         if (rc == 0) /* ENOENT shouldn't rewrite any error */
 754                                 rc = lrc;
 755                 } else if (lrc < 0) {
 756                         failed++;
 757                         rc = lrc;
 758                 }
 759                 llog_handle_put(loghandle);
 760         }
 761         if (rc)
 762                 CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
 763                        cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
 764                        rc);
 765
 766         RETURN(rc);
 767 }
 768 EXPORT_SYMBOL(llog_cat_cancel_records);
 769
 770 static int llog_cat_process_cb(const struct lu_env *env,
 771                                struct llog_handle *cat_llh,
 772                                struct llog_rec_hdr *rec, void *data)
 773 {
 774         struct llog_process_data *d = data;
 775         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
 776         struct llog_handle *llh;
 777         struct llog_log_hdr *hdr;
 778         int rc;
 779
 780         ENTRY;
 781         if (rec->lrh_type != LLOG_LOGID_MAGIC) {
 782                 CERROR("invalid record in catalog\n");
 783                 RETURN(-EINVAL);
 784         }
 785         CDEBUG(D_HA, "processing log "DOSTID":%x at index %u of catalog "
 786                DOSTID"\n", POSTID(&lir->lid_id.lgl_oi), lir->lid_id.lgl_ogen,
 787                rec->lrh_index, POSTID(&cat_llh->lgh_id.lgl_oi));
 788
 789         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
 790         if (rc) {
 791                 CERROR("%s: cannot find handle for llog "DOSTID": %d\n",
 792                        cat_llh->lgh_ctxt->loc_obd->obd_name,
 793                        POSTID(&lir->lid_id.lgl_oi), rc);
 794                 if (rc == -ENOENT || rc == -ESTALE) {
 795                         /* After a server crash, a stub of index
 796                          * record in catlog could be kept, because
 797                          * plain log destroy + catlog index record
 798                          * deletion are not atomic. So we end up with
 799                          * an index but no actual record. Destroy the
 800                          * index and move on. */
 801                         rc = llog_cat_cleanup(env, cat_llh, NULL,
 802                                               rec->lrh_index);
 803                 }
 804
 805                 RETURN(rc);
 806         }
 807
 808         /* clean old empty llogs, do not consider current llog in use */
 809         /* ignore remote (lgh_obj=NULL) llogs */
 810         hdr = llh->lgh_hdr;
 811         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 812             hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
 813             llh != cat_llh->u.chd.chd_current_log) {
 814                 rc = llog_destroy(env, llh);
 815                 if (rc)
 816                         CERROR("%s: fail to destroy empty log: rc = %d\n",
 817                                llh->lgh_ctxt->loc_obd->obd_name, rc);
 818                 GOTO(out, rc = LLOG_DEL_PLAIN);
 819         }
 820
 821         if (rec->lrh_index < d->lpd_startcat) {
 822                 /* Skip processing of the logs until startcat */
 823                 rc = 0;
 824         } else if (d->lpd_startidx > 0) {
 825                 struct llog_process_cat_data cd;
 826
 827                 cd.lpcd_first_idx = d->lpd_startidx;
 828                 cd.lpcd_last_idx = 0;
 829                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 830                                           &cd, false);
 831                 /* Continue processing the next log from idx 0 */
 832                 d->lpd_startidx = 0;
 833         } else {
 834                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 835                                           NULL, false);
 836         }
 837
 838 out:
 839         /* The empty plain log was destroyed while processing */
 840         if (rc == LLOG_DEL_PLAIN)
 841                 rc = llog_cat_cleanup(env, cat_llh, llh,
 842                                       llh->u.phd.phd_cookie.lgc_index);
 843         llog_handle_put(llh);
 844
 845         RETURN(rc);
 846 }
 847
 848 int llog_cat_process_or_fork(const struct lu_env *env,
 849                              struct llog_handle *cat_llh, llog_cb_t cat_cb,
 850                              llog_cb_t cb, void *data, int startcat,
 851                              int startidx, bool fork)
 852 {
 853         struct llog_process_data d;
 854         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
 855         int rc;
 856         ENTRY;
 857
 858         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
 859         d.lpd_data = data;
 860         d.lpd_cb = cb;
 861         d.lpd_startcat = startcat;
 862         d.lpd_startidx = startidx;
 863
 864         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
 865             llh->llh_count > 1) {
 866                 struct llog_process_cat_data cd;
 867
 868                 CWARN("catlog "DOSTID" crosses index zero\n",
 869                       POSTID(&cat_llh->lgh_id.lgl_oi));
 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 "DOSTID":%x at index %u of catalog "
 916                DOSTID"\n", POSTID(&lir->lid_id.lgl_oi), lir->lid_id.lgl_ogen,
 917                rec->lrh_index, POSTID(&cat_llh->lgh_id.lgl_oi));
 918
 919         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
 920         if (rc) {
 921                 CWARN("%s: cannot find handle for llog "DOSTID": rc = %d\n",
 922                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 923                       POSTID(&lir->lid_id.lgl_oi), rc);
 924                 RETURN(0);
 925         }
 926         size = llog_size(env, llh);
 927         *cum_size += size;
 928
 929         CDEBUG(D_INFO, "Add llog entry "DOSTID" size %llu\n",
 930                POSTID(&llh->lgh_id.lgl_oi), 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 "DOSTID":%x at index %u of catalog "
 964                DOSTID"\n", POSTID(&lir->lid_id.lgl_oi), lir->lid_id.lgl_ogen,
 965                le32_to_cpu(rec->lrh_index), POSTID(&cat_llh->lgh_id.lgl_oi));
 966
 967         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
 968         if (rc) {
 969                 CERROR("%s: cannot find handle for llog "DOSTID": %d\n",
 970                        cat_llh->lgh_ctxt->loc_obd->obd_name,
 971                        POSTID(&lir->lid_id.lgl_oi), rc);
 972                 if (rc == -ENOENT || rc == -ESTALE) {
 973                         /* After a server crash, a stub of index
 974                          * record in catlog could be kept, because
 975                          * plain log destroy + catlog index record
 976                          * deletion are not atomic. So we end up with
 977                          * an index but no actual record. Destroy the
 978                          * index and move on. */
 979                         rc = llog_cat_cleanup(env, cat_llh, NULL,
 980                                               rec->lrh_index);
 981                 }
 982
 983                 RETURN(rc);
 984         }
 985
 986         /* clean old empty llogs, do not consider current llog in use */
 987         hdr = llh->lgh_hdr;
 988         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 989             hdr->llh_count == 1 &&
 990             llh != cat_llh->u.chd.chd_current_log) {
 991                 rc = llog_destroy(env, llh);
 992                 if (rc)
 993                         CERROR("%s: fail to destroy empty log: rc = %d\n",
 994                                llh->lgh_ctxt->loc_obd->obd_name, rc);
 995                 GOTO(out, rc = LLOG_DEL_PLAIN);
 996         }
 997
 998         rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
 999
1000 out:
1001         /* The empty plain was destroyed while processing */
1002         if (rc == LLOG_DEL_PLAIN)
1003                 rc = llog_cat_cleanup(env, cat_llh, llh,
1004                                       llh->u.phd.phd_cookie.lgc_index);
1005
1006         llog_handle_put(llh);
1007         RETURN(rc);
1008 }
1009
1010 int llog_cat_reverse_process(const struct lu_env *env,
1011                              struct llog_handle *cat_llh,
1012                              llog_cb_t cb, void *data)
1013 {
1014         struct llog_process_data d;
1015         struct llog_process_cat_data cd;
1016         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
1017         int rc;
1018         ENTRY;
1019
1020         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
1021         d.lpd_data = data;
1022         d.lpd_cb = cb;
1023
1024         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
1025             llh->llh_count > 1) {
1026                 CWARN("catalog "DOSTID" crosses index zero\n",
1027                       POSTID(&cat_llh->lgh_id.lgl_oi));
1028
1029                 cd.lpcd_first_idx = 0;
1030                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
1031                 rc = llog_reverse_process(env, cat_llh,
1032                                           llog_cat_reverse_process_cb,
1033                                           &d, &cd);
1034                 if (rc != 0)
1035                         RETURN(rc);
1036
1037                 cd.lpcd_first_idx = le32_to_cpu(llh->llh_cat_idx);
1038                 cd.lpcd_last_idx = 0;
1039                 rc = llog_reverse_process(env, cat_llh,
1040                                           llog_cat_reverse_process_cb,
1041                                           &d, &cd);
1042         } else {
1043                 rc = llog_reverse_process(env, cat_llh,
1044                                           llog_cat_reverse_process_cb,
1045                                           &d, NULL);
1046         }
1047
1048         RETURN(rc);
1049 }
1050 EXPORT_SYMBOL(llog_cat_reverse_process);
1051
1052 static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
1053 {
1054         struct llog_log_hdr *llh = cathandle->lgh_hdr;
1055         int bitmap_size;
1056
1057         ENTRY;
1058
1059         bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
1060         /*
1061          * The llh_cat_idx equals to the first used index minus 1
1062          * so if we canceled the first index then llh_cat_idx
1063          * must be renewed.
1064          */
1065         if (llh->llh_cat_idx == (idx - 1)) {
1066                 llh->llh_cat_idx = idx;
1067
1068                 while (idx != cathandle->lgh_last_idx) {
1069                         idx = (idx + 1) % bitmap_size;
1070                         if (!ext2_test_bit(idx, LLOG_HDR_BITMAP(llh))) {
1071                                 /* update llh_cat_idx for each unset bit,
1072                                  * expecting the next one is set */
1073                                 llh->llh_cat_idx = idx;
1074                         } else if (idx == 0) {
1075                                 /* skip header bit */
1076                                 llh->llh_cat_idx = 0;
1077                                 continue;
1078                         } else {
1079                                 /* the first index is found */
1080                                 break;
1081                         }
1082                 }
1083
1084                 CDEBUG(D_RPCTRACE, "Set catlog "DOSTID" first idx %u,"
1085                        " (last_idx %u)\n", POSTID(&cathandle->lgh_id.lgl_oi),
1086                        llh->llh_cat_idx, cathandle->lgh_last_idx);
1087         }
1088
1089         RETURN(0);
1090 }
1091
1092 /* Cleanup deleted plain llog traces from catalog */
1093 int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
1094                      struct llog_handle *loghandle, int index)
1095 {
1096         int rc;
1097
1098         LASSERT(index);
1099         if (loghandle != NULL) {
1100                 /* remove destroyed llog from catalog list and
1101                  * chd_current_log variable */
1102                 down_write(&cathandle->lgh_lock);
1103                 if (cathandle->u.chd.chd_current_log == loghandle)
1104                         cathandle->u.chd.chd_current_log = NULL;
1105                 list_del_init(&loghandle->u.phd.phd_entry);
1106                 up_write(&cathandle->lgh_lock);
1107                 LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
1108                 /* llog was opened and keep in a list, close it now */
1109                 llog_close(env, loghandle);
1110         }
1111
1112         /* do not attempt to cleanup on-disk llog if on client side */
1113         if (cathandle->lgh_obj == NULL)
1114                 return 0;
1115
1116         /* remove plain llog entry from catalog by index */
1117         llog_cat_set_first_idx(cathandle, index);
1118         rc = llog_cancel_rec(env, cathandle, index);
1119         if (rc == 0)
1120                 CDEBUG(D_HA, "cancel plain log at index"
1121                        " %u of catalog "DOSTID"\n",
1122                        index, POSTID(&cathandle->lgh_id.lgl_oi));
1123         return rc;
1124 }