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, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/obdclass/llog_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                 if (cathandle->lgh_name == NULL) {
  89                         CWARN("%s: there are no more free slots in catalog "
  90                               DFID":%x\n",
  91                               loghandle->lgh_ctxt->loc_obd->obd_name,
  92                               PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
  93                               cathandle->lgh_id.lgl_ogen);
  94                 } else {
  95                         CWARN("%s: there are no more free slots in "
  96                               "catalog %s\n",
  97                               loghandle->lgh_ctxt->loc_obd->obd_name,
  98                               cathandle->lgh_name);
  99                 }
 100                 RETURN(-ENOSPC);
 101         }
 102
 103         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED))
 104                 RETURN(-ENOSPC);
 105
 106         if (loghandle->lgh_hdr != NULL) {
 107                 /* If llog object is remote and creation is failed, lgh_hdr
 108                  * might be left over here, free it first */
 109                 LASSERT(!llog_exist(loghandle));
 110                 OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
 111                 loghandle->lgh_hdr = NULL;
 112         }
 113
 114         if (th == NULL) {
 115                 dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 116
 117                 handle = dt_trans_create(env, dt);
 118                 if (IS_ERR(handle))
 119                         RETURN(PTR_ERR(handle));
 120
 121                 /* Create update llog object synchronously, which
 122                  * happens during inialization process see
 123                  * lod_sub_prep_llog(), to make sure the update
 124                  * llog object is created before corss-MDT writing
 125                  * updates into the llog object */
 126                 if (cathandle->lgh_ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
 127                         handle->th_sync = 1;
 128
 129                 handle->th_wait_submit = 1;
 130
 131                 rc = llog_declare_create(env, loghandle, handle);
 132                 if (rc != 0)
 133                         GOTO(out, rc);
 134
 135                 rec->lid_hdr.lrh_len = sizeof(*rec);
 136                 rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
 137                 rec->lid_id = loghandle->lgh_id;
 138                 rc = llog_declare_write_rec(env, cathandle, &rec->lid_hdr, -1,
 139                                             handle);
 140                 if (rc != 0)
 141                         GOTO(out, rc);
 142
 143                 rc = dt_trans_start_local(env, dt, handle);
 144                 if (rc != 0)
 145                         GOTO(out, rc);
 146
 147                 th = handle;
 148         }
 149
 150         rc = llog_create(env, loghandle, th);
 151         /* if llog is already created, no need to initialize it */
 152         if (rc == -EEXIST) {
 153                 GOTO(out, rc = 0);
 154         } else if (rc != 0) {
 155                 CERROR("%s: can't create new plain llog in catalog: rc = %d\n",
 156                        loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 157                 GOTO(out, rc);
 158         }
 159
 160         rc = llog_init_handle(env, loghandle,
 161                               LLOG_F_IS_PLAIN | LLOG_F_ZAP_WHEN_EMPTY,
 162                               &cathandle->lgh_hdr->llh_tgtuuid);
 163         if (rc < 0)
 164                 GOTO(out, rc);
 165
 166         /* build the record for this log in the catalog */
 167         rec->lid_hdr.lrh_len = sizeof(*rec);
 168         rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
 169         rec->lid_id = loghandle->lgh_id;
 170
 171         /* append the new record into catalog. The new index will be
 172          * assigned to the record and updated in rec header */
 173         rc = llog_write_rec(env, cathandle, &rec->lid_hdr,
 174                             &loghandle->u.phd.phd_cookie, LLOG_NEXT_IDX, th);
 175         if (rc < 0)
 176                 GOTO(out_destroy, rc);
 177
 178         CDEBUG(D_OTHER, "new plain log "DFID".%u of catalog "DFID"\n",
 179                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rec->lid_hdr.lrh_index,
 180                PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
 181
 182         loghandle->lgh_hdr->llh_cat_idx = rec->lid_hdr.lrh_index;
 183
 184         /* limit max size of plain llog so that space can be
 185          * released sooner, especially on small filesystems */
 186         /* 2MB for the cases when free space hasn't been learned yet */
 187         loghandle->lgh_max_size = 2 << 20;
 188         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 189         rc = dt_statfs(env, dt, &lgi->lgi_statfs);
 190         if (rc == 0 && lgi->lgi_statfs.os_bfree > 0) {
 191                 __u64 freespace = (lgi->lgi_statfs.os_bfree *
 192                                   lgi->lgi_statfs.os_bsize) >> 6;
 193                 if (freespace < loghandle->lgh_max_size)
 194                         loghandle->lgh_max_size = freespace;
 195                 /* shouldn't be > 128MB in any case?
 196                  * it's 256K records of 512 bytes each */
 197                 if (freespace > (128 << 20))
 198                         loghandle->lgh_max_size = 128 << 20;
 199         }
 200         rc = 0;
 201
 202 out:
 203         if (handle != NULL) {
 204                 handle->th_result = rc >= 0 ? 0 : rc;
 205                 dt_trans_stop(env, dt, handle);
 206         }
 207         RETURN(rc);
 208
 209 out_destroy:
 210         /* to signal llog_cat_close() it shouldn't try to destroy the llog,
 211          * we want to destroy it in this transaction, otherwise the object
 212          * becomes an orphan */
 213         loghandle->lgh_hdr->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 214         /* this is to mimic full log, so another llog_cat_current_log()
 215          * can skip it and ask for another onet */
 216         loghandle->lgh_last_idx = LLOG_HDR_BITMAP_SIZE(loghandle->lgh_hdr) + 1;
 217         llog_trans_destroy(env, loghandle, th);
 218         if (handle != NULL)
 219                 dt_trans_stop(env, dt, handle);
 220         RETURN(rc);
 221 }
 222
 223 static int llog_cat_refresh(const struct lu_env *env,
 224                             struct llog_handle *cathandle)
 225 {
 226         struct llog_handle *loghandle;
 227         int rc;
 228
 229         down_write(&cathandle->lgh_lock);
 230         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 231                             u.phd.phd_entry) {
 232                 if (!llog_exist(loghandle))
 233                         continue;
 234
 235                 rc = llog_read_header(env, loghandle, NULL);
 236                 if (rc)
 237                         goto unlock;
 238         }
 239
 240         rc = llog_read_header(env, cathandle, NULL);
 241 unlock:
 242         up_write(&loghandle->lgh_lock);
 243
 244         return rc;
 245 }
 246
 247 /*
 248  * prepare current/next log for catalog.
 249  *
 250  * if \a *ploghandle is NULL, open it, and declare create, NB, if \a
 251  * *ploghandle is remote, create it synchronously here, see comments
 252  * below.
 253  *
 254  * \a cathandle->lgh_lock is down_read-ed, it gets down_write-ed if \a
 255  * *ploghandle has to be opened.
 256  */
 257 static int llog_cat_prep_log(const struct lu_env *env,
 258                              struct llog_handle *cathandle,
 259                              struct llog_handle **ploghandle,
 260                              struct thandle *th)
 261 {
 262         int rc;
 263         int sem_upgraded;
 264
 265 start:
 266         rc = 0;
 267         sem_upgraded = 0;
 268         if (IS_ERR_OR_NULL(*ploghandle)) {
 269                 up_read(&cathandle->lgh_lock);
 270                 down_write(&cathandle->lgh_lock);
 271                 sem_upgraded = 1;
 272                 if (IS_ERR_OR_NULL(*ploghandle)) {
 273                         struct llog_handle *loghandle;
 274
 275                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
 276                                        NULL, NULL, LLOG_OPEN_NEW);
 277                         if (!rc) {
 278                                 *ploghandle = loghandle;
 279                                 list_add_tail(&loghandle->u.phd.phd_entry,
 280                                               &cathandle->u.chd.chd_head);
 281                         }
 282                 }
 283                 if (rc)
 284                         GOTO(out, rc);
 285         }
 286
 287         rc = llog_exist(*ploghandle);
 288         if (rc < 0)
 289                 GOTO(out, rc);
 290         if (rc)
 291                 GOTO(out, rc = 0);
 292
 293         if (dt_object_remote(cathandle->lgh_obj)) {
 294                 down_write_nested(&(*ploghandle)->lgh_lock, LLOGH_LOG);
 295                 if (!llog_exist(*ploghandle)) {
 296                         /* For remote operation, if we put the llog object
 297                          * creation in the current transaction, then the
 298                          * llog object will not be created on the remote
 299                          * target until the transaction stop, if other
 300                          * operations start before the transaction stop,
 301                          * and use the same llog object, will be dependent
 302                          * on the success of this transaction. So let's
 303                          * create the llog object synchronously here to
 304                          * remove the dependency. */
 305                         rc = llog_cat_new_log(env, cathandle, *ploghandle,
 306                                               NULL);
 307                         if (rc == -ESTALE) {
 308                                 up_write(&(*ploghandle)->lgh_lock);
 309                                 if (sem_upgraded)
 310                                         up_write(&cathandle->lgh_lock);
 311                                 else
 312                                         up_read(&cathandle->lgh_lock);
 313
 314                                 rc = llog_cat_refresh(env, cathandle);
 315                                 down_read_nested(&cathandle->lgh_lock,
 316                                                  LLOGH_CAT);
 317                                 if (rc)
 318                                         return rc;
 319                                 /* *ploghandle might become NULL, restart */
 320                                 goto start;
 321                         }
 322                 }
 323                 up_write(&(*ploghandle)->lgh_lock);
 324         } else {
 325                 struct llog_thread_info *lgi = llog_info(env);
 326                 struct llog_logid_rec *lirec = &lgi->lgi_logid;
 327
 328                 rc = llog_declare_create(env, *ploghandle, th);
 329                 if (rc)
 330                         GOTO(out, rc);
 331
 332                 lirec->lid_hdr.lrh_len = sizeof(*lirec);
 333                 rc = llog_declare_write_rec(env, cathandle, &lirec->lid_hdr, -1,
 334                                             th);
 335         }
 336
 337 out:
 338         if (sem_upgraded) {
 339                 up_write(&cathandle->lgh_lock);
 340                 down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 341                 if (rc == 0)
 342                         goto start;
 343         }
 344         return rc;
 345 }
 346
 347 /* Open an existent log handle and add it to the open list.
 348  * This log handle will be closed when all of the records in it are removed.
 349  *
 350  * Assumes caller has already pushed us into the kernel context and is locking.
 351  * We return a lock on the handle to ensure nobody yanks it from us.
 352  *
 353  * This takes extra reference on llog_handle via llog_handle_get() and require
 354  * this reference to be put by caller using llog_handle_put()
 355  */
 356 int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
 357                        struct llog_handle **res, struct llog_logid *logid)
 358 {
 359         struct llog_handle      *loghandle;
 360         enum llog_flag           fmt;
 361         int                      rc = 0;
 362
 363         ENTRY;
 364
 365         if (cathandle == NULL)
 366                 RETURN(-EBADF);
 367
 368         fmt = cathandle->lgh_hdr->llh_flags & LLOG_F_EXT_MASK;
 369         down_write(&cathandle->lgh_lock);
 370         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 371                             u.phd.phd_entry) {
 372                 struct llog_logid *cgl = &loghandle->lgh_id;
 373
 374                 if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
 375                     ostid_seq(&cgl->lgl_oi) == ostid_seq(&logid->lgl_oi)) {
 376                         if (cgl->lgl_ogen != logid->lgl_ogen) {
 377                                 CWARN("%s: log "DFID" generation %x != %x\n",
 378                                       loghandle->lgh_ctxt->loc_obd->obd_name,
 379                                       PFID(&logid->lgl_oi.oi_fid),
 380                                       cgl->lgl_ogen, logid->lgl_ogen);
 381                                 continue;
 382                         }
 383                         loghandle->u.phd.phd_cat_handle = cathandle;
 384                         up_write(&cathandle->lgh_lock);
 385                         GOTO(out, rc = 0);
 386                 }
 387         }
 388         up_write(&cathandle->lgh_lock);
 389
 390         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle, logid, NULL,
 391                        LLOG_OPEN_EXISTS);
 392         if (rc < 0) {
 393                 CERROR("%s: error opening log id "DFID":%x: rc = %d\n",
 394                        cathandle->lgh_ctxt->loc_obd->obd_name,
 395                        PFID(&logid->lgl_oi.oi_fid), logid->lgl_ogen, rc);
 396                 RETURN(rc);
 397         }
 398
 399         rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN | fmt, NULL);
 400         if (rc < 0) {
 401                 llog_close(env, loghandle);
 402                 loghandle = NULL;
 403                 RETURN(rc);
 404         }
 405
 406         down_write(&cathandle->lgh_lock);
 407         list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
 408         up_write(&cathandle->lgh_lock);
 409
 410         loghandle->u.phd.phd_cat_handle = cathandle;
 411         loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
 412         loghandle->u.phd.phd_cookie.lgc_index =
 413                                 loghandle->lgh_hdr->llh_cat_idx;
 414         EXIT;
 415 out:
 416         llog_handle_get(loghandle);
 417         *res = loghandle;
 418         return 0;
 419 }
 420
 421 int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
 422 {
 423         struct llog_handle      *loghandle, *n;
 424         int                      rc;
 425
 426         ENTRY;
 427
 428         list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
 429                                  u.phd.phd_entry) {
 430                 struct llog_log_hdr     *llh = loghandle->lgh_hdr;
 431                 int                      index;
 432
 433                 /* unlink open-not-created llogs */
 434                 list_del_init(&loghandle->u.phd.phd_entry);
 435                 llh = loghandle->lgh_hdr;
 436                 if (loghandle->lgh_obj != NULL && llh != NULL &&
 437                     (llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 438                     (llh->llh_count == 1)) {
 439                         rc = llog_destroy(env, loghandle);
 440                         if (rc)
 441                                 CERROR("%s: failure destroying log during "
 442                                        "cleanup: rc = %d\n",
 443                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 444                                        rc);
 445
 446                         index = loghandle->u.phd.phd_cookie.lgc_index;
 447                         llog_cat_cleanup(env, cathandle, NULL, index);
 448                 }
 449                 llog_close(env, loghandle);
 450         }
 451         /* if handle was stored in ctxt, remove it too */
 452         if (cathandle->lgh_ctxt->loc_handle == cathandle)
 453                 cathandle->lgh_ctxt->loc_handle = NULL;
 454         rc = llog_close(env, cathandle);
 455         RETURN(rc);
 456 }
 457 EXPORT_SYMBOL(llog_cat_close);
 458
 459 /**
 460  * lockdep markers for nested struct llog_handle::lgh_lock locking.
 461  */
 462 enum {
 463         LLOGH_CAT,
 464         LLOGH_LOG
 465 };
 466
 467 /** Return the currently active log handle.  If the current log handle doesn't
 468  * have enough space left for the current record, start a new one.
 469  *
 470  * If reclen is 0, we only want to know what the currently active log is,
 471  * otherwise we get a lock on this log so nobody can steal our space.
 472  *
 473  * Assumes caller has already pushed us into the kernel context and is locking.
 474  *
 475  * NOTE: loghandle is write-locked upon successful return
 476  */
 477 static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
 478                                                 struct thandle *th)
 479 {
 480         struct llog_handle *loghandle = NULL;
 481         ENTRY;
 482
 483
 484         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2)) {
 485                 down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 486                 GOTO(next, loghandle);
 487         }
 488
 489         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 490         loghandle = cathandle->u.chd.chd_current_log;
 491         if (loghandle) {
 492                 struct llog_log_hdr *llh;
 493
 494                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 495                 llh = loghandle->lgh_hdr;
 496                 if (llh == NULL || !llog_is_full(loghandle)) {
 497                         up_read(&cathandle->lgh_lock);
 498                         RETURN(loghandle);
 499                 } else {
 500                         up_write(&loghandle->lgh_lock);
 501                 }
 502         }
 503         up_read(&cathandle->lgh_lock);
 504
 505         /* time to use next log */
 506
 507         /* first, we have to make sure the state hasn't changed */
 508         down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 509         loghandle = cathandle->u.chd.chd_current_log;
 510         if (loghandle) {
 511                 struct llog_log_hdr *llh;
 512
 513                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 514                 llh = loghandle->lgh_hdr;
 515                 if (llh == NULL || !llog_is_full(loghandle))
 516                         GOTO(out_unlock, loghandle);
 517                 else
 518                         up_write(&loghandle->lgh_lock);
 519         }
 520
 521 next:
 522         /* Sigh, the chd_next_log and chd_current_log is initialized
 523          * in declare phase, and we do not serialize the catlog
 524          * accessing, so it might be possible the llog creation
 525          * thread (see llog_cat_declare_add_rec()) did not create
 526          * llog successfully, then the following thread might
 527          * meet this situation. */
 528         if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
 529                 CERROR("%s: next log does not exist!\n",
 530                        cathandle->lgh_ctxt->loc_obd->obd_name);
 531                 loghandle = ERR_PTR(-EIO);
 532                 if (cathandle->u.chd.chd_next_log == NULL) {
 533                         /* Store the error in chd_next_log, so
 534                          * the following process can get correct
 535                          * failure value */
 536                         cathandle->u.chd.chd_next_log = loghandle;
 537                 }
 538                 GOTO(out_unlock, loghandle);
 539         }
 540
 541         CDEBUG(D_INODE, "use next log\n");
 542
 543         loghandle = cathandle->u.chd.chd_next_log;
 544         cathandle->u.chd.chd_current_log = loghandle;
 545         cathandle->u.chd.chd_next_log = NULL;
 546         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 547
 548 out_unlock:
 549         up_write(&cathandle->lgh_lock);
 550         LASSERT(loghandle);
 551         RETURN(loghandle);
 552 }
 553
 554 /* Add a single record to the recovery log(s) using a catalog
 555  * Returns as llog_write_record
 556  *
 557  * Assumes caller has already pushed us into the kernel context.
 558  */
 559 int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
 560                      struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 561                      struct thandle *th)
 562 {
 563         struct llog_handle *loghandle;
 564         int rc, retried = 0;
 565         ENTRY;
 566
 567         LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
 568
 569 retry:
 570         loghandle = llog_cat_current_log(cathandle, th);
 571         if (IS_ERR(loghandle))
 572                 RETURN(PTR_ERR(loghandle));
 573
 574         /* loghandle is already locked by llog_cat_current_log() for us */
 575         if (!llog_exist(loghandle)) {
 576                 rc = llog_cat_new_log(env, cathandle, loghandle, th);
 577                 if (rc < 0) {
 578                         up_write(&loghandle->lgh_lock);
 579                         /* nobody should be trying to use this llog */
 580                         down_write(&cathandle->lgh_lock);
 581                         /* only reset current log if still room in catalog, to
 582                          * avoid unnecessarily and racy creation of new and
 583                          * partially initialized llog_handle
 584                          */
 585                         if ((cathandle->u.chd.chd_current_log == loghandle) &&
 586                             rc != -ENOSPC)
 587                                 cathandle->u.chd.chd_current_log = NULL;
 588                         up_write(&cathandle->lgh_lock);
 589                         RETURN(rc);
 590                 }
 591         }
 592         /* now let's try to add the record */
 593         rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
 594         if (rc < 0) {
 595                 CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
 596                              "llog_write_rec %d: lh=%p\n", rc, loghandle);
 597                 /* -ENOSPC is returned if no empty records left
 598                  * and when it's lack of space on the stogage.
 599                  * there is no point to try again if it's the second
 600                  * case. many callers (like llog test) expect ENOSPC,
 601                  * so we preserve this error code, but look for the
 602                  * actual cause here */
 603                 if (rc == -ENOSPC && llog_is_full(loghandle))
 604                         rc = -ENOBUFS;
 605         }
 606         up_write(&loghandle->lgh_lock);
 607
 608         if (rc == -ENOBUFS) {
 609                 if (retried++ == 0)
 610                         GOTO(retry, rc);
 611                 CERROR("%s: error on 2nd llog: rc = %d\n",
 612                        cathandle->lgh_ctxt->loc_obd->obd_name, rc);
 613         }
 614
 615         RETURN(rc);
 616 }
 617 EXPORT_SYMBOL(llog_cat_add_rec);
 618
 619 int llog_cat_declare_add_rec(const struct lu_env *env,
 620                              struct llog_handle *cathandle,
 621                              struct llog_rec_hdr *rec, struct thandle *th)
 622 {
 623         int rc;
 624
 625         ENTRY;
 626
 627 start:
 628         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 629         rc = llog_cat_prep_log(env, cathandle,
 630                                &cathandle->u.chd.chd_current_log, th);
 631         if (rc)
 632                 GOTO(unlock, rc);
 633
 634         rc = llog_cat_prep_log(env, cathandle, &cathandle->u.chd.chd_next_log,
 635                                th);
 636         if (rc)
 637                 GOTO(unlock, rc);
 638
 639         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
 640                                     rec, -1, th);
 641         if (rc == -ESTALE && dt_object_remote(cathandle->lgh_obj)) {
 642                 up_read(&cathandle->lgh_lock);
 643                 rc = llog_cat_refresh(env, cathandle);
 644                 if (rc)
 645                         RETURN(rc);
 646                 goto start;
 647         }
 648
 649 #if 0
 650         /*
 651          * XXX: we hope for declarations made for existing llog this might be
 652          * not correct with some backends where declarations are expected
 653          * against specific object like ZFS with full debugging enabled.
 654          */
 655         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_next_log, rec, -1,
 656                                     th);
 657 #endif
 658 unlock:
 659         up_read(&cathandle->lgh_lock);
 660         RETURN(rc);
 661 }
 662 EXPORT_SYMBOL(llog_cat_declare_add_rec);
 663
 664 int llog_cat_add(const struct lu_env *env, struct llog_handle *cathandle,
 665                  struct llog_rec_hdr *rec, struct llog_cookie *reccookie)
 666 {
 667         struct llog_ctxt        *ctxt;
 668         struct dt_device        *dt;
 669         struct thandle          *th = NULL;
 670         int                      rc;
 671
 672         ctxt = cathandle->lgh_ctxt;
 673         LASSERT(ctxt);
 674         LASSERT(ctxt->loc_exp);
 675
 676         LASSERT(cathandle->lgh_obj != NULL);
 677         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 678
 679         th = dt_trans_create(env, dt);
 680         if (IS_ERR(th))
 681                 RETURN(PTR_ERR(th));
 682
 683         rc = llog_cat_declare_add_rec(env, cathandle, rec, th);
 684         if (rc)
 685                 GOTO(out_trans, rc);
 686
 687         rc = dt_trans_start_local(env, dt, th);
 688         if (rc)
 689                 GOTO(out_trans, rc);
 690         rc = llog_cat_add_rec(env, cathandle, rec, reccookie, th);
 691 out_trans:
 692         dt_trans_stop(env, dt, th);
 693         RETURN(rc);
 694 }
 695 EXPORT_SYMBOL(llog_cat_add);
 696
 697 /* For each cookie in the cookie array, we clear the log in-use bit and either:
 698  * - the log is empty, so mark it free in the catalog header and delete it
 699  * - the log is not empty, just write out the log header
 700  *
 701  * The cookies may be in different log files, so we need to get new logs
 702  * each time.
 703  *
 704  * Assumes caller has already pushed us into the kernel context.
 705  */
 706 int llog_cat_cancel_records(const struct lu_env *env,
 707                             struct llog_handle *cathandle, int count,
 708                             struct llog_cookie *cookies)
 709 {
 710         int i, index, rc = 0, failed = 0;
 711
 712         ENTRY;
 713
 714         for (i = 0; i < count; i++, cookies++) {
 715                 struct llog_handle *loghandle;
 716                 struct llog_logid *lgl = &cookies->lgc_lgl;
 717                 int  lrc;
 718
 719                 rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
 720                 if (rc) {
 721                         CDEBUG(D_HA, "%s: cannot find llog for handle "DFID":%x"
 722                                ": rc = %d\n",
 723                                cathandle->lgh_ctxt->loc_obd->obd_name,
 724                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, rc);
 725                         failed++;
 726                         continue;
 727                 }
 728
 729                 if ((cathandle->lgh_ctxt->loc_flags &
 730                      LLOG_CTXT_FLAG_NORMAL_FID) && !llog_exist(loghandle)) {
 731                         /* For update log, some of loghandles of cathandle
 732                          * might not exist because remote llog creation might
 733                          * be failed, so let's skip the record cancellation
 734                          * for these non-exist llogs.
 735                          */
 736                         lrc = -ENOENT;
 737                         CDEBUG(D_HA, "%s: llog "DFID":%x does not exist"
 738                                ": rc = %d\n",
 739                                cathandle->lgh_ctxt->loc_obd->obd_name,
 740                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, lrc);
 741                         failed++;
 742                         if (rc == 0)
 743                                 rc = lrc;
 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                         lrc = llog_cat_cleanup(env, cathandle, loghandle,
 751                                                index);
 752                         if (rc == 0)
 753                                 rc = lrc;
 754                 } else if (lrc == -ENOENT) {
 755                         if (rc == 0) /* ENOENT shouldn't rewrite any error */
 756                                 rc = lrc;
 757                 } else if (lrc < 0) {
 758                         failed++;
 759                         if (rc == 0)
 760                                 rc = lrc;
 761                 }
 762                 llog_handle_put(loghandle);
 763         }
 764         if (rc)
 765                 CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
 766                        cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
 767                        rc);
 768
 769         RETURN(rc);
 770 }
 771 EXPORT_SYMBOL(llog_cat_cancel_records);
 772
 773 static int llog_cat_process_common(const struct lu_env *env,
 774                                    struct llog_handle *cat_llh,
 775                                    struct llog_rec_hdr *rec,
 776                                    struct llog_handle **llhp)
 777 {
 778         struct llog_logid_rec *lir = container_of(rec, typeof(*lir), lid_hdr);
 779         struct llog_log_hdr *hdr;
 780         int rc;
 781
 782         ENTRY;
 783         if (rec->lrh_type != le32_to_cpu(LLOG_LOGID_MAGIC)) {
 784                 rc = -EINVAL;
 785                 CWARN("%s: invalid record in catalog "DFID":%x: rc = %d\n",
 786                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 787                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid),
 788                       cat_llh->lgh_id.lgl_ogen, rc);
 789                 RETURN(rc);
 790         }
 791         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "DFID"\n",
 792                PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
 793                le32_to_cpu(rec->lrh_index),
 794                PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 795
 796         rc = llog_cat_id2handle(env, cat_llh, llhp, &lir->lid_id);
 797         if (rc) {
 798                 /* After a server crash, a stub of index record in catlog could
 799                  * be kept, because plain log destroy + catlog index record
 800                  * deletion are not atomic. So we end up with an index but no
 801                  * actual record. Destroy the index and move on. */
 802                 if (rc == -ENOENT || rc == -ESTALE)
 803                         rc = LLOG_DEL_RECORD;
 804                 else if (rc)
 805                         CWARN("%s: can't find llog handle "DFID":%x: rc = %d\n",
 806                               cat_llh->lgh_ctxt->loc_obd->obd_name,
 807                               PFID(&lir->lid_id.lgl_oi.oi_fid),
 808                               lir->lid_id.lgl_ogen, rc);
 809
 810                 RETURN(rc);
 811         }
 812
 813         /* clean old empty llogs, do not consider current llog in use */
 814         /* ignore remote (lgh_obj == NULL) llogs */
 815         hdr = (*llhp)->lgh_hdr;
 816         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 817             hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
 818             *llhp != cat_llh->u.chd.chd_current_log) {
 819                 rc = llog_destroy(env, *llhp);
 820                 if (rc)
 821                         CWARN("%s: can't destroy empty log "DFID": rc = %d\n",
 822                               (*llhp)->lgh_ctxt->loc_obd->obd_name,
 823                               PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
 824                 rc = LLOG_DEL_PLAIN;
 825         }
 826
 827         RETURN(rc);
 828 }
 829
 830 static int llog_cat_process_cb(const struct lu_env *env,
 831                                struct llog_handle *cat_llh,
 832                                struct llog_rec_hdr *rec, void *data)
 833 {
 834         struct llog_process_data *d = data;
 835         struct llog_handle *llh = NULL;
 836         int rc;
 837
 838         ENTRY;
 839         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 840         if (rc)
 841                 GOTO(out, rc);
 842
 843         if (rec->lrh_index < d->lpd_startcat) {
 844                 /* Skip processing of the logs until startcat */
 845                 rc = 0;
 846         } else if (d->lpd_startidx > 0) {
 847                 struct llog_process_cat_data cd;
 848
 849                 cd.lpcd_first_idx = d->lpd_startidx;
 850                 cd.lpcd_last_idx = 0;
 851                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 852                                           &cd, false);
 853                 /* Continue processing the next log from idx 0 */
 854                 d->lpd_startidx = 0;
 855         } else {
 856                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 857                                           NULL, false);
 858         }
 859
 860 out:
 861         /* The empty plain log was destroyed while processing */
 862         if (rc == LLOG_DEL_PLAIN) {
 863                 rc = llog_cat_cleanup(env, cat_llh, llh,
 864                                       llh->u.phd.phd_cookie.lgc_index);
 865         } else if (rc == LLOG_DEL_RECORD) {
 866                 /* clear wrong catalog entry */
 867                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
 868         }
 869
 870         if (llh)
 871                 llog_handle_put(llh);
 872
 873         RETURN(rc);
 874 }
 875
 876 int llog_cat_process_or_fork(const struct lu_env *env,
 877                              struct llog_handle *cat_llh, llog_cb_t cat_cb,
 878                              llog_cb_t cb, void *data, int startcat,
 879                              int startidx, bool fork)
 880 {
 881         struct llog_process_data d;
 882         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
 883         int rc;
 884         ENTRY;
 885
 886         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
 887         d.lpd_data = data;
 888         d.lpd_cb = cb;
 889         d.lpd_startcat = startcat;
 890         d.lpd_startidx = startidx;
 891
 892         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
 893             llh->llh_count > 1) {
 894                 struct llog_process_cat_data cd;
 895
 896                 CWARN("%s: catlog "DFID" crosses index zero\n",
 897                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 898                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 899
 900                 cd.lpcd_first_idx = llh->llh_cat_idx;
 901                 cd.lpcd_last_idx = 0;
 902                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 903                                           &d, &cd, fork);
 904                 if (rc != 0)
 905                         RETURN(rc);
 906
 907                 cd.lpcd_first_idx = 0;
 908                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
 909                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 910                                           &d, &cd, fork);
 911         } else {
 912                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 913                                           &d, NULL, fork);
 914         }
 915
 916         RETURN(rc);
 917 }
 918
 919 int llog_cat_process(const struct lu_env *env, struct llog_handle *cat_llh,
 920                      llog_cb_t cb, void *data, int startcat, int startidx)
 921 {
 922         return llog_cat_process_or_fork(env, cat_llh, llog_cat_process_cb,
 923                                         cb, data, startcat, startidx, false);
 924 }
 925 EXPORT_SYMBOL(llog_cat_process);
 926
 927 static int llog_cat_size_cb(const struct lu_env *env,
 928                              struct llog_handle *cat_llh,
 929                              struct llog_rec_hdr *rec, void *data)
 930 {
 931         struct llog_process_data *d = data;
 932         struct llog_handle *llh = NULL;
 933         __u64 *cum_size = d->lpd_data;
 934         __u64 size;
 935         int rc;
 936
 937         ENTRY;
 938         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 939
 940         if (rc == LLOG_DEL_PLAIN) {
 941                 /* empty log was deleted, don't count it */
 942                 rc = llog_cat_cleanup(env, cat_llh, llh,
 943                                       llh->u.phd.phd_cookie.lgc_index);
 944         } else if (rc == LLOG_DEL_RECORD) {
 945                 /* clear wrong catalog entry */
 946                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
 947         } else {
 948                 size = llog_size(env, llh);
 949                 *cum_size += size;
 950
 951                 CDEBUG(D_INFO, "Add llog entry "DFID" size=%llu, tot=%llu\n",
 952                        PFID(&llh->lgh_id.lgl_oi.oi_fid), size, *cum_size);
 953         }
 954
 955         if (llh != NULL)
 956                 llog_handle_put(llh);
 957
 958         RETURN(0);
 959 }
 960
 961 __u64 llog_cat_size(const struct lu_env *env, struct llog_handle *cat_llh)
 962 {
 963         __u64 size = llog_size(env, cat_llh);
 964
 965         llog_cat_process_or_fork(env, cat_llh, llog_cat_size_cb,
 966                                  NULL, &size, 0, 0, false);
 967
 968         return size;
 969 }
 970 EXPORT_SYMBOL(llog_cat_size);
 971
 972 /* currently returns the number of "free" entries in catalog,
 973  * ie the available entries for a new plain LLOG file creation,
 974  * even if catalog has wrapped
 975  */
 976 __u32 llog_cat_free_space(struct llog_handle *cat_llh)
 977 {
 978         /* simulate almost full Catalog */
 979         if (OBD_FAIL_CHECK(OBD_FAIL_CAT_FREE_RECORDS))
 980                 return cfs_fail_val;
 981
 982         if (cat_llh->lgh_hdr->llh_count == 1)
 983                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1;
 984
 985         if (cat_llh->lgh_last_idx > cat_llh->lgh_hdr->llh_cat_idx)
 986                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1 +
 987                        cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
 988
 989         /* catalog is presently wrapped */
 990         return cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
 991 }
 992 EXPORT_SYMBOL(llog_cat_free_space);
 993
 994 static int llog_cat_reverse_process_cb(const struct lu_env *env,
 995                                        struct llog_handle *cat_llh,
 996                                        struct llog_rec_hdr *rec, void *data)
 997 {
 998         struct llog_process_data *d = data;
 999         struct llog_handle *llh;
1000         int rc;
1001
1002         ENTRY;
1003         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
1004
1005         /* The empty plain log was destroyed while processing */
1006         if (rc == LLOG_DEL_PLAIN) {
1007                 rc = llog_cat_cleanup(env, cat_llh, llh,
1008                                       llh->u.phd.phd_cookie.lgc_index);
1009         } else if (rc == LLOG_DEL_RECORD) {
1010                 /* clear wrong catalog entry */
1011                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
1012         }
1013         if (rc)
1014                 RETURN(rc);
1015
1016         rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
1017
1018         /* The empty plain was destroyed while processing */
1019         if (rc == LLOG_DEL_PLAIN)
1020                 rc = llog_cat_cleanup(env, cat_llh, llh,
1021                                       llh->u.phd.phd_cookie.lgc_index);
1022
1023         llog_handle_put(llh);
1024         RETURN(rc);
1025 }
1026
1027 int llog_cat_reverse_process(const struct lu_env *env,
1028                              struct llog_handle *cat_llh,
1029                              llog_cb_t cb, void *data)
1030 {
1031         struct llog_process_data d;
1032         struct llog_process_cat_data cd;
1033         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
1034         int rc;
1035         ENTRY;
1036
1037         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
1038         d.lpd_data = data;
1039         d.lpd_cb = cb;
1040
1041         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
1042             llh->llh_count > 1) {
1043                 CWARN("%s: catalog "DFID" crosses index zero\n",
1044                       cat_llh->lgh_ctxt->loc_obd->obd_name,
1045                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
1046
1047                 cd.lpcd_first_idx = 0;
1048                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
1049                 rc = llog_reverse_process(env, cat_llh,
1050                                           llog_cat_reverse_process_cb,
1051                                           &d, &cd);
1052                 if (rc != 0)
1053                         RETURN(rc);
1054
1055                 cd.lpcd_first_idx = le32_to_cpu(llh->llh_cat_idx);
1056                 cd.lpcd_last_idx = 0;
1057                 rc = llog_reverse_process(env, cat_llh,
1058                                           llog_cat_reverse_process_cb,
1059                                           &d, &cd);
1060         } else {
1061                 rc = llog_reverse_process(env, cat_llh,
1062                                           llog_cat_reverse_process_cb,
1063                                           &d, NULL);
1064         }
1065
1066         RETURN(rc);
1067 }
1068 EXPORT_SYMBOL(llog_cat_reverse_process);
1069
1070 static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
1071 {
1072         struct llog_log_hdr *llh = cathandle->lgh_hdr;
1073         int bitmap_size;
1074
1075         ENTRY;
1076
1077         bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
1078         /*
1079          * The llh_cat_idx equals to the first used index minus 1
1080          * so if we canceled the first index then llh_cat_idx
1081          * must be renewed.
1082          */
1083         if (llh->llh_cat_idx == (idx - 1)) {
1084                 llh->llh_cat_idx = idx;
1085
1086                 while (idx != cathandle->lgh_last_idx) {
1087                         idx = (idx + 1) % bitmap_size;
1088                         if (!ext2_test_bit(idx, LLOG_HDR_BITMAP(llh))) {
1089                                 /* update llh_cat_idx for each unset bit,
1090                                  * expecting the next one is set */
1091                                 llh->llh_cat_idx = idx;
1092                         } else if (idx == 0) {
1093                                 /* skip header bit */
1094                                 llh->llh_cat_idx = 0;
1095                                 continue;
1096                         } else {
1097                                 /* the first index is found */
1098                                 break;
1099                         }
1100                 }
1101
1102                 CDEBUG(D_RPCTRACE, "catlog "DFID" first idx %u, last_idx %u\n",
1103                        PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
1104                        llh->llh_cat_idx, cathandle->lgh_last_idx);
1105         }
1106
1107         RETURN(0);
1108 }
1109
1110 /* Cleanup deleted plain llog traces from catalog */
1111 int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
1112                      struct llog_handle *loghandle, int index)
1113 {
1114         int rc;
1115
1116         LASSERT(index);
1117         if (loghandle != NULL) {
1118                 /* remove destroyed llog from catalog list and
1119                  * chd_current_log variable */
1120                 down_write(&cathandle->lgh_lock);
1121                 if (cathandle->u.chd.chd_current_log == loghandle)
1122                         cathandle->u.chd.chd_current_log = NULL;
1123                 list_del_init(&loghandle->u.phd.phd_entry);
1124                 up_write(&cathandle->lgh_lock);
1125                 LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
1126                 /* llog was opened and keep in a list, close it now */
1127                 llog_close(env, loghandle);
1128         }
1129
1130         /* do not attempt to cleanup on-disk llog if on client side */
1131         if (cathandle->lgh_obj == NULL)
1132                 return 0;
1133
1134         /* remove plain llog entry from catalog by index */
1135         llog_cat_set_first_idx(cathandle, index);
1136         rc = llog_cancel_rec(env, cathandle, index);
1137         if (rc == 0)
1138                 CDEBUG(D_HA, "cancel plain log at index %u of catalog "DFID"\n",
1139                        index, PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
1140         return rc;
1141 }