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
  53 /**
  54  * lockdep markers for nested struct llog_handle::lgh_lock locking.
  55  */
  56 enum {
  57         LLOGH_CAT,
  58         LLOGH_LOG,
  59 };
  60
  61 /* Create a new log handle and add it to the open list.
  62  * This log handle will be closed when all of the records in it are removed.
  63  *
  64  * Assumes caller has already pushed us into the kernel context and is locking.
  65  */
  66 static int llog_cat_new_log(const struct lu_env *env,
  67                             struct llog_handle *cathandle,
  68                             struct llog_handle *loghandle,
  69                             struct thandle *th)
  70 {
  71         struct llog_thread_info *lgi = llog_info(env);
  72         struct llog_logid_rec   *rec = &lgi->lgi_logid;
  73         struct thandle *handle = NULL;
  74         struct dt_device *dt = NULL;
  75         struct llog_log_hdr     *llh = cathandle->lgh_hdr;
  76         int                      rc, index;
  77
  78         ENTRY;
  79
  80         index = (cathandle->lgh_last_idx + 1) %
  81                 (OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ? (cfs_fail_val + 1) :
  82                                                 LLOG_HDR_BITMAP_SIZE(llh));
  83
  84         /* check that new llog index will not overlap with the first one.
  85          * - llh_cat_idx is the index just before the first/oldest still in-use
  86          *      index in catalog
  87          * - lgh_last_idx is the last/newest used index in catalog
  88          *
  89          * When catalog is not wrapped yet then lgh_last_idx is always larger
  90          * than llh_cat_idx. After the wrap around lgh_last_idx re-starts
  91          * from 0 and llh_cat_idx becomes the upper limit for it
  92          *
  93          * Check if catalog has already wrapped around or not by comparing
  94          * last_idx and cat_idx */
  95         if ((index == llh->llh_cat_idx + 1 && llh->llh_count > 1) ||
  96             (index == 0 && llh->llh_cat_idx == 0)) {
  97                 if (cathandle->lgh_name == NULL) {
  98                         CWARN("%s: there are no more free slots in catalog "
  99                               DFID":%x\n",
 100                               loghandle->lgh_ctxt->loc_obd->obd_name,
 101                               PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
 102                               cathandle->lgh_id.lgl_ogen);
 103                 } else {
 104                         CWARN("%s: there are no more free slots in "
 105                               "catalog %s\n",
 106                               loghandle->lgh_ctxt->loc_obd->obd_name,
 107                               cathandle->lgh_name);
 108                 }
 109                 RETURN(-ENOSPC);
 110         }
 111
 112         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED))
 113                 RETURN(-ENOSPC);
 114
 115         if (loghandle->lgh_hdr != NULL) {
 116                 /* If llog object is remote and creation is failed, lgh_hdr
 117                  * might be left over here, free it first */
 118                 LASSERT(!llog_exist(loghandle));
 119                 OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
 120                 loghandle->lgh_hdr = NULL;
 121         }
 122
 123         if (th == NULL) {
 124                 dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 125
 126                 handle = dt_trans_create(env, dt);
 127                 if (IS_ERR(handle))
 128                         RETURN(PTR_ERR(handle));
 129
 130                 /* Create update llog object synchronously, which
 131                  * happens during inialization process see
 132                  * lod_sub_prep_llog(), to make sure the update
 133                  * llog object is created before corss-MDT writing
 134                  * updates into the llog object */
 135                 if (cathandle->lgh_ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
 136                         handle->th_sync = 1;
 137
 138                 handle->th_wait_submit = 1;
 139
 140                 rc = llog_declare_create(env, loghandle, handle);
 141                 if (rc != 0)
 142                         GOTO(out, rc);
 143
 144                 rec->lid_hdr.lrh_len = sizeof(*rec);
 145                 rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
 146                 rec->lid_id = loghandle->lgh_id;
 147                 rc = llog_declare_write_rec(env, cathandle, &rec->lid_hdr, -1,
 148                                             handle);
 149                 if (rc != 0)
 150                         GOTO(out, rc);
 151
 152                 rc = dt_trans_start_local(env, dt, handle);
 153                 if (rc != 0)
 154                         GOTO(out, rc);
 155
 156                 th = handle;
 157         }
 158
 159         rc = llog_create(env, loghandle, th);
 160         /* if llog is already created, no need to initialize it */
 161         if (rc == -EEXIST) {
 162                 GOTO(out, rc = 0);
 163         } else if (rc != 0) {
 164                 CERROR("%s: can't create new plain llog in catalog: rc = %d\n",
 165                        loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 166                 GOTO(out, rc);
 167         }
 168
 169         rc = llog_init_handle(env, loghandle,
 170                               LLOG_F_IS_PLAIN | LLOG_F_ZAP_WHEN_EMPTY,
 171                               &cathandle->lgh_hdr->llh_tgtuuid);
 172         if (rc < 0)
 173                 GOTO(out, rc);
 174
 175         /* build the record for this log in the catalog */
 176         rec->lid_hdr.lrh_len = sizeof(*rec);
 177         rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
 178         rec->lid_id = loghandle->lgh_id;
 179
 180         /* append the new record into catalog. The new index will be
 181          * assigned to the record and updated in rec header */
 182         rc = llog_write_rec(env, cathandle, &rec->lid_hdr,
 183                             &loghandle->u.phd.phd_cookie, LLOG_NEXT_IDX, th);
 184         if (rc < 0)
 185                 GOTO(out_destroy, rc);
 186
 187         CDEBUG(D_OTHER, "new plain log "DFID".%u of catalog "DFID"\n",
 188                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rec->lid_hdr.lrh_index,
 189                PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
 190
 191         loghandle->lgh_hdr->llh_cat_idx = rec->lid_hdr.lrh_index;
 192
 193         /* limit max size of plain llog so that space can be
 194          * released sooner, especially on small filesystems */
 195         /* 2MB for the cases when free space hasn't been learned yet */
 196         loghandle->lgh_max_size = 2 << 20;
 197         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 198         rc = dt_statfs(env, dt, &lgi->lgi_statfs);
 199         if (rc == 0 && lgi->lgi_statfs.os_bfree > 0) {
 200                 __u64 freespace = (lgi->lgi_statfs.os_bfree *
 201                                   lgi->lgi_statfs.os_bsize) >> 6;
 202                 if (freespace < loghandle->lgh_max_size)
 203                         loghandle->lgh_max_size = freespace;
 204                 /* shouldn't be > 128MB in any case?
 205                  * it's 256K records of 512 bytes each */
 206                 if (freespace > (128 << 20))
 207                         loghandle->lgh_max_size = 128 << 20;
 208         }
 209         rc = 0;
 210
 211 out:
 212         if (handle != NULL) {
 213                 handle->th_result = rc >= 0 ? 0 : rc;
 214                 dt_trans_stop(env, dt, handle);
 215         }
 216         RETURN(rc);
 217
 218 out_destroy:
 219         /* to signal llog_cat_close() it shouldn't try to destroy the llog,
 220          * we want to destroy it in this transaction, otherwise the object
 221          * becomes an orphan */
 222         loghandle->lgh_hdr->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 223         /* this is to mimic full log, so another llog_cat_current_log()
 224          * can skip it and ask for another onet */
 225         loghandle->lgh_last_idx = LLOG_HDR_BITMAP_SIZE(loghandle->lgh_hdr) + 1;
 226         llog_trans_destroy(env, loghandle, th);
 227         if (handle != NULL)
 228                 dt_trans_stop(env, dt, handle);
 229         RETURN(rc);
 230 }
 231
 232 static int llog_cat_refresh(const struct lu_env *env,
 233                             struct llog_handle *cathandle)
 234 {
 235         struct llog_handle *loghandle;
 236         int rc;
 237
 238         down_write(&cathandle->lgh_lock);
 239         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 240                             u.phd.phd_entry) {
 241                 if (!llog_exist(loghandle))
 242                         continue;
 243
 244                 rc = llog_read_header(env, loghandle, NULL);
 245                 if (rc)
 246                         goto unlock;
 247         }
 248
 249         rc = llog_read_header(env, cathandle, NULL);
 250 unlock:
 251         up_write(&loghandle->lgh_lock);
 252
 253         return rc;
 254 }
 255
 256 /*
 257  * prepare current/next log for catalog.
 258  *
 259  * if log is NULL, open it, and declare create, NB, if log is remote, create it
 260  * synchronously here, see comments below.
 261  */
 262 static int llog_cat_prep_log(const struct lu_env *env,
 263                              struct llog_handle *cathandle,
 264                              struct llog_handle **ploghandle,
 265                              struct thandle *th)
 266 {
 267         int rc = 0;
 268
 269         if (IS_ERR_OR_NULL(*ploghandle)) {
 270                 down_write(&cathandle->lgh_lock);
 271                 if (IS_ERR_OR_NULL(*ploghandle)) {
 272                         struct llog_handle *loghandle;
 273
 274                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
 275                                        NULL, NULL, LLOG_OPEN_NEW);
 276                         if (!rc) {
 277                                 *ploghandle = loghandle;
 278                                 list_add_tail(&loghandle->u.phd.phd_entry,
 279                                               &cathandle->u.chd.chd_head);
 280                         }
 281                 }
 282                 up_write(&cathandle->lgh_lock);
 283
 284                 if (rc)
 285                         return rc;
 286         }
 287
 288         if (llog_exist(*ploghandle))
 289                 return 0;
 290
 291         if (dt_object_remote(cathandle->lgh_obj)) {
 292 create_remote_log:
 293                 down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 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                                 up_read(&cathandle->lgh_lock);
 310
 311                                 rc = llog_cat_refresh(env, cathandle);
 312                                 if (!rc)
 313                                         goto create_remote_log;
 314
 315                                 return rc;
 316                         }
 317                 }
 318                 up_write(&(*ploghandle)->lgh_lock);
 319                 up_read(&cathandle->lgh_lock);
 320         } else {
 321                 struct llog_thread_info *lgi = llog_info(env);
 322                 struct llog_logid_rec *lirec = &lgi->lgi_logid;
 323
 324                 rc = llog_declare_create(env, *ploghandle, th);
 325                 if (rc)
 326                         return rc;
 327
 328                 lirec->lid_hdr.lrh_len = sizeof(*lirec);
 329                 rc = llog_declare_write_rec(env, cathandle, &lirec->lid_hdr, -1,
 330                                             th);
 331         }
 332
 333         return rc;
 334 }
 335
 336 /* Open an existent log handle and add it to the open list.
 337  * This log handle will be closed when all of the records in it are removed.
 338  *
 339  * Assumes caller has already pushed us into the kernel context and is locking.
 340  * We return a lock on the handle to ensure nobody yanks it from us.
 341  *
 342  * This takes extra reference on llog_handle via llog_handle_get() and require
 343  * this reference to be put by caller using llog_handle_put()
 344  */
 345 int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
 346                        struct llog_handle **res, struct llog_logid *logid)
 347 {
 348         struct llog_handle      *loghandle;
 349         enum llog_flag           fmt;
 350         int                      rc = 0;
 351
 352         ENTRY;
 353
 354         if (cathandle == NULL)
 355                 RETURN(-EBADF);
 356
 357         fmt = cathandle->lgh_hdr->llh_flags & LLOG_F_EXT_MASK;
 358         down_write(&cathandle->lgh_lock);
 359         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 360                             u.phd.phd_entry) {
 361                 struct llog_logid *cgl = &loghandle->lgh_id;
 362
 363                 if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
 364                     ostid_seq(&cgl->lgl_oi) == ostid_seq(&logid->lgl_oi)) {
 365                         if (cgl->lgl_ogen != logid->lgl_ogen) {
 366                                 CWARN("%s: log "DFID" generation %x != %x\n",
 367                                       loghandle->lgh_ctxt->loc_obd->obd_name,
 368                                       PFID(&logid->lgl_oi.oi_fid),
 369                                       cgl->lgl_ogen, logid->lgl_ogen);
 370                                 continue;
 371                         }
 372                         loghandle->u.phd.phd_cat_handle = cathandle;
 373                         up_write(&cathandle->lgh_lock);
 374                         GOTO(out, rc = 0);
 375                 }
 376         }
 377         up_write(&cathandle->lgh_lock);
 378
 379         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle, logid, NULL,
 380                        LLOG_OPEN_EXISTS);
 381         if (rc < 0) {
 382                 CERROR("%s: error opening log id "DFID":%x: rc = %d\n",
 383                        cathandle->lgh_ctxt->loc_obd->obd_name,
 384                        PFID(&logid->lgl_oi.oi_fid), logid->lgl_ogen, rc);
 385                 RETURN(rc);
 386         }
 387
 388         rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN | fmt, NULL);
 389         if (rc < 0) {
 390                 llog_close(env, loghandle);
 391                 loghandle = NULL;
 392                 RETURN(rc);
 393         }
 394
 395         down_write(&cathandle->lgh_lock);
 396         list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
 397         up_write(&cathandle->lgh_lock);
 398
 399         loghandle->u.phd.phd_cat_handle = cathandle;
 400         loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
 401         loghandle->u.phd.phd_cookie.lgc_index =
 402                                 loghandle->lgh_hdr->llh_cat_idx;
 403         EXIT;
 404 out:
 405         llog_handle_get(loghandle);
 406         *res = loghandle;
 407         return 0;
 408 }
 409
 410 int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
 411 {
 412         struct llog_handle      *loghandle, *n;
 413         int                      rc;
 414
 415         ENTRY;
 416
 417         list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
 418                                  u.phd.phd_entry) {
 419                 struct llog_log_hdr     *llh = loghandle->lgh_hdr;
 420                 int                      index;
 421
 422                 /* unlink open-not-created llogs */
 423                 list_del_init(&loghandle->u.phd.phd_entry);
 424                 llh = loghandle->lgh_hdr;
 425                 if (loghandle->lgh_obj != NULL && llh != NULL &&
 426                     (llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 427                     (llh->llh_count == 1)) {
 428                         rc = llog_destroy(env, loghandle);
 429                         if (rc)
 430                                 CERROR("%s: failure destroying log during "
 431                                        "cleanup: rc = %d\n",
 432                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 433                                        rc);
 434
 435                         index = loghandle->u.phd.phd_cookie.lgc_index;
 436                         llog_cat_cleanup(env, cathandle, NULL, index);
 437                 }
 438                 llog_close(env, loghandle);
 439         }
 440         /* if handle was stored in ctxt, remove it too */
 441         if (cathandle->lgh_ctxt->loc_handle == cathandle)
 442                 cathandle->lgh_ctxt->loc_handle = NULL;
 443         rc = llog_close(env, cathandle);
 444         RETURN(rc);
 445 }
 446 EXPORT_SYMBOL(llog_cat_close);
 447
 448 /** Return the currently active log handle.  If the current log handle doesn't
 449  * have enough space left for the current record, start a new one.
 450  *
 451  * If reclen is 0, we only want to know what the currently active log is,
 452  * otherwise we get a lock on this log so nobody can steal our space.
 453  *
 454  * Assumes caller has already pushed us into the kernel context and is locking.
 455  *
 456  * NOTE: loghandle is write-locked upon successful return
 457  */
 458 static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
 459                                                 struct thandle *th)
 460 {
 461         struct llog_handle *loghandle = NULL;
 462         ENTRY;
 463
 464
 465         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2)) {
 466                 down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 467                 GOTO(next, loghandle);
 468         }
 469
 470         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 471         loghandle = cathandle->u.chd.chd_current_log;
 472         if (loghandle) {
 473                 struct llog_log_hdr *llh;
 474
 475                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 476                 llh = loghandle->lgh_hdr;
 477                 if (llh == NULL || !llog_is_full(loghandle)) {
 478                         up_read(&cathandle->lgh_lock);
 479                         RETURN(loghandle);
 480                 } else {
 481                         up_write(&loghandle->lgh_lock);
 482                 }
 483         }
 484         up_read(&cathandle->lgh_lock);
 485
 486         /* time to use next log */
 487
 488         /* first, we have to make sure the state hasn't changed */
 489         down_write_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                         GOTO(out_unlock, loghandle);
 498                 else
 499                         up_write(&loghandle->lgh_lock);
 500         }
 501
 502 next:
 503         /* Sigh, the chd_next_log and chd_current_log is initialized
 504          * in declare phase, and we do not serialize the catlog
 505          * accessing, so it might be possible the llog creation
 506          * thread (see llog_cat_declare_add_rec()) did not create
 507          * llog successfully, then the following thread might
 508          * meet this situation. */
 509         if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
 510                 CERROR("%s: next log does not exist!\n",
 511                        cathandle->lgh_ctxt->loc_obd->obd_name);
 512                 loghandle = ERR_PTR(-EIO);
 513                 if (cathandle->u.chd.chd_next_log == NULL) {
 514                         /* Store the error in chd_next_log, so
 515                          * the following process can get correct
 516                          * failure value */
 517                         cathandle->u.chd.chd_next_log = loghandle;
 518                 }
 519                 GOTO(out_unlock, loghandle);
 520         }
 521
 522         CDEBUG(D_INODE, "use next log\n");
 523
 524         loghandle = cathandle->u.chd.chd_next_log;
 525         cathandle->u.chd.chd_current_log = loghandle;
 526         cathandle->u.chd.chd_next_log = NULL;
 527         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 528
 529 out_unlock:
 530         up_write(&cathandle->lgh_lock);
 531         LASSERT(loghandle);
 532         RETURN(loghandle);
 533 }
 534
 535 /* Add a single record to the recovery log(s) using a catalog
 536  * Returns as llog_write_record
 537  *
 538  * Assumes caller has already pushed us into the kernel context.
 539  */
 540 int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
 541                      struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 542                      struct thandle *th)
 543 {
 544         struct llog_handle *loghandle;
 545         int rc, retried = 0;
 546         ENTRY;
 547
 548         LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
 549
 550 retry:
 551         loghandle = llog_cat_current_log(cathandle, th);
 552         if (IS_ERR(loghandle))
 553                 RETURN(PTR_ERR(loghandle));
 554
 555         /* loghandle is already locked by llog_cat_current_log() for us */
 556         if (!llog_exist(loghandle)) {
 557                 rc = llog_cat_new_log(env, cathandle, loghandle, th);
 558                 if (rc < 0) {
 559                         up_write(&loghandle->lgh_lock);
 560                         /* nobody should be trying to use this llog */
 561                         down_write(&cathandle->lgh_lock);
 562                         /* only reset current log if still room in catalog, to
 563                          * avoid unnecessarily and racy creation of new and
 564                          * partially initialized llog_handle
 565                          */
 566                         if ((cathandle->u.chd.chd_current_log == loghandle) &&
 567                             rc != -ENOSPC)
 568                                 cathandle->u.chd.chd_current_log = NULL;
 569                         up_write(&cathandle->lgh_lock);
 570                         RETURN(rc);
 571                 }
 572         }
 573         /* now let's try to add the record */
 574         rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
 575         if (rc < 0) {
 576                 CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
 577                              "llog_write_rec %d: lh=%p\n", rc, loghandle);
 578                 /* -ENOSPC is returned if no empty records left
 579                  * and when it's lack of space on the stogage.
 580                  * there is no point to try again if it's the second
 581                  * case. many callers (like llog test) expect ENOSPC,
 582                  * so we preserve this error code, but look for the
 583                  * actual cause here */
 584                 if (rc == -ENOSPC && llog_is_full(loghandle))
 585                         rc = -ENOBUFS;
 586         }
 587         up_write(&loghandle->lgh_lock);
 588
 589         if (rc == -ENOBUFS) {
 590                 if (retried++ == 0)
 591                         GOTO(retry, rc);
 592                 CERROR("%s: error on 2nd llog: rc = %d\n",
 593                        cathandle->lgh_ctxt->loc_obd->obd_name, rc);
 594         }
 595
 596         RETURN(rc);
 597 }
 598 EXPORT_SYMBOL(llog_cat_add_rec);
 599
 600 int llog_cat_declare_add_rec(const struct lu_env *env,
 601                              struct llog_handle *cathandle,
 602                              struct llog_rec_hdr *rec, struct thandle *th)
 603 {
 604         int rc;
 605
 606         ENTRY;
 607
 608         rc = llog_cat_prep_log(env, cathandle,
 609                                &cathandle->u.chd.chd_current_log, th);
 610         if (rc)
 611                 RETURN(rc);
 612
 613         rc = llog_cat_prep_log(env, cathandle, &cathandle->u.chd.chd_next_log,
 614                                th);
 615         if (rc)
 616                 RETURN(rc);
 617
 618 declare_write:
 619         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
 620                                     rec, -1, th);
 621         if (rc == -ESTALE && dt_object_remote(cathandle->lgh_obj)) {
 622                 rc = llog_cat_refresh(env, cathandle);
 623                 if (!rc)
 624                         goto declare_write;
 625         }
 626
 627 #if 0
 628         /*
 629          * XXX: we hope for declarations made for existing llog this might be
 630          * not correct with some backends where declarations are expected
 631          * against specific object like ZFS with full debugging enabled.
 632          */
 633         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_next_log, rec, -1,
 634                                     th);
 635 #endif
 636
 637         RETURN(rc);
 638 }
 639 EXPORT_SYMBOL(llog_cat_declare_add_rec);
 640
 641 int llog_cat_add(const struct lu_env *env, struct llog_handle *cathandle,
 642                  struct llog_rec_hdr *rec, struct llog_cookie *reccookie)
 643 {
 644         struct llog_ctxt        *ctxt;
 645         struct dt_device        *dt;
 646         struct thandle          *th = NULL;
 647         int                      rc;
 648
 649         ctxt = cathandle->lgh_ctxt;
 650         LASSERT(ctxt);
 651         LASSERT(ctxt->loc_exp);
 652
 653         LASSERT(cathandle->lgh_obj != NULL);
 654         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 655
 656         th = dt_trans_create(env, dt);
 657         if (IS_ERR(th))
 658                 RETURN(PTR_ERR(th));
 659
 660         rc = llog_cat_declare_add_rec(env, cathandle, rec, th);
 661         if (rc)
 662                 GOTO(out_trans, rc);
 663
 664         rc = dt_trans_start_local(env, dt, th);
 665         if (rc)
 666                 GOTO(out_trans, rc);
 667         rc = llog_cat_add_rec(env, cathandle, rec, reccookie, th);
 668 out_trans:
 669         dt_trans_stop(env, dt, th);
 670         RETURN(rc);
 671 }
 672 EXPORT_SYMBOL(llog_cat_add);
 673
 674 /* For each cookie in the cookie array, we clear the log in-use bit and either:
 675  * - the log is empty, so mark it free in the catalog header and delete it
 676  * - the log is not empty, just write out the log header
 677  *
 678  * The cookies may be in different log files, so we need to get new logs
 679  * each time.
 680  *
 681  * Assumes caller has already pushed us into the kernel context.
 682  */
 683 int llog_cat_cancel_records(const struct lu_env *env,
 684                             struct llog_handle *cathandle, int count,
 685                             struct llog_cookie *cookies)
 686 {
 687         int i, index, rc = 0, failed = 0;
 688
 689         ENTRY;
 690
 691         for (i = 0; i < count; i++, cookies++) {
 692                 struct llog_handle *loghandle;
 693                 struct llog_logid *lgl = &cookies->lgc_lgl;
 694                 int  lrc;
 695
 696                 rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
 697                 if (rc) {
 698                         CDEBUG(D_HA, "%s: cannot find llog for handle "DFID":%x"
 699                                ": rc = %d\n",
 700                                cathandle->lgh_ctxt->loc_obd->obd_name,
 701                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, rc);
 702                         failed++;
 703                         continue;
 704                 }
 705
 706                 if ((cathandle->lgh_ctxt->loc_flags &
 707                      LLOG_CTXT_FLAG_NORMAL_FID) && !llog_exist(loghandle)) {
 708                         /* For update log, some of loghandles of cathandle
 709                          * might not exist because remote llog creation might
 710                          * be failed, so let's skip the record cancellation
 711                          * for these non-exist llogs.
 712                          */
 713                         lrc = -ENOENT;
 714                         CDEBUG(D_HA, "%s: llog "DFID":%x does not exist"
 715                                ": rc = %d\n",
 716                                cathandle->lgh_ctxt->loc_obd->obd_name,
 717                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, lrc);
 718                         failed++;
 719                         if (rc == 0)
 720                                 rc = lrc;
 721                         continue;
 722                 }
 723
 724                 lrc = llog_cancel_rec(env, loghandle, cookies->lgc_index);
 725                 if (lrc == LLOG_DEL_PLAIN) { /* log has been destroyed */
 726                         index = loghandle->u.phd.phd_cookie.lgc_index;
 727                         lrc = llog_cat_cleanup(env, cathandle, loghandle,
 728                                                index);
 729                         if (rc == 0)
 730                                 rc = lrc;
 731                 } else if (lrc == -ENOENT) {
 732                         if (rc == 0) /* ENOENT shouldn't rewrite any error */
 733                                 rc = lrc;
 734                 } else if (lrc < 0) {
 735                         failed++;
 736                         if (rc == 0)
 737                                 rc = lrc;
 738                 }
 739                 llog_handle_put(loghandle);
 740         }
 741         if (rc)
 742                 CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
 743                        cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
 744                        rc);
 745
 746         RETURN(rc);
 747 }
 748 EXPORT_SYMBOL(llog_cat_cancel_records);
 749
 750 static int llog_cat_process_common(const struct lu_env *env,
 751                                    struct llog_handle *cat_llh,
 752                                    struct llog_rec_hdr *rec,
 753                                    struct llog_handle **llhp)
 754 {
 755         struct llog_logid_rec *lir = container_of(rec, typeof(*lir), lid_hdr);
 756         struct llog_log_hdr *hdr;
 757         int rc;
 758
 759         ENTRY;
 760         if (rec->lrh_type != le32_to_cpu(LLOG_LOGID_MAGIC)) {
 761                 rc = -EINVAL;
 762                 CWARN("%s: invalid record in catalog "DFID":%x: rc = %d\n",
 763                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 764                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid),
 765                       cat_llh->lgh_id.lgl_ogen, rc);
 766                 RETURN(rc);
 767         }
 768         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "DFID"\n",
 769                PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
 770                le32_to_cpu(rec->lrh_index),
 771                PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 772
 773         rc = llog_cat_id2handle(env, cat_llh, llhp, &lir->lid_id);
 774         if (rc) {
 775                 /* After a server crash, a stub of index record in catlog could
 776                  * be kept, because plain log destroy + catlog index record
 777                  * deletion are not atomic. So we end up with an index but no
 778                  * actual record. Destroy the index and move on. */
 779                 if (rc == -ENOENT || rc == -ESTALE)
 780                         rc = LLOG_DEL_RECORD;
 781                 else if (rc)
 782                         CWARN("%s: can't find llog handle "DFID":%x: rc = %d\n",
 783                               cat_llh->lgh_ctxt->loc_obd->obd_name,
 784                               PFID(&lir->lid_id.lgl_oi.oi_fid),
 785                               lir->lid_id.lgl_ogen, rc);
 786
 787                 RETURN(rc);
 788         }
 789
 790         /* clean old empty llogs, do not consider current llog in use */
 791         /* ignore remote (lgh_obj == NULL) llogs */
 792         hdr = (*llhp)->lgh_hdr;
 793         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 794             hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
 795             *llhp != cat_llh->u.chd.chd_current_log) {
 796                 rc = llog_destroy(env, *llhp);
 797                 if (rc)
 798                         CWARN("%s: can't destroy empty log "DFID": rc = %d\n",
 799                               (*llhp)->lgh_ctxt->loc_obd->obd_name,
 800                               PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
 801                 rc = LLOG_DEL_PLAIN;
 802         }
 803
 804         RETURN(rc);
 805 }
 806
 807 static int llog_cat_process_cb(const struct lu_env *env,
 808                                struct llog_handle *cat_llh,
 809                                struct llog_rec_hdr *rec, void *data)
 810 {
 811         struct llog_process_data *d = data;
 812         struct llog_handle *llh = NULL;
 813         int rc;
 814
 815         ENTRY;
 816         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 817         if (rc)
 818                 GOTO(out, rc);
 819
 820         if (rec->lrh_index < d->lpd_startcat) {
 821                 /* Skip processing of the logs until startcat */
 822                 rc = 0;
 823         } else if (d->lpd_startidx > 0) {
 824                 struct llog_process_cat_data cd;
 825
 826                 cd.lpcd_first_idx = d->lpd_startidx;
 827                 cd.lpcd_last_idx = 0;
 828                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 829                                           &cd, false);
 830                 /* Continue processing the next log from idx 0 */
 831                 d->lpd_startidx = 0;
 832         } else {
 833                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 834                                           NULL, false);
 835         }
 836
 837 out:
 838         /* The empty plain log was destroyed while processing */
 839         if (rc == LLOG_DEL_PLAIN) {
 840                 rc = llog_cat_cleanup(env, cat_llh, llh,
 841                                       llh->u.phd.phd_cookie.lgc_index);
 842         } else if (rc == LLOG_DEL_RECORD) {
 843                 /* clear wrong catalog entry */
 844                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
 845         }
 846
 847         if (llh)
 848                 llog_handle_put(llh);
 849
 850         RETURN(rc);
 851 }
 852
 853 int llog_cat_process_or_fork(const struct lu_env *env,
 854                              struct llog_handle *cat_llh, llog_cb_t cat_cb,
 855                              llog_cb_t cb, void *data, int startcat,
 856                              int startidx, bool fork)
 857 {
 858         struct llog_process_data d;
 859         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
 860         int rc;
 861         ENTRY;
 862
 863         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
 864         d.lpd_data = data;
 865         d.lpd_cb = cb;
 866         d.lpd_startcat = startcat;
 867         d.lpd_startidx = startidx;
 868
 869         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
 870             llh->llh_count > 1) {
 871                 struct llog_process_cat_data cd;
 872
 873                 CWARN("%s: catlog "DFID" crosses index zero\n",
 874                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 875                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 876
 877                 cd.lpcd_first_idx = llh->llh_cat_idx;
 878                 cd.lpcd_last_idx = 0;
 879                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 880                                           &d, &cd, fork);
 881                 if (rc != 0)
 882                         RETURN(rc);
 883
 884                 cd.lpcd_first_idx = 0;
 885                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
 886                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 887                                           &d, &cd, fork);
 888         } else {
 889                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 890                                           &d, NULL, fork);
 891         }
 892
 893         RETURN(rc);
 894 }
 895
 896 int llog_cat_process(const struct lu_env *env, struct llog_handle *cat_llh,
 897                      llog_cb_t cb, void *data, int startcat, int startidx)
 898 {
 899         return llog_cat_process_or_fork(env, cat_llh, llog_cat_process_cb,
 900                                         cb, data, startcat, startidx, false);
 901 }
 902 EXPORT_SYMBOL(llog_cat_process);
 903
 904 static int llog_cat_size_cb(const struct lu_env *env,
 905                              struct llog_handle *cat_llh,
 906                              struct llog_rec_hdr *rec, void *data)
 907 {
 908         struct llog_process_data *d = data;
 909         struct llog_handle *llh = NULL;
 910         __u64 *cum_size = d->lpd_data;
 911         __u64 size;
 912         int rc;
 913
 914         ENTRY;
 915         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 916
 917         if (rc == LLOG_DEL_PLAIN) {
 918                 /* empty log was deleted, don't count it */
 919                 rc = llog_cat_cleanup(env, cat_llh, llh,
 920                                       llh->u.phd.phd_cookie.lgc_index);
 921         } else if (rc == LLOG_DEL_RECORD) {
 922                 /* clear wrong catalog entry */
 923                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
 924         } else {
 925                 size = llog_size(env, llh);
 926                 *cum_size += size;
 927
 928                 CDEBUG(D_INFO, "Add llog entry "DFID" size=%llu, tot=%llu\n",
 929                        PFID(&llh->lgh_id.lgl_oi.oi_fid), size, *cum_size);
 930         }
 931
 932         if (llh != NULL)
 933                 llog_handle_put(llh);
 934
 935         RETURN(0);
 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 /* currently returns the number of "free" entries in catalog,
 950  * ie the available entries for a new plain LLOG file creation,
 951  * even if catalog has wrapped
 952  */
 953 __u32 llog_cat_free_space(struct llog_handle *cat_llh)
 954 {
 955         /* simulate almost full Catalog */
 956         if (OBD_FAIL_CHECK(OBD_FAIL_CAT_FREE_RECORDS))
 957                 return cfs_fail_val;
 958
 959         if (cat_llh->lgh_hdr->llh_count == 1)
 960                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1;
 961
 962         if (cat_llh->lgh_last_idx > cat_llh->lgh_hdr->llh_cat_idx)
 963                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1 +
 964                        cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
 965
 966         /* catalog is presently wrapped */
 967         return cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
 968 }
 969 EXPORT_SYMBOL(llog_cat_free_space);
 970
 971 static int llog_cat_reverse_process_cb(const struct lu_env *env,
 972                                        struct llog_handle *cat_llh,
 973                                        struct llog_rec_hdr *rec, void *data)
 974 {
 975         struct llog_process_data *d = data;
 976         struct llog_handle *llh;
 977         int rc;
 978
 979         ENTRY;
 980         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 981
 982         /* The empty plain log was destroyed while processing */
 983         if (rc == LLOG_DEL_PLAIN) {
 984                 rc = llog_cat_cleanup(env, cat_llh, llh,
 985                                       llh->u.phd.phd_cookie.lgc_index);
 986         } else if (rc == LLOG_DEL_RECORD) {
 987                 /* clear wrong catalog entry */
 988                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
 989         }
 990         if (rc)
 991                 RETURN(rc);
 992
 993         rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
 994
 995         /* The empty plain was destroyed while processing */
 996         if (rc == LLOG_DEL_PLAIN)
 997                 rc = llog_cat_cleanup(env, cat_llh, llh,
 998                                       llh->u.phd.phd_cookie.lgc_index);
 999
1000         llog_handle_put(llh);
1001         RETURN(rc);
1002 }
1003
1004 int llog_cat_reverse_process(const struct lu_env *env,
1005                              struct llog_handle *cat_llh,
1006                              llog_cb_t cb, void *data)
1007 {
1008         struct llog_process_data d;
1009         struct llog_process_cat_data cd;
1010         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
1011         int rc;
1012         ENTRY;
1013
1014         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
1015         d.lpd_data = data;
1016         d.lpd_cb = cb;
1017
1018         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
1019             llh->llh_count > 1) {
1020                 CWARN("%s: catalog "DFID" crosses index zero\n",
1021                       cat_llh->lgh_ctxt->loc_obd->obd_name,
1022                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
1023
1024                 cd.lpcd_first_idx = 0;
1025                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
1026                 rc = llog_reverse_process(env, cat_llh,
1027                                           llog_cat_reverse_process_cb,
1028                                           &d, &cd);
1029                 if (rc != 0)
1030                         RETURN(rc);
1031
1032                 cd.lpcd_first_idx = le32_to_cpu(llh->llh_cat_idx);
1033                 cd.lpcd_last_idx = 0;
1034                 rc = llog_reverse_process(env, cat_llh,
1035                                           llog_cat_reverse_process_cb,
1036                                           &d, &cd);
1037         } else {
1038                 rc = llog_reverse_process(env, cat_llh,
1039                                           llog_cat_reverse_process_cb,
1040                                           &d, NULL);
1041         }
1042
1043         RETURN(rc);
1044 }
1045 EXPORT_SYMBOL(llog_cat_reverse_process);
1046
1047 static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
1048 {
1049         struct llog_log_hdr *llh = cathandle->lgh_hdr;
1050         int bitmap_size;
1051
1052         ENTRY;
1053
1054         bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
1055         /*
1056          * The llh_cat_idx equals to the first used index minus 1
1057          * so if we canceled the first index then llh_cat_idx
1058          * must be renewed.
1059          */
1060         if (llh->llh_cat_idx == (idx - 1)) {
1061                 llh->llh_cat_idx = idx;
1062
1063                 while (idx != cathandle->lgh_last_idx) {
1064                         idx = (idx + 1) % bitmap_size;
1065                         if (!ext2_test_bit(idx, LLOG_HDR_BITMAP(llh))) {
1066                                 /* update llh_cat_idx for each unset bit,
1067                                  * expecting the next one is set */
1068                                 llh->llh_cat_idx = idx;
1069                         } else if (idx == 0) {
1070                                 /* skip header bit */
1071                                 llh->llh_cat_idx = 0;
1072                                 continue;
1073                         } else {
1074                                 /* the first index is found */
1075                                 break;
1076                         }
1077                 }
1078
1079                 CDEBUG(D_RPCTRACE, "catlog "DFID" first idx %u, last_idx %u\n",
1080                        PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
1081                        llh->llh_cat_idx, cathandle->lgh_last_idx);
1082         }
1083
1084         RETURN(0);
1085 }
1086
1087 /* Cleanup deleted plain llog traces from catalog */
1088 int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
1089                      struct llog_handle *loghandle, int index)
1090 {
1091         int rc;
1092
1093         LASSERT(index);
1094         if (loghandle != NULL) {
1095                 /* remove destroyed llog from catalog list and
1096                  * chd_current_log variable */
1097                 down_write(&cathandle->lgh_lock);
1098                 if (cathandle->u.chd.chd_current_log == loghandle)
1099                         cathandle->u.chd.chd_current_log = NULL;
1100                 list_del_init(&loghandle->u.phd.phd_entry);
1101                 up_write(&cathandle->lgh_lock);
1102                 LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
1103                 /* llog was opened and keep in a list, close it now */
1104                 llog_close(env, loghandle);
1105         }
1106
1107         /* do not attempt to cleanup on-disk llog if on client side */
1108         if (cathandle->lgh_obj == NULL)
1109                 return 0;
1110
1111         /* remove plain llog entry from catalog by index */
1112         llog_cat_set_first_idx(cathandle, index);
1113         rc = llog_cancel_rec(env, cathandle, index);
1114         if (rc == 0)
1115                 CDEBUG(D_HA, "cancel plain log at index %u of catalog "DFID"\n",
1116                        index, PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
1117         return rc;
1118 }