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 log is NULL, open it, and declare create, NB, if log is remote, create it
 251  * synchronously here, see comments below.
 252  */
 253 static int llog_cat_prep_log(const struct lu_env *env,
 254                              struct llog_handle *cathandle,
 255                              struct llog_handle **ploghandle,
 256                              struct thandle *th)
 257 {
 258         int rc = 0;
 259
 260         if (IS_ERR_OR_NULL(*ploghandle)) {
 261                 down_write(&cathandle->lgh_lock);
 262                 if (IS_ERR_OR_NULL(*ploghandle)) {
 263                         struct llog_handle *loghandle;
 264
 265                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
 266                                        NULL, NULL, LLOG_OPEN_NEW);
 267                         if (!rc) {
 268                                 *ploghandle = loghandle;
 269                                 list_add_tail(&loghandle->u.phd.phd_entry,
 270                                               &cathandle->u.chd.chd_head);
 271                         }
 272                 }
 273                 up_write(&cathandle->lgh_lock);
 274
 275                 if (rc)
 276                         return rc;
 277         }
 278
 279         if (llog_exist(*ploghandle))
 280                 return 0;
 281
 282         if (dt_object_remote(cathandle->lgh_obj)) {
 283 create_remote_log:
 284                 down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 285                 down_write_nested(&(*ploghandle)->lgh_lock, LLOGH_LOG);
 286                 if (!llog_exist(*ploghandle)) {
 287                         /* For remote operation, if we put the llog object
 288                          * creation in the current transaction, then the
 289                          * llog object will not be created on the remote
 290                          * target until the transaction stop, if other
 291                          * operations start before the transaction stop,
 292                          * and use the same llog object, will be dependent
 293                          * on the success of this transaction. So let's
 294                          * create the llog object synchronously here to
 295                          * remove the dependency. */
 296                         rc = llog_cat_new_log(env, cathandle, *ploghandle,
 297                                               NULL);
 298                         if (rc == -ESTALE) {
 299                                 up_write(&(*ploghandle)->lgh_lock);
 300                                 up_read(&cathandle->lgh_lock);
 301
 302                                 rc = llog_cat_refresh(env, cathandle);
 303                                 if (!rc)
 304                                         goto create_remote_log;
 305
 306                                 return rc;
 307                         }
 308                 }
 309                 up_write(&(*ploghandle)->lgh_lock);
 310                 up_read(&cathandle->lgh_lock);
 311         } else {
 312                 struct llog_thread_info *lgi = llog_info(env);
 313                 struct llog_logid_rec *lirec = &lgi->lgi_logid;
 314
 315                 rc = llog_declare_create(env, *ploghandle, th);
 316                 if (rc)
 317                         return rc;
 318
 319                 lirec->lid_hdr.lrh_len = sizeof(*lirec);
 320                 rc = llog_declare_write_rec(env, cathandle, &lirec->lid_hdr, -1,
 321                                             th);
 322         }
 323
 324         return rc;
 325 }
 326
 327 /* Open an existent log handle and add it to the open list.
 328  * This log handle will be closed when all of the records in it are removed.
 329  *
 330  * Assumes caller has already pushed us into the kernel context and is locking.
 331  * We return a lock on the handle to ensure nobody yanks it from us.
 332  *
 333  * This takes extra reference on llog_handle via llog_handle_get() and require
 334  * this reference to be put by caller using llog_handle_put()
 335  */
 336 int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
 337                        struct llog_handle **res, struct llog_logid *logid)
 338 {
 339         struct llog_handle      *loghandle;
 340         enum llog_flag           fmt;
 341         int                      rc = 0;
 342
 343         ENTRY;
 344
 345         if (cathandle == NULL)
 346                 RETURN(-EBADF);
 347
 348         fmt = cathandle->lgh_hdr->llh_flags & LLOG_F_EXT_MASK;
 349         down_write(&cathandle->lgh_lock);
 350         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 351                             u.phd.phd_entry) {
 352                 struct llog_logid *cgl = &loghandle->lgh_id;
 353
 354                 if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
 355                     ostid_seq(&cgl->lgl_oi) == ostid_seq(&logid->lgl_oi)) {
 356                         if (cgl->lgl_ogen != logid->lgl_ogen) {
 357                                 CWARN("%s: log "DFID" generation %x != %x\n",
 358                                       loghandle->lgh_ctxt->loc_obd->obd_name,
 359                                       PFID(&logid->lgl_oi.oi_fid),
 360                                       cgl->lgl_ogen, logid->lgl_ogen);
 361                                 continue;
 362                         }
 363                         loghandle->u.phd.phd_cat_handle = cathandle;
 364                         up_write(&cathandle->lgh_lock);
 365                         GOTO(out, rc = 0);
 366                 }
 367         }
 368         up_write(&cathandle->lgh_lock);
 369
 370         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle, logid, NULL,
 371                        LLOG_OPEN_EXISTS);
 372         if (rc < 0) {
 373                 CERROR("%s: error opening log id "DFID":%x: rc = %d\n",
 374                        cathandle->lgh_ctxt->loc_obd->obd_name,
 375                        PFID(&logid->lgl_oi.oi_fid), logid->lgl_ogen, rc);
 376                 RETURN(rc);
 377         }
 378
 379         rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN | fmt, NULL);
 380         if (rc < 0) {
 381                 llog_close(env, loghandle);
 382                 loghandle = NULL;
 383                 RETURN(rc);
 384         }
 385
 386         down_write(&cathandle->lgh_lock);
 387         list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
 388         up_write(&cathandle->lgh_lock);
 389
 390         loghandle->u.phd.phd_cat_handle = cathandle;
 391         loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
 392         loghandle->u.phd.phd_cookie.lgc_index =
 393                                 loghandle->lgh_hdr->llh_cat_idx;
 394         EXIT;
 395 out:
 396         llog_handle_get(loghandle);
 397         *res = loghandle;
 398         return 0;
 399 }
 400
 401 int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
 402 {
 403         struct llog_handle      *loghandle, *n;
 404         int                      rc;
 405
 406         ENTRY;
 407
 408         list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
 409                                  u.phd.phd_entry) {
 410                 struct llog_log_hdr     *llh = loghandle->lgh_hdr;
 411                 int                      index;
 412
 413                 /* unlink open-not-created llogs */
 414                 list_del_init(&loghandle->u.phd.phd_entry);
 415                 llh = loghandle->lgh_hdr;
 416                 if (loghandle->lgh_obj != NULL && llh != NULL &&
 417                     (llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 418                     (llh->llh_count == 1)) {
 419                         rc = llog_destroy(env, loghandle);
 420                         if (rc)
 421                                 CERROR("%s: failure destroying log during "
 422                                        "cleanup: rc = %d\n",
 423                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 424                                        rc);
 425
 426                         index = loghandle->u.phd.phd_cookie.lgc_index;
 427                         llog_cat_cleanup(env, cathandle, NULL, index);
 428                 }
 429                 llog_close(env, loghandle);
 430         }
 431         /* if handle was stored in ctxt, remove it too */
 432         if (cathandle->lgh_ctxt->loc_handle == cathandle)
 433                 cathandle->lgh_ctxt->loc_handle = NULL;
 434         rc = llog_close(env, cathandle);
 435         RETURN(rc);
 436 }
 437 EXPORT_SYMBOL(llog_cat_close);
 438
 439 /**
 440  * lockdep markers for nested struct llog_handle::lgh_lock locking.
 441  */
 442 enum {
 443         LLOGH_CAT,
 444         LLOGH_LOG
 445 };
 446
 447 /** Return the currently active log handle.  If the current log handle doesn't
 448  * have enough space left for the current record, start a new one.
 449  *
 450  * If reclen is 0, we only want to know what the currently active log is,
 451  * otherwise we get a lock on this log so nobody can steal our space.
 452  *
 453  * Assumes caller has already pushed us into the kernel context and is locking.
 454  *
 455  * NOTE: loghandle is write-locked upon successful return
 456  */
 457 static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
 458                                                 struct thandle *th)
 459 {
 460         struct llog_handle *loghandle = NULL;
 461         ENTRY;
 462
 463
 464         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2)) {
 465                 down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 466                 GOTO(next, loghandle);
 467         }
 468
 469         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
 470         loghandle = cathandle->u.chd.chd_current_log;
 471         if (loghandle) {
 472                 struct llog_log_hdr *llh;
 473
 474                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 475                 llh = loghandle->lgh_hdr;
 476                 if (llh == NULL || !llog_is_full(loghandle)) {
 477                         up_read(&cathandle->lgh_lock);
 478                         RETURN(loghandle);
 479                 } else {
 480                         up_write(&loghandle->lgh_lock);
 481                 }
 482         }
 483         up_read(&cathandle->lgh_lock);
 484
 485         /* time to use next log */
 486
 487         /* first, we have to make sure the state hasn't changed */
 488         down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
 489         loghandle = cathandle->u.chd.chd_current_log;
 490         if (loghandle) {
 491                 struct llog_log_hdr *llh;
 492
 493                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 494                 llh = loghandle->lgh_hdr;
 495                 if (llh == NULL || !llog_is_full(loghandle))
 496                         GOTO(out_unlock, loghandle);
 497                 else
 498                         up_write(&loghandle->lgh_lock);
 499         }
 500
 501 next:
 502         /* Sigh, the chd_next_log and chd_current_log is initialized
 503          * in declare phase, and we do not serialize the catlog
 504          * accessing, so it might be possible the llog creation
 505          * thread (see llog_cat_declare_add_rec()) did not create
 506          * llog successfully, then the following thread might
 507          * meet this situation. */
 508         if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
 509                 CERROR("%s: next log does not exist!\n",
 510                        cathandle->lgh_ctxt->loc_obd->obd_name);
 511                 loghandle = ERR_PTR(-EIO);
 512                 if (cathandle->u.chd.chd_next_log == NULL) {
 513                         /* Store the error in chd_next_log, so
 514                          * the following process can get correct
 515                          * failure value */
 516                         cathandle->u.chd.chd_next_log = loghandle;
 517                 }
 518                 GOTO(out_unlock, loghandle);
 519         }
 520
 521         CDEBUG(D_INODE, "use next log\n");
 522
 523         loghandle = cathandle->u.chd.chd_next_log;
 524         cathandle->u.chd.chd_current_log = loghandle;
 525         cathandle->u.chd.chd_next_log = NULL;
 526         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
 527
 528 out_unlock:
 529         up_write(&cathandle->lgh_lock);
 530         LASSERT(loghandle);
 531         RETURN(loghandle);
 532 }
 533
 534 /* Add a single record to the recovery log(s) using a catalog
 535  * Returns as llog_write_record
 536  *
 537  * Assumes caller has already pushed us into the kernel context.
 538  */
 539 int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
 540                      struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
 541                      struct thandle *th)
 542 {
 543         struct llog_handle *loghandle;
 544         int rc, retried = 0;
 545         ENTRY;
 546
 547         LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
 548
 549 retry:
 550         loghandle = llog_cat_current_log(cathandle, th);
 551         if (IS_ERR(loghandle))
 552                 RETURN(PTR_ERR(loghandle));
 553
 554         /* loghandle is already locked by llog_cat_current_log() for us */
 555         if (!llog_exist(loghandle)) {
 556                 rc = llog_cat_new_log(env, cathandle, loghandle, th);
 557                 if (rc < 0) {
 558                         up_write(&loghandle->lgh_lock);
 559                         /* nobody should be trying to use this llog */
 560                         down_write(&cathandle->lgh_lock);
 561                         /* only reset current log if still room in catalog, to
 562                          * avoid unnecessarily and racy creation of new and
 563                          * partially initialized llog_handle
 564                          */
 565                         if ((cathandle->u.chd.chd_current_log == loghandle) &&
 566                             rc != -ENOSPC)
 567                                 cathandle->u.chd.chd_current_log = NULL;
 568                         up_write(&cathandle->lgh_lock);
 569                         RETURN(rc);
 570                 }
 571         }
 572         /* now let's try to add the record */
 573         rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
 574         if (rc < 0) {
 575                 CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
 576                              "llog_write_rec %d: lh=%p\n", rc, loghandle);
 577                 /* -ENOSPC is returned if no empty records left
 578                  * and when it's lack of space on the stogage.
 579                  * there is no point to try again if it's the second
 580                  * case. many callers (like llog test) expect ENOSPC,
 581                  * so we preserve this error code, but look for the
 582                  * actual cause here */
 583                 if (rc == -ENOSPC && llog_is_full(loghandle))
 584                         rc = -ENOBUFS;
 585         }
 586         up_write(&loghandle->lgh_lock);
 587
 588         if (rc == -ENOBUFS) {
 589                 if (retried++ == 0)
 590                         GOTO(retry, rc);
 591                 CERROR("%s: error on 2nd llog: rc = %d\n",
 592                        cathandle->lgh_ctxt->loc_obd->obd_name, rc);
 593         }
 594
 595         RETURN(rc);
 596 }
 597 EXPORT_SYMBOL(llog_cat_add_rec);
 598
 599 int llog_cat_declare_add_rec(const struct lu_env *env,
 600                              struct llog_handle *cathandle,
 601                              struct llog_rec_hdr *rec, struct thandle *th)
 602 {
 603         int rc;
 604
 605         ENTRY;
 606
 607         rc = llog_cat_prep_log(env, cathandle,
 608                                &cathandle->u.chd.chd_current_log, th);
 609         if (rc)
 610                 RETURN(rc);
 611
 612         rc = llog_cat_prep_log(env, cathandle, &cathandle->u.chd.chd_next_log,
 613                                th);
 614         if (rc)
 615                 RETURN(rc);
 616
 617 declare_write:
 618         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
 619                                     rec, -1, th);
 620         if (rc == -ESTALE && dt_object_remote(cathandle->lgh_obj)) {
 621                 rc = llog_cat_refresh(env, cathandle);
 622                 if (!rc)
 623                         goto declare_write;
 624         }
 625
 626 #if 0
 627         /*
 628          * XXX: we hope for declarations made for existing llog this might be
 629          * not correct with some backends where declarations are expected
 630          * against specific object like ZFS with full debugging enabled.
 631          */
 632         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_next_log, rec, -1,
 633                                     th);
 634 #endif
 635
 636         RETURN(rc);
 637 }
 638 EXPORT_SYMBOL(llog_cat_declare_add_rec);
 639
 640 int llog_cat_add(const struct lu_env *env, struct llog_handle *cathandle,
 641                  struct llog_rec_hdr *rec, struct llog_cookie *reccookie)
 642 {
 643         struct llog_ctxt        *ctxt;
 644         struct dt_device        *dt;
 645         struct thandle          *th = NULL;
 646         int                      rc;
 647
 648         ctxt = cathandle->lgh_ctxt;
 649         LASSERT(ctxt);
 650         LASSERT(ctxt->loc_exp);
 651
 652         LASSERT(cathandle->lgh_obj != NULL);
 653         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
 654
 655         th = dt_trans_create(env, dt);
 656         if (IS_ERR(th))
 657                 RETURN(PTR_ERR(th));
 658
 659         rc = llog_cat_declare_add_rec(env, cathandle, rec, th);
 660         if (rc)
 661                 GOTO(out_trans, rc);
 662
 663         rc = dt_trans_start_local(env, dt, th);
 664         if (rc)
 665                 GOTO(out_trans, rc);
 666         rc = llog_cat_add_rec(env, cathandle, rec, reccookie, th);
 667 out_trans:
 668         dt_trans_stop(env, dt, th);
 669         RETURN(rc);
 670 }
 671 EXPORT_SYMBOL(llog_cat_add);
 672
 673 /* For each cookie in the cookie array, we clear the log in-use bit and either:
 674  * - the log is empty, so mark it free in the catalog header and delete it
 675  * - the log is not empty, just write out the log header
 676  *
 677  * The cookies may be in different log files, so we need to get new logs
 678  * each time.
 679  *
 680  * Assumes caller has already pushed us into the kernel context.
 681  */
 682 int llog_cat_cancel_records(const struct lu_env *env,
 683                             struct llog_handle *cathandle, int count,
 684                             struct llog_cookie *cookies)
 685 {
 686         int i, index, rc = 0, failed = 0;
 687
 688         ENTRY;
 689
 690         for (i = 0; i < count; i++, cookies++) {
 691                 struct llog_handle *loghandle;
 692                 struct llog_logid *lgl = &cookies->lgc_lgl;
 693                 int  lrc;
 694
 695                 rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
 696                 if (rc) {
 697                         CDEBUG(D_HA, "%s: cannot find llog for handle "DFID":%x"
 698                                ": rc = %d\n",
 699                                cathandle->lgh_ctxt->loc_obd->obd_name,
 700                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, rc);
 701                         failed++;
 702                         continue;
 703                 }
 704
 705                 if ((cathandle->lgh_ctxt->loc_flags &
 706                      LLOG_CTXT_FLAG_NORMAL_FID) && !llog_exist(loghandle)) {
 707                         /* For update log, some of loghandles of cathandle
 708                          * might not exist because remote llog creation might
 709                          * be failed, so let's skip the record cancellation
 710                          * for these non-exist llogs.
 711                          */
 712                         lrc = -ENOENT;
 713                         CDEBUG(D_HA, "%s: llog "DFID":%x does not exist"
 714                                ": rc = %d\n",
 715                                cathandle->lgh_ctxt->loc_obd->obd_name,
 716                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, lrc);
 717                         failed++;
 718                         if (rc == 0)
 719                                 rc = lrc;
 720                         continue;
 721                 }
 722
 723                 lrc = llog_cancel_rec(env, loghandle, cookies->lgc_index);
 724                 if (lrc == LLOG_DEL_PLAIN) { /* log has been destroyed */
 725                         index = loghandle->u.phd.phd_cookie.lgc_index;
 726                         lrc = llog_cat_cleanup(env, cathandle, loghandle,
 727                                                index);
 728                         if (rc == 0)
 729                                 rc = lrc;
 730                 } else if (lrc == -ENOENT) {
 731                         if (rc == 0) /* ENOENT shouldn't rewrite any error */
 732                                 rc = lrc;
 733                 } else if (lrc < 0) {
 734                         failed++;
 735                         if (rc == 0)
 736                                 rc = lrc;
 737                 }
 738                 llog_handle_put(loghandle);
 739         }
 740         if (rc)
 741                 CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
 742                        cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
 743                        rc);
 744
 745         RETURN(rc);
 746 }
 747 EXPORT_SYMBOL(llog_cat_cancel_records);
 748
 749 static int llog_cat_process_common(const struct lu_env *env,
 750                                    struct llog_handle *cat_llh,
 751                                    struct llog_rec_hdr *rec,
 752                                    struct llog_handle **llhp)
 753 {
 754         struct llog_logid_rec *lir = container_of(rec, typeof(*lir), lid_hdr);
 755         struct llog_log_hdr *hdr;
 756         int rc;
 757
 758         ENTRY;
 759         if (rec->lrh_type != le32_to_cpu(LLOG_LOGID_MAGIC)) {
 760                 rc = -EINVAL;
 761                 CWARN("%s: invalid record in catalog "DFID":%x: rc = %d\n",
 762                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 763                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid),
 764                       cat_llh->lgh_id.lgl_ogen, rc);
 765                 RETURN(rc);
 766         }
 767         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "DFID"\n",
 768                PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
 769                le32_to_cpu(rec->lrh_index),
 770                PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 771
 772         rc = llog_cat_id2handle(env, cat_llh, llhp, &lir->lid_id);
 773         if (rc) {
 774                 /* After a server crash, a stub of index record in catlog could
 775                  * be kept, because plain log destroy + catlog index record
 776                  * deletion are not atomic. So we end up with an index but no
 777                  * actual record. Destroy the index and move on. */
 778                 if (rc == -ENOENT || rc == -ESTALE)
 779                         rc = LLOG_DEL_RECORD;
 780                 else if (rc)
 781                         CWARN("%s: can't find llog handle "DFID":%x: rc = %d\n",
 782                               cat_llh->lgh_ctxt->loc_obd->obd_name,
 783                               PFID(&lir->lid_id.lgl_oi.oi_fid),
 784                               lir->lid_id.lgl_ogen, rc);
 785
 786                 RETURN(rc);
 787         }
 788
 789         /* clean old empty llogs, do not consider current llog in use */
 790         /* ignore remote (lgh_obj == NULL) llogs */
 791         hdr = (*llhp)->lgh_hdr;
 792         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 793             hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
 794             *llhp != cat_llh->u.chd.chd_current_log) {
 795                 rc = llog_destroy(env, *llhp);
 796                 if (rc)
 797                         CWARN("%s: can't destroy empty log "DFID": rc = %d\n",
 798                               (*llhp)->lgh_ctxt->loc_obd->obd_name,
 799                               PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
 800                 rc = LLOG_DEL_PLAIN;
 801         }
 802
 803         RETURN(rc);
 804 }
 805
 806 static int llog_cat_process_cb(const struct lu_env *env,
 807                                struct llog_handle *cat_llh,
 808                                struct llog_rec_hdr *rec, void *data)
 809 {
 810         struct llog_process_data *d = data;
 811         struct llog_handle *llh = NULL;
 812         int rc;
 813
 814         ENTRY;
 815         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 816         if (rc)
 817                 GOTO(out, rc);
 818
 819         if (rec->lrh_index < d->lpd_startcat) {
 820                 /* Skip processing of the logs until startcat */
 821                 rc = 0;
 822         } else if (d->lpd_startidx > 0) {
 823                 struct llog_process_cat_data cd;
 824
 825                 cd.lpcd_first_idx = d->lpd_startidx;
 826                 cd.lpcd_last_idx = 0;
 827                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 828                                           &cd, false);
 829                 /* Continue processing the next log from idx 0 */
 830                 d->lpd_startidx = 0;
 831         } else {
 832                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
 833                                           NULL, false);
 834         }
 835
 836 out:
 837         /* The empty plain log was destroyed while processing */
 838         if (rc == LLOG_DEL_PLAIN) {
 839                 rc = llog_cat_cleanup(env, cat_llh, llh,
 840                                       llh->u.phd.phd_cookie.lgc_index);
 841         } else if (rc == LLOG_DEL_RECORD) {
 842                 /* clear wrong catalog entry */
 843                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
 844         }
 845
 846         if (llh)
 847                 llog_handle_put(llh);
 848
 849         RETURN(rc);
 850 }
 851
 852 int llog_cat_process_or_fork(const struct lu_env *env,
 853                              struct llog_handle *cat_llh, llog_cb_t cat_cb,
 854                              llog_cb_t cb, void *data, int startcat,
 855                              int startidx, bool fork)
 856 {
 857         struct llog_process_data d;
 858         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
 859         int rc;
 860         ENTRY;
 861
 862         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
 863         d.lpd_data = data;
 864         d.lpd_cb = cb;
 865         d.lpd_startcat = startcat;
 866         d.lpd_startidx = startidx;
 867
 868         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
 869             llh->llh_count > 1) {
 870                 struct llog_process_cat_data cd;
 871
 872                 CWARN("%s: catlog "DFID" crosses index zero\n",
 873                       cat_llh->lgh_ctxt->loc_obd->obd_name,
 874                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
 875
 876                 cd.lpcd_first_idx = llh->llh_cat_idx;
 877                 cd.lpcd_last_idx = 0;
 878                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 879                                           &d, &cd, fork);
 880                 if (rc != 0)
 881                         RETURN(rc);
 882
 883                 cd.lpcd_first_idx = 0;
 884                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
 885                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 886                                           &d, &cd, fork);
 887         } else {
 888                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
 889                                           &d, NULL, fork);
 890         }
 891
 892         RETURN(rc);
 893 }
 894
 895 int llog_cat_process(const struct lu_env *env, struct llog_handle *cat_llh,
 896                      llog_cb_t cb, void *data, int startcat, int startidx)
 897 {
 898         return llog_cat_process_or_fork(env, cat_llh, llog_cat_process_cb,
 899                                         cb, data, startcat, startidx, false);
 900 }
 901 EXPORT_SYMBOL(llog_cat_process);
 902
 903 static int llog_cat_size_cb(const struct lu_env *env,
 904                              struct llog_handle *cat_llh,
 905                              struct llog_rec_hdr *rec, void *data)
 906 {
 907         struct llog_process_data *d = data;
 908         struct llog_handle *llh = NULL;
 909         __u64 *cum_size = d->lpd_data;
 910         __u64 size;
 911         int rc;
 912
 913         ENTRY;
 914         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 915
 916         if (rc == LLOG_DEL_PLAIN) {
 917                 /* empty log was deleted, don't count it */
 918                 rc = llog_cat_cleanup(env, cat_llh, llh,
 919                                       llh->u.phd.phd_cookie.lgc_index);
 920         } else if (rc == LLOG_DEL_RECORD) {
 921                 /* clear wrong catalog entry */
 922                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
 923         } else {
 924                 size = llog_size(env, llh);
 925                 *cum_size += size;
 926
 927                 CDEBUG(D_INFO, "Add llog entry "DFID" size=%llu, tot=%llu\n",
 928                        PFID(&llh->lgh_id.lgl_oi.oi_fid), size, *cum_size);
 929         }
 930
 931         if (llh != NULL)
 932                 llog_handle_put(llh);
 933
 934         RETURN(0);
 935 }
 936
 937 __u64 llog_cat_size(const struct lu_env *env, struct llog_handle *cat_llh)
 938 {
 939         __u64 size = llog_size(env, cat_llh);
 940
 941         llog_cat_process_or_fork(env, cat_llh, llog_cat_size_cb,
 942                                  NULL, &size, 0, 0, false);
 943
 944         return size;
 945 }
 946 EXPORT_SYMBOL(llog_cat_size);
 947
 948 /* currently returns the number of "free" entries in catalog,
 949  * ie the available entries for a new plain LLOG file creation,
 950  * even if catalog has wrapped
 951  */
 952 __u32 llog_cat_free_space(struct llog_handle *cat_llh)
 953 {
 954         /* simulate almost full Catalog */
 955         if (OBD_FAIL_CHECK(OBD_FAIL_CAT_FREE_RECORDS))
 956                 return cfs_fail_val;
 957
 958         if (cat_llh->lgh_hdr->llh_count == 1)
 959                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1;
 960
 961         if (cat_llh->lgh_last_idx > cat_llh->lgh_hdr->llh_cat_idx)
 962                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1 +
 963                        cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
 964
 965         /* catalog is presently wrapped */
 966         return cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
 967 }
 968 EXPORT_SYMBOL(llog_cat_free_space);
 969
 970 static int llog_cat_reverse_process_cb(const struct lu_env *env,
 971                                        struct llog_handle *cat_llh,
 972                                        struct llog_rec_hdr *rec, void *data)
 973 {
 974         struct llog_process_data *d = data;
 975         struct llog_handle *llh;
 976         int rc;
 977
 978         ENTRY;
 979         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 980
 981         /* The empty plain log was destroyed while processing */
 982         if (rc == LLOG_DEL_PLAIN) {
 983                 rc = llog_cat_cleanup(env, cat_llh, llh,
 984                                       llh->u.phd.phd_cookie.lgc_index);
 985         } else if (rc == LLOG_DEL_RECORD) {
 986                 /* clear wrong catalog entry */
 987                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
 988         }
 989         if (rc)
 990                 RETURN(rc);
 991
 992         rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
 993
 994         /* The empty plain was destroyed while processing */
 995         if (rc == LLOG_DEL_PLAIN)
 996                 rc = llog_cat_cleanup(env, cat_llh, llh,
 997                                       llh->u.phd.phd_cookie.lgc_index);
 998
 999         llog_handle_put(llh);
1000         RETURN(rc);
1001 }
1002
1003 int llog_cat_reverse_process(const struct lu_env *env,
1004                              struct llog_handle *cat_llh,
1005                              llog_cb_t cb, void *data)
1006 {
1007         struct llog_process_data d;
1008         struct llog_process_cat_data cd;
1009         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
1010         int rc;
1011         ENTRY;
1012
1013         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
1014         d.lpd_data = data;
1015         d.lpd_cb = cb;
1016
1017         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
1018             llh->llh_count > 1) {
1019                 CWARN("%s: catalog "DFID" crosses index zero\n",
1020                       cat_llh->lgh_ctxt->loc_obd->obd_name,
1021                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
1022
1023                 cd.lpcd_first_idx = 0;
1024                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
1025                 rc = llog_reverse_process(env, cat_llh,
1026                                           llog_cat_reverse_process_cb,
1027                                           &d, &cd);
1028                 if (rc != 0)
1029                         RETURN(rc);
1030
1031                 cd.lpcd_first_idx = le32_to_cpu(llh->llh_cat_idx);
1032                 cd.lpcd_last_idx = 0;
1033                 rc = llog_reverse_process(env, cat_llh,
1034                                           llog_cat_reverse_process_cb,
1035                                           &d, &cd);
1036         } else {
1037                 rc = llog_reverse_process(env, cat_llh,
1038                                           llog_cat_reverse_process_cb,
1039                                           &d, NULL);
1040         }
1041
1042         RETURN(rc);
1043 }
1044 EXPORT_SYMBOL(llog_cat_reverse_process);
1045
1046 static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
1047 {
1048         struct llog_log_hdr *llh = cathandle->lgh_hdr;
1049         int bitmap_size;
1050
1051         ENTRY;
1052
1053         bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
1054         /*
1055          * The llh_cat_idx equals to the first used index minus 1
1056          * so if we canceled the first index then llh_cat_idx
1057          * must be renewed.
1058          */
1059         if (llh->llh_cat_idx == (idx - 1)) {
1060                 llh->llh_cat_idx = idx;
1061
1062                 while (idx != cathandle->lgh_last_idx) {
1063                         idx = (idx + 1) % bitmap_size;
1064                         if (!ext2_test_bit(idx, LLOG_HDR_BITMAP(llh))) {
1065                                 /* update llh_cat_idx for each unset bit,
1066                                  * expecting the next one is set */
1067                                 llh->llh_cat_idx = idx;
1068                         } else if (idx == 0) {
1069                                 /* skip header bit */
1070                                 llh->llh_cat_idx = 0;
1071                                 continue;
1072                         } else {
1073                                 /* the first index is found */
1074                                 break;
1075                         }
1076                 }
1077
1078                 CDEBUG(D_RPCTRACE, "catlog "DFID" first idx %u, last_idx %u\n",
1079                        PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
1080                        llh->llh_cat_idx, cathandle->lgh_last_idx);
1081         }
1082
1083         RETURN(0);
1084 }
1085
1086 /* Cleanup deleted plain llog traces from catalog */
1087 int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
1088                      struct llog_handle *loghandle, int index)
1089 {
1090         int rc;
1091
1092         LASSERT(index);
1093         if (loghandle != NULL) {
1094                 /* remove destroyed llog from catalog list and
1095                  * chd_current_log variable */
1096                 down_write(&cathandle->lgh_lock);
1097                 if (cathandle->u.chd.chd_current_log == loghandle)
1098                         cathandle->u.chd.chd_current_log = NULL;
1099                 list_del_init(&loghandle->u.phd.phd_entry);
1100                 up_write(&cathandle->lgh_lock);
1101                 LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
1102                 /* llog was opened and keep in a list, close it now */
1103                 llog_close(env, loghandle);
1104         }
1105
1106         /* do not attempt to cleanup on-disk llog if on client side */
1107         if (cathandle->lgh_obj == NULL)
1108                 return 0;
1109
1110         /* remove plain llog entry from catalog by index */
1111         llog_cat_set_first_idx(cathandle, index);
1112         rc = llog_cancel_rec(env, cathandle, index);
1113         if (rc == 0)
1114                 CDEBUG(D_HA, "cancel plain log at index %u of catalog "DFID"\n",
1115                        index, PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
1116         return rc;
1117 }