lustre/obdclass/llog_osd.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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/llog_osd.c - low level llog routines on top of OSD API
  37  *
  38  * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
  39  * Author: Mikhail Pershin <mike.pershin@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LOG
  43
  44 #ifndef EXPORT_SYMTAB
  45 #define EXPORT_SYMTAB
  46 #endif
  47
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <lustre_fid.h>
  51 #include <dt_object.h>
  52
  53 #include "llog_internal.h"
  54 #include "local_storage.h"
  55
  56 /*
  57  * - multi-chunks or big-declaration approach
  58  * - use unique sequence instead of llog sb tracking unique ids
  59  * - re-use existing environment
  60  * - named llog support (can be used for testing only at the present)
  61  * - llog_origin_connect() work with OSD API
  62  */
  63
  64 static int llog_osd_declare_new_object(const struct lu_env *env,
  65                                        struct local_oid_storage *los,
  66                                        struct dt_object *o,
  67                                        struct thandle *th)
  68 {
  69         struct llog_thread_info *lgi = llog_info(env);
  70
  71         lgi->lgi_attr.la_valid = LA_MODE;
  72         lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
  73         lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
  74
  75         return local_object_declare_create(env, los, o, &lgi->lgi_attr,
  76                                            &lgi->lgi_dof, th);
  77 }
  78
  79 static int llog_osd_create_new_object(const struct lu_env *env,
  80                                       struct local_oid_storage *los,
  81                                       struct dt_object *o,
  82                                       struct thandle *th)
  83 {
  84         struct llog_thread_info *lgi = llog_info(env);
  85
  86         lgi->lgi_attr.la_valid = LA_MODE;
  87         lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
  88         lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
  89
  90         return local_object_create(env, los, o, &lgi->lgi_attr,
  91                                    &lgi->lgi_dof, th);
  92 }
  93
  94 static int llog_osd_pad(const struct lu_env *env, struct dt_object *o,
  95                         loff_t *off, int len, int index, struct thandle *th)
  96 {
  97         struct llog_thread_info *lgi = llog_info(env);
  98         int                      rc;
  99
 100         ENTRY;
 101
 102         LASSERT(th);
 103         LASSERT(off);
 104         LASSERT(len >= LLOG_MIN_REC_SIZE && (len & 0x7) == 0);
 105
 106         lgi->lgi_tail.lrt_len = lgi->lgi_lrh.lrh_len = len;
 107         lgi->lgi_tail.lrt_index = lgi->lgi_lrh.lrh_index = index;
 108         lgi->lgi_lrh.lrh_type = LLOG_PAD_MAGIC;
 109
 110         lgi->lgi_buf.lb_buf = &lgi->lgi_lrh;
 111         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_lrh);
 112         dt_write_lock(env, o, 0);
 113         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 114         if (rc) {
 115                 CERROR("%s: error writing padding record: rc = %d\n",
 116                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 117                 GOTO(out, rc);
 118         }
 119
 120         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 121         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 122         *off += len - sizeof(lgi->lgi_lrh) - sizeof(lgi->lgi_tail);
 123         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 124         if (rc)
 125                 CERROR("%s: error writing padding record: rc = %d\n",
 126                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 127 out:
 128         dt_write_unlock(env, o);
 129         RETURN(rc);
 130 }
 131
 132 static int llog_osd_write_blob(const struct lu_env *env, struct dt_object *o,
 133                                struct llog_rec_hdr *rec, void *buf,
 134                                loff_t *off, struct thandle *th)
 135 {
 136         struct llog_thread_info *lgi = llog_info(env);
 137         int                      buflen = rec->lrh_len;
 138         int                      rc;
 139
 140         ENTRY;
 141
 142         LASSERT(env);
 143         LASSERT(o);
 144
 145         if (buflen == 0)
 146                 CWARN("0-length record\n");
 147
 148         CDEBUG(D_OTHER, "write blob with type %x, buf %p/%u at off %llu\n",
 149                rec->lrh_type, buf, buflen, *off);
 150
 151         if (!buf) {
 152                 lgi->lgi_buf.lb_len = buflen;
 153                 lgi->lgi_buf.lb_buf = rec;
 154                 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 155                 if (rc)
 156                         CERROR("%s: error writing log record: rc = %d\n",
 157                                o->do_lu.lo_dev->ld_obd->obd_name, rc);
 158                 RETURN(rc);
 159         }
 160
 161         /* the buf case */
 162         /* protect the following 3 writes from concurrent read */
 163         dt_write_lock(env, o, 0);
 164         rec->lrh_len = sizeof(*rec) + buflen + sizeof(lgi->lgi_tail);
 165         lgi->lgi_buf.lb_len = sizeof(*rec);
 166         lgi->lgi_buf.lb_buf = rec;
 167         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 168         if (rc) {
 169                 CERROR("%s: error writing log hdr: rc = %d\n",
 170                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 171                 GOTO(out, rc);
 172         }
 173
 174         lgi->lgi_buf.lb_len = buflen;
 175         lgi->lgi_buf.lb_buf = buf;
 176         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 177         if (rc) {
 178                 CERROR("%s: error writing log buffer: rc = %d\n",
 179                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
 180                 GOTO(out, rc);
 181         }
 182
 183         lgi->lgi_tail.lrt_len = rec->lrh_len;
 184         lgi->lgi_tail.lrt_index = rec->lrh_index;
 185         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 186         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 187         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 188         if (rc)
 189                 CERROR("%s: error writing log tail: rc = %d\n",
 190                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 191 out:
 192         dt_write_unlock(env, o);
 193         RETURN(rc);
 194 }
 195
 196 static int llog_osd_read_header(const struct lu_env *env,
 197                                 struct llog_handle *handle)
 198 {
 199         struct llog_rec_hdr     *llh_hdr;
 200         struct dt_object        *o;
 201         struct llog_thread_info *lgi;
 202         int                      rc;
 203
 204         ENTRY;
 205
 206         LASSERT(sizeof(*handle->lgh_hdr) == LLOG_CHUNK_SIZE);
 207
 208         o = handle->lgh_obj;
 209         LASSERT(o);
 210
 211         lgi = llog_info(env);
 212
 213         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 214         if (rc)
 215                 RETURN(rc);
 216
 217         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 218
 219         if (lgi->lgi_attr.la_size == 0) {
 220                 CDEBUG(D_HA, "not reading header from 0-byte log\n");
 221                 RETURN(LLOG_EEMPTY);
 222         }
 223
 224         lgi->lgi_off = 0;
 225         lgi->lgi_buf.lb_buf = handle->lgh_hdr;
 226         lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE;
 227
 228         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
 229         if (rc) {
 230                 CERROR("%s: error reading log header from "DFID": rc = %d\n",
 231                        o->do_lu.lo_dev->ld_obd->obd_name,
 232                        PFID(lu_object_fid(&o->do_lu)), rc);
 233                 RETURN(rc);
 234         }
 235
 236         llh_hdr = &handle->lgh_hdr->llh_hdr;
 237         if (LLOG_REC_HDR_NEEDS_SWABBING(llh_hdr))
 238                 lustre_swab_llog_hdr(handle->lgh_hdr);
 239
 240         if (llh_hdr->lrh_type != LLOG_HDR_MAGIC) {
 241                 CERROR("%s: bad log %s "DFID" header magic: %#x "
 242                        "(expected %#x)\n", o->do_lu.lo_dev->ld_obd->obd_name,
 243                        handle->lgh_name ? handle->lgh_name : "",
 244                        PFID(lu_object_fid(&o->do_lu)),
 245                        llh_hdr->lrh_type, LLOG_HDR_MAGIC);
 246                 RETURN(-EIO);
 247         } else if (llh_hdr->lrh_len != LLOG_CHUNK_SIZE) {
 248                 CERROR("%s: incorrectly sized log %s "DFID" header: "
 249                        "%#x (expected %#x)\n"
 250                        "you may need to re-run lconf --write_conf.\n",
 251                        o->do_lu.lo_dev->ld_obd->obd_name,
 252                        handle->lgh_name ? handle->lgh_name : "",
 253                        PFID(lu_object_fid(&o->do_lu)),
 254                        llh_hdr->lrh_len, LLOG_CHUNK_SIZE);
 255                 RETURN(-EIO);
 256         }
 257
 258         handle->lgh_last_idx = handle->lgh_hdr->llh_tail.lrt_index;
 259
 260         RETURN(0);
 261 }
 262
 263 static int llog_osd_declare_write_rec(const struct lu_env *env,
 264                                       struct llog_handle *loghandle,
 265                                       struct llog_rec_hdr *rec,
 266                                       int idx, struct thandle *th)
 267 {
 268         struct llog_thread_info *lgi = llog_info(env);
 269         struct dt_object        *o;
 270         int                      rc;
 271
 272         ENTRY;
 273
 274         LASSERT(env);
 275         LASSERT(th);
 276         LASSERT(loghandle);
 277
 278         o = loghandle->lgh_obj;
 279         LASSERT(o);
 280
 281         /* each time we update header */
 282         rc = dt_declare_record_write(env, o, sizeof(struct llog_log_hdr), 0,
 283                                      th);
 284         if (rc || idx == 0) /* if error or just header */
 285                 RETURN(rc);
 286
 287         if (dt_object_exists(o)) {
 288                 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 289                 lgi->lgi_off = lgi->lgi_attr.la_size;
 290                 LASSERT(ergo(rc == 0, lgi->lgi_attr.la_valid & LA_SIZE));
 291                 if (rc)
 292                         RETURN(rc);
 293         } else {
 294                 lgi->lgi_off = 0;
 295         }
 296
 297         /* XXX: implement declared window or multi-chunks approach */
 298         rc = dt_declare_record_write(env, o, 32 * 1024, lgi->lgi_off, th);
 299
 300         RETURN(rc);
 301 }
 302
 303 /* returns negative in on error; 0 if success && reccookie == 0; 1 otherwise */
 304 /* appends if idx == -1, otherwise overwrites record idx. */
 305 static int llog_osd_write_rec(const struct lu_env *env,
 306                               struct llog_handle *loghandle,
 307                               struct llog_rec_hdr *rec,
 308                               struct llog_cookie *reccookie, int cookiecount,
 309                               void *buf, int idx, struct thandle *th)
 310 {
 311         struct llog_thread_info *lgi = llog_info(env);
 312         struct llog_log_hdr     *llh;
 313         int                      reclen = rec->lrh_len;
 314         int                      index, rc;
 315         struct llog_rec_tail    *lrt;
 316         struct dt_object        *o;
 317         size_t                   left;
 318
 319         ENTRY;
 320
 321         LASSERT(env);
 322         llh = loghandle->lgh_hdr;
 323         LASSERT(llh);
 324         o = loghandle->lgh_obj;
 325         LASSERT(o);
 326         LASSERT(th);
 327
 328         CDEBUG(D_OTHER, "new record %x to "DFID"\n",
 329                rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
 330
 331         /* record length should not bigger than LLOG_CHUNK_SIZE */
 332         if (buf)
 333                 rc = (reclen > LLOG_CHUNK_SIZE - sizeof(struct llog_rec_hdr) -
 334                       sizeof(struct llog_rec_tail)) ? -E2BIG : 0;
 335         else
 336                 rc = (reclen > LLOG_CHUNK_SIZE) ? -E2BIG : 0;
 337         if (rc)
 338                 RETURN(rc);
 339
 340         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 341         if (rc)
 342                 RETURN(rc);
 343
 344         if (buf)
 345                 /* write_blob adds header and tail to lrh_len. */
 346                 reclen = sizeof(*rec) + rec->lrh_len +
 347                          sizeof(struct llog_rec_tail);
 348
 349         if (idx != -1) {
 350                 /* no header: only allowed to insert record 1 */
 351                 if (idx != 1 && lgi->lgi_attr.la_size == 0)
 352                         LBUG();
 353
 354                 if (idx && llh->llh_size && llh->llh_size != rec->lrh_len)
 355                         RETURN(-EINVAL);
 356
 357                 if (!ext2_test_bit(idx, llh->llh_bitmap))
 358                         CERROR("%s: modify unset record %u\n",
 359                                o->do_lu.lo_dev->ld_obd->obd_name, idx);
 360                 if (idx != rec->lrh_index)
 361                         CERROR("%s: index mismatch %d %u\n",
 362                                o->do_lu.lo_dev->ld_obd->obd_name, idx,
 363                                rec->lrh_index);
 364
 365                 lgi->lgi_off = 0;
 366                 rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 367                                          &lgi->lgi_off, th);
 368                 /* we are done if we only write the header or on error */
 369                 if (rc || idx == 0)
 370                         RETURN(rc);
 371
 372                 if (buf) {
 373                         /* We assume that caller has set lgh_cur_* */
 374                         lgi->lgi_off = loghandle->lgh_cur_offset;
 375                         CDEBUG(D_OTHER,
 376                                "modify record "LPX64": idx:%d/%u/%d, len:%u "
 377                                "offset %llu\n",
 378                                loghandle->lgh_id.lgl_oid, idx, rec->lrh_index,
 379                                loghandle->lgh_cur_idx, rec->lrh_len,
 380                                (long long)(lgi->lgi_off - sizeof(*llh)));
 381                         if (rec->lrh_index != loghandle->lgh_cur_idx) {
 382                                 CERROR("%s: modify idx mismatch %u/%d\n",
 383                                        o->do_lu.lo_dev->ld_obd->obd_name, idx,
 384                                        loghandle->lgh_cur_idx);
 385                                 RETURN(-EFAULT);
 386                         }
 387                 } else {
 388                         /* Assumes constant lrh_len */
 389                         lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
 390                 }
 391
 392                 rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 393                 if (rc == 0 && reccookie) {
 394                         reccookie->lgc_lgl = loghandle->lgh_id;
 395                         reccookie->lgc_index = idx;
 396                         rc = 1;
 397                 }
 398                 RETURN(rc);
 399         }
 400
 401         /* Make sure that records don't cross a chunk boundary, so we can
 402          * process them page-at-a-time if needed.  If it will cross a chunk
 403          * boundary, write in a fake (but referenced) entry to pad the chunk.
 404          *
 405          * We know that llog_current_log() will return a loghandle that is
 406          * big enough to hold reclen, so all we care about is padding here.
 407          */
 408         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 409         lgi->lgi_off = lgi->lgi_attr.la_size;
 410         left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
 411         /* NOTE: padding is a record, but no bit is set */
 412         if (left != 0 && left != reclen &&
 413             left < (reclen + LLOG_MIN_REC_SIZE)) {
 414                 index = loghandle->lgh_last_idx + 1;
 415                 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
 416                 if (rc)
 417                         RETURN(rc);
 418                 loghandle->lgh_last_idx++; /*for pad rec*/
 419         }
 420         /* if it's the last idx in log file, then return -ENOSPC */
 421         if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
 422                 RETURN(-ENOSPC);
 423
 424         loghandle->lgh_last_idx++;
 425         index = loghandle->lgh_last_idx;
 426         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 427         rec->lrh_index = index;
 428         if (buf == NULL) {
 429                 lrt = (struct llog_rec_tail *)((char *)rec + rec->lrh_len -
 430                                                sizeof(*lrt));
 431                 lrt->lrt_len = rec->lrh_len;
 432                 lrt->lrt_index = rec->lrh_index;
 433         }
 434         /* The caller should make sure only 1 process access the lgh_last_idx,
 435          * Otherwise it might hit the assert.*/
 436         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 437         spin_lock(&loghandle->lgh_hdr_lock);
 438         if (ext2_set_bit(index, llh->llh_bitmap)) {
 439                 CERROR("%s: index %u already set in log bitmap\n",
 440                        o->do_lu.lo_dev->ld_obd->obd_name, index);
 441                 spin_unlock(&loghandle->lgh_hdr_lock);
 442                 LBUG(); /* should never happen */
 443         }
 444         llh->llh_count++;
 445         spin_unlock(&loghandle->lgh_hdr_lock);
 446         llh->llh_tail.lrt_index = index;
 447
 448         lgi->lgi_off = 0;
 449         rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL, &lgi->lgi_off,
 450                                  th);
 451         if (rc)
 452                 RETURN(rc);
 453
 454         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 455         if (rc)
 456                 RETURN(rc);
 457         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 458         lgi->lgi_off = lgi->lgi_attr.la_size;
 459
 460         rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 461         if (rc)
 462                 RETURN(rc);
 463
 464         CDEBUG(D_RPCTRACE, "added record "LPX64": idx: %u, %u\n",
 465                loghandle->lgh_id.lgl_oid, index, rec->lrh_len);
 466         if (rc == 0 && reccookie) {
 467                 reccookie->lgc_lgl = loghandle->lgh_id;
 468                 reccookie->lgc_index = index;
 469                 if ((rec->lrh_type == MDS_UNLINK_REC) ||
 470                     (rec->lrh_type == MDS_SETATTR64_REC))
 471                         reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 472                 else if (rec->lrh_type == OST_SZ_REC)
 473                         reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
 474                 else
 475                         reccookie->lgc_subsys = -1;
 476                 rc = 1;
 477         }
 478         RETURN(rc);
 479 }
 480
 481 /* We can skip reading at least as many log blocks as the number of
 482  * minimum sized log records we are skipping.  If it turns out
 483  * that we are not far enough along the log (because the
 484  * actual records are larger than minimum size) we just skip
 485  * some more records.
 486  */
 487 static void llog_skip_over(__u64 *off, int curr, int goal)
 488 {
 489         if (goal <= curr)
 490                 return;
 491         *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
 492                 ~(LLOG_CHUNK_SIZE - 1);
 493 }
 494
 495 /* sets:
 496  *  - cur_offset to the furthest point read in the log file
 497  *  - cur_idx to the log index preceeding cur_offset
 498  * returns -EIO/-EINVAL on error
 499  */
 500 static int llog_osd_next_block(const struct lu_env *env,
 501                                struct llog_handle *loghandle, int *cur_idx,
 502                                int next_idx, __u64 *cur_offset, void *buf,
 503                                int len)
 504 {
 505         struct llog_thread_info *lgi = llog_info(env);
 506         struct dt_object        *o;
 507         struct dt_device        *dt;
 508         int                      rc;
 509
 510         ENTRY;
 511
 512         LASSERT(env);
 513         LASSERT(lgi);
 514
 515         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 516                 RETURN(-EINVAL);
 517
 518         CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
 519                next_idx, *cur_idx, *cur_offset);
 520
 521         LASSERT(loghandle);
 522         LASSERT(loghandle->lgh_ctxt);
 523
 524         o = loghandle->lgh_obj;
 525         LASSERT(o);
 526         LASSERT(dt_object_exists(o));
 527         dt = lu2dt_dev(o->do_lu.lo_dev);
 528         LASSERT(dt);
 529
 530         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 531         if (rc)
 532                 GOTO(out, rc);
 533
 534         while (*cur_offset < lgi->lgi_attr.la_size) {
 535                 struct llog_rec_hdr     *rec, *last_rec;
 536                 struct llog_rec_tail    *tail;
 537
 538                 llog_skip_over(cur_offset, *cur_idx, next_idx);
 539
 540                 /* read up to next LLOG_CHUNK_SIZE block */
 541                 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
 542                                       (*cur_offset & (LLOG_CHUNK_SIZE - 1));
 543                 lgi->lgi_buf.lb_buf = buf;
 544
 545                 /* Note: read lock is not needed around la_size get above at
 546                  * the time of dt_attr_get(). There are only two cases that
 547                  * matter. Either la_size == cur_offset, in which case the
 548                  * entire read is skipped, or la_size > cur_offset and the loop
 549                  * is entered and this thread is blocked at dt_read_lock()
 550                  * until the write is completed. When the write completes, then
 551                  * the dt_read() will be done with the full length, and will
 552                  * get the full data.
 553                  */
 554                 dt_read_lock(env, o, 0);
 555                 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
 556                 dt_read_unlock(env, o);
 557                 if (rc < 0) {
 558                         CERROR("%s: can't read llog block from log "DFID
 559                                " offset "LPU64": rc = %d\n",
 560                                o->do_lu.lo_dev->ld_obd->obd_name,
 561                                PFID(lu_object_fid(&o->do_lu)), *cur_offset,
 562                                rc);
 563                         GOTO(out, rc);
 564                 }
 565
 566                 if (rc < len) {
 567                         /* signal the end of the valid buffer to
 568                          * llog_process */
 569                         memset(buf + rc, 0, len - rc);
 570                 }
 571
 572                 if (rc == 0) /* end of file, nothing to do */
 573                         GOTO(out, rc);
 574
 575                 if (rc < sizeof(*tail)) {
 576                         CERROR("%s: invalid llog block at log id "LPU64"/%u "
 577                                "offset "LPU64"\n",
 578                                o->do_lu.lo_dev->ld_obd->obd_name,
 579                                loghandle->lgh_id.lgl_oid,
 580                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 581                         GOTO(out, rc = -EINVAL);
 582                 }
 583
 584                 rec = buf;
 585                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 586                         lustre_swab_llog_rec(rec);
 587
 588                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 589                                                 sizeof(struct llog_rec_tail));
 590                 /* get the last record in block */
 591                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 592                                                    le32_to_cpu(tail->lrt_len));
 593
 594                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 595                         lustre_swab_llog_rec(last_rec);
 596                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 597
 598                 *cur_idx = tail->lrt_index;
 599
 600                 /* this shouldn't happen */
 601                 if (tail->lrt_index == 0) {
 602                         CERROR("%s: invalid llog tail at log id "LPU64"/%u "
 603                                "offset "LPU64"\n",
 604                                o->do_lu.lo_dev->ld_obd->obd_name,
 605                                loghandle->lgh_id.lgl_oid,
 606                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 607                         GOTO(out, rc = -EINVAL);
 608                 }
 609                 if (tail->lrt_index < next_idx)
 610                         continue;
 611
 612                 /* sanity check that the start of the new buffer is no farther
 613                  * than the record that we wanted.  This shouldn't happen. */
 614                 if (rec->lrh_index > next_idx) {
 615                         CERROR("%s: missed desired record? %u > %u\n",
 616                                o->do_lu.lo_dev->ld_obd->obd_name,
 617                                rec->lrh_index, next_idx);
 618                         GOTO(out, rc = -ENOENT);
 619                 }
 620                 GOTO(out, rc = 0);
 621         }
 622         GOTO(out, rc = -EIO);
 623 out:
 624         return rc;
 625 }
 626
 627 static int llog_osd_prev_block(const struct lu_env *env,
 628                                struct llog_handle *loghandle,
 629                                int prev_idx, void *buf, int len)
 630 {
 631         struct llog_thread_info *lgi = llog_info(env);
 632         struct dt_object        *o;
 633         struct dt_device        *dt;
 634         loff_t                   cur_offset;
 635         int                      rc;
 636
 637         ENTRY;
 638
 639         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 640                 RETURN(-EINVAL);
 641
 642         CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
 643
 644         LASSERT(loghandle);
 645         LASSERT(loghandle->lgh_ctxt);
 646
 647         o = loghandle->lgh_obj;
 648         LASSERT(o);
 649         LASSERT(dt_object_exists(o));
 650         dt = lu2dt_dev(o->do_lu.lo_dev);
 651         LASSERT(dt);
 652
 653         cur_offset = LLOG_CHUNK_SIZE;
 654         llog_skip_over(&cur_offset, 0, prev_idx);
 655
 656         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 657         if (rc)
 658                 GOTO(out, rc);
 659
 660         while (cur_offset < lgi->lgi_attr.la_size) {
 661                 struct llog_rec_hdr     *rec, *last_rec;
 662                 struct llog_rec_tail    *tail;
 663
 664                 lgi->lgi_buf.lb_len = len;
 665                 lgi->lgi_buf.lb_buf = buf;
 666                 /* It is OK to have locking around dt_read() only, see
 667                  * comment in llog_osd_next_block for details
 668                  */
 669                 dt_read_lock(env, o, 0);
 670                 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
 671                 dt_read_unlock(env, o);
 672                 if (rc < 0) {
 673                         CERROR("%s: can't read llog block from log "DFID
 674                                " offset "LPU64": rc = %d\n",
 675                                o->do_lu.lo_dev->ld_obd->obd_name,
 676                                PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
 677                         GOTO(out, rc);
 678                 }
 679
 680                 if (rc == 0) /* end of file, nothing to do */
 681                         GOTO(out, rc);
 682
 683                 if (rc < sizeof(*tail)) {
 684                         CERROR("%s: invalid llog block at log id "LPU64"/%u "
 685                                "offset "LPU64"\n",
 686                                o->do_lu.lo_dev->ld_obd->obd_name,
 687                                loghandle->lgh_id.lgl_oid,
 688                                loghandle->lgh_id.lgl_ogen, cur_offset);
 689                         GOTO(out, rc = -EINVAL);
 690                 }
 691
 692                 rec = buf;
 693                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 694                         lustre_swab_llog_rec(rec);
 695
 696                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 697                                                 sizeof(struct llog_rec_tail));
 698                 /* get the last record in block */
 699                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 700                                                    le32_to_cpu(tail->lrt_len));
 701
 702                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 703                         lustre_swab_llog_rec(last_rec);
 704                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 705
 706                 /* this shouldn't happen */
 707                 if (tail->lrt_index == 0) {
 708                         CERROR("%s: invalid llog tail at log id "LPU64"/%u "
 709                                "offset "LPU64"\n",
 710                                o->do_lu.lo_dev->ld_obd->obd_name,
 711                                loghandle->lgh_id.lgl_oid,
 712                                loghandle->lgh_id.lgl_ogen, cur_offset);
 713                         GOTO(out, rc = -EINVAL);
 714                 }
 715                 if (tail->lrt_index < prev_idx)
 716                         continue;
 717
 718                 /* sanity check that the start of the new buffer is no farther
 719                  * than the record that we wanted.  This shouldn't happen. */
 720                 if (rec->lrh_index > prev_idx) {
 721                         CERROR("%s: missed desired record? %u > %u\n",
 722                                o->do_lu.lo_dev->ld_obd->obd_name,
 723                                rec->lrh_index, prev_idx);
 724                         GOTO(out, rc = -ENOENT);
 725                 }
 726                 GOTO(out, rc = 0);
 727         }
 728         GOTO(out, rc = -EIO);
 729 out:
 730         return rc;
 731 }
 732
 733 struct dt_object *llog_osd_dir_get(const struct lu_env *env,
 734                                    struct llog_ctxt *ctxt)
 735 {
 736         struct dt_device        *dt;
 737         struct dt_thread_info   *dti = dt_info(env);
 738         struct dt_object        *dir;
 739         int                      rc;
 740
 741         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 742         if (ctxt->loc_dir == NULL) {
 743                 rc = dt_root_get(env, dt, &dti->dti_fid);
 744                 if (rc)
 745                         return ERR_PTR(rc);
 746                 dir = dt_locate(env, dt, &dti->dti_fid);
 747         } else {
 748                 lu_object_get(&ctxt->loc_dir->do_lu);
 749                 dir = ctxt->loc_dir;
 750         }
 751
 752         return dir;
 753 }
 754
 755 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
 756                          struct llog_logid *logid, char *name,
 757                          enum llog_open_param open_param)
 758 {
 759         struct llog_thread_info         *lgi = llog_info(env);
 760         struct llog_ctxt                *ctxt = handle->lgh_ctxt;
 761         struct dt_object                *o;
 762         struct dt_device                *dt;
 763         struct ls_device                *ls;
 764         struct local_oid_storage        *los;
 765         int                              rc = 0;
 766
 767         ENTRY;
 768
 769         LASSERT(env);
 770         LASSERT(ctxt);
 771         LASSERT(ctxt->loc_exp);
 772         LASSERT(ctxt->loc_exp->exp_obd);
 773         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 774         LASSERT(dt);
 775
 776         ls = ls_device_get(dt);
 777         if (IS_ERR(ls))
 778                 RETURN(PTR_ERR(ls));
 779
 780         mutex_lock(&ls->ls_los_mutex);
 781         los = dt_los_find(ls, FID_SEQ_LLOG);
 782         mutex_unlock(&ls->ls_los_mutex);
 783         LASSERT(los);
 784         ls_device_put(env, ls);
 785
 786         LASSERT(handle);
 787
 788         if (logid != NULL) {
 789                 logid_to_fid(logid, &lgi->lgi_fid);
 790         } else if (name) {
 791                 struct dt_object *llog_dir;
 792
 793                 llog_dir = llog_osd_dir_get(env, ctxt);
 794                 if (IS_ERR(llog_dir))
 795                         GOTO(out, rc = PTR_ERR(llog_dir));
 796                 dt_read_lock(env, llog_dir, 0);
 797                 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
 798                 dt_read_unlock(env, llog_dir);
 799                 lu_object_put(env, &llog_dir->do_lu);
 800                 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
 801                         /* generate fid for new llog */
 802                         rc = local_object_fid_generate(env, los,
 803                                                        &lgi->lgi_fid);
 804                 }
 805                 if (rc < 0)
 806                         GOTO(out, rc);
 807                 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
 808                 if (handle->lgh_name)
 809                         strcpy(handle->lgh_name, name);
 810                 else
 811                         GOTO(out, rc = -ENOMEM);
 812         } else {
 813                 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
 814                 /* generate fid for new llog */
 815                 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
 816                 if (rc < 0)
 817                         GOTO(out, rc);
 818         }
 819
 820         o = ls_locate(env, ls, &lgi->lgi_fid);
 821         if (IS_ERR(o))
 822                 GOTO(out_name, rc = PTR_ERR(o));
 823
 824         /* No new llog is expected but doesn't exist */
 825         if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
 826                 GOTO(out_put, rc = -ENOENT);
 827
 828         fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
 829         handle->lgh_obj = o;
 830         handle->private_data = los;
 831         LASSERT(handle->lgh_ctxt);
 832
 833         RETURN(rc);
 834
 835 out_put:
 836         lu_object_put(env, &o->do_lu);
 837 out_name:
 838         if (handle->lgh_name != NULL)
 839                 OBD_FREE(handle->lgh_name, strlen(name) + 1);
 840 out:
 841         dt_los_put(los);
 842         RETURN(rc);
 843 }
 844
 845 static int llog_osd_exist(struct llog_handle *handle)
 846 {
 847         LASSERT(handle->lgh_obj);
 848         return (dt_object_exists(handle->lgh_obj) &&
 849                 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
 850 }
 851
 852 static int llog_osd_declare_create(const struct lu_env *env,
 853                                    struct llog_handle *res, struct thandle *th)
 854 {
 855         struct llog_thread_info         *lgi = llog_info(env);
 856         struct local_oid_storage        *los;
 857         struct dt_object                *o;
 858         int                              rc;
 859
 860         ENTRY;
 861
 862         LASSERT(res->lgh_obj);
 863         LASSERT(th);
 864
 865         /* object can be created by another thread */
 866         o = res->lgh_obj;
 867         if (dt_object_exists(o))
 868                 RETURN(0);
 869
 870         los = res->private_data;
 871         LASSERT(los);
 872
 873         rc = llog_osd_declare_new_object(env, los, o, th);
 874         if (rc)
 875                 RETURN(rc);
 876
 877         rc = dt_declare_record_write(env, o, LLOG_CHUNK_SIZE, 0, th);
 878         if (rc)
 879                 RETURN(rc);
 880
 881         if (res->lgh_name) {
 882                 struct dt_object *llog_dir;
 883
 884                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 885                 if (IS_ERR(llog_dir))
 886                         RETURN(PTR_ERR(llog_dir));
 887                 dt_declare_ref_add(env, o, th);
 888                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 889                 rc = dt_declare_insert(env, llog_dir,
 890                                        (struct dt_rec *)&lgi->lgi_fid,
 891                                        (struct dt_key *)res->lgh_name, th);
 892                 lu_object_put(env, &llog_dir->do_lu);
 893                 if (rc)
 894                         CERROR("%s: can't declare named llog %s: rc = %d\n",
 895                                o->do_lu.lo_dev->ld_obd->obd_name,
 896                                res->lgh_name, rc);
 897         }
 898         RETURN(rc);
 899 }
 900
 901 /* This is a callback from the llog_* functions.
 902  * Assumes caller has already pushed us into the kernel context. */
 903 static int llog_osd_create(const struct lu_env *env, struct llog_handle *res,
 904                            struct thandle *th)
 905 {
 906         struct llog_thread_info *lgi = llog_info(env);
 907         struct local_oid_storage *los;
 908         struct dt_object        *o;
 909         int                      rc = 0;
 910
 911         ENTRY;
 912
 913         LASSERT(env);
 914         o = res->lgh_obj;
 915         LASSERT(o);
 916
 917         /* llog can be already created */
 918         if (dt_object_exists(o))
 919                 RETURN(-EEXIST);
 920
 921         los = res->private_data;
 922         LASSERT(los);
 923
 924         dt_write_lock(env, o, 0);
 925         if (!dt_object_exists(o))
 926                 rc = llog_osd_create_new_object(env, los, o, th);
 927         else
 928                 rc = -EEXIST;
 929         if (res->lgh_name)
 930                 dt_ref_add(env, o, th);
 931         dt_write_unlock(env, o);
 932         if (rc)
 933                 RETURN(rc);
 934
 935         if (res->lgh_name) {
 936                 struct dt_object *llog_dir;
 937
 938                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 939                 if (IS_ERR(llog_dir))
 940                         RETURN(PTR_ERR(llog_dir));
 941
 942                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 943                 dt_read_lock(env, llog_dir, 0);
 944                 rc = dt_insert(env, llog_dir,
 945                                (struct dt_rec *)&lgi->lgi_fid,
 946                                (struct dt_key *)res->lgh_name,
 947                                th, BYPASS_CAPA, 1);
 948                 dt_read_unlock(env, llog_dir);
 949                 lu_object_put(env, &llog_dir->do_lu);
 950                 if (rc)
 951                         CERROR("%s: can't create named llog %s: rc = %d\n",
 952                                o->do_lu.lo_dev->ld_obd->obd_name,
 953                                res->lgh_name, rc);
 954         }
 955         RETURN(rc);
 956 }
 957
 958 static int llog_osd_close(const struct lu_env *env, struct llog_handle *handle)
 959 {
 960         struct local_oid_storage        *los;
 961         int                              rc = 0;
 962
 963         ENTRY;
 964
 965         LASSERT(handle->lgh_obj);
 966
 967         lu_object_put(env, &handle->lgh_obj->do_lu);
 968
 969         los = handle->private_data;
 970         LASSERT(los);
 971         dt_los_put(los);
 972
 973         if (handle->lgh_name)
 974                 OBD_FREE(handle->lgh_name, strlen(handle->lgh_name) + 1);
 975
 976         RETURN(rc);
 977 }
 978
 979 static int llog_osd_destroy(const struct lu_env *env,
 980                             struct llog_handle *loghandle)
 981 {
 982         struct llog_thread_info *lgi = llog_info(env);
 983         struct llog_ctxt        *ctxt;
 984         struct dt_object        *o, *llog_dir = NULL;
 985         struct dt_device        *d;
 986         struct thandle          *th;
 987         char                    *name = NULL;
 988         int                      rc;
 989
 990         ENTRY;
 991
 992         ctxt = loghandle->lgh_ctxt;
 993         LASSERT(ctxt);
 994
 995         o = loghandle->lgh_obj;
 996         LASSERT(o);
 997
 998         d = lu2dt_dev(o->do_lu.lo_dev);
 999         LASSERT(d);
1000         LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
1001
1002         th = dt_trans_create(env, d);
1003         if (IS_ERR(th))
1004                 RETURN(PTR_ERR(th));
1005
1006         if (loghandle->lgh_name) {
1007                 llog_dir = llog_osd_dir_get(env, ctxt);
1008                 if (IS_ERR(llog_dir))
1009                         GOTO(out_trans, rc = PTR_ERR(llog_dir));
1010
1011                 dt_declare_ref_del(env, o, th);
1012                 name = loghandle->lgh_name;
1013                 rc = dt_declare_delete(env, llog_dir,
1014                                        (struct dt_key *)name, th);
1015                 if (rc)
1016                         GOTO(out_trans, rc);
1017         }
1018
1019         dt_declare_ref_del(env, o, th);
1020
1021         rc = dt_declare_destroy(env, o, th);
1022         if (rc)
1023                 GOTO(out_trans, rc);
1024
1025         rc = dt_trans_start_local(env, d, th);
1026         if (rc)
1027                 GOTO(out_trans, rc);
1028
1029         dt_write_lock(env, o, 0);
1030         if (dt_object_exists(o)) {
1031                 if (name) {
1032                         dt_ref_del(env, o, th);
1033                         dt_read_lock(env, llog_dir, 0);
1034                         rc = dt_delete(env, llog_dir,
1035                                        (struct dt_key *) name,
1036                                        th, BYPASS_CAPA);
1037                         dt_read_unlock(env, llog_dir);
1038                         if (rc) {
1039                                 CERROR("%s: can't remove llog %s: rc = %d\n",
1040                                        o->do_lu.lo_dev->ld_obd->obd_name,
1041                                        name, rc);
1042                                 GOTO(out_unlock, rc);
1043                         }
1044                 }
1045                 /*
1046                  * XXX: compatibility bits
1047                  *      on old filesystems llogs are referenced by the name
1048                  *      on the new ones they are referenced by OI and by
1049                  *      the name
1050                  */
1051                 rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
1052                 if (rc)
1053                         GOTO(out_unlock, rc);
1054                 LASSERT(lgi->lgi_attr.la_nlink < 2);
1055                 if (lgi->lgi_attr.la_nlink == 1)
1056                         dt_ref_del(env, o, th);
1057                 rc = dt_destroy(env, o, th);
1058                 if (rc)
1059                         GOTO(out_unlock, rc);
1060         }
1061 out_unlock:
1062         dt_write_unlock(env, o);
1063 out_trans:
1064         dt_trans_stop(env, d, th);
1065         if (llog_dir != NULL)
1066                 lu_object_put(env, &llog_dir->do_lu);
1067         RETURN(rc);
1068 }
1069
1070 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1071                           struct obd_llog_group *olg, int ctxt_idx,
1072                           struct obd_device *disk_obd)
1073 {
1074         struct local_oid_storage        *los;
1075         struct llog_thread_info         *lgi = llog_info(env);
1076         struct llog_ctxt                *ctxt;
1077         int                              rc = 0;
1078
1079         ENTRY;
1080
1081         LASSERT(obd);
1082         LASSERT(olg->olg_ctxts[ctxt_idx]);
1083
1084         ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1085         LASSERT(ctxt);
1086
1087         /* initialize data allowing to generate new fids,
1088          * literally we need a sequece */
1089         lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1090         lgi->lgi_fid.f_oid = 1;
1091         lgi->lgi_fid.f_ver = 0;
1092         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1093                                     &lgi->lgi_fid, &los);
1094         llog_ctxt_put(ctxt);
1095         return rc;
1096 }
1097
1098 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1099 {
1100         struct dt_device                *dt;
1101         struct ls_device                *ls;
1102         struct local_oid_storage        *los;
1103
1104         LASSERT(ctxt->loc_exp->exp_obd);
1105         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
1106         ls = ls_device_get(dt);
1107         if (IS_ERR(ls))
1108                 RETURN(PTR_ERR(ls));
1109
1110         mutex_lock(&ls->ls_los_mutex);
1111         los = dt_los_find(ls, FID_SEQ_LLOG);
1112         mutex_unlock(&ls->ls_los_mutex);
1113         if (los != NULL) {
1114                 dt_los_put(los);
1115                 local_oid_storage_fini(env, los);
1116         }
1117         ls_device_put(env, ls);
1118         return 0;
1119 }
1120
1121 struct llog_operations llog_osd_ops = {
1122         .lop_next_block         = llog_osd_next_block,
1123         .lop_prev_block         = llog_osd_prev_block,
1124         .lop_read_header        = llog_osd_read_header,
1125         .lop_destroy            = llog_osd_destroy,
1126         .lop_setup              = llog_osd_setup,
1127         .lop_cleanup            = llog_osd_cleanup,
1128         .lop_open               = llog_osd_open,
1129         .lop_exist              = llog_osd_exist,
1130         .lop_declare_create     = llog_osd_declare_create,
1131         .lop_create             = llog_osd_create,
1132         .lop_declare_write_rec  = llog_osd_declare_write_rec,
1133         .lop_write_rec          = llog_osd_write_rec,
1134         .lop_close              = llog_osd_close,
1135 };
1136 EXPORT_SYMBOL(llog_osd_ops);
1137
1138 /* reads the catalog list */
1139 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1140                           int idx, int count, struct llog_catid *idarray)
1141 {
1142         struct llog_thread_info *lgi = llog_info(env);
1143         struct dt_object        *o = NULL;
1144         struct thandle          *th;
1145         int                      rc, size;
1146
1147         ENTRY;
1148
1149         LASSERT(d);
1150
1151         size = sizeof(*idarray) * count;
1152         lgi->lgi_off = idx *  sizeof(*idarray);
1153
1154         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1155         o = dt_locate(env, d, &lgi->lgi_fid);
1156         if (IS_ERR(o))
1157                 RETURN(PTR_ERR(o));
1158
1159         if (!dt_object_exists(o)) {
1160                 th = dt_trans_create(env, d);
1161                 if (IS_ERR(th))
1162                         GOTO(out, rc = PTR_ERR(th));
1163
1164                 lgi->lgi_attr.la_valid = LA_MODE;
1165                 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1166                 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1167
1168                 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1169                                        &lgi->lgi_dof, th);
1170                 if (rc)
1171                         GOTO(out_trans, rc);
1172
1173                 rc = dt_trans_start_local(env, d, th);
1174                 if (rc)
1175                         GOTO(out_trans, rc);
1176
1177                 dt_write_lock(env, o, 0);
1178                 if (!dt_object_exists(o))
1179                         rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1180                                        &lgi->lgi_dof, th);
1181                 dt_write_unlock(env, o);
1182 out_trans:
1183                 dt_trans_stop(env, d, th);
1184                 if (rc)
1185                         GOTO(out, rc);
1186         }
1187
1188         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1189         if (rc)
1190                 GOTO(out, rc);
1191
1192         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1193                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1194                        o->do_lu.lo_dev->ld_obd->obd_name,
1195                        lgi->lgi_attr.la_mode);
1196                 GOTO(out, rc = -ENOENT);
1197         }
1198
1199         CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1200                (int)lgi->lgi_attr.la_size, size);
1201
1202         /* return just number of llogs */
1203         if (idarray == NULL) {
1204                 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1205                 GOTO(out, rc);
1206         }
1207
1208         /* read for new ost index or for empty file */
1209         memset(idarray, 0, size);
1210         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1211                 GOTO(out, rc = 0);
1212         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1213                 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1214
1215         lgi->lgi_buf.lb_buf = idarray;
1216         lgi->lgi_buf.lb_len = size;
1217         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1218         if (rc) {
1219                 CERROR("%s: error reading CATALOGS: rc = %d\n",
1220                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
1221                 GOTO(out, rc);
1222         }
1223
1224         EXIT;
1225 out:
1226         lu_object_put(env, &o->do_lu);
1227         RETURN(rc);
1228 }
1229 EXPORT_SYMBOL(llog_osd_get_cat_list);
1230
1231 /* writes the cat list */
1232 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1233                           int idx, int count, struct llog_catid *idarray)
1234 {
1235         struct llog_thread_info *lgi = llog_info(env);
1236         struct dt_object        *o = NULL;
1237         struct thandle          *th;
1238         int                      rc, size;
1239
1240         if (!count)
1241                 RETURN(0);
1242
1243         LASSERT(d);
1244
1245         size = sizeof(*idarray) * count;
1246         lgi->lgi_off = idx * sizeof(*idarray);
1247
1248         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1249         o = dt_locate(env, d, &lgi->lgi_fid);
1250         if (IS_ERR(o))
1251                 RETURN(PTR_ERR(o));
1252
1253         if (!dt_object_exists(o))
1254                 GOTO(out, rc = -ENOENT);
1255
1256         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1257         if (rc)
1258                 GOTO(out, rc);
1259
1260         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1261                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1262                        o->do_lu.lo_dev->ld_obd->obd_name,
1263                        lgi->lgi_attr.la_mode);
1264                 GOTO(out, rc = -ENOENT);
1265         }
1266
1267         th = dt_trans_create(env, d);
1268         if (IS_ERR(th))
1269                 GOTO(out, rc = PTR_ERR(th));
1270
1271         rc = dt_declare_record_write(env, o, size, lgi->lgi_off, th);
1272         if (rc)
1273                 GOTO(out, rc);
1274
1275         rc = dt_trans_start_local(env, d, th);
1276         if (rc)
1277                 GOTO(out_trans, rc);
1278
1279         lgi->lgi_buf.lb_buf = idarray;
1280         lgi->lgi_buf.lb_len = size;
1281         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1282         if (rc)
1283                 CDEBUG(D_INODE, "error writeing CATALOGS: rc = %d\n", rc);
1284 out_trans:
1285         dt_trans_stop(env, d, th);
1286 out:
1287         lu_object_put(env, &o->do_lu);
1288         RETURN(rc);
1289 }
1290 EXPORT_SYMBOL(llog_osd_put_cat_list);