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, 2013, 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         lgi->lgi_attr.la_valid = LA_SIZE;
 152         lgi->lgi_attr.la_size = *off;
 153
 154         if (!buf) {
 155                 lgi->lgi_buf.lb_len = buflen;
 156                 lgi->lgi_buf.lb_buf = rec;
 157                 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 158                 if (rc)
 159                         CERROR("%s: error writing log record: rc = %d\n",
 160                                o->do_lu.lo_dev->ld_obd->obd_name, rc);
 161                 GOTO(out, rc);
 162         }
 163
 164         /* the buf case */
 165         /* protect the following 3 writes from concurrent read */
 166         dt_write_lock(env, o, 0);
 167         rec->lrh_len = sizeof(*rec) + buflen + sizeof(lgi->lgi_tail);
 168         lgi->lgi_buf.lb_len = sizeof(*rec);
 169         lgi->lgi_buf.lb_buf = rec;
 170         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 171         if (rc) {
 172                 CERROR("%s: error writing log hdr: rc = %d\n",
 173                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 174                 GOTO(out_unlock, rc);
 175         }
 176
 177         lgi->lgi_buf.lb_len = buflen;
 178         lgi->lgi_buf.lb_buf = buf;
 179         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 180         if (rc) {
 181                 CERROR("%s: error writing log buffer: rc = %d\n",
 182                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
 183                 GOTO(out_unlock, rc);
 184         }
 185
 186         lgi->lgi_tail.lrt_len = rec->lrh_len;
 187         lgi->lgi_tail.lrt_index = rec->lrh_index;
 188         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 189         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 190         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 191         if (rc)
 192                 CERROR("%s: error writing log tail: rc = %d\n",
 193                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 194
 195 out_unlock:
 196         dt_write_unlock(env, o);
 197
 198 out:
 199         /* cleanup the content written above */
 200         if (rc) {
 201                 dt_punch(env, o, lgi->lgi_attr.la_size, OBD_OBJECT_EOF, th,
 202                          BYPASS_CAPA);
 203                 dt_attr_set(env, o, &lgi->lgi_attr, th, BYPASS_CAPA);
 204         }
 205
 206         RETURN(rc);
 207 }
 208
 209 static int llog_osd_read_header(const struct lu_env *env,
 210                                 struct llog_handle *handle)
 211 {
 212         struct llog_rec_hdr     *llh_hdr;
 213         struct dt_object        *o;
 214         struct llog_thread_info *lgi;
 215         int                      rc;
 216
 217         ENTRY;
 218
 219         LASSERT(sizeof(*handle->lgh_hdr) == LLOG_CHUNK_SIZE);
 220
 221         o = handle->lgh_obj;
 222         LASSERT(o);
 223
 224         lgi = llog_info(env);
 225
 226         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 227         if (rc)
 228                 RETURN(rc);
 229
 230         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 231
 232         if (lgi->lgi_attr.la_size == 0) {
 233                 CDEBUG(D_HA, "not reading header from 0-byte log\n");
 234                 RETURN(LLOG_EEMPTY);
 235         }
 236
 237         lgi->lgi_off = 0;
 238         lgi->lgi_buf.lb_buf = handle->lgh_hdr;
 239         lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE;
 240
 241         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
 242         if (rc) {
 243                 CERROR("%s: error reading log header from "DFID": rc = %d\n",
 244                        o->do_lu.lo_dev->ld_obd->obd_name,
 245                        PFID(lu_object_fid(&o->do_lu)), rc);
 246                 RETURN(rc);
 247         }
 248
 249         llh_hdr = &handle->lgh_hdr->llh_hdr;
 250         if (LLOG_REC_HDR_NEEDS_SWABBING(llh_hdr))
 251                 lustre_swab_llog_hdr(handle->lgh_hdr);
 252
 253         if (llh_hdr->lrh_type != LLOG_HDR_MAGIC) {
 254                 CERROR("%s: bad log %s "DFID" header magic: %#x "
 255                        "(expected %#x)\n", o->do_lu.lo_dev->ld_obd->obd_name,
 256                        handle->lgh_name ? handle->lgh_name : "",
 257                        PFID(lu_object_fid(&o->do_lu)),
 258                        llh_hdr->lrh_type, LLOG_HDR_MAGIC);
 259                 RETURN(-EIO);
 260         } else if (llh_hdr->lrh_len != LLOG_CHUNK_SIZE) {
 261                 CERROR("%s: incorrectly sized log %s "DFID" header: "
 262                        "%#x (expected %#x)\n"
 263                        "you may need to re-run lconf --write_conf.\n",
 264                        o->do_lu.lo_dev->ld_obd->obd_name,
 265                        handle->lgh_name ? handle->lgh_name : "",
 266                        PFID(lu_object_fid(&o->do_lu)),
 267                        llh_hdr->lrh_len, LLOG_CHUNK_SIZE);
 268                 RETURN(-EIO);
 269         }
 270
 271         handle->lgh_last_idx = handle->lgh_hdr->llh_tail.lrt_index;
 272
 273         RETURN(0);
 274 }
 275
 276 static int llog_osd_declare_write_rec(const struct lu_env *env,
 277                                       struct llog_handle *loghandle,
 278                                       struct llog_rec_hdr *rec,
 279                                       int idx, struct thandle *th)
 280 {
 281         struct llog_thread_info *lgi = llog_info(env);
 282         struct dt_object        *o;
 283         int                      rc;
 284
 285         ENTRY;
 286
 287         LASSERT(env);
 288         LASSERT(th);
 289         LASSERT(loghandle);
 290
 291         o = loghandle->lgh_obj;
 292         LASSERT(o);
 293
 294         /* each time we update header */
 295         rc = dt_declare_record_write(env, o, sizeof(struct llog_log_hdr), 0,
 296                                      th);
 297         if (rc || idx == 0) /* if error or just header */
 298                 RETURN(rc);
 299
 300         if (dt_object_exists(o)) {
 301                 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 302                 lgi->lgi_off = lgi->lgi_attr.la_size;
 303                 LASSERT(ergo(rc == 0, lgi->lgi_attr.la_valid & LA_SIZE));
 304                 if (rc)
 305                         RETURN(rc);
 306
 307                 rc = dt_declare_punch(env, o, lgi->lgi_off, OBD_OBJECT_EOF, th);
 308                 if (rc)
 309                         RETURN(rc);
 310         } else {
 311                 lgi->lgi_off = 0;
 312         }
 313
 314         /* XXX: implement declared window or multi-chunks approach */
 315         rc = dt_declare_record_write(env, o, 32 * 1024, lgi->lgi_off, th);
 316
 317         RETURN(rc);
 318 }
 319
 320 /* returns negative in on error; 0 if success && reccookie == 0; 1 otherwise */
 321 /* appends if idx == -1, otherwise overwrites record idx. */
 322 static int llog_osd_write_rec(const struct lu_env *env,
 323                               struct llog_handle *loghandle,
 324                               struct llog_rec_hdr *rec,
 325                               struct llog_cookie *reccookie, int cookiecount,
 326                               void *buf, int idx, struct thandle *th)
 327 {
 328         struct llog_thread_info *lgi = llog_info(env);
 329         struct llog_log_hdr     *llh;
 330         int                      reclen = rec->lrh_len;
 331         int                      index, rc, old_tail_idx;
 332         struct llog_rec_tail    *lrt;
 333         struct dt_object        *o;
 334         size_t                   left;
 335
 336         ENTRY;
 337
 338         LASSERT(env);
 339         llh = loghandle->lgh_hdr;
 340         LASSERT(llh);
 341         o = loghandle->lgh_obj;
 342         LASSERT(o);
 343         LASSERT(th);
 344
 345         CDEBUG(D_OTHER, "new record %x to "DFID"\n",
 346                rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
 347
 348         /* record length should not bigger than LLOG_CHUNK_SIZE */
 349         if (buf)
 350                 rc = (reclen > LLOG_CHUNK_SIZE - sizeof(struct llog_rec_hdr) -
 351                       sizeof(struct llog_rec_tail)) ? -E2BIG : 0;
 352         else
 353                 rc = (reclen > LLOG_CHUNK_SIZE) ? -E2BIG : 0;
 354         if (rc)
 355                 RETURN(rc);
 356
 357         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 358         if (rc)
 359                 RETURN(rc);
 360
 361         if (buf)
 362                 /* write_blob adds header and tail to lrh_len. */
 363                 reclen = sizeof(*rec) + rec->lrh_len +
 364                          sizeof(struct llog_rec_tail);
 365
 366         if (idx != -1) {
 367                 /* no header: only allowed to insert record 1 */
 368                 if (idx != 1 && lgi->lgi_attr.la_size == 0)
 369                         LBUG();
 370
 371                 if (idx && llh->llh_size && llh->llh_size != rec->lrh_len)
 372                         RETURN(-EINVAL);
 373
 374                 if (!ext2_test_bit(idx, llh->llh_bitmap))
 375                         CERROR("%s: modify unset record %u\n",
 376                                o->do_lu.lo_dev->ld_obd->obd_name, idx);
 377                 if (idx != rec->lrh_index)
 378                         CERROR("%s: index mismatch %d %u\n",
 379                                o->do_lu.lo_dev->ld_obd->obd_name, idx,
 380                                rec->lrh_index);
 381
 382                 lgi->lgi_off = 0;
 383                 rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 384                                          &lgi->lgi_off, th);
 385                 /* we are done if we only write the header or on error */
 386                 if (rc || idx == 0)
 387                         RETURN(rc);
 388
 389                 if (buf) {
 390                         /* We assume that caller has set lgh_cur_* */
 391                         lgi->lgi_off = loghandle->lgh_cur_offset;
 392                         CDEBUG(D_OTHER,
 393                                "modify record "DOSTID": idx:%d/%u/%d, len:%u "
 394                                "offset %llu\n",
 395                                POSTID(&loghandle->lgh_id.lgl_oi), idx,
 396                                rec->lrh_index,
 397                                loghandle->lgh_cur_idx, rec->lrh_len,
 398                                (long long)(lgi->lgi_off - sizeof(*llh)));
 399                         if (rec->lrh_index != loghandle->lgh_cur_idx) {
 400                                 CERROR("%s: modify idx mismatch %u/%d\n",
 401                                        o->do_lu.lo_dev->ld_obd->obd_name, idx,
 402                                        loghandle->lgh_cur_idx);
 403                                 RETURN(-EFAULT);
 404                         }
 405                 } else {
 406                         /* Assumes constant lrh_len */
 407                         lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
 408                 }
 409
 410                 rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 411                 if (rc == 0 && reccookie) {
 412                         reccookie->lgc_lgl = loghandle->lgh_id;
 413                         reccookie->lgc_index = idx;
 414                         rc = 1;
 415                 }
 416                 RETURN(rc);
 417         }
 418
 419         /* Make sure that records don't cross a chunk boundary, so we can
 420          * process them page-at-a-time if needed.  If it will cross a chunk
 421          * boundary, write in a fake (but referenced) entry to pad the chunk.
 422          *
 423          * We know that llog_current_log() will return a loghandle that is
 424          * big enough to hold reclen, so all we care about is padding here.
 425          */
 426         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 427         lgi->lgi_off = lgi->lgi_attr.la_size;
 428         left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
 429         /* NOTE: padding is a record, but no bit is set */
 430         if (left != 0 && left != reclen &&
 431             left < (reclen + LLOG_MIN_REC_SIZE)) {
 432                 index = loghandle->lgh_last_idx + 1;
 433                 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
 434                 if (rc)
 435                         RETURN(rc);
 436                 loghandle->lgh_last_idx++; /*for pad rec*/
 437         }
 438         /* if it's the last idx in log file, then return -ENOSPC */
 439         if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
 440                 RETURN(-ENOSPC);
 441
 442         loghandle->lgh_last_idx++;
 443         index = loghandle->lgh_last_idx;
 444         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 445         rec->lrh_index = index;
 446         if (buf == NULL) {
 447                 lrt = (struct llog_rec_tail *)((char *)rec + rec->lrh_len -
 448                                                sizeof(*lrt));
 449                 lrt->lrt_len = rec->lrh_len;
 450                 lrt->lrt_index = rec->lrh_index;
 451         }
 452         /* The caller should make sure only 1 process access the lgh_last_idx,
 453          * Otherwise it might hit the assert.*/
 454         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 455         spin_lock(&loghandle->lgh_hdr_lock);
 456         if (ext2_set_bit(index, llh->llh_bitmap)) {
 457                 CERROR("%s: index %u already set in log bitmap\n",
 458                        o->do_lu.lo_dev->ld_obd->obd_name, index);
 459                 spin_unlock(&loghandle->lgh_hdr_lock);
 460                 LBUG(); /* should never happen */
 461         }
 462         llh->llh_count++;
 463         spin_unlock(&loghandle->lgh_hdr_lock);
 464         old_tail_idx = llh->llh_tail.lrt_index;
 465         llh->llh_tail.lrt_index = index;
 466
 467         lgi->lgi_off = 0;
 468         rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL, &lgi->lgi_off,
 469                                  th);
 470         if (rc)
 471                 GOTO(out, rc);
 472
 473         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 474         if (rc)
 475                 GOTO(out, rc);
 476
 477         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 478         lgi->lgi_off = lgi->lgi_attr.la_size;
 479
 480         rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 481
 482 out:
 483         /* cleanup llog for error case */
 484         if (rc) {
 485                 spin_lock(&loghandle->lgh_hdr_lock);
 486                 ext2_clear_bit(index, llh->llh_bitmap);
 487                 llh->llh_count--;
 488                 spin_unlock(&loghandle->lgh_hdr_lock);
 489
 490                 /* restore the header */
 491                 loghandle->lgh_last_idx--;
 492                 llh->llh_tail.lrt_index = old_tail_idx;
 493                 lgi->lgi_off = 0;
 494                 llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 495                                     &lgi->lgi_off, th);
 496         }
 497
 498         CDEBUG(D_RPCTRACE, "added record "DOSTID": idx: %u, %u\n",
 499                POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len);
 500         if (rc == 0 && reccookie) {
 501                 reccookie->lgc_lgl = loghandle->lgh_id;
 502                 reccookie->lgc_index = index;
 503                 if ((rec->lrh_type == MDS_UNLINK_REC) ||
 504                     (rec->lrh_type == MDS_SETATTR64_REC))
 505                         reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 506                 else if (rec->lrh_type == OST_SZ_REC)
 507                         reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
 508                 else
 509                         reccookie->lgc_subsys = -1;
 510                 rc = 1;
 511         }
 512         RETURN(rc);
 513 }
 514
 515 /* We can skip reading at least as many log blocks as the number of
 516  * minimum sized log records we are skipping.  If it turns out
 517  * that we are not far enough along the log (because the
 518  * actual records are larger than minimum size) we just skip
 519  * some more records.
 520  */
 521 static void llog_skip_over(__u64 *off, int curr, int goal)
 522 {
 523         if (goal <= curr)
 524                 return;
 525         *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
 526                 ~(LLOG_CHUNK_SIZE - 1);
 527 }
 528
 529 /* sets:
 530  *  - cur_offset to the furthest point read in the log file
 531  *  - cur_idx to the log index preceeding cur_offset
 532  * returns -EIO/-EINVAL on error
 533  */
 534 static int llog_osd_next_block(const struct lu_env *env,
 535                                struct llog_handle *loghandle, int *cur_idx,
 536                                int next_idx, __u64 *cur_offset, void *buf,
 537                                int len)
 538 {
 539         struct llog_thread_info *lgi = llog_info(env);
 540         struct dt_object        *o;
 541         struct dt_device        *dt;
 542         int                      rc;
 543
 544         ENTRY;
 545
 546         LASSERT(env);
 547         LASSERT(lgi);
 548
 549         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 550                 RETURN(-EINVAL);
 551
 552         CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
 553                next_idx, *cur_idx, *cur_offset);
 554
 555         LASSERT(loghandle);
 556         LASSERT(loghandle->lgh_ctxt);
 557
 558         o = loghandle->lgh_obj;
 559         LASSERT(o);
 560         LASSERT(dt_object_exists(o));
 561         dt = lu2dt_dev(o->do_lu.lo_dev);
 562         LASSERT(dt);
 563
 564         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 565         if (rc)
 566                 GOTO(out, rc);
 567
 568         while (*cur_offset < lgi->lgi_attr.la_size) {
 569                 struct llog_rec_hdr     *rec, *last_rec;
 570                 struct llog_rec_tail    *tail;
 571
 572                 llog_skip_over(cur_offset, *cur_idx, next_idx);
 573
 574                 /* read up to next LLOG_CHUNK_SIZE block */
 575                 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
 576                                       (*cur_offset & (LLOG_CHUNK_SIZE - 1));
 577                 lgi->lgi_buf.lb_buf = buf;
 578
 579                 /* Note: read lock is not needed around la_size get above at
 580                  * the time of dt_attr_get(). There are only two cases that
 581                  * matter. Either la_size == cur_offset, in which case the
 582                  * entire read is skipped, or la_size > cur_offset and the loop
 583                  * is entered and this thread is blocked at dt_read_lock()
 584                  * until the write is completed. When the write completes, then
 585                  * the dt_read() will be done with the full length, and will
 586                  * get the full data.
 587                  */
 588                 dt_read_lock(env, o, 0);
 589                 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
 590                 dt_read_unlock(env, o);
 591                 if (rc < 0) {
 592                         CERROR("%s: can't read llog block from log "DFID
 593                                " offset "LPU64": rc = %d\n",
 594                                o->do_lu.lo_dev->ld_obd->obd_name,
 595                                PFID(lu_object_fid(&o->do_lu)), *cur_offset,
 596                                rc);
 597                         GOTO(out, rc);
 598                 }
 599
 600                 if (rc < len) {
 601                         /* signal the end of the valid buffer to
 602                          * llog_process */
 603                         memset(buf + rc, 0, len - rc);
 604                 }
 605
 606                 if (rc == 0) /* end of file, nothing to do */
 607                         GOTO(out, rc);
 608
 609                 if (rc < sizeof(*tail)) {
 610                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 611                                "offset "LPU64"\n",
 612                                o->do_lu.lo_dev->ld_obd->obd_name,
 613                                POSTID(&loghandle->lgh_id.lgl_oi),
 614                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 615                         GOTO(out, rc = -EINVAL);
 616                 }
 617
 618                 rec = buf;
 619                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 620                         lustre_swab_llog_rec(rec);
 621
 622                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 623                                                 sizeof(struct llog_rec_tail));
 624                 /* get the last record in block */
 625                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 626                                                    le32_to_cpu(tail->lrt_len));
 627
 628                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 629                         lustre_swab_llog_rec(last_rec);
 630                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 631
 632                 *cur_idx = tail->lrt_index;
 633
 634                 /* this shouldn't happen */
 635                 if (tail->lrt_index == 0) {
 636                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 637                                "offset "LPU64"\n",
 638                                o->do_lu.lo_dev->ld_obd->obd_name,
 639                                POSTID(&loghandle->lgh_id.lgl_oi),
 640                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 641                         GOTO(out, rc = -EINVAL);
 642                 }
 643                 if (tail->lrt_index < next_idx)
 644                         continue;
 645
 646                 /* sanity check that the start of the new buffer is no farther
 647                  * than the record that we wanted.  This shouldn't happen. */
 648                 if (rec->lrh_index > next_idx) {
 649                         CERROR("%s: missed desired record? %u > %u\n",
 650                                o->do_lu.lo_dev->ld_obd->obd_name,
 651                                rec->lrh_index, next_idx);
 652                         GOTO(out, rc = -ENOENT);
 653                 }
 654                 GOTO(out, rc = 0);
 655         }
 656         GOTO(out, rc = -EIO);
 657 out:
 658         return rc;
 659 }
 660
 661 static int llog_osd_prev_block(const struct lu_env *env,
 662                                struct llog_handle *loghandle,
 663                                int prev_idx, void *buf, int len)
 664 {
 665         struct llog_thread_info *lgi = llog_info(env);
 666         struct dt_object        *o;
 667         struct dt_device        *dt;
 668         loff_t                   cur_offset;
 669         int                      rc;
 670
 671         ENTRY;
 672
 673         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 674                 RETURN(-EINVAL);
 675
 676         CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
 677
 678         LASSERT(loghandle);
 679         LASSERT(loghandle->lgh_ctxt);
 680
 681         o = loghandle->lgh_obj;
 682         LASSERT(o);
 683         LASSERT(dt_object_exists(o));
 684         dt = lu2dt_dev(o->do_lu.lo_dev);
 685         LASSERT(dt);
 686
 687         cur_offset = LLOG_CHUNK_SIZE;
 688         llog_skip_over(&cur_offset, 0, prev_idx);
 689
 690         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 691         if (rc)
 692                 GOTO(out, rc);
 693
 694         while (cur_offset < lgi->lgi_attr.la_size) {
 695                 struct llog_rec_hdr     *rec, *last_rec;
 696                 struct llog_rec_tail    *tail;
 697
 698                 lgi->lgi_buf.lb_len = len;
 699                 lgi->lgi_buf.lb_buf = buf;
 700                 /* It is OK to have locking around dt_read() only, see
 701                  * comment in llog_osd_next_block for details
 702                  */
 703                 dt_read_lock(env, o, 0);
 704                 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
 705                 dt_read_unlock(env, o);
 706                 if (rc < 0) {
 707                         CERROR("%s: can't read llog block from log "DFID
 708                                " offset "LPU64": rc = %d\n",
 709                                o->do_lu.lo_dev->ld_obd->obd_name,
 710                                PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
 711                         GOTO(out, rc);
 712                 }
 713
 714                 if (rc == 0) /* end of file, nothing to do */
 715                         GOTO(out, rc);
 716
 717                 if (rc < sizeof(*tail)) {
 718                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 719                                "offset "LPU64"\n",
 720                                o->do_lu.lo_dev->ld_obd->obd_name,
 721                                POSTID(&loghandle->lgh_id.lgl_oi),
 722                                loghandle->lgh_id.lgl_ogen, cur_offset);
 723                         GOTO(out, rc = -EINVAL);
 724                 }
 725
 726                 rec = buf;
 727                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 728                         lustre_swab_llog_rec(rec);
 729
 730                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 731                                                 sizeof(struct llog_rec_tail));
 732                 /* get the last record in block */
 733                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 734                                                    le32_to_cpu(tail->lrt_len));
 735
 736                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 737                         lustre_swab_llog_rec(last_rec);
 738                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 739
 740                 /* this shouldn't happen */
 741                 if (tail->lrt_index == 0) {
 742                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 743                                "offset "LPU64"\n",
 744                                o->do_lu.lo_dev->ld_obd->obd_name,
 745                                POSTID(&loghandle->lgh_id.lgl_oi),
 746                                loghandle->lgh_id.lgl_ogen, cur_offset);
 747                         GOTO(out, rc = -EINVAL);
 748                 }
 749                 if (tail->lrt_index < prev_idx)
 750                         continue;
 751
 752                 /* sanity check that the start of the new buffer is no farther
 753                  * than the record that we wanted.  This shouldn't happen. */
 754                 if (rec->lrh_index > prev_idx) {
 755                         CERROR("%s: missed desired record? %u > %u\n",
 756                                o->do_lu.lo_dev->ld_obd->obd_name,
 757                                rec->lrh_index, prev_idx);
 758                         GOTO(out, rc = -ENOENT);
 759                 }
 760                 GOTO(out, rc = 0);
 761         }
 762         GOTO(out, rc = -EIO);
 763 out:
 764         return rc;
 765 }
 766
 767 struct dt_object *llog_osd_dir_get(const struct lu_env *env,
 768                                    struct llog_ctxt *ctxt)
 769 {
 770         struct dt_device        *dt;
 771         struct dt_thread_info   *dti = dt_info(env);
 772         struct dt_object        *dir;
 773         int                      rc;
 774
 775         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 776         if (ctxt->loc_dir == NULL) {
 777                 rc = dt_root_get(env, dt, &dti->dti_fid);
 778                 if (rc)
 779                         return ERR_PTR(rc);
 780                 dir = dt_locate(env, dt, &dti->dti_fid);
 781         } else {
 782                 lu_object_get(&ctxt->loc_dir->do_lu);
 783                 dir = ctxt->loc_dir;
 784         }
 785
 786         return dir;
 787 }
 788
 789 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
 790                          struct llog_logid *logid, char *name,
 791                          enum llog_open_param open_param)
 792 {
 793         struct llog_thread_info         *lgi = llog_info(env);
 794         struct llog_ctxt                *ctxt = handle->lgh_ctxt;
 795         struct dt_object                *o;
 796         struct dt_device                *dt;
 797         struct ls_device                *ls;
 798         struct local_oid_storage        *los;
 799         int                              rc = 0;
 800
 801         ENTRY;
 802
 803         LASSERT(env);
 804         LASSERT(ctxt);
 805         LASSERT(ctxt->loc_exp);
 806         LASSERT(ctxt->loc_exp->exp_obd);
 807         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 808         LASSERT(dt);
 809
 810         ls = ls_device_get(dt);
 811         if (IS_ERR(ls))
 812                 RETURN(PTR_ERR(ls));
 813
 814         mutex_lock(&ls->ls_los_mutex);
 815         los = dt_los_find(ls, name != NULL ? FID_SEQ_LLOG_NAME : FID_SEQ_LLOG);
 816         mutex_unlock(&ls->ls_los_mutex);
 817         LASSERT(los);
 818         ls_device_put(env, ls);
 819
 820         LASSERT(handle);
 821
 822         if (logid != NULL) {
 823                 logid_to_fid(logid, &lgi->lgi_fid);
 824         } else if (name) {
 825                 struct dt_object *llog_dir;
 826
 827                 llog_dir = llog_osd_dir_get(env, ctxt);
 828                 if (IS_ERR(llog_dir))
 829                         GOTO(out, rc = PTR_ERR(llog_dir));
 830                 dt_read_lock(env, llog_dir, 0);
 831                 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
 832                 dt_read_unlock(env, llog_dir);
 833                 lu_object_put(env, &llog_dir->do_lu);
 834                 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
 835                         /* generate fid for new llog */
 836                         rc = local_object_fid_generate(env, los,
 837                                                        &lgi->lgi_fid);
 838                 }
 839                 if (rc < 0)
 840                         GOTO(out, rc);
 841                 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
 842                 if (handle->lgh_name)
 843                         strcpy(handle->lgh_name, name);
 844                 else
 845                         GOTO(out, rc = -ENOMEM);
 846         } else {
 847                 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
 848                 /* generate fid for new llog */
 849                 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
 850                 if (rc < 0)
 851                         GOTO(out, rc);
 852         }
 853
 854         o = ls_locate(env, ls, &lgi->lgi_fid);
 855         if (IS_ERR(o))
 856                 GOTO(out_name, rc = PTR_ERR(o));
 857
 858         /* No new llog is expected but doesn't exist */
 859         if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
 860                 GOTO(out_put, rc = -ENOENT);
 861
 862         fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
 863         handle->lgh_obj = o;
 864         handle->private_data = los;
 865         LASSERT(handle->lgh_ctxt);
 866
 867         RETURN(rc);
 868
 869 out_put:
 870         lu_object_put(env, &o->do_lu);
 871 out_name:
 872         if (handle->lgh_name != NULL)
 873                 OBD_FREE(handle->lgh_name, strlen(name) + 1);
 874 out:
 875         dt_los_put(los);
 876         RETURN(rc);
 877 }
 878
 879 static int llog_osd_exist(struct llog_handle *handle)
 880 {
 881         LASSERT(handle->lgh_obj);
 882         return (dt_object_exists(handle->lgh_obj) &&
 883                 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
 884 }
 885
 886 static int llog_osd_declare_create(const struct lu_env *env,
 887                                    struct llog_handle *res, struct thandle *th)
 888 {
 889         struct llog_thread_info         *lgi = llog_info(env);
 890         struct local_oid_storage        *los;
 891         struct dt_object                *o;
 892         int                              rc;
 893
 894         ENTRY;
 895
 896         LASSERT(res->lgh_obj);
 897         LASSERT(th);
 898
 899         /* object can be created by another thread */
 900         o = res->lgh_obj;
 901         if (dt_object_exists(o))
 902                 RETURN(0);
 903
 904         los = res->private_data;
 905         LASSERT(los);
 906
 907         rc = llog_osd_declare_new_object(env, los, o, th);
 908         if (rc)
 909                 RETURN(rc);
 910
 911         rc = dt_declare_record_write(env, o, LLOG_CHUNK_SIZE, 0, th);
 912         if (rc)
 913                 RETURN(rc);
 914
 915         if (res->lgh_name) {
 916                 struct dt_object *llog_dir;
 917
 918                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 919                 if (IS_ERR(llog_dir))
 920                         RETURN(PTR_ERR(llog_dir));
 921                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 922                 rc = dt_declare_insert(env, llog_dir,
 923                                        (struct dt_rec *)&lgi->lgi_fid,
 924                                        (struct dt_key *)res->lgh_name, th);
 925                 lu_object_put(env, &llog_dir->do_lu);
 926                 if (rc)
 927                         CERROR("%s: can't declare named llog %s: rc = %d\n",
 928                                o->do_lu.lo_dev->ld_obd->obd_name,
 929                                res->lgh_name, rc);
 930         }
 931         RETURN(rc);
 932 }
 933
 934 /* This is a callback from the llog_* functions.
 935  * Assumes caller has already pushed us into the kernel context. */
 936 static int llog_osd_create(const struct lu_env *env, struct llog_handle *res,
 937                            struct thandle *th)
 938 {
 939         struct llog_thread_info *lgi = llog_info(env);
 940         struct local_oid_storage *los;
 941         struct dt_object        *o;
 942         int                      rc = 0;
 943
 944         ENTRY;
 945
 946         LASSERT(env);
 947         o = res->lgh_obj;
 948         LASSERT(o);
 949
 950         /* llog can be already created */
 951         if (dt_object_exists(o))
 952                 RETURN(-EEXIST);
 953
 954         los = res->private_data;
 955         LASSERT(los);
 956
 957         dt_write_lock(env, o, 0);
 958         if (!dt_object_exists(o))
 959                 rc = llog_osd_create_new_object(env, los, o, th);
 960         else
 961                 rc = -EEXIST;
 962
 963         dt_write_unlock(env, o);
 964         if (rc)
 965                 RETURN(rc);
 966
 967         if (res->lgh_name) {
 968                 struct dt_object *llog_dir;
 969
 970                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 971                 if (IS_ERR(llog_dir))
 972                         RETURN(PTR_ERR(llog_dir));
 973
 974                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 975                 dt_read_lock(env, llog_dir, 0);
 976                 rc = dt_insert(env, llog_dir,
 977                                (struct dt_rec *)&lgi->lgi_fid,
 978                                (struct dt_key *)res->lgh_name,
 979                                th, BYPASS_CAPA, 1);
 980                 dt_read_unlock(env, llog_dir);
 981                 lu_object_put(env, &llog_dir->do_lu);
 982                 if (rc)
 983                         CERROR("%s: can't create named llog %s: rc = %d\n",
 984                                o->do_lu.lo_dev->ld_obd->obd_name,
 985                                res->lgh_name, rc);
 986         }
 987         RETURN(rc);
 988 }
 989
 990 static int llog_osd_close(const struct lu_env *env, struct llog_handle *handle)
 991 {
 992         struct local_oid_storage        *los;
 993         int                              rc = 0;
 994
 995         ENTRY;
 996
 997         LASSERT(handle->lgh_obj);
 998
 999         lu_object_put(env, &handle->lgh_obj->do_lu);
1000
1001         los = handle->private_data;
1002         LASSERT(los);
1003         dt_los_put(los);
1004
1005         if (handle->lgh_name)
1006                 OBD_FREE(handle->lgh_name, strlen(handle->lgh_name) + 1);
1007
1008         RETURN(rc);
1009 }
1010
1011 static int llog_osd_destroy(const struct lu_env *env,
1012                             struct llog_handle *loghandle)
1013 {
1014         struct llog_ctxt        *ctxt;
1015         struct dt_object        *o, *llog_dir = NULL;
1016         struct dt_device        *d;
1017         struct thandle          *th;
1018         char                    *name = NULL;
1019         int                      rc;
1020
1021         ENTRY;
1022
1023         ctxt = loghandle->lgh_ctxt;
1024         LASSERT(ctxt);
1025
1026         o = loghandle->lgh_obj;
1027         LASSERT(o);
1028
1029         d = lu2dt_dev(o->do_lu.lo_dev);
1030         LASSERT(d);
1031         LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
1032
1033         th = dt_trans_create(env, d);
1034         if (IS_ERR(th))
1035                 RETURN(PTR_ERR(th));
1036
1037         if (loghandle->lgh_name) {
1038                 llog_dir = llog_osd_dir_get(env, ctxt);
1039                 if (IS_ERR(llog_dir))
1040                         GOTO(out_trans, rc = PTR_ERR(llog_dir));
1041
1042                 name = loghandle->lgh_name;
1043                 rc = dt_declare_delete(env, llog_dir,
1044                                        (struct dt_key *)name, th);
1045                 if (rc)
1046                         GOTO(out_trans, rc);
1047         }
1048
1049         dt_declare_ref_del(env, o, th);
1050
1051         rc = dt_declare_destroy(env, o, th);
1052         if (rc)
1053                 GOTO(out_trans, rc);
1054
1055         rc = dt_trans_start_local(env, d, th);
1056         if (rc)
1057                 GOTO(out_trans, rc);
1058
1059         dt_write_lock(env, o, 0);
1060         if (dt_object_exists(o)) {
1061                 if (name) {
1062                         dt_read_lock(env, llog_dir, 0);
1063                         rc = dt_delete(env, llog_dir,
1064                                        (struct dt_key *) name,
1065                                        th, BYPASS_CAPA);
1066                         dt_read_unlock(env, llog_dir);
1067                         if (rc) {
1068                                 CERROR("%s: can't remove llog %s: rc = %d\n",
1069                                        o->do_lu.lo_dev->ld_obd->obd_name,
1070                                        name, rc);
1071                                 GOTO(out_unlock, rc);
1072                         }
1073                 }
1074                 dt_ref_del(env, o, th);
1075                 rc = dt_destroy(env, o, th);
1076                 if (rc)
1077                         GOTO(out_unlock, rc);
1078         }
1079 out_unlock:
1080         dt_write_unlock(env, o);
1081 out_trans:
1082         dt_trans_stop(env, d, th);
1083         if (llog_dir != NULL)
1084                 lu_object_put(env, &llog_dir->do_lu);
1085         RETURN(rc);
1086 }
1087
1088 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1089                           struct obd_llog_group *olg, int ctxt_idx,
1090                           struct obd_device *disk_obd)
1091 {
1092         struct local_oid_storage        *los;
1093         struct llog_thread_info         *lgi = llog_info(env);
1094         struct llog_ctxt                *ctxt;
1095         int                              rc = 0;
1096
1097         ENTRY;
1098
1099         LASSERT(obd);
1100         LASSERT(olg->olg_ctxts[ctxt_idx]);
1101
1102         ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1103         LASSERT(ctxt);
1104
1105         /* initialize data allowing to generate new fids,
1106          * literally we need a sequece */
1107         lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1108         lgi->lgi_fid.f_oid = 1;
1109         lgi->lgi_fid.f_ver = 0;
1110         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1111                                     &lgi->lgi_fid, &los);
1112         if (rc < 0)
1113                 return rc;
1114
1115         lgi->lgi_fid.f_seq = FID_SEQ_LLOG_NAME;
1116         lgi->lgi_fid.f_oid = 1;
1117         lgi->lgi_fid.f_ver = 0;
1118         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1119                                     &lgi->lgi_fid, &los);
1120         llog_ctxt_put(ctxt);
1121         return rc;
1122 }
1123
1124 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1125 {
1126         struct dt_device                *dt;
1127         struct ls_device                *ls;
1128         struct local_oid_storage        *los, *nlos;
1129
1130         LASSERT(ctxt->loc_exp->exp_obd);
1131         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
1132         ls = ls_device_get(dt);
1133         if (IS_ERR(ls))
1134                 RETURN(PTR_ERR(ls));
1135
1136         mutex_lock(&ls->ls_los_mutex);
1137         los = dt_los_find(ls, FID_SEQ_LLOG);
1138         nlos = dt_los_find(ls, FID_SEQ_LLOG_NAME);
1139         mutex_unlock(&ls->ls_los_mutex);
1140         if (los != NULL) {
1141                 dt_los_put(los);
1142                 local_oid_storage_fini(env, los);
1143         }
1144         if (nlos != NULL) {
1145                 dt_los_put(nlos);
1146                 local_oid_storage_fini(env, nlos);
1147         }
1148         ls_device_put(env, ls);
1149         return 0;
1150 }
1151
1152 struct llog_operations llog_osd_ops = {
1153         .lop_next_block         = llog_osd_next_block,
1154         .lop_prev_block         = llog_osd_prev_block,
1155         .lop_read_header        = llog_osd_read_header,
1156         .lop_destroy            = llog_osd_destroy,
1157         .lop_setup              = llog_osd_setup,
1158         .lop_cleanup            = llog_osd_cleanup,
1159         .lop_open               = llog_osd_open,
1160         .lop_exist              = llog_osd_exist,
1161         .lop_declare_create     = llog_osd_declare_create,
1162         .lop_create             = llog_osd_create,
1163         .lop_declare_write_rec  = llog_osd_declare_write_rec,
1164         .lop_write_rec          = llog_osd_write_rec,
1165         .lop_close              = llog_osd_close,
1166 };
1167 EXPORT_SYMBOL(llog_osd_ops);
1168
1169 /* reads the catalog list */
1170 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1171                           int idx, int count, struct llog_catid *idarray)
1172 {
1173         struct llog_thread_info *lgi = llog_info(env);
1174         struct dt_object        *o = NULL;
1175         struct thandle          *th;
1176         int                      rc, size;
1177
1178         ENTRY;
1179
1180         LASSERT(d);
1181
1182         size = sizeof(*idarray) * count;
1183         lgi->lgi_off = idx *  sizeof(*idarray);
1184
1185         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1186
1187         o = dt_locate(env, d, &lgi->lgi_fid);
1188         if (IS_ERR(o))
1189                 RETURN(PTR_ERR(o));
1190
1191         if (!dt_object_exists(o)) {
1192                 th = dt_trans_create(env, d);
1193                 if (IS_ERR(th))
1194                         GOTO(out, rc = PTR_ERR(th));
1195
1196                 lgi->lgi_attr.la_valid = LA_MODE;
1197                 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1198                 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1199
1200                 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1201                                        &lgi->lgi_dof, th);
1202                 if (rc)
1203                         GOTO(out_trans, rc);
1204
1205                 rc = dt_trans_start_local(env, d, th);
1206                 if (rc)
1207                         GOTO(out_trans, rc);
1208
1209                 dt_write_lock(env, o, 0);
1210                 if (!dt_object_exists(o))
1211                         rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1212                                        &lgi->lgi_dof, th);
1213                 dt_write_unlock(env, o);
1214 out_trans:
1215                 dt_trans_stop(env, d, th);
1216                 if (rc)
1217                         GOTO(out, rc);
1218         }
1219
1220         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1221         if (rc)
1222                 GOTO(out, rc);
1223
1224         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1225                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1226                        o->do_lu.lo_dev->ld_obd->obd_name,
1227                        lgi->lgi_attr.la_mode);
1228                 GOTO(out, rc = -ENOENT);
1229         }
1230
1231         CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1232                (int)lgi->lgi_attr.la_size, size);
1233
1234         /* return just number of llogs */
1235         if (idarray == NULL) {
1236                 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1237                 GOTO(out, rc);
1238         }
1239
1240         /* read for new ost index or for empty file */
1241         memset(idarray, 0, size);
1242         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1243                 GOTO(out, rc = 0);
1244         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1245                 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1246
1247         lgi->lgi_buf.lb_buf = idarray;
1248         lgi->lgi_buf.lb_len = size;
1249         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1250         if (rc) {
1251                 CERROR("%s: error reading CATALOGS: rc = %d\n",
1252                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
1253                 GOTO(out, rc);
1254         }
1255
1256         EXIT;
1257 out:
1258         lu_object_put(env, &o->do_lu);
1259         RETURN(rc);
1260 }
1261 EXPORT_SYMBOL(llog_osd_get_cat_list);
1262
1263 /* writes the cat list */
1264 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1265                           int idx, int count, struct llog_catid *idarray)
1266 {
1267         struct llog_thread_info *lgi = llog_info(env);
1268         struct dt_object        *o = NULL;
1269         struct thandle          *th;
1270         int                      rc, size;
1271
1272         if (!count)
1273                 RETURN(0);
1274
1275         LASSERT(d);
1276
1277         size = sizeof(*idarray) * count;
1278         lgi->lgi_off = idx * sizeof(*idarray);
1279
1280         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1281
1282         o = dt_locate(env, d, &lgi->lgi_fid);
1283         if (IS_ERR(o))
1284                 RETURN(PTR_ERR(o));
1285
1286         if (!dt_object_exists(o))
1287                 GOTO(out, rc = -ENOENT);
1288
1289         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1290         if (rc)
1291                 GOTO(out, rc);
1292
1293         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1294                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1295                        o->do_lu.lo_dev->ld_obd->obd_name,
1296                        lgi->lgi_attr.la_mode);
1297                 GOTO(out, rc = -ENOENT);
1298         }
1299
1300         th = dt_trans_create(env, d);
1301         if (IS_ERR(th))
1302                 GOTO(out, rc = PTR_ERR(th));
1303
1304         rc = dt_declare_record_write(env, o, size, lgi->lgi_off, th);
1305         if (rc)
1306                 GOTO(out, rc);
1307
1308         rc = dt_trans_start_local(env, d, th);
1309         if (rc)
1310                 GOTO(out_trans, rc);
1311
1312         lgi->lgi_buf.lb_buf = idarray;
1313         lgi->lgi_buf.lb_len = size;
1314         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1315         if (rc)
1316                 CDEBUG(D_INODE, "error writeing CATALOGS: rc = %d\n", rc);
1317 out_trans:
1318         dt_trans_stop(env, d, th);
1319 out:
1320         lu_object_put(env, &o->do_lu);
1321         RETURN(rc);
1322 }
1323 EXPORT_SYMBOL(llog_osd_put_cat_list);