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         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 "LPX64": idx:%d/%u/%d, len:%u "
 394                                "offset %llu\n",
 395                                loghandle->lgh_id.lgl_oid, idx, rec->lrh_index,
 396                                loghandle->lgh_cur_idx, rec->lrh_len,
 397                                (long long)(lgi->lgi_off - sizeof(*llh)));
 398                         if (rec->lrh_index != loghandle->lgh_cur_idx) {
 399                                 CERROR("%s: modify idx mismatch %u/%d\n",
 400                                        o->do_lu.lo_dev->ld_obd->obd_name, idx,
 401                                        loghandle->lgh_cur_idx);
 402                                 RETURN(-EFAULT);
 403                         }
 404                 } else {
 405                         /* Assumes constant lrh_len */
 406                         lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
 407                 }
 408
 409                 rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 410                 if (rc == 0 && reccookie) {
 411                         reccookie->lgc_lgl = loghandle->lgh_id;
 412                         reccookie->lgc_index = idx;
 413                         rc = 1;
 414                 }
 415                 RETURN(rc);
 416         }
 417
 418         /* Make sure that records don't cross a chunk boundary, so we can
 419          * process them page-at-a-time if needed.  If it will cross a chunk
 420          * boundary, write in a fake (but referenced) entry to pad the chunk.
 421          *
 422          * We know that llog_current_log() will return a loghandle that is
 423          * big enough to hold reclen, so all we care about is padding here.
 424          */
 425         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 426         lgi->lgi_off = lgi->lgi_attr.la_size;
 427         left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
 428         /* NOTE: padding is a record, but no bit is set */
 429         if (left != 0 && left != reclen &&
 430             left < (reclen + LLOG_MIN_REC_SIZE)) {
 431                 index = loghandle->lgh_last_idx + 1;
 432                 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
 433                 if (rc)
 434                         RETURN(rc);
 435                 loghandle->lgh_last_idx++; /*for pad rec*/
 436         }
 437         /* if it's the last idx in log file, then return -ENOSPC */
 438         if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
 439                 RETURN(-ENOSPC);
 440
 441         loghandle->lgh_last_idx++;
 442         index = loghandle->lgh_last_idx;
 443         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 444         rec->lrh_index = index;
 445         if (buf == NULL) {
 446                 lrt = (struct llog_rec_tail *)((char *)rec + rec->lrh_len -
 447                                                sizeof(*lrt));
 448                 lrt->lrt_len = rec->lrh_len;
 449                 lrt->lrt_index = rec->lrh_index;
 450         }
 451         /* The caller should make sure only 1 process access the lgh_last_idx,
 452          * Otherwise it might hit the assert.*/
 453         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 454         spin_lock(&loghandle->lgh_hdr_lock);
 455         if (ext2_set_bit(index, llh->llh_bitmap)) {
 456                 CERROR("%s: index %u already set in log bitmap\n",
 457                        o->do_lu.lo_dev->ld_obd->obd_name, index);
 458                 spin_unlock(&loghandle->lgh_hdr_lock);
 459                 LBUG(); /* should never happen */
 460         }
 461         llh->llh_count++;
 462         spin_unlock(&loghandle->lgh_hdr_lock);
 463         old_tail_idx = llh->llh_tail.lrt_index;
 464         llh->llh_tail.lrt_index = index;
 465
 466         lgi->lgi_off = 0;
 467         rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL, &lgi->lgi_off,
 468                                  th);
 469         if (rc)
 470                 GOTO(out, rc);
 471
 472         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 473         if (rc)
 474                 GOTO(out, rc);
 475
 476         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 477         lgi->lgi_off = lgi->lgi_attr.la_size;
 478
 479         rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 480
 481 out:
 482         /* cleanup llog for error case */
 483         if (rc) {
 484                 spin_lock(&loghandle->lgh_hdr_lock);
 485                 ext2_clear_bit(index, llh->llh_bitmap);
 486                 llh->llh_count--;
 487                 spin_unlock(&loghandle->lgh_hdr_lock);
 488
 489                 /* restore the header */
 490                 loghandle->lgh_last_idx--;
 491                 llh->llh_tail.lrt_index = old_tail_idx;
 492                 lgi->lgi_off = 0;
 493                 llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 494                                     &lgi->lgi_off, th);
 495         }
 496
 497         CDEBUG(D_RPCTRACE, "added record "LPX64": idx: %u, %u\n",
 498                loghandle->lgh_id.lgl_oid, index, rec->lrh_len);
 499         if (rc == 0 && reccookie) {
 500                 reccookie->lgc_lgl = loghandle->lgh_id;
 501                 reccookie->lgc_index = index;
 502                 if ((rec->lrh_type == MDS_UNLINK_REC) ||
 503                     (rec->lrh_type == MDS_SETATTR64_REC))
 504                         reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 505                 else if (rec->lrh_type == OST_SZ_REC)
 506                         reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
 507                 else
 508                         reccookie->lgc_subsys = -1;
 509                 rc = 1;
 510         }
 511         RETURN(rc);
 512 }
 513
 514 /* We can skip reading at least as many log blocks as the number of
 515  * minimum sized log records we are skipping.  If it turns out
 516  * that we are not far enough along the log (because the
 517  * actual records are larger than minimum size) we just skip
 518  * some more records.
 519  */
 520 static void llog_skip_over(__u64 *off, int curr, int goal)
 521 {
 522         if (goal <= curr)
 523                 return;
 524         *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
 525                 ~(LLOG_CHUNK_SIZE - 1);
 526 }
 527
 528 /* sets:
 529  *  - cur_offset to the furthest point read in the log file
 530  *  - cur_idx to the log index preceeding cur_offset
 531  * returns -EIO/-EINVAL on error
 532  */
 533 static int llog_osd_next_block(const struct lu_env *env,
 534                                struct llog_handle *loghandle, int *cur_idx,
 535                                int next_idx, __u64 *cur_offset, void *buf,
 536                                int len)
 537 {
 538         struct llog_thread_info *lgi = llog_info(env);
 539         struct dt_object        *o;
 540         struct dt_device        *dt;
 541         int                      rc;
 542
 543         ENTRY;
 544
 545         LASSERT(env);
 546         LASSERT(lgi);
 547
 548         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 549                 RETURN(-EINVAL);
 550
 551         CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
 552                next_idx, *cur_idx, *cur_offset);
 553
 554         LASSERT(loghandle);
 555         LASSERT(loghandle->lgh_ctxt);
 556
 557         o = loghandle->lgh_obj;
 558         LASSERT(o);
 559         LASSERT(dt_object_exists(o));
 560         dt = lu2dt_dev(o->do_lu.lo_dev);
 561         LASSERT(dt);
 562
 563         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 564         if (rc)
 565                 GOTO(out, rc);
 566
 567         while (*cur_offset < lgi->lgi_attr.la_size) {
 568                 struct llog_rec_hdr     *rec, *last_rec;
 569                 struct llog_rec_tail    *tail;
 570
 571                 llog_skip_over(cur_offset, *cur_idx, next_idx);
 572
 573                 /* read up to next LLOG_CHUNK_SIZE block */
 574                 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
 575                                       (*cur_offset & (LLOG_CHUNK_SIZE - 1));
 576                 lgi->lgi_buf.lb_buf = buf;
 577
 578                 /* Note: read lock is not needed around la_size get above at
 579                  * the time of dt_attr_get(). There are only two cases that
 580                  * matter. Either la_size == cur_offset, in which case the
 581                  * entire read is skipped, or la_size > cur_offset and the loop
 582                  * is entered and this thread is blocked at dt_read_lock()
 583                  * until the write is completed. When the write completes, then
 584                  * the dt_read() will be done with the full length, and will
 585                  * get the full data.
 586                  */
 587                 dt_read_lock(env, o, 0);
 588                 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
 589                 dt_read_unlock(env, o);
 590                 if (rc < 0) {
 591                         CERROR("%s: can't read llog block from log "DFID
 592                                " offset "LPU64": rc = %d\n",
 593                                o->do_lu.lo_dev->ld_obd->obd_name,
 594                                PFID(lu_object_fid(&o->do_lu)), *cur_offset,
 595                                rc);
 596                         GOTO(out, rc);
 597                 }
 598
 599                 if (rc < len) {
 600                         /* signal the end of the valid buffer to
 601                          * llog_process */
 602                         memset(buf + rc, 0, len - rc);
 603                 }
 604
 605                 if (rc == 0) /* end of file, nothing to do */
 606                         GOTO(out, rc);
 607
 608                 if (rc < sizeof(*tail)) {
 609                         CERROR("%s: invalid llog block at log id "LPU64"/%u "
 610                                "offset "LPU64"\n",
 611                                o->do_lu.lo_dev->ld_obd->obd_name,
 612                                loghandle->lgh_id.lgl_oid,
 613                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 614                         GOTO(out, rc = -EINVAL);
 615                 }
 616
 617                 rec = buf;
 618                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 619                         lustre_swab_llog_rec(rec);
 620
 621                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 622                                                 sizeof(struct llog_rec_tail));
 623                 /* get the last record in block */
 624                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 625                                                    le32_to_cpu(tail->lrt_len));
 626
 627                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 628                         lustre_swab_llog_rec(last_rec);
 629                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 630
 631                 *cur_idx = tail->lrt_index;
 632
 633                 /* this shouldn't happen */
 634                 if (tail->lrt_index == 0) {
 635                         CERROR("%s: invalid llog tail at log id "LPU64"/%u "
 636                                "offset "LPU64"\n",
 637                                o->do_lu.lo_dev->ld_obd->obd_name,
 638                                loghandle->lgh_id.lgl_oid,
 639                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 640                         GOTO(out, rc = -EINVAL);
 641                 }
 642                 if (tail->lrt_index < next_idx)
 643                         continue;
 644
 645                 /* sanity check that the start of the new buffer is no farther
 646                  * than the record that we wanted.  This shouldn't happen. */
 647                 if (rec->lrh_index > next_idx) {
 648                         CERROR("%s: missed desired record? %u > %u\n",
 649                                o->do_lu.lo_dev->ld_obd->obd_name,
 650                                rec->lrh_index, next_idx);
 651                         GOTO(out, rc = -ENOENT);
 652                 }
 653                 GOTO(out, rc = 0);
 654         }
 655         GOTO(out, rc = -EIO);
 656 out:
 657         return rc;
 658 }
 659
 660 static int llog_osd_prev_block(const struct lu_env *env,
 661                                struct llog_handle *loghandle,
 662                                int prev_idx, void *buf, int len)
 663 {
 664         struct llog_thread_info *lgi = llog_info(env);
 665         struct dt_object        *o;
 666         struct dt_device        *dt;
 667         loff_t                   cur_offset;
 668         int                      rc;
 669
 670         ENTRY;
 671
 672         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 673                 RETURN(-EINVAL);
 674
 675         CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
 676
 677         LASSERT(loghandle);
 678         LASSERT(loghandle->lgh_ctxt);
 679
 680         o = loghandle->lgh_obj;
 681         LASSERT(o);
 682         LASSERT(dt_object_exists(o));
 683         dt = lu2dt_dev(o->do_lu.lo_dev);
 684         LASSERT(dt);
 685
 686         cur_offset = LLOG_CHUNK_SIZE;
 687         llog_skip_over(&cur_offset, 0, prev_idx);
 688
 689         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 690         if (rc)
 691                 GOTO(out, rc);
 692
 693         while (cur_offset < lgi->lgi_attr.la_size) {
 694                 struct llog_rec_hdr     *rec, *last_rec;
 695                 struct llog_rec_tail    *tail;
 696
 697                 lgi->lgi_buf.lb_len = len;
 698                 lgi->lgi_buf.lb_buf = buf;
 699                 /* It is OK to have locking around dt_read() only, see
 700                  * comment in llog_osd_next_block for details
 701                  */
 702                 dt_read_lock(env, o, 0);
 703                 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
 704                 dt_read_unlock(env, o);
 705                 if (rc < 0) {
 706                         CERROR("%s: can't read llog block from log "DFID
 707                                " offset "LPU64": rc = %d\n",
 708                                o->do_lu.lo_dev->ld_obd->obd_name,
 709                                PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
 710                         GOTO(out, rc);
 711                 }
 712
 713                 if (rc == 0) /* end of file, nothing to do */
 714                         GOTO(out, rc);
 715
 716                 if (rc < sizeof(*tail)) {
 717                         CERROR("%s: invalid llog block at log id "LPU64"/%u "
 718                                "offset "LPU64"\n",
 719                                o->do_lu.lo_dev->ld_obd->obd_name,
 720                                loghandle->lgh_id.lgl_oid,
 721                                loghandle->lgh_id.lgl_ogen, cur_offset);
 722                         GOTO(out, rc = -EINVAL);
 723                 }
 724
 725                 rec = buf;
 726                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 727                         lustre_swab_llog_rec(rec);
 728
 729                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 730                                                 sizeof(struct llog_rec_tail));
 731                 /* get the last record in block */
 732                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 733                                                    le32_to_cpu(tail->lrt_len));
 734
 735                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 736                         lustre_swab_llog_rec(last_rec);
 737                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 738
 739                 /* this shouldn't happen */
 740                 if (tail->lrt_index == 0) {
 741                         CERROR("%s: invalid llog tail at log id "LPU64"/%u "
 742                                "offset "LPU64"\n",
 743                                o->do_lu.lo_dev->ld_obd->obd_name,
 744                                loghandle->lgh_id.lgl_oid,
 745                                loghandle->lgh_id.lgl_ogen, cur_offset);
 746                         GOTO(out, rc = -EINVAL);
 747                 }
 748                 if (tail->lrt_index < prev_idx)
 749                         continue;
 750
 751                 /* sanity check that the start of the new buffer is no farther
 752                  * than the record that we wanted.  This shouldn't happen. */
 753                 if (rec->lrh_index > prev_idx) {
 754                         CERROR("%s: missed desired record? %u > %u\n",
 755                                o->do_lu.lo_dev->ld_obd->obd_name,
 756                                rec->lrh_index, prev_idx);
 757                         GOTO(out, rc = -ENOENT);
 758                 }
 759                 GOTO(out, rc = 0);
 760         }
 761         GOTO(out, rc = -EIO);
 762 out:
 763         return rc;
 764 }
 765
 766 struct dt_object *llog_osd_dir_get(const struct lu_env *env,
 767                                    struct llog_ctxt *ctxt)
 768 {
 769         struct dt_device        *dt;
 770         struct dt_thread_info   *dti = dt_info(env);
 771         struct dt_object        *dir;
 772         int                      rc;
 773
 774         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 775         if (ctxt->loc_dir == NULL) {
 776                 rc = dt_root_get(env, dt, &dti->dti_fid);
 777                 if (rc)
 778                         return ERR_PTR(rc);
 779                 dir = dt_locate(env, dt, &dti->dti_fid);
 780         } else {
 781                 lu_object_get(&ctxt->loc_dir->do_lu);
 782                 dir = ctxt->loc_dir;
 783         }
 784
 785         return dir;
 786 }
 787
 788 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
 789                          struct llog_logid *logid, char *name,
 790                          enum llog_open_param open_param)
 791 {
 792         struct llog_thread_info         *lgi = llog_info(env);
 793         struct llog_ctxt                *ctxt = handle->lgh_ctxt;
 794         struct dt_object                *o;
 795         struct dt_device                *dt;
 796         struct ls_device                *ls;
 797         struct local_oid_storage        *los;
 798         int                              rc = 0;
 799
 800         ENTRY;
 801
 802         LASSERT(env);
 803         LASSERT(ctxt);
 804         LASSERT(ctxt->loc_exp);
 805         LASSERT(ctxt->loc_exp->exp_obd);
 806         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 807         LASSERT(dt);
 808
 809         ls = ls_device_get(dt);
 810         if (IS_ERR(ls))
 811                 RETURN(PTR_ERR(ls));
 812
 813         mutex_lock(&ls->ls_los_mutex);
 814         los = dt_los_find(ls, FID_SEQ_LLOG);
 815         mutex_unlock(&ls->ls_los_mutex);
 816         LASSERT(los);
 817         ls_device_put(env, ls);
 818
 819         LASSERT(handle);
 820
 821         if (logid != NULL) {
 822                 logid_to_fid(logid, &lgi->lgi_fid);
 823         } else if (name) {
 824                 struct dt_object *llog_dir;
 825
 826                 llog_dir = llog_osd_dir_get(env, ctxt);
 827                 if (IS_ERR(llog_dir))
 828                         GOTO(out, rc = PTR_ERR(llog_dir));
 829                 dt_read_lock(env, llog_dir, 0);
 830                 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
 831                 dt_read_unlock(env, llog_dir);
 832                 lu_object_put(env, &llog_dir->do_lu);
 833                 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
 834                         /* generate fid for new llog */
 835                         rc = local_object_fid_generate(env, los,
 836                                                        &lgi->lgi_fid);
 837                 }
 838                 if (rc < 0)
 839                         GOTO(out, rc);
 840                 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
 841                 if (handle->lgh_name)
 842                         strcpy(handle->lgh_name, name);
 843                 else
 844                         GOTO(out, rc = -ENOMEM);
 845         } else {
 846                 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
 847                 /* generate fid for new llog */
 848                 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
 849                 if (rc < 0)
 850                         GOTO(out, rc);
 851         }
 852
 853         o = ls_locate(env, ls, &lgi->lgi_fid);
 854         if (IS_ERR(o))
 855                 GOTO(out_name, rc = PTR_ERR(o));
 856
 857         /* No new llog is expected but doesn't exist */
 858         if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
 859                 GOTO(out_put, rc = -ENOENT);
 860
 861         fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
 862         handle->lgh_obj = o;
 863         handle->private_data = los;
 864         LASSERT(handle->lgh_ctxt);
 865
 866         RETURN(rc);
 867
 868 out_put:
 869         lu_object_put(env, &o->do_lu);
 870 out_name:
 871         if (handle->lgh_name != NULL)
 872                 OBD_FREE(handle->lgh_name, strlen(name) + 1);
 873 out:
 874         dt_los_put(los);
 875         RETURN(rc);
 876 }
 877
 878 static int llog_osd_exist(struct llog_handle *handle)
 879 {
 880         LASSERT(handle->lgh_obj);
 881         return (dt_object_exists(handle->lgh_obj) &&
 882                 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
 883 }
 884
 885 static int llog_osd_declare_create(const struct lu_env *env,
 886                                    struct llog_handle *res, struct thandle *th)
 887 {
 888         struct llog_thread_info         *lgi = llog_info(env);
 889         struct local_oid_storage        *los;
 890         struct dt_object                *o;
 891         int                              rc;
 892
 893         ENTRY;
 894
 895         LASSERT(res->lgh_obj);
 896         LASSERT(th);
 897
 898         /* object can be created by another thread */
 899         o = res->lgh_obj;
 900         if (dt_object_exists(o))
 901                 RETURN(0);
 902
 903         los = res->private_data;
 904         LASSERT(los);
 905
 906         rc = llog_osd_declare_new_object(env, los, o, th);
 907         if (rc)
 908                 RETURN(rc);
 909
 910         rc = dt_declare_record_write(env, o, LLOG_CHUNK_SIZE, 0, th);
 911         if (rc)
 912                 RETURN(rc);
 913
 914         if (res->lgh_name) {
 915                 struct dt_object *llog_dir;
 916
 917                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 918                 if (IS_ERR(llog_dir))
 919                         RETURN(PTR_ERR(llog_dir));
 920                 dt_declare_ref_add(env, o, th);
 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         if (res->lgh_name)
 963                 dt_ref_add(env, o, th);
 964         dt_write_unlock(env, o);
 965         if (rc)
 966                 RETURN(rc);
 967
 968         if (res->lgh_name) {
 969                 struct dt_object *llog_dir;
 970
 971                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 972                 if (IS_ERR(llog_dir))
 973                         RETURN(PTR_ERR(llog_dir));
 974
 975                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 976                 dt_read_lock(env, llog_dir, 0);
 977                 rc = dt_insert(env, llog_dir,
 978                                (struct dt_rec *)&lgi->lgi_fid,
 979                                (struct dt_key *)res->lgh_name,
 980                                th, BYPASS_CAPA, 1);
 981                 dt_read_unlock(env, llog_dir);
 982                 lu_object_put(env, &llog_dir->do_lu);
 983                 if (rc)
 984                         CERROR("%s: can't create named llog %s: rc = %d\n",
 985                                o->do_lu.lo_dev->ld_obd->obd_name,
 986                                res->lgh_name, rc);
 987         }
 988         RETURN(rc);
 989 }
 990
 991 static int llog_osd_close(const struct lu_env *env, struct llog_handle *handle)
 992 {
 993         struct local_oid_storage        *los;
 994         int                              rc = 0;
 995
 996         ENTRY;
 997
 998         LASSERT(handle->lgh_obj);
 999
1000         lu_object_put(env, &handle->lgh_obj->do_lu);
1001
1002         los = handle->private_data;
1003         LASSERT(los);
1004         dt_los_put(los);
1005
1006         if (handle->lgh_name)
1007                 OBD_FREE(handle->lgh_name, strlen(handle->lgh_name) + 1);
1008
1009         RETURN(rc);
1010 }
1011
1012 static int llog_osd_destroy(const struct lu_env *env,
1013                             struct llog_handle *loghandle)
1014 {
1015         struct llog_thread_info *lgi = llog_info(env);
1016         struct llog_ctxt        *ctxt;
1017         struct dt_object        *o, *llog_dir = NULL;
1018         struct dt_device        *d;
1019         struct thandle          *th;
1020         char                    *name = NULL;
1021         int                      rc;
1022
1023         ENTRY;
1024
1025         ctxt = loghandle->lgh_ctxt;
1026         LASSERT(ctxt);
1027
1028         o = loghandle->lgh_obj;
1029         LASSERT(o);
1030
1031         d = lu2dt_dev(o->do_lu.lo_dev);
1032         LASSERT(d);
1033         LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
1034
1035         th = dt_trans_create(env, d);
1036         if (IS_ERR(th))
1037                 RETURN(PTR_ERR(th));
1038
1039         if (loghandle->lgh_name) {
1040                 llog_dir = llog_osd_dir_get(env, ctxt);
1041                 if (IS_ERR(llog_dir))
1042                         GOTO(out_trans, rc = PTR_ERR(llog_dir));
1043
1044                 dt_declare_ref_del(env, o, th);
1045                 name = loghandle->lgh_name;
1046                 rc = dt_declare_delete(env, llog_dir,
1047                                        (struct dt_key *)name, th);
1048                 if (rc)
1049                         GOTO(out_trans, rc);
1050         }
1051
1052         dt_declare_ref_del(env, o, th);
1053
1054         rc = dt_declare_destroy(env, o, th);
1055         if (rc)
1056                 GOTO(out_trans, rc);
1057
1058         rc = dt_trans_start_local(env, d, th);
1059         if (rc)
1060                 GOTO(out_trans, rc);
1061
1062         dt_write_lock(env, o, 0);
1063         if (dt_object_exists(o)) {
1064                 if (name) {
1065                         dt_ref_del(env, o, th);
1066                         dt_read_lock(env, llog_dir, 0);
1067                         rc = dt_delete(env, llog_dir,
1068                                        (struct dt_key *) name,
1069                                        th, BYPASS_CAPA);
1070                         dt_read_unlock(env, llog_dir);
1071                         if (rc) {
1072                                 CERROR("%s: can't remove llog %s: rc = %d\n",
1073                                        o->do_lu.lo_dev->ld_obd->obd_name,
1074                                        name, rc);
1075                                 GOTO(out_unlock, rc);
1076                         }
1077                 }
1078                 /*
1079                  * XXX: compatibility bits
1080                  *      on old filesystems llogs are referenced by the name
1081                  *      on the new ones they are referenced by OI and by
1082                  *      the name
1083                  */
1084                 rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
1085                 if (rc)
1086                         GOTO(out_unlock, rc);
1087                 LASSERT(lgi->lgi_attr.la_nlink < 2);
1088                 if (lgi->lgi_attr.la_nlink == 1)
1089                         dt_ref_del(env, o, th);
1090                 rc = dt_destroy(env, o, th);
1091                 if (rc)
1092                         GOTO(out_unlock, rc);
1093         }
1094 out_unlock:
1095         dt_write_unlock(env, o);
1096 out_trans:
1097         dt_trans_stop(env, d, th);
1098         if (llog_dir != NULL)
1099                 lu_object_put(env, &llog_dir->do_lu);
1100         RETURN(rc);
1101 }
1102
1103 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1104                           struct obd_llog_group *olg, int ctxt_idx,
1105                           struct obd_device *disk_obd)
1106 {
1107         struct local_oid_storage        *los;
1108         struct llog_thread_info         *lgi = llog_info(env);
1109         struct llog_ctxt                *ctxt;
1110         int                              rc = 0;
1111
1112         ENTRY;
1113
1114         LASSERT(obd);
1115         LASSERT(olg->olg_ctxts[ctxt_idx]);
1116
1117         ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1118         LASSERT(ctxt);
1119
1120         /* initialize data allowing to generate new fids,
1121          * literally we need a sequece */
1122         lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1123         lgi->lgi_fid.f_oid = 1;
1124         lgi->lgi_fid.f_ver = 0;
1125         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1126                                     &lgi->lgi_fid, &los);
1127         llog_ctxt_put(ctxt);
1128         return rc;
1129 }
1130
1131 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1132 {
1133         struct dt_device                *dt;
1134         struct ls_device                *ls;
1135         struct local_oid_storage        *los;
1136
1137         LASSERT(ctxt->loc_exp->exp_obd);
1138         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
1139         ls = ls_device_get(dt);
1140         if (IS_ERR(ls))
1141                 RETURN(PTR_ERR(ls));
1142
1143         mutex_lock(&ls->ls_los_mutex);
1144         los = dt_los_find(ls, FID_SEQ_LLOG);
1145         mutex_unlock(&ls->ls_los_mutex);
1146         if (los != NULL) {
1147                 dt_los_put(los);
1148                 local_oid_storage_fini(env, los);
1149         }
1150         ls_device_put(env, ls);
1151         return 0;
1152 }
1153
1154 struct llog_operations llog_osd_ops = {
1155         .lop_next_block         = llog_osd_next_block,
1156         .lop_prev_block         = llog_osd_prev_block,
1157         .lop_read_header        = llog_osd_read_header,
1158         .lop_destroy            = llog_osd_destroy,
1159         .lop_setup              = llog_osd_setup,
1160         .lop_cleanup            = llog_osd_cleanup,
1161         .lop_open               = llog_osd_open,
1162         .lop_exist              = llog_osd_exist,
1163         .lop_declare_create     = llog_osd_declare_create,
1164         .lop_create             = llog_osd_create,
1165         .lop_declare_write_rec  = llog_osd_declare_write_rec,
1166         .lop_write_rec          = llog_osd_write_rec,
1167         .lop_close              = llog_osd_close,
1168 };
1169 EXPORT_SYMBOL(llog_osd_ops);
1170
1171 /* reads the catalog list */
1172 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1173                           int idx, int count, struct llog_catid *idarray)
1174 {
1175         struct llog_thread_info *lgi = llog_info(env);
1176         struct dt_object        *o = NULL;
1177         struct thandle          *th;
1178         int                      rc, size;
1179
1180         ENTRY;
1181
1182         LASSERT(d);
1183
1184         size = sizeof(*idarray) * count;
1185         lgi->lgi_off = idx *  sizeof(*idarray);
1186
1187         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1188
1189         o = dt_locate(env, d, &lgi->lgi_fid);
1190         if (IS_ERR(o))
1191                 RETURN(PTR_ERR(o));
1192
1193         if (!dt_object_exists(o)) {
1194                 th = dt_trans_create(env, d);
1195                 if (IS_ERR(th))
1196                         GOTO(out, rc = PTR_ERR(th));
1197
1198                 lgi->lgi_attr.la_valid = LA_MODE;
1199                 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1200                 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1201
1202                 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1203                                        &lgi->lgi_dof, th);
1204                 if (rc)
1205                         GOTO(out_trans, rc);
1206
1207                 rc = dt_trans_start_local(env, d, th);
1208                 if (rc)
1209                         GOTO(out_trans, rc);
1210
1211                 dt_write_lock(env, o, 0);
1212                 if (!dt_object_exists(o))
1213                         rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1214                                        &lgi->lgi_dof, th);
1215                 dt_write_unlock(env, o);
1216 out_trans:
1217                 dt_trans_stop(env, d, th);
1218                 if (rc)
1219                         GOTO(out, rc);
1220         }
1221
1222         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1223         if (rc)
1224                 GOTO(out, rc);
1225
1226         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1227                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1228                        o->do_lu.lo_dev->ld_obd->obd_name,
1229                        lgi->lgi_attr.la_mode);
1230                 GOTO(out, rc = -ENOENT);
1231         }
1232
1233         CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1234                (int)lgi->lgi_attr.la_size, size);
1235
1236         /* return just number of llogs */
1237         if (idarray == NULL) {
1238                 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1239                 GOTO(out, rc);
1240         }
1241
1242         /* read for new ost index or for empty file */
1243         memset(idarray, 0, size);
1244         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1245                 GOTO(out, rc = 0);
1246         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1247                 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1248
1249         lgi->lgi_buf.lb_buf = idarray;
1250         lgi->lgi_buf.lb_len = size;
1251         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1252         if (rc) {
1253                 CERROR("%s: error reading CATALOGS: rc = %d\n",
1254                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
1255                 GOTO(out, rc);
1256         }
1257
1258         EXIT;
1259 out:
1260         lu_object_put(env, &o->do_lu);
1261         RETURN(rc);
1262 }
1263 EXPORT_SYMBOL(llog_osd_get_cat_list);
1264
1265 /* writes the cat list */
1266 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1267                           int idx, int count, struct llog_catid *idarray)
1268 {
1269         struct llog_thread_info *lgi = llog_info(env);
1270         struct dt_object        *o = NULL;
1271         struct thandle          *th;
1272         int                      rc, size;
1273
1274         if (!count)
1275                 RETURN(0);
1276
1277         LASSERT(d);
1278
1279         size = sizeof(*idarray) * count;
1280         lgi->lgi_off = idx * sizeof(*idarray);
1281
1282         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1283
1284         o = dt_locate(env, d, &lgi->lgi_fid);
1285         if (IS_ERR(o))
1286                 RETURN(PTR_ERR(o));
1287
1288         if (!dt_object_exists(o))
1289                 GOTO(out, rc = -ENOENT);
1290
1291         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1292         if (rc)
1293                 GOTO(out, rc);
1294
1295         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1296                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1297                        o->do_lu.lo_dev->ld_obd->obd_name,
1298                        lgi->lgi_attr.la_mode);
1299                 GOTO(out, rc = -ENOENT);
1300         }
1301
1302         th = dt_trans_create(env, d);
1303         if (IS_ERR(th))
1304                 GOTO(out, rc = PTR_ERR(th));
1305
1306         rc = dt_declare_record_write(env, o, size, lgi->lgi_off, th);
1307         if (rc)
1308                 GOTO(out, rc);
1309
1310         rc = dt_trans_start_local(env, d, th);
1311         if (rc)
1312                 GOTO(out_trans, rc);
1313
1314         lgi->lgi_buf.lb_buf = idarray;
1315         lgi->lgi_buf.lb_len = size;
1316         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1317         if (rc)
1318                 CDEBUG(D_INODE, "error writeing CATALOGS: rc = %d\n", rc);
1319 out_trans:
1320         dt_trans_stop(env, d, th);
1321 out:
1322         lu_object_put(env, &o->do_lu);
1323         RETURN(rc);
1324 }
1325 EXPORT_SYMBOL(llog_osd_put_cat_list);