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) 2011, 2012, Intel, Inc.
  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         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 113         if (rc) {
 114                 CERROR("%s: error writing padding record: rc = %d\n",
 115                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 116                 RETURN(rc);
 117         }
 118
 119         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 120         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 121         *off += len - sizeof(lgi->lgi_lrh) - sizeof(lgi->lgi_tail);
 122         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 123         if (rc)
 124                 CERROR("%s: error writing padding record: rc = %d\n",
 125                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 126
 127         RETURN(rc);
 128 }
 129
 130 static int llog_osd_write_blob(const struct lu_env *env, struct dt_object *o,
 131                                struct llog_rec_hdr *rec, void *buf,
 132                                loff_t *off, struct thandle *th)
 133 {
 134         struct llog_thread_info *lgi = llog_info(env);
 135         int                      buflen = rec->lrh_len;
 136         int                      rc;
 137
 138         ENTRY;
 139
 140         LASSERT(env);
 141         LASSERT(o);
 142
 143         if (buflen == 0)
 144                 CWARN("0-length record\n");
 145
 146         CDEBUG(D_OTHER, "write blob with type %x, buf %p/%u at off %llu\n",
 147                rec->lrh_type, buf, buflen, *off);
 148
 149         if (!buf) {
 150                 lgi->lgi_buf.lb_len = buflen;
 151                 lgi->lgi_buf.lb_buf = rec;
 152                 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 153                 if (rc)
 154                         CERROR("%s: error writing log record: rc = %d\n",
 155                                o->do_lu.lo_dev->ld_obd->obd_name, rc);
 156                 RETURN(rc);
 157         }
 158
 159         /* the buf case */
 160         rec->lrh_len = sizeof(*rec) + buflen + sizeof(lgi->lgi_tail);
 161         lgi->lgi_buf.lb_len = sizeof(*rec);
 162         lgi->lgi_buf.lb_buf = rec;
 163         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 164         if (rc) {
 165                 CERROR("%s: error writing log hdr: rc = %d\n",
 166                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 167                 GOTO(out, rc);
 168         }
 169
 170         lgi->lgi_buf.lb_len = buflen;
 171         lgi->lgi_buf.lb_buf = buf;
 172         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 173         if (rc) {
 174                 CERROR("%s: error writing log buffer: rc = %d\n",
 175                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
 176                 GOTO(out, rc);
 177         }
 178
 179         lgi->lgi_tail.lrt_len = rec->lrh_len;
 180         lgi->lgi_tail.lrt_index = rec->lrh_index;
 181         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 182         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 183         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 184         if (rc)
 185                 CERROR("%s: error writing log tail: rc = %d\n",
 186                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 187 out:
 188         RETURN(rc);
 189 }
 190
 191 static int llog_osd_read_header(const struct lu_env *env,
 192                                 struct llog_handle *handle)
 193 {
 194         struct llog_rec_hdr     *llh_hdr;
 195         struct dt_object        *o;
 196         struct llog_thread_info *lgi;
 197         int                      rc;
 198
 199         ENTRY;
 200
 201         LASSERT(sizeof(*handle->lgh_hdr) == LLOG_CHUNK_SIZE);
 202
 203         o = handle->lgh_obj;
 204         LASSERT(o);
 205
 206         lgi = llog_info(env);
 207
 208         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 209         if (rc)
 210                 RETURN(rc);
 211
 212         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 213
 214         if (lgi->lgi_attr.la_size == 0) {
 215                 CDEBUG(D_HA, "not reading header from 0-byte log\n");
 216                 RETURN(LLOG_EEMPTY);
 217         }
 218
 219         lgi->lgi_off = 0;
 220         lgi->lgi_buf.lb_buf = handle->lgh_hdr;
 221         lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE;
 222
 223         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
 224         if (rc) {
 225                 CERROR("%s: error reading log header from "DFID": rc = %d\n",
 226                        o->do_lu.lo_dev->ld_obd->obd_name,
 227                        PFID(lu_object_fid(&o->do_lu)), rc);
 228                 RETURN(rc);
 229         }
 230
 231         llh_hdr = &handle->lgh_hdr->llh_hdr;
 232         if (LLOG_REC_HDR_NEEDS_SWABBING(llh_hdr))
 233                 lustre_swab_llog_hdr(handle->lgh_hdr);
 234
 235         if (llh_hdr->lrh_type != LLOG_HDR_MAGIC) {
 236                 CERROR("%s: bad log %s "DFID" header magic: %#x "
 237                        "(expected %#x)\n", o->do_lu.lo_dev->ld_obd->obd_name,
 238                        handle->lgh_name ? handle->lgh_name : "",
 239                        PFID(lu_object_fid(&o->do_lu)),
 240                        llh_hdr->lrh_type, LLOG_HDR_MAGIC);
 241                 RETURN(-EIO);
 242         } else if (llh_hdr->lrh_len != LLOG_CHUNK_SIZE) {
 243                 CERROR("%s: incorrectly sized log %s "DFID" header: "
 244                        "%#x (expected %#x)\n"
 245                        "you may need to re-run lconf --write_conf.\n",
 246                        o->do_lu.lo_dev->ld_obd->obd_name,
 247                        handle->lgh_name ? handle->lgh_name : "",
 248                        PFID(lu_object_fid(&o->do_lu)),
 249                        llh_hdr->lrh_len, LLOG_CHUNK_SIZE);
 250                 RETURN(-EIO);
 251         }
 252
 253         handle->lgh_last_idx = handle->lgh_hdr->llh_tail.lrt_index;
 254
 255         RETURN(0);
 256 }
 257
 258 static int llog_osd_declare_write_rec(const struct lu_env *env,
 259                                       struct llog_handle *loghandle,
 260                                       struct llog_rec_hdr *rec,
 261                                       int idx, struct thandle *th)
 262 {
 263         struct llog_thread_info *lgi = llog_info(env);
 264         struct dt_object        *o;
 265         int                      rc;
 266
 267         ENTRY;
 268
 269         LASSERT(env);
 270         LASSERT(th);
 271         LASSERT(loghandle);
 272
 273         o = loghandle->lgh_obj;
 274         LASSERT(o);
 275
 276         /* each time we update header */
 277         rc = dt_declare_record_write(env, o, sizeof(struct llog_log_hdr), 0,
 278                                      th);
 279         if (rc || idx == 0) /* if error or just header */
 280                 RETURN(rc);
 281
 282         if (dt_object_exists(o)) {
 283                 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 284                 lgi->lgi_off = lgi->lgi_attr.la_size;
 285                 LASSERT(ergo(rc == 0, lgi->lgi_attr.la_valid & LA_SIZE));
 286                 if (rc)
 287                         RETURN(rc);
 288         } else {
 289                 lgi->lgi_off = 0;
 290         }
 291
 292         /* XXX: implement declared window or multi-chunks approach */
 293         rc = dt_declare_record_write(env, o, 32 * 1024, lgi->lgi_off, th);
 294
 295         RETURN(rc);
 296 }
 297
 298 /* returns negative in on error; 0 if success && reccookie == 0; 1 otherwise */
 299 /* appends if idx == -1, otherwise overwrites record idx. */
 300 static int llog_osd_write_rec(const struct lu_env *env,
 301                               struct llog_handle *loghandle,
 302                               struct llog_rec_hdr *rec,
 303                               struct llog_cookie *reccookie, int cookiecount,
 304                               void *buf, int idx, struct thandle *th)
 305 {
 306         struct llog_thread_info *lgi = llog_info(env);
 307         struct llog_log_hdr     *llh;
 308         int                      reclen = rec->lrh_len;
 309         int                      index, rc;
 310         struct llog_rec_tail    *lrt;
 311         struct dt_object        *o;
 312         size_t                   left;
 313
 314         ENTRY;
 315
 316         LASSERT(env);
 317         llh = loghandle->lgh_hdr;
 318         LASSERT(llh);
 319         o = loghandle->lgh_obj;
 320         LASSERT(o);
 321         LASSERT(th);
 322
 323         CDEBUG(D_OTHER, "new record %x to "DFID"\n",
 324                rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
 325
 326         /* record length should not bigger than LLOG_CHUNK_SIZE */
 327         if (buf)
 328                 rc = (reclen > LLOG_CHUNK_SIZE - sizeof(struct llog_rec_hdr) -
 329                       sizeof(struct llog_rec_tail)) ? -E2BIG : 0;
 330         else
 331                 rc = (reclen > LLOG_CHUNK_SIZE) ? -E2BIG : 0;
 332         if (rc)
 333                 RETURN(rc);
 334
 335         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 336         if (rc)
 337                 RETURN(rc);
 338
 339         if (buf)
 340                 /* write_blob adds header and tail to lrh_len. */
 341                 reclen = sizeof(*rec) + rec->lrh_len +
 342                          sizeof(struct llog_rec_tail);
 343
 344         if (idx != -1) {
 345                 /* no header: only allowed to insert record 1 */
 346                 if (idx != 1 && lgi->lgi_attr.la_size == 0)
 347                         LBUG();
 348
 349                 if (idx && llh->llh_size && llh->llh_size != rec->lrh_len)
 350                         RETURN(-EINVAL);
 351
 352                 if (!ext2_test_bit(idx, llh->llh_bitmap))
 353                         CERROR("%s: modify unset record %u\n",
 354                                o->do_lu.lo_dev->ld_obd->obd_name, idx);
 355                 if (idx != rec->lrh_index)
 356                         CERROR("%s: index mismatch %d %u\n",
 357                                o->do_lu.lo_dev->ld_obd->obd_name, idx,
 358                                rec->lrh_index);
 359
 360                 lgi->lgi_off = 0;
 361                 rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 362                                          &lgi->lgi_off, th);
 363                 /* we are done if we only write the header or on error */
 364                 if (rc || idx == 0)
 365                         RETURN(rc);
 366
 367                 if (buf) {
 368                         /* We assume that caller has set lgh_cur_* */
 369                         lgi->lgi_off = loghandle->lgh_cur_offset;
 370                         CDEBUG(D_OTHER,
 371                                "modify record "LPX64": idx:%d/%u/%d, len:%u "
 372                                "offset %llu\n",
 373                                loghandle->lgh_id.lgl_oid, idx, rec->lrh_index,
 374                                loghandle->lgh_cur_idx, rec->lrh_len,
 375                                (long long)(lgi->lgi_off - sizeof(*llh)));
 376                         if (rec->lrh_index != loghandle->lgh_cur_idx) {
 377                                 CERROR("%s: modify idx mismatch %u/%d\n",
 378                                        o->do_lu.lo_dev->ld_obd->obd_name, idx,
 379                                        loghandle->lgh_cur_idx);
 380                                 RETURN(-EFAULT);
 381                         }
 382                 } else {
 383                         /* Assumes constant lrh_len */
 384                         lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
 385                 }
 386
 387                 rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 388                 if (rc == 0 && reccookie) {
 389                         reccookie->lgc_lgl = loghandle->lgh_id;
 390                         reccookie->lgc_index = idx;
 391                         rc = 1;
 392                 }
 393                 RETURN(rc);
 394         }
 395
 396         /* Make sure that records don't cross a chunk boundary, so we can
 397          * process them page-at-a-time if needed.  If it will cross a chunk
 398          * boundary, write in a fake (but referenced) entry to pad the chunk.
 399          *
 400          * We know that llog_current_log() will return a loghandle that is
 401          * big enough to hold reclen, so all we care about is padding here.
 402          */
 403         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 404         lgi->lgi_off = lgi->lgi_attr.la_size;
 405         left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
 406         /* NOTE: padding is a record, but no bit is set */
 407         if (left != 0 && left != reclen &&
 408             left < (reclen + LLOG_MIN_REC_SIZE)) {
 409                 index = loghandle->lgh_last_idx + 1;
 410                 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
 411                 if (rc)
 412                         RETURN(rc);
 413                 loghandle->lgh_last_idx++; /*for pad rec*/
 414         }
 415         /* if it's the last idx in log file, then return -ENOSPC */
 416         if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
 417                 RETURN(-ENOSPC);
 418
 419         loghandle->lgh_last_idx++;
 420         index = loghandle->lgh_last_idx;
 421         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 422         rec->lrh_index = index;
 423         if (buf == NULL) {
 424                 lrt = (struct llog_rec_tail *)((char *)rec + rec->lrh_len -
 425                                                sizeof(*lrt));
 426                 lrt->lrt_len = rec->lrh_len;
 427                 lrt->lrt_index = rec->lrh_index;
 428         }
 429         /* The caller should make sure only 1 process access the lgh_last_idx,
 430          * Otherwise it might hit the assert.*/
 431         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 432         cfs_spin_lock(&loghandle->lgh_hdr_lock);
 433         if (ext2_set_bit(index, llh->llh_bitmap)) {
 434                 CERROR("%s: index %u already set in log bitmap\n",
 435                        o->do_lu.lo_dev->ld_obd->obd_name, index);
 436                 cfs_spin_unlock(&loghandle->lgh_hdr_lock);
 437                 LBUG(); /* should never happen */
 438         }
 439         llh->llh_count++;
 440         cfs_spin_unlock(&loghandle->lgh_hdr_lock);
 441         llh->llh_tail.lrt_index = index;
 442
 443         lgi->lgi_off = 0;
 444         rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL, &lgi->lgi_off,
 445                                  th);
 446         if (rc)
 447                 RETURN(rc);
 448
 449         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 450         if (rc)
 451                 RETURN(rc);
 452         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 453         lgi->lgi_off = lgi->lgi_attr.la_size;
 454
 455         rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 456         if (rc)
 457                 RETURN(rc);
 458
 459         CDEBUG(D_RPCTRACE, "added record "LPX64": idx: %u, %u\n",
 460                loghandle->lgh_id.lgl_oid, index, rec->lrh_len);
 461         if (rc == 0 && reccookie) {
 462                 reccookie->lgc_lgl = loghandle->lgh_id;
 463                 reccookie->lgc_index = index;
 464                 if ((rec->lrh_type == MDS_UNLINK_REC) ||
 465                     (rec->lrh_type == MDS_SETATTR64_REC))
 466                         reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 467                 else if (rec->lrh_type == OST_SZ_REC)
 468                         reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
 469                 else
 470                         reccookie->lgc_subsys = -1;
 471                 rc = 1;
 472         }
 473         RETURN(rc);
 474 }
 475
 476 /* We can skip reading at least as many log blocks as the number of
 477  * minimum sized log records we are skipping.  If it turns out
 478  * that we are not far enough along the log (because the
 479  * actual records are larger than minimum size) we just skip
 480  * some more records.
 481  */
 482 static void llog_skip_over(__u64 *off, int curr, int goal)
 483 {
 484         if (goal <= curr)
 485                 return;
 486         *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
 487                 ~(LLOG_CHUNK_SIZE - 1);
 488 }
 489
 490 /* sets:
 491  *  - cur_offset to the furthest point read in the log file
 492  *  - cur_idx to the log index preceeding cur_offset
 493  * returns -EIO/-EINVAL on error
 494  */
 495 static int llog_osd_next_block(const struct lu_env *env,
 496                                struct llog_handle *loghandle, int *cur_idx,
 497                                int next_idx, __u64 *cur_offset, void *buf,
 498                                int len)
 499 {
 500         struct llog_thread_info *lgi = llog_info(env);
 501         struct dt_object        *o;
 502         struct dt_device        *dt;
 503         int                      rc;
 504
 505         ENTRY;
 506
 507         LASSERT(env);
 508         LASSERT(lgi);
 509
 510         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 511                 RETURN(-EINVAL);
 512
 513         CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
 514                next_idx, *cur_idx, *cur_offset);
 515
 516         LASSERT(loghandle);
 517         LASSERT(loghandle->lgh_ctxt);
 518
 519         o = loghandle->lgh_obj;
 520         LASSERT(o);
 521         LASSERT(dt_object_exists(o));
 522         dt = lu2dt_dev(o->do_lu.lo_dev);
 523         LASSERT(dt);
 524
 525         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 526         if (rc)
 527                 GOTO(out, rc);
 528
 529         while (*cur_offset < lgi->lgi_attr.la_size) {
 530                 struct llog_rec_hdr     *rec, *last_rec;
 531                 struct llog_rec_tail    *tail;
 532
 533                 llog_skip_over(cur_offset, *cur_idx, next_idx);
 534
 535                 /* read up to next LLOG_CHUNK_SIZE block */
 536                 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
 537                                       (*cur_offset & (LLOG_CHUNK_SIZE - 1));
 538                 lgi->lgi_buf.lb_buf = buf;
 539                 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
 540                 if (rc < 0) {
 541                         CERROR("%s: can't read llog block from log "DFID
 542                                " offset "LPU64": rc = %d\n",
 543                                o->do_lu.lo_dev->ld_obd->obd_name,
 544                                PFID(lu_object_fid(&o->do_lu)), *cur_offset,
 545                                rc);
 546                         GOTO(out, rc);
 547                 }
 548
 549                 if (rc < len) {
 550                         /* signal the end of the valid buffer to
 551                          * llog_process */
 552                         memset(buf + rc, 0, len - rc);
 553                 }
 554
 555                 if (rc == 0) /* end of file, nothing to do */
 556                         GOTO(out, rc);
 557
 558                 if (rc < sizeof(*tail)) {
 559                         CERROR("%s: invalid llog block at log id "LPU64"/%u "
 560                                "offset "LPU64"\n",
 561                                o->do_lu.lo_dev->ld_obd->obd_name,
 562                                loghandle->lgh_id.lgl_oid,
 563                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 564                         GOTO(out, rc = -EINVAL);
 565                 }
 566
 567                 rec = buf;
 568                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 569                         lustre_swab_llog_rec(rec);
 570
 571                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 572                                                 sizeof(struct llog_rec_tail));
 573                 /* get the last record in block */
 574                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 575                                                    le32_to_cpu(tail->lrt_len));
 576
 577                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 578                         lustre_swab_llog_rec(last_rec);
 579                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 580
 581                 *cur_idx = tail->lrt_index;
 582
 583                 /* this shouldn't happen */
 584                 if (tail->lrt_index == 0) {
 585                         CERROR("%s: invalid llog tail at log id "LPU64"/%u "
 586                                "offset "LPU64"\n",
 587                                o->do_lu.lo_dev->ld_obd->obd_name,
 588                                loghandle->lgh_id.lgl_oid,
 589                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 590                         GOTO(out, rc = -EINVAL);
 591                 }
 592                 if (tail->lrt_index < next_idx)
 593                         continue;
 594
 595                 /* sanity check that the start of the new buffer is no farther
 596                  * than the record that we wanted.  This shouldn't happen. */
 597                 if (rec->lrh_index > next_idx) {
 598                         CERROR("%s: missed desired record? %u > %u\n",
 599                                o->do_lu.lo_dev->ld_obd->obd_name,
 600                                rec->lrh_index, next_idx);
 601                         GOTO(out, rc = -ENOENT);
 602                 }
 603                 GOTO(out, rc = 0);
 604         }
 605         GOTO(out, rc = -EIO);
 606 out:
 607         return rc;
 608 }
 609
 610 static int llog_osd_prev_block(const struct lu_env *env,
 611                                struct llog_handle *loghandle,
 612                                int prev_idx, void *buf, int len)
 613 {
 614         struct llog_thread_info *lgi = llog_info(env);
 615         struct dt_object        *o;
 616         struct dt_device        *dt;
 617         loff_t                   cur_offset;
 618         int                      rc;
 619
 620         ENTRY;
 621
 622         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 623                 RETURN(-EINVAL);
 624
 625         CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
 626
 627         LASSERT(loghandle);
 628         LASSERT(loghandle->lgh_ctxt);
 629
 630         o = loghandle->lgh_obj;
 631         LASSERT(o);
 632         LASSERT(dt_object_exists(o));
 633         dt = lu2dt_dev(o->do_lu.lo_dev);
 634         LASSERT(dt);
 635
 636         cur_offset = LLOG_CHUNK_SIZE;
 637         llog_skip_over(&cur_offset, 0, prev_idx);
 638
 639         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 640         if (rc)
 641                 GOTO(out, rc);
 642
 643         while (cur_offset < lgi->lgi_attr.la_size) {
 644                 struct llog_rec_hdr     *rec, *last_rec;
 645                 struct llog_rec_tail    *tail;
 646
 647                 lgi->lgi_buf.lb_len = len;
 648                 lgi->lgi_buf.lb_buf = buf;
 649                 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
 650                 if (rc < 0) {
 651                         CERROR("%s: can't read llog block from log "DFID
 652                                " offset "LPU64": rc = %d\n",
 653                                o->do_lu.lo_dev->ld_obd->obd_name,
 654                                PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
 655                         GOTO(out, rc);
 656                 }
 657
 658                 if (rc == 0) /* end of file, nothing to do */
 659                         GOTO(out, rc);
 660
 661                 if (rc < sizeof(*tail)) {
 662                         CERROR("%s: invalid llog block at log id "LPU64"/%u "
 663                                "offset "LPU64"\n",
 664                                o->do_lu.lo_dev->ld_obd->obd_name,
 665                                loghandle->lgh_id.lgl_oid,
 666                                loghandle->lgh_id.lgl_ogen, cur_offset);
 667                         GOTO(out, rc = -EINVAL);
 668                 }
 669
 670                 rec = buf;
 671                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 672                         lustre_swab_llog_rec(rec);
 673
 674                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 675                                                 sizeof(struct llog_rec_tail));
 676                 /* get the last record in block */
 677                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 678                                                    le32_to_cpu(tail->lrt_len));
 679
 680                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 681                         lustre_swab_llog_rec(last_rec);
 682                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 683
 684                 /* this shouldn't happen */
 685                 if (tail->lrt_index == 0) {
 686                         CERROR("%s: invalid llog tail at log id "LPU64"/%u "
 687                                "offset "LPU64"\n",
 688                                o->do_lu.lo_dev->ld_obd->obd_name,
 689                                loghandle->lgh_id.lgl_oid,
 690                                loghandle->lgh_id.lgl_ogen, cur_offset);
 691                         GOTO(out, rc = -EINVAL);
 692                 }
 693                 if (tail->lrt_index < prev_idx)
 694                         continue;
 695
 696                 /* sanity check that the start of the new buffer is no farther
 697                  * than the record that we wanted.  This shouldn't happen. */
 698                 if (rec->lrh_index > prev_idx) {
 699                         CERROR("%s: missed desired record? %u > %u\n",
 700                                o->do_lu.lo_dev->ld_obd->obd_name,
 701                                rec->lrh_index, prev_idx);
 702                         GOTO(out, rc = -ENOENT);
 703                 }
 704                 GOTO(out, rc = 0);
 705         }
 706         GOTO(out, rc = -EIO);
 707 out:
 708         return rc;
 709 }
 710
 711 struct dt_object *llog_osd_dir_get(const struct lu_env *env,
 712                                    struct llog_ctxt *ctxt)
 713 {
 714         struct dt_device        *dt;
 715         struct dt_thread_info   *dti = dt_info(env);
 716         struct dt_object        *dir;
 717         int                      rc;
 718
 719         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 720         if (ctxt->loc_dir == NULL) {
 721                 rc = dt_root_get(env, dt, &dti->dti_fid);
 722                 if (rc)
 723                         return ERR_PTR(rc);
 724                 dir = dt_locate(env, dt, &dti->dti_fid);
 725         } else {
 726                 lu_object_get(&ctxt->loc_dir->do_lu);
 727                 dir = ctxt->loc_dir;
 728         }
 729
 730         return dir;
 731 }
 732
 733 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
 734                          struct llog_logid *logid, char *name,
 735                          enum llog_open_param open_param)
 736 {
 737         struct llog_thread_info         *lgi = llog_info(env);
 738         struct llog_ctxt                *ctxt = handle->lgh_ctxt;
 739         struct dt_object                *o;
 740         struct dt_device                *dt;
 741         struct ls_device                *ls;
 742         struct local_oid_storage        *los;
 743         int                              rc = 0;
 744
 745         ENTRY;
 746
 747         LASSERT(env);
 748         LASSERT(ctxt);
 749         LASSERT(ctxt->loc_exp);
 750         LASSERT(ctxt->loc_exp->exp_obd);
 751         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 752         LASSERT(dt);
 753
 754         ls = ls_device_get(dt);
 755         if (IS_ERR(ls))
 756                 RETURN(PTR_ERR(ls));
 757
 758         cfs_mutex_lock(&ls->ls_los_mutex);
 759         los = dt_los_find(ls, FID_SEQ_LLOG);
 760         cfs_mutex_unlock(&ls->ls_los_mutex);
 761         LASSERT(los);
 762         ls_device_put(env, ls);
 763
 764         LASSERT(handle);
 765
 766         if (logid != NULL) {
 767                 logid_to_fid(logid, &lgi->lgi_fid);
 768         } else if (name) {
 769                 struct dt_object *llog_dir;
 770
 771                 llog_dir = llog_osd_dir_get(env, ctxt);
 772                 if (IS_ERR(llog_dir))
 773                         GOTO(out, rc = PTR_ERR(llog_dir));
 774                 dt_read_lock(env, llog_dir, 0);
 775                 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
 776                 dt_read_unlock(env, llog_dir);
 777                 lu_object_put(env, &llog_dir->do_lu);
 778                 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
 779                         /* generate fid for new llog */
 780                         rc = local_object_fid_generate(env, los,
 781                                                        &lgi->lgi_fid);
 782                 }
 783                 if (rc < 0)
 784                         GOTO(out, rc);
 785                 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
 786                 if (handle->lgh_name)
 787                         strcpy(handle->lgh_name, name);
 788                 else
 789                         GOTO(out, rc = -ENOMEM);
 790         } else {
 791                 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
 792                 /* generate fid for new llog */
 793                 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
 794                 if (rc < 0)
 795                         GOTO(out, rc);
 796         }
 797
 798         o = ls_locate(env, ls, &lgi->lgi_fid);
 799         if (IS_ERR(o))
 800                 GOTO(out_name, rc = PTR_ERR(o));
 801
 802         /* No new llog is expected but doesn't exist */
 803         if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
 804                 GOTO(out_put, rc = -ENOENT);
 805
 806         fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
 807         handle->lgh_obj = o;
 808         handle->private_data = los;
 809         LASSERT(handle->lgh_ctxt);
 810
 811         RETURN(rc);
 812
 813 out_put:
 814         lu_object_put(env, &o->do_lu);
 815 out_name:
 816         if (handle->lgh_name != NULL)
 817                 OBD_FREE(handle->lgh_name, strlen(name) + 1);
 818 out:
 819         dt_los_put(los);
 820         RETURN(rc);
 821 }
 822
 823 static int llog_osd_exist(struct llog_handle *handle)
 824 {
 825         LASSERT(handle->lgh_obj);
 826         return (dt_object_exists(handle->lgh_obj) &&
 827                 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
 828 }
 829
 830 static int llog_osd_declare_create(const struct lu_env *env,
 831                                    struct llog_handle *res, struct thandle *th)
 832 {
 833         struct llog_thread_info         *lgi = llog_info(env);
 834         struct local_oid_storage        *los;
 835         struct dt_object                *o;
 836         int                              rc;
 837
 838         ENTRY;
 839
 840         LASSERT(res->lgh_obj);
 841         LASSERT(th);
 842
 843         /* object can be created by another thread */
 844         o = res->lgh_obj;
 845         if (dt_object_exists(o))
 846                 RETURN(0);
 847
 848         los = res->private_data;
 849         LASSERT(los);
 850
 851         rc = llog_osd_declare_new_object(env, los, o, th);
 852         if (rc)
 853                 RETURN(rc);
 854
 855         rc = dt_declare_record_write(env, o, LLOG_CHUNK_SIZE, 0, th);
 856         if (rc)
 857                 RETURN(rc);
 858
 859         if (res->lgh_name) {
 860                 struct dt_object *llog_dir;
 861
 862                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 863                 if (IS_ERR(llog_dir))
 864                         RETURN(PTR_ERR(llog_dir));
 865                 dt_declare_ref_add(env, o, th);
 866                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 867                 rc = dt_declare_insert(env, llog_dir,
 868                                        (struct dt_rec *)&lgi->lgi_fid,
 869                                        (struct dt_key *)res->lgh_name, th);
 870                 lu_object_put(env, &llog_dir->do_lu);
 871                 if (rc)
 872                         CERROR("%s: can't declare named llog %s: rc = %d\n",
 873                                o->do_lu.lo_dev->ld_obd->obd_name,
 874                                res->lgh_name, rc);
 875         }
 876         RETURN(rc);
 877 }
 878
 879 /* This is a callback from the llog_* functions.
 880  * Assumes caller has already pushed us into the kernel context. */
 881 static int llog_osd_create(const struct lu_env *env, struct llog_handle *res,
 882                            struct thandle *th)
 883 {
 884         struct llog_thread_info *lgi = llog_info(env);
 885         struct local_oid_storage *los;
 886         struct dt_object        *o;
 887         int                      rc = 0;
 888
 889         ENTRY;
 890
 891         LASSERT(env);
 892         o = res->lgh_obj;
 893         LASSERT(o);
 894
 895         /* llog can be already created */
 896         if (dt_object_exists(o))
 897                 RETURN(-EEXIST);
 898
 899         los = res->private_data;
 900         LASSERT(los);
 901
 902         dt_write_lock(env, o, 0);
 903         if (!dt_object_exists(o))
 904                 rc = llog_osd_create_new_object(env, los, o, th);
 905         else
 906                 rc = -EEXIST;
 907         if (res->lgh_name)
 908                 dt_ref_add(env, o, th);
 909         dt_write_unlock(env, o);
 910         if (rc)
 911                 RETURN(rc);
 912
 913         if (res->lgh_name) {
 914                 struct dt_object *llog_dir;
 915
 916                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 917                 if (IS_ERR(llog_dir))
 918                         RETURN(PTR_ERR(llog_dir));
 919
 920                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 921                 dt_read_lock(env, llog_dir, 0);
 922                 rc = dt_insert(env, llog_dir,
 923                                (struct dt_rec *)&lgi->lgi_fid,
 924                                (struct dt_key *)res->lgh_name,
 925                                th, BYPASS_CAPA, 1);
 926                 dt_read_unlock(env, llog_dir);
 927                 lu_object_put(env, &llog_dir->do_lu);
 928                 if (rc)
 929                         CERROR("%s: can't create named llog %s: rc = %d\n",
 930                                o->do_lu.lo_dev->ld_obd->obd_name,
 931                                res->lgh_name, rc);
 932         }
 933         RETURN(rc);
 934 }
 935
 936 static int llog_osd_close(const struct lu_env *env, struct llog_handle *handle)
 937 {
 938         struct local_oid_storage        *los;
 939         int                              rc = 0;
 940
 941         ENTRY;
 942
 943         LASSERT(handle->lgh_obj);
 944
 945         lu_object_put(env, &handle->lgh_obj->do_lu);
 946
 947         los = handle->private_data;
 948         LASSERT(los);
 949         dt_los_put(los);
 950
 951         if (handle->lgh_name)
 952                 OBD_FREE(handle->lgh_name, strlen(handle->lgh_name) + 1);
 953
 954         RETURN(rc);
 955 }
 956
 957 static int llog_osd_destroy(const struct lu_env *env,
 958                             struct llog_handle *loghandle)
 959 {
 960         struct llog_thread_info *lgi = llog_info(env);
 961         struct llog_ctxt        *ctxt;
 962         struct dt_object        *o, *llog_dir = NULL;
 963         struct dt_device        *d;
 964         struct thandle          *th;
 965         char                    *name = NULL;
 966         int                      rc;
 967
 968         ENTRY;
 969
 970         ctxt = loghandle->lgh_ctxt;
 971         LASSERT(ctxt);
 972
 973         o = loghandle->lgh_obj;
 974         LASSERT(o);
 975
 976         d = lu2dt_dev(o->do_lu.lo_dev);
 977         LASSERT(d);
 978         LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
 979
 980         th = dt_trans_create(env, d);
 981         if (IS_ERR(th))
 982                 RETURN(PTR_ERR(th));
 983
 984         if (loghandle->lgh_name) {
 985                 llog_dir = llog_osd_dir_get(env, ctxt);
 986                 if (IS_ERR(llog_dir))
 987                         GOTO(out_trans, rc = PTR_ERR(llog_dir));
 988
 989                 dt_declare_ref_del(env, o, th);
 990                 name = loghandle->lgh_name;
 991                 rc = dt_declare_delete(env, llog_dir,
 992                                        (struct dt_key *)name, th);
 993                 if (rc)
 994                         GOTO(out_trans, rc);
 995         }
 996
 997         dt_declare_ref_del(env, o, th);
 998
 999         rc = dt_declare_destroy(env, o, th);
1000         if (rc)
1001                 GOTO(out_trans, rc);
1002
1003         rc = dt_trans_start_local(env, d, th);
1004         if (rc)
1005                 GOTO(out_trans, rc);
1006
1007         dt_write_lock(env, o, 0);
1008         if (dt_object_exists(o)) {
1009                 if (name) {
1010                         dt_ref_del(env, o, th);
1011                         dt_read_lock(env, llog_dir, 0);
1012                         rc = dt_delete(env, llog_dir,
1013                                        (struct dt_key *) name,
1014                                        th, BYPASS_CAPA);
1015                         dt_read_unlock(env, llog_dir);
1016                         if (rc) {
1017                                 CERROR("%s: can't remove llog %s: rc = %d\n",
1018                                        o->do_lu.lo_dev->ld_obd->obd_name,
1019                                        name, rc);
1020                                 GOTO(out_unlock, rc);
1021                         }
1022                 }
1023                 /*
1024                  * XXX: compatibility bits
1025                  *      on old filesystems llogs are referenced by the name
1026                  *      on the new ones they are referenced by OI and by
1027                  *      the name
1028                  */
1029                 rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
1030                 if (rc)
1031                         GOTO(out_unlock, rc);
1032                 LASSERT(lgi->lgi_attr.la_nlink < 2);
1033                 if (lgi->lgi_attr.la_nlink == 1)
1034                         dt_ref_del(env, o, th);
1035                 rc = dt_destroy(env, o, th);
1036                 if (rc)
1037                         GOTO(out_unlock, rc);
1038         }
1039 out_unlock:
1040         dt_write_unlock(env, o);
1041 out_trans:
1042         dt_trans_stop(env, d, th);
1043         if (llog_dir != NULL)
1044                 lu_object_put(env, &llog_dir->do_lu);
1045         RETURN(rc);
1046 }
1047
1048 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1049                           struct obd_llog_group *olg, int ctxt_idx,
1050                           struct obd_device *disk_obd)
1051 {
1052         struct local_oid_storage        *los;
1053         struct llog_thread_info         *lgi = llog_info(env);
1054         struct llog_ctxt                *ctxt;
1055         int                              rc = 0;
1056
1057         ENTRY;
1058
1059         LASSERT(obd);
1060         LASSERT(olg->olg_ctxts[ctxt_idx]);
1061
1062         ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1063         LASSERT(ctxt);
1064
1065         /* initialize data allowing to generate new fids,
1066          * literally we need a sequece */
1067         lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1068         lgi->lgi_fid.f_oid = 1;
1069         lgi->lgi_fid.f_ver = 0;
1070         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1071                                     &lgi->lgi_fid, &los);
1072         llog_ctxt_put(ctxt);
1073         return rc;
1074 }
1075
1076 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1077 {
1078         struct dt_device                *dt;
1079         struct ls_device                *ls;
1080         struct local_oid_storage        *los;
1081
1082         LASSERT(ctxt->loc_exp->exp_obd);
1083         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
1084         ls = ls_device_get(dt);
1085         if (IS_ERR(ls))
1086                 RETURN(PTR_ERR(ls));
1087
1088         cfs_mutex_lock(&ls->ls_los_mutex);
1089         los = dt_los_find(ls, FID_SEQ_LLOG);
1090         cfs_mutex_unlock(&ls->ls_los_mutex);
1091         if (los != NULL) {
1092                 dt_los_put(los);
1093                 local_oid_storage_fini(env, los);
1094         }
1095         ls_device_put(env, ls);
1096         return 0;
1097 }
1098
1099 struct llog_operations llog_osd_ops = {
1100         .lop_next_block         = llog_osd_next_block,
1101         .lop_prev_block         = llog_osd_prev_block,
1102         .lop_read_header        = llog_osd_read_header,
1103         .lop_destroy            = llog_osd_destroy,
1104         .lop_setup              = llog_osd_setup,
1105         .lop_cleanup            = llog_osd_cleanup,
1106         .lop_open               = llog_osd_open,
1107         .lop_exist              = llog_osd_exist,
1108         .lop_declare_create     = llog_osd_declare_create,
1109         .lop_create             = llog_osd_create,
1110         .lop_declare_write_rec  = llog_osd_declare_write_rec,
1111         .lop_write_rec          = llog_osd_write_rec,
1112         .lop_close              = llog_osd_close,
1113 };
1114 EXPORT_SYMBOL(llog_osd_ops);
1115
1116 /* reads the catalog list */
1117 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1118                           int idx, int count, struct llog_catid *idarray)
1119 {
1120         struct llog_thread_info *lgi = llog_info(env);
1121         struct dt_object        *o = NULL;
1122         struct thandle          *th;
1123         int                      rc, size;
1124
1125         ENTRY;
1126
1127         LASSERT(d);
1128
1129         size = sizeof(*idarray) * count;
1130         lgi->lgi_off = idx *  sizeof(*idarray);
1131
1132         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1133         o = dt_locate(env, d, &lgi->lgi_fid);
1134         if (IS_ERR(o))
1135                 RETURN(PTR_ERR(o));
1136
1137         if (!dt_object_exists(o)) {
1138                 th = dt_trans_create(env, d);
1139                 if (IS_ERR(th))
1140                         GOTO(out, rc = PTR_ERR(th));
1141
1142                 lgi->lgi_attr.la_valid = LA_MODE;
1143                 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1144                 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1145
1146                 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1147                                        &lgi->lgi_dof, th);
1148                 if (rc)
1149                         GOTO(out_trans, rc);
1150
1151                 rc = dt_trans_start_local(env, d, th);
1152                 if (rc)
1153                         GOTO(out_trans, rc);
1154
1155                 dt_write_lock(env, o, 0);
1156                 if (!dt_object_exists(o))
1157                         rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1158                                        &lgi->lgi_dof, th);
1159                 dt_write_unlock(env, o);
1160 out_trans:
1161                 dt_trans_stop(env, d, th);
1162                 if (rc)
1163                         GOTO(out, rc);
1164         }
1165
1166         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1167         if (rc)
1168                 GOTO(out, rc);
1169
1170         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1171                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1172                        o->do_lu.lo_dev->ld_obd->obd_name,
1173                        lgi->lgi_attr.la_mode);
1174                 GOTO(out, rc = -ENOENT);
1175         }
1176
1177         CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1178                (int)lgi->lgi_attr.la_size, size);
1179
1180         /* return just number of llogs */
1181         if (idarray == NULL) {
1182                 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1183                 GOTO(out, rc);
1184         }
1185
1186         /* read for new ost index or for empty file */
1187         memset(idarray, 0, size);
1188         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1189                 GOTO(out, rc = 0);
1190         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1191                 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1192
1193         lgi->lgi_buf.lb_buf = idarray;
1194         lgi->lgi_buf.lb_len = size;
1195         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1196         if (rc) {
1197                 CERROR("%s: error reading CATALOGS: rc = %d\n",
1198                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
1199                 GOTO(out, rc);
1200         }
1201
1202         EXIT;
1203 out:
1204         lu_object_put(env, &o->do_lu);
1205         RETURN(rc);
1206 }
1207 EXPORT_SYMBOL(llog_osd_get_cat_list);
1208
1209 /* writes the cat list */
1210 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1211                           int idx, int count, struct llog_catid *idarray)
1212 {
1213         struct llog_thread_info *lgi = llog_info(env);
1214         struct dt_object        *o = NULL;
1215         struct thandle          *th;
1216         int                      rc, size;
1217
1218         if (!count)
1219                 RETURN(0);
1220
1221         LASSERT(d);
1222
1223         size = sizeof(*idarray) * count;
1224         lgi->lgi_off = idx * sizeof(*idarray);
1225
1226         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1227         o = dt_locate(env, d, &lgi->lgi_fid);
1228         if (IS_ERR(o))
1229                 RETURN(PTR_ERR(o));
1230
1231         if (!dt_object_exists(o))
1232                 GOTO(out, rc = -ENOENT);
1233
1234         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1235         if (rc)
1236                 GOTO(out, rc);
1237
1238         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1239                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1240                        o->do_lu.lo_dev->ld_obd->obd_name,
1241                        lgi->lgi_attr.la_mode);
1242                 GOTO(out, rc = -ENOENT);
1243         }
1244
1245         th = dt_trans_create(env, d);
1246         if (IS_ERR(th))
1247                 GOTO(out, rc = PTR_ERR(th));
1248
1249         rc = dt_declare_record_write(env, o, size, lgi->lgi_off, th);
1250         if (rc)
1251                 GOTO(out, rc);
1252
1253         rc = dt_trans_start_local(env, d, th);
1254         if (rc)
1255                 GOTO(out_trans, rc);
1256
1257         lgi->lgi_buf.lb_buf = idarray;
1258         lgi->lgi_buf.lb_len = size;
1259         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1260         if (rc)
1261                 CDEBUG(D_INODE, "error writeing CATALOGS: rc = %d\n", rc);
1262 out_trans:
1263         dt_trans_stop(env, d, th);
1264 out:
1265         lu_object_put(env, &o->do_lu);
1266         RETURN(rc);
1267 }
1268 EXPORT_SYMBOL(llog_osd_put_cat_list);