lustre/obdclass/llog_osd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/llog_osd.c - low level llog routines on top of OSD API
  37  *
  38  * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
  39  * Author: Mikhail Pershin <mike.pershin@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LOG
  43
  44 #include <obd.h>
  45 #include <obd_class.h>
  46 #include <lustre_fid.h>
  47 #include <dt_object.h>
  48
  49 #include "llog_internal.h"
  50 #include "local_storage.h"
  51
  52 /*
  53  * - multi-chunks or big-declaration approach
  54  * - use unique sequence instead of llog sb tracking unique ids
  55  * - re-use existing environment
  56  * - named llog support (can be used for testing only at the present)
  57  * - llog_origin_connect() work with OSD API
  58  */
  59
  60 static int llog_osd_declare_new_object(const struct lu_env *env,
  61                                        struct local_oid_storage *los,
  62                                        struct dt_object *o,
  63                                        struct thandle *th)
  64 {
  65         struct llog_thread_info *lgi = llog_info(env);
  66
  67         lgi->lgi_attr.la_valid = LA_MODE;
  68         lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
  69         lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
  70
  71         return local_object_declare_create(env, los, o, &lgi->lgi_attr,
  72                                            &lgi->lgi_dof, th);
  73 }
  74
  75 static int llog_osd_create_new_object(const struct lu_env *env,
  76                                       struct local_oid_storage *los,
  77                                       struct dt_object *o,
  78                                       struct thandle *th)
  79 {
  80         struct llog_thread_info *lgi = llog_info(env);
  81
  82         lgi->lgi_attr.la_valid = LA_MODE;
  83         lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
  84         lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
  85
  86         return local_object_create(env, los, o, &lgi->lgi_attr,
  87                                    &lgi->lgi_dof, th);
  88 }
  89
  90 static int llog_osd_pad(const struct lu_env *env, struct dt_object *o,
  91                         loff_t *off, int len, int index, struct thandle *th)
  92 {
  93         struct llog_thread_info *lgi = llog_info(env);
  94         int                      rc;
  95
  96         ENTRY;
  97
  98         LASSERT(th);
  99         LASSERT(off);
 100         LASSERT(len >= LLOG_MIN_REC_SIZE && (len & 0x7) == 0);
 101
 102         lgi->lgi_tail.lrt_len = lgi->lgi_lrh.lrh_len = len;
 103         lgi->lgi_tail.lrt_index = lgi->lgi_lrh.lrh_index = index;
 104         lgi->lgi_lrh.lrh_type = LLOG_PAD_MAGIC;
 105
 106         lgi->lgi_buf.lb_buf = &lgi->lgi_lrh;
 107         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_lrh);
 108         dt_write_lock(env, o, 0);
 109         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 110         if (rc) {
 111                 CERROR("%s: error writing padding record: rc = %d\n",
 112                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 113                 GOTO(out, rc);
 114         }
 115
 116         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 117         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 118         *off += len - sizeof(lgi->lgi_lrh) - sizeof(lgi->lgi_tail);
 119         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 120         if (rc)
 121                 CERROR("%s: error writing padding record: rc = %d\n",
 122                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 123 out:
 124         dt_write_unlock(env, o);
 125         RETURN(rc);
 126 }
 127
 128 static int llog_osd_write_blob(const struct lu_env *env, struct dt_object *o,
 129                                struct llog_rec_hdr *rec, void *buf,
 130                                loff_t *off, struct thandle *th)
 131 {
 132         struct llog_thread_info *lgi = llog_info(env);
 133         int                      buflen = rec->lrh_len;
 134         int                      rc;
 135
 136         ENTRY;
 137
 138         LASSERT(env);
 139         LASSERT(o);
 140
 141         if (buflen == 0)
 142                 CWARN("0-length record\n");
 143
 144         CDEBUG(D_OTHER, "write blob with type %x, buf %p/%u at off %llu\n",
 145                rec->lrh_type, buf, buflen, *off);
 146
 147         lgi->lgi_attr.la_valid = LA_SIZE;
 148         lgi->lgi_attr.la_size = *off;
 149
 150         if (!buf) {
 151                 lgi->lgi_buf.lb_len = buflen;
 152                 lgi->lgi_buf.lb_buf = rec;
 153                 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 154                 if (rc)
 155                         CERROR("%s: error writing log record: rc = %d\n",
 156                                o->do_lu.lo_dev->ld_obd->obd_name, rc);
 157                 GOTO(out, rc);
 158         }
 159
 160         /* the buf case */
 161         /* protect the following 3 writes from concurrent read */
 162         dt_write_lock(env, o, 0);
 163         rec->lrh_len = sizeof(*rec) + buflen + sizeof(lgi->lgi_tail);
 164         lgi->lgi_buf.lb_len = sizeof(*rec);
 165         lgi->lgi_buf.lb_buf = rec;
 166         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 167         if (rc) {
 168                 CERROR("%s: error writing log hdr: rc = %d\n",
 169                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 170                 GOTO(out_unlock, rc);
 171         }
 172
 173         lgi->lgi_buf.lb_len = buflen;
 174         lgi->lgi_buf.lb_buf = buf;
 175         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 176         if (rc) {
 177                 CERROR("%s: error writing log buffer: rc = %d\n",
 178                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
 179                 GOTO(out_unlock, rc);
 180         }
 181
 182         lgi->lgi_tail.lrt_len = rec->lrh_len;
 183         lgi->lgi_tail.lrt_index = rec->lrh_index;
 184         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 185         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 186         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 187         if (rc)
 188                 CERROR("%s: error writing log tail: rc = %d\n",
 189                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 190
 191 out_unlock:
 192         dt_write_unlock(env, o);
 193
 194 out:
 195         /* cleanup the content written above */
 196         if (rc) {
 197                 dt_punch(env, o, lgi->lgi_attr.la_size, OBD_OBJECT_EOF, th,
 198                          BYPASS_CAPA);
 199                 dt_attr_set(env, o, &lgi->lgi_attr, th, BYPASS_CAPA);
 200         }
 201
 202         RETURN(rc);
 203 }
 204
 205 static int llog_osd_read_header(const struct lu_env *env,
 206                                 struct llog_handle *handle)
 207 {
 208         struct llog_rec_hdr     *llh_hdr;
 209         struct dt_object        *o;
 210         struct llog_thread_info *lgi;
 211         int                      rc;
 212
 213         ENTRY;
 214
 215         LASSERT(sizeof(*handle->lgh_hdr) == LLOG_CHUNK_SIZE);
 216
 217         o = handle->lgh_obj;
 218         LASSERT(o);
 219
 220         lgi = llog_info(env);
 221
 222         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 223         if (rc)
 224                 RETURN(rc);
 225
 226         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 227
 228         if (lgi->lgi_attr.la_size == 0) {
 229                 CDEBUG(D_HA, "not reading header from 0-byte log\n");
 230                 RETURN(LLOG_EEMPTY);
 231         }
 232
 233         lgi->lgi_off = 0;
 234         lgi->lgi_buf.lb_buf = handle->lgh_hdr;
 235         lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE;
 236
 237         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
 238         if (rc) {
 239                 CERROR("%s: error reading log header from "DFID": rc = %d\n",
 240                        o->do_lu.lo_dev->ld_obd->obd_name,
 241                        PFID(lu_object_fid(&o->do_lu)), rc);
 242                 RETURN(rc);
 243         }
 244
 245         llh_hdr = &handle->lgh_hdr->llh_hdr;
 246         if (LLOG_REC_HDR_NEEDS_SWABBING(llh_hdr))
 247                 lustre_swab_llog_hdr(handle->lgh_hdr);
 248
 249         if (llh_hdr->lrh_type != LLOG_HDR_MAGIC) {
 250                 CERROR("%s: bad log %s "DFID" header magic: %#x "
 251                        "(expected %#x)\n", o->do_lu.lo_dev->ld_obd->obd_name,
 252                        handle->lgh_name ? handle->lgh_name : "",
 253                        PFID(lu_object_fid(&o->do_lu)),
 254                        llh_hdr->lrh_type, LLOG_HDR_MAGIC);
 255                 RETURN(-EIO);
 256         } else if (llh_hdr->lrh_len != LLOG_CHUNK_SIZE) {
 257                 CERROR("%s: incorrectly sized log %s "DFID" header: "
 258                        "%#x (expected %#x)\n"
 259                        "you may need to re-run lconf --write_conf.\n",
 260                        o->do_lu.lo_dev->ld_obd->obd_name,
 261                        handle->lgh_name ? handle->lgh_name : "",
 262                        PFID(lu_object_fid(&o->do_lu)),
 263                        llh_hdr->lrh_len, LLOG_CHUNK_SIZE);
 264                 RETURN(-EIO);
 265         }
 266
 267         handle->lgh_last_idx = handle->lgh_hdr->llh_tail.lrt_index;
 268
 269         RETURN(0);
 270 }
 271
 272 static int llog_osd_declare_write_rec(const struct lu_env *env,
 273                                       struct llog_handle *loghandle,
 274                                       struct llog_rec_hdr *rec,
 275                                       int idx, struct thandle *th)
 276 {
 277         struct llog_thread_info *lgi = llog_info(env);
 278         struct dt_object        *o;
 279         int                      rc;
 280
 281         ENTRY;
 282
 283         LASSERT(env);
 284         LASSERT(th);
 285         LASSERT(loghandle);
 286
 287         o = loghandle->lgh_obj;
 288         LASSERT(o);
 289
 290         /* each time we update header */
 291         rc = dt_declare_record_write(env, o, sizeof(struct llog_log_hdr), 0,
 292                                      th);
 293         if (rc || idx == 0) /* if error or just header */
 294                 RETURN(rc);
 295
 296         if (dt_object_exists(o)) {
 297                 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 298                 lgi->lgi_off = lgi->lgi_attr.la_size;
 299                 LASSERT(ergo(rc == 0, lgi->lgi_attr.la_valid & LA_SIZE));
 300                 if (rc)
 301                         RETURN(rc);
 302
 303                 rc = dt_declare_punch(env, o, lgi->lgi_off, OBD_OBJECT_EOF, th);
 304                 if (rc)
 305                         RETURN(rc);
 306         } else {
 307                 lgi->lgi_off = 0;
 308         }
 309
 310         /* XXX: implement declared window or multi-chunks approach */
 311         rc = dt_declare_record_write(env, o, 32 * 1024, lgi->lgi_off, th);
 312
 313         RETURN(rc);
 314 }
 315
 316 /* returns negative in on error; 0 if success && reccookie == 0; 1 otherwise */
 317 /* appends if idx == -1, otherwise overwrites record idx. */
 318 static int llog_osd_write_rec(const struct lu_env *env,
 319                               struct llog_handle *loghandle,
 320                               struct llog_rec_hdr *rec,
 321                               struct llog_cookie *reccookie, int cookiecount,
 322                               void *buf, int idx, struct thandle *th)
 323 {
 324         struct llog_thread_info *lgi = llog_info(env);
 325         struct llog_log_hdr     *llh;
 326         int                      reclen = rec->lrh_len;
 327         int                      index, rc, old_tail_idx;
 328         struct llog_rec_tail    *lrt;
 329         struct dt_object        *o;
 330         size_t                   left;
 331
 332         ENTRY;
 333
 334         LASSERT(env);
 335         llh = loghandle->lgh_hdr;
 336         LASSERT(llh);
 337         o = loghandle->lgh_obj;
 338         LASSERT(o);
 339         LASSERT(th);
 340
 341         CDEBUG(D_OTHER, "new record %x to "DFID"\n",
 342                rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
 343
 344         /* record length should not bigger than LLOG_CHUNK_SIZE */
 345         if (buf)
 346                 rc = (reclen > LLOG_CHUNK_SIZE - sizeof(struct llog_rec_hdr) -
 347                       sizeof(struct llog_rec_tail)) ? -E2BIG : 0;
 348         else
 349                 rc = (reclen > LLOG_CHUNK_SIZE) ? -E2BIG : 0;
 350         if (rc)
 351                 RETURN(rc);
 352
 353         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 354         if (rc)
 355                 RETURN(rc);
 356
 357         if (buf)
 358                 /* write_blob adds header and tail to lrh_len. */
 359                 reclen = sizeof(*rec) + rec->lrh_len +
 360                          sizeof(struct llog_rec_tail);
 361
 362         if (idx != -1) {
 363                 /* no header: only allowed to insert record 1 */
 364                 if (idx != 1 && lgi->lgi_attr.la_size == 0)
 365                         LBUG();
 366
 367                 if (idx && llh->llh_size && llh->llh_size != rec->lrh_len)
 368                         RETURN(-EINVAL);
 369
 370                 if (!ext2_test_bit(idx, llh->llh_bitmap))
 371                         CERROR("%s: modify unset record %u\n",
 372                                o->do_lu.lo_dev->ld_obd->obd_name, idx);
 373                 if (idx != rec->lrh_index)
 374                         CERROR("%s: index mismatch %d %u\n",
 375                                o->do_lu.lo_dev->ld_obd->obd_name, idx,
 376                                rec->lrh_index);
 377
 378                 lgi->lgi_off = 0;
 379                 rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 380                                          &lgi->lgi_off, th);
 381                 /* we are done if we only write the header or on error */
 382                 if (rc || idx == 0)
 383                         RETURN(rc);
 384
 385                 if (buf) {
 386                         /* We assume that caller has set lgh_cur_* */
 387                         lgi->lgi_off = loghandle->lgh_cur_offset;
 388                         CDEBUG(D_OTHER,
 389                                "modify record "DOSTID": idx:%d/%u/%d, len:%u "
 390                                "offset %llu\n",
 391                                POSTID(&loghandle->lgh_id.lgl_oi), idx,
 392                                rec->lrh_index,
 393                                loghandle->lgh_cur_idx, rec->lrh_len,
 394                                (long long)(lgi->lgi_off - sizeof(*llh)));
 395                         if (rec->lrh_index != loghandle->lgh_cur_idx) {
 396                                 CERROR("%s: modify idx mismatch %u/%d\n",
 397                                        o->do_lu.lo_dev->ld_obd->obd_name, idx,
 398                                        loghandle->lgh_cur_idx);
 399                                 RETURN(-EFAULT);
 400                         }
 401                 } else {
 402                         /* Assumes constant lrh_len */
 403                         lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
 404                 }
 405
 406                 rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 407                 if (rc == 0 && reccookie) {
 408                         reccookie->lgc_lgl = loghandle->lgh_id;
 409                         reccookie->lgc_index = idx;
 410                         rc = 1;
 411                 }
 412                 RETURN(rc);
 413         }
 414
 415         /* Make sure that records don't cross a chunk boundary, so we can
 416          * process them page-at-a-time if needed.  If it will cross a chunk
 417          * boundary, write in a fake (but referenced) entry to pad the chunk.
 418          *
 419          * We know that llog_current_log() will return a loghandle that is
 420          * big enough to hold reclen, so all we care about is padding here.
 421          */
 422         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 423         lgi->lgi_off = lgi->lgi_attr.la_size;
 424         left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
 425         /* NOTE: padding is a record, but no bit is set */
 426         if (left != 0 && left != reclen &&
 427             left < (reclen + LLOG_MIN_REC_SIZE)) {
 428                 index = loghandle->lgh_last_idx + 1;
 429                 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
 430                 if (rc)
 431                         RETURN(rc);
 432                 loghandle->lgh_last_idx++; /*for pad rec*/
 433         }
 434         /* if it's the last idx in log file, then return -ENOSPC */
 435         if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
 436                 RETURN(-ENOSPC);
 437
 438         loghandle->lgh_last_idx++;
 439         index = loghandle->lgh_last_idx;
 440         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 441         rec->lrh_index = index;
 442         if (buf == NULL) {
 443                 lrt = (struct llog_rec_tail *)((char *)rec + rec->lrh_len -
 444                                                sizeof(*lrt));
 445                 lrt->lrt_len = rec->lrh_len;
 446                 lrt->lrt_index = rec->lrh_index;
 447         }
 448         /* The caller should make sure only 1 process access the lgh_last_idx,
 449          * Otherwise it might hit the assert.*/
 450         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 451         spin_lock(&loghandle->lgh_hdr_lock);
 452         if (ext2_set_bit(index, llh->llh_bitmap)) {
 453                 CERROR("%s: index %u already set in log bitmap\n",
 454                        o->do_lu.lo_dev->ld_obd->obd_name, index);
 455                 spin_unlock(&loghandle->lgh_hdr_lock);
 456                 LBUG(); /* should never happen */
 457         }
 458         llh->llh_count++;
 459         spin_unlock(&loghandle->lgh_hdr_lock);
 460         old_tail_idx = llh->llh_tail.lrt_index;
 461         llh->llh_tail.lrt_index = index;
 462
 463         lgi->lgi_off = 0;
 464         rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL, &lgi->lgi_off,
 465                                  th);
 466         if (rc)
 467                 GOTO(out, rc);
 468
 469         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 470         if (rc)
 471                 GOTO(out, rc);
 472
 473         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 474         lgi->lgi_off = lgi->lgi_attr.la_size;
 475
 476         rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 477
 478 out:
 479         /* cleanup llog for error case */
 480         if (rc) {
 481                 spin_lock(&loghandle->lgh_hdr_lock);
 482                 ext2_clear_bit(index, llh->llh_bitmap);
 483                 llh->llh_count--;
 484                 spin_unlock(&loghandle->lgh_hdr_lock);
 485
 486                 /* restore the header */
 487                 loghandle->lgh_last_idx--;
 488                 llh->llh_tail.lrt_index = old_tail_idx;
 489                 lgi->lgi_off = 0;
 490                 llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 491                                     &lgi->lgi_off, th);
 492         }
 493
 494         CDEBUG(D_RPCTRACE, "added record "DOSTID": idx: %u, %u\n",
 495                POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len);
 496         if (rc == 0 && reccookie) {
 497                 reccookie->lgc_lgl = loghandle->lgh_id;
 498                 reccookie->lgc_index = index;
 499                 if ((rec->lrh_type == MDS_UNLINK_REC) ||
 500                     (rec->lrh_type == MDS_SETATTR64_REC))
 501                         reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 502                 else if (rec->lrh_type == OST_SZ_REC)
 503                         reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
 504                 else
 505                         reccookie->lgc_subsys = -1;
 506                 rc = 1;
 507         }
 508         RETURN(rc);
 509 }
 510
 511 /* We can skip reading at least as many log blocks as the number of
 512  * minimum sized log records we are skipping.  If it turns out
 513  * that we are not far enough along the log (because the
 514  * actual records are larger than minimum size) we just skip
 515  * some more records.
 516  */
 517 static void llog_skip_over(__u64 *off, int curr, int goal)
 518 {
 519         if (goal <= curr)
 520                 return;
 521         *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
 522                 ~(LLOG_CHUNK_SIZE - 1);
 523 }
 524
 525 /* sets:
 526  *  - cur_offset to the furthest point read in the log file
 527  *  - cur_idx to the log index preceeding cur_offset
 528  * returns -EIO/-EINVAL on error
 529  */
 530 static int llog_osd_next_block(const struct lu_env *env,
 531                                struct llog_handle *loghandle, int *cur_idx,
 532                                int next_idx, __u64 *cur_offset, void *buf,
 533                                int len)
 534 {
 535         struct llog_thread_info *lgi = llog_info(env);
 536         struct dt_object        *o;
 537         struct dt_device        *dt;
 538         int                      rc;
 539
 540         ENTRY;
 541
 542         LASSERT(env);
 543         LASSERT(lgi);
 544
 545         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 546                 RETURN(-EINVAL);
 547
 548         CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
 549                next_idx, *cur_idx, *cur_offset);
 550
 551         LASSERT(loghandle);
 552         LASSERT(loghandle->lgh_ctxt);
 553
 554         o = loghandle->lgh_obj;
 555         LASSERT(o);
 556         LASSERT(dt_object_exists(o));
 557         dt = lu2dt_dev(o->do_lu.lo_dev);
 558         LASSERT(dt);
 559
 560         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 561         if (rc)
 562                 GOTO(out, rc);
 563
 564         while (*cur_offset < lgi->lgi_attr.la_size) {
 565                 struct llog_rec_hdr     *rec, *last_rec;
 566                 struct llog_rec_tail    *tail;
 567
 568                 llog_skip_over(cur_offset, *cur_idx, next_idx);
 569
 570                 /* read up to next LLOG_CHUNK_SIZE block */
 571                 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
 572                                       (*cur_offset & (LLOG_CHUNK_SIZE - 1));
 573                 lgi->lgi_buf.lb_buf = buf;
 574
 575                 /* Note: read lock is not needed around la_size get above at
 576                  * the time of dt_attr_get(). There are only two cases that
 577                  * matter. Either la_size == cur_offset, in which case the
 578                  * entire read is skipped, or la_size > cur_offset and the loop
 579                  * is entered and this thread is blocked at dt_read_lock()
 580                  * until the write is completed. When the write completes, then
 581                  * the dt_read() will be done with the full length, and will
 582                  * get the full data.
 583                  */
 584                 dt_read_lock(env, o, 0);
 585                 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
 586                 dt_read_unlock(env, o);
 587                 if (rc < 0) {
 588                         CERROR("%s: can't read llog block from log "DFID
 589                                " offset "LPU64": rc = %d\n",
 590                                o->do_lu.lo_dev->ld_obd->obd_name,
 591                                PFID(lu_object_fid(&o->do_lu)), *cur_offset,
 592                                rc);
 593                         GOTO(out, rc);
 594                 }
 595
 596                 if (rc < len) {
 597                         /* signal the end of the valid buffer to
 598                          * llog_process */
 599                         memset(buf + rc, 0, len - rc);
 600                 }
 601
 602                 if (rc == 0) /* end of file, nothing to do */
 603                         GOTO(out, rc);
 604
 605                 if (rc < sizeof(*tail)) {
 606                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 607                                "offset "LPU64"\n",
 608                                o->do_lu.lo_dev->ld_obd->obd_name,
 609                                POSTID(&loghandle->lgh_id.lgl_oi),
 610                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 611                         GOTO(out, rc = -EINVAL);
 612                 }
 613
 614                 rec = buf;
 615                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 616                         lustre_swab_llog_rec(rec);
 617
 618                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 619                                                 sizeof(struct llog_rec_tail));
 620                 /* get the last record in block */
 621                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 622                                                    le32_to_cpu(tail->lrt_len));
 623
 624                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 625                         lustre_swab_llog_rec(last_rec);
 626                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 627
 628                 *cur_idx = tail->lrt_index;
 629
 630                 /* this shouldn't happen */
 631                 if (tail->lrt_index == 0) {
 632                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 633                                "offset "LPU64"\n",
 634                                o->do_lu.lo_dev->ld_obd->obd_name,
 635                                POSTID(&loghandle->lgh_id.lgl_oi),
 636                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 637                         GOTO(out, rc = -EINVAL);
 638                 }
 639                 if (tail->lrt_index < next_idx)
 640                         continue;
 641
 642                 /* sanity check that the start of the new buffer is no farther
 643                  * than the record that we wanted.  This shouldn't happen. */
 644                 if (rec->lrh_index > next_idx) {
 645                         CERROR("%s: missed desired record? %u > %u\n",
 646                                o->do_lu.lo_dev->ld_obd->obd_name,
 647                                rec->lrh_index, next_idx);
 648                         GOTO(out, rc = -ENOENT);
 649                 }
 650                 GOTO(out, rc = 0);
 651         }
 652         GOTO(out, rc = -EIO);
 653 out:
 654         return rc;
 655 }
 656
 657 static int llog_osd_prev_block(const struct lu_env *env,
 658                                struct llog_handle *loghandle,
 659                                int prev_idx, void *buf, int len)
 660 {
 661         struct llog_thread_info *lgi = llog_info(env);
 662         struct dt_object        *o;
 663         struct dt_device        *dt;
 664         loff_t                   cur_offset;
 665         int                      rc;
 666
 667         ENTRY;
 668
 669         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 670                 RETURN(-EINVAL);
 671
 672         CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
 673
 674         LASSERT(loghandle);
 675         LASSERT(loghandle->lgh_ctxt);
 676
 677         o = loghandle->lgh_obj;
 678         LASSERT(o);
 679         LASSERT(dt_object_exists(o));
 680         dt = lu2dt_dev(o->do_lu.lo_dev);
 681         LASSERT(dt);
 682
 683         cur_offset = LLOG_CHUNK_SIZE;
 684         llog_skip_over(&cur_offset, 0, prev_idx);
 685
 686         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 687         if (rc)
 688                 GOTO(out, rc);
 689
 690         while (cur_offset < lgi->lgi_attr.la_size) {
 691                 struct llog_rec_hdr     *rec, *last_rec;
 692                 struct llog_rec_tail    *tail;
 693
 694                 lgi->lgi_buf.lb_len = len;
 695                 lgi->lgi_buf.lb_buf = buf;
 696                 /* It is OK to have locking around dt_read() only, see
 697                  * comment in llog_osd_next_block for details
 698                  */
 699                 dt_read_lock(env, o, 0);
 700                 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
 701                 dt_read_unlock(env, o);
 702                 if (rc < 0) {
 703                         CERROR("%s: can't read llog block from log "DFID
 704                                " offset "LPU64": rc = %d\n",
 705                                o->do_lu.lo_dev->ld_obd->obd_name,
 706                                PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
 707                         GOTO(out, rc);
 708                 }
 709
 710                 if (rc == 0) /* end of file, nothing to do */
 711                         GOTO(out, rc);
 712
 713                 if (rc < sizeof(*tail)) {
 714                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 715                                "offset "LPU64"\n",
 716                                o->do_lu.lo_dev->ld_obd->obd_name,
 717                                POSTID(&loghandle->lgh_id.lgl_oi),
 718                                loghandle->lgh_id.lgl_ogen, cur_offset);
 719                         GOTO(out, rc = -EINVAL);
 720                 }
 721
 722                 rec = buf;
 723                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 724                         lustre_swab_llog_rec(rec);
 725
 726                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 727                                                 sizeof(struct llog_rec_tail));
 728                 /* get the last record in block */
 729                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 730                                                    le32_to_cpu(tail->lrt_len));
 731
 732                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 733                         lustre_swab_llog_rec(last_rec);
 734                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 735
 736                 /* this shouldn't happen */
 737                 if (tail->lrt_index == 0) {
 738                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 739                                "offset "LPU64"\n",
 740                                o->do_lu.lo_dev->ld_obd->obd_name,
 741                                POSTID(&loghandle->lgh_id.lgl_oi),
 742                                loghandle->lgh_id.lgl_ogen, cur_offset);
 743                         GOTO(out, rc = -EINVAL);
 744                 }
 745                 if (tail->lrt_index < prev_idx)
 746                         continue;
 747
 748                 /* sanity check that the start of the new buffer is no farther
 749                  * than the record that we wanted.  This shouldn't happen. */
 750                 if (rec->lrh_index > prev_idx) {
 751                         CERROR("%s: missed desired record? %u > %u\n",
 752                                o->do_lu.lo_dev->ld_obd->obd_name,
 753                                rec->lrh_index, prev_idx);
 754                         GOTO(out, rc = -ENOENT);
 755                 }
 756                 GOTO(out, rc = 0);
 757         }
 758         GOTO(out, rc = -EIO);
 759 out:
 760         return rc;
 761 }
 762
 763 struct dt_object *llog_osd_dir_get(const struct lu_env *env,
 764                                    struct llog_ctxt *ctxt)
 765 {
 766         struct dt_device        *dt;
 767         struct dt_thread_info   *dti = dt_info(env);
 768         struct dt_object        *dir;
 769         int                      rc;
 770
 771         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 772         if (ctxt->loc_dir == NULL) {
 773                 rc = dt_root_get(env, dt, &dti->dti_fid);
 774                 if (rc)
 775                         return ERR_PTR(rc);
 776                 dir = dt_locate(env, dt, &dti->dti_fid);
 777
 778                 if (!IS_ERR(dir) && !dt_try_as_dir(env, dir)) {
 779                         lu_object_put(env, &dir->do_lu);
 780                         return ERR_PTR(-ENOTDIR);
 781                 }
 782         } else {
 783                 lu_object_get(&ctxt->loc_dir->do_lu);
 784                 dir = ctxt->loc_dir;
 785         }
 786
 787         return dir;
 788 }
 789
 790 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
 791                          struct llog_logid *logid, char *name,
 792                          enum llog_open_param open_param)
 793 {
 794         struct llog_thread_info         *lgi = llog_info(env);
 795         struct llog_ctxt                *ctxt = handle->lgh_ctxt;
 796         struct dt_object                *o;
 797         struct dt_device                *dt;
 798         struct ls_device                *ls;
 799         struct local_oid_storage        *los;
 800         int                              rc = 0;
 801
 802         ENTRY;
 803
 804         LASSERT(env);
 805         LASSERT(ctxt);
 806         LASSERT(ctxt->loc_exp);
 807         LASSERT(ctxt->loc_exp->exp_obd);
 808         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 809         LASSERT(dt);
 810
 811         ls = ls_device_get(dt);
 812         if (IS_ERR(ls))
 813                 RETURN(PTR_ERR(ls));
 814
 815         mutex_lock(&ls->ls_los_mutex);
 816         los = dt_los_find(ls, name != NULL ? FID_SEQ_LLOG_NAME : FID_SEQ_LLOG);
 817         mutex_unlock(&ls->ls_los_mutex);
 818         LASSERT(los);
 819         ls_device_put(env, ls);
 820
 821         LASSERT(handle);
 822
 823         if (logid != NULL) {
 824                 logid_to_fid(logid, &lgi->lgi_fid);
 825         } else if (name) {
 826                 struct dt_object *llog_dir;
 827
 828                 llog_dir = llog_osd_dir_get(env, ctxt);
 829                 if (IS_ERR(llog_dir))
 830                         GOTO(out, rc = PTR_ERR(llog_dir));
 831                 dt_read_lock(env, llog_dir, 0);
 832                 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
 833                 dt_read_unlock(env, llog_dir);
 834                 lu_object_put(env, &llog_dir->do_lu);
 835                 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
 836                         /* generate fid for new llog */
 837                         rc = local_object_fid_generate(env, los,
 838                                                        &lgi->lgi_fid);
 839                 }
 840                 if (rc < 0)
 841                         GOTO(out, rc);
 842                 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
 843                 if (handle->lgh_name)
 844                         strcpy(handle->lgh_name, name);
 845                 else
 846                         GOTO(out, rc = -ENOMEM);
 847         } else {
 848                 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
 849                 /* generate fid for new llog */
 850                 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
 851                 if (rc < 0)
 852                         GOTO(out, rc);
 853         }
 854
 855         o = ls_locate(env, ls, &lgi->lgi_fid);
 856         if (IS_ERR(o))
 857                 GOTO(out_name, rc = PTR_ERR(o));
 858
 859         /* No new llog is expected but doesn't exist */
 860         if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
 861                 GOTO(out_put, rc = -ENOENT);
 862
 863         fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
 864         handle->lgh_obj = o;
 865         handle->private_data = los;
 866         LASSERT(handle->lgh_ctxt);
 867
 868         RETURN(rc);
 869
 870 out_put:
 871         lu_object_put(env, &o->do_lu);
 872 out_name:
 873         if (handle->lgh_name != NULL)
 874                 OBD_FREE(handle->lgh_name, strlen(name) + 1);
 875 out:
 876         dt_los_put(los);
 877         RETURN(rc);
 878 }
 879
 880 static int llog_osd_exist(struct llog_handle *handle)
 881 {
 882         LASSERT(handle->lgh_obj);
 883         return (dt_object_exists(handle->lgh_obj) &&
 884                 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
 885 }
 886
 887 static int llog_osd_declare_create(const struct lu_env *env,
 888                                    struct llog_handle *res, struct thandle *th)
 889 {
 890         struct llog_thread_info         *lgi = llog_info(env);
 891         struct local_oid_storage        *los;
 892         struct dt_object                *o;
 893         int                              rc;
 894
 895         ENTRY;
 896
 897         LASSERT(res->lgh_obj);
 898         LASSERT(th);
 899
 900         /* object can be created by another thread */
 901         o = res->lgh_obj;
 902         if (dt_object_exists(o))
 903                 RETURN(0);
 904
 905         los = res->private_data;
 906         LASSERT(los);
 907
 908         rc = llog_osd_declare_new_object(env, los, o, th);
 909         if (rc)
 910                 RETURN(rc);
 911
 912         rc = dt_declare_record_write(env, o, LLOG_CHUNK_SIZE, 0, th);
 913         if (rc)
 914                 RETURN(rc);
 915
 916         if (res->lgh_name) {
 917                 struct dt_object *llog_dir;
 918
 919                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 920                 if (IS_ERR(llog_dir))
 921                         RETURN(PTR_ERR(llog_dir));
 922                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 923                 rc = dt_declare_insert(env, llog_dir,
 924                                        (struct dt_rec *)&lgi->lgi_fid,
 925                                        (struct dt_key *)res->lgh_name, th);
 926                 lu_object_put(env, &llog_dir->do_lu);
 927                 if (rc)
 928                         CERROR("%s: can't declare named llog %s: rc = %d\n",
 929                                o->do_lu.lo_dev->ld_obd->obd_name,
 930                                res->lgh_name, rc);
 931         }
 932         RETURN(rc);
 933 }
 934
 935 /* This is a callback from the llog_* functions.
 936  * Assumes caller has already pushed us into the kernel context. */
 937 static int llog_osd_create(const struct lu_env *env, struct llog_handle *res,
 938                            struct thandle *th)
 939 {
 940         struct llog_thread_info *lgi = llog_info(env);
 941         struct local_oid_storage *los;
 942         struct dt_object        *o;
 943         int                      rc = 0;
 944
 945         ENTRY;
 946
 947         LASSERT(env);
 948         o = res->lgh_obj;
 949         LASSERT(o);
 950
 951         /* llog can be already created */
 952         if (dt_object_exists(o))
 953                 RETURN(-EEXIST);
 954
 955         los = res->private_data;
 956         LASSERT(los);
 957
 958         dt_write_lock(env, o, 0);
 959         if (!dt_object_exists(o))
 960                 rc = llog_osd_create_new_object(env, los, o, th);
 961         else
 962                 rc = -EEXIST;
 963
 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_ctxt        *ctxt;
1016         struct dt_object        *o, *llog_dir = NULL;
1017         struct dt_device        *d;
1018         struct thandle          *th;
1019         char                    *name = NULL;
1020         int                      rc;
1021
1022         ENTRY;
1023
1024         ctxt = loghandle->lgh_ctxt;
1025         LASSERT(ctxt);
1026
1027         o = loghandle->lgh_obj;
1028         LASSERT(o);
1029
1030         d = lu2dt_dev(o->do_lu.lo_dev);
1031         LASSERT(d);
1032         LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
1033
1034         th = dt_trans_create(env, d);
1035         if (IS_ERR(th))
1036                 RETURN(PTR_ERR(th));
1037
1038         if (loghandle->lgh_name) {
1039                 llog_dir = llog_osd_dir_get(env, ctxt);
1040                 if (IS_ERR(llog_dir))
1041                         GOTO(out_trans, rc = PTR_ERR(llog_dir));
1042
1043                 name = loghandle->lgh_name;
1044                 rc = dt_declare_delete(env, llog_dir,
1045                                        (struct dt_key *)name, th);
1046                 if (rc)
1047                         GOTO(out_trans, rc);
1048         }
1049
1050         dt_declare_ref_del(env, o, th);
1051
1052         rc = dt_declare_destroy(env, o, th);
1053         if (rc)
1054                 GOTO(out_trans, rc);
1055
1056         rc = dt_trans_start_local(env, d, th);
1057         if (rc)
1058                 GOTO(out_trans, rc);
1059
1060         dt_write_lock(env, o, 0);
1061         if (dt_object_exists(o)) {
1062                 if (name) {
1063                         dt_read_lock(env, llog_dir, 0);
1064                         rc = dt_delete(env, llog_dir,
1065                                        (struct dt_key *) name,
1066                                        th, BYPASS_CAPA);
1067                         dt_read_unlock(env, llog_dir);
1068                         if (rc) {
1069                                 CERROR("%s: can't remove llog %s: rc = %d\n",
1070                                        o->do_lu.lo_dev->ld_obd->obd_name,
1071                                        name, rc);
1072                                 GOTO(out_unlock, rc);
1073                         }
1074                 }
1075                 dt_ref_del(env, o, th);
1076                 rc = dt_destroy(env, o, th);
1077                 if (rc)
1078                         GOTO(out_unlock, rc);
1079         }
1080 out_unlock:
1081         dt_write_unlock(env, o);
1082 out_trans:
1083         dt_trans_stop(env, d, th);
1084         if (llog_dir != NULL)
1085                 lu_object_put(env, &llog_dir->do_lu);
1086         RETURN(rc);
1087 }
1088
1089 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1090                           struct obd_llog_group *olg, int ctxt_idx,
1091                           struct obd_device *disk_obd)
1092 {
1093         struct local_oid_storage        *los;
1094         struct llog_thread_info         *lgi = llog_info(env);
1095         struct llog_ctxt                *ctxt;
1096         int                              rc = 0;
1097
1098         ENTRY;
1099
1100         LASSERT(obd);
1101         LASSERT(olg->olg_ctxts[ctxt_idx]);
1102
1103         ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1104         LASSERT(ctxt);
1105
1106         /* initialize data allowing to generate new fids,
1107          * literally we need a sequece */
1108         lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1109         lgi->lgi_fid.f_oid = 1;
1110         lgi->lgi_fid.f_ver = 0;
1111         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1112                                     &lgi->lgi_fid, &los);
1113         if (rc < 0)
1114                 return rc;
1115
1116         lgi->lgi_fid.f_seq = FID_SEQ_LLOG_NAME;
1117         lgi->lgi_fid.f_oid = 1;
1118         lgi->lgi_fid.f_ver = 0;
1119         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1120                                     &lgi->lgi_fid, &los);
1121         llog_ctxt_put(ctxt);
1122         return rc;
1123 }
1124
1125 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1126 {
1127         struct dt_device                *dt;
1128         struct ls_device                *ls;
1129         struct local_oid_storage        *los, *nlos;
1130
1131         LASSERT(ctxt->loc_exp->exp_obd);
1132         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
1133         ls = ls_device_get(dt);
1134         if (IS_ERR(ls))
1135                 RETURN(PTR_ERR(ls));
1136
1137         mutex_lock(&ls->ls_los_mutex);
1138         los = dt_los_find(ls, FID_SEQ_LLOG);
1139         nlos = dt_los_find(ls, FID_SEQ_LLOG_NAME);
1140         mutex_unlock(&ls->ls_los_mutex);
1141         if (los != NULL) {
1142                 dt_los_put(los);
1143                 local_oid_storage_fini(env, los);
1144         }
1145         if (nlos != NULL) {
1146                 dt_los_put(nlos);
1147                 local_oid_storage_fini(env, nlos);
1148         }
1149         ls_device_put(env, ls);
1150         return 0;
1151 }
1152
1153 struct llog_operations llog_osd_ops = {
1154         .lop_next_block         = llog_osd_next_block,
1155         .lop_prev_block         = llog_osd_prev_block,
1156         .lop_read_header        = llog_osd_read_header,
1157         .lop_destroy            = llog_osd_destroy,
1158         .lop_setup              = llog_osd_setup,
1159         .lop_cleanup            = llog_osd_cleanup,
1160         .lop_open               = llog_osd_open,
1161         .lop_exist              = llog_osd_exist,
1162         .lop_declare_create     = llog_osd_declare_create,
1163         .lop_create             = llog_osd_create,
1164         .lop_declare_write_rec  = llog_osd_declare_write_rec,
1165         .lop_write_rec          = llog_osd_write_rec,
1166         .lop_close              = llog_osd_close,
1167 };
1168 EXPORT_SYMBOL(llog_osd_ops);
1169
1170 /* reads the catalog list */
1171 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1172                           int idx, int count, struct llog_catid *idarray)
1173 {
1174         struct llog_thread_info *lgi = llog_info(env);
1175         struct dt_object        *o = NULL;
1176         struct thandle          *th;
1177         int                      rc, size;
1178
1179         ENTRY;
1180
1181         LASSERT(d);
1182
1183         size = sizeof(*idarray) * count;
1184         lgi->lgi_off = idx *  sizeof(*idarray);
1185
1186         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1187
1188         o = dt_locate(env, d, &lgi->lgi_fid);
1189         if (IS_ERR(o))
1190                 RETURN(PTR_ERR(o));
1191
1192         if (!dt_object_exists(o)) {
1193                 th = dt_trans_create(env, d);
1194                 if (IS_ERR(th))
1195                         GOTO(out, rc = PTR_ERR(th));
1196
1197                 lgi->lgi_attr.la_valid = LA_MODE;
1198                 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1199                 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1200
1201                 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1202                                        &lgi->lgi_dof, th);
1203                 if (rc)
1204                         GOTO(out_trans, rc);
1205
1206                 rc = dt_trans_start_local(env, d, th);
1207                 if (rc)
1208                         GOTO(out_trans, rc);
1209
1210                 dt_write_lock(env, o, 0);
1211                 if (!dt_object_exists(o))
1212                         rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1213                                        &lgi->lgi_dof, th);
1214                 dt_write_unlock(env, o);
1215 out_trans:
1216                 dt_trans_stop(env, d, th);
1217                 if (rc)
1218                         GOTO(out, rc);
1219         }
1220
1221         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1222         if (rc)
1223                 GOTO(out, rc);
1224
1225         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1226                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1227                        o->do_lu.lo_dev->ld_obd->obd_name,
1228                        lgi->lgi_attr.la_mode);
1229                 GOTO(out, rc = -ENOENT);
1230         }
1231
1232         CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1233                (int)lgi->lgi_attr.la_size, size);
1234
1235         /* return just number of llogs */
1236         if (idarray == NULL) {
1237                 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1238                 GOTO(out, rc);
1239         }
1240
1241         /* read for new ost index or for empty file */
1242         memset(idarray, 0, size);
1243         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1244                 GOTO(out, rc = 0);
1245         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1246                 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1247
1248         lgi->lgi_buf.lb_buf = idarray;
1249         lgi->lgi_buf.lb_len = size;
1250         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1251         if (rc) {
1252                 CERROR("%s: error reading CATALOGS: rc = %d\n",
1253                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
1254                 GOTO(out, rc);
1255         }
1256
1257         EXIT;
1258 out:
1259         lu_object_put(env, &o->do_lu);
1260         RETURN(rc);
1261 }
1262 EXPORT_SYMBOL(llog_osd_get_cat_list);
1263
1264 /* writes the cat list */
1265 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1266                           int idx, int count, struct llog_catid *idarray)
1267 {
1268         struct llog_thread_info *lgi = llog_info(env);
1269         struct dt_object        *o = NULL;
1270         struct thandle          *th;
1271         int                      rc, size;
1272
1273         if (!count)
1274                 RETURN(0);
1275
1276         LASSERT(d);
1277
1278         size = sizeof(*idarray) * count;
1279         lgi->lgi_off = idx * sizeof(*idarray);
1280
1281         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1282
1283         o = dt_locate(env, d, &lgi->lgi_fid);
1284         if (IS_ERR(o))
1285                 RETURN(PTR_ERR(o));
1286
1287         if (!dt_object_exists(o))
1288                 GOTO(out, rc = -ENOENT);
1289
1290         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1291         if (rc)
1292                 GOTO(out, rc);
1293
1294         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1295                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1296                        o->do_lu.lo_dev->ld_obd->obd_name,
1297                        lgi->lgi_attr.la_mode);
1298                 GOTO(out, rc = -ENOENT);
1299         }
1300
1301         th = dt_trans_create(env, d);
1302         if (IS_ERR(th))
1303                 GOTO(out, rc = PTR_ERR(th));
1304
1305         rc = dt_declare_record_write(env, o, size, lgi->lgi_off, th);
1306         if (rc)
1307                 GOTO(out, rc);
1308
1309         rc = dt_trans_start_local(env, d, th);
1310         if (rc)
1311                 GOTO(out_trans, rc);
1312
1313         lgi->lgi_buf.lb_buf = idarray;
1314         lgi->lgi_buf.lb_len = size;
1315         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1316         if (rc)
1317                 CDEBUG(D_INODE, "error writeing CATALOGS: rc = %d\n", rc);
1318 out_trans:
1319         dt_trans_stop(env, d, th);
1320 out:
1321         lu_object_put(env, &o->do_lu);
1322         RETURN(rc);
1323 }
1324 EXPORT_SYMBOL(llog_osd_put_cat_list);