drivers/staging/lustre/lustre/obdclass/llog_osd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/llog_osd.c - low level llog routines on top of OSD API
  37  *
  38  * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
  39  * Author: Mikhail Pershin <mike.pershin@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LOG
  43
  44 #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         LASSERT(th);
  97         LASSERT(off);
  98         LASSERT(len >= LLOG_MIN_REC_SIZE && (len & 0x7) == 0);
  99
 100         lgi->lgi_tail.lrt_len = lgi->lgi_lrh.lrh_len = len;
 101         lgi->lgi_tail.lrt_index = lgi->lgi_lrh.lrh_index = index;
 102         lgi->lgi_lrh.lrh_type = LLOG_PAD_MAGIC;
 103
 104         lgi->lgi_buf.lb_buf = &lgi->lgi_lrh;
 105         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_lrh);
 106         dt_write_lock(env, o, 0);
 107         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 108         if (rc) {
 109                 CERROR("%s: error writing padding record: rc = %d\n",
 110                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 111                 GOTO(out, rc);
 112         }
 113
 114         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 115         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 116         *off += len - sizeof(lgi->lgi_lrh) - sizeof(lgi->lgi_tail);
 117         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 118         if (rc)
 119                 CERROR("%s: error writing padding record: rc = %d\n",
 120                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 121 out:
 122         dt_write_unlock(env, o);
 123         return rc;
 124 }
 125
 126 static int llog_osd_write_blob(const struct lu_env *env, struct dt_object *o,
 127                                struct llog_rec_hdr *rec, void *buf,
 128                                loff_t *off, struct thandle *th)
 129 {
 130         struct llog_thread_info *lgi = llog_info(env);
 131         int                      buflen = rec->lrh_len;
 132         int                      rc;
 133
 134         LASSERT(env);
 135         LASSERT(o);
 136
 137         if (buflen == 0)
 138                 CWARN("0-length record\n");
 139
 140         CDEBUG(D_OTHER, "write blob with type %x, buf %p/%u at off %llu\n",
 141                rec->lrh_type, buf, buflen, *off);
 142
 143         lgi->lgi_attr.la_valid = LA_SIZE;
 144         lgi->lgi_attr.la_size = *off;
 145
 146         if (!buf) {
 147                 lgi->lgi_buf.lb_len = buflen;
 148                 lgi->lgi_buf.lb_buf = rec;
 149                 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 150                 if (rc)
 151                         CERROR("%s: error writing log record: rc = %d\n",
 152                                o->do_lu.lo_dev->ld_obd->obd_name, rc);
 153                 GOTO(out, rc);
 154         }
 155
 156         /* the buf case */
 157         /* protect the following 3 writes from concurrent read */
 158         dt_write_lock(env, o, 0);
 159         rec->lrh_len = sizeof(*rec) + buflen + sizeof(lgi->lgi_tail);
 160         lgi->lgi_buf.lb_len = sizeof(*rec);
 161         lgi->lgi_buf.lb_buf = rec;
 162         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 163         if (rc) {
 164                 CERROR("%s: error writing log hdr: rc = %d\n",
 165                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 166                 GOTO(out_unlock, rc);
 167         }
 168
 169         lgi->lgi_buf.lb_len = buflen;
 170         lgi->lgi_buf.lb_buf = buf;
 171         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 172         if (rc) {
 173                 CERROR("%s: error writing log buffer: rc = %d\n",
 174                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
 175                 GOTO(out_unlock, rc);
 176         }
 177
 178         lgi->lgi_tail.lrt_len = rec->lrh_len;
 179         lgi->lgi_tail.lrt_index = rec->lrh_index;
 180         lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
 181         lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
 182         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 183         if (rc)
 184                 CERROR("%s: error writing log tail: rc = %d\n",
 185                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 186
 187 out_unlock:
 188         dt_write_unlock(env, o);
 189
 190 out:
 191         /* cleanup the content written above */
 192         if (rc) {
 193                 dt_punch(env, o, lgi->lgi_attr.la_size, OBD_OBJECT_EOF, th,
 194                          BYPASS_CAPA);
 195                 dt_attr_set(env, o, &lgi->lgi_attr, th, BYPASS_CAPA);
 196         }
 197
 198         return rc;
 199 }
 200
 201 static int llog_osd_read_header(const struct lu_env *env,
 202                                 struct llog_handle *handle)
 203 {
 204         struct llog_rec_hdr     *llh_hdr;
 205         struct dt_object        *o;
 206         struct llog_thread_info *lgi;
 207         int                      rc;
 208
 209         LASSERT(sizeof(*handle->lgh_hdr) == LLOG_CHUNK_SIZE);
 210
 211         o = handle->lgh_obj;
 212         LASSERT(o);
 213
 214         lgi = llog_info(env);
 215
 216         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 217         if (rc)
 218                 return rc;
 219
 220         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 221
 222         if (lgi->lgi_attr.la_size == 0) {
 223                 CDEBUG(D_HA, "not reading header from 0-byte log\n");
 224                 return LLOG_EEMPTY;
 225         }
 226
 227         lgi->lgi_off = 0;
 228         lgi->lgi_buf.lb_buf = handle->lgh_hdr;
 229         lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE;
 230
 231         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
 232         if (rc) {
 233                 CERROR("%s: error reading log header from "DFID": rc = %d\n",
 234                        o->do_lu.lo_dev->ld_obd->obd_name,
 235                        PFID(lu_object_fid(&o->do_lu)), rc);
 236                 return rc;
 237         }
 238
 239         llh_hdr = &handle->lgh_hdr->llh_hdr;
 240         if (LLOG_REC_HDR_NEEDS_SWABBING(llh_hdr))
 241                 lustre_swab_llog_hdr(handle->lgh_hdr);
 242
 243         if (llh_hdr->lrh_type != LLOG_HDR_MAGIC) {
 244                 CERROR("%s: bad log %s "DFID" header magic: %#x "
 245                        "(expected %#x)\n", o->do_lu.lo_dev->ld_obd->obd_name,
 246                        handle->lgh_name ? handle->lgh_name : "",
 247                        PFID(lu_object_fid(&o->do_lu)),
 248                        llh_hdr->lrh_type, LLOG_HDR_MAGIC);
 249                 return -EIO;
 250         } else if (llh_hdr->lrh_len != LLOG_CHUNK_SIZE) {
 251                 CERROR("%s: incorrectly sized log %s "DFID" header: "
 252                        "%#x (expected %#x)\n"
 253                        "you may need to re-run lconf --write_conf.\n",
 254                        o->do_lu.lo_dev->ld_obd->obd_name,
 255                        handle->lgh_name ? handle->lgh_name : "",
 256                        PFID(lu_object_fid(&o->do_lu)),
 257                        llh_hdr->lrh_len, LLOG_CHUNK_SIZE);
 258                 return -EIO;
 259         }
 260
 261         handle->lgh_last_idx = handle->lgh_hdr->llh_tail.lrt_index;
 262
 263         return 0;
 264 }
 265
 266 static int llog_osd_declare_write_rec(const struct lu_env *env,
 267                                       struct llog_handle *loghandle,
 268                                       struct llog_rec_hdr *rec,
 269                                       int idx, struct thandle *th)
 270 {
 271         struct llog_thread_info *lgi = llog_info(env);
 272         struct dt_object        *o;
 273         int                      rc;
 274
 275         LASSERT(env);
 276         LASSERT(th);
 277         LASSERT(loghandle);
 278
 279         o = loghandle->lgh_obj;
 280         LASSERT(o);
 281
 282         /* each time we update header */
 283         rc = dt_declare_record_write(env, o, sizeof(struct llog_log_hdr), 0,
 284                                      th);
 285         if (rc || idx == 0) /* if error or just header */
 286                 return rc;
 287
 288         if (dt_object_exists(o)) {
 289                 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 290                 lgi->lgi_off = lgi->lgi_attr.la_size;
 291                 LASSERT(ergo(rc == 0, lgi->lgi_attr.la_valid & LA_SIZE));
 292                 if (rc)
 293                         return rc;
 294
 295                 rc = dt_declare_punch(env, o, lgi->lgi_off, OBD_OBJECT_EOF, th);
 296                 if (rc)
 297                         return rc;
 298         } else {
 299                 lgi->lgi_off = 0;
 300         }
 301
 302         /* XXX: implement declared window or multi-chunks approach */
 303         rc = dt_declare_record_write(env, o, 32 * 1024, lgi->lgi_off, th);
 304
 305         return rc;
 306 }
 307
 308 /* returns negative in on error; 0 if success && reccookie == 0; 1 otherwise */
 309 /* appends if idx == -1, otherwise overwrites record idx. */
 310 static int llog_osd_write_rec(const struct lu_env *env,
 311                               struct llog_handle *loghandle,
 312                               struct llog_rec_hdr *rec,
 313                               struct llog_cookie *reccookie, int cookiecount,
 314                               void *buf, int idx, struct thandle *th)
 315 {
 316         struct llog_thread_info *lgi = llog_info(env);
 317         struct llog_log_hdr     *llh;
 318         int                      reclen = rec->lrh_len;
 319         int                      index, rc, old_tail_idx;
 320         struct llog_rec_tail    *lrt;
 321         struct dt_object        *o;
 322         size_t                   left;
 323
 324         LASSERT(env);
 325         llh = loghandle->lgh_hdr;
 326         LASSERT(llh);
 327         o = loghandle->lgh_obj;
 328         LASSERT(o);
 329         LASSERT(th);
 330
 331         CDEBUG(D_OTHER, "new record %x to "DFID"\n",
 332                rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
 333
 334         /* record length should not bigger than LLOG_CHUNK_SIZE */
 335         if (buf)
 336                 rc = (reclen > LLOG_CHUNK_SIZE - sizeof(struct llog_rec_hdr) -
 337                       sizeof(struct llog_rec_tail)) ? -E2BIG : 0;
 338         else
 339                 rc = (reclen > LLOG_CHUNK_SIZE) ? -E2BIG : 0;
 340         if (rc)
 341                 return rc;
 342
 343         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 344         if (rc)
 345                 return rc;
 346
 347         if (buf)
 348                 /* write_blob adds header and tail to lrh_len. */
 349                 reclen = sizeof(*rec) + rec->lrh_len +
 350                          sizeof(struct llog_rec_tail);
 351
 352         if (idx != -1) {
 353                 /* no header: only allowed to insert record 1 */
 354                 if (idx != 1 && lgi->lgi_attr.la_size == 0)
 355                         LBUG();
 356
 357                 if (idx && llh->llh_size && llh->llh_size != rec->lrh_len)
 358                         return -EINVAL;
 359
 360                 if (!ext2_test_bit(idx, llh->llh_bitmap))
 361                         CERROR("%s: modify unset record %u\n",
 362                                o->do_lu.lo_dev->ld_obd->obd_name, idx);
 363                 if (idx != rec->lrh_index)
 364                         CERROR("%s: index mismatch %d %u\n",
 365                                o->do_lu.lo_dev->ld_obd->obd_name, idx,
 366                                rec->lrh_index);
 367
 368                 lgi->lgi_off = 0;
 369                 rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 370                                          &lgi->lgi_off, th);
 371                 /* we are done if we only write the header or on error */
 372                 if (rc || idx == 0)
 373                         return rc;
 374
 375                 if (buf) {
 376                         /* We assume that caller has set lgh_cur_* */
 377                         lgi->lgi_off = loghandle->lgh_cur_offset;
 378                         CDEBUG(D_OTHER,
 379                                "modify record "DOSTID": idx:%d/%u/%d, len:%u "
 380                                "offset %llu\n",
 381                                POSTID(&loghandle->lgh_id.lgl_oi), idx,
 382                                rec->lrh_index,
 383                                loghandle->lgh_cur_idx, rec->lrh_len,
 384                                (long long)(lgi->lgi_off - sizeof(*llh)));
 385                         if (rec->lrh_index != loghandle->lgh_cur_idx) {
 386                                 CERROR("%s: modify idx mismatch %u/%d\n",
 387                                        o->do_lu.lo_dev->ld_obd->obd_name, idx,
 388                                        loghandle->lgh_cur_idx);
 389                                 return -EFAULT;
 390                         }
 391                 } else {
 392                         /* Assumes constant lrh_len */
 393                         lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
 394                 }
 395
 396                 rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 397                 if (rc == 0 && reccookie) {
 398                         reccookie->lgc_lgl = loghandle->lgh_id;
 399                         reccookie->lgc_index = idx;
 400                         rc = 1;
 401                 }
 402                 return rc;
 403         }
 404
 405         /* Make sure that records don't cross a chunk boundary, so we can
 406          * process them page-at-a-time if needed.  If it will cross a chunk
 407          * boundary, write in a fake (but referenced) entry to pad the chunk.
 408          *
 409          * We know that llog_current_log() will return a loghandle that is
 410          * big enough to hold reclen, so all we care about is padding here.
 411          */
 412         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 413         lgi->lgi_off = lgi->lgi_attr.la_size;
 414         left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
 415         /* NOTE: padding is a record, but no bit is set */
 416         if (left != 0 && left != reclen &&
 417             left < (reclen + LLOG_MIN_REC_SIZE)) {
 418                 index = loghandle->lgh_last_idx + 1;
 419                 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
 420                 if (rc)
 421                         return rc;
 422                 loghandle->lgh_last_idx++; /*for pad rec*/
 423         }
 424         /* if it's the last idx in log file, then return -ENOSPC */
 425         if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
 426                 return -ENOSPC;
 427
 428         loghandle->lgh_last_idx++;
 429         index = loghandle->lgh_last_idx;
 430         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 431         rec->lrh_index = index;
 432         if (buf == NULL) {
 433                 lrt = (struct llog_rec_tail *)((char *)rec + rec->lrh_len -
 434                                                sizeof(*lrt));
 435                 lrt->lrt_len = rec->lrh_len;
 436                 lrt->lrt_index = rec->lrh_index;
 437         }
 438         /* The caller should make sure only 1 process access the lgh_last_idx,
 439          * Otherwise it might hit the assert.*/
 440         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 441         spin_lock(&loghandle->lgh_hdr_lock);
 442         if (ext2_set_bit(index, llh->llh_bitmap)) {
 443                 CERROR("%s: index %u already set in log bitmap\n",
 444                        o->do_lu.lo_dev->ld_obd->obd_name, index);
 445                 spin_unlock(&loghandle->lgh_hdr_lock);
 446                 LBUG(); /* should never happen */
 447         }
 448         llh->llh_count++;
 449         spin_unlock(&loghandle->lgh_hdr_lock);
 450         old_tail_idx = llh->llh_tail.lrt_index;
 451         llh->llh_tail.lrt_index = index;
 452
 453         lgi->lgi_off = 0;
 454         rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL, &lgi->lgi_off,
 455                                  th);
 456         if (rc)
 457                 GOTO(out, rc);
 458
 459         rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
 460         if (rc)
 461                 GOTO(out, rc);
 462
 463         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 464         lgi->lgi_off = lgi->lgi_attr.la_size;
 465
 466         rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
 467
 468 out:
 469         /* cleanup llog for error case */
 470         if (rc) {
 471                 spin_lock(&loghandle->lgh_hdr_lock);
 472                 ext2_clear_bit(index, llh->llh_bitmap);
 473                 llh->llh_count--;
 474                 spin_unlock(&loghandle->lgh_hdr_lock);
 475
 476                 /* restore the header */
 477                 loghandle->lgh_last_idx--;
 478                 llh->llh_tail.lrt_index = old_tail_idx;
 479                 lgi->lgi_off = 0;
 480                 llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
 481                                     &lgi->lgi_off, th);
 482         }
 483
 484         CDEBUG(D_RPCTRACE, "added record "DOSTID": idx: %u, %u\n",
 485                POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len);
 486         if (rc == 0 && reccookie) {
 487                 reccookie->lgc_lgl = loghandle->lgh_id;
 488                 reccookie->lgc_index = index;
 489                 if ((rec->lrh_type == MDS_UNLINK_REC) ||
 490                     (rec->lrh_type == MDS_SETATTR64_REC))
 491                         reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 492                 else if (rec->lrh_type == OST_SZ_REC)
 493                         reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
 494                 else
 495                         reccookie->lgc_subsys = -1;
 496                 rc = 1;
 497         }
 498         return rc;
 499 }
 500
 501 /* We can skip reading at least as many log blocks as the number of
 502  * minimum sized log records we are skipping.  If it turns out
 503  * that we are not far enough along the log (because the
 504  * actual records are larger than minimum size) we just skip
 505  * some more records.
 506  */
 507 static void llog_skip_over(__u64 *off, int curr, int goal)
 508 {
 509         if (goal <= curr)
 510                 return;
 511         *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
 512                 ~(LLOG_CHUNK_SIZE - 1);
 513 }
 514
 515 /* sets:
 516  *  - cur_offset to the furthest point read in the log file
 517  *  - cur_idx to the log index preceeding cur_offset
 518  * returns -EIO/-EINVAL on error
 519  */
 520 static int llog_osd_next_block(const struct lu_env *env,
 521                                struct llog_handle *loghandle, int *cur_idx,
 522                                int next_idx, __u64 *cur_offset, void *buf,
 523                                int len)
 524 {
 525         struct llog_thread_info *lgi = llog_info(env);
 526         struct dt_object        *o;
 527         struct dt_device        *dt;
 528         int                      rc;
 529
 530         LASSERT(env);
 531         LASSERT(lgi);
 532
 533         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 534                 return -EINVAL;
 535
 536         CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
 537                next_idx, *cur_idx, *cur_offset);
 538
 539         LASSERT(loghandle);
 540         LASSERT(loghandle->lgh_ctxt);
 541
 542         o = loghandle->lgh_obj;
 543         LASSERT(o);
 544         LASSERT(dt_object_exists(o));
 545         dt = lu2dt_dev(o->do_lu.lo_dev);
 546         LASSERT(dt);
 547
 548         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 549         if (rc)
 550                 GOTO(out, rc);
 551
 552         while (*cur_offset < lgi->lgi_attr.la_size) {
 553                 struct llog_rec_hdr     *rec, *last_rec;
 554                 struct llog_rec_tail    *tail;
 555
 556                 llog_skip_over(cur_offset, *cur_idx, next_idx);
 557
 558                 /* read up to next LLOG_CHUNK_SIZE block */
 559                 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
 560                                       (*cur_offset & (LLOG_CHUNK_SIZE - 1));
 561                 lgi->lgi_buf.lb_buf = buf;
 562
 563                 /* Note: read lock is not needed around la_size get above at
 564                  * the time of dt_attr_get(). There are only two cases that
 565                  * matter. Either la_size == cur_offset, in which case the
 566                  * entire read is skipped, or la_size > cur_offset and the loop
 567                  * is entered and this thread is blocked at dt_read_lock()
 568                  * until the write is completed. When the write completes, then
 569                  * the dt_read() will be done with the full length, and will
 570                  * get the full data.
 571                  */
 572                 dt_read_lock(env, o, 0);
 573                 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
 574                 dt_read_unlock(env, o);
 575                 if (rc < 0) {
 576                         CERROR("%s: can't read llog block from log "DFID
 577                                " offset "LPU64": rc = %d\n",
 578                                o->do_lu.lo_dev->ld_obd->obd_name,
 579                                PFID(lu_object_fid(&o->do_lu)), *cur_offset,
 580                                rc);
 581                         GOTO(out, rc);
 582                 }
 583
 584                 if (rc < len) {
 585                         /* signal the end of the valid buffer to
 586                          * llog_process */
 587                         memset(buf + rc, 0, len - rc);
 588                 }
 589
 590                 if (rc == 0) /* end of file, nothing to do */
 591                         GOTO(out, rc);
 592
 593                 if (rc < sizeof(*tail)) {
 594                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 595                                "offset "LPU64"\n",
 596                                o->do_lu.lo_dev->ld_obd->obd_name,
 597                                POSTID(&loghandle->lgh_id.lgl_oi),
 598                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 599                         GOTO(out, rc = -EINVAL);
 600                 }
 601
 602                 rec = buf;
 603                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 604                         lustre_swab_llog_rec(rec);
 605
 606                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 607                                                 sizeof(struct llog_rec_tail));
 608                 /* get the last record in block */
 609                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 610                                                    le32_to_cpu(tail->lrt_len));
 611
 612                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 613                         lustre_swab_llog_rec(last_rec);
 614                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 615
 616                 *cur_idx = tail->lrt_index;
 617
 618                 /* this shouldn't happen */
 619                 if (tail->lrt_index == 0) {
 620                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 621                                "offset "LPU64"\n",
 622                                o->do_lu.lo_dev->ld_obd->obd_name,
 623                                POSTID(&loghandle->lgh_id.lgl_oi),
 624                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 625                         GOTO(out, rc = -EINVAL);
 626                 }
 627                 if (tail->lrt_index < next_idx)
 628                         continue;
 629
 630                 /* sanity check that the start of the new buffer is no farther
 631                  * than the record that we wanted.  This shouldn't happen. */
 632                 if (rec->lrh_index > next_idx) {
 633                         CERROR("%s: missed desired record? %u > %u\n",
 634                                o->do_lu.lo_dev->ld_obd->obd_name,
 635                                rec->lrh_index, next_idx);
 636                         GOTO(out, rc = -ENOENT);
 637                 }
 638                 GOTO(out, rc = 0);
 639         }
 640         GOTO(out, rc = -EIO);
 641 out:
 642         return rc;
 643 }
 644
 645 static int llog_osd_prev_block(const struct lu_env *env,
 646                                struct llog_handle *loghandle,
 647                                int prev_idx, void *buf, int len)
 648 {
 649         struct llog_thread_info *lgi = llog_info(env);
 650         struct dt_object        *o;
 651         struct dt_device        *dt;
 652         loff_t                   cur_offset;
 653         int                      rc;
 654
 655         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 656                 return -EINVAL;
 657
 658         CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
 659
 660         LASSERT(loghandle);
 661         LASSERT(loghandle->lgh_ctxt);
 662
 663         o = loghandle->lgh_obj;
 664         LASSERT(o);
 665         LASSERT(dt_object_exists(o));
 666         dt = lu2dt_dev(o->do_lu.lo_dev);
 667         LASSERT(dt);
 668
 669         cur_offset = LLOG_CHUNK_SIZE;
 670         llog_skip_over(&cur_offset, 0, prev_idx);
 671
 672         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
 673         if (rc)
 674                 GOTO(out, rc);
 675
 676         while (cur_offset < lgi->lgi_attr.la_size) {
 677                 struct llog_rec_hdr     *rec, *last_rec;
 678                 struct llog_rec_tail    *tail;
 679
 680                 lgi->lgi_buf.lb_len = len;
 681                 lgi->lgi_buf.lb_buf = buf;
 682                 /* It is OK to have locking around dt_read() only, see
 683                  * comment in llog_osd_next_block for details
 684                  */
 685                 dt_read_lock(env, o, 0);
 686                 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
 687                 dt_read_unlock(env, o);
 688                 if (rc < 0) {
 689                         CERROR("%s: can't read llog block from log "DFID
 690                                " offset "LPU64": rc = %d\n",
 691                                o->do_lu.lo_dev->ld_obd->obd_name,
 692                                PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
 693                         GOTO(out, rc);
 694                 }
 695
 696                 if (rc == 0) /* end of file, nothing to do */
 697                         GOTO(out, rc);
 698
 699                 if (rc < sizeof(*tail)) {
 700                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 701                                "offset "LPU64"\n",
 702                                o->do_lu.lo_dev->ld_obd->obd_name,
 703                                POSTID(&loghandle->lgh_id.lgl_oi),
 704                                loghandle->lgh_id.lgl_ogen, cur_offset);
 705                         GOTO(out, rc = -EINVAL);
 706                 }
 707
 708                 rec = buf;
 709                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 710                         lustre_swab_llog_rec(rec);
 711
 712                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 713                                                 sizeof(struct llog_rec_tail));
 714                 /* get the last record in block */
 715                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 716                                                    le32_to_cpu(tail->lrt_len));
 717
 718                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 719                         lustre_swab_llog_rec(last_rec);
 720                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 721
 722                 /* this shouldn't happen */
 723                 if (tail->lrt_index == 0) {
 724                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 725                                "offset "LPU64"\n",
 726                                o->do_lu.lo_dev->ld_obd->obd_name,
 727                                POSTID(&loghandle->lgh_id.lgl_oi),
 728                                loghandle->lgh_id.lgl_ogen, cur_offset);
 729                         GOTO(out, rc = -EINVAL);
 730                 }
 731                 if (tail->lrt_index < prev_idx)
 732                         continue;
 733
 734                 /* sanity check that the start of the new buffer is no farther
 735                  * than the record that we wanted.  This shouldn't happen. */
 736                 if (rec->lrh_index > prev_idx) {
 737                         CERROR("%s: missed desired record? %u > %u\n",
 738                                o->do_lu.lo_dev->ld_obd->obd_name,
 739                                rec->lrh_index, prev_idx);
 740                         GOTO(out, rc = -ENOENT);
 741                 }
 742                 GOTO(out, rc = 0);
 743         }
 744         GOTO(out, rc = -EIO);
 745 out:
 746         return rc;
 747 }
 748
 749 struct dt_object *llog_osd_dir_get(const struct lu_env *env,
 750                                    struct llog_ctxt *ctxt)
 751 {
 752         struct dt_device        *dt;
 753         struct dt_thread_info   *dti = dt_info(env);
 754         struct dt_object        *dir;
 755         int                      rc;
 756
 757         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 758         if (ctxt->loc_dir == NULL) {
 759                 rc = dt_root_get(env, dt, &dti->dti_fid);
 760                 if (rc)
 761                         return ERR_PTR(rc);
 762                 dir = dt_locate(env, dt, &dti->dti_fid);
 763         } else {
 764                 lu_object_get(&ctxt->loc_dir->do_lu);
 765                 dir = ctxt->loc_dir;
 766         }
 767
 768         return dir;
 769 }
 770
 771 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
 772                          struct llog_logid *logid, char *name,
 773                          enum llog_open_param open_param)
 774 {
 775         struct llog_thread_info         *lgi = llog_info(env);
 776         struct llog_ctxt                *ctxt = handle->lgh_ctxt;
 777         struct dt_object                *o;
 778         struct dt_device                *dt;
 779         struct ls_device                *ls;
 780         struct local_oid_storage        *los;
 781         int                              rc = 0;
 782
 783         LASSERT(env);
 784         LASSERT(ctxt);
 785         LASSERT(ctxt->loc_exp);
 786         LASSERT(ctxt->loc_exp->exp_obd);
 787         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 788         LASSERT(dt);
 789
 790         ls = ls_device_get(dt);
 791         if (IS_ERR(ls))
 792                 return PTR_ERR(ls);
 793
 794         mutex_lock(&ls->ls_los_mutex);
 795         los = dt_los_find(ls, name != NULL ? FID_SEQ_LLOG_NAME : FID_SEQ_LLOG);
 796         mutex_unlock(&ls->ls_los_mutex);
 797         LASSERT(los);
 798         ls_device_put(env, ls);
 799
 800         LASSERT(handle);
 801
 802         if (logid != NULL) {
 803                 logid_to_fid(logid, &lgi->lgi_fid);
 804         } else if (name) {
 805                 struct dt_object *llog_dir;
 806
 807                 llog_dir = llog_osd_dir_get(env, ctxt);
 808                 if (IS_ERR(llog_dir))
 809                         GOTO(out, rc = PTR_ERR(llog_dir));
 810                 dt_read_lock(env, llog_dir, 0);
 811                 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
 812                 dt_read_unlock(env, llog_dir);
 813                 lu_object_put(env, &llog_dir->do_lu);
 814                 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
 815                         /* generate fid for new llog */
 816                         rc = local_object_fid_generate(env, los,
 817                                                        &lgi->lgi_fid);
 818                 }
 819                 if (rc < 0)
 820                         GOTO(out, rc);
 821                 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
 822                 if (handle->lgh_name)
 823                         strcpy(handle->lgh_name, name);
 824                 else
 825                         GOTO(out, rc = -ENOMEM);
 826         } else {
 827                 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
 828                 /* generate fid for new llog */
 829                 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
 830                 if (rc < 0)
 831                         GOTO(out, rc);
 832         }
 833
 834         o = ls_locate(env, ls, &lgi->lgi_fid);
 835         if (IS_ERR(o))
 836                 GOTO(out_name, rc = PTR_ERR(o));
 837
 838         /* No new llog is expected but doesn't exist */
 839         if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
 840                 GOTO(out_put, rc = -ENOENT);
 841
 842         fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
 843         handle->lgh_obj = o;
 844         handle->private_data = los;
 845         LASSERT(handle->lgh_ctxt);
 846
 847         return rc;
 848
 849 out_put:
 850         lu_object_put(env, &o->do_lu);
 851 out_name:
 852         if (handle->lgh_name != NULL)
 853                 OBD_FREE(handle->lgh_name, strlen(name) + 1);
 854 out:
 855         dt_los_put(los);
 856         return rc;
 857 }
 858
 859 static int llog_osd_exist(struct llog_handle *handle)
 860 {
 861         LASSERT(handle->lgh_obj);
 862         return (dt_object_exists(handle->lgh_obj) &&
 863                 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
 864 }
 865
 866 static int llog_osd_declare_create(const struct lu_env *env,
 867                                    struct llog_handle *res, struct thandle *th)
 868 {
 869         struct llog_thread_info         *lgi = llog_info(env);
 870         struct local_oid_storage        *los;
 871         struct dt_object                *o;
 872         int                              rc;
 873
 874         LASSERT(res->lgh_obj);
 875         LASSERT(th);
 876
 877         /* object can be created by another thread */
 878         o = res->lgh_obj;
 879         if (dt_object_exists(o))
 880                 return 0;
 881
 882         los = res->private_data;
 883         LASSERT(los);
 884
 885         rc = llog_osd_declare_new_object(env, los, o, th);
 886         if (rc)
 887                 return rc;
 888
 889         rc = dt_declare_record_write(env, o, LLOG_CHUNK_SIZE, 0, th);
 890         if (rc)
 891                 return rc;
 892
 893         if (res->lgh_name) {
 894                 struct dt_object *llog_dir;
 895
 896                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 897                 if (IS_ERR(llog_dir))
 898                         return PTR_ERR(llog_dir);
 899                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 900                 rc = dt_declare_insert(env, llog_dir,
 901                                        (struct dt_rec *)&lgi->lgi_fid,
 902                                        (struct dt_key *)res->lgh_name, th);
 903                 lu_object_put(env, &llog_dir->do_lu);
 904                 if (rc)
 905                         CERROR("%s: can't declare named llog %s: rc = %d\n",
 906                                o->do_lu.lo_dev->ld_obd->obd_name,
 907                                res->lgh_name, rc);
 908         }
 909         return rc;
 910 }
 911
 912 /* This is a callback from the llog_* functions.
 913  * Assumes caller has already pushed us into the kernel context. */
 914 static int llog_osd_create(const struct lu_env *env, struct llog_handle *res,
 915                            struct thandle *th)
 916 {
 917         struct llog_thread_info *lgi = llog_info(env);
 918         struct local_oid_storage *los;
 919         struct dt_object        *o;
 920         int                   rc = 0;
 921
 922         LASSERT(env);
 923         o = res->lgh_obj;
 924         LASSERT(o);
 925
 926         /* llog can be already created */
 927         if (dt_object_exists(o))
 928                 return -EEXIST;
 929
 930         los = res->private_data;
 931         LASSERT(los);
 932
 933         dt_write_lock(env, o, 0);
 934         if (!dt_object_exists(o))
 935                 rc = llog_osd_create_new_object(env, los, o, th);
 936         else
 937                 rc = -EEXIST;
 938
 939         dt_write_unlock(env, o);
 940         if (rc)
 941                 return rc;
 942
 943         if (res->lgh_name) {
 944                 struct dt_object *llog_dir;
 945
 946                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
 947                 if (IS_ERR(llog_dir))
 948                         return PTR_ERR(llog_dir);
 949
 950                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
 951                 dt_read_lock(env, llog_dir, 0);
 952                 rc = dt_insert(env, llog_dir,
 953                                (struct dt_rec *)&lgi->lgi_fid,
 954                                (struct dt_key *)res->lgh_name,
 955                                th, BYPASS_CAPA, 1);
 956                 dt_read_unlock(env, llog_dir);
 957                 lu_object_put(env, &llog_dir->do_lu);
 958                 if (rc)
 959                         CERROR("%s: can't create named llog %s: rc = %d\n",
 960                                o->do_lu.lo_dev->ld_obd->obd_name,
 961                                res->lgh_name, rc);
 962         }
 963         return rc;
 964 }
 965
 966 static int llog_osd_close(const struct lu_env *env, struct llog_handle *handle)
 967 {
 968         struct local_oid_storage        *los;
 969         int                              rc = 0;
 970
 971         LASSERT(handle->lgh_obj);
 972
 973         lu_object_put(env, &handle->lgh_obj->do_lu);
 974
 975         los = handle->private_data;
 976         LASSERT(los);
 977         dt_los_put(los);
 978
 979         if (handle->lgh_name)
 980                 OBD_FREE(handle->lgh_name, strlen(handle->lgh_name) + 1);
 981
 982         return rc;
 983 }
 984
 985 static int llog_osd_destroy(const struct lu_env *env,
 986                             struct llog_handle *loghandle)
 987 {
 988         struct llog_ctxt        *ctxt;
 989         struct dt_object        *o, *llog_dir = NULL;
 990         struct dt_device        *d;
 991         struct thandle          *th;
 992         char                    *name = NULL;
 993         int                      rc;
 994
 995         ctxt = loghandle->lgh_ctxt;
 996         LASSERT(ctxt);
 997
 998         o = loghandle->lgh_obj;
 999         LASSERT(o);
1000
1001         d = lu2dt_dev(o->do_lu.lo_dev);
1002         LASSERT(d);
1003         LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
1004
1005         th = dt_trans_create(env, d);
1006         if (IS_ERR(th))
1007                 return PTR_ERR(th);
1008
1009         if (loghandle->lgh_name) {
1010                 llog_dir = llog_osd_dir_get(env, ctxt);
1011                 if (IS_ERR(llog_dir))
1012                         GOTO(out_trans, rc = PTR_ERR(llog_dir));
1013
1014                 name = loghandle->lgh_name;
1015                 rc = dt_declare_delete(env, llog_dir,
1016                                        (struct dt_key *)name, th);
1017                 if (rc)
1018                         GOTO(out_trans, rc);
1019         }
1020
1021         dt_declare_ref_del(env, o, th);
1022
1023         rc = dt_declare_destroy(env, o, th);
1024         if (rc)
1025                 GOTO(out_trans, rc);
1026
1027         rc = dt_trans_start_local(env, d, th);
1028         if (rc)
1029                 GOTO(out_trans, rc);
1030
1031         dt_write_lock(env, o, 0);
1032         if (dt_object_exists(o)) {
1033                 if (name) {
1034                         dt_read_lock(env, llog_dir, 0);
1035                         rc = dt_delete(env, llog_dir,
1036                                        (struct dt_key *) name,
1037                                        th, BYPASS_CAPA);
1038                         dt_read_unlock(env, llog_dir);
1039                         if (rc) {
1040                                 CERROR("%s: can't remove llog %s: rc = %d\n",
1041                                        o->do_lu.lo_dev->ld_obd->obd_name,
1042                                        name, rc);
1043                                 GOTO(out_unlock, rc);
1044                         }
1045                 }
1046                 dt_ref_del(env, o, th);
1047                 rc = dt_destroy(env, o, th);
1048                 if (rc)
1049                         GOTO(out_unlock, rc);
1050         }
1051 out_unlock:
1052         dt_write_unlock(env, o);
1053 out_trans:
1054         dt_trans_stop(env, d, th);
1055         if (llog_dir != NULL)
1056                 lu_object_put(env, &llog_dir->do_lu);
1057         return rc;
1058 }
1059
1060 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1061                           struct obd_llog_group *olg, int ctxt_idx,
1062                           struct obd_device *disk_obd)
1063 {
1064         struct local_oid_storage        *los;
1065         struct llog_thread_info         *lgi = llog_info(env);
1066         struct llog_ctxt                *ctxt;
1067         int                              rc = 0;
1068
1069         LASSERT(obd);
1070         LASSERT(olg->olg_ctxts[ctxt_idx]);
1071
1072         ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1073         LASSERT(ctxt);
1074
1075         /* initialize data allowing to generate new fids,
1076          * literally we need a sequece */
1077         lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1078         lgi->lgi_fid.f_oid = 1;
1079         lgi->lgi_fid.f_ver = 0;
1080         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1081                                     &lgi->lgi_fid, &los);
1082         if (rc < 0)
1083                 return rc;
1084
1085         lgi->lgi_fid.f_seq = FID_SEQ_LLOG_NAME;
1086         lgi->lgi_fid.f_oid = 1;
1087         lgi->lgi_fid.f_ver = 0;
1088         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1089                                     &lgi->lgi_fid, &los);
1090         llog_ctxt_put(ctxt);
1091         return rc;
1092 }
1093
1094 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1095 {
1096         struct dt_device                *dt;
1097         struct ls_device                *ls;
1098         struct local_oid_storage        *los, *nlos;
1099
1100         LASSERT(ctxt->loc_exp->exp_obd);
1101         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
1102         ls = ls_device_get(dt);
1103         if (IS_ERR(ls))
1104                 return PTR_ERR(ls);
1105
1106         mutex_lock(&ls->ls_los_mutex);
1107         los = dt_los_find(ls, FID_SEQ_LLOG);
1108         nlos = dt_los_find(ls, FID_SEQ_LLOG_NAME);
1109         mutex_unlock(&ls->ls_los_mutex);
1110         if (los != NULL) {
1111                 dt_los_put(los);
1112                 local_oid_storage_fini(env, los);
1113         }
1114         if (nlos != NULL) {
1115                 dt_los_put(nlos);
1116                 local_oid_storage_fini(env, nlos);
1117         }
1118         ls_device_put(env, ls);
1119         return 0;
1120 }
1121
1122 struct llog_operations llog_osd_ops = {
1123         .lop_next_block         = llog_osd_next_block,
1124         .lop_prev_block         = llog_osd_prev_block,
1125         .lop_read_header        = llog_osd_read_header,
1126         .lop_destroy            = llog_osd_destroy,
1127         .lop_setup              = llog_osd_setup,
1128         .lop_cleanup            = llog_osd_cleanup,
1129         .lop_open               = llog_osd_open,
1130         .lop_exist              = llog_osd_exist,
1131         .lop_declare_create     = llog_osd_declare_create,
1132         .lop_create             = llog_osd_create,
1133         .lop_declare_write_rec  = llog_osd_declare_write_rec,
1134         .lop_write_rec          = llog_osd_write_rec,
1135         .lop_close              = llog_osd_close,
1136 };
1137 EXPORT_SYMBOL(llog_osd_ops);
1138
1139 /* reads the catalog list */
1140 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1141                           int idx, int count, struct llog_catid *idarray)
1142 {
1143         struct llog_thread_info *lgi = llog_info(env);
1144         struct dt_object        *o = NULL;
1145         struct thandle          *th;
1146         int                      rc, size;
1147
1148         LASSERT(d);
1149
1150         size = sizeof(*idarray) * count;
1151         lgi->lgi_off = idx *  sizeof(*idarray);
1152
1153         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1154
1155         o = dt_locate(env, d, &lgi->lgi_fid);
1156         if (IS_ERR(o))
1157                 return PTR_ERR(o);
1158
1159         if (!dt_object_exists(o)) {
1160                 th = dt_trans_create(env, d);
1161                 if (IS_ERR(th))
1162                         GOTO(out, rc = PTR_ERR(th));
1163
1164                 lgi->lgi_attr.la_valid = LA_MODE;
1165                 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1166                 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1167
1168                 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1169                                        &lgi->lgi_dof, th);
1170                 if (rc)
1171                         GOTO(out_trans, rc);
1172
1173                 rc = dt_trans_start_local(env, d, th);
1174                 if (rc)
1175                         GOTO(out_trans, rc);
1176
1177                 dt_write_lock(env, o, 0);
1178                 if (!dt_object_exists(o))
1179                         rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1180                                        &lgi->lgi_dof, th);
1181                 dt_write_unlock(env, o);
1182 out_trans:
1183                 dt_trans_stop(env, d, th);
1184                 if (rc)
1185                         GOTO(out, rc);
1186         }
1187
1188         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1189         if (rc)
1190                 GOTO(out, rc);
1191
1192         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1193                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1194                        o->do_lu.lo_dev->ld_obd->obd_name,
1195                        lgi->lgi_attr.la_mode);
1196                 GOTO(out, rc = -ENOENT);
1197         }
1198
1199         CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1200                (int)lgi->lgi_attr.la_size, size);
1201
1202         /* return just number of llogs */
1203         if (idarray == NULL) {
1204                 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1205                 GOTO(out, rc);
1206         }
1207
1208         /* read for new ost index or for empty file */
1209         memset(idarray, 0, size);
1210         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1211                 GOTO(out, rc = 0);
1212         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1213                 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1214
1215         lgi->lgi_buf.lb_buf = idarray;
1216         lgi->lgi_buf.lb_len = size;
1217         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1218         if (rc) {
1219                 CERROR("%s: error reading CATALOGS: rc = %d\n",
1220                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
1221                 GOTO(out, rc);
1222         }
1223
1224 out:
1225         lu_object_put(env, &o->do_lu);
1226         return rc;
1227 }
1228 EXPORT_SYMBOL(llog_osd_get_cat_list);
1229
1230 /* writes the cat list */
1231 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1232                           int idx, int count, struct llog_catid *idarray)
1233 {
1234         struct llog_thread_info *lgi = llog_info(env);
1235         struct dt_object        *o = NULL;
1236         struct thandle          *th;
1237         int                      rc, size;
1238
1239         if (!count)
1240                 return 0;
1241
1242         LASSERT(d);
1243
1244         size = sizeof(*idarray) * count;
1245         lgi->lgi_off = idx * sizeof(*idarray);
1246
1247         lu_local_obj_fid(&lgi->lgi_fid, LLOG_CATALOGS_OID);
1248
1249         o = dt_locate(env, d, &lgi->lgi_fid);
1250         if (IS_ERR(o))
1251                 return PTR_ERR(o);
1252
1253         if (!dt_object_exists(o))
1254                 GOTO(out, rc = -ENOENT);
1255
1256         rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1257         if (rc)
1258                 GOTO(out, rc);
1259
1260         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1261                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1262                        o->do_lu.lo_dev->ld_obd->obd_name,
1263                        lgi->lgi_attr.la_mode);
1264                 GOTO(out, rc = -ENOENT);
1265         }
1266
1267         th = dt_trans_create(env, d);
1268         if (IS_ERR(th))
1269                 GOTO(out, rc = PTR_ERR(th));
1270
1271         rc = dt_declare_record_write(env, o, size, lgi->lgi_off, th);
1272         if (rc)
1273                 GOTO(out, rc);
1274
1275         rc = dt_trans_start_local(env, d, th);
1276         if (rc)
1277                 GOTO(out_trans, rc);
1278
1279         lgi->lgi_buf.lb_buf = idarray;
1280         lgi->lgi_buf.lb_len = size;
1281         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1282         if (rc)
1283                 CDEBUG(D_INODE, "error writeing CATALOGS: rc = %d\n", rc);
1284 out_trans:
1285         dt_trans_stop(env, d, th);
1286 out:
1287         lu_object_put(env, &o->do_lu);
1288         return rc;
1289 }
1290 EXPORT_SYMBOL(llog_osd_put_cat_list);