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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2014 Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32 /*
  33  * lustre/obdclass/llog_osd.c
  34  *
  35  * Low level llog routines on top of OSD API
  36  *
  37  * This file provides set of methods for llog operations on top of
  38  * dt_device. It contains all supported llog_operations interfaces and
  39  * supplimental functions.
  40  *
  41  * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
  42  * Author: Mikhail Pershin <mike.pershin@intel.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_LOG
  46
  47 #include <obd.h>
  48 #include <obd_class.h>
  49 #include <lustre_fid.h>
  50 #include <dt_object.h>
  51
  52 #include "llog_internal.h"
  53 #include "local_storage.h"
  54
  55 /**
  56  * Implementation of the llog_operations::lop_declare_create
  57  *
  58  * This function is a wrapper over local_storage API function
  59  * local_object_declare_create().
  60  *
  61  * \param[in] env       execution environment
  62  * \param[in] los       local_storage for bottom storage device
  63  * \param[in] o         dt_object to create
  64  * \param[in] th        current transaction handle
  65  *
  66  * \retval              0 on successful declaration of the new object
  67  * \retval              negative error if declaration was failed
  68  */
  69 static int llog_osd_declare_new_object(const struct lu_env *env,
  70                                        struct local_oid_storage *los,
  71                                        struct dt_object *o,
  72                                        struct thandle *th)
  73 {
  74         struct llog_thread_info *lgi = llog_info(env);
  75
  76         lgi->lgi_attr.la_valid = LA_MODE;
  77         lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
  78         lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
  79
  80         return local_object_declare_create(env, los, o, &lgi->lgi_attr,
  81                                            &lgi->lgi_dof, th);
  82 }
  83
  84 /**
  85  * Implementation of the llog_operations::lop_create
  86  *
  87  * This function is a wrapper over local_storage API function
  88  * local_object_create().
  89  *
  90  * \param[in] env       execution environment
  91  * \param[in] los       local_storage for bottom storage device
  92  * \param[in] o         dt_object to create
  93  * \param[in] th        current transaction handle
  94  *
  95  * \retval              0 on successful creation of the new object
  96  * \retval              negative error if creation was failed
  97  */
  98 static int llog_osd_create_new_object(const struct lu_env *env,
  99                                       struct local_oid_storage *los,
 100                                       struct dt_object *o,
 101                                       struct thandle *th)
 102 {
 103         struct llog_thread_info *lgi = llog_info(env);
 104
 105         lgi->lgi_attr.la_valid = LA_MODE;
 106         lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
 107         lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
 108
 109         return local_object_create(env, los, o, &lgi->lgi_attr,
 110                                    &lgi->lgi_dof, th);
 111 }
 112
 113 /**
 114  * Write a padding record to the llog
 115  *
 116  * This function writes a padding record to the end of llog. That may
 117  * be needed if llog contains records of variable size, e.g. config logs
 118  * or changelogs.
 119  * The padding record just aligns llog to the LLOG_CHUNK_SIZE boundary if
 120  * the current record doesn't fit in the remaining space.
 121  *
 122  * It allocates full length to avoid two separate writes for header and tail.
 123  * Such 2-steps scheme needs extra protection and complex error handling.
 124  *
 125  * \param[in]     env   execution environment
 126  * \param[in]     o     dt_object to create
 127  * \param[in,out] off   pointer to the padding start offset
 128  * \param[in]     len   padding length
 129  * \param[in]     index index of the padding record in a llog
 130  * \param[in]     th    current transaction handle
 131  *
 132  * \retval              0 on successful padding write
 133  * \retval              negative error if write failed
 134  */
 135 static int llog_osd_pad(const struct lu_env *env, struct dt_object *o,
 136                         loff_t *off, int len, int index, struct thandle *th)
 137 {
 138         struct llog_thread_info *lgi = llog_info(env);
 139         struct llog_rec_hdr     *rec;
 140         struct llog_rec_tail    *tail;
 141         int                      rc;
 142
 143         ENTRY;
 144
 145         LASSERT(th);
 146         LASSERT(off);
 147         LASSERT(len >= LLOG_MIN_REC_SIZE && (len & 0x7) == 0);
 148
 149         OBD_ALLOC(rec, len);
 150         if (rec == NULL)
 151                 RETURN(-ENOMEM);
 152
 153         rec->lrh_len = len;
 154         rec->lrh_index = index;
 155         rec->lrh_type = LLOG_PAD_MAGIC;
 156
 157         tail = rec_tail(rec);
 158         tail->lrt_len = len;
 159         tail->lrt_index = index;
 160
 161         lgi->lgi_buf.lb_buf = rec;
 162         lgi->lgi_buf.lb_len = len;
 163         rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
 164         if (rc)
 165                 CERROR("%s: error writing padding record: rc = %d\n",
 166                        o->do_lu.lo_dev->ld_obd->obd_name, rc);
 167
 168         OBD_FREE(rec, len);
 169         RETURN(rc);
 170 }
 171
 172 /**
 173  * Implementation of the llog_operations::lop_read_header
 174  *
 175  * This function reads the current llog header from the bottom storage
 176  * device.
 177  *
 178  * \param[in] env       execution environment
 179  * \param[in] handle    llog handle of the current llog
 180  *
 181  * \retval              0 on successful header read
 182  * \retval              negative error if read failed
 183  */
 184 static int llog_osd_read_header(const struct lu_env *env,
 185                                 struct llog_handle *handle)
 186 {
 187         struct llog_rec_hdr     *llh_hdr;
 188         struct dt_object        *o;
 189         struct llog_thread_info *lgi;
 190         enum llog_flag           flags;
 191         int                      rc;
 192
 193         ENTRY;
 194
 195         LASSERT(sizeof(*handle->lgh_hdr) == LLOG_CHUNK_SIZE);
 196
 197         o = handle->lgh_obj;
 198         LASSERT(o);
 199
 200         lgi = llog_info(env);
 201
 202         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 203         if (rc)
 204                 RETURN(rc);
 205
 206         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 207
 208         if (lgi->lgi_attr.la_size == 0) {
 209                 CDEBUG(D_HA, "not reading header from 0-byte log\n");
 210                 RETURN(LLOG_EEMPTY);
 211         }
 212
 213         flags = handle->lgh_hdr->llh_flags;
 214
 215         lgi->lgi_off = 0;
 216         lgi->lgi_buf.lb_buf = handle->lgh_hdr;
 217         lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE;
 218
 219         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
 220         if (rc) {
 221                 CERROR("%s: error reading log header from "DFID": rc = %d\n",
 222                        o->do_lu.lo_dev->ld_obd->obd_name,
 223                        PFID(lu_object_fid(&o->do_lu)), rc);
 224                 RETURN(rc);
 225         }
 226
 227         llh_hdr = &handle->lgh_hdr->llh_hdr;
 228         if (LLOG_REC_HDR_NEEDS_SWABBING(llh_hdr))
 229                 lustre_swab_llog_hdr(handle->lgh_hdr);
 230
 231         if (llh_hdr->lrh_type != LLOG_HDR_MAGIC) {
 232                 CERROR("%s: bad log %s "DFID" header magic: %#x "
 233                        "(expected %#x)\n", o->do_lu.lo_dev->ld_obd->obd_name,
 234                        handle->lgh_name ? handle->lgh_name : "",
 235                        PFID(lu_object_fid(&o->do_lu)),
 236                        llh_hdr->lrh_type, LLOG_HDR_MAGIC);
 237                 RETURN(-EIO);
 238         } else if (llh_hdr->lrh_len != LLOG_CHUNK_SIZE) {
 239                 CERROR("%s: incorrectly sized log %s "DFID" header: "
 240                        "%#x (expected %#x)\n"
 241                        "you may need to re-run lconf --write_conf.\n",
 242                        o->do_lu.lo_dev->ld_obd->obd_name,
 243                        handle->lgh_name ? handle->lgh_name : "",
 244                        PFID(lu_object_fid(&o->do_lu)),
 245                        llh_hdr->lrh_len, LLOG_CHUNK_SIZE);
 246                 RETURN(-EIO);
 247         }
 248
 249         handle->lgh_hdr->llh_flags |= (flags & LLOG_F_EXT_MASK);
 250         handle->lgh_last_idx = handle->lgh_hdr->llh_tail.lrt_index;
 251
 252         RETURN(0);
 253 }
 254
 255 /**
 256  * Implementation of the llog_operations::lop_declare_write
 257  *
 258  * This function declares the new record write.
 259  *
 260  * \param[in] env       execution environment
 261  * \param[in] loghandle llog handle of the current llog
 262  * \param[in] rec       llog record header. This is a real header of the full
 263  *                      llog record to write. This is the beginning of buffer
 264  *                      to write, the length of buffer is stored in
 265  *                      \a rec::lrh_len
 266  * \param[in] idx       index of the llog record. If \a idx == -1 then this is
 267  *                      append case, otherwise \a idx is the index of record
 268  *                      to modify
 269  * \param[in] th        current transaction handle
 270  *
 271  * \retval              0 on successful declaration
 272  * \retval              negative error if declaration failed
 273  */
 274 static int llog_osd_declare_write_rec(const struct lu_env *env,
 275                                       struct llog_handle *loghandle,
 276                                       struct llog_rec_hdr *rec,
 277                                       int idx, struct thandle *th)
 278 {
 279         struct llog_thread_info *lgi = llog_info(env);
 280         struct dt_object        *o;
 281         int                      rc;
 282
 283         ENTRY;
 284
 285         LASSERT(env);
 286         LASSERT(th);
 287         LASSERT(loghandle);
 288         LASSERT(rec);
 289         LASSERT(rec->lrh_len <= LLOG_CHUNK_SIZE);
 290
 291         o = loghandle->lgh_obj;
 292         LASSERT(o);
 293
 294         lgi->lgi_buf.lb_len = sizeof(struct llog_log_hdr);
 295         lgi->lgi_buf.lb_buf = NULL;
 296         /* each time we update header */
 297         rc = dt_declare_record_write(env, o, &lgi->lgi_buf, 0,
 298                                      th);
 299         if (rc || idx == 0) /* if error or just header */
 300                 RETURN(rc);
 301
 302         /**
 303          * the pad record can be inserted so take into account double
 304          * record size
 305          */
 306         lgi->lgi_buf.lb_len = rec->lrh_len * 2;
 307         lgi->lgi_buf.lb_buf = NULL;
 308         /* XXX: implement declared window or multi-chunks approach */
 309         rc = dt_declare_record_write(env, o, &lgi->lgi_buf, -1, th);
 310
 311         RETURN(rc);
 312 }
 313
 314 /**
 315  * Implementation of the llog_operations::lop_write
 316  *
 317  * This function writes the new record in the llog or modify the existed one.
 318  *
 319  * \param[in]  env              execution environment
 320  * \param[in]  loghandle        llog handle of the current llog
 321  * \param[in]  rec              llog record header. This is a real header of
 322  *                              the full llog record to write. This is
 323  *                              the beginning of buffer to write, the length
 324  *                              of buffer is stored in \a rec::lrh_len
 325  * \param[out] reccookie        pointer to the cookie to return back if needed.
 326  *                              It is used for further cancel of this llog
 327  *                              record.
 328  * \param[in]  idx              index of the llog record. If \a idx == -1 then
 329  *                              this is append case, otherwise \a idx is
 330  *                              the index of record to modify
 331  * \param[in]  th               current transaction handle
 332  *
 333  * \retval                      0 on successful write && \a reccookie == NULL
 334  *                              1 on successful write && \a reccookie != NULL
 335  * \retval                      negative error if write failed
 336  */
 337 static int llog_osd_write_rec(const struct lu_env *env,
 338                               struct llog_handle *loghandle,
 339                               struct llog_rec_hdr *rec,
 340                               struct llog_cookie *reccookie,
 341                               int idx, struct thandle *th)
 342 {
 343         struct llog_thread_info *lgi = llog_info(env);
 344         struct llog_log_hdr     *llh;
 345         int                      reclen = rec->lrh_len;
 346         int                      index, rc;
 347         struct llog_rec_tail    *lrt;
 348         struct dt_object        *o;
 349         size_t                   left;
 350         bool                     header_is_updated = false;
 351
 352         ENTRY;
 353
 354         LASSERT(env);
 355         llh = loghandle->lgh_hdr;
 356         LASSERT(llh);
 357         o = loghandle->lgh_obj;
 358         LASSERT(o);
 359         LASSERT(th);
 360
 361         CDEBUG(D_OTHER, "new record %x to "DFID"\n",
 362                rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
 363
 364         /* record length should not bigger than LLOG_CHUNK_SIZE */
 365         if (reclen > LLOG_CHUNK_SIZE)
 366                 RETURN(-E2BIG);
 367
 368         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 369         if (rc)
 370                 RETURN(rc);
 371
 372         /**
 373          * The modification case.
 374          * If idx set then the record with that index must be modified.
 375          * There are three cases possible:
 376          * 1) the common case is the llog header update (idx == 0)
 377          * 2) the llog record modification during llog process.
 378          *    This is indicated by the \a loghandle::lgh_cur_idx > 0.
 379          *    In that case the \a loghandle::lgh_cur_offset
 380          * 3) otherwise this is assumed that llog consist of records of
 381          *    fixed size, i.e. catalog. The llog header must has llh_size
 382          *    field equal to record size. The record offset is calculated
 383          *    just by /a idx value
 384          *
 385          * During modification we don't need extra header update because
 386          * the bitmap and record count are not changed. The record header
 387          * and tail remains the same too.
 388          */
 389         if (idx != LLOG_NEXT_IDX) {
 390                 /* llog can be empty only when first record is being written */
 391                 LASSERT(ergo(idx > 0, lgi->lgi_attr.la_size > 0));
 392
 393                 if (!ext2_test_bit(idx, llh->llh_bitmap)) {
 394                         CERROR("%s: modify unset record %u\n",
 395                                o->do_lu.lo_dev->ld_obd->obd_name, idx);
 396                         RETURN(-ENOENT);
 397                 }
 398
 399                 if (idx != rec->lrh_index) {
 400                         CERROR("%s: modify index mismatch %d %u\n",
 401                                o->do_lu.lo_dev->ld_obd->obd_name, idx,
 402                                rec->lrh_index);
 403                         RETURN(-EFAULT);
 404                 }
 405
 406                 if (idx == LLOG_HEADER_IDX) {
 407                         /* llog header update */
 408                         LASSERT(reclen == sizeof(struct llog_log_hdr));
 409                         LASSERT(rec == &llh->llh_hdr);
 410
 411                         lgi->lgi_off = 0;
 412                         lgi->lgi_buf.lb_len = reclen;
 413                         lgi->lgi_buf.lb_buf = rec;
 414                         rc = dt_record_write(env, o, &lgi->lgi_buf,
 415                                              &lgi->lgi_off, th);
 416                         RETURN(rc);
 417                 } else if (loghandle->lgh_cur_idx > 0) {
 418                         /**
 419                          * The lgh_cur_offset can be used only if index is
 420                          * the same.
 421                          */
 422                         if (idx != loghandle->lgh_cur_idx) {
 423                                 CERROR("%s: modify index mismatch %d %d\n",
 424                                        o->do_lu.lo_dev->ld_obd->obd_name, idx,
 425                                        loghandle->lgh_cur_idx);
 426                                 RETURN(-EFAULT);
 427                         }
 428
 429                         lgi->lgi_off = loghandle->lgh_cur_offset;
 430                         CDEBUG(D_OTHER, "modify record "DOSTID": idx:%d, "
 431                                "len:%u offset %llu\n",
 432                                POSTID(&loghandle->lgh_id.lgl_oi), idx,
 433                                rec->lrh_len, (long long)lgi->lgi_off);
 434                 } else if (llh->llh_size > 0) {
 435                         if (llh->llh_size != rec->lrh_len) {
 436                                 CERROR("%s: wrong record size, llh_size is %u"
 437                                        " but record size is %u\n",
 438                                        o->do_lu.lo_dev->ld_obd->obd_name,
 439                                        llh->llh_size, rec->lrh_len);
 440                                 RETURN(-EINVAL);
 441                         }
 442                         lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
 443                 } else {
 444                         /* This can be result of lgh_cur_idx is not set during
 445                          * llog processing or llh_size is not set to proper
 446                          * record size for fixed records llog. Therefore it is
 447                          * impossible to get record offset. */
 448                         CERROR("%s: can't get record offset, idx:%d, "
 449                                "len:%u.\n", o->do_lu.lo_dev->ld_obd->obd_name,
 450                                idx, rec->lrh_len);
 451                         RETURN(-EFAULT);
 452                 }
 453
 454                 /* update only data, header and tail remain the same */
 455                 lgi->lgi_off += sizeof(struct llog_rec_hdr);
 456                 lgi->lgi_buf.lb_len = REC_DATA_LEN(rec);
 457                 lgi->lgi_buf.lb_buf = REC_DATA(rec);
 458                 rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 459                 if (rc == 0 && reccookie) {
 460                         reccookie->lgc_lgl = loghandle->lgh_id;
 461                         reccookie->lgc_index = idx;
 462                         rc = 1;
 463                 }
 464                 RETURN(rc);
 465         }
 466
 467         /**
 468          * The append case.
 469          * The most common case of using llog. The new index is assigned to
 470          * the new record, new bit is set in llog bitmap and llog count is
 471          * incremented.
 472          *
 473          * Make sure that records don't cross a chunk boundary, so we can
 474          * process them page-at-a-time if needed.  If it will cross a chunk
 475          * boundary, write in a fake (but referenced) entry to pad the chunk.
 476          */
 477         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 478         lgi->lgi_off = lgi->lgi_attr.la_size;
 479         left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
 480         /* NOTE: padding is a record, but no bit is set */
 481         if (left != 0 && left != reclen &&
 482             left < (reclen + LLOG_MIN_REC_SIZE)) {
 483                 index = loghandle->lgh_last_idx + 1;
 484                 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
 485                 if (rc)
 486                         RETURN(rc);
 487                 loghandle->lgh_last_idx++; /* for pad rec */
 488         }
 489         /* if it's the last idx in log file, then return -ENOSPC */
 490         if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
 491                 RETURN(-ENOSPC);
 492
 493         /* increment the last_idx along with llh_tail index, they should
 494          * be equal for a llog lifetime */
 495         loghandle->lgh_last_idx++;
 496         index = loghandle->lgh_last_idx;
 497         llh->llh_tail.lrt_index = index;
 498         /**
 499          * NB: the caller should make sure only 1 process access
 500          * the lgh_last_idx, e.g. append should be exclusive.
 501          * Otherwise it might hit the assert.
 502          */
 503         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 504         rec->lrh_index = index;
 505         lrt = rec_tail(rec);
 506         lrt->lrt_len = rec->lrh_len;
 507         lrt->lrt_index = rec->lrh_index;
 508
 509         /* the lgh_hdr_lock protects llog header data from concurrent
 510          * update/cancel, the llh_count and llh_bitmap are protected */
 511         spin_lock(&loghandle->lgh_hdr_lock);
 512         if (ext2_set_bit(index, llh->llh_bitmap)) {
 513                 CERROR("%s: index %u already set in log bitmap\n",
 514                        o->do_lu.lo_dev->ld_obd->obd_name, index);
 515                 spin_unlock(&loghandle->lgh_hdr_lock);
 516                 LBUG(); /* should never happen */
 517         }
 518         llh->llh_count++;
 519         spin_unlock(&loghandle->lgh_hdr_lock);
 520
 521         lgi->lgi_off = 0;
 522         lgi->lgi_buf.lb_len = llh->llh_hdr.lrh_len;
 523         lgi->lgi_buf.lb_buf = &llh->llh_hdr;
 524         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 525         if (rc)
 526                 GOTO(out, rc);
 527
 528         header_is_updated = true;
 529         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 530         if (rc)
 531                 GOTO(out, rc);
 532
 533         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 534         lgi->lgi_off = lgi->lgi_attr.la_size;
 535         lgi->lgi_buf.lb_len = reclen;
 536         lgi->lgi_buf.lb_buf = rec;
 537         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 538         if (rc < 0)
 539                 GOTO(out, rc);
 540
 541         CDEBUG(D_OTHER, "added record "DOSTID": idx: %u, %u\n",
 542                POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len);
 543         if (reccookie != NULL) {
 544                 reccookie->lgc_lgl = loghandle->lgh_id;
 545                 reccookie->lgc_index = index;
 546                 if ((rec->lrh_type == MDS_UNLINK_REC) ||
 547                     (rec->lrh_type == MDS_SETATTR64_REC))
 548                         reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 549                 else if (rec->lrh_type == OST_SZ_REC)
 550                         reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
 551                 else
 552                         reccookie->lgc_subsys = -1;
 553                 rc = 1;
 554         }
 555         RETURN(rc);
 556 out:
 557         /* cleanup llog for error case */
 558         spin_lock(&loghandle->lgh_hdr_lock);
 559         ext2_clear_bit(index, llh->llh_bitmap);
 560         llh->llh_count--;
 561         spin_unlock(&loghandle->lgh_hdr_lock);
 562
 563         /* restore llog last_idx */
 564         loghandle->lgh_last_idx--;
 565         llh->llh_tail.lrt_index = loghandle->lgh_last_idx;
 566
 567         /* restore the header on disk if it was written */
 568         if (header_is_updated) {
 569                 lgi->lgi_off = 0;
 570                 lgi->lgi_buf.lb_len = llh->llh_hdr.lrh_len;
 571                 lgi->lgi_buf.lb_buf = &llh->llh_hdr;
 572                 dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 573         }
 574
 575         RETURN(rc);
 576 }
 577
 578 /**
 579  * We can skip reading at least as many log blocks as the number of
 580  * minimum sized log records we are skipping.  If it turns out
 581  * that we are not far enough along the log (because the
 582  * actual records are larger than minimum size) we just skip
 583  * some more records.
 584  */
 585 static inline void llog_skip_over(__u64 *off, int curr, int goal)
 586 {
 587         if (goal <= curr)
 588                 return;
 589         *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
 590                 ~(LLOG_CHUNK_SIZE - 1);
 591 }
 592
 593 /**
 594  * Remove optional fields that the client doesn't expect.
 595  * This is typically in order to ensure compatibility with older clients.
 596  * It is assumed that since we exclusively remove fields, the block will be
 597  * big enough to handle the remapped records. It is also assumed that records
 598  * of a block have the same format (i.e.: the same features enabled).
 599  *
 600  * \param[in,out]    hdr        Header of the block of records to remap.
 601  * \param[in,out]    last_hdr   Last header, don't read past this point.
 602  * \param[in]        flags      Flags describing the fields to keep.
 603  */
 604 static void changelog_block_trim_ext(struct llog_rec_hdr *hdr,
 605                                      struct llog_rec_hdr *last_hdr,
 606                                      enum changelog_rec_flags flags)
 607 {
 608         if (hdr->lrh_type != CHANGELOG_REC)
 609                 return;
 610
 611         do {
 612                 struct changelog_rec *rec = (struct changelog_rec *)(hdr + 1);
 613
 614                 changelog_remap_rec(rec, rec->cr_flags & flags);
 615                 hdr = llog_rec_hdr_next(hdr);
 616         } while ((char *)hdr <= (char *)last_hdr);
 617 }
 618
 619 /**
 620  * Implementation of the llog_operations::lop_next_block
 621  *
 622  * This function finds the the next llog block to return which contains
 623  * record with required index. It is main part of llog processing.
 624  *
 625  * \param[in]     env           execution environment
 626  * \param[in]     loghandle     llog handle of the current llog
 627  * \param[in,out] cur_idx       index preceeding cur_offset
 628  * \param[in]     next_idx      target index to find
 629  * \param[in,out] cur_offset    furtherst point read in the file
 630  * \param[in]     buf           pointer to data buffer to fill
 631  * \param[in]     len           required len to read, it is
 632  *                              LLOG_CHUNK_SIZE usually.
 633  *
 634  * \retval                      0 on successful buffer read
 635  * \retval                      negative value on error
 636  */
 637 static int llog_osd_next_block(const struct lu_env *env,
 638                                struct llog_handle *loghandle, int *cur_idx,
 639                                int next_idx, __u64 *cur_offset, void *buf,
 640                                int len)
 641 {
 642         struct llog_thread_info *lgi = llog_info(env);
 643         struct dt_object        *o;
 644         struct dt_device        *dt;
 645         int                      rc;
 646
 647         ENTRY;
 648
 649         LASSERT(env);
 650         LASSERT(lgi);
 651
 652         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 653                 RETURN(-EINVAL);
 654
 655         CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
 656                next_idx, *cur_idx, *cur_offset);
 657
 658         LASSERT(loghandle);
 659         LASSERT(loghandle->lgh_ctxt);
 660
 661         o = loghandle->lgh_obj;
 662         LASSERT(o);
 663         LASSERT(dt_object_exists(o));
 664         dt = lu2dt_dev(o->do_lu.lo_dev);
 665         LASSERT(dt);
 666
 667         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 668         if (rc)
 669                 GOTO(out, rc);
 670
 671         while (*cur_offset < lgi->lgi_attr.la_size) {
 672                 struct llog_rec_hdr     *rec, *last_rec;
 673                 struct llog_rec_tail    *tail;
 674
 675                 llog_skip_over(cur_offset, *cur_idx, next_idx);
 676
 677                 /* read up to next LLOG_CHUNK_SIZE block */
 678                 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
 679                                       (*cur_offset & (LLOG_CHUNK_SIZE - 1));
 680                 lgi->lgi_buf.lb_buf = buf;
 681
 682                 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
 683                 if (rc < 0) {
 684                         CERROR("%s: can't read llog block from log "DFID
 685                                " offset "LPU64": rc = %d\n",
 686                                o->do_lu.lo_dev->ld_obd->obd_name,
 687                                PFID(lu_object_fid(&o->do_lu)), *cur_offset,
 688                                rc);
 689                         GOTO(out, rc);
 690                 }
 691
 692                 if (rc < len) {
 693                         /* signal the end of the valid buffer to
 694                          * llog_process */
 695                         memset(buf + rc, 0, len - rc);
 696                 }
 697
 698                 if (rc == 0) /* end of file, nothing to do */
 699                         GOTO(out, rc);
 700
 701                 if (rc < sizeof(*tail)) {
 702                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 703                                "offset "LPU64"\n",
 704                                o->do_lu.lo_dev->ld_obd->obd_name,
 705                                POSTID(&loghandle->lgh_id.lgl_oi),
 706                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 707                         GOTO(out, rc = -EINVAL);
 708                 }
 709
 710                 rec = buf;
 711                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 712                         lustre_swab_llog_rec(rec);
 713
 714                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 715                                                 sizeof(struct llog_rec_tail));
 716                 /* get the last record in block */
 717                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 718                                                    tail->lrt_len);
 719
 720                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 721                         lustre_swab_llog_rec(last_rec);
 722                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 723
 724                 *cur_idx = tail->lrt_index;
 725
 726                 /* this shouldn't happen */
 727                 if (tail->lrt_index == 0) {
 728                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 729                                "offset "LPU64"\n",
 730                                o->do_lu.lo_dev->ld_obd->obd_name,
 731                                POSTID(&loghandle->lgh_id.lgl_oi),
 732                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 733                         GOTO(out, rc = -EINVAL);
 734                 }
 735                 if (tail->lrt_index < next_idx)
 736                         continue;
 737
 738                 /* sanity check that the start of the new buffer is no farther
 739                  * than the record that we wanted.  This shouldn't happen. */
 740                 if (rec->lrh_index > next_idx) {
 741                         CERROR("%s: missed desired record? %u > %u\n",
 742                                o->do_lu.lo_dev->ld_obd->obd_name,
 743                                rec->lrh_index, next_idx);
 744                         GOTO(out, rc = -ENOENT);
 745                 }
 746
 747                 /* Trim unsupported extensions for compat w/ older clients */
 748                 if (!(loghandle->lgh_hdr->llh_flags & LLOG_F_EXT_JOBID))
 749                         changelog_block_trim_ext(rec, last_rec,
 750                                                  CLF_VERSION | CLF_RENAME);
 751
 752                 GOTO(out, rc = 0);
 753         }
 754         GOTO(out, rc = -EIO);
 755 out:
 756         return rc;
 757 }
 758
 759 /**
 760  * Implementation of the llog_operations::lop_prev_block
 761  *
 762  * This function finds the llog block to return which contains
 763  * record with required index but in reverse order - from end of llog
 764  * to the beginning.
 765  * It is main part of reverse llog processing.
 766  *
 767  * \param[in] env       execution environment
 768  * \param[in] loghandle llog handle of the current llog
 769  * \param[in] prev_idx  target index to find
 770  * \param[in] buf       pointer to data buffer to fill
 771  * \param[in] len       required len to read, it is LLOG_CHUNK_SIZE usually.
 772  *
 773  * \retval              0 on successful buffer read
 774  * \retval              negative value on error
 775  */
 776 static int llog_osd_prev_block(const struct lu_env *env,
 777                                struct llog_handle *loghandle,
 778                                int prev_idx, void *buf, int len)
 779 {
 780         struct llog_thread_info *lgi = llog_info(env);
 781         struct dt_object        *o;
 782         struct dt_device        *dt;
 783         loff_t                   cur_offset;
 784         int                      rc;
 785
 786         ENTRY;
 787
 788         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 789                 RETURN(-EINVAL);
 790
 791         CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
 792
 793         LASSERT(loghandle);
 794         LASSERT(loghandle->lgh_ctxt);
 795
 796         o = loghandle->lgh_obj;
 797         LASSERT(o);
 798         LASSERT(dt_object_exists(o));
 799         dt = lu2dt_dev(o->do_lu.lo_dev);
 800         LASSERT(dt);
 801
 802         cur_offset = LLOG_CHUNK_SIZE;
 803         llog_skip_over(&cur_offset, 0, prev_idx);
 804
 805         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 806         if (rc)
 807                 GOTO(out, rc);
 808
 809         while (cur_offset < lgi->lgi_attr.la_size) {
 810                 struct llog_rec_hdr     *rec, *last_rec;
 811                 struct llog_rec_tail    *tail;
 812
 813                 lgi->lgi_buf.lb_len = len;
 814                 lgi->lgi_buf.lb_buf = buf;
 815                 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
 816                 if (rc < 0) {
 817                         CERROR("%s: can't read llog block from log "DFID
 818                                " offset "LPU64": rc = %d\n",
 819                                o->do_lu.lo_dev->ld_obd->obd_name,
 820                                PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
 821                         GOTO(out, rc);
 822                 }
 823
 824                 if (rc == 0) /* end of file, nothing to do */
 825                         GOTO(out, rc);
 826
 827                 if (rc < sizeof(*tail)) {
 828                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 829                                "offset "LPU64"\n",
 830                                o->do_lu.lo_dev->ld_obd->obd_name,
 831                                POSTID(&loghandle->lgh_id.lgl_oi),
 832                                loghandle->lgh_id.lgl_ogen, cur_offset);
 833                         GOTO(out, rc = -EINVAL);
 834                 }
 835
 836                 rec = buf;
 837                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 838                         lustre_swab_llog_rec(rec);
 839
 840                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 841                                                 sizeof(struct llog_rec_tail));
 842                 /* get the last record in block */
 843                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 844                                                    le32_to_cpu(tail->lrt_len));
 845
 846                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 847                         lustre_swab_llog_rec(last_rec);
 848                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 849
 850                 /* this shouldn't happen */
 851                 if (tail->lrt_index == 0) {
 852                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 853                                "offset "LPU64"\n",
 854                                o->do_lu.lo_dev->ld_obd->obd_name,
 855                                POSTID(&loghandle->lgh_id.lgl_oi),
 856                                loghandle->lgh_id.lgl_ogen, cur_offset);
 857                         GOTO(out, rc = -EINVAL);
 858                 }
 859                 if (tail->lrt_index < prev_idx)
 860                         continue;
 861
 862                 /* sanity check that the start of the new buffer is no farther
 863                  * than the record that we wanted.  This shouldn't happen. */
 864                 if (rec->lrh_index > prev_idx) {
 865                         CERROR("%s: missed desired record? %u > %u\n",
 866                                o->do_lu.lo_dev->ld_obd->obd_name,
 867                                rec->lrh_index, prev_idx);
 868                         GOTO(out, rc = -ENOENT);
 869                 }
 870
 871                 /* Trim unsupported extensions for compat w/ older clients */
 872                 if (!(loghandle->lgh_hdr->llh_flags & LLOG_F_EXT_JOBID))
 873                         changelog_block_trim_ext(rec, last_rec,
 874                                                  CLF_VERSION | CLF_RENAME);
 875
 876                 GOTO(out, rc = 0);
 877         }
 878         GOTO(out, rc = -EIO);
 879 out:
 880         return rc;
 881 }
 882
 883 /**
 884  * This is helper function to get llog directory object. It is used by named
 885  * llog operations to find/insert/delete llog entry from llog directory.
 886  *
 887  * \param[in] env       execution environment
 888  * \param[in] ctxt      llog context
 889  *
 890  * \retval              dt_object of llog directory
 891  * \retval              ERR_PTR of negative value on error
 892  */
 893 static struct dt_object *llog_osd_dir_get(const struct lu_env *env,
 894                                           struct llog_ctxt *ctxt)
 895 {
 896         struct dt_device        *dt;
 897         struct dt_thread_info   *dti = dt_info(env);
 898         struct dt_object        *dir;
 899         int                      rc;
 900
 901         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 902         if (ctxt->loc_dir == NULL) {
 903                 rc = dt_root_get(env, dt, &dti->dti_fid);
 904                 if (rc)
 905                         return ERR_PTR(rc);
 906                 dir = dt_locate(env, dt, &dti->dti_fid);
 907
 908                 if (!IS_ERR(dir) && !dt_try_as_dir(env, dir)) {
 909                         lu_object_put(env, &dir->do_lu);
 910                         return ERR_PTR(-ENOTDIR);
 911                 }
 912         } else {
 913                 lu_object_get(&ctxt->loc_dir->do_lu);
 914                 dir = ctxt->loc_dir;
 915         }
 916
 917         return dir;
 918 }
 919
 920 /**
 921  * Implementation of the llog_operations::lop_open
 922  *
 923  * This function opens the llog by its logid or by name, it may open also
 924  * non existent llog and assing then new id to it.
 925  * The llog_open/llog_close pair works similar to lu_object_find/put,
 926  * the object may not exist prior open. The result of open is just dt_object
 927  * in the llog header.
 928  *
 929  * \param[in] env               execution environment
 930  * \param[in] handle            llog handle of the current llog
 931  * \param[in] logid             logid of llog to open (nameless llog)
 932  * \param[in] name              name of llog to open (named llog)
 933  * \param[in] open_param
 934  *                              LLOG_OPEN_NEW - new llog, may not exist
 935  *                              LLOG_OPEN_EXIST - old llog, must exist
 936  *
 937  * \retval                      0 on successful open, llog_handle::lgh_obj
 938  *                              contains the dt_object of the llog.
 939  * \retval                      negative value on error
 940  */
 941 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
 942                          struct llog_logid *logid, char *name,
 943                          enum llog_open_param open_param)
 944 {
 945         struct llog_thread_info         *lgi = llog_info(env);
 946         struct llog_ctxt                *ctxt = handle->lgh_ctxt;
 947         struct dt_object                *o;
 948         struct dt_device                *dt;
 949         struct ls_device                *ls;
 950         struct local_oid_storage        *los;
 951         int                              rc = 0;
 952
 953         ENTRY;
 954
 955         LASSERT(env);
 956         LASSERT(ctxt);
 957         LASSERT(ctxt->loc_exp);
 958         LASSERT(ctxt->loc_exp->exp_obd);
 959         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 960         LASSERT(dt);
 961
 962         ls = ls_device_get(dt);
 963         if (IS_ERR(ls))
 964                 RETURN(PTR_ERR(ls));
 965
 966         mutex_lock(&ls->ls_los_mutex);
 967         los = dt_los_find(ls, name != NULL ? FID_SEQ_LLOG_NAME : FID_SEQ_LLOG);
 968         mutex_unlock(&ls->ls_los_mutex);
 969         LASSERT(los);
 970         ls_device_put(env, ls);
 971
 972         LASSERT(handle);
 973
 974         if (logid != NULL) {
 975                 logid_to_fid(logid, &lgi->lgi_fid);
 976         } else if (name) {
 977                 struct dt_object *llog_dir;
 978
 979                 llog_dir = llog_osd_dir_get(env, ctxt);
 980                 if (IS_ERR(llog_dir))
 981                         GOTO(out, rc = PTR_ERR(llog_dir));
 982                 dt_read_lock(env, llog_dir, 0);
 983                 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
 984                 dt_read_unlock(env, llog_dir);
 985                 lu_object_put(env, &llog_dir->do_lu);
 986                 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
 987                         /* generate fid for new llog */
 988                         rc = local_object_fid_generate(env, los,
 989                                                        &lgi->lgi_fid);
 990                 }
 991                 if (rc < 0)
 992                         GOTO(out, rc);
 993                 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
 994                 if (handle->lgh_name)
 995                         strcpy(handle->lgh_name, name);
 996                 else
 997                         GOTO(out, rc = -ENOMEM);
 998         } else {
 999                 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
1000                 /* generate fid for new llog */
1001                 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
1002                 if (rc < 0)
1003                         GOTO(out, rc);
1004         }
1005
1006         o = ls_locate(env, ls, &lgi->lgi_fid, NULL);
1007         if (IS_ERR(o))
1008                 GOTO(out_name, rc = PTR_ERR(o));
1009
1010         /* No new llog is expected but doesn't exist */
1011         if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
1012                 GOTO(out_put, rc = -ENOENT);
1013
1014         fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
1015         handle->lgh_obj = o;
1016         handle->private_data = los;
1017         LASSERT(handle->lgh_ctxt);
1018
1019         RETURN(rc);
1020
1021 out_put:
1022         lu_object_put(env, &o->do_lu);
1023 out_name:
1024         if (handle->lgh_name != NULL)
1025                 OBD_FREE(handle->lgh_name, strlen(name) + 1);
1026 out:
1027         dt_los_put(los);
1028         RETURN(rc);
1029 }
1030
1031 /**
1032  * Implementation of the llog_operations::lop_exist
1033  *
1034  * This function checks that llog exists on storage.
1035  *
1036  * \param[in] handle    llog handle of the current llog
1037  *
1038  * \retval              true if llog object exists and is not just destroyed
1039  * \retval              false if llog doesn't exist or just destroyed
1040  */
1041 static int llog_osd_exist(struct llog_handle *handle)
1042 {
1043         LASSERT(handle->lgh_obj);
1044         return (dt_object_exists(handle->lgh_obj) &&
1045                 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
1046 }
1047
1048 /**
1049  * Implementation of the llog_operations::lop_declare_create
1050  *
1051  * This function declares the llog create. It declares also name insert
1052  * into llog directory in case of named llog.
1053  *
1054  * \param[in] env       execution environment
1055  * \param[in] res       llog handle of the current llog
1056  * \param[in] th        current transaction handle
1057  *
1058  * \retval              0 on successful create declaration
1059  * \retval              negative value on error
1060  */
1061 static int llog_osd_declare_create(const struct lu_env *env,
1062                                    struct llog_handle *res, struct thandle *th)
1063 {
1064         struct llog_thread_info         *lgi = llog_info(env);
1065         struct dt_insert_rec            *rec = &lgi->lgi_dt_rec;
1066         struct local_oid_storage        *los;
1067         struct dt_object                *o;
1068         int                              rc;
1069
1070         ENTRY;
1071
1072         LASSERT(res->lgh_obj);
1073         LASSERT(th);
1074
1075         /* object can be created by another thread */
1076         o = res->lgh_obj;
1077         if (dt_object_exists(o))
1078                 RETURN(0);
1079
1080         los = res->private_data;
1081         LASSERT(los);
1082
1083         rc = llog_osd_declare_new_object(env, los, o, th);
1084         if (rc)
1085                 RETURN(rc);
1086
1087         /* do not declare header initialization here as it's declared
1088          * in llog_osd_declare_write_rec() which is always called */
1089
1090         if (res->lgh_name) {
1091                 struct dt_object *llog_dir;
1092
1093                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
1094                 if (IS_ERR(llog_dir))
1095                         RETURN(PTR_ERR(llog_dir));
1096                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
1097                 rec->rec_fid = &lgi->lgi_fid;
1098                 rec->rec_type = S_IFREG;
1099                 rc = dt_declare_insert(env, llog_dir,
1100                                        (struct dt_rec *)rec,
1101                                        (struct dt_key *)res->lgh_name, th);
1102                 lu_object_put(env, &llog_dir->do_lu);
1103                 if (rc)
1104                         CERROR("%s: can't declare named llog %s: rc = %d\n",
1105                                o->do_lu.lo_dev->ld_obd->obd_name,
1106                                res->lgh_name, rc);
1107         }
1108         RETURN(rc);
1109 }
1110
1111 /**
1112  * Implementation of the llog_operations::lop_create
1113  *
1114  * This function creates the llog according with llog_handle::lgh_obj
1115  * and llog_handle::lgh_name.
1116  *
1117  * \param[in] env       execution environment
1118  * \param[in] res       llog handle of the current llog
1119  * \param[in] th        current transaction handle
1120  *
1121  * \retval              0 on successful create
1122  * \retval              negative value on error
1123  */
1124 static int llog_osd_create(const struct lu_env *env, struct llog_handle *res,
1125                            struct thandle *th)
1126 {
1127         struct llog_thread_info *lgi = llog_info(env);
1128         struct dt_insert_rec    *rec = &lgi->lgi_dt_rec;
1129         struct local_oid_storage *los;
1130         struct dt_object        *o;
1131         int                      rc = 0;
1132
1133         ENTRY;
1134
1135         LASSERT(env);
1136         o = res->lgh_obj;
1137         LASSERT(o);
1138
1139         /* llog can be already created */
1140         if (dt_object_exists(o))
1141                 RETURN(-EEXIST);
1142
1143         los = res->private_data;
1144         LASSERT(los);
1145
1146         dt_write_lock(env, o, 0);
1147         if (!dt_object_exists(o))
1148                 rc = llog_osd_create_new_object(env, los, o, th);
1149         else
1150                 rc = -EEXIST;
1151
1152         dt_write_unlock(env, o);
1153         if (rc)
1154                 RETURN(rc);
1155
1156         if (res->lgh_name) {
1157                 struct dt_object *llog_dir;
1158
1159                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
1160                 if (IS_ERR(llog_dir))
1161                         RETURN(PTR_ERR(llog_dir));
1162
1163                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
1164                 rec->rec_fid = &lgi->lgi_fid;
1165                 rec->rec_type = S_IFREG;
1166                 dt_read_lock(env, llog_dir, 0);
1167                 rc = dt_insert(env, llog_dir, (struct dt_rec *)rec,
1168                                (struct dt_key *)res->lgh_name,
1169                                th, 1);
1170                 dt_read_unlock(env, llog_dir);
1171                 lu_object_put(env, &llog_dir->do_lu);
1172                 if (rc)
1173                         CERROR("%s: can't create named llog %s: rc = %d\n",
1174                                o->do_lu.lo_dev->ld_obd->obd_name,
1175                                res->lgh_name, rc);
1176         }
1177         RETURN(rc);
1178 }
1179
1180 /**
1181  * Implementation of the llog_operations::lop_close
1182  *
1183  * This function closes the llog. It just put llog object and referenced
1184  * local storage.
1185  *
1186  * \param[in] env       execution environment
1187  * \param[in] handle    llog handle of the current llog
1188  *
1189  * \retval              0 on successful llog close
1190  * \retval              negative value on error
1191  */
1192 static int llog_osd_close(const struct lu_env *env, struct llog_handle *handle)
1193 {
1194         struct local_oid_storage        *los;
1195         int                              rc = 0;
1196
1197         ENTRY;
1198
1199         LASSERT(handle->lgh_obj);
1200
1201         lu_object_put(env, &handle->lgh_obj->do_lu);
1202
1203         los = handle->private_data;
1204         LASSERT(los);
1205         dt_los_put(los);
1206
1207         if (handle->lgh_name)
1208                 OBD_FREE(handle->lgh_name, strlen(handle->lgh_name) + 1);
1209
1210         RETURN(rc);
1211 }
1212
1213 /**
1214  * Implementation of the llog_operations::lop_destroy
1215  *
1216  * This function destroys the llog and deletes also entry in the
1217  * llog directory in case of named llog. Llog should be opened prior that.
1218  * Destroy method is not part of external transaction and does everything
1219  * inside.
1220  *
1221  * \param[in] env               execution environment
1222  * \param[in] loghandle llog handle of the current llog
1223  *
1224  * \retval              0 on successful destroy
1225  * \retval              negative value on error
1226  */
1227 static int llog_osd_destroy(const struct lu_env *env,
1228                             struct llog_handle *loghandle)
1229 {
1230         struct llog_ctxt        *ctxt;
1231         struct dt_object        *o, *llog_dir = NULL;
1232         struct dt_device        *d;
1233         struct thandle          *th;
1234         char                    *name = NULL;
1235         int                      rc;
1236
1237         ENTRY;
1238
1239         ctxt = loghandle->lgh_ctxt;
1240         LASSERT(ctxt);
1241
1242         o = loghandle->lgh_obj;
1243         LASSERT(o);
1244
1245         d = lu2dt_dev(o->do_lu.lo_dev);
1246         LASSERT(d);
1247         LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
1248
1249         th = dt_trans_create(env, d);
1250         if (IS_ERR(th))
1251                 RETURN(PTR_ERR(th));
1252
1253         if (loghandle->lgh_name) {
1254                 llog_dir = llog_osd_dir_get(env, ctxt);
1255                 if (IS_ERR(llog_dir))
1256                         GOTO(out_trans, rc = PTR_ERR(llog_dir));
1257
1258                 name = loghandle->lgh_name;
1259                 rc = dt_declare_delete(env, llog_dir,
1260                                        (struct dt_key *)name, th);
1261                 if (rc)
1262                         GOTO(out_trans, rc);
1263         }
1264
1265         rc = dt_declare_ref_del(env, o, th);
1266         if (rc < 0)
1267                 GOTO(out_trans, rc);
1268
1269         rc = dt_declare_destroy(env, o, th);
1270         if (rc)
1271                 GOTO(out_trans, rc);
1272
1273         rc = dt_trans_start_local(env, d, th);
1274         if (rc)
1275                 GOTO(out_trans, rc);
1276
1277         dt_write_lock(env, o, 0);
1278         if (dt_object_exists(o)) {
1279                 if (name) {
1280                         dt_read_lock(env, llog_dir, 0);
1281                         rc = dt_delete(env, llog_dir,
1282                                        (struct dt_key *) name,
1283                                        th);
1284                         dt_read_unlock(env, llog_dir);
1285                         if (rc) {
1286                                 CERROR("%s: can't remove llog %s: rc = %d\n",
1287                                        o->do_lu.lo_dev->ld_obd->obd_name,
1288                                        name, rc);
1289                                 GOTO(out_unlock, rc);
1290                         }
1291                 }
1292                 dt_ref_del(env, o, th);
1293                 rc = dt_destroy(env, o, th);
1294                 if (rc)
1295                         GOTO(out_unlock, rc);
1296         }
1297 out_unlock:
1298         dt_write_unlock(env, o);
1299 out_trans:
1300         dt_trans_stop(env, d, th);
1301         if (llog_dir != NULL)
1302                 lu_object_put(env, &llog_dir->do_lu);
1303         RETURN(rc);
1304 }
1305
1306 /**
1307  * Implementation of the llog_operations::lop_setup
1308  *
1309  * This function setup the llog on local storage.
1310  *
1311  * \param[in] env       execution environment
1312  * \param[in] obd       obd device the llog belongs to
1313  * \param[in] olg       the llog group, it is always zero group now.
1314  * \param[in] ctxt_idx  the llog index, it defines the purpose of this llog.
1315  *                      Every new llog type have to use own index.
1316  * \param[in] disk_obd  the storage obd, where llog is stored.
1317  *
1318  * \retval              0 on successful llog setup
1319  * \retval              negative value on error
1320  */
1321 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1322                           struct obd_llog_group *olg, int ctxt_idx,
1323                           struct obd_device *disk_obd)
1324 {
1325         struct llog_thread_info         *lgi = llog_info(env);
1326         struct llog_ctxt                *ctxt;
1327         int                              rc = 0;
1328         ENTRY;
1329
1330         LASSERT(obd);
1331         LASSERT(olg->olg_ctxts[ctxt_idx]);
1332
1333         ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1334         LASSERT(ctxt);
1335
1336         /* initialize data allowing to generate new fids,
1337          * literally we need a sequece */
1338         lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1339         lgi->lgi_fid.f_oid = 1;
1340         lgi->lgi_fid.f_ver = 0;
1341         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1342                                     &lgi->lgi_fid,
1343                                     &ctxt->loc_los_nameless);
1344         if (rc != 0)
1345                 GOTO(out, rc);
1346
1347         lgi->lgi_fid.f_seq = FID_SEQ_LLOG_NAME;
1348         lgi->lgi_fid.f_oid = 1;
1349         lgi->lgi_fid.f_ver = 0;
1350         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1351                                     &lgi->lgi_fid,
1352                                     &ctxt->loc_los_named);
1353         if (rc != 0) {
1354                 local_oid_storage_fini(env, ctxt->loc_los_nameless);
1355                 ctxt->loc_los_nameless = NULL;
1356         }
1357
1358         GOTO(out, rc);
1359
1360 out:
1361         llog_ctxt_put(ctxt);
1362         return rc;
1363 }
1364
1365 /**
1366  * Implementation of the llog_operations::lop_cleanup
1367  *
1368  * This function cleanups the llog on local storage.
1369  *
1370  * \param[in] env       execution environment
1371  * \param[in] ctxt      the llog context
1372  *
1373  * \retval              0
1374  */
1375 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1376 {
1377         if (ctxt->loc_los_nameless != NULL) {
1378                 local_oid_storage_fini(env, ctxt->loc_los_nameless);
1379                 ctxt->loc_los_nameless = NULL;
1380         }
1381
1382         if (ctxt->loc_los_named != NULL) {
1383                 local_oid_storage_fini(env, ctxt->loc_los_named);
1384                 ctxt->loc_los_named = NULL;
1385         }
1386
1387         return 0;
1388 }
1389
1390 struct llog_operations llog_osd_ops = {
1391         .lop_next_block         = llog_osd_next_block,
1392         .lop_prev_block         = llog_osd_prev_block,
1393         .lop_read_header        = llog_osd_read_header,
1394         .lop_destroy            = llog_osd_destroy,
1395         .lop_setup              = llog_osd_setup,
1396         .lop_cleanup            = llog_osd_cleanup,
1397         .lop_open               = llog_osd_open,
1398         .lop_exist              = llog_osd_exist,
1399         .lop_declare_create     = llog_osd_declare_create,
1400         .lop_create             = llog_osd_create,
1401         .lop_declare_write_rec  = llog_osd_declare_write_rec,
1402         .lop_write_rec          = llog_osd_write_rec,
1403         .lop_close              = llog_osd_close,
1404 };
1405 EXPORT_SYMBOL(llog_osd_ops);
1406
1407 /**
1408  * Read the special file which contains the list of llog catalogs IDs
1409  *
1410  * This function reads the CATALOGS file which contains the array of llog
1411  * catalogs IDs. The main purpose of this file is to store OSP llogs indexed
1412  * by OST/MDT number.
1413  *
1414  * \param[in]  env              execution environment
1415  * \param[in]  d                corresponding storage device
1416  * \param[in]  idx              position to start from, usually OST/MDT index
1417  * \param[in]  count            how many catalog IDs to read
1418  * \param[out] idarray          the buffer for the data. If it is NULL then
1419  *                              function returns just number of catalog IDs
1420  *                              in the file.
1421  * \param[in]  fid              LLOG_CATALOGS_OID for CATALOG object
1422  *
1423  * \retval                      0 on successful read of catalog IDs
1424  * \retval                      negative value on error
1425  * \retval                      positive value which is number of records in
1426  *                              the file if \a idarray is NULL
1427  */
1428 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1429                           int idx, int count, struct llog_catid *idarray,
1430                           const struct lu_fid *fid)
1431 {
1432         struct llog_thread_info *lgi = llog_info(env);
1433         struct dt_object        *o = NULL;
1434         struct thandle          *th;
1435         int                      rc, size;
1436
1437         ENTRY;
1438
1439         LASSERT(d);
1440
1441         size = sizeof(*idarray) * count;
1442         lgi->lgi_off = idx *  sizeof(*idarray);
1443
1444         lgi->lgi_fid = *fid;
1445         o = dt_locate(env, d, &lgi->lgi_fid);
1446         if (IS_ERR(o))
1447                 RETURN(PTR_ERR(o));
1448
1449         if (!dt_object_exists(o)) {
1450                 th = dt_trans_create(env, d);
1451                 if (IS_ERR(th))
1452                         GOTO(out, rc = PTR_ERR(th));
1453
1454                 lgi->lgi_attr.la_valid = LA_MODE;
1455                 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1456                 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1457
1458                 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1459                                        &lgi->lgi_dof, th);
1460                 if (rc)
1461                         GOTO(out_trans, rc);
1462
1463                 rc = dt_trans_start_local(env, d, th);
1464                 if (rc)
1465                         GOTO(out_trans, rc);
1466
1467                 dt_write_lock(env, o, 0);
1468                 if (!dt_object_exists(o))
1469                         rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1470                                        &lgi->lgi_dof, th);
1471                 dt_write_unlock(env, o);
1472 out_trans:
1473                 dt_trans_stop(env, d, th);
1474                 if (rc)
1475                         GOTO(out, rc);
1476         }
1477
1478         rc = dt_attr_get(env, o, &lgi->lgi_attr);
1479         if (rc)
1480                 GOTO(out, rc);
1481
1482         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1483                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1484                        o->do_lu.lo_dev->ld_obd->obd_name,
1485                        lgi->lgi_attr.la_mode);
1486                 GOTO(out, rc = -ENOENT);
1487         }
1488
1489         CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1490                (int)lgi->lgi_attr.la_size, size);
1491
1492         /* return just number of llogs */
1493         if (idarray == NULL) {
1494                 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1495                 GOTO(out, rc);
1496         }
1497
1498         /* read for new ost index or for empty file */
1499         memset(idarray, 0, size);
1500         if (lgi->lgi_attr.la_size <= lgi->lgi_off)
1501                 GOTO(out, rc = 0);
1502         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1503                 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1504
1505         lgi->lgi_buf.lb_buf = idarray;
1506         lgi->lgi_buf.lb_len = size;
1507         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1508         /* -EFAULT means the llog is a sparse file. This is not an error
1509          * after arbitrary OST index is supported. */
1510         if (rc < 0 && rc != -EFAULT) {
1511                 CERROR("%s: error reading CATALOGS: rc = %d\n",
1512                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
1513                 GOTO(out, rc);
1514         }
1515
1516         EXIT;
1517 out:
1518         lu_object_put(env, &o->do_lu);
1519         RETURN(rc);
1520 }
1521 EXPORT_SYMBOL(llog_osd_get_cat_list);
1522
1523 /**
1524  * Write the special file which contains the list of llog catalogs IDs
1525  *
1526  * This function writes the CATALOG file which contains the array of llog
1527  * catalogs IDs. It is used mostly to store OSP llogs indexed by OST/MDT
1528  * number.
1529  *
1530  * \param[in]  env      execution environment
1531  * \param[in]  d        corresponding storage device
1532  * \param[in]  idx      position to start from, usually OST/MDT index
1533  * \param[in]  count    how many catalog IDs to write
1534  * \param[out] idarray  the buffer with the data to write.
1535  * \param[in]  fid      LLOG_CATALOGS_OID for CATALOG object
1536  *
1537  * \retval              0 on successful write of catalog IDs
1538  * \retval              negative value on error
1539  */
1540 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1541                           int idx, int count, struct llog_catid *idarray,
1542                           const struct lu_fid *fid)
1543 {
1544         struct llog_thread_info *lgi = llog_info(env);
1545         struct dt_object        *o = NULL;
1546         struct thandle          *th;
1547         int                      rc, size;
1548
1549         if (count == 0)
1550                 RETURN(0);
1551
1552         LASSERT(d);
1553
1554         size = sizeof(*idarray) * count;
1555         lgi->lgi_off = idx * sizeof(*idarray);
1556         lgi->lgi_fid = *fid;
1557
1558         o = dt_locate(env, d, &lgi->lgi_fid);
1559         if (IS_ERR(o))
1560                 RETURN(PTR_ERR(o));
1561
1562         if (!dt_object_exists(o))
1563                 GOTO(out, rc = -ENOENT);
1564
1565         rc = dt_attr_get(env, o, &lgi->lgi_attr);
1566         if (rc)
1567                 GOTO(out, rc);
1568
1569         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1570                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1571                        o->do_lu.lo_dev->ld_obd->obd_name,
1572                        lgi->lgi_attr.la_mode);
1573                 GOTO(out, rc = -ENOENT);
1574         }
1575
1576         th = dt_trans_create(env, d);
1577         if (IS_ERR(th))
1578                 GOTO(out, rc = PTR_ERR(th));
1579
1580         lgi->lgi_buf.lb_len = size;
1581         lgi->lgi_buf.lb_buf = idarray;
1582         rc = dt_declare_record_write(env, o, &lgi->lgi_buf, lgi->lgi_off, th);
1583         if (rc)
1584                 GOTO(out, rc);
1585
1586         rc = dt_trans_start_local(env, d, th);
1587         if (rc)
1588                 GOTO(out_trans, rc);
1589
1590         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1591         if (rc)
1592                 CDEBUG(D_INODE, "can't write CATALOGS at index %d: rc = %d\n",
1593                        idx, rc);
1594 out_trans:
1595         dt_trans_stop(env, d, th);
1596 out:
1597         lu_object_put(env, &o->do_lu);
1598         RETURN(rc);
1599 }
1600 EXPORT_SYMBOL(llog_osd_put_cat_list);