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
 351         ENTRY;
 352
 353         LASSERT(env);
 354         llh = loghandle->lgh_hdr;
 355         LASSERT(llh);
 356         o = loghandle->lgh_obj;
 357         LASSERT(o);
 358         LASSERT(th);
 359
 360         CDEBUG(D_OTHER, "new record %x to "DFID"\n",
 361                rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
 362
 363         /* record length should not bigger than LLOG_CHUNK_SIZE */
 364         if (reclen > LLOG_CHUNK_SIZE)
 365                 RETURN(-E2BIG);
 366
 367         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 368         if (rc)
 369                 RETURN(rc);
 370
 371         /**
 372          * The modification case.
 373          * If idx set then the record with that index must be modified.
 374          * There are three cases possible:
 375          * 1) the common case is the llog header update (idx == 0)
 376          * 2) the llog record modification during llog process.
 377          *    This is indicated by the \a loghandle::lgh_cur_idx > 0.
 378          *    In that case the \a loghandle::lgh_cur_offset
 379          * 3) otherwise this is assumed that llog consist of records of
 380          *    fixed size, i.e. catalog. The llog header must has llh_size
 381          *    field equal to record size. The record offset is calculated
 382          *    just by /a idx value
 383          *
 384          * During modification we don't need extra header update because
 385          * the bitmap and record count are not changed. The record header
 386          * and tail remains the same too.
 387          */
 388         if (idx != LLOG_NEXT_IDX) {
 389                 /* llog can be empty only when first record is being written */
 390                 LASSERT(ergo(idx > 0, lgi->lgi_attr.la_size > 0));
 391
 392                 if (!ext2_test_bit(idx, llh->llh_bitmap)) {
 393                         CERROR("%s: modify unset record %u\n",
 394                                o->do_lu.lo_dev->ld_obd->obd_name, idx);
 395                         RETURN(-ENOENT);
 396                 }
 397
 398                 if (idx != rec->lrh_index) {
 399                         CERROR("%s: modify index mismatch %d %u\n",
 400                                o->do_lu.lo_dev->ld_obd->obd_name, idx,
 401                                rec->lrh_index);
 402                         RETURN(-EFAULT);
 403                 }
 404
 405                 if (idx == LLOG_HEADER_IDX) {
 406                         /* llog header update */
 407                         LASSERT(reclen == sizeof(struct llog_log_hdr));
 408                         LASSERT(rec == &llh->llh_hdr);
 409
 410                         lgi->lgi_off = 0;
 411                         lgi->lgi_buf.lb_len = reclen;
 412                         lgi->lgi_buf.lb_buf = rec;
 413                         rc = dt_record_write(env, o, &lgi->lgi_buf,
 414                                              &lgi->lgi_off, th);
 415                         RETURN(rc);
 416                 } else if (loghandle->lgh_cur_idx > 0) {
 417                         /**
 418                          * The lgh_cur_offset can be used only if index is
 419                          * the same.
 420                          */
 421                         if (idx != loghandle->lgh_cur_idx) {
 422                                 CERROR("%s: modify index mismatch %d %d\n",
 423                                        o->do_lu.lo_dev->ld_obd->obd_name, idx,
 424                                        loghandle->lgh_cur_idx);
 425                                 RETURN(-EFAULT);
 426                         }
 427
 428                         lgi->lgi_off = loghandle->lgh_cur_offset;
 429                         CDEBUG(D_OTHER, "modify record "DOSTID": idx:%d, "
 430                                "len:%u offset %llu\n",
 431                                POSTID(&loghandle->lgh_id.lgl_oi), idx,
 432                                rec->lrh_len, (long long)lgi->lgi_off);
 433                 } else if (llh->llh_size > 0) {
 434                         if (llh->llh_size != rec->lrh_len) {
 435                                 CERROR("%s: wrong record size, llh_size is %u"
 436                                        " but record size is %u\n",
 437                                        o->do_lu.lo_dev->ld_obd->obd_name,
 438                                        llh->llh_size, rec->lrh_len);
 439                                 RETURN(-EINVAL);
 440                         }
 441                         lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
 442                 } else {
 443                         /* This can be result of lgh_cur_idx is not set during
 444                          * llog processing or llh_size is not set to proper
 445                          * record size for fixed records llog. Therefore it is
 446                          * impossible to get record offset. */
 447                         CERROR("%s: can't get record offset, idx:%d, "
 448                                "len:%u.\n", o->do_lu.lo_dev->ld_obd->obd_name,
 449                                idx, rec->lrh_len);
 450                         RETURN(-EFAULT);
 451                 }
 452
 453                 /* update only data, header and tail remain the same */
 454                 lgi->lgi_off += sizeof(struct llog_rec_hdr);
 455                 lgi->lgi_buf.lb_len = REC_DATA_LEN(rec);
 456                 lgi->lgi_buf.lb_buf = REC_DATA(rec);
 457                 rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 458                 if (rc == 0 && reccookie) {
 459                         reccookie->lgc_lgl = loghandle->lgh_id;
 460                         reccookie->lgc_index = idx;
 461                         rc = 1;
 462                 }
 463                 RETURN(rc);
 464         }
 465
 466         /**
 467          * The append case.
 468          * The most common case of using llog. The new index is assigned to
 469          * the new record, new bit is set in llog bitmap and llog count is
 470          * incremented.
 471          *
 472          * Make sure that records don't cross a chunk boundary, so we can
 473          * process them page-at-a-time if needed.  If it will cross a chunk
 474          * boundary, write in a fake (but referenced) entry to pad the chunk.
 475          */
 476         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 477         lgi->lgi_off = lgi->lgi_attr.la_size;
 478         left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
 479         /* NOTE: padding is a record, but no bit is set */
 480         if (left != 0 && left != reclen &&
 481             left < (reclen + LLOG_MIN_REC_SIZE)) {
 482                 index = loghandle->lgh_last_idx + 1;
 483                 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
 484                 if (rc)
 485                         RETURN(rc);
 486                 loghandle->lgh_last_idx++; /* for pad rec */
 487         }
 488         /* if it's the last idx in log file, then return -ENOSPC */
 489         if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
 490                 RETURN(-ENOSPC);
 491
 492         /* increment the last_idx along with llh_tail index, they should
 493          * be equal for a llog lifetime */
 494         loghandle->lgh_last_idx++;
 495         index = loghandle->lgh_last_idx;
 496         llh->llh_tail.lrt_index = index;
 497         /**
 498          * NB: the caller should make sure only 1 process access
 499          * the lgh_last_idx, e.g. append should be exclusive.
 500          * Otherwise it might hit the assert.
 501          */
 502         LASSERT(index < LLOG_BITMAP_SIZE(llh));
 503         rec->lrh_index = index;
 504         lrt = rec_tail(rec);
 505         lrt->lrt_len = rec->lrh_len;
 506         lrt->lrt_index = rec->lrh_index;
 507
 508         /* the lgh_hdr_lock protects llog header data from concurrent
 509          * update/cancel, the llh_count and llh_bitmap are protected */
 510         spin_lock(&loghandle->lgh_hdr_lock);
 511         if (ext2_set_bit(index, llh->llh_bitmap)) {
 512                 CERROR("%s: index %u already set in log bitmap\n",
 513                        o->do_lu.lo_dev->ld_obd->obd_name, index);
 514                 spin_unlock(&loghandle->lgh_hdr_lock);
 515                 LBUG(); /* should never happen */
 516         }
 517         llh->llh_count++;
 518         spin_unlock(&loghandle->lgh_hdr_lock);
 519
 520         if (lgi->lgi_attr.la_size == 0) {
 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 != 0)
 526                         GOTO(out, rc);
 527         } else {
 528                 /* Note: If this is not initialization (size == 0), then do not
 529                  * write the whole header (8k bytes), only update header/tail
 530                  * and bits needs to be updated. Because this update might be
 531                  * part of cross-MDT operation, which needs to write these
 532                  * updates into the update log(32KB limit) and also pack inside
 533                  * the RPC (1MB limit), if we write 8K for each operation, which
 534                  * will cost a lot space, and keep us adding more updates to one
 535                  * update log.*/
 536                 lgi->lgi_off = offsetof(typeof(*llh), llh_count);
 537                 lgi->lgi_buf.lb_len = sizeof(llh->llh_count);
 538                 lgi->lgi_buf.lb_buf = &llh->llh_count;
 539                 rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 540                 if (rc != 0)
 541                         GOTO(out, rc);
 542
 543                 lgi->lgi_off = offsetof(typeof(*llh),
 544                         llh_bitmap[index / (sizeof(*llh->llh_bitmap) * 8)]);
 545                 lgi->lgi_buf.lb_len = sizeof(*llh->llh_bitmap);
 546                 lgi->lgi_buf.lb_buf =
 547                         &llh->llh_bitmap[index/(sizeof(*llh->llh_bitmap)*8)];
 548                 rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 549                 if (rc != 0)
 550                         GOTO(out, rc);
 551
 552                 lgi->lgi_off = offsetof(typeof(*llh), llh_tail);
 553                 lgi->lgi_buf.lb_len = sizeof(llh->llh_tail);
 554                 lgi->lgi_buf.lb_buf = &llh->llh_tail;
 555                 rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 556                 if (rc != 0)
 557                         GOTO(out, rc);
 558         }
 559
 560         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 561         if (rc)
 562                 GOTO(out, rc);
 563
 564         LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
 565         lgi->lgi_off = lgi->lgi_attr.la_size;
 566         lgi->lgi_buf.lb_len = reclen;
 567         lgi->lgi_buf.lb_buf = rec;
 568         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
 569         if (rc < 0)
 570                 GOTO(out, rc);
 571
 572         CDEBUG(D_OTHER, "added record "DOSTID": idx: %u, %u\n",
 573                POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len);
 574         if (reccookie != NULL) {
 575                 reccookie->lgc_lgl = loghandle->lgh_id;
 576                 reccookie->lgc_index = index;
 577                 if ((rec->lrh_type == MDS_UNLINK_REC) ||
 578                     (rec->lrh_type == MDS_SETATTR64_REC))
 579                         reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 580                 else if (rec->lrh_type == OST_SZ_REC)
 581                         reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
 582                 else
 583                         reccookie->lgc_subsys = -1;
 584                 rc = 1;
 585         }
 586         RETURN(rc);
 587 out:
 588         /* cleanup llog for error case */
 589         spin_lock(&loghandle->lgh_hdr_lock);
 590         ext2_clear_bit(index, llh->llh_bitmap);
 591         llh->llh_count--;
 592         spin_unlock(&loghandle->lgh_hdr_lock);
 593
 594         /* restore llog last_idx */
 595         loghandle->lgh_last_idx--;
 596         llh->llh_tail.lrt_index = loghandle->lgh_last_idx;
 597
 598         RETURN(rc);
 599 }
 600
 601 /**
 602  * We can skip reading at least as many log blocks as the number of
 603  * minimum sized log records we are skipping.  If it turns out
 604  * that we are not far enough along the log (because the
 605  * actual records are larger than minimum size) we just skip
 606  * some more records.
 607  */
 608 static inline void llog_skip_over(__u64 *off, int curr, int goal)
 609 {
 610         if (goal <= curr)
 611                 return;
 612         *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
 613                 ~(LLOG_CHUNK_SIZE - 1);
 614 }
 615
 616 /**
 617  * Remove optional fields that the client doesn't expect.
 618  * This is typically in order to ensure compatibility with older clients.
 619  * It is assumed that since we exclusively remove fields, the block will be
 620  * big enough to handle the remapped records. It is also assumed that records
 621  * of a block have the same format (i.e.: the same features enabled).
 622  *
 623  * \param[in,out]    hdr        Header of the block of records to remap.
 624  * \param[in,out]    last_hdr   Last header, don't read past this point.
 625  * \param[in]        flags      Flags describing the fields to keep.
 626  */
 627 static void changelog_block_trim_ext(struct llog_rec_hdr *hdr,
 628                                      struct llog_rec_hdr *last_hdr,
 629                                      enum changelog_rec_flags flags)
 630 {
 631         if (hdr->lrh_type != CHANGELOG_REC)
 632                 return;
 633
 634         do {
 635                 struct changelog_rec *rec = (struct changelog_rec *)(hdr + 1);
 636
 637                 changelog_remap_rec(rec, rec->cr_flags & flags);
 638                 hdr = llog_rec_hdr_next(hdr);
 639         } while ((char *)hdr <= (char *)last_hdr);
 640 }
 641
 642 /**
 643  * Implementation of the llog_operations::lop_next_block
 644  *
 645  * This function finds the the next llog block to return which contains
 646  * record with required index. It is main part of llog processing.
 647  *
 648  * \param[in]     env           execution environment
 649  * \param[in]     loghandle     llog handle of the current llog
 650  * \param[in,out] cur_idx       index preceeding cur_offset
 651  * \param[in]     next_idx      target index to find
 652  * \param[in,out] cur_offset    furtherst point read in the file
 653  * \param[in]     buf           pointer to data buffer to fill
 654  * \param[in]     len           required len to read, it is
 655  *                              LLOG_CHUNK_SIZE usually.
 656  *
 657  * \retval                      0 on successful buffer read
 658  * \retval                      negative value on error
 659  */
 660 static int llog_osd_next_block(const struct lu_env *env,
 661                                struct llog_handle *loghandle, int *cur_idx,
 662                                int next_idx, __u64 *cur_offset, void *buf,
 663                                int len)
 664 {
 665         struct llog_thread_info *lgi = llog_info(env);
 666         struct dt_object        *o;
 667         struct dt_device        *dt;
 668         int                      rc;
 669
 670         ENTRY;
 671
 672         LASSERT(env);
 673         LASSERT(lgi);
 674
 675         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 676                 RETURN(-EINVAL);
 677
 678         CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
 679                next_idx, *cur_idx, *cur_offset);
 680
 681         LASSERT(loghandle);
 682         LASSERT(loghandle->lgh_ctxt);
 683
 684         o = loghandle->lgh_obj;
 685         LASSERT(o);
 686         LASSERT(dt_object_exists(o));
 687         dt = lu2dt_dev(o->do_lu.lo_dev);
 688         LASSERT(dt);
 689
 690         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 691         if (rc)
 692                 GOTO(out, rc);
 693
 694         while (*cur_offset < lgi->lgi_attr.la_size) {
 695                 struct llog_rec_hdr     *rec, *last_rec;
 696                 struct llog_rec_tail    *tail;
 697
 698                 llog_skip_over(cur_offset, *cur_idx, next_idx);
 699
 700                 /* read up to next LLOG_CHUNK_SIZE block */
 701                 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
 702                                       (*cur_offset & (LLOG_CHUNK_SIZE - 1));
 703                 lgi->lgi_buf.lb_buf = buf;
 704
 705                 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
 706                 if (rc < 0) {
 707                         CERROR("%s: can't read llog block from log "DFID
 708                                " offset "LPU64": rc = %d\n",
 709                                o->do_lu.lo_dev->ld_obd->obd_name,
 710                                PFID(lu_object_fid(&o->do_lu)), *cur_offset,
 711                                rc);
 712                         GOTO(out, rc);
 713                 }
 714
 715                 if (rc < len) {
 716                         /* signal the end of the valid buffer to
 717                          * llog_process */
 718                         memset(buf + rc, 0, len - rc);
 719                 }
 720
 721                 if (rc == 0) /* end of file, nothing to do */
 722                         GOTO(out, rc);
 723
 724                 if (rc < sizeof(*tail)) {
 725                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 726                                "offset "LPU64"\n",
 727                                o->do_lu.lo_dev->ld_obd->obd_name,
 728                                POSTID(&loghandle->lgh_id.lgl_oi),
 729                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 730                         GOTO(out, rc = -EINVAL);
 731                 }
 732
 733                 rec = buf;
 734                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 735                         lustre_swab_llog_rec(rec);
 736
 737                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 738                                                 sizeof(struct llog_rec_tail));
 739                 /* get the last record in block */
 740                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 741                                                    tail->lrt_len);
 742
 743                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 744                         lustre_swab_llog_rec(last_rec);
 745                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 746
 747                 *cur_idx = tail->lrt_index;
 748
 749                 /* this shouldn't happen */
 750                 if (tail->lrt_index == 0) {
 751                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 752                                "offset "LPU64"\n",
 753                                o->do_lu.lo_dev->ld_obd->obd_name,
 754                                POSTID(&loghandle->lgh_id.lgl_oi),
 755                                loghandle->lgh_id.lgl_ogen, *cur_offset);
 756                         GOTO(out, rc = -EINVAL);
 757                 }
 758                 if (tail->lrt_index < next_idx)
 759                         continue;
 760
 761                 /* sanity check that the start of the new buffer is no farther
 762                  * than the record that we wanted.  This shouldn't happen. */
 763                 if (rec->lrh_index > next_idx) {
 764                         CERROR("%s: missed desired record? %u > %u\n",
 765                                o->do_lu.lo_dev->ld_obd->obd_name,
 766                                rec->lrh_index, next_idx);
 767                         GOTO(out, rc = -ENOENT);
 768                 }
 769
 770                 /* Trim unsupported extensions for compat w/ older clients */
 771                 if (!(loghandle->lgh_hdr->llh_flags & LLOG_F_EXT_JOBID))
 772                         changelog_block_trim_ext(rec, last_rec,
 773                                                  CLF_VERSION | CLF_RENAME);
 774
 775                 GOTO(out, rc = 0);
 776         }
 777         GOTO(out, rc = -EIO);
 778 out:
 779         return rc;
 780 }
 781
 782 /**
 783  * Implementation of the llog_operations::lop_prev_block
 784  *
 785  * This function finds the llog block to return which contains
 786  * record with required index but in reverse order - from end of llog
 787  * to the beginning.
 788  * It is main part of reverse llog processing.
 789  *
 790  * \param[in] env       execution environment
 791  * \param[in] loghandle llog handle of the current llog
 792  * \param[in] prev_idx  target index to find
 793  * \param[in] buf       pointer to data buffer to fill
 794  * \param[in] len       required len to read, it is LLOG_CHUNK_SIZE usually.
 795  *
 796  * \retval              0 on successful buffer read
 797  * \retval              negative value on error
 798  */
 799 static int llog_osd_prev_block(const struct lu_env *env,
 800                                struct llog_handle *loghandle,
 801                                int prev_idx, void *buf, int len)
 802 {
 803         struct llog_thread_info *lgi = llog_info(env);
 804         struct dt_object        *o;
 805         struct dt_device        *dt;
 806         loff_t                   cur_offset;
 807         int                      rc;
 808
 809         ENTRY;
 810
 811         if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
 812                 RETURN(-EINVAL);
 813
 814         CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
 815
 816         LASSERT(loghandle);
 817         LASSERT(loghandle->lgh_ctxt);
 818
 819         o = loghandle->lgh_obj;
 820         LASSERT(o);
 821         LASSERT(dt_object_exists(o));
 822         dt = lu2dt_dev(o->do_lu.lo_dev);
 823         LASSERT(dt);
 824
 825         cur_offset = LLOG_CHUNK_SIZE;
 826         llog_skip_over(&cur_offset, 0, prev_idx);
 827
 828         rc = dt_attr_get(env, o, &lgi->lgi_attr);
 829         if (rc)
 830                 GOTO(out, rc);
 831
 832         while (cur_offset < lgi->lgi_attr.la_size) {
 833                 struct llog_rec_hdr     *rec, *last_rec;
 834                 struct llog_rec_tail    *tail;
 835
 836                 lgi->lgi_buf.lb_len = len;
 837                 lgi->lgi_buf.lb_buf = buf;
 838                 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
 839                 if (rc < 0) {
 840                         CERROR("%s: can't read llog block from log "DFID
 841                                " offset "LPU64": rc = %d\n",
 842                                o->do_lu.lo_dev->ld_obd->obd_name,
 843                                PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
 844                         GOTO(out, rc);
 845                 }
 846
 847                 if (rc == 0) /* end of file, nothing to do */
 848                         GOTO(out, rc);
 849
 850                 if (rc < sizeof(*tail)) {
 851                         CERROR("%s: invalid llog block at log id "DOSTID"/%u "
 852                                "offset "LPU64"\n",
 853                                o->do_lu.lo_dev->ld_obd->obd_name,
 854                                POSTID(&loghandle->lgh_id.lgl_oi),
 855                                loghandle->lgh_id.lgl_ogen, cur_offset);
 856                         GOTO(out, rc = -EINVAL);
 857                 }
 858
 859                 rec = buf;
 860                 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 861                         lustre_swab_llog_rec(rec);
 862
 863                 tail = (struct llog_rec_tail *)((char *)buf + rc -
 864                                                 sizeof(struct llog_rec_tail));
 865                 /* get the last record in block */
 866                 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
 867                                                    le32_to_cpu(tail->lrt_len));
 868
 869                 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
 870                         lustre_swab_llog_rec(last_rec);
 871                 LASSERT(last_rec->lrh_index == tail->lrt_index);
 872
 873                 /* this shouldn't happen */
 874                 if (tail->lrt_index == 0) {
 875                         CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
 876                                "offset "LPU64"\n",
 877                                o->do_lu.lo_dev->ld_obd->obd_name,
 878                                POSTID(&loghandle->lgh_id.lgl_oi),
 879                                loghandle->lgh_id.lgl_ogen, cur_offset);
 880                         GOTO(out, rc = -EINVAL);
 881                 }
 882                 if (tail->lrt_index < prev_idx)
 883                         continue;
 884
 885                 /* sanity check that the start of the new buffer is no farther
 886                  * than the record that we wanted.  This shouldn't happen. */
 887                 if (rec->lrh_index > prev_idx) {
 888                         CERROR("%s: missed desired record? %u > %u\n",
 889                                o->do_lu.lo_dev->ld_obd->obd_name,
 890                                rec->lrh_index, prev_idx);
 891                         GOTO(out, rc = -ENOENT);
 892                 }
 893
 894                 /* Trim unsupported extensions for compat w/ older clients */
 895                 if (!(loghandle->lgh_hdr->llh_flags & LLOG_F_EXT_JOBID))
 896                         changelog_block_trim_ext(rec, last_rec,
 897                                                  CLF_VERSION | CLF_RENAME);
 898
 899                 GOTO(out, rc = 0);
 900         }
 901         GOTO(out, rc = -EIO);
 902 out:
 903         return rc;
 904 }
 905
 906 /**
 907  * This is helper function to get llog directory object. It is used by named
 908  * llog operations to find/insert/delete llog entry from llog directory.
 909  *
 910  * \param[in] env       execution environment
 911  * \param[in] ctxt      llog context
 912  *
 913  * \retval              dt_object of llog directory
 914  * \retval              ERR_PTR of negative value on error
 915  */
 916 static struct dt_object *llog_osd_dir_get(const struct lu_env *env,
 917                                           struct llog_ctxt *ctxt)
 918 {
 919         struct dt_device        *dt;
 920         struct dt_thread_info   *dti = dt_info(env);
 921         struct dt_object        *dir;
 922         int                      rc;
 923
 924         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 925         if (ctxt->loc_dir == NULL) {
 926                 rc = dt_root_get(env, dt, &dti->dti_fid);
 927                 if (rc)
 928                         return ERR_PTR(rc);
 929                 dir = dt_locate(env, dt, &dti->dti_fid);
 930
 931                 if (!IS_ERR(dir) && !dt_try_as_dir(env, dir)) {
 932                         lu_object_put(env, &dir->do_lu);
 933                         return ERR_PTR(-ENOTDIR);
 934                 }
 935         } else {
 936                 lu_object_get(&ctxt->loc_dir->do_lu);
 937                 dir = ctxt->loc_dir;
 938         }
 939
 940         return dir;
 941 }
 942
 943 /**
 944  * Implementation of the llog_operations::lop_open
 945  *
 946  * This function opens the llog by its logid or by name, it may open also
 947  * non existent llog and assing then new id to it.
 948  * The llog_open/llog_close pair works similar to lu_object_find/put,
 949  * the object may not exist prior open. The result of open is just dt_object
 950  * in the llog header.
 951  *
 952  * \param[in] env               execution environment
 953  * \param[in] handle            llog handle of the current llog
 954  * \param[in] logid             logid of llog to open (nameless llog)
 955  * \param[in] name              name of llog to open (named llog)
 956  * \param[in] open_param
 957  *                              LLOG_OPEN_NEW - new llog, may not exist
 958  *                              LLOG_OPEN_EXIST - old llog, must exist
 959  *
 960  * \retval                      0 on successful open, llog_handle::lgh_obj
 961  *                              contains the dt_object of the llog.
 962  * \retval                      negative value on error
 963  */
 964 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
 965                          struct llog_logid *logid, char *name,
 966                          enum llog_open_param open_param)
 967 {
 968         struct llog_thread_info         *lgi = llog_info(env);
 969         struct llog_ctxt                *ctxt = handle->lgh_ctxt;
 970         struct dt_object                *o;
 971         struct dt_device                *dt;
 972         struct ls_device                *ls;
 973         struct local_oid_storage        *los;
 974         int                              rc = 0;
 975
 976         ENTRY;
 977
 978         LASSERT(env);
 979         LASSERT(ctxt);
 980         LASSERT(ctxt->loc_exp);
 981         LASSERT(ctxt->loc_exp->exp_obd);
 982         dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
 983         LASSERT(dt);
 984
 985         ls = ls_device_get(dt);
 986         if (IS_ERR(ls))
 987                 RETURN(PTR_ERR(ls));
 988
 989         mutex_lock(&ls->ls_los_mutex);
 990         los = dt_los_find(ls, name != NULL ? FID_SEQ_LLOG_NAME : FID_SEQ_LLOG);
 991         mutex_unlock(&ls->ls_los_mutex);
 992         LASSERT(los);
 993         ls_device_put(env, ls);
 994
 995         LASSERT(handle);
 996
 997         if (logid != NULL) {
 998                 logid_to_fid(logid, &lgi->lgi_fid);
 999         } else if (name) {
1000                 struct dt_object *llog_dir;
1001
1002                 llog_dir = llog_osd_dir_get(env, ctxt);
1003                 if (IS_ERR(llog_dir))
1004                         GOTO(out, rc = PTR_ERR(llog_dir));
1005                 dt_read_lock(env, llog_dir, 0);
1006                 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
1007                 dt_read_unlock(env, llog_dir);
1008                 lu_object_put(env, &llog_dir->do_lu);
1009                 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
1010                         /* generate fid for new llog */
1011                         rc = local_object_fid_generate(env, los,
1012                                                        &lgi->lgi_fid);
1013                 }
1014                 if (rc < 0)
1015                         GOTO(out, rc);
1016                 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
1017                 if (handle->lgh_name)
1018                         strcpy(handle->lgh_name, name);
1019                 else
1020                         GOTO(out, rc = -ENOMEM);
1021         } else {
1022                 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
1023                 /* generate fid for new llog */
1024                 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
1025                 if (rc < 0)
1026                         GOTO(out, rc);
1027         }
1028
1029         o = ls_locate(env, ls, &lgi->lgi_fid, NULL);
1030         if (IS_ERR(o))
1031                 GOTO(out_name, rc = PTR_ERR(o));
1032
1033         /* No new llog is expected but doesn't exist */
1034         if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
1035                 GOTO(out_put, rc = -ENOENT);
1036
1037         fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
1038         handle->lgh_obj = o;
1039         handle->private_data = los;
1040         LASSERT(handle->lgh_ctxt);
1041
1042         RETURN(rc);
1043
1044 out_put:
1045         lu_object_put(env, &o->do_lu);
1046 out_name:
1047         if (handle->lgh_name != NULL)
1048                 OBD_FREE(handle->lgh_name, strlen(name) + 1);
1049 out:
1050         dt_los_put(los);
1051         RETURN(rc);
1052 }
1053
1054 /**
1055  * Implementation of the llog_operations::lop_exist
1056  *
1057  * This function checks that llog exists on storage.
1058  *
1059  * \param[in] handle    llog handle of the current llog
1060  *
1061  * \retval              true if llog object exists and is not just destroyed
1062  * \retval              false if llog doesn't exist or just destroyed
1063  */
1064 static int llog_osd_exist(struct llog_handle *handle)
1065 {
1066         LASSERT(handle->lgh_obj);
1067         return (dt_object_exists(handle->lgh_obj) &&
1068                 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
1069 }
1070
1071 /**
1072  * Implementation of the llog_operations::lop_declare_create
1073  *
1074  * This function declares the llog create. It declares also name insert
1075  * into llog directory in case of named llog.
1076  *
1077  * \param[in] env       execution environment
1078  * \param[in] res       llog handle of the current llog
1079  * \param[in] th        current transaction handle
1080  *
1081  * \retval              0 on successful create declaration
1082  * \retval              negative value on error
1083  */
1084 static int llog_osd_declare_create(const struct lu_env *env,
1085                                    struct llog_handle *res, struct thandle *th)
1086 {
1087         struct llog_thread_info         *lgi = llog_info(env);
1088         struct dt_insert_rec            *rec = &lgi->lgi_dt_rec;
1089         struct local_oid_storage        *los;
1090         struct dt_object                *o;
1091         int                              rc;
1092
1093         ENTRY;
1094
1095         LASSERT(res->lgh_obj);
1096         LASSERT(th);
1097
1098         /* object can be created by another thread */
1099         o = res->lgh_obj;
1100         if (dt_object_exists(o))
1101                 RETURN(0);
1102
1103         los = res->private_data;
1104         LASSERT(los);
1105
1106         rc = llog_osd_declare_new_object(env, los, o, th);
1107         if (rc)
1108                 RETURN(rc);
1109
1110         /* do not declare header initialization here as it's declared
1111          * in llog_osd_declare_write_rec() which is always called */
1112
1113         if (res->lgh_name) {
1114                 struct dt_object *llog_dir;
1115
1116                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
1117                 if (IS_ERR(llog_dir))
1118                         RETURN(PTR_ERR(llog_dir));
1119                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
1120                 rec->rec_fid = &lgi->lgi_fid;
1121                 rec->rec_type = S_IFREG;
1122                 rc = dt_declare_insert(env, llog_dir,
1123                                        (struct dt_rec *)rec,
1124                                        (struct dt_key *)res->lgh_name, th);
1125                 lu_object_put(env, &llog_dir->do_lu);
1126                 if (rc)
1127                         CERROR("%s: can't declare named llog %s: rc = %d\n",
1128                                o->do_lu.lo_dev->ld_obd->obd_name,
1129                                res->lgh_name, rc);
1130         }
1131         RETURN(rc);
1132 }
1133
1134 /**
1135  * Implementation of the llog_operations::lop_create
1136  *
1137  * This function creates the llog according with llog_handle::lgh_obj
1138  * and llog_handle::lgh_name.
1139  *
1140  * \param[in] env       execution environment
1141  * \param[in] res       llog handle of the current llog
1142  * \param[in] th        current transaction handle
1143  *
1144  * \retval              0 on successful create
1145  * \retval              negative value on error
1146  */
1147 static int llog_osd_create(const struct lu_env *env, struct llog_handle *res,
1148                            struct thandle *th)
1149 {
1150         struct llog_thread_info *lgi = llog_info(env);
1151         struct dt_insert_rec    *rec = &lgi->lgi_dt_rec;
1152         struct local_oid_storage *los;
1153         struct dt_object        *o;
1154         int                      rc = 0;
1155
1156         ENTRY;
1157
1158         LASSERT(env);
1159         o = res->lgh_obj;
1160         LASSERT(o);
1161
1162         /* llog can be already created */
1163         if (dt_object_exists(o))
1164                 RETURN(-EEXIST);
1165
1166         los = res->private_data;
1167         LASSERT(los);
1168
1169         dt_write_lock(env, o, 0);
1170         if (!dt_object_exists(o))
1171                 rc = llog_osd_create_new_object(env, los, o, th);
1172         else
1173                 rc = -EEXIST;
1174
1175         dt_write_unlock(env, o);
1176         if (rc)
1177                 RETURN(rc);
1178
1179         if (res->lgh_name) {
1180                 struct dt_object *llog_dir;
1181
1182                 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
1183                 if (IS_ERR(llog_dir))
1184                         RETURN(PTR_ERR(llog_dir));
1185
1186                 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
1187                 rec->rec_fid = &lgi->lgi_fid;
1188                 rec->rec_type = S_IFREG;
1189                 dt_read_lock(env, llog_dir, 0);
1190                 rc = dt_insert(env, llog_dir, (struct dt_rec *)rec,
1191                                (struct dt_key *)res->lgh_name,
1192                                th, 1);
1193                 dt_read_unlock(env, llog_dir);
1194                 lu_object_put(env, &llog_dir->do_lu);
1195                 if (rc)
1196                         CERROR("%s: can't create named llog %s: rc = %d\n",
1197                                o->do_lu.lo_dev->ld_obd->obd_name,
1198                                res->lgh_name, rc);
1199         }
1200         RETURN(rc);
1201 }
1202
1203 /**
1204  * Implementation of the llog_operations::lop_close
1205  *
1206  * This function closes the llog. It just put llog object and referenced
1207  * local storage.
1208  *
1209  * \param[in] env       execution environment
1210  * \param[in] handle    llog handle of the current llog
1211  *
1212  * \retval              0 on successful llog close
1213  * \retval              negative value on error
1214  */
1215 static int llog_osd_close(const struct lu_env *env, struct llog_handle *handle)
1216 {
1217         struct local_oid_storage        *los;
1218         int                              rc = 0;
1219
1220         ENTRY;
1221
1222         LASSERT(handle->lgh_obj);
1223
1224         lu_object_put(env, &handle->lgh_obj->do_lu);
1225
1226         los = handle->private_data;
1227         LASSERT(los);
1228         dt_los_put(los);
1229
1230         if (handle->lgh_name)
1231                 OBD_FREE(handle->lgh_name, strlen(handle->lgh_name) + 1);
1232
1233         RETURN(rc);
1234 }
1235
1236 /**
1237  * Implementation of the llog_operations::lop_destroy
1238  *
1239  * This function destroys the llog and deletes also entry in the
1240  * llog directory in case of named llog. Llog should be opened prior that.
1241  * Destroy method is not part of external transaction and does everything
1242  * inside.
1243  *
1244  * \param[in] env               execution environment
1245  * \param[in] loghandle llog handle of the current llog
1246  *
1247  * \retval              0 on successful destroy
1248  * \retval              negative value on error
1249  */
1250 static int llog_osd_destroy(const struct lu_env *env,
1251                             struct llog_handle *loghandle)
1252 {
1253         struct llog_ctxt        *ctxt;
1254         struct dt_object        *o, *llog_dir = NULL;
1255         struct dt_device        *d;
1256         struct thandle          *th;
1257         char                    *name = NULL;
1258         int                      rc;
1259
1260         ENTRY;
1261
1262         ctxt = loghandle->lgh_ctxt;
1263         LASSERT(ctxt);
1264
1265         o = loghandle->lgh_obj;
1266         LASSERT(o);
1267
1268         d = lu2dt_dev(o->do_lu.lo_dev);
1269         LASSERT(d);
1270         LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
1271
1272         th = dt_trans_create(env, d);
1273         if (IS_ERR(th))
1274                 RETURN(PTR_ERR(th));
1275
1276         if (loghandle->lgh_name) {
1277                 llog_dir = llog_osd_dir_get(env, ctxt);
1278                 if (IS_ERR(llog_dir))
1279                         GOTO(out_trans, rc = PTR_ERR(llog_dir));
1280
1281                 name = loghandle->lgh_name;
1282                 rc = dt_declare_delete(env, llog_dir,
1283                                        (struct dt_key *)name, th);
1284                 if (rc)
1285                         GOTO(out_trans, rc);
1286         }
1287
1288         rc = dt_declare_ref_del(env, o, th);
1289         if (rc < 0)
1290                 GOTO(out_trans, rc);
1291
1292         rc = dt_declare_destroy(env, o, th);
1293         if (rc)
1294                 GOTO(out_trans, rc);
1295
1296         rc = dt_trans_start_local(env, d, th);
1297         if (rc)
1298                 GOTO(out_trans, rc);
1299
1300         dt_write_lock(env, o, 0);
1301         if (dt_object_exists(o)) {
1302                 if (name) {
1303                         dt_read_lock(env, llog_dir, 0);
1304                         rc = dt_delete(env, llog_dir,
1305                                        (struct dt_key *) name,
1306                                        th);
1307                         dt_read_unlock(env, llog_dir);
1308                         if (rc) {
1309                                 CERROR("%s: can't remove llog %s: rc = %d\n",
1310                                        o->do_lu.lo_dev->ld_obd->obd_name,
1311                                        name, rc);
1312                                 GOTO(out_unlock, rc);
1313                         }
1314                 }
1315                 dt_ref_del(env, o, th);
1316                 rc = dt_destroy(env, o, th);
1317                 if (rc)
1318                         GOTO(out_unlock, rc);
1319         }
1320 out_unlock:
1321         dt_write_unlock(env, o);
1322 out_trans:
1323         dt_trans_stop(env, d, th);
1324         if (llog_dir != NULL)
1325                 lu_object_put(env, &llog_dir->do_lu);
1326         RETURN(rc);
1327 }
1328
1329 /**
1330  * Implementation of the llog_operations::lop_setup
1331  *
1332  * This function setup the llog on local storage.
1333  *
1334  * \param[in] env       execution environment
1335  * \param[in] obd       obd device the llog belongs to
1336  * \param[in] olg       the llog group, it is always zero group now.
1337  * \param[in] ctxt_idx  the llog index, it defines the purpose of this llog.
1338  *                      Every new llog type have to use own index.
1339  * \param[in] disk_obd  the storage obd, where llog is stored.
1340  *
1341  * \retval              0 on successful llog setup
1342  * \retval              negative value on error
1343  */
1344 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1345                           struct obd_llog_group *olg, int ctxt_idx,
1346                           struct obd_device *disk_obd)
1347 {
1348         struct llog_thread_info         *lgi = llog_info(env);
1349         struct llog_ctxt                *ctxt;
1350         int                              rc = 0;
1351         ENTRY;
1352
1353         LASSERT(obd);
1354         LASSERT(olg->olg_ctxts[ctxt_idx]);
1355
1356         ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1357         LASSERT(ctxt);
1358
1359         /* initialize data allowing to generate new fids,
1360          * literally we need a sequece */
1361         lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1362         lgi->lgi_fid.f_oid = 1;
1363         lgi->lgi_fid.f_ver = 0;
1364         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1365                                     &lgi->lgi_fid,
1366                                     &ctxt->loc_los_nameless);
1367         if (rc != 0)
1368                 GOTO(out, rc);
1369
1370         lgi->lgi_fid.f_seq = FID_SEQ_LLOG_NAME;
1371         lgi->lgi_fid.f_oid = 1;
1372         lgi->lgi_fid.f_ver = 0;
1373         rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1374                                     &lgi->lgi_fid,
1375                                     &ctxt->loc_los_named);
1376         if (rc != 0) {
1377                 local_oid_storage_fini(env, ctxt->loc_los_nameless);
1378                 ctxt->loc_los_nameless = NULL;
1379         }
1380
1381         GOTO(out, rc);
1382
1383 out:
1384         llog_ctxt_put(ctxt);
1385         return rc;
1386 }
1387
1388 /**
1389  * Implementation of the llog_operations::lop_cleanup
1390  *
1391  * This function cleanups the llog on local storage.
1392  *
1393  * \param[in] env       execution environment
1394  * \param[in] ctxt      the llog context
1395  *
1396  * \retval              0
1397  */
1398 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1399 {
1400         if (ctxt->loc_los_nameless != NULL) {
1401                 local_oid_storage_fini(env, ctxt->loc_los_nameless);
1402                 ctxt->loc_los_nameless = NULL;
1403         }
1404
1405         if (ctxt->loc_los_named != NULL) {
1406                 local_oid_storage_fini(env, ctxt->loc_los_named);
1407                 ctxt->loc_los_named = NULL;
1408         }
1409
1410         return 0;
1411 }
1412
1413 struct llog_operations llog_osd_ops = {
1414         .lop_next_block         = llog_osd_next_block,
1415         .lop_prev_block         = llog_osd_prev_block,
1416         .lop_read_header        = llog_osd_read_header,
1417         .lop_destroy            = llog_osd_destroy,
1418         .lop_setup              = llog_osd_setup,
1419         .lop_cleanup            = llog_osd_cleanup,
1420         .lop_open               = llog_osd_open,
1421         .lop_exist              = llog_osd_exist,
1422         .lop_declare_create     = llog_osd_declare_create,
1423         .lop_create             = llog_osd_create,
1424         .lop_declare_write_rec  = llog_osd_declare_write_rec,
1425         .lop_write_rec          = llog_osd_write_rec,
1426         .lop_close              = llog_osd_close,
1427 };
1428 EXPORT_SYMBOL(llog_osd_ops);
1429
1430 /**
1431  * Read the special file which contains the list of llog catalogs IDs
1432  *
1433  * This function reads the CATALOGS file which contains the array of llog
1434  * catalogs IDs. The main purpose of this file is to store OSP llogs indexed
1435  * by OST/MDT number.
1436  *
1437  * \param[in]  env              execution environment
1438  * \param[in]  d                corresponding storage device
1439  * \param[in]  idx              position to start from, usually OST/MDT index
1440  * \param[in]  count            how many catalog IDs to read
1441  * \param[out] idarray          the buffer for the data. If it is NULL then
1442  *                              function returns just number of catalog IDs
1443  *                              in the file.
1444  * \param[in]  fid              LLOG_CATALOGS_OID for CATALOG object
1445  *
1446  * \retval                      0 on successful read of catalog IDs
1447  * \retval                      negative value on error
1448  * \retval                      positive value which is number of records in
1449  *                              the file if \a idarray is NULL
1450  */
1451 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1452                           int idx, int count, struct llog_catid *idarray,
1453                           const struct lu_fid *fid)
1454 {
1455         struct llog_thread_info *lgi = llog_info(env);
1456         struct dt_object        *o = NULL;
1457         struct thandle          *th;
1458         int                      rc, size;
1459
1460         ENTRY;
1461
1462         LASSERT(d);
1463
1464         size = sizeof(*idarray) * count;
1465         lgi->lgi_off = idx *  sizeof(*idarray);
1466
1467         lgi->lgi_fid = *fid;
1468         o = dt_locate(env, d, &lgi->lgi_fid);
1469         if (IS_ERR(o))
1470                 RETURN(PTR_ERR(o));
1471
1472         if (!dt_object_exists(o)) {
1473                 th = dt_trans_create(env, d);
1474                 if (IS_ERR(th))
1475                         GOTO(out, rc = PTR_ERR(th));
1476
1477                 lgi->lgi_attr.la_valid = LA_MODE;
1478                 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1479                 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1480
1481                 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1482                                        &lgi->lgi_dof, th);
1483                 if (rc)
1484                         GOTO(out_trans, rc);
1485
1486                 rc = dt_trans_start_local(env, d, th);
1487                 if (rc)
1488                         GOTO(out_trans, rc);
1489
1490                 dt_write_lock(env, o, 0);
1491                 if (!dt_object_exists(o))
1492                         rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1493                                        &lgi->lgi_dof, th);
1494                 dt_write_unlock(env, o);
1495 out_trans:
1496                 dt_trans_stop(env, d, th);
1497                 if (rc)
1498                         GOTO(out, rc);
1499         }
1500
1501         rc = dt_attr_get(env, o, &lgi->lgi_attr);
1502         if (rc)
1503                 GOTO(out, rc);
1504
1505         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1506                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1507                        o->do_lu.lo_dev->ld_obd->obd_name,
1508                        lgi->lgi_attr.la_mode);
1509                 GOTO(out, rc = -ENOENT);
1510         }
1511
1512         CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1513                (int)lgi->lgi_attr.la_size, size);
1514
1515         /* return just number of llogs */
1516         if (idarray == NULL) {
1517                 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1518                 GOTO(out, rc);
1519         }
1520
1521         /* read for new ost index or for empty file */
1522         memset(idarray, 0, size);
1523         if (lgi->lgi_attr.la_size <= lgi->lgi_off)
1524                 GOTO(out, rc = 0);
1525         if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1526                 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1527
1528         lgi->lgi_buf.lb_buf = idarray;
1529         lgi->lgi_buf.lb_len = size;
1530         rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1531         /* -EFAULT means the llog is a sparse file. This is not an error
1532          * after arbitrary OST index is supported. */
1533         if (rc < 0 && rc != -EFAULT) {
1534                 CERROR("%s: error reading CATALOGS: rc = %d\n",
1535                        o->do_lu.lo_dev->ld_obd->obd_name,  rc);
1536                 GOTO(out, rc);
1537         }
1538
1539         EXIT;
1540 out:
1541         lu_object_put(env, &o->do_lu);
1542         RETURN(rc);
1543 }
1544 EXPORT_SYMBOL(llog_osd_get_cat_list);
1545
1546 /**
1547  * Write the special file which contains the list of llog catalogs IDs
1548  *
1549  * This function writes the CATALOG file which contains the array of llog
1550  * catalogs IDs. It is used mostly to store OSP llogs indexed by OST/MDT
1551  * number.
1552  *
1553  * \param[in]  env      execution environment
1554  * \param[in]  d        corresponding storage device
1555  * \param[in]  idx      position to start from, usually OST/MDT index
1556  * \param[in]  count    how many catalog IDs to write
1557  * \param[out] idarray  the buffer with the data to write.
1558  * \param[in]  fid      LLOG_CATALOGS_OID for CATALOG object
1559  *
1560  * \retval              0 on successful write of catalog IDs
1561  * \retval              negative value on error
1562  */
1563 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1564                           int idx, int count, struct llog_catid *idarray,
1565                           const struct lu_fid *fid)
1566 {
1567         struct llog_thread_info *lgi = llog_info(env);
1568         struct dt_object        *o = NULL;
1569         struct thandle          *th;
1570         int                      rc, size;
1571
1572         if (count == 0)
1573                 RETURN(0);
1574
1575         LASSERT(d);
1576
1577         size = sizeof(*idarray) * count;
1578         lgi->lgi_off = idx * sizeof(*idarray);
1579         lgi->lgi_fid = *fid;
1580
1581         o = dt_locate(env, d, &lgi->lgi_fid);
1582         if (IS_ERR(o))
1583                 RETURN(PTR_ERR(o));
1584
1585         if (!dt_object_exists(o))
1586                 GOTO(out, rc = -ENOENT);
1587
1588         rc = dt_attr_get(env, o, &lgi->lgi_attr);
1589         if (rc)
1590                 GOTO(out, rc);
1591
1592         if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1593                 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1594                        o->do_lu.lo_dev->ld_obd->obd_name,
1595                        lgi->lgi_attr.la_mode);
1596                 GOTO(out, rc = -ENOENT);
1597         }
1598
1599         th = dt_trans_create(env, d);
1600         if (IS_ERR(th))
1601                 GOTO(out, rc = PTR_ERR(th));
1602
1603         lgi->lgi_buf.lb_len = size;
1604         lgi->lgi_buf.lb_buf = idarray;
1605         rc = dt_declare_record_write(env, o, &lgi->lgi_buf, lgi->lgi_off, th);
1606         if (rc)
1607                 GOTO(out, rc);
1608
1609         rc = dt_trans_start_local(env, d, th);
1610         if (rc)
1611                 GOTO(out_trans, rc);
1612
1613         rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1614         if (rc)
1615                 CDEBUG(D_INODE, "can't write CATALOGS at index %d: rc = %d\n",
1616                        idx, rc);
1617 out_trans:
1618         dt_trans_stop(env, d, th);
1619 out:
1620         lu_object_put(env, &o->do_lu);
1621         RETURN(rc);
1622 }
1623 EXPORT_SYMBOL(llog_osd_put_cat_list);