lustre/obdclass/llog.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, 2017, 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  * lustre/obdclass/llog.c
  33  *
  34  * OST<->MDS recovery logging infrastructure.
  35  * Invariants in implementation:
  36  * - we do not share logs among different OST<->MDS connections, so that
  37  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  38  *
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  41  * Author: Mikhail Pershin <tappro@whamcloud.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LOG
  45
  46 #include <linux/pid_namespace.h>
  47 #include <linux/kthread.h>
  48 #include <llog_swab.h>
  49 #include <lustre_log.h>
  50 #include <obd_support.h>
  51 #include <obd_class.h>
  52 #include "llog_internal.h"
  53 /*
  54  * Allocate a new log or catalog handle
  55  * Used inside llog_open().
  56  */
  57 static struct llog_handle *llog_alloc_handle(void)
  58 {
  59         struct llog_handle *loghandle;
  60
  61         OBD_ALLOC_PTR(loghandle);
  62         if (loghandle == NULL)
  63                 return NULL;
  64
  65         init_rwsem(&loghandle->lgh_lock);
  66         mutex_init(&loghandle->lgh_hdr_mutex);
  67         init_rwsem(&loghandle->lgh_last_sem);
  68         INIT_LIST_HEAD(&loghandle->u.phd.phd_entry);
  69         atomic_set(&loghandle->lgh_refcount, 1);
  70
  71         return loghandle;
  72 }
  73
  74 /*
  75  * Free llog handle and header data if exists. Used in llog_close() only
  76  */
  77 static void llog_free_handle(struct llog_handle *loghandle)
  78 {
  79         LASSERT(loghandle != NULL);
  80
  81         /* failed llog_init_handle */
  82         if (loghandle->lgh_hdr == NULL)
  83                 goto out;
  84
  85         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  86                 LASSERT(list_empty(&loghandle->u.phd.phd_entry));
  87         else if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  88                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  89         OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
  90 out:
  91         OBD_FREE_PTR(loghandle);
  92 }
  93
  94 void llog_handle_get(struct llog_handle *loghandle)
  95 {
  96         atomic_inc(&loghandle->lgh_refcount);
  97 }
  98
  99 void llog_handle_put(struct llog_handle *loghandle)
 100 {
 101         LASSERT(atomic_read(&loghandle->lgh_refcount) > 0);
 102         if (atomic_dec_and_test(&loghandle->lgh_refcount))
 103                 llog_free_handle(loghandle);
 104 }
 105
 106 static int llog_declare_destroy(const struct lu_env *env,
 107                                 struct llog_handle *handle,
 108                                 struct thandle *th)
 109 {
 110         struct llog_operations *lop;
 111         int rc;
 112
 113         ENTRY;
 114
 115         rc = llog_handle2ops(handle, &lop);
 116         if (rc)
 117                 RETURN(rc);
 118         if (lop->lop_declare_destroy == NULL)
 119                 RETURN(-EOPNOTSUPP);
 120
 121         rc = lop->lop_declare_destroy(env, handle, th);
 122
 123         RETURN(rc);
 124 }
 125
 126 int llog_trans_destroy(const struct lu_env *env, struct llog_handle *handle,
 127                        struct thandle *th)
 128 {
 129         struct llog_operations  *lop;
 130         int rc;
 131         ENTRY;
 132
 133         rc = llog_handle2ops(handle, &lop);
 134         if (rc < 0)
 135                 RETURN(rc);
 136         if (lop->lop_destroy == NULL)
 137                 RETURN(-EOPNOTSUPP);
 138
 139         LASSERT(handle->lgh_obj != NULL);
 140         if (!dt_object_exists(handle->lgh_obj))
 141                 RETURN(0);
 142
 143         rc = lop->lop_destroy(env, handle, th);
 144
 145         RETURN(rc);
 146 }
 147
 148 int llog_destroy(const struct lu_env *env, struct llog_handle *handle)
 149 {
 150         struct llog_operations  *lop;
 151         struct dt_device        *dt;
 152         struct thandle          *th;
 153         int rc;
 154
 155         ENTRY;
 156
 157         rc = llog_handle2ops(handle, &lop);
 158         if (rc < 0)
 159                 RETURN(rc);
 160         if (lop->lop_destroy == NULL)
 161                 RETURN(-EOPNOTSUPP);
 162
 163         if (handle->lgh_obj == NULL) {
 164                 /* if lgh_obj == NULL, then it is from client side destroy */
 165                 rc = lop->lop_destroy(env, handle, NULL);
 166                 RETURN(rc);
 167         }
 168
 169         if (!dt_object_exists(handle->lgh_obj))
 170                 RETURN(0);
 171
 172         dt = lu2dt_dev(handle->lgh_obj->do_lu.lo_dev);
 173
 174         if (unlikely(unlikely(dt->dd_rdonly)))
 175                 RETURN(-EROFS);
 176
 177         th = dt_trans_create(env, dt);
 178         if (IS_ERR(th))
 179                 RETURN(PTR_ERR(th));
 180
 181         rc = llog_declare_destroy(env, handle, th);
 182         if (rc != 0)
 183                 GOTO(out_trans, rc);
 184
 185         rc = dt_trans_start_local(env, dt, th);
 186         if (rc < 0)
 187                 GOTO(out_trans, rc);
 188
 189         rc = lop->lop_destroy(env, handle, th);
 190
 191 out_trans:
 192         dt_trans_stop(env, dt, th);
 193
 194         RETURN(rc);
 195 }
 196 EXPORT_SYMBOL(llog_destroy);
 197
 198 /* returns negative on error; 0 if success; 1 if success & log destroyed */
 199 int llog_cancel_arr_rec(const struct lu_env *env, struct llog_handle *loghandle,
 200                         int num, int *index)
 201 {
 202         struct llog_thread_info *lgi = llog_info(env);
 203         struct dt_device        *dt;
 204         struct llog_log_hdr     *llh;
 205         struct thandle          *th;
 206         __u32                    tmp_lgc_index;
 207         int                      rc, i = 0;
 208         int rc1;
 209         bool subtract_count = false;
 210
 211         ENTRY;
 212
 213         LASSERT(loghandle != NULL);
 214         LASSERT(loghandle->lgh_ctxt != NULL);
 215         LASSERT(loghandle->lgh_obj != NULL);
 216
 217         llh = loghandle->lgh_hdr;
 218
 219         CDEBUG(D_RPCTRACE, "Canceling %d records, first %d in log "DFID"\n",
 220                num, index[0], PFID(&loghandle->lgh_id.lgl_oi.oi_fid));
 221
 222         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 223
 224         if (unlikely(unlikely(dt->dd_rdonly)))
 225                 RETURN(0);
 226
 227         th = dt_trans_create(env, dt);
 228         if (IS_ERR(th))
 229                 RETURN(PTR_ERR(th));
 230
 231         rc = llog_declare_write_rec(env, loghandle, &llh->llh_hdr, 0, th);
 232         if (rc < 0)
 233                 GOTO(out_trans, rc);
 234
 235         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY)) {
 236                 rc = llog_declare_destroy(env, loghandle, th);
 237                 if (rc < 0)
 238                         GOTO(out_trans, rc);
 239         }
 240
 241         th->th_wait_submit = 1;
 242         rc = dt_trans_start_local(env, dt, th);
 243         if (rc < 0)
 244                 GOTO(out_trans, rc);
 245
 246         down_write(&loghandle->lgh_lock);
 247         /* clear bitmap */
 248         mutex_lock(&loghandle->lgh_hdr_mutex);
 249         for (i = 0; i < num; ++i) {
 250                 if (index[i] == 0) {
 251                         CERROR("Can't cancel index 0 which is header\n");
 252                         GOTO(out_unlock, rc = -EINVAL);
 253                 }
 254                 if (!ext2_clear_bit(index[i], LLOG_HDR_BITMAP(llh))) {
 255                         CDEBUG(D_RPCTRACE, "Catalog index %u already clear?\n",
 256                                index[i]);
 257                         GOTO(out_unlock, rc = -ENOENT);
 258                 }
 259         }
 260         loghandle->lgh_hdr->llh_count -= num;
 261         subtract_count = true;
 262
 263         /* Since llog_process_thread use lgi_cookie, it`s better to save them
 264          * and restore after using
 265          */
 266         tmp_lgc_index = lgi->lgi_cookie.lgc_index;
 267         /* Pass this index to llog_osd_write_rec(), which will use the index
 268          * to only update the necesary bitmap. */
 269         lgi->lgi_cookie.lgc_index = index[0];
 270         /* update header */
 271         rc = llog_write_rec(env, loghandle, &llh->llh_hdr, (num != 1 ? NULL :
 272                             &lgi->lgi_cookie), LLOG_HEADER_IDX, th);
 273         lgi->lgi_cookie.lgc_index = tmp_lgc_index;
 274
 275         if (rc != 0)
 276                 GOTO(out_unlock, rc);
 277
 278         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 279             (llh->llh_count == 1) &&
 280             ((loghandle->lgh_last_idx == LLOG_HDR_BITMAP_SIZE(llh) - 1) ||
 281              (loghandle->u.phd.phd_cat_handle != NULL &&
 282               loghandle->u.phd.phd_cat_handle->u.chd.chd_current_log !=
 283                 loghandle))) {
 284                 /* never try to destroy it again */
 285                 llh->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 286                 rc = llog_trans_destroy(env, loghandle, th);
 287                 if (rc < 0) {
 288                         /* Sigh, can not destroy the final plain llog, but
 289                          * the bitmap has been clearly, so the record can not
 290                          * be accessed anymore, let's return 0 for now, and
 291                          * the orphan will be handled by LFSCK. */
 292                         CERROR("%s: can't destroy empty llog "DFID": rc = %d\n",
 293                                loghandle->lgh_ctxt->loc_obd->obd_name,
 294                                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rc);
 295                         GOTO(out_unlock, rc = 0);
 296                 }
 297                 rc = LLOG_DEL_PLAIN;
 298         }
 299
 300 out_unlock:
 301         mutex_unlock(&loghandle->lgh_hdr_mutex);
 302         up_write(&loghandle->lgh_lock);
 303 out_trans:
 304         rc1 = dt_trans_stop(env, dt, th);
 305         if (rc == 0)
 306                 rc = rc1;
 307         if (rc < 0) {
 308                 mutex_lock(&loghandle->lgh_hdr_mutex);
 309                 if (subtract_count)
 310                         loghandle->lgh_hdr->llh_count += num;
 311                 for (i = i - 1; i >= 0; i--)
 312                         ext2_set_bit(index[i], LLOG_HDR_BITMAP(llh));
 313                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 314         }
 315         RETURN(rc);
 316 }
 317
 318 int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
 319                     int index)
 320 {
 321         return llog_cancel_arr_rec(env, loghandle, 1, &index);
 322 }
 323
 324 int llog_read_header(const struct lu_env *env, struct llog_handle *handle,
 325                      const struct obd_uuid *uuid)
 326 {
 327         struct llog_operations *lop;
 328         int rc;
 329         ENTRY;
 330
 331         rc = llog_handle2ops(handle, &lop);
 332         if (rc)
 333                 RETURN(rc);
 334
 335         if (lop->lop_read_header == NULL)
 336                 RETURN(-EOPNOTSUPP);
 337
 338         rc = lop->lop_read_header(env, handle);
 339         if (rc == LLOG_EEMPTY) {
 340                 struct llog_log_hdr *llh = handle->lgh_hdr;
 341
 342                 /* lrh_len should be initialized in llog_init_handle */
 343                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 344                 llh->llh_count = 1;         /* for the header record */
 345                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 346                 LASSERT(handle->lgh_ctxt->loc_chunk_size >=
 347                                                 LLOG_MIN_CHUNK_SIZE);
 348                 llh->llh_hdr.lrh_len = handle->lgh_ctxt->loc_chunk_size;
 349                 llh->llh_hdr.lrh_index = 0;
 350                 llh->llh_timestamp = ktime_get_real_seconds();
 351                 if (uuid)
 352                         memcpy(&llh->llh_tgtuuid, uuid,
 353                                sizeof(llh->llh_tgtuuid));
 354                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 355                 /* Since update llog header might also call this function,
 356                  * let's reset the bitmap to 0 here */
 357                 memset(LLOG_HDR_BITMAP(llh), 0, llh->llh_hdr.lrh_len -
 358                                                 llh->llh_bitmap_offset -
 359                                                 sizeof(llh->llh_tail));
 360                 ext2_set_bit(0, LLOG_HDR_BITMAP(llh));
 361                 LLOG_HDR_TAIL(llh)->lrt_len = llh->llh_hdr.lrh_len;
 362                 LLOG_HDR_TAIL(llh)->lrt_index = llh->llh_hdr.lrh_index;
 363                 rc = 0;
 364         }
 365         RETURN(rc);
 366 }
 367 EXPORT_SYMBOL(llog_read_header);
 368
 369 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 370                      int flags, struct obd_uuid *uuid)
 371 {
 372         struct llog_log_hdr     *llh;
 373         enum llog_flag           fmt = flags & LLOG_F_EXT_MASK;
 374         int                      rc;
 375         int                     chunk_size = handle->lgh_ctxt->loc_chunk_size;
 376         ENTRY;
 377
 378         LASSERT(handle->lgh_hdr == NULL);
 379
 380         LASSERT(chunk_size >= LLOG_MIN_CHUNK_SIZE);
 381         OBD_ALLOC_LARGE(llh, chunk_size);
 382         if (llh == NULL)
 383                 RETURN(-ENOMEM);
 384
 385         handle->lgh_hdr = llh;
 386         handle->lgh_hdr_size = chunk_size;
 387         /* first assign flags to use llog_client_ops */
 388         llh->llh_flags = flags;
 389         rc = llog_read_header(env, handle, uuid);
 390         if (rc == 0) {
 391                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 392                               flags & LLOG_F_IS_CAT) ||
 393                              (llh->llh_flags & LLOG_F_IS_CAT &&
 394                               flags & LLOG_F_IS_PLAIN))) {
 395                         CERROR("%s: llog type is %s but initializing %s\n",
 396                                handle->lgh_ctxt->loc_obd->obd_name,
 397                                llh->llh_flags & LLOG_F_IS_CAT ?
 398                                "catalog" : "plain",
 399                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 400                         GOTO(out, rc = -EINVAL);
 401                 } else if (llh->llh_flags &
 402                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 403                         /*
 404                          * it is possible to open llog without specifying llog
 405                          * type so it is taken from llh_flags
 406                          */
 407                         flags = llh->llh_flags;
 408                 } else {
 409                         /* for some reason the llh_flags has no type set */
 410                         CERROR("llog type is not specified!\n");
 411                         GOTO(out, rc = -EINVAL);
 412                 }
 413                 if (unlikely(uuid &&
 414                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 415                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 416                                handle->lgh_ctxt->loc_obd->obd_name,
 417                                (char *)uuid->uuid,
 418                                (char *)llh->llh_tgtuuid.uuid);
 419                         GOTO(out, rc = -EEXIST);
 420                 }
 421         }
 422         if (flags & LLOG_F_IS_CAT) {
 423                 LASSERT(list_empty(&handle->u.chd.chd_head));
 424                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 425                 llh->llh_size = sizeof(struct llog_logid_rec);
 426                 llh->llh_flags |= LLOG_F_IS_FIXSIZE;
 427         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 428                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 429                        handle->lgh_ctxt->loc_obd->obd_name,
 430                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 431                 rc = -EINVAL;
 432         }
 433         llh->llh_flags |= fmt;
 434 out:
 435         if (rc) {
 436                 OBD_FREE_LARGE(llh, chunk_size);
 437                 handle->lgh_hdr = NULL;
 438         }
 439         RETURN(rc);
 440 }
 441 EXPORT_SYMBOL(llog_init_handle);
 442
 443 static int llog_process_thread(void *arg)
 444 {
 445         struct llog_process_info        *lpi = arg;
 446         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 447         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 448         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 449         struct llog_thread_info         *lti;
 450         char                            *buf;
 451         size_t                           chunk_size;
 452         __u64                            cur_offset;
 453         int                              rc = 0, index = 1, last_index;
 454         int                              saved_index = 0;
 455         int                              last_called_index = 0;
 456         bool                             repeated = false;
 457
 458         ENTRY;
 459
 460         if (llh == NULL)
 461                 RETURN(-EINVAL);
 462
 463         lti = lpi->lpi_env == NULL ? NULL : llog_info(lpi->lpi_env);
 464
 465         cur_offset = chunk_size = llh->llh_hdr.lrh_len;
 466         /* expect chunk_size to be power of two */
 467         LASSERT(is_power_of_2(chunk_size));
 468
 469         OBD_ALLOC_LARGE(buf, chunk_size);
 470         if (buf == NULL) {
 471                 lpi->lpi_rc = -ENOMEM;
 472                 RETURN(0);
 473         }
 474
 475         if (cd != NULL) {
 476                 last_called_index = cd->lpcd_first_idx;
 477                 index = cd->lpcd_first_idx + 1;
 478         }
 479         if (cd != NULL && cd->lpcd_last_idx)
 480                 last_index = cd->lpcd_last_idx;
 481         else
 482                 last_index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 483
 484         while (rc == 0) {
 485                 struct llog_rec_hdr *rec;
 486                 off_t chunk_offset = 0;
 487                 unsigned int buf_offset = 0;
 488                 bool partial_chunk;
 489                 int     lh_last_idx;
 490                 int     synced_idx = 0;
 491
 492                 /* skip records not set in bitmap */
 493                 while (index <= last_index &&
 494                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 495                         ++index;
 496
 497                 /* There are no indices prior the last_index */
 498                 if (index > last_index)
 499                         break;
 500
 501                 CDEBUG(D_OTHER, "index: %d last_index %d\n", index,
 502                        last_index);
 503
 504 repeat:
 505                 /* get the buf with our target record; avoid old garbage */
 506                 memset(buf, 0, chunk_size);
 507                 /* the record index for outdated chunk data */
 508                 /* it is safe to process buffer until saved lgh_last_idx */
 509                 lh_last_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 510                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 511                                      index, &cur_offset, buf, chunk_size);
 512                 if (repeated && rc)
 513                         CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
 514                                " buf_offset %u, rc = %d\n", cur_offset,
 515                                (__u64)chunk_offset, buf_offset, rc);
 516                 /* we`ve tried to reread the chunk, but there is no
 517                  * new records */
 518                 if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
 519                     cur_offset)
 520                         GOTO(out, rc = 0);
 521                 if (rc != 0)
 522                         GOTO(out, rc);
 523
 524                 /* NB: after llog_next_block() call the cur_offset is the
 525                  * offset of the next block after read one.
 526                  * The absolute offset of the current chunk is calculated
 527                  * from cur_offset value and stored in chunk_offset variable.
 528                  */
 529                 if ((cur_offset & (chunk_size - 1)) != 0) {
 530                         partial_chunk = true;
 531                         chunk_offset = cur_offset & ~(chunk_size - 1);
 532                 } else {
 533                         partial_chunk = false;
 534                         chunk_offset = cur_offset - chunk_size;
 535                 }
 536
 537                 /* NB: when rec->lrh_len is accessed it is already swabbed
 538                  * since it is used at the "end" of the loop and the rec
 539                  * swabbing is done at the beginning of the loop. */
 540                 for (rec = (struct llog_rec_hdr *)(buf + buf_offset);
 541                      (char *)rec < buf + chunk_size;
 542                      rec = llog_rec_hdr_next(rec)) {
 543
 544                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 545                                rec, rec->lrh_type);
 546
 547                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 548                                 lustre_swab_llog_rec(rec);
 549
 550                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 551                                rec->lrh_type, rec->lrh_index);
 552
 553                         if (index == (synced_idx + 1) &&
 554                             synced_idx == LLOG_HDR_TAIL(llh)->lrt_index)
 555                                 GOTO(out, rc = 0);
 556
 557                         if (OBD_FAIL_PRECHECK(OBD_FAIL_LLOG_PROCESS_TIMEOUT) &&
 558                                 cfs_fail_val == (unsigned int)
 559                                         (loghandle->lgh_id.lgl_oi.oi.oi_id &
 560                                          0xFFFFFFFF)) {
 561                                 OBD_RACE(OBD_FAIL_LLOG_PROCESS_TIMEOUT);
 562                         }
 563
 564                         /* the bitmap could be changed during processing
 565                          * records from the chunk. For wrapped catalog
 566                          * it means we can read deleted record and try to
 567                          * process it. Check this case and reread the chunk.
 568                          * It is safe to process to lh_last_idx, including
 569                          * lh_last_idx if it was synced. We can not do <=
 570                          * comparison, cause for wrapped catalog lgh_last_idx
 571                          * could be less than index. So we detect last index
 572                          * for processing as index == lh_last_idx+1. But when
 573                          * catalog is wrapped and full lgh_last_idx=llh_cat_idx,
 574                          * the first processing index is llh_cat_idx+1.
 575                          */
 576
 577                         if ((index == lh_last_idx && synced_idx != index) ||
 578                             (index == (lh_last_idx + 1) &&
 579                              !(index == (llh->llh_cat_idx + 1) &&
 580                                (llh->llh_flags & LLOG_F_IS_CAT))) ||
 581                             (rec->lrh_index == 0 && !repeated)) {
 582
 583                                 /* save offset inside buffer for the re-read */
 584                                 buf_offset = (char *)rec - (char *)buf;
 585                                 cur_offset = chunk_offset;
 586                                 repeated = true;
 587                                 /* We need to be sure lgh_last_idx
 588                                  * record was saved to disk
 589                                  */
 590                                 down_read(&loghandle->lgh_last_sem);
 591                                 synced_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 592                                 up_read(&loghandle->lgh_last_sem);
 593                                 CDEBUG(D_OTHER, "synced_idx: %d\n", synced_idx);
 594                                 goto repeat;
 595
 596                         }
 597
 598                         repeated = false;
 599
 600                         if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
 601                                 CWARN("%s: invalid length %d in llog "DFID
 602                                       "record for index %d/%d\n",
 603                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 604                                        rec->lrh_len,
 605                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 606                                        rec->lrh_index, index);
 607
 608                                 GOTO(out, rc = -EINVAL);
 609                         }
 610
 611                         if (rec->lrh_index < index) {
 612                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 613                                        rec->lrh_index);
 614                                 continue;
 615                         }
 616
 617                         if (rec->lrh_index != index) {
 618                                 CERROR("%s: "DFID" Invalid record: index %u"
 619                                        " but expected %u\n",
 620                                        loghandle->lgh_ctxt->loc_obd->obd_name,
 621                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 622                                        rec->lrh_index, index);
 623                                 GOTO(out, rc = -ERANGE);
 624                         }
 625
 626                         CDEBUG(D_OTHER,
 627                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 628                                rec->lrh_index, rec->lrh_len,
 629                                (int)(buf + chunk_size - (char *)rec));
 630
 631                         /* lgh_cur_offset is used only at llog_test_3 */
 632                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 633                                                     chunk_offset;
 634
 635                         /* if set, process the callback on this record */
 636                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 637                                 struct llog_cookie *lgc;
 638                                 __u64   tmp_off;
 639                                 int     tmp_idx;
 640
 641                                 CDEBUG(D_OTHER, "index: %d, lh_last_idx: %d "
 642                                        "synced_idx: %d lgh_last_idx: %d\n",
 643                                        index, lh_last_idx, synced_idx,
 644                                        loghandle->lgh_last_idx);
 645
 646                                 if (lti != NULL) {
 647                                         lgc = &lti->lgi_cookie;
 648                                         /* store lu_env for recursive calls */
 649                                         tmp_off = lgc->lgc_offset;
 650                                         tmp_idx = lgc->lgc_index;
 651
 652                                         lgc->lgc_offset = (char *)rec -
 653                                                 (char *)buf + chunk_offset;
 654                                         lgc->lgc_index = rec->lrh_index;
 655                                 }
 656                                 /* using lu_env for passing record offset to
 657                                  * llog_write through various callbacks */
 658                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 659                                                  lpi->lpi_cbdata);
 660                                 last_called_index = index;
 661
 662                                 if (lti != NULL) {
 663                                         lgc->lgc_offset = tmp_off;
 664                                         lgc->lgc_index = tmp_idx;
 665                                 }
 666
 667                                 if (rc == LLOG_PROC_BREAK) {
 668                                         GOTO(out, rc);
 669                                 } else if (rc == LLOG_DEL_RECORD) {
 670                                         rc = llog_cancel_rec(lpi->lpi_env,
 671                                                              loghandle,
 672                                                              rec->lrh_index);
 673                                 }
 674                                 if (rc)
 675                                         GOTO(out, rc);
 676                                 /* some stupid callbacks directly cancel records
 677                                  * and delete llog. Check it and stop
 678                                  * processing. */
 679                                 if (loghandle->lgh_hdr == NULL ||
 680                                     loghandle->lgh_hdr->llh_count == 1)
 681                                         GOTO(out, rc = 0);
 682                         }
 683                         /* exit if the last index is reached */
 684                         if (index >= last_index)
 685                                 GOTO(out, rc = 0);
 686                         ++index;
 687                 }
 688         }
 689
 690 out:
 691         if (cd != NULL)
 692                 cd->lpcd_last_idx = last_called_index;
 693
 694         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 695                 if (dt_object_remote(loghandle->lgh_obj)) {
 696                         /* If it is remote object, then -EIO might means
 697                          * disconnection or eviction, let's return -EAGAIN,
 698                          * so for update recovery log processing, it will
 699                          * retry until the umount or abort recovery, see
 700                          * lod_sub_recovery_thread() */
 701                         CERROR("%s retry remote llog process\n",
 702                                loghandle->lgh_ctxt->loc_obd->obd_name);
 703                         rc = -EAGAIN;
 704                 } else {
 705                         /* something bad happened to the processing of a local
 706                          * llog file, probably I/O error or the log got
 707                          * corrupted to be able to finally release the log we
 708                          * discard any remaining bits in the header */
 709                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 710                                " %s index %d count %d\n",
 711                                loghandle->lgh_ctxt->loc_obd->obd_name,
 712                                POSTID(&loghandle->lgh_id.lgl_oi),
 713                                loghandle->lgh_id.lgl_ogen,
 714                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 715                                 "plain"), index, llh->llh_count);
 716
 717                         while (index <= last_index) {
 718                                 if (ext2_test_bit(index,
 719                                                   LLOG_HDR_BITMAP(llh)) != 0)
 720                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 721                                                         index);
 722                                 index++;
 723                         }
 724                         rc = 0;
 725                 }
 726         }
 727
 728         OBD_FREE_LARGE(buf, chunk_size);
 729         lpi->lpi_rc = rc;
 730         return 0;
 731 }
 732
 733 static int llog_process_thread_daemonize(void *arg)
 734 {
 735         struct llog_process_info        *lpi = arg;
 736         struct lu_env                    env;
 737         int                              rc;
 738         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 739
 740         task_lock(lpi->lpi_reftask);
 741         new_ns = lpi->lpi_reftask->nsproxy;
 742         if (curr_ns != new_ns) {
 743                 get_nsproxy(new_ns);
 744
 745                 current->nsproxy = new_ns;
 746                 /* XXX: we should call put_nsproxy() instead of
 747                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 748                  * used outside of the kernel itself, because it calls
 749                  * free_nsproxy() which is not exported by the kernel
 750                  * (defined in kernel/nsproxy.c) */
 751                 atomic_dec(&curr_ns->count);
 752         }
 753         task_unlock(lpi->lpi_reftask);
 754
 755         unshare_fs_struct();
 756
 757         /* client env has no keys, tags is just 0 */
 758         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 759         if (rc)
 760                 goto out;
 761         lpi->lpi_env = &env;
 762
 763         rc = llog_process_thread(arg);
 764
 765         lu_env_fini(&env);
 766 out:
 767         complete(&lpi->lpi_completion);
 768         return rc;
 769 }
 770
 771 int llog_process_or_fork(const struct lu_env *env,
 772                          struct llog_handle *loghandle,
 773                          llog_cb_t cb, void *data, void *catdata, bool fork)
 774 {
 775         struct llog_process_info *lpi;
 776         int                      rc;
 777
 778         ENTRY;
 779
 780         OBD_ALLOC_PTR(lpi);
 781         if (lpi == NULL) {
 782                 CERROR("cannot alloc pointer\n");
 783                 RETURN(-ENOMEM);
 784         }
 785         lpi->lpi_loghandle = loghandle;
 786         lpi->lpi_cb        = cb;
 787         lpi->lpi_cbdata    = data;
 788         lpi->lpi_catdata   = catdata;
 789
 790         if (fork) {
 791                 struct task_struct *task;
 792
 793                 /* The new thread can't use parent env,
 794                  * init the new one in llog_process_thread_daemonize. */
 795                 lpi->lpi_env = NULL;
 796                 init_completion(&lpi->lpi_completion);
 797                 /* take reference to current, so that
 798                  * llog_process_thread_daemonize() can use it to switch to
 799                  * namespace associated with current  */
 800                 lpi->lpi_reftask = current;
 801                 task = kthread_run(llog_process_thread_daemonize, lpi,
 802                                    "llog_process_thread");
 803                 if (IS_ERR(task)) {
 804                         rc = PTR_ERR(task);
 805                         CERROR("%s: cannot start thread: rc = %d\n",
 806                                loghandle->lgh_ctxt->loc_obd->obd_name, rc);
 807                         GOTO(out_lpi, rc);
 808                 }
 809                 wait_for_completion(&lpi->lpi_completion);
 810         } else {
 811                 lpi->lpi_env = env;
 812                 llog_process_thread(lpi);
 813         }
 814         rc = lpi->lpi_rc;
 815
 816 out_lpi:
 817         OBD_FREE_PTR(lpi);
 818         RETURN(rc);
 819 }
 820 EXPORT_SYMBOL(llog_process_or_fork);
 821
 822 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 823                  llog_cb_t cb, void *data, void *catdata)
 824 {
 825         int rc;
 826         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 827         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 828 }
 829 EXPORT_SYMBOL(llog_process);
 830
 831 int llog_reverse_process(const struct lu_env *env,
 832                          struct llog_handle *loghandle, llog_cb_t cb,
 833                          void *data, void *catdata)
 834 {
 835         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 836         struct llog_process_cat_data *cd = catdata;
 837         void *buf;
 838         int rc = 0, first_index = 1, index, idx;
 839         __u32   chunk_size = llh->llh_hdr.lrh_len;
 840         ENTRY;
 841
 842         OBD_ALLOC_LARGE(buf, chunk_size);
 843         if (buf == NULL)
 844                 RETURN(-ENOMEM);
 845
 846         if (cd != NULL)
 847                 first_index = cd->lpcd_first_idx + 1;
 848         if (cd != NULL && cd->lpcd_last_idx)
 849                 index = cd->lpcd_last_idx;
 850         else
 851                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 852
 853         while (rc == 0) {
 854                 struct llog_rec_hdr *rec;
 855                 struct llog_rec_tail *tail;
 856
 857                 /* skip records not set in bitmap */
 858                 while (index >= first_index &&
 859                        !ext2_test_bit(index, LLOG_HDR_BITMAP(llh)))
 860                         --index;
 861
 862                 LASSERT(index >= first_index - 1);
 863                 if (index == first_index - 1)
 864                         break;
 865
 866                 /* get the buf with our target record; avoid old garbage */
 867                 memset(buf, 0, chunk_size);
 868                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 869                 if (rc)
 870                         GOTO(out, rc);
 871
 872                 rec = buf;
 873                 idx = rec->lrh_index;
 874                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 875                 while (idx < index) {
 876                         rec = (void *)rec + rec->lrh_len;
 877                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 878                                 lustre_swab_llog_rec(rec);
 879                         idx ++;
 880                 }
 881                 LASSERT(idx == index);
 882                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 883
 884                 /* process records in buffer, starting where we found one */
 885                 while ((void *)tail > buf) {
 886                         if (tail->lrt_index == 0)
 887                                 GOTO(out, rc = 0); /* no more records */
 888
 889                         /* if set, process the callback on this record */
 890                         if (ext2_test_bit(index, LLOG_HDR_BITMAP(llh))) {
 891                                 rec = (void *)tail - tail->lrt_len +
 892                                       sizeof(*tail);
 893
 894                                 rc = cb(env, loghandle, rec, data);
 895                                 if (rc == LLOG_PROC_BREAK) {
 896                                         GOTO(out, rc);
 897                                 } else if (rc == LLOG_DEL_RECORD) {
 898                                         rc = llog_cancel_rec(env, loghandle,
 899                                                              tail->lrt_index);
 900                                 }
 901                                 if (rc)
 902                                         GOTO(out, rc);
 903                         }
 904
 905                         /* previous record, still in buffer? */
 906                         --index;
 907                         if (index < first_index)
 908                                 GOTO(out, rc = 0);
 909                         tail = (void *)tail - tail->lrt_len;
 910                 }
 911         }
 912
 913 out:
 914         if (buf != NULL)
 915                 OBD_FREE_LARGE(buf, chunk_size);
 916         RETURN(rc);
 917 }
 918 EXPORT_SYMBOL(llog_reverse_process);
 919
 920 /**
 921  * new llog API
 922  *
 923  * API functions:
 924  *      llog_open - open llog, may not exist
 925  *      llog_exist - check if llog exists
 926  *      llog_close - close opened llog, pair for open, frees llog_handle
 927  *      llog_declare_create - declare llog creation
 928  *      llog_create - create new llog on disk, need transaction handle
 929  *      llog_declare_write_rec - declaration of llog write
 930  *      llog_write_rec - write llog record on disk, need transaction handle
 931  *      llog_declare_add - declare llog catalog record addition
 932  *      llog_add - add llog record in catalog, need transaction handle
 933  */
 934 int llog_exist(struct llog_handle *loghandle)
 935 {
 936         struct llog_operations  *lop;
 937         int                      rc;
 938
 939         ENTRY;
 940
 941         rc = llog_handle2ops(loghandle, &lop);
 942         if (rc)
 943                 RETURN(rc);
 944         if (lop->lop_exist == NULL)
 945                 RETURN(-EOPNOTSUPP);
 946
 947         rc = lop->lop_exist(loghandle);
 948         RETURN(rc);
 949 }
 950 EXPORT_SYMBOL(llog_exist);
 951
 952 int llog_declare_create(const struct lu_env *env,
 953                         struct llog_handle *loghandle, struct thandle *th)
 954 {
 955         struct llog_operations  *lop;
 956         int                      raised, rc;
 957
 958         ENTRY;
 959
 960         rc = llog_handle2ops(loghandle, &lop);
 961         if (rc)
 962                 RETURN(rc);
 963         if (lop->lop_declare_create == NULL)
 964                 RETURN(-EOPNOTSUPP);
 965
 966         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 967         if (!raised)
 968                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 969         rc = lop->lop_declare_create(env, loghandle, th);
 970         if (!raised)
 971                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 972         RETURN(rc);
 973 }
 974
 975 int llog_create(const struct lu_env *env, struct llog_handle *handle,
 976                 struct thandle *th)
 977 {
 978         struct llog_operations  *lop;
 979         int                      raised, rc;
 980
 981         ENTRY;
 982
 983         rc = llog_handle2ops(handle, &lop);
 984         if (rc)
 985                 RETURN(rc);
 986         if (lop->lop_create == NULL)
 987                 RETURN(-EOPNOTSUPP);
 988
 989         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
 990         if (!raised)
 991                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 992         rc = lop->lop_create(env, handle, th);
 993         if (!raised)
 994                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 995         RETURN(rc);
 996 }
 997
 998 int llog_declare_write_rec(const struct lu_env *env,
 999                            struct llog_handle *handle,
1000                            struct llog_rec_hdr *rec, int idx,
1001                            struct thandle *th)
1002 {
1003         struct llog_operations  *lop;
1004         int                      raised, rc;
1005
1006         ENTRY;
1007
1008         rc = llog_handle2ops(handle, &lop);
1009         if (rc)
1010                 RETURN(rc);
1011         LASSERT(lop);
1012         if (lop->lop_declare_write_rec == NULL)
1013                 RETURN(-EOPNOTSUPP);
1014
1015         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1016         if (!raised)
1017                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1018         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
1019         if (!raised)
1020                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1021         RETURN(rc);
1022 }
1023
1024 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
1025                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1026                    int idx, struct thandle *th)
1027 {
1028         struct llog_operations  *lop;
1029         int                      raised, rc, buflen;
1030
1031         ENTRY;
1032
1033         /* API sanity checks */
1034         if (handle == NULL) {
1035                 CERROR("loghandle is missed\n");
1036                 RETURN(-EPROTO);
1037         } else if (handle->lgh_obj == NULL) {
1038                 CERROR("loghandle %p with NULL object\n",
1039                         handle);
1040                 RETURN(-EPROTO);
1041         } else if (th == NULL) {
1042                 CERROR("%s: missed transaction handle\n",
1043                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name);
1044                 RETURN(-EPROTO);
1045         } else if (handle->lgh_hdr == NULL) {
1046                 CERROR("%s: loghandle %p with no header\n",
1047                         handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name,
1048                         handle);
1049                 RETURN(-EPROTO);
1050         }
1051
1052         rc = llog_handle2ops(handle, &lop);
1053         if (rc)
1054                 RETURN(rc);
1055
1056         if (lop->lop_write_rec == NULL)
1057                 RETURN(-EOPNOTSUPP);
1058
1059         buflen = rec->lrh_len;
1060         LASSERT(cfs_size_round(buflen) == buflen);
1061
1062         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1063         if (!raised)
1064                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1065         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1066         if (!raised)
1067                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1068         RETURN(rc);
1069 }
1070
1071 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1072              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1073              struct thandle *th)
1074 {
1075         int raised, rc;
1076
1077         ENTRY;
1078
1079         if (lgh->lgh_logops->lop_add == NULL)
1080                 RETURN(-EOPNOTSUPP);
1081
1082         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1083         if (!raised)
1084                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1085         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1086         if (!raised)
1087                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1088         RETURN(rc);
1089 }
1090 EXPORT_SYMBOL(llog_add);
1091
1092 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1093                      struct llog_rec_hdr *rec, struct thandle *th)
1094 {
1095         int raised, rc;
1096
1097         ENTRY;
1098
1099         if (lgh->lgh_logops->lop_declare_add == NULL)
1100                 RETURN(-EOPNOTSUPP);
1101
1102         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1103         if (!raised)
1104                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1105         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1106         if (!raised)
1107                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1108         RETURN(rc);
1109 }
1110 EXPORT_SYMBOL(llog_declare_add);
1111
1112 /**
1113  * Helper function to open llog or create it if doesn't exist.
1114  * It hides all transaction handling from caller.
1115  */
1116 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1117                      struct llog_handle **res, struct llog_logid *logid,
1118                      char *name)
1119 {
1120         struct dt_device        *d;
1121         struct thandle          *th;
1122         int                      rc;
1123
1124         ENTRY;
1125
1126         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1127         if (rc)
1128                 RETURN(rc);
1129
1130         if (llog_exist(*res))
1131                 RETURN(0);
1132
1133         LASSERT((*res)->lgh_obj != NULL);
1134
1135         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1136
1137         if (unlikely(unlikely(d->dd_rdonly)))
1138                 RETURN(-EROFS);
1139
1140         th = dt_trans_create(env, d);
1141         if (IS_ERR(th))
1142                 GOTO(out, rc = PTR_ERR(th));
1143
1144         /* Create update llog object synchronously, which
1145          * happens during inialization process see
1146          * lod_sub_prep_llog(), to make sure the update
1147          * llog object is created before corss-MDT writing
1148          * updates into the llog object */
1149         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1150                 th->th_sync = 1;
1151
1152         th->th_wait_submit = 1;
1153         rc = llog_declare_create(env, *res, th);
1154         if (rc == 0) {
1155                 rc = dt_trans_start_local(env, d, th);
1156                 if (rc == 0)
1157                         rc = llog_create(env, *res, th);
1158         }
1159         dt_trans_stop(env, d, th);
1160 out:
1161         if (rc)
1162                 llog_close(env, *res);
1163         RETURN(rc);
1164 }
1165 EXPORT_SYMBOL(llog_open_create);
1166
1167 /**
1168  * Helper function to delete existent llog.
1169  */
1170 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1171                struct llog_logid *logid, char *name)
1172 {
1173         struct llog_handle      *handle;
1174         int                      rc = 0, rc2;
1175
1176         ENTRY;
1177
1178         /* nothing to erase */
1179         if (name == NULL && logid == NULL)
1180                 RETURN(0);
1181
1182         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1183         if (rc < 0)
1184                 RETURN(rc);
1185
1186         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1187         if (rc == 0)
1188                 rc = llog_destroy(env, handle);
1189
1190         rc2 = llog_close(env, handle);
1191         if (rc == 0)
1192                 rc = rc2;
1193         RETURN(rc);
1194 }
1195 EXPORT_SYMBOL(llog_erase);
1196
1197 /*
1198  * Helper function for write record in llog.
1199  * It hides all transaction handling from caller.
1200  * Valid only with local llog.
1201  */
1202 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1203                struct llog_rec_hdr *rec, int idx)
1204 {
1205         struct dt_device        *dt;
1206         struct thandle          *th;
1207         bool                    need_cookie;
1208         int                     rc;
1209
1210         ENTRY;
1211
1212         LASSERT(loghandle);
1213         LASSERT(loghandle->lgh_ctxt);
1214         LASSERT(loghandle->lgh_obj != NULL);
1215
1216         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1217
1218         if (unlikely(unlikely(dt->dd_rdonly)))
1219                 RETURN(-EROFS);
1220
1221         th = dt_trans_create(env, dt);
1222         if (IS_ERR(th))
1223                 RETURN(PTR_ERR(th));
1224
1225         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1226         if (rc)
1227                 GOTO(out_trans, rc);
1228
1229         th->th_wait_submit = 1;
1230         rc = dt_trans_start_local(env, dt, th);
1231         if (rc)
1232                 GOTO(out_trans, rc);
1233
1234         need_cookie = !(idx == LLOG_HEADER_IDX || idx == LLOG_NEXT_IDX);
1235
1236         down_write(&loghandle->lgh_lock);
1237         if (need_cookie) {
1238                 struct llog_thread_info *lti = llog_info(env);
1239
1240                 /* cookie comes from llog_process_thread */
1241                 rc = llog_write_rec(env, loghandle, rec, &lti->lgi_cookie,
1242                                     rec->lrh_index, th);
1243                 /* upper layer didn`t pass cookie so change rc */
1244                 rc = (rc == 1 ? 0 : rc);
1245         } else {
1246                 rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1247         }
1248
1249         up_write(&loghandle->lgh_lock);
1250 out_trans:
1251         dt_trans_stop(env, dt, th);
1252         RETURN(rc);
1253 }
1254 EXPORT_SYMBOL(llog_write);
1255
1256 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1257               struct llog_handle **lgh, struct llog_logid *logid,
1258               char *name, enum llog_open_param open_param)
1259 {
1260         int      raised;
1261         int      rc;
1262
1263         ENTRY;
1264
1265         LASSERT(ctxt);
1266         LASSERT(ctxt->loc_logops);
1267
1268         if (ctxt->loc_logops->lop_open == NULL) {
1269                 *lgh = NULL;
1270                 RETURN(-EOPNOTSUPP);
1271         }
1272
1273         *lgh = llog_alloc_handle();
1274         if (*lgh == NULL)
1275                 RETURN(-ENOMEM);
1276         (*lgh)->lgh_ctxt = ctxt;
1277         (*lgh)->lgh_logops = ctxt->loc_logops;
1278
1279         raised = cfs_cap_raised(CFS_CAP_SYS_RESOURCE);
1280         if (!raised)
1281                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1282         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1283         if (!raised)
1284                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1285         if (rc) {
1286                 llog_free_handle(*lgh);
1287                 *lgh = NULL;
1288         }
1289         RETURN(rc);
1290 }
1291 EXPORT_SYMBOL(llog_open);
1292
1293 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1294 {
1295         struct llog_operations  *lop;
1296         int                      rc;
1297
1298         ENTRY;
1299
1300         rc = llog_handle2ops(loghandle, &lop);
1301         if (rc)
1302                 GOTO(out, rc);
1303         if (lop->lop_close == NULL)
1304                 GOTO(out, rc = -EOPNOTSUPP);
1305         rc = lop->lop_close(env, loghandle);
1306 out:
1307         llog_handle_put(loghandle);
1308         RETURN(rc);
1309 }
1310 EXPORT_SYMBOL(llog_close);
1311
1312 /**
1313  * Helper function to get the llog size in records. It is used by MGS
1314  * mostly to check that config llog exists and contains data.
1315  *
1316  * \param[in] env       execution environment
1317  * \param[in] ctxt      llog context
1318  * \param[in] name      llog name
1319  *
1320  * \retval              true if there are records in llog besides a header
1321  * \retval              false on error or llog without records
1322  */
1323 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1324                   char *name)
1325 {
1326         struct llog_handle      *llh;
1327         int                      rc = 0;
1328
1329         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1330         if (rc < 0) {
1331                 if (likely(rc == -ENOENT))
1332                         rc = 0;
1333                 GOTO(out, rc);
1334         }
1335
1336         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1337         if (rc)
1338                 GOTO(out_close, rc);
1339         rc = llog_get_size(llh);
1340
1341 out_close:
1342         llog_close(env, llh);
1343 out:
1344         /* The header is record 1, the llog is still considered as empty
1345          * if there is only header */
1346         return (rc <= 1);
1347 }
1348 EXPORT_SYMBOL(llog_is_empty);
1349
1350 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1351                       struct llog_rec_hdr *rec, void *data)
1352 {
1353         struct llog_handle      *copy_llh = data;
1354
1355         /* Append all records */
1356         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1357 }
1358
1359 /* backup plain llog */
1360 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1361                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1362                 char *name, char *backup)
1363 {
1364         struct llog_handle      *llh, *bllh;
1365         int                      rc;
1366
1367         ENTRY;
1368
1369         /* open original log */
1370         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1371         if (rc < 0) {
1372                 /* the -ENOENT case is also reported to the caller
1373                  * but silently so it should handle that if needed.
1374                  */
1375                 if (rc != -ENOENT)
1376                         CERROR("%s: failed to open log %s: rc = %d\n",
1377                                obd->obd_name, name, rc);
1378                 RETURN(rc);
1379         }
1380
1381         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1382         if (rc)
1383                 GOTO(out_close, rc);
1384
1385         /* Make sure there's no old backup log */
1386         rc = llog_erase(env, bctxt, NULL, backup);
1387         if (rc < 0 && rc != -ENOENT)
1388                 GOTO(out_close, rc);
1389
1390         /* open backup log */
1391         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1392         if (rc) {
1393                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1394                        obd->obd_name, backup, rc);
1395                 GOTO(out_close, rc);
1396         }
1397
1398         /* check that backup llog is not the same object as original one */
1399         if (llh->lgh_obj == bllh->lgh_obj) {
1400                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1401                        obd->obd_name, name, backup, llh->lgh_obj,
1402                        bllh->lgh_obj);
1403                 GOTO(out_backup, rc = -EEXIST);
1404         }
1405
1406         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1407         if (rc)
1408                 GOTO(out_backup, rc);
1409
1410         /* Copy log record by record */
1411         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1412                                   NULL, false);
1413         if (rc)
1414                 CERROR("%s: failed to backup log %s: rc = %d\n",
1415                        obd->obd_name, name, rc);
1416 out_backup:
1417         llog_close(env, bllh);
1418 out_close:
1419         llog_close(env, llh);
1420         RETURN(rc);
1421 }
1422 EXPORT_SYMBOL(llog_backup);
1423
1424 /* Get size of llog */
1425 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1426 {
1427         int rc;
1428         struct lu_attr la;
1429
1430         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1431         if (rc) {
1432                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1433                        llh->lgh_ctxt->loc_obd->obd_name,
1434                        PFID(&llh->lgh_id.lgl_oi.oi_fid), rc);
1435                 return 0;
1436         }
1437
1438         return la.la_size;
1439 }
1440 EXPORT_SYMBOL(llog_size);
1441