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  *
  31  * lustre/obdclass/llog.c
  32  *
  33  * OST<->MDS recovery logging infrastructure.
  34  * Invariants in implementation:
  35  * - we do not share logs among different OST<->MDS connections, so that
  36  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  37  *
  38  * Author: Andreas Dilger <adilger@clusterfs.com>
  39  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  40  * Author: Mikhail Pershin <tappro@whamcloud.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_LOG
  44
  45 #include <linux/pid_namespace.h>
  46 #include <linux/kthread.h>
  47 #include <llog_swab.h>
  48 #include <lustre_log.h>
  49 #include <obd_support.h>
  50 #include <obd_class.h>
  51 #include "llog_internal.h"
  52
  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         refcount_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 struct llog_handle *llog_handle_get(struct llog_handle *loghandle)
  95 {
  96         if (refcount_inc_not_zero(&loghandle->lgh_refcount))
  97                 return loghandle;
  98         return NULL;
  99 }
 100
 101 int llog_handle_put(const struct lu_env *env, struct llog_handle *loghandle)
 102 {
 103         int rc = 0;
 104
 105         if (refcount_dec_and_test(&loghandle->lgh_refcount)) {
 106                 const struct llog_operations *lop;
 107
 108                 rc = llog_handle2ops(loghandle, &lop);
 109                 if (!rc) {
 110                         if (lop->lop_close)
 111                                 rc = lop->lop_close(env, loghandle);
 112                         else
 113                                 rc = -EOPNOTSUPP;
 114                 }
 115                 llog_free_handle(loghandle);
 116         }
 117         return rc;
 118 }
 119
 120 static int llog_declare_destroy(const struct lu_env *env,
 121                                 struct llog_handle *handle,
 122                                 struct thandle *th)
 123 {
 124         const struct llog_operations *lop;
 125         int rc;
 126
 127         ENTRY;
 128
 129         rc = llog_handle2ops(handle, &lop);
 130         if (rc)
 131                 RETURN(rc);
 132         if (lop->lop_declare_destroy == NULL)
 133                 RETURN(-EOPNOTSUPP);
 134
 135         rc = lop->lop_declare_destroy(env, handle, th);
 136
 137         RETURN(rc);
 138 }
 139
 140 int llog_trans_destroy(const struct lu_env *env, struct llog_handle *handle,
 141                        struct thandle *th)
 142 {
 143         const struct llog_operations *lop;
 144         int rc;
 145         ENTRY;
 146
 147         rc = llog_handle2ops(handle, &lop);
 148         if (rc < 0)
 149                 RETURN(rc);
 150         if (lop->lop_destroy == NULL)
 151                 RETURN(-EOPNOTSUPP);
 152
 153         LASSERT(handle->lgh_obj != NULL);
 154         if (!llog_exist(handle))
 155                 RETURN(0);
 156
 157         rc = lop->lop_destroy(env, handle, th);
 158
 159         RETURN(rc);
 160 }
 161
 162 int llog_destroy(const struct lu_env *env, struct llog_handle *handle)
 163 {
 164         const struct llog_operations *lop;
 165         struct dt_device *dt;
 166         struct thandle *th;
 167         int rc;
 168
 169         ENTRY;
 170
 171         rc = llog_handle2ops(handle, &lop);
 172         if (rc < 0)
 173                 RETURN(rc);
 174         if (lop->lop_destroy == NULL)
 175                 RETURN(-EOPNOTSUPP);
 176
 177         if (handle->lgh_obj == NULL) {
 178                 /* if lgh_obj == NULL, then it is from client side destroy */
 179                 rc = lop->lop_destroy(env, handle, NULL);
 180                 RETURN(rc);
 181         }
 182
 183         if (!llog_exist(handle))
 184                 RETURN(0);
 185
 186         dt = lu2dt_dev(handle->lgh_obj->do_lu.lo_dev);
 187
 188         if (unlikely(unlikely(dt->dd_rdonly)))
 189                 RETURN(-EROFS);
 190
 191         th = dt_trans_create(env, dt);
 192         if (IS_ERR(th))
 193                 RETURN(PTR_ERR(th));
 194
 195         rc = llog_declare_destroy(env, handle, th);
 196         if (rc != 0)
 197                 GOTO(out_trans, rc);
 198
 199         rc = dt_trans_start_local(env, dt, th);
 200         if (rc < 0)
 201                 GOTO(out_trans, rc);
 202
 203         rc = lop->lop_destroy(env, handle, th);
 204
 205 out_trans:
 206         dt_trans_stop(env, dt, th);
 207
 208         RETURN(rc);
 209 }
 210 EXPORT_SYMBOL(llog_destroy);
 211
 212 /* returns negative on error; 0 if success; 1 if success & log destroyed */
 213 int llog_cancel_arr_rec(const struct lu_env *env, struct llog_handle *loghandle,
 214                         int num, int *index)
 215 {
 216         struct llog_thread_info *lgi = llog_info(env);
 217         struct dt_device        *dt;
 218         struct llog_log_hdr     *llh;
 219         struct thandle          *th;
 220         __u32                    tmp_lgc_index;
 221         int                      rc, i = 0;
 222         int rc1;
 223         bool subtract_count = false;
 224
 225         ENTRY;
 226
 227         LASSERT(loghandle != NULL);
 228         LASSERT(loghandle->lgh_ctxt != NULL);
 229         LASSERT(loghandle->lgh_obj != NULL);
 230
 231         llh = loghandle->lgh_hdr;
 232
 233         CDEBUG(D_RPCTRACE, "Canceling %d records, first %d in log "DFID"\n",
 234                num, index[0], PFID(&loghandle->lgh_id.lgl_oi.oi_fid));
 235
 236         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 237
 238         if (unlikely(unlikely(dt->dd_rdonly)))
 239                 RETURN(0);
 240
 241         th = dt_trans_create(env, dt);
 242         if (IS_ERR(th))
 243                 RETURN(PTR_ERR(th));
 244
 245         rc = llog_declare_write_rec(env, loghandle, &llh->llh_hdr, 0, th);
 246         if (rc < 0)
 247                 GOTO(out_trans, rc);
 248
 249         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY)) {
 250                 rc = llog_declare_destroy(env, loghandle, th);
 251                 if (rc < 0)
 252                         GOTO(out_trans, rc);
 253         }
 254
 255         th->th_wait_submit = 1;
 256         rc = dt_trans_start_local(env, dt, th);
 257         if (rc < 0)
 258                 GOTO(out_trans, rc);
 259
 260         down_write(&loghandle->lgh_lock);
 261         /* clear bitmap */
 262         mutex_lock(&loghandle->lgh_hdr_mutex);
 263         for (i = 0; i < num; ++i) {
 264                 if (index[i] == 0) {
 265                         CERROR("Can't cancel index 0 which is header\n");
 266                         GOTO(out_unlock, rc = -EINVAL);
 267                 }
 268                 if (!__test_and_clear_bit_le(index[i], LLOG_HDR_BITMAP(llh))) {
 269                         CDEBUG(D_RPCTRACE, "Catalog index %u already clear?\n",
 270                                index[i]);
 271                         GOTO(out_unlock, rc = -ENOENT);
 272                 }
 273         }
 274         loghandle->lgh_hdr->llh_count -= num;
 275         subtract_count = true;
 276
 277         /* Since llog_process_thread use lgi_cookie, it`s better to save them
 278          * and restore after using
 279          */
 280         tmp_lgc_index = lgi->lgi_cookie.lgc_index;
 281         /* Pass this index to llog_osd_write_rec(), which will use the index
 282          * to only update the necesary bitmap. */
 283         lgi->lgi_cookie.lgc_index = index[0];
 284         /* update header */
 285         rc = llog_write_rec(env, loghandle, &llh->llh_hdr, (num != 1 ? NULL :
 286                             &lgi->lgi_cookie), LLOG_HEADER_IDX, th);
 287         lgi->lgi_cookie.lgc_index = tmp_lgc_index;
 288
 289         if (rc != 0)
 290                 GOTO(out_unlock, rc);
 291
 292         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 293             (llh->llh_count == 1) &&
 294             ((loghandle->lgh_last_idx == LLOG_HDR_BITMAP_SIZE(llh) - 1) ||
 295              (loghandle->u.phd.phd_cat_handle != NULL &&
 296               loghandle->u.phd.phd_cat_handle->u.chd.chd_current_log !=
 297                 loghandle))) {
 298                 /* never try to destroy it again */
 299                 llh->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 300                 rc = llog_trans_destroy(env, loghandle, th);
 301                 if (rc < 0) {
 302                         /* Sigh, can not destroy the final plain llog, but
 303                          * the bitmap has been clearly, so the record can not
 304                          * be accessed anymore, let's return 0 for now, and
 305                          * the orphan will be handled by LFSCK. */
 306                         CERROR("%s: can't destroy empty llog "DFID": rc = %d\n",
 307                                loghandle2name(loghandle),
 308                                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rc);
 309                         GOTO(out_unlock, rc = 0);
 310                 }
 311                 rc = LLOG_DEL_PLAIN;
 312         }
 313
 314 out_unlock:
 315         if (rc < 0) {
 316                 /* restore bitmap while holding a mutex */
 317                 if (subtract_count) {
 318                         loghandle->lgh_hdr->llh_count += num;
 319                         subtract_count = false;
 320                 }
 321                 for (i = i - 1; i >= 0; i--)
 322                         set_bit_le(index[i], LLOG_HDR_BITMAP(llh));
 323         }
 324         mutex_unlock(&loghandle->lgh_hdr_mutex);
 325         up_write(&loghandle->lgh_lock);
 326 out_trans:
 327         rc1 = dt_trans_stop(env, dt, th);
 328         if (rc == 0)
 329                 rc = rc1;
 330         if (rc1 < 0) {
 331                 mutex_lock(&loghandle->lgh_hdr_mutex);
 332                 if (subtract_count)
 333                         loghandle->lgh_hdr->llh_count += num;
 334                 for (i = i - 1; i >= 0; i--)
 335                         set_bit_le(index[i], LLOG_HDR_BITMAP(llh));
 336                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 337         }
 338         RETURN(rc);
 339 }
 340
 341 int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
 342                     int index)
 343 {
 344         return llog_cancel_arr_rec(env, loghandle, 1, &index);
 345 }
 346
 347 int llog_read_header(const struct lu_env *env, struct llog_handle *handle,
 348                      const struct obd_uuid *uuid)
 349 {
 350         const struct llog_operations *lop;
 351         int rc;
 352         ENTRY;
 353
 354         rc = llog_handle2ops(handle, &lop);
 355         if (rc)
 356                 RETURN(rc);
 357
 358         if (lop->lop_read_header == NULL)
 359                 RETURN(-EOPNOTSUPP);
 360
 361         rc = lop->lop_read_header(env, handle);
 362         if (rc == LLOG_EEMPTY) {
 363                 struct llog_log_hdr *llh = handle->lgh_hdr;
 364
 365                 /* lrh_len should be initialized in llog_init_handle */
 366                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 367                 llh->llh_count = 1;         /* for the header record */
 368                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 369                 LASSERT(handle->lgh_ctxt->loc_chunk_size >=
 370                                                 LLOG_MIN_CHUNK_SIZE);
 371                 llh->llh_hdr.lrh_len = handle->lgh_ctxt->loc_chunk_size;
 372                 llh->llh_hdr.lrh_index = 0;
 373                 llh->llh_timestamp = ktime_get_real_seconds();
 374                 if (uuid)
 375                         memcpy(&llh->llh_tgtuuid, uuid,
 376                                sizeof(llh->llh_tgtuuid));
 377                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 378                 /* Since update llog header might also call this function,
 379                  * let's reset the bitmap to 0 here */
 380                 memset(LLOG_HDR_BITMAP(llh), 0, llh->llh_hdr.lrh_len -
 381                                                 llh->llh_bitmap_offset -
 382                                                 sizeof(llh->llh_tail));
 383                 set_bit_le(0, LLOG_HDR_BITMAP(llh));
 384                 LLOG_HDR_TAIL(llh)->lrt_len = llh->llh_hdr.lrh_len;
 385                 LLOG_HDR_TAIL(llh)->lrt_index = llh->llh_hdr.lrh_index;
 386                 rc = 0;
 387         }
 388         RETURN(rc);
 389 }
 390 EXPORT_SYMBOL(llog_read_header);
 391
 392 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 393                      int flags, struct obd_uuid *uuid)
 394 {
 395         struct llog_log_hdr     *llh;
 396         enum llog_flag           fmt = flags & LLOG_F_EXT_MASK;
 397         int                      rc;
 398         int                     chunk_size = handle->lgh_ctxt->loc_chunk_size;
 399         ENTRY;
 400
 401         LASSERT(handle->lgh_hdr == NULL);
 402
 403         LASSERT(chunk_size >= LLOG_MIN_CHUNK_SIZE);
 404         OBD_ALLOC_LARGE(llh, chunk_size);
 405         if (llh == NULL)
 406                 RETURN(-ENOMEM);
 407
 408         handle->lgh_hdr = llh;
 409         handle->lgh_hdr_size = chunk_size;
 410         /* first assign flags to use llog_client_ops */
 411         llh->llh_flags = flags;
 412         rc = llog_read_header(env, handle, uuid);
 413         if (rc == 0) {
 414                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 415                               flags & LLOG_F_IS_CAT) ||
 416                              (llh->llh_flags & LLOG_F_IS_CAT &&
 417                               flags & LLOG_F_IS_PLAIN))) {
 418                         CERROR("%s: llog type is %s but initializing %s\n",
 419                                loghandle2name(handle),
 420                                llh->llh_flags & LLOG_F_IS_CAT ?
 421                                "catalog" : "plain",
 422                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 423                         GOTO(out, rc = -EINVAL);
 424                 } else if (llh->llh_flags &
 425                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 426                         /*
 427                          * it is possible to open llog without specifying llog
 428                          * type so it is taken from llh_flags
 429                          */
 430                         flags = llh->llh_flags;
 431                 } else {
 432                         /* for some reason the llh_flags has no type set */
 433                         CERROR("llog type is not specified!\n");
 434                         GOTO(out, rc = -EINVAL);
 435                 }
 436                 if (unlikely(uuid &&
 437                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 438                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 439                                loghandle2name(handle),
 440                                (char *)uuid->uuid,
 441                                (char *)llh->llh_tgtuuid.uuid);
 442                         GOTO(out, rc = -EEXIST);
 443                 }
 444         }
 445         if (flags & LLOG_F_IS_CAT) {
 446                 LASSERT(list_empty(&handle->u.chd.chd_head));
 447                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 448                 llh->llh_size = sizeof(struct llog_logid_rec);
 449                 llh->llh_flags |= LLOG_F_IS_FIXSIZE;
 450         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 451                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 452                        loghandle2name(handle), flags, LLOG_F_IS_CAT,
 453                        LLOG_F_IS_PLAIN);
 454                 rc = -EINVAL;
 455         }
 456         llh->llh_flags |= fmt;
 457 out:
 458         if (rc) {
 459                 OBD_FREE_LARGE(llh, chunk_size);
 460                 handle->lgh_hdr = NULL;
 461         }
 462         RETURN(rc);
 463 }
 464 EXPORT_SYMBOL(llog_init_handle);
 465
 466 int llog_verify_record(const struct llog_handle *llh, struct llog_rec_hdr *rec)
 467 {
 468         int chunk_size = llh->lgh_hdr->llh_hdr.lrh_len;
 469
 470         if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
 471                 CERROR("%s: record is too large: %d > %d\n",
 472                        loghandle2name(llh), rec->lrh_len, chunk_size);
 473                 return -EINVAL;
 474         }
 475         if (rec->lrh_index >= LLOG_HDR_BITMAP_SIZE(llh->lgh_hdr)) {
 476                 CERROR("%s: index is too high: %d\n",
 477                        loghandle2name(llh), rec->lrh_index);
 478                 return -EINVAL;
 479         }
 480         if ((rec->lrh_type & LLOG_OP_MASK) != LLOG_OP_MAGIC) {
 481                 CERROR("%s: magic %x is bad\n",
 482                        loghandle2name(llh), rec->lrh_type);
 483                 return -EINVAL;
 484         }
 485
 486         return 0;
 487 }
 488 EXPORT_SYMBOL(llog_verify_record);
 489
 490 static int llog_process_thread(void *arg)
 491 {
 492         struct llog_process_info        *lpi = arg;
 493         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 494         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 495         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 496         struct llog_thread_info         *lti;
 497         char                            *buf;
 498         size_t                           chunk_size;
 499         __u64                            cur_offset;
 500         int                              rc = 0, index = 1, last_index;
 501         int                              saved_index = 0;
 502         int                              last_called_index = 0;
 503         bool                             repeated = false;
 504         bool                            refresh_idx = false;
 505
 506         ENTRY;
 507
 508         if (llh == NULL)
 509                 RETURN(-EINVAL);
 510
 511         lti = lpi->lpi_env == NULL ? NULL : llog_info(lpi->lpi_env);
 512
 513         cur_offset = chunk_size = llh->llh_hdr.lrh_len;
 514         /* expect chunk_size to be power of two */
 515         LASSERT(is_power_of_2(chunk_size));
 516
 517         OBD_ALLOC_LARGE(buf, chunk_size);
 518         if (buf == NULL) {
 519                 lpi->lpi_rc = -ENOMEM;
 520                 RETURN(0);
 521         }
 522
 523         if (cd != NULL) {
 524                 last_called_index = cd->lpcd_first_idx;
 525                 index = cd->lpcd_first_idx + 1;
 526         }
 527         if (cd != NULL && cd->lpcd_last_idx)
 528                 last_index = cd->lpcd_last_idx;
 529         else
 530                 last_index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 531
 532         while (rc == 0) {
 533                 struct llog_rec_hdr *rec;
 534                 off_t chunk_offset = 0;
 535                 unsigned int buf_offset = 0;
 536                 int     lh_last_idx;
 537                 int     synced_idx = 0;
 538
 539                 /* skip records not set in bitmap */
 540                 while (index <= last_index &&
 541                        !test_bit_le(index, LLOG_HDR_BITMAP(llh)))
 542                         ++index;
 543
 544                 /* There are no indices prior the last_index */
 545                 if (index > last_index)
 546                         break;
 547
 548                 CDEBUG(D_OTHER, "index: %d last_index %d\n", index,
 549                        last_index);
 550
 551 repeat:
 552                 /* get the buf with our target record; avoid old garbage */
 553                 memset(buf, 0, chunk_size);
 554                 /* the record index for outdated chunk data */
 555                 /* it is safe to process buffer until saved lgh_last_idx */
 556                 lh_last_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 557                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 558                                      index, &cur_offset, buf, chunk_size);
 559                 if (repeated && rc)
 560                         CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
 561                                " buf_offset %u, rc = %d\n", cur_offset,
 562                                (__u64)chunk_offset, buf_offset, rc);
 563                 /* we`ve tried to reread the chunk, but there is no
 564                  * new records */
 565                 if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
 566                     cur_offset)
 567                         GOTO(out, rc = 0);
 568                 if (rc != 0)
 569                         GOTO(out, rc);
 570
 571                 /* NB: after llog_next_block() call the cur_offset is the
 572                  * offset of the next block after read one.
 573                  * The absolute offset of the current chunk is calculated
 574                  * from cur_offset value and stored in chunk_offset variable.
 575                  */
 576                 if ((cur_offset & (chunk_size - 1)) != 0)
 577                         chunk_offset = cur_offset & ~(chunk_size - 1);
 578                 else
 579                         chunk_offset = cur_offset - chunk_size;
 580
 581                 /* NB: when rec->lrh_len is accessed it is already swabbed
 582                  * since it is used at the "end" of the loop and the rec
 583                  * swabbing is done at the beginning of the loop. */
 584                 for (rec = (struct llog_rec_hdr *)(buf + buf_offset);
 585                      (char *)rec < buf + chunk_size;
 586                      rec = llog_rec_hdr_next(rec)) {
 587
 588                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 589                                rec, rec->lrh_type);
 590
 591                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 592                                 lustre_swab_llog_rec(rec);
 593
 594                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 595                                rec->lrh_type, rec->lrh_index);
 596
 597                         if (index == (synced_idx + 1) &&
 598                             synced_idx == LLOG_HDR_TAIL(llh)->lrt_index)
 599                                 GOTO(out, rc = 0);
 600
 601                         if (OBD_FAIL_PRECHECK(OBD_FAIL_LLOG_PROCESS_TIMEOUT) &&
 602                                 cfs_fail_val == (unsigned int)
 603                                         (loghandle->lgh_id.lgl_oi.oi.oi_id &
 604                                          0xFFFFFFFF)) {
 605                                 OBD_RACE(OBD_FAIL_LLOG_PROCESS_TIMEOUT);
 606                         }
 607
 608                         /* the bitmap could be changed during processing
 609                          * records from the chunk. For wrapped catalog
 610                          * it means we can read deleted record and try to
 611                          * process it. Check this case and reread the chunk.
 612                          * It is safe to process to lh_last_idx, including
 613                          * lh_last_idx if it was synced. We can not do <=
 614                          * comparison, cause for wrapped catalog lgh_last_idx
 615                          * could be less than index. So we detect last index
 616                          * for processing as index == lh_last_idx+1. But when
 617                          * catalog is wrapped and full lgh_last_idx=llh_cat_idx,
 618                          * the first processing index is llh_cat_idx+1.The
 619                          * exception is !(lgh_last_idx == llh_cat_idx &&
 620                          * index == llh_cat_idx + 1), and after simplification
 621                          * it turns to
 622                          * lh_last_idx != LLOG_HDR_TAIL(llh)->lrt_index
 623                          * This exception is working for catalog only.
 624                          */
 625
 626                         if ((index == lh_last_idx && synced_idx != index) ||
 627                             (index == (lh_last_idx + 1) &&
 628                              lh_last_idx != LLOG_HDR_TAIL(llh)->lrt_index) ||
 629                             (rec->lrh_index == 0 && !repeated)) {
 630
 631                                 /* save offset inside buffer for the re-read */
 632                                 buf_offset = (char *)rec - (char *)buf;
 633                                 cur_offset = chunk_offset;
 634                                 repeated = true;
 635                                 /* We need to be sure lgh_last_idx
 636                                  * record was saved to disk
 637                                  */
 638                                 down_read(&loghandle->lgh_last_sem);
 639                                 synced_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 640                                 up_read(&loghandle->lgh_last_sem);
 641                                 CDEBUG(D_OTHER, "synced_idx: %d\n", synced_idx);
 642                                 goto repeat;
 643
 644                         }
 645
 646                         repeated = false;
 647
 648                         rc = llog_verify_record(loghandle, rec);
 649                         if (rc) {
 650                                 CERROR("%s: invalid record in llog "DFID
 651                                        " record for index %d/%d: rc = %d\n",
 652                                        loghandle2name(loghandle),
 653                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 654                                        rec->lrh_index, index, rc);
 655                                 /*
 656                                  * the block seem to be corrupted, let's try
 657                                  * with the next one. reset rc to go to the
 658                                  * next chunk.
 659                                  */
 660                                 refresh_idx = true;
 661                                 index = 0;
 662                                 GOTO(repeat, rc = 0);
 663                         }
 664
 665                         if (rec->lrh_index < index) {
 666                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 667                                        rec->lrh_index);
 668                                 continue;
 669                         }
 670
 671                         if (rec->lrh_index != index) {
 672                                 /*
 673                                  * the last time we couldn't parse the block due
 674                                  * to corruption, thus has no idea about the
 675                                  * next index, take it from the block, once.
 676                                  */
 677                                 if (refresh_idx) {
 678                                         refresh_idx = false;
 679                                         index = rec->lrh_index;
 680                                 } else {
 681                                         CERROR("%s: "DFID" Invalid record: index"
 682                                                " %u but expected %u\n",
 683                                                loghandle2name(loghandle),
 684                                                PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 685                                                rec->lrh_index, index);
 686                                         GOTO(out, rc = -ERANGE);
 687                                 }
 688                         }
 689
 690                         CDEBUG(D_OTHER,
 691                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 692                                rec->lrh_index, rec->lrh_len,
 693                                (int)(buf + chunk_size - (char *)rec));
 694
 695                         /* lgh_cur_offset is used only at llog_test_3 */
 696                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 697                                                     chunk_offset;
 698
 699                         /* if set, process the callback on this record */
 700                         if (test_bit_le(index, LLOG_HDR_BITMAP(llh))) {
 701                                 struct llog_cookie *lgc;
 702                                 __u64   tmp_off;
 703                                 int     tmp_idx;
 704
 705                                 CDEBUG((llh->llh_flags & LLOG_F_IS_CAT ?
 706                                         D_HA : D_OTHER),
 707                                        "index: %d, lh_last_idx: %d "
 708                                        "synced_idx: %d lgh_last_idx: %d\n",
 709                                        index, lh_last_idx, synced_idx,
 710                                        loghandle->lgh_last_idx);
 711
 712                                 if (lti != NULL) {
 713                                         lgc = &lti->lgi_cookie;
 714                                         /* store lu_env for recursive calls */
 715                                         tmp_off = lgc->lgc_offset;
 716                                         tmp_idx = lgc->lgc_index;
 717
 718                                         lgc->lgc_offset = (char *)rec -
 719                                                 (char *)buf + chunk_offset;
 720                                         lgc->lgc_index = rec->lrh_index;
 721                                 }
 722                                 /* using lu_env for passing record offset to
 723                                  * llog_write through various callbacks */
 724                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 725                                                  lpi->lpi_cbdata);
 726                                 last_called_index = index;
 727
 728                                 if (lti != NULL) {
 729                                         lgc->lgc_offset = tmp_off;
 730                                         lgc->lgc_index = tmp_idx;
 731                                 }
 732
 733                                 if (rc == LLOG_PROC_BREAK) {
 734                                         GOTO(out, rc);
 735                                 } else if (rc == LLOG_DEL_RECORD) {
 736                                         rc = llog_cancel_rec(lpi->lpi_env,
 737                                                              loghandle,
 738                                                              rec->lrh_index);
 739                                         /* Allow parallel cancelling, ENOENT
 740                                          * means record was canceled at another
 741                                          * processing thread or callback
 742                                          */
 743                                         if (rc == -ENOENT)
 744                                                 rc = 0;
 745                                 }
 746                                 if (rc)
 747                                         GOTO(out, rc);
 748                                 /* some stupid callbacks directly cancel records
 749                                  * and delete llog. Check it and stop
 750                                  * processing. */
 751                                 if (loghandle->lgh_hdr == NULL ||
 752                                     loghandle->lgh_hdr->llh_count == 1)
 753                                         GOTO(out, rc = 0);
 754                         }
 755                         /* exit if the last index is reached */
 756                         if (index >= last_index)
 757                                 GOTO(out, rc = 0);
 758                         ++index;
 759                 }
 760         }
 761
 762 out:
 763         CDEBUG(D_HA, "stop processing %s "DOSTID":%x index %d count %d\n",
 764                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" : "plain"),
 765                POSTID(&loghandle->lgh_id.lgl_oi), loghandle->lgh_id.lgl_ogen,
 766                index, llh->llh_count);
 767
 768         if (cd != NULL)
 769                 cd->lpcd_last_idx = last_called_index;
 770
 771         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 772                 if (dt_object_remote(loghandle->lgh_obj)) {
 773                         /* If it is remote object, then -EIO might means
 774                          * disconnection or eviction, let's return -EAGAIN,
 775                          * so for update recovery log processing, it will
 776                          * retry until the umount or abort recovery, see
 777                          * lod_sub_recovery_thread() */
 778                         CERROR("%s retry remote llog process\n",
 779                                loghandle2name(loghandle));
 780                         rc = -EAGAIN;
 781                 } else {
 782                         /* something bad happened to the processing of a local
 783                          * llog file, probably I/O error or the log got
 784                          * corrupted to be able to finally release the log we
 785                          * discard any remaining bits in the header */
 786                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 787                                " %s index %d count %d\n",
 788                                loghandle2name(loghandle),
 789                                POSTID(&loghandle->lgh_id.lgl_oi),
 790                                loghandle->lgh_id.lgl_ogen,
 791                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 792                                 "plain"), index, llh->llh_count);
 793
 794                         while (index <= last_index) {
 795                                 if (test_bit_le(index,
 796                                                   LLOG_HDR_BITMAP(llh)) != 0)
 797                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 798                                                         index);
 799                                 index++;
 800                         }
 801                         rc = 0;
 802                 }
 803         }
 804
 805         OBD_FREE_LARGE(buf, chunk_size);
 806         lpi->lpi_rc = rc;
 807         return 0;
 808 }
 809
 810 static int llog_process_thread_daemonize(void *arg)
 811 {
 812         struct llog_process_info        *lpi = arg;
 813         struct lu_env                    env;
 814         int                              rc;
 815         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 816
 817         task_lock(lpi->lpi_reftask);
 818         new_ns = lpi->lpi_reftask->nsproxy;
 819         if (curr_ns != new_ns) {
 820                 get_nsproxy(new_ns);
 821
 822                 current->nsproxy = new_ns;
 823                 /* XXX: we should call put_nsproxy() instead of
 824                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 825                  * used outside of the kernel itself, because it calls
 826                  * free_nsproxy() which is not exported by the kernel
 827                  * (defined in kernel/nsproxy.c) */
 828                 if (curr_ns)
 829                         atomic_dec(&curr_ns->count);
 830         }
 831         task_unlock(lpi->lpi_reftask);
 832
 833         /* client env has no keys, tags is just 0 */
 834         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 835         if (rc)
 836                 goto out;
 837         lpi->lpi_env = &env;
 838
 839         rc = llog_process_thread(arg);
 840
 841         lu_env_fini(&env);
 842 out:
 843         complete(&lpi->lpi_completion);
 844         return rc;
 845 }
 846
 847 int llog_process_or_fork(const struct lu_env *env,
 848                          struct llog_handle *loghandle,
 849                          llog_cb_t cb, void *data, void *catdata, bool fork)
 850 {
 851         struct llog_process_info *lpi;
 852         struct llog_process_data *d = data;
 853         struct llog_process_cat_data *cd = catdata;
 854         __u32 flags = loghandle->lgh_hdr->llh_flags;
 855         int rc;
 856
 857         ENTRY;
 858
 859         OBD_ALLOC_PTR(lpi);
 860         if (lpi == NULL) {
 861                 CERROR("cannot alloc pointer\n");
 862                 RETURN(-ENOMEM);
 863         }
 864         lpi->lpi_loghandle = loghandle;
 865         lpi->lpi_cb        = cb;
 866         lpi->lpi_cbdata    = data;
 867         lpi->lpi_catdata   = catdata;
 868
 869         CDEBUG(D_OTHER, "Processing "DFID" flags 0x%03x startcat %d startidx %d first_idx %d last_idx %d\n",
 870                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), flags,
 871                (flags & LLOG_F_IS_CAT) && d ? d->lpd_startcat : -1,
 872                (flags & LLOG_F_IS_CAT) && d ? d->lpd_startidx : -1,
 873                cd ? cd->lpcd_first_idx : -1, cd ? cd->lpcd_last_idx : -1);
 874         if (fork) {
 875                 struct task_struct *task;
 876
 877                 /* The new thread can't use parent env,
 878                  * init the new one in llog_process_thread_daemonize. */
 879                 lpi->lpi_env = NULL;
 880                 init_completion(&lpi->lpi_completion);
 881                 /* take reference to current, so that
 882                  * llog_process_thread_daemonize() can use it to switch to
 883                  * namespace associated with current  */
 884                 lpi->lpi_reftask = current;
 885                 task = kthread_run(llog_process_thread_daemonize, lpi,
 886                                    "llog_process_thread");
 887                 if (IS_ERR(task)) {
 888                         rc = PTR_ERR(task);
 889                         CERROR("%s: cannot start thread: rc = %d\n",
 890                                loghandle2name(loghandle), rc);
 891                         GOTO(out_lpi, rc);
 892                 }
 893                 wait_for_completion(&lpi->lpi_completion);
 894         } else {
 895                 lpi->lpi_env = env;
 896                 llog_process_thread(lpi);
 897         }
 898         rc = lpi->lpi_rc;
 899
 900 out_lpi:
 901         OBD_FREE_PTR(lpi);
 902         RETURN(rc);
 903 }
 904 EXPORT_SYMBOL(llog_process_or_fork);
 905
 906 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 907                  llog_cb_t cb, void *data, void *catdata)
 908 {
 909         int rc;
 910         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 911         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 912 }
 913 EXPORT_SYMBOL(llog_process);
 914
 915 static inline const struct cred *llog_raise_resource(void)
 916 {
 917         struct cred *cred = NULL;
 918
 919         if (cap_raised(current_cap(), CAP_SYS_RESOURCE))
 920                 return cred;
 921
 922         cred = prepare_creds();
 923         if (!cred)
 924                 return cred;
 925
 926         cap_raise(cred->cap_effective, CAP_SYS_RESOURCE);
 927         return override_creds(cred);
 928 }
 929
 930 static inline void llog_restore_resource(const struct cred *old_cred)
 931 {
 932         if (old_cred)
 933                 revert_creds(old_cred);
 934 }
 935
 936 int llog_reverse_process(const struct lu_env *env,
 937                          struct llog_handle *loghandle, llog_cb_t cb,
 938                          void *data, void *catdata)
 939 {
 940         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 941         struct llog_process_cat_data *cd = catdata;
 942         void *buf;
 943         int rc = 0, first_index = 1, index, idx;
 944         __u32   chunk_size = llh->llh_hdr.lrh_len;
 945         ENTRY;
 946
 947         OBD_ALLOC_LARGE(buf, chunk_size);
 948         if (buf == NULL)
 949                 RETURN(-ENOMEM);
 950
 951         if (cd != NULL)
 952                 first_index = cd->lpcd_first_idx + 1;
 953         if (cd != NULL && cd->lpcd_last_idx)
 954                 index = cd->lpcd_last_idx;
 955         else
 956                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 957
 958         while (rc == 0) {
 959                 struct llog_rec_hdr *rec;
 960                 struct llog_rec_tail *tail;
 961
 962                 /* skip records not set in bitmap */
 963                 while (index >= first_index &&
 964                        !test_bit_le(index, LLOG_HDR_BITMAP(llh)))
 965                         --index;
 966
 967                 LASSERT(index >= first_index - 1);
 968                 if (index == first_index - 1)
 969                         break;
 970
 971                 /* get the buf with our target record; avoid old garbage */
 972                 memset(buf, 0, chunk_size);
 973                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 974                 if (rc)
 975                         GOTO(out, rc);
 976
 977                 rec = buf;
 978                 idx = rec->lrh_index;
 979                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 980                 while (idx < index) {
 981                         rec = (void *)rec + rec->lrh_len;
 982                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 983                                 lustre_swab_llog_rec(rec);
 984                         idx ++;
 985                 }
 986                 LASSERT(idx == index);
 987                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 988
 989                 /* process records in buffer, starting where we found one */
 990                 while ((void *)tail > buf) {
 991                         if (tail->lrt_index == 0)
 992                                 GOTO(out, rc = 0); /* no more records */
 993
 994                         /* if set, process the callback on this record */
 995                         if (test_bit_le(index, LLOG_HDR_BITMAP(llh))) {
 996                                 rec = (void *)tail - tail->lrt_len +
 997                                       sizeof(*tail);
 998
 999                                 rc = cb(env, loghandle, rec, data);
1000                                 if (rc == LLOG_PROC_BREAK) {
1001                                         GOTO(out, rc);
1002                                 } else if (rc == LLOG_DEL_RECORD) {
1003                                         rc = llog_cancel_rec(env, loghandle,
1004                                                              tail->lrt_index);
1005                                 }
1006                                 if (rc)
1007                                         GOTO(out, rc);
1008                         }
1009
1010                         /* previous record, still in buffer? */
1011                         --index;
1012                         if (index < first_index)
1013                                 GOTO(out, rc = 0);
1014                         tail = (void *)tail - tail->lrt_len;
1015                 }
1016         }
1017
1018 out:
1019         if (buf != NULL)
1020                 OBD_FREE_LARGE(buf, chunk_size);
1021         RETURN(rc);
1022 }
1023 EXPORT_SYMBOL(llog_reverse_process);
1024
1025 /**
1026  * new llog API
1027  *
1028  * API functions:
1029  *      llog_open - open llog, may not exist
1030  *      llog_exist - check if llog exists
1031  *      llog_close - close opened llog, pair for open, frees llog_handle
1032  *      llog_declare_create - declare llog creation
1033  *      llog_create - create new llog on disk, need transaction handle
1034  *      llog_declare_write_rec - declaration of llog write
1035  *      llog_write_rec - write llog record on disk, need transaction handle
1036  *      llog_declare_add - declare llog catalog record addition
1037  *      llog_add - add llog record in catalog, need transaction handle
1038  */
1039 int llog_exist(struct llog_handle *loghandle)
1040 {
1041         const struct llog_operations *lop;
1042         int rc;
1043
1044         ENTRY;
1045
1046         rc = llog_handle2ops(loghandle, &lop);
1047         if (rc)
1048                 RETURN(rc);
1049         if (lop->lop_exist == NULL)
1050                 RETURN(-EOPNOTSUPP);
1051
1052         rc = lop->lop_exist(loghandle);
1053         RETURN(rc);
1054 }
1055 EXPORT_SYMBOL(llog_exist);
1056
1057 int llog_declare_create(const struct lu_env *env,
1058                         struct llog_handle *loghandle, struct thandle *th)
1059 {
1060         const struct cred *old_cred;
1061         const struct llog_operations *lop;
1062         int rc;
1063
1064         ENTRY;
1065
1066         rc = llog_handle2ops(loghandle, &lop);
1067         if (rc)
1068                 RETURN(rc);
1069         if (lop->lop_declare_create == NULL)
1070                 RETURN(-EOPNOTSUPP);
1071
1072         old_cred = llog_raise_resource();
1073         rc = lop->lop_declare_create(env, loghandle, th);
1074         llog_restore_resource(old_cred);
1075         RETURN(rc);
1076 }
1077
1078 int llog_create(const struct lu_env *env, struct llog_handle *handle,
1079                 struct thandle *th)
1080 {
1081         const struct cred *old_cred;
1082         const struct llog_operations *lop;
1083         int rc;
1084
1085         ENTRY;
1086
1087         rc = llog_handle2ops(handle, &lop);
1088         if (rc)
1089                 RETURN(rc);
1090         if (lop->lop_create == NULL)
1091                 RETURN(-EOPNOTSUPP);
1092
1093         old_cred = llog_raise_resource();
1094         rc = lop->lop_create(env, handle, th);
1095         llog_restore_resource(old_cred);
1096         RETURN(rc);
1097 }
1098
1099 int llog_declare_write_rec(const struct lu_env *env,
1100                            struct llog_handle *handle,
1101                            struct llog_rec_hdr *rec, int idx,
1102                            struct thandle *th)
1103 {
1104         const struct cred *old_cred;
1105         const struct llog_operations *lop;
1106         int rc;
1107
1108         ENTRY;
1109
1110         rc = llog_handle2ops(handle, &lop);
1111         if (rc)
1112                 RETURN(rc);
1113         LASSERT(lop);
1114         if (lop->lop_declare_write_rec == NULL)
1115                 RETURN(-EOPNOTSUPP);
1116
1117         old_cred = llog_raise_resource();
1118         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
1119         llog_restore_resource(old_cred);
1120         RETURN(rc);
1121 }
1122
1123 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
1124                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1125                    int idx, struct thandle *th)
1126 {
1127         const struct cred *old_cred;
1128         const struct llog_operations *lop;
1129         int rc, buflen;
1130
1131         ENTRY;
1132
1133         /* API sanity checks */
1134         if (handle == NULL) {
1135                 CERROR("loghandle is missed\n");
1136                 RETURN(-EPROTO);
1137         } else if (handle->lgh_obj == NULL) {
1138                 CERROR("loghandle %p with NULL object\n",
1139                         handle);
1140                 RETURN(-EPROTO);
1141         } else if (th == NULL) {
1142                 CERROR("%s: missed transaction handle\n",
1143                        loghandle2name(handle));
1144                 RETURN(-EPROTO);
1145         } else if (handle->lgh_hdr == NULL) {
1146                 CERROR("%s: loghandle %p with no header\n",
1147                        loghandle2name(handle), handle);
1148                 RETURN(-EPROTO);
1149         }
1150
1151         rc = llog_handle2ops(handle, &lop);
1152         if (rc)
1153                 RETURN(rc);
1154
1155         if (lop->lop_write_rec == NULL)
1156                 RETURN(-EOPNOTSUPP);
1157
1158         buflen = rec->lrh_len;
1159         LASSERT(cfs_size_round(buflen) == buflen);
1160
1161         old_cred = llog_raise_resource();
1162         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1163         llog_restore_resource(old_cred);
1164         RETURN(rc);
1165 }
1166
1167 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1168              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1169              struct thandle *th)
1170 {
1171         const struct cred *old_cred;
1172         int rc;
1173
1174         ENTRY;
1175
1176         if (lgh->lgh_logops->lop_add == NULL)
1177                 RETURN(-EOPNOTSUPP);
1178
1179         old_cred = llog_raise_resource();
1180         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1181         llog_restore_resource(old_cred);
1182         RETURN(rc);
1183 }
1184 EXPORT_SYMBOL(llog_add);
1185
1186 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1187                      struct llog_rec_hdr *rec, struct thandle *th)
1188 {
1189         const struct cred *old_cred;
1190         int rc;
1191
1192         ENTRY;
1193
1194         if (lgh->lgh_logops->lop_declare_add == NULL)
1195                 RETURN(-EOPNOTSUPP);
1196
1197         old_cred = llog_raise_resource();
1198         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1199         llog_restore_resource(old_cred);
1200         RETURN(rc);
1201 }
1202 EXPORT_SYMBOL(llog_declare_add);
1203
1204 /**
1205  * Helper function to open llog or create it if doesn't exist.
1206  * It hides all transaction handling from caller.
1207  */
1208 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1209                      struct llog_handle **res, struct llog_logid *logid,
1210                      char *name)
1211 {
1212         struct dt_device        *d;
1213         struct thandle          *th;
1214         int                      rc;
1215
1216         ENTRY;
1217
1218         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1219         if (rc)
1220                 RETURN(rc);
1221
1222         if (llog_exist(*res))
1223                 RETURN(0);
1224
1225         LASSERT((*res)->lgh_obj != NULL);
1226
1227         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1228
1229         if (unlikely(unlikely(d->dd_rdonly)))
1230                 RETURN(-EROFS);
1231
1232         th = dt_trans_create(env, d);
1233         if (IS_ERR(th))
1234                 GOTO(out, rc = PTR_ERR(th));
1235
1236         /* Create update llog object synchronously, which
1237          * happens during inialization process see
1238          * lod_sub_prep_llog(), to make sure the update
1239          * llog object is created before corss-MDT writing
1240          * updates into the llog object */
1241         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1242                 th->th_sync = 1;
1243
1244         th->th_wait_submit = 1;
1245         rc = llog_declare_create(env, *res, th);
1246         if (rc == 0) {
1247                 rc = dt_trans_start_local(env, d, th);
1248                 if (rc == 0)
1249                         rc = llog_create(env, *res, th);
1250         }
1251         dt_trans_stop(env, d, th);
1252 out:
1253         if (rc)
1254                 llog_close(env, *res);
1255         RETURN(rc);
1256 }
1257 EXPORT_SYMBOL(llog_open_create);
1258
1259 /**
1260  * Helper function to delete existent llog.
1261  */
1262 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1263                struct llog_logid *logid, char *name)
1264 {
1265         struct llog_handle      *handle;
1266         int                      rc = 0, rc2;
1267
1268         ENTRY;
1269
1270         /* nothing to erase */
1271         if (name == NULL && logid == NULL)
1272                 RETURN(0);
1273
1274         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1275         if (rc < 0)
1276                 RETURN(rc);
1277
1278         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1279         if (rc == 0)
1280                 rc = llog_destroy(env, handle);
1281
1282         rc2 = llog_close(env, handle);
1283         if (rc == 0)
1284                 rc = rc2;
1285         RETURN(rc);
1286 }
1287 EXPORT_SYMBOL(llog_erase);
1288
1289 /*
1290  * Helper function for write record in llog.
1291  * It hides all transaction handling from caller.
1292  * Valid only with local llog.
1293  */
1294 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1295                struct llog_rec_hdr *rec, int idx)
1296 {
1297         struct dt_device        *dt;
1298         struct thandle          *th;
1299         bool                    need_cookie;
1300         int                     rc;
1301
1302         ENTRY;
1303
1304         LASSERT(loghandle);
1305         LASSERT(loghandle->lgh_ctxt);
1306         LASSERT(loghandle->lgh_obj != NULL);
1307
1308         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1309
1310         if (unlikely(unlikely(dt->dd_rdonly)))
1311                 RETURN(-EROFS);
1312
1313         th = dt_trans_create(env, dt);
1314         if (IS_ERR(th))
1315                 RETURN(PTR_ERR(th));
1316
1317         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1318         if (rc)
1319                 GOTO(out_trans, rc);
1320
1321         th->th_wait_submit = 1;
1322         rc = dt_trans_start_local(env, dt, th);
1323         if (rc)
1324                 GOTO(out_trans, rc);
1325
1326         need_cookie = !(idx == LLOG_HEADER_IDX || idx == LLOG_NEXT_IDX);
1327
1328         down_write(&loghandle->lgh_lock);
1329         if (need_cookie) {
1330                 struct llog_thread_info *lti = llog_info(env);
1331
1332                 /* cookie comes from llog_process_thread */
1333                 rc = llog_write_rec(env, loghandle, rec, &lti->lgi_cookie,
1334                                     rec->lrh_index, th);
1335                 /* upper layer didn`t pass cookie so change rc */
1336                 rc = (rc == 1 ? 0 : rc);
1337         } else {
1338                 rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1339         }
1340
1341         up_write(&loghandle->lgh_lock);
1342 out_trans:
1343         dt_trans_stop(env, dt, th);
1344         RETURN(rc);
1345 }
1346 EXPORT_SYMBOL(llog_write);
1347
1348 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1349               struct llog_handle **lgh, struct llog_logid *logid,
1350               char *name, enum llog_open_param open_param)
1351 {
1352         const struct cred *old_cred;
1353         int      rc;
1354
1355         ENTRY;
1356
1357         LASSERT(ctxt);
1358         LASSERT(ctxt->loc_logops);
1359
1360         if (ctxt->loc_logops->lop_open == NULL) {
1361                 *lgh = NULL;
1362                 RETURN(-EOPNOTSUPP);
1363         }
1364
1365         *lgh = llog_alloc_handle();
1366         if (*lgh == NULL)
1367                 RETURN(-ENOMEM);
1368         (*lgh)->lgh_ctxt = ctxt;
1369         (*lgh)->lgh_logops = ctxt->loc_logops;
1370
1371         old_cred = llog_raise_resource();
1372         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1373         llog_restore_resource(old_cred);
1374         if (rc) {
1375                 llog_free_handle(*lgh);
1376                 *lgh = NULL;
1377         }
1378         RETURN(rc);
1379 }
1380 EXPORT_SYMBOL(llog_open);
1381
1382 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1383 {
1384         return llog_handle_put(env, loghandle);
1385 }
1386 EXPORT_SYMBOL(llog_close);
1387
1388 /**
1389  * Helper function to get the llog size in records. It is used by MGS
1390  * mostly to check that config llog exists and contains data.
1391  *
1392  * \param[in] env       execution environment
1393  * \param[in] ctxt      llog context
1394  * \param[in] name      llog name
1395  *
1396  * \retval              true if there are records in llog besides a header
1397  * \retval              false on error or llog without records
1398  */
1399 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1400                   char *name)
1401 {
1402         struct llog_handle      *llh;
1403         int                      rc = 0;
1404
1405         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1406         if (rc < 0) {
1407                 if (likely(rc == -ENOENT))
1408                         rc = 0;
1409                 GOTO(out, rc);
1410         }
1411
1412         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1413         if (rc)
1414                 GOTO(out_close, rc);
1415         rc = llog_get_size(llh);
1416
1417 out_close:
1418         llog_close(env, llh);
1419 out:
1420         /* The header is record 1, the llog is still considered as empty
1421          * if there is only header */
1422         return (rc <= 1);
1423 }
1424 EXPORT_SYMBOL(llog_is_empty);
1425
1426 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1427                       struct llog_rec_hdr *rec, void *data)
1428 {
1429         struct llog_handle      *copy_llh = data;
1430
1431         /* Append all records */
1432         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1433 }
1434
1435 /* backup plain llog */
1436 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1437                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1438                 char *name, char *backup)
1439 {
1440         struct llog_handle      *llh, *bllh;
1441         int                      rc;
1442
1443         ENTRY;
1444
1445         /* open original log */
1446         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1447         if (rc < 0) {
1448                 /* the -ENOENT case is also reported to the caller
1449                  * but silently so it should handle that if needed.
1450                  */
1451                 if (rc != -ENOENT)
1452                         CERROR("%s: failed to open log %s: rc = %d\n",
1453                                obd->obd_name, name, rc);
1454                 RETURN(rc);
1455         }
1456
1457         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1458         if (rc)
1459                 GOTO(out_close, rc);
1460
1461         /* Make sure there's no old backup log */
1462         rc = llog_erase(env, bctxt, NULL, backup);
1463         if (rc < 0 && rc != -ENOENT)
1464                 GOTO(out_close, rc);
1465
1466         /* open backup log */
1467         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1468         if (rc) {
1469                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1470                        obd->obd_name, backup, rc);
1471                 GOTO(out_close, rc);
1472         }
1473
1474         /* check that backup llog is not the same object as original one */
1475         if (llh->lgh_obj == bllh->lgh_obj) {
1476                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1477                        obd->obd_name, name, backup, llh->lgh_obj,
1478                        bllh->lgh_obj);
1479                 GOTO(out_backup, rc = -EEXIST);
1480         }
1481
1482         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1483         if (rc)
1484                 GOTO(out_backup, rc);
1485
1486         /* Copy log record by record */
1487         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1488                                   NULL, false);
1489         if (rc)
1490                 CERROR("%s: failed to backup log %s: rc = %d\n",
1491                        obd->obd_name, name, rc);
1492 out_backup:
1493         llog_close(env, bllh);
1494 out_close:
1495         llog_close(env, llh);
1496         RETURN(rc);
1497 }
1498 EXPORT_SYMBOL(llog_backup);
1499
1500 /* Get size of llog */
1501 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1502 {
1503         int rc;
1504         struct lu_attr la;
1505
1506         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1507         if (rc) {
1508                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1509                        loghandle2name(llh), PFID(&llh->lgh_id.lgl_oi.oi_fid),
1510                        rc);
1511                 return 0;
1512         }
1513
1514         return la.la_size;
1515 }
1516 EXPORT_SYMBOL(llog_size);
1517