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/fs_struct.h>
  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 /*
  55  * Allocate a new log or catalog handle
  56  * Used inside llog_open().
  57  */
  58 static struct llog_handle *llog_alloc_handle(void)
  59 {
  60         struct llog_handle *loghandle;
  61
  62         OBD_ALLOC_PTR(loghandle);
  63         if (loghandle == NULL)
  64                 return NULL;
  65
  66         init_rwsem(&loghandle->lgh_lock);
  67         mutex_init(&loghandle->lgh_hdr_mutex);
  68         init_rwsem(&loghandle->lgh_last_sem);
  69         INIT_LIST_HEAD(&loghandle->u.phd.phd_entry);
  70         refcount_set(&loghandle->lgh_refcount, 1);
  71
  72         return loghandle;
  73 }
  74
  75 /*
  76  * Free llog handle and header data if exists. Used in llog_close() only
  77  */
  78 static void llog_free_handle(struct llog_handle *loghandle)
  79 {
  80         LASSERT(loghandle != NULL);
  81
  82         /* failed llog_init_handle */
  83         if (loghandle->lgh_hdr == NULL)
  84                 goto out;
  85
  86         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  87                 LASSERT(list_empty(&loghandle->u.phd.phd_entry));
  88         else if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  89                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  90         OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
  91 out:
  92         OBD_FREE_PTR(loghandle);
  93 }
  94
  95 struct llog_handle *llog_handle_get(struct llog_handle *loghandle)
  96 {
  97         if (refcount_inc_not_zero(&loghandle->lgh_refcount))
  98                 return loghandle;
  99         return NULL;
 100 }
 101
 102 int llog_handle_put(const struct lu_env *env, struct llog_handle *loghandle)
 103 {
 104         int rc = 0;
 105
 106         if (refcount_dec_and_test(&loghandle->lgh_refcount)) {
 107                 const struct llog_operations *lop;
 108
 109                 rc = llog_handle2ops(loghandle, &lop);
 110                 if (!rc) {
 111                         if (lop->lop_close)
 112                                 rc = lop->lop_close(env, loghandle);
 113                         else
 114                                 rc = -EOPNOTSUPP;
 115                 }
 116                 llog_free_handle(loghandle);
 117         }
 118         return rc;
 119 }
 120
 121 static int llog_declare_destroy(const struct lu_env *env,
 122                                 struct llog_handle *handle,
 123                                 struct thandle *th)
 124 {
 125         const struct llog_operations *lop;
 126         int rc;
 127
 128         ENTRY;
 129
 130         rc = llog_handle2ops(handle, &lop);
 131         if (rc)
 132                 RETURN(rc);
 133         if (lop->lop_declare_destroy == NULL)
 134                 RETURN(-EOPNOTSUPP);
 135
 136         rc = lop->lop_declare_destroy(env, handle, th);
 137
 138         RETURN(rc);
 139 }
 140
 141 int llog_trans_destroy(const struct lu_env *env, struct llog_handle *handle,
 142                        struct thandle *th)
 143 {
 144         const struct llog_operations *lop;
 145         int rc;
 146         ENTRY;
 147
 148         rc = llog_handle2ops(handle, &lop);
 149         if (rc < 0)
 150                 RETURN(rc);
 151         if (lop->lop_destroy == NULL)
 152                 RETURN(-EOPNOTSUPP);
 153
 154         LASSERT(handle->lgh_obj != NULL);
 155         if (!llog_exist(handle))
 156                 RETURN(0);
 157
 158         rc = lop->lop_destroy(env, handle, th);
 159
 160         RETURN(rc);
 161 }
 162
 163 int llog_destroy(const struct lu_env *env, struct llog_handle *handle)
 164 {
 165         const struct llog_operations *lop;
 166         struct dt_device *dt;
 167         struct thandle *th;
 168         int rc;
 169
 170         ENTRY;
 171
 172         rc = llog_handle2ops(handle, &lop);
 173         if (rc < 0)
 174                 RETURN(rc);
 175         if (lop->lop_destroy == NULL)
 176                 RETURN(-EOPNOTSUPP);
 177
 178         if (handle->lgh_obj == NULL) {
 179                 /* if lgh_obj == NULL, then it is from client side destroy */
 180                 rc = lop->lop_destroy(env, handle, NULL);
 181                 RETURN(rc);
 182         }
 183
 184         if (!llog_exist(handle))
 185                 RETURN(0);
 186
 187         dt = lu2dt_dev(handle->lgh_obj->do_lu.lo_dev);
 188
 189         if (unlikely(unlikely(dt->dd_rdonly)))
 190                 RETURN(-EROFS);
 191
 192         th = dt_trans_create(env, dt);
 193         if (IS_ERR(th))
 194                 RETURN(PTR_ERR(th));
 195
 196         rc = llog_declare_destroy(env, handle, th);
 197         if (rc != 0)
 198                 GOTO(out_trans, rc);
 199
 200         rc = dt_trans_start_local(env, dt, th);
 201         if (rc < 0)
 202                 GOTO(out_trans, rc);
 203
 204         rc = lop->lop_destroy(env, handle, th);
 205
 206 out_trans:
 207         dt_trans_stop(env, dt, th);
 208
 209         RETURN(rc);
 210 }
 211 EXPORT_SYMBOL(llog_destroy);
 212
 213 /* returns negative on error; 0 if success; 1 if success & log destroyed */
 214 int llog_cancel_arr_rec(const struct lu_env *env, struct llog_handle *loghandle,
 215                         int num, int *index)
 216 {
 217         struct llog_thread_info *lgi = llog_info(env);
 218         struct dt_device        *dt;
 219         struct llog_log_hdr     *llh;
 220         struct thandle          *th;
 221         __u32                    tmp_lgc_index;
 222         int                      rc, i = 0;
 223         int rc1;
 224         bool subtract_count = false;
 225
 226         ENTRY;
 227
 228         LASSERT(loghandle != NULL);
 229         LASSERT(loghandle->lgh_ctxt != NULL);
 230         LASSERT(loghandle->lgh_obj != NULL);
 231
 232         llh = loghandle->lgh_hdr;
 233
 234         CDEBUG(D_RPCTRACE, "Canceling %d records, first %d in log "DFID"\n",
 235                num, index[0], PFID(&loghandle->lgh_id.lgl_oi.oi_fid));
 236
 237         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
 238
 239         if (unlikely(unlikely(dt->dd_rdonly)))
 240                 RETURN(0);
 241
 242         th = dt_trans_create(env, dt);
 243         if (IS_ERR(th))
 244                 RETURN(PTR_ERR(th));
 245
 246         rc = llog_declare_write_rec(env, loghandle, &llh->llh_hdr, 0, th);
 247         if (rc < 0)
 248                 GOTO(out_trans, rc);
 249
 250         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY)) {
 251                 rc = llog_declare_destroy(env, loghandle, th);
 252                 if (rc < 0)
 253                         GOTO(out_trans, rc);
 254         }
 255
 256         th->th_wait_submit = 1;
 257         rc = dt_trans_start_local(env, dt, th);
 258         if (rc < 0)
 259                 GOTO(out_trans, rc);
 260
 261         down_write(&loghandle->lgh_lock);
 262         /* clear bitmap */
 263         mutex_lock(&loghandle->lgh_hdr_mutex);
 264         for (i = 0; i < num; ++i) {
 265                 if (index[i] == 0) {
 266                         CERROR("Can't cancel index 0 which is header\n");
 267                         GOTO(out_unlock, rc = -EINVAL);
 268                 }
 269                 if (!__test_and_clear_bit_le(index[i], LLOG_HDR_BITMAP(llh))) {
 270                         CDEBUG(D_RPCTRACE, "Catalog index %u already clear?\n",
 271                                index[i]);
 272                         GOTO(out_unlock, rc = -ENOENT);
 273                 }
 274         }
 275         loghandle->lgh_hdr->llh_count -= num;
 276         subtract_count = true;
 277
 278         /* Since llog_process_thread use lgi_cookie, it`s better to save them
 279          * and restore after using
 280          */
 281         tmp_lgc_index = lgi->lgi_cookie.lgc_index;
 282         /* Pass this index to llog_osd_write_rec(), which will use the index
 283          * to only update the necesary bitmap. */
 284         lgi->lgi_cookie.lgc_index = index[0];
 285         /* update header */
 286         rc = llog_write_rec(env, loghandle, &llh->llh_hdr, (num != 1 ? NULL :
 287                             &lgi->lgi_cookie), LLOG_HEADER_IDX, th);
 288         lgi->lgi_cookie.lgc_index = tmp_lgc_index;
 289
 290         if (rc != 0)
 291                 GOTO(out_unlock, rc);
 292
 293         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 294             (llh->llh_count == 1) &&
 295             ((loghandle->lgh_last_idx == LLOG_HDR_BITMAP_SIZE(llh) - 1) ||
 296              (loghandle->u.phd.phd_cat_handle != NULL &&
 297               loghandle->u.phd.phd_cat_handle->u.chd.chd_current_log !=
 298                 loghandle))) {
 299                 /* never try to destroy it again */
 300                 llh->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
 301                 rc = llog_trans_destroy(env, loghandle, th);
 302                 if (rc < 0) {
 303                         /* Sigh, can not destroy the final plain llog, but
 304                          * the bitmap has been clearly, so the record can not
 305                          * be accessed anymore, let's return 0 for now, and
 306                          * the orphan will be handled by LFSCK. */
 307                         CERROR("%s: can't destroy empty llog "DFID": rc = %d\n",
 308                                loghandle2name(loghandle),
 309                                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rc);
 310                         GOTO(out_unlock, rc = 0);
 311                 }
 312                 rc = LLOG_DEL_PLAIN;
 313         }
 314
 315 out_unlock:
 316         if (rc < 0) {
 317                 /* restore bitmap while holding a mutex */
 318                 if (subtract_count) {
 319                         loghandle->lgh_hdr->llh_count += num;
 320                         subtract_count = false;
 321                 }
 322                 for (i = i - 1; i >= 0; i--)
 323                         set_bit_le(index[i], LLOG_HDR_BITMAP(llh));
 324         }
 325         mutex_unlock(&loghandle->lgh_hdr_mutex);
 326         up_write(&loghandle->lgh_lock);
 327 out_trans:
 328         rc1 = dt_trans_stop(env, dt, th);
 329         if (rc == 0)
 330                 rc = rc1;
 331         if (rc1 < 0) {
 332                 mutex_lock(&loghandle->lgh_hdr_mutex);
 333                 if (subtract_count)
 334                         loghandle->lgh_hdr->llh_count += num;
 335                 for (i = i - 1; i >= 0; i--)
 336                         set_bit_le(index[i], LLOG_HDR_BITMAP(llh));
 337                 mutex_unlock(&loghandle->lgh_hdr_mutex);
 338         }
 339         RETURN(rc);
 340 }
 341
 342 int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
 343                     int index)
 344 {
 345         return llog_cancel_arr_rec(env, loghandle, 1, &index);
 346 }
 347
 348 int llog_read_header(const struct lu_env *env, struct llog_handle *handle,
 349                      const struct obd_uuid *uuid)
 350 {
 351         const struct llog_operations *lop;
 352         int rc;
 353         ENTRY;
 354
 355         rc = llog_handle2ops(handle, &lop);
 356         if (rc)
 357                 RETURN(rc);
 358
 359         if (lop->lop_read_header == NULL)
 360                 RETURN(-EOPNOTSUPP);
 361
 362         rc = lop->lop_read_header(env, handle);
 363         if (rc == LLOG_EEMPTY) {
 364                 struct llog_log_hdr *llh = handle->lgh_hdr;
 365
 366                 /* lrh_len should be initialized in llog_init_handle */
 367                 handle->lgh_last_idx = 0; /* header is record with index 0 */
 368                 llh->llh_count = 1;         /* for the header record */
 369                 llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 370                 LASSERT(handle->lgh_ctxt->loc_chunk_size >=
 371                                                 LLOG_MIN_CHUNK_SIZE);
 372                 llh->llh_hdr.lrh_len = handle->lgh_ctxt->loc_chunk_size;
 373                 llh->llh_hdr.lrh_index = 0;
 374                 llh->llh_timestamp = ktime_get_real_seconds();
 375                 if (uuid)
 376                         memcpy(&llh->llh_tgtuuid, uuid,
 377                                sizeof(llh->llh_tgtuuid));
 378                 llh->llh_bitmap_offset = offsetof(typeof(*llh), llh_bitmap);
 379                 /* Since update llog header might also call this function,
 380                  * let's reset the bitmap to 0 here */
 381                 memset(LLOG_HDR_BITMAP(llh), 0, llh->llh_hdr.lrh_len -
 382                                                 llh->llh_bitmap_offset -
 383                                                 sizeof(llh->llh_tail));
 384                 set_bit_le(0, LLOG_HDR_BITMAP(llh));
 385                 LLOG_HDR_TAIL(llh)->lrt_len = llh->llh_hdr.lrh_len;
 386                 LLOG_HDR_TAIL(llh)->lrt_index = llh->llh_hdr.lrh_index;
 387                 rc = 0;
 388         }
 389         RETURN(rc);
 390 }
 391 EXPORT_SYMBOL(llog_read_header);
 392
 393 int llog_init_handle(const struct lu_env *env, struct llog_handle *handle,
 394                      int flags, struct obd_uuid *uuid)
 395 {
 396         struct llog_log_hdr     *llh;
 397         enum llog_flag           fmt = flags & LLOG_F_EXT_MASK;
 398         int                      rc;
 399         int                     chunk_size = handle->lgh_ctxt->loc_chunk_size;
 400         ENTRY;
 401
 402         LASSERT(handle->lgh_hdr == NULL);
 403
 404         LASSERT(chunk_size >= LLOG_MIN_CHUNK_SIZE);
 405         OBD_ALLOC_LARGE(llh, chunk_size);
 406         if (llh == NULL)
 407                 RETURN(-ENOMEM);
 408
 409         handle->lgh_hdr = llh;
 410         handle->lgh_hdr_size = chunk_size;
 411         /* first assign flags to use llog_client_ops */
 412         llh->llh_flags = flags;
 413         rc = llog_read_header(env, handle, uuid);
 414         if (rc == 0) {
 415                 if (unlikely((llh->llh_flags & LLOG_F_IS_PLAIN &&
 416                               flags & LLOG_F_IS_CAT) ||
 417                              (llh->llh_flags & LLOG_F_IS_CAT &&
 418                               flags & LLOG_F_IS_PLAIN))) {
 419                         CERROR("%s: llog type is %s but initializing %s\n",
 420                                loghandle2name(handle),
 421                                llh->llh_flags & LLOG_F_IS_CAT ?
 422                                "catalog" : "plain",
 423                                flags & LLOG_F_IS_CAT ? "catalog" : "plain");
 424                         GOTO(out, rc = -EINVAL);
 425                 } else if (llh->llh_flags &
 426                            (LLOG_F_IS_PLAIN | LLOG_F_IS_CAT)) {
 427                         /*
 428                          * it is possible to open llog without specifying llog
 429                          * type so it is taken from llh_flags
 430                          */
 431                         flags = llh->llh_flags;
 432                 } else {
 433                         /* for some reason the llh_flags has no type set */
 434                         CERROR("llog type is not specified!\n");
 435                         GOTO(out, rc = -EINVAL);
 436                 }
 437                 if (unlikely(uuid &&
 438                              !obd_uuid_equals(uuid, &llh->llh_tgtuuid))) {
 439                         CERROR("%s: llog uuid mismatch: %s/%s\n",
 440                                loghandle2name(handle),
 441                                (char *)uuid->uuid,
 442                                (char *)llh->llh_tgtuuid.uuid);
 443                         GOTO(out, rc = -EEXIST);
 444                 }
 445         }
 446         if (flags & LLOG_F_IS_CAT) {
 447                 LASSERT(list_empty(&handle->u.chd.chd_head));
 448                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 449                 llh->llh_size = sizeof(struct llog_logid_rec);
 450                 llh->llh_flags |= LLOG_F_IS_FIXSIZE;
 451         } else if (!(flags & LLOG_F_IS_PLAIN)) {
 452                 CERROR("%s: unknown flags: %#x (expected %#x or %#x)\n",
 453                        loghandle2name(handle), flags, LLOG_F_IS_CAT,
 454                        LLOG_F_IS_PLAIN);
 455                 rc = -EINVAL;
 456         }
 457         llh->llh_flags |= fmt;
 458 out:
 459         if (rc) {
 460                 OBD_FREE_LARGE(llh, chunk_size);
 461                 handle->lgh_hdr = NULL;
 462         }
 463         RETURN(rc);
 464 }
 465 EXPORT_SYMBOL(llog_init_handle);
 466
 467 int llog_verify_record(const struct llog_handle *llh, struct llog_rec_hdr *rec)
 468 {
 469         int chunk_size = llh->lgh_hdr->llh_hdr.lrh_len;
 470
 471         if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
 472                 CERROR("%s: record is too large: %d > %d\n",
 473                        loghandle2name(llh), rec->lrh_len, chunk_size);
 474                 return -EINVAL;
 475         }
 476         if (rec->lrh_index >= LLOG_HDR_BITMAP_SIZE(llh->lgh_hdr)) {
 477                 CERROR("%s: index is too high: %d\n",
 478                        loghandle2name(llh), rec->lrh_index);
 479                 return -EINVAL;
 480         }
 481         if ((rec->lrh_type & LLOG_OP_MASK) != LLOG_OP_MAGIC) {
 482                 CERROR("%s: magic %x is bad\n",
 483                        loghandle2name(llh), rec->lrh_type);
 484                 return -EINVAL;
 485         }
 486
 487         return 0;
 488 }
 489 EXPORT_SYMBOL(llog_verify_record);
 490
 491 static int llog_process_thread(void *arg)
 492 {
 493         struct llog_process_info        *lpi = arg;
 494         struct llog_handle              *loghandle = lpi->lpi_loghandle;
 495         struct llog_log_hdr             *llh = loghandle->lgh_hdr;
 496         struct llog_process_cat_data    *cd  = lpi->lpi_catdata;
 497         struct llog_thread_info         *lti;
 498         char                            *buf;
 499         size_t                           chunk_size;
 500         __u64                            cur_offset;
 501         int                              rc = 0, index = 1, last_index;
 502         int                              saved_index = 0;
 503         int                              last_called_index = 0;
 504         bool                             repeated = false;
 505         bool                            refresh_idx = false;
 506
 507         ENTRY;
 508
 509         if (llh == NULL)
 510                 RETURN(-EINVAL);
 511
 512         lti = lpi->lpi_env == NULL ? NULL : llog_info(lpi->lpi_env);
 513
 514         cur_offset = chunk_size = llh->llh_hdr.lrh_len;
 515         /* expect chunk_size to be power of two */
 516         LASSERT(is_power_of_2(chunk_size));
 517
 518         OBD_ALLOC_LARGE(buf, chunk_size);
 519         if (buf == NULL) {
 520                 lpi->lpi_rc = -ENOMEM;
 521                 RETURN(0);
 522         }
 523
 524         if (cd != NULL) {
 525                 last_called_index = cd->lpcd_first_idx;
 526                 index = cd->lpcd_first_idx + 1;
 527         }
 528         if (cd != NULL && cd->lpcd_last_idx)
 529                 last_index = cd->lpcd_last_idx;
 530         else
 531                 last_index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 532
 533         while (rc == 0) {
 534                 struct llog_rec_hdr *rec;
 535                 off_t chunk_offset = 0;
 536                 unsigned int buf_offset = 0;
 537                 int     lh_last_idx;
 538                 int     synced_idx = 0;
 539
 540                 /* skip records not set in bitmap */
 541                 while (index <= last_index &&
 542                        !test_bit_le(index, LLOG_HDR_BITMAP(llh)))
 543                         ++index;
 544
 545                 /* There are no indices prior the last_index */
 546                 if (index > last_index)
 547                         break;
 548
 549                 CDEBUG(D_OTHER, "index: %d last_index %d\n", index,
 550                        last_index);
 551
 552 repeat:
 553                 /* get the buf with our target record; avoid old garbage */
 554                 memset(buf, 0, chunk_size);
 555                 /* the record index for outdated chunk data */
 556                 /* it is safe to process buffer until saved lgh_last_idx */
 557                 lh_last_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 558                 rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
 559                                      index, &cur_offset, buf, chunk_size);
 560                 if (repeated && rc)
 561                         CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
 562                                " buf_offset %u, rc = %d\n", cur_offset,
 563                                (__u64)chunk_offset, buf_offset, rc);
 564                 if (rc == -ESTALE)
 565                         GOTO(out, rc = 0);
 566                 /* we`ve tried to reread the chunk, but there is no
 567                  * new records */
 568                 if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
 569                     cur_offset)
 570                         GOTO(out, rc = 0);
 571                 if (rc != 0)
 572                         GOTO(out, rc);
 573
 574                 /* NB: after llog_next_block() call the cur_offset is the
 575                  * offset of the next block after read one.
 576                  * The absolute offset of the current chunk is calculated
 577                  * from cur_offset value and stored in chunk_offset variable.
 578                  */
 579                 if ((cur_offset & (chunk_size - 1)) != 0)
 580                         chunk_offset = cur_offset & ~(chunk_size - 1);
 581                 else
 582                         chunk_offset = cur_offset - chunk_size;
 583
 584                 /* NB: when rec->lrh_len is accessed it is already swabbed
 585                  * since it is used at the "end" of the loop and the rec
 586                  * swabbing is done at the beginning of the loop. */
 587                 for (rec = (struct llog_rec_hdr *)(buf + buf_offset);
 588                      (char *)rec < buf + chunk_size;
 589                      rec = llog_rec_hdr_next(rec)) {
 590
 591                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 592                                rec, rec->lrh_type);
 593
 594                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 595                                 lustre_swab_llog_rec(rec);
 596
 597                         CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
 598                                rec->lrh_type, rec->lrh_index);
 599
 600                         if (index == (synced_idx + 1) &&
 601                             synced_idx == LLOG_HDR_TAIL(llh)->lrt_index)
 602                                 GOTO(out, rc = 0);
 603
 604                         if (OBD_FAIL_PRECHECK(OBD_FAIL_LLOG_PROCESS_TIMEOUT) &&
 605                                 cfs_fail_val == (unsigned int)
 606                                         (loghandle->lgh_id.lgl_oi.oi.oi_id &
 607                                          0xFFFFFFFF)) {
 608                                 OBD_RACE(OBD_FAIL_LLOG_PROCESS_TIMEOUT);
 609                         }
 610
 611                         /* the bitmap could be changed during processing
 612                          * records from the chunk. For wrapped catalog
 613                          * it means we can read deleted record and try to
 614                          * process it. Check this case and reread the chunk.
 615                          * It is safe to process to lh_last_idx, including
 616                          * lh_last_idx if it was synced. We can not do <=
 617                          * comparison, cause for wrapped catalog lgh_last_idx
 618                          * could be less than index. So we detect last index
 619                          * for processing as index == lh_last_idx+1. But when
 620                          * catalog is wrapped and full lgh_last_idx=llh_cat_idx,
 621                          * the first processing index is llh_cat_idx+1.The
 622                          * exception is !(lgh_last_idx == llh_cat_idx &&
 623                          * index == llh_cat_idx + 1), and after simplification
 624                          * it turns to
 625                          * lh_last_idx != LLOG_HDR_TAIL(llh)->lrt_index
 626                          * This exception is working for catalog only.
 627                          */
 628
 629                         if ((index == lh_last_idx && synced_idx != index) ||
 630                             (index == (lh_last_idx + 1) &&
 631                              lh_last_idx != LLOG_HDR_TAIL(llh)->lrt_index) ||
 632                             (rec->lrh_index == 0 && !repeated)) {
 633
 634                                 /* save offset inside buffer for the re-read */
 635                                 buf_offset = (char *)rec - (char *)buf;
 636                                 cur_offset = chunk_offset;
 637                                 repeated = true;
 638                                 /* We need to be sure lgh_last_idx
 639                                  * record was saved to disk
 640                                  */
 641                                 down_read(&loghandle->lgh_last_sem);
 642                                 synced_idx = LLOG_HDR_TAIL(llh)->lrt_index;
 643                                 up_read(&loghandle->lgh_last_sem);
 644                                 CDEBUG(D_OTHER, "synced_idx: %d\n", synced_idx);
 645                                 goto repeat;
 646
 647                         }
 648
 649                         repeated = false;
 650
 651                         rc = llog_verify_record(loghandle, rec);
 652                         if (rc) {
 653                                 CERROR("%s: invalid record in llog "DFID
 654                                        " record for index %d/%d: rc = %d\n",
 655                                        loghandle2name(loghandle),
 656                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 657                                        rec->lrh_index, index, rc);
 658                                 /*
 659                                  * the block seem to be corrupted, let's try
 660                                  * with the next one. reset rc to go to the
 661                                  * next chunk.
 662                                  */
 663                                 refresh_idx = true;
 664                                 index = 0;
 665                                 GOTO(repeat, rc = 0);
 666                         }
 667
 668                         if (rec->lrh_index < index) {
 669                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 670                                        rec->lrh_index);
 671                                 continue;
 672                         }
 673
 674                         if (rec->lrh_index > index) {
 675                                 /* the record itself looks good, but we met a
 676                                  * gap which can be result of old bugs, just
 677                                  * keep going */
 678                                 CERROR("%s: "DFID" index %u, expected %u\n",
 679                                        loghandle2name(loghandle),
 680                                        PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
 681                                        rec->lrh_index, index);
 682                                 index = rec->lrh_index;
 683                         }
 684
 685                         CDEBUG(D_OTHER,
 686                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 687                                rec->lrh_index, rec->lrh_len,
 688                                (int)(buf + chunk_size - (char *)rec));
 689
 690                         /* lgh_cur_offset is used only at llog_test_3 */
 691                         loghandle->lgh_cur_offset = (char *)rec - (char *)buf +
 692                                                     chunk_offset;
 693
 694                         /* if set, process the callback on this record */
 695                         if (test_bit_le(index, LLOG_HDR_BITMAP(llh))) {
 696                                 struct llog_cookie *lgc;
 697                                 __u64   tmp_off;
 698                                 int     tmp_idx;
 699
 700                                 CDEBUG((llh->llh_flags & LLOG_F_IS_CAT ?
 701                                         D_HA : D_OTHER),
 702                                        "index: %d, lh_last_idx: %d "
 703                                        "synced_idx: %d lgh_last_idx: %d\n",
 704                                        index, lh_last_idx, synced_idx,
 705                                        loghandle->lgh_last_idx);
 706
 707                                 if (lti != NULL) {
 708                                         lgc = &lti->lgi_cookie;
 709                                         /* store lu_env for recursive calls */
 710                                         tmp_off = lgc->lgc_offset;
 711                                         tmp_idx = lgc->lgc_index;
 712
 713                                         lgc->lgc_offset = (char *)rec -
 714                                                 (char *)buf + chunk_offset;
 715                                         lgc->lgc_index = rec->lrh_index;
 716                                 }
 717                                 /* using lu_env for passing record offset to
 718                                  * llog_write through various callbacks */
 719                                 rc = lpi->lpi_cb(lpi->lpi_env, loghandle, rec,
 720                                                  lpi->lpi_cbdata);
 721                                 last_called_index = index;
 722
 723                                 if (lti != NULL) {
 724                                         lgc->lgc_offset = tmp_off;
 725                                         lgc->lgc_index = tmp_idx;
 726                                 }
 727
 728                                 if (rc == LLOG_PROC_BREAK) {
 729                                         GOTO(out, rc);
 730                                 } else if (rc == LLOG_DEL_RECORD) {
 731                                         rc = llog_cancel_rec(lpi->lpi_env,
 732                                                              loghandle,
 733                                                              rec->lrh_index);
 734                                         /* Allow parallel cancelling, ENOENT
 735                                          * means record was canceled at another
 736                                          * processing thread or callback
 737                                          */
 738                                         if (rc == -ENOENT)
 739                                                 rc = 0;
 740                                 }
 741                                 if (rc)
 742                                         GOTO(out, rc);
 743                                 /* some stupid callbacks directly cancel records
 744                                  * and delete llog. Check it and stop
 745                                  * processing. */
 746                                 if (loghandle->lgh_hdr == NULL ||
 747                                     loghandle->lgh_hdr->llh_count == 1)
 748                                         GOTO(out, rc = 0);
 749                         }
 750                         /* exit if the last index is reached */
 751                         if (index >= last_index)
 752                                 GOTO(out, rc = 0);
 753                         ++index;
 754                 }
 755         }
 756
 757 out:
 758         CDEBUG(D_HA, "stop processing %s "DOSTID":%x index %d count %d\n",
 759                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" : "plain"),
 760                POSTID(&loghandle->lgh_id.lgl_oi), loghandle->lgh_id.lgl_ogen,
 761                index, llh->llh_count);
 762
 763         if (cd != NULL)
 764                 cd->lpcd_last_idx = last_called_index;
 765
 766         if (unlikely(rc == -EIO && loghandle->lgh_obj != NULL)) {
 767                 if (dt_object_remote(loghandle->lgh_obj)) {
 768                         /* If it is remote object, then -EIO might means
 769                          * disconnection or eviction, let's return -EAGAIN,
 770                          * so for update recovery log processing, it will
 771                          * retry until the umount or abort recovery, see
 772                          * lod_sub_recovery_thread() */
 773                         CERROR("%s retry remote llog process\n",
 774                                loghandle2name(loghandle));
 775                         rc = -EAGAIN;
 776                 } else {
 777                         /* something bad happened to the processing of a local
 778                          * llog file, probably I/O error or the log got
 779                          * corrupted to be able to finally release the log we
 780                          * discard any remaining bits in the header */
 781                         CERROR("%s: Local llog found corrupted #"DOSTID":%x"
 782                                " %s index %d count %d\n",
 783                                loghandle2name(loghandle),
 784                                POSTID(&loghandle->lgh_id.lgl_oi),
 785                                loghandle->lgh_id.lgl_ogen,
 786                                ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
 787                                 "plain"), index, llh->llh_count);
 788
 789                         while (index <= last_index) {
 790                                 if (test_bit_le(index,
 791                                                   LLOG_HDR_BITMAP(llh)) != 0)
 792                                         llog_cancel_rec(lpi->lpi_env, loghandle,
 793                                                         index);
 794                                 index++;
 795                         }
 796                         rc = 0;
 797                 }
 798         }
 799
 800         OBD_FREE_LARGE(buf, chunk_size);
 801         lpi->lpi_rc = rc;
 802         return 0;
 803 }
 804
 805 static int llog_process_thread_daemonize(void *arg)
 806 {
 807         struct llog_process_info        *lpi = arg;
 808         struct lu_env                    env;
 809         int                              rc;
 810         struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
 811
 812         task_lock(lpi->lpi_reftask);
 813         new_ns = lpi->lpi_reftask->nsproxy;
 814         if (curr_ns != new_ns) {
 815                 get_nsproxy(new_ns);
 816
 817                 current->nsproxy = new_ns;
 818                 /* XXX: we should call put_nsproxy() instead of
 819                  * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
 820                  * used outside of the kernel itself, because it calls
 821                  * free_nsproxy() which is not exported by the kernel
 822                  * (defined in kernel/nsproxy.c) */
 823                 if (curr_ns)
 824                         atomic_dec(&curr_ns->count);
 825         }
 826         task_unlock(lpi->lpi_reftask);
 827
 828         unshare_fs_struct();
 829         /* client env has no keys, tags is just 0 */
 830         rc = lu_env_init(&env, LCT_LOCAL | LCT_MG_THREAD);
 831         if (rc)
 832                 goto out;
 833         lpi->lpi_env = &env;
 834
 835         rc = llog_process_thread(arg);
 836
 837         lu_env_fini(&env);
 838 out:
 839         complete(&lpi->lpi_completion);
 840         return rc;
 841 }
 842
 843 int llog_process_or_fork(const struct lu_env *env,
 844                          struct llog_handle *loghandle,
 845                          llog_cb_t cb, void *data, void *catdata, bool fork)
 846 {
 847         struct llog_process_info *lpi;
 848         struct llog_process_data *d = data;
 849         struct llog_process_cat_data *cd = catdata;
 850         __u32 flags = loghandle->lgh_hdr->llh_flags;
 851         int rc;
 852
 853         ENTRY;
 854
 855         OBD_ALLOC_PTR(lpi);
 856         if (lpi == NULL) {
 857                 CERROR("cannot alloc pointer\n");
 858                 RETURN(-ENOMEM);
 859         }
 860         lpi->lpi_loghandle = loghandle;
 861         lpi->lpi_cb        = cb;
 862         lpi->lpi_cbdata    = data;
 863         lpi->lpi_catdata   = catdata;
 864
 865         CDEBUG(D_OTHER, "Processing "DFID" flags 0x%03x startcat %d startidx %d first_idx %d last_idx %d\n",
 866                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), flags,
 867                (flags & LLOG_F_IS_CAT) && d ? d->lpd_startcat : -1,
 868                (flags & LLOG_F_IS_CAT) && d ? d->lpd_startidx : -1,
 869                cd ? cd->lpcd_first_idx : -1, cd ? cd->lpcd_last_idx : -1);
 870         if (fork) {
 871                 struct task_struct *task;
 872
 873                 /* The new thread can't use parent env,
 874                  * init the new one in llog_process_thread_daemonize. */
 875                 lpi->lpi_env = NULL;
 876                 init_completion(&lpi->lpi_completion);
 877                 /* take reference to current, so that
 878                  * llog_process_thread_daemonize() can use it to switch to
 879                  * namespace associated with current  */
 880                 lpi->lpi_reftask = current;
 881                 task = kthread_run(llog_process_thread_daemonize, lpi,
 882                                    "llog_process_thread");
 883                 if (IS_ERR(task)) {
 884                         rc = PTR_ERR(task);
 885                         CERROR("%s: cannot start thread: rc = %d\n",
 886                                loghandle2name(loghandle), rc);
 887                         GOTO(out_lpi, rc);
 888                 }
 889                 wait_for_completion(&lpi->lpi_completion);
 890         } else {
 891                 lpi->lpi_env = env;
 892                 llog_process_thread(lpi);
 893         }
 894         rc = lpi->lpi_rc;
 895
 896 out_lpi:
 897         OBD_FREE_PTR(lpi);
 898         RETURN(rc);
 899 }
 900 EXPORT_SYMBOL(llog_process_or_fork);
 901
 902 int llog_process(const struct lu_env *env, struct llog_handle *loghandle,
 903                  llog_cb_t cb, void *data, void *catdata)
 904 {
 905         int rc;
 906         rc = llog_process_or_fork(env, loghandle, cb, data, catdata, true);
 907         return rc == LLOG_DEL_PLAIN ? 0 : rc;
 908 }
 909 EXPORT_SYMBOL(llog_process);
 910
 911 static inline const struct cred *llog_raise_resource(void)
 912 {
 913         struct cred *cred = NULL;
 914
 915         if (cap_raised(current_cap(), CAP_SYS_RESOURCE))
 916                 return cred;
 917
 918         cred = prepare_creds();
 919         if (!cred)
 920                 return cred;
 921
 922         cap_raise(cred->cap_effective, CAP_SYS_RESOURCE);
 923         return override_creds(cred);
 924 }
 925
 926 static inline void llog_restore_resource(const struct cred *old_cred)
 927 {
 928         if (old_cred)
 929                 revert_creds(old_cred);
 930 }
 931
 932 int llog_reverse_process(const struct lu_env *env,
 933                          struct llog_handle *loghandle, llog_cb_t cb,
 934                          void *data, void *catdata)
 935 {
 936         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 937         struct llog_process_cat_data *cd = catdata;
 938         void *buf;
 939         int rc = 0, first_index = 1, index, idx;
 940         __u32   chunk_size = llh->llh_hdr.lrh_len;
 941         ENTRY;
 942
 943         OBD_ALLOC_LARGE(buf, chunk_size);
 944         if (buf == NULL)
 945                 RETURN(-ENOMEM);
 946
 947         if (cd != NULL)
 948                 first_index = cd->lpcd_first_idx + 1;
 949         if (cd != NULL && cd->lpcd_last_idx)
 950                 index = cd->lpcd_last_idx;
 951         else
 952                 index = LLOG_HDR_BITMAP_SIZE(llh) - 1;
 953
 954         while (rc == 0) {
 955                 struct llog_rec_hdr *rec;
 956                 struct llog_rec_tail *tail;
 957
 958                 /* skip records not set in bitmap */
 959                 while (index >= first_index &&
 960                        !test_bit_le(index, LLOG_HDR_BITMAP(llh)))
 961                         --index;
 962
 963                 LASSERT(index >= first_index - 1);
 964                 if (index == first_index - 1)
 965                         break;
 966
 967                 /* get the buf with our target record; avoid old garbage */
 968                 memset(buf, 0, chunk_size);
 969                 rc = llog_prev_block(env, loghandle, index, buf, chunk_size);
 970                 if (rc)
 971                         GOTO(out, rc);
 972
 973                 rec = buf;
 974                 idx = rec->lrh_index;
 975                 CDEBUG(D_RPCTRACE, "index %u : idx %u\n", index, idx);
 976                 while (idx < index) {
 977                         rec = (void *)rec + rec->lrh_len;
 978                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 979                                 lustre_swab_llog_rec(rec);
 980                         idx ++;
 981                 }
 982                 LASSERT(idx == index);
 983                 tail = (void *)rec + rec->lrh_len - sizeof(*tail);
 984
 985                 /* process records in buffer, starting where we found one */
 986                 while ((void *)tail > buf) {
 987                         if (tail->lrt_index == 0)
 988                                 GOTO(out, rc = 0); /* no more records */
 989
 990                         /* if set, process the callback on this record */
 991                         if (test_bit_le(index, LLOG_HDR_BITMAP(llh))) {
 992                                 rec = (void *)tail - tail->lrt_len +
 993                                       sizeof(*tail);
 994
 995                                 rc = cb(env, loghandle, rec, data);
 996                                 if (rc == LLOG_PROC_BREAK) {
 997                                         GOTO(out, rc);
 998                                 } else if (rc == LLOG_DEL_RECORD) {
 999                                         rc = llog_cancel_rec(env, loghandle,
1000                                                              tail->lrt_index);
1001                                 }
1002                                 if (rc)
1003                                         GOTO(out, rc);
1004                         }
1005
1006                         /* previous record, still in buffer? */
1007                         --index;
1008                         if (index < first_index)
1009                                 GOTO(out, rc = 0);
1010                         tail = (void *)tail - tail->lrt_len;
1011                 }
1012         }
1013
1014 out:
1015         if (buf != NULL)
1016                 OBD_FREE_LARGE(buf, chunk_size);
1017         RETURN(rc);
1018 }
1019 EXPORT_SYMBOL(llog_reverse_process);
1020
1021 /**
1022  * new llog API
1023  *
1024  * API functions:
1025  *      llog_open - open llog, may not exist
1026  *      llog_exist - check if llog exists
1027  *      llog_close - close opened llog, pair for open, frees llog_handle
1028  *      llog_declare_create - declare llog creation
1029  *      llog_create - create new llog on disk, need transaction handle
1030  *      llog_declare_write_rec - declaration of llog write
1031  *      llog_write_rec - write llog record on disk, need transaction handle
1032  *      llog_declare_add - declare llog catalog record addition
1033  *      llog_add - add llog record in catalog, need transaction handle
1034  */
1035 int llog_exist(struct llog_handle *loghandle)
1036 {
1037         const struct llog_operations *lop;
1038         int rc;
1039
1040         ENTRY;
1041
1042         rc = llog_handle2ops(loghandle, &lop);
1043         if (rc)
1044                 RETURN(rc);
1045         if (lop->lop_exist == NULL)
1046                 RETURN(-EOPNOTSUPP);
1047
1048         rc = lop->lop_exist(loghandle);
1049         RETURN(rc);
1050 }
1051 EXPORT_SYMBOL(llog_exist);
1052
1053 int llog_declare_create(const struct lu_env *env,
1054                         struct llog_handle *loghandle, struct thandle *th)
1055 {
1056         const struct cred *old_cred;
1057         const struct llog_operations *lop;
1058         int rc;
1059
1060         ENTRY;
1061
1062         rc = llog_handle2ops(loghandle, &lop);
1063         if (rc)
1064                 RETURN(rc);
1065         if (lop->lop_declare_create == NULL)
1066                 RETURN(-EOPNOTSUPP);
1067
1068         old_cred = llog_raise_resource();
1069         rc = lop->lop_declare_create(env, loghandle, th);
1070         llog_restore_resource(old_cred);
1071         RETURN(rc);
1072 }
1073
1074 int llog_create(const struct lu_env *env, struct llog_handle *handle,
1075                 struct thandle *th)
1076 {
1077         const struct cred *old_cred;
1078         const struct llog_operations *lop;
1079         int rc;
1080
1081         ENTRY;
1082
1083         rc = llog_handle2ops(handle, &lop);
1084         if (rc)
1085                 RETURN(rc);
1086         if (lop->lop_create == NULL)
1087                 RETURN(-EOPNOTSUPP);
1088
1089         old_cred = llog_raise_resource();
1090         rc = lop->lop_create(env, handle, th);
1091         llog_restore_resource(old_cred);
1092         RETURN(rc);
1093 }
1094
1095 int llog_declare_write_rec(const struct lu_env *env,
1096                            struct llog_handle *handle,
1097                            struct llog_rec_hdr *rec, int idx,
1098                            struct thandle *th)
1099 {
1100         const struct cred *old_cred;
1101         const struct llog_operations *lop;
1102         int rc;
1103
1104         ENTRY;
1105
1106         rc = llog_handle2ops(handle, &lop);
1107         if (rc)
1108                 RETURN(rc);
1109         LASSERT(lop);
1110         if (lop->lop_declare_write_rec == NULL)
1111                 RETURN(-EOPNOTSUPP);
1112
1113         old_cred = llog_raise_resource();
1114         rc = lop->lop_declare_write_rec(env, handle, rec, idx, th);
1115         llog_restore_resource(old_cred);
1116         RETURN(rc);
1117 }
1118
1119 int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
1120                    struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1121                    int idx, struct thandle *th)
1122 {
1123         const struct cred *old_cred;
1124         const struct llog_operations *lop;
1125         int rc, buflen;
1126
1127         ENTRY;
1128
1129         /* API sanity checks */
1130         if (handle == NULL) {
1131                 CERROR("loghandle is missed\n");
1132                 RETURN(-EPROTO);
1133         } else if (handle->lgh_obj == NULL) {
1134                 CERROR("loghandle %p with NULL object\n",
1135                         handle);
1136                 RETURN(-EPROTO);
1137         } else if (th == NULL) {
1138                 CERROR("%s: missed transaction handle\n",
1139                        loghandle2name(handle));
1140                 RETURN(-EPROTO);
1141         } else if (handle->lgh_hdr == NULL) {
1142                 CERROR("%s: loghandle %p with no header\n",
1143                        loghandle2name(handle), handle);
1144                 RETURN(-EPROTO);
1145         }
1146
1147         rc = llog_handle2ops(handle, &lop);
1148         if (rc)
1149                 RETURN(rc);
1150
1151         if (lop->lop_write_rec == NULL)
1152                 RETURN(-EOPNOTSUPP);
1153
1154         buflen = rec->lrh_len;
1155         LASSERT(cfs_size_round(buflen) == buflen);
1156
1157         old_cred = llog_raise_resource();
1158         rc = lop->lop_write_rec(env, handle, rec, logcookies, idx, th);
1159         llog_restore_resource(old_cred);
1160         RETURN(rc);
1161 }
1162
1163 int llog_add(const struct lu_env *env, struct llog_handle *lgh,
1164              struct llog_rec_hdr *rec, struct llog_cookie *logcookies,
1165              struct thandle *th)
1166 {
1167         const struct cred *old_cred;
1168         int rc;
1169
1170         ENTRY;
1171
1172         if (lgh->lgh_logops->lop_add == NULL)
1173                 RETURN(-EOPNOTSUPP);
1174
1175         old_cred = llog_raise_resource();
1176         rc = lgh->lgh_logops->lop_add(env, lgh, rec, logcookies, th);
1177         llog_restore_resource(old_cred);
1178         RETURN(rc);
1179 }
1180 EXPORT_SYMBOL(llog_add);
1181
1182 int llog_declare_add(const struct lu_env *env, struct llog_handle *lgh,
1183                      struct llog_rec_hdr *rec, struct thandle *th)
1184 {
1185         const struct cred *old_cred;
1186         int rc;
1187
1188         ENTRY;
1189
1190         if (lgh->lgh_logops->lop_declare_add == NULL)
1191                 RETURN(-EOPNOTSUPP);
1192
1193         old_cred = llog_raise_resource();
1194         rc = lgh->lgh_logops->lop_declare_add(env, lgh, rec, th);
1195         llog_restore_resource(old_cred);
1196         RETURN(rc);
1197 }
1198 EXPORT_SYMBOL(llog_declare_add);
1199
1200 /**
1201  * Helper function to open llog or create it if doesn't exist.
1202  * It hides all transaction handling from caller.
1203  */
1204 int llog_open_create(const struct lu_env *env, struct llog_ctxt *ctxt,
1205                      struct llog_handle **res, struct llog_logid *logid,
1206                      char *name)
1207 {
1208         struct dt_device        *d;
1209         struct thandle          *th;
1210         int                      rc;
1211
1212         ENTRY;
1213
1214         rc = llog_open(env, ctxt, res, logid, name, LLOG_OPEN_NEW);
1215         if (rc)
1216                 RETURN(rc);
1217
1218         if (llog_exist(*res))
1219                 RETURN(0);
1220
1221         LASSERT((*res)->lgh_obj != NULL);
1222
1223         d = lu2dt_dev((*res)->lgh_obj->do_lu.lo_dev);
1224
1225         if (unlikely(unlikely(d->dd_rdonly)))
1226                 RETURN(-EROFS);
1227
1228         th = dt_trans_create(env, d);
1229         if (IS_ERR(th))
1230                 GOTO(out, rc = PTR_ERR(th));
1231
1232         /* Create update llog object synchronously, which
1233          * happens during inialization process see
1234          * lod_sub_prep_llog(), to make sure the update
1235          * llog object is created before corss-MDT writing
1236          * updates into the llog object */
1237         if (ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
1238                 th->th_sync = 1;
1239
1240         th->th_wait_submit = 1;
1241         rc = llog_declare_create(env, *res, th);
1242         if (rc == 0) {
1243                 rc = dt_trans_start_local(env, d, th);
1244                 if (rc == 0)
1245                         rc = llog_create(env, *res, th);
1246         }
1247         dt_trans_stop(env, d, th);
1248 out:
1249         if (rc)
1250                 llog_close(env, *res);
1251         RETURN(rc);
1252 }
1253 EXPORT_SYMBOL(llog_open_create);
1254
1255 /**
1256  * Helper function to delete existent llog.
1257  */
1258 int llog_erase(const struct lu_env *env, struct llog_ctxt *ctxt,
1259                struct llog_logid *logid, char *name)
1260 {
1261         struct llog_handle      *handle;
1262         int                      rc = 0, rc2;
1263
1264         ENTRY;
1265
1266         /* nothing to erase */
1267         if (name == NULL && logid == NULL)
1268                 RETURN(0);
1269
1270         rc = llog_open(env, ctxt, &handle, logid, name, LLOG_OPEN_EXISTS);
1271         if (rc < 0)
1272                 RETURN(rc);
1273
1274         rc = llog_init_handle(env, handle, LLOG_F_IS_PLAIN, NULL);
1275         if (rc == 0)
1276                 rc = llog_destroy(env, handle);
1277
1278         rc2 = llog_close(env, handle);
1279         if (rc == 0)
1280                 rc = rc2;
1281         RETURN(rc);
1282 }
1283 EXPORT_SYMBOL(llog_erase);
1284
1285 /*
1286  * Helper function for write record in llog.
1287  * It hides all transaction handling from caller.
1288  * Valid only with local llog.
1289  */
1290 int llog_write(const struct lu_env *env, struct llog_handle *loghandle,
1291                struct llog_rec_hdr *rec, int idx)
1292 {
1293         struct dt_device        *dt;
1294         struct thandle          *th;
1295         bool                    need_cookie;
1296         int                     rc;
1297
1298         ENTRY;
1299
1300         LASSERT(loghandle);
1301         LASSERT(loghandle->lgh_ctxt);
1302         LASSERT(loghandle->lgh_obj != NULL);
1303
1304         dt = lu2dt_dev(loghandle->lgh_obj->do_lu.lo_dev);
1305
1306         if (unlikely(unlikely(dt->dd_rdonly)))
1307                 RETURN(-EROFS);
1308
1309         th = dt_trans_create(env, dt);
1310         if (IS_ERR(th))
1311                 RETURN(PTR_ERR(th));
1312
1313         rc = llog_declare_write_rec(env, loghandle, rec, idx, th);
1314         if (rc)
1315                 GOTO(out_trans, rc);
1316
1317         th->th_wait_submit = 1;
1318         rc = dt_trans_start_local(env, dt, th);
1319         if (rc)
1320                 GOTO(out_trans, rc);
1321
1322         need_cookie = !(idx == LLOG_HEADER_IDX || idx == LLOG_NEXT_IDX);
1323
1324         down_write(&loghandle->lgh_lock);
1325         if (need_cookie) {
1326                 struct llog_thread_info *lti = llog_info(env);
1327
1328                 /* cookie comes from llog_process_thread */
1329                 rc = llog_write_rec(env, loghandle, rec, &lti->lgi_cookie,
1330                                     rec->lrh_index, th);
1331                 /* upper layer didn`t pass cookie so change rc */
1332                 rc = (rc == 1 ? 0 : rc);
1333         } else {
1334                 rc = llog_write_rec(env, loghandle, rec, NULL, idx, th);
1335         }
1336
1337         up_write(&loghandle->lgh_lock);
1338 out_trans:
1339         dt_trans_stop(env, dt, th);
1340         RETURN(rc);
1341 }
1342 EXPORT_SYMBOL(llog_write);
1343
1344 int llog_open(const struct lu_env *env, struct llog_ctxt *ctxt,
1345               struct llog_handle **lgh, struct llog_logid *logid,
1346               char *name, enum llog_open_param open_param)
1347 {
1348         const struct cred *old_cred;
1349         int      rc;
1350
1351         ENTRY;
1352
1353         LASSERT(ctxt);
1354         LASSERT(ctxt->loc_logops);
1355
1356         if (ctxt->loc_logops->lop_open == NULL) {
1357                 *lgh = NULL;
1358                 RETURN(-EOPNOTSUPP);
1359         }
1360
1361         *lgh = llog_alloc_handle();
1362         if (*lgh == NULL)
1363                 RETURN(-ENOMEM);
1364         (*lgh)->lgh_ctxt = ctxt;
1365         (*lgh)->lgh_logops = ctxt->loc_logops;
1366
1367         old_cred = llog_raise_resource();
1368         rc = ctxt->loc_logops->lop_open(env, *lgh, logid, name, open_param);
1369         llog_restore_resource(old_cred);
1370         if (rc) {
1371                 llog_free_handle(*lgh);
1372                 *lgh = NULL;
1373         }
1374         RETURN(rc);
1375 }
1376 EXPORT_SYMBOL(llog_open);
1377
1378 int llog_close(const struct lu_env *env, struct llog_handle *loghandle)
1379 {
1380         return llog_handle_put(env, loghandle);
1381 }
1382 EXPORT_SYMBOL(llog_close);
1383
1384 /**
1385  * Helper function to get the llog size in records. It is used by MGS
1386  * mostly to check that config llog exists and contains data.
1387  *
1388  * \param[in] env       execution environment
1389  * \param[in] ctxt      llog context
1390  * \param[in] name      llog name
1391  *
1392  * \retval              true if there are records in llog besides a header
1393  * \retval              false on error or llog without records
1394  */
1395 int llog_is_empty(const struct lu_env *env, struct llog_ctxt *ctxt,
1396                   char *name)
1397 {
1398         struct llog_handle      *llh;
1399         int                      rc = 0;
1400
1401         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1402         if (rc < 0) {
1403                 if (likely(rc == -ENOENT))
1404                         rc = 0;
1405                 GOTO(out, rc);
1406         }
1407
1408         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1409         if (rc)
1410                 GOTO(out_close, rc);
1411         rc = llog_get_size(llh);
1412
1413 out_close:
1414         llog_close(env, llh);
1415 out:
1416         /* The header is record 1, the llog is still considered as empty
1417          * if there is only header */
1418         return (rc <= 1);
1419 }
1420 EXPORT_SYMBOL(llog_is_empty);
1421
1422 int llog_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1423                       struct llog_rec_hdr *rec, void *data)
1424 {
1425         struct llog_handle      *copy_llh = data;
1426
1427         /* Append all records */
1428         return llog_write(env, copy_llh, rec, LLOG_NEXT_IDX);
1429 }
1430
1431 /* backup plain llog */
1432 int llog_backup(const struct lu_env *env, struct obd_device *obd,
1433                 struct llog_ctxt *ctxt, struct llog_ctxt *bctxt,
1434                 char *name, char *backup)
1435 {
1436         struct llog_handle      *llh, *bllh;
1437         int                      rc;
1438
1439         ENTRY;
1440
1441         /* open original log */
1442         rc = llog_open(env, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1443         if (rc < 0) {
1444                 /* the -ENOENT case is also reported to the caller
1445                  * but silently so it should handle that if needed.
1446                  */
1447                 if (rc != -ENOENT)
1448                         CERROR("%s: failed to open log %s: rc = %d\n",
1449                                obd->obd_name, name, rc);
1450                 RETURN(rc);
1451         }
1452
1453         rc = llog_init_handle(env, llh, LLOG_F_IS_PLAIN, NULL);
1454         if (rc)
1455                 GOTO(out_close, rc);
1456
1457         /* Make sure there's no old backup log */
1458         rc = llog_erase(env, bctxt, NULL, backup);
1459         if (rc < 0 && rc != -ENOENT)
1460                 GOTO(out_close, rc);
1461
1462         /* open backup log */
1463         rc = llog_open_create(env, bctxt, &bllh, NULL, backup);
1464         if (rc) {
1465                 CERROR("%s: failed to open backup logfile %s: rc = %d\n",
1466                        obd->obd_name, backup, rc);
1467                 GOTO(out_close, rc);
1468         }
1469
1470         /* check that backup llog is not the same object as original one */
1471         if (llh->lgh_obj == bllh->lgh_obj) {
1472                 CERROR("%s: backup llog %s to itself (%s), objects %p/%p\n",
1473                        obd->obd_name, name, backup, llh->lgh_obj,
1474                        bllh->lgh_obj);
1475                 GOTO(out_backup, rc = -EEXIST);
1476         }
1477
1478         rc = llog_init_handle(env, bllh, LLOG_F_IS_PLAIN, NULL);
1479         if (rc)
1480                 GOTO(out_backup, rc);
1481
1482         /* Copy log record by record */
1483         rc = llog_process_or_fork(env, llh, llog_copy_handler, (void *)bllh,
1484                                   NULL, false);
1485         if (rc)
1486                 CERROR("%s: failed to backup log %s: rc = %d\n",
1487                        obd->obd_name, name, rc);
1488 out_backup:
1489         llog_close(env, bllh);
1490 out_close:
1491         llog_close(env, llh);
1492         RETURN(rc);
1493 }
1494 EXPORT_SYMBOL(llog_backup);
1495
1496 /* Get size of llog */
1497 __u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
1498 {
1499         int rc;
1500         struct lu_attr la;
1501
1502         rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
1503         if (rc) {
1504                 CERROR("%s: attr_get failed for "DFID": rc = %d\n",
1505                        loghandle2name(llh), PFID(&llh->lgh_id.lgl_oi.oi_fid),
1506                        rc);
1507                 return 0;
1508         }
1509
1510         return la.la_size;
1511 }
1512 EXPORT_SYMBOL(llog_size);
1513