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