Whamcloud - gitweb
LU-17705 ptlrpc: replace synchronize_rcu() with rcu_barrier()
[fs/lustre-release.git] / lustre / obdclass / llog_cat.c
index 05d2f26..3d9cceb 100644 (file)
  *
  * You should have received a copy of the GNU General Public License
  * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
  *
  * GPL HEADER END
  */
  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, 2013, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
  *
  * lustre/obdclass/llog_cat.c
  *
 
 #define DEBUG_SUBSYSTEM S_LOG
 
-#ifndef __KERNEL__
-#include <liblustre.h>
-#endif
 
 #include <obd_class.h>
 
 #include "llog_internal.h"
 
+
+/**
+ * lockdep markers for nested struct llog_handle::lgh_lock locking.
+ */
+enum {
+       LLOGH_CAT,
+       LLOGH_LOG,
+};
+
 /* Create a new log handle and add it to the open list.
  * This log handle will be closed when all of the records in it are removed.
  *
@@ -68,28 +69,101 @@ static int llog_cat_new_log(const struct lu_env *env,
 {
        struct llog_thread_info *lgi = llog_info(env);
        struct llog_logid_rec   *rec = &lgi->lgi_logid;
-       int                      rc;
+       struct thandle *handle = NULL;
+       struct dt_device *dt = NULL;
+       struct llog_log_hdr     *llh = cathandle->lgh_hdr;
+       int                      rc, index;
 
        ENTRY;
 
-       if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED))
+       index = (cathandle->lgh_last_idx + 1) % (llog_max_idx(llh) + 1);
+
+       /* check that new llog index will not overlap with the first one.
+        * - llh_cat_idx is the index just before the first/oldest still in-use
+        *      index in catalog
+        * - lgh_last_idx is the last/newest used index in catalog
+        *
+        * When catalog is not wrapped yet then lgh_last_idx is always larger
+        * than llh_cat_idx. After the wrap around lgh_last_idx re-starts
+        * from 0 and llh_cat_idx becomes the upper limit for it
+        *
+        * Check if catalog has already wrapped around or not by comparing
+        * last_idx and cat_idx */
+       if ((index == llh->llh_cat_idx + 1 && llh->llh_count > 1) ||
+           (index == 0 && llh->llh_cat_idx == 0)) {
+               if (cathandle->lgh_name == NULL) {
+                       CWARN("%s: there are no more free slots in catalog "DFID"\n",
+                             loghandle2name(loghandle),
+                             PLOGID(&cathandle->lgh_id));
+               } else {
+                       CWARN("%s: there are no more free slots in catalog %s\n",
+                             loghandle2name(loghandle), cathandle->lgh_name);
+               }
+               RETURN(-ENOSPC);
+       }
+
+       if (CFS_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED))
                RETURN(-ENOSPC);
 
+       if (loghandle->lgh_hdr != NULL) {
+               /* If llog object is remote and creation is failed, lgh_hdr
+                * might be left over here, free it first */
+               LASSERT(!llog_exist(loghandle));
+               OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
+               loghandle->lgh_hdr = NULL;
+       }
+
+       if (th == NULL) {
+               dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
+
+               handle = dt_trans_create(env, dt);
+               if (IS_ERR(handle))
+                       RETURN(PTR_ERR(handle));
+
+               /* Create update llog object synchronously, which
+                * happens during inialization process see
+                * lod_sub_prep_llog(), to make sure the update
+                * llog object is created before corss-MDT writing
+                * updates into the llog object */
+               if (cathandle->lgh_ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
+                       handle->th_sync = 1;
+
+               handle->th_wait_submit = 1;
+
+               rc = llog_declare_create(env, loghandle, handle);
+               if (rc != 0)
+                       GOTO(out, rc);
+
+               rec->lid_hdr.lrh_len = sizeof(*rec);
+               rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
+               rec->lid_id = loghandle->lgh_id;
+               rc = llog_declare_write_rec(env, cathandle, &rec->lid_hdr, -1,
+                                           handle);
+               if (rc != 0)
+                       GOTO(out, rc);
+
+               rc = dt_trans_start_local(env, dt, handle);
+               if (rc != 0)
+                       GOTO(out, rc);
+
+               th = handle;
+       }
+
        rc = llog_create(env, loghandle, th);
        /* if llog is already created, no need to initialize it */
        if (rc == -EEXIST) {
-               RETURN(0);
+               GOTO(out, rc = 0);
        } else if (rc != 0) {
                CERROR("%s: can't create new plain llog in catalog: rc = %d\n",
-                      loghandle->lgh_ctxt->loc_obd->obd_name, rc);
-               RETURN(rc);
+                      loghandle2name(loghandle), rc);
+               GOTO(out, rc);
        }
 
        rc = llog_init_handle(env, loghandle,
                              LLOG_F_IS_PLAIN | LLOG_F_ZAP_WHEN_EMPTY,
                              &cathandle->lgh_hdr->llh_tgtuuid);
-       if (rc)
-               GOTO(out_destroy, rc);
+       if (rc < 0)
+               GOTO(out, rc);
 
        /* build the record for this log in the catalog */
        rec->lid_hdr.lrh_len = sizeof(*rec);
@@ -103,18 +177,184 @@ static int llog_cat_new_log(const struct lu_env *env,
        if (rc < 0)
                GOTO(out_destroy, rc);
 
-       CDEBUG(D_OTHER, "new recovery log "DOSTID":%x for index %u of catalog"
-              DOSTID"\n", POSTID(&loghandle->lgh_id.lgl_oi),
-              loghandle->lgh_id.lgl_ogen, rec->lid_hdr.lrh_index,
-              POSTID(&cathandle->lgh_id.lgl_oi));
+       CDEBUG(D_OTHER, "new plain log "DFID".%u of catalog "DFID"\n",
+              PLOGID(&loghandle->lgh_id), rec->lid_hdr.lrh_index,
+              PLOGID(&cathandle->lgh_id));
 
        loghandle->lgh_hdr->llh_cat_idx = rec->lid_hdr.lrh_index;
-       RETURN(0);
+
+       /* limit max size of plain llog so that space can be
+        * released sooner, especially on small filesystems */
+       /* 2MB for the cases when free space hasn't been learned yet */
+       loghandle->lgh_max_size = 2 << 20;
+       dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
+       rc = dt_statfs(env, dt, &lgi->lgi_statfs);
+       if (rc == 0 && lgi->lgi_statfs.os_bfree > 0) {
+               __u64 freespace = (lgi->lgi_statfs.os_bfree *
+                                 lgi->lgi_statfs.os_bsize) >> 6;
+               if (freespace < loghandle->lgh_max_size)
+                       loghandle->lgh_max_size = freespace;
+               /* shouldn't be > 128MB in any case?
+                * it's 256K records of 512 bytes each */
+               if (freespace > (128 << 20))
+                       loghandle->lgh_max_size = 128 << 20;
+       }
+       if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PLAIN_RECORDS) ||
+                    CFS_FAIL_PRECHECK(OBD_FAIL_CATALOG_FULL_CHECK))) {
+               // limit the numer of plain records for test
+               loghandle->lgh_max_size = loghandle->lgh_hdr_size +
+                      cfs_fail_val * 64;
+       }
+
+       rc = 0;
+
+out:
+       if (handle != NULL) {
+               handle->th_result = rc >= 0 ? 0 : rc;
+               dt_trans_stop(env, dt, handle);
+       }
+       RETURN(rc);
+
 out_destroy:
-       llog_destroy(env, loghandle);
+       /* to signal llog_cat_close() it shouldn't try to destroy the llog,
+        * we want to destroy it in this transaction, otherwise the object
+        * becomes an orphan */
+       loghandle->lgh_hdr->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
+       /* this is to mimic full log, so another llog_cat_current_log()
+        * can skip it and ask for another onet */
+       loghandle->lgh_last_idx = llog_max_idx(loghandle->lgh_hdr) + 1;
+       llog_trans_destroy(env, loghandle, th);
+       if (handle != NULL)
+               dt_trans_stop(env, dt, handle);
        RETURN(rc);
 }
 
+static int llog_cat_refresh(const struct lu_env *env,
+                           struct llog_handle *cathandle)
+{
+       struct llog_handle *loghandle;
+       int rc;
+
+       down_write(&cathandle->lgh_lock);
+       list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
+                           u.phd.phd_entry) {
+               if (!llog_exist(loghandle))
+                       continue;
+
+               down_write(&loghandle->lgh_lock);
+               rc = llog_read_header(env, loghandle, NULL);
+               up_write(&loghandle->lgh_lock);
+               if (rc)
+                       goto unlock;
+       }
+
+       rc = llog_read_header(env, cathandle, NULL);
+unlock:
+       up_write(&cathandle->lgh_lock);
+
+       return rc;
+}
+
+/*
+ * prepare current/next log for catalog.
+ *
+ * if \a *ploghandle is NULL, open it, and declare create, NB, if \a
+ * *ploghandle is remote, create it synchronously here, see comments
+ * below.
+ *
+ * \a cathandle->lgh_lock is down_read-ed, it gets down_write-ed if \a
+ * *ploghandle has to be opened.
+ */
+static int llog_cat_prep_log(const struct lu_env *env,
+                            struct llog_handle *cathandle,
+                            struct llog_handle **ploghandle,
+                            struct thandle *th)
+{
+       int rc;
+       int sem_upgraded;
+
+start:
+       rc = 0;
+       sem_upgraded = 0;
+       if (IS_ERR_OR_NULL(*ploghandle)) {
+               up_read(&cathandle->lgh_lock);
+               down_write(&cathandle->lgh_lock);
+               sem_upgraded = 1;
+               if (IS_ERR_OR_NULL(*ploghandle)) {
+                       struct llog_handle *loghandle;
+
+                       rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
+                                      NULL, NULL, LLOG_OPEN_NEW);
+                       if (!rc) {
+                               *ploghandle = loghandle;
+                               list_add_tail(&loghandle->u.phd.phd_entry,
+                                             &cathandle->u.chd.chd_head);
+                       }
+               }
+               if (rc)
+                       GOTO(out, rc);
+       }
+
+       rc = llog_exist(*ploghandle);
+       if (rc < 0)
+               GOTO(out, rc);
+       if (rc)
+               GOTO(out, rc = 0);
+
+       if (dt_object_remote(cathandle->lgh_obj)) {
+               down_write_nested(&(*ploghandle)->lgh_lock, LLOGH_LOG);
+               if (!llog_exist(*ploghandle)) {
+                       /* For remote operation, if we put the llog object
+                        * creation in the current transaction, then the
+                        * llog object will not be created on the remote
+                        * target until the transaction stop, if other
+                        * operations start before the transaction stop,
+                        * and use the same llog object, will be dependent
+                        * on the success of this transaction. So let's
+                        * create the llog object synchronously here to
+                        * remove the dependency. */
+                       rc = llog_cat_new_log(env, cathandle, *ploghandle,
+                                             NULL);
+                       if (rc == -ESTALE) {
+                               up_write(&(*ploghandle)->lgh_lock);
+                               if (sem_upgraded)
+                                       up_write(&cathandle->lgh_lock);
+                               else
+                                       up_read(&cathandle->lgh_lock);
+
+                               rc = llog_cat_refresh(env, cathandle);
+                               down_read_nested(&cathandle->lgh_lock,
+                                                LLOGH_CAT);
+                               if (rc)
+                                       return rc;
+                               /* *ploghandle might become NULL, restart */
+                               goto start;
+                       }
+               }
+               up_write(&(*ploghandle)->lgh_lock);
+       } else {
+               struct llog_thread_info *lgi = llog_info(env);
+               struct llog_logid_rec *lirec = &lgi->lgi_logid;
+
+               rc = llog_declare_create(env, *ploghandle, th);
+               if (rc)
+                       GOTO(out, rc);
+
+               lirec->lid_hdr.lrh_len = sizeof(*lirec);
+               rc = llog_declare_write_rec(env, cathandle, &lirec->lid_hdr, -1,
+                                           th);
+       }
+
+out:
+       if (sem_upgraded) {
+               up_write(&cathandle->lgh_lock);
+               down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
+               if (rc == 0)
+                       goto start;
+       }
+       return rc;
+}
+
 /* Open an existent log handle and add it to the open list.
  * This log handle will be closed when all of the records in it are removed.
  *
@@ -128,6 +368,7 @@ int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
                       struct llog_handle **res, struct llog_logid *logid)
 {
        struct llog_handle      *loghandle;
+       enum llog_flag           fmt;
        int                      rc = 0;
 
        ENTRY;
@@ -135,6 +376,7 @@ int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
        if (cathandle == NULL)
                RETURN(-EBADF);
 
+       fmt = cathandle->lgh_hdr->llh_flags & LLOG_F_EXT_MASK;
        down_write(&cathandle->lgh_lock);
        list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
                            u.phd.phd_entry) {
@@ -142,16 +384,16 @@ int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
 
                if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
                    ostid_seq(&cgl->lgl_oi) == ostid_seq(&logid->lgl_oi)) {
-                       if (cgl->lgl_ogen != logid->lgl_ogen) {
-                               CERROR("%s: log "DOSTID" generation %x != %x\n",
-                                      loghandle->lgh_ctxt->loc_obd->obd_name,
-                                      POSTID(&logid->lgl_oi), cgl->lgl_ogen,
-                                      logid->lgl_ogen);
+                       *res = llog_handle_get(loghandle);
+                       if (!*res) {
+                               CERROR("%s: log "DFID" refcount is zero!\n",
+                                      loghandle2name(loghandle),
+                                      PLOGID(logid));
                                continue;
                        }
                        loghandle->u.phd.phd_cat_handle = cathandle;
                        up_write(&cathandle->lgh_lock);
-                       GOTO(out, rc = 0);
+                       RETURN(rc);
                }
        }
        up_write(&cathandle->lgh_lock);
@@ -159,32 +401,30 @@ int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
        rc = llog_open(env, cathandle->lgh_ctxt, &loghandle, logid, NULL,
                       LLOG_OPEN_EXISTS);
        if (rc < 0) {
-               CERROR("%s: error opening log id "DOSTID":%x: rc = %d\n",
-                      cathandle->lgh_ctxt->loc_obd->obd_name,
-                      POSTID(&logid->lgl_oi), logid->lgl_ogen, rc);
+               CERROR("%s: error opening log id "DFID": rc = %d\n",
+                      loghandle2name(cathandle), PLOGID(logid), rc);
                RETURN(rc);
        }
 
-       rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN, NULL);
+       rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN |
+                             LLOG_F_ZAP_WHEN_EMPTY | fmt, NULL);
        if (rc < 0) {
                llog_close(env, loghandle);
-               loghandle = NULL;
+               *res = NULL;
                RETURN(rc);
        }
 
+       *res = llog_handle_get(loghandle);
+       LASSERT(*res);
        down_write(&cathandle->lgh_lock);
-       list_add(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
+       list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
        up_write(&cathandle->lgh_lock);
 
        loghandle->u.phd.phd_cat_handle = cathandle;
        loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
        loghandle->u.phd.phd_cookie.lgc_index =
                                loghandle->lgh_hdr->llh_cat_idx;
-       EXIT;
-out:
-       llog_handle_get(loghandle);
-       *res = loghandle;
-       return 0;
+       RETURN(0);
 }
 
 int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
@@ -209,8 +449,7 @@ int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
                        if (rc)
                                CERROR("%s: failure destroying log during "
                                       "cleanup: rc = %d\n",
-                                      loghandle->lgh_ctxt->loc_obd->obd_name,
-                                      rc);
+                                      loghandle2name(loghandle), rc);
 
                        index = loghandle->u.phd.phd_cookie.lgc_index;
                        llog_cat_cleanup(env, cathandle, NULL, index);
@@ -225,14 +464,6 @@ int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
 }
 EXPORT_SYMBOL(llog_cat_close);
 
-/**
- * lockdep markers for nested struct llog_handle::lgh_lock locking.
- */
-enum {
-        LLOGH_CAT,
-        LLOGH_LOG
-};
-
 /** Return the currently active log handle.  If the current log handle doesn't
  * have enough space left for the current record, start a new one.
  *
@@ -249,6 +480,12 @@ static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
         struct llog_handle *loghandle = NULL;
         ENTRY;
 
+
+       if (CFS_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2)) {
+               down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
+               GOTO(next, loghandle);
+       }
+
        down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
         loghandle = cathandle->u.chd.chd_current_log;
         if (loghandle) {
@@ -256,8 +493,7 @@ static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
 
                down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
                llh = loghandle->lgh_hdr;
-               if (llh == NULL ||
-                   loghandle->lgh_last_idx < LLOG_BITMAP_SIZE(llh) - 1) {
+               if (llh == NULL || !llog_is_full(loghandle)) {
                        up_read(&cathandle->lgh_lock);
                         RETURN(loghandle);
                 } else {
@@ -276,14 +512,31 @@ static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
 
                down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
                llh = loghandle->lgh_hdr;
-               LASSERT(llh);
-                if (loghandle->lgh_last_idx < LLOG_BITMAP_SIZE(llh) - 1) {
-                       up_write(&cathandle->lgh_lock);
-                        RETURN(loghandle);
-                } else {
+               if (llh == NULL || !llog_is_full(loghandle))
+                       GOTO(out_unlock, loghandle);
+               else
                        up_write(&loghandle->lgh_lock);
-                }
-        }
+       }
+
+next:
+       /* Sigh, the chd_next_log and chd_current_log is initialized
+        * in declare phase, and we do not serialize the catlog
+        * accessing, so it might be possible the llog creation
+        * thread (see llog_cat_declare_add_rec()) did not create
+        * llog successfully, then the following thread might
+        * meet this situation. */
+       if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
+               CERROR("%s: next log does not exist!\n",
+                      loghandle2name(cathandle));
+               loghandle = ERR_PTR(-EIO);
+               if (cathandle->u.chd.chd_next_log == NULL) {
+                       /* Store the error in chd_next_log, so
+                        * the following process can get correct
+                        * failure value */
+                       cathandle->u.chd.chd_next_log = loghandle;
+               }
+               GOTO(out_unlock, loghandle);
+       }
 
        CDEBUG(D_INODE, "use next log\n");
 
@@ -291,6 +544,8 @@ static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
        cathandle->u.chd.chd_current_log = loghandle;
        cathandle->u.chd.chd_next_log = NULL;
        down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
+
+out_unlock:
        up_write(&cathandle->lgh_lock);
        LASSERT(loghandle);
        RETURN(loghandle);
@@ -306,45 +561,50 @@ int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
                     struct thandle *th)
 {
         struct llog_handle *loghandle;
-        int rc;
-        ENTRY;
+       int rc, retried = 0;
+       ENTRY;
 
-        LASSERT(rec->lrh_len <= LLOG_CHUNK_SIZE);
+       LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
+
+retry:
        loghandle = llog_cat_current_log(cathandle, th);
-       LASSERT(!IS_ERR(loghandle));
+       if (IS_ERR(loghandle))
+               RETURN(PTR_ERR(loghandle));
 
        /* loghandle is already locked by llog_cat_current_log() for us */
        if (!llog_exist(loghandle)) {
                rc = llog_cat_new_log(env, cathandle, loghandle, th);
                if (rc < 0) {
                        up_write(&loghandle->lgh_lock);
+                       /* nobody should be trying to use this llog */
+                       down_write(&cathandle->lgh_lock);
+                       if (cathandle->u.chd.chd_current_log == loghandle)
+                               cathandle->u.chd.chd_current_log = NULL;
+                       up_write(&cathandle->lgh_lock);
                        RETURN(rc);
                }
        }
        /* now let's try to add the record */
        rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
-       if (rc < 0)
+       if (rc < 0) {
                CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
                             "llog_write_rec %d: lh=%p\n", rc, loghandle);
+               /* -ENOSPC is returned if no empty records left
+                * and when it's lack of space on the stogage.
+                * there is no point to try again if it's the second
+                * case. many callers (like llog test) expect ENOSPC,
+                * so we preserve this error code, but look for the
+                * actual cause here */
+               if (rc == -ENOSPC && llog_is_full(loghandle))
+                       rc = -ENOBUFS;
+       }
        up_write(&loghandle->lgh_lock);
-        if (rc == -ENOSPC) {
-               /* try to use next log */
-               loghandle = llog_cat_current_log(cathandle, th);
-               LASSERT(!IS_ERR(loghandle));
-               /* new llog can be created concurrently */
-               if (!llog_exist(loghandle)) {
-                       rc = llog_cat_new_log(env, cathandle, loghandle, th);
-                       if (rc < 0) {
-                               up_write(&loghandle->lgh_lock);
-                               RETURN(rc);
-                       }
-               }
-               /* now let's try to add the record */
-               rc = llog_write_rec(env, loghandle, rec, reccookie,
-                                   LLOG_NEXT_IDX, th);
-               if (rc < 0)
-                       CERROR("llog_write_rec %d: lh=%p\n", rc, loghandle);
-               up_write(&loghandle->lgh_lock);
+
+       if (rc == -ENOBUFS) {
+               if (retried++ == 0)
+                       GOTO(retry, rc);
+               CERROR("%s: error on 2nd llog: rc = %d\n",
+                      loghandle2name(cathandle), rc);
        }
 
        RETURN(rc);
@@ -355,72 +615,43 @@ int llog_cat_declare_add_rec(const struct lu_env *env,
                             struct llog_handle *cathandle,
                             struct llog_rec_hdr *rec, struct thandle *th)
 {
-       struct llog_thread_info *lgi = llog_info(env);
-       struct llog_logid_rec   *lirec = &lgi->lgi_logid;
-       struct llog_handle      *loghandle, *next;
-       int                      rc = 0;
+       int rc;
 
        ENTRY;
 
-       if (cathandle->u.chd.chd_current_log == NULL) {
-               /* declare new plain llog */
-               down_write(&cathandle->lgh_lock);
-               if (cathandle->u.chd.chd_current_log == NULL) {
-                       rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
-                                      NULL, NULL, LLOG_OPEN_NEW);
-                       if (rc == 0) {
-                               cathandle->u.chd.chd_current_log = loghandle;
-                               list_add_tail(&loghandle->u.phd.phd_entry,
-                                             &cathandle->u.chd.chd_head);
-                       }
-               }
-               up_write(&cathandle->lgh_lock);
-       } else if (cathandle->u.chd.chd_next_log == NULL) {
-               /* declare next plain llog */
-               down_write(&cathandle->lgh_lock);
-               if (cathandle->u.chd.chd_next_log == NULL) {
-                       rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
-                                      NULL, NULL, LLOG_OPEN_NEW);
-                       if (rc == 0) {
-                               cathandle->u.chd.chd_next_log = loghandle;
-                               list_add_tail(&loghandle->u.phd.phd_entry,
-                                             &cathandle->u.chd.chd_head);
-                       }
-               }
-               up_write(&cathandle->lgh_lock);
-       }
+start:
+       down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
+       rc = llog_cat_prep_log(env, cathandle,
+                              &cathandle->u.chd.chd_current_log, th);
        if (rc)
-               GOTO(out, rc);
+               GOTO(unlock, rc);
 
-       lirec->lid_hdr.lrh_len = sizeof(*lirec);
+       rc = llog_cat_prep_log(env, cathandle, &cathandle->u.chd.chd_next_log,
+                              th);
+       if (rc)
+               GOTO(unlock, rc);
 
-       if (!llog_exist(cathandle->u.chd.chd_current_log)) {
-               rc = llog_declare_create(env, cathandle->u.chd.chd_current_log,
-                                        th);
-               if (rc)
-                       GOTO(out, rc);
-               llog_declare_write_rec(env, cathandle, &lirec->lid_hdr, -1, th);
-       }
-       /* declare records in the llogs */
        rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
                                    rec, -1, th);
-       if (rc)
-               GOTO(out, rc);
-
-       next = cathandle->u.chd.chd_next_log;
-       if (next) {
-               if (!llog_exist(next)) {
-                       rc = llog_declare_create(env, next, th);
-                       llog_declare_write_rec(env, cathandle, &lirec->lid_hdr,
-                                              -1, th);
-               }
-               /* XXX: we hope for declarations made for existing llog
-                *      this might be not correct with some backends
-                *      where declarations are expected against specific
-                *      object like ZFS with full debugging enabled */
-               /*llog_declare_write_rec(env, next, rec, -1, th);*/
+       if (rc == -ESTALE && dt_object_remote(cathandle->lgh_obj)) {
+               up_read(&cathandle->lgh_lock);
+               rc = llog_cat_refresh(env, cathandle);
+               if (rc)
+                       RETURN(rc);
+               goto start;
        }
-out:
+
+#if 0
+       /*
+        * XXX: we hope for declarations made for existing llog this might be
+        * not correct with some backends where declarations are expected
+        * against specific object like ZFS with full debugging enabled.
+        */
+       rc = llog_declare_write_rec(env, cathandle->u.chd.chd_next_log, rec, -1,
+                                   th);
+#endif
+unlock:
+       up_read(&cathandle->lgh_lock);
        RETURN(rc);
 }
 EXPORT_SYMBOL(llog_cat_declare_add_rec);
@@ -458,6 +689,57 @@ out_trans:
 }
 EXPORT_SYMBOL(llog_cat_add);
 
+int llog_cat_cancel_arr_rec(const struct lu_env *env,
+                           struct llog_handle *cathandle,
+                           struct llog_logid *lgl, int count, int *index)
+{
+       struct llog_handle *loghandle;
+       int  rc;
+
+       ENTRY;
+       rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
+       if (rc) {
+               CDEBUG(D_HA, "%s: can't find llog handle for "DFID": rc = %d\n",
+                      loghandle2name(cathandle), PLOGID(lgl), rc);
+               RETURN(rc);
+       }
+
+       if ((cathandle->lgh_ctxt->loc_flags &
+            LLOG_CTXT_FLAG_NORMAL_FID) && !llog_exist(loghandle)) {
+               /* For update log, some of loghandles of cathandle
+                * might not exist because remote llog creation might
+                * be failed, so let's skip the record cancellation
+                * for these non-exist llogs.
+                */
+               rc = -ENOENT;
+               CDEBUG(D_HA, "%s: llog "DFID" does not exist: rc = %d\n",
+                      loghandle2name(cathandle), PLOGID(lgl), rc);
+               llog_handle_put(env, loghandle);
+               RETURN(rc);
+       }
+
+       rc = llog_cancel_arr_rec(env, loghandle, count, index);
+       if (rc == LLOG_DEL_PLAIN) { /* log has been destroyed */
+               int cat_index;
+
+               cat_index = loghandle->u.phd.phd_cookie.lgc_index;
+               rc = llog_cat_cleanup(env, cathandle, loghandle, cat_index);
+               if (rc)
+                       CERROR("%s: fail to cancel catalog record: rc = %d\n",
+                              loghandle2name(cathandle), rc);
+               rc = 0;
+
+       }
+       llog_handle_put(env, loghandle);
+
+       if (rc)
+               CERROR("%s: fail to cancel %d llog-records: rc = %d\n",
+                      loghandle2name(cathandle), count, rc);
+
+       RETURN(rc);
+}
+EXPORT_SYMBOL(llog_cat_cancel_arr_rec);
+
 /* For each cookie in the cookie array, we clear the log in-use bit and either:
  * - the log is empty, so mark it free in the catalog header and delete it
  * - the log is not empty, just write out the log header
@@ -471,104 +753,111 @@ int llog_cat_cancel_records(const struct lu_env *env,
                            struct llog_handle *cathandle, int count,
                            struct llog_cookie *cookies)
 {
-       int i, index, rc = 0, failed = 0;
+       int i, rc = 0, failed = 0;
 
        ENTRY;
 
        for (i = 0; i < count; i++, cookies++) {
-               struct llog_handle      *loghandle;
-               struct llog_logid       *lgl = &cookies->lgc_lgl;
-               int                      lrc;
-
-               rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
-               if (rc) {
-                       CERROR("%s: cannot find handle for llog "DOSTID": %d\n",
-                              cathandle->lgh_ctxt->loc_obd->obd_name,
-                              POSTID(&lgl->lgl_oi), rc);
-                       failed++;
-                       continue;
-               }
+               int lrc;
 
-               lrc = llog_cancel_rec(env, loghandle, cookies->lgc_index);
-               if (lrc == LLOG_DEL_PLAIN) { /* log has been destroyed */
-                       index = loghandle->u.phd.phd_cookie.lgc_index;
-                       rc = llog_cat_cleanup(env, cathandle, loghandle,
-                                             index);
-               } else if (lrc == -ENOENT) {
-                       if (rc == 0) /* ENOENT shouldn't rewrite any error */
-                               rc = lrc;
-               } else if (lrc < 0) {
+               lrc = llog_cat_cancel_arr_rec(env, cathandle, &cookies->lgc_lgl,
+                                            1, &cookies->lgc_index);
+               if (lrc) {
                        failed++;
-                       rc = lrc;
+                       if (!rc)
+                               rc = lrc;
                }
-               llog_handle_put(loghandle);
        }
-       if (rc)
+       if (failed)
                CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
-                      cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
-                      rc);
-
+                      loghandle2name(cathandle), failed, count, rc);
        RETURN(rc);
 }
 EXPORT_SYMBOL(llog_cat_cancel_records);
 
-int llog_cat_process_cb(const struct lu_env *env, struct llog_handle *cat_llh,
-                       struct llog_rec_hdr *rec, void *data)
+static int llog_cat_process_common(const struct lu_env *env,
+                                  struct llog_handle *cat_llh,
+                                  struct llog_rec_hdr *rec,
+                                  struct llog_handle **llhp)
 {
-        struct llog_process_data *d = data;
-        struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
-        struct llog_handle *llh;
+       struct llog_logid_rec *lir = container_of(rec, typeof(*lir), lid_hdr);
        struct llog_log_hdr *hdr;
-        int rc;
+       int rc;
 
-        ENTRY;
-        if (rec->lrh_type != LLOG_LOGID_MAGIC) {
-                CERROR("invalid record in catalog\n");
-                RETURN(-EINVAL);
-        }
-       CDEBUG(D_HA, "processing log "DOSTID":%x at index %u of catalog "
-              DOSTID"\n", POSTID(&lir->lid_id.lgl_oi), lir->lid_id.lgl_ogen,
-              rec->lrh_index, POSTID(&cat_llh->lgh_id.lgl_oi));
+       ENTRY;
+       if (rec->lrh_type != le32_to_cpu(LLOG_LOGID_MAGIC)) {
+               rc = -EINVAL;
+               CWARN("%s: invalid record in catalog "DFID": rc = %d\n",
+                     loghandle2name(cat_llh), PLOGID(&cat_llh->lgh_id), rc);
+               RETURN(rc);
+       }
+       CDEBUG(D_HA, "processing log "DFID" at index %u of catalog "DFID"\n",
+              PLOGID(&lir->lid_id), le32_to_cpu(rec->lrh_index),
+              PLOGID(&cat_llh->lgh_id));
 
-       rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
+       rc = llog_cat_id2handle(env, cat_llh, llhp, &lir->lid_id);
        if (rc) {
-               CERROR("%s: cannot find handle for llog "DOSTID": %d\n",
-                      cat_llh->lgh_ctxt->loc_obd->obd_name,
-                      POSTID(&lir->lid_id.lgl_oi), rc);
-               if (rc == -ENOENT || rc == -ESTALE) {
-                       /* After a server crash, a stub of index
-                        * record in catlog could be kept, because
-                        * plain log destroy + catlog index record
-                        * deletion are not atomic. So we end up with
-                        * an index but no actual record. Destroy the
-                        * index and move on. */
-                       rc = llog_cat_cleanup(env, cat_llh, NULL,
-                                             rec->lrh_index);
-               }
+               /* After a server crash, a stub of index record in catlog could
+                * be kept, because plain log destroy + catlog index record
+                * deletion are not atomic. So we end up with an index but no
+                * actual record. Destroy the index and move on. */
+               if (rc == -ENOENT || rc == -ESTALE)
+                       rc = LLOG_DEL_RECORD;
+               else if (rc)
+                       CWARN("%s: can't find llog handle "DFID": rc = %d\n",
+                             loghandle2name(cat_llh), PLOGID(&lir->lid_id),
+                             rc);
 
                RETURN(rc);
        }
 
        /* clean old empty llogs, do not consider current llog in use */
-       hdr = llh->lgh_hdr;
+       /* ignore remote (lgh_obj == NULL) llogs */
+       hdr = (*llhp)->lgh_hdr;
        if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
-           hdr->llh_count == 1 &&
-           llh != cat_llh->u.chd.chd_current_log) {
-               rc = llog_destroy(env, llh);
+           hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
+           *llhp != cat_llh->u.chd.chd_current_log &&
+           *llhp != cat_llh->u.chd.chd_next_log) {
+               rc = llog_destroy(env, *llhp);
                if (rc)
-                       CERROR("%s: fail to destroy empty log: rc = %d\n",
-                              llh->lgh_ctxt->loc_obd->obd_name, rc);
-               GOTO(out, rc = LLOG_DEL_PLAIN);
+                       CWARN("%s: can't destroy empty log "DFID": rc = %d\n",
+                             loghandle2name((*llhp)), PLOGID(&lir->lid_id),
+                             rc);
+               rc = LLOG_DEL_PLAIN;
        }
 
-       if (rec->lrh_index < d->lpd_startcat) {
-               /* Skip processing of the logs until startcat */
-               rc = 0;
-       } else if (d->lpd_startidx > 0) {
-                struct llog_process_cat_data cd;
+       RETURN(rc);
+}
+
+static int llog_cat_process_cb(const struct lu_env *env,
+                              struct llog_handle *cat_llh,
+                              struct llog_rec_hdr *rec, void *data)
+{
+       struct llog_process_data *d = data;
+       struct llog_handle *llh = NULL;
+       int rc;
+
+       ENTRY;
+
+       /* Skip processing of the logs until startcat */
+       if (rec->lrh_index < d->lpd_startcat)
+               RETURN(0);
+
+       rc = llog_cat_process_common(env, cat_llh, rec, &llh);
+       if (rc)
+               GOTO(out, rc);
+
+       if (d->lpd_startidx > 0) {
+               struct llog_process_cat_data cd = {
+                       .lpcd_first_idx = 0,
+                       .lpcd_last_idx = 0,
+                       .lpcd_read_mode = LLOG_READ_MODE_NORMAL,
+               };
+
+               /* startidx is always associated with a catalog index */
+               if (d->lpd_startcat == rec->lrh_index)
+                       cd.lpcd_first_idx = d->lpd_startidx;
 
-                cd.lpcd_first_idx = d->lpd_startidx;
-                cd.lpcd_last_idx = 0;
                rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
                                          &cd, false);
                /* Continue processing the next log from idx 0 */
@@ -577,125 +866,234 @@ int llog_cat_process_cb(const struct lu_env *env, struct llog_handle *cat_llh,
                rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
                                          NULL, false);
        }
+       if (rc == -ENOENT && (cat_llh->lgh_hdr->llh_flags & LLOG_F_RM_ON_ERR)) {
+               /*
+                * plain llog is reported corrupted, so better to just remove
+                * it if the caller is fine with that.
+                */
+               CERROR("%s: remove corrupted/missing llog "DFID"\n",
+                      loghandle2name(cat_llh), PLOGID(&llh->lgh_id));
+               rc = LLOG_DEL_PLAIN;
+       }
 
 out:
        /* The empty plain log was destroyed while processing */
-       if (rc == LLOG_DEL_PLAIN)
-               rc = llog_cat_cleanup(env, cat_llh, llh,
-                                     llh->u.phd.phd_cookie.lgc_index);
-       llog_handle_put(llh);
+       if (rc == LLOG_DEL_PLAIN || rc == LLOG_DEL_RECORD)
+               /* clear wrong catalog entry */
+               rc = llog_cat_cleanup(env, cat_llh, llh, rec->lrh_index);
+       else if (rc == LLOG_SKIP_PLAIN)
+               /* processing callback ask to skip the llog -> continue */
+               rc = 0;
+
+       if (llh)
+               llog_handle_put(env, llh);
 
        RETURN(rc);
 }
 
 int llog_cat_process_or_fork(const struct lu_env *env,
-                            struct llog_handle *cat_llh,
+                            struct llog_handle *cat_llh, llog_cb_t cat_cb,
                             llog_cb_t cb, void *data, int startcat,
                             int startidx, bool fork)
 {
-        struct llog_process_data d;
-        struct llog_log_hdr *llh = cat_llh->lgh_hdr;
-        int rc;
-        ENTRY;
+       struct llog_log_hdr *llh = cat_llh->lgh_hdr;
+       struct llog_process_data d;
+       struct llog_process_cat_data cd;
+       int rc;
 
-        LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
-        d.lpd_data = data;
-        d.lpd_cb = cb;
-        d.lpd_startcat = startcat;
-        d.lpd_startidx = startidx;
+       ENTRY;
 
-        if (llh->llh_cat_idx > cat_llh->lgh_last_idx) {
-                struct llog_process_cat_data cd;
+       LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
+       d.lpd_data = data;
+       d.lpd_cb = cb;
+
+       /* default: start from the oldest record */
+       d.lpd_startidx = 0;
+       d.lpd_startcat = llh->llh_cat_idx + 1;
+       cd.lpcd_first_idx = llh->llh_cat_idx;
+       cd.lpcd_last_idx = 0;
+       cd.lpcd_read_mode = LLOG_READ_MODE_NORMAL;
+
+       if (startcat > 0 && startcat <= llog_max_idx(llh)) {
+               /* start from a custom catalog/llog plain indexes*/
+               d.lpd_startidx = startidx;
+               d.lpd_startcat = startcat;
+               cd.lpcd_first_idx = startcat - 1;
+       } else if (startcat != 0) {
+               CWARN("%s: startcat %d out of range for catlog "DFID"\n",
+                     loghandle2name(cat_llh), startcat,
+                     PLOGID(&cat_llh->lgh_id));
+               RETURN(-EINVAL);
+       }
 
-                CWARN("catlog "DOSTID" crosses index zero\n",
-                      POSTID(&cat_llh->lgh_id.lgl_oi));
+       startcat = d.lpd_startcat;
 
-                cd.lpcd_first_idx = llh->llh_cat_idx;
-                cd.lpcd_last_idx = 0;
-               rc = llog_process_or_fork(env, cat_llh, llog_cat_process_cb,
-                                         &d, &cd, fork);
-               if (rc != 0)
+       /* if startcat <= lgh_last_idx, we only need to process the first part
+        * of the catalog (from startcat).
+        */
+       if (llog_cat_is_wrapped(cat_llh) && startcat > cat_llh->lgh_last_idx) {
+               int cat_idx_origin = llh->llh_cat_idx;
+
+               CWARN("%s: catlog "DFID" crosses index zero\n",
+                     loghandle2name(cat_llh),
+                     PLOGID(&cat_llh->lgh_id));
+
+               /* processing the catalog part at the end */
+               rc = llog_process_or_fork(env, cat_llh, cat_cb, &d, &cd, fork);
+               if (rc)
                        RETURN(rc);
 
+               /* Reset the startcat because it has already reached catalog
+                * bottom.
+                * lgh_last_idx value could be increased during processing. So
+                * we process the remaining of catalog entries to be sure.
+                */
+               d.lpd_startcat = 1;
+               d.lpd_startidx = 0;
                cd.lpcd_first_idx = 0;
-               cd.lpcd_last_idx = cat_llh->lgh_last_idx;
-               rc = llog_process_or_fork(env, cat_llh, llog_cat_process_cb,
-                                         &d, &cd, fork);
-        } else {
-               rc = llog_process_or_fork(env, cat_llh, llog_cat_process_cb,
-                                         &d, NULL, fork);
-        }
+               cd.lpcd_last_idx = max(cat_idx_origin, cat_llh->lgh_last_idx);
+       } else if (llog_cat_is_wrapped(cat_llh)) {
+               /* only process 1st part -> stop before reaching 2sd part */
+               cd.lpcd_last_idx = llh->llh_cat_idx;
+       }
 
-        RETURN(rc);
+       /* processing the catalog part at the begining */
+       rc = llog_process_or_fork(env, cat_llh, cat_cb, &d, &cd, fork);
+
+       RETURN(rc);
 }
 EXPORT_SYMBOL(llog_cat_process_or_fork);
 
+/**
+ * Process catalog records with a callback
+ *
+ * \note
+ * If "starcat = 0", this is the default processing. "startidx" argument is
+ * ignored and processing begin from the oldest record.
+ * If "startcat > 0", this is a custom starting point. Processing begin with
+ * the llog plain defined in the catalog record at index "startcat". The first
+ * llog plain record to process is at index "startidx + 1".
+ *
+ * \param env          Lustre environnement
+ * \param cat_llh      Catalog llog handler
+ * \param cb           Callback executed for each records (in llog plain files)
+ * \param data         Callback data argument
+ * \param startcat     Catalog index of the llog plain to start with.
+ * \param startidx     Index of the llog plain to start processing. The first
+ *                     record to process is at startidx + 1.
+ *
+ * \retval 0 processing successfully completed
+ * \retval LLOG_PROC_BREAK processing was stopped by the callback.
+ * \retval -errno on error.
+ */
 int llog_cat_process(const struct lu_env *env, struct llog_handle *cat_llh,
                     llog_cb_t cb, void *data, int startcat, int startidx)
 {
-       return llog_cat_process_or_fork(env, cat_llh, cb, data, startcat,
-                                       startidx, false);
+       return llog_cat_process_or_fork(env, cat_llh, llog_cat_process_cb,
+                                       cb, data, startcat, startidx, false);
 }
 EXPORT_SYMBOL(llog_cat_process);
 
+static int llog_cat_size_cb(const struct lu_env *env,
+                            struct llog_handle *cat_llh,
+                            struct llog_rec_hdr *rec, void *data)
+{
+       struct llog_process_data *d = data;
+       struct llog_handle *llh = NULL;
+       __u64 *cum_size = d->lpd_data;
+       __u64 size;
+       int rc;
+
+       ENTRY;
+       rc = llog_cat_process_common(env, cat_llh, rec, &llh);
+
+       if (rc == LLOG_DEL_PLAIN) {
+               /* empty log was deleted, don't count it */
+               rc = llog_cat_cleanup(env, cat_llh, llh,
+                                     llh->u.phd.phd_cookie.lgc_index);
+       } else if (rc == LLOG_DEL_RECORD) {
+               /* clear wrong catalog entry */
+               rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
+       } else {
+               size = llog_size(env, llh);
+               *cum_size += size;
+
+               CDEBUG(D_INFO, "Add llog entry "DFID" size=%llu, tot=%llu\n",
+                      PLOGID(&llh->lgh_id), size, *cum_size);
+       }
+
+       if (llh != NULL)
+               llog_handle_put(env, llh);
+
+       RETURN(0);
+}
+
+__u64 llog_cat_size(const struct lu_env *env, struct llog_handle *cat_llh)
+{
+       __u64 size = llog_size(env, cat_llh);
+
+       llog_cat_process_or_fork(env, cat_llh, llog_cat_size_cb,
+                                NULL, &size, 0, 0, false);
+
+       return size;
+}
+EXPORT_SYMBOL(llog_cat_size);
+
+/* currently returns the number of "free" entries in catalog,
+ * ie the available entries for a new plain LLOG file creation,
+ * even if catalog has wrapped
+ */
+__u32 llog_cat_free_space(struct llog_handle *cat_llh)
+{
+       /* simulate almost full Catalog */
+       if (CFS_FAIL_CHECK(OBD_FAIL_CAT_FREE_RECORDS))
+               return cfs_fail_val;
+
+       if (cat_llh->lgh_hdr->llh_count == 1)
+               return llog_max_idx(cat_llh->lgh_hdr);
+
+       if (cat_llh->lgh_last_idx > cat_llh->lgh_hdr->llh_cat_idx)
+               return llog_max_idx(cat_llh->lgh_hdr) +
+                      cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
+
+       /* catalog is presently wrapped */
+       return cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
+}
+EXPORT_SYMBOL(llog_cat_free_space);
+
 static int llog_cat_reverse_process_cb(const struct lu_env *env,
                                       struct llog_handle *cat_llh,
                                       struct llog_rec_hdr *rec, void *data)
 {
        struct llog_process_data *d = data;
-       struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
        struct llog_handle *llh;
-       struct llog_log_hdr *hdr;
        int rc;
 
-       if (le32_to_cpu(rec->lrh_type) != LLOG_LOGID_MAGIC) {
-               CERROR("invalid record in catalog\n");
-               RETURN(-EINVAL);
-       }
-       CDEBUG(D_HA, "processing log "DOSTID":%x at index %u of catalog "
-              DOSTID"\n", POSTID(&lir->lid_id.lgl_oi), lir->lid_id.lgl_ogen,
-              le32_to_cpu(rec->lrh_index), POSTID(&cat_llh->lgh_id.lgl_oi));
-
-       rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
-       if (rc) {
-               CERROR("%s: cannot find handle for llog "DOSTID": %d\n",
-                      cat_llh->lgh_ctxt->loc_obd->obd_name,
-                      POSTID(&lir->lid_id.lgl_oi), rc);
-               if (rc == -ENOENT || rc == -ESTALE) {
-                       /* After a server crash, a stub of index
-                        * record in catlog could be kept, because
-                        * plain log destroy + catlog index record
-                        * deletion are not atomic. So we end up with
-                        * an index but no actual record. Destroy the
-                        * index and move on. */
-                       rc = llog_cat_cleanup(env, cat_llh, NULL,
-                                             rec->lrh_index);
-               }
-
-               RETURN(rc);
-       }
+       ENTRY;
+       rc = llog_cat_process_common(env, cat_llh, rec, &llh);
 
-       /* clean old empty llogs, do not consider current llog in use */
-       hdr = llh->lgh_hdr;
-       if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
-           hdr->llh_count == 1 &&
-           llh != cat_llh->u.chd.chd_current_log) {
-               rc = llog_destroy(env, llh);
-               if (rc)
-                       CERROR("%s: fail to destroy empty log: rc = %d\n",
-                              llh->lgh_ctxt->loc_obd->obd_name, rc);
-               GOTO(out, rc = LLOG_DEL_PLAIN);
+       /* The empty plain log was destroyed while processing */
+       if (rc == LLOG_DEL_PLAIN) {
+               rc = llog_cat_cleanup(env, cat_llh, llh,
+                                     llh->u.phd.phd_cookie.lgc_index);
+       } else if (rc == LLOG_DEL_RECORD) {
+               /* clear wrong catalog entry */
+               rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
+       } else if (rc == LLOG_SKIP_PLAIN) {
+               /* processing callback ask to skip the llog -> continue */
+               rc = 0;
        }
+       if (rc)
+               RETURN(rc);
 
        rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
 
-out:
        /* The empty plain was destroyed while processing */
        if (rc == LLOG_DEL_PLAIN)
                rc = llog_cat_cleanup(env, cat_llh, llh,
                                      llh->u.phd.phd_cookie.lgc_index);
 
-       llog_handle_put(llh);
+       llog_handle_put(env, llh);
        RETURN(rc);
 }
 
@@ -710,15 +1108,18 @@ int llog_cat_reverse_process(const struct lu_env *env,
         ENTRY;
 
         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
+       cd.lpcd_read_mode = LLOG_READ_MODE_NORMAL;
         d.lpd_data = data;
         d.lpd_cb = cb;
 
-       if (llh->llh_cat_idx > cat_llh->lgh_last_idx) {
-               CWARN("catalog "DOSTID" crosses index zero\n",
-                     POSTID(&cat_llh->lgh_id.lgl_oi));
+       if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
+           llh->llh_count > 1) {
+               CWARN("%s: catalog "DFID" crosses index zero\n",
+                     loghandle2name(cat_llh),
+                     PLOGID(&cat_llh->lgh_id));
 
-                cd.lpcd_first_idx = 0;
-                cd.lpcd_last_idx = cat_llh->lgh_last_idx;
+               cd.lpcd_first_idx = 0;
+               cd.lpcd_last_idx = cat_llh->lgh_last_idx;
                rc = llog_reverse_process(env, cat_llh,
                                          llog_cat_reverse_process_cb,
                                          &d, &cd);
@@ -740,33 +1141,41 @@ int llog_cat_reverse_process(const struct lu_env *env,
 }
 EXPORT_SYMBOL(llog_cat_reverse_process);
 
-int llog_cat_set_first_idx(struct llog_handle *cathandle, int index)
+static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
 {
-        struct llog_log_hdr *llh = cathandle->lgh_hdr;
-        int i, bitmap_size, idx;
-        ENTRY;
+       struct llog_log_hdr *llh = cathandle->lgh_hdr;
+       int idx_nbr;
 
-        bitmap_size = LLOG_BITMAP_SIZE(llh);
-        if (llh->llh_cat_idx == (index - 1)) {
-                idx = llh->llh_cat_idx + 1;
-                llh->llh_cat_idx = idx;
-                if (idx == cathandle->lgh_last_idx)
-                        goto out;
-                for (i = (index + 1) % bitmap_size;
-                     i != cathandle->lgh_last_idx;
-                     i = (i + 1) % bitmap_size) {
-                        if (!ext2_test_bit(i, llh->llh_bitmap)) {
-                                idx = llh->llh_cat_idx + 1;
-                                llh->llh_cat_idx = idx;
-                        } else if (i == 0) {
-                                llh->llh_cat_idx = 0;
-                        } else {
-                                break;
-                        }
-                }
-out:
-               CDEBUG(D_RPCTRACE, "set catlog "DOSTID" first idx %u\n",
-                      POSTID(&cathandle->lgh_id.lgl_oi), llh->llh_cat_idx);
+       ENTRY;
+
+       idx_nbr = llog_max_idx(llh) + 1;
+       /*
+        * The llh_cat_idx equals to the first used index minus 1
+        * so if we canceled the first index then llh_cat_idx
+        * must be renewed.
+        */
+       if (llh->llh_cat_idx == (idx - 1)) {
+               llh->llh_cat_idx = idx;
+
+               while (idx != cathandle->lgh_last_idx) {
+                       idx = (idx + 1) % idx_nbr;
+                       if (!test_bit_le(idx, LLOG_HDR_BITMAP(llh))) {
+                               /* update llh_cat_idx for each unset bit,
+                                * expecting the next one is set */
+                               llh->llh_cat_idx = idx;
+                       } else if (idx == 0) {
+                               /* skip header bit */
+                               llh->llh_cat_idx = 0;
+                               continue;
+                       } else {
+                               /* the first index is found */
+                               break;
+                       }
+               }
+
+               CDEBUG(D_HA, "catlog "DFID" first idx %u, last_idx %u\n",
+                      PLOGID(&cathandle->lgh_id), llh->llh_cat_idx,
+                      cathandle->lgh_last_idx);
        }
 
        RETURN(0);
@@ -787,79 +1196,48 @@ int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
                        cathandle->u.chd.chd_current_log = NULL;
                list_del_init(&loghandle->u.phd.phd_entry);
                up_write(&cathandle->lgh_lock);
-               LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
+               LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index ||
+                       loghandle->u.phd.phd_cookie.lgc_index == 0);
                /* llog was opened and keep in a list, close it now */
                llog_close(env, loghandle);
        }
+
+       /* do not attempt to cleanup on-disk llog if on client side */
+       if (cathandle->lgh_obj == NULL)
+               return 0;
+
        /* remove plain llog entry from catalog by index */
        llog_cat_set_first_idx(cathandle, index);
        rc = llog_cancel_rec(env, cathandle, index);
-       if (rc == 0)
-               CDEBUG(D_HA, "cancel plain log at index"
-                      " %u of catalog "DOSTID"\n",
-                      index, POSTID(&cathandle->lgh_id.lgl_oi));
+       if (!rc && loghandle)
+               CDEBUG(D_HA,
+                      "cancel plain log "DFID" at index %u of catalog "DFID"\n",
+                      PLOGID(&loghandle->lgh_id), index,
+                      PLOGID(&cathandle->lgh_id));
        return rc;
 }
 
-int cat_cancel_cb(const struct lu_env *env, struct llog_handle *cathandle,
-                 struct llog_rec_hdr *rec, void *data)
+/* retain log in catalog, and zap it if log is empty */
+int llog_cat_retain_cb(const struct lu_env *env, struct llog_handle *cat,
+                      struct llog_rec_hdr *rec, void *data)
 {
-       struct llog_logid_rec   *lir = (struct llog_logid_rec *)rec;
-       struct llog_handle      *loghandle;
-       struct llog_log_hdr     *llh;
-       int                      rc;
-
-       ENTRY;
-
-       if (rec->lrh_type != LLOG_LOGID_MAGIC) {
-               CERROR("invalid record in catalog\n");
-               RETURN(-EINVAL);
-       }
+       struct llog_handle *log = NULL;
+       int rc;
 
-       CDEBUG(D_HA, "processing log "DOSTID":%x at index %u of catalog "
-              DOSTID"\n", POSTID(&lir->lid_id.lgl_oi), lir->lid_id.lgl_ogen,
-              rec->lrh_index, POSTID(&cathandle->lgh_id.lgl_oi));
+       rc = llog_cat_process_common(env, cat, rec, &log);
 
-       rc = llog_cat_id2handle(env, cathandle, &loghandle, &lir->lid_id);
-       if (rc) {
-               CERROR("%s: cannot find handle for llog "DOSTID": %d\n",
-                      cathandle->lgh_ctxt->loc_obd->obd_name,
-                      POSTID(&lir->lid_id.lgl_oi), rc);
-               if (rc == -ENOENT || rc == -ESTALE) {
-                       /* remove index from catalog */
-                       llog_cat_cleanup(env, cathandle, NULL, rec->lrh_index);
-               }
-               RETURN(rc);
-       }
-
-       llh = loghandle->lgh_hdr;
-       if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
-           (llh->llh_count == 1)) {
-               rc = llog_destroy(env, loghandle);
-               if (rc)
-                       CERROR("%s: fail to destroy empty log: rc = %d\n",
-                              loghandle->lgh_ctxt->loc_obd->obd_name, rc);
+       /* The empty plain log was destroyed while processing */
+       if (rc == LLOG_DEL_PLAIN || rc == LLOG_DEL_RECORD)
+               /* clear wrong catalog entry */
+               rc = llog_cat_cleanup(env, cat, log, rec->lrh_index);
+       else if (!rc)
+               llog_retain(env, log);
 
-               llog_cat_cleanup(env, cathandle, loghandle,
-                                loghandle->u.phd.phd_cookie.lgc_index);
-       }
-       llog_handle_put(loghandle);
+       if (log)
+               llog_handle_put(env, log);
 
-       RETURN(rc);
+       return rc;
 }
-EXPORT_SYMBOL(cat_cancel_cb);
-
-/* helper to initialize catalog llog and process it to cancel */
-int llog_cat_init_and_process(const struct lu_env *env,
-                             struct llog_handle *llh)
-{
-       int rc;
+EXPORT_SYMBOL(llog_cat_retain_cb);
 
-       rc = llog_init_handle(env, llh, LLOG_F_IS_CAT, NULL);
-       if (rc)
-               RETURN(rc);
-
-       RETURN(0);
-}
-EXPORT_SYMBOL(llog_cat_init_and_process);