Whamcloud - gitweb
LU-6838 llog: limit file size of plain logs
[fs/lustre-release.git] / lustre / obdclass / llog_cat.c
index e5cafa7..86a2047 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
  */
@@ -27,7 +23,7 @@
  * 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, 2015, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -48,9 +44,6 @@
 
 #define DEBUG_SUBSYSTEM S_LOG
 
-#ifndef __KERNEL__
-#include <liblustre.h>
-#endif
 
 #include <obd_class.h>
 
@@ -68,28 +61,97 @@ 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;
 
+       index = (cathandle->lgh_last_idx + 1) %
+               (OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ? (cfs_fail_val + 1) :
+                                               LLOG_HDR_BITMAP_SIZE(llh));
+
+       /* 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)) {
+               CWARN("%s: there are no more free slots in catalog\n",
+                     loghandle->lgh_ctxt->loc_obd->obd_name);
+               RETURN(-ENOSPC);
+       }
+
        if (OBD_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);
+               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);
@@ -103,15 +165,47 @@ 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"
+       CDEBUG(D_OTHER, "new plain 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));
 
        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;
+       }
+       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_HDR_BITMAP_SIZE(llh) + 1;
+       llog_trans_destroy(env, loghandle, th);
        RETURN(rc);
 }
 
@@ -128,6 +222,7 @@ 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;
@@ -135,9 +230,10 @@ 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);
-       cfs_list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
-                               u.phd.phd_entry) {
+       list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
+                           u.phd.phd_entry) {
                struct llog_logid *cgl = &loghandle->lgh_id;
 
                if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
@@ -165,7 +261,7 @@ int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
                RETURN(rc);
        }
 
-       rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN, NULL);
+       rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN | fmt, NULL);
        if (rc < 0) {
                llog_close(env, loghandle);
                loghandle = NULL;
@@ -173,7 +269,7 @@ int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
        }
 
        down_write(&cathandle->lgh_lock);
-       cfs_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;
@@ -194,13 +290,13 @@ int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
 
        ENTRY;
 
-       cfs_list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
-                                    u.phd.phd_entry) {
+       list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
+                                u.phd.phd_entry) {
                struct llog_log_hdr     *llh = loghandle->lgh_hdr;
                int                      index;
 
                /* unlink open-not-created llogs */
-               cfs_list_del_init(&loghandle->u.phd.phd_entry);
+               list_del_init(&loghandle->u.phd.phd_entry);
                llh = loghandle->lgh_hdr;
                if (loghandle->lgh_obj != NULL && llh != NULL &&
                    (llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
@@ -249,6 +345,12 @@ static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
         struct llog_handle *loghandle = NULL;
         ENTRY;
 
+
+       if (OBD_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 +358,7 @@ 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 {
@@ -277,13 +378,31 @@ 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 (!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",
+                      cathandle->lgh_ctxt->loc_obd->obd_name);
+               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");
 
@@ -291,11 +410,47 @@ 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);
 }
 
+static int llog_cat_update_header(const struct lu_env *env,
+                          struct llog_handle *cathandle)
+{
+       struct llog_handle *loghandle;
+       int rc;
+       ENTRY;
+
+       /* refresh llog */
+       down_write(&cathandle->lgh_lock);
+       if (!cathandle->lgh_stale) {
+               up_write(&cathandle->lgh_lock);
+               RETURN(0);
+       }
+       list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
+                           u.phd.phd_entry) {
+               if (!llog_exist(loghandle))
+                       continue;
+
+               rc = llog_read_header(env, loghandle, NULL);
+               if (rc != 0) {
+                       up_write(&cathandle->lgh_lock);
+                       GOTO(out, rc);
+               }
+       }
+       rc = llog_read_header(env, cathandle, NULL);
+       if (rc == 0)
+               cathandle->lgh_stale = 0;
+       up_write(&cathandle->lgh_lock);
+       if (rc != 0)
+               GOTO(out, rc);
+out:
+       RETURN(rc);
+}
+
 /* Add a single record to the recovery log(s) using a catalog
  * Returns as llog_write_record
  *
@@ -306,45 +461,50 @@ 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",
+                      cathandle->lgh_ctxt->loc_obd->obd_name, rc);
        }
 
        RETURN(rc);
@@ -370,21 +530,23 @@ int llog_cat_declare_add_rec(const struct lu_env *env,
                                       NULL, NULL, LLOG_OPEN_NEW);
                        if (rc == 0) {
                                cathandle->u.chd.chd_current_log = loghandle;
-                               cfs_list_add_tail(&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);
-       } else if (cathandle->u.chd.chd_next_log == NULL) {
+       } else if (cathandle->u.chd.chd_next_log == NULL ||
+                  IS_ERR(cathandle->u.chd.chd_next_log)) {
                /* declare next plain llog */
                down_write(&cathandle->lgh_lock);
-               if (cathandle->u.chd.chd_next_log == NULL) {
+               if (cathandle->u.chd.chd_next_log == NULL ||
+                   IS_ERR(cathandle->u.chd.chd_next_log)) {
                        rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
                                       NULL, NULL, LLOG_OPEN_NEW);
                        if (rc == 0) {
                                cathandle->u.chd.chd_next_log = loghandle;
-                               cfs_list_add_tail(&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);
@@ -395,24 +557,117 @@ int llog_cat_declare_add_rec(const struct lu_env *env,
        lirec->lid_hdr.lrh_len = sizeof(*lirec);
 
        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);
+               if (dt_object_remote(cathandle->lgh_obj)) {
+                       /* 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. */
+create_again:
+                       down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
+                       loghandle = cathandle->u.chd.chd_current_log;
+                       down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
+                       if (cathandle->lgh_stale) {
+                               up_write(&loghandle->lgh_lock);
+                               up_read(&cathandle->lgh_lock);
+                               GOTO(out, rc = -EIO);
+                       }
+                       if (!llog_exist(loghandle)) {
+                               rc = llog_cat_new_log(env, cathandle, loghandle,
+                                                     NULL);
+                               if (rc == -ESTALE)
+                                       cathandle->lgh_stale = 1;
+                       }
+                       up_write(&loghandle->lgh_lock);
+                       up_read(&cathandle->lgh_lock);
+                       if (rc == -ESTALE) {
+                               rc = llog_cat_update_header(env, cathandle);
+                               if (rc != 0)
+                                       GOTO(out, rc);
+                               goto create_again;
+                       } else if (rc < 0) {
+                               GOTO(out, rc);
+                       }
+               } else {
+                       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);
+               }
        }
+
+write_again:
        /* declare records in the llogs */
        rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
                                    rec, -1, th);
-       if (rc)
+       if (rc == -ESTALE) {
+               down_write(&cathandle->lgh_lock);
+               if (cathandle->lgh_stale) {
+                       up_write(&cathandle->lgh_lock);
+                       GOTO(out, rc = -EIO);
+               }
+
+               cathandle->lgh_stale = 1;
+               up_write(&cathandle->lgh_lock);
+               rc = llog_cat_update_header(env, cathandle);
+               if (rc != 0)
+                       GOTO(out, rc);
+               goto write_again;
+       } else if (rc < 0) {
                GOTO(out, rc);
+       }
 
        next = cathandle->u.chd.chd_next_log;
-       if (next) {
+       if (!IS_ERR_OR_NULL(next)) {
                if (!llog_exist(next)) {
-                       rc = llog_declare_create(env, next, th);
-                       llog_declare_write_rec(env, cathandle, &lirec->lid_hdr,
-                                              -1, th);
+                       if (dt_object_remote(cathandle->lgh_obj)) {
+                               /* 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. */
+                               down_write_nested(&cathandle->lgh_lock,
+                                                LLOGH_CAT);
+                               next = cathandle->u.chd.chd_next_log;
+                               if (IS_ERR_OR_NULL(next)) {
+                                       /* Sigh, another thread just tried,
+                                        * let's fail as well */
+                                       up_write(&cathandle->lgh_lock);
+                                       if (next == NULL)
+                                               rc = -EIO;
+                                       else
+                                               rc = PTR_ERR(next);
+                                       GOTO(out, rc);
+                               }
+
+                               down_write_nested(&next->lgh_lock, LLOGH_LOG);
+                               if (!llog_exist(next)) {
+                                       rc = llog_cat_new_log(env, cathandle,
+                                                             next, NULL);
+                                       if (rc < 0)
+                                               cathandle->u.chd.chd_next_log =
+                                                               ERR_PTR(rc);
+                               }
+                               up_write(&next->lgh_lock);
+                               up_write(&cathandle->lgh_lock);
+                               if (rc < 0)
+                                       GOTO(out, rc);
+                       } else {
+                               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
@@ -512,12 +767,14 @@ int llog_cat_cancel_records(const struct lu_env *env,
 }
 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_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;
 
         ENTRY;
@@ -534,9 +791,33 @@ int llog_cat_process_cb(const struct lu_env *env, struct llog_handle *cat_llh,
                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);
        }
 
+       /* clean old empty llogs, do not consider current llog in use */
+       /* ignore remote (lgh_obj=NULL) llogs */
+       hdr = llh->lgh_hdr;
+       if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
+           hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
+           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);
+       }
+
        if (rec->lrh_index < d->lpd_startcat) {
                /* Skip processing of the logs until startcat */
                rc = 0;
@@ -554,6 +835,7 @@ int llog_cat_process_cb(const struct lu_env *env, struct llog_handle *cat_llh,
                                          NULL, false);
        }
 
+out:
        /* The empty plain log was destroyed while processing */
        if (rc == LLOG_DEL_PLAIN)
                rc = llog_cat_cleanup(env, cat_llh, llh,
@@ -564,7 +846,7 @@ int llog_cat_process_cb(const struct lu_env *env, struct llog_handle *cat_llh,
 }
 
 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)
 {
@@ -579,7 +861,8 @@ int llog_cat_process_or_fork(const struct lu_env *env,
         d.lpd_startcat = startcat;
         d.lpd_startidx = startidx;
 
-        if (llh->llh_cat_idx > cat_llh->lgh_last_idx) {
+       if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
+           llh->llh_count > 1) {
                 struct llog_process_cat_data cd;
 
                 CWARN("catlog "DOSTID" crosses index zero\n",
@@ -587,32 +870,82 @@ int llog_cat_process_or_fork(const struct lu_env *env,
 
                 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,
+               rc = llog_process_or_fork(env, cat_llh, cat_cb,
                                          &d, &cd, fork);
                if (rc != 0)
                        RETURN(rc);
 
                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,
+               rc = llog_process_or_fork(env, cat_llh, cat_cb,
                                          &d, &cd, fork);
         } else {
-               rc = llog_process_or_fork(env, cat_llh, llog_cat_process_cb,
+               rc = llog_process_or_fork(env, cat_llh, cat_cb,
                                          &d, NULL, fork);
         }
 
         RETURN(rc);
 }
-EXPORT_SYMBOL(llog_cat_process_or_fork);
 
 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_logid_rec *lir = (struct llog_logid_rec *)rec;
+       struct llog_handle *llh;
+       int rc;
+       __u64 *cum_size = d->lpd_data;
+       __u64 size;
+
+       ENTRY;
+       if (rec->lrh_type != LLOG_LOGID_MAGIC) {
+               CERROR("%s: invalid record in catalog, rc = %d\n",
+                      cat_llh->lgh_ctxt->loc_obd->obd_name, -EINVAL);
+               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));
+
+       rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
+       if (rc) {
+               CWARN("%s: cannot find handle for llog "DOSTID": rc = %d\n",
+                     cat_llh->lgh_ctxt->loc_obd->obd_name,
+                     POSTID(&lir->lid_id.lgl_oi), rc);
+               RETURN(0);
+       }
+       size = llog_size(env, llh);
+       *cum_size += size;
+
+       CDEBUG(D_INFO, "Add llog entry "DOSTID" size %llu\n",
+              POSTID(&llh->lgh_id.lgl_oi), size);
+
+       llog_handle_put(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);
+
 static int llog_cat_reverse_process_cb(const struct lu_env *env,
                                       struct llog_handle *cat_llh,
                                       struct llog_rec_hdr *rec, void *data)
@@ -620,6 +953,7 @@ static int llog_cat_reverse_process_cb(const struct lu_env *env,
        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) {
@@ -635,11 +969,35 @@ static int llog_cat_reverse_process_cb(const struct lu_env *env,
                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);
        }
 
+       /* 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);
+       }
+
        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,
@@ -663,7 +1021,8 @@ int llog_cat_reverse_process(const struct lu_env *env,
         d.lpd_data = data;
         d.lpd_cb = cb;
 
-       if (llh->llh_cat_idx > cat_llh->lgh_last_idx) {
+       if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
+           llh->llh_count > 1) {
                CWARN("catalog "DOSTID" crosses index zero\n",
                      POSTID(&cat_llh->lgh_id.lgl_oi));
 
@@ -690,33 +1049,41 @@ 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 bitmap_size;
 
-        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;
+
+       bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
+       /*
+        * 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) % bitmap_size;
+                       if (!ext2_test_bit(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_RPCTRACE, "Set catlog "DOSTID" first idx %u,"
+                      " (last_idx %u)\n", POSTID(&cathandle->lgh_id.lgl_oi),
+                      llh->llh_cat_idx, cathandle->lgh_last_idx);
        }
 
        RETURN(0);
@@ -735,12 +1102,17 @@ int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
                down_write(&cathandle->lgh_lock);
                if (cathandle->u.chd.chd_current_log == loghandle)
                        cathandle->u.chd.chd_current_log = NULL;
-               cfs_list_del_init(&loghandle->u.phd.phd_entry);
+               list_del_init(&loghandle->u.phd.phd_entry);
                up_write(&cathandle->lgh_lock);
                LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
                /* 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);
@@ -750,70 +1122,3 @@ int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
                       index, POSTID(&cathandle->lgh_id.lgl_oi));
        return rc;
 }
-
-int cat_cancel_cb(const struct lu_env *env, struct llog_handle *cathandle,
-                 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);
-       }
-
-       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_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);
-
-               llog_cat_cleanup(env, cathandle, loghandle,
-                                loghandle->u.phd.phd_cookie.lgc_index);
-       }
-       llog_handle_put(loghandle);
-
-       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;
-
-       rc = llog_init_handle(env, llh, LLOG_F_IS_CAT, NULL);
-       if (rc)
-               RETURN(rc);
-
-       rc = llog_process_or_fork(env, llh, cat_cancel_cb, NULL, NULL, false);
-       if (rc)
-               CERROR("%s: llog_process() with cat_cancel_cb failed: rc = "
-                      "%d\n", llh->lgh_ctxt->loc_obd->obd_name, rc);
-       RETURN(0);
-}
-EXPORT_SYMBOL(llog_cat_init_and_process);
-