cluster/dht: use readdir for fix-layout in rebalance

fixes: #2241
Change-Id: I5fe2ecea25a399ad58e31a2e322caf69fc7f49eb
Signed-off-by: Pranith Kumar K <[email protected]>

cli/src/cli-rpc-ops.c

@@ -1576,7 +1576,7 @@ gf_cli_print_rebalance_status(dict_t *dict, enum gf_task_types task_type)
         sec = ((uint64_t)elapsed % 3600) % 60;
 
         if (fix_layout) {
-            cli_out("%35s %50s %8d:%d:%d", node_name, status_str, hrs, min,
+            cli_out("%35s %50s %8d:%02d:%02d", node_name, status_str, hrs, min,
                     sec);
         } else {
             if (size_str) {

xlators/cluster/dht/src/dht-common.c

@@ -6649,8 +6649,10 @@ dht_queue_readdir(call_frame_t *frame, xlator_t *xl, off_t offset,
 {
     dht_local_t *local;
     int32_t queue;
+    xlator_t *this;
 
     local = frame->local;
+    this = frame->this;
 
     local->queue_xl = xl;
     local->queue_offset = offset;
@@ -6677,7 +6679,7 @@ dht_queue_readdir(call_frame_t *frame, xlator_t *xl, off_t offset,
             /* A negative value means that an unwind has been called before
              * returning from the previous wind. This means that 'local' is
              * not needed anymore and must be destroyed. */
-            dht_local_wipe(frame->this, local);
+            dht_local_wipe(this, local);
         }
     }
 }
@@ -6690,8 +6692,10 @@ dht_queue_readdirp(call_frame_t *frame, xlator_t *xl, off_t offset,
 {
     dht_local_t *local;
     int32_t queue;
+    xlator_t *this;
 
     local = frame->local;
+    this = frame->this;
 
     local->queue_xl = xl;
     local->queue_offset = offset;
@@ -6705,7 +6709,7 @@ dht_queue_readdirp(call_frame_t *frame, xlator_t *xl, off_t offset,
         } while ((queue = uatomic_sub_return(&local->queue, 1)) > 0);
 
         if (queue < 0) {
-            dht_local_wipe(frame->this, local);
+            dht_local_wipe(this, local);
         }
     }
 }
@@ -7030,6 +7034,8 @@ dht_readdir_cbk(call_frame_t *frame, void *cookie, xlator_t *this, int op_ret,
     dht_conf_t *conf = NULL;
     dht_methods_t *methods = NULL;
     gf_boolean_t skip_hashed_check = _gf_false;
+    gf_boolean_t readdir_optimize = _gf_false;
+    gf_boolean_t add = _gf_false;
 
     INIT_LIST_HEAD(&entries.list);
 
@@ -7039,6 +7045,7 @@ dht_readdir_cbk(call_frame_t *frame, void *cookie, xlator_t *this, int op_ret,
     conf = this->private;
     GF_VALIDATE_OR_GOTO(this->name, conf, done);
 
+    readdir_optimize = conf->readdir_optimize;
     methods = &(conf->methods);
 
     if (op_ret <= 0)
@@ -7062,12 +7069,24 @@ dht_readdir_cbk(call_frame_t *frame, void *cookie, xlator_t *this, int op_ret,
     {
         next_offset = orig_entry->d_off;
 
-        gf_msg_debug(this->name, 0, "%s: entry = %s, type = %d", prev->name,
-                     orig_entry->d_name, orig_entry->d_type);
-
         subvol = methods->layout_search(this, layout, orig_entry->d_name);
+        gf_log(this->name, GF_LOG_INFO, "%s: entry = %s, type = %d %p, %p", prev->name,
+               orig_entry->d_name, orig_entry->d_type, subvol, prev);
+
+        /* a) If rebalance is running, pick from first_up_subvol
+         */
+        if (DT_ISDIR(orig_entry->d_type) && readdir_optimize) {
+            if (prev == local->first_up_subvol) {
+                add = _gf_true;
+            } else {
+                continue;
+            }
+        } else if (!subvol || (subvol == prev)) {
+            add = _gf_true;
+        }
 
-        if (!subvol || (subvol == prev)) {
+        if (add) {
+            add = _gf_false;
             entry = gf_dirent_for_name(orig_entry->d_name);
             if (!entry) {
                 gf_msg(this->name, GF_LOG_ERROR, ENOMEM, DHT_MSG_NO_MEMORY,
@@ -7080,7 +7099,7 @@ dht_readdir_cbk(call_frame_t *frame, void *cookie, xlator_t *this, int op_ret,
             entry->d_type = orig_entry->d_type;
             entry->d_len = orig_entry->d_len;
 
-            gf_msg_debug(this->name, 0, "%s: Adding = entry %s", prev->name,
+            gf_log(this->name, GF_LOG_INFO, "%s: Adding = entry %s", prev->name,
                          entry->d_name);
 
             list_add_tail(&entry->list, &entries.list);

xlators/cluster/dht/src/dht-rebalance.c

@@ -3639,6 +3639,12 @@ gf_defrag_fix_layout(xlator_t *this, gf_defrag_info_t *defrag, loc_t *loc,
         goto out;
     }
 
+    linked_inode = inode_link(loc->inode, loc->parent, loc->name, &iatt);
+
+    inode = loc->inode;
+    loc->inode = linked_inode;
+    inode_unref(inode);
+
     fd = fd_create(loc->inode, defrag->pid);
     if (!fd) {
         gf_log(this->name, GF_LOG_ERROR, "Failed to create fd");
@@ -3671,8 +3677,8 @@ gf_defrag_fix_layout(xlator_t *this, gf_defrag_info_t *defrag, loc_t *loc,
     fd_bind(fd);
     INIT_LIST_HEAD(&entries.list);
 
-    while ((ret = syncop_readdirp(this, fd, 131072, offset, &entries, NULL,
-                                  NULL)) != 0) {
+    while ((ret = syncop_readdir(this, fd, 131072, offset, &entries, NULL,
+                                 NULL)) != 0) {
         if (ret < 0) {
             if (-ret == ENOENT || -ret == ESTALE) {
                 if (conf->decommission_subvols_cnt) {
@@ -3703,11 +3709,14 @@ gf_defrag_fix_layout(xlator_t *this, gf_defrag_info_t *defrag, loc_t *loc,
                 goto out;
             }
 
+            if (__is_root_gfid (fd->inode->gfid)) {
+                gf_log(this->name, GF_LOG_INFO, "Entry: %s", entry->d_name);
+            }
             offset = entry->d_off;
 
             if (!strcmp(entry->d_name, ".") || !strcmp(entry->d_name, ".."))
                 continue;
-            if (!IA_ISDIR(entry->d_stat.ia_type)) {
+            if (!DT_ISDIR(entry->d_type)) {
                 continue;
             }
             loc_wipe(&entry_loc);
@@ -3730,40 +3739,6 @@ gf_defrag_fix_layout(xlator_t *this, gf_defrag_info_t *defrag, loc_t *loc,
                 }
             }
 
-            if (gf_uuid_is_null(entry->d_stat.ia_gfid)) {
-                gf_log(this->name, GF_LOG_ERROR,
-                       "%s/%s"
-                       " gfid not present",
-                       loc->path, entry->d_name);
-                continue;
-            }
-
-            gf_uuid_copy(entry_loc.gfid, entry->d_stat.ia_gfid);
-
-            /*In case the gfid stored in the inode by inode_link
-             * and the gfid obtained in the lookup differs, then
-             * client3_3_lookup_cbk will return ESTALE and proper
-             * error will be captured
-             */
-
-            linked_inode = inode_link(entry_loc.inode, loc->inode,
-                                      entry->d_name, &entry->d_stat);
-
-            inode = entry_loc.inode;
-            entry_loc.inode = linked_inode;
-            inode_unref(inode);
-
-            if (gf_uuid_is_null(loc->gfid)) {
-                gf_log(this->name, GF_LOG_ERROR,
-                       "%s/%s"
-                       " gfid not present",
-                       loc->path, entry->d_name);
-                defrag->total_failures++;
-                continue;
-            }
-
-            gf_uuid_copy(entry_loc.pargfid, loc->gfid);
-
             /* A return value of 2 means, either process_dir or
              * lookup of a dir failed. Hence, don't commit hash
              * for the current directory*/

xlators/mgmt/glusterd/src/glusterd-rebalance.c

@@ -289,7 +289,7 @@ glusterd_handle_defrag_start(glusterd_volinfo_t *volinfo, char *op_errstr,
 
     runner_add_args(
         &runner, SBIN_DIR "/glusterfs", "-s", volfileserver, "--volfile-id",
-        volname, "--xlator-option", "*dht.use-readdirp=yes", "--xlator-option",
+        volname, "--xlator-option", "*dht.use-readdirp=no", "--xlator-option",
         "*dht.lookup-unhashed=yes", "--xlator-option",
         "*dht.assert-no-child-down=yes", "--xlator-option",
         "*dht.readdir-optimize=on", "--process-name", "rebalance", NULL);