]> sourceware.org Git - lvm2.git/commitdiff
Fix pvmove abort when temporary mirror fails to be cluster-aware.
authorMilan Broz <mbroz@redhat.com>
Wed, 27 Jan 2010 13:29:11 +0000 (13:29 +0000)
committerMilan Broz <mbroz@redhat.com>
Wed, 27 Jan 2010 13:29:11 +0000 (13:29 +0000)
When activation of pvmove mirror fails on cluster, some nodes
still possibly succeeded in activation.

 - Explicitly deactivate that mirror to be sure
 - properly pair suspend/resume calls to not cause memory lock problems in clvmd

Code cannot simply call _finish_pvmove on cluster in this situation, because
changed LVs are suspended twice (causing memory inbalance) and also temporary
mirror is activated when it is not expected (and we know that it failed already).

Patch prepares special function which remove temporary mirror references from
metadata and then resumes changed LVs.

WHATS_NEW
tools/pvmove.c

index 55d329ebc4a937172a0f69c9a1fa322de18f0ef7..af4c5ec22bb90a1db0792d98f667511e6aa62825 100644 (file)
--- a/WHATS_NEW
+++ b/WHATS_NEW
@@ -1,5 +1,6 @@
 Version 2.02.60 - 23rd January 2010
 ===================================
+  Fix pvmove abort workaround to be cluster-aware when temporary mirror activation fails.
   Always query device by using uuid only and not name in clvmd.
   Add missing metadata reverts in pvmove error path.
   Unlock shared lock in clvmd if activation calls failed.
index a286db16d7bb90da106057f223d5668e15d6f89f..b91d153472bebde7a09c683e7b4da0a8db6cc25a 100644 (file)
@@ -1,6 +1,6 @@
 /*
  * Copyright (C) 2003-2004 Sistina Software, Inc. All rights reserved.
- * Copyright (C) 2004-2007 Red Hat, Inc. All rights reserved.
+ * Copyright (C) 2004-2010 Red Hat, Inc. All rights reserved.
  *
  * This file is part of LVM2.
  *
@@ -284,9 +284,26 @@ static int _activate_lv(struct cmd_context *cmd, struct logical_volume *lv_mirr,
        return r;
 }
 
-static int _finish_pvmove(struct cmd_context *cmd, struct volume_group *vg,
-                         struct logical_volume *lv_mirr,
-                         struct dm_list *lvs_changed);
+static int _detach_pvmove_mirror(struct cmd_context *cmd,
+                                struct logical_volume *lv_mirr)
+{
+       struct dm_list lvs_completed;
+       struct lv_list *lvl;
+
+       /* Update metadata to remove mirror segments and break dependencies */
+       dm_list_init(&lvs_completed);
+       if (!lv_remove_mirrors(cmd, lv_mirr, 1, 0, NULL, PVMOVE) ||
+           !remove_layers_for_segments_all(cmd, lv_mirr, PVMOVE,
+                                           &lvs_completed)) {
+               return 0;
+       }
+
+       dm_list_iterate_items(lvl, &lvs_completed)
+               /* FIXME Assumes only one pvmove at a time! */
+               lvl->lv->status &= ~LOCKED;
+
+       return 1;
+}
 
 static int _update_metadata(struct cmd_context *cmd, struct volume_group *vg,
                            struct logical_volume *lv_mirr,
@@ -341,12 +358,27 @@ static int _update_metadata(struct cmd_context *cmd, struct volume_group *vg,
                        }
 
                        /*
+                        * FIXME: review ordering of operations above,
+                        * temporary mirror should be preloaded in suspend.
+                        * Also banned operation here when suspended.
                         * Nothing changed yet, try to revert pvmove.
                         */
                        log_error("Temporary pvmove mirror activation failed.");
-                       if (!_finish_pvmove(cmd, vg, lv_mirr, lvs_changed))
+
+                       /* Ensure that temporary mrror is deactivate even on other nodes. */
+                       (void)deactivate_lv(cmd, lv_mirr);
+
+                       /* Revert metadata */
+                       if (!_detach_pvmove_mirror(cmd, lv_mirr) ||
+                           !lv_remove(lv_mirr) ||
+                           !vg_write(vg) || !vg_commit(vg))
                                log_error("ABORTING: Restoring original configuration "
                                          "before pvmove failed. Run pvmove --abort.");
+
+                       /* Unsuspend LVs */
+                       if(!resume_lvs(cmd, lvs_changed))
+                               stack;
+
                        goto out;
                }
        } else if (!resume_lv(cmd, lv_mirr)) {
@@ -487,22 +519,12 @@ static int _finish_pvmove(struct cmd_context *cmd, struct volume_group *vg,
                          struct dm_list *lvs_changed)
 {
        int r = 1;
-       struct dm_list lvs_completed;
-       struct lv_list *lvl;
 
-       /* Update metadata to remove mirror segments and break dependencies */
-       dm_list_init(&lvs_completed);
-       if (!lv_remove_mirrors(cmd, lv_mirr, 1, 0, NULL, PVMOVE) ||
-           !remove_layers_for_segments_all(cmd, lv_mirr, PVMOVE,
-                                           &lvs_completed)) {
+       if (!_detach_pvmove_mirror(cmd, lv_mirr)) {
                log_error("ABORTING: Removal of temporary mirror failed");
                return 0;
        }
 
-       dm_list_iterate_items(lvl, &lvs_completed)
-               /* FIXME Assumes only one pvmove at a time! */
-               lvl->lv->status &= ~LOCKED;
-
        /* Store metadata without dependencies on mirror segments */
        if (!vg_write(vg)) {
                log_error("ABORTING: Failed to write new data locations "
This page took 0.047427 seconds and 5 git commands to generate.