LU-8359 ldlm: Wrong evict during failover

There is a race between setting obd_fail & OBD_OPT_FAILOVER.
tgt_client_del() checks only OBD_OPT_FAILOVER,
class_disconnect_export_list() is called with flags copied
from obd, and umount can start while disconnect
is in progress.

It is better to rely only on obd_fail.
We shouldn't evict during failover at all, it should
be handled on a new server.
Such wrong evict can happen when server can't send CP AST
to the client because failover has started already.

Change-Id: I649d35d180b2239fe558b375872d3805629968a9
Seagate-bug-id: MRP-3604
Signed-off-by: Andriy Skulysh <andriy.skulysh@seagate.com>
Reviewed-on: https://review.whamcloud.com/21114
Tested-by: Jenkins
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Ben Evans <bevans@cray.com>
Reviewed-by: Alexander Zarochentsev <alexander.zarochentsev@seagate.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>

diff --git a/lustre/include/lustre_export.h b/lustre/include/lustre_export.h
index 7299b68..7aa1dff 100644 (file)
--- a/lustre/include/lustre_export.h
+++ b/lustre/include/lustre_export.h
@@ -164,7 +164,6 @@ do {                                                                           \
 
 enum obd_option {
         OBD_OPT_FORCE =         0x0001,
-        OBD_OPT_FAILOVER =      0x0002,
         OBD_OPT_ABORT_RECOV =   0x0004,
 };
 

diff --git a/lustre/include/obd_class.h b/lustre/include/obd_class.h
index 991586a..0fafed0 100644 (file)
--- a/lustre/include/obd_class.h
+++ b/lustre/include/obd_class.h
@@ -317,8 +317,7 @@ void class_disconnect_stale_exports(struct obd_device *,
                                     int (*test_export)(struct obd_export *));
 static inline enum obd_option exp_flags_from_obd(struct obd_device *obd)
 {
-        return ((obd->obd_fail ? OBD_OPT_FAILOVER : 0) |
-                (obd->obd_force ? OBD_OPT_FORCE : 0) |
+        return ((obd->obd_force ? OBD_OPT_FORCE : 0) |
                 (obd->obd_abort_recovery ? OBD_OPT_ABORT_RECOV : 0) |
                 0);
 }

diff --git a/lustre/mdt/mdt_handler.c b/lustre/mdt/mdt_handler.c
index 20a2f8e..e281117 100644 (file)
--- a/lustre/mdt/mdt_handler.c
+++ b/lustre/mdt/mdt_handler.c
@@ -5502,7 +5502,7 @@ static int mdt_export_cleanup(struct obd_export *exp)
                                rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
 
                        /* Don't unlink orphan on failover umount, LU-184 */
-                       if (exp->exp_flags & OBD_OPT_FAILOVER) {
+                       if (exp->exp_obd->obd_fail) {
                                ma->ma_valid = MA_FLAGS;
                                ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
                        }
@@ -5511,9 +5511,7 @@ static int mdt_export_cleanup(struct obd_export *exp)
         }
         info->mti_mdt = NULL;
         /* cleanup client slot early */
-        /* Do not erase record for recoverable client. */
-        if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
-               tgt_client_del(&env, exp);
+       tgt_client_del(&env, exp);
         lu_env_fini(&env);
 
         RETURN(rc);

diff --git a/lustre/ofd/ofd_obd.c b/lustre/ofd/ofd_obd.c
index 33d6b4f..ebd74a8 100644 (file)
--- a/lustre/ofd/ofd_obd.c
+++ b/lustre/ofd/ofd_obd.c
@@ -436,9 +436,7 @@ int ofd_obd_disconnect(struct obd_export *exp)
 
        tgt_grant_discard(exp);
 
-       /* Do not erase record for recoverable client. */
-       if (exp->exp_obd->obd_replayable &&
-           (!exp->exp_obd->obd_fail || exp->exp_failed)) {
+       if (exp->exp_obd->obd_replayable) {
                rc = lu_env_init(&env, LCT_DT_THREAD);
                if (rc)
                        GOTO(out, rc);

diff --git a/lustre/target/tgt_lastrcvd.c b/lustre/target/tgt_lastrcvd.c
index c7aecdf..3a55026 100644 (file)
--- a/lustre/target/tgt_lastrcvd.c
+++ b/lustre/target/tgt_lastrcvd.c
@@ -1079,6 +1079,10 @@ int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
                RETURN(-EINVAL);
        }
 
+       /* Do not erase record for recoverable client. */
+       if (exp->exp_obd->obd_fail)
+               RETURN(0);
+
        /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
        if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
                    (char *)tgt->lut_obd->obd_uuid.uuid) ||
@@ -1105,9 +1109,6 @@ int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
                LBUG();
        }
 
-       /* Do not erase record for recoverable client. */
-       if (exp->exp_flags & OBD_OPT_FAILOVER)
-               RETURN(0);
 
        if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_DEL))
                RETURN(0);