struct ptlrpc_connect_async_args {
__u64 pcaa_peer_committed;
int pcaa_initial_connect;
- int pcaa_was_invalid;
};
/* A CLOSED import should remain so. */
int ptlrpc_init_import(struct obd_import *imp)
{
unsigned long flags;
-
+
spin_lock_irqsave(&imp->imp_lock, flags);
imp->imp_generation++;
{
unsigned long flags;
int rc = 0;
-
+
spin_lock_irqsave(&imp->imp_lock, flags);
if (imp->imp_state == LUSTRE_IMP_FULL) {
IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
- spin_unlock_irqrestore(&imp->imp_lock, flags);
+ spin_unlock_irqrestore(&imp->imp_lock, flags);
obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
rc = 1;
} else {
spin_unlock_irqrestore(&imp->imp_lock, flags);
CDEBUG(D_HA, "%p %s: import already not connected: %s\n",
- imp,imp->imp_client->cli_name,
+ imp,imp->imp_client->cli_name,
ptlrpc_import_state_name(imp->imp_state));
}
return rc;
}
-void ptlrpc_invalidate_import(struct obd_import *imp)
+/*
+ * This acts as a barrier; all existing requests are rejected, and
+ * no new requests will be accepted until the import is valid again.
+ */
+void ptlrpc_deactivate_import(struct obd_import *imp)
{
- struct obd_device *obd = imp->imp_obd;
unsigned long flags;
ENTRY;
spin_lock_irqsave(&imp->imp_lock, flags);
- /* This is a bit of a hack, but invalidating replayable
- * imports makes a temporary reconnect failure into a much more
- * ugly -- and hard to remedy -- situation. */
- if (!imp->imp_replayable) {
- CDEBUG(D_HA, "setting import %s INVALID\n",
- imp->imp_target_uuid.uuid);
- imp->imp_invalid = 1;
- }
+ CDEBUG(D_HA, "setting import %s INVALID\n",
+ imp->imp_target_uuid.uuid);
+ imp->imp_invalid = 1;
imp->imp_generation++;
spin_unlock_irqrestore(&imp->imp_lock, flags);
ptlrpc_abort_inflight(imp);
- obd_import_event(obd, imp, IMP_EVENT_INVALIDATE);
+ obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
+}
+
+/*
+ * This function will invalidate the import, if necessary, then block
+ * for all the RPC completions, and finally notify the obd to
+ * invalidate its state (ie cancel locks, clear pending requests,
+ * etc).
+ *
+ * in_rpc: true if this is called while processing an rpc, like
+ * CONNECT. It will allow for one RPC to be inflight while
+ * waiting for requests to complete. Ugly, yes, but I don't see an
+ * cleaner way right now.
+ */
+void ptlrpc_invalidate_import(struct obd_import *imp, int in_rpc)
+{
+ struct l_wait_info lwi;
+ int inflight = 0;
+ int rc;
+
+ if (!imp->imp_invalid)
+ ptlrpc_deactivate_import(imp);
+
+ LASSERT(imp->imp_invalid);
+
+ if (in_rpc)
+ inflight = 1;
+ /* wait for all requests to error out and call completion callbacks */
+ lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), NULL,
+ NULL, NULL);
+ rc = l_wait_event(imp->imp_recovery_waitq,
+ (atomic_read(&imp->imp_inflight) == inflight),
+ &lwi);
+
+ if (rc)
+ CERROR("%s: rc = %d waiting for callback (%d != %d)\n",
+ imp->imp_target_uuid.uuid, rc,
+ atomic_read(&imp->imp_inflight), inflight);
+
+ obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
}
-void ptlrpc_validate_import(struct obd_import *imp)
+static void ptlrpc_activate_import(struct obd_import *imp)
{
struct obd_device *obd = imp->imp_obd;
unsigned long flags;
imp->imp_target_uuid.uuid,
imp->imp_connection->c_remote_uuid.uuid,
imp->imp_obd->obd_name);
- ptlrpc_invalidate_import(imp);
+ ptlrpc_deactivate_import(imp);
}
-
- CDEBUG(D_HA, "%s: waking up pinger\n",
+
+ CDEBUG(D_HA, "%s: waking up pinger\n",
imp->imp_target_uuid.uuid);
-
+
spin_lock_irqsave(&imp->imp_lock, flags);
imp->imp_force_verify = 1;
spin_unlock_irqrestore(&imp->imp_lock, flags);
-
+
ptlrpc_pinger_wake_up();
-
}
EXIT;
}
int initial_connect = 0;
int rc;
__u64 committed_before_reconnect = 0;
- int was_invalid = 0;
struct ptlrpc_request *request;
int size[] = {sizeof(imp->imp_target_uuid),
sizeof(obd->obd_uuid),
IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
- imp->imp_conn_cnt++;
+ imp->imp_conn_cnt++;
imp->imp_last_replay_transno = 0;
if (imp->imp_remote_handle.cookie == 0) {
}
- if (imp->imp_invalid) {
- imp->imp_invalid = 0;
- was_invalid = 1;
- }
spin_unlock_irqrestore(&imp->imp_lock, flags);
imp->imp_connection = conn;
dlmexp = class_conn2export(&imp->imp_dlm_handle);
-
+
LASSERT(dlmexp != NULL);
if (dlmexp->exp_connection)
aa->pcaa_peer_committed = committed_before_reconnect;
aa->pcaa_initial_connect = initial_connect;
- aa->pcaa_was_invalid = was_invalid;
if (aa->pcaa_initial_connect)
imp->imp_replayable = 1;
unsigned long flags;
int msg_flags;
ENTRY;
-
+
spin_lock_irqsave(&imp->imp_lock, flags);
if (imp->imp_state == LUSTRE_IMP_CLOSED) {
spin_unlock_irqrestore(&imp->imp_lock, flags);
imp->imp_remote_handle = request->rq_repmsg->handle;
} else {
CERROR("reconnected to %s@%s after partition\n",
- imp->imp_target_uuid.uuid,
+ imp->imp_target_uuid.uuid,
imp->imp_connection->c_remote_uuid.uuid);
}
- IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
- }
- else if (MSG_CONNECT_RECOVERING & msg_flags) {
+
+ if (imp->imp_invalid)
+ IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
+ else
+ IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
+ } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
LASSERT(imp->imp_replayable);
- imp->imp_state = LUSTRE_IMP_RECOVER;
imp->imp_remote_handle = request->rq_repmsg->handle;
IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
- }
- else {
+ } else {
imp->imp_remote_handle = request->rq_repmsg->handle;
IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
}
-
+
/* Sanity checks for a reconnected import. */
- if (!(imp->imp_replayable) !=
- !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
+ if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
CERROR("imp_replayable flag does not match server "
"after reconnect. We should LBUG right here.\n");
}
finish:
rc = ptlrpc_import_recovery_state_machine(imp);
if (rc != 0) {
- if (aa->pcaa_was_invalid)
- ptlrpc_invalidate_import(imp);
-
if (rc == -ENOTCONN) {
CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
"invalidating and reconnecting\n",
imp->imp_connection->c_remote_uuid.uuid);
ptlrpc_connect_import(imp, NULL);
RETURN(0);
- }
+ }
}
out:
if (rc != 0) {
IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
if (aa->pcaa_initial_connect && !imp->imp_initial_recov) {
- ptlrpc_invalidate_import(imp);
+ ptlrpc_deactivate_import(imp);
}
CDEBUG(D_ERROR, "recovery of %s on %s failed (%d)\n",
imp->imp_target_uuid.uuid,
req->rq_replen = lustre_msg_size(0, NULL);
req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
- req->rq_timeout *= 3;
+ req->rq_timeout *= 3;
req->rq_interpret_reply = completed_replay_interpret;
ptlrpcd_add_req(req);
RETURN(0);
}
-
int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
{
int rc = 0;
CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
imp->imp_target_uuid.uuid,
imp->imp_connection->c_remote_uuid.uuid);
- ptlrpc_invalidate_import(imp);
+
+ ptlrpc_invalidate_import(imp, 1);
+
IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
- }
-
+ }
+
if (imp->imp_state == LUSTRE_IMP_REPLAY) {
CDEBUG(D_HA, "replay requested by %s\n",
imp->imp_target_uuid.uuid);
rc = ptlrpc_replay_next(imp, &inflight);
- if (inflight == 0 &&
+ if (inflight == 0 &&
atomic_read(&imp->imp_replay_inflight) == 0) {
IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
rc = ldlm_replay_locks(imp);
if (rc)
GOTO(out, rc);
IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
- ptlrpc_validate_import(imp);
- }
+ ptlrpc_activate_import(imp);
+ }
if (imp->imp_state == LUSTRE_IMP_FULL) {
wake_up(&imp->imp_recovery_waitq);
RETURN(rc);
}
-static int back_to_sleep(void *unused)
+static int back_to_sleep(void *unused)
{
return 0;
}
if (ptlrpc_import_in_recovery(imp)) {
struct l_wait_info lwi;
- lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), back_to_sleep,
+ lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), back_to_sleep,
NULL, NULL);
- rc = l_wait_event(imp->imp_recovery_waitq,
+ rc = l_wait_event(imp->imp_recovery_waitq,
!ptlrpc_import_in_recovery(imp), &lwi);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff