#include <linux/lustre_lite.h>
#include <linux/lustre_ha.h>
-#if 0
-/* FIXME: reference to mdc_getstatus causes dependency problems */
-static int ll_reconnect(struct ll_sb_info *sbi)
+int ll_reconnect(struct ptlrpc_connection *conn)
{
- struct ll_fid rootfid;
- __u64 last_committed;
- __u64 last_xid;
- int err;
- struct ptlrpc_request *request;
- struct ptlrpc_connection *conn = sbi2mdc(sbi)->cl_import.imp_connection;
-
- ptlrpc_readdress_connection(conn, "mds");
+ struct list_head *tmp;
+ int rc = -EINVAL;
+ /* XXX c_lock semantics! */
conn->c_level = LUSTRE_CONN_CON;
- /* XXX: need to store the last_* values somewhere */
- err = mdc_getstatus(&sbi->ll_mdc_conn, &rootfid, &last_committed,
- &last_xid, &request);
- if (err) {
- CERROR("cannot mds_connect: rc = %d\n", err);
- GOTO(out_disc, err = -ENOTCONN);
+ /* XXX this code MUST be shared with class_obd_connect! */
+ list_for_each(tmp, &conn->c_imports) {
+ struct obd_import *imp = list_entry(tmp, struct obd_import,
+ imp_chain);
+ struct obd_device *obd = imp->imp_obd;
+ struct client_obd *cli = &obd->u.cli;
+ int rq_opc = (obd->obd_type->typ_ops->o_brw)
+ ? OST_CONNECT : MDS_CONNECT;
+ int size[] = { sizeof(cli->cl_target_uuid),
+ sizeof(obd->obd_uuid) };
+ char *tmp[] = {cli->cl_target_uuid, obd->obd_uuid };
+ struct lustre_handle old_hdl;
+ struct ptlrpc_request *request;
+ struct obd_export *ldlmexp;
+
+ LASSERT(imp->imp_connection == conn);
+ request = ptlrpc_prep_req(imp, rq_opc, 2, size, tmp);
+ request->rq_level = LUSTRE_CONN_NEW;
+ request->rq_replen = lustre_msg_size(0, NULL);
+ /*
+ * This address is the export that represents our client-side
+ * LDLM service (for ASTs). We should only have one on this
+ * list, so we just grab the first one.
+ *
+ * XXX tear down export, call class_obd_connect!
+ */
+ ldlmexp = list_entry(obd->obd_exports.next, struct obd_export,
+ exp_obd_chain);
+ request->rq_reqmsg->addr = (__u64)(unsigned long)ldlmexp;
+ request->rq_reqmsg->cookie = ldlmexp->exp_cookie;
+ rc = ptlrpc_queue_wait(request);
+ rc = ptlrpc_check_status(request, rc);
+ if (rc) {
+ CERROR("cannot connect to %s@%s: rc = %d\n",
+ cli->cl_target_uuid, conn->c_remote_uuid, rc);
+ ptlrpc_free_req(request);
+ GOTO(out_disc, rc = -ENOTCONN);
+ }
+
+ old_hdl = imp->imp_handle;
+ imp->imp_handle.addr = request->rq_repmsg->addr;
+ imp->imp_handle.cookie = request->rq_repmsg->cookie;
+ CDEBUG(D_HA, "reconnected to %s@%s (%Lx/%Lx, was %Lx/%Lx)!\n",
+ cli->cl_target_uuid, conn->c_remote_uuid,
+ imp->imp_handle.addr, imp->imp_handle.cookie,
+ old_hdl.addr, old_hdl.cookie);
+ ptlrpc_free_req(request);
}
- conn->c_last_xid = last_xid;
conn->c_level = LUSTRE_CONN_RECOVD;
out_disc:
- return err;
+ return rc;
}
-#endif
static int ll_recover_upcall(struct ptlrpc_connection *conn)
{
char *argv[3];
char *envp[3];
+ int rc;
ENTRY;
conn->c_level = LUSTRE_CONN_RECOVD;
envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
envp[2] = NULL;
- RETURN(call_usermodehelper(argv[0], argv, envp));
+ rc = call_usermodehelper(argv[0], argv, envp);
+ if (rc < 0) {
+ /*
+ * Tragically, this will never be run, because call_umh doesn't
+ * report errors like -ENOENT to its caller.
+ */
+ CERROR("Error invoking recovery upcall (%s): %d\n",
+ obd_recovery_upcall, rc);
+ CERROR("Check /proc/sys/lustre/recovery_upcall?\n");
+ } else {
+ CDEBUG(D_HA, "Invoked upcall %s for connection %s\n",
+ argv[0], argv[1]);
+ }
+ RETURN(rc);
}
static int ll_recover_reconnect(struct ptlrpc_connection *conn)
{
- RETURN(-ENOSYS);
-#if 0
- /* XXXshaver this code needs to know about connection-driven recovery! */
-
- struct ptlrpc_request *req;
- struct list_head *tmp, *pos;
- struct ll_sb_info *sbi = cli->cli_data;
- struct ptlrpc_connection *conn = cli->cli_connection;
int rc = 0;
+ struct list_head *tmp, *pos;
+ struct ptlrpc_request *req;
ENTRY;
/* 1. reconnect */
- ll_reconnect(sbi);
+ rc = ll_reconnect(conn);
+ if (rc)
+ RETURN(rc);
/* 2. walk the request list */
spin_lock(&conn->c_lock);
+
+ CDEBUG(D_HA, "connection %p to %s has last_xid "LPD64"\n",
+ conn, conn->c_remote_uuid, conn->c_last_xid);
+
list_for_each_safe(tmp, pos, &conn->c_sending_head) {
req = list_entry(tmp, struct ptlrpc_request, rq_list);
/* replay what needs to be replayed */
if (req->rq_flags & PTL_RPC_FL_REPLAY) {
- CDEBUG(D_INODE, "req %Ld needs replay [last rcvd %Ld]\n",
- req->rq_xid, conn->c_last_xid);
+ CDEBUG(D_HA, "FL_REPLAY: xid "LPD64" op %d @ %d\n",
+ req->rq_xid, req->rq_reqmsg->opc,
+ req->rq_import->imp_client->cli_request_portal);
+ rc = ptlrpc_replay_req(req);
+#if 0
#error We should not hold a spinlock over such a lengthy operation.
#error If necessary, drop spinlock, do operation, re-get spinlock, restart loop.
#error If we need to avoid re-processint items, then delete them from the list
#error as they are replayed and re-add at the tail of this list, so the next
#error item to process will always be at the head of the list.
- rc = ptlrpc_replay_req(req);
+#endif
if (rc) {
CERROR("recovery replay error %d for req %Ld\n",
rc, req->rq_xid);
/* server has seen req, we have reply: skip */
if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
req->rq_xid <= conn->c_last_xid) {
- CDEBUG(D_INODE,
- "req %Ld was complete: skip [last rcvd %Ld]\n",
- req->rq_xid, conn->c_last_xid);
+ CDEBUG(D_HA, "REPLIED SKIP: xid "LPD64" op %d @ %d\n",
+ req->rq_xid, req->rq_reqmsg->opc,
+ req->rq_import->imp_client->cli_request_portal);
continue;
}
/* server has lost req, we have reply: resend, ign reply */
if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
req->rq_xid > conn->c_last_xid) {
- CDEBUG(D_INODE, "lost req %Ld have rep: replay [last "
- "rcvd %Ld]\n", req->rq_xid, conn->c_last_xid);
+ CDEBUG(D_HA, "REPLIED RESEND: xid "LPD64" op %d @ %d\n",
+ req->rq_xid, req->rq_reqmsg->opc,
+ req->rq_import->imp_client->cli_request_portal);
rc = ptlrpc_replay_req(req);
if (rc) {
CERROR("request resend error %d for req %Ld\n",
/* server has seen req, we have lost reply: -ERESTARTSYS */
if ( !(req->rq_flags & PTL_RPC_FL_REPLIED) &&
req->rq_xid <= conn->c_last_xid) {
- CDEBUG(D_INODE, "lost rep %Ld srv did req: restart "
- "[last rcvd %Ld]\n",
- req->rq_xid, conn->c_last_xid);
+ CDEBUG(D_HA, "RESTARTSYS: xid "LPD64" op %d @ %d\n",
+ req->rq_xid, req->rq_reqmsg->opc,
+ req->rq_import->imp_client->cli_request_portal);
ptlrpc_restart_req(req);
}
/* service has not seen req, no reply: resend */
if ( !(req->rq_flags & PTL_RPC_FL_REPLIED) &&
req->rq_xid > conn->c_last_xid) {
- CDEBUG(D_INODE,
- "lost rep/req %Ld: resend [last rcvd %Ld]\n",
- req->rq_xid, conn->c_last_xid);
+ CDEBUG(D_HA, "RESEND: xid "LPD64" op %d @ %d\n",
+ req->rq_xid, req->rq_reqmsg->opc,
+ req->rq_import->imp_client->cli_request_portal);
ptlrpc_resend_req(req);
}
}
- sbi2mdc(sbi)->cl_conn->c_level = LUSTRE_CONN_FULL;
+ conn->c_level = LUSTRE_CONN_FULL;
recovd_conn_fixed(conn);
+ CERROR("recovery complete on conn %p(%s), waking delayed reqs\n",
+ conn, conn->c_remote_uuid);
/* Finally, continue what we delayed since recovery started */
list_for_each_safe(tmp, pos, &conn->c_delayed_head) {
req = list_entry(tmp, struct ptlrpc_request, rq_list);
out:
spin_unlock(&conn->c_lock);
return rc;
+}
+
+static int ll_retry_recovery(struct ptlrpc_connection *conn)
+{
+ CDEBUG(D_HA, "Recovery has failed on conn %p\n", conn);
+#if 0
+ /* XXX use a timer, sideshow bob */
+ recovd_conn_fail(conn);
+ /* XXX this is disabled until I fix it so that we don't just keep
+ * XXX retrying in the case of a missing upcall.
+ */
#endif
+ return 0;
}
int ll_recover(struct recovd_data *rd, int phase)
case PTLRPC_RECOVD_PHASE_RECOVER:
RETURN(ll_recover_reconnect(conn));
case PTLRPC_RECOVD_PHASE_FAILURE:
- fixme();
- RETURN(0);
+ RETURN(ll_retry_recovery(conn));
}
LBUG();