/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Lustre Light Super operations
+ * Portal-RPC reconnection and replay operations, for use in recovery.
*
* This code is issued under the GNU General Public License.
* See the file COPYING in this distribution
*
- * Copryright (C) 1996 Peter J. Braam <braam@stelias.com>
- * Copryright (C) 1999 Stelias Computing Inc. <braam@stelias.com>
- * Copryright (C) 1999 Seagate Technology Inc.
- * Copryright (C) 2001 Mountain View Data, Inc.
- * Copryright (C) 2002 Cluster File Systems, Inc.
+ * Copyright (C) 1996 Peter J. Braam <braam@stelias.com>
+ * Copyright (C) 1999 Stelias Computing Inc. <braam@stelias.com>
+ * Copyright (C) 1999 Seagate Technology Inc.
+ * Copyright (C) 2001 Mountain View Data, Inc.
+ * Copyright (C) 2002 Cluster File Systems, Inc.
*
*/
#include <linux/module.h>
#include <linux/kmod.h>
-#define DEBUG_SUBSYSTEM S_LLITE
+#define DEBUG_SUBSYSTEM S_RPC
-#include <linux/lustre_lite.h>
#include <linux/lustre_ha.h>
+#include <linux/lustre_net.h>
+#include <linux/obd.h>
-int ll_reconnect(struct ptlrpc_connection *conn)
+int ptlrpc_reconnect_import(struct obd_import *imp, int rq_opc)
{
- struct ptlrpc_request *request;
- struct list_head *tmp;
- int rc = -EINVAL;
-
- /* XXX c_lock semantics! */
- conn->c_level = LUSTRE_CONN_CON;
-
- /* 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;
-
- 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);
- /* XXX are (addr, cookie) right? */
- request->rq_reqmsg->addr = imp->imp_handle.addr;
- request->rq_reqmsg->cookie = imp->imp_handle.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);
+ struct obd_device *obd = imp->imp_obd;
+ struct client_obd *cli = &obd->u.cli;
+ int size[] = { sizeof(cli->cl_target_uuid), sizeof(obd->obd_uuid) };
+ char *tmp[] = {cli->cl_target_uuid, obd->obd_uuid };
+ struct ptlrpc_connection *conn = imp->imp_connection;
+ struct lustre_handle old_hdl;
+ struct ptlrpc_request *request;
+ struct obd_export *ldlmexp;
+ int rc;
+
+ 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);
+ switch (rc) {
+ case EALREADY:
+ case -EALREADY:
+ /* already connected! */
+ memset(&old_hdl, 0, sizeof(old_hdl));
+ if (!memcmp(&old_hdl.addr, &request->rq_repmsg->addr,
+ sizeof (old_hdl.addr)) &&
+ !memcmp(&old_hdl.cookie, &request->rq_repmsg->cookie,
+ sizeof (old_hdl.cookie))) {
+ CERROR("%s@%s didn't like our handle %Lx/%Lx, failed\n",
+ cli->cl_target_uuid, conn->c_remote_uuid,
+ (__u64)(unsigned long)ldlmexp,
+ ldlmexp->exp_cookie);
GOTO(out_disc, rc = -ENOTCONN);
}
+ old_hdl.addr = request->rq_repmsg->addr;
+ old_hdl.cookie = request->rq_repmsg->cookie;
+ if (memcmp(&imp->imp_handle, &old_hdl, sizeof(old_hdl))) {
+ CERROR("%s@%s changed handle from %Lx/%Lx to %Lx/%Lx; "
+ "copying, but this may foreshadow disaster\n",
+ cli->cl_target_uuid, conn->c_remote_uuid,
+ old_hdl.addr, old_hdl.cookie,
+ imp->imp_handle.addr, imp->imp_handle.cookie);
+ imp->imp_handle.addr = request->rq_repmsg->addr;
+ imp->imp_handle.cookie = request->rq_repmsg->cookie;
+ GOTO(out_disc, rc = EALREADY);
+ }
+
+ CERROR("reconnected to %s@%s after partition\n",
+ cli->cl_target_uuid, conn->c_remote_uuid);
+ GOTO(out_disc, rc = EALREADY);
+ case 0:
old_hdl = imp->imp_handle;
imp->imp_handle.addr = request->rq_repmsg->addr;
imp->imp_handle.cookie = request->rq_repmsg->cookie;
- CERROR("reconnected to %s@%s (%Lx/%Lx, was %Lx/%Lx)!\n",
+ CERROR("now connected 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);
+ GOTO(out_disc, rc = 0);
+ default:
+ CERROR("cannot connect to %s@%s: rc = %d\n",
+ cli->cl_target_uuid, conn->c_remote_uuid, rc);
+ GOTO(out_disc, rc = -ENOTCONN); /* XXX preserve rc? */
}
- conn->c_level = LUSTRE_CONN_RECOVD;
out_disc:
+ ptlrpc_req_finished(request);
return rc;
}
-static int ll_recover_upcall(struct ptlrpc_connection *conn)
+int ptlrpc_run_recovery_upcall(struct ptlrpc_connection *conn)
{
char *argv[3];
char *envp[3];
int rc;
ENTRY;
- conn->c_level = LUSTRE_CONN_RECOVD;
-
argv[0] = obd_recovery_upcall;
argv[1] = conn->c_remote_uuid;
argv[2] = NULL;
rc = call_usermodehelper(argv[0], argv, envp);
if (rc < 0) {
- CERROR("Error invoking recovery upcall (%s): %d\n",
- obd_recovery_upcall, rc);
+ CERROR("Error invoking recovery upcall %s for %s: %d\n",
+ argv[0], argv[1], rc);
CERROR("Check /proc/sys/lustre/recovery_upcall?\n");
+ } else {
+ CERROR("Invoked upcall %s for connection %s\n",
+ argv[0], argv[1]);
}
- RETURN(rc);
+
+ /*
+ * We don't want to make this a "failed" recovery, because the system
+ * administrator -- or, perhaps, tester -- may well be able to rescue
+ * things by running the correct upcall.
+ */
+ RETURN(0);
}
-static int ll_recover_reconnect(struct ptlrpc_connection *conn)
+int ptlrpc_replay(struct obd_import *imp, int send_last_flag)
{
int rc = 0;
struct list_head *tmp, *pos;
struct ptlrpc_request *req;
+ __u64 committed = imp->imp_peer_committed_transno;
ENTRY;
- /* 1. reconnect */
- rc = ll_reconnect(conn);
- if (rc)
- RETURN(rc);
-
- /* 2. walk the request list */
- spin_lock(&conn->c_lock);
+ /* It might have committed some after we last spoke, so make sure we
+ * get rid of them now.
+ */
+ spin_lock(&imp->imp_lock);
- 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_NET, "req %Ld needs replay [last rcvd %Ld]\n",
- req->rq_xid, conn->c_last_xid);
- 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.
-#endif
- if (rc) {
- CERROR("recovery replay error %d for req %Ld\n",
- rc, req->rq_xid);
- GOTO(out, rc);
- }
- }
+ ptlrpc_free_committed(imp);
- /* 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_NET,
- "req %Ld was complete: skip [last rcvd %Ld]\n",
- req->rq_xid, conn->c_last_xid);
- continue;
- }
+ CDEBUG(D_HA, "import %p from %s has committed "LPD64"\n",
+ imp, imp->imp_obd->u.cli.cl_target_uuid, committed);
- /* 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_NET, "lost req %Ld have rep: replay [last "
- "rcvd %Ld]\n", req->rq_xid, conn->c_last_xid);
- rc = ptlrpc_replay_req(req);
- if (rc) {
- CERROR("request resend error %d for req %Ld\n",
- rc, req->rq_xid);
- GOTO(out, rc);
- }
- }
+ list_for_each(tmp, &imp->imp_replay_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_list);
+ DEBUG_REQ(D_HA, req, "RETAINED: ");
+ }
- /* 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_NET, "lost rep %Ld srv did req: restart "
- "[last rcvd %Ld]\n",
- req->rq_xid, conn->c_last_xid);
- ptlrpc_restart_req(req);
- }
+ list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_list);
- /* 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_NET,
- "lost rep/req %Ld: resend [last rcvd %Ld]\n",
- req->rq_xid, conn->c_last_xid);
- ptlrpc_resend_req(req);
+ if (req->rq_transno == imp->imp_max_transno &&
+ send_last_flag) {
+ req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
+ DEBUG_REQ(D_HA, req, "LAST_REPLAY:");
+ } else {
+ DEBUG_REQ(D_HA, req, "REPLAY:");
}
- }
-
- conn->c_level = LUSTRE_CONN_FULL;
- recovd_conn_fixed(conn);
+ rc = ptlrpc_replay_req(req);
+ req->rq_reqmsg->flags &= ~MSG_LAST_REPLAY;
- CDEBUG(D_NET, "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);
- ptlrpc_continue_req(req);
+ if (rc) {
+ CERROR("recovery replay error %d for req %Ld\n",
+ rc, req->rq_xid);
+ GOTO(out, rc);
+ }
}
- EXIT;
out:
- spin_unlock(&conn->c_lock);
+ spin_unlock(&imp->imp_lock);
return rc;
}
-static int ll_retry_recovery(struct ptlrpc_connection *conn)
+#define NO_RESEND 0 /* No action required. */
+#define RESEND 1 /* Resend required. */
+#define RESEND_IGNORE 2 /* Resend, ignore the reply (already saw it). */
+#define RESTART 3 /* Have to restart the call, sorry! */
+
+static int resend_type(struct ptlrpc_request *req, __u64 committed)
{
-#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;
+ if (req->rq_transno < committed) {
+ if (req->rq_flags & PTL_RPC_FL_REPLIED) {
+ /* Saw the reply and it was committed, no biggie. */
+ DEBUG_REQ(D_HA, req, "NO_RESEND");
+ return NO_RESEND;
+ }
+ /* Request committed, but no reply: have to restart. */
+ return RESTART;
+ }
+
+ if (req->rq_flags & PTL_RPC_FL_REPLIED) {
+ /* Saw reply, so resend and ignore new reply. */
+ return RESEND_IGNORE;
+ }
+
+ /* Didn't see reply either, so resend. */
+ return RESEND;
+
}
-int ll_recover(struct recovd_data *rd, int phase)
+int ptlrpc_resend(struct obd_import *imp)
{
- struct ptlrpc_connection *conn = class_rd2conn(rd);
+ int rc = 0;
+ struct list_head *tmp, *pos;
+ struct ptlrpc_request *req;
+ __u64 committed = imp->imp_peer_committed_transno;
- LASSERT(conn);
ENTRY;
- switch (phase) {
- case PTLRPC_RECOVD_PHASE_PREPARE:
- RETURN(ll_recover_upcall(conn));
- case PTLRPC_RECOVD_PHASE_RECOVER:
- RETURN(ll_recover_reconnect(conn));
- case PTLRPC_RECOVD_PHASE_FAILURE:
- RETURN(ll_retry_recovery(conn));
+ spin_lock(&imp->imp_lock);
+ list_for_each(tmp, &imp->imp_sending_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_list);
+ DEBUG_REQ(D_HA, req, "SENDING: ");
}
- LBUG();
- RETURN(-ENOSYS);
+ list_for_each_safe(tmp, pos, &imp->imp_sending_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_list);
+
+ switch(resend_type(req, committed)) {
+ case NO_RESEND:
+ break;
+
+ case RESTART:
+ DEBUG_REQ(D_HA, req, "RESTART:");
+ ptlrpc_restart_req(req);
+ break;
+
+ case RESEND_IGNORE:
+ DEBUG_REQ(D_HA, req, "RESEND_IGNORE:");
+ rc = ptlrpc_replay_req(req);
+ if (rc) {
+ DEBUG_REQ(D_ERROR, req, "error %d resending:",
+ rc);
+ ptlrpc_restart_req(req); /* might as well */
+ }
+ break;
+
+ case RESEND:
+ DEBUG_REQ(D_HA, req, "RESEND:");
+ ptlrpc_resend_req(req);
+ break;
+
+ default:
+ LBUG();
+ }
+ }
+
+ RETURN(rc);
+}
+
+void ptlrpc_wake_delayed(struct obd_import *imp)
+{
+ struct list_head *tmp, *pos;
+ struct ptlrpc_request *req;
+
+ spin_lock(&imp->imp_lock);
+ list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_list);
+ DEBUG_REQ(D_HA, req, "waking:");
+ wake_up(&req->rq_wait_for_rep);
+ }
+ spin_unlock(&imp->imp_lock);
}