*
* 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
+ * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
+ * Author: Mike Shaver <shaver@clusterfs.com>
*
- * 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.
+ * This file is part of the Lustre file system, http://www.lustre.org
+ * Lustre is a trademark of Cluster File Systems, Inc.
*
+ * You may have signed or agreed to another license before downloading
+ * this software. If so, you are bound by the terms and conditions
+ * of that agreement, and the following does not apply to you. See the
+ * LICENSE file included with this distribution for more information.
+ *
+ * If you did not agree to a different license, then this copy of Lustre
+ * is open source software; you can redistribute it and/or modify it
+ * under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * In either case, Lustre is distributed in the hope that it will be
+ * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
+ * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * license text for more details.
*/
-#include <linux/config.h>
-#include <linux/module.h>
-#include <linux/kmod.h>
-
#define DEBUG_SUBSYSTEM S_RPC
+#ifdef __KERNEL__
+# include <libcfs/libcfs.h>
+#else
+# include <liblustre.h>
+#endif
+
+#include <obd_support.h>
+#include <lustre_ha.h>
+#include <lustre_net.h>
+#include <lustre_import.h>
+#include <lustre_export.h>
+#include <obd.h>
+#include <obd_ost.h>
+#include <obd_class.h>
+#include <obd_lov.h> /* for IOC_LOV_SET_OSC_ACTIVE */
+#include <libcfs/list.h>
+
+#include "ptlrpc_internal.h"
-#include <linux/lustre_ha.h>
-#include <linux/lustre_net.h>
-#include <linux/obd.h>
+static int ptlrpc_recover_import_no_retry(struct obd_import *, char *);
-int ptlrpc_reconnect_import(struct obd_import *imp, int rq_opc)
+void ptlrpc_initiate_recovery(struct obd_import *imp)
{
- 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;
+ ENTRY;
- 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.
+ CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
+ ptlrpc_connect_import(imp, NULL);
+
+ EXIT;
+}
+
+int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
+{
+ int rc = 0;
+ struct list_head *tmp, *pos;
+ struct ptlrpc_request *req = NULL;
+ __u64 last_transno;
+ ENTRY;
+
+ *inflight = 0;
+
+ /* It might have committed some after we last spoke, so make sure we
+ * get rid of them now.
+ */
+ spin_lock(&imp->imp_lock);
+ imp->imp_last_transno_checked = 0;
+ ptlrpc_free_committed(imp);
+ last_transno = imp->imp_last_replay_transno;
+ spin_unlock(&imp->imp_lock);
+
+ CDEBUG(D_HA, "import %p from %s committed "LPU64" last "LPU64"\n",
+ imp, obd2cli_tgt(imp->imp_obd),
+ imp->imp_peer_committed_transno, last_transno);
+
+ /* Do I need to hold a lock across this iteration? We shouldn't be
+ * racing with any additions to the list, because we're in recovery
+ * and are therefore not processing additional requests to add. Calls
+ * to ptlrpc_free_committed might commit requests, but nothing "newer"
+ * than the one we're replaying (it can't be committed until it's
+ * replayed, and we're doing that here). l_f_e_safe protects against
+ * problems with the current request being committed, in the unlikely
+ * event of that race. So, in conclusion, I think that it's safe to
+ * perform this list-walk without the imp_lock held.
*
- * XXX tear down export, call class_obd_connect?
+ * But, the {mdc,osc}_replay_open callbacks both iterate
+ * request lists, and have comments saying they assume the
+ * imp_lock is being held by ptlrpc_replay, but it's not. it's
+ * just a little race...
*/
- 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);
+ list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
+
+ /* If need to resend the last sent transno (because a
+ reconnect has occurred), then stop on the matching
+ req and send it again. If, however, the last sent
+ transno has been committed then we continue replay
+ from the next request. */
+ if (imp->imp_resend_replay &&
+ req->rq_transno == last_transno) {
+ lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
+ break;
}
- 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);
+ if (req->rq_transno > last_transno) {
+ imp->imp_last_replay_transno = req->rq_transno;
+ break;
}
-
- 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("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);
- 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? */
+
+ req = NULL;
}
- out_disc:
- ptlrpc_req_finished(request);
- return rc;
+ spin_lock(&imp->imp_lock);
+ imp->imp_resend_replay = 0;
+ spin_unlock(&imp->imp_lock);
+
+ if (req != NULL) {
+ rc = ptlrpc_replay_req(req);
+ if (rc) {
+ CERROR("recovery replay error %d for req "
+ LPD64"\n", rc, req->rq_xid);
+ RETURN(rc);
+ }
+ *inflight = 1;
+ }
+ RETURN(rc);
}
-int ptlrpc_run_recovery_upcall(struct ptlrpc_connection *conn)
+int ptlrpc_resend(struct obd_import *imp)
{
- char *argv[3];
- char *envp[3];
- int rc;
+ struct ptlrpc_request *req, *next;
ENTRY;
- conn->c_level = LUSTRE_CONN_RECOVD;
-
- argv[0] = obd_recovery_upcall;
- argv[1] = conn->c_remote_uuid;
- argv[2] = NULL;
-
- envp[0] = "HOME=/";
- envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
- envp[2] = NULL;
-
- rc = call_usermodehelper(argv[0], argv, envp);
- if (rc < 0) {
- 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]);
- }
- /*
- * 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.
+ /* As long as we're in recovery, nothing should be added to the sending
+ * list, so we don't need to hold the lock during this iteration and
+ * resend process.
+ */
+ /* Well... what if lctl recover is called twice at the same time?
*/
+ spin_lock(&imp->imp_lock);
+ if (imp->imp_state != LUSTRE_IMP_RECOVER) {
+ spin_unlock(&imp->imp_lock);
+ RETURN(-1);
+ }
+
+ list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
+ LASSERTF((long)req > CFS_PAGE_SIZE && req != LP_POISON,
+ "req %p bad\n", req);
+ LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
+ if (!req->rq_no_resend)
+ ptlrpc_resend_req(req);
+ }
+ spin_unlock(&imp->imp_lock);
+
RETURN(0);
}
-#define REPLAY_COMMITTED 0 /* Fully processed (commit + reply). */
-#define REPLAY_REPLAY 1 /* Forced-replay (e.g. open). */
-#define REPLAY_RESEND 2 /* Resend required. */
-#define REPLAY_RESEND_IGNORE 3 /* Resend, ignore the reply (already saw it). */
-#define REPLAY_RESTART 4 /* Have to restart the call, sorry! */
-
-static int replay_state(struct ptlrpc_request *req, __u64 committed)
+void ptlrpc_wake_delayed(struct obd_import *imp)
{
- /* This request must always be replayed. */
- if (req->rq_flags & PTL_RPC_FL_REPLAY)
- return REPLAY_REPLAY;
-
- /* Uncommitted request */
- if (req->rq_transno > committed) {
- if (req->rq_flags & PTL_RPC_FL_REPLIED) {
- /* Saw reply, so resend and ignore new reply. */
- return REPLAY_RESEND_IGNORE;
- }
-
- /* Didn't see reply either, so resend. */
- return REPLAY_RESEND;
- }
+ struct list_head *tmp, *pos;
+ struct ptlrpc_request *req;
- /* This request has been committed and we saw the reply. Goodbye! */
- if (req->rq_flags & PTL_RPC_FL_REPLIED)
- return REPLAY_COMMITTED;
+ spin_lock(&imp->imp_lock);
+ list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_list);
- /* Request committed, but we didn't see the reply: have to restart. */
- return REPLAY_RESTART;
+ DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
+ ptlrpc_wake_client_req(req);
+ }
+ spin_unlock(&imp->imp_lock);
}
-static char *replay_state2str(int state) {
- static char *state_strings[] = {
- "COMMITTED", "REPLAY", "RESEND", "RESEND_IGNORE", "RESTART",
- };
- static char *unknown_state = "UNKNOWN";
+void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
+{
+ struct obd_import *imp = failed_req->rq_import;
+ ENTRY;
- if (state < 0 ||
- state > (sizeof(state_strings) / sizeof(state_strings[0]))) {
- return unknown_state;
+ CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
+ imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
+ imp->imp_connection->c_remote_uuid.uuid);
+
+ if (ptlrpc_set_import_discon(imp,
+ lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
+ if (!imp->imp_replayable) {
+ CDEBUG(D_HA, "import %s@%s for %s not replayable, "
+ "auto-deactivating\n",
+ obd2cli_tgt(imp->imp_obd),
+ imp->imp_connection->c_remote_uuid.uuid,
+ imp->imp_obd->obd_name);
+ ptlrpc_deactivate_import(imp);
+ }
+ /* to control recovery via lctl {disable|enable}_recovery */
+ if (imp->imp_deactive == 0)
+ ptlrpc_connect_import(imp, NULL);
}
- return state_strings[state];
+ /* Wait for recovery to complete and resend. If evicted, then
+ this request will be errored out later.*/
+ spin_lock(&failed_req->rq_lock);
+ if (!failed_req->rq_no_resend)
+ failed_req->rq_resend = 1;
+ spin_unlock(&failed_req->rq_lock);
+
+ EXIT;
}
-int ptlrpc_replay(struct obd_import *imp, int unreplied_only)
+/*
+ * Administratively active/deactive a client.
+ * This should only be called by the ioctl interface, currently
+ * - the lctl deactivate and activate commands
+ * - echo 0/1 >> /proc/osc/XXX/active
+ * - client umount -f (ll_umount_begin)
+ */
+int ptlrpc_set_import_active(struct obd_import *imp, int active)
{
- int rc = 0, state;
- struct list_head *tmp, *pos;
- struct ptlrpc_request *req;
- struct ptlrpc_connection *conn = imp->imp_connection;
- __u64 committed = imp->imp_peer_committed_transno;
- ENTRY;
-
- spin_lock(&imp->imp_lock);
-
- CDEBUG(D_HA, "import %p from %s has committed "LPD64"\n",
- imp, imp->imp_obd->u.cli.cl_target_uuid, committed);
+ struct obd_device *obd = imp->imp_obd;
+ int rc = 0;
- list_for_each(tmp, &imp->imp_request_list) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- state = replay_state(req, committed);
- DEBUG_REQ(D_HA, req, "SENDING: %s: ", replay_state2str(state));
+ ENTRY;
+ LASSERT(obd);
+
+ /* When deactivating, mark import invalid, and abort in-flight
+ * requests. */
+ if (!active) {
+ LCONSOLE_WARN("setting import %s INACTIVE by administrator "
+ "request\n", obd2cli_tgt(imp->imp_obd));
+ ptlrpc_invalidate_import(imp);
+
+ spin_lock(&imp->imp_lock);
+ imp->imp_deactive = 1;
+ spin_unlock(&imp->imp_lock);
}
- list_for_each(tmp, &conn->c_delayed_head) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- state = replay_state(req, committed);
- DEBUG_REQ(D_HA, req, "DELAYED: %s: ", replay_state2str(state));
+ /* When activating, mark import valid, and attempt recovery */
+ if (active) {
+ spin_lock(&imp->imp_lock);
+ imp->imp_deactive = 0;
+ spin_unlock(&imp->imp_lock);
+
+ CDEBUG(D_HA, "setting import %s VALID\n",
+ obd2cli_tgt(imp->imp_obd));
+ rc = ptlrpc_recover_import(imp, NULL);
}
- list_for_each_safe(tmp, pos, &imp->imp_request_list) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
+ RETURN(rc);
+}
- if (unreplied_only) {
- if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
- DEBUG_REQ(D_HA, req, "UNREPLIED:");
- ptlrpc_restart_req(req);
- }
- continue;
- }
+/* Attempt to reconnect an import */
+int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid)
+{
+ int rc;
+ ENTRY;
- state = replay_state(req, committed);
+ /* force import to be disconnected. */
+ ptlrpc_set_import_discon(imp, 0);
- if (req->rq_transno == imp->imp_max_transno) {
- req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
- DEBUG_REQ(D_HA, req, "last for replay");
- LASSERT(state != REPLAY_COMMITTED);
- }
+ spin_lock(&imp->imp_lock);
+ imp->imp_deactive = 0;
+ spin_unlock(&imp->imp_lock);
- switch (state) {
- case REPLAY_REPLAY:
- DEBUG_REQ(D_HA, req, "REPLAY:");
- 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);
- }
- break;
+ rc = ptlrpc_recover_import_no_retry(imp, new_uuid);
- case REPLAY_COMMITTED:
- DEBUG_REQ(D_ERROR, req, "COMMITTED:");
- /* XXX commit now? */
- break;
+ RETURN(rc);
+}
- case REPLAY_RESEND_IGNORE:
- DEBUG_REQ(D_HA, req, "RESEND_IGNORE:");
- rc = ptlrpc_replay_req(req);
- if (rc) {
- CERROR("request resend error %d for req %Ld\n",
- rc, req->rq_xid);
- GOTO(out, rc);
- }
- break;
+int ptlrpc_import_in_recovery(struct obd_import *imp)
+{
+ int in_recovery = 1;
+ spin_lock(&imp->imp_lock);
+ if (imp->imp_state == LUSTRE_IMP_FULL ||
+ imp->imp_state == LUSTRE_IMP_CLOSED ||
+ imp->imp_state == LUSTRE_IMP_DISCON)
+ in_recovery = 0;
+ spin_unlock(&imp->imp_lock);
+ return in_recovery;
+}
- case REPLAY_RESTART:
- DEBUG_REQ(D_HA, req, "RESTART:");
- ptlrpc_restart_req(req);
- break;
+static int ptlrpc_recover_import_no_retry(struct obd_import *imp,
+ char *new_uuid)
+{
+ int rc;
+ int in_recovery = 0;
+ struct l_wait_info lwi;
+ ENTRY;
- case REPLAY_RESEND:
- DEBUG_REQ(D_HA, req, "RESEND:");
- ptlrpc_resend_req(req);
- break;
+ /* Check if reconnect is already in progress */
+ spin_lock(&imp->imp_lock);
+ if (imp->imp_state != LUSTRE_IMP_DISCON) {
+ in_recovery = 1;
+ }
+ spin_unlock(&imp->imp_lock);
- default:
- LBUG();
- }
+ if (in_recovery == 1)
+ RETURN(-EALREADY);
- }
+ rc = ptlrpc_connect_import(imp, new_uuid);
+ if (rc)
+ RETURN(rc);
- conn->c_level = LUSTRE_CONN_FULL;
- recovd_conn_fixed(conn);
+ CDEBUG(D_HA, "%s: recovery started, waiting\n",
+ obd2cli_tgt(imp->imp_obd));
- CERROR("recovery complete on conn %p(%s), waking delayed reqs\n",
- conn, conn->c_remote_uuid);
- /* Finally, continue processing requests that blocked for recovery. */
- list_for_each_safe(tmp, pos, &conn->c_delayed_head) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_HA, req, "WAKING: ");
- ptlrpc_continue_req(req);
- }
+ lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
+ NULL, NULL);
+ rc = l_wait_event(imp->imp_recovery_waitq,
+ !ptlrpc_import_in_recovery(imp), &lwi);
+ CDEBUG(D_HA, "%s: recovery finished\n",
+ obd2cli_tgt(imp->imp_obd));
- EXIT;
- out:
- spin_unlock(&conn->c_lock);
- return rc;
+ RETURN(rc);
}