lustre/ptlrpc/recover.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Portal-RPC reconnection and replay operations, for use in recovery.
   5  *
   6  * This code is issued under the GNU General Public License.
   7  * See the file COPYING in this distribution
   8  *
   9  * Copryright (C) 1996 Peter J. Braam <braam@stelias.com>
  10  * Copryright (C) 1999 Stelias Computing Inc. <braam@stelias.com>
  11  * Copryright (C) 1999 Seagate Technology Inc.
  12  * Copryright (C) 2001 Mountain View Data, Inc.
  13  * Copryright (C) 2002 Cluster File Systems, Inc.
  14  *
  15  */
  16
  17 #include <linux/config.h>
  18 #include <linux/module.h>
  19 #include <linux/kmod.h>
  20
  21 #define DEBUG_SUBSYSTEM S_RPC
  22
  23 #include <linux/lustre_ha.h>
  24 #include <linux/lustre_net.h>
  25 #include <linux/obd.h>
  26
  27 int ptlrpc_reconnect_import(struct obd_import *imp, int rq_opc)
  28 {
  29         struct obd_device *obd = imp->imp_obd;
  30         struct client_obd *cli = &obd->u.cli;
  31         int size[] = { sizeof(cli->cl_target_uuid), sizeof(obd->obd_uuid) };
  32         char *tmp[] = {cli->cl_target_uuid, obd->obd_uuid };
  33         struct ptlrpc_connection *conn = imp->imp_connection;
  34         struct lustre_handle old_hdl;
  35         struct ptlrpc_request *request;
  36         struct obd_export *ldlmexp;
  37         int rc;
  38
  39         request = ptlrpc_prep_req(imp, rq_opc, 2, size, tmp);
  40         request->rq_level = LUSTRE_CONN_NEW;
  41         request->rq_replen = lustre_msg_size(0, NULL);
  42         /*
  43
  44          * This address is the export that represents our client-side LDLM
  45          * service (for ASTs).  We should only have one on this list, so we
  46          * just grab the first one.
  47          *
  48          * XXX tear down export, call class_obd_connect?
  49          */
  50         ldlmexp = list_entry(obd->obd_exports.next, struct obd_export,
  51                              exp_obd_chain);
  52         request->rq_reqmsg->addr = (__u64)(unsigned long)ldlmexp;
  53         request->rq_reqmsg->cookie = ldlmexp->exp_cookie;
  54         rc = ptlrpc_queue_wait(request);
  55         rc = ptlrpc_check_status(request, rc);
  56         if (rc) {
  57                 CERROR("cannot connect to %s@%s: rc = %d\n",
  58                        cli->cl_target_uuid, conn->c_remote_uuid, rc);
  59                 ptlrpc_free_req(request);
  60                 GOTO(out_disc, rc = -ENOTCONN);
  61         }
  62
  63         old_hdl = imp->imp_handle;
  64         imp->imp_handle.addr = request->rq_repmsg->addr;
  65         imp->imp_handle.cookie = request->rq_repmsg->cookie;
  66         CERROR("reconnected to %s@%s (%Lx/%Lx, was %Lx/%Lx)!\n",
  67                cli->cl_target_uuid, conn->c_remote_uuid,
  68                imp->imp_handle.addr, imp->imp_handle.cookie,
  69                old_hdl.addr, old_hdl.cookie);
  70         ptlrpc_req_finished(request);
  71
  72  out_disc:
  73         return rc;
  74 }
  75
  76 int ptlrpc_run_recovery_upcall(struct ptlrpc_connection *conn)
  77 {
  78         char *argv[3];
  79         char *envp[3];
  80         int rc;
  81
  82         ENTRY;
  83         conn->c_level = LUSTRE_CONN_RECOVD;
  84
  85         argv[0] = obd_recovery_upcall;
  86         argv[1] = conn->c_remote_uuid;
  87         argv[2] = NULL;
  88
  89         envp[0] = "HOME=/";
  90         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
  91         envp[2] = NULL;
  92
  93         rc = call_usermodehelper(argv[0], argv, envp);
  94         if (rc < 0) {
  95                 CERROR("Error invoking recovery upcall %s for %s: %d\n",
  96                        argv[0], argv[1], rc);
  97                 CERROR("Check /proc/sys/lustre/recovery_upcall?\n");
  98         } else {
  99                 CERROR("Invoked upcall %s for connection %s\n",
 100                        argv[0], argv[1]);
 101         }
 102
 103         /*
 104          * We don't want to make this a "failed" recovery, because the system
 105          * administrator -- or, perhaps, tester -- may well be able to rescue
 106          * things by running the correct upcall.
 107          */
 108         RETURN(0);
 109 }
 110
 111 int ptlrpc_replay(struct ptlrpc_connection *conn)
 112 {
 113         int rc = 0;
 114         struct list_head *tmp, *pos;
 115         struct ptlrpc_request *req;
 116         ENTRY;
 117
 118         spin_lock(&conn->c_lock);
 119
 120         CDEBUG(D_HA, "connection %p to %s has last_xid "LPD64"\n",
 121                conn, conn->c_remote_uuid, conn->c_last_xid);
 122
 123         list_for_each_safe(tmp, pos, &conn->c_sending_head) {
 124                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 125
 126                 /* replay what needs to be replayed */
 127                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 128                         CDEBUG(D_HA, "FL_REPLAY: xid "LPD64" transno "LPD64" op %d @ %d\n",
 129                                req->rq_xid, req->rq_repmsg->transno, req->rq_reqmsg->opc,
 130                                req->rq_import->imp_client->cli_request_portal);
 131                         rc = ptlrpc_replay_req(req);
 132 #if 0
 133 #error We should not hold a spinlock over such a lengthy operation.
 134 #error If necessary, drop spinlock, do operation, re-get spinlock, restart loop.
 135 #error If we need to avoid re-processint items, then delete them from the list
 136 #error as they are replayed and re-add at the tail of this list, so the next
 137 #error item to process will always be at the head of the list.
 138 #endif
 139                         if (rc) {
 140                                 CERROR("recovery replay error %d for req %Ld\n",
 141                                        rc, req->rq_xid);
 142                                 GOTO(out, rc);
 143                         }
 144                 }
 145
 146                 /* server has seen req, we have reply: skip */
 147                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 148                     req->rq_xid <= conn->c_last_xid) {
 149                         CDEBUG(D_HA, "REPLIED SKIP: xid "LPD64" transno "
 150                                LPD64" op %d @ %d\n",
 151                                req->rq_xid, req->rq_repmsg->transno,
 152                                req->rq_reqmsg->opc,
 153                                req->rq_import->imp_client->cli_request_portal);
 154                         continue;
 155                 }
 156
 157                 /* server has lost req, we have reply: resend, ign reply */
 158                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 159                     req->rq_xid > conn->c_last_xid) {
 160                         CDEBUG(D_HA, "REPLIED RESEND: xid "LPD64" transno "
 161                                LPD64" op %d @ %d\n",
 162                                req->rq_xid, req->rq_repmsg->transno,
 163                                req->rq_reqmsg->opc,
 164                                req->rq_import->imp_client->cli_request_portal);
 165                         rc = ptlrpc_replay_req(req);
 166                         if (rc) {
 167                                 CERROR("request resend error %d for req %Ld\n",
 168                                        rc, req->rq_xid);
 169                                 GOTO(out, rc);
 170                         }
 171                 }
 172
 173                 /* server has seen req, we have lost reply: -ERESTARTSYS */
 174                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 175                      req->rq_xid <= conn->c_last_xid) {
 176                         CDEBUG(D_HA, "RESTARTSYS: xid "LPD64" op %d @ %d\n",
 177                                req->rq_xid, req->rq_reqmsg->opc,
 178                                req->rq_import->imp_client->cli_request_portal);
 179                         ptlrpc_restart_req(req);
 180                 }
 181
 182                 /* service has not seen req, no reply: resend */
 183                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 184                      req->rq_xid > conn->c_last_xid) {
 185                         CDEBUG(D_HA, "RESEND: xid "LPD64" transno "LPD64
 186                                " op %d @ %d\n", req->rq_xid,
 187                                req->rq_repmsg ? req->rq_repmsg->transno : 0,
 188                                req->rq_reqmsg->opc,
 189                                req->rq_import->imp_client->cli_request_portal);
 190                         ptlrpc_resend_req(req);
 191                 }
 192
 193         }
 194
 195         conn->c_level = LUSTRE_CONN_FULL;
 196         recovd_conn_fixed(conn);
 197
 198         CERROR("recovery complete on conn %p(%s), waking delayed reqs\n",
 199                conn, conn->c_remote_uuid);
 200         /* Finally, continue what we delayed since recovery started */
 201         list_for_each_safe(tmp, pos, &conn->c_delayed_head) {
 202                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 203                 ptlrpc_continue_req(req);
 204         }
 205
 206         EXIT;
 207  out:
 208         spin_unlock(&conn->c_lock);
 209         return rc;
 210 }