lustre/ptlrpc/recover.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Light Super operations
   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_LLITE
  22
  23 #include <linux/lustre_lite.h>
  24 #include <linux/lustre_ha.h>
  25
  26 int ll_reconnect(struct ptlrpc_connection *conn)
  27 {
  28         struct list_head *tmp;
  29         int rc = -EINVAL;
  30
  31         /* XXX c_lock semantics! */
  32         conn->c_level = LUSTRE_CONN_CON;
  33
  34         /* XXX this code MUST be shared with class_obd_connect! */
  35         list_for_each(tmp, &conn->c_imports) {
  36                 struct obd_import *imp = list_entry(tmp, struct obd_import,
  37                                                     imp_chain);
  38                 struct obd_device *obd = imp->imp_obd;
  39                 struct client_obd *cli = &obd->u.cli;
  40                 int rq_opc = (obd->obd_type->typ_ops->o_brw)
  41                         ? OST_CONNECT : MDS_CONNECT;
  42                 int size[] = { sizeof(cli->cl_target_uuid),
  43                                sizeof(obd->obd_uuid) };
  44                 char *tmp[] = {cli->cl_target_uuid, obd->obd_uuid };
  45                 struct lustre_handle old_hdl;
  46                 struct ptlrpc_request *request;
  47                 struct obd_export *ldlmexp;
  48
  49                 LASSERT(imp->imp_connection == conn);
  50                 request = ptlrpc_prep_req(imp, rq_opc, 2, size, tmp);
  51                 request->rq_level = LUSTRE_CONN_NEW;
  52                 request->rq_replen = lustre_msg_size(0, NULL);
  53                 /*
  54                  * This address is the export that represents our client-side
  55                  * LDLM service (for ASTs).  We should only have one on this
  56                  * list, so we just grab the first one.
  57                  *
  58                  * XXX tear down export, call class_obd_connect!
  59                  */
  60                 ldlmexp = list_entry(obd->obd_exports.next, struct obd_export,
  61                                      exp_obd_chain);
  62                 request->rq_reqmsg->addr = (__u64)(unsigned long)ldlmexp;
  63                 request->rq_reqmsg->cookie = ldlmexp->exp_cookie;
  64                 rc = ptlrpc_queue_wait(request);
  65                 rc = ptlrpc_check_status(request, rc);
  66                 if (rc) {
  67                         CERROR("cannot connect to %s@%s: rc = %d\n",
  68                                cli->cl_target_uuid, conn->c_remote_uuid, rc);
  69                         ptlrpc_free_req(request);
  70                         GOTO(out_disc, rc = -ENOTCONN);
  71                 }
  72
  73                 old_hdl = imp->imp_handle;
  74                 imp->imp_handle.addr = request->rq_repmsg->addr;
  75                 imp->imp_handle.cookie = request->rq_repmsg->cookie;
  76                 CDEBUG(D_HA, "reconnected to %s@%s (%Lx/%Lx, was %Lx/%Lx)!\n",
  77                        cli->cl_target_uuid, conn->c_remote_uuid,
  78                        imp->imp_handle.addr, imp->imp_handle.cookie,
  79                        old_hdl.addr, old_hdl.cookie);
  80                 ptlrpc_free_req(request);
  81         }
  82         conn->c_level = LUSTRE_CONN_RECOVD;
  83
  84  out_disc:
  85         return rc;
  86 }
  87
  88 static int ll_recover_upcall(struct ptlrpc_connection *conn)
  89 {
  90         char *argv[3];
  91         char *envp[3];
  92         int rc;
  93
  94         ENTRY;
  95         conn->c_level = LUSTRE_CONN_RECOVD;
  96
  97         argv[0] = obd_recovery_upcall;
  98         argv[1] = conn->c_remote_uuid;
  99         argv[2] = NULL;
 100
 101         envp[0] = "HOME=/";
 102         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
 103         envp[2] = NULL;
 104
 105         rc = call_usermodehelper(argv[0], argv, envp);
 106         if (rc < 0) {
 107                 /*
 108                  * Tragically, this will never be run, because call_umh doesn't
 109                  * report errors like -ENOENT to its caller.
 110                  */
 111                 CERROR("Error invoking recovery upcall (%s): %d\n",
 112                        obd_recovery_upcall, rc);
 113                 CERROR("Check /proc/sys/lustre/recovery_upcall?\n");
 114         } else {
 115                 CDEBUG(D_HA, "Invoked upcall %s for connection %s\n",
 116                        argv[0], argv[1]);
 117         }
 118         RETURN(rc);
 119 }
 120
 121 static int ll_recover_reconnect(struct ptlrpc_connection *conn)
 122 {
 123         int rc = 0;
 124         struct list_head *tmp, *pos;
 125         struct ptlrpc_request *req;
 126         ENTRY;
 127
 128         /* 1. reconnect */
 129         rc = ll_reconnect(conn);
 130         if (rc)
 131                 RETURN(rc);
 132
 133         /* 2. walk the request list */
 134         spin_lock(&conn->c_lock);
 135
 136         CDEBUG(D_HA, "connection %p to %s has last_xid "LPD64"\n",
 137                conn, conn->c_remote_uuid, conn->c_last_xid);
 138
 139         list_for_each_safe(tmp, pos, &conn->c_sending_head) {
 140                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 141
 142                 /* replay what needs to be replayed */
 143                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 144                         CDEBUG(D_HA, "FL_REPLAY: xid "LPD64" op %d @ %d\n",
 145                                req->rq_xid, req->rq_reqmsg->opc,
 146                                req->rq_import->imp_client->cli_request_portal);
 147                         rc = ptlrpc_replay_req(req);
 148 #if 0
 149 #error We should not hold a spinlock over such a lengthy operation.
 150 #error If necessary, drop spinlock, do operation, re-get spinlock, restart loop.
 151 #error If we need to avoid re-processint items, then delete them from the list
 152 #error as they are replayed and re-add at the tail of this list, so the next
 153 #error item to process will always be at the head of the list.
 154 #endif
 155                         if (rc) {
 156                                 CERROR("recovery replay error %d for req %Ld\n",
 157                                        rc, req->rq_xid);
 158                                 GOTO(out, rc);
 159                         }
 160                 }
 161
 162                 /* server has seen req, we have reply: skip */
 163                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 164                     req->rq_xid <= conn->c_last_xid) {
 165                         CDEBUG(D_HA, "REPLIED SKIP: xid "LPD64" op %d @ %d\n",
 166                                req->rq_xid, req->rq_reqmsg->opc,
 167                                req->rq_import->imp_client->cli_request_portal);
 168                         continue;
 169                 }
 170
 171                 /* server has lost req, we have reply: resend, ign reply */
 172                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 173                     req->rq_xid > conn->c_last_xid) {
 174                         CDEBUG(D_HA, "REPLIED RESEND: xid "LPD64" op %d @ %d\n",
 175                                req->rq_xid, req->rq_reqmsg->opc,
 176                                req->rq_import->imp_client->cli_request_portal);
 177                         rc = ptlrpc_replay_req(req);
 178                         if (rc) {
 179                                 CERROR("request resend error %d for req %Ld\n",
 180                                        rc, req->rq_xid);
 181                                 GOTO(out, rc);
 182                         }
 183                 }
 184
 185                 /* server has seen req, we have lost reply: -ERESTARTSYS */
 186                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 187                      req->rq_xid <= conn->c_last_xid) {
 188                         CDEBUG(D_HA, "RESTARTSYS: xid "LPD64" op %d @ %d\n",
 189                                req->rq_xid, req->rq_reqmsg->opc,
 190                                req->rq_import->imp_client->cli_request_portal);
 191                         ptlrpc_restart_req(req);
 192                 }
 193
 194                 /* service has not seen req, no reply: resend */
 195                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 196                      req->rq_xid > conn->c_last_xid) {
 197                         CDEBUG(D_HA, "RESEND: xid "LPD64" op %d @ %d\n",
 198                                req->rq_xid, req->rq_reqmsg->opc,
 199                                req->rq_import->imp_client->cli_request_portal);
 200                         ptlrpc_resend_req(req);
 201                 }
 202
 203         }
 204
 205         conn->c_level = LUSTRE_CONN_FULL;
 206         recovd_conn_fixed(conn);
 207
 208         CERROR("recovery complete on conn %p(%s), waking delayed reqs\n",
 209                conn, conn->c_remote_uuid);
 210         /* Finally, continue what we delayed since recovery started */
 211         list_for_each_safe(tmp, pos, &conn->c_delayed_head) {
 212                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 213                 ptlrpc_continue_req(req);
 214         }
 215
 216         EXIT;
 217  out:
 218         spin_unlock(&conn->c_lock);
 219         return rc;
 220 }
 221
 222 static int ll_retry_recovery(struct ptlrpc_connection *conn)
 223 {
 224         CDEBUG(D_HA, "Recovery has failed on conn %p\n", conn);
 225 #if 0
 226         /* XXX use a timer, sideshow bob */
 227         recovd_conn_fail(conn);
 228         /* XXX this is disabled until I fix it so that we don't just keep
 229          * XXX retrying in the case of a missing upcall.
 230          */
 231 #endif
 232         return 0;
 233 }
 234
 235 int ll_recover(struct recovd_data *rd, int phase)
 236 {
 237         struct ptlrpc_connection *conn = class_rd2conn(rd);
 238
 239         LASSERT(conn);
 240         ENTRY;
 241
 242         switch (phase) {
 243             case PTLRPC_RECOVD_PHASE_PREPARE:
 244                 RETURN(ll_recover_upcall(conn));
 245             case PTLRPC_RECOVD_PHASE_RECOVER:
 246                 RETURN(ll_recover_reconnect(conn));
 247             case PTLRPC_RECOVD_PHASE_FAILURE:
 248                 RETURN(ll_retry_recovery(conn));
 249         }
 250
 251         LBUG();
 252         RETURN(-ENOSYS);
 253 }