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 ptlrpc_request *request;
  29         struct list_head *tmp;
  30         int rc = -EINVAL;
  31
  32         /* XXX c_lock semantics! */
  33         conn->c_level = LUSTRE_CONN_CON;
  34
  35         /* XXX this code MUST be shared with class_obd_connect! */
  36         list_for_each(tmp, &conn->c_imports) {
  37                 struct obd_import *imp = list_entry(tmp, struct obd_import,
  38                                                     imp_chain);
  39                 struct obd_device *obd = imp->imp_obd;
  40                 struct client_obd *cli = &obd->u.cli;
  41                 int rq_opc = (obd->obd_type->typ_ops->o_brw)
  42                         ? OST_CONNECT : MDS_CONNECT;
  43                 int size[] = { sizeof(cli->cl_target_uuid),
  44                                sizeof(obd->obd_uuid) };
  45                 char *tmp[] = {cli->cl_target_uuid, obd->obd_uuid };
  46                 struct lustre_handle old_hdl;
  47
  48                 LASSERT(imp->imp_connection == conn);
  49                 request = ptlrpc_prep_req(imp, rq_opc, 2, size, tmp);
  50                 request->rq_level = LUSTRE_CONN_NEW;
  51                 request->rq_replen = lustre_msg_size(0, NULL);
  52                 /* XXX are (addr, cookie) right? */
  53                 request->rq_reqmsg->addr = imp->imp_handle.addr;
  54                 request->rq_reqmsg->cookie = imp->imp_handle.cookie;
  55                 rc = ptlrpc_queue_wait(request);
  56                 rc = ptlrpc_check_status(request, rc);
  57                 if (rc) {
  58                         CERROR("cannot connect to %s@%s: rc = %d\n",
  59                                cli->cl_target_uuid, conn->c_remote_uuid, rc);
  60                         ptlrpc_free_req(request);
  61                         GOTO(out_disc, rc = -ENOTCONN);
  62                 }
  63
  64                 old_hdl = imp->imp_handle;
  65                 imp->imp_handle.addr = request->rq_repmsg->addr;
  66                 imp->imp_handle.cookie = request->rq_repmsg->cookie;
  67                 CERROR("reconnected to %s@%s (%Lx/%Lx, was %Lx/%Lx)!\n",
  68                        cli->cl_target_uuid, conn->c_remote_uuid,
  69                        imp->imp_handle.addr, imp->imp_handle.cookie,
  70                        old_hdl.addr, old_hdl.cookie);
  71                 ptlrpc_free_req(request);
  72         }
  73         conn->c_level = LUSTRE_CONN_RECOVD;
  74
  75  out_disc:
  76         return rc;
  77 }
  78
  79 static int ll_recover_upcall(struct ptlrpc_connection *conn)
  80 {
  81         char *argv[3];
  82         char *envp[3];
  83         int rc;
  84
  85         ENTRY;
  86         conn->c_level = LUSTRE_CONN_RECOVD;
  87
  88         argv[0] = obd_recovery_upcall;
  89         argv[1] = conn->c_remote_uuid;
  90         argv[2] = NULL;
  91
  92         envp[0] = "HOME=/";
  93         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
  94         envp[2] = NULL;
  95
  96         rc = call_usermodehelper(argv[0], argv, envp);
  97         if (rc < 0) {
  98                 CERROR("Error invoking recovery upcall (%s): %d\n",
  99                        obd_recovery_upcall, rc);
 100                 CERROR("Check /proc/sys/lustre/recovery_upcall?\n");
 101         } else {
 102                 CERROR("Invoked upcall %s for connection %s\n",
 103                        argv[0], argv[1]);
 104         }
 105         RETURN(rc);
 106 }
 107
 108 static int ll_recover_reconnect(struct ptlrpc_connection *conn)
 109 {
 110         int rc = 0;
 111         struct list_head *tmp, *pos;
 112         struct ptlrpc_request *req;
 113         ENTRY;
 114
 115         /* 1. reconnect */
 116         rc = ll_reconnect(conn);
 117         if (rc)
 118                 RETURN(rc);
 119
 120         /* 2. walk the request list */
 121         spin_lock(&conn->c_lock);
 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_NET, "req %Ld needs replay [last rcvd %Ld]\n",
 129                                req->rq_xid, conn->c_last_xid);
 130                         rc = ptlrpc_replay_req(req);
 131 #if 0
 132 #error We should not hold a spinlock over such a lengthy operation.
 133 #error If necessary, drop spinlock, do operation, re-get spinlock, restart loop.
 134 #error If we need to avoid re-processint items, then delete them from the list
 135 #error as they are replayed and re-add at the tail of this list, so the next
 136 #error item to process will always be at the head of the list.
 137 #endif
 138                         if (rc) {
 139                                 CERROR("recovery replay error %d for req %Ld\n",
 140                                        rc, req->rq_xid);
 141                                 GOTO(out, rc);
 142                         }
 143                 }
 144
 145                 /* server has seen req, we have reply: skip */
 146                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 147                     req->rq_xid <= conn->c_last_xid) {
 148                         CDEBUG(D_NET,
 149                                "req %Ld was complete: skip [last rcvd %Ld]\n",
 150                                req->rq_xid, conn->c_last_xid);
 151                         continue;
 152                 }
 153
 154                 /* server has lost req, we have reply: resend, ign reply */
 155                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 156                     req->rq_xid > conn->c_last_xid) {
 157                         CDEBUG(D_NET, "lost req %Ld have rep: replay [last "
 158                                "rcvd %Ld]\n", req->rq_xid, conn->c_last_xid);
 159                         rc = ptlrpc_replay_req(req);
 160                         if (rc) {
 161                                 CERROR("request resend error %d for req %Ld\n",
 162                                        rc, req->rq_xid);
 163                                 GOTO(out, rc);
 164                         }
 165                 }
 166
 167                 /* server has seen req, we have lost reply: -ERESTARTSYS */
 168                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 169                      req->rq_xid <= conn->c_last_xid) {
 170                         CDEBUG(D_NET, "lost rep %Ld srv did req: restart "
 171                                "[last rcvd %Ld]\n",
 172                                req->rq_xid, conn->c_last_xid);
 173                         ptlrpc_restart_req(req);
 174                 }
 175
 176                 /* service has not seen req, no reply: resend */
 177                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 178                      req->rq_xid > conn->c_last_xid) {
 179                         CDEBUG(D_NET,
 180                                "lost rep/req %Ld: resend [last rcvd %Ld]\n",
 181                                req->rq_xid, conn->c_last_xid);
 182                         ptlrpc_resend_req(req);
 183                 }
 184
 185         }
 186
 187         conn->c_level = LUSTRE_CONN_FULL;
 188         recovd_conn_fixed(conn);
 189
 190         CERROR("recovery complete on conn %p(%s), waking delayed reqs\n",
 191                conn, conn->c_remote_uuid);
 192         /* Finally, continue what we delayed since recovery started */
 193         list_for_each_safe(tmp, pos, &conn->c_delayed_head) {
 194                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 195                 ptlrpc_continue_req(req);
 196         }
 197
 198         EXIT;
 199  out:
 200         spin_unlock(&conn->c_lock);
 201         return rc;
 202 }
 203
 204 static int ll_retry_recovery(struct ptlrpc_connection *conn)
 205 {
 206         CERROR("Recovery has failed on conn %p\n", conn);
 207 #if 0
 208         /* XXX use a timer, sideshow bob */
 209         recovd_conn_fail(conn);
 210         /* XXX this is disabled until I fix it so that we don't just keep
 211          * XXX retrying in the case of a missing upcall.
 212          */
 213 #endif
 214         return 0;
 215 }
 216
 217 int ll_recover(struct recovd_data *rd, int phase)
 218 {
 219         struct ptlrpc_connection *conn = class_rd2conn(rd);
 220
 221         LASSERT(conn);
 222         ENTRY;
 223
 224         switch (phase) {
 225             case PTLRPC_RECOVD_PHASE_PREPARE:
 226                 RETURN(ll_recover_upcall(conn));
 227             case PTLRPC_RECOVD_PHASE_RECOVER:
 228                 RETURN(ll_recover_reconnect(conn));
 229             case PTLRPC_RECOVD_PHASE_FAILURE:
 230                 RETURN(ll_retry_recovery(conn));
 231         }
 232
 233         LBUG();
 234         RETURN(-ENOSYS);
 235 }