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 #define REPLAY_COMMITTED     0 /* Fully processed (commit + reply). */
 112 #define REPLAY_REPLAY        1 /* Forced-replay (e.g. open). */
 113 #define REPLAY_RESEND        2 /* Resend required. */
 114 #define REPLAY_RESEND_IGNORE 3 /* Resend, ignore the reply (already saw it). */
 115 #define REPLAY_RESTART       4 /* Have to restart the call, sorry! */
 116 #define REPLAY_NO_STATE      5 /* Request doesn't change MDS state: skip. */
 117
 118 static int replay_state(struct ptlrpc_request *req, __u64 last_xid)
 119 {
 120         /* This request must always be replayed. */
 121         if (req->rq_flags & PTL_RPC_FL_REPLAY)
 122                 return REPLAY_REPLAY;
 123
 124         /* Uncommitted request */
 125         if (req->rq_xid > last_xid) {
 126                 if (req->rq_flags & PTL_RPC_FL_REPLIED) {
 127                         if (req->rq_transno == 0) {
 128                                 /* If no transno was returned, no state was
 129                                    altered on the MDS. */
 130                                 return REPLAY_NO_STATE;
 131                         }
 132
 133                         /* Saw reply, so resend and ignore new reply. */
 134                         return REPLAY_RESEND_IGNORE;
 135                 }
 136
 137                 /* Didn't see reply either, so resend. */
 138                 return REPLAY_RESEND;
 139         }
 140
 141         /* This request has been committed and we saw the reply.  Goodbye! */
 142         if (req->rq_flags & PTL_RPC_FL_REPLIED)
 143                 return REPLAY_COMMITTED;
 144
 145         /* Request committed, but we didn't see the reply: have to restart. */
 146         return REPLAY_RESTART;
 147 }
 148
 149 static char *replay_state2str(int state) {
 150         static char *state_strings[] = {
 151                 "COMMITTED", "REPLAY", "RESEND", "RESEND_IGNORE", "RESTART",
 152                 "NO_STATE"
 153         };
 154         static char *unknown_state = "UNKNOWN";
 155
 156         if (state < 0 ||
 157             state > (sizeof(state_strings) / sizeof(state_strings[0]))) {
 158                 return unknown_state;
 159         }
 160
 161         return state_strings[state];
 162 }
 163
 164 int ptlrpc_replay(struct ptlrpc_connection *conn)
 165 {
 166         int rc = 0;
 167         struct list_head *tmp, *pos;
 168         struct ptlrpc_request *req;
 169         ENTRY;
 170
 171         spin_lock(&conn->c_lock);
 172
 173         CDEBUG(D_HA, "connection %p to %s has last_xid "LPD64"\n",
 174                conn, conn->c_remote_uuid, conn->c_last_xid);
 175
 176         list_for_each(tmp, &conn->c_sending_head) {
 177                 int state;
 178                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 179                 state = replay_state(req, conn->c_last_xid);
 180                 DEBUG_REQ(D_HA, req, "SENDING: %s: ", replay_state2str(state));
 181         }
 182
 183         list_for_each(tmp, &conn->c_delayed_head) {
 184                 int state;
 185                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 186                 state = replay_state(req, conn->c_last_xid);
 187                 DEBUG_REQ(D_HA, req, "DELAYED: ");
 188         }
 189
 190         list_for_each_safe(tmp, pos, &conn->c_sending_head) {
 191                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 192
 193                 switch (replay_state(req, conn->c_last_xid)) {
 194                     case REPLAY_REPLAY:
 195                         DEBUG_REQ(D_HA, req, "REPLAY:");
 196                         rc = ptlrpc_replay_req(req);
 197 #if 0
 198 #error We should not hold a spinlock over such a lengthy operation.
 199 #error If necessary, drop spinlock, do operation, re-get spinlock, restart loop.
 200 #error If we need to avoid re-processint items, then delete them from the list
 201 #error as they are replayed and re-add at the tail of this list, so the next
 202 #error item to process will always be at the head of the list.
 203 #endif
 204                         if (rc) {
 205                                 CERROR("recovery replay error %d for req %Ld\n",
 206                                        rc, req->rq_xid);
 207                                 GOTO(out, rc);
 208                         }
 209                         break;
 210
 211
 212                     case REPLAY_COMMITTED:
 213                         DEBUG_REQ(D_HA, req, "COMMITTED:");
 214                         /* XXX commit now? */
 215                         break;
 216
 217                     case REPLAY_NO_STATE:
 218                         DEBUG_REQ(D_HA, req, "NO_STATE:");
 219                         /* XXX commit now? */
 220                         break;
 221
 222                     case REPLAY_RESEND_IGNORE:
 223                         DEBUG_REQ(D_HA, req, "RESEND_IGNORE:");
 224                         rc = ptlrpc_replay_req(req);
 225                         if (rc) {
 226                                 CERROR("request resend error %d for req %Ld\n",
 227                                        rc, req->rq_xid);
 228                                 GOTO(out, rc);
 229                         }
 230                         break;
 231
 232                     case REPLAY_RESTART:
 233                         DEBUG_REQ(D_HA, req, "RESTART:");
 234                         ptlrpc_restart_req(req);
 235                         break;
 236
 237                     case REPLAY_RESEND:
 238                         DEBUG_REQ(D_HA, req, "RESEND:");
 239                         ptlrpc_resend_req(req);
 240                         break;
 241
 242                     default:
 243                         LBUG();
 244                 }
 245
 246         }
 247
 248         conn->c_level = LUSTRE_CONN_FULL;
 249         recovd_conn_fixed(conn);
 250
 251         CERROR("recovery complete on conn %p(%s), waking delayed reqs\n",
 252                conn, conn->c_remote_uuid);
 253         /* Finally, continue processing requests that blocked for recovery. */
 254         list_for_each_safe(tmp, pos, &conn->c_delayed_head) {
 255                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 256                 DEBUG_REQ(D_HA, req, "WAKING: ");
 257                 ptlrpc_continue_req(req);
 258         }
 259
 260         EXIT;
 261  out:
 262         spin_unlock(&conn->c_lock);
 263         return rc;
 264 }