lustre/llite/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 static int ll_reconnect(struct ll_sb_info *sbi)
  27 {
  28         struct ll_fid rootfid;
  29         __u64 last_committed;
  30         __u64 last_xid;
  31         int err;
  32         struct ptlrpc_request *request;
  33         struct ptlrpc_connection *conn = sbi2mdc(sbi)->cl_import.imp_connection;
  34
  35         ptlrpc_readdress_connection(conn, "mds");
  36
  37         conn->c_level = LUSTRE_CONN_CON;
  38
  39         /* XXX: need to store the last_* values somewhere */
  40         err = mdc_getstatus(&sbi->ll_mdc_conn, &rootfid, &last_committed,
  41                             &last_xid, &request);
  42         if (err) {
  43                 CERROR("cannot mds_connect: rc = %d\n", err);
  44                 GOTO(out_disc, err = -ENOTCONN);
  45         }
  46         conn->c_last_xid = last_xid;
  47         conn->c_level = LUSTRE_CONN_RECOVD;
  48
  49  out_disc:
  50         return err;
  51 }
  52
  53 static int ll_recover_upcall(struct ptlrpc_connection *conn)
  54 {
  55         char *argv[3];
  56         char *envp[3];
  57
  58         ENTRY;
  59         conn->c_level = LUSTRE_CONN_RECOVD;
  60
  61         argv[0] = obd_recovery_upcall;
  62         argv[1] = conn->c_remote_uuid;
  63         argv[2] = NULL;
  64
  65         envp[0] = "HOME=/";
  66         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
  67         envp[2] = NULL;
  68
  69         RETURN(call_usermodehelper(argv[0], argv, envp));
  70 }
  71
  72 static int ll_recover_reconnect(struct ptlrpc_connection *conn)
  73 {
  74         RETURN(-ENOSYS);
  75 #if 0
  76         /* XXXshaver this code needs to know about connection-driven recovery! */
  77
  78         struct ptlrpc_request *req;
  79         struct list_head *tmp, *pos;
  80         struct ll_sb_info *sbi = cli->cli_data;
  81         struct ptlrpc_connection *conn = cli->cli_connection;
  82         int rc = 0;
  83         ENTRY;
  84
  85         /* 1. reconnect */
  86         ll_reconnect(sbi);
  87
  88         /* 2. walk the request list */
  89         spin_lock(&conn->c_lock);
  90         list_for_each_safe(tmp, pos, &conn->c_sending_head) {
  91                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
  92
  93                 /* replay what needs to be replayed */
  94                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
  95                         CDEBUG(D_INODE, "req %Ld needs replay [last rcvd %Ld]\n",
  96                                req->rq_xid, conn->c_last_xid);
  97 #error We should not hold a spinlock over such a lengthy operation.
  98 #error If necessary, drop spinlock, do operation, re-get spinlock, restart loop.
  99 #error If we need to avoid re-processint items, then delete them from the list
 100 #error as they are replayed and re-add at the tail of this list, so the next
 101 #error item to process will always be at the head of the list.
 102                         rc = ptlrpc_replay_req(req);
 103                         if (rc) {
 104                                 CERROR("recovery replay error %d for req %Ld\n",
 105                                        rc, req->rq_xid);
 106                                 GOTO(out, rc);
 107                         }
 108                 }
 109
 110                 /* server has seen req, we have reply: skip */
 111                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 112                     req->rq_xid <= conn->c_last_xid) {
 113                         CDEBUG(D_INODE,
 114                                "req %Ld was complete: skip [last rcvd %Ld]\n",
 115                                req->rq_xid, conn->c_last_xid);
 116                         continue;
 117                 }
 118
 119                 /* server has lost req, we have reply: resend, ign reply */
 120                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 121                     req->rq_xid > conn->c_last_xid) {
 122                         CDEBUG(D_INODE, "lost req %Ld have rep: replay [last "
 123                                "rcvd %Ld]\n", req->rq_xid, conn->c_last_xid);
 124                         rc = ptlrpc_replay_req(req);
 125                         if (rc) {
 126                                 CERROR("request resend error %d for req %Ld\n",
 127                                        rc, req->rq_xid);
 128                                 GOTO(out, rc);
 129                         }
 130                 }
 131
 132                 /* server has seen req, we have lost reply: -ERESTARTSYS */
 133                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 134                      req->rq_xid <= conn->c_last_xid) {
 135                         CDEBUG(D_INODE, "lost rep %Ld srv did req: restart "
 136                                "[last rcvd %Ld]\n",
 137                                req->rq_xid, conn->c_last_xid);
 138                         ptlrpc_restart_req(req);
 139                 }
 140
 141                 /* service has not seen req, no reply: resend */
 142                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 143                      req->rq_xid > conn->c_last_xid) {
 144                         CDEBUG(D_INODE,
 145                                "lost rep/req %Ld: resend [last rcvd %Ld]\n",
 146                                req->rq_xid, conn->c_last_xid);
 147                         ptlrpc_resend_req(req);
 148                 }
 149
 150         }
 151
 152         sbi2mdc(sbi)->cl_conn->c_level = LUSTRE_CONN_FULL;
 153         recovd_conn_fixed(conn);
 154
 155         /* Finally, continue what we delayed since recovery started */
 156         list_for_each_safe(tmp, pos, &conn->c_delayed_head) {
 157                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 158                 ptlrpc_continue_req(req);
 159         }
 160
 161         EXIT;
 162  out:
 163         spin_unlock(&conn->c_lock);
 164         return rc;
 165 #endif
 166 }
 167
 168 int ll_recover(struct recovd_data *rd, int phase)
 169 {
 170         struct ptlrpc_connection *conn = class_rd2conn(rd);
 171
 172         LASSERT(conn);
 173         ENTRY;
 174
 175         switch (phase) {
 176             case PTLRPC_RECOVD_PHASE_PREPARE:
 177                 RETURN(ll_recover_upcall(conn));
 178             case PTLRPC_RECOVD_PHASE_RECOVER:
 179                 RETURN(ll_recover_reconnect(conn));
 180             case PTLRPC_RECOVD_PHASE_FAILURE:
 181                 fixme();
 182                 RETURN(0);
 183         }
 184
 185         LBUG();
 186         RETURN(-ENOSYS);
 187 }