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 #if 0
  27 /* FIXME: reference to mdc_getstatus causes dependency problems */
  28 static int ll_reconnect(struct ll_sb_info *sbi)
  29 {
  30         struct ll_fid rootfid;
  31         __u64 last_committed;
  32         __u64 last_xid;
  33         int err;
  34         struct ptlrpc_request *request;
  35         struct ptlrpc_connection *conn = sbi2mdc(sbi)->cl_import.imp_connection;
  36
  37         ptlrpc_readdress_connection(conn, "mds");
  38
  39         conn->c_level = LUSTRE_CONN_CON;
  40
  41         /* XXX: need to store the last_* values somewhere */
  42         err = mdc_getstatus(&sbi->ll_mdc_conn, &rootfid, &last_committed,
  43                             &last_xid, &request);
  44         if (err) {
  45                 CERROR("cannot mds_connect: rc = %d\n", err);
  46                 GOTO(out_disc, err = -ENOTCONN);
  47         }
  48         conn->c_last_xid = last_xid;
  49         conn->c_level = LUSTRE_CONN_RECOVD;
  50
  51  out_disc:
  52         return err;
  53 }
  54 #endif
  55
  56 static int ll_recover_upcall(struct ptlrpc_connection *conn)
  57 {
  58         char *argv[3];
  59         char *envp[3];
  60
  61         ENTRY;
  62         conn->c_level = LUSTRE_CONN_RECOVD;
  63
  64         argv[0] = obd_recovery_upcall;
  65         argv[1] = conn->c_remote_uuid;
  66         argv[2] = NULL;
  67
  68         envp[0] = "HOME=/";
  69         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
  70         envp[2] = NULL;
  71
  72         RETURN(call_usermodehelper(argv[0], argv, envp));
  73 }
  74
  75 static int ll_recover_reconnect(struct ptlrpc_connection *conn)
  76 {
  77         RETURN(-ENOSYS);
  78 #if 0
  79         /* XXXshaver this code needs to know about connection-driven recovery! */
  80
  81         struct ptlrpc_request *req;
  82         struct list_head *tmp, *pos;
  83         struct ll_sb_info *sbi = cli->cli_data;
  84         struct ptlrpc_connection *conn = cli->cli_connection;
  85         int rc = 0;
  86         ENTRY;
  87
  88         /* 1. reconnect */
  89         ll_reconnect(sbi);
  90
  91         /* 2. walk the request list */
  92         spin_lock(&conn->c_lock);
  93         list_for_each_safe(tmp, pos, &conn->c_sending_head) {
  94                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
  95
  96                 /* replay what needs to be replayed */
  97                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
  98                         CDEBUG(D_INODE, "req %Ld needs replay [last rcvd %Ld]\n",
  99                                req->rq_xid, conn->c_last_xid);
 100 #error We should not hold a spinlock over such a lengthy operation.
 101 #error If necessary, drop spinlock, do operation, re-get spinlock, restart loop.
 102 #error If we need to avoid re-processint items, then delete them from the list
 103 #error as they are replayed and re-add at the tail of this list, so the next
 104 #error item to process will always be at the head of the list.
 105                         rc = ptlrpc_replay_req(req);
 106                         if (rc) {
 107                                 CERROR("recovery replay error %d for req %Ld\n",
 108                                        rc, req->rq_xid);
 109                                 GOTO(out, rc);
 110                         }
 111                 }
 112
 113                 /* server has seen req, we have reply: skip */
 114                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 115                     req->rq_xid <= conn->c_last_xid) {
 116                         CDEBUG(D_INODE,
 117                                "req %Ld was complete: skip [last rcvd %Ld]\n",
 118                                req->rq_xid, conn->c_last_xid);
 119                         continue;
 120                 }
 121
 122                 /* server has lost req, we have reply: resend, ign reply */
 123                 if ((req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 124                     req->rq_xid > conn->c_last_xid) {
 125                         CDEBUG(D_INODE, "lost req %Ld have rep: replay [last "
 126                                "rcvd %Ld]\n", req->rq_xid, conn->c_last_xid);
 127                         rc = ptlrpc_replay_req(req);
 128                         if (rc) {
 129                                 CERROR("request resend error %d for req %Ld\n",
 130                                        rc, req->rq_xid);
 131                                 GOTO(out, rc);
 132                         }
 133                 }
 134
 135                 /* server has seen req, we have lost reply: -ERESTARTSYS */
 136                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 137                      req->rq_xid <= conn->c_last_xid) {
 138                         CDEBUG(D_INODE, "lost rep %Ld srv did req: restart "
 139                                "[last rcvd %Ld]\n",
 140                                req->rq_xid, conn->c_last_xid);
 141                         ptlrpc_restart_req(req);
 142                 }
 143
 144                 /* service has not seen req, no reply: resend */
 145                 if ( !(req->rq_flags & PTL_RPC_FL_REPLIED)  &&
 146                      req->rq_xid > conn->c_last_xid) {
 147                         CDEBUG(D_INODE,
 148                                "lost rep/req %Ld: resend [last rcvd %Ld]\n",
 149                                req->rq_xid, conn->c_last_xid);
 150                         ptlrpc_resend_req(req);
 151                 }
 152
 153         }
 154
 155         sbi2mdc(sbi)->cl_conn->c_level = LUSTRE_CONN_FULL;
 156         recovd_conn_fixed(conn);
 157
 158         /* Finally, continue what we delayed since recovery started */
 159         list_for_each_safe(tmp, pos, &conn->c_delayed_head) {
 160                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 161                 ptlrpc_continue_req(req);
 162         }
 163
 164         EXIT;
 165  out:
 166         spin_unlock(&conn->c_lock);
 167         return rc;
 168 #endif
 169 }
 170
 171 int ll_recover(struct recovd_data *rd, int phase)
 172 {
 173         struct ptlrpc_connection *conn = class_rd2conn(rd);
 174
 175         LASSERT(conn);
 176         ENTRY;
 177
 178         switch (phase) {
 179             case PTLRPC_RECOVD_PHASE_PREPARE:
 180                 RETURN(ll_recover_upcall(conn));
 181             case PTLRPC_RECOVD_PHASE_RECOVER:
 182                 RETURN(ll_recover_reconnect(conn));
 183             case PTLRPC_RECOVD_PHASE_FAILURE:
 184                 fixme();
 185                 RETURN(0);
 186         }
 187
 188         LBUG();
 189         RETURN(-ENOSYS);
 190 }