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  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   7  *   Author: Mike Shaver <shaver@clusterfs.com>
   8  *
   9  *   This file is part of Lustre, http://www.lustre.org.
  10  *
  11  *   Lustre is free software; you can redistribute it and/or
  12  *   modify it under the terms of version 2 of the GNU General Public
  13  *   License as published by the Free Software Foundation.
  14  *
  15  *   Lustre is distributed in the hope that it will be useful,
  16  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  *   GNU General Public License for more details.
  19  *
  20  *   You should have received a copy of the GNU General Public License
  21  *   along with Lustre; if not, write to the Free Software
  22  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  */
  24
  25 #include <linux/config.h>
  26 #include <linux/module.h>
  27 #include <linux/kmod.h>
  28
  29 #define DEBUG_SUBSYSTEM S_RPC
  30
  31 #include <linux/lustre_ha.h>
  32 #include <linux/lustre_net.h>
  33 #include <linux/obd.h>
  34
  35 int ptlrpc_reconnect_import(struct obd_import *imp, int rq_opc,
  36                             struct ptlrpc_request **reqptr)
  37 {
  38         struct obd_device *obd = imp->imp_obd;
  39         struct client_obd *cli = &obd->u.cli;
  40         int size[] = { sizeof(cli->cl_target_uuid), sizeof(obd->obd_uuid) };
  41         char *tmp[] = {cli->cl_target_uuid.uuid, obd->obd_uuid.uuid};
  42         struct ptlrpc_connection *conn = imp->imp_connection;
  43         struct ptlrpc_request *req;
  44         struct obd_export *ldlmexp;
  45         struct lustre_handle old_hdl;
  46         int rc;
  47
  48         req = ptlrpc_prep_req(imp, rq_opc, 2, size, tmp);
  49         if (!req)
  50                 RETURN(-ENOMEM);
  51         req->rq_level = LUSTRE_CONN_NEW;
  52         req->rq_replen = lustre_msg_size(0, NULL);
  53         /*
  54          * This address is the export that represents our client-side LDLM
  55          * service (for ASTs).  We should only have one on this list, so we
  56          * 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         req->rq_reqmsg->addr = (__u64)(unsigned long)ldlmexp;
  63         req->rq_reqmsg->cookie = ldlmexp->exp_cookie;
  64         rc = ptlrpc_queue_wait(req);
  65         if (rc) {
  66                 CERROR("cannot connect to %s@%s: rc = %d\n",
  67                        cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid, rc);
  68                 GOTO(out_disc, rc);
  69         }
  70         if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT) {
  71                 memset(&old_hdl, 0, sizeof(old_hdl));
  72                 if (!memcmp(&old_hdl.addr, &req->rq_repmsg->addr,
  73                             sizeof (old_hdl.addr)) &&
  74                     !memcmp(&old_hdl.cookie, &req->rq_repmsg->cookie,
  75                             sizeof (old_hdl.cookie))) {
  76                         CERROR("%s@%s didn't like our handle "LPX64"/"LPX64
  77                                ", failed\n", cli->cl_target_uuid.uuid,
  78                                conn->c_remote_uuid.uuid,
  79                                (__u64)(unsigned long)ldlmexp,
  80                                ldlmexp->exp_cookie);
  81                         GOTO(out_disc, rc = -ENOTCONN);
  82                 }
  83
  84                 old_hdl.addr = req->rq_repmsg->addr;
  85                 old_hdl.cookie = req->rq_repmsg->cookie;
  86                 if (memcmp(&imp->imp_handle, &old_hdl, sizeof(old_hdl))) {
  87                         CERROR("%s@%s changed handle from "LPX64"/"LPX64
  88                                " to "LPX64"/"LPX64"; "
  89                                "copying, but this may foreshadow disaster\n",
  90                                cli->cl_target_uuid.uuid,
  91                                conn->c_remote_uuid.uuid,
  92                                old_hdl.addr, old_hdl.cookie,
  93                                imp->imp_handle.addr, imp->imp_handle.cookie);
  94                         imp->imp_handle.addr = req->rq_repmsg->addr;
  95                         imp->imp_handle.cookie = req->rq_repmsg->cookie;
  96                         GOTO(out_disc, rc = 0);
  97                 }
  98
  99                 CERROR("reconnected to %s@%s after partition\n",
 100                        cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid);
 101                 GOTO(out_disc, rc = 0);
 102         }
 103
 104         old_hdl = imp->imp_handle;
 105         imp->imp_handle.addr = req->rq_repmsg->addr;
 106         imp->imp_handle.cookie = req->rq_repmsg->cookie;
 107         CERROR("reconnected to %s@%s ("LPX64"/"LPX64", was "LPX64"/"
 108                LPX64")!\n", cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid,
 109                imp->imp_handle.addr, imp->imp_handle.cookie,
 110                old_hdl.addr, old_hdl.cookie);
 111         GOTO(out_disc, rc = 0);
 112
 113  out_disc:
 114         *reqptr = req;
 115         return rc;
 116 }
 117
 118 int ptlrpc_run_recovery_upcall(struct ptlrpc_connection *conn)
 119 {
 120         char *argv[3];
 121         char *envp[3];
 122         int rc;
 123
 124         ENTRY;
 125         argv[0] = obd_recovery_upcall;
 126         argv[1] = conn->c_remote_uuid.uuid;
 127         argv[2] = NULL;
 128
 129         envp[0] = "HOME=/";
 130         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
 131         envp[2] = NULL;
 132
 133         rc = call_usermodehelper(argv[0], argv, envp);
 134         if (rc < 0) {
 135                 CERROR("Error invoking recovery upcall %s for %s: %d\n",
 136                        argv[0], argv[1], rc);
 137                 CERROR("Check /proc/sys/lustre/recovery_upcall?\n");
 138         } else {
 139                 CERROR("Invoked upcall %s for connection %s\n",
 140                        argv[0], argv[1]);
 141         }
 142
 143         /*
 144          * We don't want to make this a "failed" recovery, because the system
 145          * administrator -- or, perhaps, tester -- may well be able to rescue
 146          * things by running the correct upcall.
 147          */
 148         RETURN(0);
 149 }
 150
 151 int ptlrpc_replay(struct obd_import *imp)
 152 {
 153         int rc = 0;
 154         struct list_head *tmp, *pos;
 155         struct ptlrpc_request *req;
 156         unsigned long flags;
 157         __u64 committed = imp->imp_peer_committed_transno;
 158         ENTRY;
 159
 160         /* It might have committed some after we last spoke, so make sure we
 161          * get rid of them now.
 162          */
 163         spin_lock_irqsave(&imp->imp_lock, flags);
 164
 165         ptlrpc_free_committed(imp);
 166
 167         CDEBUG(D_HA, "import %p from %s has committed "LPD64"\n",
 168                imp, imp->imp_obd->u.cli.cl_target_uuid.uuid, committed);
 169
 170         list_for_each(tmp, &imp->imp_replay_list) {
 171                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 172                 DEBUG_REQ(D_HA, req, "RETAINED: ");
 173         }
 174
 175         list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 176                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 177
 178                 DEBUG_REQ(D_HA, req, "REPLAY:");
 179
 180                 /* XXX locking WRT failure during replay? */
 181                 rc = ptlrpc_replay_req(req);
 182
 183                 if (rc) {
 184                         CERROR("recovery replay error %d for req "LPD64"\n",
 185                                rc, req->rq_xid);
 186                         GOTO(out, rc);
 187                 }
 188         }
 189
 190  out:
 191         spin_unlock_irqrestore(&imp->imp_lock, flags);
 192         return rc;
 193 }
 194
 195 #define NO_RESEND     0 /* No action required. */
 196 #define RESEND        1 /* Resend required. */
 197 #define RESEND_IGNORE 2 /* Resend, ignore the reply (already saw it). */
 198 #define RESTART       3 /* Have to restart the call, sorry! */
 199
 200 static int resend_type(struct ptlrpc_request *req, __u64 committed)
 201 {
 202         if (req->rq_transno && req->rq_transno < committed) {
 203                 if (req->rq_flags & PTL_RPC_FL_REPLIED) {
 204                         /* Saw the reply and it was committed, no biggie. */
 205                         DEBUG_REQ(D_HA, req, "NO_RESEND");
 206                         return NO_RESEND;
 207                 }
 208                 /* Request committed, but no reply: have to restart. */
 209                 return RESTART;
 210         }
 211
 212         if (req->rq_flags & PTL_RPC_FL_REPLIED) {
 213                 /* Saw reply, so resend and ignore new reply. */
 214                 return RESEND_IGNORE;
 215         }
 216
 217         /* Didn't see reply either, so resend. */
 218         return RESEND;
 219
 220 }
 221
 222 int ptlrpc_resend(struct obd_import *imp)
 223 {
 224         int rc = 0;
 225         struct list_head *tmp, *pos;
 226         struct ptlrpc_request *req;
 227         unsigned long flags;
 228         __u64 committed = imp->imp_peer_committed_transno;
 229
 230         ENTRY;
 231
 232         spin_lock_irqsave(&imp->imp_lock, flags);
 233         list_for_each(tmp, &imp->imp_sending_list) {
 234                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 235                 DEBUG_REQ(D_HA, req, "SENDING: ");
 236         }
 237
 238         list_for_each_safe(tmp, pos, &imp->imp_sending_list) {
 239                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 240
 241                 switch(resend_type(req, committed)) {
 242                     case NO_RESEND:
 243                         break;
 244
 245                     case RESTART:
 246                         DEBUG_REQ(D_HA, req, "RESTART:");
 247                         ptlrpc_restart_req(req);
 248                         break;
 249
 250                     case RESEND_IGNORE:
 251                         DEBUG_REQ(D_HA, req, "RESEND_IGNORE:");
 252                         rc = ptlrpc_replay_req(req);
 253                         if (rc) {
 254                                 DEBUG_REQ(D_ERROR, req, "error %d resending:",
 255                                           rc);
 256                                 ptlrpc_restart_req(req); /* might as well */
 257                         }
 258                         break;
 259
 260                     case RESEND:
 261                         DEBUG_REQ(D_HA, req, "RESEND:");
 262                         ptlrpc_resend_req(req);
 263                         break;
 264
 265                     default:
 266                         LBUG();
 267                 }
 268         }
 269
 270         spin_unlock_irqrestore(&imp->imp_lock, flags);
 271         RETURN(rc);
 272 }
 273
 274 void ptlrpc_wake_delayed(struct obd_import *imp)
 275 {
 276         unsigned long flags;
 277         struct list_head *tmp, *pos;
 278         struct ptlrpc_request *req;
 279
 280         spin_lock_irqsave(&imp->imp_lock, flags);
 281         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 282                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 283                 DEBUG_REQ(D_HA, req, "waking:");
 284                 wake_up(&req->rq_wait_for_rep);
 285         }
 286         spin_unlock_irqrestore(&imp->imp_lock, flags);
 287 }