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 #define DEBUG_SUBSYSTEM S_RPC
  26 #ifdef __KERNEL__
  27 #include <linux/config.h>
  28 #include <linux/module.h>
  29 #include <linux/kmod.h>
  30 #else
  31 #include <liblustre.h>
  32 #endif
  33
  34 #include <linux/lustre_ha.h>
  35 #include <linux/lustre_net.h>
  36 #include <linux/obd.h>
  37
  38 int ptlrpc_reconnect_import(struct obd_import *imp, int rq_opc,
  39                             struct ptlrpc_request **reqptr)
  40 {
  41         struct obd_device *obd = imp->imp_obd;
  42         struct client_obd *cli = &obd->u.cli;
  43         int size[] = { sizeof(cli->cl_target_uuid), sizeof(obd->obd_uuid) };
  44         char *tmp[] = {cli->cl_target_uuid.uuid, obd->obd_uuid.uuid};
  45         struct ptlrpc_connection *conn = imp->imp_connection;
  46         struct ptlrpc_request *req;
  47         struct obd_export *ldlmexp;
  48         struct lustre_handle old_hdl;
  49         int rc;
  50
  51         req = ptlrpc_prep_req(imp, rq_opc, 2, size, tmp);
  52         if (!req)
  53                 RETURN(-ENOMEM);
  54         req->rq_level = LUSTRE_CONN_NEW;
  55         req->rq_replen = lustre_msg_size(0, NULL);
  56         /*
  57          * This address is the export that represents our client-side LDLM
  58          * service (for ASTs).  We should only have one on this list, so we
  59          * just grab the first one.
  60          *
  61          * XXX tear down export, call class_obd_connect?
  62          */
  63         ldlmexp = list_entry(obd->obd_exports.next, struct obd_export,
  64                              exp_obd_chain);
  65         req->rq_reqmsg->addr = (__u64)(unsigned long)ldlmexp;
  66         req->rq_reqmsg->cookie = ldlmexp->exp_cookie;
  67         rc = ptlrpc_queue_wait(req);
  68         if (rc) {
  69                 CERROR("cannot connect to %s@%s: rc = %d\n",
  70                        cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid, rc);
  71                 GOTO(out_disc, rc);
  72         }
  73         if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT) {
  74                 memset(&old_hdl, 0, sizeof(old_hdl));
  75                 if (!memcmp(&old_hdl.addr, &req->rq_repmsg->addr,
  76                             sizeof (old_hdl.addr)) &&
  77                     !memcmp(&old_hdl.cookie, &req->rq_repmsg->cookie,
  78                             sizeof (old_hdl.cookie))) {
  79                         CERROR("%s@%s didn't like our handle "LPX64"/"LPX64
  80                                ", failed\n", cli->cl_target_uuid.uuid,
  81                                conn->c_remote_uuid.uuid,
  82                                (__u64)(unsigned long)ldlmexp,
  83                                ldlmexp->exp_cookie);
  84                         GOTO(out_disc, rc = -ENOTCONN);
  85                 }
  86
  87                 old_hdl.addr = req->rq_repmsg->addr;
  88                 old_hdl.cookie = req->rq_repmsg->cookie;
  89                 if (memcmp(&imp->imp_handle, &old_hdl, sizeof(old_hdl))) {
  90                         CERROR("%s@%s changed handle from "LPX64"/"LPX64
  91                                " to "LPX64"/"LPX64"; "
  92                                "copying, but this may foreshadow disaster\n",
  93                                cli->cl_target_uuid.uuid,
  94                                conn->c_remote_uuid.uuid,
  95                                old_hdl.addr, old_hdl.cookie,
  96                                imp->imp_handle.addr, imp->imp_handle.cookie);
  97                         imp->imp_handle.addr = req->rq_repmsg->addr;
  98                         imp->imp_handle.cookie = req->rq_repmsg->cookie;
  99                         GOTO(out_disc, rc = 0);
 100                 }
 101
 102                 CERROR("reconnected to %s@%s after partition\n",
 103                        cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid);
 104                 GOTO(out_disc, rc = 0);
 105         }
 106
 107         old_hdl = imp->imp_handle;
 108         imp->imp_handle.addr = req->rq_repmsg->addr;
 109         imp->imp_handle.cookie = req->rq_repmsg->cookie;
 110         CERROR("reconnected to %s@%s ("LPX64"/"LPX64", was "LPX64"/"
 111                LPX64")!\n", cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid,
 112                imp->imp_handle.addr, imp->imp_handle.cookie,
 113                old_hdl.addr, old_hdl.cookie);
 114         GOTO(out_disc, rc = 0);
 115
 116  out_disc:
 117         *reqptr = req;
 118         return rc;
 119 }
 120
 121 int ptlrpc_run_recovery_upcall(struct ptlrpc_connection *conn)
 122 {
 123         char *argv[3];
 124         char *envp[3];
 125         int rc;
 126
 127         ENTRY;
 128         argv[0] = obd_recovery_upcall;
 129         argv[1] = conn->c_remote_uuid.uuid;
 130         argv[2] = NULL;
 131
 132         envp[0] = "HOME=/";
 133         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
 134         envp[2] = NULL;
 135
 136         rc = call_usermodehelper(argv[0], argv, envp);
 137         if (rc < 0) {
 138                 CERROR("Error invoking recovery upcall %s for %s: %d\n",
 139                        argv[0], argv[1], rc);
 140                 CERROR("Check /proc/sys/lustre/recovery_upcall?\n");
 141         } else {
 142                 CERROR("Invoked upcall %s for connection %s\n",
 143                        argv[0], argv[1]);
 144         }
 145
 146         /*
 147          * We don't want to make this a "failed" recovery, because the system
 148          * administrator -- or, perhaps, tester -- may well be able to rescue
 149          * things by running the correct upcall.
 150          */
 151         RETURN(0);
 152 }
 153
 154 int ptlrpc_replay(struct obd_import *imp)
 155 {
 156         int rc = 0;
 157         struct list_head *tmp, *pos;
 158         struct ptlrpc_request *req;
 159         unsigned long flags;
 160         __u64 committed = imp->imp_peer_committed_transno;
 161         ENTRY;
 162
 163         /* It might have committed some after we last spoke, so make sure we
 164          * get rid of them now.
 165          */
 166         spin_lock_irqsave(&imp->imp_lock, flags);
 167
 168         ptlrpc_free_committed(imp);
 169
 170         CDEBUG(D_HA, "import %p from %s has committed "LPD64"\n",
 171                imp, imp->imp_obd->u.cli.cl_target_uuid.uuid, committed);
 172
 173         list_for_each(tmp, &imp->imp_replay_list) {
 174                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 175                 DEBUG_REQ(D_HA, req, "RETAINED: ");
 176         }
 177
 178         list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 179                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 180
 181                 DEBUG_REQ(D_HA, req, "REPLAY:");
 182
 183                 /* XXX locking WRT failure during replay? */
 184                 rc = ptlrpc_replay_req(req);
 185
 186                 if (rc) {
 187                         CERROR("recovery replay error %d for req "LPD64"\n",
 188                                rc, req->rq_xid);
 189                         GOTO(out, rc);
 190                 }
 191         }
 192
 193  out:
 194         spin_unlock_irqrestore(&imp->imp_lock, flags);
 195         return rc;
 196 }
 197
 198 #define NO_RESEND     0 /* No action required. */
 199 #define RESEND        1 /* Resend required. */
 200 #define RESEND_IGNORE 2 /* Resend, ignore the reply (already saw it). */
 201 #define RESTART       3 /* Have to restart the call, sorry! */
 202
 203 static int resend_type(struct ptlrpc_request *req, __u64 committed)
 204 {
 205         if (req->rq_transno && req->rq_transno < committed) {
 206                 if (req->rq_flags & PTL_RPC_FL_REPLIED) {
 207                         /* Saw the reply and it was committed, no biggie. */
 208                         DEBUG_REQ(D_HA, req, "NO_RESEND");
 209                         return NO_RESEND;
 210                 }
 211                 /* Request committed, but no reply: have to restart. */
 212                 return RESTART;
 213         }
 214
 215         if (req->rq_flags & PTL_RPC_FL_REPLIED) {
 216                 /* Saw reply, so resend and ignore new reply. */
 217                 return RESEND_IGNORE;
 218         }
 219
 220         /* Didn't see reply either, so resend. */
 221         return RESEND;
 222
 223 }
 224
 225 int ptlrpc_resend(struct obd_import *imp)
 226 {
 227         int rc = 0;
 228         struct list_head *tmp, *pos;
 229         struct ptlrpc_request *req;
 230         unsigned long flags;
 231         __u64 committed = imp->imp_peer_committed_transno;
 232
 233         ENTRY;
 234
 235         spin_lock_irqsave(&imp->imp_lock, flags);
 236         list_for_each_safe(tmp, pos, &imp->imp_sending_list) {
 237                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 238
 239                 switch(resend_type(req, committed)) {
 240                     case NO_RESEND:
 241                         break;
 242
 243                     case RESTART:
 244                         ptlrpc_restart_req(req);
 245                         break;
 246
 247                     case RESEND_IGNORE:
 248                         rc = ptlrpc_replay_req(req);
 249                         if (rc) {
 250                                 DEBUG_REQ(D_ERROR, req, "error %d resending:",
 251                                           rc);
 252                                 ptlrpc_restart_req(req); /* might as well */
 253                         }
 254                         break;
 255
 256                     case RESEND:
 257                         ptlrpc_resend_req(req);
 258                         break;
 259
 260                     default:
 261                         LBUG();
 262                 }
 263         }
 264
 265         spin_unlock_irqrestore(&imp->imp_lock, flags);
 266         RETURN(rc);
 267 }
 268
 269 void ptlrpc_wake_delayed(struct obd_import *imp)
 270 {
 271         unsigned long flags;
 272         struct list_head *tmp, *pos;
 273         struct ptlrpc_request *req;
 274
 275         spin_lock_irqsave(&imp->imp_lock, flags);
 276         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 277                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 278                 DEBUG_REQ(D_HA, req, "waking:");
 279                 wake_up(&req->rq_wait_for_rep);
 280         }
 281         spin_unlock_irqrestore(&imp->imp_lock, flags);
 282 }