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/obd_support.h>
  35 #include <linux/lustre_ha.h>
  36 #include <linux/lustre_net.h>
  37 #include <linux/lustre_import.h>
  38 #include <linux/lustre_export.h>
  39 #include <linux/obd.h>
  40 #include <linux/obd_ost.h>
  41 #include <linux/obd_class.h>
  42 #include <linux/obd_lov.h> /* for IOC_LOV_SET_OSC_ACTIVE */
  43
  44 #include "ptlrpc_internal.h"
  45
  46 static int ptlrpc_recover_import_no_retry(struct obd_import *, char *);
  47
  48 void ptlrpc_run_recovery_over_upcall(struct obd_device *obd)
  49 {
  50         char *argv[4];
  51         char *envp[3];
  52         int rc;
  53         ENTRY;
  54
  55         argv[0] = obd_lustre_upcall;
  56         argv[1] = "RECOVERY_OVER";
  57         argv[2] = obd->obd_uuid.uuid;
  58         argv[3] = NULL;
  59
  60         envp[0] = "HOME=/";
  61         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
  62         envp[2] = NULL;
  63
  64         rc = USERMODEHELPER(argv[0], argv, envp);
  65         if (rc < 0) {
  66                 CERROR("Error invoking recovery upcall %s %s %s: %d; check "
  67                        "/proc/sys/lustre/upcall\n",
  68                        argv[0], argv[1], argv[2], rc);
  69
  70         } else {
  71                 CERROR("Invoked upcall %s %s %s\n",
  72                        argv[0], argv[1], argv[2]);
  73         }
  74 }
  75
  76 void ptlrpc_run_failed_import_upcall(struct obd_import* imp)
  77 {
  78 #ifdef __KERNEL__
  79         unsigned long flags;
  80         char *argv[7];
  81         char *envp[3];
  82         int rc;
  83         ENTRY;
  84
  85         spin_lock_irqsave(&imp->imp_lock, flags);
  86         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
  87                 spin_unlock_irqrestore(&imp->imp_lock, flags);
  88                 EXIT;
  89                 return;
  90         }
  91         spin_unlock_irqrestore(&imp->imp_lock, flags);
  92
  93         argv[0] = obd_lustre_upcall;
  94         argv[1] = "FAILED_IMPORT";
  95         argv[2] = imp->imp_target_uuid.uuid;
  96         argv[3] = imp->imp_obd->obd_name;
  97         argv[4] = imp->imp_connection->c_remote_uuid.uuid;
  98         argv[5] = imp->imp_obd->obd_uuid.uuid;
  99         argv[6] = NULL;
 100
 101         envp[0] = "HOME=/";
 102         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
 103         envp[2] = NULL;
 104
 105         rc = USERMODEHELPER(argv[0], argv, envp);
 106         if (rc < 0) {
 107                 CERROR("Error invoking recovery upcall %s %s %s %s %s: %d; "
 108                        "check /proc/sys/lustre/lustre_upcall\n",
 109                        argv[0], argv[1], argv[2], argv[3], argv[4],rc);
 110
 111         } else {
 112                 CERROR("Invoked upcall %s %s %s %s %s\n",
 113                        argv[0], argv[1], argv[2], argv[3], argv[4]);
 114         }
 115 #else
 116         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 117                 EXIT;
 118                 return;
 119         }
 120         ptlrpc_recover_import(imp, NULL);
 121 #endif
 122 }
 123
 124 int ptlrpc_replay_next(struct obd_import *imp)
 125 {
 126         int rc = 0;
 127         struct list_head *tmp, *pos;
 128         struct ptlrpc_request *req;
 129         unsigned long flags;
 130         __u64 last_transno;
 131         int sent_req = 0;
 132         ENTRY;
 133
 134         /* It might have committed some after we last spoke, so make sure we
 135          * get rid of them now.
 136          */
 137         spin_lock_irqsave(&imp->imp_lock, flags);
 138         ptlrpc_free_committed(imp);
 139         last_transno = imp->imp_last_replay_transno;
 140         spin_unlock_irqrestore(&imp->imp_lock, flags);
 141
 142         CDEBUG(D_HA, "import %p from %s has committed "LPD64"\n",
 143                imp, imp->imp_target_uuid.uuid, imp->imp_peer_committed_transno);
 144         /* Do I need to hold a lock across this iteration?  We shouldn't be
 145          * racing with any additions to the list, because we're in recovery
 146          * and are therefore not processing additional requests to add.  Calls
 147          * to ptlrpc_free_committed might commit requests, but nothing "newer"
 148          * than the one we're replaying (it can't be committed until it's
 149          * replayed, and we're doing that here).  l_f_e_safe protects against
 150          * problems with the current request being committed, in the unlikely
 151          * event of that race.  So, in conclusion, I think that it's safe to
 152          * perform this list-walk without the imp_lock held.
 153          *
 154          * But, the {mdc,osc}_replay_open callbacks both iterate
 155          * request lists, and have comments saying they assume the
 156          * imp_lock is being held by ptlrpc_replay, but it's not. it's
 157          * just a little race...
 158          */
 159         list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 160                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 161                 if (req->rq_transno > last_transno) {
 162                         /* remove from list so ptlrpcd can send the
 163                            req, it should be reinserted after it is
 164                            sent and replied.  Perhaps better solution
 165                            would be to add req->rq_replay_list so the
 166                            req can be saved for replay and still go
 167                            through the normal send thread. */
 168                         rc = ptlrpc_replay_req(req);
 169                         if (rc) {
 170                                 CERROR("recovery replay error %d for req "LPD64"\n",
 171                                        rc, req->rq_xid);
 172                                 RETURN(rc);
 173                         }
 174                         sent_req = 1;
 175                         break;
 176                 }
 177
 178         }
 179
 180         RETURN(sent_req);
 181 }
 182
 183 int ptlrpc_resend(struct obd_import *imp)
 184 {
 185         struct list_head *tmp, *pos;
 186         struct ptlrpc_request *req;
 187         unsigned long flags;
 188
 189         ENTRY;
 190
 191         /* As long as we're in recovery, nothing should be added to the sending
 192          * list, so we don't need to hold the lock during this iteration and
 193          * resend process.
 194          */
 195         /* Well... what if lctl recover is called twice at the same time?
 196          */
 197         spin_lock_irqsave(&imp->imp_lock, flags);
 198         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 199                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 200                 RETURN(-1);
 201         }
 202         spin_unlock_irqrestore(&imp->imp_lock, flags);
 203
 204         list_for_each_safe(tmp, pos, &imp->imp_sending_list) {
 205                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 206                 ptlrpc_resend_req(req);
 207         }
 208
 209         RETURN(0);
 210 }
 211
 212 void ptlrpc_wake_delayed(struct obd_import *imp)
 213 {
 214         unsigned long flags;
 215         struct list_head *tmp, *pos;
 216         struct ptlrpc_request *req;
 217
 218         spin_lock_irqsave(&imp->imp_lock, flags);
 219         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 220                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 221
 222                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 223                 ptlrpc_wake_client_req(req);
 224         }
 225         spin_unlock_irqrestore(&imp->imp_lock, flags);
 226 }
 227
 228 inline void ptlrpc_invalidate_import_state(struct obd_import *imp)
 229 {
 230         struct obd_device *obd = imp->imp_obd;
 231         struct ldlm_namespace *ns = obd->obd_namespace;
 232
 233         ptlrpc_abort_inflight(imp);
 234
 235         obd_invalidate_import(obd, imp);
 236
 237         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
 238 }
 239
 240 void ptlrpc_handle_failed_import(struct obd_import *imp)
 241 {
 242         ENTRY;
 243
 244         if (!imp->imp_replayable) {
 245                 CDEBUG(D_HA,
 246                        "import %s@%s for %s not replayable, deactivating\n",
 247                        imp->imp_target_uuid.uuid,
 248                        imp->imp_connection->c_remote_uuid.uuid,
 249                        imp->imp_obd->obd_name);
 250                 ptlrpc_set_import_active(imp, 0);
 251         }
 252
 253         ptlrpc_run_failed_import_upcall(imp);
 254         EXIT;
 255 }
 256
 257 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 258 {
 259         int rc;
 260         struct obd_import *imp= failed_req->rq_import;
 261         unsigned long flags;
 262         ENTRY;
 263
 264         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 265                imp->imp_obd->obd_name,
 266                imp->imp_target_uuid.uuid,
 267                imp->imp_connection->c_remote_uuid.uuid);
 268
 269         ptlrpc_set_import_discon(imp);
 270
 271         rc = ptlrpc_connect_import(imp, NULL);
 272
 273         /* Wait for recovery to complete and resend. If evicted, then
 274            this request will be errored out later.*/
 275         spin_lock_irqsave(&failed_req->rq_lock, flags);
 276         if (!failed_req->rq_no_resend)
 277                 failed_req->rq_resend = 1;
 278         spin_unlock_irqrestore(&failed_req->rq_lock, flags);
 279
 280         EXIT;
 281 }
 282
 283 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 284 {
 285         struct obd_device *obd = imp->imp_obd;
 286         unsigned long flags;
 287
 288         LASSERT(obd);
 289
 290         /* When deactivating, mark import invalid, and abort in-flight
 291          * requests. */
 292         if (!active) {
 293                 spin_lock_irqsave(&imp->imp_lock, flags);
 294                 /* This is a bit of a hack, but invalidating replayable
 295                  * imports makes a temporary reconnect failure into a much more
 296                  * ugly -- and hard to remedy -- situation. */
 297                 if (!imp->imp_replayable) {
 298                         CDEBUG(D_HA, "setting import %s INVALID\n",
 299                                imp->imp_target_uuid.uuid);
 300                         imp->imp_invalid = 1;
 301                 }
 302                 imp->imp_generation++;
 303                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 304                 ptlrpc_invalidate_import_state(imp);
 305         }
 306
 307         /* When activating, mark import valid */
 308         if (active) {
 309                 CDEBUG(D_HA, "setting import %s VALID\n",
 310                        imp->imp_target_uuid.uuid);
 311                 spin_lock_irqsave(&imp->imp_lock, flags);
 312                 imp->imp_invalid = 0;
 313                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 314         }
 315
 316         if (obd->obd_observer)
 317                 RETURN(obd_notify(obd->obd_observer, obd, active));
 318
 319         RETURN(0);
 320 }
 321
 322 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid)
 323 {
 324         int rc;
 325         ENTRY;
 326
 327         /* force import to be disconnected. */
 328         ptlrpc_set_import_discon(imp);
 329
 330         rc = ptlrpc_recover_import_no_retry(imp, new_uuid);
 331
 332         RETURN(rc);
 333 }
 334
 335 int ptlrpc_import_in_recovery(struct obd_import *imp)
 336 {
 337         unsigned long flags;
 338         int in_recovery = 1;
 339         spin_lock_irqsave(&imp->imp_lock, flags);
 340         if (imp->imp_state == LUSTRE_IMP_FULL ||
 341             imp->imp_state == LUSTRE_IMP_CLOSED ||
 342             imp->imp_state == LUSTRE_IMP_DISCON)
 343                 in_recovery = 0;
 344         spin_unlock_irqrestore(&imp->imp_lock, flags);
 345         return in_recovery;
 346 }
 347
 348 static int ptlrpc_recover_import_no_retry(struct obd_import *imp,
 349                                           char *new_uuid)
 350 {
 351         int rc;
 352         unsigned long flags;
 353         int in_recovery = 0;
 354         struct l_wait_info lwi;
 355         ENTRY;
 356
 357         spin_lock_irqsave(&imp->imp_lock, flags);
 358         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 359                 in_recovery = 1;
 360         }
 361         spin_unlock_irqrestore(&imp->imp_lock, flags);
 362
 363         if (in_recovery == 1)
 364                 RETURN(-EALREADY);
 365
 366
 367         rc = ptlrpc_connect_import(imp, new_uuid);
 368         if (rc)
 369                 RETURN(rc);
 370
 371         CDEBUG(D_ERROR, "%s: recovery started, waiting\n",
 372                imp->imp_client->cli_name);
 373
 374         lwi = LWI_TIMEOUT(MAX(obd_timeout * HZ, 1), NULL, NULL);
 375         rc = l_wait_event(imp->imp_recovery_waitq,
 376                           !ptlrpc_import_in_recovery(imp), &lwi);
 377         CDEBUG(D_ERROR, "%s: recovery finished\n",
 378                imp->imp_client->cli_name);
 379
 380         RETURN(rc);
 381
 382 }
 383
 384 void ptlrpc_fail_export(struct obd_export *exp)
 385 {
 386         int rc, already_failed;
 387         unsigned long flags;
 388
 389         spin_lock_irqsave(&exp->exp_lock, flags);
 390         already_failed = exp->exp_failed;
 391         exp->exp_failed = 1;
 392         spin_unlock_irqrestore(&exp->exp_lock, flags);
 393
 394         if (already_failed) {
 395                 CDEBUG(D_HA, "disconnecting dead export %p/%s; skipping\n",
 396                        exp, exp->exp_client_uuid.uuid);
 397                 return;
 398         }
 399
 400         CDEBUG(D_HA, "disconnecting export %p/%s\n",
 401                exp, exp->exp_client_uuid.uuid);
 402
 403         /* Most callers into obd_disconnect are removing their own reference
 404          * (request, for example) in addition to the one from the hash table.
 405          * We don't have such a reference here, so make one. */
 406         class_export_get(exp);
 407         rc = obd_disconnect(exp, 0);
 408         if (rc)
 409                 CERROR("disconnecting export %p failed: %d\n", exp, rc);
 410 }