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/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 /* This might block waiting for the upcall to start, so it should
 125  * not be called from a thread that shouldn't block. (Like ptlrpcd) */
 126 void ptlrpc_initiate_recovery(struct obd_import *imp)
 127 {
 128         ENTRY;
 129
 130         LASSERT (obd_lustre_upcall != NULL);
 131
 132         if (strcmp(obd_lustre_upcall, "DEFAULT") == 0) {
 133                 CDEBUG(D_HA, "%s: starting recovery without upcall\n",
 134                         imp->imp_target_uuid.uuid);
 135                 ptlrpc_connect_import(imp, NULL);
 136         }
 137         else if (strcmp(obd_lustre_upcall, "NONE") == 0) {
 138                 CDEBUG(D_HA, "%s: recovery disabled\n",
 139                         imp->imp_target_uuid.uuid);
 140         }
 141         else {
 142                 CDEBUG(D_HA, "%s: calling upcall to start recovery\n",
 143                         imp->imp_target_uuid.uuid);
 144                 ptlrpc_run_failed_import_upcall(imp);
 145         }
 146
 147         EXIT;
 148 }
 149
 150 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
 151 {
 152         int rc = 0;
 153         struct list_head *tmp, *pos;
 154         struct ptlrpc_request *req = NULL;
 155         unsigned long flags;
 156         __u64 last_transno;
 157         ENTRY;
 158
 159         *inflight = 0;
 160
 161         /* It might have committed some after we last spoke, so make sure we
 162          * get rid of them now.
 163          */
 164         spin_lock_irqsave(&imp->imp_lock, flags);
 165         ptlrpc_free_committed(imp);
 166         last_transno = imp->imp_last_replay_transno;
 167         spin_unlock_irqrestore(&imp->imp_lock, flags);
 168
 169         CDEBUG(D_HA, "import %p from %s committed "LPU64" last "LPU64"\n",
 170                imp, imp->imp_target_uuid.uuid, imp->imp_peer_committed_transno,
 171                last_transno);
 172
 173         /* Do I need to hold a lock across this iteration?  We shouldn't be
 174          * racing with any additions to the list, because we're in recovery
 175          * and are therefore not processing additional requests to add.  Calls
 176          * to ptlrpc_free_committed might commit requests, but nothing "newer"
 177          * than the one we're replaying (it can't be committed until it's
 178          * replayed, and we're doing that here).  l_f_e_safe protects against
 179          * problems with the current request being committed, in the unlikely
 180          * event of that race.  So, in conclusion, I think that it's safe to
 181          * perform this list-walk without the imp_lock held.
 182          *
 183          * But, the {mdc,osc}_replay_open callbacks both iterate
 184          * request lists, and have comments saying they assume the
 185          * imp_lock is being held by ptlrpc_replay, but it's not. it's
 186          * just a little race...
 187          */
 188         list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 189                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 190
 191                 /* If need to resend, stop on the matching one first. It's
 192                    possible though it's already been committed, so in that case
 193                    we'll just continue with replay */
 194                 if (imp->imp_resend_replay &&
 195                     req->rq_transno == last_transno) {
 196                         lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 197                         break;
 198                 }
 199
 200                 if (req->rq_transno > last_transno) {
 201                         imp->imp_last_replay_transno = req->rq_transno;
 202                         break;
 203                 }
 204
 205                 req = NULL;
 206         }
 207
 208         imp->imp_resend_replay = 0;
 209
 210         if (req != NULL) {
 211                 rc = ptlrpc_replay_req(req);
 212                 if (rc) {
 213                         CERROR("recovery replay error %d for req "
 214                                LPD64"\n", rc, req->rq_xid);
 215                         RETURN(rc);
 216                 }
 217                 *inflight = 1;
 218         }
 219         RETURN(rc);
 220 }
 221
 222 int ptlrpc_resend(struct obd_import *imp)
 223 {
 224         struct list_head *tmp, *pos;
 225         struct ptlrpc_request *req;
 226         unsigned long flags;
 227
 228         ENTRY;
 229
 230         /* As long as we're in recovery, nothing should be added to the sending
 231          * list, so we don't need to hold the lock during this iteration and
 232          * resend process.
 233          */
 234         /* Well... what if lctl recover is called twice at the same time?
 235          */
 236         spin_lock_irqsave(&imp->imp_lock, flags);
 237         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 238                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 239                 RETURN(-1);
 240         }
 241         spin_unlock_irqrestore(&imp->imp_lock, flags);
 242
 243         list_for_each_safe(tmp, pos, &imp->imp_sending_list) {
 244                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 245                 ptlrpc_resend_req(req);
 246         }
 247
 248         RETURN(0);
 249 }
 250
 251 void ptlrpc_wake_delayed(struct obd_import *imp)
 252 {
 253         unsigned long flags;
 254         struct list_head *tmp, *pos;
 255         struct ptlrpc_request *req;
 256
 257         spin_lock_irqsave(&imp->imp_lock, flags);
 258         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 259                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 260
 261                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 262                 ptlrpc_wake_client_req(req);
 263         }
 264         spin_unlock_irqrestore(&imp->imp_lock, flags);
 265 }
 266
 267 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 268 {
 269         int rc;
 270         struct obd_import *imp= failed_req->rq_import;
 271         unsigned long flags;
 272         ENTRY;
 273
 274         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 275                imp->imp_obd->obd_name,
 276                imp->imp_target_uuid.uuid,
 277                imp->imp_connection->c_remote_uuid.uuid);
 278
 279         if (ptlrpc_set_import_discon(imp)) {
 280                 if (!imp->imp_replayable) {
 281                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 282                                "auto-deactivating\n",
 283                                imp->imp_target_uuid.uuid,
 284                                imp->imp_connection->c_remote_uuid.uuid,
 285                                imp->imp_obd->obd_name);
 286                         ptlrpc_deactivate_import(imp);
 287                 }
 288
 289                 rc = ptlrpc_connect_import(imp, NULL);
 290         }
 291
 292
 293         /* Wait for recovery to complete and resend. If evicted, then
 294            this request will be errored out later.*/
 295         spin_lock_irqsave(&failed_req->rq_lock, flags);
 296         failed_req->rq_resend = 1;
 297         spin_unlock_irqrestore(&failed_req->rq_lock, flags);
 298
 299         EXIT;
 300 }
 301
 302 /*
 303  * This should only be called by the ioctl interface, currently
 304  * with the lctl deactivate and activate commands.
 305  */
 306 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 307 {
 308         struct obd_device *obd = imp->imp_obd;
 309         int rc = 0;
 310
 311         LASSERT(obd);
 312
 313         /* When deactivating, mark import invalid, and abort in-flight
 314          * requests. */
 315         if (!active) {
 316                 ptlrpc_invalidate_import(imp, 0);
 317         }
 318
 319         /* When activating, mark import valid, and attempt recovery */
 320         if (active) {
 321                 CDEBUG(D_HA, "setting import %s VALID\n",
 322                        imp->imp_target_uuid.uuid);
 323                 rc = ptlrpc_recover_import(imp, NULL);
 324         }
 325
 326         RETURN(rc);
 327 }
 328
 329 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid)
 330 {
 331         int rc;
 332         ENTRY;
 333
 334         /* force import to be disconnected. */
 335         ptlrpc_set_import_discon(imp);
 336
 337         rc = ptlrpc_recover_import_no_retry(imp, new_uuid);
 338
 339         RETURN(rc);
 340 }
 341
 342 int ptlrpc_import_in_recovery(struct obd_import *imp)
 343 {
 344         unsigned long flags;
 345         int in_recovery = 1;
 346         spin_lock_irqsave(&imp->imp_lock, flags);
 347         if (imp->imp_state == LUSTRE_IMP_FULL ||
 348             imp->imp_state == LUSTRE_IMP_CLOSED ||
 349             imp->imp_state == LUSTRE_IMP_DISCON)
 350                 in_recovery = 0;
 351         spin_unlock_irqrestore(&imp->imp_lock, flags);
 352         return in_recovery;
 353 }
 354
 355 static int ptlrpc_recover_import_no_retry(struct obd_import *imp,
 356                                           char *new_uuid)
 357 {
 358         int rc;
 359         unsigned long flags;
 360         int in_recovery = 0;
 361         struct l_wait_info lwi;
 362         ENTRY;
 363
 364         spin_lock_irqsave(&imp->imp_lock, flags);
 365         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 366                 in_recovery = 1;
 367         }
 368         spin_unlock_irqrestore(&imp->imp_lock, flags);
 369
 370         if (in_recovery == 1)
 371                 RETURN(-EALREADY);
 372
 373         rc = ptlrpc_connect_import(imp, new_uuid);
 374         if (rc)
 375                 RETURN(rc);
 376
 377         CDEBUG(D_HA, "%s: recovery started, waiting\n",
 378                imp->imp_target_uuid.uuid);
 379
 380         lwi = LWI_TIMEOUT(MAX(obd_timeout * HZ, 1), NULL, NULL);
 381         rc = l_wait_event(imp->imp_recovery_waitq,
 382                           !ptlrpc_import_in_recovery(imp), &lwi);
 383         CDEBUG(D_HA, "%s: recovery finished\n",
 384                imp->imp_target_uuid.uuid);
 385
 386         RETURN(rc);
 387 }
 388
 389 void ptlrpc_fail_export(struct obd_export *exp)
 390 {
 391         int rc, already_failed;
 392         unsigned long flags;
 393
 394         spin_lock_irqsave(&exp->exp_lock, flags);
 395         already_failed = exp->exp_failed;
 396         exp->exp_failed = 1;
 397         spin_unlock_irqrestore(&exp->exp_lock, flags);
 398
 399         if (already_failed) {
 400                 CDEBUG(D_HA, "disconnecting dead export %p/%s; skipping\n",
 401                        exp, exp->exp_client_uuid.uuid);
 402                 return;
 403         }
 404
 405         CDEBUG(D_HA, "disconnecting export %p/%s\n",
 406                exp, exp->exp_client_uuid.uuid);
 407
 408         if (obd_dump_on_timeout)
 409                 portals_debug_dumplog();
 410
 411         /* Most callers into obd_disconnect are removing their own reference
 412          * (request, for example) in addition to the one from the hash table.
 413          * We don't have such a reference here, so make one. */
 414         class_export_get(exp);
 415         rc = obd_disconnect(exp, 0);
 416         if (rc)
 417                 CERROR("disconnecting export %p failed: %d\n", exp, rc);
 418 }