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