lustre/ptlrpc/recov_thread.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2003 Cluster File Systems, Inc.
   5  *   Author: Andreas Dilger <adilger@clusterfs.com>
   6  *
   7  *   This file is part of the Lustre file system, http://www.lustre.org
   8  *   Lustre is a trademark of Cluster File Systems, Inc.
   9  *
  10  *   You may have signed or agreed to another license before downloading
  11  *   this software.  If so, you are bound by the terms and conditions
  12  *   of that agreement, and the following does not apply to you.  See the
  13  *   LICENSE file included with this distribution for more information.
  14  *
  15  *   If you did not agree to a different license, then this copy of Lustre
  16  *   is open source software; you can redistribute it and/or modify it
  17  *   under the terms of version 2 of the GNU General Public License as
  18  *   published by the Free Software Foundation.
  19  *
  20  *   In either case, Lustre is distributed in the hope that it will be
  21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  23  *   license text for more details.
  24  *
  25  * OST<->MDS recovery logging thread.
  26  *
  27  * Invariants in implementation:
  28  * - we do not share logs among different OST<->MDS connections, so that
  29  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_LOG
  33
  34 #ifndef EXPORT_SYMTAB
  35 # define EXPORT_SYMTAB
  36 #endif
  37
  38 #ifdef __KERNEL__
  39 # include <libcfs/libcfs.h>
  40 #else
  41 # include <libcfs/list.h>
  42 # include <liblustre.h>
  43 #endif
  44
  45 #include <obd_class.h>
  46 #include <lustre_commit_confd.h>
  47 #include <obd_support.h>
  48 #include <obd_class.h>
  49 #include <lustre_net.h>
  50 #include <lnet/types.h>
  51 #include <libcfs/list.h>
  52 #include <lustre_log.h>
  53 #include "ptlrpc_internal.h"
  54
  55 #ifdef __KERNEL__
  56
  57 /* Allocate new commit structs in case we do not have enough.
  58  * Make the llcd size small enough that it fits into a single page when we
  59  * are sending/receiving it. */
  60 static int llcd_alloc(struct llog_commit_master *lcm)
  61 {
  62         struct llog_canceld_ctxt *llcd;
  63         int llcd_size;
  64
  65         /* payload of lustre_msg V2 is bigger */
  66         llcd_size = 4096 - lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
  67         OBD_ALLOC(llcd,
  68                   llcd_size + offsetof(struct llog_canceld_ctxt, llcd_cookies));
  69         if (llcd == NULL)
  70                 return -ENOMEM;
  71
  72         llcd->llcd_size = llcd_size;
  73         llcd->llcd_lcm = lcm;
  74
  75         spin_lock(&lcm->lcm_llcd_lock);
  76         list_add(&llcd->llcd_list, &lcm->lcm_llcd_free);
  77         atomic_inc(&lcm->lcm_llcd_numfree);
  78         spin_unlock(&lcm->lcm_llcd_lock);
  79
  80         return 0;
  81 }
  82
  83 /* Get a free cookie struct from the list */
  84 static struct llog_canceld_ctxt *llcd_grab(struct llog_commit_master *lcm)
  85 {
  86         struct llog_canceld_ctxt *llcd;
  87
  88 repeat:
  89         spin_lock(&lcm->lcm_llcd_lock);
  90         if (list_empty(&lcm->lcm_llcd_free)) {
  91                 spin_unlock(&lcm->lcm_llcd_lock);
  92                 if (llcd_alloc(lcm) < 0) {
  93                         CERROR("unable to allocate log commit data!\n");
  94                         return NULL;
  95                 }
  96                 /* check new llcd wasn't grabbed while lock dropped, b=7407 */
  97                 goto repeat;
  98         }
  99
 100         llcd = list_entry(lcm->lcm_llcd_free.next, typeof(*llcd), llcd_list);
 101         list_del(&llcd->llcd_list);
 102         atomic_dec(&lcm->lcm_llcd_numfree);
 103         spin_unlock(&lcm->lcm_llcd_lock);
 104
 105         llcd->llcd_cookiebytes = 0;
 106
 107         return llcd;
 108 }
 109
 110 static void llcd_put(struct llog_canceld_ctxt *llcd)
 111 {
 112         struct llog_commit_master *lcm = llcd->llcd_lcm;
 113
 114         llog_ctxt_put(llcd->llcd_ctxt);
 115         if (atomic_read(&lcm->lcm_llcd_numfree) >= lcm->lcm_llcd_maxfree) {
 116                 int llcd_size = llcd->llcd_size +
 117                          offsetof(struct llog_canceld_ctxt, llcd_cookies);
 118                 OBD_FREE(llcd, llcd_size);
 119         } else {
 120                 spin_lock(&lcm->lcm_llcd_lock);
 121                 list_add(&llcd->llcd_list, &lcm->lcm_llcd_free);
 122                 atomic_inc(&lcm->lcm_llcd_numfree);
 123                 spin_unlock(&lcm->lcm_llcd_lock);
 124         }
 125 }
 126
 127 /* Send some cookies to the appropriate target */
 128 static void llcd_send(struct llog_canceld_ctxt *llcd)
 129 {
 130         if (!(llcd->llcd_lcm->lcm_flags & LLOG_LCM_FL_EXIT)) {
 131                 spin_lock(&llcd->llcd_lcm->lcm_llcd_lock);
 132                 list_add_tail(&llcd->llcd_list,
 133                               &llcd->llcd_lcm->lcm_llcd_pending);
 134                 spin_unlock(&llcd->llcd_lcm->lcm_llcd_lock);
 135         }
 136         cfs_waitq_signal_nr(&llcd->llcd_lcm->lcm_waitq, 1);
 137 }
 138
 139 /**
 140  * Grab llcd and assign it to passed @ctxt. Also set up backward link
 141  * and get ref on @ctxt.
 142  */
 143 static struct llog_canceld_ctxt *ctxt_llcd_grab(struct llog_ctxt *ctxt)
 144 {
 145         struct llog_canceld_ctxt *llcd;
 146
 147         LASSERT_SEM_LOCKED(&ctxt->loc_sem);
 148         llcd = llcd_grab(ctxt->loc_lcm);
 149         if (llcd == NULL)
 150                 return NULL;
 151
 152         llcd->llcd_ctxt = llog_ctxt_get(ctxt);
 153         ctxt->loc_llcd = llcd;
 154
 155         CDEBUG(D_RPCTRACE,"grab llcd %p:%p\n", ctxt->loc_llcd, ctxt);
 156         return llcd;
 157 }
 158
 159 /**
 160  * Put llcd in passed @ctxt. Set ->loc_llcd to NULL.
 161  */
 162 static void ctxt_llcd_put(struct llog_ctxt *ctxt)
 163 {
 164         mutex_down(&ctxt->loc_sem);
 165         if (ctxt->loc_llcd != NULL) {
 166                 CDEBUG(D_RPCTRACE,"put llcd %p:%p\n", ctxt->loc_llcd, ctxt);
 167                 llcd_put(ctxt->loc_llcd);
 168                 ctxt->loc_llcd = NULL;
 169         }
 170         if (ctxt->loc_imp) {
 171                 class_import_put(ctxt->loc_imp);
 172                 ctxt->loc_imp = NULL;
 173         }
 174         mutex_up(&ctxt->loc_sem);
 175 }
 176
 177 /* deleted objects have a commit callback that cancels the MDS
 178  * log record for the deletion.  The commit callback calls this
 179  * function
 180  */
 181 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
 182                          struct lov_stripe_md *lsm, int count,
 183                          struct llog_cookie *cookies, int flags)
 184 {
 185         struct llog_canceld_ctxt *llcd;
 186         int rc = 0;
 187         ENTRY;
 188
 189         LASSERT(ctxt);
 190
 191         mutex_down(&ctxt->loc_sem);
 192         llcd = ctxt->loc_llcd;
 193
 194         if (ctxt->loc_imp == NULL) {
 195                 CDEBUG(D_RPCTRACE, "no import for ctxt %p\n", ctxt);
 196                 GOTO(out, rc = 0);
 197         }
 198
 199         if (count > 0 && cookies != NULL) {
 200                 if (llcd == NULL) {
 201                         llcd = ctxt_llcd_grab(ctxt);
 202                         if (llcd == NULL) {
 203                                 CERROR("couldn't get an llcd - dropped "LPX64
 204                                        ":%x+%u\n",
 205                                        cookies->lgc_lgl.lgl_oid,
 206                                        cookies->lgc_lgl.lgl_ogen,
 207                                        cookies->lgc_index);
 208                                 GOTO(out, rc = -ENOMEM);
 209                         }
 210                 }
 211
 212                 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes,
 213                        cookies, sizeof(*cookies));
 214                 llcd->llcd_cookiebytes += sizeof(*cookies);
 215         } else {
 216                 if (llcd == NULL || !(flags & OBD_LLOG_FL_SENDNOW))
 217                         GOTO(out, rc);
 218         }
 219
 220         if ((llcd->llcd_size - llcd->llcd_cookiebytes) < sizeof(*cookies) ||
 221             (flags & OBD_LLOG_FL_SENDNOW)) {
 222                 CDEBUG(D_RPCTRACE, "send llcd %p:%p\n", llcd, llcd->llcd_ctxt);
 223                 ctxt->loc_llcd = NULL;
 224                 llcd_send(llcd);
 225         }
 226 out:
 227         mutex_up(&ctxt->loc_sem);
 228         return rc;
 229 }
 230 EXPORT_SYMBOL(llog_obd_repl_cancel);
 231
 232 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
 233 {
 234         int rc = 0;
 235         ENTRY;
 236
 237         if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
 238                 CDEBUG(D_RPCTRACE,"reverse import disconnect\n");
 239                 /*
 240                  * We put llcd because it is not going to sending list and
 241                  * thus, its refc will not be handled. We will handle it here.
 242                  */
 243                 ctxt_llcd_put(ctxt);
 244         } else {
 245                 /*
 246                  * Sending cancel. This means that ctxt->loc_llcd wil be
 247                  * put on sending list in llog_obd_repl_cancel() and in
 248                  * this case recovery thread will take care of it refc.
 249                  */
 250                 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
 251         }
 252         RETURN(rc);
 253 }
 254 EXPORT_SYMBOL(llog_obd_repl_sync);
 255
 256 static inline void stop_log_commit(struct llog_commit_master *lcm,
 257                                    struct llog_commit_daemon *lcd,
 258                                    int rc)
 259 {
 260         CERROR("error preparing commit: rc %d\n", rc);
 261
 262         spin_lock(&lcm->lcm_llcd_lock);
 263         list_splice_init(&lcd->lcd_llcd_list, &lcm->lcm_llcd_resend);
 264         spin_unlock(&lcm->lcm_llcd_lock);
 265 }
 266
 267 static int log_commit_thread(void *arg)
 268 {
 269         struct llog_commit_master *lcm = arg;
 270         struct llog_commit_daemon *lcd;
 271         struct llog_canceld_ctxt *llcd, *n;
 272         struct obd_import *import = NULL;
 273         ENTRY;
 274
 275         OBD_ALLOC(lcd, sizeof(*lcd));
 276         if (lcd == NULL)
 277                 RETURN(-ENOMEM);
 278
 279         spin_lock(&lcm->lcm_thread_lock);
 280         THREAD_NAME(cfs_curproc_comm(), CFS_CURPROC_COMM_MAX - 1,
 281                     "ll_log_comt_%02d", atomic_read(&lcm->lcm_thread_total));
 282         atomic_inc(&lcm->lcm_thread_total);
 283         spin_unlock(&lcm->lcm_thread_lock);
 284
 285         ptlrpc_daemonize(cfs_curproc_comm()); /* thread never needs to do IO */
 286
 287         CFS_INIT_LIST_HEAD(&lcd->lcd_lcm_list);
 288         CFS_INIT_LIST_HEAD(&lcd->lcd_llcd_list);
 289         lcd->lcd_lcm = lcm;
 290
 291         CDEBUG(D_HA, "%s started\n", cfs_curproc_comm());
 292         do {
 293                 struct ptlrpc_request *request;
 294                 struct list_head *sending_list;
 295                 int rc = 0;
 296
 297                 if (import)
 298                         class_import_put(import);
 299                 import = NULL;
 300
 301                 /* If we do not have enough pages available, allocate some */
 302                 while (atomic_read(&lcm->lcm_llcd_numfree) <
 303                        lcm->lcm_llcd_minfree) {
 304                         if (llcd_alloc(lcm) < 0)
 305                                 break;
 306                 }
 307
 308                 spin_lock(&lcm->lcm_thread_lock);
 309                 atomic_inc(&lcm->lcm_thread_numidle);
 310                 list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_idle);
 311                 spin_unlock(&lcm->lcm_thread_lock);
 312
 313                 wait_event_interruptible(lcm->lcm_waitq,
 314                                          !list_empty(&lcm->lcm_llcd_pending) ||
 315                                          lcm->lcm_flags & LLOG_LCM_FL_EXIT);
 316
 317                 /* If we are the last available thread, start a new one in case
 318                  * we get blocked on an RPC (nobody else will start a new one)*/
 319                 spin_lock(&lcm->lcm_thread_lock);
 320                 atomic_dec(&lcm->lcm_thread_numidle);
 321                 list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_busy);
 322                 spin_unlock(&lcm->lcm_thread_lock);
 323
 324                 sending_list = &lcm->lcm_llcd_pending;
 325         resend:
 326                 if (import)
 327                         class_import_put(import);
 328                 import = NULL;
 329                 if (lcm->lcm_flags & LLOG_LCM_FL_EXIT) {
 330                         lcm->lcm_llcd_maxfree = 0;
 331                         lcm->lcm_llcd_minfree = 0;
 332                         lcm->lcm_thread_max = 0;
 333
 334                         if (list_empty(&lcm->lcm_llcd_pending) ||
 335                             lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE)
 336                                 break;
 337                 }
 338
 339                 if (atomic_read(&lcm->lcm_thread_numidle) <= 1 &&
 340                     atomic_read(&lcm->lcm_thread_total) < lcm->lcm_thread_max) {
 341                         rc = llog_start_commit_thread(lcm);
 342                         if (rc < 0)
 343                                 CERROR("error starting thread: rc %d\n", rc);
 344                 }
 345
 346                 /* Move all of the pending cancels from the same OST off of
 347                  * the list, so we don't get multiple threads blocked and/or
 348                  * doing upcalls on the same OST in case of failure. */
 349                 spin_lock(&lcm->lcm_llcd_lock);
 350                 if (!list_empty(sending_list)) {
 351                         list_move_tail(sending_list->next,
 352                                        &lcd->lcd_llcd_list);
 353                         llcd = list_entry(lcd->lcd_llcd_list.next,
 354                                           typeof(*llcd), llcd_list);
 355                         LASSERT(llcd->llcd_lcm == lcm);
 356                         import = llcd->llcd_ctxt->loc_imp;
 357                         if (import)
 358                                 class_import_get(import);
 359                 }
 360                 list_for_each_entry_safe(llcd, n, sending_list, llcd_list) {
 361                         LASSERT(llcd->llcd_lcm == lcm);
 362                         if (import == llcd->llcd_ctxt->loc_imp)
 363                                 list_move_tail(&llcd->llcd_list,
 364                                                &lcd->lcd_llcd_list);
 365                 }
 366                 if (sending_list != &lcm->lcm_llcd_resend) {
 367                         list_for_each_entry_safe(llcd, n, &lcm->lcm_llcd_resend,
 368                                                  llcd_list) {
 369                                 LASSERT(llcd->llcd_lcm == lcm);
 370                                 if (import == llcd->llcd_ctxt->loc_imp)
 371                                         list_move_tail(&llcd->llcd_list,
 372                                                        &lcd->lcd_llcd_list);
 373                         }
 374                 }
 375                 spin_unlock(&lcm->lcm_llcd_lock);
 376
 377                 /* We are the only one manipulating our local list - no lock */
 378                 list_for_each_entry_safe(llcd,n, &lcd->lcd_llcd_list,llcd_list){
 379                         char *bufs[2] = { NULL, (char *)llcd->llcd_cookies };
 380
 381                         list_del(&llcd->llcd_list);
 382                         if (llcd->llcd_cookiebytes == 0) {
 383                                 CDEBUG(D_RPCTRACE, "put empty llcd %p:%p\n",
 384                                        llcd, llcd->llcd_ctxt);
 385                                 llcd_put(llcd);
 386                                 continue;
 387                         }
 388
 389                         mutex_down(&llcd->llcd_ctxt->loc_sem);
 390                         if (llcd->llcd_ctxt->loc_imp == NULL) {
 391                                 mutex_up(&llcd->llcd_ctxt->loc_sem);
 392                                 CWARN("import will be destroyed, put "
 393                                       "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
 394                                 llcd_put(llcd);
 395                                 continue;
 396                         }
 397                         mutex_up(&llcd->llcd_ctxt->loc_sem);
 398
 399                         if (!import || (import == LP_POISON) ||
 400                             (import->imp_client == LP_POISON)) {
 401                                 CERROR("No import %p (llcd=%p, ctxt=%p)\n",
 402                                        import, llcd, llcd->llcd_ctxt);
 403                                 llcd_put(llcd);
 404                                 continue;
 405                         }
 406
 407                         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10);
 408
 409                         request = ptlrpc_request_alloc(import, &RQF_LOG_CANCEL);
 410                         if (request == NULL) {
 411                                 rc = -ENOMEM;
 412                                 stop_log_commit(lcm, lcd, rc);
 413                                 break;
 414                         }
 415
 416                         req_capsule_set_size(&request->rq_pill, &RMF_LOGCOOKIES,
 417                                              RCL_CLIENT,llcd->llcd_cookiebytes);
 418
 419                         rc = ptlrpc_request_bufs_pack(request,
 420                                                       LUSTRE_LOG_VERSION,
 421                                                       OBD_LOG_CANCEL, bufs,
 422                                                       NULL);
 423                         if (rc) {
 424                                 ptlrpc_request_free(request);
 425                                 stop_log_commit(lcm, lcd, rc);
 426                                 break;
 427                         }
 428
 429                         /* bug 5515 */
 430                         request->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 431                         request->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 432                         ptlrpc_at_set_req_timeout(request);
 433
 434                         ptlrpc_request_set_replen(request);
 435                         mutex_down(&llcd->llcd_ctxt->loc_sem);
 436                         if (llcd->llcd_ctxt->loc_imp == NULL) {
 437                                 mutex_up(&llcd->llcd_ctxt->loc_sem);
 438                                 CWARN("import will be destroyed, put "
 439                                       "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
 440                                 llcd_put(llcd);
 441                                 ptlrpc_req_finished(request);
 442                                 continue;
 443                         }
 444                         mutex_up(&llcd->llcd_ctxt->loc_sem);
 445                         rc = ptlrpc_queue_wait(request);
 446                         ptlrpc_req_finished(request);
 447
 448                         /* If the RPC failed, we put this and the remaining
 449                          * messages onto the resend list for another time. */
 450                         if (rc == 0) {
 451                                 llcd_put(llcd);
 452                                 continue;
 453                         }
 454
 455                         CERROR("commit %p:%p drop %d cookies: rc %d\n",
 456                                llcd, llcd->llcd_ctxt,
 457                                (int)(llcd->llcd_cookiebytes /
 458                                      sizeof(*llcd->llcd_cookies)), rc);
 459                         llcd_put(llcd);
 460                 }
 461
 462                 if (rc == 0) {
 463                         sending_list = &lcm->lcm_llcd_resend;
 464                         if (!list_empty(sending_list))
 465                                 goto resend;
 466                 }
 467         } while(1);
 468
 469         if (import)
 470                 class_import_put(import);
 471
 472         /* If we are force exiting, just drop all of the cookies. */
 473         if (lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE) {
 474                 spin_lock(&lcm->lcm_llcd_lock);
 475                 list_splice_init(&lcm->lcm_llcd_pending, &lcd->lcd_llcd_list);
 476                 list_splice_init(&lcm->lcm_llcd_resend, &lcd->lcd_llcd_list);
 477                 list_splice_init(&lcm->lcm_llcd_free, &lcd->lcd_llcd_list);
 478                 spin_unlock(&lcm->lcm_llcd_lock);
 479
 480                 list_for_each_entry_safe(llcd, n, &lcd->lcd_llcd_list,llcd_list)
 481                         llcd_put(llcd);
 482         }
 483
 484         spin_lock(&lcm->lcm_thread_lock);
 485         list_del(&lcd->lcd_lcm_list);
 486         spin_unlock(&lcm->lcm_thread_lock);
 487         OBD_FREE(lcd, sizeof(*lcd));
 488
 489         CDEBUG(D_HA, "%s exiting\n", cfs_curproc_comm());
 490
 491         spin_lock(&lcm->lcm_thread_lock);
 492         atomic_dec(&lcm->lcm_thread_total);
 493         spin_unlock(&lcm->lcm_thread_lock);
 494         cfs_waitq_signal(&lcm->lcm_waitq);
 495
 496         return 0;
 497 }
 498
 499 int llog_start_commit_thread(struct llog_commit_master *lcm)
 500 {
 501         int rc;
 502         ENTRY;
 503
 504         if (atomic_read(&lcm->lcm_thread_total) >= lcm->lcm_thread_max)
 505                 RETURN(0);
 506
 507         rc = cfs_kernel_thread(log_commit_thread, lcm, CLONE_VM | CLONE_FILES);
 508         if (rc < 0) {
 509                 CERROR("error starting thread #%d: %d\n",
 510                        atomic_read(&lcm->lcm_thread_total), rc);
 511                 RETURN(rc);
 512         }
 513
 514         RETURN(0);
 515 }
 516 EXPORT_SYMBOL(llog_start_commit_thread);
 517
 518 static struct llog_process_args {
 519         struct semaphore         llpa_sem;
 520         struct llog_ctxt        *llpa_ctxt;
 521         void                    *llpa_cb;
 522         void                    *llpa_arg;
 523 } llpa;
 524
 525 int llog_init_commit_master(struct llog_commit_master *lcm)
 526 {
 527         CFS_INIT_LIST_HEAD(&lcm->lcm_thread_busy);
 528         CFS_INIT_LIST_HEAD(&lcm->lcm_thread_idle);
 529         spin_lock_init(&lcm->lcm_thread_lock);
 530         atomic_set(&lcm->lcm_thread_numidle, 0);
 531         cfs_waitq_init(&lcm->lcm_waitq);
 532         CFS_INIT_LIST_HEAD(&lcm->lcm_llcd_pending);
 533         CFS_INIT_LIST_HEAD(&lcm->lcm_llcd_resend);
 534         CFS_INIT_LIST_HEAD(&lcm->lcm_llcd_free);
 535         spin_lock_init(&lcm->lcm_llcd_lock);
 536         atomic_set(&lcm->lcm_llcd_numfree, 0);
 537         lcm->lcm_llcd_minfree = 0;
 538         lcm->lcm_thread_max = 5;
 539         /* FIXME initialize semaphore for llog_process_args */
 540         sema_init(&llpa.llpa_sem, 1);
 541         return 0;
 542 }
 543 EXPORT_SYMBOL(llog_init_commit_master);
 544
 545 int llog_cleanup_commit_master(struct llog_commit_master *lcm,
 546                                int force)
 547 {
 548         lcm->lcm_flags |= LLOG_LCM_FL_EXIT;
 549         if (force)
 550                 lcm->lcm_flags |= LLOG_LCM_FL_EXIT_FORCE;
 551         cfs_waitq_signal(&lcm->lcm_waitq);
 552
 553         wait_event_interruptible(lcm->lcm_waitq,
 554                                  atomic_read(&lcm->lcm_thread_total) == 0);
 555         return 0;
 556 }
 557 EXPORT_SYMBOL(llog_cleanup_commit_master);
 558
 559 static int log_process_thread(void *args)
 560 {
 561         struct llog_process_args *data = args;
 562         struct llog_ctxt *ctxt = data->llpa_ctxt;
 563         void   *cb = data->llpa_cb;
 564         struct llog_logid logid = *(struct llog_logid *)(data->llpa_arg);
 565         struct llog_handle *llh = NULL;
 566         int rc;
 567         ENTRY;
 568
 569         mutex_up(&data->llpa_sem);
 570         ptlrpc_daemonize("llog_process");     /* thread does IO to log files */
 571
 572         rc = llog_create(ctxt, &llh, &logid, NULL);
 573         if (rc) {
 574                 CERROR("llog_create failed %d\n", rc);
 575                 GOTO(out, rc);
 576         }
 577         rc = llog_init_handle(llh, LLOG_F_IS_CAT, NULL);
 578         if (rc) {
 579                 CERROR("llog_init_handle failed %d\n", rc);
 580                 GOTO(release_llh, rc);
 581         }
 582
 583         if (cb) {
 584                 rc = llog_cat_process(llh, (llog_cb_t)cb, NULL);
 585                 if (rc != LLOG_PROC_BREAK)
 586                         CERROR("llog_cat_process failed %d\n", rc);
 587         } else {
 588                 CWARN("no callback function for recovery\n");
 589         }
 590
 591         CDEBUG(D_HA, "send llcd %p:%p forcibly after recovery\n",
 592                ctxt->loc_llcd, ctxt);
 593         llog_sync(ctxt, NULL);
 594
 595 release_llh:
 596         rc = llog_cat_put(llh);
 597         if (rc)
 598                 CERROR("llog_cat_put failed %d\n", rc);
 599 out:
 600         llog_ctxt_put(ctxt);
 601         RETURN(rc);
 602 }
 603
 604 static int llog_recovery_generic(struct llog_ctxt *ctxt, void *handle,void *arg)
 605 {
 606         struct obd_device *obd = ctxt->loc_obd;
 607         int rc;
 608         ENTRY;
 609
 610         if (obd->obd_stopping)
 611                 RETURN(-ENODEV);
 612
 613         mutex_down(&llpa.llpa_sem);
 614         llpa.llpa_cb = handle;
 615         llpa.llpa_arg = arg;
 616         llpa.llpa_ctxt = llog_ctxt_get(ctxt);
 617         if (!llpa.llpa_ctxt) {
 618                 up(&llpa.llpa_sem);
 619                 RETURN(-ENODEV);
 620         }
 621         rc = cfs_kernel_thread(log_process_thread, &llpa, CLONE_VM | CLONE_FILES);
 622         if (rc < 0) {
 623                 llog_ctxt_put(ctxt);
 624                 CERROR("error starting log_process_thread: %d\n", rc);
 625         } else {
 626                 CDEBUG(D_HA, "log_process_thread: %d\n", rc);
 627                 rc = 0;
 628         }
 629
 630         RETURN(rc);
 631 }
 632
 633 int llog_repl_connect(struct llog_ctxt *ctxt, int count,
 634                       struct llog_logid *logid, struct llog_gen *gen,
 635                       struct obd_uuid *uuid)
 636 {
 637         struct llog_canceld_ctxt *llcd;
 638         int rc;
 639         ENTRY;
 640
 641         /* send back llcd before recovery from llog */
 642         if (ctxt->loc_llcd != NULL) {
 643                 CWARN("llcd %p:%p not empty\n", ctxt->loc_llcd, ctxt);
 644                 llog_sync(ctxt, NULL);
 645         }
 646
 647         mutex_down(&ctxt->loc_sem);
 648         ctxt->loc_gen = *gen;
 649         llcd = ctxt_llcd_grab(ctxt);
 650         if (llcd == NULL) {
 651                 CERROR("couldn't get an llcd\n");
 652                 mutex_up(&ctxt->loc_sem);
 653                 RETURN(-ENOMEM);
 654         }
 655         mutex_up(&ctxt->loc_sem);
 656
 657         rc = llog_recovery_generic(ctxt, ctxt->llog_proc_cb, logid);
 658         if (rc != 0) {
 659                 ctxt_llcd_put(ctxt);
 660                 CERROR("error recovery process: %d\n", rc);
 661         }
 662         RETURN(rc);
 663 }
 664 EXPORT_SYMBOL(llog_repl_connect);
 665
 666 #else /* !__KERNEL__ */
 667
 668 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
 669                          struct lov_stripe_md *lsm, int count,
 670                          struct llog_cookie *cookies, int flags)
 671 {
 672         return 0;
 673 }
 674 #endif