lustre/osp/osp_sync.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osp/osp_sync.c
  37  *
  38  * Lustre OST Proxy Device
  39  *
  40  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  41  * Author: Mikhail Pershin <mike.pershin@intel.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_MDS
  45
  46 #include <lustre_log.h>
  47 #include <lustre_update.h>
  48 #include "osp_internal.h"
  49
  50 static int osp_sync_id_traction_init(struct osp_device *d);
  51 static void osp_sync_id_traction_fini(struct osp_device *d);
  52 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
  53 static void osp_sync_remove_from_tracker(struct osp_device *d);
  54
  55 /*
  56  * this is a components of OSP implementing synchronization between MDS and OST
  57  * it llogs all interesting changes (currently it's uig/gid change and object
  58  * destroy) atomically, then makes sure changes hit OST storage
  59  *
  60  * we have 4 queues of work:
  61  *
  62  * the first queue is llog itself, once read a change is stored in 2nd queue
  63  * in form of RPC (but RPC isn't fired yet).
  64  *
  65  * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
  66  * commit. once change is committed locally it migrates onto 3rd queue.
  67  *
  68  * the third queue (opd_syn_committed_here) holds changes committed locally,
  69  * but not sent to OST (as the pipe can be full). once pipe becomes non-full
  70  * we take a change from the queue and fire corresponded RPC.
  71  *
  72  * once RPC is reported committed by OST (using regular last_committed mech.)
  73  * the change jumps into 4th queue (opd_syn_committed_there), now we can
  74  * cancel corresponded llog record and release RPC
  75  *
  76  * opd_syn_changes is a number of unread llog records (to be processed).
  77  * notice this number doesn't include llog records from previous boots.
  78  * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
  79  *
  80  * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
  81  * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
  82  * too much memory -- how to deal with 1000th OSTs ? batching could help?
  83  *
  84  * opd_syn_rpc_in_flight is a number of RPC in flight.
  85  * we control this with OSP_MAX_IN_FLIGHT
  86  */
  87
  88 /* XXX: do math to learn reasonable threshold
  89  * should it be ~ number of changes fitting bulk? */
  90
  91 #define OSP_SYN_THRESHOLD       10
  92 #define OSP_MAX_IN_FLIGHT       8
  93 #define OSP_MAX_IN_PROGRESS     4096
  94
  95 #define OSP_JOB_MAGIC           0x26112005
  96
  97 static inline int osp_sync_running(struct osp_device *d)
  98 {
  99         return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
 100 }
 101
 102 static inline int osp_sync_stopped(struct osp_device *d)
 103 {
 104         return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
 105 }
 106
 107 static inline int osp_sync_has_new_job(struct osp_device *d)
 108 {
 109         return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
 110                 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
 111                 || (d->opd_syn_prev_done == 0);
 112 }
 113
 114 static inline int osp_sync_low_in_progress(struct osp_device *d)
 115 {
 116         return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
 117 }
 118
 119 static inline int osp_sync_low_in_flight(struct osp_device *d)
 120 {
 121         return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
 122 }
 123
 124 static inline int osp_sync_has_work(struct osp_device *d)
 125 {
 126         /* has new/old changes and low in-progress? */
 127         if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
 128             osp_sync_low_in_flight(d) && d->opd_imp_connected)
 129                 return 1;
 130
 131         /* has remotely committed? */
 132         if (!list_empty(&d->opd_syn_committed_there))
 133                 return 1;
 134
 135         return 0;
 136 }
 137
 138 #define osp_sync_check_for_work(d)                      \
 139 {                                                       \
 140         if (osp_sync_has_work(d)) {                     \
 141                 wake_up(&d->opd_syn_waitq);    \
 142         }                                               \
 143 }
 144
 145 void __osp_sync_check_for_work(struct osp_device *d)
 146 {
 147         osp_sync_check_for_work(d);
 148 }
 149
 150 static inline int osp_sync_can_process_new(struct osp_device *d,
 151                                            struct llog_rec_hdr *rec)
 152 {
 153         LASSERT(d);
 154
 155         if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
 156                 return 0;
 157         if (!osp_sync_low_in_progress(d))
 158                 return 0;
 159         if (!osp_sync_low_in_flight(d))
 160                 return 0;
 161         if (!d->opd_imp_connected)
 162                 return 0;
 163         if (d->opd_syn_prev_done == 0)
 164                 return 1;
 165         if (d->opd_syn_changes == 0)
 166                 return 0;
 167         if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
 168                 return 1;
 169         return 0;
 170 }
 171
 172 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
 173                          llog_op_type type, struct thandle *th)
 174 {
 175         struct osp_thread_info  *osi = osp_env_info(env);
 176         struct osp_device       *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
 177         struct llog_ctxt        *ctxt;
 178         int                      rc;
 179
 180         ENTRY;
 181
 182         /* it's a layering violation, to access internals of th,
 183          * but we can do this as a sanity check, for a while */
 184         LASSERT(th->th_dev == d->opd_storage);
 185
 186         switch (type) {
 187         case MDS_UNLINK64_REC:
 188                 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
 189                 break;
 190         case MDS_SETATTR64_REC:
 191                 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
 192                 break;
 193         default:
 194                 LBUG();
 195         }
 196
 197         /* we want ->dt_trans_start() to allocate per-thandle structure */
 198         th->th_tags |= LCT_OSP_THREAD;
 199
 200         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 201         LASSERT(ctxt);
 202
 203         rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
 204         llog_ctxt_put(ctxt);
 205
 206         RETURN(rc);
 207 }
 208
 209 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
 210                             const struct lu_fid *fid, llog_op_type type,
 211                             int count, struct thandle *th,
 212                             const struct lu_attr *attr)
 213 {
 214         struct osp_thread_info  *osi = osp_env_info(env);
 215         struct llog_ctxt        *ctxt;
 216         struct osp_txn_info     *txn;
 217         int                      rc;
 218
 219         ENTRY;
 220
 221         /* it's a layering violation, to access internals of th,
 222          * but we can do this as a sanity check, for a while */
 223         LASSERT(th->th_dev == d->opd_storage);
 224
 225         switch (type) {
 226         case MDS_UNLINK64_REC:
 227                 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
 228                 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
 229                 osi->osi_unlink.lur_fid  = *fid;
 230                 osi->osi_unlink.lur_count = count;
 231                 break;
 232         case MDS_SETATTR64_REC:
 233                 rc = fid_to_ostid(fid, &osi->osi_oi);
 234                 LASSERT(rc == 0);
 235                 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
 236                 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
 237                 osi->osi_setattr.lsr_oi  = osi->osi_oi;
 238                 LASSERT(attr);
 239                 osi->osi_setattr.lsr_uid = attr->la_uid;
 240                 osi->osi_setattr.lsr_gid = attr->la_gid;
 241                 osi->osi_setattr.lsr_valid =
 242                         ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
 243                         ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0);
 244                 break;
 245         default:
 246                 LBUG();
 247         }
 248
 249         txn = osp_txn_info(&th->th_ctx);
 250         LASSERT(txn);
 251
 252         txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
 253         osi->osi_hdr.lrh_id = txn->oti_current_id;
 254
 255         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 256         if (ctxt == NULL)
 257                 RETURN(-ENOMEM);
 258         rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
 259                       th);
 260         llog_ctxt_put(ctxt);
 261
 262         CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
 263                d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
 264                (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
 265                (unsigned long) osi->osi_cookie.lgc_index, rc);
 266
 267         if (rc > 0)
 268                 rc = 0;
 269
 270         if (likely(rc == 0)) {
 271                 spin_lock(&d->opd_syn_lock);
 272                 d->opd_syn_changes++;
 273                 spin_unlock(&d->opd_syn_lock);
 274         }
 275
 276         RETURN(rc);
 277 }
 278
 279 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
 280                  llog_op_type type, struct thandle *th,
 281                  const struct lu_attr *attr)
 282 {
 283         return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
 284                                 lu_object_fid(&o->opo_obj.do_lu), type, 1,
 285                                 th, attr);
 286 }
 287
 288 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
 289                  struct lu_fid *fid, int lost, struct thandle *th)
 290 {
 291         return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
 292 }
 293
 294 /*
 295  * it's quite obvious we can't maintain all the structures in the memory:
 296  * while OST is down, MDS can be processing thousands and thousands of unlinks
 297  * filling persistent llogs and in-core respresentation
 298  *
 299  * this doesn't scale at all. so we need basically the following:
 300  * a) destroy/setattr append llog records
 301  * b) once llog has grown to X records, we process first Y committed records
 302  *
 303  *  once record R is found via llog_process(), it becomes committed after any
 304  *  subsequent commit callback (at the most)
 305  */
 306
 307 /*
 308  * called for each atomic on-disk change (not once per transaction batch)
 309  * and goes over the list
 310  * XXX: should be optimized?
 311  */
 312
 313 /**
 314  * called for each RPC reported committed
 315  */
 316 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
 317 {
 318         struct osp_device *d = req->rq_cb_data;
 319
 320         CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
 321
 322         if (unlikely(req->rq_transno == 0))
 323                 return;
 324
 325         /* do not do any opd_dyn_rpc_* accounting here
 326          * it's done in osp_sync_interpret sooner or later */
 327
 328         LASSERT(d);
 329         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 330         LASSERT(list_empty(&req->rq_exp_list));
 331
 332         ptlrpc_request_addref(req);
 333
 334         spin_lock(&d->opd_syn_lock);
 335         list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
 336         spin_unlock(&d->opd_syn_lock);
 337
 338         /* XXX: some batching wouldn't hurt */
 339         wake_up(&d->opd_syn_waitq);
 340 }
 341
 342 static int osp_sync_interpret(const struct lu_env *env,
 343                               struct ptlrpc_request *req, void *aa, int rc)
 344 {
 345         struct osp_device *d = req->rq_cb_data;
 346
 347         if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
 348                 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
 349         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 350         LASSERT(d);
 351
 352         CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
 353                atomic_read(&req->rq_refcount),
 354                rc, (unsigned) req->rq_transno);
 355         LASSERT(rc || req->rq_transno);
 356
 357         if (rc == -ENOENT) {
 358                 /*
 359                  * we tried to destroy object or update attributes,
 360                  * but object doesn't exist anymore - cancell llog record
 361                  */
 362                 LASSERT(req->rq_transno == 0);
 363                 LASSERT(list_empty(&req->rq_exp_list));
 364
 365                 ptlrpc_request_addref(req);
 366
 367                 spin_lock(&d->opd_syn_lock);
 368                 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
 369                 spin_unlock(&d->opd_syn_lock);
 370
 371                 wake_up(&d->opd_syn_waitq);
 372         } else if (rc) {
 373                 struct obd_import *imp = req->rq_import;
 374                 /*
 375                  * error happened, we'll try to repeat on next boot ?
 376                  */
 377                 LASSERTF(req->rq_transno == 0 ||
 378                          req->rq_import_generation < imp->imp_generation,
 379                          "transno "LPU64", rc %d, gen: req %d, imp %d\n",
 380                          req->rq_transno, rc, req->rq_import_generation,
 381                          imp->imp_generation);
 382                 if (req->rq_transno == 0) {
 383                         /* this is the last time we see the request
 384                          * if transno is not zero, then commit cb
 385                          * will be called at some point */
 386                         LASSERT(d->opd_syn_rpc_in_progress > 0);
 387                         spin_lock(&d->opd_syn_lock);
 388                         d->opd_syn_rpc_in_progress--;
 389                         spin_unlock(&d->opd_syn_lock);
 390                 }
 391
 392                 wake_up(&d->opd_syn_waitq);
 393         } else if (d->opd_pre != NULL &&
 394                    unlikely(d->opd_pre_status == -ENOSPC)) {
 395                 /*
 396                  * if current status is -ENOSPC (lack of free space on OST)
 397                  * then we should poll OST immediately once object destroy
 398                  * is replied
 399                  */
 400                 osp_statfs_need_now(d);
 401         }
 402
 403         LASSERT(d->opd_syn_rpc_in_flight > 0);
 404         spin_lock(&d->opd_syn_lock);
 405         d->opd_syn_rpc_in_flight--;
 406         spin_unlock(&d->opd_syn_lock);
 407         if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
 408                 wake_up(&d->opd_syn_barrier_waitq);
 409         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 410                d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 411                d->opd_syn_rpc_in_progress);
 412
 413         osp_sync_check_for_work(d);
 414
 415         return 0;
 416 }
 417
 418 /*
 419  * the function walks through list of committed locally changes
 420  * and send them to RPC until the pipe is full
 421  */
 422 static void osp_sync_send_new_rpc(struct osp_device *d,
 423                                   struct ptlrpc_request *req)
 424 {
 425         LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
 426         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 427
 428         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 429 }
 430
 431 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
 432                                                struct llog_handle *llh,
 433                                                struct llog_rec_hdr *h,
 434                                                ost_cmd_t op,
 435                                                const struct req_format *format)
 436 {
 437         struct ptlrpc_request   *req;
 438         struct ost_body         *body;
 439         struct obd_import       *imp;
 440         int                      rc;
 441
 442         /* Prepare the request */
 443         imp = d->opd_obd->u.cli.cl_import;
 444         LASSERT(imp);
 445         req = ptlrpc_request_alloc(imp, format);
 446         if (req == NULL)
 447                 RETURN(ERR_PTR(-ENOMEM));
 448
 449         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
 450         if (rc) {
 451                 ptlrpc_req_finished(req);
 452                 return ERR_PTR(rc);
 453         }
 454
 455         /*
 456          * this is a trick: to save on memory allocations we put cookie
 457          * into the request, but don't set corresponded flag in o_valid
 458          * so that OST doesn't interpret this cookie. once the request
 459          * is committed on OST we take cookie from the request and cancel
 460          */
 461         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 462         LASSERT(body);
 463         body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
 464         body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 465         body->oa.o_lcookie.lgc_index = h->lrh_index;
 466         INIT_LIST_HEAD(&req->rq_exp_list);
 467         req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
 468
 469         req->rq_interpret_reply = osp_sync_interpret;
 470         req->rq_commit_cb = osp_sync_request_commit_cb;
 471         req->rq_cb_data = d;
 472
 473         ptlrpc_request_set_replen(req);
 474
 475         return req;
 476 }
 477
 478 static int osp_sync_new_setattr_job(struct osp_device *d,
 479                                     struct llog_handle *llh,
 480                                     struct llog_rec_hdr *h)
 481 {
 482         struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
 483         struct ptlrpc_request           *req;
 484         struct ost_body                 *body;
 485
 486         ENTRY;
 487         LASSERT(h->lrh_type == MDS_SETATTR64_REC);
 488
 489         /* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
 490          * so no bits other than these should be set. */
 491         if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
 492                 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
 493                        d->opd_obd->obd_name, rec->lsr_valid);
 494                 /* return 0 so that sync thread can continue processing
 495                  * other records. */
 496                 RETURN(0);
 497         }
 498
 499         req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
 500         if (IS_ERR(req))
 501                 RETURN(PTR_ERR(req));
 502
 503         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 504         LASSERT(body);
 505         body->oa.o_oi = rec->lsr_oi;
 506         body->oa.o_uid = rec->lsr_uid;
 507         body->oa.o_gid = rec->lsr_gid;
 508         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 509         /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
 510          * we assume that both UID and GID are valid in that case. */
 511         if (rec->lsr_valid == 0)
 512                 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
 513         else
 514                 body->oa.o_valid |= rec->lsr_valid;
 515
 516         osp_sync_send_new_rpc(d, req);
 517         RETURN(1);
 518 }
 519
 520 static int osp_sync_new_unlink_job(struct osp_device *d,
 521                                    struct llog_handle *llh,
 522                                    struct llog_rec_hdr *h)
 523 {
 524         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 525         struct ptlrpc_request   *req;
 526         struct ost_body         *body;
 527
 528         ENTRY;
 529         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 530
 531         req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
 532         if (IS_ERR(req))
 533                 RETURN(PTR_ERR(req));
 534
 535         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 536         LASSERT(body);
 537         ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
 538         ostid_set_id(&body->oa.o_oi, rec->lur_oid);
 539         body->oa.o_misc = rec->lur_count;
 540         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 541         if (rec->lur_count)
 542                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 543
 544         osp_sync_send_new_rpc(d, req);
 545         RETURN(1);
 546 }
 547
 548 static int osp_prep_unlink_update_req(const struct lu_env *env,
 549                                       struct osp_device *osp,
 550                                       struct llog_handle *llh,
 551                                       struct llog_rec_hdr *h,
 552                                       struct ptlrpc_request **reqp)
 553 {
 554         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 555         struct dt_update_request        *update = NULL;
 556         struct ptlrpc_request           *req;
 557         struct llog_cookie              lcookie;
 558         const void                      *buf;
 559         __u16                           size;
 560         int                             rc;
 561         ENTRY;
 562
 563         update = dt_update_request_create(&osp->opd_dt_dev);
 564         if (IS_ERR(update))
 565                 RETURN(PTR_ERR(update));
 566
 567         /* This can only happens for unlink slave directory, so decrease
 568          * ref for ".." and "." */
 569         rc = out_update_pack(env, &update->dur_buf, OUT_REF_DEL, &rec->lur_fid,
 570                              0, NULL, NULL, 0);
 571         if (rc != 0)
 572                 GOTO(out, rc);
 573
 574         rc = out_update_pack(env, &update->dur_buf, OUT_REF_DEL, &rec->lur_fid,
 575                              0, NULL, NULL, 0);
 576         if (rc != 0)
 577                 GOTO(out, rc);
 578
 579         lcookie.lgc_lgl = llh->lgh_id;
 580         lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 581         lcookie.lgc_index = h->lrh_index;
 582         size = sizeof(lcookie);
 583         buf = &lcookie;
 584
 585         rc = out_update_pack(env, &update->dur_buf, OUT_DESTROY, &rec->lur_fid,
 586                              1, &size, &buf, 0);
 587         if (rc != 0)
 588                 GOTO(out, rc);
 589
 590         rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
 591                                  update->dur_buf.ub_req, &req);
 592         if (rc != 0)
 593                 GOTO(out, rc);
 594
 595         INIT_LIST_HEAD(&req->rq_exp_list);
 596         req->rq_svc_thread = (void *)OSP_JOB_MAGIC;
 597
 598         req->rq_interpret_reply = osp_sync_interpret;
 599         req->rq_commit_cb = osp_sync_request_commit_cb;
 600         req->rq_cb_data = osp;
 601
 602         ptlrpc_request_set_replen(req);
 603         *reqp = req;
 604 out:
 605         if (update != NULL)
 606                 dt_update_request_destroy(update);
 607
 608         RETURN(rc);
 609 }
 610
 611 static int osp_sync_new_unlink64_job(const struct lu_env *env,
 612                                      struct osp_device *d,
 613                                      struct llog_handle *llh,
 614                                      struct llog_rec_hdr *h)
 615 {
 616         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 617         struct ptlrpc_request           *req = NULL;
 618         struct ost_body                 *body;
 619         int                              rc;
 620
 621         ENTRY;
 622         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 623
 624         if (d->opd_connect_mdt) {
 625                 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
 626                 if (rc != 0)
 627                         RETURN(rc);
 628         } else {
 629                 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
 630                                        &RQF_OST_DESTROY);
 631                 if (IS_ERR(req))
 632                         RETURN(PTR_ERR(req));
 633
 634                 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 635                 if (body == NULL)
 636                         RETURN(-EFAULT);
 637                 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
 638                 if (rc < 0)
 639                         RETURN(rc);
 640                 body->oa.o_misc = rec->lur_count;
 641                 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 642                                    OBD_MD_FLOBJCOUNT;
 643         }
 644         osp_sync_send_new_rpc(d, req);
 645         RETURN(1);
 646 }
 647
 648 static int osp_sync_process_record(const struct lu_env *env,
 649                                    struct osp_device *d,
 650                                    struct llog_handle *llh,
 651                                    struct llog_rec_hdr *rec)
 652 {
 653         struct llog_cookie       cookie;
 654         int                      rc = 0;
 655
 656         cookie.lgc_lgl = llh->lgh_id;
 657         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 658         cookie.lgc_index = rec->lrh_index;
 659
 660         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 661                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 662
 663                 /* we're waiting for the record generated by this instance */
 664                 LASSERT(d->opd_syn_prev_done == 0);
 665                 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
 666                             sizeof(gen->lgr_gen))) {
 667                         CDEBUG(D_HA, "processed all old entries\n");
 668                         d->opd_syn_prev_done = 1;
 669                 }
 670
 671                 /* cancel any generation record */
 672                 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
 673                                              1, &cookie);
 674
 675                 return rc;
 676         }
 677
 678         /*
 679          * now we prepare and fill requests to OST, put them on the queue
 680          * and fire after next commit callback
 681          */
 682
 683         /* notice we increment counters before sending RPC, to be consistent
 684          * in RPC interpret callback which may happen very quickly */
 685         spin_lock(&d->opd_syn_lock);
 686         d->opd_syn_rpc_in_flight++;
 687         d->opd_syn_rpc_in_progress++;
 688         spin_unlock(&d->opd_syn_lock);
 689
 690         switch (rec->lrh_type) {
 691         /* case MDS_UNLINK_REC is kept for compatibility */
 692         case MDS_UNLINK_REC:
 693                 rc = osp_sync_new_unlink_job(d, llh, rec);
 694                 break;
 695         case MDS_UNLINK64_REC:
 696                 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
 697                 break;
 698         case MDS_SETATTR64_REC:
 699                 rc = osp_sync_new_setattr_job(d, llh, rec);
 700                 break;
 701         default:
 702                 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
 703                        rec->lrh_type);
 704                 /* we should continue processing */
 705         }
 706
 707         /* rc > 0 means sync RPC being added to the queue */
 708         if (likely(rc > 0)) {
 709                 spin_lock(&d->opd_syn_lock);
 710                 if (d->opd_syn_prev_done) {
 711                         LASSERT(d->opd_syn_changes > 0);
 712                         LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
 713                         /*
 714                          * NOTE: it's possible to meet same id if
 715                          * OST stores few stripes of same file
 716                          */
 717                         if (rec->lrh_id > d->opd_syn_last_processed_id) {
 718                                 d->opd_syn_last_processed_id = rec->lrh_id;
 719                                 wake_up(&d->opd_syn_barrier_waitq);
 720                         }
 721
 722                         d->opd_syn_changes--;
 723                 }
 724                 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 725                        d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 726                        d->opd_syn_rpc_in_progress);
 727                 spin_unlock(&d->opd_syn_lock);
 728                 rc = 0;
 729         } else {
 730                 spin_lock(&d->opd_syn_lock);
 731                 d->opd_syn_rpc_in_flight--;
 732                 d->opd_syn_rpc_in_progress--;
 733                 spin_unlock(&d->opd_syn_lock);
 734         }
 735
 736         CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
 737                rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
 738         return rc;
 739 }
 740
 741 static void osp_sync_process_committed(const struct lu_env *env,
 742                                        struct osp_device *d)
 743 {
 744         struct obd_device       *obd = d->opd_obd;
 745         struct obd_import       *imp = obd->u.cli.cl_import;
 746         struct ost_body         *body;
 747         struct ptlrpc_request   *req, *tmp;
 748         struct llog_ctxt        *ctxt;
 749         struct llog_handle      *llh;
 750         struct list_head         list;
 751         int                      rc, done = 0;
 752
 753         ENTRY;
 754
 755         if (list_empty(&d->opd_syn_committed_there))
 756                 return;
 757
 758         /*
 759          * if current status is -ENOSPC (lack of free space on OST)
 760          * then we should poll OST immediately once object destroy
 761          * is committed.
 762          * notice: we do this upon commit as well because some backends
 763          * (like DMU) do not release space right away.
 764          */
 765         if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
 766                 osp_statfs_need_now(d);
 767
 768         /*
 769          * now cancel them all
 770          * XXX: can we improve this using some batching?
 771          *      with batch RPC that'll happen automatically?
 772          * XXX: can we store ctxt in lod_device and save few cycles ?
 773          */
 774         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
 775         LASSERT(ctxt);
 776
 777         llh = ctxt->loc_handle;
 778         LASSERT(llh);
 779
 780         INIT_LIST_HEAD(&list);
 781         spin_lock(&d->opd_syn_lock);
 782         list_splice(&d->opd_syn_committed_there, &list);
 783         INIT_LIST_HEAD(&d->opd_syn_committed_there);
 784         spin_unlock(&d->opd_syn_lock);
 785
 786         list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
 787                 struct llog_cookie *lcookie = NULL;
 788
 789                 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 790                 list_del_init(&req->rq_exp_list);
 791
 792                 if (d->opd_connect_mdt) {
 793                         struct object_update_request *ureq;
 794                         struct object_update *update;
 795                         ureq = req_capsule_client_get(&req->rq_pill,
 796                                                       &RMF_OUT_UPDATE);
 797                         LASSERT(ureq != NULL &&
 798                                 ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
 799
 800                         /* 1st/2nd is for decref . and .., 3rd one is for
 801                          * destroy, where the log cookie is stored.
 802                          * See osp_prep_unlink_update_req */
 803                         update = object_update_request_get(ureq, 2, NULL);
 804                         LASSERT(update != NULL);
 805                         lcookie = object_update_param_get(update, 0, NULL);
 806                         LASSERT(lcookie != NULL);
 807                 } else {
 808                         body = req_capsule_client_get(&req->rq_pill,
 809                                                       &RMF_OST_BODY);
 810                         LASSERT(body);
 811                         lcookie = &body->oa.o_lcookie;
 812                 }
 813                 /* import can be closing, thus all commit cb's are
 814                  * called we can check committness directly */
 815                 if (req->rq_transno <= imp->imp_peer_committed_transno) {
 816                         rc = llog_cat_cancel_records(env, llh, 1, lcookie);
 817                         if (rc)
 818                                 CERROR("%s: can't cancel record: %d\n",
 819                                        obd->obd_name, rc);
 820                 } else {
 821                         DEBUG_REQ(D_HA, req, "not committed");
 822                 }
 823
 824                 ptlrpc_req_finished(req);
 825                 done++;
 826         }
 827
 828         llog_ctxt_put(ctxt);
 829
 830         LASSERT(d->opd_syn_rpc_in_progress >= done);
 831         spin_lock(&d->opd_syn_lock);
 832         d->opd_syn_rpc_in_progress -= done;
 833         spin_unlock(&d->opd_syn_lock);
 834         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 835                d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 836                d->opd_syn_rpc_in_progress);
 837
 838         osp_sync_check_for_work(d);
 839
 840         /* wake up the thread if requested to stop:
 841          * it might be waiting for in-progress to complete */
 842         if (unlikely(osp_sync_running(d) == 0))
 843                 wake_up(&d->opd_syn_waitq);
 844
 845         EXIT;
 846 }
 847
 848 /*
 849  * this is where most of queues processing happens
 850  */
 851 static int osp_sync_process_queues(const struct lu_env *env,
 852                                    struct llog_handle *llh,
 853                                    struct llog_rec_hdr *rec,
 854                                    void *data)
 855 {
 856         struct osp_device       *d = data;
 857         int                      rc;
 858
 859         do {
 860                 struct l_wait_info lwi = { 0 };
 861
 862                 if (!osp_sync_running(d)) {
 863                         CDEBUG(D_HA, "stop llog processing\n");
 864                         return LLOG_PROC_BREAK;
 865                 }
 866
 867                 /* process requests committed by OST */
 868                 osp_sync_process_committed(env, d);
 869
 870                 /* if we there are changes to be processed and we have
 871                  * resources for this ... do now */
 872                 if (osp_sync_can_process_new(d, rec)) {
 873                         if (llh == NULL) {
 874                                 /* ask llog for another record */
 875                                 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
 876                                        d->opd_syn_changes,
 877                                        d->opd_syn_rpc_in_progress,
 878                                        d->opd_syn_rpc_in_flight);
 879                                 return 0;
 880                         }
 881
 882                         /*
 883                          * try to send, in case of disconnection, suspend
 884                          * processing till we can send this request
 885                          */
 886                         do {
 887                                 rc = osp_sync_process_record(env, d, llh, rec);
 888                                 /*
 889                                  * XXX: probably different handling is needed
 890                                  * for some bugs, like immediate exit or if
 891                                  * OSP gets inactive
 892                                  */
 893                                 if (rc) {
 894                                         CERROR("can't send: %d\n", rc);
 895                                         l_wait_event(d->opd_syn_waitq,
 896                                                      !osp_sync_running(d) ||
 897                                                      osp_sync_has_work(d),
 898                                                      &lwi);
 899                                 }
 900                         } while (rc != 0 && osp_sync_running(d));
 901
 902                         llh = NULL;
 903                         rec = NULL;
 904                 }
 905
 906                 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
 907                         osp_sync_remove_from_tracker(d);
 908
 909                 l_wait_event(d->opd_syn_waitq,
 910                              !osp_sync_running(d) ||
 911                              osp_sync_can_process_new(d, rec) ||
 912                              !list_empty(&d->opd_syn_committed_there),
 913                              &lwi);
 914         } while (1);
 915 }
 916
 917 /*
 918  * this thread runs llog_cat_process() scanner calling our callback
 919  * to process llog records. in the callback we implement tricky
 920  * state machine as we don't want to start scanning of the llog again
 921  * and again, also we don't want to process too many records and send
 922  * too many RPCs a time. so, depending on current load (num of changes
 923  * being synced to OST) the callback can suspend awaiting for some
 924  * new conditions, like syncs completed.
 925  *
 926  * in order to process llog records left by previous boots and to allow
 927  * llog_process_thread() to find something (otherwise it'd just exit
 928  * immediately) we add a special GENERATATION record on each boot.
 929  */
 930 static int osp_sync_thread(void *_arg)
 931 {
 932         struct osp_device       *d = _arg;
 933         struct ptlrpc_thread    *thread = &d->opd_syn_thread;
 934         struct l_wait_info       lwi = { 0 };
 935         struct llog_ctxt        *ctxt;
 936         struct obd_device       *obd = d->opd_obd;
 937         struct llog_handle      *llh;
 938         struct lu_env            env;
 939         int                      rc, count;
 940
 941         ENTRY;
 942
 943         rc = lu_env_init(&env, LCT_LOCAL);
 944         if (rc) {
 945                 CERROR("%s: can't initialize env: rc = %d\n",
 946                        obd->obd_name, rc);
 947                 RETURN(rc);
 948         }
 949
 950         spin_lock(&d->opd_syn_lock);
 951         thread->t_flags = SVC_RUNNING;
 952         spin_unlock(&d->opd_syn_lock);
 953         wake_up(&thread->t_ctl_waitq);
 954
 955         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
 956         if (ctxt == NULL) {
 957                 CERROR("can't get appropriate context\n");
 958                 GOTO(out, rc = -EINVAL);
 959         }
 960
 961         llh = ctxt->loc_handle;
 962         if (llh == NULL) {
 963                 CERROR("can't get llh\n");
 964                 llog_ctxt_put(ctxt);
 965                 GOTO(out, rc = -EINVAL);
 966         }
 967
 968         rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
 969         LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
 970                  "%lu changes, %u in progress, %u in flight: %d\n",
 971                  d->opd_syn_changes, d->opd_syn_rpc_in_progress,
 972                  d->opd_syn_rpc_in_flight, rc);
 973
 974         /* we don't expect llog_process_thread() to exit till umount */
 975         LASSERTF(thread->t_flags != SVC_RUNNING,
 976                  "%lu changes, %u in progress, %u in flight\n",
 977                  d->opd_syn_changes, d->opd_syn_rpc_in_progress,
 978                  d->opd_syn_rpc_in_flight);
 979
 980         /* wait till all the requests are completed */
 981         count = 0;
 982         while (d->opd_syn_rpc_in_progress > 0) {
 983                 osp_sync_process_committed(&env, d);
 984
 985                 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
 986                 rc = l_wait_event(d->opd_syn_waitq,
 987                                   d->opd_syn_rpc_in_progress == 0,
 988                                   &lwi);
 989                 if (rc == -ETIMEDOUT)
 990                         count++;
 991                 LASSERTF(count < 10, "%s: %d %d %sempty\n",
 992                          d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
 993                          d->opd_syn_rpc_in_flight,
 994                          list_empty(&d->opd_syn_committed_there) ? "" : "!");
 995
 996         }
 997
 998         llog_cat_close(&env, llh);
 999         rc = llog_cleanup(&env, ctxt);
1000         if (rc)
1001                 CERROR("can't cleanup llog: %d\n", rc);
1002 out:
1003         LASSERTF(d->opd_syn_rpc_in_progress == 0,
1004                  "%s: %d %d %sempty\n",
1005                  d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1006                  d->opd_syn_rpc_in_flight,
1007                  list_empty(&d->opd_syn_committed_there) ? "" : "!");
1008
1009         thread->t_flags = SVC_STOPPED;
1010
1011         wake_up(&thread->t_ctl_waitq);
1012
1013         lu_env_fini(&env);
1014
1015         RETURN(0);
1016 }
1017
1018 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1019 {
1020         struct osp_thread_info  *osi = osp_env_info(env);
1021         struct lu_fid           *fid = &osi->osi_fid;
1022         struct llog_handle      *lgh = NULL;
1023         struct obd_device       *obd = d->opd_obd;
1024         struct llog_ctxt        *ctxt;
1025         int                     rc;
1026
1027         ENTRY;
1028
1029         LASSERT(obd);
1030
1031         /*
1032          * open llog corresponding to our OST
1033          */
1034         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1035         obd->obd_lvfs_ctxt.dt = d->opd_storage;
1036
1037         if (d->opd_connect_mdt)
1038                 lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
1039         else
1040                 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1041
1042         rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1043                                    &osi->osi_cid, fid);
1044         if (rc) {
1045                 CERROR("%s: can't get id from catalogs: rc = %d\n",
1046                        obd->obd_name, rc);
1047                 RETURN(rc);
1048         }
1049
1050         CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1051                obd->obd_name, d->opd_index,
1052                POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1053                osi->osi_cid.lci_logid.lgl_ogen);
1054
1055         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1056                         &osp_mds_ost_orig_logops);
1057         if (rc)
1058                 RETURN(rc);
1059
1060         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1061         LASSERT(ctxt);
1062
1063         if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1064                 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1065                                LLOG_OPEN_EXISTS);
1066                 /* re-create llog if it is missing */
1067                 if (rc == -ENOENT)
1068                         logid_set_id(&osi->osi_cid.lci_logid, 0);
1069                 else if (rc < 0)
1070                         GOTO(out_cleanup, rc);
1071         }
1072
1073         if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1074                 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1075                 if (rc < 0)
1076                         GOTO(out_cleanup, rc);
1077                 osi->osi_cid.lci_logid = lgh->lgh_id;
1078         }
1079
1080         LASSERT(lgh != NULL);
1081         ctxt->loc_handle = lgh;
1082
1083         rc = llog_cat_init_and_process(env, lgh);
1084         if (rc)
1085                 GOTO(out_close, rc);
1086
1087         rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1088                                    &osi->osi_cid, fid);
1089         if (rc)
1090                 GOTO(out_close, rc);
1091
1092         /*
1093          * put a mark in the llog till which we'll be processing
1094          * old records restless
1095          */
1096         d->opd_syn_generation.mnt_cnt = cfs_time_current();
1097         d->opd_syn_generation.conn_cnt = cfs_time_current();
1098
1099         osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1100         osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1101
1102         memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1103                sizeof(osi->osi_gen.lgr_gen));
1104
1105         rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1106         if (rc < 0)
1107                 GOTO(out_close, rc);
1108         llog_ctxt_put(ctxt);
1109         RETURN(0);
1110 out_close:
1111         llog_cat_close(env, lgh);
1112 out_cleanup:
1113         llog_cleanup(env, ctxt);
1114         RETURN(rc);
1115 }
1116
1117 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1118 {
1119         struct llog_ctxt *ctxt;
1120
1121         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1122         if (ctxt != NULL)
1123                 llog_cat_close(env, ctxt->loc_handle);
1124         llog_cleanup(env, ctxt);
1125 }
1126
1127 /*
1128  * initializes sync component of OSP
1129  */
1130 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1131 {
1132         struct l_wait_info       lwi = { 0 };
1133         struct task_struct      *task;
1134         int                      rc;
1135
1136         ENTRY;
1137
1138         rc = osp_sync_id_traction_init(d);
1139         if (rc)
1140                 RETURN(rc);
1141
1142         /*
1143          * initialize llog storing changes
1144          */
1145         rc = osp_sync_llog_init(env, d);
1146         if (rc) {
1147                 CERROR("%s: can't initialize llog: rc = %d\n",
1148                        d->opd_obd->obd_name, rc);
1149                 GOTO(err_id, rc);
1150         }
1151
1152         /*
1153          * Start synchronization thread
1154          */
1155         d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1156         d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1157         spin_lock_init(&d->opd_syn_lock);
1158         init_waitqueue_head(&d->opd_syn_waitq);
1159         init_waitqueue_head(&d->opd_syn_barrier_waitq);
1160         init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1161         INIT_LIST_HEAD(&d->opd_syn_committed_there);
1162
1163         task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1164                            d->opd_index, d->opd_group);
1165         if (IS_ERR(task)) {
1166                 rc = PTR_ERR(task);
1167                 CERROR("%s: cannot start sync thread: rc = %d\n",
1168                        d->opd_obd->obd_name, rc);
1169                 GOTO(err_llog, rc);
1170         }
1171
1172         l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1173                      osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1174
1175         RETURN(0);
1176 err_llog:
1177         osp_sync_llog_fini(env, d);
1178 err_id:
1179         osp_sync_id_traction_fini(d);
1180         return rc;
1181 }
1182
1183 int osp_sync_fini(struct osp_device *d)
1184 {
1185         struct ptlrpc_thread *thread = &d->opd_syn_thread;
1186
1187         ENTRY;
1188
1189         thread->t_flags = SVC_STOPPING;
1190         wake_up(&d->opd_syn_waitq);
1191         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1192
1193         /*
1194          * unregister transaction callbacks only when sync thread
1195          * has finished operations with llog
1196          */
1197         osp_sync_id_traction_fini(d);
1198
1199         RETURN(0);
1200 }
1201
1202 static DEFINE_MUTEX(osp_id_tracker_sem);
1203 static struct list_head osp_id_tracker_list =
1204                 LIST_HEAD_INIT(osp_id_tracker_list);
1205
1206 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1207 {
1208         struct osp_id_tracker   *tr = cookie;
1209         struct osp_device       *d;
1210         struct osp_txn_info     *txn;
1211
1212         LASSERT(tr);
1213
1214         txn = osp_txn_info(&th->th_ctx);
1215         if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1216                 return;
1217
1218         spin_lock(&tr->otr_lock);
1219         if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1220                 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1221                        tr->otr_committed_id, txn->oti_current_id);
1222                 tr->otr_committed_id = txn->oti_current_id;
1223
1224                 list_for_each_entry(d, &tr->otr_wakeup_list,
1225                                     opd_syn_ontrack) {
1226                         d->opd_syn_last_committed_id = tr->otr_committed_id;
1227                         wake_up(&d->opd_syn_waitq);
1228                 }
1229         }
1230         spin_unlock(&tr->otr_lock);
1231 }
1232
1233 static int osp_sync_id_traction_init(struct osp_device *d)
1234 {
1235         struct osp_id_tracker   *tr, *found = NULL;
1236         int                      rc = 0;
1237
1238         LASSERT(d);
1239         LASSERT(d->opd_storage);
1240         LASSERT(d->opd_syn_tracker == NULL);
1241         INIT_LIST_HEAD(&d->opd_syn_ontrack);
1242
1243         mutex_lock(&osp_id_tracker_sem);
1244         list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1245                 if (tr->otr_dev == d->opd_storage) {
1246                         LASSERT(atomic_read(&tr->otr_refcount));
1247                         atomic_inc(&tr->otr_refcount);
1248                         d->opd_syn_tracker = tr;
1249                         found = tr;
1250                         break;
1251                 }
1252         }
1253
1254         if (found == NULL) {
1255                 rc = -ENOMEM;
1256                 OBD_ALLOC_PTR(tr);
1257                 if (tr) {
1258                         d->opd_syn_tracker = tr;
1259                         spin_lock_init(&tr->otr_lock);
1260                         tr->otr_dev = d->opd_storage;
1261                         tr->otr_next_id = 1;
1262                         tr->otr_committed_id = 0;
1263                         atomic_set(&tr->otr_refcount, 1);
1264                         INIT_LIST_HEAD(&tr->otr_wakeup_list);
1265                         list_add(&tr->otr_list, &osp_id_tracker_list);
1266                         tr->otr_tx_cb.dtc_txn_commit =
1267                                                 osp_sync_tracker_commit_cb;
1268                         tr->otr_tx_cb.dtc_cookie = tr;
1269                         tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1270                         dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1271                         rc = 0;
1272                 }
1273         }
1274         mutex_unlock(&osp_id_tracker_sem);
1275
1276         return rc;
1277 }
1278
1279 static void osp_sync_id_traction_fini(struct osp_device *d)
1280 {
1281         struct osp_id_tracker *tr;
1282
1283         ENTRY;
1284
1285         LASSERT(d);
1286         tr = d->opd_syn_tracker;
1287         if (tr == NULL) {
1288                 EXIT;
1289                 return;
1290         }
1291
1292         osp_sync_remove_from_tracker(d);
1293
1294         mutex_lock(&osp_id_tracker_sem);
1295         if (atomic_dec_and_test(&tr->otr_refcount)) {
1296                 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1297                 LASSERT(list_empty(&tr->otr_wakeup_list));
1298                 list_del(&tr->otr_list);
1299                 OBD_FREE_PTR(tr);
1300                 d->opd_syn_tracker = NULL;
1301         }
1302         mutex_unlock(&osp_id_tracker_sem);
1303
1304         EXIT;
1305 }
1306
1307 /*
1308  * generates id for the tracker
1309  */
1310 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1311 {
1312         struct osp_id_tracker *tr;
1313
1314         tr = d->opd_syn_tracker;
1315         LASSERT(tr);
1316
1317         /* XXX: we can improve this introducing per-cpu preallocated ids? */
1318         spin_lock(&tr->otr_lock);
1319         if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1320                 spin_unlock(&tr->otr_lock);
1321                 CERROR("%s: next %u, last synced %lu\n",
1322                        d->opd_obd->obd_name, tr->otr_next_id,
1323                        d->opd_syn_last_used_id);
1324                 LBUG();
1325         }
1326
1327         if (id == 0)
1328                 id = tr->otr_next_id++;
1329         if (id > d->opd_syn_last_used_id)
1330                 d->opd_syn_last_used_id = id;
1331         if (list_empty(&d->opd_syn_ontrack))
1332                 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1333         spin_unlock(&tr->otr_lock);
1334         CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1335
1336         return id;
1337 }
1338
1339 static void osp_sync_remove_from_tracker(struct osp_device *d)
1340 {
1341         struct osp_id_tracker *tr;
1342
1343         tr = d->opd_syn_tracker;
1344         LASSERT(tr);
1345
1346         if (list_empty(&d->opd_syn_ontrack))
1347                 return;
1348
1349         spin_lock(&tr->otr_lock);
1350         list_del_init(&d->opd_syn_ontrack);
1351         spin_unlock(&tr->otr_lock);
1352 }
1353