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 "osp_internal.h"
  48
  49 static int osp_sync_id_traction_init(struct osp_device *d);
  50 static void osp_sync_id_traction_fini(struct osp_device *d);
  51 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
  52 static void osp_sync_remove_from_tracker(struct osp_device *d);
  53
  54 /*
  55  * this is a components of OSP implementing synchronization between MDS and OST
  56  * it llogs all interesting changes (currently it's uig/gid change and object
  57  * destroy) atomically, then makes sure changes hit OST storage
  58  *
  59  * we have 4 queues of work:
  60  *
  61  * the first queue is llog itself, once read a change is stored in 2nd queue
  62  * in form of RPC (but RPC isn't fired yet).
  63  *
  64  * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
  65  * commit. once change is committed locally it migrates onto 3rd queue.
  66  *
  67  * the third queue (opd_syn_committed_here) holds changes committed locally,
  68  * but not sent to OST (as the pipe can be full). once pipe becomes non-full
  69  * we take a change from the queue and fire corresponded RPC.
  70  *
  71  * once RPC is reported committed by OST (using regular last_committed mech.)
  72  * the change jumps into 4th queue (opd_syn_committed_there), now we can
  73  * cancel corresponded llog record and release RPC
  74  *
  75  * opd_syn_changes is a number of unread llog records (to be processed).
  76  * notice this number doesn't include llog records from previous boots.
  77  * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
  78  *
  79  * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
  80  * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
  81  * too much memory -- how to deal with 1000th OSTs ? batching could help?
  82  *
  83  * opd_syn_rpc_in_flight is a number of RPC in flight.
  84  * we control this with OSP_MAX_IN_FLIGHT
  85  */
  86
  87 /* XXX: do math to learn reasonable threshold
  88  * should it be ~ number of changes fitting bulk? */
  89
  90 #define OSP_SYN_THRESHOLD       10
  91 #define OSP_MAX_IN_FLIGHT       8
  92 #define OSP_MAX_IN_PROGRESS     4096
  93
  94 #define OSP_JOB_MAGIC           0x26112005
  95
  96 static inline int osp_sync_running(struct osp_device *d)
  97 {
  98         return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
  99 }
 100
 101 static inline int osp_sync_stopped(struct osp_device *d)
 102 {
 103         return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
 104 }
 105
 106 static inline int osp_sync_has_new_job(struct osp_device *d)
 107 {
 108         return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
 109                 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
 110                 || (d->opd_syn_prev_done == 0);
 111 }
 112
 113 static inline int osp_sync_low_in_progress(struct osp_device *d)
 114 {
 115         return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
 116 }
 117
 118 static inline int osp_sync_low_in_flight(struct osp_device *d)
 119 {
 120         return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
 121 }
 122
 123 static inline int osp_sync_has_work(struct osp_device *d)
 124 {
 125         /* has new/old changes and low in-progress? */
 126         if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
 127             osp_sync_low_in_flight(d) && d->opd_imp_connected)
 128                 return 1;
 129
 130         /* has remotely committed? */
 131         if (!list_empty(&d->opd_syn_committed_there))
 132                 return 1;
 133
 134         return 0;
 135 }
 136
 137 #define osp_sync_check_for_work(d)                      \
 138 {                                                       \
 139         if (osp_sync_has_work(d)) {                     \
 140                 wake_up(&d->opd_syn_waitq);    \
 141         }                                               \
 142 }
 143
 144 void __osp_sync_check_for_work(struct osp_device *d)
 145 {
 146         osp_sync_check_for_work(d);
 147 }
 148
 149 static inline int osp_sync_can_process_new(struct osp_device *d,
 150                                            struct llog_rec_hdr *rec)
 151 {
 152         LASSERT(d);
 153
 154         if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
 155                 return 0;
 156         if (!osp_sync_low_in_progress(d))
 157                 return 0;
 158         if (!osp_sync_low_in_flight(d))
 159                 return 0;
 160         if (!d->opd_imp_connected)
 161                 return 0;
 162         if (d->opd_syn_prev_done == 0)
 163                 return 1;
 164         if (d->opd_syn_changes == 0)
 165                 return 0;
 166         if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
 167                 return 1;
 168         return 0;
 169 }
 170
 171 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
 172                          llog_op_type type, struct thandle *th)
 173 {
 174         struct osp_thread_info  *osi = osp_env_info(env);
 175         struct osp_device       *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
 176         struct llog_ctxt        *ctxt;
 177         int                      rc;
 178
 179         ENTRY;
 180
 181         /* it's a layering violation, to access internals of th,
 182          * but we can do this as a sanity check, for a while */
 183         LASSERT(th->th_dev == d->opd_storage);
 184
 185         switch (type) {
 186         case MDS_UNLINK64_REC:
 187                 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
 188                 break;
 189         case MDS_SETATTR64_REC:
 190                 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
 191                 break;
 192         default:
 193                 LBUG();
 194         }
 195
 196         /* we want ->dt_trans_start() to allocate per-thandle structure */
 197         th->th_tags |= LCT_OSP_THREAD;
 198
 199         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 200         LASSERT(ctxt);
 201
 202         rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
 203         llog_ctxt_put(ctxt);
 204
 205         RETURN(rc);
 206 }
 207
 208 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
 209                             const struct lu_fid *fid, llog_op_type type,
 210                             int count, struct thandle *th,
 211                             const struct lu_attr *attr)
 212 {
 213         struct osp_thread_info  *osi = osp_env_info(env);
 214         struct llog_ctxt        *ctxt;
 215         struct osp_txn_info     *txn;
 216         int                      rc;
 217
 218         ENTRY;
 219
 220         /* it's a layering violation, to access internals of th,
 221          * but we can do this as a sanity check, for a while */
 222         LASSERT(th->th_dev == d->opd_storage);
 223
 224         switch (type) {
 225         case MDS_UNLINK64_REC:
 226                 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
 227                 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
 228                 osi->osi_unlink.lur_fid  = *fid;
 229                 osi->osi_unlink.lur_count = count;
 230                 break;
 231         case MDS_SETATTR64_REC:
 232                 rc = fid_to_ostid(fid, &osi->osi_oi);
 233                 LASSERT(rc == 0);
 234                 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
 235                 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
 236                 osi->osi_setattr.lsr_oi  = osi->osi_oi;
 237                 LASSERT(attr);
 238                 osi->osi_setattr.lsr_uid = attr->la_uid;
 239                 osi->osi_setattr.lsr_gid = attr->la_gid;
 240                 osi->osi_setattr.lsr_valid = attr->la_valid;
 241                 break;
 242         default:
 243                 LBUG();
 244         }
 245
 246         txn = osp_txn_info(&th->th_ctx);
 247         LASSERT(txn);
 248
 249         txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
 250         osi->osi_hdr.lrh_id = txn->oti_current_id;
 251
 252         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 253         if (ctxt == NULL)
 254                 RETURN(-ENOMEM);
 255         rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
 256                       th);
 257         llog_ctxt_put(ctxt);
 258
 259         CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
 260                d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
 261                (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
 262                (unsigned long) osi->osi_cookie.lgc_index, rc);
 263
 264         if (rc > 0)
 265                 rc = 0;
 266
 267         if (likely(rc == 0)) {
 268                 spin_lock(&d->opd_syn_lock);
 269                 d->opd_syn_changes++;
 270                 spin_unlock(&d->opd_syn_lock);
 271         }
 272
 273         RETURN(rc);
 274 }
 275
 276 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
 277                  llog_op_type type, struct thandle *th,
 278                  const struct lu_attr *attr)
 279 {
 280         return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
 281                                 lu_object_fid(&o->opo_obj.do_lu), type, 1,
 282                                 th, attr);
 283 }
 284
 285 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
 286                  struct lu_fid *fid, int lost, struct thandle *th)
 287 {
 288         return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
 289 }
 290
 291 /*
 292  * it's quite obvious we can't maintain all the structures in the memory:
 293  * while OST is down, MDS can be processing thousands and thousands of unlinks
 294  * filling persistent llogs and in-core respresentation
 295  *
 296  * this doesn't scale at all. so we need basically the following:
 297  * a) destroy/setattr append llog records
 298  * b) once llog has grown to X records, we process first Y committed records
 299  *
 300  *  once record R is found via llog_process(), it becomes committed after any
 301  *  subsequent commit callback (at the most)
 302  */
 303
 304 /*
 305  * called for each atomic on-disk change (not once per transaction batch)
 306  * and goes over the list
 307  * XXX: should be optimized?
 308  */
 309
 310 /**
 311  * called for each RPC reported committed
 312  */
 313 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
 314 {
 315         struct osp_device *d = req->rq_cb_data;
 316
 317         CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
 318
 319         if (unlikely(req->rq_transno == 0))
 320                 return;
 321
 322         /* do not do any opd_dyn_rpc_* accounting here
 323          * it's done in osp_sync_interpret sooner or later */
 324
 325         LASSERT(d);
 326         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 327         LASSERT(list_empty(&req->rq_exp_list));
 328
 329         ptlrpc_request_addref(req);
 330
 331         spin_lock(&d->opd_syn_lock);
 332         list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
 333         spin_unlock(&d->opd_syn_lock);
 334
 335         /* XXX: some batching wouldn't hurt */
 336         wake_up(&d->opd_syn_waitq);
 337 }
 338
 339 static int osp_sync_interpret(const struct lu_env *env,
 340                               struct ptlrpc_request *req, void *aa, int rc)
 341 {
 342         struct osp_device *d = req->rq_cb_data;
 343
 344         if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
 345                 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
 346         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 347         LASSERT(d);
 348
 349         CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
 350                atomic_read(&req->rq_refcount),
 351                rc, (unsigned) req->rq_transno);
 352         LASSERT(rc || req->rq_transno);
 353
 354         if (rc == -ENOENT) {
 355                 /*
 356                  * we tried to destroy object or update attributes,
 357                  * but object doesn't exist anymore - cancell llog record
 358                  */
 359                 LASSERT(req->rq_transno == 0);
 360                 LASSERT(list_empty(&req->rq_exp_list));
 361
 362                 ptlrpc_request_addref(req);
 363
 364                 spin_lock(&d->opd_syn_lock);
 365                 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
 366                 spin_unlock(&d->opd_syn_lock);
 367
 368                 wake_up(&d->opd_syn_waitq);
 369         } else if (rc) {
 370                 struct obd_import *imp = req->rq_import;
 371                 /*
 372                  * error happened, we'll try to repeat on next boot ?
 373                  */
 374                 LASSERTF(req->rq_transno == 0 ||
 375                          req->rq_import_generation < imp->imp_generation,
 376                          "transno "LPU64", rc %d, gen: req %d, imp %d\n",
 377                          req->rq_transno, rc, req->rq_import_generation,
 378                          imp->imp_generation);
 379                 if (req->rq_transno == 0) {
 380                         /* this is the last time we see the request
 381                          * if transno is not zero, then commit cb
 382                          * will be called at some point */
 383                         LASSERT(d->opd_syn_rpc_in_progress > 0);
 384                         spin_lock(&d->opd_syn_lock);
 385                         d->opd_syn_rpc_in_progress--;
 386                         spin_unlock(&d->opd_syn_lock);
 387                 }
 388
 389                 wake_up(&d->opd_syn_waitq);
 390         } else if (d->opd_pre != NULL &&
 391                    unlikely(d->opd_pre_status == -ENOSPC)) {
 392                 /*
 393                  * if current status is -ENOSPC (lack of free space on OST)
 394                  * then we should poll OST immediately once object destroy
 395                  * is replied
 396                  */
 397                 osp_statfs_need_now(d);
 398         }
 399
 400         LASSERT(d->opd_syn_rpc_in_flight > 0);
 401         spin_lock(&d->opd_syn_lock);
 402         d->opd_syn_rpc_in_flight--;
 403         spin_unlock(&d->opd_syn_lock);
 404         if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
 405                 wake_up(&d->opd_syn_barrier_waitq);
 406         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 407                d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 408                d->opd_syn_rpc_in_progress);
 409
 410         osp_sync_check_for_work(d);
 411
 412         return 0;
 413 }
 414
 415 /*
 416  * the function walks through list of committed locally changes
 417  * and send them to RPC until the pipe is full
 418  */
 419 static void osp_sync_send_new_rpc(struct osp_device *d,
 420                                   struct ptlrpc_request *req)
 421 {
 422         LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
 423         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 424
 425         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 426 }
 427
 428 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
 429                                                struct llog_handle *llh,
 430                                                struct llog_rec_hdr *h,
 431                                                ost_cmd_t op,
 432                                                const struct req_format *format)
 433 {
 434         struct ptlrpc_request   *req;
 435         struct ost_body         *body;
 436         struct obd_import       *imp;
 437         int                      rc;
 438
 439         /* Prepare the request */
 440         imp = d->opd_obd->u.cli.cl_import;
 441         LASSERT(imp);
 442         req = ptlrpc_request_alloc(imp, format);
 443         if (req == NULL)
 444                 RETURN(ERR_PTR(-ENOMEM));
 445
 446         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
 447         if (rc) {
 448                 ptlrpc_req_finished(req);
 449                 return ERR_PTR(rc);
 450         }
 451
 452         /*
 453          * this is a trick: to save on memory allocations we put cookie
 454          * into the request, but don't set corresponded flag in o_valid
 455          * so that OST doesn't interpret this cookie. once the request
 456          * is committed on OST we take cookie from the request and cancel
 457          */
 458         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 459         LASSERT(body);
 460         body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
 461         body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 462         body->oa.o_lcookie.lgc_index = h->lrh_index;
 463         INIT_LIST_HEAD(&req->rq_exp_list);
 464         req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
 465
 466         req->rq_interpret_reply = osp_sync_interpret;
 467         req->rq_commit_cb = osp_sync_request_commit_cb;
 468         req->rq_cb_data = d;
 469
 470         ptlrpc_request_set_replen(req);
 471
 472         return req;
 473 }
 474
 475 static int osp_sync_new_setattr_job(struct osp_device *d,
 476                                     struct llog_handle *llh,
 477                                     struct llog_rec_hdr *h)
 478 {
 479         struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
 480         struct ptlrpc_request           *req;
 481         struct ost_body                 *body;
 482
 483         ENTRY;
 484         LASSERT(h->lrh_type == MDS_SETATTR64_REC);
 485
 486         /* lsr_valid can only be 0 or LA_UID/GID set */
 487         if ((rec->lsr_valid & ~(LA_UID | LA_GID)) != 0) {
 488                 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
 489                        d->opd_obd->obd_name, rec->lsr_valid);
 490                 /* return 0 so that sync thread can continue processing
 491                  * other records. */
 492                 RETURN(0);
 493         }
 494
 495         req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
 496         if (IS_ERR(req))
 497                 RETURN(PTR_ERR(req));
 498
 499         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 500         LASSERT(body);
 501         body->oa.o_oi = rec->lsr_oi;
 502         body->oa.o_uid = rec->lsr_uid;
 503         body->oa.o_gid = rec->lsr_gid;
 504         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 505         /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
 506          * we assume that both UID and GID are valid in that case. */
 507         if (rec->lsr_valid == 0) {
 508                 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
 509         } else {
 510                 if (rec->lsr_valid & LA_UID)
 511                         body->oa.o_valid |= OBD_MD_FLUID;
 512                 if (rec->lsr_valid & LA_GID)
 513                         body->oa.o_valid |= OBD_MD_FLGID;
 514         }
 515
 516         osp_sync_send_new_rpc(d, req);
 517         RETURN(0);
 518 }
 519
 520 /* Old records may be in old format, so we handle that too */
 521 static int osp_sync_new_unlink_job(struct osp_device *d,
 522                                    struct llog_handle *llh,
 523                                    struct llog_rec_hdr *h)
 524 {
 525         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 526         struct ptlrpc_request   *req;
 527         struct ost_body         *body;
 528
 529         ENTRY;
 530         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 531
 532         req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
 533         if (IS_ERR(req))
 534                 RETURN(PTR_ERR(req));
 535
 536         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 537         LASSERT(body);
 538         ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
 539         ostid_set_id(&body->oa.o_oi, rec->lur_oid);
 540         body->oa.o_misc = rec->lur_count;
 541         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 542         if (rec->lur_count)
 543                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 544
 545         osp_sync_send_new_rpc(d, req);
 546         RETURN(0);
 547 }
 548
 549 static int osp_prep_unlink_update_req(const struct lu_env *env,
 550                                       struct osp_device *osp,
 551                                       struct llog_handle *llh,
 552                                       struct llog_rec_hdr *h,
 553                                       struct ptlrpc_request **reqp)
 554 {
 555         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 556         struct dt_update_request        *update = NULL;
 557         struct ptlrpc_request           *req;
 558         const char                      *buf;
 559         struct llog_cookie              lcookie;
 560         int                             size;
 561         int                             rc;
 562         ENTRY;
 563
 564         update = out_create_update_req(&osp->opd_dt_dev);
 565         if (IS_ERR(update))
 566                 RETURN(PTR_ERR(update));
 567
 568         /* This can only happens for unlink slave directory, so decrease
 569          * ref for ".." and "." */
 570         rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
 571                                NULL, NULL);
 572         if (rc != 0)
 573                 GOTO(out, rc);
 574
 575         rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
 576                                NULL, NULL);
 577         if (rc != 0)
 578                 GOTO(out, rc);
 579
 580         lcookie.lgc_lgl = llh->lgh_id;
 581         lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 582         lcookie.lgc_index = h->lrh_index;
 583         size = sizeof(lcookie);
 584         buf = (const char *)&lcookie;
 585
 586         rc = out_insert_update(env, update, OUT_DESTROY, &rec->lur_fid, 1,
 587                                &size, &buf);
 588         if (rc != 0)
 589                 GOTO(out, rc);
 590
 591         rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
 592                                  update->dur_req, &req);
 593         if (rc != 0)
 594                 GOTO(out, rc);
 595
 596         INIT_LIST_HEAD(&req->rq_exp_list);
 597         req->rq_svc_thread = (void *)OSP_JOB_MAGIC;
 598
 599         req->rq_interpret_reply = osp_sync_interpret;
 600         req->rq_commit_cb = osp_sync_request_commit_cb;
 601         req->rq_cb_data = osp;
 602
 603         ptlrpc_request_set_replen(req);
 604         *reqp = req;
 605 out:
 606         if (update != NULL)
 607                 out_destroy_update_req(update);
 608
 609         RETURN(rc);
 610 }
 611
 612 static int osp_sync_new_unlink64_job(const struct lu_env *env,
 613                                      struct osp_device *d,
 614                                      struct llog_handle *llh,
 615                                      struct llog_rec_hdr *h)
 616 {
 617         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 618         struct ptlrpc_request           *req = NULL;
 619         struct ost_body                 *body;
 620         int                              rc;
 621
 622         ENTRY;
 623         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 624
 625         if (d->opd_connect_mdt) {
 626                 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
 627                 if (rc != 0)
 628                         RETURN(rc);
 629         } else {
 630                 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
 631                                        &RQF_OST_DESTROY);
 632                 if (IS_ERR(req))
 633                         RETURN(PTR_ERR(req));
 634
 635                 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 636                 if (body == NULL)
 637                         RETURN(-EFAULT);
 638                 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
 639                 if (rc < 0)
 640                         RETURN(rc);
 641                 body->oa.o_misc = rec->lur_count;
 642                 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 643                                    OBD_MD_FLOBJCOUNT;
 644         }
 645         osp_sync_send_new_rpc(d, req);
 646         RETURN(0);
 647 }
 648
 649 static int osp_sync_process_record(const struct lu_env *env,
 650                                    struct osp_device *d,
 651                                    struct llog_handle *llh,
 652                                    struct llog_rec_hdr *rec)
 653 {
 654         struct llog_cookie       cookie;
 655         int                      rc = 0;
 656
 657         cookie.lgc_lgl = llh->lgh_id;
 658         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 659         cookie.lgc_index = rec->lrh_index;
 660
 661         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 662                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 663
 664                 /* we're waiting for the record generated by this instance */
 665                 LASSERT(d->opd_syn_prev_done == 0);
 666                 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
 667                             sizeof(gen->lgr_gen))) {
 668                         CDEBUG(D_HA, "processed all old entries\n");
 669                         d->opd_syn_prev_done = 1;
 670                 }
 671
 672                 /* cancel any generation record */
 673                 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
 674                                              1, &cookie);
 675
 676                 return rc;
 677         }
 678
 679         /*
 680          * now we prepare and fill requests to OST, put them on the queue
 681          * and fire after next commit callback
 682          */
 683
 684         /* notice we increment counters before sending RPC, to be consistent
 685          * in RPC interpret callback which may happen very quickly */
 686         spin_lock(&d->opd_syn_lock);
 687         d->opd_syn_rpc_in_flight++;
 688         d->opd_syn_rpc_in_progress++;
 689         spin_unlock(&d->opd_syn_lock);
 690
 691         switch (rec->lrh_type) {
 692         /* case MDS_UNLINK_REC is kept for compatibility */
 693         case MDS_UNLINK_REC:
 694                 rc = osp_sync_new_unlink_job(d, llh, rec);
 695                 break;
 696         case MDS_UNLINK64_REC:
 697                 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
 698                 break;
 699         case MDS_SETATTR64_REC:
 700                 rc = osp_sync_new_setattr_job(d, llh, rec);
 701                 break;
 702         default:
 703                 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
 704                        rec->lrh_type);
 705                 /* we should continue processing */
 706                 return 0;
 707         }
 708
 709         if (likely(rc == 0)) {
 710                 spin_lock(&d->opd_syn_lock);
 711                 if (d->opd_syn_prev_done) {
 712                         LASSERT(d->opd_syn_changes > 0);
 713                         LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
 714                         /*
 715                          * NOTE: it's possible to meet same id if
 716                          * OST stores few stripes of same file
 717                          */
 718                         if (rec->lrh_id > d->opd_syn_last_processed_id) {
 719                                 d->opd_syn_last_processed_id = rec->lrh_id;
 720                                 wake_up(&d->opd_syn_barrier_waitq);
 721                         }
 722
 723                         d->opd_syn_changes--;
 724                 }
 725                 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 726                        d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 727                        d->opd_syn_rpc_in_progress);
 728                 spin_unlock(&d->opd_syn_lock);
 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