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, 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 #ifndef EXPORT_SYMTAB
  45 # define EXPORT_SYMTAB
  46 #endif
  47 #define DEBUG_SUBSYSTEM S_MDS
  48
  49 #include <lustre_log.h>
  50 #include "osp_internal.h"
  51
  52 static int osp_sync_id_traction_init(struct osp_device *d);
  53 static void osp_sync_id_traction_fini(struct osp_device *d);
  54 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
  55 static void osp_sync_remove_from_tracker(struct osp_device *d);
  56
  57 /*
  58  * this is a components of OSP implementing synchronization between MDS and OST
  59  * it llogs all interesting changes (currently it's uig/gid change and object
  60  * destroy) atomically, then makes sure changes hit OST storage
  61  *
  62  * we have 4 queues of work:
  63  *
  64  * the first queue is llog itself, once read a change is stored in 2nd queue
  65  * in form of RPC (but RPC isn't fired yet).
  66  *
  67  * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
  68  * commit. once change is committed locally it migrates onto 3rd queue.
  69  *
  70  * the third queue (opd_syn_committed_here) holds changes committed locally,
  71  * but not sent to OST (as the pipe can be full). once pipe becomes non-full
  72  * we take a change from the queue and fire corresponded RPC.
  73  *
  74  * once RPC is reported committed by OST (using regular last_committed mech.)
  75  * the change jumps into 4th queue (opd_syn_committed_there), now we can
  76  * cancel corresponded llog record and release RPC
  77  *
  78  * opd_syn_changes is a number of unread llog records (to be processed).
  79  * notice this number doesn't include llog records from previous boots.
  80  * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
  81  *
  82  * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
  83  * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
  84  * too much memory -- how to deal with 1000th OSTs ? batching could help?
  85  *
  86  * opd_syn_rpc_in_flight is a number of RPC in flight.
  87  * we control this with OSP_MAX_IN_FLIGHT
  88  */
  89
  90 /* XXX: do math to learn reasonable threshold
  91  * should it be ~ number of changes fitting bulk? */
  92
  93 #define OSP_SYN_THRESHOLD       10
  94 #define OSP_MAX_IN_FLIGHT       8
  95 #define OSP_MAX_IN_PROGRESS     4096
  96
  97 #define OSP_JOB_MAGIC           0x26112005
  98
  99 static inline int osp_sync_running(struct osp_device *d)
 100 {
 101         return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
 102 }
 103
 104 static inline int osp_sync_stopped(struct osp_device *d)
 105 {
 106         return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
 107 }
 108
 109 static inline int osp_sync_has_new_job(struct osp_device *d)
 110 {
 111         return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
 112                 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
 113                 || (d->opd_syn_prev_done == 0);
 114 }
 115
 116 static inline int osp_sync_low_in_progress(struct osp_device *d)
 117 {
 118         return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
 119 }
 120
 121 static inline int osp_sync_low_in_flight(struct osp_device *d)
 122 {
 123         return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
 124 }
 125
 126 static inline int osp_sync_has_work(struct osp_device *d)
 127 {
 128         /* has new/old changes and low in-progress? */
 129         if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
 130             osp_sync_low_in_flight(d) && d->opd_imp_connected)
 131                 return 1;
 132
 133         /* has remotely committed? */
 134         if (!cfs_list_empty(&d->opd_syn_committed_there))
 135                 return 1;
 136
 137         return 0;
 138 }
 139
 140 #define osp_sync_check_for_work(d)                      \
 141 {                                                       \
 142         if (osp_sync_has_work(d)) {                     \
 143                 cfs_waitq_signal(&d->opd_syn_waitq);    \
 144         }                                               \
 145 }
 146
 147 void __osp_sync_check_for_work(struct osp_device *d)
 148 {
 149         osp_sync_check_for_work(d);
 150 }
 151
 152 static inline int osp_sync_can_process_new(struct osp_device *d,
 153                                            struct llog_rec_hdr *rec)
 154 {
 155         LASSERT(d);
 156
 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_ostid_pack(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_oid  = osi->osi_oi.oi_id;
 238                 osi->osi_setattr.lsr_oseq = osi->osi_oi.oi_seq;
 239                 LASSERT(attr);
 240                 osi->osi_setattr.lsr_uid = attr->la_uid;
 241                 osi->osi_setattr.lsr_gid = attr->la_gid;
 242                 break;
 243         default:
 244                 LBUG();
 245         }
 246
 247         txn = osp_txn_info(&th->th_ctx);
 248         LASSERT(txn);
 249
 250         txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
 251         osi->osi_hdr.lrh_id = txn->oti_current_id;
 252
 253         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 254         if (ctxt == NULL)
 255                 RETURN(-ENOMEM);
 256         rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
 257                       NULL, th);
 258         llog_ctxt_put(ctxt);
 259
 260         CDEBUG(D_OTHER, "%s: new record %lu:%lu:%lu/%lu: %d\n",
 261                d->opd_obd->obd_name,
 262                (unsigned long) osi->osi_cookie.lgc_lgl.lgl_oid,
 263                (unsigned long) osi->osi_cookie.lgc_lgl.lgl_oseq,
 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(cfs_list_empty(&req->rq_exp_list));
 331
 332         ptlrpc_request_addref(req);
 333
 334         spin_lock(&d->opd_syn_lock);
 335         cfs_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         cfs_waitq_signal(&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                cfs_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(cfs_list_empty(&req->rq_exp_list));
 364
 365                 ptlrpc_request_addref(req);
 366
 367                 spin_lock(&d->opd_syn_lock);
 368                 cfs_list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
 369                 spin_unlock(&d->opd_syn_lock);
 370
 371                 cfs_waitq_signal(&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                 cfs_waitq_signal(&d->opd_syn_waitq);
 393         } else if (unlikely(d->opd_pre_status == -ENOSPC)) {
 394                 /*
 395                  * if current status is -ENOSPC (lack of free space on OST)
 396                  * then we should poll OST immediately once object destroy
 397                  * is replied
 398                  */
 399                 osp_statfs_need_now(d);
 400         }
 401
 402         LASSERT(d->opd_syn_rpc_in_flight > 0);
 403         spin_lock(&d->opd_syn_lock);
 404         d->opd_syn_rpc_in_flight--;
 405         spin_unlock(&d->opd_syn_lock);
 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         CFS_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         req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
 487         if (IS_ERR(req))
 488                 RETURN(PTR_ERR(req));
 489
 490         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 491         LASSERT(body);
 492         body->oa.o_id  = rec->lsr_oid;
 493         body->oa.o_seq = rec->lsr_oseq;
 494         body->oa.o_uid = rec->lsr_uid;
 495         body->oa.o_gid = rec->lsr_gid;
 496         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 497                            OBD_MD_FLUID | OBD_MD_FLGID;
 498
 499         osp_sync_send_new_rpc(d, req);
 500         RETURN(0);
 501 }
 502
 503 /* Old records may be in old format, so we handle that too */
 504 static int osp_sync_new_unlink_job(struct osp_device *d,
 505                                    struct llog_handle *llh,
 506                                    struct llog_rec_hdr *h)
 507 {
 508         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 509         struct ptlrpc_request   *req;
 510         struct ost_body         *body;
 511
 512         ENTRY;
 513         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 514
 515         req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
 516         if (IS_ERR(req))
 517                 RETURN(PTR_ERR(req));
 518
 519         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 520         LASSERT(body);
 521         body->oa.o_id  = rec->lur_oid;
 522         body->oa.o_seq = rec->lur_oseq;
 523         body->oa.o_misc = rec->lur_count;
 524         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 525         if (rec->lur_count)
 526                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 527
 528         osp_sync_send_new_rpc(d, req);
 529         RETURN(0);
 530 }
 531
 532 static int osp_sync_new_unlink64_job(struct osp_device *d,
 533                                      struct llog_handle *llh,
 534                                      struct llog_rec_hdr *h)
 535 {
 536         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 537         struct ptlrpc_request           *req;
 538         struct ost_body                 *body;
 539         int                              rc;
 540
 541         ENTRY;
 542         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 543
 544         req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
 545         if (IS_ERR(req))
 546                 RETURN(PTR_ERR(req));
 547
 548         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 549         if (body == NULL)
 550                 RETURN(-EFAULT);
 551         rc = fid_ostid_pack(&rec->lur_fid, &body->oa.o_oi);
 552         if (rc < 0)
 553                 RETURN(rc);
 554         body->oa.o_misc = rec->lur_count;
 555         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID | OBD_MD_FLOBJCOUNT;
 556
 557         osp_sync_send_new_rpc(d, req);
 558         RETURN(0);
 559 }
 560
 561 static int osp_sync_process_record(const struct lu_env *env,
 562                                    struct osp_device *d,
 563                                    struct llog_handle *llh,
 564                                    struct llog_rec_hdr *rec)
 565 {
 566         struct llog_cookie       cookie;
 567         int                      rc = 0;
 568
 569         cookie.lgc_lgl = llh->lgh_id;
 570         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 571         cookie.lgc_index = rec->lrh_index;
 572
 573         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 574                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 575
 576                 /* we're waiting for the record generated by this instance */
 577                 LASSERT(d->opd_syn_prev_done == 0);
 578                 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
 579                             sizeof(gen->lgr_gen))) {
 580                         CDEBUG(D_HA, "processed all old entries\n");
 581                         d->opd_syn_prev_done = 1;
 582                 }
 583
 584                 /* cancel any generation record */
 585                 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
 586                                              1, &cookie);
 587
 588                 return rc;
 589         }
 590
 591         /*
 592          * now we prepare and fill requests to OST, put them on the queue
 593          * and fire after next commit callback
 594          */
 595
 596         /* notice we increment counters before sending RPC, to be consistent
 597          * in RPC interpret callback which may happen very quickly */
 598         spin_lock(&d->opd_syn_lock);
 599         d->opd_syn_rpc_in_flight++;
 600         d->opd_syn_rpc_in_progress++;
 601         spin_unlock(&d->opd_syn_lock);
 602
 603         switch (rec->lrh_type) {
 604         /* case MDS_UNLINK_REC is kept for compatibility */
 605         case MDS_UNLINK_REC:
 606                 rc = osp_sync_new_unlink_job(d, llh, rec);
 607                 break;
 608         case MDS_UNLINK64_REC:
 609                 rc = osp_sync_new_unlink64_job(d, llh, rec);
 610                 break;
 611         case MDS_SETATTR64_REC:
 612                 rc = osp_sync_new_setattr_job(d, llh, rec);
 613                 break;
 614         default:
 615                 CERROR("unknown record type: %x\n", rec->lrh_type);
 616                        rc = -EINVAL;
 617                        break;
 618         }
 619
 620         if (likely(rc == 0)) {
 621                 spin_lock(&d->opd_syn_lock);
 622                 if (d->opd_syn_prev_done) {
 623                         LASSERT(d->opd_syn_changes > 0);
 624                         LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
 625                         /*
 626                          * NOTE: it's possible to meet same id if
 627                          * OST stores few stripes of same file
 628                          */
 629                         if (rec->lrh_id > d->opd_syn_last_processed_id)
 630                                 d->opd_syn_last_processed_id = rec->lrh_id;
 631
 632                         d->opd_syn_changes--;
 633                 }
 634                 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 635                        d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 636                        d->opd_syn_rpc_in_progress);
 637                 spin_unlock(&d->opd_syn_lock);
 638         } else {
 639                 spin_lock(&d->opd_syn_lock);
 640                 d->opd_syn_rpc_in_flight--;
 641                 d->opd_syn_rpc_in_progress--;
 642                 spin_unlock(&d->opd_syn_lock);
 643         }
 644
 645         CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
 646                rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
 647         return rc;
 648 }
 649
 650 static void osp_sync_process_committed(const struct lu_env *env,
 651                                        struct osp_device *d)
 652 {
 653         struct obd_device       *obd = d->opd_obd;
 654         struct obd_import       *imp = obd->u.cli.cl_import;
 655         struct ost_body         *body;
 656         struct ptlrpc_request   *req, *tmp;
 657         struct llog_ctxt        *ctxt;
 658         struct llog_handle      *llh;
 659         cfs_list_t               list;
 660         int                      rc, done = 0;
 661
 662         ENTRY;
 663
 664         if (cfs_list_empty(&d->opd_syn_committed_there))
 665                 return;
 666
 667         /*
 668          * if current status is -ENOSPC (lack of free space on OST)
 669          * then we should poll OST immediately once object destroy
 670          * is committed.
 671          * notice: we do this upon commit as well because some backends
 672          * (like DMU) do not release space right away.
 673          */
 674         if (unlikely(d->opd_pre_status == -ENOSPC))
 675                 osp_statfs_need_now(d);
 676
 677         /*
 678          * now cancel them all
 679          * XXX: can we improve this using some batching?
 680          *      with batch RPC that'll happen automatically?
 681          * XXX: can we store ctxt in lod_device and save few cycles ?
 682          */
 683         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
 684         LASSERT(ctxt);
 685
 686         llh = ctxt->loc_handle;
 687         LASSERT(llh);
 688
 689         CFS_INIT_LIST_HEAD(&list);
 690         spin_lock(&d->opd_syn_lock);
 691         cfs_list_splice(&d->opd_syn_committed_there, &list);
 692         CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
 693         spin_unlock(&d->opd_syn_lock);
 694
 695         cfs_list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
 696                 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 697                 cfs_list_del_init(&req->rq_exp_list);
 698
 699                 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 700                 LASSERT(body);
 701
 702                 /* import can be closing, thus all commit cb's are
 703                  * called we can check committness directly */
 704                 if (req->rq_transno <= imp->imp_peer_committed_transno) {
 705                         rc = llog_cat_cancel_records(env, llh, 1,
 706                                                      &body->oa.o_lcookie);
 707                         if (rc)
 708                                 CERROR("%s: can't cancel record: %d\n",
 709                                        obd->obd_name, rc);
 710                 } else {
 711                         DEBUG_REQ(D_HA, req, "not committed");
 712                 }
 713
 714                 ptlrpc_req_finished(req);
 715                 done++;
 716         }
 717
 718         llog_ctxt_put(ctxt);
 719
 720         LASSERT(d->opd_syn_rpc_in_progress >= done);
 721         spin_lock(&d->opd_syn_lock);
 722         d->opd_syn_rpc_in_progress -= done;
 723         spin_unlock(&d->opd_syn_lock);
 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
 728         osp_sync_check_for_work(d);
 729
 730         /* wake up the thread if requested to stop:
 731          * it might be waiting for in-progress to complete */
 732         if (unlikely(osp_sync_running(d) == 0))
 733                 cfs_waitq_signal(&d->opd_syn_waitq);
 734
 735         EXIT;
 736 }
 737
 738 /*
 739  * this is where most of queues processing happens
 740  */
 741 static int osp_sync_process_queues(const struct lu_env *env,
 742                                    struct llog_handle *llh,
 743                                    struct llog_rec_hdr *rec,
 744                                    void *data)
 745 {
 746         struct osp_device       *d = data;
 747         int                      rc;
 748
 749         do {
 750                 struct l_wait_info lwi = { 0 };
 751
 752                 if (!osp_sync_running(d)) {
 753                         CDEBUG(D_HA, "stop llog processing\n");
 754                         return LLOG_PROC_BREAK;
 755                 }
 756
 757                 /* process requests committed by OST */
 758                 osp_sync_process_committed(env, d);
 759
 760                 /* if we there are changes to be processed and we have
 761                  * resources for this ... do now */
 762                 if (osp_sync_can_process_new(d, rec)) {
 763                         if (llh == NULL) {
 764                                 /* ask llog for another record */
 765                                 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
 766                                        d->opd_syn_changes,
 767                                        d->opd_syn_rpc_in_progress,
 768                                        d->opd_syn_rpc_in_flight);
 769                                 return 0;
 770                         }
 771
 772                         /*
 773                          * try to send, in case of disconnection, suspend
 774                          * processing till we can send this request
 775                          */
 776                         do {
 777                                 rc = osp_sync_process_record(env, d, llh, rec);
 778                                 /*
 779                                  * XXX: probably different handling is needed
 780                                  * for some bugs, like immediate exit or if
 781                                  * OSP gets inactive
 782                                  */
 783                                 if (rc) {
 784                                         CERROR("can't send: %d\n", rc);
 785                                         l_wait_event(d->opd_syn_waitq,
 786                                                      !osp_sync_running(d) ||
 787                                                      osp_sync_has_work(d),
 788                                                      &lwi);
 789                                 }
 790                         } while (rc != 0 && osp_sync_running(d));
 791
 792                         llh = NULL;
 793                         rec = NULL;
 794                 }
 795
 796                 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
 797                         osp_sync_remove_from_tracker(d);
 798
 799                 l_wait_event(d->opd_syn_waitq,
 800                              !osp_sync_running(d) ||
 801                              osp_sync_can_process_new(d, rec) ||
 802                              !cfs_list_empty(&d->opd_syn_committed_there),
 803                              &lwi);
 804         } while (1);
 805 }
 806
 807 /*
 808  * this thread runs llog_cat_process() scanner calling our callback
 809  * to process llog records. in the callback we implement tricky
 810  * state machine as we don't want to start scanning of the llog again
 811  * and again, also we don't want to process too many records and send
 812  * too many RPCs a time. so, depending on current load (num of changes
 813  * being synced to OST) the callback can suspend awaiting for some
 814  * new conditions, like syncs completed.
 815  *
 816  * in order to process llog records left by previous boots and to allow
 817  * llog_process_thread() to find something (otherwise it'd just exit
 818  * immediately) we add a special GENERATATION record on each boot.
 819  */
 820 static int osp_sync_thread(void *_arg)
 821 {
 822         struct osp_device       *d = _arg;
 823         struct ptlrpc_thread    *thread = &d->opd_syn_thread;
 824         struct l_wait_info       lwi = { 0 };
 825         struct llog_ctxt        *ctxt;
 826         struct obd_device       *obd = d->opd_obd;
 827         struct llog_handle      *llh;
 828         struct lu_env            env;
 829         int                      rc, count;
 830         char                     pname[16];
 831
 832         ENTRY;
 833
 834         rc = lu_env_init(&env, LCT_LOCAL);
 835         if (rc) {
 836                 CERROR("%s: can't initialize env: rc = %d\n",
 837                        obd->obd_name, rc);
 838                 RETURN(rc);
 839         }
 840
 841         sprintf(pname, "osp-syn-%u", d->opd_index);
 842         cfs_daemonize(pname);
 843
 844         spin_lock(&d->opd_syn_lock);
 845         thread->t_flags = SVC_RUNNING;
 846         spin_unlock(&d->opd_syn_lock);
 847         cfs_waitq_signal(&thread->t_ctl_waitq);
 848
 849         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
 850         if (ctxt == NULL) {
 851                 CERROR("can't get appropriate context\n");
 852                 GOTO(out, rc = -EINVAL);
 853         }
 854
 855         llh = ctxt->loc_handle;
 856         if (llh == NULL) {
 857                 CERROR("can't get llh\n");
 858                 llog_ctxt_put(ctxt);
 859                 GOTO(out, rc = -EINVAL);
 860         }
 861
 862         rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
 863         LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
 864                  "%lu changes, %u in progress, %u in flight: %d\n",
 865                  d->opd_syn_changes, d->opd_syn_rpc_in_progress,
 866                  d->opd_syn_rpc_in_flight, rc);
 867
 868         /* we don't expect llog_process_thread() to exit till umount */
 869         LASSERTF(thread->t_flags != SVC_RUNNING,
 870                  "%lu changes, %u in progress, %u in flight\n",
 871                  d->opd_syn_changes, d->opd_syn_rpc_in_progress,
 872                  d->opd_syn_rpc_in_flight);
 873
 874         /* wait till all the requests are completed */
 875         count = 0;
 876         while (d->opd_syn_rpc_in_progress > 0) {
 877                 osp_sync_process_committed(&env, d);
 878
 879                 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
 880                 rc = l_wait_event(d->opd_syn_waitq,
 881                                   d->opd_syn_rpc_in_progress == 0,
 882                                   &lwi);
 883                 if (rc == -ETIMEDOUT)
 884                         count++;
 885                 LASSERTF(count < 10, "%s: %d %d %sempty\n",
 886                          d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
 887                          d->opd_syn_rpc_in_flight,
 888                          cfs_list_empty(&d->opd_syn_committed_there) ? "" :"!");
 889
 890         }
 891
 892         llog_cat_close(&env, llh);
 893         rc = llog_cleanup(&env, ctxt);
 894         if (rc)
 895                 CERROR("can't cleanup llog: %d\n", rc);
 896 out:
 897         thread->t_flags = SVC_STOPPED;
 898
 899         cfs_waitq_signal(&thread->t_ctl_waitq);
 900         LASSERTF(d->opd_syn_rpc_in_progress == 0,
 901                  "%s: %d %d %sempty\n",
 902                  d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
 903                  d->opd_syn_rpc_in_flight,
 904                  cfs_list_empty(&d->opd_syn_committed_there) ? "" : "!");
 905
 906         lu_env_fini(&env);
 907
 908         RETURN(0);
 909 }
 910
 911 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
 912 {
 913         struct osp_thread_info *osi = osp_env_info(env);
 914         struct llog_handle     *lgh;
 915         struct obd_device      *obd = d->opd_obd;
 916         struct llog_ctxt       *ctxt;
 917         int                     rc;
 918
 919         ENTRY;
 920
 921         LASSERT(obd);
 922
 923         /*
 924          * open llog corresponding to our OST
 925          */
 926         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 927         obd->obd_lvfs_ctxt.dt = d->opd_storage;
 928
 929         rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
 930                                    &osi->osi_cid);
 931         if (rc) {
 932                 CERROR("%s: can't get id from catalogs: rc = %d\n",
 933                        obd->obd_name, rc);
 934                 RETURN(rc);
 935         }
 936
 937         CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
 938                obd->obd_name, d->opd_index, osi->osi_cid.lci_logid.lgl_oid,
 939                osi->osi_cid.lci_logid.lgl_oseq,
 940                osi->osi_cid.lci_logid.lgl_ogen);
 941
 942         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
 943                         &osp_mds_ost_orig_logops);
 944         if (rc)
 945                 RETURN(rc);
 946
 947         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
 948         LASSERT(ctxt);
 949
 950         if (likely(osi->osi_cid.lci_logid.lgl_oid != 0)) {
 951                 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
 952                                LLOG_OPEN_EXISTS);
 953                 /* re-create llog if it is missing */
 954                 if (rc == -ENOENT)
 955                         osi->osi_cid.lci_logid.lgl_oid = 0;
 956                 else if (rc < 0)
 957                         GOTO(out_cleanup, rc);
 958         }
 959
 960         if (unlikely(osi->osi_cid.lci_logid.lgl_oid == 0)) {
 961                 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
 962                 if (rc < 0)
 963                         GOTO(out_cleanup, rc);
 964                 osi->osi_cid.lci_logid = lgh->lgh_id;
 965         }
 966
 967         ctxt->loc_handle = lgh;
 968
 969         rc = llog_cat_init_and_process(env, lgh);
 970         if (rc)
 971                 GOTO(out_close, rc);
 972
 973         rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
 974                                    &osi->osi_cid);
 975         if (rc)
 976                 GOTO(out_close, rc);
 977
 978         /*
 979          * put a mark in the llog till which we'll be processing
 980          * old records restless
 981          */
 982         d->opd_syn_generation.mnt_cnt = cfs_time_current();
 983         d->opd_syn_generation.conn_cnt = cfs_time_current();
 984
 985         osi->osi_hdr.lrh_type = LLOG_GEN_REC;
 986         osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
 987
 988         memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
 989                sizeof(osi->osi_gen.lgr_gen));
 990
 991         rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie,
 992                           NULL);
 993         if (rc < 0)
 994                 GOTO(out_close, rc);
 995         llog_ctxt_put(ctxt);
 996         RETURN(0);
 997 out_close:
 998         llog_cat_close(env, lgh);
 999 out_cleanup:
1000         llog_cleanup(env, ctxt);
1001         RETURN(rc);
1002 }
1003
1004 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1005 {
1006         struct llog_ctxt *ctxt;
1007
1008         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1009         llog_cat_close(env, ctxt->loc_handle);
1010         llog_cleanup(env, ctxt);
1011 }
1012
1013 /*
1014  * initializes sync component of OSP
1015  */
1016 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1017 {
1018         struct l_wait_info       lwi = { 0 };
1019         int                      rc;
1020
1021         ENTRY;
1022
1023         rc = osp_sync_id_traction_init(d);
1024         if (rc)
1025                 RETURN(rc);
1026
1027         /*
1028          * initialize llog storing changes
1029          */
1030         rc = osp_sync_llog_init(env, d);
1031         if (rc) {
1032                 CERROR("%s: can't initialize llog: rc = %d\n",
1033                        d->opd_obd->obd_name, rc);
1034                 GOTO(err_id, rc);
1035         }
1036
1037         /*
1038          * Start synchronization thread
1039          */
1040         d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1041         d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1042         spin_lock_init(&d->opd_syn_lock);
1043         cfs_waitq_init(&d->opd_syn_waitq);
1044         cfs_waitq_init(&d->opd_syn_thread.t_ctl_waitq);
1045         CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
1046
1047         rc = cfs_create_thread(osp_sync_thread, d, 0);
1048         if (rc < 0) {
1049                 CERROR("%s: can't start sync thread: rc = %d\n",
1050                        d->opd_obd->obd_name, rc);
1051                 GOTO(err_llog, rc);
1052         }
1053
1054         l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1055                      osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1056
1057         RETURN(0);
1058 err_llog:
1059         osp_sync_llog_fini(env, d);
1060 err_id:
1061         osp_sync_id_traction_fini(d);
1062         return rc;
1063 }
1064
1065 int osp_sync_fini(struct osp_device *d)
1066 {
1067         struct ptlrpc_thread *thread = &d->opd_syn_thread;
1068
1069         ENTRY;
1070
1071         thread->t_flags = SVC_STOPPING;
1072         cfs_waitq_signal(&d->opd_syn_waitq);
1073         cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1074
1075         /*
1076          * unregister transaction callbacks only when sync thread
1077          * has finished operations with llog
1078          */
1079         osp_sync_id_traction_fini(d);
1080
1081         RETURN(0);
1082 }
1083
1084 static DEFINE_MUTEX(osp_id_tracker_sem);
1085 static CFS_LIST_HEAD(osp_id_tracker_list);
1086
1087 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1088 {
1089         struct osp_id_tracker   *tr = cookie;
1090         struct osp_device       *d;
1091         struct osp_txn_info     *txn;
1092
1093         LASSERT(tr);
1094
1095         txn = osp_txn_info(&th->th_ctx);
1096         if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1097                 return;
1098
1099         spin_lock(&tr->otr_lock);
1100         if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1101                 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1102                        tr->otr_committed_id, txn->oti_current_id);
1103                 tr->otr_committed_id = txn->oti_current_id;
1104
1105                 cfs_list_for_each_entry(d, &tr->otr_wakeup_list,
1106                                         opd_syn_ontrack) {
1107                         d->opd_syn_last_committed_id = tr->otr_committed_id;
1108                         cfs_waitq_signal(&d->opd_syn_waitq);
1109                 }
1110         }
1111         spin_unlock(&tr->otr_lock);
1112 }
1113
1114 static int osp_sync_id_traction_init(struct osp_device *d)
1115 {
1116         struct osp_id_tracker   *tr, *found = NULL;
1117         int                      rc = 0;
1118
1119         LASSERT(d);
1120         LASSERT(d->opd_storage);
1121         LASSERT(d->opd_syn_tracker == NULL);
1122         CFS_INIT_LIST_HEAD(&d->opd_syn_ontrack);
1123
1124         mutex_lock(&osp_id_tracker_sem);
1125         cfs_list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1126                 if (tr->otr_dev == d->opd_storage) {
1127                         LASSERT(cfs_atomic_read(&tr->otr_refcount));
1128                         cfs_atomic_inc(&tr->otr_refcount);
1129                         d->opd_syn_tracker = tr;
1130                         found = tr;
1131                         break;
1132                 }
1133         }
1134
1135         if (found == NULL) {
1136                 rc = -ENOMEM;
1137                 OBD_ALLOC_PTR(tr);
1138                 if (tr) {
1139                         d->opd_syn_tracker = tr;
1140                         spin_lock_init(&tr->otr_lock);
1141                         tr->otr_dev = d->opd_storage;
1142                         tr->otr_next_id = 1;
1143                         tr->otr_committed_id = 0;
1144                         cfs_atomic_set(&tr->otr_refcount, 1);
1145                         CFS_INIT_LIST_HEAD(&tr->otr_wakeup_list);
1146                         cfs_list_add(&tr->otr_list, &osp_id_tracker_list);
1147                         tr->otr_tx_cb.dtc_txn_commit =
1148                                                 osp_sync_tracker_commit_cb;
1149                         tr->otr_tx_cb.dtc_cookie = tr;
1150                         tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1151                         dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1152                         rc = 0;
1153                 }
1154         }
1155         mutex_unlock(&osp_id_tracker_sem);
1156
1157         return rc;
1158 }
1159
1160 static void osp_sync_id_traction_fini(struct osp_device *d)
1161 {
1162         struct osp_id_tracker *tr;
1163
1164         ENTRY;
1165
1166         LASSERT(d);
1167         tr = d->opd_syn_tracker;
1168         if (tr == NULL) {
1169                 EXIT;
1170                 return;
1171         }
1172
1173         osp_sync_remove_from_tracker(d);
1174
1175         mutex_lock(&osp_id_tracker_sem);
1176         if (cfs_atomic_dec_and_test(&tr->otr_refcount)) {
1177                 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1178                 LASSERT(cfs_list_empty(&tr->otr_wakeup_list));
1179                 cfs_list_del(&tr->otr_list);
1180                 OBD_FREE_PTR(tr);
1181                 d->opd_syn_tracker = NULL;
1182         }
1183         mutex_unlock(&osp_id_tracker_sem);
1184
1185         EXIT;
1186 }
1187
1188 /*
1189  * generates id for the tracker
1190  */
1191 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1192 {
1193         struct osp_id_tracker *tr;
1194
1195         tr = d->opd_syn_tracker;
1196         LASSERT(tr);
1197
1198         /* XXX: we can improve this introducing per-cpu preallocated ids? */
1199         spin_lock(&tr->otr_lock);
1200         if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1201                 spin_unlock(&tr->otr_lock);
1202                 CERROR("%s: next %u, last synced %lu\n",
1203                        d->opd_obd->obd_name, tr->otr_next_id,
1204                        d->opd_syn_last_used_id);
1205                 LBUG();
1206         }
1207
1208         if (id == 0)
1209                 id = tr->otr_next_id++;
1210         if (id > d->opd_syn_last_used_id)
1211                 d->opd_syn_last_used_id = id;
1212         if (cfs_list_empty(&d->opd_syn_ontrack))
1213                 cfs_list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1214         spin_unlock(&tr->otr_lock);
1215         CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1216
1217         return id;
1218 }
1219
1220 static void osp_sync_remove_from_tracker(struct osp_device *d)
1221 {
1222         struct osp_id_tracker *tr;
1223
1224         tr = d->opd_syn_tracker;
1225         LASSERT(tr);
1226
1227         if (cfs_list_empty(&d->opd_syn_ontrack))
1228                 return;
1229
1230         spin_lock(&tr->otr_lock);
1231         cfs_list_del_init(&d->opd_syn_ontrack);
1232         spin_unlock(&tr->otr_lock);
1233 }
1234