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