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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osp/osp_sync.c
  33  *
  34  * Lustre OST Proxy Device
  35  *
  36  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  37  * Author: Mikhail Pershin <mike.pershin@intel.com>
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_MDS
  41
  42 #include <linux/kthread.h>
  43 #include <lustre_log.h>
  44 #include <lustre_update.h>
  45 #include "osp_internal.h"
  46
  47 static int osp_sync_id_traction_init(struct osp_device *d);
  48 static void osp_sync_id_traction_fini(struct osp_device *d);
  49 static __u64 osp_sync_id_get(struct osp_device *d, __u64 id);
  50 static void osp_sync_remove_from_tracker(struct osp_device *d);
  51
  52 /*
  53  * this is a components of OSP implementing synchronization between MDS and OST
  54  * it llogs all interesting changes (currently it's uig/gid change and object
  55  * destroy) atomically, then makes sure changes hit OST storage
  56  *
  57  * we have 4 queues of work:
  58  *
  59  * the first queue is llog itself, once read a change is stored in 2nd queue
  60  * in form of RPC (but RPC isn't fired yet).
  61  *
  62  * the second queue (opd_sync_waiting_for_commit) holds changes awaiting local
  63  * commit. once change is committed locally it migrates onto 3rd queue.
  64  *
  65  * the third queue (opd_sync_committed_here) holds changes committed locally,
  66  * but not sent to OST (as the pipe can be full). once pipe becomes non-full
  67  * we take a change from the queue and fire corresponded RPC.
  68  *
  69  * once RPC is reported committed by OST (using regular last_committed mech.)
  70  * the change jumps into 4th queue (opd_sync_committed_there), now we can
  71  * cancel corresponded llog record and release RPC
  72  *
  73  * opd_sync_changes is a number of unread llog records (to be processed).
  74  * notice this number doesn't include llog records from previous boots.
  75  * with OSP_SYNC_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
  76  *
  77  * opd_sync_rpcs_in_progress is total number of requests in above 2-4 queues.
  78  * we control this with OSP_MAX_RPCS_IN_PROGRESS so that OSP don't consume
  79  * too much memory -- how to deal with 1000th OSTs ? batching could help?
  80  *
  81  * opd_sync_rpcs_in_flight is a number of RPC in flight.
  82  * we control this with OSP_MAX_RPCS_IN_FLIGHT
  83  */
  84
  85 /* XXX: do math to learn reasonable threshold
  86  * should it be ~ number of changes fitting bulk? */
  87
  88 #define OSP_SYNC_THRESHOLD              10
  89 #define OSP_MAX_RPCS_IN_FLIGHT          8
  90 #define OSP_MAX_RPCS_IN_PROGRESS        4096
  91
  92 #define OSP_JOB_MAGIC           0x26112005
  93
  94 struct osp_job_req_args {
  95         /** bytes reserved for ptlrpc_replay_req() */
  96         struct ptlrpc_replay_async_args jra_raa;
  97         struct list_head                jra_committed_link;
  98         struct list_head                jra_in_flight_link;
  99         struct llog_cookie              jra_lcookie;
 100         __u32                           jra_magic;
 101 };
 102
 103 static inline int osp_sync_running(struct osp_device *d)
 104 {
 105         return !!(d->opd_sync_thread.t_flags & SVC_RUNNING);
 106 }
 107
 108 /**
 109  * Check status: whether OSP thread has stopped
 110  *
 111  * \param[in] d         OSP device
 112  *
 113  * \retval 0            still running
 114  * \retval 1            stopped
 115  */
 116 static inline int osp_sync_stopped(struct osp_device *d)
 117 {
 118         return !!(d->opd_sync_thread.t_flags & SVC_STOPPED);
 119 }
 120
 121 /*
 122  ** Check for new changes to sync
 123  *
 124  * \param[in] d         OSP device
 125  *
 126  * \retval 1            there are changes
 127  * \retval 0            there are no changes
 128  */
 129 static inline int osp_sync_has_new_job(struct osp_device *d)
 130 {
 131         return ((d->opd_sync_last_processed_id < d->opd_sync_last_used_id) &&
 132                 (d->opd_sync_last_processed_id < d->opd_sync_last_committed_id))
 133                 || (d->opd_sync_prev_done == 0);
 134 }
 135
 136 static inline int osp_sync_in_flight_conflict(struct osp_device *d,
 137                                              struct llog_rec_hdr *h)
 138 {
 139         struct osp_job_req_args *jra;
 140         struct ost_id            ostid;
 141         int                      conflict = 0;
 142
 143         if (h == NULL || h->lrh_type == LLOG_GEN_REC ||
 144             list_empty(&d->opd_sync_in_flight_list))
 145                 return conflict;
 146
 147         memset(&ostid, 0, sizeof(ostid));
 148         switch (h->lrh_type) {
 149         case MDS_UNLINK_REC: {
 150                 struct llog_unlink_rec *unlink = (struct llog_unlink_rec *)h;
 151
 152                 ostid_set_seq(&ostid, unlink->lur_oseq);
 153                 if (ostid_set_id(&ostid, unlink->lur_oid)) {
 154                         CERROR("Bad %llu to set " DOSTID "\n",
 155                                (unsigned long long)(unlink->lur_oid),
 156                                POSTID(&ostid));
 157                         return 1;
 158                 }
 159                 }
 160                 break;
 161         case MDS_UNLINK64_REC:
 162                 fid_to_ostid(&((struct llog_unlink64_rec *)h)->lur_fid, &ostid);
 163                 break;
 164         case MDS_SETATTR64_REC:
 165                 ostid = ((struct llog_setattr64_rec *)h)->lsr_oi;
 166                 break;
 167         default:
 168                 LBUG();
 169         }
 170
 171         spin_lock(&d->opd_sync_lock);
 172         list_for_each_entry(jra, &d->opd_sync_in_flight_list,
 173                             jra_in_flight_link) {
 174                 struct ptlrpc_request   *req;
 175                 struct ost_body         *body;
 176
 177                 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
 178
 179                 req = container_of((void *)jra, struct ptlrpc_request,
 180                                    rq_async_args);
 181                 body = req_capsule_client_get(&req->rq_pill,
 182                                               &RMF_OST_BODY);
 183                 LASSERT(body);
 184
 185                 if (memcmp(&ostid, &body->oa.o_oi, sizeof(ostid)) == 0) {
 186                         conflict = 1;
 187                         break;
 188                 }
 189         }
 190         spin_unlock(&d->opd_sync_lock);
 191
 192         return conflict;
 193 }
 194
 195 static inline int osp_sync_rpcs_in_progress_low(struct osp_device *d)
 196 {
 197         return atomic_read(&d->opd_sync_rpcs_in_progress) <
 198                 d->opd_sync_max_rpcs_in_progress;
 199 }
 200
 201 /**
 202  * Check for room in the network pipe to OST
 203  *
 204  * \param[in] d         OSP device
 205  *
 206  * \retval 1            there is room
 207  * \retval 0            no room, the pipe is full
 208  */
 209 static inline int osp_sync_rpcs_in_flight_low(struct osp_device *d)
 210 {
 211         return atomic_read(&d->opd_sync_rpcs_in_flight) <
 212                 d->opd_sync_max_rpcs_in_flight;
 213 }
 214
 215 /**
 216  * Wake up check for the main sync thread
 217  *
 218  * \param[in] d         OSP device
 219  *
 220  * \retval 1            time to wake up
 221  * \retval 0            no need to wake up
 222  */
 223 static inline int osp_sync_has_work(struct osp_device *osp)
 224 {
 225         /* has new/old changes and low in-progress? */
 226         if (osp_sync_has_new_job(osp) && osp_sync_rpcs_in_progress_low(osp) &&
 227             osp_sync_rpcs_in_flight_low(osp) && osp->opd_imp_connected)
 228                 return 1;
 229
 230         /* has remotely committed? */
 231         if (!list_empty(&osp->opd_sync_committed_there))
 232                 return 1;
 233
 234         return 0;
 235 }
 236
 237 void osp_sync_check_for_work(struct osp_device *osp)
 238 {
 239         if (osp_sync_has_work(osp))
 240                 wake_up(&osp->opd_sync_waitq);
 241 }
 242
 243 static inline __u64 osp_sync_correct_id(struct osp_device *d,
 244                                         struct llog_rec_hdr *rec)
 245 {
 246         /*
 247          * llog use cyclic store with 32 bit lrh_id
 248          * so overflow lrh_id is possible. Range between
 249          * last_processed and last_committed is less than
 250          * 64745 ^ 2 and less than 2^32 - 1
 251          */
 252         __u64 correct_id = d->opd_sync_last_committed_id;
 253
 254         if ((correct_id & 0xffffffffULL) < rec->lrh_id)
 255                 correct_id -= 0x100000000ULL;
 256
 257         correct_id &= ~0xffffffffULL;
 258         correct_id |= rec->lrh_id;
 259
 260         return correct_id;
 261 }
 262 /**
 263  * Check and return ready-for-new status.
 264  *
 265  * The thread processing llog record uses this function to check whether
 266  * it's time to take another record and process it. The number of conditions
 267  * must be met: the connection should be ready, RPCs in flight not exceeding
 268  * the limit, the record is committed locally, etc (see the lines below).
 269  *
 270  * \param[in] d         OSP device
 271  * \param[in] rec       next llog record to process
 272  *
 273  * \retval 0            not ready
 274  * \retval 1            ready
 275  */
 276 static inline int osp_sync_can_process_new(struct osp_device *d,
 277                                            struct llog_rec_hdr *rec)
 278 {
 279         LASSERT(d);
 280
 281         if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
 282                 return 0;
 283         if (unlikely(osp_sync_in_flight_conflict(d, rec)))
 284                 return 0;
 285         if (!osp_sync_rpcs_in_progress_low(d))
 286                 return 0;
 287         if (!osp_sync_rpcs_in_flight_low(d))
 288                 return 0;
 289         if (!d->opd_imp_connected)
 290                 return 0;
 291         if (d->opd_sync_prev_done == 0)
 292                 return 1;
 293         if (atomic_read(&d->opd_sync_changes) == 0)
 294                 return 0;
 295         if (rec == NULL ||
 296             osp_sync_correct_id(d, rec) <= d->opd_sync_last_committed_id)
 297                 return 1;
 298         return 0;
 299 }
 300
 301 /**
 302  * Declare intention to add a new change.
 303  *
 304  * With regard to OSD API, we have to declare any changes ahead. In this
 305  * case we declare an intention to add a llog record representing the
 306  * change on the local storage.
 307  *
 308  * \param[in] env       LU environment provided by the caller
 309  * \param[in] o         OSP object
 310  * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 311  * \param[in] th        transaction handle (local)
 312  *
 313  * \retval 0            on success
 314  * \retval negative     negated errno on error
 315  */
 316 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
 317                          llog_op_type type, struct thandle *th)
 318 {
 319         struct osp_thread_info  *osi = osp_env_info(env);
 320         struct osp_device       *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
 321         struct llog_ctxt        *ctxt;
 322         struct thandle          *storage_th;
 323         int                      rc;
 324
 325         ENTRY;
 326
 327         /* it's a layering violation, to access internals of th,
 328          * but we can do this as a sanity check, for a while */
 329         LASSERT(th->th_top != NULL);
 330         storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
 331         if (IS_ERR(storage_th))
 332                 RETURN(PTR_ERR(storage_th));
 333
 334         switch (type) {
 335         case MDS_UNLINK64_REC:
 336                 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
 337                 break;
 338         case MDS_SETATTR64_REC:
 339                 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec_v2);
 340                 break;
 341         default:
 342                 LBUG();
 343         }
 344
 345         /* we want ->dt_trans_start() to allocate per-thandle structure */
 346         storage_th->th_tags |= LCT_OSP_THREAD;
 347
 348         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 349         LASSERT(ctxt);
 350
 351         rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
 352                               storage_th);
 353         llog_ctxt_put(ctxt);
 354
 355         RETURN(rc);
 356 }
 357
 358 /**
 359  * Generate a llog record for a given change.
 360  *
 361  * Generates a llog record for the change passed. The change can be of two
 362  * types: unlink and setattr. The record gets an ID which later will be
 363  * used to track commit status of the change. For unlink changes, the caller
 364  * can supply a starting FID and the count of the objects to destroy. For
 365  * setattr the caller should apply attributes to apply.
 366  *
 367  *
 368  * \param[in] env       LU environment provided by the caller
 369  * \param[in] d         OSP device
 370  * \param[in] fid       fid of the object the change should be applied to
 371  * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 372  * \param[in] count     count of objects to destroy
 373  * \param[in] th        transaction handle (local)
 374  * \param[in] attr      attributes for setattr
 375  *
 376  * \retval 0            on success
 377  * \retval negative     negated errno on error
 378  */
 379 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
 380                             const struct lu_fid *fid, llog_op_type type,
 381                             int count, struct thandle *th,
 382                             const struct lu_attr *attr)
 383 {
 384         struct osp_thread_info  *osi = osp_env_info(env);
 385         struct llog_ctxt        *ctxt;
 386         struct osp_txn_info     *txn;
 387         struct thandle          *storage_th;
 388         int                      rc;
 389
 390         ENTRY;
 391
 392         /* it's a layering violation, to access internals of th,
 393          * but we can do this as a sanity check, for a while */
 394         LASSERT(th->th_top != NULL);
 395         storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
 396         if (IS_ERR(storage_th))
 397                 RETURN(PTR_ERR(storage_th));
 398
 399         switch (type) {
 400         case MDS_UNLINK64_REC:
 401                 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
 402                 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
 403                 osi->osi_unlink.lur_fid  = *fid;
 404                 osi->osi_unlink.lur_count = count;
 405                 break;
 406         case MDS_SETATTR64_REC:
 407                 rc = fid_to_ostid(fid, &osi->osi_oi);
 408                 LASSERT(rc == 0);
 409                 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
 410                 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
 411                 osi->osi_setattr.lsr_oi  = osi->osi_oi;
 412                 LASSERT(attr);
 413                 osi->osi_setattr.lsr_uid = attr->la_uid;
 414                 osi->osi_setattr.lsr_gid = attr->la_gid;
 415                 osi->osi_setattr.lsr_projid = attr->la_projid;
 416                 osi->osi_setattr.lsr_valid =
 417                         ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
 418                         ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0) |
 419                         ((attr->la_valid & LA_PROJID) ? OBD_MD_FLPROJID : 0);
 420                 break;
 421         default:
 422                 LBUG();
 423         }
 424
 425         txn = osp_txn_info(&storage_th->th_ctx);
 426         LASSERT(txn);
 427
 428         txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
 429         osi->osi_hdr.lrh_id = (txn->oti_current_id & 0xffffffffULL);
 430         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 431         if (ctxt == NULL)
 432                 RETURN(-ENOMEM);
 433
 434         rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
 435                       storage_th);
 436         llog_ctxt_put(ctxt);
 437
 438         if (likely(rc >= 0)) {
 439                 CDEBUG(D_OTHER, "%s: new record "DFID":%x.%u: rc = %d\n",
 440                        d->opd_obd->obd_name,
 441                        PFID(&osi->osi_cookie.lgc_lgl.lgl_oi.oi_fid),
 442                        osi->osi_cookie.lgc_lgl.lgl_ogen,
 443                        osi->osi_cookie.lgc_index, rc);
 444                 atomic_inc(&d->opd_sync_changes);
 445         }
 446         /* return 0 always here, error case just cause no llog record */
 447         RETURN(0);
 448 }
 449
 450 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
 451                  llog_op_type type, struct thandle *th,
 452                  const struct lu_attr *attr)
 453 {
 454         return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
 455                                 lu_object_fid(&o->opo_obj.do_lu), type, 1,
 456                                 th, attr);
 457 }
 458
 459 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
 460                         struct lu_fid *fid, int lost, struct thandle *th)
 461 {
 462         return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
 463 }
 464
 465 /*
 466  * it's quite obvious we can't maintain all the structures in the memory:
 467  * while OST is down, MDS can be processing thousands and thousands of unlinks
 468  * filling persistent llogs and in-core respresentation
 469  *
 470  * this doesn't scale at all. so we need basically the following:
 471  * a) destroy/setattr append llog records
 472  * b) once llog has grown to X records, we process first Y committed records
 473  *
 474  *  once record R is found via llog_process(), it becomes committed after any
 475  *  subsequent commit callback (at the most)
 476  */
 477
 478 /**
 479  * ptlrpc commit callback.
 480  *
 481  * The callback is called by PTLRPC when a RPC is reported committed by the
 482  * target (OST). We register the callback for the every RPC applying a change
 483  * from the llog. This way we know then the llog records can be cancelled.
 484  * Notice the callback can be called when OSP is finishing. We can detect this
 485  * checking that actual transno in the request is less or equal of known
 486  * committed transno (see osp_sync_process_committed() for the details).
 487  * XXX: this is pretty expensive and can be improved later using batching.
 488  *
 489  * \param[in] req       request
 490  */
 491 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
 492 {
 493         struct osp_device *d = req->rq_cb_data;
 494         struct osp_job_req_args *jra;
 495
 496         CDEBUG(D_HA, "commit req %p, transno %llu\n", req, req->rq_transno);
 497
 498         if (unlikely(req->rq_transno == 0))
 499                 return;
 500
 501         /* do not do any opd_sync_rpcs_* accounting here
 502          * it's done in osp_sync_interpret sooner or later */
 503         LASSERT(d);
 504
 505         jra = ptlrpc_req_async_args(req);
 506         LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
 507         LASSERT(list_empty(&jra->jra_committed_link));
 508
 509         ptlrpc_request_addref(req);
 510
 511         spin_lock(&d->opd_sync_lock);
 512         list_add(&jra->jra_committed_link, &d->opd_sync_committed_there);
 513         spin_unlock(&d->opd_sync_lock);
 514
 515         /* XXX: some batching wouldn't hurt */
 516         wake_up(&d->opd_sync_waitq);
 517 }
 518
 519 /**
 520  * RPC interpretation callback.
 521  *
 522  * The callback is called by ptlrpc when RPC is replied. Now we have to decide
 523  * whether we should:
 524  *  - put request on a special list to wait until it's committed by the target,
 525  *    if the request is successful
 526  *  - schedule llog record cancel if no target object is found
 527  *  - try later (essentially after reboot) in case of unexpected error
 528  *
 529  * \param[in] env       LU environment provided by the caller
 530  * \param[in] req       request replied
 531  * \param[in] aa        callback data
 532  * \param[in] rc        result of RPC
 533  *
 534  * \retval 0            always
 535  */
 536 static int osp_sync_interpret(const struct lu_env *env,
 537                               struct ptlrpc_request *req, void *aa, int rc)
 538 {
 539         struct osp_device *d = req->rq_cb_data;
 540         struct osp_job_req_args *jra = aa;
 541
 542         if (jra->jra_magic != OSP_JOB_MAGIC) {
 543                 DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
 544                 LBUG();
 545         }
 546         LASSERT(d);
 547
 548         CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
 549                atomic_read(&req->rq_refcount),
 550                rc, (unsigned) req->rq_transno);
 551         LASSERT(rc || req->rq_transno);
 552
 553         if (rc == -ENOENT) {
 554                 /*
 555                  * we tried to destroy object or update attributes,
 556                  * but object doesn't exist anymore - cancell llog record
 557                  */
 558                 LASSERT(req->rq_transno == 0);
 559                 LASSERT(list_empty(&jra->jra_committed_link));
 560
 561                 ptlrpc_request_addref(req);
 562
 563                 spin_lock(&d->opd_sync_lock);
 564                 list_add(&jra->jra_committed_link,
 565                          &d->opd_sync_committed_there);
 566                 spin_unlock(&d->opd_sync_lock);
 567
 568                 wake_up(&d->opd_sync_waitq);
 569         } else if (rc) {
 570                 struct obd_import *imp = req->rq_import;
 571                 /*
 572                  * error happened, we'll try to repeat on next boot ?
 573                  */
 574                 LASSERTF(req->rq_transno == 0 ||
 575                          req->rq_import_generation < imp->imp_generation,
 576                          "transno %llu, rc %d, gen: req %d, imp %d\n",
 577                          req->rq_transno, rc, req->rq_import_generation,
 578                          imp->imp_generation);
 579                 if (req->rq_transno == 0) {
 580                         /* this is the last time we see the request
 581                          * if transno is not zero, then commit cb
 582                          * will be called at some point */
 583                         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) > 0);
 584                         atomic_dec(&d->opd_sync_rpcs_in_progress);
 585                 }
 586
 587                 wake_up(&d->opd_sync_waitq);
 588         } else if (d->opd_pre != NULL &&
 589                    unlikely(d->opd_pre_status == -ENOSPC)) {
 590                 /*
 591                  * if current status is -ENOSPC (lack of free space on OST)
 592                  * then we should poll OST immediately once object destroy
 593                  * is replied
 594                  */
 595                 osp_statfs_need_now(d);
 596         }
 597
 598         spin_lock(&d->opd_sync_lock);
 599         list_del_init(&jra->jra_in_flight_link);
 600         spin_unlock(&d->opd_sync_lock);
 601         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) > 0);
 602         atomic_dec(&d->opd_sync_rpcs_in_flight);
 603         if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
 604                 wake_up(&d->opd_sync_barrier_waitq);
 605         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 606                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 607                atomic_read(&d->opd_sync_rpcs_in_progress));
 608
 609         osp_sync_check_for_work(d);
 610
 611         return 0;
 612 }
 613
 614 /*
 615  ** Add request to ptlrpc queue.
 616  *
 617  * This is just a tiny helper function to put the request on the sending list
 618  *
 619  * \param[in] d         OSP device
 620  * \param[in] llh       llog handle where the record is stored
 621  * \param[in] h         llog record
 622  * \param[in] req       request
 623  */
 624 static void osp_sync_send_new_rpc(struct osp_device *d,
 625                                   struct llog_handle *llh,
 626                                   struct llog_rec_hdr *h,
 627                                   struct ptlrpc_request *req)
 628 {
 629         struct osp_job_req_args *jra;
 630
 631         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) <=
 632                 d->opd_sync_max_rpcs_in_flight);
 633
 634         jra = ptlrpc_req_async_args(req);
 635         jra->jra_magic = OSP_JOB_MAGIC;
 636         jra->jra_lcookie.lgc_lgl = llh->lgh_id;
 637         jra->jra_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 638         jra->jra_lcookie.lgc_index = h->lrh_index;
 639         INIT_LIST_HEAD(&jra->jra_committed_link);
 640         spin_lock(&d->opd_sync_lock);
 641         list_add_tail(&jra->jra_in_flight_link, &d->opd_sync_in_flight_list);
 642         spin_unlock(&d->opd_sync_lock);
 643
 644         ptlrpcd_add_req(req);
 645 }
 646
 647
 648 /**
 649  * Allocate and prepare RPC for a new change.
 650  *
 651  * The function allocates and initializes an RPC which will be sent soon to
 652  * apply the change to the target OST. The request is initialized from the
 653  * llog record passed. Notice only the fields common to all type of changes
 654  * are initialized.
 655  *
 656  * \param[in] d         OSP device
 657  * \param[in] op        type of the change
 658  * \param[in] format    request format to be used
 659  *
 660  * \retval pointer              new request on success
 661  * \retval ERR_PTR(errno)       on error
 662  */
 663 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
 664                                                ost_cmd_t op,
 665                                                const struct req_format *format)
 666 {
 667         struct ptlrpc_request   *req;
 668         struct obd_import       *imp;
 669         int                      rc;
 670
 671         /* Prepare the request */
 672         imp = d->opd_obd->u.cli.cl_import;
 673         LASSERT(imp);
 674
 675         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
 676                 RETURN(ERR_PTR(-ENOMEM));
 677
 678         req = ptlrpc_request_alloc(imp, format);
 679         if (req == NULL)
 680                 RETURN(ERR_PTR(-ENOMEM));
 681
 682         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
 683         if (rc) {
 684                 ptlrpc_req_finished(req);
 685                 return ERR_PTR(rc);
 686         }
 687
 688         req->rq_interpret_reply = osp_sync_interpret;
 689         req->rq_commit_cb = osp_sync_request_commit_cb;
 690         req->rq_cb_data = d;
 691
 692         ptlrpc_request_set_replen(req);
 693
 694         return req;
 695 }
 696
 697 /**
 698  * Generate a request for setattr change.
 699  *
 700  * The function prepares a new RPC, initializes it with setattr specific
 701  * bits and send the RPC.
 702  *
 703  * \param[in] d         OSP device
 704  * \param[in] llh       llog handle where the record is stored
 705  * \param[in] h         llog record
 706  *
 707  * \retval 0            on success
 708  * \retval 1            on invalid record
 709  * \retval negative     negated errno on error
 710  */
 711 static int osp_sync_new_setattr_job(struct osp_device *d,
 712                                     struct llog_handle *llh,
 713                                     struct llog_rec_hdr *h)
 714 {
 715         struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
 716         struct ptlrpc_request           *req;
 717         struct ost_body                 *body;
 718
 719         ENTRY;
 720         LASSERT(h->lrh_type == MDS_SETATTR64_REC);
 721
 722         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
 723                 RETURN(1);
 724
 725         /* lsr_valid can only be 0 or HAVE OBD_MD_{FLUID, FLGID, FLPROJID} set,
 726          * so no bits other than these should be set. */
 727         if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID |
 728             OBD_MD_FLPROJID)) != 0) {
 729                 CERROR("%s: invalid setattr record, lsr_valid:%llu\n",
 730                         d->opd_obd->obd_name, rec->lsr_valid);
 731                 /* return 1 on invalid record */
 732                 RETURN(1);
 733         }
 734
 735         req = osp_sync_new_job(d, OST_SETATTR, &RQF_OST_SETATTR);
 736         if (IS_ERR(req))
 737                 RETURN(PTR_ERR(req));
 738
 739         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 740         LASSERT(body);
 741         body->oa.o_oi = rec->lsr_oi;
 742         body->oa.o_uid = rec->lsr_uid;
 743         body->oa.o_gid = rec->lsr_gid;
 744         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 745         if (h->lrh_len > sizeof(struct llog_setattr64_rec))
 746                 body->oa.o_projid = ((struct llog_setattr64_rec_v2 *)
 747                                       rec)->lsr_projid;
 748
 749         /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
 750          * we assume that both UID and GID are valid in that case. */
 751         if (rec->lsr_valid == 0)
 752                 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
 753         else
 754                 body->oa.o_valid |= rec->lsr_valid;
 755
 756         osp_sync_send_new_rpc(d, llh, h, req);
 757         RETURN(0);
 758 }
 759
 760 /**
 761  * Generate a request for unlink change.
 762  *
 763  * The function prepares a new RPC, initializes it with unlink(destroy)
 764  * specific bits and sends the RPC. The function is used to handle
 765  * llog_unlink_rec which were used in the older versions of Lustre.
 766  * Current version uses llog_unlink_rec64.
 767  *
 768  * \param[in] d         OSP device
 769  * \param[in] llh       llog handle where the record is stored
 770  * \param[in] h         llog record
 771  *
 772  * \retval 0            on success
 773  * \retval negative     negated errno on error
 774  */
 775 static int osp_sync_new_unlink_job(struct osp_device *d,
 776                                    struct llog_handle *llh,
 777                                    struct llog_rec_hdr *h)
 778 {
 779         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 780         struct ptlrpc_request   *req;
 781         struct ost_body         *body;
 782         int rc;
 783
 784         ENTRY;
 785         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 786
 787         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 788         if (IS_ERR(req))
 789                 RETURN(PTR_ERR(req));
 790
 791         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 792         LASSERT(body);
 793         ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
 794         rc = ostid_set_id(&body->oa.o_oi, rec->lur_oid);
 795         if (rc)
 796                 return rc;
 797         body->oa.o_misc = rec->lur_count;
 798         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 799         if (rec->lur_count)
 800                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 801
 802         osp_sync_send_new_rpc(d, llh, h, req);
 803         RETURN(0);
 804 }
 805
 806 /**
 807  * Generate a request for unlink change.
 808  *
 809  * The function prepares a new RPC, initializes it with unlink(destroy)
 810  * specific bits and sends the RPC. Depending on the target (MDT or OST)
 811  * two different protocols are used. For MDT we use OUT (basically OSD API
 812  * updates transferred via a network). For OST we still use the old
 813  * protocol (OBD?), originally for compatibility. Later we can start to
 814  * use OUT for OST as well, this will allow batching and better code
 815  * unification.
 816  *
 817  * \param[in] d         OSP device
 818  * \param[in] llh       llog handle where the record is stored
 819  * \param[in] h         llog record
 820  *
 821  * \retval 0            on success
 822  * \retval negative     negated errno on error
 823  */
 824 static int osp_sync_new_unlink64_job(struct osp_device *d,
 825                                      struct llog_handle *llh,
 826                                      struct llog_rec_hdr *h)
 827 {
 828         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 829         struct ptlrpc_request           *req = NULL;
 830         struct ost_body                 *body;
 831         int                              rc;
 832
 833         ENTRY;
 834         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 835         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 836         if (IS_ERR(req))
 837                 RETURN(PTR_ERR(req));
 838
 839         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 840         if (body == NULL)
 841                 RETURN(-EFAULT);
 842         rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
 843         if (rc < 0)
 844                 RETURN(rc);
 845         body->oa.o_misc = rec->lur_count;
 846         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 847                            OBD_MD_FLOBJCOUNT;
 848         osp_sync_send_new_rpc(d, llh, h, req);
 849         RETURN(0);
 850 }
 851
 852 /**
 853  * Process llog records.
 854  *
 855  * This function is called to process the llog records committed locally.
 856  * In the recovery model used by OSP we can apply a change to a remote
 857  * target once corresponding transaction (like posix unlink) is committed
 858  * locally so can't revert.
 859  * Depending on the llog record type, a given handler is called that is
 860  * responsible for preparing and sending the RPC to apply the change.
 861  * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
 862  *
 863  * \param[in] env       LU environment provided by the caller
 864  * \param[in] d         OSP device
 865  * \param[in] llh       llog handle where the record is stored
 866  * \param[in] rec       llog record
 867  */
 868 static void osp_sync_process_record(const struct lu_env *env,
 869                                     struct osp_device *d,
 870                                     struct llog_handle *llh,
 871                                     struct llog_rec_hdr *rec)
 872 {
 873         struct llog_handle      *cathandle = llh->u.phd.phd_cat_handle;
 874         struct llog_cookie       cookie;
 875         int                      rc = 0;
 876
 877         ENTRY;
 878
 879         cookie.lgc_lgl = llh->lgh_id;
 880         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 881         cookie.lgc_index = rec->lrh_index;
 882
 883         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 884                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 885
 886                 /* we're waiting for the record generated by this instance */
 887                 LASSERT(d->opd_sync_prev_done == 0);
 888                 if (!memcmp(&d->opd_sync_generation, &gen->lgr_gen,
 889                             sizeof(gen->lgr_gen))) {
 890                         CDEBUG(D_HA, "processed all old entries\n");
 891                         d->opd_sync_prev_done = 1;
 892                 }
 893
 894                 /* cancel any generation record */
 895                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 896
 897                 RETURN_EXIT;
 898         }
 899
 900         /*
 901          * now we prepare and fill requests to OST, put them on the queue
 902          * and fire after next commit callback
 903          */
 904
 905         /* notice we increment counters before sending RPC, to be consistent
 906          * in RPC interpret callback which may happen very quickly */
 907         atomic_inc(&d->opd_sync_rpcs_in_flight);
 908         atomic_inc(&d->opd_sync_rpcs_in_progress);
 909
 910         switch (rec->lrh_type) {
 911         /* case MDS_UNLINK_REC is kept for compatibility */
 912         case MDS_UNLINK_REC:
 913                 rc = osp_sync_new_unlink_job(d, llh, rec);
 914                 break;
 915         case MDS_UNLINK64_REC:
 916                 rc = osp_sync_new_unlink64_job(d, llh, rec);
 917                 break;
 918         case MDS_SETATTR64_REC:
 919                 rc = osp_sync_new_setattr_job(d, llh, rec);
 920                 break;
 921         default:
 922                 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
 923                        rec->lrh_type);
 924                 /* treat "unknown record type" as "invalid" */
 925                 rc = 1;
 926                 break;
 927         }
 928
 929         /* For all kinds of records, not matter successful or not,
 930          * we should decrease changes and bump last_processed_id.
 931          */
 932         if (d->opd_sync_prev_done) {
 933                 __u64 correct_id = osp_sync_correct_id(d, rec);
 934                 LASSERT(atomic_read(&d->opd_sync_changes) > 0);
 935                 LASSERT(correct_id <= d->opd_sync_last_committed_id);
 936                 /* NOTE: it's possible to meet same id if
 937                  * OST stores few stripes of same file
 938                  */
 939                 while (1) {
 940                         /* another thread may be trying to set new value */
 941                         rmb();
 942                         if (correct_id > d->opd_sync_last_processed_id) {
 943                                 d->opd_sync_last_processed_id = correct_id;
 944                                 wake_up(&d->opd_sync_barrier_waitq);
 945                         } else
 946                                 break;
 947                 }
 948                 atomic_dec(&d->opd_sync_changes);
 949         }
 950         if (rc != 0) {
 951                 atomic_dec(&d->opd_sync_rpcs_in_flight);
 952                 atomic_dec(&d->opd_sync_rpcs_in_progress);
 953         }
 954
 955         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 956                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 957                atomic_read(&d->opd_sync_rpcs_in_progress));
 958
 959         /* Delete the invalid record */
 960         if (rc == 1) {
 961                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 962                 if (rc != 0)
 963                         CERROR("%s: can't delete invalid record: "
 964                                "fid = "DFID", rec_id = %u, rc = %d\n",
 965                                d->opd_obd->obd_name,
 966                                PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
 967                                rec->lrh_id, rc);
 968         }
 969
 970         CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
 971                rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);
 972
 973         RETURN_EXIT;
 974 }
 975
 976 /**
 977  * Cancel llog records for the committed changes.
 978  *
 979  * The function walks through the list of the committed RPCs and cancels
 980  * corresponding llog records. see osp_sync_request_commit_cb() for the
 981  * details.
 982  *
 983  * \param[in] env       LU environment provided by the caller
 984  * \param[in] d         OSP device
 985  */
 986 static void osp_sync_process_committed(const struct lu_env *env,
 987                                        struct osp_device *d)
 988 {
 989         struct obd_device       *obd = d->opd_obd;
 990         struct obd_import       *imp = obd->u.cli.cl_import;
 991         struct ost_body         *body;
 992         struct ptlrpc_request   *req;
 993         struct llog_ctxt        *ctxt;
 994         struct llog_handle      *llh;
 995         struct list_head         list;
 996         int                      rc, done = 0;
 997
 998         ENTRY;
 999
1000         if (list_empty(&d->opd_sync_committed_there))
1001                 return;
1002
1003         /*
1004          * if current status is -ENOSPC (lack of free space on OST)
1005          * then we should poll OST immediately once object destroy
1006          * is committed.
1007          * notice: we do this upon commit as well because some backends
1008          * (like DMU) do not release space right away.
1009          */
1010         if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
1011                 osp_statfs_need_now(d);
1012
1013         /*
1014          * now cancel them all
1015          * XXX: can we improve this using some batching?
1016          *      with batch RPC that'll happen automatically?
1017          * XXX: can we store ctxt in lod_device and save few cycles ?
1018          */
1019         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1020         LASSERT(ctxt);
1021
1022         llh = ctxt->loc_handle;
1023         LASSERT(llh);
1024
1025         INIT_LIST_HEAD(&list);
1026         spin_lock(&d->opd_sync_lock);
1027         list_splice(&d->opd_sync_committed_there, &list);
1028         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1029         spin_unlock(&d->opd_sync_lock);
1030
1031         while (!list_empty(&list)) {
1032                 struct osp_job_req_args *jra;
1033
1034                 jra = list_entry(list.next, struct osp_job_req_args,
1035                                  jra_committed_link);
1036                 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1037                 list_del_init(&jra->jra_committed_link);
1038
1039                 req = container_of((void *)jra, struct ptlrpc_request,
1040                                    rq_async_args);
1041                 body = req_capsule_client_get(&req->rq_pill,
1042                                               &RMF_OST_BODY);
1043                 LASSERT(body);
1044                 /* import can be closing, thus all commit cb's are
1045                  * called we can check committness directly */
1046                 if (req->rq_import_generation == imp->imp_generation) {
1047                         rc = llog_cat_cancel_records(env, llh, 1,
1048                                                      &jra->jra_lcookie);
1049                         if (rc)
1050                                 CERROR("%s: can't cancel record: %d\n",
1051                                        obd->obd_name, rc);
1052                 } else {
1053                         DEBUG_REQ(D_OTHER, req, "imp_committed = %llu",
1054                                   imp->imp_peer_committed_transno);
1055                 }
1056                 ptlrpc_req_finished(req);
1057                 done++;
1058         }
1059
1060         llog_ctxt_put(ctxt);
1061
1062         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) >= done);
1063         atomic_sub(done, &d->opd_sync_rpcs_in_progress);
1064         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
1065                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
1066                atomic_read(&d->opd_sync_rpcs_in_progress));
1067
1068         osp_sync_check_for_work(d);
1069
1070         /* wake up the thread if requested to stop:
1071          * it might be waiting for in-progress to complete */
1072         if (unlikely(osp_sync_running(d) == 0))
1073                 wake_up(&d->opd_sync_waitq);
1074
1075         EXIT;
1076 }
1077
1078 /**
1079  * The core of the syncing mechanism.
1080  *
1081  * This is a callback called by the llog processing function. Essentially it
1082  * suspends llog processing until there is a record to process (it's supposed
1083  * to be committed locally). The function handles RPCs committed by the target
1084  * and cancels corresponding llog records.
1085  *
1086  * \param[in] env       LU environment provided by the caller
1087  * \param[in] llh       llog handle we're processing
1088  * \param[in] rec       current llog record
1089  * \param[in] data      callback data containing a pointer to the device
1090  *
1091  * \retval 0                    to ask the caller (llog_process()) to continue
1092  * \retval LLOG_PROC_BREAK      to ask the caller to break
1093  */
1094 static int osp_sync_process_queues(const struct lu_env *env,
1095                                    struct llog_handle *llh,
1096                                    struct llog_rec_hdr *rec,
1097                                    void *data)
1098 {
1099         struct osp_device       *d = data;
1100
1101         do {
1102                 struct l_wait_info lwi = { 0 };
1103
1104                 if (!osp_sync_running(d)) {
1105                         CDEBUG(D_HA, "stop llog processing\n");
1106                         return LLOG_PROC_BREAK;
1107                 }
1108
1109                 /* process requests committed by OST */
1110                 osp_sync_process_committed(env, d);
1111
1112                 /* if we there are changes to be processed and we have
1113                  * resources for this ... do now */
1114                 if (osp_sync_can_process_new(d, rec)) {
1115                         if (llh == NULL) {
1116                                 /* ask llog for another record */
1117                                 CDEBUG(D_HA, "%u changes, %u in progress,"
1118                                      " %u in flight\n",
1119                                      atomic_read(&d->opd_sync_changes),
1120                                      atomic_read(&d->opd_sync_rpcs_in_progress),
1121                                      atomic_read(&d->opd_sync_rpcs_in_flight));
1122                                 return 0;
1123                         }
1124                         osp_sync_process_record(env, d, llh, rec);
1125                         llh = NULL;
1126                         rec = NULL;
1127                 }
1128
1129                 if (d->opd_sync_last_processed_id == d->opd_sync_last_used_id)
1130                         osp_sync_remove_from_tracker(d);
1131
1132                 l_wait_event(d->opd_sync_waitq,
1133                              !osp_sync_running(d) ||
1134                              osp_sync_can_process_new(d, rec) ||
1135                              !list_empty(&d->opd_sync_committed_there),
1136                              &lwi);
1137         } while (1);
1138 }
1139
1140 /**
1141  * OSP sync thread.
1142  *
1143  * This thread runs llog_cat_process() scanner calling our callback
1144  * to process llog records. in the callback we implement tricky
1145  * state machine as we don't want to start scanning of the llog again
1146  * and again, also we don't want to process too many records and send
1147  * too many RPCs a time. so, depending on current load (num of changes
1148  * being synced to OST) the callback can suspend awaiting for some
1149  * new conditions, like syncs completed.
1150  *
1151  * In order to process llog records left by previous boots and to allow
1152  * llog_process_thread() to find something (otherwise it'd just exit
1153  * immediately) we add a special GENERATATION record on each boot.
1154  *
1155  * \param[in] _arg      a pointer to thread's arguments
1156  *
1157  * \retval 0            on success
1158  * \retval negative     negated errno on error
1159  */
1160 static int osp_sync_thread(void *_arg)
1161 {
1162         struct osp_device       *d = _arg;
1163         struct ptlrpc_thread    *thread = &d->opd_sync_thread;
1164         struct l_wait_info       lwi = { 0 };
1165         struct llog_ctxt        *ctxt;
1166         struct obd_device       *obd = d->opd_obd;
1167         struct llog_handle      *llh;
1168         struct lu_env            env;
1169         int                      rc, count;
1170
1171         ENTRY;
1172
1173         rc = lu_env_init(&env, LCT_LOCAL);
1174         if (rc) {
1175                 CERROR("%s: can't initialize env: rc = %d\n",
1176                        obd->obd_name, rc);
1177
1178                 spin_lock(&d->opd_sync_lock);
1179                 thread->t_flags = SVC_STOPPED;
1180                 spin_unlock(&d->opd_sync_lock);
1181                 wake_up(&thread->t_ctl_waitq);
1182
1183                 RETURN(rc);
1184         }
1185
1186         spin_lock(&d->opd_sync_lock);
1187         thread->t_flags = SVC_RUNNING;
1188         spin_unlock(&d->opd_sync_lock);
1189         wake_up(&thread->t_ctl_waitq);
1190
1191         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1192         if (ctxt == NULL) {
1193                 CERROR("can't get appropriate context\n");
1194                 GOTO(out, rc = -EINVAL);
1195         }
1196
1197         llh = ctxt->loc_handle;
1198         if (llh == NULL) {
1199                 CERROR("can't get llh\n");
1200                 llog_ctxt_put(ctxt);
1201                 GOTO(out, rc = -EINVAL);
1202         }
1203
1204         rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
1205         if (rc < 0) {
1206                 CERROR("%s: llog process with osp_sync_process_queues "
1207                        "failed: %d\n", d->opd_obd->obd_name, rc);
1208                 GOTO(close, rc);
1209         }
1210         LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1211                  "%u changes, %u in progress, %u in flight: %d\n",
1212                  atomic_read(&d->opd_sync_changes),
1213                  atomic_read(&d->opd_sync_rpcs_in_progress),
1214                  atomic_read(&d->opd_sync_rpcs_in_flight), rc);
1215
1216         /* we don't expect llog_process_thread() to exit till umount */
1217         LASSERTF(thread->t_flags != SVC_RUNNING,
1218                  "%u changes, %u in progress, %u in flight\n",
1219                  atomic_read(&d->opd_sync_changes),
1220                  atomic_read(&d->opd_sync_rpcs_in_progress),
1221                  atomic_read(&d->opd_sync_rpcs_in_flight));
1222
1223         /* wait till all the requests are completed */
1224         count = 0;
1225         while (atomic_read(&d->opd_sync_rpcs_in_progress) > 0) {
1226                 osp_sync_process_committed(&env, d);
1227
1228                 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1229                 rc = l_wait_event(d->opd_sync_waitq,
1230                                 atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1231                                   &lwi);
1232                 if (rc == -ETIMEDOUT)
1233                         count++;
1234                 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1235                          d->opd_obd->obd_name,
1236                          atomic_read(&d->opd_sync_rpcs_in_progress),
1237                          atomic_read(&d->opd_sync_rpcs_in_flight),
1238                          list_empty(&d->opd_sync_committed_there) ? "" : "!");
1239
1240         }
1241
1242 close:
1243         llog_cat_close(&env, llh);
1244         rc = llog_cleanup(&env, ctxt);
1245         if (rc)
1246                 CERROR("can't cleanup llog: %d\n", rc);
1247 out:
1248         LASSERTF(atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1249                  "%s: %d %d %sempty\n", d->opd_obd->obd_name,
1250                  atomic_read(&d->opd_sync_rpcs_in_progress),
1251                  atomic_read(&d->opd_sync_rpcs_in_flight),
1252                  list_empty(&d->opd_sync_committed_there) ? "" : "!");
1253
1254         thread->t_flags = SVC_STOPPED;
1255
1256         wake_up(&thread->t_ctl_waitq);
1257
1258         lu_env_fini(&env);
1259
1260         RETURN(0);
1261 }
1262
1263 /**
1264  * Initialize llog.
1265  *
1266  * Initializes the llog. Specific llog to be used depends on the type of the
1267  * target OSP represents (OST or MDT). The function adds appends a new llog
1268  * record to mark the place where the records associated with this boot
1269  * start.
1270  *
1271  * \param[in] env       LU environment provided by the caller
1272  * \param[in] d         OSP device
1273  *
1274  * \retval 0            on success
1275  * \retval negative     negated errno on error
1276  */
1277 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1278 {
1279         struct osp_thread_info  *osi = osp_env_info(env);
1280         struct lu_fid           *fid = &osi->osi_fid;
1281         struct llog_handle      *lgh = NULL;
1282         struct obd_device       *obd = d->opd_obd;
1283         struct llog_ctxt        *ctxt;
1284         int                     rc;
1285
1286         ENTRY;
1287
1288         LASSERT(obd);
1289
1290         /*
1291          * open llog corresponding to our OST
1292          */
1293         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1294         obd->obd_lvfs_ctxt.dt = d->opd_storage;
1295
1296         lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1297
1298         rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1299                                    &osi->osi_cid, fid);
1300         if (rc < 0) {
1301                 if (rc != -EFAULT) {
1302                         CERROR("%s: can't get id from catalogs: rc = %d\n",
1303                                obd->obd_name, rc);
1304                         RETURN(rc);
1305                 }
1306
1307                 /* After sparse OST indices is supported, the CATALOG file
1308                  * may become a sparse file that results in failure on
1309                  * reading. Skip this error as the llog will be created
1310                  * later */
1311                 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1312                 rc = 0;
1313         }
1314
1315         CDEBUG(D_INFO, "%s: Init llog for %d - catid "DFID":%x\n",
1316                obd->obd_name, d->opd_index,
1317                PFID(&osi->osi_cid.lci_logid.lgl_oi.oi_fid),
1318                osi->osi_cid.lci_logid.lgl_ogen);
1319
1320         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT,
1321                         d->opd_storage->dd_lu_dev.ld_obd,
1322                         &osp_mds_ost_orig_logops);
1323         if (rc)
1324                 RETURN(rc);
1325
1326         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1327         LASSERT(ctxt);
1328
1329         if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1330                 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1331                                LLOG_OPEN_EXISTS);
1332                 /* re-create llog if it is missing */
1333                 if (rc == -ENOENT)
1334                         logid_set_id(&osi->osi_cid.lci_logid, 0);
1335                 else if (rc < 0)
1336                         GOTO(out_cleanup, rc);
1337         }
1338
1339         if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1340                 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1341                 if (rc < 0)
1342                         GOTO(out_cleanup, rc);
1343                 osi->osi_cid.lci_logid = lgh->lgh_id;
1344         }
1345
1346         LASSERT(lgh != NULL);
1347         ctxt->loc_handle = lgh;
1348
1349         rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
1350         if (rc)
1351                 GOTO(out_close, rc);
1352
1353         rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1354                                    &osi->osi_cid, fid);
1355         if (rc)
1356                 GOTO(out_close, rc);
1357
1358         /*
1359          * put a mark in the llog till which we'll be processing
1360          * old records restless
1361          */
1362         d->opd_sync_generation.mnt_cnt = cfs_time_current();
1363         d->opd_sync_generation.conn_cnt = cfs_time_current();
1364
1365         osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1366         osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1367
1368         memcpy(&osi->osi_gen.lgr_gen, &d->opd_sync_generation,
1369                sizeof(osi->osi_gen.lgr_gen));
1370
1371         rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1372         if (rc < 0)
1373                 GOTO(out_close, rc);
1374         llog_ctxt_put(ctxt);
1375         RETURN(0);
1376 out_close:
1377         llog_cat_close(env, lgh);
1378 out_cleanup:
1379         llog_cleanup(env, ctxt);
1380         RETURN(rc);
1381 }
1382
1383 /**
1384  * Cleanup llog used for syncing.
1385  *
1386  * Closes and cleanups the llog. The function is called when the device is
1387  * shutting down.
1388  *
1389  * \param[in] env       LU environment provided by the caller
1390  * \param[in] d         OSP device
1391  */
1392 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1393 {
1394         struct llog_ctxt *ctxt;
1395
1396         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1397         if (ctxt) {
1398                 llog_cat_close(env, ctxt->loc_handle);
1399                 llog_cleanup(env, ctxt);
1400         }
1401 }
1402
1403 /**
1404  * Initialization of the sync component of OSP.
1405  *
1406  * Initializes the llog and starts a new thread to handle the changes to
1407  * the remote target (OST or MDT).
1408  *
1409  * \param[in] env       LU environment provided by the caller
1410  * \param[in] d         OSP device
1411  *
1412  * \retval 0            on success
1413  * \retval negative     negated errno on error
1414  */
1415 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1416 {
1417         struct l_wait_info       lwi = { 0 };
1418         struct task_struct      *task;
1419         int                      rc;
1420
1421         ENTRY;
1422
1423         d->opd_sync_max_rpcs_in_flight = OSP_MAX_RPCS_IN_FLIGHT;
1424         d->opd_sync_max_rpcs_in_progress = OSP_MAX_RPCS_IN_PROGRESS;
1425         spin_lock_init(&d->opd_sync_lock);
1426         init_waitqueue_head(&d->opd_sync_waitq);
1427         init_waitqueue_head(&d->opd_sync_barrier_waitq);
1428         thread_set_flags(&d->opd_sync_thread, SVC_INIT);
1429         init_waitqueue_head(&d->opd_sync_thread.t_ctl_waitq);
1430         INIT_LIST_HEAD(&d->opd_sync_in_flight_list);
1431         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1432
1433         if (d->opd_storage->dd_rdonly)
1434                 RETURN(0);
1435
1436         rc = osp_sync_id_traction_init(d);
1437         if (rc)
1438                 RETURN(rc);
1439
1440         /*
1441          * initialize llog storing changes
1442          */
1443         rc = osp_sync_llog_init(env, d);
1444         if (rc) {
1445                 CERROR("%s: can't initialize llog: rc = %d\n",
1446                        d->opd_obd->obd_name, rc);
1447                 GOTO(err_id, rc);
1448         }
1449
1450         /*
1451          * Start synchronization thread
1452          */
1453         task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1454                            d->opd_index, d->opd_group);
1455         if (IS_ERR(task)) {
1456                 rc = PTR_ERR(task);
1457                 CERROR("%s: cannot start sync thread: rc = %d\n",
1458                        d->opd_obd->obd_name, rc);
1459                 GOTO(err_llog, rc);
1460         }
1461
1462         l_wait_event(d->opd_sync_thread.t_ctl_waitq,
1463                      osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1464
1465         RETURN(0);
1466 err_llog:
1467         osp_sync_llog_fini(env, d);
1468 err_id:
1469         osp_sync_id_traction_fini(d);
1470         return rc;
1471 }
1472
1473 /**
1474  * Stop the syncing thread.
1475  *
1476  * Asks the syncing thread to stop and wait until it's stopped.
1477  *
1478  * \param[in] d         OSP device
1479  *
1480  * \retval              0
1481  */
1482 int osp_sync_fini(struct osp_device *d)
1483 {
1484         struct ptlrpc_thread *thread = &d->opd_sync_thread;
1485
1486         ENTRY;
1487
1488         if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1489                 thread->t_flags = SVC_STOPPING;
1490                 wake_up(&d->opd_sync_waitq);
1491                 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1492         }
1493
1494         /*
1495          * unregister transaction callbacks only when sync thread
1496          * has finished operations with llog
1497          */
1498         osp_sync_id_traction_fini(d);
1499
1500         RETURN(0);
1501 }
1502
1503 static DEFINE_MUTEX(osp_id_tracker_sem);
1504 static struct list_head osp_id_tracker_list =
1505                 LIST_HEAD_INIT(osp_id_tracker_list);
1506
1507 /**
1508  * OSD commit callback.
1509  *
1510  * The function is used as a local OSD commit callback to track the highest
1511  * committed llog record id. see osp_sync_id_traction_init() for the details.
1512  *
1513  * \param[in] th        local transaction handle committed
1514  * \param[in] cookie    commit callback data (our private structure)
1515  */
1516 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1517 {
1518         struct osp_id_tracker   *tr = cookie;
1519         struct osp_device       *d;
1520         struct osp_txn_info     *txn;
1521
1522         LASSERT(tr);
1523
1524         txn = osp_txn_info(&th->th_ctx);
1525         if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1526                 return;
1527
1528         spin_lock(&tr->otr_lock);
1529         if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1530                 CDEBUG(D_OTHER, "committed: %llu -> %llu\n",
1531                        tr->otr_committed_id, txn->oti_current_id);
1532                 tr->otr_committed_id = txn->oti_current_id;
1533
1534                 list_for_each_entry(d, &tr->otr_wakeup_list,
1535                                     opd_sync_ontrack) {
1536                         d->opd_sync_last_committed_id = tr->otr_committed_id;
1537                         wake_up(&d->opd_sync_waitq);
1538                 }
1539         }
1540         spin_unlock(&tr->otr_lock);
1541 }
1542
1543 /**
1544  * Initialize commit tracking mechanism.
1545  *
1546  * Some setups may have thousands of OSTs and each will be represented by OSP.
1547  * Meaning order of magnitute many more changes to apply every second. In order
1548  * to keep the number of commit callbacks low this mechanism was introduced.
1549  * The mechanism is very similar to transno used by MDT service: it's an single
1550  * ID stream which can be assigned by any OSP to its llog records. The tricky
1551  * part is that ID is stored in per-transaction data and re-used by all the OSPs
1552  * involved in that transaction. Then all these OSPs are woken up utilizing a single OSD commit callback.
1553  *
1554  * The function initializes the data used by the tracker described above.
1555  * A singler tracker per OSD device is created.
1556  *
1557  * \param[in] d         OSP device
1558  *
1559  * \retval 0            on success
1560  * \retval negative     negated errno on error
1561  */
1562 static int osp_sync_id_traction_init(struct osp_device *d)
1563 {
1564         struct osp_id_tracker   *tr, *found = NULL;
1565         int                      rc = 0;
1566
1567         LASSERT(d);
1568         LASSERT(d->opd_storage);
1569         LASSERT(d->opd_sync_tracker == NULL);
1570         INIT_LIST_HEAD(&d->opd_sync_ontrack);
1571
1572         mutex_lock(&osp_id_tracker_sem);
1573         list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1574                 if (tr->otr_dev == d->opd_storage) {
1575                         LASSERT(atomic_read(&tr->otr_refcount));
1576                         atomic_inc(&tr->otr_refcount);
1577                         d->opd_sync_tracker = tr;
1578                         found = tr;
1579                         break;
1580                 }
1581         }
1582
1583         if (found == NULL) {
1584                 rc = -ENOMEM;
1585                 OBD_ALLOC_PTR(tr);
1586                 if (tr) {
1587                         d->opd_sync_tracker = tr;
1588                         spin_lock_init(&tr->otr_lock);
1589                         tr->otr_dev = d->opd_storage;
1590                         tr->otr_next_id = 1;
1591                         tr->otr_committed_id = 0;
1592                         atomic_set(&tr->otr_refcount, 1);
1593                         INIT_LIST_HEAD(&tr->otr_wakeup_list);
1594                         list_add(&tr->otr_list, &osp_id_tracker_list);
1595                         tr->otr_tx_cb.dtc_txn_commit =
1596                                                 osp_sync_tracker_commit_cb;
1597                         tr->otr_tx_cb.dtc_cookie = tr;
1598                         tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1599                         dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1600                         rc = 0;
1601                 }
1602         }
1603         mutex_unlock(&osp_id_tracker_sem);
1604
1605         return rc;
1606 }
1607
1608 /**
1609  * Release commit tracker.
1610  *
1611  * Decrease a refcounter on the tracker used by the given OSP device \a d.
1612  * If no more users left, then the tracker is released.
1613  *
1614  * \param[in] d         OSP device
1615  */
1616 static void osp_sync_id_traction_fini(struct osp_device *d)
1617 {
1618         struct osp_id_tracker *tr;
1619
1620         ENTRY;
1621
1622         LASSERT(d);
1623         tr = d->opd_sync_tracker;
1624         if (tr == NULL) {
1625                 EXIT;
1626                 return;
1627         }
1628
1629         osp_sync_remove_from_tracker(d);
1630
1631         mutex_lock(&osp_id_tracker_sem);
1632         if (atomic_dec_and_test(&tr->otr_refcount)) {
1633                 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1634                 LASSERT(list_empty(&tr->otr_wakeup_list));
1635                 list_del(&tr->otr_list);
1636                 OBD_FREE_PTR(tr);
1637                 d->opd_sync_tracker = NULL;
1638         }
1639         mutex_unlock(&osp_id_tracker_sem);
1640
1641         EXIT;
1642 }
1643
1644 /**
1645  * Generate a new ID on a tracker.
1646  *
1647  * Generates a new ID using the tracker associated with the given OSP device
1648  * \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
1649  * the wakeup list, so OSP won't miss when a transaction using the ID is
1650  * committed.
1651  *
1652  * \param[in] d         OSP device
1653  * \param[in] id        0 or ID generated previously
1654  *
1655  * \retval              ID the caller should use
1656  */
1657 static __u64 osp_sync_id_get(struct osp_device *d, __u64 id)
1658 {
1659         struct osp_id_tracker *tr;
1660
1661         tr = d->opd_sync_tracker;
1662         LASSERT(tr);
1663
1664         /* XXX: we can improve this introducing per-cpu preallocated ids? */
1665         spin_lock(&tr->otr_lock);
1666         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_TRACK_OVERFLOW))
1667                 tr->otr_next_id = 0xfffffff0;
1668
1669         if (unlikely(tr->otr_next_id <= d->opd_sync_last_used_id)) {
1670                 spin_unlock(&tr->otr_lock);
1671                 CERROR("%s: next %llu, last synced %llu\n",
1672                        d->opd_obd->obd_name, tr->otr_next_id,
1673                        d->opd_sync_last_used_id);
1674                 LBUG();
1675         }
1676
1677         if (id == 0)
1678                 id = tr->otr_next_id++;
1679         if (id > d->opd_sync_last_used_id)
1680                 d->opd_sync_last_used_id = id;
1681         if (list_empty(&d->opd_sync_ontrack))
1682                 list_add(&d->opd_sync_ontrack, &tr->otr_wakeup_list);
1683         spin_unlock(&tr->otr_lock);
1684         CDEBUG(D_OTHER, "new id %llu\n", id);
1685
1686         return id;
1687 }
1688
1689 /**
1690  * Stop to propagate commit status to OSP.
1691  *
1692  * If the OSP does not have any llog records she's waiting to commit, then
1693  * it is possible to unsubscribe from wakeups from the tracking using this
1694  * method.
1695  *
1696  * \param[in] d         OSP device not willing to wakeup
1697  */
1698 static void osp_sync_remove_from_tracker(struct osp_device *d)
1699 {
1700         struct osp_id_tracker *tr;
1701
1702         tr = d->opd_sync_tracker;
1703         LASSERT(tr);
1704
1705         if (list_empty(&d->opd_sync_ontrack))
1706                 return;
1707
1708         spin_lock(&tr->otr_lock);
1709         list_del_init(&d->opd_sync_ontrack);
1710         spin_unlock(&tr->otr_lock);
1711 }
1712