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 /*
  48  * this is a components of OSP implementing synchronization between MDS and OST
  49  * it llogs all interesting changes (currently it's uig/gid change and object
  50  * destroy) atomically, then makes sure changes hit OST storage
  51  *
  52  * we have 4 queues of work:
  53  *
  54  * the first queue is llog itself, once read a change is stored in 2nd queue
  55  * in form of RPC (but RPC isn't fired yet).
  56  *
  57  * the second queue (opd_sync_waiting_for_commit) holds changes awaiting local
  58  * commit. once change is committed locally it migrates onto 3rd queue.
  59  *
  60  * the third queue (opd_sync_committed_here) holds changes committed locally,
  61  * but not sent to OST (as the pipe can be full). once pipe becomes non-full
  62  * we take a change from the queue and fire corresponded RPC.
  63  *
  64  * once RPC is reported committed by OST (using regular last_committed mech.)
  65  * the change jumps into 4th queue (opd_sync_committed_there), now we can
  66  * cancel corresponded llog record and release RPC
  67  *
  68  * opd_sync_changes is a number of unread llog records (to be processed).
  69  * notice this number doesn't include llog records from previous boots.
  70  * with OSP_SYNC_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
  71  *
  72  * opd_sync_rpcs_in_progress is total number of requests in above 2-4 queues.
  73  * we control this with OSP_MAX_RPCS_IN_PROGRESS so that OSP don't consume
  74  * too much memory -- how to deal with 1000th OSTs ? batching could help?
  75  *
  76  * opd_sync_rpcs_in_flight is a number of RPC in flight.
  77  * we control this with OSP_MAX_RPCS_IN_FLIGHT
  78  */
  79
  80 /* XXX: do math to learn reasonable threshold
  81  * should it be ~ number of changes fitting bulk? */
  82
  83 #define OSP_SYNC_THRESHOLD              10
  84 #define OSP_MAX_RPCS_IN_FLIGHT          8
  85 #define OSP_MAX_RPCS_IN_PROGRESS        4096
  86
  87 #define OSP_JOB_MAGIC           0x26112005
  88
  89 struct osp_job_req_args {
  90         /** bytes reserved for ptlrpc_replay_req() */
  91         struct ptlrpc_replay_async_args jra_raa;
  92         struct list_head                jra_committed_link;
  93         struct list_head                jra_in_flight_link;
  94         struct llog_cookie              jra_lcookie;
  95         __u32                           jra_magic;
  96 };
  97
  98 static inline int osp_sync_running(struct osp_device *d)
  99 {
 100         return !!(d->opd_sync_thread.t_flags & SVC_RUNNING);
 101 }
 102
 103 /**
 104  * Check status: whether OSP thread has stopped
 105  *
 106  * \param[in] d         OSP device
 107  *
 108  * \retval 0            still running
 109  * \retval 1            stopped
 110  */
 111 static inline int osp_sync_stopped(struct osp_device *d)
 112 {
 113         return !!(d->opd_sync_thread.t_flags & SVC_STOPPED);
 114 }
 115
 116 /*
 117  ** Check for new changes to sync
 118  *
 119  * \param[in] d         OSP device
 120  *
 121  * \retval 1            there are changes
 122  * \retval 0            there are no changes
 123  */
 124 static inline int osp_sync_has_new_job(struct osp_device *d)
 125 {
 126         return atomic_read(&d->opd_sync_changes) > 0 ||
 127                 d->opd_sync_prev_done == 0;
 128 }
 129
 130 static inline int osp_sync_in_flight_conflict(struct osp_device *d,
 131                                              struct llog_rec_hdr *h)
 132 {
 133         struct osp_job_req_args *jra;
 134         struct ost_id            ostid;
 135         int                      conflict = 0;
 136
 137         if (h == NULL || h->lrh_type == LLOG_GEN_REC ||
 138             list_empty(&d->opd_sync_in_flight_list))
 139                 return conflict;
 140
 141         memset(&ostid, 0, sizeof(ostid));
 142         switch (h->lrh_type) {
 143         case MDS_UNLINK_REC: {
 144                 struct llog_unlink_rec *unlink = (struct llog_unlink_rec *)h;
 145
 146                 ostid_set_seq(&ostid, unlink->lur_oseq);
 147                 if (ostid_set_id(&ostid, unlink->lur_oid)) {
 148                         CERROR("Bad %llu to set " DOSTID "\n",
 149                                (unsigned long long)(unlink->lur_oid),
 150                                POSTID(&ostid));
 151                         return 1;
 152                 }
 153                 }
 154                 break;
 155         case MDS_UNLINK64_REC:
 156                 fid_to_ostid(&((struct llog_unlink64_rec *)h)->lur_fid, &ostid);
 157                 break;
 158         case MDS_SETATTR64_REC:
 159                 ostid = ((struct llog_setattr64_rec *)h)->lsr_oi;
 160                 break;
 161         default:
 162                 LBUG();
 163         }
 164
 165         spin_lock(&d->opd_sync_lock);
 166         list_for_each_entry(jra, &d->opd_sync_in_flight_list,
 167                             jra_in_flight_link) {
 168                 struct ptlrpc_request   *req;
 169                 struct ost_body         *body;
 170
 171                 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
 172
 173                 req = container_of((void *)jra, struct ptlrpc_request,
 174                                    rq_async_args);
 175                 body = req_capsule_client_get(&req->rq_pill,
 176                                               &RMF_OST_BODY);
 177                 LASSERT(body);
 178
 179                 if (memcmp(&ostid, &body->oa.o_oi, sizeof(ostid)) == 0) {
 180                         conflict = 1;
 181                         break;
 182                 }
 183         }
 184         spin_unlock(&d->opd_sync_lock);
 185
 186         return conflict;
 187 }
 188
 189 static inline int osp_sync_rpcs_in_progress_low(struct osp_device *d)
 190 {
 191         return atomic_read(&d->opd_sync_rpcs_in_progress) <
 192                 d->opd_sync_max_rpcs_in_progress;
 193 }
 194
 195 /**
 196  * Check for room in the network pipe to OST
 197  *
 198  * \param[in] d         OSP device
 199  *
 200  * \retval 1            there is room
 201  * \retval 0            no room, the pipe is full
 202  */
 203 static inline int osp_sync_rpcs_in_flight_low(struct osp_device *d)
 204 {
 205         return atomic_read(&d->opd_sync_rpcs_in_flight) <
 206                 d->opd_sync_max_rpcs_in_flight;
 207 }
 208
 209 /**
 210  * Wake up check for the main sync thread
 211  *
 212  * \param[in] d         OSP device
 213  *
 214  * \retval 1            time to wake up
 215  * \retval 0            no need to wake up
 216  */
 217 static inline int osp_sync_has_work(struct osp_device *osp)
 218 {
 219         /* has new/old changes and low in-progress? */
 220         if (osp_sync_has_new_job(osp) && osp_sync_rpcs_in_progress_low(osp) &&
 221             osp_sync_rpcs_in_flight_low(osp) && osp->opd_imp_connected)
 222                 return 1;
 223
 224         /* has remotely committed? */
 225         if (!list_empty(&osp->opd_sync_committed_there))
 226                 return 1;
 227
 228         return 0;
 229 }
 230
 231 void osp_sync_check_for_work(struct osp_device *osp)
 232 {
 233         if (osp_sync_has_work(osp))
 234                 wake_up(&osp->opd_sync_waitq);
 235 }
 236
 237 static inline __u64 osp_sync_correct_id(struct osp_device *d,
 238                                         struct llog_rec_hdr *rec)
 239 {
 240         /*
 241          * llog use cyclic store with 32 bit lrh_id
 242          * so overflow lrh_id is possible. Range between
 243          * last_processed and last_committed is less than
 244          * 64745 ^ 2 and less than 2^32 - 1
 245          */
 246         __u64 correct_id = d->opd_sync_last_committed_id;
 247
 248         if ((correct_id & 0xffffffffULL) < rec->lrh_id)
 249                 correct_id -= 0x100000000ULL;
 250
 251         correct_id &= ~0xffffffffULL;
 252         correct_id |= rec->lrh_id;
 253
 254         return correct_id;
 255 }
 256
 257 /**
 258  * Check and return ready-for-new status.
 259  *
 260  * The thread processing llog record uses this function to check whether
 261  * it's time to take another record and process it. The number of conditions
 262  * must be met: the connection should be ready, RPCs in flight not exceeding
 263  * the limit, the record is committed locally, etc (see the lines below).
 264  *
 265  * \param[in] d         OSP device
 266  * \param[in] rec       next llog record to process
 267  *
 268  * \retval 0            not ready
 269  * \retval 1            ready
 270  */
 271 static inline int osp_sync_can_process_new(struct osp_device *d,
 272                                            struct llog_rec_hdr *rec)
 273 {
 274         LASSERT(d);
 275
 276         if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
 277                 return 0;
 278         if (unlikely(osp_sync_in_flight_conflict(d, rec)))
 279                 return 0;
 280         if (!osp_sync_rpcs_in_progress_low(d))
 281                 return 0;
 282         if (!osp_sync_rpcs_in_flight_low(d))
 283                 return 0;
 284         if (!d->opd_imp_connected)
 285                 return 0;
 286         if (d->opd_sync_prev_done == 0)
 287                 return 1;
 288         if (atomic_read(&d->opd_sync_changes) == 0)
 289                 return 0;
 290         if (rec == NULL)
 291                 return 1;
 292         /* notice "<" not "<=" */
 293         if (osp_sync_correct_id(d, rec) < d->opd_sync_last_committed_id)
 294                 return 1;
 295         return 0;
 296 }
 297
 298 /**
 299  * Declare intention to add a new change.
 300  *
 301  * With regard to OSD API, we have to declare any changes ahead. In this
 302  * case we declare an intention to add a llog record representing the
 303  * change on the local storage.
 304  *
 305  * \param[in] env       LU environment provided by the caller
 306  * \param[in] o         OSP object
 307  * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 308  * \param[in] th        transaction handle (local)
 309  *
 310  * \retval 0            on success
 311  * \retval negative     negated errno on error
 312  */
 313 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
 314                          llog_op_type type, struct thandle *th)
 315 {
 316         struct osp_thread_info  *osi = osp_env_info(env);
 317         struct osp_device       *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
 318         struct llog_ctxt        *ctxt;
 319         struct thandle          *storage_th;
 320         int                      rc;
 321
 322         ENTRY;
 323
 324         /* it's a layering violation, to access internals of th,
 325          * but we can do this as a sanity check, for a while */
 326         LASSERT(th->th_top != NULL);
 327         storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
 328         if (IS_ERR(storage_th))
 329                 RETURN(PTR_ERR(storage_th));
 330
 331         switch (type) {
 332         case MDS_UNLINK64_REC:
 333                 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
 334                 break;
 335         case MDS_SETATTR64_REC:
 336                 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec_v2);
 337                 break;
 338         default:
 339                 LBUG();
 340         }
 341
 342         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 343         LASSERT(ctxt);
 344
 345         rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
 346                               storage_th);
 347         llog_ctxt_put(ctxt);
 348
 349         RETURN(rc);
 350 }
 351
 352 /* add the commit callback every second */
 353 int osp_sync_add_commit_cb_1s(const struct lu_env *env, struct osp_device *d,
 354                               struct thandle *th)
 355 {
 356         int add = 0;
 357
 358         /* fast path */
 359         if (cfs_time_before(cfs_time_current(), d->opd_sync_next_commit_cb))
 360                 return 0;
 361
 362         spin_lock(&d->opd_sync_lock);
 363         if (cfs_time_aftereq(cfs_time_current(), d->opd_sync_next_commit_cb))
 364                 add = 1;
 365         d->opd_sync_next_commit_cb = cfs_time_shift(1);
 366         spin_unlock(&d->opd_sync_lock);
 367
 368         if (add == 0)
 369                 return 0;
 370         return osp_sync_add_commit_cb(env, d, th);
 371 }
 372
 373
 374 /**
 375  * Generate a llog record for a given change.
 376  *
 377  * Generates a llog record for the change passed. The change can be of two
 378  * types: unlink and setattr. The record gets an ID which later will be
 379  * used to track commit status of the change. For unlink changes, the caller
 380  * can supply a starting FID and the count of the objects to destroy. For
 381  * setattr the caller should apply attributes to apply.
 382  *
 383  *
 384  * \param[in] env       LU environment provided by the caller
 385  * \param[in] d         OSP device
 386  * \param[in] fid       fid of the object the change should be applied to
 387  * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 388  * \param[in] count     count of objects to destroy
 389  * \param[in] th        transaction handle (local)
 390  * \param[in] attr      attributes for setattr
 391  *
 392  * \retval 0            on success
 393  * \retval negative     negated errno on error
 394  */
 395 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
 396                             const struct lu_fid *fid, llog_op_type type,
 397                             int count, struct thandle *th,
 398                             const struct lu_attr *attr)
 399 {
 400         struct osp_thread_info  *osi = osp_env_info(env);
 401         struct llog_ctxt        *ctxt;
 402         struct thandle          *storage_th;
 403         int                      rc;
 404
 405         ENTRY;
 406
 407         /* it's a layering violation, to access internals of th,
 408          * but we can do this as a sanity check, for a while */
 409         LASSERT(th->th_top != NULL);
 410         storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
 411         if (IS_ERR(storage_th))
 412                 RETURN(PTR_ERR(storage_th));
 413
 414         switch (type) {
 415         case MDS_UNLINK64_REC:
 416                 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
 417                 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
 418                 osi->osi_unlink.lur_fid  = *fid;
 419                 osi->osi_unlink.lur_count = count;
 420                 break;
 421         case MDS_SETATTR64_REC:
 422                 rc = fid_to_ostid(fid, &osi->osi_oi);
 423                 LASSERT(rc == 0);
 424                 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
 425                 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
 426                 osi->osi_setattr.lsr_oi  = osi->osi_oi;
 427                 LASSERT(attr);
 428                 osi->osi_setattr.lsr_uid = attr->la_uid;
 429                 osi->osi_setattr.lsr_gid = attr->la_gid;
 430                 osi->osi_setattr.lsr_projid = attr->la_projid;
 431                 osi->osi_setattr.lsr_valid =
 432                         ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
 433                         ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0) |
 434                         ((attr->la_valid & LA_PROJID) ? OBD_MD_FLPROJID : 0);
 435                 break;
 436         default:
 437                 LBUG();
 438         }
 439
 440         /* we keep the same id, but increment it when the callback
 441          * is registered, so that all records upto the one taken
 442          * by the callback are subject to processing */
 443         spin_lock(&d->opd_sync_lock);
 444         osi->osi_hdr.lrh_id = d->opd_sync_last_used_id;
 445         spin_unlock(&d->opd_sync_lock);
 446
 447         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 448         if (ctxt == NULL)
 449                 RETURN(-ENOMEM);
 450
 451         rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
 452                       storage_th);
 453         llog_ctxt_put(ctxt);
 454
 455         if (likely(rc >= 0)) {
 456                 CDEBUG(D_OTHER, "%s: new record "DFID":%x.%u: rc = %d\n",
 457                        d->opd_obd->obd_name,
 458                        PFID(&osi->osi_cookie.lgc_lgl.lgl_oi.oi_fid),
 459                        osi->osi_cookie.lgc_lgl.lgl_ogen,
 460                        osi->osi_cookie.lgc_index, rc);
 461                 atomic_inc(&d->opd_sync_changes);
 462         }
 463
 464         rc = osp_sync_add_commit_cb_1s(env, d, th);
 465
 466         /* return 0 always here, error case just cause no llog record */
 467         RETURN(0);
 468 }
 469
 470 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
 471                  llog_op_type type, struct thandle *th,
 472                  const struct lu_attr *attr)
 473 {
 474         return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
 475                                 lu_object_fid(&o->opo_obj.do_lu), type, 1,
 476                                 th, attr);
 477 }
 478
 479 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
 480                         struct lu_fid *fid, int lost, struct thandle *th)
 481 {
 482         return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
 483 }
 484
 485 /*
 486  * it's quite obvious we can't maintain all the structures in the memory:
 487  * while OST is down, MDS can be processing thousands and thousands of unlinks
 488  * filling persistent llogs and in-core respresentation
 489  *
 490  * this doesn't scale at all. so we need basically the following:
 491  * a) destroy/setattr append llog records
 492  * b) once llog has grown to X records, we process first Y committed records
 493  *
 494  *  once record R is found via llog_process(), it becomes committed after any
 495  *  subsequent commit callback (at the most)
 496  */
 497
 498 /**
 499  * ptlrpc commit callback.
 500  *
 501  * The callback is called by PTLRPC when a RPC is reported committed by the
 502  * target (OST). We register the callback for the every RPC applying a change
 503  * from the llog. This way we know then the llog records can be cancelled.
 504  * Notice the callback can be called when OSP is finishing. We can detect this
 505  * checking that actual transno in the request is less or equal of known
 506  * committed transno (see osp_sync_process_committed() for the details).
 507  * XXX: this is pretty expensive and can be improved later using batching.
 508  *
 509  * \param[in] req       request
 510  */
 511 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
 512 {
 513         struct osp_device *d = req->rq_cb_data;
 514         struct osp_job_req_args *jra;
 515
 516         CDEBUG(D_HA, "commit req %p, transno %llu\n", req, req->rq_transno);
 517
 518         if (unlikely(req->rq_transno == 0))
 519                 return;
 520
 521         /* do not do any opd_sync_rpcs_* accounting here
 522          * it's done in osp_sync_interpret sooner or later */
 523         LASSERT(d);
 524
 525         jra = ptlrpc_req_async_args(req);
 526         LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
 527         LASSERT(list_empty(&jra->jra_committed_link));
 528
 529         ptlrpc_request_addref(req);
 530
 531         spin_lock(&d->opd_sync_lock);
 532         list_add(&jra->jra_committed_link, &d->opd_sync_committed_there);
 533         spin_unlock(&d->opd_sync_lock);
 534
 535         /* XXX: some batching wouldn't hurt */
 536         wake_up(&d->opd_sync_waitq);
 537 }
 538
 539 /**
 540  * RPC interpretation callback.
 541  *
 542  * The callback is called by ptlrpc when RPC is replied. Now we have to decide
 543  * whether we should:
 544  *  - put request on a special list to wait until it's committed by the target,
 545  *    if the request is successful
 546  *  - schedule llog record cancel if no target object is found
 547  *  - try later (essentially after reboot) in case of unexpected error
 548  *
 549  * \param[in] env       LU environment provided by the caller
 550  * \param[in] req       request replied
 551  * \param[in] aa        callback data
 552  * \param[in] rc        result of RPC
 553  *
 554  * \retval 0            always
 555  */
 556 static int osp_sync_interpret(const struct lu_env *env,
 557                               struct ptlrpc_request *req, void *aa, int rc)
 558 {
 559         struct osp_device *d = req->rq_cb_data;
 560         struct osp_job_req_args *jra = aa;
 561
 562         if (jra->jra_magic != OSP_JOB_MAGIC) {
 563                 DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
 564                 LBUG();
 565         }
 566         LASSERT(d);
 567
 568         CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
 569                atomic_read(&req->rq_refcount),
 570                rc, (unsigned) req->rq_transno);
 571         LASSERT(rc || req->rq_transno);
 572
 573         if (rc == -ENOENT) {
 574                 /*
 575                  * we tried to destroy object or update attributes,
 576                  * but object doesn't exist anymore - cancell llog record
 577                  */
 578                 LASSERT(req->rq_transno == 0);
 579                 LASSERT(list_empty(&jra->jra_committed_link));
 580
 581                 ptlrpc_request_addref(req);
 582
 583                 spin_lock(&d->opd_sync_lock);
 584                 list_add(&jra->jra_committed_link,
 585                          &d->opd_sync_committed_there);
 586                 spin_unlock(&d->opd_sync_lock);
 587
 588                 wake_up(&d->opd_sync_waitq);
 589         } else if (rc) {
 590                 struct obd_import *imp = req->rq_import;
 591                 /*
 592                  * error happened, we'll try to repeat on next boot ?
 593                  */
 594                 LASSERTF(req->rq_transno == 0 ||
 595                          req->rq_import_generation < imp->imp_generation,
 596                          "transno %llu, rc %d, gen: req %d, imp %d\n",
 597                          req->rq_transno, rc, req->rq_import_generation,
 598                          imp->imp_generation);
 599                 if (req->rq_transno == 0) {
 600                         /* this is the last time we see the request
 601                          * if transno is not zero, then commit cb
 602                          * will be called at some point */
 603                         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) > 0);
 604                         atomic_dec(&d->opd_sync_rpcs_in_progress);
 605                 }
 606
 607                 wake_up(&d->opd_sync_waitq);
 608         } else if (d->opd_pre != NULL &&
 609                    unlikely(d->opd_pre_status == -ENOSPC)) {
 610                 /*
 611                  * if current status is -ENOSPC (lack of free space on OST)
 612                  * then we should poll OST immediately once object destroy
 613                  * is replied
 614                  */
 615                 osp_statfs_need_now(d);
 616         }
 617
 618         spin_lock(&d->opd_sync_lock);
 619         list_del_init(&jra->jra_in_flight_link);
 620         spin_unlock(&d->opd_sync_lock);
 621         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) > 0);
 622         atomic_dec(&d->opd_sync_rpcs_in_flight);
 623         if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
 624                 wake_up(&d->opd_sync_barrier_waitq);
 625         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 626                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 627                atomic_read(&d->opd_sync_rpcs_in_progress));
 628
 629         osp_sync_check_for_work(d);
 630
 631         return 0;
 632 }
 633
 634 /*
 635  ** Add request to ptlrpc queue.
 636  *
 637  * This is just a tiny helper function to put the request on the sending list
 638  *
 639  * \param[in] d         OSP device
 640  * \param[in] llh       llog handle where the record is stored
 641  * \param[in] h         llog record
 642  * \param[in] req       request
 643  */
 644 static void osp_sync_send_new_rpc(struct osp_device *d,
 645                                   struct llog_handle *llh,
 646                                   struct llog_rec_hdr *h,
 647                                   struct ptlrpc_request *req)
 648 {
 649         struct osp_job_req_args *jra;
 650
 651         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) <=
 652                 d->opd_sync_max_rpcs_in_flight);
 653
 654         jra = ptlrpc_req_async_args(req);
 655         jra->jra_magic = OSP_JOB_MAGIC;
 656         jra->jra_lcookie.lgc_lgl = llh->lgh_id;
 657         jra->jra_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 658         jra->jra_lcookie.lgc_index = h->lrh_index;
 659         INIT_LIST_HEAD(&jra->jra_committed_link);
 660         spin_lock(&d->opd_sync_lock);
 661         list_add_tail(&jra->jra_in_flight_link, &d->opd_sync_in_flight_list);
 662         spin_unlock(&d->opd_sync_lock);
 663
 664         ptlrpcd_add_req(req);
 665 }
 666
 667
 668 /**
 669  * Allocate and prepare RPC for a new change.
 670  *
 671  * The function allocates and initializes an RPC which will be sent soon to
 672  * apply the change to the target OST. The request is initialized from the
 673  * llog record passed. Notice only the fields common to all type of changes
 674  * are initialized.
 675  *
 676  * \param[in] d         OSP device
 677  * \param[in] op        type of the change
 678  * \param[in] format    request format to be used
 679  *
 680  * \retval pointer              new request on success
 681  * \retval ERR_PTR(errno)       on error
 682  */
 683 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
 684                                                ost_cmd_t op,
 685                                                const struct req_format *format)
 686 {
 687         struct ptlrpc_request   *req;
 688         struct obd_import       *imp;
 689         int                      rc;
 690
 691         /* Prepare the request */
 692         imp = d->opd_obd->u.cli.cl_import;
 693         LASSERT(imp);
 694
 695         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
 696                 RETURN(ERR_PTR(-ENOMEM));
 697
 698         req = ptlrpc_request_alloc(imp, format);
 699         if (req == NULL)
 700                 RETURN(ERR_PTR(-ENOMEM));
 701
 702         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
 703         if (rc) {
 704                 ptlrpc_req_finished(req);
 705                 return ERR_PTR(rc);
 706         }
 707
 708         req->rq_interpret_reply = osp_sync_interpret;
 709         req->rq_commit_cb = osp_sync_request_commit_cb;
 710         req->rq_cb_data = d;
 711
 712         ptlrpc_request_set_replen(req);
 713
 714         return req;
 715 }
 716
 717 /**
 718  * Generate a request for setattr change.
 719  *
 720  * The function prepares a new RPC, initializes it with setattr specific
 721  * bits and send the RPC.
 722  *
 723  * \param[in] d         OSP device
 724  * \param[in] llh       llog handle where the record is stored
 725  * \param[in] h         llog record
 726  *
 727  * \retval 0            on success
 728  * \retval 1            on invalid record
 729  * \retval negative     negated errno on error
 730  */
 731 static int osp_sync_new_setattr_job(struct osp_device *d,
 732                                     struct llog_handle *llh,
 733                                     struct llog_rec_hdr *h)
 734 {
 735         struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
 736         struct ptlrpc_request           *req;
 737         struct ost_body                 *body;
 738
 739         ENTRY;
 740         LASSERT(h->lrh_type == MDS_SETATTR64_REC);
 741
 742         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
 743                 RETURN(1);
 744
 745         /* lsr_valid can only be 0 or HAVE OBD_MD_{FLUID, FLGID, FLPROJID} set,
 746          * so no bits other than these should be set. */
 747         if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID |
 748             OBD_MD_FLPROJID)) != 0) {
 749                 CERROR("%s: invalid setattr record, lsr_valid:%llu\n",
 750                         d->opd_obd->obd_name, rec->lsr_valid);
 751                 /* return 1 on invalid record */
 752                 RETURN(1);
 753         }
 754
 755         req = osp_sync_new_job(d, OST_SETATTR, &RQF_OST_SETATTR);
 756         if (IS_ERR(req))
 757                 RETURN(PTR_ERR(req));
 758
 759         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 760         LASSERT(body);
 761         body->oa.o_oi = rec->lsr_oi;
 762         body->oa.o_uid = rec->lsr_uid;
 763         body->oa.o_gid = rec->lsr_gid;
 764         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 765         if (h->lrh_len > sizeof(struct llog_setattr64_rec))
 766                 body->oa.o_projid = ((struct llog_setattr64_rec_v2 *)
 767                                       rec)->lsr_projid;
 768
 769         /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
 770          * we assume that both UID and GID are valid in that case. */
 771         if (rec->lsr_valid == 0)
 772                 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
 773         else
 774                 body->oa.o_valid |= rec->lsr_valid;
 775
 776         osp_sync_send_new_rpc(d, llh, h, req);
 777         RETURN(0);
 778 }
 779
 780 /**
 781  * Generate a request for unlink change.
 782  *
 783  * The function prepares a new RPC, initializes it with unlink(destroy)
 784  * specific bits and sends the RPC. The function is used to handle
 785  * llog_unlink_rec which were used in the older versions of Lustre.
 786  * Current version uses llog_unlink_rec64.
 787  *
 788  * \param[in] d         OSP device
 789  * \param[in] llh       llog handle where the record is stored
 790  * \param[in] h         llog record
 791  *
 792  * \retval 0            on success
 793  * \retval negative     negated errno on error
 794  */
 795 static int osp_sync_new_unlink_job(struct osp_device *d,
 796                                    struct llog_handle *llh,
 797                                    struct llog_rec_hdr *h)
 798 {
 799         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 800         struct ptlrpc_request   *req;
 801         struct ost_body         *body;
 802         int rc;
 803
 804         ENTRY;
 805         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 806
 807         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 808         if (IS_ERR(req))
 809                 RETURN(PTR_ERR(req));
 810
 811         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 812         LASSERT(body);
 813         ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
 814         rc = ostid_set_id(&body->oa.o_oi, rec->lur_oid);
 815         if (rc)
 816                 return rc;
 817         body->oa.o_misc = rec->lur_count;
 818         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 819         if (rec->lur_count)
 820                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 821
 822         osp_sync_send_new_rpc(d, llh, h, req);
 823         RETURN(0);
 824 }
 825
 826 /**
 827  * Generate a request for unlink change.
 828  *
 829  * The function prepares a new RPC, initializes it with unlink(destroy)
 830  * specific bits and sends the RPC. Depending on the target (MDT or OST)
 831  * two different protocols are used. For MDT we use OUT (basically OSD API
 832  * updates transferred via a network). For OST we still use the old
 833  * protocol (OBD?), originally for compatibility. Later we can start to
 834  * use OUT for OST as well, this will allow batching and better code
 835  * unification.
 836  *
 837  * \param[in] d         OSP device
 838  * \param[in] llh       llog handle where the record is stored
 839  * \param[in] h         llog record
 840  *
 841  * \retval 0            on success
 842  * \retval negative     negated errno on error
 843  */
 844 static int osp_sync_new_unlink64_job(struct osp_device *d,
 845                                      struct llog_handle *llh,
 846                                      struct llog_rec_hdr *h)
 847 {
 848         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 849         struct ptlrpc_request           *req = NULL;
 850         struct ost_body                 *body;
 851         int                              rc;
 852
 853         ENTRY;
 854         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 855         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 856         if (IS_ERR(req))
 857                 RETURN(PTR_ERR(req));
 858
 859         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 860         if (body == NULL)
 861                 RETURN(-EFAULT);
 862         rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
 863         if (rc < 0)
 864                 RETURN(rc);
 865         body->oa.o_misc = rec->lur_count;
 866         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 867                            OBD_MD_FLOBJCOUNT;
 868         osp_sync_send_new_rpc(d, llh, h, req);
 869         RETURN(0);
 870 }
 871
 872 /**
 873  * Process llog records.
 874  *
 875  * This function is called to process the llog records committed locally.
 876  * In the recovery model used by OSP we can apply a change to a remote
 877  * target once corresponding transaction (like posix unlink) is committed
 878  * locally so can't revert.
 879  * Depending on the llog record type, a given handler is called that is
 880  * responsible for preparing and sending the RPC to apply the change.
 881  * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
 882  *
 883  * \param[in] env       LU environment provided by the caller
 884  * \param[in] d         OSP device
 885  * \param[in] llh       llog handle where the record is stored
 886  * \param[in] rec       llog record
 887  */
 888 static void osp_sync_process_record(const struct lu_env *env,
 889                                     struct osp_device *d,
 890                                     struct llog_handle *llh,
 891                                     struct llog_rec_hdr *rec)
 892 {
 893         struct llog_handle      *cathandle = llh->u.phd.phd_cat_handle;
 894         struct llog_cookie       cookie;
 895         int                      rc = 0;
 896
 897         ENTRY;
 898
 899         cookie.lgc_lgl = llh->lgh_id;
 900         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 901         cookie.lgc_index = rec->lrh_index;
 902
 903         d->opd_sync_last_catalog_idx = llh->lgh_hdr->llh_cat_idx;
 904
 905         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 906                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 907
 908                 /* we're waiting for the record generated by this instance */
 909                 LASSERT(d->opd_sync_prev_done == 0);
 910                 if (!memcmp(&d->opd_sync_generation, &gen->lgr_gen,
 911                             sizeof(gen->lgr_gen))) {
 912                         CDEBUG(D_HA, "processed all old entries\n");
 913                         d->opd_sync_prev_done = 1;
 914                 }
 915
 916                 /* cancel any generation record */
 917                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 918
 919                 RETURN_EXIT;
 920         }
 921
 922         /*
 923          * now we prepare and fill requests to OST, put them on the queue
 924          * and fire after next commit callback
 925          */
 926
 927         /* notice we increment counters before sending RPC, to be consistent
 928          * in RPC interpret callback which may happen very quickly */
 929         atomic_inc(&d->opd_sync_rpcs_in_flight);
 930         atomic_inc(&d->opd_sync_rpcs_in_progress);
 931
 932         switch (rec->lrh_type) {
 933         /* case MDS_UNLINK_REC is kept for compatibility */
 934         case MDS_UNLINK_REC:
 935                 rc = osp_sync_new_unlink_job(d, llh, rec);
 936                 break;
 937         case MDS_UNLINK64_REC:
 938                 rc = osp_sync_new_unlink64_job(d, llh, rec);
 939                 break;
 940         case MDS_SETATTR64_REC:
 941                 rc = osp_sync_new_setattr_job(d, llh, rec);
 942                 break;
 943         default:
 944                 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
 945                        rec->lrh_type);
 946                 /* treat "unknown record type" as "invalid" */
 947                 rc = 1;
 948                 break;
 949         }
 950
 951         /* For all kinds of records, not matter successful or not,
 952          * we should decrease changes and bump last_processed_id.
 953          */
 954         if (d->opd_sync_prev_done) {
 955                 LASSERT(atomic_read(&d->opd_sync_changes) > 0);
 956                 atomic_dec(&d->opd_sync_changes);
 957                 wake_up(&d->opd_sync_barrier_waitq);
 958         }
 959         atomic64_inc(&d->opd_sync_processed_recs);
 960         if (rc != 0) {
 961                 atomic_dec(&d->opd_sync_rpcs_in_flight);
 962                 atomic_dec(&d->opd_sync_rpcs_in_progress);
 963         }
 964
 965         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 966                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 967                atomic_read(&d->opd_sync_rpcs_in_progress));
 968
 969         /* Delete the invalid record */
 970         if (rc == 1) {
 971                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 972                 if (rc != 0)
 973                         CERROR("%s: can't delete invalid record: "
 974                                "fid = "DFID", rec_id = %u, rc = %d\n",
 975                                d->opd_obd->obd_name,
 976                                PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
 977                                rec->lrh_id, rc);
 978         }
 979
 980         CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
 981                rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);
 982
 983         RETURN_EXIT;
 984 }
 985
 986 /**
 987  * Cancel llog records for the committed changes.
 988  *
 989  * The function walks through the list of the committed RPCs and cancels
 990  * corresponding llog records. see osp_sync_request_commit_cb() for the
 991  * details.
 992  *
 993  * \param[in] env       LU environment provided by the caller
 994  * \param[in] d         OSP device
 995  */
 996 static void osp_sync_process_committed(const struct lu_env *env,
 997                                        struct osp_device *d)
 998 {
 999         struct obd_device       *obd = d->opd_obd;
1000         struct obd_import       *imp = obd->u.cli.cl_import;
1001         struct ost_body         *body;
1002         struct ptlrpc_request   *req;
1003         struct llog_ctxt        *ctxt;
1004         struct llog_handle      *llh;
1005         struct list_head         list;
1006         int                      rc, done = 0;
1007
1008         ENTRY;
1009
1010         if (list_empty(&d->opd_sync_committed_there))
1011                 return;
1012
1013         /*
1014          * if current status is -ENOSPC (lack of free space on OST)
1015          * then we should poll OST immediately once object destroy
1016          * is committed.
1017          * notice: we do this upon commit as well because some backends
1018          * (like DMU) do not release space right away.
1019          */
1020         if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
1021                 osp_statfs_need_now(d);
1022
1023         /*
1024          * now cancel them all
1025          * XXX: can we improve this using some batching?
1026          *      with batch RPC that'll happen automatically?
1027          * XXX: can we store ctxt in lod_device and save few cycles ?
1028          */
1029         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1030         LASSERT(ctxt);
1031
1032         llh = ctxt->loc_handle;
1033         LASSERT(llh);
1034
1035         INIT_LIST_HEAD(&list);
1036         spin_lock(&d->opd_sync_lock);
1037         list_splice(&d->opd_sync_committed_there, &list);
1038         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1039         spin_unlock(&d->opd_sync_lock);
1040
1041         while (!list_empty(&list)) {
1042                 struct osp_job_req_args *jra;
1043
1044                 jra = list_entry(list.next, struct osp_job_req_args,
1045                                  jra_committed_link);
1046                 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1047                 list_del_init(&jra->jra_committed_link);
1048
1049                 req = container_of((void *)jra, struct ptlrpc_request,
1050                                    rq_async_args);
1051                 body = req_capsule_client_get(&req->rq_pill,
1052                                               &RMF_OST_BODY);
1053                 LASSERT(body);
1054                 /* import can be closing, thus all commit cb's are
1055                  * called we can check committness directly */
1056                 if (req->rq_import_generation == imp->imp_generation) {
1057                         rc = llog_cat_cancel_records(env, llh, 1,
1058                                                      &jra->jra_lcookie);
1059                         if (rc)
1060                                 CERROR("%s: can't cancel record: %d\n",
1061                                        obd->obd_name, rc);
1062                 } else {
1063                         DEBUG_REQ(D_OTHER, req, "imp_committed = %llu",
1064                                   imp->imp_peer_committed_transno);
1065                 }
1066                 ptlrpc_req_finished(req);
1067                 done++;
1068         }
1069
1070         llog_ctxt_put(ctxt);
1071
1072         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) >= done);
1073         atomic_sub(done, &d->opd_sync_rpcs_in_progress);
1074         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
1075                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
1076                atomic_read(&d->opd_sync_rpcs_in_progress));
1077
1078         osp_sync_check_for_work(d);
1079
1080         /* wake up the thread if requested to stop:
1081          * it might be waiting for in-progress to complete */
1082         if (unlikely(osp_sync_running(d) == 0))
1083                 wake_up(&d->opd_sync_waitq);
1084
1085         EXIT;
1086 }
1087
1088 /**
1089  * The core of the syncing mechanism.
1090  *
1091  * This is a callback called by the llog processing function. Essentially it
1092  * suspends llog processing until there is a record to process (it's supposed
1093  * to be committed locally). The function handles RPCs committed by the target
1094  * and cancels corresponding llog records.
1095  *
1096  * \param[in] env       LU environment provided by the caller
1097  * \param[in] llh       llog handle we're processing
1098  * \param[in] rec       current llog record
1099  * \param[in] data      callback data containing a pointer to the device
1100  *
1101  * \retval 0                    to ask the caller (llog_process()) to continue
1102  * \retval LLOG_PROC_BREAK      to ask the caller to break
1103  */
1104 static int osp_sync_process_queues(const struct lu_env *env,
1105                                    struct llog_handle *llh,
1106                                    struct llog_rec_hdr *rec,
1107                                    void *data)
1108 {
1109         struct osp_device       *d = data;
1110
1111         do {
1112                 struct l_wait_info lwi = { 0 };
1113
1114                 if (!osp_sync_running(d)) {
1115                         CDEBUG(D_HA, "stop llog processing\n");
1116                         return LLOG_PROC_BREAK;
1117                 }
1118
1119                 /* process requests committed by OST */
1120                 osp_sync_process_committed(env, d);
1121
1122                 /* if we there are changes to be processed and we have
1123                  * resources for this ... do now */
1124                 if (osp_sync_can_process_new(d, rec)) {
1125                         if (llh == NULL) {
1126                                 /* ask llog for another record */
1127                                 CDEBUG(D_HA, "%u changes, %u in progress,"
1128                                      " %u in flight\n",
1129                                      atomic_read(&d->opd_sync_changes),
1130                                      atomic_read(&d->opd_sync_rpcs_in_progress),
1131                                      atomic_read(&d->opd_sync_rpcs_in_flight));
1132                                 return 0;
1133                         }
1134                         osp_sync_process_record(env, d, llh, rec);
1135                         llh = NULL;
1136                         rec = NULL;
1137                 }
1138
1139                 l_wait_event(d->opd_sync_waitq,
1140                              !osp_sync_running(d) ||
1141                              osp_sync_can_process_new(d, rec) ||
1142                              !list_empty(&d->opd_sync_committed_there),
1143                              &lwi);
1144         } while (1);
1145 }
1146
1147 /**
1148  * OSP sync thread.
1149  *
1150  * This thread runs llog_cat_process() scanner calling our callback
1151  * to process llog records. in the callback we implement tricky
1152  * state machine as we don't want to start scanning of the llog again
1153  * and again, also we don't want to process too many records and send
1154  * too many RPCs a time. so, depending on current load (num of changes
1155  * being synced to OST) the callback can suspend awaiting for some
1156  * new conditions, like syncs completed.
1157  *
1158  * In order to process llog records left by previous boots and to allow
1159  * llog_process_thread() to find something (otherwise it'd just exit
1160  * immediately) we add a special GENERATATION record on each boot.
1161  *
1162  * \param[in] _arg      a pointer to thread's arguments
1163  *
1164  * \retval 0            on success
1165  * \retval negative     negated errno on error
1166  */
1167 static int osp_sync_thread(void *_arg)
1168 {
1169         struct osp_device       *d = _arg;
1170         struct ptlrpc_thread    *thread = &d->opd_sync_thread;
1171         struct l_wait_info       lwi = { 0 };
1172         struct llog_ctxt        *ctxt;
1173         struct obd_device       *obd = d->opd_obd;
1174         struct llog_handle      *llh;
1175         struct lu_env            env;
1176         int                      rc, count;
1177         bool                     wrapped;
1178
1179         ENTRY;
1180
1181         rc = lu_env_init(&env, LCT_LOCAL);
1182         if (rc) {
1183                 CERROR("%s: can't initialize env: rc = %d\n",
1184                        obd->obd_name, rc);
1185
1186                 spin_lock(&d->opd_sync_lock);
1187                 thread->t_flags = SVC_STOPPED;
1188                 spin_unlock(&d->opd_sync_lock);
1189                 wake_up(&thread->t_ctl_waitq);
1190
1191                 RETURN(rc);
1192         }
1193
1194         spin_lock(&d->opd_sync_lock);
1195         thread->t_flags = SVC_RUNNING;
1196         spin_unlock(&d->opd_sync_lock);
1197         wake_up(&thread->t_ctl_waitq);
1198
1199         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1200         if (ctxt == NULL) {
1201                 CERROR("can't get appropriate context\n");
1202                 GOTO(out, rc = -EINVAL);
1203         }
1204
1205         llh = ctxt->loc_handle;
1206         if (llh == NULL) {
1207                 CERROR("can't get llh\n");
1208                 llog_ctxt_put(ctxt);
1209                 GOTO(out, rc = -EINVAL);
1210         }
1211
1212         /*
1213          * Catalog processing stops when it processed last catalog record
1214          * with index equal to the end of catalog bitmap. Or if it is wrapped,
1215          * processing stops with index equal to the lgh_last_idx. We need to
1216          * continue processing.
1217          */
1218         d->opd_sync_last_catalog_idx = 0;
1219         do {
1220                 int     size;
1221
1222                 wrapped = (llh->lgh_hdr->llh_cat_idx >= llh->lgh_last_idx &&
1223                            llh->lgh_hdr->llh_count > 1);
1224
1225                 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d,
1226                                       d->opd_sync_last_catalog_idx, 0);
1227
1228                 size = OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ?
1229                        cfs_fail_val : (LLOG_HDR_BITMAP_SIZE(llh->lgh_hdr) - 1);
1230                 /* processing reaches catalog bottom */
1231                 if (d->opd_sync_last_catalog_idx == size)
1232                         d->opd_sync_last_catalog_idx = 0;
1233                 else if (wrapped)
1234                         /* If catalog is wrapped we can`t predict last index of
1235                          * processing because lgh_last_idx could be changed.
1236                          * Starting form the next one */
1237                         d->opd_sync_last_catalog_idx++;
1238
1239         } while (rc == 0 && (wrapped || d->opd_sync_last_catalog_idx == 0));
1240
1241         if (rc < 0) {
1242                 CERROR("%s: llog process with osp_sync_process_queues "
1243                        "failed: %d\n", d->opd_obd->obd_name, rc);
1244                 GOTO(close, rc);
1245         }
1246         LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1247                  "%u changes, %u in progress, %u in flight: %d\n",
1248                  atomic_read(&d->opd_sync_changes),
1249                  atomic_read(&d->opd_sync_rpcs_in_progress),
1250                  atomic_read(&d->opd_sync_rpcs_in_flight), rc);
1251
1252         /* we don't expect llog_process_thread() to exit till umount */
1253         LASSERTF(thread->t_flags != SVC_RUNNING,
1254                  "%u changes, %u in progress, %u in flight\n",
1255                  atomic_read(&d->opd_sync_changes),
1256                  atomic_read(&d->opd_sync_rpcs_in_progress),
1257                  atomic_read(&d->opd_sync_rpcs_in_flight));
1258
1259         /* wait till all the requests are completed */
1260         count = 0;
1261         while (atomic_read(&d->opd_sync_rpcs_in_progress) > 0) {
1262                 osp_sync_process_committed(&env, d);
1263
1264                 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1265                 rc = l_wait_event(d->opd_sync_waitq,
1266                                 atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1267                                   &lwi);
1268                 if (rc == -ETIMEDOUT)
1269                         count++;
1270                 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1271                          d->opd_obd->obd_name,
1272                          atomic_read(&d->opd_sync_rpcs_in_progress),
1273                          atomic_read(&d->opd_sync_rpcs_in_flight),
1274                          list_empty(&d->opd_sync_committed_there) ? "" : "!");
1275
1276         }
1277
1278 close:
1279         llog_cat_close(&env, llh);
1280         rc = llog_cleanup(&env, ctxt);
1281         if (rc)
1282                 CERROR("can't cleanup llog: %d\n", rc);
1283 out:
1284         LASSERTF(atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1285                  "%s: %d %d %sempty\n", d->opd_obd->obd_name,
1286                  atomic_read(&d->opd_sync_rpcs_in_progress),
1287                  atomic_read(&d->opd_sync_rpcs_in_flight),
1288                  list_empty(&d->opd_sync_committed_there) ? "" : "!");
1289
1290         thread->t_flags = SVC_STOPPED;
1291
1292         wake_up(&thread->t_ctl_waitq);
1293
1294         lu_env_fini(&env);
1295
1296         RETURN(0);
1297 }
1298
1299 /**
1300  * Initialize llog.
1301  *
1302  * Initializes the llog. Specific llog to be used depends on the type of the
1303  * target OSP represents (OST or MDT). The function adds appends a new llog
1304  * record to mark the place where the records associated with this boot
1305  * start.
1306  *
1307  * \param[in] env       LU environment provided by the caller
1308  * \param[in] d         OSP device
1309  *
1310  * \retval 0            on success
1311  * \retval negative     negated errno on error
1312  */
1313 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1314 {
1315         struct osp_thread_info  *osi = osp_env_info(env);
1316         struct lu_fid           *fid = &osi->osi_fid;
1317         struct llog_handle      *lgh = NULL;
1318         struct obd_device       *obd = d->opd_obd;
1319         struct llog_ctxt        *ctxt;
1320         int                     rc;
1321
1322         ENTRY;
1323
1324         LASSERT(obd);
1325
1326         /*
1327          * open llog corresponding to our OST
1328          */
1329         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1330         obd->obd_lvfs_ctxt.dt = d->opd_storage;
1331
1332         lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1333
1334         rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1335                                    &osi->osi_cid, fid);
1336         if (rc < 0) {
1337                 if (rc != -EFAULT) {
1338                         CERROR("%s: can't get id from catalogs: rc = %d\n",
1339                                obd->obd_name, rc);
1340                         RETURN(rc);
1341                 }
1342
1343                 /* After sparse OST indices is supported, the CATALOG file
1344                  * may become a sparse file that results in failure on
1345                  * reading. Skip this error as the llog will be created
1346                  * later */
1347                 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1348                 rc = 0;
1349         }
1350
1351         CDEBUG(D_INFO, "%s: Init llog for %d - catid "DFID":%x\n",
1352                obd->obd_name, d->opd_index,
1353                PFID(&osi->osi_cid.lci_logid.lgl_oi.oi_fid),
1354                osi->osi_cid.lci_logid.lgl_ogen);
1355
1356         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT,
1357                         d->opd_storage->dd_lu_dev.ld_obd,
1358                         &osp_mds_ost_orig_logops);
1359         if (rc)
1360                 RETURN(rc);
1361
1362         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1363         LASSERT(ctxt);
1364
1365         if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1366                 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1367                                LLOG_OPEN_EXISTS);
1368                 /* re-create llog if it is missing */
1369                 if (rc == -ENOENT)
1370                         logid_set_id(&osi->osi_cid.lci_logid, 0);
1371                 else if (rc < 0)
1372                         GOTO(out_cleanup, rc);
1373         }
1374
1375         if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1376                 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1377                 if (rc < 0)
1378                         GOTO(out_cleanup, rc);
1379                 osi->osi_cid.lci_logid = lgh->lgh_id;
1380         }
1381
1382         LASSERT(lgh != NULL);
1383         ctxt->loc_handle = lgh;
1384
1385         rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
1386         if (rc)
1387                 GOTO(out_close, rc);
1388
1389         rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1390                                    &osi->osi_cid, fid);
1391         if (rc)
1392                 GOTO(out_close, rc);
1393
1394         /*
1395          * put a mark in the llog till which we'll be processing
1396          * old records restless
1397          */
1398         d->opd_sync_generation.mnt_cnt = cfs_time_current();
1399         d->opd_sync_generation.conn_cnt = cfs_time_current();
1400
1401         osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1402         osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1403
1404         memcpy(&osi->osi_gen.lgr_gen, &d->opd_sync_generation,
1405                sizeof(osi->osi_gen.lgr_gen));
1406
1407         rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1408         if (rc < 0)
1409                 GOTO(out_close, rc);
1410         llog_ctxt_put(ctxt);
1411         RETURN(0);
1412 out_close:
1413         llog_cat_close(env, lgh);
1414 out_cleanup:
1415         llog_cleanup(env, ctxt);
1416         RETURN(rc);
1417 }
1418
1419 /**
1420  * Cleanup llog used for syncing.
1421  *
1422  * Closes and cleanups the llog. The function is called when the device is
1423  * shutting down.
1424  *
1425  * \param[in] env       LU environment provided by the caller
1426  * \param[in] d         OSP device
1427  */
1428 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1429 {
1430         struct llog_ctxt *ctxt;
1431
1432         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1433         if (ctxt) {
1434                 llog_cat_close(env, ctxt->loc_handle);
1435                 llog_cleanup(env, ctxt);
1436         }
1437 }
1438
1439 /**
1440  * Initialization of the sync component of OSP.
1441  *
1442  * Initializes the llog and starts a new thread to handle the changes to
1443  * the remote target (OST or MDT).
1444  *
1445  * \param[in] env       LU environment provided by the caller
1446  * \param[in] d         OSP device
1447  *
1448  * \retval 0            on success
1449  * \retval negative     negated errno on error
1450  */
1451 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1452 {
1453         struct l_wait_info       lwi = { 0 };
1454         struct task_struct      *task;
1455         int                      rc;
1456
1457         ENTRY;
1458
1459         d->opd_sync_max_rpcs_in_flight = OSP_MAX_RPCS_IN_FLIGHT;
1460         d->opd_sync_max_rpcs_in_progress = OSP_MAX_RPCS_IN_PROGRESS;
1461         spin_lock_init(&d->opd_sync_lock);
1462         init_waitqueue_head(&d->opd_sync_waitq);
1463         init_waitqueue_head(&d->opd_sync_barrier_waitq);
1464         thread_set_flags(&d->opd_sync_thread, SVC_INIT);
1465         init_waitqueue_head(&d->opd_sync_thread.t_ctl_waitq);
1466         INIT_LIST_HEAD(&d->opd_sync_in_flight_list);
1467         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1468
1469         if (d->opd_storage->dd_rdonly)
1470                 RETURN(0);
1471
1472         /*
1473          * initialize llog storing changes
1474          */
1475         rc = osp_sync_llog_init(env, d);
1476         if (rc) {
1477                 CERROR("%s: can't initialize llog: rc = %d\n",
1478                        d->opd_obd->obd_name, rc);
1479                 GOTO(err_id, rc);
1480         }
1481
1482         /*
1483          * Start synchronization thread
1484          */
1485         task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1486                            d->opd_index, d->opd_group);
1487         if (IS_ERR(task)) {
1488                 rc = PTR_ERR(task);
1489                 CERROR("%s: cannot start sync thread: rc = %d\n",
1490                        d->opd_obd->obd_name, rc);
1491                 GOTO(err_llog, rc);
1492         }
1493
1494         l_wait_event(d->opd_sync_thread.t_ctl_waitq,
1495                      osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1496
1497         RETURN(0);
1498 err_llog:
1499         osp_sync_llog_fini(env, d);
1500 err_id:
1501         return rc;
1502 }
1503
1504 /**
1505  * Stop the syncing thread.
1506  *
1507  * Asks the syncing thread to stop and wait until it's stopped.
1508  *
1509  * \param[in] d         OSP device
1510  *
1511  * \retval              0
1512  */
1513 int osp_sync_fini(struct osp_device *d)
1514 {
1515         struct ptlrpc_thread *thread = &d->opd_sync_thread;
1516
1517         ENTRY;
1518
1519         if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1520                 thread->t_flags = SVC_STOPPING;
1521                 wake_up(&d->opd_sync_waitq);
1522                 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1523         }
1524
1525         RETURN(0);
1526 }
1527
1528 struct osp_last_committed_cb {
1529         struct dt_txn_commit_cb  ospc_cb;
1530         struct osp_device       *ospc_dev;
1531         __u64                    ospc_transno;
1532 };
1533
1534 void osp_sync_local_commit_cb(struct lu_env *env, struct thandle *th,
1535                               struct dt_txn_commit_cb *dcb, int err)
1536 {
1537         struct osp_last_committed_cb    *cb;
1538         struct osp_device               *d;
1539
1540         cb = container_of0(dcb, struct osp_last_committed_cb, ospc_cb);
1541         d = cb->ospc_dev;
1542
1543         CDEBUG(D_HA, "%s: %llu committed\n", d->opd_obd->obd_name,
1544                cb->ospc_transno);
1545
1546         spin_lock(&d->opd_sync_lock);
1547         if (cb->ospc_transno > d->opd_sync_last_committed_id)
1548                 d->opd_sync_last_committed_id = cb->ospc_transno;
1549         spin_unlock(&d->opd_sync_lock);
1550
1551         osp_sync_check_for_work(d);
1552         lu_device_put(osp2lu_dev(d));
1553         if (atomic_dec_and_test(&d->opd_commits_registered))
1554                 wake_up(&d->opd_sync_waitq);
1555
1556         OBD_FREE_PTR(cb);
1557 }
1558
1559 int osp_sync_add_commit_cb(const struct lu_env *env, struct osp_device *d,
1560                            struct thandle *th)
1561 {
1562         struct osp_last_committed_cb    *cb;
1563         struct dt_txn_commit_cb         *dcb;
1564         int                              rc = 0;
1565
1566         OBD_ALLOC_PTR(cb);
1567         if (cb == NULL)
1568                 return -ENOMEM;
1569         cb->ospc_dev = d;
1570         dcb = &cb->ospc_cb;
1571         dcb->dcb_func = osp_sync_local_commit_cb;
1572         spin_lock(&d->opd_sync_lock);
1573         cb->ospc_transno = ++d->opd_sync_last_used_id;
1574         spin_unlock(&d->opd_sync_lock);
1575
1576         rc = dt_trans_cb_add(th, dcb);
1577         CDEBUG(D_HA, "%s: add commit cb at %llu, next at %llu, rc = %d\n",
1578                d->opd_obd->obd_name, (unsigned long long) cfs_time_current(),
1579                (unsigned long long) d->opd_sync_next_commit_cb, rc);
1580
1581         if (likely(rc == 0)) {
1582                 lu_device_get(osp2lu_dev(d));
1583                 atomic_inc(&d->opd_commits_registered);
1584         } else
1585                 OBD_FREE_PTR(cb);
1586
1587         return rc;
1588 }
1589
1590 /*
1591  * generate an empty transaction and hook the commit callback in
1592  * then force transaction commit
1593  */
1594 void osp_sync_force(const struct lu_env *env, struct osp_device *d)
1595 {
1596         struct thandle *th;
1597         int rc;
1598
1599         th = dt_trans_create(env, d->opd_storage);
1600         if (IS_ERR(th)) {
1601                 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
1602                 return;
1603         }
1604         rc = dt_trans_start_local(env, d->opd_storage, th);
1605         if (rc == 0) {
1606                 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
1607                        d->opd_obd->obd_name,
1608                        atomic_read(&d->opd_sync_changes));
1609                 rc = osp_sync_add_commit_cb(env, d, th);
1610                 dt_trans_stop(env, d->opd_storage, th);
1611         }
1612
1613         dt_commit_async(env, d->opd_storage);
1614 }