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
 572         if (rc == -ENOENT) {
 573                 /*
 574                  * we tried to destroy object or update attributes,
 575                  * but object doesn't exist anymore - cancell llog record
 576                  */
 577                 LASSERT(req->rq_transno == 0);
 578                 LASSERT(list_empty(&jra->jra_committed_link));
 579
 580                 ptlrpc_request_addref(req);
 581
 582                 spin_lock(&d->opd_sync_lock);
 583                 list_add(&jra->jra_committed_link,
 584                          &d->opd_sync_committed_there);
 585                 spin_unlock(&d->opd_sync_lock);
 586
 587                 wake_up(&d->opd_sync_waitq);
 588         } else if (rc) {
 589                 struct obd_import *imp = req->rq_import;
 590                 /*
 591                  * error happened, we'll try to repeat on next boot ?
 592                  */
 593                 LASSERTF(req->rq_transno == 0 || rc == -EIO ||
 594                          req->rq_import_generation < imp->imp_generation,
 595                          "transno %llu, rc %d, gen: req %d, imp %d\n",
 596                          req->rq_transno, rc, req->rq_import_generation,
 597                          imp->imp_generation);
 598                 if (req->rq_transno == 0) {
 599                         /* this is the last time we see the request
 600                          * if transno is not zero, then commit cb
 601                          * will be called at some point */
 602                         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) > 0);
 603                         atomic_dec(&d->opd_sync_rpcs_in_progress);
 604                 }
 605
 606                 wake_up(&d->opd_sync_waitq);
 607         } else if (d->opd_pre != NULL &&
 608                    unlikely(d->opd_pre_status == -ENOSPC)) {
 609                 /*
 610                  * if current status is -ENOSPC (lack of free space on OST)
 611                  * then we should poll OST immediately once object destroy
 612                  * is replied
 613                  */
 614                 osp_statfs_need_now(d);
 615         }
 616
 617         spin_lock(&d->opd_sync_lock);
 618         list_del_init(&jra->jra_in_flight_link);
 619         spin_unlock(&d->opd_sync_lock);
 620         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) > 0);
 621         atomic_dec(&d->opd_sync_rpcs_in_flight);
 622         if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
 623                 wake_up(&d->opd_sync_barrier_waitq);
 624         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 625                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 626                atomic_read(&d->opd_sync_rpcs_in_progress));
 627
 628         osp_sync_check_for_work(d);
 629
 630         return 0;
 631 }
 632
 633 /*
 634  ** Add request to ptlrpc queue.
 635  *
 636  * This is just a tiny helper function to put the request on the sending list
 637  *
 638  * \param[in] d         OSP device
 639  * \param[in] llh       llog handle where the record is stored
 640  * \param[in] h         llog record
 641  * \param[in] req       request
 642  */
 643 static void osp_sync_send_new_rpc(struct osp_device *d,
 644                                   struct llog_handle *llh,
 645                                   struct llog_rec_hdr *h,
 646                                   struct ptlrpc_request *req)
 647 {
 648         struct osp_job_req_args *jra;
 649
 650         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) <=
 651                 d->opd_sync_max_rpcs_in_flight);
 652
 653         jra = ptlrpc_req_async_args(req);
 654         jra->jra_magic = OSP_JOB_MAGIC;
 655         jra->jra_lcookie.lgc_lgl = llh->lgh_id;
 656         jra->jra_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 657         jra->jra_lcookie.lgc_index = h->lrh_index;
 658         INIT_LIST_HEAD(&jra->jra_committed_link);
 659         spin_lock(&d->opd_sync_lock);
 660         list_add_tail(&jra->jra_in_flight_link, &d->opd_sync_in_flight_list);
 661         spin_unlock(&d->opd_sync_lock);
 662
 663         ptlrpcd_add_req(req);
 664 }
 665
 666
 667 /**
 668  * Allocate and prepare RPC for a new change.
 669  *
 670  * The function allocates and initializes an RPC which will be sent soon to
 671  * apply the change to the target OST. The request is initialized from the
 672  * llog record passed. Notice only the fields common to all type of changes
 673  * are initialized.
 674  *
 675  * \param[in] d         OSP device
 676  * \param[in] op        type of the change
 677  * \param[in] format    request format to be used
 678  *
 679  * \retval pointer              new request on success
 680  * \retval ERR_PTR(errno)       on error
 681  */
 682 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
 683                                                ost_cmd_t op,
 684                                                const struct req_format *format)
 685 {
 686         struct ptlrpc_request   *req;
 687         struct obd_import       *imp;
 688         int                      rc;
 689
 690         /* Prepare the request */
 691         imp = d->opd_obd->u.cli.cl_import;
 692         LASSERT(imp);
 693
 694         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
 695                 RETURN(ERR_PTR(-ENOMEM));
 696
 697         req = ptlrpc_request_alloc(imp, format);
 698         if (req == NULL)
 699                 RETURN(ERR_PTR(-ENOMEM));
 700
 701         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
 702         if (rc) {
 703                 ptlrpc_req_finished(req);
 704                 return ERR_PTR(rc);
 705         }
 706
 707         req->rq_interpret_reply = osp_sync_interpret;
 708         req->rq_commit_cb = osp_sync_request_commit_cb;
 709         req->rq_cb_data = d;
 710
 711         ptlrpc_request_set_replen(req);
 712
 713         return req;
 714 }
 715
 716 /**
 717  * Generate a request for setattr change.
 718  *
 719  * The function prepares a new RPC, initializes it with setattr specific
 720  * bits and send the RPC.
 721  *
 722  * \param[in] d         OSP device
 723  * \param[in] llh       llog handle where the record is stored
 724  * \param[in] h         llog record
 725  *
 726  * \retval 0            on success
 727  * \retval 1            on invalid record
 728  * \retval negative     negated errno on error
 729  */
 730 static int osp_sync_new_setattr_job(struct osp_device *d,
 731                                     struct llog_handle *llh,
 732                                     struct llog_rec_hdr *h)
 733 {
 734         struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
 735         struct ptlrpc_request           *req;
 736         struct ost_body                 *body;
 737
 738         ENTRY;
 739         LASSERT(h->lrh_type == MDS_SETATTR64_REC);
 740
 741         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
 742                 RETURN(1);
 743
 744         /* lsr_valid can only be 0 or HAVE OBD_MD_{FLUID, FLGID, FLPROJID} set,
 745          * so no bits other than these should be set. */
 746         if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID |
 747             OBD_MD_FLPROJID)) != 0) {
 748                 CERROR("%s: invalid setattr record, lsr_valid:%llu\n",
 749                         d->opd_obd->obd_name, rec->lsr_valid);
 750                 /* return 1 on invalid record */
 751                 RETURN(1);
 752         }
 753
 754         req = osp_sync_new_job(d, OST_SETATTR, &RQF_OST_SETATTR);
 755         if (IS_ERR(req))
 756                 RETURN(PTR_ERR(req));
 757
 758         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 759         LASSERT(body);
 760         body->oa.o_oi = rec->lsr_oi;
 761         body->oa.o_uid = rec->lsr_uid;
 762         body->oa.o_gid = rec->lsr_gid;
 763         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 764         if (h->lrh_len > sizeof(struct llog_setattr64_rec))
 765                 body->oa.o_projid = ((struct llog_setattr64_rec_v2 *)
 766                                       rec)->lsr_projid;
 767
 768         /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
 769          * we assume that both UID and GID are valid in that case. */
 770         if (rec->lsr_valid == 0)
 771                 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
 772         else
 773                 body->oa.o_valid |= rec->lsr_valid;
 774
 775         osp_sync_send_new_rpc(d, llh, h, req);
 776         RETURN(0);
 777 }
 778
 779 /**
 780  * Generate a request for unlink change.
 781  *
 782  * The function prepares a new RPC, initializes it with unlink(destroy)
 783  * specific bits and sends the RPC. The function is used to handle
 784  * llog_unlink_rec which were used in the older versions of Lustre.
 785  * Current version uses llog_unlink_rec64.
 786  *
 787  * \param[in] d         OSP device
 788  * \param[in] llh       llog handle where the record is stored
 789  * \param[in] h         llog record
 790  *
 791  * \retval 0            on success
 792  * \retval negative     negated errno on error
 793  */
 794 static int osp_sync_new_unlink_job(struct osp_device *d,
 795                                    struct llog_handle *llh,
 796                                    struct llog_rec_hdr *h)
 797 {
 798         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 799         struct ptlrpc_request   *req;
 800         struct ost_body         *body;
 801         int rc;
 802
 803         ENTRY;
 804         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 805
 806         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 807         if (IS_ERR(req))
 808                 RETURN(PTR_ERR(req));
 809
 810         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 811         LASSERT(body);
 812         ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
 813         rc = ostid_set_id(&body->oa.o_oi, rec->lur_oid);
 814         if (rc)
 815                 return rc;
 816         body->oa.o_misc = rec->lur_count;
 817         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 818         if (rec->lur_count)
 819                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 820
 821         osp_sync_send_new_rpc(d, llh, h, req);
 822         RETURN(0);
 823 }
 824
 825 /**
 826  * Generate a request for unlink change.
 827  *
 828  * The function prepares a new RPC, initializes it with unlink(destroy)
 829  * specific bits and sends the RPC. Depending on the target (MDT or OST)
 830  * two different protocols are used. For MDT we use OUT (basically OSD API
 831  * updates transferred via a network). For OST we still use the old
 832  * protocol (OBD?), originally for compatibility. Later we can start to
 833  * use OUT for OST as well, this will allow batching and better code
 834  * unification.
 835  *
 836  * \param[in] d         OSP device
 837  * \param[in] llh       llog handle where the record is stored
 838  * \param[in] h         llog record
 839  *
 840  * \retval 0            on success
 841  * \retval negative     negated errno on error
 842  */
 843 static int osp_sync_new_unlink64_job(struct osp_device *d,
 844                                      struct llog_handle *llh,
 845                                      struct llog_rec_hdr *h)
 846 {
 847         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 848         struct ptlrpc_request           *req = NULL;
 849         struct ost_body                 *body;
 850         int                              rc;
 851
 852         ENTRY;
 853         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 854         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 855         if (IS_ERR(req))
 856                 RETURN(PTR_ERR(req));
 857
 858         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 859         if (body == NULL)
 860                 RETURN(-EFAULT);
 861         rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
 862         if (rc < 0)
 863                 RETURN(rc);
 864         body->oa.o_misc = rec->lur_count;
 865         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 866                            OBD_MD_FLOBJCOUNT;
 867         osp_sync_send_new_rpc(d, llh, h, req);
 868         RETURN(0);
 869 }
 870
 871 /**
 872  * Process llog records.
 873  *
 874  * This function is called to process the llog records committed locally.
 875  * In the recovery model used by OSP we can apply a change to a remote
 876  * target once corresponding transaction (like posix unlink) is committed
 877  * locally so can't revert.
 878  * Depending on the llog record type, a given handler is called that is
 879  * responsible for preparing and sending the RPC to apply the change.
 880  * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
 881  *
 882  * \param[in] env       LU environment provided by the caller
 883  * \param[in] d         OSP device
 884  * \param[in] llh       llog handle where the record is stored
 885  * \param[in] rec       llog record
 886  */
 887 static void osp_sync_process_record(const struct lu_env *env,
 888                                     struct osp_device *d,
 889                                     struct llog_handle *llh,
 890                                     struct llog_rec_hdr *rec)
 891 {
 892         struct llog_handle      *cathandle = llh->u.phd.phd_cat_handle;
 893         struct llog_cookie       cookie;
 894         int                      rc = 0;
 895
 896         ENTRY;
 897
 898         cookie.lgc_lgl = llh->lgh_id;
 899         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 900         cookie.lgc_index = rec->lrh_index;
 901
 902         d->opd_sync_last_catalog_idx = llh->lgh_hdr->llh_cat_idx;
 903
 904         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 905                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 906
 907                 /* we're waiting for the record generated by this instance */
 908                 LASSERT(d->opd_sync_prev_done == 0);
 909                 if (!memcmp(&d->opd_sync_generation, &gen->lgr_gen,
 910                             sizeof(gen->lgr_gen))) {
 911                         CDEBUG(D_HA, "processed all old entries\n");
 912                         d->opd_sync_prev_done = 1;
 913                 }
 914
 915                 /* cancel any generation record */
 916                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 917
 918                 RETURN_EXIT;
 919         }
 920
 921         /*
 922          * now we prepare and fill requests to OST, put them on the queue
 923          * and fire after next commit callback
 924          */
 925
 926         /* notice we increment counters before sending RPC, to be consistent
 927          * in RPC interpret callback which may happen very quickly */
 928         atomic_inc(&d->opd_sync_rpcs_in_flight);
 929         atomic_inc(&d->opd_sync_rpcs_in_progress);
 930
 931         switch (rec->lrh_type) {
 932         /* case MDS_UNLINK_REC is kept for compatibility */
 933         case MDS_UNLINK_REC:
 934                 rc = osp_sync_new_unlink_job(d, llh, rec);
 935                 break;
 936         case MDS_UNLINK64_REC:
 937                 rc = osp_sync_new_unlink64_job(d, llh, rec);
 938                 break;
 939         case MDS_SETATTR64_REC:
 940                 rc = osp_sync_new_setattr_job(d, llh, rec);
 941                 break;
 942         default:
 943                 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
 944                        rec->lrh_type);
 945                 /* treat "unknown record type" as "invalid" */
 946                 rc = 1;
 947                 break;
 948         }
 949
 950         /* For all kinds of records, not matter successful or not,
 951          * we should decrease changes and bump last_processed_id.
 952          */
 953         if (d->opd_sync_prev_done) {
 954                 LASSERT(atomic_read(&d->opd_sync_changes) > 0);
 955                 atomic_dec(&d->opd_sync_changes);
 956                 wake_up(&d->opd_sync_barrier_waitq);
 957         }
 958         atomic64_inc(&d->opd_sync_processed_recs);
 959         if (rc != 0) {
 960                 atomic_dec(&d->opd_sync_rpcs_in_flight);
 961                 atomic_dec(&d->opd_sync_rpcs_in_progress);
 962         }
 963
 964         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 965                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 966                atomic_read(&d->opd_sync_rpcs_in_progress));
 967
 968         /* Delete the invalid record */
 969         if (rc == 1) {
 970                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 971                 if (rc != 0)
 972                         CERROR("%s: can't delete invalid record: "
 973                                "fid = "DFID", rec_id = %u, rc = %d\n",
 974                                d->opd_obd->obd_name,
 975                                PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
 976                                rec->lrh_id, rc);
 977         }
 978
 979         CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
 980                rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);
 981
 982         RETURN_EXIT;
 983 }
 984
 985 /**
 986  * Cancel llog records for the committed changes.
 987  *
 988  * The function walks through the list of the committed RPCs and cancels
 989  * corresponding llog records. see osp_sync_request_commit_cb() for the
 990  * details.
 991  *
 992  * \param[in] env       LU environment provided by the caller
 993  * \param[in] d         OSP device
 994  */
 995 static void osp_sync_process_committed(const struct lu_env *env,
 996                                        struct osp_device *d)
 997 {
 998         struct obd_device       *obd = d->opd_obd;
 999         struct obd_import       *imp = obd->u.cli.cl_import;
1000         struct ost_body         *body;
1001         struct ptlrpc_request   *req;
1002         struct llog_ctxt        *ctxt;
1003         struct llog_handle      *llh;
1004         struct list_head         list;
1005         int                      rc, done = 0;
1006
1007         ENTRY;
1008
1009         if (list_empty(&d->opd_sync_committed_there))
1010                 return;
1011
1012         /*
1013          * if current status is -ENOSPC (lack of free space on OST)
1014          * then we should poll OST immediately once object destroy
1015          * is committed.
1016          * notice: we do this upon commit as well because some backends
1017          * (like DMU) do not release space right away.
1018          */
1019         if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
1020                 osp_statfs_need_now(d);
1021
1022         /*
1023          * now cancel them all
1024          * XXX: can we improve this using some batching?
1025          *      with batch RPC that'll happen automatically?
1026          * XXX: can we store ctxt in lod_device and save few cycles ?
1027          */
1028         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1029         LASSERT(ctxt);
1030
1031         llh = ctxt->loc_handle;
1032         LASSERT(llh);
1033
1034         INIT_LIST_HEAD(&list);
1035         spin_lock(&d->opd_sync_lock);
1036         list_splice(&d->opd_sync_committed_there, &list);
1037         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1038         spin_unlock(&d->opd_sync_lock);
1039
1040         while (!list_empty(&list)) {
1041                 struct osp_job_req_args *jra;
1042
1043                 jra = list_entry(list.next, struct osp_job_req_args,
1044                                  jra_committed_link);
1045                 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1046                 list_del_init(&jra->jra_committed_link);
1047
1048                 req = container_of((void *)jra, struct ptlrpc_request,
1049                                    rq_async_args);
1050                 body = req_capsule_client_get(&req->rq_pill,
1051                                               &RMF_OST_BODY);
1052                 LASSERT(body);
1053                 /* import can be closing, thus all commit cb's are
1054                  * called we can check committness directly */
1055                 if (req->rq_import_generation == imp->imp_generation) {
1056                         rc = llog_cat_cancel_records(env, llh, 1,
1057                                                      &jra->jra_lcookie);
1058                         if (rc)
1059                                 CERROR("%s: can't cancel record: %d\n",
1060                                        obd->obd_name, rc);
1061                 } else {
1062                         DEBUG_REQ(D_OTHER, req, "imp_committed = %llu",
1063                                   imp->imp_peer_committed_transno);
1064                 }
1065                 ptlrpc_req_finished(req);
1066                 done++;
1067         }
1068
1069         llog_ctxt_put(ctxt);
1070
1071         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) >= done);
1072         atomic_sub(done, &d->opd_sync_rpcs_in_progress);
1073         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
1074                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
1075                atomic_read(&d->opd_sync_rpcs_in_progress));
1076
1077         osp_sync_check_for_work(d);
1078
1079         /* wake up the thread if requested to stop:
1080          * it might be waiting for in-progress to complete */
1081         if (unlikely(osp_sync_running(d) == 0))
1082                 wake_up(&d->opd_sync_waitq);
1083
1084         EXIT;
1085 }
1086
1087 /**
1088  * The core of the syncing mechanism.
1089  *
1090  * This is a callback called by the llog processing function. Essentially it
1091  * suspends llog processing until there is a record to process (it's supposed
1092  * to be committed locally). The function handles RPCs committed by the target
1093  * and cancels corresponding llog records.
1094  *
1095  * \param[in] env       LU environment provided by the caller
1096  * \param[in] llh       llog handle we're processing
1097  * \param[in] rec       current llog record
1098  * \param[in] data      callback data containing a pointer to the device
1099  *
1100  * \retval 0                    to ask the caller (llog_process()) to continue
1101  * \retval LLOG_PROC_BREAK      to ask the caller to break
1102  */
1103 static int osp_sync_process_queues(const struct lu_env *env,
1104                                    struct llog_handle *llh,
1105                                    struct llog_rec_hdr *rec,
1106                                    void *data)
1107 {
1108         struct osp_device       *d = data;
1109
1110         do {
1111                 struct l_wait_info lwi = { 0 };
1112
1113                 if (!osp_sync_running(d)) {
1114                         CDEBUG(D_HA, "stop llog processing\n");
1115                         return LLOG_PROC_BREAK;
1116                 }
1117
1118                 /* process requests committed by OST */
1119                 osp_sync_process_committed(env, d);
1120
1121                 /* if we there are changes to be processed and we have
1122                  * resources for this ... do now */
1123                 if (osp_sync_can_process_new(d, rec)) {
1124                         if (llh == NULL) {
1125                                 /* ask llog for another record */
1126                                 CDEBUG(D_HA, "%u changes, %u in progress,"
1127                                      " %u in flight\n",
1128                                      atomic_read(&d->opd_sync_changes),
1129                                      atomic_read(&d->opd_sync_rpcs_in_progress),
1130                                      atomic_read(&d->opd_sync_rpcs_in_flight));
1131                                 return 0;
1132                         }
1133                         osp_sync_process_record(env, d, llh, rec);
1134                         llh = NULL;
1135                         rec = NULL;
1136                 }
1137
1138                 l_wait_event(d->opd_sync_waitq,
1139                              !osp_sync_running(d) ||
1140                              osp_sync_can_process_new(d, rec) ||
1141                              !list_empty(&d->opd_sync_committed_there),
1142                              &lwi);
1143         } while (1);
1144 }
1145
1146 /**
1147  * OSP sync thread.
1148  *
1149  * This thread runs llog_cat_process() scanner calling our callback
1150  * to process llog records. in the callback we implement tricky
1151  * state machine as we don't want to start scanning of the llog again
1152  * and again, also we don't want to process too many records and send
1153  * too many RPCs a time. so, depending on current load (num of changes
1154  * being synced to OST) the callback can suspend awaiting for some
1155  * new conditions, like syncs completed.
1156  *
1157  * In order to process llog records left by previous boots and to allow
1158  * llog_process_thread() to find something (otherwise it'd just exit
1159  * immediately) we add a special GENERATATION record on each boot.
1160  *
1161  * \param[in] _arg      a pointer to thread's arguments
1162  *
1163  * \retval 0            on success
1164  * \retval negative     negated errno on error
1165  */
1166 static int osp_sync_thread(void *_arg)
1167 {
1168         struct osp_device       *d = _arg;
1169         struct ptlrpc_thread    *thread = &d->opd_sync_thread;
1170         struct l_wait_info       lwi = { 0 };
1171         struct llog_ctxt        *ctxt;
1172         struct obd_device       *obd = d->opd_obd;
1173         struct llog_handle      *llh;
1174         struct lu_env            env;
1175         int                      rc, count;
1176         bool                     wrapped;
1177
1178         ENTRY;
1179
1180         rc = lu_env_init(&env, LCT_LOCAL);
1181         if (rc) {
1182                 CERROR("%s: can't initialize env: rc = %d\n",
1183                        obd->obd_name, rc);
1184
1185                 spin_lock(&d->opd_sync_lock);
1186                 thread->t_flags = SVC_STOPPED;
1187                 spin_unlock(&d->opd_sync_lock);
1188                 wake_up(&thread->t_ctl_waitq);
1189
1190                 RETURN(rc);
1191         }
1192
1193         spin_lock(&d->opd_sync_lock);
1194         thread->t_flags = SVC_RUNNING;
1195         spin_unlock(&d->opd_sync_lock);
1196         wake_up(&thread->t_ctl_waitq);
1197
1198         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1199         if (ctxt == NULL) {
1200                 CERROR("can't get appropriate context\n");
1201                 GOTO(out, rc = -EINVAL);
1202         }
1203
1204         llh = ctxt->loc_handle;
1205         if (llh == NULL) {
1206                 CERROR("can't get llh\n");
1207                 llog_ctxt_put(ctxt);
1208                 GOTO(out, rc = -EINVAL);
1209         }
1210
1211         /*
1212          * Catalog processing stops when it processed last catalog record
1213          * with index equal to the end of catalog bitmap. Or if it is wrapped,
1214          * processing stops with index equal to the lgh_last_idx. We need to
1215          * continue processing.
1216          */
1217         d->opd_sync_last_catalog_idx = 0;
1218         do {
1219                 int     size;
1220
1221                 wrapped = (llh->lgh_hdr->llh_cat_idx >= llh->lgh_last_idx &&
1222                            llh->lgh_hdr->llh_count > 1);
1223
1224                 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d,
1225                                       d->opd_sync_last_catalog_idx, 0);
1226
1227                 size = OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ?
1228                        cfs_fail_val : (LLOG_HDR_BITMAP_SIZE(llh->lgh_hdr) - 1);
1229                 /* processing reaches catalog bottom */
1230                 if (d->opd_sync_last_catalog_idx == size)
1231                         d->opd_sync_last_catalog_idx = 0;
1232                 else if (wrapped)
1233                         /* If catalog is wrapped we can`t predict last index of
1234                          * processing because lgh_last_idx could be changed.
1235                          * Starting form the next one */
1236                         d->opd_sync_last_catalog_idx++;
1237
1238         } while (rc == 0 && (wrapped || d->opd_sync_last_catalog_idx == 0));
1239
1240         if (rc < 0) {
1241                 CERROR("%s: llog process with osp_sync_process_queues "
1242                        "failed: %d\n", d->opd_obd->obd_name, rc);
1243                 GOTO(close, rc);
1244         }
1245         LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1246                  "%u changes, %u in progress, %u in flight: %d\n",
1247                  atomic_read(&d->opd_sync_changes),
1248                  atomic_read(&d->opd_sync_rpcs_in_progress),
1249                  atomic_read(&d->opd_sync_rpcs_in_flight), rc);
1250
1251         /* we don't expect llog_process_thread() to exit till umount */
1252         LASSERTF(thread->t_flags != SVC_RUNNING,
1253                  "%u changes, %u in progress, %u in flight\n",
1254                  atomic_read(&d->opd_sync_changes),
1255                  atomic_read(&d->opd_sync_rpcs_in_progress),
1256                  atomic_read(&d->opd_sync_rpcs_in_flight));
1257
1258         /* wait till all the requests are completed */
1259         count = 0;
1260         while (atomic_read(&d->opd_sync_rpcs_in_progress) > 0) {
1261                 osp_sync_process_committed(&env, d);
1262
1263                 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1264                 rc = l_wait_event(d->opd_sync_waitq,
1265                                 atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1266                                   &lwi);
1267                 if (rc == -ETIMEDOUT)
1268                         count++;
1269                 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1270                          d->opd_obd->obd_name,
1271                          atomic_read(&d->opd_sync_rpcs_in_progress),
1272                          atomic_read(&d->opd_sync_rpcs_in_flight),
1273                          list_empty(&d->opd_sync_committed_there) ? "" : "!");
1274
1275         }
1276
1277 close:
1278         llog_cat_close(&env, llh);
1279         rc = llog_cleanup(&env, ctxt);
1280         if (rc)
1281                 CERROR("can't cleanup llog: %d\n", rc);
1282 out:
1283         LASSERTF(atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1284                  "%s: %d %d %sempty\n", d->opd_obd->obd_name,
1285                  atomic_read(&d->opd_sync_rpcs_in_progress),
1286                  atomic_read(&d->opd_sync_rpcs_in_flight),
1287                  list_empty(&d->opd_sync_committed_there) ? "" : "!");
1288
1289         thread->t_flags = SVC_STOPPED;
1290
1291         wake_up(&thread->t_ctl_waitq);
1292
1293         lu_env_fini(&env);
1294
1295         RETURN(0);
1296 }
1297
1298 /**
1299  * Initialize llog.
1300  *
1301  * Initializes the llog. Specific llog to be used depends on the type of the
1302  * target OSP represents (OST or MDT). The function adds appends a new llog
1303  * record to mark the place where the records associated with this boot
1304  * start.
1305  *
1306  * \param[in] env       LU environment provided by the caller
1307  * \param[in] d         OSP device
1308  *
1309  * \retval 0            on success
1310  * \retval negative     negated errno on error
1311  */
1312 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1313 {
1314         struct osp_thread_info  *osi = osp_env_info(env);
1315         struct lu_fid           *fid = &osi->osi_fid;
1316         struct llog_handle      *lgh = NULL;
1317         struct obd_device       *obd = d->opd_obd;
1318         struct llog_ctxt        *ctxt;
1319         int                     rc;
1320
1321         ENTRY;
1322
1323         LASSERT(obd);
1324
1325         /*
1326          * open llog corresponding to our OST
1327          */
1328         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1329         obd->obd_lvfs_ctxt.dt = d->opd_storage;
1330
1331         lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1332
1333         rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1334                                    &osi->osi_cid, fid);
1335         if (rc < 0) {
1336                 if (rc != -EFAULT) {
1337                         CERROR("%s: can't get id from catalogs: rc = %d\n",
1338                                obd->obd_name, rc);
1339                         RETURN(rc);
1340                 }
1341
1342                 /* After sparse OST indices is supported, the CATALOG file
1343                  * may become a sparse file that results in failure on
1344                  * reading. Skip this error as the llog will be created
1345                  * later */
1346                 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1347                 rc = 0;
1348         }
1349
1350         CDEBUG(D_INFO, "%s: Init llog for %d - catid "DFID":%x\n",
1351                obd->obd_name, d->opd_index,
1352                PFID(&osi->osi_cid.lci_logid.lgl_oi.oi_fid),
1353                osi->osi_cid.lci_logid.lgl_ogen);
1354
1355         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT,
1356                         d->opd_storage->dd_lu_dev.ld_obd,
1357                         &osp_mds_ost_orig_logops);
1358         if (rc)
1359                 RETURN(rc);
1360
1361         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1362         LASSERT(ctxt);
1363
1364         if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1365                 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1366                                LLOG_OPEN_EXISTS);
1367                 /* re-create llog if it is missing */
1368                 if (rc == -ENOENT)
1369                         logid_set_id(&osi->osi_cid.lci_logid, 0);
1370                 else if (rc < 0)
1371                         GOTO(out_cleanup, rc);
1372         }
1373
1374         if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1375                 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1376                 if (rc < 0)
1377                         GOTO(out_cleanup, rc);
1378                 osi->osi_cid.lci_logid = lgh->lgh_id;
1379         }
1380
1381         LASSERT(lgh != NULL);
1382         ctxt->loc_handle = lgh;
1383
1384         rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
1385         if (rc)
1386                 GOTO(out_close, rc);
1387
1388         rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1389                                    &osi->osi_cid, fid);
1390         if (rc)
1391                 GOTO(out_close, rc);
1392
1393         /*
1394          * put a mark in the llog till which we'll be processing
1395          * old records restless
1396          */
1397         d->opd_sync_generation.mnt_cnt = cfs_time_current();
1398         d->opd_sync_generation.conn_cnt = cfs_time_current();
1399
1400         osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1401         osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1402
1403         memcpy(&osi->osi_gen.lgr_gen, &d->opd_sync_generation,
1404                sizeof(osi->osi_gen.lgr_gen));
1405
1406         rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1407         if (rc < 0)
1408                 GOTO(out_close, rc);
1409         llog_ctxt_put(ctxt);
1410         RETURN(0);
1411 out_close:
1412         llog_cat_close(env, lgh);
1413 out_cleanup:
1414         llog_cleanup(env, ctxt);
1415         RETURN(rc);
1416 }
1417
1418 /**
1419  * Cleanup llog used for syncing.
1420  *
1421  * Closes and cleanups the llog. The function is called when the device is
1422  * shutting down.
1423  *
1424  * \param[in] env       LU environment provided by the caller
1425  * \param[in] d         OSP device
1426  */
1427 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1428 {
1429         struct llog_ctxt *ctxt;
1430
1431         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1432         if (ctxt) {
1433                 llog_cat_close(env, ctxt->loc_handle);
1434                 llog_cleanup(env, ctxt);
1435         }
1436 }
1437
1438 /**
1439  * Initialization of the sync component of OSP.
1440  *
1441  * Initializes the llog and starts a new thread to handle the changes to
1442  * the remote target (OST or MDT).
1443  *
1444  * \param[in] env       LU environment provided by the caller
1445  * \param[in] d         OSP device
1446  *
1447  * \retval 0            on success
1448  * \retval negative     negated errno on error
1449  */
1450 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1451 {
1452         struct l_wait_info       lwi = { 0 };
1453         struct task_struct      *task;
1454         int                      rc;
1455
1456         ENTRY;
1457
1458         d->opd_sync_max_rpcs_in_flight = OSP_MAX_RPCS_IN_FLIGHT;
1459         d->opd_sync_max_rpcs_in_progress = OSP_MAX_RPCS_IN_PROGRESS;
1460         spin_lock_init(&d->opd_sync_lock);
1461         init_waitqueue_head(&d->opd_sync_waitq);
1462         init_waitqueue_head(&d->opd_sync_barrier_waitq);
1463         thread_set_flags(&d->opd_sync_thread, SVC_INIT);
1464         init_waitqueue_head(&d->opd_sync_thread.t_ctl_waitq);
1465         INIT_LIST_HEAD(&d->opd_sync_in_flight_list);
1466         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1467
1468         if (d->opd_storage->dd_rdonly)
1469                 RETURN(0);
1470
1471         /*
1472          * initialize llog storing changes
1473          */
1474         rc = osp_sync_llog_init(env, d);
1475         if (rc) {
1476                 CERROR("%s: can't initialize llog: rc = %d\n",
1477                        d->opd_obd->obd_name, rc);
1478                 GOTO(err_id, rc);
1479         }
1480
1481         /*
1482          * Start synchronization thread
1483          */
1484         task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1485                            d->opd_index, d->opd_group);
1486         if (IS_ERR(task)) {
1487                 rc = PTR_ERR(task);
1488                 CERROR("%s: cannot start sync thread: rc = %d\n",
1489                        d->opd_obd->obd_name, rc);
1490                 GOTO(err_llog, rc);
1491         }
1492
1493         l_wait_event(d->opd_sync_thread.t_ctl_waitq,
1494                      osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1495
1496         RETURN(0);
1497 err_llog:
1498         osp_sync_llog_fini(env, d);
1499 err_id:
1500         return rc;
1501 }
1502
1503 /**
1504  * Stop the syncing thread.
1505  *
1506  * Asks the syncing thread to stop and wait until it's stopped.
1507  *
1508  * \param[in] d         OSP device
1509  *
1510  * \retval              0
1511  */
1512 int osp_sync_fini(struct osp_device *d)
1513 {
1514         struct ptlrpc_thread *thread = &d->opd_sync_thread;
1515
1516         ENTRY;
1517
1518         if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1519                 thread->t_flags = SVC_STOPPING;
1520                 wake_up(&d->opd_sync_waitq);
1521                 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1522         }
1523
1524         RETURN(0);
1525 }
1526
1527 struct osp_last_committed_cb {
1528         struct dt_txn_commit_cb  ospc_cb;
1529         struct osp_device       *ospc_dev;
1530         __u64                    ospc_transno;
1531 };
1532
1533 void osp_sync_local_commit_cb(struct lu_env *env, struct thandle *th,
1534                               struct dt_txn_commit_cb *dcb, int err)
1535 {
1536         struct osp_last_committed_cb    *cb;
1537         struct osp_device               *d;
1538
1539         cb = container_of0(dcb, struct osp_last_committed_cb, ospc_cb);
1540         d = cb->ospc_dev;
1541
1542         CDEBUG(D_HA, "%s: %llu committed\n", d->opd_obd->obd_name,
1543                cb->ospc_transno);
1544
1545         spin_lock(&d->opd_sync_lock);
1546         if (cb->ospc_transno > d->opd_sync_last_committed_id)
1547                 d->opd_sync_last_committed_id = cb->ospc_transno;
1548         spin_unlock(&d->opd_sync_lock);
1549
1550         osp_sync_check_for_work(d);
1551         lu_device_put(osp2lu_dev(d));
1552         if (atomic_dec_and_test(&d->opd_commits_registered))
1553                 wake_up(&d->opd_sync_waitq);
1554
1555         OBD_FREE_PTR(cb);
1556 }
1557
1558 int osp_sync_add_commit_cb(const struct lu_env *env, struct osp_device *d,
1559                            struct thandle *th)
1560 {
1561         struct osp_last_committed_cb    *cb;
1562         struct dt_txn_commit_cb         *dcb;
1563         int                              rc = 0;
1564
1565         OBD_ALLOC_PTR(cb);
1566         if (cb == NULL)
1567                 return -ENOMEM;
1568         cb->ospc_dev = d;
1569         dcb = &cb->ospc_cb;
1570         dcb->dcb_func = osp_sync_local_commit_cb;
1571         spin_lock(&d->opd_sync_lock);
1572         cb->ospc_transno = ++d->opd_sync_last_used_id;
1573         spin_unlock(&d->opd_sync_lock);
1574
1575         rc = dt_trans_cb_add(th, dcb);
1576         CDEBUG(D_HA, "%s: add commit cb at %llu, next at %llu, rc = %d\n",
1577                d->opd_obd->obd_name, (unsigned long long) cfs_time_current(),
1578                (unsigned long long) d->opd_sync_next_commit_cb, rc);
1579
1580         if (likely(rc == 0)) {
1581                 lu_device_get(osp2lu_dev(d));
1582                 atomic_inc(&d->opd_commits_registered);
1583         } else
1584                 OBD_FREE_PTR(cb);
1585
1586         return rc;
1587 }
1588
1589 /*
1590  * generate an empty transaction and hook the commit callback in
1591  * then force transaction commit
1592  */
1593 void osp_sync_force(const struct lu_env *env, struct osp_device *d)
1594 {
1595         struct thandle *th;
1596         int rc;
1597
1598         th = dt_trans_create(env, d->opd_storage);
1599         if (IS_ERR(th)) {
1600                 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
1601                 return;
1602         }
1603         rc = dt_trans_start_local(env, d->opd_storage, th);
1604         if (rc == 0) {
1605                 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
1606                        d->opd_obd->obd_name,
1607                        atomic_read(&d->opd_sync_changes));
1608                 rc = osp_sync_add_commit_cb(env, d, th);
1609                 dt_trans_stop(env, d->opd_storage, th);
1610         }
1611
1612         dt_commit_async(env, d->opd_storage);
1613 }