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, 2017, 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 int osp_sync_add_commit_cb(const struct lu_env *env,
  99                                   struct osp_device *d, struct thandle *th);
 100
 101 static inline int osp_sync_running(struct osp_device *d)
 102 {
 103         return !!(d->opd_sync_thread.t_flags & SVC_RUNNING);
 104 }
 105
 106 /**
 107  * Check status: whether OSP thread has stopped
 108  *
 109  * \param[in] d         OSP device
 110  *
 111  * \retval 0            still running
 112  * \retval 1            stopped
 113  */
 114 static inline int osp_sync_stopped(struct osp_device *d)
 115 {
 116         return !!(d->opd_sync_thread.t_flags & SVC_STOPPED);
 117 }
 118
 119 /*
 120  ** Check for new changes to sync
 121  *
 122  * \param[in] d         OSP device
 123  *
 124  * \retval 1            there are changes
 125  * \retval 0            there are no changes
 126  */
 127 static inline int osp_sync_has_new_job(struct osp_device *d)
 128 {
 129         return atomic_read(&d->opd_sync_changes) > 0 ||
 130                 d->opd_sync_prev_done == 0;
 131 }
 132
 133 static inline int osp_sync_in_flight_conflict(struct osp_device *d,
 134                                              struct llog_rec_hdr *h)
 135 {
 136         struct osp_job_req_args *jra;
 137         struct ost_id            ostid;
 138         int                      conflict = 0;
 139
 140         if (h == NULL || h->lrh_type == LLOG_GEN_REC ||
 141             list_empty(&d->opd_sync_in_flight_list))
 142                 return conflict;
 143
 144         memset(&ostid, 0, sizeof(ostid));
 145         switch (h->lrh_type) {
 146         case MDS_UNLINK_REC: {
 147                 struct llog_unlink_rec *unlink = (struct llog_unlink_rec *)h;
 148
 149                 ostid_set_seq(&ostid, unlink->lur_oseq);
 150                 if (ostid_set_id(&ostid, unlink->lur_oid)) {
 151                         CERROR("Bad %llu to set " DOSTID "\n",
 152                                (unsigned long long)(unlink->lur_oid),
 153                                POSTID(&ostid));
 154                         return 1;
 155                 }
 156                 }
 157                 break;
 158         case MDS_UNLINK64_REC:
 159                 fid_to_ostid(&((struct llog_unlink64_rec *)h)->lur_fid, &ostid);
 160                 break;
 161         case MDS_SETATTR64_REC:
 162                 ostid = ((struct llog_setattr64_rec *)h)->lsr_oi;
 163                 break;
 164         default:
 165                 LBUG();
 166         }
 167
 168         spin_lock(&d->opd_sync_lock);
 169         list_for_each_entry(jra, &d->opd_sync_in_flight_list,
 170                             jra_in_flight_link) {
 171                 struct ptlrpc_request   *req;
 172                 struct ost_body         *body;
 173
 174                 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
 175
 176                 req = container_of((void *)jra, struct ptlrpc_request,
 177                                    rq_async_args);
 178                 body = req_capsule_client_get(&req->rq_pill,
 179                                               &RMF_OST_BODY);
 180                 LASSERT(body);
 181
 182                 if (memcmp(&ostid, &body->oa.o_oi, sizeof(ostid)) == 0) {
 183                         conflict = 1;
 184                         break;
 185                 }
 186         }
 187         spin_unlock(&d->opd_sync_lock);
 188
 189         return conflict;
 190 }
 191
 192 static inline int osp_sync_rpcs_in_progress_low(struct osp_device *d)
 193 {
 194         return atomic_read(&d->opd_sync_rpcs_in_progress) <
 195                 d->opd_sync_max_rpcs_in_progress;
 196 }
 197
 198 /**
 199  * Check for room in the network pipe to OST
 200  *
 201  * \param[in] d         OSP device
 202  *
 203  * \retval 1            there is room
 204  * \retval 0            no room, the pipe is full
 205  */
 206 static inline int osp_sync_rpcs_in_flight_low(struct osp_device *d)
 207 {
 208         return atomic_read(&d->opd_sync_rpcs_in_flight) <
 209                 d->opd_sync_max_rpcs_in_flight;
 210 }
 211
 212 /**
 213  * Wake up check for the main sync thread
 214  *
 215  * \param[in] d         OSP device
 216  *
 217  * \retval 1            time to wake up
 218  * \retval 0            no need to wake up
 219  */
 220 static inline int osp_sync_has_work(struct osp_device *osp)
 221 {
 222         /* has new/old changes and low in-progress? */
 223         if (osp_sync_has_new_job(osp) && osp_sync_rpcs_in_progress_low(osp) &&
 224             osp_sync_rpcs_in_flight_low(osp) && osp->opd_imp_connected)
 225                 return 1;
 226
 227         /* has remotely committed? */
 228         if (!list_empty(&osp->opd_sync_committed_there))
 229                 return 1;
 230
 231         return 0;
 232 }
 233
 234 void osp_sync_check_for_work(struct osp_device *osp)
 235 {
 236         if (osp_sync_has_work(osp))
 237                 wake_up(&osp->opd_sync_waitq);
 238 }
 239
 240 static inline __u64 osp_sync_correct_id(struct osp_device *d,
 241                                         struct llog_rec_hdr *rec)
 242 {
 243         /*
 244          * llog use cyclic store with 32 bit lrh_id
 245          * so overflow lrh_id is possible. Range between
 246          * last_processed and last_committed is less than
 247          * 64745 ^ 2 and less than 2^32 - 1
 248          */
 249         __u64 correct_id = d->opd_sync_last_committed_id;
 250
 251         if ((correct_id & 0xffffffffULL) < rec->lrh_id)
 252                 correct_id -= 0x100000000ULL;
 253
 254         correct_id &= ~0xffffffffULL;
 255         correct_id |= rec->lrh_id;
 256
 257         return correct_id;
 258 }
 259
 260 /**
 261  * Check and return ready-for-new status.
 262  *
 263  * The thread processing llog record uses this function to check whether
 264  * it's time to take another record and process it. The number of conditions
 265  * must be met: the connection should be ready, RPCs in flight not exceeding
 266  * the limit, the record is committed locally, etc (see the lines below).
 267  *
 268  * \param[in] d         OSP device
 269  * \param[in] rec       next llog record to process
 270  *
 271  * \retval 0            not ready
 272  * \retval 1            ready
 273  */
 274 static inline int osp_sync_can_process_new(struct osp_device *d,
 275                                            struct llog_rec_hdr *rec)
 276 {
 277         LASSERT(d);
 278
 279         if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
 280                 return 0;
 281         if (unlikely(osp_sync_in_flight_conflict(d, rec)))
 282                 return 0;
 283         if (!osp_sync_rpcs_in_progress_low(d))
 284                 return 0;
 285         if (!osp_sync_rpcs_in_flight_low(d))
 286                 return 0;
 287         if (!d->opd_imp_connected)
 288                 return 0;
 289         if (d->opd_sync_prev_done == 0)
 290                 return 1;
 291         if (atomic_read(&d->opd_sync_changes) == 0)
 292                 return 0;
 293         if (rec == NULL)
 294                 return 1;
 295         /* notice "<" not "<=" */
 296         if (osp_sync_correct_id(d, rec) < d->opd_sync_last_committed_id)
 297                 return 1;
 298         return 0;
 299 }
 300
 301 /**
 302  * Declare intention to add a new change.
 303  *
 304  * With regard to OSD API, we have to declare any changes ahead. In this
 305  * case we declare an intention to add a llog record representing the
 306  * change on the local storage.
 307  *
 308  * \param[in] env       LU environment provided by the caller
 309  * \param[in] o         OSP object
 310  * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 311  * \param[in] th        transaction handle (local)
 312  *
 313  * \retval 0            on success
 314  * \retval negative     negated errno on error
 315  */
 316 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
 317                          enum llog_op_type type, struct thandle *th)
 318 {
 319         struct osp_thread_info  *osi = osp_env_info(env);
 320         struct osp_device       *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
 321         struct llog_ctxt        *ctxt;
 322         struct thandle          *storage_th;
 323         int                      rc;
 324
 325         ENTRY;
 326
 327         /* it's a layering violation, to access internals of th,
 328          * but we can do this as a sanity check, for a while */
 329         LASSERT(th->th_top != NULL);
 330         storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
 331         if (IS_ERR(storage_th))
 332                 RETURN(PTR_ERR(storage_th));
 333
 334         switch (type) {
 335         case MDS_UNLINK64_REC:
 336                 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
 337                 break;
 338         case MDS_SETATTR64_REC:
 339                 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec_v2);
 340                 break;
 341         default:
 342                 LBUG();
 343         }
 344
 345         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 346         LASSERT(ctxt);
 347
 348         rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
 349                               storage_th);
 350         llog_ctxt_put(ctxt);
 351
 352         RETURN(rc);
 353 }
 354
 355 /**
 356  * Generate a llog record for a given change.
 357  *
 358  * Generates a llog record for the change passed. The change can be of two
 359  * types: unlink and setattr. The record gets an ID which later will be
 360  * used to track commit status of the change. For unlink changes, the caller
 361  * can supply a starting FID and the count of the objects to destroy. For
 362  * setattr the caller should apply attributes to apply.
 363  *
 364  *
 365  * \param[in] env       LU environment provided by the caller
 366  * \param[in] d         OSP device
 367  * \param[in] fid       fid of the object the change should be applied to
 368  * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 369  * \param[in] count     count of objects to destroy
 370  * \param[in] th        transaction handle (local)
 371  * \param[in] attr      attributes for setattr
 372  *
 373  * \retval 0            on success
 374  * \retval negative     negated errno on error
 375  */
 376 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
 377                             const struct lu_fid *fid, enum llog_op_type type,
 378                             int count, struct thandle *th,
 379                             const struct lu_attr *attr)
 380 {
 381         struct osp_thread_info  *osi = osp_env_info(env);
 382         struct llog_ctxt        *ctxt;
 383         struct thandle          *storage_th;
 384         bool                     immediate_commit_cb = false;
 385         int                      rc;
 386
 387         ENTRY;
 388
 389         /* it's a layering violation, to access internals of th,
 390          * but we can do this as a sanity check, for a while */
 391         LASSERT(th->th_top != NULL);
 392         storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
 393         if (IS_ERR(storage_th))
 394                 RETURN(PTR_ERR(storage_th));
 395
 396         switch (type) {
 397         case MDS_UNLINK64_REC:
 398                 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
 399                 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
 400                 osi->osi_unlink.lur_fid  = *fid;
 401                 osi->osi_unlink.lur_count = count;
 402                 break;
 403         case MDS_SETATTR64_REC:
 404                 rc = fid_to_ostid(fid, &osi->osi_oi);
 405                 LASSERT(rc == 0);
 406                 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
 407                 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
 408                 osi->osi_setattr.lsr_oi  = osi->osi_oi;
 409                 LASSERT(attr);
 410                 osi->osi_setattr.lsr_uid = attr->la_uid;
 411                 osi->osi_setattr.lsr_gid = attr->la_gid;
 412                 osi->osi_setattr.lsr_layout_version = attr->la_layout_version;
 413                 osi->osi_setattr.lsr_projid = attr->la_projid;
 414                 osi->osi_setattr.lsr_valid =
 415                         ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
 416                         ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0) |
 417                         ((attr->la_valid & LA_PROJID) ? OBD_MD_FLPROJID : 0);
 418                 if (attr->la_valid & LA_LAYOUT_VERSION) {
 419                         osi->osi_setattr.lsr_valid |= OBD_MD_LAYOUT_VERSION;
 420
 421                         /* FLR: the layout version has to be transferred to
 422                          * OST objects ASAP, otherwise clients will have to
 423                          * experience delay to be able to write OST objects. */
 424                         immediate_commit_cb = true;
 425                 }
 426                 break;
 427         default:
 428                 LBUG();
 429         }
 430
 431         /* we keep the same id, but increment it when the callback
 432          * is registered, so that all records upto the one taken
 433          * by the callback are subject to processing */
 434         spin_lock(&d->opd_sync_lock);
 435         osi->osi_hdr.lrh_id = d->opd_sync_last_used_id;
 436         spin_unlock(&d->opd_sync_lock);
 437
 438         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 439         if (ctxt == NULL)
 440                 RETURN(-ENOMEM);
 441
 442         rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
 443                       storage_th);
 444         llog_ctxt_put(ctxt);
 445
 446         if (likely(rc >= 0)) {
 447                 CDEBUG(D_OTHER, "%s: new record "DFID":%x.%u: rc = %d\n",
 448                        d->opd_obd->obd_name,
 449                        PFID(&osi->osi_cookie.lgc_lgl.lgl_oi.oi_fid),
 450                        osi->osi_cookie.lgc_lgl.lgl_ogen,
 451                        osi->osi_cookie.lgc_index, rc);
 452                 atomic_inc(&d->opd_sync_changes);
 453         }
 454
 455         if (immediate_commit_cb)
 456                 rc = osp_sync_add_commit_cb(env, d, th);
 457         else
 458                 rc = osp_sync_add_commit_cb_1s(env, d, th);
 459
 460         /* return 0 always here, error case just cause no llog record */
 461         RETURN(0);
 462 }
 463
 464 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
 465                  enum llog_op_type type, struct thandle *th,
 466                  const struct lu_attr *attr)
 467 {
 468         return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
 469                                 lu_object_fid(&o->opo_obj.do_lu), type, 1,
 470                                 th, attr);
 471 }
 472
 473 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
 474                         struct lu_fid *fid, int lost, struct thandle *th)
 475 {
 476         return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
 477 }
 478
 479 /*
 480  * it's quite obvious we can't maintain all the structures in the memory:
 481  * while OST is down, MDS can be processing thousands and thousands of unlinks
 482  * filling persistent llogs and in-core respresentation
 483  *
 484  * this doesn't scale at all. so we need basically the following:
 485  * a) destroy/setattr append llog records
 486  * b) once llog has grown to X records, we process first Y committed records
 487  *
 488  *  once record R is found via llog_process(), it becomes committed after any
 489  *  subsequent commit callback (at the most)
 490  */
 491
 492 /**
 493  * ptlrpc commit callback.
 494  *
 495  * The callback is called by PTLRPC when a RPC is reported committed by the
 496  * target (OST). We register the callback for the every RPC applying a change
 497  * from the llog. This way we know then the llog records can be cancelled.
 498  * Notice the callback can be called when OSP is finishing. We can detect this
 499  * checking that actual transno in the request is less or equal of known
 500  * committed transno (see osp_sync_process_committed() for the details).
 501  * XXX: this is pretty expensive and can be improved later using batching.
 502  *
 503  * \param[in] req       request
 504  */
 505 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
 506 {
 507         struct osp_device *d = req->rq_cb_data;
 508         struct osp_job_req_args *jra;
 509
 510         CDEBUG(D_HA, "commit req %p, transno %llu\n", req, req->rq_transno);
 511
 512         if (unlikely(req->rq_transno == 0))
 513                 return;
 514
 515         /* do not do any opd_sync_rpcs_* accounting here
 516          * it's done in osp_sync_interpret sooner or later */
 517         LASSERT(d);
 518
 519         jra = ptlrpc_req_async_args(req);
 520         LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
 521         LASSERT(list_empty(&jra->jra_committed_link));
 522
 523         ptlrpc_request_addref(req);
 524
 525         spin_lock(&d->opd_sync_lock);
 526         list_add(&jra->jra_committed_link, &d->opd_sync_committed_there);
 527         spin_unlock(&d->opd_sync_lock);
 528
 529         /* XXX: some batching wouldn't hurt */
 530         wake_up(&d->opd_sync_waitq);
 531 }
 532
 533 /**
 534  * RPC interpretation callback.
 535  *
 536  * The callback is called by ptlrpc when RPC is replied. Now we have to decide
 537  * whether we should:
 538  *  - put request on a special list to wait until it's committed by the target,
 539  *    if the request is successful
 540  *  - schedule llog record cancel if no target object is found
 541  *  - try later (essentially after reboot) in case of unexpected error
 542  *
 543  * \param[in] env       LU environment provided by the caller
 544  * \param[in] req       request replied
 545  * \param[in] aa        callback data
 546  * \param[in] rc        result of RPC
 547  *
 548  * \retval 0            always
 549  */
 550 static int osp_sync_interpret(const struct lu_env *env,
 551                               struct ptlrpc_request *req, void *args, int rc)
 552 {
 553         struct osp_job_req_args *jra = args;
 554         struct osp_device *d = req->rq_cb_data;
 555
 556         if (jra->jra_magic != OSP_JOB_MAGIC) {
 557                 DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
 558                 LBUG();
 559         }
 560         LASSERT(d);
 561
 562         CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
 563                atomic_read(&req->rq_refcount),
 564                rc, (unsigned) req->rq_transno);
 565
 566         if (rc == -ENOENT) {
 567                 /*
 568                  * we tried to destroy object or update attributes,
 569                  * but object doesn't exist anymore - cancell llog record
 570                  */
 571                 LASSERT(req->rq_transno == 0);
 572                 LASSERT(list_empty(&jra->jra_committed_link));
 573
 574                 ptlrpc_request_addref(req);
 575
 576                 spin_lock(&d->opd_sync_lock);
 577                 list_add(&jra->jra_committed_link,
 578                          &d->opd_sync_committed_there);
 579                 spin_unlock(&d->opd_sync_lock);
 580
 581                 wake_up(&d->opd_sync_waitq);
 582         } else if (rc) {
 583                 struct obd_import *imp = req->rq_import;
 584                 /*
 585                  * error happened, we'll try to repeat on next boot ?
 586                  */
 587                 LASSERTF(req->rq_transno == 0 || rc == -EIO ||
 588                          req->rq_import_generation < imp->imp_generation,
 589                          "transno %llu, rc %d, gen: req %d, imp %d\n",
 590                          req->rq_transno, rc, req->rq_import_generation,
 591                          imp->imp_generation);
 592                 if (req->rq_transno == 0) {
 593                         /* this is the last time we see the request
 594                          * if transno is not zero, then commit cb
 595                          * will be called at some point */
 596                         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) > 0);
 597                         atomic_dec(&d->opd_sync_rpcs_in_progress);
 598                 }
 599
 600                 wake_up(&d->opd_sync_waitq);
 601         } else if (d->opd_pre != NULL &&
 602                    unlikely(d->opd_pre_status == -ENOSPC)) {
 603                 /*
 604                  * if current status is -ENOSPC (lack of free space on OST)
 605                  * then we should poll OST immediately once object destroy
 606                  * is replied
 607                  */
 608                 osp_statfs_need_now(d);
 609         }
 610
 611         spin_lock(&d->opd_sync_lock);
 612         list_del_init(&jra->jra_in_flight_link);
 613         spin_unlock(&d->opd_sync_lock);
 614         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) > 0);
 615         atomic_dec(&d->opd_sync_rpcs_in_flight);
 616         if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
 617                 wake_up(&d->opd_sync_barrier_waitq);
 618         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 619                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 620                atomic_read(&d->opd_sync_rpcs_in_progress));
 621
 622         osp_sync_check_for_work(d);
 623
 624         return 0;
 625 }
 626
 627 /*
 628  ** Add request to ptlrpc queue.
 629  *
 630  * This is just a tiny helper function to put the request on the sending list
 631  *
 632  * \param[in] d         OSP device
 633  * \param[in] llh       llog handle where the record is stored
 634  * \param[in] h         llog record
 635  * \param[in] req       request
 636  */
 637 static void osp_sync_send_new_rpc(struct osp_device *d,
 638                                   struct llog_handle *llh,
 639                                   struct llog_rec_hdr *h,
 640                                   struct ptlrpc_request *req)
 641 {
 642         struct osp_job_req_args *jra;
 643
 644         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) <=
 645                 d->opd_sync_max_rpcs_in_flight);
 646
 647         jra = ptlrpc_req_async_args(req);
 648         jra->jra_magic = OSP_JOB_MAGIC;
 649         jra->jra_lcookie.lgc_lgl = llh->lgh_id;
 650         jra->jra_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 651         jra->jra_lcookie.lgc_index = h->lrh_index;
 652         INIT_LIST_HEAD(&jra->jra_committed_link);
 653         spin_lock(&d->opd_sync_lock);
 654         list_add_tail(&jra->jra_in_flight_link, &d->opd_sync_in_flight_list);
 655         spin_unlock(&d->opd_sync_lock);
 656
 657         ptlrpcd_add_req(req);
 658 }
 659
 660
 661 /**
 662  * Allocate and prepare RPC for a new change.
 663  *
 664  * The function allocates and initializes an RPC which will be sent soon to
 665  * apply the change to the target OST. The request is initialized from the
 666  * llog record passed. Notice only the fields common to all type of changes
 667  * are initialized.
 668  *
 669  * \param[in] d         OSP device
 670  * \param[in] op        type of the change
 671  * \param[in] format    request format to be used
 672  *
 673  * \retval pointer              new request on success
 674  * \retval ERR_PTR(errno)       on error
 675  */
 676 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
 677                                                enum ost_cmd op,
 678                                                const struct req_format *format)
 679 {
 680         struct ptlrpc_request   *req;
 681         struct obd_import       *imp;
 682         int                      rc;
 683
 684         /* Prepare the request */
 685         imp = d->opd_obd->u.cli.cl_import;
 686         LASSERT(imp);
 687
 688         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
 689                 RETURN(ERR_PTR(-ENOMEM));
 690
 691         req = ptlrpc_request_alloc(imp, format);
 692         if (req == NULL)
 693                 RETURN(ERR_PTR(-ENOMEM));
 694
 695         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
 696         if (rc) {
 697                 ptlrpc_req_finished(req);
 698                 return ERR_PTR(rc);
 699         }
 700
 701         req->rq_interpret_reply = osp_sync_interpret;
 702         req->rq_commit_cb = osp_sync_request_commit_cb;
 703         req->rq_cb_data = d;
 704
 705         ptlrpc_request_set_replen(req);
 706
 707         return req;
 708 }
 709
 710 /**
 711  * Generate a request for setattr change.
 712  *
 713  * The function prepares a new RPC, initializes it with setattr specific
 714  * bits and send the RPC.
 715  *
 716  * \param[in] d         OSP device
 717  * \param[in] llh       llog handle where the record is stored
 718  * \param[in] h         llog record
 719  *
 720  * \retval 0            on success
 721  * \retval 1            on invalid record
 722  * \retval negative     negated errno on error
 723  */
 724 static int osp_sync_new_setattr_job(struct osp_device *d,
 725                                     struct llog_handle *llh,
 726                                     struct llog_rec_hdr *h)
 727 {
 728         struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
 729         struct ptlrpc_request           *req;
 730         struct ost_body                 *body;
 731
 732         ENTRY;
 733         LASSERT(h->lrh_type == MDS_SETATTR64_REC);
 734
 735         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
 736                 RETURN(1);
 737
 738         /* lsr_valid can only be 0 or HAVE OBD_MD_{FLUID, FLGID, FLPROJID} set,
 739          * so no bits other than these should be set. */
 740         if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID |
 741             OBD_MD_FLPROJID | OBD_MD_LAYOUT_VERSION)) != 0) {
 742                 CERROR("%s: invalid setattr record, lsr_valid:%llu\n",
 743                         d->opd_obd->obd_name, rec->lsr_valid);
 744                 /* return 1 on invalid record */
 745                 RETURN(1);
 746         }
 747
 748         req = osp_sync_new_job(d, OST_SETATTR, &RQF_OST_SETATTR);
 749         if (IS_ERR(req))
 750                 RETURN(PTR_ERR(req));
 751
 752         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 753         LASSERT(body);
 754         body->oa.o_oi = rec->lsr_oi;
 755         body->oa.o_uid = rec->lsr_uid;
 756         body->oa.o_gid = rec->lsr_gid;
 757         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 758         if (h->lrh_len > sizeof(struct llog_setattr64_rec)) {
 759                 struct llog_setattr64_rec_v2 *rec_v2 = (typeof(rec_v2))rec;
 760                 body->oa.o_projid = rec_v2->lsr_projid;
 761                 body->oa.o_layout_version = rec_v2->lsr_layout_version;
 762         }
 763
 764         /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
 765          * we assume that both UID and GID are valid in that case. */
 766         if (rec->lsr_valid == 0)
 767                 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
 768         else
 769                 body->oa.o_valid |= rec->lsr_valid;
 770
 771         if (body->oa.o_valid & OBD_MD_LAYOUT_VERSION) {
 772                 OBD_FAIL_TIMEOUT(OBD_FAIL_FLR_LV_DELAY, cfs_fail_val);
 773                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_FLR_LV_INC)))
 774                         ++body->oa.o_layout_version;
 775         }
 776
 777         osp_sync_send_new_rpc(d, llh, h, req);
 778         RETURN(0);
 779 }
 780
 781 /**
 782  * Generate a request for unlink change.
 783  *
 784  * The function prepares a new RPC, initializes it with unlink(destroy)
 785  * specific bits and sends the RPC. The function is used to handle
 786  * llog_unlink_rec which were used in the older versions of Lustre.
 787  * Current version uses llog_unlink_rec64.
 788  *
 789  * \param[in] d         OSP device
 790  * \param[in] llh       llog handle where the record is stored
 791  * \param[in] h         llog record
 792  *
 793  * \retval 0            on success
 794  * \retval negative     negated errno on error
 795  */
 796 static int osp_sync_new_unlink_job(struct osp_device *d,
 797                                    struct llog_handle *llh,
 798                                    struct llog_rec_hdr *h)
 799 {
 800         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 801         struct ptlrpc_request   *req;
 802         struct ost_body         *body;
 803         int rc;
 804
 805         ENTRY;
 806         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 807
 808         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 809         if (IS_ERR(req))
 810                 RETURN(PTR_ERR(req));
 811
 812         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 813         LASSERT(body);
 814         ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
 815         rc = ostid_set_id(&body->oa.o_oi, rec->lur_oid);
 816         if (rc)
 817                 return rc;
 818         body->oa.o_misc = rec->lur_count;
 819         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 820         if (rec->lur_count)
 821                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 822
 823         osp_sync_send_new_rpc(d, llh, h, req);
 824         RETURN(0);
 825 }
 826
 827 /**
 828  * Generate a request for unlink change.
 829  *
 830  * The function prepares a new RPC, initializes it with unlink(destroy)
 831  * specific bits and sends the RPC. Depending on the target (MDT or OST)
 832  * two different protocols are used. For MDT we use OUT (basically OSD API
 833  * updates transferred via a network). For OST we still use the old
 834  * protocol (OBD?), originally for compatibility. Later we can start to
 835  * use OUT for OST as well, this will allow batching and better code
 836  * unification.
 837  *
 838  * \param[in] d         OSP device
 839  * \param[in] llh       llog handle where the record is stored
 840  * \param[in] h         llog record
 841  *
 842  * \retval 0            on success
 843  * \retval negative     negated errno on error
 844  */
 845 static int osp_sync_new_unlink64_job(struct osp_device *d,
 846                                      struct llog_handle *llh,
 847                                      struct llog_rec_hdr *h)
 848 {
 849         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 850         struct ptlrpc_request           *req = NULL;
 851         struct ost_body                 *body;
 852         int                              rc;
 853
 854         ENTRY;
 855         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 856         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 857         if (IS_ERR(req))
 858                 RETURN(PTR_ERR(req));
 859
 860         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 861         if (body == NULL)
 862                 RETURN(-EFAULT);
 863         rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
 864         if (rc < 0)
 865                 RETURN(rc);
 866         body->oa.o_misc = rec->lur_count;
 867         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 868                            OBD_MD_FLOBJCOUNT;
 869         osp_sync_send_new_rpc(d, llh, h, req);
 870         RETURN(0);
 871 }
 872
 873 /**
 874  * Process llog records.
 875  *
 876  * This function is called to process the llog records committed locally.
 877  * In the recovery model used by OSP we can apply a change to a remote
 878  * target once corresponding transaction (like posix unlink) is committed
 879  * locally so can't revert.
 880  * Depending on the llog record type, a given handler is called that is
 881  * responsible for preparing and sending the RPC to apply the change.
 882  * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
 883  *
 884  * \param[in] env       LU environment provided by the caller
 885  * \param[in] d         OSP device
 886  * \param[in] llh       llog handle where the record is stored
 887  * \param[in] rec       llog record
 888  */
 889 static void osp_sync_process_record(const struct lu_env *env,
 890                                     struct osp_device *d,
 891                                     struct llog_handle *llh,
 892                                     struct llog_rec_hdr *rec)
 893 {
 894         struct llog_handle      *cathandle = llh->u.phd.phd_cat_handle;
 895         struct llog_cookie       cookie;
 896         int                      rc = 0;
 897
 898         ENTRY;
 899
 900         cookie.lgc_lgl = llh->lgh_id;
 901         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 902         cookie.lgc_index = rec->lrh_index;
 903
 904         d->opd_sync_last_catalog_idx = llh->lgh_hdr->llh_cat_idx;
 905
 906         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 907                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 908
 909                 /* we're waiting for the record generated by this instance */
 910                 LASSERT(d->opd_sync_prev_done == 0);
 911                 if (!memcmp(&d->opd_sync_generation, &gen->lgr_gen,
 912                             sizeof(gen->lgr_gen))) {
 913                         CDEBUG(D_HA, "processed all old entries\n");
 914                         d->opd_sync_prev_done = 1;
 915                 }
 916
 917                 /* cancel any generation record */
 918                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 919
 920                 /* flush all pending records ASAP */
 921                 osp_sync_force(env, d);
 922
 923                 RETURN_EXIT;
 924         }
 925
 926         /*
 927          * now we prepare and fill requests to OST, put them on the queue
 928          * and fire after next commit callback
 929          */
 930
 931         /* notice we increment counters before sending RPC, to be consistent
 932          * in RPC interpret callback which may happen very quickly */
 933         atomic_inc(&d->opd_sync_rpcs_in_flight);
 934         atomic_inc(&d->opd_sync_rpcs_in_progress);
 935
 936         switch (rec->lrh_type) {
 937         /* case MDS_UNLINK_REC is kept for compatibility */
 938         case MDS_UNLINK_REC:
 939                 rc = osp_sync_new_unlink_job(d, llh, rec);
 940                 break;
 941         case MDS_UNLINK64_REC:
 942                 rc = osp_sync_new_unlink64_job(d, llh, rec);
 943                 break;
 944         case MDS_SETATTR64_REC:
 945                 rc = osp_sync_new_setattr_job(d, llh, rec);
 946                 break;
 947         default:
 948                 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
 949                        rec->lrh_type);
 950                 /* treat "unknown record type" as "invalid" */
 951                 rc = 1;
 952                 break;
 953         }
 954
 955         /* For all kinds of records, not matter successful or not,
 956          * we should decrease changes and bump last_processed_id.
 957          */
 958         if (d->opd_sync_prev_done) {
 959                 LASSERT(atomic_read(&d->opd_sync_changes) > 0);
 960                 atomic_dec(&d->opd_sync_changes);
 961                 wake_up(&d->opd_sync_barrier_waitq);
 962         }
 963         atomic64_inc(&d->opd_sync_processed_recs);
 964         if (rc != 0) {
 965                 atomic_dec(&d->opd_sync_rpcs_in_flight);
 966                 atomic_dec(&d->opd_sync_rpcs_in_progress);
 967         }
 968
 969         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 970                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 971                atomic_read(&d->opd_sync_rpcs_in_progress));
 972
 973         /* Delete the invalid record */
 974         if (rc == 1) {
 975                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 976                 if (rc != 0)
 977                         CERROR("%s: can't delete invalid record: "
 978                                "fid = "DFID", rec_id = %u, rc = %d\n",
 979                                d->opd_obd->obd_name,
 980                                PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
 981                                rec->lrh_id, rc);
 982         }
 983
 984         CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
 985                rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);
 986
 987         RETURN_EXIT;
 988 }
 989
 990 /**
 991  * Cancel llog records for the committed changes.
 992  *
 993  * The function walks through the list of the committed RPCs and cancels
 994  * corresponding llog records. see osp_sync_request_commit_cb() for the
 995  * details.
 996  *
 997  * \param[in] env       LU environment provided by the caller
 998  * \param[in] d         OSP device
 999  */
1000 static void osp_sync_process_committed(const struct lu_env *env,
1001                                        struct osp_device *d)
1002 {
1003         struct obd_device       *obd = d->opd_obd;
1004         struct obd_import       *imp = obd->u.cli.cl_import;
1005         struct ost_body         *body;
1006         struct ptlrpc_request   *req;
1007         struct llog_ctxt        *ctxt;
1008         struct llog_handle      *llh;
1009         int                     *arr;
1010         struct list_head         list, *le;
1011         struct llog_logid        lgid;
1012         int                      rc, i, count = 0, done = 0;
1013
1014         ENTRY;
1015
1016         if (list_empty(&d->opd_sync_committed_there))
1017                 return;
1018
1019         /*
1020          * if current status is -ENOSPC (lack of free space on OST)
1021          * then we should poll OST immediately once object destroy
1022          * is committed.
1023          * notice: we do this upon commit as well because some backends
1024          * (like DMU) do not release space right away.
1025          */
1026         if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
1027                 osp_statfs_need_now(d);
1028
1029         /*
1030          * now cancel them all
1031          * XXX: can we improve this using some batching?
1032          *      with batch RPC that'll happen automatically?
1033          * XXX: can we store ctxt in lod_device and save few cycles ?
1034          */
1035         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1036         LASSERT(ctxt);
1037
1038         llh = ctxt->loc_handle;
1039         LASSERT(llh);
1040
1041         INIT_LIST_HEAD(&list);
1042         spin_lock(&d->opd_sync_lock);
1043         list_splice(&d->opd_sync_committed_there, &list);
1044         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1045         spin_unlock(&d->opd_sync_lock);
1046
1047         list_for_each(le, &list)
1048                 count++;
1049         if (count > 2)
1050                 OBD_ALLOC_WAIT(arr, sizeof(int) * count);
1051         else
1052                 arr = NULL;
1053         i = 0;
1054         while (!list_empty(&list)) {
1055                 struct osp_job_req_args *jra;
1056
1057                 jra = list_entry(list.next, struct osp_job_req_args,
1058                                  jra_committed_link);
1059                 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1060                 list_del_init(&jra->jra_committed_link);
1061
1062                 req = container_of((void *)jra, struct ptlrpc_request,
1063                                    rq_async_args);
1064                 body = req_capsule_client_get(&req->rq_pill,
1065                                               &RMF_OST_BODY);
1066                 LASSERT(body);
1067                 /* import can be closing, thus all commit cb's are
1068                  * called we can check committness directly */
1069                 if (req->rq_import_generation == imp->imp_generation) {
1070                         if (arr && (!i ||
1071                                     !memcmp(&jra->jra_lcookie.lgc_lgl, &lgid,
1072                                            sizeof(lgid)))) {
1073                                 if (unlikely(!i))
1074                                         lgid = jra->jra_lcookie.lgc_lgl;
1075
1076                                 arr[i++] = jra->jra_lcookie.lgc_index;
1077                         } else {
1078                                 rc = llog_cat_cancel_records(env, llh, 1,
1079                                                              &jra->jra_lcookie);
1080                                 if (rc)
1081                                         CERROR("%s: can't cancel record: %d\n",
1082                                                obd->obd_name, rc);
1083                         }
1084                 } else {
1085                         DEBUG_REQ(D_OTHER, req, "imp_committed = %llu",
1086                                   imp->imp_peer_committed_transno);
1087                 }
1088                 ptlrpc_req_finished(req);
1089                 done++;
1090         }
1091         if (arr && i > 0) {
1092                 rc = llog_cat_cancel_arr_rec(env, llh, &lgid, i, arr);
1093
1094                 if (rc)
1095                         CERROR("%s: can't cancel %d records rc: %d\n",
1096                                obd->obd_name, i, rc);
1097                 else
1098                         CDEBUG(D_OTHER, "%s: massive records cancel id "DFID\
1099                                " num %d\n", obd->obd_name,
1100                                PFID(&lgid.lgl_oi.oi_fid), i);
1101         }
1102         if (arr)
1103                 OBD_FREE(arr, sizeof(int) * count);
1104
1105         llog_ctxt_put(ctxt);
1106
1107         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) >= done);
1108         atomic_sub(done, &d->opd_sync_rpcs_in_progress);
1109         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress, done %d\n",
1110                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
1111                atomic_read(&d->opd_sync_rpcs_in_progress), done);
1112
1113         osp_sync_check_for_work(d);
1114
1115         /* wake up the thread if requested to stop:
1116          * it might be waiting for in-progress to complete */
1117         if (unlikely(osp_sync_running(d) == 0))
1118                 wake_up(&d->opd_sync_waitq);
1119
1120         EXIT;
1121 }
1122
1123 /**
1124  * The core of the syncing mechanism.
1125  *
1126  * This is a callback called by the llog processing function. Essentially it
1127  * suspends llog processing until there is a record to process (it's supposed
1128  * to be committed locally). The function handles RPCs committed by the target
1129  * and cancels corresponding llog records.
1130  *
1131  * \param[in] env       LU environment provided by the caller
1132  * \param[in] llh       llog handle we're processing
1133  * \param[in] rec       current llog record
1134  * \param[in] data      callback data containing a pointer to the device
1135  *
1136  * \retval 0                    to ask the caller (llog_process()) to continue
1137  * \retval LLOG_PROC_BREAK      to ask the caller to break
1138  */
1139 static int osp_sync_process_queues(const struct lu_env *env,
1140                                    struct llog_handle *llh,
1141                                    struct llog_rec_hdr *rec,
1142                                    void *data)
1143 {
1144         struct osp_device       *d = data;
1145
1146         do {
1147                 struct l_wait_info lwi = { 0 };
1148
1149                 if (!osp_sync_running(d)) {
1150                         CDEBUG(D_HA, "stop llog processing\n");
1151                         return LLOG_PROC_BREAK;
1152                 }
1153
1154                 /* process requests committed by OST */
1155                 osp_sync_process_committed(env, d);
1156
1157                 /* if we there are changes to be processed and we have
1158                  * resources for this ... do now */
1159                 if (osp_sync_can_process_new(d, rec)) {
1160                         if (llh == NULL) {
1161                                 /* ask llog for another record */
1162                                 CDEBUG(D_HA, "%u changes, %u in progress,"
1163                                      " %u in flight\n",
1164                                      atomic_read(&d->opd_sync_changes),
1165                                      atomic_read(&d->opd_sync_rpcs_in_progress),
1166                                      atomic_read(&d->opd_sync_rpcs_in_flight));
1167                                 return 0;
1168                         }
1169                         osp_sync_process_record(env, d, llh, rec);
1170                         llh = NULL;
1171                         rec = NULL;
1172                 }
1173
1174                 l_wait_event(d->opd_sync_waitq,
1175                              !osp_sync_running(d) ||
1176                              osp_sync_can_process_new(d, rec) ||
1177                              !list_empty(&d->opd_sync_committed_there),
1178                              &lwi);
1179         } while (1);
1180 }
1181
1182 /**
1183  * OSP sync thread.
1184  *
1185  * This thread runs llog_cat_process() scanner calling our callback
1186  * to process llog records. in the callback we implement tricky
1187  * state machine as we don't want to start scanning of the llog again
1188  * and again, also we don't want to process too many records and send
1189  * too many RPCs a time. so, depending on current load (num of changes
1190  * being synced to OST) the callback can suspend awaiting for some
1191  * new conditions, like syncs completed.
1192  *
1193  * In order to process llog records left by previous boots and to allow
1194  * llog_process_thread() to find something (otherwise it'd just exit
1195  * immediately) we add a special GENERATATION record on each boot.
1196  *
1197  * \param[in] _arg      a pointer to thread's arguments
1198  *
1199  * \retval 0            on success
1200  * \retval negative     negated errno on error
1201  */
1202 static int osp_sync_thread(void *_arg)
1203 {
1204         struct osp_device       *d = _arg;
1205         struct ptlrpc_thread    *thread = &d->opd_sync_thread;
1206         struct l_wait_info       lwi = { 0 };
1207         struct llog_ctxt        *ctxt;
1208         struct obd_device       *obd = d->opd_obd;
1209         struct llog_handle      *llh;
1210         struct lu_env            env;
1211         int                      rc, count;
1212         bool                     wrapped;
1213
1214         ENTRY;
1215
1216         rc = lu_env_init(&env, LCT_LOCAL);
1217         if (rc) {
1218                 CERROR("%s: can't initialize env: rc = %d\n",
1219                        obd->obd_name, rc);
1220
1221                 spin_lock(&d->opd_sync_lock);
1222                 thread->t_flags = SVC_STOPPED;
1223                 spin_unlock(&d->opd_sync_lock);
1224                 wake_up(&thread->t_ctl_waitq);
1225
1226                 RETURN(rc);
1227         }
1228
1229         spin_lock(&d->opd_sync_lock);
1230         thread->t_flags = SVC_RUNNING;
1231         spin_unlock(&d->opd_sync_lock);
1232         wake_up(&thread->t_ctl_waitq);
1233
1234         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1235         if (ctxt == NULL) {
1236                 CERROR("can't get appropriate context\n");
1237                 GOTO(out, rc = -EINVAL);
1238         }
1239
1240         llh = ctxt->loc_handle;
1241         if (llh == NULL) {
1242                 CERROR("can't get llh\n");
1243                 llog_ctxt_put(ctxt);
1244                 GOTO(out, rc = -EINVAL);
1245         }
1246
1247         /*
1248          * Catalog processing stops when it processed last catalog record
1249          * with index equal to the end of catalog bitmap. Or if it is wrapped,
1250          * processing stops with index equal to the lgh_last_idx. We need to
1251          * continue processing.
1252          */
1253         d->opd_sync_last_catalog_idx = 0;
1254         do {
1255                 int     size;
1256
1257                 wrapped = (llh->lgh_hdr->llh_cat_idx >= llh->lgh_last_idx &&
1258                            llh->lgh_hdr->llh_count > 1);
1259
1260                 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d,
1261                                       d->opd_sync_last_catalog_idx, 0);
1262
1263                 size = OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ?
1264                        cfs_fail_val : (LLOG_HDR_BITMAP_SIZE(llh->lgh_hdr) - 1);
1265                 /* processing reaches catalog bottom */
1266                 if (d->opd_sync_last_catalog_idx == size)
1267                         d->opd_sync_last_catalog_idx = LLOG_CAT_FIRST;
1268                 else if (wrapped)
1269                         /* If catalog is wrapped we can`t predict last index of
1270                          * processing because lgh_last_idx could be changed.
1271                          * Starting form the next one */
1272                         d->opd_sync_last_catalog_idx++;
1273
1274         } while (rc == 0 && (wrapped ||
1275                              d->opd_sync_last_catalog_idx == LLOG_CAT_FIRST));
1276
1277         if (rc < 0) {
1278                 CERROR("%s: llog process with osp_sync_process_queues "
1279                        "failed: %d\n", d->opd_obd->obd_name, rc);
1280                 GOTO(close, rc);
1281         }
1282         LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1283                  "%u changes, %u in progress, %u in flight: %d\n",
1284                  atomic_read(&d->opd_sync_changes),
1285                  atomic_read(&d->opd_sync_rpcs_in_progress),
1286                  atomic_read(&d->opd_sync_rpcs_in_flight), rc);
1287
1288         /* we don't expect llog_process_thread() to exit till umount */
1289         LASSERTF(thread->t_flags != SVC_RUNNING,
1290                  "%u changes, %u in progress, %u in flight\n",
1291                  atomic_read(&d->opd_sync_changes),
1292                  atomic_read(&d->opd_sync_rpcs_in_progress),
1293                  atomic_read(&d->opd_sync_rpcs_in_flight));
1294
1295         /* wait till all the requests are completed */
1296         count = 0;
1297         while (atomic_read(&d->opd_sync_rpcs_in_progress) > 0) {
1298                 osp_sync_process_committed(&env, d);
1299
1300                 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1301                 rc = l_wait_event(d->opd_sync_waitq,
1302                                 atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1303                                   &lwi);
1304                 if (rc == -ETIMEDOUT)
1305                         count++;
1306                 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1307                          d->opd_obd->obd_name,
1308                          atomic_read(&d->opd_sync_rpcs_in_progress),
1309                          atomic_read(&d->opd_sync_rpcs_in_flight),
1310                          list_empty(&d->opd_sync_committed_there) ? "" : "!");
1311
1312         }
1313
1314 close:
1315         llog_cat_close(&env, llh);
1316         rc = llog_cleanup(&env, ctxt);
1317         if (rc)
1318                 CERROR("can't cleanup llog: %d\n", rc);
1319 out:
1320         LASSERTF(atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1321                  "%s: %d %d %sempty\n", d->opd_obd->obd_name,
1322                  atomic_read(&d->opd_sync_rpcs_in_progress),
1323                  atomic_read(&d->opd_sync_rpcs_in_flight),
1324                  list_empty(&d->opd_sync_committed_there) ? "" : "!");
1325
1326         thread->t_flags = SVC_STOPPED;
1327
1328         wake_up(&thread->t_ctl_waitq);
1329
1330         lu_env_fini(&env);
1331
1332         RETURN(0);
1333 }
1334
1335 /**
1336  * Initialize llog.
1337  *
1338  * Initializes the llog. Specific llog to be used depends on the type of the
1339  * target OSP represents (OST or MDT). The function adds appends a new llog
1340  * record to mark the place where the records associated with this boot
1341  * start.
1342  *
1343  * \param[in] env       LU environment provided by the caller
1344  * \param[in] d         OSP device
1345  *
1346  * \retval 0            on success
1347  * \retval negative     negated errno on error
1348  */
1349 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1350 {
1351         struct osp_thread_info  *osi = osp_env_info(env);
1352         struct lu_fid           *fid = &osi->osi_fid;
1353         struct llog_handle      *lgh = NULL;
1354         struct obd_device       *obd = d->opd_obd;
1355         struct llog_ctxt        *ctxt;
1356         int                     rc;
1357
1358         ENTRY;
1359
1360         LASSERT(obd);
1361
1362         /*
1363          * open llog corresponding to our OST
1364          */
1365         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1366         obd->obd_lvfs_ctxt.dt = d->opd_storage;
1367
1368         lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1369
1370         rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1371                                    &osi->osi_cid, fid);
1372         if (rc < 0) {
1373                 if (rc != -EFAULT) {
1374                         CERROR("%s: can't get id from catalogs: rc = %d\n",
1375                                obd->obd_name, rc);
1376                         RETURN(rc);
1377                 }
1378
1379                 /* After sparse OST indices is supported, the CATALOG file
1380                  * may become a sparse file that results in failure on
1381                  * reading. Skip this error as the llog will be created
1382                  * later */
1383                 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1384                 rc = 0;
1385         }
1386
1387         CDEBUG(D_INFO, "%s: Init llog for %d - catid "DFID":%x\n",
1388                obd->obd_name, d->opd_index,
1389                PFID(&osi->osi_cid.lci_logid.lgl_oi.oi_fid),
1390                osi->osi_cid.lci_logid.lgl_ogen);
1391
1392         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT,
1393                         d->opd_storage->dd_lu_dev.ld_obd,
1394                         &llog_common_cat_ops);
1395         if (rc)
1396                 RETURN(rc);
1397
1398         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1399         LASSERT(ctxt);
1400
1401         if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1402                 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1403                                LLOG_OPEN_EXISTS);
1404                 /* re-create llog if it is missing */
1405                 if (rc == -ENOENT)
1406                         logid_set_id(&osi->osi_cid.lci_logid, 0);
1407                 else if (rc < 0)
1408                         GOTO(out_cleanup, rc);
1409         }
1410
1411         if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1412                 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1413                 if (rc < 0)
1414                         GOTO(out_cleanup, rc);
1415                 osi->osi_cid.lci_logid = lgh->lgh_id;
1416         }
1417
1418         LASSERT(lgh != NULL);
1419         ctxt->loc_handle = lgh;
1420
1421         rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
1422         if (rc)
1423                 GOTO(out_close, rc);
1424
1425         rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1426                                    &osi->osi_cid, fid);
1427         if (rc)
1428                 GOTO(out_close, rc);
1429
1430         /*
1431          * put a mark in the llog till which we'll be processing
1432          * old records restless
1433          */
1434         d->opd_sync_generation.mnt_cnt = ktime_get_ns();
1435         d->opd_sync_generation.conn_cnt = ktime_get_ns();
1436
1437         osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1438         osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1439
1440         memcpy(&osi->osi_gen.lgr_gen, &d->opd_sync_generation,
1441                sizeof(osi->osi_gen.lgr_gen));
1442
1443         rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1444         if (rc < 0)
1445                 GOTO(out_close, rc);
1446         llog_ctxt_put(ctxt);
1447         RETURN(0);
1448 out_close:
1449         llog_cat_close(env, lgh);
1450 out_cleanup:
1451         llog_cleanup(env, ctxt);
1452         RETURN(rc);
1453 }
1454
1455 /**
1456  * Cleanup llog used for syncing.
1457  *
1458  * Closes and cleanups the llog. The function is called when the device is
1459  * shutting down.
1460  *
1461  * \param[in] env       LU environment provided by the caller
1462  * \param[in] d         OSP device
1463  */
1464 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1465 {
1466         struct llog_ctxt *ctxt;
1467
1468         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1469         if (ctxt) {
1470                 llog_cat_close(env, ctxt->loc_handle);
1471                 llog_cleanup(env, ctxt);
1472         }
1473 }
1474
1475 /**
1476  * Initialization of the sync component of OSP.
1477  *
1478  * Initializes the llog and starts a new thread to handle the changes to
1479  * the remote target (OST or MDT).
1480  *
1481  * \param[in] env       LU environment provided by the caller
1482  * \param[in] d         OSP device
1483  *
1484  * \retval 0            on success
1485  * \retval negative     negated errno on error
1486  */
1487 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1488 {
1489         struct l_wait_info       lwi = { 0 };
1490         struct task_struct      *task;
1491         int                      rc;
1492
1493         ENTRY;
1494
1495         d->opd_sync_max_rpcs_in_flight = OSP_MAX_RPCS_IN_FLIGHT;
1496         d->opd_sync_max_rpcs_in_progress = OSP_MAX_RPCS_IN_PROGRESS;
1497         spin_lock_init(&d->opd_sync_lock);
1498         init_waitqueue_head(&d->opd_sync_waitq);
1499         init_waitqueue_head(&d->opd_sync_barrier_waitq);
1500         thread_set_flags(&d->opd_sync_thread, SVC_INIT);
1501         init_waitqueue_head(&d->opd_sync_thread.t_ctl_waitq);
1502         INIT_LIST_HEAD(&d->opd_sync_in_flight_list);
1503         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1504
1505         if (d->opd_storage->dd_rdonly)
1506                 RETURN(0);
1507
1508         /*
1509          * initialize llog storing changes
1510          */
1511         rc = osp_sync_llog_init(env, d);
1512         if (rc) {
1513                 CERROR("%s: can't initialize llog: rc = %d\n",
1514                        d->opd_obd->obd_name, rc);
1515                 GOTO(err_id, rc);
1516         }
1517
1518         /*
1519          * Start synchronization thread
1520          */
1521         task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1522                            d->opd_index, d->opd_group);
1523         if (IS_ERR(task)) {
1524                 rc = PTR_ERR(task);
1525                 CERROR("%s: cannot start sync thread: rc = %d\n",
1526                        d->opd_obd->obd_name, rc);
1527                 GOTO(err_llog, rc);
1528         }
1529
1530         l_wait_event(d->opd_sync_thread.t_ctl_waitq,
1531                      osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1532
1533         RETURN(0);
1534 err_llog:
1535         osp_sync_llog_fini(env, d);
1536 err_id:
1537         return rc;
1538 }
1539
1540 /**
1541  * Stop the syncing thread.
1542  *
1543  * Asks the syncing thread to stop and wait until it's stopped.
1544  *
1545  * \param[in] d         OSP device
1546  *
1547  * \retval              0
1548  */
1549 int osp_sync_fini(struct osp_device *d)
1550 {
1551         struct ptlrpc_thread *thread = &d->opd_sync_thread;
1552
1553         ENTRY;
1554
1555         if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1556                 thread->t_flags = SVC_STOPPING;
1557                 wake_up(&d->opd_sync_waitq);
1558                 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1559         }
1560
1561         RETURN(0);
1562 }
1563
1564 struct osp_last_committed_cb {
1565         struct dt_txn_commit_cb  ospc_cb;
1566         struct osp_device       *ospc_dev;
1567         __u64                    ospc_transno;
1568 };
1569
1570 void osp_sync_local_commit_cb(struct lu_env *env, struct thandle *th,
1571                               struct dt_txn_commit_cb *dcb, int err)
1572 {
1573         struct osp_last_committed_cb    *cb;
1574         struct osp_device               *d;
1575
1576         cb = container_of0(dcb, struct osp_last_committed_cb, ospc_cb);
1577         d = cb->ospc_dev;
1578
1579         CDEBUG(D_HA, "%s: %llu committed\n", d->opd_obd->obd_name,
1580                cb->ospc_transno);
1581
1582         spin_lock(&d->opd_sync_lock);
1583         if (cb->ospc_transno > d->opd_sync_last_committed_id)
1584                 d->opd_sync_last_committed_id = cb->ospc_transno;
1585         spin_unlock(&d->opd_sync_lock);
1586
1587         osp_sync_check_for_work(d);
1588         lu_device_put(osp2lu_dev(d));
1589         if (atomic_dec_and_test(&d->opd_commits_registered))
1590                 wake_up(&d->opd_sync_waitq);
1591
1592         OBD_FREE_PTR(cb);
1593 }
1594
1595 static int osp_sync_add_commit_cb(const struct lu_env *env,
1596                                   struct osp_device *d, struct thandle *th)
1597 {
1598         struct osp_last_committed_cb    *cb;
1599         struct dt_txn_commit_cb         *dcb;
1600         int                              rc = 0;
1601
1602         OBD_ALLOC_PTR(cb);
1603         if (cb == NULL)
1604                 return -ENOMEM;
1605         cb->ospc_dev = d;
1606         dcb = &cb->ospc_cb;
1607         dcb->dcb_func = osp_sync_local_commit_cb;
1608         spin_lock(&d->opd_sync_lock);
1609         cb->ospc_transno = ++d->opd_sync_last_used_id;
1610         spin_unlock(&d->opd_sync_lock);
1611
1612         rc = dt_trans_cb_add(th, dcb);
1613         CDEBUG(D_HA, "%s: add commit cb at %lluns, next at %lluns, rc = %d\n",
1614                d->opd_obd->obd_name, ktime_get_ns(),
1615                ktime_to_ns(d->opd_sync_next_commit_cb), rc);
1616
1617         if (likely(rc == 0)) {
1618                 lu_device_get(osp2lu_dev(d));
1619                 atomic_inc(&d->opd_commits_registered);
1620         } else
1621                 OBD_FREE_PTR(cb);
1622
1623         return rc;
1624 }
1625
1626 /* add the commit callback every second */
1627 int osp_sync_add_commit_cb_1s(const struct lu_env *env, struct osp_device *d,
1628                               struct thandle *th)
1629 {
1630         ktime_t now = ktime_get();
1631         bool add = false;
1632
1633         /* fast path */
1634         if (ktime_before(now, d->opd_sync_next_commit_cb))
1635                 return 0;
1636
1637         spin_lock(&d->opd_sync_lock);
1638         if (ktime_before(d->opd_sync_next_commit_cb, now)) {
1639                 add = true;
1640                 d->opd_sync_next_commit_cb = ktime_add_ns(now, NSEC_PER_SEC);
1641         }
1642         spin_unlock(&d->opd_sync_lock);
1643
1644         if (!add)
1645                 return 0;
1646
1647         return osp_sync_add_commit_cb(env, d, th);
1648 }
1649
1650 /*
1651  * generate an empty transaction and hook the commit callback in
1652  * then force transaction commit
1653  */
1654 void osp_sync_force(const struct lu_env *env, struct osp_device *d)
1655 {
1656         struct thandle *th;
1657         int rc;
1658
1659         th = dt_trans_create(env, d->opd_storage);
1660         if (IS_ERR(th)) {
1661                 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
1662                 return;
1663         }
1664         rc = dt_trans_start_local(env, d->opd_storage, th);
1665         if (rc == 0) {
1666                 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
1667                        d->opd_obd->obd_name,
1668                        atomic_read(&d->opd_sync_changes));
1669                 rc = osp_sync_add_commit_cb(env, d, th);
1670                 dt_trans_stop(env, d->opd_storage, th);
1671         }
1672
1673         dt_commit_async(env, d->opd_storage);
1674 }