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         if (!ctxt) {
 347                 /* for a reason OSP wasn't able to open llog,
 348                  * just skip logging this operation and hope
 349                  * LFSCK will fix it eventually */
 350                 CERROR("logging isn't available, run LFSCK\n");
 351                 RETURN(0);
 352         }
 353
 354         rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
 355                               storage_th);
 356         llog_ctxt_put(ctxt);
 357
 358         RETURN(rc);
 359 }
 360
 361 /**
 362  * Generate a llog record for a given change.
 363  *
 364  * Generates a llog record for the change passed. The change can be of two
 365  * types: unlink and setattr. The record gets an ID which later will be
 366  * used to track commit status of the change. For unlink changes, the caller
 367  * can supply a starting FID and the count of the objects to destroy. For
 368  * setattr the caller should apply attributes to apply.
 369  *
 370  *
 371  * \param[in] env       LU environment provided by the caller
 372  * \param[in] d         OSP device
 373  * \param[in] fid       fid of the object the change should be applied to
 374  * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 375  * \param[in] count     count of objects to destroy
 376  * \param[in] th        transaction handle (local)
 377  * \param[in] attr      attributes for setattr
 378  *
 379  * \retval 0            on success
 380  * \retval negative     negated errno on error
 381  */
 382 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
 383                             const struct lu_fid *fid, enum llog_op_type type,
 384                             int count, struct thandle *th,
 385                             const struct lu_attr *attr)
 386 {
 387         struct osp_thread_info  *osi = osp_env_info(env);
 388         struct llog_ctxt        *ctxt;
 389         struct thandle          *storage_th;
 390         bool                     immediate_commit_cb = false;
 391         int                      rc;
 392
 393         ENTRY;
 394
 395         /* it's a layering violation, to access internals of th,
 396          * but we can do this as a sanity check, for a while */
 397         LASSERT(th->th_top != NULL);
 398         storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
 399         if (IS_ERR(storage_th))
 400                 RETURN(PTR_ERR(storage_th));
 401
 402         switch (type) {
 403         case MDS_UNLINK64_REC:
 404                 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
 405                 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
 406                 osi->osi_unlink.lur_fid  = *fid;
 407                 osi->osi_unlink.lur_count = count;
 408                 break;
 409         case MDS_SETATTR64_REC:
 410                 rc = fid_to_ostid(fid, &osi->osi_oi);
 411                 LASSERT(rc == 0);
 412                 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
 413                 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
 414                 osi->osi_setattr.lsr_oi  = osi->osi_oi;
 415                 LASSERT(attr);
 416                 osi->osi_setattr.lsr_uid = attr->la_uid;
 417                 osi->osi_setattr.lsr_gid = attr->la_gid;
 418                 osi->osi_setattr.lsr_layout_version = attr->la_layout_version;
 419                 osi->osi_setattr.lsr_projid = attr->la_projid;
 420                 osi->osi_setattr.lsr_valid =
 421                         ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
 422                         ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0) |
 423                         ((attr->la_valid & LA_PROJID) ? OBD_MD_FLPROJID : 0);
 424                 if (attr->la_valid & LA_LAYOUT_VERSION) {
 425                         osi->osi_setattr.lsr_valid |= OBD_MD_LAYOUT_VERSION;
 426
 427                         /* FLR: the layout version has to be transferred to
 428                          * OST objects ASAP, otherwise clients will have to
 429                          * experience delay to be able to write OST objects. */
 430                         immediate_commit_cb = true;
 431                 }
 432                 break;
 433         default:
 434                 LBUG();
 435         }
 436
 437         /* we keep the same id, but increment it when the callback
 438          * is registered, so that all records upto the one taken
 439          * by the callback are subject to processing */
 440         spin_lock(&d->opd_sync_lock);
 441         osi->osi_hdr.lrh_id = d->opd_sync_last_used_id;
 442         spin_unlock(&d->opd_sync_lock);
 443
 444         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 445         if (ctxt == NULL) {
 446                 /* see comment in osp_sync_declare_add() */
 447                 RETURN(0);
 448         }
 449
 450         rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
 451                       storage_th);
 452         llog_ctxt_put(ctxt);
 453
 454         if (likely(rc >= 0)) {
 455                 CDEBUG(D_OTHER, "%s: new record "DFID":%x.%u: rc = %d\n",
 456                        d->opd_obd->obd_name,
 457                        PFID(&osi->osi_cookie.lgc_lgl.lgl_oi.oi_fid),
 458                        osi->osi_cookie.lgc_lgl.lgl_ogen,
 459                        osi->osi_cookie.lgc_index, rc);
 460                 atomic_inc(&d->opd_sync_changes);
 461         }
 462
 463         if (immediate_commit_cb)
 464                 rc = osp_sync_add_commit_cb(env, d, th);
 465         else
 466                 rc = osp_sync_add_commit_cb_1s(env, d, th);
 467
 468         /* return 0 always here, error case just cause no llog record */
 469         RETURN(0);
 470 }
 471
 472 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
 473                  enum llog_op_type type, struct thandle *th,
 474                  const struct lu_attr *attr)
 475 {
 476         return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
 477                                 lu_object_fid(&o->opo_obj.do_lu), type, 1,
 478                                 th, attr);
 479 }
 480
 481 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
 482                         struct lu_fid *fid, int lost, struct thandle *th)
 483 {
 484         return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
 485 }
 486
 487 /*
 488  * it's quite obvious we can't maintain all the structures in the memory:
 489  * while OST is down, MDS can be processing thousands and thousands of unlinks
 490  * filling persistent llogs and in-core respresentation
 491  *
 492  * this doesn't scale at all. so we need basically the following:
 493  * a) destroy/setattr append llog records
 494  * b) once llog has grown to X records, we process first Y committed records
 495  *
 496  *  once record R is found via llog_process(), it becomes committed after any
 497  *  subsequent commit callback (at the most)
 498  */
 499
 500 /**
 501  * ptlrpc commit callback.
 502  *
 503  * The callback is called by PTLRPC when a RPC is reported committed by the
 504  * target (OST). We register the callback for the every RPC applying a change
 505  * from the llog. This way we know then the llog records can be cancelled.
 506  * Notice the callback can be called when OSP is finishing. We can detect this
 507  * checking that actual transno in the request is less or equal of known
 508  * committed transno (see osp_sync_process_committed() for the details).
 509  * XXX: this is pretty expensive and can be improved later using batching.
 510  *
 511  * \param[in] req       request
 512  */
 513 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
 514 {
 515         struct osp_device *d = req->rq_cb_data;
 516         struct osp_job_req_args *jra;
 517
 518         CDEBUG(D_HA, "commit req %p, transno %llu\n", req, req->rq_transno);
 519
 520         if (unlikely(req->rq_transno == 0))
 521                 return;
 522
 523         /* do not do any opd_sync_rpcs_* accounting here
 524          * it's done in osp_sync_interpret sooner or later */
 525         LASSERT(d);
 526
 527         jra = ptlrpc_req_async_args(jra, req);
 528         LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
 529         LASSERT(list_empty(&jra->jra_committed_link));
 530
 531         ptlrpc_request_addref(req);
 532
 533         spin_lock(&d->opd_sync_lock);
 534         list_add(&jra->jra_committed_link, &d->opd_sync_committed_there);
 535         spin_unlock(&d->opd_sync_lock);
 536
 537         /* XXX: some batching wouldn't hurt */
 538         wake_up(&d->opd_sync_waitq);
 539 }
 540
 541 /**
 542  * RPC interpretation callback.
 543  *
 544  * The callback is called by ptlrpc when RPC is replied. Now we have to decide
 545  * whether we should:
 546  *  - put request on a special list to wait until it's committed by the target,
 547  *    if the request is successful
 548  *  - schedule llog record cancel if no target object is found
 549  *  - try later (essentially after reboot) in case of unexpected error
 550  *
 551  * \param[in] env       LU environment provided by the caller
 552  * \param[in] req       request replied
 553  * \param[in] aa        callback data
 554  * \param[in] rc        result of RPC
 555  *
 556  * \retval 0            always
 557  */
 558 static int osp_sync_interpret(const struct lu_env *env,
 559                               struct ptlrpc_request *req, void *args, int rc)
 560 {
 561         struct osp_job_req_args *jra = args;
 562         struct osp_device *d = req->rq_cb_data;
 563
 564         if (jra->jra_magic != OSP_JOB_MAGIC) {
 565                 DEBUG_REQ(D_ERROR, req, "bad magic %u", jra->jra_magic);
 566                 LBUG();
 567         }
 568         LASSERT(d);
 569
 570         CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
 571                atomic_read(&req->rq_refcount),
 572                rc, (unsigned) req->rq_transno);
 573
 574         if (rc == -ENOENT) {
 575                 /*
 576                  * we tried to destroy object or update attributes,
 577                  * but object doesn't exist anymore - cancell llog record
 578                  */
 579                 LASSERT(req->rq_transno == 0);
 580                 LASSERT(list_empty(&jra->jra_committed_link));
 581
 582                 ptlrpc_request_addref(req);
 583
 584                 spin_lock(&d->opd_sync_lock);
 585                 list_add(&jra->jra_committed_link,
 586                          &d->opd_sync_committed_there);
 587                 spin_unlock(&d->opd_sync_lock);
 588
 589                 wake_up(&d->opd_sync_waitq);
 590         } else if (rc) {
 591                 struct obd_import *imp = req->rq_import;
 592                 /*
 593                  * error happened, we'll try to repeat on next boot ?
 594                  */
 595                 LASSERTF(req->rq_transno == 0 || rc == -EIO ||
 596                          req->rq_import_generation < imp->imp_generation,
 597                          "transno %llu, rc %d, gen: req %d, imp %d\n",
 598                          req->rq_transno, rc, req->rq_import_generation,
 599                          imp->imp_generation);
 600                 if (req->rq_transno == 0) {
 601                         /* this is the last time we see the request
 602                          * if transno is not zero, then commit cb
 603                          * will be called at some point */
 604                         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) > 0);
 605                         atomic_dec(&d->opd_sync_rpcs_in_progress);
 606                 }
 607
 608                 wake_up(&d->opd_sync_waitq);
 609         } else if (d->opd_pre != NULL &&
 610                    unlikely(d->opd_pre_status == -ENOSPC)) {
 611                 /*
 612                  * if current status is -ENOSPC (lack of free space on OST)
 613                  * then we should poll OST immediately once object destroy
 614                  * is replied
 615                  */
 616                 osp_statfs_need_now(d);
 617         }
 618
 619         spin_lock(&d->opd_sync_lock);
 620         list_del_init(&jra->jra_in_flight_link);
 621         spin_unlock(&d->opd_sync_lock);
 622         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) > 0);
 623         atomic_dec(&d->opd_sync_rpcs_in_flight);
 624         if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
 625                 wake_up(&d->opd_sync_barrier_waitq);
 626         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 627                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 628                atomic_read(&d->opd_sync_rpcs_in_progress));
 629
 630         osp_sync_check_for_work(d);
 631
 632         return 0;
 633 }
 634
 635 /*
 636  ** Add request to ptlrpc queue.
 637  *
 638  * This is just a tiny helper function to put the request on the sending list
 639  *
 640  * \param[in] d         OSP device
 641  * \param[in] llh       llog handle where the record is stored
 642  * \param[in] h         llog record
 643  * \param[in] req       request
 644  */
 645 static void osp_sync_send_new_rpc(struct osp_device *d,
 646                                   struct llog_handle *llh,
 647                                   struct llog_rec_hdr *h,
 648                                   struct ptlrpc_request *req)
 649 {
 650         struct osp_job_req_args *jra;
 651
 652         LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) <=
 653                 d->opd_sync_max_rpcs_in_flight);
 654
 655         jra = ptlrpc_req_async_args(jra, req);
 656         jra->jra_magic = OSP_JOB_MAGIC;
 657         jra->jra_lcookie.lgc_lgl = llh->lgh_id;
 658         jra->jra_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 659         jra->jra_lcookie.lgc_index = h->lrh_index;
 660         INIT_LIST_HEAD(&jra->jra_committed_link);
 661         spin_lock(&d->opd_sync_lock);
 662         list_add_tail(&jra->jra_in_flight_link, &d->opd_sync_in_flight_list);
 663         spin_unlock(&d->opd_sync_lock);
 664
 665         ptlrpcd_add_req(req);
 666 }
 667
 668
 669 /**
 670  * Allocate and prepare RPC for a new change.
 671  *
 672  * The function allocates and initializes an RPC which will be sent soon to
 673  * apply the change to the target OST. The request is initialized from the
 674  * llog record passed. Notice only the fields common to all type of changes
 675  * are initialized.
 676  *
 677  * \param[in] d         OSP device
 678  * \param[in] op        type of the change
 679  * \param[in] format    request format to be used
 680  *
 681  * \retval pointer              new request on success
 682  * \retval ERR_PTR(errno)       on error
 683  */
 684 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
 685                                                enum ost_cmd op,
 686                                                const struct req_format *format)
 687 {
 688         struct ptlrpc_request   *req;
 689         struct obd_import       *imp;
 690         int                      rc;
 691
 692         /* Prepare the request */
 693         imp = d->opd_obd->u.cli.cl_import;
 694         LASSERT(imp);
 695
 696         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
 697                 RETURN(ERR_PTR(-ENOMEM));
 698
 699         req = ptlrpc_request_alloc(imp, format);
 700         if (req == NULL)
 701                 RETURN(ERR_PTR(-ENOMEM));
 702
 703         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
 704         if (rc) {
 705                 ptlrpc_req_finished(req);
 706                 return ERR_PTR(rc);
 707         }
 708
 709         req->rq_interpret_reply = osp_sync_interpret;
 710         req->rq_commit_cb = osp_sync_request_commit_cb;
 711         req->rq_cb_data = d;
 712
 713         ptlrpc_request_set_replen(req);
 714
 715         return req;
 716 }
 717
 718 /**
 719  * Generate a request for setattr change.
 720  *
 721  * The function prepares a new RPC, initializes it with setattr specific
 722  * bits and send the RPC.
 723  *
 724  * \param[in] d         OSP device
 725  * \param[in] llh       llog handle where the record is stored
 726  * \param[in] h         llog record
 727  *
 728  * \retval 0            on success
 729  * \retval 1            on invalid record
 730  * \retval negative     negated errno on error
 731  */
 732 static int osp_sync_new_setattr_job(struct osp_device *d,
 733                                     struct llog_handle *llh,
 734                                     struct llog_rec_hdr *h)
 735 {
 736         struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
 737         struct ptlrpc_request           *req;
 738         struct ost_body                 *body;
 739
 740         ENTRY;
 741         LASSERT(h->lrh_type == MDS_SETATTR64_REC);
 742
 743         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
 744                 RETURN(1);
 745
 746         /* lsr_valid can only be 0 or HAVE OBD_MD_{FLUID, FLGID, FLPROJID} set,
 747          * so no bits other than these should be set. */
 748         if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID |
 749             OBD_MD_FLPROJID | OBD_MD_LAYOUT_VERSION)) != 0) {
 750                 CERROR("%s: invalid setattr record, lsr_valid:%llu\n",
 751                         d->opd_obd->obd_name, rec->lsr_valid);
 752                 /* return 1 on invalid record */
 753                 RETURN(1);
 754         }
 755
 756         req = osp_sync_new_job(d, OST_SETATTR, &RQF_OST_SETATTR);
 757         if (IS_ERR(req))
 758                 RETURN(PTR_ERR(req));
 759
 760         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 761         LASSERT(body);
 762         body->oa.o_oi = rec->lsr_oi;
 763         body->oa.o_uid = rec->lsr_uid;
 764         body->oa.o_gid = rec->lsr_gid;
 765         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 766         if (h->lrh_len > sizeof(struct llog_setattr64_rec)) {
 767                 struct llog_setattr64_rec_v2 *rec_v2 = (typeof(rec_v2))rec;
 768                 body->oa.o_projid = rec_v2->lsr_projid;
 769                 body->oa.o_layout_version = rec_v2->lsr_layout_version;
 770         }
 771
 772         /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
 773          * we assume that both UID and GID are valid in that case. */
 774         if (rec->lsr_valid == 0)
 775                 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
 776         else
 777                 body->oa.o_valid |= rec->lsr_valid;
 778
 779         if (body->oa.o_valid & OBD_MD_LAYOUT_VERSION) {
 780                 OBD_FAIL_TIMEOUT(OBD_FAIL_FLR_LV_DELAY, cfs_fail_val);
 781                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_FLR_LV_INC)))
 782                         ++body->oa.o_layout_version;
 783         }
 784
 785         osp_sync_send_new_rpc(d, llh, h, req);
 786         RETURN(0);
 787 }
 788
 789 /**
 790  * Generate a request for unlink change.
 791  *
 792  * The function prepares a new RPC, initializes it with unlink(destroy)
 793  * specific bits and sends the RPC. The function is used to handle
 794  * llog_unlink_rec which were used in the older versions of Lustre.
 795  * Current version uses llog_unlink_rec64.
 796  *
 797  * \param[in] d         OSP device
 798  * \param[in] llh       llog handle where the record is stored
 799  * \param[in] h         llog record
 800  *
 801  * \retval 0            on success
 802  * \retval negative     negated errno on error
 803  */
 804 static int osp_sync_new_unlink_job(struct osp_device *d,
 805                                    struct llog_handle *llh,
 806                                    struct llog_rec_hdr *h)
 807 {
 808         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 809         struct ptlrpc_request   *req;
 810         struct ost_body         *body;
 811         int rc;
 812
 813         ENTRY;
 814         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 815
 816         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 817         if (IS_ERR(req))
 818                 RETURN(PTR_ERR(req));
 819
 820         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 821         LASSERT(body);
 822         ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
 823         rc = ostid_set_id(&body->oa.o_oi, rec->lur_oid);
 824         if (rc)
 825                 return rc;
 826         body->oa.o_misc = rec->lur_count;
 827         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 828         if (rec->lur_count)
 829                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 830
 831         osp_sync_send_new_rpc(d, llh, h, req);
 832         RETURN(0);
 833 }
 834
 835 /**
 836  * Generate a request for unlink change.
 837  *
 838  * The function prepares a new RPC, initializes it with unlink(destroy)
 839  * specific bits and sends the RPC. Depending on the target (MDT or OST)
 840  * two different protocols are used. For MDT we use OUT (basically OSD API
 841  * updates transferred via a network). For OST we still use the old
 842  * protocol (OBD?), originally for compatibility. Later we can start to
 843  * use OUT for OST as well, this will allow batching and better code
 844  * unification.
 845  *
 846  * \param[in] d         OSP device
 847  * \param[in] llh       llog handle where the record is stored
 848  * \param[in] h         llog record
 849  *
 850  * \retval 0            on success
 851  * \retval negative     negated errno on error
 852  */
 853 static int osp_sync_new_unlink64_job(struct osp_device *d,
 854                                      struct llog_handle *llh,
 855                                      struct llog_rec_hdr *h)
 856 {
 857         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 858         struct ptlrpc_request           *req = NULL;
 859         struct ost_body                 *body;
 860         int                              rc;
 861
 862         ENTRY;
 863         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 864         req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
 865         if (IS_ERR(req))
 866                 RETURN(PTR_ERR(req));
 867
 868         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 869         if (body == NULL)
 870                 RETURN(-EFAULT);
 871         rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
 872         if (rc < 0)
 873                 RETURN(rc);
 874         body->oa.o_misc = rec->lur_count;
 875         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 876                            OBD_MD_FLOBJCOUNT;
 877         osp_sync_send_new_rpc(d, llh, h, req);
 878         RETURN(0);
 879 }
 880
 881 /**
 882  * Process llog records.
 883  *
 884  * This function is called to process the llog records committed locally.
 885  * In the recovery model used by OSP we can apply a change to a remote
 886  * target once corresponding transaction (like posix unlink) is committed
 887  * locally so can't revert.
 888  * Depending on the llog record type, a given handler is called that is
 889  * responsible for preparing and sending the RPC to apply the change.
 890  * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
 891  *
 892  * \param[in] env       LU environment provided by the caller
 893  * \param[in] d         OSP device
 894  * \param[in] llh       llog handle where the record is stored
 895  * \param[in] rec       llog record
 896  */
 897 static void osp_sync_process_record(const struct lu_env *env,
 898                                     struct osp_device *d,
 899                                     struct llog_handle *llh,
 900                                     struct llog_rec_hdr *rec)
 901 {
 902         struct llog_handle      *cathandle = llh->u.phd.phd_cat_handle;
 903         struct llog_cookie       cookie;
 904         int                      rc = 0;
 905
 906         ENTRY;
 907
 908         cookie.lgc_lgl = llh->lgh_id;
 909         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 910         cookie.lgc_index = rec->lrh_index;
 911
 912         d->opd_sync_last_catalog_idx = llh->lgh_hdr->llh_cat_idx;
 913
 914         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 915                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 916
 917                 /* we're waiting for the record generated by this instance */
 918                 LASSERT(d->opd_sync_prev_done == 0);
 919                 if (!memcmp(&d->opd_sync_generation, &gen->lgr_gen,
 920                             sizeof(gen->lgr_gen))) {
 921                         CDEBUG(D_HA, "processed all old entries\n");
 922                         d->opd_sync_prev_done = 1;
 923                 }
 924
 925                 /* cancel any generation record */
 926                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 927
 928                 /* flush all pending records ASAP */
 929                 osp_sync_force(env, d);
 930
 931                 RETURN_EXIT;
 932         }
 933
 934         /*
 935          * now we prepare and fill requests to OST, put them on the queue
 936          * and fire after next commit callback
 937          */
 938
 939         /* notice we increment counters before sending RPC, to be consistent
 940          * in RPC interpret callback which may happen very quickly */
 941         atomic_inc(&d->opd_sync_rpcs_in_flight);
 942         atomic_inc(&d->opd_sync_rpcs_in_progress);
 943
 944         switch (rec->lrh_type) {
 945         /* case MDS_UNLINK_REC is kept for compatibility */
 946         case MDS_UNLINK_REC:
 947                 rc = osp_sync_new_unlink_job(d, llh, rec);
 948                 break;
 949         case MDS_UNLINK64_REC:
 950                 rc = osp_sync_new_unlink64_job(d, llh, rec);
 951                 break;
 952         case MDS_SETATTR64_REC:
 953                 rc = osp_sync_new_setattr_job(d, llh, rec);
 954                 break;
 955         default:
 956                 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
 957                        rec->lrh_type);
 958                 /* treat "unknown record type" as "invalid" */
 959                 rc = 1;
 960                 break;
 961         }
 962
 963         /* For all kinds of records, not matter successful or not,
 964          * we should decrease changes and bump last_processed_id.
 965          */
 966         if (d->opd_sync_prev_done) {
 967                 LASSERT(atomic_read(&d->opd_sync_changes) > 0);
 968                 atomic_dec(&d->opd_sync_changes);
 969                 wake_up(&d->opd_sync_barrier_waitq);
 970         }
 971         atomic64_inc(&d->opd_sync_processed_recs);
 972         if (rc != 0) {
 973                 atomic_dec(&d->opd_sync_rpcs_in_flight);
 974                 atomic_dec(&d->opd_sync_rpcs_in_progress);
 975         }
 976
 977         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 978                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
 979                atomic_read(&d->opd_sync_rpcs_in_progress));
 980
 981         /* Delete the invalid record */
 982         if (rc == 1) {
 983                 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
 984                 if (rc != 0)
 985                         CERROR("%s: can't delete invalid record: "
 986                                "fid = "DFID", rec_id = %u, rc = %d\n",
 987                                d->opd_obd->obd_name,
 988                                PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
 989                                rec->lrh_id, rc);
 990         }
 991
 992         CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
 993                rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);
 994
 995         RETURN_EXIT;
 996 }
 997
 998 /**
 999  * Cancel llog records for the committed changes.
1000  *
1001  * The function walks through the list of the committed RPCs and cancels
1002  * corresponding llog records. see osp_sync_request_commit_cb() for the
1003  * details.
1004  *
1005  * \param[in] env       LU environment provided by the caller
1006  * \param[in] d         OSP device
1007  */
1008 static void osp_sync_process_committed(const struct lu_env *env,
1009                                        struct osp_device *d)
1010 {
1011         struct obd_device       *obd = d->opd_obd;
1012         struct obd_import       *imp = obd->u.cli.cl_import;
1013         struct ost_body         *body;
1014         struct ptlrpc_request   *req;
1015         struct llog_ctxt        *ctxt;
1016         struct llog_handle      *llh;
1017         int                     *arr;
1018         struct list_head         list, *le;
1019         struct llog_logid        lgid;
1020         int                      rc, i, count = 0, done = 0;
1021
1022         ENTRY;
1023
1024         if (list_empty(&d->opd_sync_committed_there))
1025                 return;
1026
1027         /*
1028          * if current status is -ENOSPC (lack of free space on OST)
1029          * then we should poll OST immediately once object destroy
1030          * is committed.
1031          * notice: we do this upon commit as well because some backends
1032          * (like DMU) do not release space right away.
1033          */
1034         if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
1035                 osp_statfs_need_now(d);
1036
1037         /*
1038          * now cancel them all
1039          * XXX: can we improve this using some batching?
1040          *      with batch RPC that'll happen automatically?
1041          * XXX: can we store ctxt in lod_device and save few cycles ?
1042          */
1043         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1044         LASSERT(ctxt);
1045
1046         llh = ctxt->loc_handle;
1047         LASSERT(llh);
1048
1049         INIT_LIST_HEAD(&list);
1050         spin_lock(&d->opd_sync_lock);
1051         list_splice(&d->opd_sync_committed_there, &list);
1052         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1053         spin_unlock(&d->opd_sync_lock);
1054
1055         list_for_each(le, &list)
1056                 count++;
1057         if (count > 2)
1058                 OBD_ALLOC_WAIT(arr, sizeof(int) * count);
1059         else
1060                 arr = NULL;
1061         i = 0;
1062         while (!list_empty(&list)) {
1063                 struct osp_job_req_args *jra;
1064
1065                 jra = list_entry(list.next, struct osp_job_req_args,
1066                                  jra_committed_link);
1067                 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1068                 list_del_init(&jra->jra_committed_link);
1069
1070                 req = container_of((void *)jra, struct ptlrpc_request,
1071                                    rq_async_args);
1072                 body = req_capsule_client_get(&req->rq_pill,
1073                                               &RMF_OST_BODY);
1074                 LASSERT(body);
1075                 /* import can be closing, thus all commit cb's are
1076                  * called we can check committness directly */
1077                 if (req->rq_import_generation == imp->imp_generation) {
1078                         if (arr && (!i ||
1079                                     !memcmp(&jra->jra_lcookie.lgc_lgl, &lgid,
1080                                            sizeof(lgid)))) {
1081                                 if (unlikely(!i))
1082                                         lgid = jra->jra_lcookie.lgc_lgl;
1083
1084                                 arr[i++] = jra->jra_lcookie.lgc_index;
1085                         } else {
1086                                 rc = llog_cat_cancel_records(env, llh, 1,
1087                                                              &jra->jra_lcookie);
1088                                 if (rc)
1089                                         CERROR("%s: can't cancel record: %d\n",
1090                                                obd->obd_name, rc);
1091                         }
1092                 } else {
1093                         DEBUG_REQ(D_OTHER, req, "imp_committed = %llu",
1094                                   imp->imp_peer_committed_transno);
1095                 }
1096                 ptlrpc_req_finished(req);
1097                 done++;
1098         }
1099         if (arr && i > 0) {
1100                 rc = llog_cat_cancel_arr_rec(env, llh, &lgid, i, arr);
1101
1102                 if (rc)
1103                         CERROR("%s: can't cancel %d records rc: %d\n",
1104                                obd->obd_name, i, rc);
1105                 else
1106                         CDEBUG(D_OTHER, "%s: massive records cancel id "DFID\
1107                                " num %d\n", obd->obd_name,
1108                                PFID(&lgid.lgl_oi.oi_fid), i);
1109         }
1110         if (arr)
1111                 OBD_FREE(arr, sizeof(int) * count);
1112
1113         llog_ctxt_put(ctxt);
1114
1115         LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) >= done);
1116         atomic_sub(done, &d->opd_sync_rpcs_in_progress);
1117         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress, done %d\n",
1118                d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
1119                atomic_read(&d->opd_sync_rpcs_in_progress), done);
1120
1121         osp_sync_check_for_work(d);
1122
1123         /* wake up the thread if requested to stop:
1124          * it might be waiting for in-progress to complete */
1125         if (unlikely(osp_sync_running(d) == 0))
1126                 wake_up(&d->opd_sync_waitq);
1127
1128         EXIT;
1129 }
1130
1131 /**
1132  * The core of the syncing mechanism.
1133  *
1134  * This is a callback called by the llog processing function. Essentially it
1135  * suspends llog processing until there is a record to process (it's supposed
1136  * to be committed locally). The function handles RPCs committed by the target
1137  * and cancels corresponding llog records.
1138  *
1139  * \param[in] env       LU environment provided by the caller
1140  * \param[in] llh       llog handle we're processing
1141  * \param[in] rec       current llog record
1142  * \param[in] data      callback data containing a pointer to the device
1143  *
1144  * \retval 0                    to ask the caller (llog_process()) to continue
1145  * \retval LLOG_PROC_BREAK      to ask the caller to break
1146  */
1147 static int osp_sync_process_queues(const struct lu_env *env,
1148                                    struct llog_handle *llh,
1149                                    struct llog_rec_hdr *rec,
1150                                    void *data)
1151 {
1152         struct osp_device       *d = data;
1153
1154         do {
1155                 if (!osp_sync_running(d)) {
1156                         CDEBUG(D_HA, "stop llog processing\n");
1157                         return LLOG_PROC_BREAK;
1158                 }
1159
1160                 /* process requests committed by OST */
1161                 osp_sync_process_committed(env, d);
1162
1163                 /* if we there are changes to be processed and we have
1164                  * resources for this ... do now */
1165                 if (osp_sync_can_process_new(d, rec)) {
1166                         if (llh == NULL) {
1167                                 /* ask llog for another record */
1168                                 CDEBUG(D_HA, "%u changes, %u in progress,"
1169                                      " %u in flight\n",
1170                                      atomic_read(&d->opd_sync_changes),
1171                                      atomic_read(&d->opd_sync_rpcs_in_progress),
1172                                      atomic_read(&d->opd_sync_rpcs_in_flight));
1173                                 return 0;
1174                         }
1175                         osp_sync_process_record(env, d, llh, rec);
1176                         llh = NULL;
1177                         rec = NULL;
1178                 }
1179
1180                 wait_event_idle(d->opd_sync_waitq,
1181                                 !osp_sync_running(d) ||
1182                                 osp_sync_can_process_new(d, rec) ||
1183                                 !list_empty(&d->opd_sync_committed_there));
1184         } while (1);
1185 }
1186
1187 /**
1188  * OSP sync thread.
1189  *
1190  * This thread runs llog_cat_process() scanner calling our callback
1191  * to process llog records. in the callback we implement tricky
1192  * state machine as we don't want to start scanning of the llog again
1193  * and again, also we don't want to process too many records and send
1194  * too many RPCs a time. so, depending on current load (num of changes
1195  * being synced to OST) the callback can suspend awaiting for some
1196  * new conditions, like syncs completed.
1197  *
1198  * In order to process llog records left by previous boots and to allow
1199  * llog_process_thread() to find something (otherwise it'd just exit
1200  * immediately) we add a special GENERATATION record on each boot.
1201  *
1202  * \param[in] _arg      a pointer to thread's arguments
1203  *
1204  * \retval 0            on success
1205  * \retval negative     negated errno on error
1206  */
1207 static int osp_sync_thread(void *_arg)
1208 {
1209         struct osp_device       *d = _arg;
1210         struct ptlrpc_thread    *thread = &d->opd_sync_thread;
1211         struct l_wait_info       lwi = { 0 };
1212         struct llog_ctxt        *ctxt;
1213         struct obd_device       *obd = d->opd_obd;
1214         struct llog_handle      *llh;
1215         struct lu_env            env;
1216         int                      rc, count;
1217         bool                     wrapped;
1218
1219         ENTRY;
1220
1221         rc = lu_env_init(&env, LCT_LOCAL);
1222         if (rc) {
1223                 CERROR("%s: can't initialize env: rc = %d\n",
1224                        obd->obd_name, rc);
1225
1226                 spin_lock(&d->opd_sync_lock);
1227                 thread->t_flags = SVC_STOPPED;
1228                 spin_unlock(&d->opd_sync_lock);
1229                 wake_up(&thread->t_ctl_waitq);
1230
1231                 RETURN(rc);
1232         }
1233
1234         spin_lock(&d->opd_sync_lock);
1235         thread->t_flags = SVC_RUNNING;
1236         spin_unlock(&d->opd_sync_lock);
1237         wake_up(&thread->t_ctl_waitq);
1238
1239 again:
1240         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1241         if (ctxt == NULL) {
1242                 CERROR("can't get appropriate context\n");
1243                 GOTO(out, rc = -EINVAL);
1244         }
1245
1246         llh = ctxt->loc_handle;
1247         if (llh == NULL) {
1248                 CERROR("can't get llh\n");
1249                 llog_ctxt_put(ctxt);
1250                 GOTO(out, rc = -EINVAL);
1251         }
1252
1253         /*
1254          * Catalog processing stops when it processed last catalog record
1255          * with index equal to the end of catalog bitmap. Or if it is wrapped,
1256          * processing stops with index equal to the lgh_last_idx. We need to
1257          * continue processing.
1258          */
1259         d->opd_sync_last_catalog_idx = 0;
1260         do {
1261                 int     size;
1262
1263                 wrapped = (llh->lgh_hdr->llh_cat_idx >= llh->lgh_last_idx &&
1264                            llh->lgh_hdr->llh_count > 1);
1265
1266                 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CANT_PROCESS_LLOG)) {
1267                         rc = -EINPROGRESS;
1268                         goto next;
1269                 }
1270                 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d,
1271                                       d->opd_sync_last_catalog_idx, 0);
1272
1273 next:
1274                 size = OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ?
1275                        cfs_fail_val : (LLOG_HDR_BITMAP_SIZE(llh->lgh_hdr) - 1);
1276                 /* processing reaches catalog bottom */
1277                 if (d->opd_sync_last_catalog_idx == size)
1278                         d->opd_sync_last_catalog_idx = LLOG_CAT_FIRST;
1279                 else if (wrapped)
1280                         /* If catalog is wrapped we can`t predict last index of
1281                          * processing because lgh_last_idx could be changed.
1282                          * Starting form the next one */
1283                         d->opd_sync_last_catalog_idx++;
1284
1285         } while (rc == 0 && (wrapped ||
1286                              d->opd_sync_last_catalog_idx == LLOG_CAT_FIRST));
1287
1288         if (rc < 0) {
1289                 if (rc == -EINPROGRESS) {
1290                         /* can't access the llog now - OI scrub is trying to fix
1291                          * underlying issue. let's wait and try again */
1292                         llog_cat_close(&env, llh);
1293                         rc = llog_cleanup(&env, ctxt);
1294                         if (rc)
1295                                 GOTO(out, rc);
1296                         schedule_timeout_interruptible(HZ * 5);
1297                         goto again;
1298                 }
1299
1300                 CERROR("%s: llog process with osp_sync_process_queues "
1301                        "failed: %d\n", d->opd_obd->obd_name, rc);
1302                 GOTO(close, rc);
1303         }
1304         LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1305                  "%u changes, %u in progress, %u in flight: %d\n",
1306                  atomic_read(&d->opd_sync_changes),
1307                  atomic_read(&d->opd_sync_rpcs_in_progress),
1308                  atomic_read(&d->opd_sync_rpcs_in_flight), rc);
1309
1310         /* we don't expect llog_process_thread() to exit till umount */
1311         LASSERTF(thread->t_flags != SVC_RUNNING,
1312                  "%u changes, %u in progress, %u in flight\n",
1313                  atomic_read(&d->opd_sync_changes),
1314                  atomic_read(&d->opd_sync_rpcs_in_progress),
1315                  atomic_read(&d->opd_sync_rpcs_in_flight));
1316
1317         /* wait till all the requests are completed */
1318         count = 0;
1319         while (atomic_read(&d->opd_sync_rpcs_in_progress) > 0) {
1320                 osp_sync_process_committed(&env, d);
1321
1322                 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1323                 rc = l_wait_event(d->opd_sync_waitq,
1324                                 atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1325                                   &lwi);
1326                 if (rc == -ETIMEDOUT)
1327                         count++;
1328                 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1329                          d->opd_obd->obd_name,
1330                          atomic_read(&d->opd_sync_rpcs_in_progress),
1331                          atomic_read(&d->opd_sync_rpcs_in_flight),
1332                          list_empty(&d->opd_sync_committed_there) ? "" : "!");
1333
1334         }
1335
1336 close:
1337         llog_cat_close(&env, llh);
1338         rc = llog_cleanup(&env, ctxt);
1339         if (rc)
1340                 CERROR("can't cleanup llog: %d\n", rc);
1341 out:
1342         LASSERTF(atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1343                  "%s: %d %d %sempty\n", d->opd_obd->obd_name,
1344                  atomic_read(&d->opd_sync_rpcs_in_progress),
1345                  atomic_read(&d->opd_sync_rpcs_in_flight),
1346                  list_empty(&d->opd_sync_committed_there) ? "" : "!");
1347
1348         thread->t_flags = SVC_STOPPED;
1349
1350         wake_up(&thread->t_ctl_waitq);
1351
1352         lu_env_fini(&env);
1353
1354         RETURN(0);
1355 }
1356
1357 /**
1358  * Initialize llog.
1359  *
1360  * Initializes the llog. Specific llog to be used depends on the type of the
1361  * target OSP represents (OST or MDT). The function adds appends a new llog
1362  * record to mark the place where the records associated with this boot
1363  * start.
1364  *
1365  * \param[in] env       LU environment provided by the caller
1366  * \param[in] d         OSP device
1367  *
1368  * \retval 0            on success
1369  * \retval negative     negated errno on error
1370  */
1371 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1372 {
1373         struct osp_thread_info  *osi = osp_env_info(env);
1374         struct lu_fid           *fid = &osi->osi_fid;
1375         struct llog_handle      *lgh = NULL;
1376         struct obd_device       *obd = d->opd_obd;
1377         struct llog_ctxt        *ctxt;
1378         int                     rc;
1379
1380         ENTRY;
1381
1382         LASSERT(obd);
1383
1384         /*
1385          * open llog corresponding to our OST
1386          */
1387         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1388         obd->obd_lvfs_ctxt.dt = d->opd_storage;
1389
1390         lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1391
1392         rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1393                                    &osi->osi_cid, fid);
1394         if (rc < 0) {
1395                 if (rc != -EFAULT) {
1396                         CERROR("%s: can't get id from catalogs: rc = %d\n",
1397                                obd->obd_name, rc);
1398                         RETURN(rc);
1399                 }
1400
1401                 /* After sparse OST indices is supported, the CATALOG file
1402                  * may become a sparse file that results in failure on
1403                  * reading. Skip this error as the llog will be created
1404                  * later */
1405                 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1406                 rc = 0;
1407         }
1408
1409         CDEBUG(D_INFO, "%s: Init llog for %d - catid "DFID":%x\n",
1410                obd->obd_name, d->opd_index,
1411                PFID(&osi->osi_cid.lci_logid.lgl_oi.oi_fid),
1412                osi->osi_cid.lci_logid.lgl_ogen);
1413
1414         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT,
1415                         d->opd_storage->dd_lu_dev.ld_obd,
1416                         &llog_common_cat_ops);
1417         if (rc)
1418                 RETURN(rc);
1419
1420         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1421         LASSERT(ctxt);
1422
1423         if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1424                 struct lu_fid fid_temp;
1425
1426                 if (CFS_FAIL_CHECK(OBD_FAIL_OSP_INVALID_LOGID)) {
1427                         memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1428                         logid_set_id(&osi->osi_cid.lci_logid, cfs_fail_val);
1429                 }
1430
1431                 logid_to_fid(&osi->osi_cid.lci_logid, &fid_temp);
1432                 if (fid_is_sane(&fid_temp)) {
1433                         rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid,
1434                                        NULL, LLOG_OPEN_EXISTS);
1435
1436                         /* re-create llog if it is missing */
1437                         if (rc == -ENOENT)
1438                                 logid_set_id(&osi->osi_cid.lci_logid, 0);
1439                         else if (rc < 0)
1440                                 GOTO(out_cleanup, rc);
1441                 } else {
1442                         CERROR("%s: the catid "DFID" for init llog %d is bad\n",
1443                                obd->obd_name, PFID(&fid_temp), d->opd_index);
1444
1445                         /* it will be recreated later */
1446                         logid_set_id(&osi->osi_cid.lci_logid, 0);
1447                 }
1448         }
1449
1450         if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1451                 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1452                 if (rc < 0)
1453                         GOTO(out_cleanup, rc);
1454                 osi->osi_cid.lci_logid = lgh->lgh_id;
1455         }
1456
1457         LASSERT(lgh != NULL);
1458         ctxt->loc_handle = lgh;
1459
1460         rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
1461         if (rc)
1462                 GOTO(out_close, rc);
1463
1464         rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1465                                    &osi->osi_cid, fid);
1466         if (rc)
1467                 GOTO(out_close, rc);
1468
1469         /*
1470          * put a mark in the llog till which we'll be processing
1471          * old records restless
1472          */
1473         d->opd_sync_generation.mnt_cnt = ktime_get_ns();
1474         d->opd_sync_generation.conn_cnt = ktime_get_ns();
1475
1476         osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1477         osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1478
1479         memcpy(&osi->osi_gen.lgr_gen, &d->opd_sync_generation,
1480                sizeof(osi->osi_gen.lgr_gen));
1481
1482         rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1483         if (rc < 0)
1484                 GOTO(out_close, rc);
1485         llog_ctxt_put(ctxt);
1486         RETURN(0);
1487 out_close:
1488         llog_cat_close(env, lgh);
1489 out_cleanup:
1490         llog_cleanup(env, ctxt);
1491         RETURN(rc);
1492 }
1493
1494 /**
1495  * Cleanup llog used for syncing.
1496  *
1497  * Closes and cleanups the llog. The function is called when the device is
1498  * shutting down.
1499  *
1500  * \param[in] env       LU environment provided by the caller
1501  * \param[in] d         OSP device
1502  */
1503 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1504 {
1505         struct llog_ctxt *ctxt;
1506
1507         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1508         if (ctxt) {
1509                 llog_cat_close(env, ctxt->loc_handle);
1510                 llog_cleanup(env, ctxt);
1511         }
1512 }
1513
1514 /**
1515  * Initialization of the sync component of OSP.
1516  *
1517  * Initializes the llog and starts a new thread to handle the changes to
1518  * the remote target (OST or MDT).
1519  *
1520  * \param[in] env       LU environment provided by the caller
1521  * \param[in] d         OSP device
1522  *
1523  * \retval 0            on success
1524  * \retval negative     negated errno on error
1525  */
1526 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1527 {
1528         struct task_struct      *task;
1529         int                      rc;
1530
1531         ENTRY;
1532
1533         d->opd_sync_max_rpcs_in_flight = OSP_MAX_RPCS_IN_FLIGHT;
1534         d->opd_sync_max_rpcs_in_progress = OSP_MAX_RPCS_IN_PROGRESS;
1535         spin_lock_init(&d->opd_sync_lock);
1536         init_waitqueue_head(&d->opd_sync_waitq);
1537         init_waitqueue_head(&d->opd_sync_barrier_waitq);
1538         thread_set_flags(&d->opd_sync_thread, SVC_INIT);
1539         init_waitqueue_head(&d->opd_sync_thread.t_ctl_waitq);
1540         INIT_LIST_HEAD(&d->opd_sync_in_flight_list);
1541         INIT_LIST_HEAD(&d->opd_sync_committed_there);
1542
1543         if (d->opd_storage->dd_rdonly)
1544                 RETURN(0);
1545
1546         /*
1547          * initialize llog storing changes
1548          */
1549         rc = osp_sync_llog_init(env, d);
1550         if (rc) {
1551                 CERROR("%s: can't initialize llog: rc = %d\n",
1552                        d->opd_obd->obd_name, rc);
1553                 GOTO(err_id, rc);
1554         }
1555
1556         /*
1557          * Start synchronization thread
1558          */
1559         task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1560                            d->opd_index, d->opd_group);
1561         if (IS_ERR(task)) {
1562                 rc = PTR_ERR(task);
1563                 CERROR("%s: cannot start sync thread: rc = %d\n",
1564                        d->opd_obd->obd_name, rc);
1565                 GOTO(err_llog, rc);
1566         }
1567
1568         wait_event_idle(d->opd_sync_thread.t_ctl_waitq,
1569                         osp_sync_running(d) || osp_sync_stopped(d));
1570
1571         RETURN(0);
1572 err_llog:
1573         osp_sync_llog_fini(env, d);
1574 err_id:
1575         return rc;
1576 }
1577
1578 /**
1579  * Stop the syncing thread.
1580  *
1581  * Asks the syncing thread to stop and wait until it's stopped.
1582  *
1583  * \param[in] d         OSP device
1584  *
1585  * \retval              0
1586  */
1587 int osp_sync_fini(struct osp_device *d)
1588 {
1589         struct ptlrpc_thread *thread = &d->opd_sync_thread;
1590
1591         ENTRY;
1592
1593         if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1594                 thread->t_flags = SVC_STOPPING;
1595                 wake_up(&d->opd_sync_waitq);
1596                 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1597         }
1598
1599         RETURN(0);
1600 }
1601
1602 struct osp_last_committed_cb {
1603         struct dt_txn_commit_cb  ospc_cb;
1604         struct osp_device       *ospc_dev;
1605         __u64                    ospc_transno;
1606 };
1607
1608 void osp_sync_local_commit_cb(struct lu_env *env, struct thandle *th,
1609                               struct dt_txn_commit_cb *dcb, int err)
1610 {
1611         struct osp_last_committed_cb    *cb;
1612         struct osp_device               *d;
1613
1614         cb = container_of0(dcb, struct osp_last_committed_cb, ospc_cb);
1615         d = cb->ospc_dev;
1616
1617         CDEBUG(D_HA, "%s: %llu committed\n", d->opd_obd->obd_name,
1618                cb->ospc_transno);
1619
1620         spin_lock(&d->opd_sync_lock);
1621         if (cb->ospc_transno > d->opd_sync_last_committed_id)
1622                 d->opd_sync_last_committed_id = cb->ospc_transno;
1623         spin_unlock(&d->opd_sync_lock);
1624
1625         osp_sync_check_for_work(d);
1626         lu_device_put(osp2lu_dev(d));
1627         if (atomic_dec_and_test(&d->opd_commits_registered))
1628                 wake_up(&d->opd_sync_waitq);
1629
1630         OBD_FREE_PTR(cb);
1631 }
1632
1633 static int osp_sync_add_commit_cb(const struct lu_env *env,
1634                                   struct osp_device *d, struct thandle *th)
1635 {
1636         struct osp_last_committed_cb    *cb;
1637         struct dt_txn_commit_cb         *dcb;
1638         int                              rc = 0;
1639
1640         OBD_ALLOC_PTR(cb);
1641         if (cb == NULL)
1642                 return -ENOMEM;
1643         cb->ospc_dev = d;
1644         dcb = &cb->ospc_cb;
1645         dcb->dcb_func = osp_sync_local_commit_cb;
1646         spin_lock(&d->opd_sync_lock);
1647         cb->ospc_transno = ++d->opd_sync_last_used_id;
1648         spin_unlock(&d->opd_sync_lock);
1649
1650         rc = dt_trans_cb_add(th, dcb);
1651         CDEBUG(D_HA, "%s: add commit cb at %lluns, next at %lluns, rc = %d\n",
1652                d->opd_obd->obd_name, ktime_get_ns(),
1653                ktime_to_ns(d->opd_sync_next_commit_cb), rc);
1654
1655         if (likely(rc == 0)) {
1656                 lu_device_get(osp2lu_dev(d));
1657                 atomic_inc(&d->opd_commits_registered);
1658         } else
1659                 OBD_FREE_PTR(cb);
1660
1661         return rc;
1662 }
1663
1664 /* add the commit callback every second */
1665 int osp_sync_add_commit_cb_1s(const struct lu_env *env, struct osp_device *d,
1666                               struct thandle *th)
1667 {
1668         ktime_t now = ktime_get();
1669         bool add = false;
1670
1671         /* fast path */
1672         if (ktime_before(now, d->opd_sync_next_commit_cb))
1673                 return 0;
1674
1675         spin_lock(&d->opd_sync_lock);
1676         if (ktime_before(d->opd_sync_next_commit_cb, now)) {
1677                 add = true;
1678                 d->opd_sync_next_commit_cb = ktime_add_ns(now, NSEC_PER_SEC);
1679         }
1680         spin_unlock(&d->opd_sync_lock);
1681
1682         if (!add)
1683                 return 0;
1684
1685         return osp_sync_add_commit_cb(env, d, th);
1686 }
1687
1688 /*
1689  * generate an empty transaction and hook the commit callback in
1690  * then force transaction commit
1691  */
1692 void osp_sync_force(const struct lu_env *env, struct osp_device *d)
1693 {
1694         struct thandle *th;
1695         int rc;
1696
1697         th = dt_trans_create(env, d->opd_storage);
1698         if (IS_ERR(th)) {
1699                 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
1700                 return;
1701         }
1702         rc = dt_trans_start_local(env, d->opd_storage, th);
1703         if (rc == 0) {
1704                 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
1705                        d->opd_obd->obd_name,
1706                        atomic_read(&d->opd_sync_changes));
1707                 rc = osp_sync_add_commit_cb(env, d, th);
1708                 dt_trans_stop(env, d->opd_storage, th);
1709         }
1710
1711         dt_commit_async(env, d->opd_storage);
1712 }