lustre/ofd/ofd_io.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) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ofd/ofd_io.c
  33  *
  34  * This file provides functions to handle IO requests from clients and
  35  * also LFSCK routines to check parent file identifier (PFID) consistency.
  36  *
  37  * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
  38  * Author: Fan Yong <fan.yong@intel.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_FILTER
  42
  43 #include <linux/kthread.h>
  44 #include "ofd_internal.h"
  45 #include <lustre_nodemap.h>
  46
  47 struct ofd_inconsistency_item {
  48         struct list_head         oii_list;
  49         struct ofd_object       *oii_obj;
  50         struct filter_fid        oii_ff;
  51 };
  52
  53 /**
  54  * Verify single object for parent FID consistency.
  55  *
  56  * Part of LFSCK processing which checks single object PFID stored in extended
  57  * attribute (XATTR) against real FID of MDT parent object received by LFSCK.
  58  * This verifies that the OST object is being referenced by only a single MDT
  59  * object.
  60  *
  61  * \param[in] env       execution environment
  62  * \param[in] ofd       OFD device
  63  * \param[in] oii       object-related local data
  64  * \param[in] lrl       LFSCK request data
  65  */
  66 static void ofd_inconsistency_verify_one(const struct lu_env *env,
  67                                          struct ofd_device *ofd,
  68                                          struct ofd_inconsistency_item *oii,
  69                                          struct lfsck_req_local *lrl)
  70 {
  71         struct ofd_object *fo = oii->oii_obj;
  72         struct filter_fid *client_ff = &oii->oii_ff;
  73         struct filter_fid *local_ff = &fo->ofo_ff;
  74         int rc;
  75
  76         LASSERT(fo->ofo_pfid_checking);
  77         LASSERT(!fo->ofo_pfid_verified);
  78
  79         lrl->lrl_fid = fo->ofo_header.loh_fid; /* OST-object itself FID. */
  80         lrl->lrl_ff_client = *client_ff; /* client given PFID. */
  81         lrl->lrl_ff_local = *local_ff; /* OST local stored PFID. */
  82
  83         rc = lfsck_in_notify_local(env, ofd->ofd_osd, lrl, NULL);
  84         ofd_write_lock(env, fo);
  85         switch (lrl->lrl_status) {
  86         case LPVS_INIT:
  87                 LASSERT(rc <= 0);
  88
  89                 if (rc < 0)
  90                         CDEBUG(D_LFSCK, "%s: fail to verify OST local stored "
  91                                "PFID xattr for "DFID", the client given PFID "
  92                                DFID", OST local stored PFID "DFID": rc = %d\n",
  93                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
  94                                PFID(&client_ff->ff_parent),
  95                                PFID(&local_ff->ff_parent), rc);
  96                 else
  97                         fo->ofo_pfid_verified = 1;
  98                 break;
  99         case LPVS_INCONSISTENT:
 100                 LASSERT(rc != 0);
 101
 102                 ofd->ofd_inconsistency_self_detected++;
 103                 if (rc < 0)
 104                         CDEBUG(D_LFSCK, "%s: fail to verify the client given "
 105                                "PFID for "DFID", the client given PFID "DFID
 106                                ", local stored PFID "DFID": rc = %d\n",
 107                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
 108                                PFID(&client_ff->ff_parent),
 109                                PFID(&local_ff->ff_parent), rc);
 110                 else
 111                         CDEBUG(D_LFSCK, "%s: both the client given PFID and "
 112                                "the OST local stored PFID are stale for the "
 113                                "OST-object "DFID", client given PFID is "DFID
 114                                ", local stored PFID is "DFID"\n",
 115                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
 116                                PFID(&client_ff->ff_parent),
 117                                PFID(&local_ff->ff_parent));
 118                 break;
 119         case LPVS_INCONSISTENT_TOFIX:
 120                 ofd->ofd_inconsistency_self_detected++;
 121                 if (rc == 0) {
 122                         ofd->ofd_inconsistency_self_repaired++;
 123                         CDEBUG(D_LFSCK, "%s: fixed the staled OST PFID xattr "
 124                                "for "DFID", with the client given PFID "DFID
 125                                ", the old stored PFID "DFID"\n",
 126                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
 127                                PFID(&client_ff->ff_parent),
 128                                PFID(&local_ff->ff_parent));
 129                 } else if (rc < 0) {
 130                         CDEBUG(D_LFSCK, "%s: fail to fix the OST PFID xattr "
 131                                "for "DFID", client given PFID "DFID", local "
 132                                "stored PFID "DFID": rc = %d\n",
 133                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
 134                                PFID(&client_ff->ff_parent),
 135                                PFID(&local_ff->ff_parent), rc);
 136                 }
 137                 local_ff->ff_parent = client_ff->ff_parent;
 138                 fo->ofo_pfid_verified = 1;
 139                 break;
 140         default:
 141                 break;
 142         }
 143         fo->ofo_pfid_checking = 0;
 144         ofd_write_unlock(env, fo);
 145
 146         ofd_object_put(env, fo);
 147         OBD_FREE_PTR(oii);
 148 }
 149
 150 /**
 151  * Verification thread to check parent FID consistency.
 152  *
 153  * Kernel thread to check consistency of parent FID for any
 154  * new item added for checking by ofd_add_inconsistency_item().
 155  *
 156  * \param[in] args      OFD device
 157  *
 158  * \retval              0 on successful thread termination
 159  * \retval              negative value if thread can't start
 160  */
 161 static int ofd_inconsistency_verification_main(void *args)
 162 {
 163         struct lu_env env;
 164         struct ofd_device *ofd = args;
 165         struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread;
 166         struct ofd_inconsistency_item *oii;
 167         struct lfsck_req_local *lrl = NULL;
 168         struct l_wait_info lwi = { 0 };
 169         int rc;
 170         ENTRY;
 171
 172         rc = lu_env_init(&env, LCT_DT_THREAD);
 173         spin_lock(&ofd->ofd_inconsistency_lock);
 174         thread_set_flags(thread, rc ? SVC_STOPPED : SVC_RUNNING);
 175         wake_up_all(&thread->t_ctl_waitq);
 176         spin_unlock(&ofd->ofd_inconsistency_lock);
 177         if (rc)
 178                 RETURN(rc);
 179
 180         OBD_ALLOC_PTR(lrl);
 181         if (unlikely(!lrl))
 182                 GOTO(out_unlocked, rc = -ENOMEM);
 183
 184         lrl->lrl_event = LEL_PAIRS_VERIFY_LOCAL;
 185         lrl->lrl_active = LFSCK_TYPE_LAYOUT;
 186
 187         spin_lock(&ofd->ofd_inconsistency_lock);
 188         while (1) {
 189                 if (unlikely(!thread_is_running(thread)))
 190                         break;
 191
 192                 while (!list_empty(&ofd->ofd_inconsistency_list)) {
 193                         oii = list_entry(ofd->ofd_inconsistency_list.next,
 194                                          struct ofd_inconsistency_item,
 195                                          oii_list);
 196                         list_del_init(&oii->oii_list);
 197                         spin_unlock(&ofd->ofd_inconsistency_lock);
 198                         ofd_inconsistency_verify_one(&env, ofd, oii, lrl);
 199                         spin_lock(&ofd->ofd_inconsistency_lock);
 200                 }
 201
 202                 spin_unlock(&ofd->ofd_inconsistency_lock);
 203                 l_wait_event(thread->t_ctl_waitq,
 204                              !list_empty(&ofd->ofd_inconsistency_list) ||
 205                              !thread_is_running(thread),
 206                              &lwi);
 207                 spin_lock(&ofd->ofd_inconsistency_lock);
 208         }
 209
 210         while (!list_empty(&ofd->ofd_inconsistency_list)) {
 211                 struct ofd_object *fo;
 212
 213                 oii = list_entry(ofd->ofd_inconsistency_list.next,
 214                                  struct ofd_inconsistency_item,
 215                                  oii_list);
 216                 list_del_init(&oii->oii_list);
 217                 fo = oii->oii_obj;
 218                 spin_unlock(&ofd->ofd_inconsistency_lock);
 219
 220                 ofd_write_lock(&env, fo);
 221                 fo->ofo_pfid_checking = 0;
 222                 ofd_write_unlock(&env, fo);
 223
 224                 ofd_object_put(&env, fo);
 225                 OBD_FREE_PTR(oii);
 226                 spin_lock(&ofd->ofd_inconsistency_lock);
 227         }
 228
 229         OBD_FREE_PTR(lrl);
 230
 231         GOTO(out, rc = 0);
 232
 233 out_unlocked:
 234         spin_lock(&ofd->ofd_inconsistency_lock);
 235 out:
 236         thread_set_flags(thread, SVC_STOPPED);
 237         wake_up_all(&thread->t_ctl_waitq);
 238         spin_unlock(&ofd->ofd_inconsistency_lock);
 239         lu_env_fini(&env);
 240
 241         return rc;
 242 }
 243
 244 /**
 245  * Start parent FID verification thread.
 246  *
 247  * See ofd_inconsistency_verification_main().
 248  *
 249  * \param[in] ofd       OFD device
 250  *
 251  * \retval              0 on successful start of thread
 252  * \retval              negative value on error
 253  */
 254 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd)
 255 {
 256         struct ptlrpc_thread    *thread = &ofd->ofd_inconsistency_thread;
 257         struct l_wait_info       lwi    = { 0 };
 258         struct task_struct      *task;
 259         int                      rc;
 260
 261         spin_lock(&ofd->ofd_inconsistency_lock);
 262         if (unlikely(thread_is_running(thread))) {
 263                 spin_unlock(&ofd->ofd_inconsistency_lock);
 264
 265                 return -EALREADY;
 266         }
 267
 268         thread_set_flags(thread, 0);
 269         spin_unlock(&ofd->ofd_inconsistency_lock);
 270         task = kthread_run(ofd_inconsistency_verification_main, ofd,
 271                            "inconsistency_verification");
 272         if (IS_ERR(task)) {
 273                 rc = PTR_ERR(task);
 274                 CERROR("%s: cannot start self_repair thread: rc = %d\n",
 275                        ofd_name(ofd), rc);
 276         } else {
 277                 rc = 0;
 278                 l_wait_event(thread->t_ctl_waitq,
 279                              thread_is_running(thread) ||
 280                              thread_is_stopped(thread),
 281                              &lwi);
 282         }
 283
 284         return rc;
 285 }
 286
 287 /**
 288  * Stop parent FID verification thread.
 289  *
 290  * \param[in] ofd       OFD device
 291  *
 292  * \retval              0 on successful start of thread
 293  * \retval              -EALREADY if thread is already stopped
 294  */
 295 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd)
 296 {
 297         struct ptlrpc_thread    *thread = &ofd->ofd_inconsistency_thread;
 298         struct l_wait_info       lwi    = { 0 };
 299
 300         spin_lock(&ofd->ofd_inconsistency_lock);
 301         if (thread_is_init(thread) || thread_is_stopped(thread)) {
 302                 spin_unlock(&ofd->ofd_inconsistency_lock);
 303
 304                 return -EALREADY;
 305         }
 306
 307         thread_set_flags(thread, SVC_STOPPING);
 308         spin_unlock(&ofd->ofd_inconsistency_lock);
 309         wake_up_all(&thread->t_ctl_waitq);
 310         l_wait_event(thread->t_ctl_waitq,
 311                      thread_is_stopped(thread),
 312                      &lwi);
 313
 314         return 0;
 315 }
 316
 317 /**
 318  * Add new item for parent FID verification.
 319  *
 320  * Prepare new verification item and pass it to the dedicated
 321  * verification thread for further processing.
 322  *
 323  * \param[in] env       execution environment
 324  * \param[in] fo        OFD object
 325  * \param[in] oa        OBDO structure with PFID
 326  */
 327 static void ofd_add_inconsistency_item(const struct lu_env *env,
 328                                        struct ofd_object *fo, struct obdo *oa)
 329 {
 330         struct ofd_device *ofd = ofd_obj2dev(fo);
 331         struct ofd_inconsistency_item *oii;
 332         struct filter_fid *ff;
 333         bool wakeup = false;
 334
 335         OBD_ALLOC_PTR(oii);
 336         if (oii == NULL)
 337                 return;
 338
 339         INIT_LIST_HEAD(&oii->oii_list);
 340         lu_object_get(&fo->ofo_obj.do_lu);
 341         oii->oii_obj = fo;
 342         ff = &oii->oii_ff;
 343         ff->ff_parent.f_seq = oa->o_parent_seq;
 344         ff->ff_parent.f_oid = oa->o_parent_oid;
 345         ff->ff_parent.f_stripe_idx = oa->o_stripe_idx;
 346         ff->ff_layout = oa->o_layout;
 347
 348         spin_lock(&ofd->ofd_inconsistency_lock);
 349         if (fo->ofo_pfid_checking || fo->ofo_pfid_verified) {
 350                 spin_unlock(&ofd->ofd_inconsistency_lock);
 351                 OBD_FREE_PTR(oii);
 352
 353                 return;
 354         }
 355
 356         fo->ofo_pfid_checking = 1;
 357         if (list_empty(&ofd->ofd_inconsistency_list))
 358                 wakeup = true;
 359         list_add_tail(&oii->oii_list, &ofd->ofd_inconsistency_list);
 360         spin_unlock(&ofd->ofd_inconsistency_lock);
 361         if (wakeup)
 362                 wake_up_all(&ofd->ofd_inconsistency_thread.t_ctl_waitq);
 363
 364         /* XXX: When the found inconsistency exceeds some threshold,
 365          *      we can trigger the LFSCK to scan part of the system
 366          *      or the whole system, which depends on how to define
 367          *      the threshold, a simple way maybe like that: define
 368          *      the absolute value of how many inconsisteny allowed
 369          *      to be repaired via self detect/repair mechanism, if
 370          *      exceeded, then trigger the LFSCK to scan the layout
 371          *      inconsistency within the whole system. */
 372 }
 373
 374 /**
 375  * Verify parent FID of an object.
 376  *
 377  * Check the parent FID is sane and start extended
 378  * verification procedure otherwise.
 379  *
 380  * \param[in] env       execution environment
 381  * \param[in] fo        OFD object
 382  * \param[in] oa        OBDO structure with PFID
 383  *
 384  * \retval              0 on successful verification
 385  * \retval              -EINPROGRESS if PFID is being repaired
 386  * \retval              -EPERM if PFID was verified but still insane
 387  */
 388 int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
 389                   struct obdo *oa)
 390 {
 391         struct lu_fid *pfid = &fo->ofo_ff.ff_parent;
 392         int rc = 0;
 393         ENTRY;
 394
 395         if (fid_is_sane(pfid)) {
 396                 if (likely(oa->o_parent_seq == pfid->f_seq &&
 397                            oa->o_parent_oid == pfid->f_oid &&
 398                            oa->o_stripe_idx == pfid->f_stripe_idx))
 399                         RETURN(0);
 400
 401                 if (fo->ofo_pfid_verified)
 402                         RETURN(-EPERM);
 403         }
 404
 405         /* The OST-object may be inconsistent, and we need further verification.
 406          * To avoid block the RPC service thread, return -EINPROGRESS to client
 407          * and make it retry later. */
 408         if (fo->ofo_pfid_checking)
 409                 RETURN(-EINPROGRESS);
 410
 411         rc = ofd_object_ff_load(env, fo);
 412         if (rc == -ENODATA)
 413                 RETURN(0);
 414
 415         if (rc < 0)
 416                 RETURN(rc);
 417
 418         if (likely(oa->o_parent_seq == pfid->f_seq &&
 419                    oa->o_parent_oid == pfid->f_oid &&
 420                    oa->o_stripe_idx == pfid->f_stripe_idx))
 421                 RETURN(0);
 422
 423         /* Push it to the dedicated thread for further verification. */
 424         ofd_add_inconsistency_item(env, fo, oa);
 425
 426         RETURN(-EINPROGRESS);
 427 }
 428
 429 /**
 430  * Prepare buffers for read request processing.
 431  *
 432  * This function converts remote buffers from client to local buffers
 433  * and prepares the latter.
 434  *
 435  * \param[in] env       execution environment
 436  * \param[in] exp       OBD export of client
 437  * \param[in] ofd       OFD device
 438  * \param[in] fid       FID of object
 439  * \param[in] la        object attributes
 440  * \param[in] oa        OBDO structure from client
 441  * \param[in] niocount  number of remote buffers
 442  * \param[in] rnb       remote buffers
 443  * \param[in] nr_local  number of local buffers
 444  * \param[in] lnb       local buffers
 445  * \param[in] jobid     job ID name
 446  *
 447  * \retval              0 on successful prepare
 448  * \retval              negative value on error
 449  */
 450 static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
 451                            struct ofd_device *ofd, const struct lu_fid *fid,
 452                            struct lu_attr *la, struct obdo *oa, int niocount,
 453                            struct niobuf_remote *rnb, int *nr_local,
 454                            struct niobuf_local *lnb, char *jobid)
 455 {
 456         struct ofd_object *fo;
 457         int i, j, rc, tot_bytes = 0;
 458         enum dt_bufs_type dbt = DT_BUFS_TYPE_READ;
 459
 460         ENTRY;
 461         LASSERT(env != NULL);
 462
 463         fo = ofd_object_find(env, ofd, fid);
 464         if (IS_ERR(fo))
 465                 RETURN(PTR_ERR(fo));
 466         LASSERT(fo != NULL);
 467
 468         ofd_read_lock(env, fo);
 469         if (!ofd_object_exists(fo))
 470                 GOTO(unlock, rc = -ENOENT);
 471
 472         if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
 473                 rc = ofd_verify_ff(env, fo, oa);
 474                 if (rc != 0)
 475                         GOTO(unlock, rc);
 476         }
 477
 478         if (ptlrpc_connection_is_local(exp->exp_connection))
 479                 dbt |= DT_BUFS_TYPE_LOCAL;
 480
 481         for (*nr_local = 0, i = 0, j = 0; i < niocount; i++) {
 482                 rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
 483                                  lnb + j, dbt);
 484                 if (unlikely(rc < 0))
 485                         GOTO(buf_put, rc);
 486                 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
 487                 /* correct index for local buffers to continue with */
 488                 j += rc;
 489                 *nr_local += rc;
 490                 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
 491                 tot_bytes += rnb[i].rnb_len;
 492         }
 493
 494         LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
 495         rc = dt_attr_get(env, ofd_object_child(fo), la);
 496         if (unlikely(rc))
 497                 GOTO(buf_put, rc);
 498
 499         rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
 500         if (unlikely(rc))
 501                 GOTO(buf_put, rc);
 502
 503         ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes);
 504         RETURN(0);
 505
 506 buf_put:
 507         dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
 508 unlock:
 509         ofd_read_unlock(env, fo);
 510         ofd_object_put(env, fo);
 511         return rc;
 512 }
 513
 514 /**
 515  * Prepare buffers for write request processing.
 516  *
 517  * This function converts remote buffers from client to local buffers
 518  * and prepares the latter. If there is recovery in progress and required
 519  * object is missing then it can be re-created before write.
 520  *
 521  * \param[in] env       execution environment
 522  * \param[in] exp       OBD export of client
 523  * \param[in] ofd       OFD device
 524  * \param[in] fid       FID of object
 525  * \param[in] la        object attributes
 526  * \param[in] oa        OBDO structure from client
 527  * \param[in] objcount  always 1
 528  * \param[in] obj       object data
 529  * \param[in] rnb       remote buffers
 530  * \param[in] nr_local  number of local buffers
 531  * \param[in] lnb       local buffers
 532  * \param[in] jobid     job ID name
 533  *
 534  * \retval              0 on successful prepare
 535  * \retval              negative value on error
 536  */
 537 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
 538                             struct ofd_device *ofd, const struct lu_fid *fid,
 539                             struct lu_attr *la, struct obdo *oa,
 540                             int objcount, struct obd_ioobj *obj,
 541                             struct niobuf_remote *rnb, int *nr_local,
 542                             struct niobuf_local *lnb, char *jobid)
 543 {
 544         struct ofd_object *fo;
 545         int i, j, k, rc = 0, tot_bytes = 0;
 546         enum dt_bufs_type dbt = DT_BUFS_TYPE_WRITE;
 547
 548         ENTRY;
 549         LASSERT(env != NULL);
 550         LASSERT(objcount == 1);
 551
 552         if (unlikely(exp->exp_obd->obd_recovering)) {
 553                 u64 seq = fid_seq(fid);
 554                 u64 oid = fid_oid(fid);
 555                 struct ofd_seq *oseq;
 556
 557                 oseq = ofd_seq_load(env, ofd, seq);
 558                 if (IS_ERR(oseq)) {
 559                         CERROR("%s: Can't find FID Sequence %#llx: rc = %d\n",
 560                                ofd_name(ofd), seq, (int)PTR_ERR(oseq));
 561                         GOTO(out, rc = -EINVAL);
 562                 }
 563
 564                 if (oid > ofd_seq_last_oid(oseq)) {
 565                         int sync = 0;
 566                         int diff;
 567
 568                         mutex_lock(&oseq->os_create_lock);
 569                         diff = oid - ofd_seq_last_oid(oseq);
 570
 571                         /* Do sync create if the seq is about to used up */
 572                         if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) {
 573                                 if (unlikely(oid >= IDIF_MAX_OID - 1))
 574                                         sync = 1;
 575                         } else if (fid_seq_is_norm(seq)) {
 576                                 if (unlikely(oid >=
 577                                              LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
 578                                         sync = 1;
 579                         } else {
 580                                 CERROR("%s : invalid o_seq "DOSTID"\n",
 581                                        ofd_name(ofd), POSTID(&oa->o_oi));
 582                                 mutex_unlock(&oseq->os_create_lock);
 583                                 ofd_seq_put(env, oseq);
 584                                 GOTO(out, rc = -EINVAL);
 585                         }
 586
 587                         while (diff > 0) {
 588                                 u64 next_id = ofd_seq_last_oid(oseq) + 1;
 589                                 int count = ofd_precreate_batch(ofd, diff);
 590
 591                                 rc = ofd_precreate_objects(env, ofd, next_id,
 592                                                            oseq, count, sync);
 593                                 if (rc < 0) {
 594                                         mutex_unlock(&oseq->os_create_lock);
 595                                         ofd_seq_put(env, oseq);
 596                                         GOTO(out, rc);
 597                                 }
 598
 599                                 diff -= rc;
 600                         }
 601
 602                         mutex_unlock(&oseq->os_create_lock);
 603                 }
 604
 605                 ofd_seq_put(env, oseq);
 606         }
 607
 608         fo = ofd_object_find(env, ofd, fid);
 609         if (IS_ERR(fo))
 610                 GOTO(out, rc = PTR_ERR(fo));
 611         LASSERT(fo != NULL);
 612
 613         ofd_read_lock(env, fo);
 614         if (!ofd_object_exists(fo)) {
 615                 CERROR("%s: BRW to missing obj "DOSTID"\n",
 616                        exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
 617                 ofd_read_unlock(env, fo);
 618                 ofd_object_put(env, fo);
 619                 GOTO(out, rc = -ENOENT);
 620         }
 621
 622         if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
 623                 rc = ofd_verify_ff(env, fo, oa);
 624                 if (rc != 0) {
 625                         ofd_read_unlock(env, fo);
 626                         ofd_object_put(env, fo);
 627                         GOTO(out, rc);
 628                 }
 629         }
 630
 631         /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
 632          * space back if possible */
 633         tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
 634
 635         if (ptlrpc_connection_is_local(exp->exp_connection))
 636                 dbt |= DT_BUFS_TYPE_LOCAL;
 637
 638         /* parse remote buffers to local buffers and prepare the latter */
 639         for (*nr_local = 0, i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
 640                 rc = dt_bufs_get(env, ofd_object_child(fo),
 641                                  rnb + i, lnb + j, dbt);
 642                 if (unlikely(rc < 0))
 643                         GOTO(err, rc);
 644                 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
 645                 /* correct index for local buffers to continue with */
 646                 for (k = 0; k < rc; k++) {
 647                         lnb[j+k].lnb_flags = rnb[i].rnb_flags;
 648                         lnb[j+k].lnb_flags &= ~OBD_BRW_LOCALS;
 649                         if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
 650                                 lnb[j+k].lnb_rc = -ENOSPC;
 651                 }
 652                 j += rc;
 653                 *nr_local += rc;
 654                 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
 655                 tot_bytes += rnb[i].rnb_len;
 656         }
 657         LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
 658
 659         rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
 660         if (unlikely(rc != 0))
 661                 GOTO(err, rc);
 662
 663         ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes);
 664         RETURN(0);
 665 err:
 666         dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
 667         ofd_read_unlock(env, fo);
 668         ofd_object_put(env, fo);
 669         /* tgt_grant_prepare_write() was called, so we must commit */
 670         tgt_grant_commit(exp, oa->o_grant_used, rc);
 671 out:
 672         /* let's still process incoming grant information packed in the oa,
 673          * but without enforcing grant since we won't proceed with the write.
 674          * Just like a read request actually. */
 675         tgt_grant_prepare_read(env, exp, oa);
 676         return rc;
 677 }
 678
 679 /**
 680  * Prepare bulk IO requests for processing.
 681  *
 682  * This function does initial checks of IO and calls corresponding
 683  * functions for read/write processing.
 684  *
 685  * \param[in] env       execution environment
 686  * \param[in] cmd       IO type (read/write)
 687  * \param[in] exp       OBD export of client
 688  * \param[in] oa        OBDO structure from request
 689  * \param[in] objcount  always 1
 690  * \param[in] obj       object data
 691  * \param[in] rnb       remote buffers
 692  * \param[in] nr_local  number of local buffers
 693  * \param[in] lnb       local buffers
 694  *
 695  * \retval              0 on successful prepare
 696  * \retval              negative value on error
 697  */
 698 int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
 699                struct obdo *oa, int objcount, struct obd_ioobj *obj,
 700                struct niobuf_remote *rnb, int *nr_local,
 701                struct niobuf_local *lnb)
 702 {
 703         struct tgt_session_info *tsi = tgt_ses_info(env);
 704         struct ofd_device       *ofd = ofd_exp(exp);
 705         struct ofd_thread_info  *info;
 706         char                    *jobid;
 707         const struct lu_fid     *fid = &oa->o_oi.oi_fid;
 708         int                      rc = 0;
 709
 710         if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
 711                 CERROR("%s: bulk has too many pages %d, which exceeds the"
 712                        "maximum pages per RPC of %d\n",
 713                        exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
 714                 RETURN(-EPROTO);
 715         }
 716
 717         if (tgt_ses_req(tsi) == NULL) { /* echo client case */
 718                 info = ofd_info_init(env, exp);
 719                 jobid = NULL;
 720         } else {
 721                 info = tsi2ofd_info(tsi);
 722                 jobid = tsi->tsi_jobid;
 723         }
 724
 725         LASSERT(oa != NULL);
 726
 727         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) {
 728                 struct ofd_seq          *oseq;
 729
 730                 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
 731                 if (IS_ERR(oseq)) {
 732                         CERROR("%s: Can not find seq for "DOSTID
 733                                ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
 734                                PTR_ERR(oseq));
 735                         RETURN(-EINVAL);
 736                 }
 737
 738                 if (oseq->os_destroys_in_progress == 0) {
 739                         /* don't fail lookups for orphan recovery, it causes
 740                          * later LBUGs when objects still exist during
 741                          * precreate */
 742                         ofd_seq_put(env, oseq);
 743                         RETURN(-ENOENT);
 744                 }
 745                 ofd_seq_put(env, oseq);
 746         }
 747
 748         LASSERT(objcount == 1);
 749         LASSERT(obj->ioo_bufcnt > 0);
 750
 751         if (cmd == OBD_BRW_WRITE) {
 752                 la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
 753                 rc = ofd_preprw_write(env, exp, ofd, fid, &info->fti_attr, oa,
 754                                       objcount, obj, rnb, nr_local, lnb, jobid);
 755         } else if (cmd == OBD_BRW_READ) {
 756                 tgt_grant_prepare_read(env, exp, oa);
 757                 rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa,
 758                                      obj->ioo_bufcnt, rnb, nr_local, lnb,
 759                                      jobid);
 760                 obdo_from_la(oa, &info->fti_attr, LA_ATIME);
 761         } else {
 762                 CERROR("%s: wrong cmd %d received!\n",
 763                        exp->exp_obd->obd_name, cmd);
 764                 rc = -EPROTO;
 765         }
 766         RETURN(rc);
 767 }
 768
 769 /**
 770  * Drop reference on local buffers for read bulk IO.
 771  *
 772  * This will free all local buffers use by this read request.
 773  *
 774  * \param[in] env       execution environment
 775  * \param[in] ofd       OFD device
 776  * \param[in] fid       object FID
 777  * \param[in] objcount  always 1
 778  * \param[in] niocount  number of local buffers
 779  * \param[in] lnb       local buffers
 780  *
 781  * \retval              0 on successful execution
 782  * \retval              negative value on error
 783  */
 784 static int
 785 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
 786                   const struct lu_fid *fid, int objcount, int niocount,
 787                   struct niobuf_local *lnb)
 788 {
 789         struct ofd_object *fo;
 790
 791         ENTRY;
 792
 793         LASSERT(niocount > 0);
 794
 795         fo = ofd_object_find(env, ofd, fid);
 796         if (IS_ERR(fo))
 797                 RETURN(PTR_ERR(fo));
 798         LASSERT(fo != NULL);
 799         LASSERT(ofd_object_exists(fo));
 800         dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
 801
 802         ofd_read_unlock(env, fo);
 803         ofd_object_put(env, fo);
 804         /* second put is pair to object_get in ofd_preprw_read */
 805         ofd_object_put(env, fo);
 806
 807         RETURN(0);
 808 }
 809
 810 /**
 811  * Set attributes of object during write bulk IO processing.
 812  *
 813  * Change object attributes and write parent FID into extended
 814  * attributes when needed.
 815  *
 816  * \param[in] env       execution environment
 817  * \param[in] ofd       OFD device
 818  * \param[in] ofd_obj   OFD object
 819  * \param[in] la        object attributes
 820  * \param[in] ff        parent FID
 821  *
 822  * \retval              0 on successful attributes update
 823  * \retval              negative value on error
 824  */
 825 static int
 826 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
 827                    struct ofd_object *ofd_obj, struct lu_attr *la,
 828                    struct filter_fid *ff)
 829 {
 830         struct ofd_thread_info  *info = ofd_info(env);
 831         __u64                    valid = la->la_valid;
 832         int                      rc;
 833         struct thandle          *th;
 834         struct dt_object        *dt_obj;
 835         int                      ff_needed = 0;
 836
 837         ENTRY;
 838
 839         LASSERT(la);
 840
 841         dt_obj = ofd_object_child(ofd_obj);
 842         LASSERT(dt_obj != NULL);
 843
 844         la->la_valid &= LA_UID | LA_GID | LA_PROJID;
 845
 846         rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
 847         if (rc != 0)
 848                 GOTO(out, rc);
 849
 850         if (ff != NULL) {
 851                 rc = ofd_object_ff_load(env, ofd_obj);
 852                 if (rc == -ENODATA)
 853                         ff_needed = 1;
 854                 else if (rc < 0)
 855                         GOTO(out, rc);
 856         }
 857
 858         if (!la->la_valid && !ff_needed)
 859                 /* no attributes to set */
 860                 GOTO(out, rc = 0);
 861
 862         th = ofd_trans_create(env, ofd);
 863         if (IS_ERR(th))
 864                 GOTO(out, rc = PTR_ERR(th));
 865
 866         if (la->la_valid) {
 867                 rc = dt_declare_attr_set(env, dt_obj, la, th);
 868                 if (rc)
 869                         GOTO(out_tx, rc);
 870         }
 871
 872         if (ff_needed) {
 873                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
 874                         ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
 875                 else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
 876                         le32_add_cpu(&ff->ff_parent.f_oid, -1);
 877
 878                 info->fti_buf.lb_buf = ff;
 879                 info->fti_buf.lb_len = sizeof(*ff);
 880                 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
 881                                           XATTR_NAME_FID, 0, th);
 882                 if (rc)
 883                         GOTO(out_tx, rc);
 884         }
 885
 886         /* We don't need a transno for this operation which will be re-executed
 887          * anyway when the OST_WRITE (with a transno assigned) is replayed */
 888         rc = dt_trans_start_local(env, ofd->ofd_osd , th);
 889         if (rc)
 890                 GOTO(out_tx, rc);
 891
 892         /* set uid/gid/projid */
 893         if (la->la_valid) {
 894                 rc = dt_attr_set(env, dt_obj, la, th);
 895                 if (rc)
 896                         GOTO(out_tx, rc);
 897         }
 898
 899         /* set filter fid EA */
 900         if (ff_needed) {
 901                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
 902                         GOTO(out_tx, rc);
 903
 904                 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID,
 905                                   0, th);
 906                 if (!rc)
 907                         filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff));
 908         }
 909
 910         GOTO(out_tx, rc);
 911
 912 out_tx:
 913         dt_trans_stop(env, ofd->ofd_osd, th);
 914 out:
 915         la->la_valid = valid;
 916         return rc;
 917 }
 918
 919 struct ofd_soft_sync_callback {
 920         struct dt_txn_commit_cb  ossc_cb;
 921         struct obd_export       *ossc_exp;
 922 };
 923
 924 /**
 925  * Callback function for "soft sync" update.
 926  *
 927  * Reset fed_soft_sync_count upon committing the "soft_sync" update.
 928  * See ofd_soft_sync_cb_add() below for more details on soft sync.
 929  *
 930  * \param[in] env       execution environment
 931  * \param[in] th        transaction handle
 932  * \param[in] cb        callback data
 933  * \param[in] err       error code
 934  */
 935 static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th,
 936                              struct dt_txn_commit_cb *cb, int err)
 937 {
 938         struct ofd_soft_sync_callback   *ossc;
 939
 940         ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb);
 941
 942         CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp);
 943         atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0);
 944
 945         class_export_cb_put(ossc->ossc_exp);
 946         OBD_FREE_PTR(ossc);
 947 }
 948
 949 /**
 950  * Add callback for "soft sync" processing.
 951  *
 952  * The "soft sync" mechanism does asynchronous commit when OBD_BRW_SOFT_SYNC
 953  * flag is set in client buffers. The intention is for this operation to
 954  * commit pages belonging to a client which has "too many" outstanding
 955  * unstable pages in its cache. See LU-2139 for details.
 956  *
 957  * This function adds callback to be called when commit is done.
 958  *
 959  * \param[in] th        transaction handle
 960  * \param[in] exp       OBD export of client
 961  *
 962  * \retval              0 on successful callback adding
 963  * \retval              negative value on error
 964  */
 965 static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
 966 {
 967         struct ofd_soft_sync_callback           *ossc;
 968         struct dt_txn_commit_cb                 *dcb;
 969         int                                      rc;
 970
 971         OBD_ALLOC_PTR(ossc);
 972         if (ossc == NULL)
 973                 return -ENOMEM;
 974
 975         ossc->ossc_exp = class_export_cb_get(exp);
 976
 977         dcb = &ossc->ossc_cb;
 978         dcb->dcb_func = ofd_cb_soft_sync;
 979         INIT_LIST_HEAD(&dcb->dcb_linkage);
 980         strlcpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name));
 981
 982         rc = dt_trans_cb_add(th, dcb);
 983         if (rc) {
 984                 class_export_cb_put(exp);
 985                 OBD_FREE_PTR(ossc);
 986         }
 987
 988         return rc;
 989 }
 990
 991 /**
 992  * Commit bulk IO buffers to the storage.
 993  *
 994  * This function finalizes write IO processing by writing data to the disk.
 995  * That write can be synchronous or asynchronous depending on buffers flags.
 996  *
 997  * \param[in] env       execution environment
 998  * \param[in] exp       OBD export of client
 999  * \param[in] ofd       OFD device
1000  * \param[in] fid       FID of object
1001  * \param[in] la        object attributes
1002  * \param[in] ff        parent FID of object
1003  * \param[in] objcount  always 1
1004  * \param[in] niocount  number of local buffers
1005  * \param[in] lnb       local buffers
1006  * \param[in] granted   grant space consumed for the bulk I/O
1007  * \param[in] old_rc    result of processing at this point
1008  *
1009  * \retval              0 on successful commit
1010  * \retval              negative value on error
1011  */
1012 static int
1013 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
1014                    struct ofd_device *ofd, const struct lu_fid *fid,
1015                    struct lu_attr *la, struct filter_fid *ff, int objcount,
1016                    int niocount, struct niobuf_local *lnb,
1017                    unsigned long granted, int old_rc)
1018 {
1019         struct filter_export_data *fed = &exp->exp_filter_data;
1020         struct ofd_object *fo;
1021         struct dt_object *o;
1022         struct thandle *th;
1023         int rc = 0;
1024         int rc2 = 0;
1025         int retries = 0;
1026         int i;
1027         bool soft_sync = false;
1028         bool cb_registered = false;
1029         bool fake_write = false;
1030
1031         ENTRY;
1032
1033         LASSERT(objcount == 1);
1034
1035         fo = ofd_object_find(env, ofd, fid);
1036         LASSERT(fo != NULL);
1037         LASSERT(ofd_object_exists(fo));
1038
1039         o = ofd_object_child(fo);
1040         LASSERT(o != NULL);
1041
1042         if (old_rc)
1043                 GOTO(out, rc = old_rc);
1044
1045         /*
1046          * The first write to each object must set some attributes.  It is
1047          * important to set the uid/gid before calling
1048          * dt_declare_write_commit() since quota enforcement is now handled in
1049          * declare phases.
1050          */
1051         rc = ofd_write_attr_set(env, ofd, fo, la, ff);
1052         if (rc)
1053                 GOTO(out, rc);
1054
1055         la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
1056
1057         /* do fake write, to simulate the write case for performance testing */
1058         if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW)) {
1059                 struct niobuf_local *last = &lnb[niocount - 1];
1060                 __u64 file_size = last->lnb_file_offset + last->lnb_len;
1061                 __u64 valid = la->la_valid;
1062
1063                 la->la_valid = LA_SIZE;
1064                 la->la_size = 0;
1065                 rc = dt_attr_get(env, o, la);
1066                 if (rc < 0 && rc != -ENOENT)
1067                         GOTO(out, rc);
1068
1069                 if (file_size < la->la_size)
1070                         file_size = la->la_size;
1071
1072                 /* dirty inode by setting file size */
1073                 la->la_valid = valid | LA_SIZE;
1074                 la->la_size = file_size;
1075
1076                 fake_write = true;
1077         }
1078
1079 retry:
1080         th = ofd_trans_create(env, ofd);
1081         if (IS_ERR(th))
1082                 GOTO(out, rc = PTR_ERR(th));
1083
1084         th->th_sync |= ofd->ofd_syncjournal;
1085         if (th->th_sync == 0) {
1086                 for (i = 0; i < niocount; i++) {
1087                         if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
1088                                 th->th_sync = 1;
1089                                 break;
1090                         }
1091                         if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC)
1092                                 soft_sync = true;
1093                 }
1094         }
1095
1096         if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
1097                 GOTO(out_stop, rc = -EINPROGRESS);
1098
1099         if (likely(!fake_write)) {
1100                 rc = dt_declare_write_commit(env, o, lnb, niocount, th);
1101                 if (rc)
1102                         GOTO(out_stop, rc);
1103         }
1104
1105         if (la->la_valid) {
1106                 /* update [mac]time if needed */
1107                 rc = dt_declare_attr_set(env, o, la, th);
1108                 if (rc)
1109                         GOTO(out_stop, rc);
1110         }
1111
1112         rc = ofd_trans_start(env, ofd, fo, th);
1113         if (rc)
1114                 GOTO(out_stop, rc);
1115
1116         if (likely(!fake_write)) {
1117                 rc = dt_write_commit(env, o, lnb, niocount, th);
1118                 if (rc)
1119                         GOTO(out_stop, rc);
1120         }
1121
1122         if (la->la_valid) {
1123                 rc = dt_attr_set(env, o, la, th);
1124                 if (rc)
1125                         GOTO(out_stop, rc);
1126         }
1127
1128         /* get attr to return */
1129         rc = dt_attr_get(env, o, la);
1130
1131 out_stop:
1132         /* Force commit to make the just-deleted blocks
1133          * reusable. LU-456 */
1134         if (rc == -ENOSPC)
1135                 th->th_sync = 1;
1136
1137         /* do this before trans stop in case commit has finished */
1138         if (!th->th_sync && soft_sync && !cb_registered) {
1139                 ofd_soft_sync_cb_add(th, exp);
1140                 cb_registered = true;
1141         }
1142
1143         if (rc == 0 && granted > 0) {
1144                 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
1145                         granted = 0;
1146         }
1147
1148         rc2 = ofd_trans_stop(env, ofd, th, rc);
1149         if (!rc)
1150                 rc = rc2;
1151         if (rc == -ENOSPC && retries++ < 3) {
1152                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
1153                        retries);
1154                 goto retry;
1155         }
1156
1157         if (!soft_sync)
1158                 /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
1159                 atomic_set(&fed->fed_soft_sync_count, 0);
1160         else if (atomic_inc_return(&fed->fed_soft_sync_count) ==
1161                  ofd->ofd_soft_sync_limit)
1162                 dt_commit_async(env, ofd->ofd_osd);
1163
1164 out:
1165         dt_bufs_put(env, o, lnb, niocount);
1166         ofd_read_unlock(env, fo);
1167         ofd_object_put(env, fo);
1168         /* second put is pair to object_get in ofd_preprw_write */
1169         ofd_object_put(env, fo);
1170         if (granted > 0)
1171                 tgt_grant_commit(exp, granted, old_rc);
1172         RETURN(rc);
1173 }
1174
1175 /**
1176  * Commit bulk IO to the storage.
1177  *
1178  * This is companion function to the ofd_preprw(). It finishes bulk IO
1179  * request processing by committing buffers to the storage (WRITE) and/or
1180  * freeing those buffers (read/write). See ofd_commitrw_read() and
1181  * ofd_commitrw_write() for details about each type of IO.
1182  *
1183  * \param[in] env       execution environment
1184  * \param[in] cmd       IO type (READ/WRITE)
1185  * \param[in] exp       OBD export of client
1186  * \param[in] oa        OBDO structure from client
1187  * \param[in] objcount  always 1
1188  * \param[in] obj       object data
1189  * \param[in] rnb       remote buffers
1190  * \param[in] npages    number of local buffers
1191  * \param[in] lnb       local buffers
1192  * \param[in] old_rc    result of processing at this point
1193  *
1194  * \retval              0 on successful commit
1195  * \retval              negative value on error
1196  */
1197 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
1198                  struct obdo *oa, int objcount, struct obd_ioobj *obj,
1199                  struct niobuf_remote *rnb, int npages,
1200                  struct niobuf_local *lnb, int old_rc)
1201 {
1202         struct ofd_thread_info  *info = ofd_info(env);
1203         struct ofd_mod_data     *fmd;
1204         __u64                    valid;
1205         struct ofd_device       *ofd = ofd_exp(exp);
1206         struct filter_fid       *ff = NULL;
1207         const struct lu_fid     *fid = &oa->o_oi.oi_fid;
1208         int                      rc = 0;
1209
1210         LASSERT(npages > 0);
1211
1212         if (cmd == OBD_BRW_WRITE) {
1213                 struct lu_nodemap *nodemap;
1214
1215                 /* Don't update timestamps if this write is older than a
1216                  * setattr which modifies the timestamps. b=10150 */
1217
1218                 /* XXX when we start having persistent reservations this needs
1219                  * to be changed to ofd_fmd_get() to create the fmd if it
1220                  * doesn't already exist so we can store the reservation handle
1221                  * there. */
1222                 valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID;
1223                 fmd = ofd_fmd_find(exp, fid);
1224                 if (!fmd || fmd->fmd_mactime_xid < info->fti_xid)
1225                         valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
1226                                  OBD_MD_FLCTIME;
1227                 ofd_fmd_put(exp, fmd);
1228                 la_from_obdo(&info->fti_attr, oa, valid);
1229
1230                 if (oa->o_valid & OBD_MD_FLFID) {
1231                         ff = &info->fti_mds_fid;
1232                         ofd_prepare_fidea(ff, oa);
1233                 }
1234
1235                 rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
1236                                         ff, objcount, npages, lnb,
1237                                         oa->o_grant_used, old_rc);
1238                 if (rc == 0)
1239                         obdo_from_la(oa, &info->fti_attr,
1240                                      OFD_VALID_FLAGS | LA_GID | LA_UID |
1241                                      LA_PROJID);
1242                 else
1243                         obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID |
1244                                      LA_PROJID);
1245
1246                 /* don't report overquota flag if we failed before reaching
1247                  * commit */
1248                 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
1249                         /* return the overquota flags to client */
1250                         if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
1251                                 if (oa->o_valid & OBD_MD_FLFLAGS)
1252                                         oa->o_flags |= OBD_FL_NO_USRQUOTA;
1253                                 else
1254                                         oa->o_flags = OBD_FL_NO_USRQUOTA;
1255                         }
1256
1257                         if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
1258                                 if (oa->o_valid & OBD_MD_FLFLAGS)
1259                                         oa->o_flags |= OBD_FL_NO_GRPQUOTA;
1260                                 else
1261                                         oa->o_flags = OBD_FL_NO_GRPQUOTA;
1262                         }
1263                         if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
1264                                 if (oa->o_valid & OBD_MD_FLFLAGS)
1265                                         oa->o_flags |= OBD_FL_NO_PRJQUOTA;
1266                                 else
1267                                         oa->o_flags = OBD_FL_NO_PRJQUOTA;
1268                         }
1269
1270                         oa->o_valid |= OBD_MD_FLFLAGS;
1271                         oa->o_valid |= OBD_MD_FLALLQUOTA;
1272                 }
1273
1274                 /* Convert back to client IDs. LU-9671.
1275                  * nodemap_get_from_exp() may fail due to nodemap deactivated,
1276                  * server ID will be returned back to client in that case. */
1277                 nodemap = nodemap_get_from_exp(exp);
1278                 if (nodemap != NULL && !IS_ERR(nodemap)) {
1279                         oa->o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
1280                                                    NODEMAP_FS_TO_CLIENT,
1281                                                    oa->o_uid);
1282                         oa->o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
1283                                                    NODEMAP_FS_TO_CLIENT,
1284                                                    oa->o_gid);
1285                         nodemap_putref(nodemap);
1286                 }
1287         } else if (cmd == OBD_BRW_READ) {
1288                 struct ldlm_namespace *ns = ofd->ofd_namespace;
1289
1290                 /* If oa != NULL then ofd_preprw_read updated the inode
1291                  * atime and we should update the lvb so that other glimpses
1292                  * will also get the updated value. bug 5972 */
1293                 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
1294                          struct ldlm_resource *rs = NULL;
1295
1296                         ost_fid_build_resid(fid, &info->fti_resid);
1297                         rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
1298                                                LDLM_EXTENT, 0);
1299                         if (!IS_ERR(rs)) {
1300                                 ldlm_res_lvbo_update(rs, NULL, 1);
1301                                 ldlm_resource_putref(rs);
1302                         }
1303                 }
1304                 rc = ofd_commitrw_read(env, ofd, fid, objcount,
1305                                        npages, lnb);
1306                 if (old_rc)
1307                         rc = old_rc;
1308         } else {
1309                 LBUG();
1310                 rc = -EPROTO;
1311         }
1312
1313         RETURN(rc);
1314 }