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
 459         ENTRY;
 460         LASSERT(env != NULL);
 461
 462         fo = ofd_object_find(env, ofd, fid);
 463         if (IS_ERR(fo))
 464                 RETURN(PTR_ERR(fo));
 465         LASSERT(fo != NULL);
 466
 467         ofd_read_lock(env, fo);
 468         if (!ofd_object_exists(fo))
 469                 GOTO(unlock, rc = -ENOENT);
 470
 471         if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
 472                 rc = ofd_verify_ff(env, fo, oa);
 473                 if (rc != 0)
 474                         GOTO(unlock, rc);
 475         }
 476
 477         *nr_local = 0;
 478         for (i = 0, j = 0; i < niocount; i++) {
 479                 rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
 480                                  lnb + j, 0);
 481                 if (unlikely(rc < 0))
 482                         GOTO(buf_put, rc);
 483                 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
 484                 /* correct index for local buffers to continue with */
 485                 j += rc;
 486                 *nr_local += rc;
 487                 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
 488                 tot_bytes += rnb[i].rnb_len;
 489         }
 490
 491         LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
 492         rc = dt_attr_get(env, ofd_object_child(fo), la);
 493         if (unlikely(rc))
 494                 GOTO(buf_put, rc);
 495
 496         rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
 497         if (unlikely(rc))
 498                 GOTO(buf_put, rc);
 499
 500         ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes);
 501         RETURN(0);
 502
 503 buf_put:
 504         dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
 505 unlock:
 506         ofd_read_unlock(env, fo);
 507         ofd_object_put(env, fo);
 508         return rc;
 509 }
 510
 511 /**
 512  * Prepare buffers for write request processing.
 513  *
 514  * This function converts remote buffers from client to local buffers
 515  * and prepares the latter. If there is recovery in progress and required
 516  * object is missing then it can be re-created before write.
 517  *
 518  * \param[in] env       execution environment
 519  * \param[in] exp       OBD export of client
 520  * \param[in] ofd       OFD device
 521  * \param[in] fid       FID of object
 522  * \param[in] la        object attributes
 523  * \param[in] oa        OBDO structure from client
 524  * \param[in] objcount  always 1
 525  * \param[in] obj       object data
 526  * \param[in] rnb       remote buffers
 527  * \param[in] nr_local  number of local buffers
 528  * \param[in] lnb       local buffers
 529  * \param[in] jobid     job ID name
 530  *
 531  * \retval              0 on successful prepare
 532  * \retval              negative value on error
 533  */
 534 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
 535                             struct ofd_device *ofd, const struct lu_fid *fid,
 536                             struct lu_attr *la, struct obdo *oa,
 537                             int objcount, struct obd_ioobj *obj,
 538                             struct niobuf_remote *rnb, int *nr_local,
 539                             struct niobuf_local *lnb, char *jobid)
 540 {
 541         struct ofd_object       *fo;
 542         int                      i, j, k, rc = 0, tot_bytes = 0;
 543
 544         ENTRY;
 545         LASSERT(env != NULL);
 546         LASSERT(objcount == 1);
 547
 548         if (unlikely(exp->exp_obd->obd_recovering)) {
 549                 u64 seq = fid_seq(fid);
 550                 u64 oid = fid_oid(fid);
 551                 struct ofd_seq *oseq;
 552
 553                 oseq = ofd_seq_load(env, ofd, seq);
 554                 if (IS_ERR(oseq)) {
 555                         CERROR("%s: Can't find FID Sequence %#llx: rc = %d\n",
 556                                ofd_name(ofd), seq, (int)PTR_ERR(oseq));
 557                         GOTO(out, rc = -EINVAL);
 558                 }
 559
 560                 if (oid > ofd_seq_last_oid(oseq)) {
 561                         int sync = 0;
 562                         int diff;
 563
 564                         mutex_lock(&oseq->os_create_lock);
 565                         diff = oid - ofd_seq_last_oid(oseq);
 566
 567                         /* Do sync create if the seq is about to used up */
 568                         if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) {
 569                                 if (unlikely(oid >= IDIF_MAX_OID - 1))
 570                                         sync = 1;
 571                         } else if (fid_seq_is_norm(seq)) {
 572                                 if (unlikely(oid >=
 573                                              LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
 574                                         sync = 1;
 575                         } else {
 576                                 CERROR("%s : invalid o_seq "DOSTID"\n",
 577                                        ofd_name(ofd), POSTID(&oa->o_oi));
 578                                 mutex_unlock(&oseq->os_create_lock);
 579                                 ofd_seq_put(env, oseq);
 580                                 GOTO(out, rc = -EINVAL);
 581                         }
 582
 583                         while (diff > 0) {
 584                                 u64 next_id = ofd_seq_last_oid(oseq) + 1;
 585                                 int count = ofd_precreate_batch(ofd, diff);
 586
 587                                 rc = ofd_precreate_objects(env, ofd, next_id,
 588                                                            oseq, count, sync);
 589                                 if (rc < 0) {
 590                                         mutex_unlock(&oseq->os_create_lock);
 591                                         ofd_seq_put(env, oseq);
 592                                         GOTO(out, rc);
 593                                 }
 594
 595                                 diff -= rc;
 596                         }
 597
 598                         mutex_unlock(&oseq->os_create_lock);
 599                 }
 600
 601                 ofd_seq_put(env, oseq);
 602         }
 603
 604         fo = ofd_object_find(env, ofd, fid);
 605         if (IS_ERR(fo))
 606                 GOTO(out, rc = PTR_ERR(fo));
 607         LASSERT(fo != NULL);
 608
 609         ofd_read_lock(env, fo);
 610         if (!ofd_object_exists(fo)) {
 611                 CERROR("%s: BRW to missing obj "DOSTID"\n",
 612                        exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
 613                 ofd_read_unlock(env, fo);
 614                 ofd_object_put(env, fo);
 615                 GOTO(out, rc = -ENOENT);
 616         }
 617
 618         if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
 619                 rc = ofd_verify_ff(env, fo, oa);
 620                 if (rc != 0) {
 621                         ofd_read_unlock(env, fo);
 622                         ofd_object_put(env, fo);
 623                         GOTO(out, rc);
 624                 }
 625         }
 626
 627         /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
 628          * space back if possible */
 629         tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
 630
 631         /* parse remote buffers to local buffers and prepare the latter */
 632         *nr_local = 0;
 633         for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
 634                 rc = dt_bufs_get(env, ofd_object_child(fo),
 635                                  rnb + i, lnb + j, 1);
 636                 if (unlikely(rc < 0))
 637                         GOTO(err, rc);
 638                 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
 639                 /* correct index for local buffers to continue with */
 640                 for (k = 0; k < rc; k++) {
 641                         lnb[j+k].lnb_flags = rnb[i].rnb_flags;
 642                         if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
 643                                 lnb[j+k].lnb_rc = -ENOSPC;
 644                 }
 645                 j += rc;
 646                 *nr_local += rc;
 647                 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
 648                 tot_bytes += rnb[i].rnb_len;
 649         }
 650         LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
 651
 652         rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
 653         if (unlikely(rc != 0))
 654                 GOTO(err, rc);
 655
 656         ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes);
 657         RETURN(0);
 658 err:
 659         dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
 660         ofd_read_unlock(env, fo);
 661         ofd_object_put(env, fo);
 662         /* tgt_grant_prepare_write() was called, so we must commit */
 663         tgt_grant_commit(exp, oa->o_grant_used, rc);
 664 out:
 665         /* let's still process incoming grant information packed in the oa,
 666          * but without enforcing grant since we won't proceed with the write.
 667          * Just like a read request actually. */
 668         tgt_grant_prepare_read(env, exp, oa);
 669         return rc;
 670 }
 671
 672 /**
 673  * Prepare bulk IO requests for processing.
 674  *
 675  * This function does initial checks of IO and calls corresponding
 676  * functions for read/write processing.
 677  *
 678  * \param[in] env       execution environment
 679  * \param[in] cmd       IO type (read/write)
 680  * \param[in] exp       OBD export of client
 681  * \param[in] oa        OBDO structure from request
 682  * \param[in] objcount  always 1
 683  * \param[in] obj       object data
 684  * \param[in] rnb       remote buffers
 685  * \param[in] nr_local  number of local buffers
 686  * \param[in] lnb       local buffers
 687  *
 688  * \retval              0 on successful prepare
 689  * \retval              negative value on error
 690  */
 691 int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
 692                struct obdo *oa, int objcount, struct obd_ioobj *obj,
 693                struct niobuf_remote *rnb, int *nr_local,
 694                struct niobuf_local *lnb)
 695 {
 696         struct tgt_session_info *tsi = tgt_ses_info(env);
 697         struct ofd_device       *ofd = ofd_exp(exp);
 698         struct ofd_thread_info  *info;
 699         char                    *jobid;
 700         const struct lu_fid     *fid = &oa->o_oi.oi_fid;
 701         int                      rc = 0;
 702
 703         if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
 704                 CERROR("%s: bulk has too many pages %d, which exceeds the"
 705                        "maximum pages per RPC of %d\n",
 706                        exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
 707                 RETURN(-EPROTO);
 708         }
 709
 710         if (tgt_ses_req(tsi) == NULL) { /* echo client case */
 711                 info = ofd_info_init(env, exp);
 712                 jobid = NULL;
 713         } else {
 714                 info = tsi2ofd_info(tsi);
 715                 jobid = tsi->tsi_jobid;
 716         }
 717
 718         LASSERT(oa != NULL);
 719
 720         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) {
 721                 struct ofd_seq          *oseq;
 722
 723                 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
 724                 if (IS_ERR(oseq)) {
 725                         CERROR("%s: Can not find seq for "DOSTID
 726                                ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
 727                                PTR_ERR(oseq));
 728                         RETURN(-EINVAL);
 729                 }
 730
 731                 if (oseq->os_destroys_in_progress == 0) {
 732                         /* don't fail lookups for orphan recovery, it causes
 733                          * later LBUGs when objects still exist during
 734                          * precreate */
 735                         ofd_seq_put(env, oseq);
 736                         RETURN(-ENOENT);
 737                 }
 738                 ofd_seq_put(env, oseq);
 739         }
 740
 741         LASSERT(objcount == 1);
 742         LASSERT(obj->ioo_bufcnt > 0);
 743
 744         if (cmd == OBD_BRW_WRITE) {
 745                 la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
 746                 rc = ofd_preprw_write(env, exp, ofd, fid, &info->fti_attr, oa,
 747                                       objcount, obj, rnb, nr_local, lnb, jobid);
 748         } else if (cmd == OBD_BRW_READ) {
 749                 tgt_grant_prepare_read(env, exp, oa);
 750                 rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa,
 751                                      obj->ioo_bufcnt, rnb, nr_local, lnb,
 752                                      jobid);
 753                 obdo_from_la(oa, &info->fti_attr, LA_ATIME);
 754         } else {
 755                 CERROR("%s: wrong cmd %d received!\n",
 756                        exp->exp_obd->obd_name, cmd);
 757                 rc = -EPROTO;
 758         }
 759         RETURN(rc);
 760 }
 761
 762 /**
 763  * Drop reference on local buffers for read bulk IO.
 764  *
 765  * This will free all local buffers use by this read request.
 766  *
 767  * \param[in] env       execution environment
 768  * \param[in] ofd       OFD device
 769  * \param[in] fid       object FID
 770  * \param[in] objcount  always 1
 771  * \param[in] niocount  number of local buffers
 772  * \param[in] lnb       local buffers
 773  *
 774  * \retval              0 on successful execution
 775  * \retval              negative value on error
 776  */
 777 static int
 778 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
 779                   const struct lu_fid *fid, int objcount, int niocount,
 780                   struct niobuf_local *lnb)
 781 {
 782         struct ofd_object *fo;
 783
 784         ENTRY;
 785
 786         LASSERT(niocount > 0);
 787
 788         fo = ofd_object_find(env, ofd, fid);
 789         if (IS_ERR(fo))
 790                 RETURN(PTR_ERR(fo));
 791         LASSERT(fo != NULL);
 792         LASSERT(ofd_object_exists(fo));
 793         dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
 794
 795         ofd_read_unlock(env, fo);
 796         ofd_object_put(env, fo);
 797         /* second put is pair to object_get in ofd_preprw_read */
 798         ofd_object_put(env, fo);
 799
 800         RETURN(0);
 801 }
 802
 803 /**
 804  * Set attributes of object during write bulk IO processing.
 805  *
 806  * Change object attributes and write parent FID into extended
 807  * attributes when needed.
 808  *
 809  * \param[in] env       execution environment
 810  * \param[in] ofd       OFD device
 811  * \param[in] ofd_obj   OFD object
 812  * \param[in] la        object attributes
 813  * \param[in] ff        parent FID
 814  *
 815  * \retval              0 on successful attributes update
 816  * \retval              negative value on error
 817  */
 818 static int
 819 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
 820                    struct ofd_object *ofd_obj, struct lu_attr *la,
 821                    struct filter_fid *ff)
 822 {
 823         struct ofd_thread_info  *info = ofd_info(env);
 824         __u64                    valid = la->la_valid;
 825         int                      rc;
 826         struct thandle          *th;
 827         struct dt_object        *dt_obj;
 828         int                      ff_needed = 0;
 829
 830         ENTRY;
 831
 832         LASSERT(la);
 833
 834         dt_obj = ofd_object_child(ofd_obj);
 835         LASSERT(dt_obj != NULL);
 836
 837         la->la_valid &= LA_UID | LA_GID | LA_PROJID;
 838
 839         rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
 840         if (rc != 0)
 841                 GOTO(out, rc);
 842
 843         if (ff != NULL) {
 844                 rc = ofd_object_ff_load(env, ofd_obj);
 845                 if (rc == -ENODATA)
 846                         ff_needed = 1;
 847                 else if (rc < 0)
 848                         GOTO(out, rc);
 849         }
 850
 851         if (!la->la_valid && !ff_needed)
 852                 /* no attributes to set */
 853                 GOTO(out, rc = 0);
 854
 855         th = ofd_trans_create(env, ofd);
 856         if (IS_ERR(th))
 857                 GOTO(out, rc = PTR_ERR(th));
 858
 859         if (la->la_valid) {
 860                 rc = dt_declare_attr_set(env, dt_obj, la, th);
 861                 if (rc)
 862                         GOTO(out_tx, rc);
 863         }
 864
 865         if (ff_needed) {
 866                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
 867                         ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
 868                 else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
 869                         le32_add_cpu(&ff->ff_parent.f_oid, -1);
 870
 871                 info->fti_buf.lb_buf = ff;
 872                 info->fti_buf.lb_len = sizeof(*ff);
 873                 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
 874                                           XATTR_NAME_FID, 0, th);
 875                 if (rc)
 876                         GOTO(out_tx, rc);
 877         }
 878
 879         /* We don't need a transno for this operation which will be re-executed
 880          * anyway when the OST_WRITE (with a transno assigned) is replayed */
 881         rc = dt_trans_start_local(env, ofd->ofd_osd , th);
 882         if (rc)
 883                 GOTO(out_tx, rc);
 884
 885         /* set uid/gid/projid */
 886         if (la->la_valid) {
 887                 rc = dt_attr_set(env, dt_obj, la, th);
 888                 if (rc)
 889                         GOTO(out_tx, rc);
 890         }
 891
 892         /* set filter fid EA */
 893         if (ff_needed) {
 894                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
 895                         GOTO(out_tx, rc);
 896
 897                 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID,
 898                                   0, th);
 899                 if (!rc)
 900                         filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff));
 901         }
 902
 903         GOTO(out_tx, rc);
 904
 905 out_tx:
 906         dt_trans_stop(env, ofd->ofd_osd, th);
 907 out:
 908         la->la_valid = valid;
 909         return rc;
 910 }
 911
 912 struct ofd_soft_sync_callback {
 913         struct dt_txn_commit_cb  ossc_cb;
 914         struct obd_export       *ossc_exp;
 915 };
 916
 917 /**
 918  * Callback function for "soft sync" update.
 919  *
 920  * Reset fed_soft_sync_count upon committing the "soft_sync" update.
 921  * See ofd_soft_sync_cb_add() below for more details on soft sync.
 922  *
 923  * \param[in] env       execution environment
 924  * \param[in] th        transaction handle
 925  * \param[in] cb        callback data
 926  * \param[in] err       error code
 927  */
 928 static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th,
 929                              struct dt_txn_commit_cb *cb, int err)
 930 {
 931         struct ofd_soft_sync_callback   *ossc;
 932
 933         ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb);
 934
 935         CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp);
 936         atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0);
 937
 938         class_export_cb_put(ossc->ossc_exp);
 939         OBD_FREE_PTR(ossc);
 940 }
 941
 942 /**
 943  * Add callback for "soft sync" processing.
 944  *
 945  * The "soft sync" mechanism does asynchronous commit when OBD_BRW_SOFT_SYNC
 946  * flag is set in client buffers. The intention is for this operation to
 947  * commit pages belonging to a client which has "too many" outstanding
 948  * unstable pages in its cache. See LU-2139 for details.
 949  *
 950  * This function adds callback to be called when commit is done.
 951  *
 952  * \param[in] th        transaction handle
 953  * \param[in] exp       OBD export of client
 954  *
 955  * \retval              0 on successful callback adding
 956  * \retval              negative value on error
 957  */
 958 static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
 959 {
 960         struct ofd_soft_sync_callback           *ossc;
 961         struct dt_txn_commit_cb                 *dcb;
 962         int                                      rc;
 963
 964         OBD_ALLOC_PTR(ossc);
 965         if (ossc == NULL)
 966                 return -ENOMEM;
 967
 968         ossc->ossc_exp = class_export_cb_get(exp);
 969
 970         dcb = &ossc->ossc_cb;
 971         dcb->dcb_func = ofd_cb_soft_sync;
 972         INIT_LIST_HEAD(&dcb->dcb_linkage);
 973         strlcpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name));
 974
 975         rc = dt_trans_cb_add(th, dcb);
 976         if (rc) {
 977                 class_export_cb_put(exp);
 978                 OBD_FREE_PTR(ossc);
 979         }
 980
 981         return rc;
 982 }
 983
 984 /**
 985  * Commit bulk IO buffers to the storage.
 986  *
 987  * This function finalizes write IO processing by writing data to the disk.
 988  * That write can be synchronous or asynchronous depending on buffers flags.
 989  *
 990  * \param[in] env       execution environment
 991  * \param[in] exp       OBD export of client
 992  * \param[in] ofd       OFD device
 993  * \param[in] fid       FID of object
 994  * \param[in] la        object attributes
 995  * \param[in] ff        parent FID of object
 996  * \param[in] objcount  always 1
 997  * \param[in] niocount  number of local buffers
 998  * \param[in] lnb       local buffers
 999  * \param[in] granted   grant space consumed for the bulk I/O
1000  * \param[in] old_rc    result of processing at this point
1001  *
1002  * \retval              0 on successful commit
1003  * \retval              negative value on error
1004  */
1005 static int
1006 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
1007                    struct ofd_device *ofd, const struct lu_fid *fid,
1008                    struct lu_attr *la, struct filter_fid *ff, int objcount,
1009                    int niocount, struct niobuf_local *lnb,
1010                    unsigned long granted, int old_rc)
1011 {
1012         struct filter_export_data *fed = &exp->exp_filter_data;
1013         struct ofd_object *fo;
1014         struct dt_object *o;
1015         struct thandle *th;
1016         int rc = 0;
1017         int rc2 = 0;
1018         int retries = 0;
1019         int i;
1020         bool soft_sync = false;
1021         bool cb_registered = false;
1022         bool fake_write = false;
1023
1024         ENTRY;
1025
1026         LASSERT(objcount == 1);
1027
1028         fo = ofd_object_find(env, ofd, fid);
1029         LASSERT(fo != NULL);
1030         LASSERT(ofd_object_exists(fo));
1031
1032         o = ofd_object_child(fo);
1033         LASSERT(o != NULL);
1034
1035         if (old_rc)
1036                 GOTO(out, rc = old_rc);
1037
1038         /*
1039          * The first write to each object must set some attributes.  It is
1040          * important to set the uid/gid before calling
1041          * dt_declare_write_commit() since quota enforcement is now handled in
1042          * declare phases.
1043          */
1044         rc = ofd_write_attr_set(env, ofd, fo, la, ff);
1045         if (rc)
1046                 GOTO(out, rc);
1047
1048         la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
1049
1050         /* do fake write, to simulate the write case for performance testing */
1051         if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW)) {
1052                 struct niobuf_local *last = &lnb[niocount - 1];
1053                 __u64 file_size = last->lnb_file_offset + last->lnb_len;
1054                 __u64 valid = la->la_valid;
1055
1056                 la->la_valid = LA_SIZE;
1057                 la->la_size = 0;
1058                 rc = dt_attr_get(env, o, la);
1059                 if (rc < 0 && rc != -ENOENT)
1060                         GOTO(out, rc);
1061
1062                 if (file_size < la->la_size)
1063                         file_size = la->la_size;
1064
1065                 /* dirty inode by setting file size */
1066                 la->la_valid = valid | LA_SIZE;
1067                 la->la_size = file_size;
1068
1069                 fake_write = true;
1070         }
1071
1072 retry:
1073         th = ofd_trans_create(env, ofd);
1074         if (IS_ERR(th))
1075                 GOTO(out, rc = PTR_ERR(th));
1076
1077         th->th_sync |= ofd->ofd_syncjournal;
1078         if (th->th_sync == 0) {
1079                 for (i = 0; i < niocount; i++) {
1080                         if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
1081                                 th->th_sync = 1;
1082                                 break;
1083                         }
1084                         if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC)
1085                                 soft_sync = true;
1086                 }
1087         }
1088
1089         if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
1090                 GOTO(out_stop, rc = -EINPROGRESS);
1091
1092         if (likely(!fake_write)) {
1093                 rc = dt_declare_write_commit(env, o, lnb, niocount, th);
1094                 if (rc)
1095                         GOTO(out_stop, rc);
1096         }
1097
1098         if (la->la_valid) {
1099                 /* update [mac]time if needed */
1100                 rc = dt_declare_attr_set(env, o, la, th);
1101                 if (rc)
1102                         GOTO(out_stop, rc);
1103         }
1104
1105         rc = ofd_trans_start(env, ofd, fo, th);
1106         if (rc)
1107                 GOTO(out_stop, rc);
1108
1109         if (likely(!fake_write)) {
1110                 rc = dt_write_commit(env, o, lnb, niocount, th);
1111                 if (rc)
1112                         GOTO(out_stop, rc);
1113         }
1114
1115         if (la->la_valid) {
1116                 rc = dt_attr_set(env, o, la, th);
1117                 if (rc)
1118                         GOTO(out_stop, rc);
1119         }
1120
1121         /* get attr to return */
1122         rc = dt_attr_get(env, o, la);
1123
1124 out_stop:
1125         /* Force commit to make the just-deleted blocks
1126          * reusable. LU-456 */
1127         if (rc == -ENOSPC)
1128                 th->th_sync = 1;
1129
1130         /* do this before trans stop in case commit has finished */
1131         if (!th->th_sync && soft_sync && !cb_registered) {
1132                 ofd_soft_sync_cb_add(th, exp);
1133                 cb_registered = true;
1134         }
1135
1136         if (rc == 0 && granted > 0) {
1137                 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
1138                         granted = 0;
1139         }
1140
1141         rc2 = ofd_trans_stop(env, ofd, th, rc);
1142         if (!rc)
1143                 rc = rc2;
1144         if (rc == -ENOSPC && retries++ < 3) {
1145                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
1146                        retries);
1147                 goto retry;
1148         }
1149
1150         if (!soft_sync)
1151                 /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
1152                 atomic_set(&fed->fed_soft_sync_count, 0);
1153         else if (atomic_inc_return(&fed->fed_soft_sync_count) ==
1154                  ofd->ofd_soft_sync_limit)
1155                 dt_commit_async(env, ofd->ofd_osd);
1156
1157 out:
1158         dt_bufs_put(env, o, lnb, niocount);
1159         ofd_read_unlock(env, fo);
1160         ofd_object_put(env, fo);
1161         /* second put is pair to object_get in ofd_preprw_write */
1162         ofd_object_put(env, fo);
1163         if (granted > 0)
1164                 tgt_grant_commit(exp, granted, old_rc);
1165         RETURN(rc);
1166 }
1167
1168 /**
1169  * Commit bulk IO to the storage.
1170  *
1171  * This is companion function to the ofd_preprw(). It finishes bulk IO
1172  * request processing by committing buffers to the storage (WRITE) and/or
1173  * freeing those buffers (read/write). See ofd_commitrw_read() and
1174  * ofd_commitrw_write() for details about each type of IO.
1175  *
1176  * \param[in] env       execution environment
1177  * \param[in] cmd       IO type (READ/WRITE)
1178  * \param[in] exp       OBD export of client
1179  * \param[in] oa        OBDO structure from client
1180  * \param[in] objcount  always 1
1181  * \param[in] obj       object data
1182  * \param[in] rnb       remote buffers
1183  * \param[in] npages    number of local buffers
1184  * \param[in] lnb       local buffers
1185  * \param[in] old_rc    result of processing at this point
1186  *
1187  * \retval              0 on successful commit
1188  * \retval              negative value on error
1189  */
1190 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
1191                  struct obdo *oa, int objcount, struct obd_ioobj *obj,
1192                  struct niobuf_remote *rnb, int npages,
1193                  struct niobuf_local *lnb, int old_rc)
1194 {
1195         struct ofd_thread_info  *info = ofd_info(env);
1196         struct ofd_mod_data     *fmd;
1197         __u64                    valid;
1198         struct ofd_device       *ofd = ofd_exp(exp);
1199         struct filter_fid       *ff = NULL;
1200         const struct lu_fid     *fid = &oa->o_oi.oi_fid;
1201         int                      rc = 0;
1202
1203         LASSERT(npages > 0);
1204
1205         if (cmd == OBD_BRW_WRITE) {
1206                 struct lu_nodemap *nodemap;
1207
1208                 /* Don't update timestamps if this write is older than a
1209                  * setattr which modifies the timestamps. b=10150 */
1210
1211                 /* XXX when we start having persistent reservations this needs
1212                  * to be changed to ofd_fmd_get() to create the fmd if it
1213                  * doesn't already exist so we can store the reservation handle
1214                  * there. */
1215                 valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID;
1216                 fmd = ofd_fmd_find(exp, fid);
1217                 if (!fmd || fmd->fmd_mactime_xid < info->fti_xid)
1218                         valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
1219                                  OBD_MD_FLCTIME;
1220                 ofd_fmd_put(exp, fmd);
1221                 la_from_obdo(&info->fti_attr, oa, valid);
1222
1223                 if (oa->o_valid & OBD_MD_FLFID) {
1224                         ff = &info->fti_mds_fid;
1225                         ofd_prepare_fidea(ff, oa);
1226                 }
1227
1228                 rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
1229                                         ff, objcount, npages, lnb,
1230                                         oa->o_grant_used, old_rc);
1231                 if (rc == 0)
1232                         obdo_from_la(oa, &info->fti_attr,
1233                                      OFD_VALID_FLAGS | LA_GID | LA_UID |
1234                                      LA_PROJID);
1235                 else
1236                         obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID |
1237                                      LA_PROJID);
1238
1239                 /* don't report overquota flag if we failed before reaching
1240                  * commit */
1241                 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
1242                         /* return the overquota flags to client */
1243                         if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
1244                                 if (oa->o_valid & OBD_MD_FLFLAGS)
1245                                         oa->o_flags |= OBD_FL_NO_USRQUOTA;
1246                                 else
1247                                         oa->o_flags = OBD_FL_NO_USRQUOTA;
1248                         }
1249
1250                         if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
1251                                 if (oa->o_valid & OBD_MD_FLFLAGS)
1252                                         oa->o_flags |= OBD_FL_NO_GRPQUOTA;
1253                                 else
1254                                         oa->o_flags = OBD_FL_NO_GRPQUOTA;
1255                         }
1256                         if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
1257                                 if (oa->o_valid & OBD_MD_FLFLAGS)
1258                                         oa->o_flags |= OBD_FL_NO_PRJQUOTA;
1259                                 else
1260                                         oa->o_flags = OBD_FL_NO_PRJQUOTA;
1261                         }
1262
1263                         oa->o_valid |= OBD_MD_FLFLAGS;
1264                         oa->o_valid |= OBD_MD_FLALLQUOTA;
1265                 }
1266
1267                 /* Convert back to client IDs. LU-9671.
1268                  * nodemap_get_from_exp() may fail due to nodemap deactivated,
1269                  * server ID will be returned back to client in that case. */
1270                 nodemap = nodemap_get_from_exp(exp);
1271                 if (nodemap != NULL && !IS_ERR(nodemap)) {
1272                         oa->o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
1273                                                    NODEMAP_FS_TO_CLIENT,
1274                                                    oa->o_uid);
1275                         oa->o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
1276                                                    NODEMAP_FS_TO_CLIENT,
1277                                                    oa->o_gid);
1278                         nodemap_putref(nodemap);
1279                 }
1280         } else if (cmd == OBD_BRW_READ) {
1281                 struct ldlm_namespace *ns = ofd->ofd_namespace;
1282
1283                 /* If oa != NULL then ofd_preprw_read updated the inode
1284                  * atime and we should update the lvb so that other glimpses
1285                  * will also get the updated value. bug 5972 */
1286                 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
1287                          struct ldlm_resource *rs = NULL;
1288
1289                         ost_fid_build_resid(fid, &info->fti_resid);
1290                         rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
1291                                                LDLM_EXTENT, 0);
1292                         if (!IS_ERR(rs)) {
1293                                 ldlm_res_lvbo_update(rs, NULL, 1);
1294                                 ldlm_resource_putref(rs);
1295                         }
1296                 }
1297                 rc = ofd_commitrw_read(env, ofd, fid, objcount,
1298                                        npages, lnb);
1299                 if (old_rc)
1300                         rc = old_rc;
1301         } else {
1302                 LBUG();
1303                 rc = -EPROTO;
1304         }
1305
1306         RETURN(rc);
1307 }