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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ofd/ofd_io.c
  37  *
  38  * Author: Alex Tomas <bzzz@whamcloud.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_FILTER
  42
  43 #include "ofd_internal.h"
  44
  45 struct ofd_inconsistency_item {
  46         struct list_head         oii_list;
  47         struct ofd_object       *oii_obj;
  48         struct lu_fid            oii_pfid;
  49 };
  50
  51 static void ofd_inconsistency_verify_one(const struct lu_env *env,
  52                                          struct ofd_device *ofd,
  53                                          struct ofd_inconsistency_item *oii,
  54                                          struct lfsck_request *lr)
  55 {
  56         struct ofd_object       *fo     = oii->oii_obj;
  57         struct lu_fid           *pfid   = &fo->ofo_pfid;
  58         int                      rc;
  59
  60         LASSERT(fo->ofo_pfid_checking);
  61         LASSERT(!fo->ofo_pfid_verified);
  62
  63         lr->lr_fid = fo->ofo_header.loh_fid; /* OST-object itself FID. */
  64         lr->lr_fid2 = oii->oii_pfid; /* client given PFID. */
  65         lr->lr_fid3 = *pfid; /* OST local stored PFID. */
  66
  67         rc = lfsck_in_notify(env, ofd->ofd_osd, lr);
  68         ofd_write_lock(env, fo);
  69         switch (lr->lr_status) {
  70         case LPVS_INIT:
  71                 LASSERT(rc <= 0);
  72
  73                 if (rc < 0)
  74                         CDEBUG(D_LFSCK, "%s: fail to verify OST local stored "
  75                                "PFID xattr for "DFID", the client given PFID "
  76                                DFID", OST local stored PFID "DFID": rc = %d\n",
  77                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
  78                                PFID(&oii->oii_pfid), PFID(pfid), rc);
  79                 else
  80                         fo->ofo_pfid_verified = 1;
  81                 break;
  82         case LPVS_INCONSISTENT:
  83                 LASSERT(rc != 0);
  84
  85                 ofd->ofd_inconsistency_self_detected++;
  86                 if (rc < 0)
  87                         CDEBUG(D_LFSCK, "%s: fail to verify the client given "
  88                                "PFID for "DFID", the client given PFID "DFID
  89                                ", local stored PFID "DFID": rc = %d\n",
  90                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
  91                                PFID(&oii->oii_pfid), PFID(pfid), rc);
  92                 else
  93                         CDEBUG(D_LFSCK, "%s: both the client given PFID and "
  94                                "the OST local stored PFID are stale for the "
  95                                "OST-object "DFID", client given PFID is "DFID
  96                                ", local stored PFID is "DFID"\n",
  97                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
  98                                PFID(&oii->oii_pfid), PFID(pfid));
  99                 break;
 100         case LPVS_INCONSISTENT_TOFIX:
 101                 ofd->ofd_inconsistency_self_detected++;
 102                 if (rc == 0) {
 103                         ofd->ofd_inconsistency_self_repaired++;
 104                         CDEBUG(D_LFSCK, "%s: fixed the staled OST PFID xattr "
 105                                "for "DFID", with the client given PFID "DFID
 106                                ", the old stored PFID "DFID"\n",
 107                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
 108                                PFID(&oii->oii_pfid), PFID(pfid));
 109                 } else {
 110                         CDEBUG(D_LFSCK, "%s: fail to fix the OST PFID xattr "
 111                                "for "DFID", client given PFID "DFID", local "
 112                                "stored PFID "DFID": rc = %d\n",
 113                                ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
 114                                PFID(&oii->oii_pfid), PFID(pfid), rc);
 115                 }
 116                 *pfid = oii->oii_pfid;
 117                 fo->ofo_pfid_verified = 1;
 118                 break;
 119         default:
 120                 break;
 121         }
 122         fo->ofo_pfid_checking = 0;
 123         ofd_write_unlock(env, fo);
 124
 125         lu_object_put(env, &fo->ofo_obj.do_lu);
 126         OBD_FREE_PTR(oii);
 127 }
 128
 129 static int ofd_inconsistency_verification_main(void *args)
 130 {
 131         struct lu_env                  env;
 132         struct ofd_device             *ofd    = args;
 133         struct ptlrpc_thread          *thread = &ofd->ofd_inconsistency_thread;
 134         struct ofd_inconsistency_item *oii;
 135         struct lfsck_request          *lr     = NULL;
 136         struct l_wait_info             lwi    = { 0 };
 137         int                            rc;
 138         ENTRY;
 139
 140         rc = lu_env_init(&env, LCT_DT_THREAD);
 141         spin_lock(&ofd->ofd_inconsistency_lock);
 142         thread_set_flags(thread, rc != 0 ? SVC_STOPPED : SVC_RUNNING);
 143         wake_up_all(&thread->t_ctl_waitq);
 144         spin_unlock(&ofd->ofd_inconsistency_lock);
 145         if (rc != 0)
 146                 RETURN(rc);
 147
 148         OBD_ALLOC_PTR(lr);
 149         if (unlikely(lr == NULL))
 150                 GOTO(out, rc = -ENOMEM);
 151
 152         lr->lr_event = LE_PAIRS_VERIFY;
 153         lr->lr_active = LFSCK_TYPE_LAYOUT;
 154
 155         spin_lock(&ofd->ofd_inconsistency_lock);
 156         while (1) {
 157                 if (unlikely(!thread_is_running(thread)))
 158                         break;
 159
 160                 while (!list_empty(&ofd->ofd_inconsistency_list)) {
 161                         oii = list_entry(ofd->ofd_inconsistency_list.next,
 162                                          struct ofd_inconsistency_item,
 163                                          oii_list);
 164                         list_del_init(&oii->oii_list);
 165                         spin_unlock(&ofd->ofd_inconsistency_lock);
 166                         ofd_inconsistency_verify_one(&env, ofd, oii, lr);
 167                         spin_lock(&ofd->ofd_inconsistency_lock);
 168                 }
 169
 170                 spin_unlock(&ofd->ofd_inconsistency_lock);
 171                 l_wait_event(thread->t_ctl_waitq,
 172                              !list_empty(&ofd->ofd_inconsistency_list) ||
 173                              !thread_is_running(thread),
 174                              &lwi);
 175                 spin_lock(&ofd->ofd_inconsistency_lock);
 176         }
 177
 178         while (!list_empty(&ofd->ofd_inconsistency_list)) {
 179                 struct ofd_object *fo;
 180
 181                 oii = list_entry(ofd->ofd_inconsistency_list.next,
 182                                  struct ofd_inconsistency_item,
 183                                  oii_list);
 184                 list_del_init(&oii->oii_list);
 185                 fo = oii->oii_obj;
 186                 spin_unlock(&ofd->ofd_inconsistency_lock);
 187
 188                 ofd_write_lock(&env, fo);
 189                 fo->ofo_pfid_checking = 0;
 190                 ofd_write_unlock(&env, fo);
 191
 192                 lu_object_put(&env, &fo->ofo_obj.do_lu);
 193                 OBD_FREE_PTR(oii);
 194                 spin_lock(&ofd->ofd_inconsistency_lock);
 195         }
 196
 197         OBD_FREE_PTR(lr);
 198
 199         GOTO(out, rc = 0);
 200
 201 out:
 202         thread_set_flags(thread, SVC_STOPPED);
 203         wake_up_all(&thread->t_ctl_waitq);
 204         spin_unlock(&ofd->ofd_inconsistency_lock);
 205         lu_env_fini(&env);
 206
 207         return rc;
 208 }
 209
 210 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd)
 211 {
 212         struct ptlrpc_thread    *thread = &ofd->ofd_inconsistency_thread;
 213         struct l_wait_info       lwi    = { 0 };
 214         struct task_struct      *task;
 215         int                      rc;
 216
 217         spin_lock(&ofd->ofd_inconsistency_lock);
 218         if (unlikely(thread_is_running(thread))) {
 219                 spin_unlock(&ofd->ofd_inconsistency_lock);
 220
 221                 return -EALREADY;
 222         }
 223
 224         thread_set_flags(thread, 0);
 225         spin_unlock(&ofd->ofd_inconsistency_lock);
 226         task = kthread_run(ofd_inconsistency_verification_main, ofd,
 227                            "inconsistency_verification");
 228         if (IS_ERR(task)) {
 229                 rc = PTR_ERR(task);
 230                 CERROR("%s: cannot start self_repair thread: rc = %d\n",
 231                        ofd_name(ofd), rc);
 232         } else {
 233                 rc = 0;
 234                 l_wait_event(thread->t_ctl_waitq,
 235                              thread_is_running(thread) ||
 236                              thread_is_stopped(thread),
 237                              &lwi);
 238         }
 239
 240         return rc;
 241 }
 242
 243 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd)
 244 {
 245         struct ptlrpc_thread    *thread = &ofd->ofd_inconsistency_thread;
 246         struct l_wait_info       lwi    = { 0 };
 247
 248         spin_lock(&ofd->ofd_inconsistency_lock);
 249         if (thread_is_init(thread) || thread_is_stopped(thread)) {
 250                 spin_unlock(&ofd->ofd_inconsistency_lock);
 251
 252                 return -EALREADY;
 253         }
 254
 255         thread_set_flags(thread, SVC_STOPPING);
 256         spin_unlock(&ofd->ofd_inconsistency_lock);
 257         wake_up_all(&thread->t_ctl_waitq);
 258         l_wait_event(thread->t_ctl_waitq,
 259                      thread_is_stopped(thread),
 260                      &lwi);
 261
 262         return 0;
 263 }
 264
 265 static void ofd_add_inconsistency_item(const struct lu_env *env,
 266                                        struct ofd_object *fo, struct obdo *oa)
 267 {
 268         struct ofd_device               *ofd    = ofd_obj2dev(fo);
 269         struct ofd_inconsistency_item   *oii;
 270         bool                             wakeup = false;
 271
 272         OBD_ALLOC_PTR(oii);
 273         if (oii == NULL)
 274                 return;
 275
 276         INIT_LIST_HEAD(&oii->oii_list);
 277         lu_object_get(&fo->ofo_obj.do_lu);
 278         oii->oii_obj = fo;
 279         oii->oii_pfid.f_seq = oa->o_parent_seq;
 280         oii->oii_pfid.f_oid = oa->o_parent_oid;
 281         oii->oii_pfid.f_stripe_idx = oa->o_stripe_idx;
 282
 283         spin_lock(&ofd->ofd_inconsistency_lock);
 284         if (fo->ofo_pfid_checking || fo->ofo_pfid_verified) {
 285                 spin_unlock(&ofd->ofd_inconsistency_lock);
 286                 OBD_FREE_PTR(oii);
 287
 288                 return;
 289         }
 290
 291         fo->ofo_pfid_checking = 1;
 292         if (list_empty(&ofd->ofd_inconsistency_list))
 293                 wakeup = true;
 294         list_add_tail(&oii->oii_list, &ofd->ofd_inconsistency_list);
 295         spin_unlock(&ofd->ofd_inconsistency_lock);
 296         if (wakeup)
 297                 wake_up_all(&ofd->ofd_inconsistency_thread.t_ctl_waitq);
 298
 299         /* XXX: When the found inconsistency exceeds some threshold,
 300          *      we can trigger the LFSCK to scan part of the system
 301          *      or the whole system, which depends on how to define
 302          *      the threshold, a simple way maybe like that: define
 303          *      the absolute value of how many inconsisteny allowed
 304          *      to be repaired via self detect/repair mechanism, if
 305          *      exceeded, then trigger the LFSCK to scan the layout
 306          *      inconsistency within the whole system. */
 307 }
 308
 309 int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
 310                   struct obdo *oa)
 311 {
 312         struct lu_fid   *pfid   = &fo->ofo_pfid;
 313         int              rc     = 0;
 314         ENTRY;
 315
 316         if (fid_is_sane(pfid)) {
 317                 if (likely(oa->o_parent_seq == pfid->f_seq &&
 318                            oa->o_parent_oid == pfid->f_oid &&
 319                            oa->o_stripe_idx == pfid->f_stripe_idx))
 320                         RETURN(0);
 321
 322                 if (fo->ofo_pfid_verified)
 323                         RETURN(-EPERM);
 324         }
 325
 326         /* The OST-object may be inconsistent, and we need further verification.
 327          * To avoid block the RPC service thread, return -EINPROGRESS to client
 328          * and make it retry later. */
 329         if (fo->ofo_pfid_checking)
 330                 RETURN(-EINPROGRESS);
 331
 332         rc = ofd_object_ff_load(env, fo);
 333         if (rc == -ENODATA)
 334                 RETURN(0);
 335
 336         if (rc < 0)
 337                 RETURN(rc);
 338
 339         if (likely(oa->o_parent_seq == pfid->f_seq &&
 340                    oa->o_parent_oid == pfid->f_oid &&
 341                    oa->o_stripe_idx == pfid->f_stripe_idx))
 342                 RETURN(0);
 343
 344         /* Push it to the dedicated thread for further verification. */
 345         ofd_add_inconsistency_item(env, fo, oa);
 346
 347         RETURN(-EINPROGRESS);
 348 }
 349
 350 static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
 351                            struct ofd_device *ofd, const struct lu_fid *fid,
 352                            struct lu_attr *la, struct obdo *oa, int niocount,
 353                            struct niobuf_remote *rnb, int *nr_local,
 354                            struct niobuf_local *lnb, char *jobid)
 355 {
 356         struct ofd_object       *fo;
 357         int                      i, j, rc, tot_bytes = 0;
 358
 359         ENTRY;
 360         LASSERT(env != NULL);
 361
 362         fo = ofd_object_find(env, ofd, fid);
 363         if (IS_ERR(fo))
 364                 RETURN(PTR_ERR(fo));
 365         LASSERT(fo != NULL);
 366
 367         ofd_read_lock(env, fo);
 368         if (!ofd_object_exists(fo))
 369                 GOTO(unlock, rc = -ENOENT);
 370
 371         if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
 372                 rc = ofd_verify_ff(env, fo, oa);
 373                 if (rc != 0)
 374                         GOTO(unlock, rc);
 375         }
 376
 377         /* parse remote buffers to local buffers and prepare the latter */
 378         *nr_local = 0;
 379         for (i = 0, j = 0; i < niocount; i++) {
 380                 rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
 381                                  lnb + j, 0, ofd_object_capa(env, fo));
 382                 if (unlikely(rc < 0))
 383                         GOTO(buf_put, rc);
 384                 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
 385                 /* correct index for local buffers to continue with */
 386                 j += rc;
 387                 *nr_local += rc;
 388                 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
 389                 tot_bytes += rnb[i].rnb_len;
 390         }
 391
 392         LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
 393         rc = dt_attr_get(env, ofd_object_child(fo), la,
 394                          ofd_object_capa(env, fo));
 395         if (unlikely(rc))
 396                 GOTO(buf_put, rc);
 397
 398         rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
 399         if (unlikely(rc))
 400                 GOTO(buf_put, rc);
 401
 402         ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes);
 403         RETURN(0);
 404
 405 buf_put:
 406         dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
 407 unlock:
 408         ofd_read_unlock(env, fo);
 409         ofd_object_put(env, fo);
 410         return rc;
 411 }
 412
 413 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
 414                             struct ofd_device *ofd, const struct lu_fid *fid,
 415                             struct lu_attr *la, struct obdo *oa,
 416                             int objcount, struct obd_ioobj *obj,
 417                             struct niobuf_remote *rnb, int *nr_local,
 418                             struct niobuf_local *lnb, char *jobid)
 419 {
 420         struct ofd_object       *fo;
 421         int                      i, j, k, rc = 0, tot_bytes = 0;
 422
 423         ENTRY;
 424         LASSERT(env != NULL);
 425         LASSERT(objcount == 1);
 426
 427         if (unlikely(exp->exp_obd->obd_recovering)) {
 428                 obd_seq seq = fid_seq(fid);
 429                 obd_id  oid = fid_oid(fid);
 430                 struct ofd_seq *oseq;
 431
 432                 oseq = ofd_seq_load(env, ofd, seq);
 433                 if (IS_ERR(oseq)) {
 434                         CERROR("%s: Can't find FID Sequence "LPX64": rc = %d\n",
 435                                ofd_name(ofd), seq, (int)PTR_ERR(oseq));
 436                         GOTO(out, rc = -EINVAL);
 437                 }
 438
 439                 if (oid > ofd_seq_last_oid(oseq)) {
 440                         int sync = 0;
 441                         int diff;
 442
 443                         mutex_lock(&oseq->os_create_lock);
 444                         diff = oid - ofd_seq_last_oid(oseq);
 445
 446                         /* Do sync create if the seq is about to used up */
 447                         if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) {
 448                                 if (unlikely(oid >= IDIF_MAX_OID - 1))
 449                                         sync = 1;
 450                         } else if (fid_seq_is_norm(seq)) {
 451                                 if (unlikely(oid >=
 452                                              LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
 453                                         sync = 1;
 454                         } else {
 455                                 CERROR("%s : invalid o_seq "DOSTID"\n",
 456                                        ofd_name(ofd), POSTID(&oa->o_oi));
 457                                 mutex_unlock(&oseq->os_create_lock);
 458                                 ofd_seq_put(env, oseq);
 459                                 GOTO(out, rc = -EINVAL);
 460                         }
 461
 462                         while (diff > 0) {
 463                                 obd_id next_id = ofd_seq_last_oid(oseq) + 1;
 464                                 int count = ofd_precreate_batch(ofd, diff);
 465
 466                                 rc = ofd_precreate_objects(env, ofd, next_id,
 467                                                            oseq, count, sync);
 468                                 if (rc < 0) {
 469                                         mutex_unlock(&oseq->os_create_lock);
 470                                         ofd_seq_put(env, oseq);
 471                                         GOTO(out, rc);
 472                                 }
 473
 474                                 diff -= rc;
 475                         }
 476
 477                         mutex_unlock(&oseq->os_create_lock);
 478                 }
 479
 480                 ofd_seq_put(env, oseq);
 481         }
 482
 483         fo = ofd_object_find(env, ofd, fid);
 484         if (IS_ERR(fo))
 485                 GOTO(out, rc = PTR_ERR(fo));
 486         LASSERT(fo != NULL);
 487
 488         ofd_read_lock(env, fo);
 489         if (!ofd_object_exists(fo)) {
 490                 CERROR("%s: BRW to missing obj "DOSTID"\n",
 491                        exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
 492                 ofd_read_unlock(env, fo);
 493                 ofd_object_put(env, fo);
 494                 GOTO(out, rc = -ENOENT);
 495         }
 496
 497         if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
 498                 rc = ofd_verify_ff(env, fo, oa);
 499                 if (rc != 0) {
 500                         ofd_read_unlock(env, fo);
 501                         ofd_object_put(env, fo);
 502                         GOTO(out, rc);
 503                 }
 504         }
 505
 506         /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
 507          * space back if possible */
 508         ofd_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
 509
 510         /* parse remote buffers to local buffers and prepare the latter */
 511         *nr_local = 0;
 512         for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
 513                 rc = dt_bufs_get(env, ofd_object_child(fo),
 514                                  rnb + i, lnb + j, 1,
 515                                  ofd_object_capa(env, fo));
 516                 if (unlikely(rc < 0))
 517                         GOTO(err, rc);
 518                 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
 519                 /* correct index for local buffers to continue with */
 520                 for (k = 0; k < rc; k++) {
 521                         lnb[j+k].lnb_flags = rnb[i].rnb_flags;
 522                         if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
 523                                 lnb[j+k].lnb_rc = -ENOSPC;
 524
 525                         /* remote client can't break through quota */
 526                         if (exp_connect_rmtclient(exp))
 527                                 lnb[j+k].lnb_flags &= ~OBD_BRW_NOQUOTA;
 528                 }
 529                 j += rc;
 530                 *nr_local += rc;
 531                 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
 532                 tot_bytes += rnb[i].rnb_len;
 533         }
 534         LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
 535
 536         rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
 537         if (unlikely(rc != 0))
 538                 GOTO(err, rc);
 539
 540         ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes);
 541         RETURN(0);
 542 err:
 543         dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
 544         ofd_read_unlock(env, fo);
 545         /* ofd_grant_prepare_write() was called, so we must commit */
 546         ofd_grant_commit(env, exp, rc);
 547 out:
 548         /* let's still process incoming grant information packed in the oa,
 549          * but without enforcing grant since we won't proceed with the write.
 550          * Just like a read request actually. */
 551         ofd_grant_prepare_read(env, exp, oa);
 552         return rc;
 553 }
 554
 555 int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
 556                struct obdo *oa, int objcount, struct obd_ioobj *obj,
 557                struct niobuf_remote *rnb, int *nr_local,
 558                struct niobuf_local *lnb, struct obd_trans_info *oti,
 559                struct lustre_capa *capa)
 560 {
 561         struct tgt_session_info *tsi = tgt_ses_info(env);
 562         struct ofd_device       *ofd = ofd_exp(exp);
 563         struct ofd_thread_info  *info;
 564         char                    *jobid;
 565         const struct lu_fid     *fid = &oa->o_oi.oi_fid;
 566         int                      rc = 0;
 567
 568         if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
 569                 CERROR("%s: bulk has too many pages %d, which exceeds the"
 570                        "maximum pages per RPC of %d\n",
 571                        exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
 572                 RETURN(-EPROTO);
 573         }
 574
 575         if (tgt_ses_req(tsi) == NULL) { /* echo client case */
 576                 LASSERT(oti != NULL);
 577                 info = ofd_info_init(env, exp);
 578                 ofd_oti2info(info, oti);
 579                 jobid = NULL;
 580         } else {
 581                 info = tsi2ofd_info(tsi);
 582                 jobid = tsi->tsi_jobid;
 583         }
 584
 585         LASSERT(oa != NULL);
 586
 587         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT)) {
 588                 struct ofd_seq          *oseq;
 589
 590                 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
 591                 if (IS_ERR(oseq)) {
 592                         CERROR("%s: Can not find seq for "DOSTID
 593                                ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
 594                                PTR_ERR(oseq));
 595                         RETURN(-EINVAL);
 596                 }
 597
 598                 if (oseq->os_destroys_in_progress == 0) {
 599                         /* don't fail lookups for orphan recovery, it causes
 600                          * later LBUGs when objects still exist during
 601                          * precreate */
 602                         ofd_seq_put(env, oseq);
 603                         RETURN(-ENOENT);
 604                 }
 605                 ofd_seq_put(env, oseq);
 606         }
 607
 608         LASSERT(objcount == 1);
 609         LASSERT(obj->ioo_bufcnt > 0);
 610
 611         if (cmd == OBD_BRW_WRITE) {
 612                 rc = ofd_auth_capa(exp, fid, ostid_seq(&oa->o_oi),
 613                                    capa, CAPA_OPC_OSS_WRITE);
 614                 if (rc == 0) {
 615                         la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
 616                         rc = ofd_preprw_write(env, exp, ofd, fid,
 617                                               &info->fti_attr, oa, objcount,
 618                                               obj, rnb, nr_local, lnb, jobid);
 619                 }
 620         } else if (cmd == OBD_BRW_READ) {
 621                 rc = ofd_auth_capa(exp, fid, ostid_seq(&oa->o_oi),
 622                                    capa, CAPA_OPC_OSS_READ);
 623                 if (rc == 0) {
 624                         ofd_grant_prepare_read(env, exp, oa);
 625                         rc = ofd_preprw_read(env, exp, ofd, fid,
 626                                              &info->fti_attr, oa,
 627                                              obj->ioo_bufcnt, rnb, nr_local,
 628                                              lnb, jobid);
 629                         obdo_from_la(oa, &info->fti_attr, LA_ATIME);
 630                 }
 631         } else {
 632                 CERROR("%s: wrong cmd %d received!\n",
 633                        exp->exp_obd->obd_name, cmd);
 634                 rc = -EPROTO;
 635         }
 636         RETURN(rc);
 637 }
 638
 639 static int
 640 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
 641                   const struct lu_fid *fid, int objcount, int niocount,
 642                   struct niobuf_local *lnb)
 643 {
 644         struct ofd_object *fo;
 645
 646         ENTRY;
 647
 648         LASSERT(niocount > 0);
 649
 650         fo = ofd_object_find(env, ofd, fid);
 651         if (IS_ERR(fo))
 652                 RETURN(PTR_ERR(fo));
 653         LASSERT(fo != NULL);
 654         LASSERT(ofd_object_exists(fo));
 655         dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
 656
 657         ofd_read_unlock(env, fo);
 658         ofd_object_put(env, fo);
 659         /* second put is pair to object_get in ofd_preprw_read */
 660         ofd_object_put(env, fo);
 661
 662         RETURN(0);
 663 }
 664
 665 static int
 666 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
 667                    struct ofd_object *ofd_obj, struct lu_attr *la,
 668                    struct filter_fid *ff)
 669 {
 670         struct ofd_thread_info  *info = ofd_info(env);
 671         __u64                    valid = la->la_valid;
 672         int                      rc;
 673         struct thandle          *th;
 674         struct dt_object        *dt_obj;
 675         int                      ff_needed = 0;
 676
 677         ENTRY;
 678
 679         LASSERT(la);
 680
 681         dt_obj = ofd_object_child(ofd_obj);
 682         LASSERT(dt_obj != NULL);
 683
 684         la->la_valid &= LA_UID | LA_GID;
 685
 686         rc = ofd_attr_handle_ugid(env, ofd_obj, la, 0 /* !is_setattr */);
 687         if (rc != 0)
 688                 GOTO(out, rc);
 689
 690         if (ff != NULL) {
 691                 rc = ofd_object_ff_load(env, ofd_obj);
 692                 if (rc == -ENODATA)
 693                         ff_needed = 1;
 694                 else if (rc < 0)
 695                         GOTO(out, rc);
 696         }
 697
 698         if (!la->la_valid && !ff_needed)
 699                 /* no attributes to set */
 700                 GOTO(out, rc = 0);
 701
 702         th = ofd_trans_create(env, ofd);
 703         if (IS_ERR(th))
 704                 GOTO(out, rc = PTR_ERR(th));
 705
 706         if (la->la_valid) {
 707                 rc = dt_declare_attr_set(env, dt_obj, la, th);
 708                 if (rc)
 709                         GOTO(out_tx, rc);
 710         }
 711
 712         if (ff_needed) {
 713                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
 714                         ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
 715                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
 716                         ff->ff_parent.f_oid =
 717                         cpu_to_le32(le32_to_cpu(ff->ff_parent.f_oid) - 1);
 718
 719                 info->fti_buf.lb_buf = ff;
 720                 info->fti_buf.lb_len = sizeof(*ff);
 721                 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
 722                                           XATTR_NAME_FID, 0, th);
 723                 if (rc)
 724                         GOTO(out_tx, rc);
 725         }
 726
 727         /* We don't need a transno for this operation which will be re-executed
 728          * anyway when the OST_WRITE (with a transno assigned) is replayed */
 729         rc = dt_trans_start_local(env, ofd->ofd_osd , th);
 730         if (rc)
 731                 GOTO(out_tx, rc);
 732
 733         /* set uid/gid */
 734         if (la->la_valid) {
 735                 rc = dt_attr_set(env, dt_obj, la, th,
 736                                  ofd_object_capa(env, ofd_obj));
 737                 if (rc)
 738                         GOTO(out_tx, rc);
 739         }
 740
 741         /* set filter fid EA */
 742         if (ff_needed) {
 743                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
 744                         GOTO(out_tx, rc);
 745
 746                 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID,
 747                                   0, th, BYPASS_CAPA);
 748                 if (rc == 0) {
 749                         ofd_obj->ofo_pfid.f_seq = le64_to_cpu(ff->ff_parent.f_seq);
 750                         ofd_obj->ofo_pfid.f_oid = le32_to_cpu(ff->ff_parent.f_oid);
 751                         /* Currently, the filter_fid::ff_parent::f_ver is not
 752                          * the real parent MDT-object's FID::f_ver, instead it
 753                          * is the OST-object index in its parent MDT-object's
 754                          * layout EA. */
 755                         ofd_obj->ofo_pfid.f_stripe_idx =
 756                                         le32_to_cpu(ff->ff_parent.f_stripe_idx);
 757                 }
 758         }
 759
 760         GOTO(out_tx, rc);
 761
 762 out_tx:
 763         dt_trans_stop(env, ofd->ofd_osd, th);
 764 out:
 765         la->la_valid = valid;
 766         return rc;
 767 }
 768
 769 struct ofd_soft_sync_callback {
 770         struct dt_txn_commit_cb  ossc_cb;
 771         struct obd_export       *ossc_exp;
 772 };
 773
 774 static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th,
 775                              struct dt_txn_commit_cb *cb, int err)
 776 {
 777         struct ofd_soft_sync_callback   *ossc;
 778
 779         ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb);
 780
 781         CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp);
 782         atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0);
 783
 784         class_export_cb_put(ossc->ossc_exp);
 785         OBD_FREE_PTR(ossc);
 786 }
 787
 788 static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
 789 {
 790         struct ofd_soft_sync_callback           *ossc;
 791         struct dt_txn_commit_cb                 *dcb;
 792         int                                      rc;
 793
 794         OBD_ALLOC_PTR(ossc);
 795         if (ossc == NULL)
 796                 return -ENOMEM;
 797
 798         ossc->ossc_exp = class_export_cb_get(exp);
 799
 800         dcb = &ossc->ossc_cb;
 801         dcb->dcb_func = ofd_cb_soft_sync;
 802         CFS_INIT_LIST_HEAD(&dcb->dcb_linkage);
 803         strlcpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name));
 804
 805         rc = dt_trans_cb_add(th, dcb);
 806         if (rc) {
 807                 class_export_cb_put(exp);
 808                 OBD_FREE_PTR(ossc);
 809         }
 810
 811         return rc;
 812 }
 813
 814 static int
 815 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
 816                    struct ofd_device *ofd, const struct lu_fid *fid,
 817                    struct lu_attr *la, struct filter_fid *ff, int objcount,
 818                    int niocount, struct niobuf_local *lnb, int old_rc)
 819 {
 820         struct ofd_thread_info  *info = ofd_info(env);
 821         struct ofd_object       *fo;
 822         struct dt_object        *o;
 823         struct thandle          *th;
 824         int                      rc = 0;
 825         int                      retries = 0;
 826         int                      i;
 827         struct filter_export_data *fed = &exp->exp_filter_data;
 828         bool                     soft_sync = false;
 829         bool                     cb_registered = false;
 830
 831         ENTRY;
 832
 833         LASSERT(objcount == 1);
 834
 835         fo = ofd_object_find(env, ofd, fid);
 836         LASSERT(fo != NULL);
 837         LASSERT(ofd_object_exists(fo));
 838
 839         o = ofd_object_child(fo);
 840         LASSERT(o != NULL);
 841
 842         if (old_rc)
 843                 GOTO(out, rc = old_rc);
 844
 845         /*
 846          * The first write to each object must set some attributes.  It is
 847          * important to set the uid/gid before calling
 848          * dt_declare_write_commit() since quota enforcement is now handled in
 849          * declare phases.
 850          */
 851         rc = ofd_write_attr_set(env, ofd, fo, la, ff);
 852         if (rc)
 853                 GOTO(out, rc);
 854
 855         la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
 856
 857 retry:
 858         th = ofd_trans_create(env, ofd);
 859         if (IS_ERR(th))
 860                 GOTO(out, rc = PTR_ERR(th));
 861
 862         th->th_sync |= ofd->ofd_syncjournal;
 863         if (th->th_sync == 0) {
 864                 for (i = 0; i < niocount; i++) {
 865                         if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
 866                                 th->th_sync = 1;
 867                                 break;
 868                         }
 869                         if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC)
 870                                 soft_sync = true;
 871                 }
 872         }
 873
 874         if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
 875                 GOTO(out_stop, rc = -EINPROGRESS);
 876
 877         rc = dt_declare_write_commit(env, o, lnb, niocount, th);
 878         if (rc)
 879                 GOTO(out_stop, rc);
 880
 881         if (la->la_valid) {
 882                 /* update [mac]time if needed */
 883                 rc = dt_declare_attr_set(env, o, la, th);
 884                 if (rc)
 885                         GOTO(out_stop, rc);
 886         }
 887
 888         rc = ofd_trans_start(env, ofd, fo, th);
 889         if (rc)
 890                 GOTO(out_stop, rc);
 891
 892         rc = dt_write_commit(env, o, lnb, niocount, th);
 893         if (rc)
 894                 GOTO(out_stop, rc);
 895
 896         if (la->la_valid) {
 897                 rc = dt_attr_set(env, o, la, th, ofd_object_capa(env, fo));
 898                 if (rc)
 899                         GOTO(out_stop, rc);
 900         }
 901
 902         /* get attr to return */
 903         rc = dt_attr_get(env, o, la, ofd_object_capa(env, fo));
 904
 905 out_stop:
 906         /* Force commit to make the just-deleted blocks
 907          * reusable. LU-456 */
 908         if (rc == -ENOSPC)
 909                 th->th_sync = 1;
 910
 911         /* do this before trans stop in case commit has finished */
 912         if (!th->th_sync && soft_sync && !cb_registered) {
 913                 ofd_soft_sync_cb_add(th, exp);
 914                 cb_registered = true;
 915         }
 916
 917         ofd_trans_stop(env, ofd, th, rc);
 918         if (rc == -ENOSPC && retries++ < 3) {
 919                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
 920                        retries);
 921                 goto retry;
 922         }
 923
 924         if (!soft_sync)
 925                 /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
 926                 atomic_set(&fed->fed_soft_sync_count, 0);
 927         else if (atomic_inc_return(&fed->fed_soft_sync_count) ==
 928                  ofd->ofd_soft_sync_limit)
 929                 dt_commit_async(env, ofd->ofd_osd);
 930
 931 out:
 932         dt_bufs_put(env, o, lnb, niocount);
 933         ofd_read_unlock(env, fo);
 934         ofd_object_put(env, fo);
 935         /* second put is pair to object_get in ofd_preprw_write */
 936         ofd_object_put(env, fo);
 937         ofd_grant_commit(env, info->fti_exp, old_rc);
 938         RETURN(rc);
 939 }
 940
 941 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
 942                  struct obdo *oa, int objcount, struct obd_ioobj *obj,
 943                  struct niobuf_remote *rnb, int npages,
 944                  struct niobuf_local *lnb, struct obd_trans_info *oti,
 945                  int old_rc)
 946 {
 947         struct ofd_thread_info  *info = ofd_info(env);
 948         struct ofd_mod_data     *fmd;
 949         __u64                    valid;
 950         struct ofd_device       *ofd = ofd_exp(exp);
 951         struct filter_fid       *ff = NULL;
 952         const struct lu_fid     *fid = &oa->o_oi.oi_fid;
 953         int                      rc = 0;
 954
 955         LASSERT(npages > 0);
 956
 957         if (cmd == OBD_BRW_WRITE) {
 958                 /* Don't update timestamps if this write is older than a
 959                  * setattr which modifies the timestamps. b=10150 */
 960
 961                 /* XXX when we start having persistent reservations this needs
 962                  * to be changed to ofd_fmd_get() to create the fmd if it
 963                  * doesn't already exist so we can store the reservation handle
 964                  * there. */
 965                 valid = OBD_MD_FLUID | OBD_MD_FLGID;
 966                 fmd = ofd_fmd_find(exp, fid);
 967                 if (!fmd || fmd->fmd_mactime_xid < info->fti_xid)
 968                         valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
 969                                  OBD_MD_FLCTIME;
 970                 ofd_fmd_put(exp, fmd);
 971                 la_from_obdo(&info->fti_attr, oa, valid);
 972
 973                 if (oa->o_valid & OBD_MD_FLFID) {
 974                         ff = &info->fti_mds_fid;
 975                         ofd_prepare_fidea(ff, oa);
 976                 }
 977
 978                 rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
 979                                         ff, objcount, npages, lnb, old_rc);
 980                 if (rc == 0)
 981                         obdo_from_la(oa, &info->fti_attr,
 982                                      OFD_VALID_FLAGS | LA_GID | LA_UID);
 983                 else
 984                         obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID);
 985
 986                 /* don't report overquota flag if we failed before reaching
 987                  * commit */
 988                 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
 989                         /* return the overquota flags to client */
 990                         if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
 991                                 if (oa->o_valid & OBD_MD_FLFLAGS)
 992                                         oa->o_flags |= OBD_FL_NO_USRQUOTA;
 993                                 else
 994                                         oa->o_flags = OBD_FL_NO_USRQUOTA;
 995                         }
 996
 997                         if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
 998                                 if (oa->o_valid & OBD_MD_FLFLAGS)
 999                                         oa->o_flags |= OBD_FL_NO_GRPQUOTA;
1000                                 else
1001                                         oa->o_flags = OBD_FL_NO_GRPQUOTA;
1002                         }
1003
1004                         oa->o_valid |= OBD_MD_FLFLAGS;
1005                         oa->o_valid |= OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA;
1006                 }
1007         } else if (cmd == OBD_BRW_READ) {
1008                 struct ldlm_namespace *ns = ofd->ofd_namespace;
1009
1010                 /* If oa != NULL then ofd_preprw_read updated the inode
1011                  * atime and we should update the lvb so that other glimpses
1012                  * will also get the updated value. bug 5972 */
1013                 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
1014                          struct ldlm_resource *rs = NULL;
1015
1016                         ost_fid_build_resid(fid, &info->fti_resid);
1017                         rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
1018                                                LDLM_EXTENT, 0);
1019                         if (!IS_ERR(rs)) {
1020                                 ns->ns_lvbo->lvbo_update(rs, NULL, 1);
1021                                 ldlm_resource_putref(rs);
1022                         }
1023                 }
1024                 rc = ofd_commitrw_read(env, ofd, fid, objcount,
1025                                        npages, lnb);
1026                 if (old_rc)
1027                         rc = old_rc;
1028         } else {
1029                 LBUG();
1030                 rc = -EPROTO;
1031         }
1032
1033         if (oti != NULL)
1034                 ofd_info2oti(info, oti);
1035         RETURN(rc);
1036 }