lustre/osd-ldiskfs/osd_handler.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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, 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/osd/osd_handler.c
  37  *
  38  * Top-level entry points into osd module
  39  *
  40  * Author: Nikita Danilov <nikita@clusterfs.com>
  41  *         Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_OSD
  45
  46 #include <linux/module.h>
  47 #include <linux/user_namespace.h>
  48 #ifdef HAVE_UIDGID_HEADER
  49 # include <linux/uidgid.h>
  50 #endif
  51
  52 /* LUSTRE_VERSION_CODE */
  53 #include <lustre_ver.h>
  54 /* prerequisite for linux/xattr.h */
  55 #include <linux/types.h>
  56 /* prerequisite for linux/xattr.h */
  57 #include <linux/fs.h>
  58 /* XATTR_{REPLACE,CREATE} */
  59 #include <linux/xattr.h>
  60
  61 #include <ldiskfs/ldiskfs.h>
  62 #include <ldiskfs/xattr.h>
  63 #undef ENTRY
  64 /*
  65  * struct OBD_{ALLOC,FREE}*()
  66  * OBD_FAIL_CHECK
  67  */
  68 #include <obd_support.h>
  69 /* struct ptlrpc_thread */
  70 #include <lustre_net.h>
  71 #include <lustre_fid.h>
  72 /* process_config */
  73 #include <lustre_param.h>
  74
  75 #include "osd_internal.h"
  76 #include "osd_dynlocks.h"
  77
  78 /* llo_* api support */
  79 #include <md_object.h>
  80 #include <lustre_quota.h>
  81
  82 #include <lustre_linkea.h>
  83
  84 int ldiskfs_pdo = 1;
  85 CFS_MODULE_PARM(ldiskfs_pdo, "i", int, 0644,
  86                 "ldiskfs with parallel directory operations");
  87
  88 int ldiskfs_track_declares_assert;
  89 CFS_MODULE_PARM(ldiskfs_track_declares_assert, "i", int, 0644,
  90                 "LBUG during tracking of declares");
  91
  92 /* Slab to allocate dynlocks */
  93 struct kmem_cache *dynlock_cachep;
  94
  95 /* Slab to allocate osd_it_ea */
  96 struct kmem_cache *osd_itea_cachep;
  97
  98 static struct lu_kmem_descr ldiskfs_caches[] = {
  99         {
 100                 .ckd_cache = &dynlock_cachep,
 101                 .ckd_name  = "dynlock_cache",
 102                 .ckd_size  = sizeof(struct dynlock_handle)
 103         },
 104         {
 105                 .ckd_cache = &osd_itea_cachep,
 106                 .ckd_name  = "osd_itea_cache",
 107                 .ckd_size  = sizeof(struct osd_it_ea)
 108         },
 109         {
 110                 .ckd_cache = NULL
 111         }
 112 };
 113
 114 static const char dot[] = ".";
 115 static const char dotdot[] = "..";
 116 static const char remote_obj_dir[] = "REM_OBJ_DIR";
 117
 118 static const struct lu_object_operations      osd_lu_obj_ops;
 119 static const struct dt_object_operations      osd_obj_ops;
 120 static const struct dt_object_operations      osd_obj_ea_ops;
 121 static const struct dt_object_operations      osd_obj_otable_it_ops;
 122 static const struct dt_index_operations       osd_index_iam_ops;
 123 static const struct dt_index_operations       osd_index_ea_ops;
 124
 125 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
 126                           const struct lu_fid *fid);
 127 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
 128                                                 struct osd_device *osd);
 129
 130 int osd_trans_declare_op2rb[] = {
 131         [OSD_OT_ATTR_SET]       = OSD_OT_ATTR_SET,
 132         [OSD_OT_PUNCH]          = OSD_OT_MAX,
 133         [OSD_OT_XATTR_SET]      = OSD_OT_XATTR_SET,
 134         [OSD_OT_CREATE]         = OSD_OT_DESTROY,
 135         [OSD_OT_DESTROY]        = OSD_OT_CREATE,
 136         [OSD_OT_REF_ADD]        = OSD_OT_REF_DEL,
 137         [OSD_OT_REF_DEL]        = OSD_OT_REF_ADD,
 138         [OSD_OT_WRITE]          = OSD_OT_WRITE,
 139         [OSD_OT_INSERT]         = OSD_OT_DELETE,
 140         [OSD_OT_DELETE]         = OSD_OT_INSERT,
 141         [OSD_OT_QUOTA]          = OSD_OT_MAX,
 142 };
 143
 144 static int osd_has_index(const struct osd_object *obj)
 145 {
 146         return obj->oo_dt.do_index_ops != NULL;
 147 }
 148
 149 static int osd_object_invariant(const struct lu_object *l)
 150 {
 151         return osd_invariant(osd_obj(l));
 152 }
 153
 154 /*
 155  * Concurrency: doesn't matter
 156  */
 157
 158 /*
 159  * Concurrency: doesn't matter
 160  */
 161 static int osd_write_locked(const struct lu_env *env, struct osd_object *o)
 162 {
 163         struct osd_thread_info *oti = osd_oti_get(env);
 164         return oti->oti_w_locks > 0 && o->oo_owner == env;
 165 }
 166
 167 /*
 168  * Concurrency: doesn't access mutable data
 169  */
 170 static int osd_root_get(const struct lu_env *env,
 171                         struct dt_device *dev, struct lu_fid *f)
 172 {
 173         lu_local_obj_fid(f, OSD_FS_ROOT_OID);
 174         return 0;
 175 }
 176
 177 /*
 178  * OSD object methods.
 179  */
 180
 181 /*
 182  * Concurrency: no concurrent access is possible that early in object
 183  * life-cycle.
 184  */
 185 static struct lu_object *osd_object_alloc(const struct lu_env *env,
 186                                           const struct lu_object_header *hdr,
 187                                           struct lu_device *d)
 188 {
 189         struct osd_object *mo;
 190
 191         OBD_ALLOC_PTR(mo);
 192         if (mo != NULL) {
 193                 struct lu_object *l;
 194
 195                 l = &mo->oo_dt.do_lu;
 196                 dt_object_init(&mo->oo_dt, NULL, d);
 197                 mo->oo_dt.do_ops = &osd_obj_ea_ops;
 198                 l->lo_ops = &osd_lu_obj_ops;
 199                 init_rwsem(&mo->oo_sem);
 200                 init_rwsem(&mo->oo_ext_idx_sem);
 201                 spin_lock_init(&mo->oo_guard);
 202                 return l;
 203         } else {
 204                 return NULL;
 205         }
 206 }
 207
 208 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
 209                 struct dentry *dentry, struct lustre_mdt_attrs *lma)
 210 {
 211         int rc;
 212
 213         CLASSERT(LMA_OLD_SIZE >= sizeof(*lma));
 214         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
 215                              info->oti_mdt_attrs_old, LMA_OLD_SIZE);
 216         if (rc > 0) {
 217                 if ((void *)lma != (void *)info->oti_mdt_attrs_old)
 218                         memcpy(lma, info->oti_mdt_attrs_old, sizeof(*lma));
 219                 rc = 0;
 220                 lustre_lma_swab(lma);
 221                 /* Check LMA compatibility */
 222                 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
 223                         CWARN("%.16s: unsupported incompat LMA feature(s) %#x "
 224                               "for fid = "DFID", ino = %lu\n",
 225                               LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
 226                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
 227                               PFID(&lma->lma_self_fid), inode->i_ino);
 228                         rc = -EOPNOTSUPP;
 229                 }
 230         } else if (rc == 0) {
 231                 rc = -ENODATA;
 232         }
 233
 234         return rc;
 235 }
 236
 237 /*
 238  * retrieve object from backend ext fs.
 239  **/
 240 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
 241                        struct osd_inode_id *id)
 242 {
 243         struct inode *inode = NULL;
 244
 245         /* if we look for an inode withing a running
 246          * transaction, then we risk to deadlock */
 247         /* osd_dirent_check_repair() breaks this */
 248         /*LASSERT(current->journal_info == NULL);*/
 249
 250         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
 251         if (IS_ERR(inode)) {
 252                 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
 253                        id->oii_ino, PTR_ERR(inode));
 254         } else if (id->oii_gen != OSD_OII_NOGEN &&
 255                    inode->i_generation != id->oii_gen) {
 256                 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
 257                        "i_generation = %u\n",
 258                        id->oii_ino, id->oii_gen, inode->i_generation);
 259                 iput(inode);
 260                 inode = ERR_PTR(-ESTALE);
 261         } else if (inode->i_nlink == 0) {
 262                 /* due to parallel readdir and unlink,
 263                 * we can have dead inode here. */
 264                 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
 265                 iput(inode);
 266                 inode = ERR_PTR(-ESTALE);
 267         } else if (is_bad_inode(inode)) {
 268                 CWARN("%.16s: bad inode: ino = %u\n",
 269                 LDISKFS_SB(osd_sb(dev))->s_es->s_volume_name, id->oii_ino);
 270                 iput(inode);
 271                 inode = ERR_PTR(-ENOENT);
 272         } else {
 273                 if (id->oii_gen == OSD_OII_NOGEN)
 274                         osd_id_gen(id, inode->i_ino, inode->i_generation);
 275
 276                 /* Do not update file c/mtime in ldiskfs.
 277                  * NB: we don't have any lock to protect this because we don't
 278                  * have reference on osd_object now, but contention with
 279                  * another lookup + attr_set can't happen in the tiny window
 280                  * between if (...) and set S_NOCMTIME. */
 281                 if (!(inode->i_flags & S_NOCMTIME))
 282                         inode->i_flags |= S_NOCMTIME;
 283         }
 284         return inode;
 285 }
 286
 287 static struct inode *
 288 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
 289              struct osd_inode_id *id, struct lu_fid *fid)
 290 {
 291         struct lustre_mdt_attrs *lma   = &info->oti_mdt_attrs;
 292         struct inode            *inode;
 293         int                      rc;
 294
 295         inode = osd_iget(info, dev, id);
 296         if (IS_ERR(inode))
 297                 return inode;
 298
 299         rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
 300         if (rc == 0) {
 301                 *fid = lma->lma_self_fid;
 302         } else if (rc == -ENODATA) {
 303                 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
 304                         lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
 305                 else
 306                         lu_igif_build(fid, inode->i_ino, inode->i_generation);
 307         } else {
 308                 iput(inode);
 309                 inode = ERR_PTR(rc);
 310         }
 311         return inode;
 312 }
 313
 314 static struct inode *osd_iget_check(struct osd_thread_info *info,
 315                                     struct osd_device *dev,
 316                                     const struct lu_fid *fid,
 317                                     struct osd_inode_id *id,
 318                                     bool cached)
 319 {
 320         struct inode    *inode;
 321         int              rc     = 0;
 322         ENTRY;
 323
 324         /* The cached OI mapping is trustable. If we cannot locate the inode
 325          * via the cached OI mapping, then return the failure to the caller
 326          * directly without further OI checking. */
 327
 328         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
 329         if (IS_ERR(inode)) {
 330                 rc = PTR_ERR(inode);
 331                 if (cached || (rc != -ENOENT && rc != -ESTALE)) {
 332                         CDEBUG(D_INODE, "no inode: ino = %u, rc = %d\n",
 333                                id->oii_ino, rc);
 334
 335                         GOTO(put, rc);
 336                 }
 337
 338                 goto check_oi;
 339         }
 340
 341         if (is_bad_inode(inode)) {
 342                 rc = -ENOENT;
 343                 if (cached) {
 344                         CDEBUG(D_INODE, "bad inode: ino = %u\n", id->oii_ino);
 345
 346                         GOTO(put, rc);
 347                 }
 348
 349                 goto check_oi;
 350         }
 351
 352         if (id->oii_gen != OSD_OII_NOGEN &&
 353             inode->i_generation != id->oii_gen) {
 354                 rc = -ESTALE;
 355                 if (cached) {
 356                         CDEBUG(D_INODE, "unmatched inode: ino = %u, "
 357                                "oii_gen = %u, i_generation = %u\n",
 358                                id->oii_ino, id->oii_gen, inode->i_generation);
 359
 360                         GOTO(put, rc);
 361                 }
 362
 363                 goto check_oi;
 364         }
 365
 366         if (inode->i_nlink == 0) {
 367                 rc = -ENOENT;
 368                 if (cached) {
 369                         CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
 370
 371                         GOTO(put, rc);
 372                 }
 373
 374                 goto check_oi;
 375         }
 376
 377 check_oi:
 378         if (rc != 0) {
 379                 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
 380
 381                 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
 382                 /* XXX: There are four possible cases:
 383                  *      1. rc = 0.
 384                  *         Backup/restore caused the OI invalid.
 385                  *      2. rc = 0.
 386                  *         Someone unlinked the object but NOT removed
 387                  *         the OI mapping, such as mount target device
 388                  *         as ldiskfs, and modify something directly.
 389                  *      3. rc = -ENOENT.
 390                  *         Someone just removed the object between the
 391                  *         former oi_lookup and the iget. It is normal.
 392                  *      4. Other failure cases.
 393                  *
 394                  *      Generally, when the device is mounted, it will
 395                  *      auto check whether the system is restored from
 396                  *      file-level backup or not. We trust such detect
 397                  *      to distinguish the 1st case from the 2nd case. */
 398                 if (rc == 0) {
 399                         if (!IS_ERR(inode) && inode->i_generation != 0 &&
 400                             inode->i_generation == id->oii_gen)
 401                                 /* "id->oii_gen != OSD_OII_NOGEN" is for
 402                                  * "@cached == false" case. */
 403                                 rc = -ENOENT;
 404                         else
 405                                 rc = -EREMCHG;
 406                 } else {
 407                         /* If the OI mapping was in OI file before the
 408                          * osd_iget_check(), but now, it is disappear,
 409                          * then it must be removed by race. That is a
 410                          * normal race case. */
 411                 }
 412         } else {
 413                 if (id->oii_gen == OSD_OII_NOGEN)
 414                         osd_id_gen(id, inode->i_ino, inode->i_generation);
 415
 416                 /* Do not update file c/mtime in ldiskfs.
 417                  * NB: we don't have any lock to protect this because we don't
 418                  * have reference on osd_object now, but contention with
 419                  * another lookup + attr_set can't happen in the tiny window
 420                  * between if (...) and set S_NOCMTIME. */
 421                 if (!(inode->i_flags & S_NOCMTIME))
 422                         inode->i_flags |= S_NOCMTIME;
 423         }
 424
 425         GOTO(put, rc);
 426
 427 put:
 428         if (rc != 0) {
 429                 if (!IS_ERR(inode))
 430                         iput(inode);
 431
 432                 inode = ERR_PTR(rc);
 433         }
 434
 435         return inode;
 436 }
 437
 438 /**
 439  * \retval +v: new filter_fid, does not contain self-fid
 440  * \retval 0:  filter_fid_old, contains self-fid
 441  * \retval -v: other failure cases
 442  */
 443 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
 444                  struct dentry *dentry, struct lu_fid *fid)
 445 {
 446         struct filter_fid_old   *ff     = &info->oti_ff;
 447         struct ost_id           *ostid  = &info->oti_ostid;
 448         int                      rc;
 449
 450         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
 451         if (rc == sizeof(*ff)) {
 452                 rc = 0;
 453                 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
 454                 ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
 455                 /* XXX: should use real OST index in the future. LU-3569 */
 456                 ostid_to_fid(fid, ostid, 0);
 457         } else if (rc == sizeof(struct filter_fid)) {
 458                 rc = 1;
 459         } else if (rc >= 0) {
 460                 rc = -EINVAL;
 461         }
 462
 463         return rc;
 464 }
 465
 466 static int osd_lma_self_repair(struct osd_thread_info *info,
 467                                struct osd_device *osd, struct inode *inode,
 468                                const struct lu_fid *fid, __u32 compat)
 469 {
 470         handle_t *jh;
 471         int       rc;
 472
 473         LASSERT(current->journal_info == NULL);
 474
 475         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
 476                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
 477         if (IS_ERR(jh)) {
 478                 rc = PTR_ERR(jh);
 479                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
 480                       osd_name(osd), rc);
 481                 return rc;
 482         }
 483
 484         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
 485         if (rc != 0)
 486                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
 487                       osd_name(osd), rc);
 488         ldiskfs_journal_stop(jh);
 489         return rc;
 490 }
 491
 492 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
 493 {
 494         struct osd_thread_info  *info   = osd_oti_get(env);
 495         struct osd_device       *osd    = osd_obj2dev(obj);
 496         struct lustre_mdt_attrs *lma    = &info->oti_mdt_attrs;
 497         struct inode            *inode  = obj->oo_inode;
 498         struct dentry           *dentry = &info->oti_obj_dentry;
 499         struct lu_fid           *fid    = NULL;
 500         const struct lu_fid     *rfid   = lu_object_fid(&obj->oo_dt.do_lu);
 501         int                      rc;
 502         ENTRY;
 503
 504         CLASSERT(LMA_OLD_SIZE >= sizeof(*lma));
 505         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
 506                              info->oti_mdt_attrs_old, LMA_OLD_SIZE);
 507         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
 508                 fid = &lma->lma_self_fid;
 509                 rc = osd_get_idif(info, inode, dentry, fid);
 510                 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
 511                         /* For the given OST-object, if it has neither LMA nor
 512                          * FID in XATTR_NAME_FID, then the given FID (which is
 513                          * contained in the @obj, from client RPC for locating
 514                          * the OST-object) is trusted. We use it to generate
 515                          * the LMA. */
 516                         osd_lma_self_repair(info, osd, inode, rfid,
 517                                             LMAC_FID_ON_OST);
 518                         RETURN(0);
 519                 }
 520         }
 521
 522         if (rc < 0)
 523                 RETURN(rc);
 524
 525         if (rc > 0) {
 526                 rc = 0;
 527                 lustre_lma_swab(lma);
 528                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
 529                              CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
 530                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
 531                               "fid = "DFID", ino = %lu\n", osd_name(osd),
 532                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
 533                               PFID(rfid), inode->i_ino);
 534                         rc = -EOPNOTSUPP;
 535                 } else {
 536                         fid = &lma->lma_self_fid;
 537                 }
 538         }
 539
 540         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
 541                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
 542                         struct ost_id   *oi   = &info->oti_ostid;
 543                         struct lu_fid   *fid1 = &info->oti_fid3;
 544                         __u32            idx  = fid_idif_ost_idx(rfid);
 545
 546                         /* For old IDIF, the OST index is not part of the IDIF,
 547                          * Means that different OSTs may have the same IDIFs.
 548                          * Under such case, we need to make some compatible
 549                          * check to make sure to trigger OI scrub properly. */
 550                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
 551                                 /* Given @rfid is new, LMA is old. */
 552                                 fid_to_ostid(fid, oi);
 553                                 ostid_to_fid(fid1, oi, idx);
 554                                 if (lu_fid_eq(fid1, rfid)) {
 555                                         if (osd->od_index_in_idif)
 556                                                 osd_lma_self_repair(info, osd,
 557                                                         inode, rfid,
 558                                                         LMAC_FID_ON_OST);
 559                                         RETURN(0);
 560                                 }
 561                         }
 562                 }
 563
 564                 rc = -EREMCHG;
 565         }
 566
 567         RETURN(rc);
 568 }
 569
 570 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
 571                           const struct lu_fid *fid,
 572                           const struct lu_object_conf *conf)
 573 {
 574         struct osd_thread_info *info;
 575         struct lu_device       *ldev   = obj->oo_dt.do_lu.lo_dev;
 576         struct osd_device      *dev;
 577         struct osd_idmap_cache *oic;
 578         struct osd_inode_id    *id;
 579         struct inode           *inode;
 580         struct osd_scrub       *scrub;
 581         struct scrub_file      *sf;
 582         int                     result;
 583         int                     saved  = 0;
 584         bool                    cached  = true;
 585         bool                    triggered = false;
 586         ENTRY;
 587
 588         LINVRNT(osd_invariant(obj));
 589         LASSERT(obj->oo_inode == NULL);
 590         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
 591
 592         dev = osd_dev(ldev);
 593         scrub = &dev->od_scrub;
 594         sf = &scrub->os_file;
 595         info = osd_oti_get(env);
 596         LASSERT(info);
 597         oic = &info->oti_cache;
 598
 599         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT))
 600                 RETURN(-ENOENT);
 601
 602         /* For the object is created as locking anchor, or for the object to
 603          * be created on disk. No need to osd_oi_lookup() at here because FID
 604          * shouldn't never be re-used, if it's really a duplicate FID from
 605          * unexpected reason, we should be able to detect it later by calling
 606          * do_create->osd_oi_insert(). */
 607         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 608                 GOTO(out, result = 0);
 609
 610         /* Search order: 1. per-thread cache. */
 611         if (lu_fid_eq(fid, &oic->oic_fid) &&
 612             likely(oic->oic_dev == dev)) {
 613                 id = &oic->oic_lid;
 614                 goto iget;
 615         }
 616
 617         id = &info->oti_id;
 618         if (!list_empty(&scrub->os_inconsistent_items)) {
 619                 /* Search order: 2. OI scrub pending list. */
 620                 result = osd_oii_lookup(dev, fid, id);
 621                 if (result == 0)
 622                         goto iget;
 623         }
 624
 625         cached = false;
 626         /* Search order: 3. OI files. */
 627         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
 628         if (result == -ENOENT) {
 629                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
 630                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
 631                     !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
 632                                       sf->sf_oi_bitmap))
 633                         GOTO(out, result = 0);
 634
 635                 goto trigger;
 636         }
 637
 638         if (result != 0)
 639                 GOTO(out, result);
 640
 641 iget:
 642         inode = osd_iget_check(info, dev, fid, id, cached);
 643         if (IS_ERR(inode)) {
 644                 result = PTR_ERR(inode);
 645                 if (result == -ENOENT || result == -ESTALE)
 646                         GOTO(out, result = -ENOENT);
 647
 648                 if (result == -EREMCHG) {
 649
 650 trigger:
 651                         if (unlikely(triggered))
 652                                 GOTO(out, result = saved);
 653
 654                         triggered = true;
 655                         if (thread_is_running(&scrub->os_thread)) {
 656                                 result = -EINPROGRESS;
 657                         } else if (!dev->od_noscrub) {
 658                                 result = osd_scrub_start(dev, SS_AUTO_FULL |
 659                                         SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
 660                                 LCONSOLE_WARN("%.16s: trigger OI scrub by RPC "
 661                                               "for "DFID", rc = %d [1]\n",
 662                                               osd_name(dev), PFID(fid), result);
 663                                 if (result == 0 || result == -EALREADY)
 664                                         result = -EINPROGRESS;
 665                                 else
 666                                         result = -EREMCHG;
 667                         }
 668
 669                         if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD))
 670                                 GOTO(out, result);
 671
 672                         /* We still have chance to get the valid inode: for the
 673                          * object which is referenced by remote name entry, the
 674                          * object on the local MDT will be linked under the dir
 675                          * of "/REMOTE_PARENT_DIR" with its FID string as name.
 676                          *
 677                          * We do not know whether the object for the given FID
 678                          * is referenced by some remote name entry or not, and
 679                          * especially for DNE II, a multiple-linked object may
 680                          * have many name entries reside on many MDTs.
 681                          *
 682                          * To simplify the operation, OSD will not distinguish
 683                          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
 684                          * only happened for the RPC from other MDT during the
 685                          * OI scrub, or for the client side RPC with FID only,
 686                          * such as FID to path, or from old connected client. */
 687                         saved = result;
 688                         result = osd_lookup_in_remote_parent(info, dev,
 689                                                              fid, id);
 690                         if (result == 0) {
 691                                 cached = true;
 692                                 goto iget;
 693                         }
 694
 695                         result = saved;
 696                 }
 697
 698                 GOTO(out, result);
 699         }
 700
 701         obj->oo_inode = inode;
 702         LASSERT(obj->oo_inode->i_sb == osd_sb(dev));
 703
 704         result = osd_check_lma(env, obj);
 705         if (result != 0) {
 706                 if (result == -ENODATA) {
 707                         if (cached) {
 708                                 result = osd_oi_lookup(info, dev, fid, id,
 709                                                        OI_CHECK_FLD);
 710                                 if (result != 0) {
 711                                         /* result == -ENOENT means that the OI
 712                                          * mapping has been removed by race,
 713                                          * the target inode belongs to other
 714                                          * object.
 715                                          *
 716                                          * Others error also can be returned
 717                                          * directly. */
 718                                         iput(inode);
 719                                         obj->oo_inode = NULL;
 720                                         GOTO(out, result);
 721                                 } else {
 722                                         /* result == 0 means the cached OI
 723                                          * mapping is still in the OI file,
 724                                          * the target the inode is valid. */
 725                                 }
 726                         } else {
 727                                 /* The current OI mapping is from the OI file,
 728                                  * since the inode has been found via
 729                                  * osd_iget_check(), no need recheck OI. */
 730                         }
 731
 732                         goto found;
 733                 }
 734
 735                 iput(inode);
 736                 obj->oo_inode = NULL;
 737                 if (result != -EREMCHG)
 738                         GOTO(out, result);
 739
 740                 if (cached) {
 741                         result = osd_oi_lookup(info, dev, fid, id,
 742                                                OI_CHECK_FLD);
 743                         /* result == -ENOENT means the cached OI mapping
 744                          * has been removed from the OI file by race,
 745                          * above target inode belongs to other object.
 746                          *
 747                          * Others error also can be returned directly. */
 748                         if (result != 0)
 749                                 GOTO(out, result);
 750
 751                         /* result == 0, goto trigger */
 752                 } else {
 753                         /* The current OI mapping is from the OI file,
 754                          * since the inode has been found via
 755                          * osd_iget_check(), no need recheck OI. */
 756                 }
 757
 758                 goto trigger;
 759         }
 760
 761 found:
 762         obj->oo_compat_dot_created = 1;
 763         obj->oo_compat_dotdot_created = 1;
 764
 765         if (!S_ISDIR(inode->i_mode) || !ldiskfs_pdo) /* done */
 766                 GOTO(out, result = 0);
 767
 768         LASSERT(obj->oo_hl_head == NULL);
 769         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
 770         if (obj->oo_hl_head == NULL) {
 771                 obj->oo_inode = NULL;
 772                 iput(inode);
 773                 GOTO(out, result = -ENOMEM);
 774         }
 775         GOTO(out, result = 0);
 776
 777 out:
 778         if (result != 0 && cached)
 779                 fid_zero(&oic->oic_fid);
 780
 781         LINVRNT(osd_invariant(obj));
 782         return result;
 783 }
 784
 785 /*
 786  * Concurrency: shouldn't matter.
 787  */
 788 static void osd_object_init0(struct osd_object *obj)
 789 {
 790         LASSERT(obj->oo_inode != NULL);
 791         obj->oo_dt.do_body_ops = &osd_body_ops;
 792         obj->oo_dt.do_lu.lo_header->loh_attr |=
 793                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
 794 }
 795
 796 /*
 797  * Concurrency: no concurrent access is possible that early in object
 798  * life-cycle.
 799  */
 800 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
 801                            const struct lu_object_conf *conf)
 802 {
 803         struct osd_object *obj = osd_obj(l);
 804         int result;
 805
 806         LINVRNT(osd_invariant(obj));
 807
 808         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
 809                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
 810                 l->lo_header->loh_attr |= LOHA_EXISTS;
 811                 return 0;
 812         }
 813
 814         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
 815         obj->oo_dt.do_body_ops = &osd_body_ops_new;
 816         if (result == 0 && obj->oo_inode != NULL)
 817                 osd_object_init0(obj);
 818
 819         LINVRNT(osd_invariant(obj));
 820         return result;
 821 }
 822
 823 /*
 824  * Concurrency: no concurrent access is possible that late in object
 825  * life-cycle.
 826  */
 827 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
 828 {
 829         struct osd_object *obj = osd_obj(l);
 830
 831         LINVRNT(osd_invariant(obj));
 832
 833         dt_object_fini(&obj->oo_dt);
 834         if (obj->oo_hl_head != NULL)
 835                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
 836         OBD_FREE_PTR(obj);
 837 }
 838
 839 /*
 840  * Concurrency: no concurrent access is possible that late in object
 841  * life-cycle.
 842  */
 843 static void osd_index_fini(struct osd_object *o)
 844 {
 845         struct iam_container *bag;
 846
 847         if (o->oo_dir != NULL) {
 848                 bag = &o->oo_dir->od_container;
 849                 if (o->oo_inode != NULL) {
 850                         if (bag->ic_object == o->oo_inode)
 851                                 iam_container_fini(bag);
 852                 }
 853                 OBD_FREE_PTR(o->oo_dir);
 854                 o->oo_dir = NULL;
 855         }
 856 }
 857
 858 /*
 859  * Concurrency: no concurrent access is possible that late in object
 860  * life-cycle (for all existing callers, that is. New callers have to provide
 861  * their own locking.)
 862  */
 863 static int osd_inode_unlinked(const struct inode *inode)
 864 {
 865         return inode->i_nlink == 0;
 866 }
 867
 868 enum {
 869         OSD_TXN_OI_DELETE_CREDITS    = 20,
 870         OSD_TXN_INODE_DELETE_CREDITS = 20
 871 };
 872
 873 /*
 874  * Journal
 875  */
 876
 877 #if OSD_THANDLE_STATS
 878 /**
 879  * Set time when the handle is allocated
 880  */
 881 static void osd_th_alloced(struct osd_thandle *oth)
 882 {
 883         oth->oth_alloced = cfs_time_current();
 884 }
 885
 886 /**
 887  * Set time when the handle started
 888  */
 889 static void osd_th_started(struct osd_thandle *oth)
 890 {
 891         oth->oth_started = cfs_time_current();
 892 }
 893
 894 /**
 895  * Helper function to convert time interval to microseconds packed in
 896  * long int.
 897  */
 898 static long interval_to_usec(cfs_time_t start, cfs_time_t end)
 899 {
 900         struct timeval val;
 901
 902         cfs_duration_usec(cfs_time_sub(end, start), &val);
 903         return val.tv_sec * 1000000 + val.tv_usec;
 904 }
 905
 906 /**
 907  * Check whether the we deal with this handle for too long.
 908  */
 909 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
 910                                 cfs_time_t alloced, cfs_time_t started,
 911                                 cfs_time_t closed)
 912 {
 913         cfs_time_t now = cfs_time_current();
 914
 915         LASSERT(dev != NULL);
 916
 917         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
 918                             interval_to_usec(alloced, started));
 919         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
 920                             interval_to_usec(started, closed));
 921         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
 922                             interval_to_usec(closed, now));
 923
 924         if (cfs_time_before(cfs_time_add(alloced, cfs_time_seconds(30)), now)) {
 925                 CWARN("transaction handle %p was open for too long: "
 926                       "now "CFS_TIME_T" ,"
 927                       "alloced "CFS_TIME_T" ,"
 928                       "started "CFS_TIME_T" ,"
 929                       "closed "CFS_TIME_T"\n",
 930                       oth, now, alloced, started, closed);
 931                 libcfs_debug_dumpstack(NULL);
 932         }
 933 }
 934
 935 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
 936 {                                                                       \
 937         cfs_time_t __closed = cfs_time_current();                       \
 938         cfs_time_t __alloced = oth->oth_alloced;                        \
 939         cfs_time_t __started = oth->oth_started;                        \
 940                                                                         \
 941         expr;                                                           \
 942         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
 943 }
 944
 945 #else /* OSD_THANDLE_STATS */
 946
 947 #define osd_th_alloced(h)                  do {} while(0)
 948 #define osd_th_started(h)                  do {} while(0)
 949 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
 950
 951 #endif /* OSD_THANDLE_STATS */
 952
 953 /*
 954  * Concurrency: doesn't access mutable data.
 955  */
 956 static int osd_param_is_not_sane(const struct osd_device *dev,
 957                                  const struct thandle *th)
 958 {
 959         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
 960
 961         return oh->ot_credits > osd_transaction_size(dev);
 962 }
 963
 964 /*
 965  * Concurrency: shouldn't matter.
 966  */
 967 static void osd_trans_commit_cb(struct super_block *sb,
 968                                 struct ldiskfs_journal_cb_entry *jcb, int error)
 969 {
 970         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
 971         struct thandle     *th  = &oh->ot_super;
 972         struct lu_device   *lud = &th->th_dev->dd_lu_dev;
 973         struct dt_txn_commit_cb *dcb, *tmp;
 974
 975         LASSERT(oh->ot_handle == NULL);
 976
 977         if (error)
 978                 CERROR("transaction @0x%p commit error: %d\n", th, error);
 979
 980         dt_txn_hook_commit(th);
 981
 982         /* call per-transaction callbacks if any */
 983         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
 984                                  dcb_linkage) {
 985                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
 986                          "commit callback entry: magic=%x name='%s'\n",
 987                          dcb->dcb_magic, dcb->dcb_name);
 988                 list_del_init(&dcb->dcb_linkage);
 989                 dcb->dcb_func(NULL, th, dcb, error);
 990         }
 991
 992         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
 993         lu_device_put(lud);
 994         th->th_dev = NULL;
 995
 996         lu_context_exit(&th->th_ctx);
 997         lu_context_fini(&th->th_ctx);
 998         OBD_FREE_PTR(oh);
 999 }
1000
1001 static struct thandle *osd_trans_create(const struct lu_env *env,
1002                                         struct dt_device *d)
1003 {
1004         struct osd_thread_info  *oti = osd_oti_get(env);
1005         struct osd_iobuf        *iobuf = &oti->oti_iobuf;
1006         struct osd_thandle      *oh;
1007         struct thandle          *th;
1008         ENTRY;
1009
1010         /* on pending IO in this thread should left from prev. request */
1011         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1012
1013         th = ERR_PTR(-ENOMEM);
1014         OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1015         if (oh != NULL) {
1016                 oh->ot_quota_trans = &oti->oti_quota_trans;
1017                 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1018                 th = &oh->ot_super;
1019                 th->th_dev = d;
1020                 th->th_result = 0;
1021                 th->th_tags = LCT_TX_HANDLE;
1022                 oh->ot_credits = 0;
1023                 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1024                 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1025                 osd_th_alloced(oh);
1026
1027                 memset(oti->oti_declare_ops, 0,
1028                        sizeof(oti->oti_declare_ops));
1029                 memset(oti->oti_declare_ops_cred, 0,
1030                        sizeof(oti->oti_declare_ops_cred));
1031                 memset(oti->oti_declare_ops_used, 0,
1032                        sizeof(oti->oti_declare_ops_used));
1033         }
1034         RETURN(th);
1035 }
1036
1037 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1038 {
1039         struct osd_thread_info  *oti = osd_oti_get(env);
1040         struct osd_thandle      *oh;
1041
1042         oh = container_of0(th, struct osd_thandle, ot_super);
1043         LASSERT(oh != NULL);
1044
1045         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1046               oti->oti_declare_ops[OSD_OT_CREATE],
1047               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1048               oti->oti_declare_ops_used[OSD_OT_CREATE],
1049               oti->oti_declare_ops[OSD_OT_DESTROY],
1050               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1051               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1052         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1053               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1054               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1055               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1056               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1057               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1058               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1059         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1060               oti->oti_declare_ops[OSD_OT_WRITE],
1061               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1062               oti->oti_declare_ops_used[OSD_OT_WRITE],
1063               oti->oti_declare_ops[OSD_OT_PUNCH],
1064               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1065               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1066               oti->oti_declare_ops[OSD_OT_QUOTA],
1067               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1068               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1069         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1070               oti->oti_declare_ops[OSD_OT_INSERT],
1071               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1072               oti->oti_declare_ops_used[OSD_OT_INSERT],
1073               oti->oti_declare_ops[OSD_OT_DELETE],
1074               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1075               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1076         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1077               oti->oti_declare_ops[OSD_OT_REF_ADD],
1078               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1079               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1080               oti->oti_declare_ops[OSD_OT_REF_DEL],
1081               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1082               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1083 }
1084
1085 /*
1086  * Concurrency: shouldn't matter.
1087  */
1088 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1089                            struct thandle *th)
1090 {
1091         struct osd_thread_info *oti = osd_oti_get(env);
1092         struct osd_device  *dev = osd_dt_dev(d);
1093         handle_t           *jh;
1094         struct osd_thandle *oh;
1095         int rc;
1096
1097         ENTRY;
1098
1099         LASSERT(current->journal_info == NULL);
1100
1101         oh = container_of0(th, struct osd_thandle, ot_super);
1102         LASSERT(oh != NULL);
1103         LASSERT(oh->ot_handle == NULL);
1104
1105         rc = dt_txn_hook_start(env, d, th);
1106         if (rc != 0)
1107                 GOTO(out, rc);
1108
1109         if (unlikely(osd_param_is_not_sane(dev, th))) {
1110                 static unsigned long last_printed;
1111                 static int last_credits;
1112
1113                 CWARN("%.16s: too many transaction credits (%d > %d)\n",
1114                       LDISKFS_SB(osd_sb(dev))->s_es->s_volume_name,
1115                       oh->ot_credits,
1116                       osd_journal(dev)->j_max_transaction_buffers);
1117
1118                 osd_trans_dump_creds(env, th);
1119
1120                 if (last_credits != oh->ot_credits &&
1121                     time_after(jiffies, last_printed +
1122                                msecs_to_jiffies(60 * MSEC_PER_SEC))) {
1123                         libcfs_debug_dumpstack(NULL);
1124                         last_credits = oh->ot_credits;
1125                         last_printed = jiffies;
1126                 }
1127                 /* XXX Limit the credits to 'max_transaction_buffers', and
1128                  *     let the underlying filesystem to catch the error if
1129                  *     we really need so many credits.
1130                  *
1131                  *     This should be removed when we can calculate the
1132                  *     credits precisely. */
1133                 oh->ot_credits = osd_transaction_size(dev);
1134         }
1135
1136         /*
1137          * XXX temporary stuff. Some abstraction layer should
1138          * be used.
1139          */
1140         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1141         osd_th_started(oh);
1142         if (!IS_ERR(jh)) {
1143                 oh->ot_handle = jh;
1144                 LASSERT(oti->oti_txns == 0);
1145                 lu_context_init(&th->th_ctx, th->th_tags);
1146                 lu_context_enter(&th->th_ctx);
1147
1148                 lu_device_get(&d->dd_lu_dev);
1149                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1150                               "osd-tx", th);
1151                 oti->oti_txns++;
1152                 rc = 0;
1153         } else {
1154                 rc = PTR_ERR(jh);
1155         }
1156 out:
1157         RETURN(rc);
1158 }
1159
1160 static int osd_seq_exists(const struct lu_env *env,
1161                           struct osd_device *osd, u64 seq)
1162 {
1163         struct lu_seq_range     *range = &osd_oti_get(env)->oti_seq_range;
1164         struct seq_server_site  *ss = osd_seq_site(osd);
1165         int                     rc;
1166         ENTRY;
1167
1168         LASSERT(ss != NULL);
1169         LASSERT(ss->ss_server_fld != NULL);
1170
1171         rc = osd_fld_lookup(env, osd, seq, range);
1172         if (rc != 0) {
1173                 if (rc != -ENOENT)
1174                         CERROR("%s: can't lookup FLD sequence "LPX64
1175                                ": rc = %d\n", osd_name(osd), seq, rc);
1176                 RETURN(0);
1177         }
1178
1179         RETURN(ss->ss_node_id == range->lsr_index);
1180 }
1181
1182 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1183 {
1184         struct dt_txn_commit_cb *dcb;
1185         struct dt_txn_commit_cb *tmp;
1186
1187         /* call per-transaction stop callbacks if any */
1188         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1189                                  dcb_linkage) {
1190                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1191                          "commit callback entry: magic=%x name='%s'\n",
1192                          dcb->dcb_magic, dcb->dcb_name);
1193                 list_del_init(&dcb->dcb_linkage);
1194                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1195         }
1196 }
1197
1198 /*
1199  * Concurrency: shouldn't matter.
1200  */
1201 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1202                           struct thandle *th)
1203 {
1204         int                     rc = 0, remove_agents = 0;
1205         struct osd_thandle     *oh;
1206         struct osd_thread_info *oti = osd_oti_get(env);
1207         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1208         struct osd_device      *osd = osd_dt_dev(th->th_dev);
1209         struct qsd_instance    *qsd = osd->od_quota_slave;
1210         struct lquota_trans    *qtrans;
1211         ENTRY;
1212
1213         oh = container_of0(th, struct osd_thandle, ot_super);
1214
1215         remove_agents = oh->ot_remove_agents;
1216
1217         qtrans = oh->ot_quota_trans;
1218         oh->ot_quota_trans = NULL;
1219
1220         if (oh->ot_handle != NULL) {
1221                 handle_t *hdl = oh->ot_handle;
1222
1223                 /*
1224                  * add commit callback
1225                  * notice we don't do this in osd_trans_start()
1226                  * as underlying transaction can change during truncate
1227                  */
1228                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1229                                          &oh->ot_jcb);
1230
1231                 LASSERT(oti->oti_txns == 1);
1232                 oti->oti_txns--;
1233
1234                 rc = dt_txn_hook_stop(env, th);
1235                 if (rc != 0)
1236                         CERROR("%s: failed in transaction hook: rc = %d\n",
1237                                osd_name(osd), rc);
1238
1239                 osd_trans_stop_cb(oh, rc);
1240                 /* hook functions might modify th_sync */
1241                 hdl->h_sync = th->th_sync;
1242
1243                 oh->ot_handle = NULL;
1244                 OSD_CHECK_SLOW_TH(oh, osd, rc = ldiskfs_journal_stop(hdl));
1245                 if (rc != 0)
1246                         CERROR("%s: failed to stop transaction: rc = %d\n",
1247                                osd_name(osd), rc);
1248         } else {
1249                 osd_trans_stop_cb(oh, th->th_result);
1250                 OBD_FREE_PTR(oh);
1251         }
1252
1253         /* inform the quota slave device that the transaction is stopping */
1254         qsd_op_end(env, qsd, qtrans);
1255
1256         /* as we want IO to journal and data IO be concurrent, we don't block
1257          * awaiting data IO completion in osd_do_bio(), instead we wait here
1258          * once transaction is submitted to the journal. all reqular requests
1259          * don't do direct IO (except read/write), thus this wait_event becomes
1260          * no-op for them.
1261          *
1262          * IMPORTANT: we have to wait till any IO submited by the thread is
1263          * completed otherwise iobuf may be corrupted by different request
1264          */
1265         wait_event(iobuf->dr_wait,
1266                        atomic_read(&iobuf->dr_numreqs) == 0);
1267         osd_fini_iobuf(osd, iobuf);
1268         if (!rc)
1269                 rc = iobuf->dr_error;
1270
1271         if (unlikely(remove_agents != 0))
1272                 osd_process_scheduled_agent_removals(env, osd);
1273
1274         RETURN(rc);
1275 }
1276
1277 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1278 {
1279         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1280                                                ot_super);
1281
1282         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1283         LASSERT(&dcb->dcb_func != NULL);
1284         if (dcb->dcb_flags & DCB_TRANS_STOP)
1285                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1286         else
1287                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1288
1289         return 0;
1290 }
1291
1292 /*
1293  * Called just before object is freed. Releases all resources except for
1294  * object itself (that is released by osd_object_free()).
1295  *
1296  * Concurrency: no concurrent access is possible that late in object
1297  * life-cycle.
1298  */
1299 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1300 {
1301         struct osd_object *obj   = osd_obj(l);
1302         struct inode      *inode = obj->oo_inode;
1303
1304         LINVRNT(osd_invariant(obj));
1305
1306         /*
1307          * If object is unlinked remove fid->ino mapping from object index.
1308          */
1309
1310         osd_index_fini(obj);
1311         if (inode != NULL) {
1312                 struct qsd_instance     *qsd = osd_obj2dev(obj)->od_quota_slave;
1313                 qid_t                    uid = i_uid_read(inode);
1314                 qid_t                    gid = i_gid_read(inode);
1315
1316                 iput(inode);
1317                 obj->oo_inode = NULL;
1318
1319                 if (qsd != NULL) {
1320                         struct osd_thread_info  *info = osd_oti_get(env);
1321                         struct lquota_id_info   *qi = &info->oti_qi;
1322
1323                         /* Release granted quota to master if necessary */
1324                         qi->lqi_id.qid_uid = uid;
1325                         qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1326
1327                         qi->lqi_id.qid_uid = gid;
1328                         qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1329                 }
1330         }
1331 }
1332
1333 /*
1334  * Concurrency: ->loo_object_release() is called under site spin-lock.
1335  */
1336 static void osd_object_release(const struct lu_env *env,
1337                                struct lu_object *l)
1338 {
1339 }
1340
1341 /*
1342  * Concurrency: shouldn't matter.
1343  */
1344 static int osd_object_print(const struct lu_env *env, void *cookie,
1345                             lu_printer_t p, const struct lu_object *l)
1346 {
1347         struct osd_object *o = osd_obj(l);
1348         struct iam_descr  *d;
1349
1350         if (o->oo_dir != NULL)
1351                 d = o->oo_dir->od_container.ic_descr;
1352         else
1353                 d = NULL;
1354         return (*p)(env, cookie,
1355                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
1356                     o, o->oo_inode,
1357                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
1358                     o->oo_inode ? o->oo_inode->i_generation : 0,
1359                     d ? d->id_ops->id_name : "plain");
1360 }
1361
1362 #define GRANT_FOR_LOCAL_OIDS 32 /* 128kB for last_rcvd, quota files, ... */
1363
1364 /*
1365  * Concurrency: shouldn't matter.
1366  */
1367 int osd_statfs(const struct lu_env *env, struct dt_device *d,
1368                struct obd_statfs *sfs)
1369 {
1370         struct osd_device  *osd = osd_dt_dev(d);
1371         struct super_block *sb = osd_sb(osd);
1372         struct kstatfs     *ksfs;
1373         int result = 0;
1374
1375         if (unlikely(osd->od_mnt == NULL))
1376                 return -EINPROGRESS;
1377
1378         /* osd_lproc.c call this without env, allocate ksfs for that case */
1379         if (unlikely(env == NULL)) {
1380                 OBD_ALLOC_PTR(ksfs);
1381                 if (ksfs == NULL)
1382                         return -ENOMEM;
1383         } else {
1384                 ksfs = &osd_oti_get(env)->oti_ksfs;
1385         }
1386
1387         spin_lock(&osd->od_osfs_lock);
1388         result = sb->s_op->statfs(sb->s_root, ksfs);
1389         if (likely(result == 0)) { /* N.B. statfs can't really fail */
1390                 statfs_pack(sfs, ksfs);
1391                 if (unlikely(sb->s_flags & MS_RDONLY))
1392                         sfs->os_state = OS_STATE_READONLY;
1393                 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1394                                               LDISKFS_FEATURE_INCOMPAT_EXTENTS))
1395                         sfs->os_maxbytes = sb->s_maxbytes;
1396                 else
1397                         sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
1398         }
1399         spin_unlock(&osd->od_osfs_lock);
1400
1401         if (unlikely(env == NULL))
1402                 OBD_FREE_PTR(ksfs);
1403
1404         /* Reserve a small amount of space for local objects like last_rcvd,
1405          * llog, quota files, ... */
1406         if (sfs->os_bavail <= GRANT_FOR_LOCAL_OIDS) {
1407                 sfs->os_bavail = 0;
1408         } else {
1409                 sfs->os_bavail -= GRANT_FOR_LOCAL_OIDS;
1410                 /** Take out metadata overhead for indirect blocks */
1411                 sfs->os_bavail -= sfs->os_bavail >> (sb->s_blocksize_bits - 3);
1412         }
1413
1414         return result;
1415 }
1416
1417 /**
1418  * Estimate space needed for file creations. We assume the largest filename
1419  * which is 2^64 - 1, hence a filename of 20 chars.
1420  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
1421  */
1422 #ifdef __LDISKFS_DIR_REC_LEN
1423 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
1424 #else
1425 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
1426 #endif
1427
1428 /*
1429  * Concurrency: doesn't access mutable data.
1430  */
1431 static void osd_conf_get(const struct lu_env *env,
1432                          const struct dt_device *dev,
1433                          struct dt_device_param *param)
1434 {
1435         struct super_block *sb = osd_sb(osd_dt_dev(dev));
1436         int                ea_overhead;
1437
1438         /*
1439          * XXX should be taken from not-yet-existing fs abstraction layer.
1440          */
1441         param->ddp_max_name_len = LDISKFS_NAME_LEN;
1442         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
1443         param->ddp_block_shift  = sb->s_blocksize_bits;
1444         param->ddp_mount_type     = LDD_MT_LDISKFS;
1445         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
1446                 param->ddp_maxbytes = sb->s_maxbytes;
1447         else
1448                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
1449         /* Overhead estimate should be fairly accurate, so we really take a tiny
1450          * error margin which also avoids fragmenting the filesystem too much */
1451         param->ddp_grant_reserved = 2; /* end up to be 1.9% after conversion */
1452         /* inode are statically allocated, so per-inode space consumption
1453          * is the space consumed by the directory entry */
1454         param->ddp_inodespace     = PER_OBJ_USAGE;
1455         /* per-fragment overhead to be used by the client code */
1456         param->ddp_grant_frag     = 6 * LDISKFS_BLOCK_SIZE(sb);
1457         param->ddp_mntopts      = 0;
1458         if (test_opt(sb, XATTR_USER))
1459                 param->ddp_mntopts |= MNTOPT_USERXATTR;
1460         if (test_opt(sb, POSIX_ACL))
1461                 param->ddp_mntopts |= MNTOPT_ACL;
1462
1463         /* LOD might calculate the max stripe count based on max_ea_size,
1464          * so we need take account in the overhead as well,
1465          * xattr_header + magic + xattr_entry_head */
1466         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
1467                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
1468
1469 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
1470         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
1471                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
1472                                                                 ea_overhead;
1473         else
1474 #endif
1475                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
1476 }
1477
1478 /*
1479  * Concurrency: shouldn't matter.
1480  */
1481 static int osd_sync(const struct lu_env *env, struct dt_device *d)
1482 {
1483         int rc;
1484
1485         CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
1486
1487         rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
1488
1489         CDEBUG(D_CACHE, "synced OSD %s: rc = %d\n",
1490                LUSTRE_OSD_LDISKFS_NAME, rc);
1491
1492         return rc;
1493 }
1494
1495 /**
1496  * Start commit for OSD device.
1497  *
1498  * An implementation of dt_commit_async method for OSD device.
1499  * Asychronously starts underlayng fs sync and thereby a transaction
1500  * commit.
1501  *
1502  * \param env environment
1503  * \param d dt device
1504  *
1505  * \see dt_device_operations
1506  */
1507 static int osd_commit_async(const struct lu_env *env,
1508                             struct dt_device *d)
1509 {
1510         struct super_block *s = osd_sb(osd_dt_dev(d));
1511         ENTRY;
1512
1513         CDEBUG(D_HA, "async commit OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
1514         RETURN(s->s_op->sync_fs(s, 0));
1515 }
1516
1517 /*
1518  * Concurrency: shouldn't matter.
1519  */
1520
1521 static int osd_ro(const struct lu_env *env, struct dt_device *d)
1522 {
1523         struct super_block *sb = osd_sb(osd_dt_dev(d));
1524         struct block_device *dev = sb->s_bdev;
1525 #ifdef HAVE_DEV_SET_RDONLY
1526         struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
1527         int rc = 0;
1528 #else
1529         int rc = -EOPNOTSUPP;
1530 #endif
1531         ENTRY;
1532
1533 #ifdef HAVE_DEV_SET_RDONLY
1534         CERROR("*** setting %s read-only ***\n", osd_dt_dev(d)->od_svname);
1535
1536         if (jdev && (jdev != dev)) {
1537                 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
1538                        (long)jdev);
1539                 dev_set_rdonly(jdev);
1540         }
1541         CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
1542         dev_set_rdonly(dev);
1543 #else
1544         CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
1545                osd_dt_dev(d)->od_svname, (long)dev, rc);
1546 #endif
1547         RETURN(rc);
1548 }
1549
1550 /**
1551  * Note: we do not count into QUOTA here.
1552  * If we mount with --data_journal we may need more.
1553  */
1554 const int osd_dto_credits_noquota[DTO_NR] = {
1555         /**
1556          * Insert.
1557          * INDEX_EXTRA_TRANS_BLOCKS(8) +
1558          * SINGLEDATA_TRANS_BLOCKS(8)
1559          * XXX Note: maybe iam need more, since iam have more level than
1560          *           EXT3 htree.
1561          */
1562         [DTO_INDEX_INSERT]  = 16,
1563         /**
1564          * Delete
1565          * just modify a single entry, probably merge few within a block
1566          */
1567         [DTO_INDEX_DELETE]  = 1,
1568         /**
1569          * Used for OI scrub
1570          */
1571         [DTO_INDEX_UPDATE]  = 16,
1572         /**
1573          * 4(inode, inode bits, groups, GDT)
1574          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
1575          */
1576         [DTO_OBJECT_CREATE] = 4,
1577         /**
1578          * 4(inode, inode bits, groups, GDT)
1579          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
1580          */
1581         [DTO_OBJECT_DELETE] = 4,
1582         /**
1583          * Attr set credits (inode)
1584          */
1585         [DTO_ATTR_SET_BASE] = 1,
1586         /**
1587          * Xattr set. The same as xattr of EXT3.
1588          * DATA_TRANS_BLOCKS(14)
1589          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
1590          * are also counted in. Do not know why?
1591          */
1592         [DTO_XATTR_SET]     = 14,
1593         /**
1594          * credits for inode change during write.
1595          */
1596         [DTO_WRITE_BASE]    = 3,
1597         /**
1598          * credits for single block write.
1599          */
1600         [DTO_WRITE_BLOCK]   = 14,
1601         /**
1602          * Attr set credits for chown.
1603          * This is extra credits for setattr, and it is null without quota
1604          */
1605         [DTO_ATTR_SET_CHOWN] = 0
1606 };
1607
1608 static const struct dt_device_operations osd_dt_ops = {
1609         .dt_root_get       = osd_root_get,
1610         .dt_statfs         = osd_statfs,
1611         .dt_trans_create   = osd_trans_create,
1612         .dt_trans_start    = osd_trans_start,
1613         .dt_trans_stop     = osd_trans_stop,
1614         .dt_trans_cb_add   = osd_trans_cb_add,
1615         .dt_conf_get       = osd_conf_get,
1616         .dt_sync           = osd_sync,
1617         .dt_ro             = osd_ro,
1618         .dt_commit_async   = osd_commit_async,
1619 };
1620
1621 static void osd_object_read_lock(const struct lu_env *env,
1622                                  struct dt_object *dt, unsigned role)
1623 {
1624         struct osd_object *obj = osd_dt_obj(dt);
1625         struct osd_thread_info *oti = osd_oti_get(env);
1626
1627         LINVRNT(osd_invariant(obj));
1628
1629         LASSERT(obj->oo_owner != env);
1630         down_read_nested(&obj->oo_sem, role);
1631
1632         LASSERT(obj->oo_owner == NULL);
1633         oti->oti_r_locks++;
1634 }
1635
1636 static void osd_object_write_lock(const struct lu_env *env,
1637                                   struct dt_object *dt, unsigned role)
1638 {
1639         struct osd_object *obj = osd_dt_obj(dt);
1640         struct osd_thread_info *oti = osd_oti_get(env);
1641
1642         LINVRNT(osd_invariant(obj));
1643
1644         LASSERT(obj->oo_owner != env);
1645         down_write_nested(&obj->oo_sem, role);
1646
1647         LASSERT(obj->oo_owner == NULL);
1648         obj->oo_owner = env;
1649         oti->oti_w_locks++;
1650 }
1651
1652 static void osd_object_read_unlock(const struct lu_env *env,
1653                                    struct dt_object *dt)
1654 {
1655         struct osd_object *obj = osd_dt_obj(dt);
1656         struct osd_thread_info *oti = osd_oti_get(env);
1657
1658         LINVRNT(osd_invariant(obj));
1659
1660         LASSERT(oti->oti_r_locks > 0);
1661         oti->oti_r_locks--;
1662         up_read(&obj->oo_sem);
1663 }
1664
1665 static void osd_object_write_unlock(const struct lu_env *env,
1666                                     struct dt_object *dt)
1667 {
1668         struct osd_object *obj = osd_dt_obj(dt);
1669         struct osd_thread_info *oti = osd_oti_get(env);
1670
1671         LINVRNT(osd_invariant(obj));
1672
1673         LASSERT(obj->oo_owner == env);
1674         LASSERT(oti->oti_w_locks > 0);
1675         oti->oti_w_locks--;
1676         obj->oo_owner = NULL;
1677         up_write(&obj->oo_sem);
1678 }
1679
1680 static int osd_object_write_locked(const struct lu_env *env,
1681                                    struct dt_object *dt)
1682 {
1683         struct osd_object *obj = osd_dt_obj(dt);
1684
1685         LINVRNT(osd_invariant(obj));
1686
1687         return obj->oo_owner == env;
1688 }
1689
1690 static struct timespec *osd_inode_time(const struct lu_env *env,
1691                                        struct inode *inode, __u64 seconds)
1692 {
1693         struct osd_thread_info  *oti = osd_oti_get(env);
1694         struct timespec         *t   = &oti->oti_time;
1695
1696         t->tv_sec = seconds;
1697         t->tv_nsec = 0;
1698         *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
1699         return t;
1700 }
1701
1702
1703 static void osd_inode_getattr(const struct lu_env *env,
1704                               struct inode *inode, struct lu_attr *attr)
1705 {
1706         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
1707                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
1708                            LA_FLAGS | LA_NLINK | LA_RDEV | LA_BLKSIZE |
1709                            LA_TYPE;
1710
1711         attr->la_atime   = LTIME_S(inode->i_atime);
1712         attr->la_mtime   = LTIME_S(inode->i_mtime);
1713         attr->la_ctime   = LTIME_S(inode->i_ctime);
1714         attr->la_mode    = inode->i_mode;
1715         attr->la_size    = i_size_read(inode);
1716         attr->la_blocks  = inode->i_blocks;
1717         attr->la_uid     = i_uid_read(inode);
1718         attr->la_gid     = i_gid_read(inode);
1719         attr->la_flags   = LDISKFS_I(inode)->i_flags;
1720         attr->la_nlink   = inode->i_nlink;
1721         attr->la_rdev    = inode->i_rdev;
1722         attr->la_blksize = 1 << inode->i_blkbits;
1723         attr->la_blkbits = inode->i_blkbits;
1724 }
1725
1726 static int osd_attr_get(const struct lu_env *env,
1727                         struct dt_object *dt,
1728                         struct lu_attr *attr)
1729 {
1730         struct osd_object *obj = osd_dt_obj(dt);
1731
1732         if (!dt_object_exists(dt))
1733                 return -ENOENT;
1734
1735         LASSERT(!dt_object_remote(dt));
1736         LINVRNT(osd_invariant(obj));
1737
1738         spin_lock(&obj->oo_guard);
1739         osd_inode_getattr(env, obj->oo_inode, attr);
1740         spin_unlock(&obj->oo_guard);
1741         return 0;
1742 }
1743
1744 static int osd_declare_attr_set(const struct lu_env *env,
1745                                 struct dt_object *dt,
1746                                 const struct lu_attr *attr,
1747                                 struct thandle *handle)
1748 {
1749         struct osd_thandle     *oh;
1750         struct osd_object      *obj;
1751         struct osd_thread_info *info = osd_oti_get(env);
1752         struct lquota_id_info  *qi = &info->oti_qi;
1753         qid_t                   uid;
1754         qid_t                   gid;
1755         long long               bspace;
1756         int                     rc = 0;
1757         bool                    enforce;
1758         ENTRY;
1759
1760         LASSERT(dt != NULL);
1761         LASSERT(handle != NULL);
1762
1763         obj = osd_dt_obj(dt);
1764         LASSERT(osd_invariant(obj));
1765
1766         oh = container_of0(handle, struct osd_thandle, ot_super);
1767         LASSERT(oh->ot_handle == NULL);
1768
1769         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
1770                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
1771
1772         if (attr == NULL || obj->oo_inode == NULL)
1773                 RETURN(rc);
1774
1775         bspace   = obj->oo_inode->i_blocks;
1776         bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
1777         bspace   = toqb(bspace);
1778
1779         /* Changing ownership is always preformed by super user, it should not
1780          * fail with EDQUOT.
1781          *
1782          * We still need to call the osd_declare_qid() to calculate the journal
1783          * credits for updating quota accounting files and to trigger quota
1784          * space adjustment once the operation is completed.*/
1785         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
1786                 /* USERQUOTA */
1787                 uid = i_uid_read(obj->oo_inode);
1788                 qi->lqi_type = USRQUOTA;
1789                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
1790                 /* inode accounting */
1791                 qi->lqi_is_blk = false;
1792
1793                 /* one more inode for the new uid ... */
1794                 qi->lqi_id.qid_uid = attr->la_uid;
1795                 qi->lqi_space      = 1;
1796                 /* Reserve credits for the new uid */
1797                 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
1798                 if (rc == -EDQUOT || rc == -EINPROGRESS)
1799                         rc = 0;
1800                 if (rc)
1801                         RETURN(rc);
1802
1803                 /* and one less inode for the current uid */
1804                 qi->lqi_id.qid_uid = uid;
1805                 qi->lqi_space      = -1;
1806                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
1807                 if (rc == -EDQUOT || rc == -EINPROGRESS)
1808                         rc = 0;
1809                 if (rc)
1810                         RETURN(rc);
1811
1812                 /* block accounting */
1813                 qi->lqi_is_blk = true;
1814
1815                 /* more blocks for the new uid ... */
1816                 qi->lqi_id.qid_uid = attr->la_uid;
1817                 qi->lqi_space      = bspace;
1818                 /*
1819                  * Credits for the new uid has been reserved, re-use "obj"
1820                  * to save credit reservation.
1821                  */
1822                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
1823                 if (rc == -EDQUOT || rc == -EINPROGRESS)
1824                         rc = 0;
1825                 if (rc)
1826                         RETURN(rc);
1827
1828                 /* and finally less blocks for the current uid */
1829                 qi->lqi_id.qid_uid = uid;
1830                 qi->lqi_space      = -bspace;
1831                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
1832                 if (rc == -EDQUOT || rc == -EINPROGRESS)
1833                         rc = 0;
1834                 if (rc)
1835                         RETURN(rc);
1836
1837                 /* GROUP QUOTA */
1838                 gid = i_gid_read(obj->oo_inode);
1839                 qi->lqi_type = GRPQUOTA;
1840                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
1841
1842                 /* inode accounting */
1843                 qi->lqi_is_blk = false;
1844
1845                 /* one more inode for the new gid ... */
1846                 qi->lqi_id.qid_gid = attr->la_gid;
1847                 qi->lqi_space      = 1;
1848                 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
1849                 if (rc == -EDQUOT || rc == -EINPROGRESS)
1850                         rc = 0;
1851                 if (rc)
1852                         RETURN(rc);
1853
1854                 /* and one less inode for the current gid */
1855                 qi->lqi_id.qid_gid = gid;
1856                 qi->lqi_space      = -1;
1857                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
1858                 if (rc == -EDQUOT || rc == -EINPROGRESS)
1859                         rc = 0;
1860                 if (rc)
1861                         RETURN(rc);
1862
1863                 /* block accounting */
1864                 qi->lqi_is_blk = true;
1865
1866                 /* more blocks for the new gid ... */
1867                 qi->lqi_id.qid_gid = attr->la_gid;
1868                 qi->lqi_space      = bspace;
1869                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
1870                 if (rc == -EDQUOT || rc == -EINPROGRESS)
1871                         rc = 0;
1872                 if (rc)
1873                         RETURN(rc);
1874
1875                 /* and finally less blocks for the current gid */
1876                 qi->lqi_id.qid_gid = gid;
1877                 qi->lqi_space      = -bspace;
1878                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
1879                 if (rc == -EDQUOT || rc == -EINPROGRESS)
1880                         rc = 0;
1881                 if (rc)
1882                         RETURN(rc);
1883         }
1884
1885         RETURN(rc);
1886 }
1887
1888 static int osd_inode_setattr(const struct lu_env *env,
1889                              struct inode *inode, const struct lu_attr *attr)
1890 {
1891         __u64 bits = attr->la_valid;
1892
1893         /* Only allow set size for regular file */
1894         if (!S_ISREG(inode->i_mode))
1895                 bits &= ~(LA_SIZE | LA_BLOCKS);
1896
1897         if (bits == 0)
1898                 return 0;
1899
1900         if (bits & LA_ATIME)
1901                 inode->i_atime  = *osd_inode_time(env, inode, attr->la_atime);
1902         if (bits & LA_CTIME)
1903                 inode->i_ctime  = *osd_inode_time(env, inode, attr->la_ctime);
1904         if (bits & LA_MTIME)
1905                 inode->i_mtime  = *osd_inode_time(env, inode, attr->la_mtime);
1906         if (bits & LA_SIZE) {
1907                 LDISKFS_I(inode)->i_disksize = attr->la_size;
1908                 i_size_write(inode, attr->la_size);
1909         }
1910
1911 #if 0
1912         /* OSD should not change "i_blocks" which is used by quota.
1913          * "i_blocks" should be changed by ldiskfs only. */
1914         if (bits & LA_BLOCKS)
1915                 inode->i_blocks = attr->la_blocks;
1916 #endif
1917         if (bits & LA_MODE)
1918                 inode->i_mode = (inode->i_mode & S_IFMT) |
1919                                 (attr->la_mode & ~S_IFMT);
1920         if (bits & LA_UID)
1921                 i_uid_write(inode, attr->la_uid);
1922         if (bits & LA_GID)
1923                 i_gid_write(inode, attr->la_gid);
1924         if (bits & LA_NLINK)
1925                 set_nlink(inode, attr->la_nlink);
1926         if (bits & LA_RDEV)
1927                 inode->i_rdev = attr->la_rdev;
1928
1929         if (bits & LA_FLAGS) {
1930                 /* always keep S_NOCMTIME */
1931                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
1932                                  S_NOCMTIME;
1933         }
1934         return 0;
1935 }
1936
1937 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
1938 {
1939         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
1940             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
1941                 struct iattr    iattr;
1942                 int             rc;
1943
1944                 ll_vfs_dq_init(inode);
1945                 iattr.ia_valid = 0;
1946                 if (attr->la_valid & LA_UID)
1947                         iattr.ia_valid |= ATTR_UID;
1948                 if (attr->la_valid & LA_GID)
1949                         iattr.ia_valid |= ATTR_GID;
1950                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
1951                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
1952
1953                 rc = ll_vfs_dq_transfer(inode, &iattr);
1954                 if (rc) {
1955                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
1956                                "enforcement enabled on the ldiskfs "
1957                                "filesystem?\n", inode->i_sb->s_id, rc);
1958                         return rc;
1959                 }
1960         }
1961         return 0;
1962 }
1963
1964 static int osd_attr_set(const struct lu_env *env,
1965                         struct dt_object *dt,
1966                         const struct lu_attr *attr,
1967                         struct thandle *handle)
1968 {
1969         struct osd_object *obj = osd_dt_obj(dt);
1970         struct inode      *inode;
1971         int rc;
1972
1973         if (!dt_object_exists(dt))
1974                 return -ENOENT;
1975
1976         LASSERT(handle != NULL);
1977         LASSERT(!dt_object_remote(dt));
1978         LASSERT(osd_invariant(obj));
1979
1980         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
1981
1982         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
1983                 struct osd_thread_info  *oti  = osd_oti_get(env);
1984                 const struct lu_fid     *fid0 = lu_object_fid(&dt->do_lu);
1985                 struct lu_fid           *fid1 = &oti->oti_fid;
1986                 struct osd_inode_id     *id   = &oti->oti_id;
1987                 struct iam_path_descr   *ipd;
1988                 struct iam_container    *bag;
1989                 struct osd_thandle      *oh;
1990                 int                      rc;
1991
1992                 fid_cpu_to_be(fid1, fid0);
1993                 memset(id, 1, sizeof(*id));
1994                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
1995                                   fid0)->oi_dir.od_container;
1996                 ipd = osd_idx_ipd_get(env, bag);
1997                 if (unlikely(ipd == NULL))
1998                         RETURN(-ENOMEM);
1999
2000                 oh = container_of0(handle, struct osd_thandle, ot_super);
2001                 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2002                                 (const struct iam_rec *)id, ipd);
2003                 osd_ipd_put(env, bag, ipd);
2004                 return(rc > 0 ? 0 : rc);
2005         }
2006
2007         inode = obj->oo_inode;
2008
2009         rc = osd_quota_transfer(inode, attr);
2010         if (rc)
2011                 return rc;
2012
2013         spin_lock(&obj->oo_guard);
2014         rc = osd_inode_setattr(env, inode, attr);
2015         spin_unlock(&obj->oo_guard);
2016
2017         if (!rc)
2018                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2019
2020         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2021
2022         return rc;
2023 }
2024
2025 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2026                                            struct osd_object *obj,
2027                                            const char *name, const int namelen)
2028 {
2029         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2030 }
2031
2032 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2033                       umode_t mode, struct dt_allocation_hint *hint,
2034                       struct thandle *th)
2035 {
2036         int result;
2037         struct osd_device  *osd = osd_obj2dev(obj);
2038         struct osd_thandle *oth;
2039         struct dt_object   *parent = NULL;
2040         struct inode       *inode;
2041
2042         LINVRNT(osd_invariant(obj));
2043         LASSERT(obj->oo_inode == NULL);
2044         LASSERT(obj->oo_hl_head == NULL);
2045
2046         if (S_ISDIR(mode) && ldiskfs_pdo) {
2047                 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2048                 if (obj->oo_hl_head == NULL)
2049                         return -ENOMEM;
2050         }
2051
2052         oth = container_of(th, struct osd_thandle, ot_super);
2053         LASSERT(oth->ot_handle->h_transaction != NULL);
2054
2055         if (hint != NULL && hint->dah_parent != NULL &&
2056             !dt_object_remote(hint->dah_parent))
2057                 parent = hint->dah_parent;
2058
2059         inode = ldiskfs_create_inode(oth->ot_handle,
2060                                      parent ? osd_dt_obj(parent)->oo_inode :
2061                                               osd_sb(osd)->s_root->d_inode,
2062                                      mode);
2063         if (!IS_ERR(inode)) {
2064                 /* Do not update file c/mtime in ldiskfs. */
2065                 inode->i_flags |= S_NOCMTIME;
2066
2067                 /* For new created object, it must be consistent,
2068                  * and it is unnecessary to scrub against it. */
2069                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2070
2071                 obj->oo_inode = inode;
2072                 result = 0;
2073         } else {
2074                 if (obj->oo_hl_head != NULL) {
2075                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2076                         obj->oo_hl_head = NULL;
2077                 }
2078                 result = PTR_ERR(inode);
2079         }
2080         LINVRNT(osd_invariant(obj));
2081         return result;
2082 }
2083
2084 enum {
2085         OSD_NAME_LEN = 255
2086 };
2087
2088 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2089                      struct lu_attr *attr,
2090                      struct dt_allocation_hint *hint,
2091                      struct dt_object_format *dof,
2092                      struct thandle *th)
2093 {
2094         int result;
2095         struct osd_thandle *oth;
2096         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX));
2097
2098         LASSERT(S_ISDIR(attr->la_mode));
2099
2100         oth = container_of(th, struct osd_thandle, ot_super);
2101         LASSERT(oth->ot_handle->h_transaction != NULL);
2102         result = osd_mkfile(info, obj, mode, hint, th);
2103
2104         return result;
2105 }
2106
2107 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2108                         struct lu_attr *attr,
2109                         struct dt_allocation_hint *hint,
2110                         struct dt_object_format *dof,
2111                         struct thandle *th)
2112 {
2113         int result;
2114         struct osd_thandle *oth;
2115         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2116
2117         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2118
2119         LASSERT(S_ISREG(attr->la_mode));
2120
2121         oth = container_of(th, struct osd_thandle, ot_super);
2122         LASSERT(oth->ot_handle->h_transaction != NULL);
2123
2124         result = osd_mkfile(info, obj, mode, hint, th);
2125         if (result == 0) {
2126                 LASSERT(obj->oo_inode != NULL);
2127                 if (feat->dif_flags & DT_IND_VARKEY)
2128                         result = iam_lvar_create(obj->oo_inode,
2129                                                  feat->dif_keysize_max,
2130                                                  feat->dif_ptrsize,
2131                                                  feat->dif_recsize_max,
2132                                                  oth->ot_handle);
2133                 else
2134                         result = iam_lfix_create(obj->oo_inode,
2135                                                  feat->dif_keysize_max,
2136                                                  feat->dif_ptrsize,
2137                                                  feat->dif_recsize_max,
2138                                                  oth->ot_handle);
2139
2140         }
2141         return result;
2142 }
2143
2144 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2145                      struct lu_attr *attr,
2146                      struct dt_allocation_hint *hint,
2147                      struct dt_object_format *dof,
2148                      struct thandle *th)
2149 {
2150         LASSERT(S_ISREG(attr->la_mode));
2151         return osd_mkfile(info, obj, (attr->la_mode &
2152                                (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2153 }
2154
2155 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2156                      struct lu_attr *attr,
2157                      struct dt_allocation_hint *hint,
2158                      struct dt_object_format *dof,
2159                      struct thandle *th)
2160 {
2161         LASSERT(S_ISLNK(attr->la_mode));
2162         return osd_mkfile(info, obj, (attr->la_mode &
2163                               (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2164 }
2165
2166 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2167                      struct lu_attr *attr,
2168                      struct dt_allocation_hint *hint,
2169                      struct dt_object_format *dof,
2170                      struct thandle *th)
2171 {
2172         umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2173         int result;
2174
2175         LINVRNT(osd_invariant(obj));
2176         LASSERT(obj->oo_inode == NULL);
2177         LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2178                 S_ISFIFO(mode) || S_ISSOCK(mode));
2179
2180         result = osd_mkfile(info, obj, mode, hint, th);
2181         if (result == 0) {
2182                 LASSERT(obj->oo_inode != NULL);
2183                 /*
2184                  * This inode should be marked dirty for i_rdev.  Currently
2185                  * that is done in the osd_attr_init().
2186                  */
2187                 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2188                                    attr->la_rdev);
2189         }
2190         LINVRNT(osd_invariant(obj));
2191         return result;
2192 }
2193
2194 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2195                               struct lu_attr *,
2196                               struct dt_allocation_hint *hint,
2197                               struct dt_object_format *dof,
2198                               struct thandle *);
2199
2200 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2201 {
2202         osd_obj_type_f result;
2203
2204         switch (type) {
2205         case DFT_DIR:
2206                 result = osd_mkdir;
2207                 break;
2208         case DFT_REGULAR:
2209                 result = osd_mkreg;
2210                 break;
2211         case DFT_SYM:
2212                 result = osd_mksym;
2213                 break;
2214         case DFT_NODE:
2215                 result = osd_mknod;
2216                 break;
2217         case DFT_INDEX:
2218                 result = osd_mk_index;
2219                 break;
2220
2221         default:
2222                 LBUG();
2223                 break;
2224         }
2225         return result;
2226 }
2227
2228
2229 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2230                         struct dt_object *parent, struct dt_object *child,
2231                         umode_t child_mode)
2232 {
2233         LASSERT(ah);
2234
2235         ah->dah_parent = parent;
2236         ah->dah_mode = child_mode;
2237 }
2238
2239 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2240                           struct lu_attr *attr, struct dt_object_format *dof)
2241 {
2242         struct inode   *inode = obj->oo_inode;
2243         __u64           valid = attr->la_valid;
2244         int             result;
2245
2246         attr->la_valid &= ~(LA_TYPE | LA_MODE);
2247
2248         if (dof->dof_type != DFT_NODE)
2249                 attr->la_valid &= ~LA_RDEV;
2250         if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2251                 attr->la_valid &= ~LA_ATIME;
2252         if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2253                 attr->la_valid &= ~LA_CTIME;
2254         if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2255                 attr->la_valid &= ~LA_MTIME;
2256
2257         result = osd_quota_transfer(inode, attr);
2258         if (result)
2259                 return;
2260
2261         if (attr->la_valid != 0) {
2262                 result = osd_inode_setattr(info->oti_env, inode, attr);
2263                 /*
2264                  * The osd_inode_setattr() should always succeed here.  The
2265                  * only error that could be returned is EDQUOT when we are
2266                  * trying to change the UID or GID of the inode. However, this
2267                  * should not happen since quota enforcement is no longer
2268                  * enabled on ldiskfs (lquota takes care of it).
2269                  */
2270                 LASSERTF(result == 0, "%d\n", result);
2271                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2272         }
2273
2274         attr->la_valid = valid;
2275 }
2276
2277 /**
2278  * Helper function for osd_object_create()
2279  *
2280  * \retval 0, on success
2281  */
2282 static int __osd_object_create(struct osd_thread_info *info,
2283                                struct osd_object *obj, struct lu_attr *attr,
2284                                struct dt_allocation_hint *hint,
2285                                struct dt_object_format *dof,
2286                                struct thandle *th)
2287 {
2288         int     result;
2289         __u32   umask;
2290
2291         osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
2292
2293         /* we drop umask so that permissions we pass are not affected */
2294         umask = current->fs->umask;
2295         current->fs->umask = 0;
2296
2297         result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
2298                                                   th);
2299         if (likely(obj->oo_inode != NULL)) {
2300                 LASSERT(obj->oo_inode->i_state & I_NEW);
2301
2302                 /* Unlock the inode before attr initialization to avoid
2303                  * unnecessary dqget operations. LU-6378 */
2304                 unlock_new_inode(obj->oo_inode);
2305         }
2306
2307         if (likely(result == 0)) {
2308                 osd_attr_init(info, obj, attr, dof);
2309                 osd_object_init0(obj);
2310         }
2311
2312         /* restore previous umask value */
2313         current->fs->umask = umask;
2314
2315         osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
2316
2317         return result;
2318 }
2319
2320 /**
2321  * Helper function for osd_object_create()
2322  *
2323  * \retval 0, on success
2324  */
2325 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
2326                            const struct lu_fid *fid, struct thandle *th)
2327 {
2328         struct osd_thread_info *info = osd_oti_get(env);
2329         struct osd_inode_id    *id   = &info->oti_id;
2330         struct osd_device      *osd  = osd_obj2dev(obj);
2331         struct osd_thandle     *oh;
2332         int                     rc;
2333
2334         LASSERT(obj->oo_inode != NULL);
2335
2336         oh = container_of0(th, struct osd_thandle, ot_super);
2337         LASSERT(oh->ot_handle);
2338         osd_trans_exec_op(env, th, OSD_OT_INSERT);
2339
2340         osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
2341         rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle, OI_CHECK_FLD);
2342         osd_trans_exec_check(env, th, OSD_OT_INSERT);
2343
2344         return rc;
2345 }
2346
2347 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
2348                    u64 seq, struct lu_seq_range *range)
2349 {
2350         struct seq_server_site  *ss = osd_seq_site(osd);
2351
2352         if (fid_seq_is_idif(seq)) {
2353                 fld_range_set_ost(range);
2354                 range->lsr_index = idif_ost_idx(seq);
2355                 return 0;
2356         }
2357
2358         if (!fid_seq_in_fldb(seq)) {
2359                 fld_range_set_mdt(range);
2360                 if (ss != NULL)
2361                         /* FIXME: If ss is NULL, it suppose not get lsr_index
2362                          * at all */
2363                         range->lsr_index = ss->ss_node_id;
2364                 return 0;
2365         }
2366
2367         LASSERT(ss != NULL);
2368         fld_range_set_any(range);
2369         /* OSD will only do local fld lookup */
2370         return fld_local_lookup(env, ss->ss_server_fld, seq, range);
2371 }
2372
2373 /*
2374  * Concurrency: no external locking is necessary.
2375  */
2376 static int osd_declare_object_create(const struct lu_env *env,
2377                                      struct dt_object *dt,
2378                                      struct lu_attr *attr,
2379                                      struct dt_allocation_hint *hint,
2380                                      struct dt_object_format *dof,
2381                                      struct thandle *handle)
2382 {
2383         struct osd_thandle      *oh;
2384         int                      rc;
2385         ENTRY;
2386
2387         LASSERT(handle != NULL);
2388
2389         oh = container_of0(handle, struct osd_thandle, ot_super);
2390         LASSERT(oh->ot_handle == NULL);
2391
2392         /* EA object consumes more credits than regular object: osd_mk_index
2393          * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
2394          * leaf_node, could involves the block, block bitmap, groups, GDT
2395          * change for each block, so add 4 * 2 credits in that case. */
2396         osd_trans_declare_op(env, oh, OSD_OT_CREATE,
2397                              osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
2398                              (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
2399         /* Reuse idle OI block may cause additional one OI block
2400          * to be changed. */
2401         osd_trans_declare_op(env, oh, OSD_OT_INSERT,
2402                              osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
2403
2404         if (!attr)
2405                 RETURN(0);
2406
2407         rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid, 1, oh,
2408                                    osd_dt_obj(dt), false, NULL, false);
2409         if (rc != 0)
2410                 RETURN(rc);
2411
2412         RETURN(rc);
2413 }
2414
2415 static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
2416                              struct lu_attr *attr,
2417                              struct dt_allocation_hint *hint,
2418                              struct dt_object_format *dof, struct thandle *th)
2419 {
2420         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
2421         struct osd_object       *obj    = osd_dt_obj(dt);
2422         struct osd_thread_info  *info   = osd_oti_get(env);
2423         int result;
2424         ENTRY;
2425
2426         if (dt_object_exists(dt))
2427                 return -EEXIST;
2428
2429         LINVRNT(osd_invariant(obj));
2430         LASSERT(!dt_object_remote(dt));
2431         LASSERT(osd_write_locked(env, obj));
2432         LASSERT(th != NULL);
2433
2434         if (unlikely(fid_is_acct(fid)))
2435                 /* Quota files can't be created from the kernel any more,
2436                  * 'tune2fs -O quota' will take care of creating them */
2437                 RETURN(-EPERM);
2438
2439         result = __osd_object_create(info, obj, attr, hint, dof, th);
2440         if (result == 0) {
2441                 result = __osd_oi_insert(env, obj, fid, th);
2442                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
2443                         obj->oo_dt.do_body_ops = &osd_body_ops;
2444         }
2445         LASSERT(ergo(result == 0,
2446                 dt_object_exists(dt) && !dt_object_remote(dt)));
2447
2448         LASSERT(osd_invariant(obj));
2449         RETURN(result);
2450 }
2451
2452 /**
2453  * Called to destroy on-disk representation of the object
2454  *
2455  * Concurrency: must be locked
2456  */
2457 static int osd_declare_object_destroy(const struct lu_env *env,
2458                                       struct dt_object *dt,
2459                                       struct thandle *th)
2460 {
2461         struct osd_object  *obj = osd_dt_obj(dt);
2462         struct inode       *inode = obj->oo_inode;
2463         struct osd_thandle *oh;
2464         int                 rc;
2465         ENTRY;
2466
2467         oh = container_of0(th, struct osd_thandle, ot_super);
2468         LASSERT(oh->ot_handle == NULL);
2469         LASSERT(inode);
2470
2471         osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
2472                              osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
2473         /* Recycle idle OI leaf may cause additional three OI blocks
2474          * to be changed. */
2475         osd_trans_declare_op(env, oh, OSD_OT_DELETE,
2476                              osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
2477         /* one less inode */
2478         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2479                                    -1, oh, obj, false, NULL, false);
2480         if (rc)
2481                 RETURN(rc);
2482         /* data to be truncated */
2483         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2484                                    0, oh, obj, true, NULL, false);
2485         RETURN(rc);
2486 }
2487
2488 static int osd_object_destroy(const struct lu_env *env,
2489                               struct dt_object *dt,
2490                               struct thandle *th)
2491 {
2492         const struct lu_fid    *fid = lu_object_fid(&dt->do_lu);
2493         struct osd_object      *obj = osd_dt_obj(dt);
2494         struct inode           *inode = obj->oo_inode;
2495         struct osd_device      *osd = osd_obj2dev(obj);
2496         struct osd_thandle     *oh;
2497         int                     result;
2498         ENTRY;
2499
2500         oh = container_of0(th, struct osd_thandle, ot_super);
2501         LASSERT(oh->ot_handle);
2502         LASSERT(inode);
2503         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
2504
2505         if (unlikely(fid_is_acct(fid)))
2506                 RETURN(-EPERM);
2507
2508         if (S_ISDIR(inode->i_mode)) {
2509                 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
2510                         inode->i_nlink == 2);
2511                 /* it will check/delete the inode from remote parent,
2512                  * how to optimize it? unlink performance impaction XXX */
2513                 result = osd_delete_from_remote_parent(env, osd, obj, oh);
2514                 if (result != 0 && result != -ENOENT) {
2515                         CERROR("%s: delete inode "DFID": rc = %d\n",
2516                                osd_name(osd), PFID(fid), result);
2517                 }
2518                 spin_lock(&obj->oo_guard);
2519                 clear_nlink(inode);
2520                 spin_unlock(&obj->oo_guard);
2521                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2522         }
2523
2524         osd_trans_exec_op(env, th, OSD_OT_DESTROY);
2525
2526         result = osd_oi_delete(osd_oti_get(env), osd, fid, oh->ot_handle,
2527                                OI_CHECK_FLD);
2528
2529         osd_trans_exec_check(env, th, OSD_OT_DESTROY);
2530         /* XXX: add to ext3 orphan list */
2531         /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
2532
2533         /* not needed in the cache anymore */
2534         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
2535         obj->oo_destroyed = 1;
2536
2537         RETURN(0);
2538 }
2539
2540 /**
2541  * Put the fid into lustre_mdt_attrs, and then place the structure
2542  * inode's ea. This fid should not be altered during the life time
2543  * of the inode.
2544  *
2545  * \retval +ve, on success
2546  * \retval -ve, on error
2547  *
2548  * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
2549  */
2550 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
2551                    const struct lu_fid *fid, __u32 compat, __u32 incompat)
2552 {
2553         struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
2554         int                      rc;
2555         ENTRY;
2556
2557         if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
2558                 RETURN(0);
2559
2560         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
2561                 rc = -ENOMEM;
2562
2563         lustre_lma_init(lma, fid, compat, incompat);
2564         lustre_lma_swab(lma);
2565
2566         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma, sizeof(*lma),
2567                              XATTR_CREATE);
2568         /* LMA may already exist, but we need to check that all the
2569          * desired compat/incompat flags have been added. */
2570         if (unlikely(rc == -EEXIST)) {
2571                 if (compat == 0 && incompat == 0)
2572                         RETURN(0);
2573
2574                 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
2575                                      XATTR_NAME_LMA, info->oti_mdt_attrs_old,
2576                                      LMA_OLD_SIZE);
2577                 if (rc <= 0)
2578                         RETURN(-EINVAL);
2579
2580                 lustre_lma_swab(lma);
2581                 if (!(~lma->lma_compat & compat) &&
2582                     !(~lma->lma_incompat & incompat))
2583                         RETURN(0);
2584
2585                 lma->lma_compat |= compat;
2586                 lma->lma_incompat |= incompat;
2587                 lustre_lma_swab(lma);
2588                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
2589                                      sizeof(*lma), XATTR_REPLACE);
2590         }
2591
2592         RETURN(rc);
2593 }
2594
2595 /**
2596  * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
2597  * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
2598  * To have compatilibility with 1.8 ldiskfs driver we need to have
2599  * magic number at start of fid data.
2600  * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
2601  * its inmemory API.
2602  */
2603 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
2604                                          const struct lu_fid *fid)
2605 {
2606         if (!fid_is_namespace_visible(fid) ||
2607             OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
2608                 param->edp_magic = 0;
2609                 return;
2610         }
2611
2612         param->edp_magic = LDISKFS_LUFID_MAGIC;
2613         param->edp_len =  sizeof(struct lu_fid) + 1;
2614         fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
2615 }
2616
2617 /**
2618  * Try to read the fid from inode ea into dt_rec.
2619  *
2620  * \param fid object fid.
2621  *
2622  * \retval 0 on success
2623  */
2624 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
2625                           __u32 ino, struct lu_fid *fid,
2626                           struct osd_inode_id *id)
2627 {
2628         struct osd_thread_info *info  = osd_oti_get(env);
2629         struct inode           *inode;
2630         ENTRY;
2631
2632         osd_id_gen(id, ino, OSD_OII_NOGEN);
2633         inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
2634         if (IS_ERR(inode))
2635                 RETURN(PTR_ERR(inode));
2636
2637         iput(inode);
2638         RETURN(0);
2639 }
2640
2641 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
2642                                         struct inode *dir,
2643                                         struct inode *parent_dir,
2644                                         const struct lu_fid *dot_fid,
2645                                         const struct lu_fid *dot_dot_fid,
2646                                         struct osd_thandle *oth)
2647 {
2648         struct ldiskfs_dentry_param *dot_ldp;
2649         struct ldiskfs_dentry_param *dot_dot_ldp;
2650
2651         dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
2652         osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
2653
2654         dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
2655         dot_ldp->edp_magic = 0;
2656
2657         return ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
2658                                       dir, dot_ldp, dot_dot_ldp);
2659 }
2660
2661 /**
2662  * Create an local agent inode for remote entry
2663  */
2664 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
2665                                                   struct osd_device *osd,
2666                                                   struct osd_object *pobj,
2667                                                   const struct lu_fid *fid,
2668                                                   __u32 type,
2669                                                   struct thandle *th)
2670 {
2671         struct osd_thread_info  *info = osd_oti_get(env);
2672         struct inode            *local;
2673         struct osd_thandle      *oh;
2674         int                     rc;
2675         ENTRY;
2676
2677         LASSERT(th);
2678         oh = container_of(th, struct osd_thandle, ot_super);
2679         LASSERT(oh->ot_handle->h_transaction != NULL);
2680
2681         local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
2682         if (IS_ERR(local)) {
2683                 CERROR("%s: create local error %d\n", osd_name(osd),
2684                        (int)PTR_ERR(local));
2685                 RETURN(local);
2686         }
2687
2688         ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
2689         unlock_new_inode(local);
2690
2691         /* Set special LMA flag for local agent inode */
2692         rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
2693         if (rc != 0) {
2694                 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
2695                        osd_name(osd), PFID(fid), rc);
2696                 RETURN(ERR_PTR(rc));
2697         }
2698
2699         if (!S_ISDIR(type))
2700                 RETURN(local);
2701
2702         rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
2703                                          lu_object_fid(&pobj->oo_dt.do_lu),
2704                                          fid, oh);
2705         if (rc != 0) {
2706                 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
2707                         osd_name(osd), PFID(fid), rc);
2708                 RETURN(ERR_PTR(rc));
2709         }
2710
2711         RETURN(local);
2712 }
2713
2714 /**
2715  * when direntry is deleted, we have to take care of possible agent inode
2716  * referenced by that. unfortunately we can't do this at that point:
2717  * iget() within a running transaction leads to deadlock and we better do
2718  * not call that every delete declaration to save performance. so we put
2719  * a potention agent inode on a list and process that once the transaction
2720  * is over. Notice it's not any worse in terms of real orphans as regular
2721  * object destroy doesn't put inodes on the on-disk orphan list. this should
2722  * be addressed separately
2723  */
2724 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
2725                                             struct osd_thandle *oh,
2726                                             __u32 ino)
2727 {
2728         struct osd_device      *osd = osd_dt_dev(oh->ot_super.th_dev);
2729         struct osd_obj_orphan *oor;
2730
2731         OBD_ALLOC_PTR(oor);
2732         if (oor == NULL)
2733                 return -ENOMEM;
2734
2735         oor->oor_ino = ino;
2736         oor->oor_env = (struct lu_env *)env;
2737         spin_lock(&osd->od_osfs_lock);
2738         list_add(&oor->oor_list, &osd->od_orphan_list);
2739         spin_unlock(&osd->od_osfs_lock);
2740
2741         oh->ot_remove_agents = 1;
2742
2743         return 0;
2744
2745 }
2746
2747 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
2748                                                 struct osd_device *osd)
2749 {
2750         struct osd_thread_info *info = osd_oti_get(env);
2751         struct osd_obj_orphan *oor, *tmp;
2752         struct osd_inode_id id;
2753         struct list_head list;
2754         struct inode *inode;
2755         struct lu_fid fid;
2756         handle_t *jh;
2757         __u32 ino;
2758
2759         INIT_LIST_HEAD(&list);
2760
2761         spin_lock(&osd->od_osfs_lock);
2762         list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
2763                 if (oor->oor_env == env) {
2764                         list_del(&oor->oor_list);
2765                         list_add(&oor->oor_list, &list);
2766                 }
2767         }
2768         spin_unlock(&osd->od_osfs_lock);
2769
2770         list_for_each_entry_safe(oor, tmp, &list, oor_list) {
2771
2772                 ino = oor->oor_ino;
2773
2774                 list_del(&oor->oor_list);
2775                 OBD_FREE_PTR(oor);
2776
2777                 osd_id_gen(&id, ino, OSD_OII_NOGEN);
2778                 inode = osd_iget_fid(info, osd, &id, &fid);
2779                 if (IS_ERR(inode))
2780                         continue;
2781
2782                 if (!osd_remote_fid(env, osd, &fid)) {
2783                         iput(inode);
2784                         continue;
2785                 }
2786
2787                 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
2788                 clear_nlink(inode);
2789                 mark_inode_dirty(inode);
2790                 ldiskfs_journal_stop(jh);
2791                 iput(inode);
2792         }
2793
2794         return 0;
2795 }
2796
2797 /**
2798  * OSD layer object create function for interoperability mode (b11826).
2799  * This is mostly similar to osd_object_create(). Only difference being, fid is
2800  * inserted into inode ea here.
2801  *
2802  * \retval   0, on success
2803  * \retval -ve, on error
2804  */
2805 static int osd_object_ea_create(const struct lu_env *env, struct dt_object *dt,
2806                                 struct lu_attr *attr,
2807                                 struct dt_allocation_hint *hint,
2808                                 struct dt_object_format *dof,
2809                                 struct thandle *th)
2810 {
2811         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
2812         struct osd_object       *obj    = osd_dt_obj(dt);
2813         struct osd_thread_info  *info   = osd_oti_get(env);
2814         int                      result;
2815
2816         ENTRY;
2817
2818         if (dt_object_exists(dt))
2819                 RETURN(-EEXIST);
2820
2821         LASSERT(osd_invariant(obj));
2822         LASSERT(!dt_object_remote(dt));
2823         LASSERT(osd_write_locked(env, obj));
2824         LASSERT(th != NULL);
2825
2826         if (unlikely(fid_is_acct(fid)))
2827                 /* Quota files can't be created from the kernel any more,
2828                  * 'tune2fs -O quota' will take care of creating them */
2829                 RETURN(-EPERM);
2830
2831         result = __osd_object_create(info, obj, attr, hint, dof, th);
2832         if (result == 0) {
2833                 if (fid_is_idif(fid) &&
2834                     !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
2835                         struct lu_fid *tfid = &info->oti_fid;
2836                         struct ost_id *oi   = &info->oti_ostid;
2837
2838                         fid_to_ostid(fid, oi);
2839                         ostid_to_fid(tfid, oi, 0);
2840                         result = osd_ea_fid_set(info, obj->oo_inode, tfid,
2841                                                 LMAC_FID_ON_OST, 0);
2842                 } else {
2843                         result = osd_ea_fid_set(info, obj->oo_inode, fid,
2844                                 fid_is_on_ost(info, osd_obj2dev(obj),
2845                                               fid, OI_CHECK_FLD) ?
2846                                 LMAC_FID_ON_OST : 0, 0);
2847                 }
2848                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
2849                         obj->oo_dt.do_body_ops = &osd_body_ops;
2850         }
2851
2852         if (result == 0)
2853                 result = __osd_oi_insert(env, obj, fid, th);
2854
2855         LASSERT(ergo(result == 0,
2856                      dt_object_exists(dt) && !dt_object_remote(dt)));
2857         LINVRNT(osd_invariant(obj));
2858         RETURN(result);
2859 }
2860
2861 static int osd_declare_object_ref_add(const struct lu_env *env,
2862                                       struct dt_object *dt,
2863                                       struct thandle *handle)
2864 {
2865         struct osd_thandle       *oh;
2866
2867         /* it's possible that object doesn't exist yet */
2868         LASSERT(handle != NULL);
2869
2870         oh = container_of0(handle, struct osd_thandle, ot_super);
2871         LASSERT(oh->ot_handle == NULL);
2872
2873         osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
2874                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2875
2876         return 0;
2877 }
2878
2879 /*
2880  * Concurrency: @dt is write locked.
2881  */
2882 static int osd_object_ref_add(const struct lu_env *env,
2883                               struct dt_object *dt, struct thandle *th)
2884 {
2885         struct osd_object  *obj = osd_dt_obj(dt);
2886         struct inode       *inode = obj->oo_inode;
2887         struct osd_thandle *oh;
2888         int                 rc = 0;
2889
2890         if (!dt_object_exists(dt))
2891                 return -ENOENT;
2892
2893         LINVRNT(osd_invariant(obj));
2894         LASSERT(!dt_object_remote(dt));
2895         LASSERT(osd_write_locked(env, obj));
2896         LASSERT(th != NULL);
2897
2898         oh = container_of0(th, struct osd_thandle, ot_super);
2899         LASSERT(oh->ot_handle != NULL);
2900
2901         osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
2902
2903         CDEBUG(D_INODE, DFID" increase nlink %d\n",
2904                PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
2905         /*
2906          * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
2907          * (65000) subdirectories by storing "1" in i_nlink if the link count
2908          * would otherwise overflow. Directory tranversal tools understand
2909          * that (st_nlink == 1) indicates that the filesystem dose not track
2910          * hard links count on the directory, and will not abort subdirectory
2911          * scanning early once (st_nlink - 2) subdirs have been found.
2912          *
2913          * This also has to properly handle the case of inodes with nlink == 0
2914          * in case they are being linked into the PENDING directory
2915          */
2916         spin_lock(&obj->oo_guard);
2917         if (unlikely(inode->i_nlink == 0))
2918                 /* inc_nlink from 0 may cause WARN_ON */
2919                 set_nlink(inode, 1);
2920         else {
2921                 ldiskfs_inc_count(oh->ot_handle, inode);
2922                 if (!S_ISDIR(inode->i_mode))
2923                         LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
2924         }
2925         spin_unlock(&obj->oo_guard);
2926
2927         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2928         LINVRNT(osd_invariant(obj));
2929
2930         osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
2931
2932         return rc;
2933 }
2934
2935 static int osd_declare_object_ref_del(const struct lu_env *env,
2936                                       struct dt_object *dt,
2937                                       struct thandle *handle)
2938 {
2939         struct osd_thandle *oh;
2940
2941         LASSERT(!dt_object_remote(dt));
2942         LASSERT(handle != NULL);
2943
2944         oh = container_of0(handle, struct osd_thandle, ot_super);
2945         LASSERT(oh->ot_handle == NULL);
2946
2947         osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
2948                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2949
2950         return 0;
2951 }
2952
2953 /*
2954  * Concurrency: @dt is write locked.
2955  */
2956 static int osd_object_ref_del(const struct lu_env *env, struct dt_object *dt,
2957                               struct thandle *th)
2958 {
2959         struct osd_object       *obj = osd_dt_obj(dt);
2960         struct inode            *inode = obj->oo_inode;
2961         struct osd_device       *osd = osd_dev(dt->do_lu.lo_dev);
2962         struct osd_thandle      *oh;
2963
2964         if (!dt_object_exists(dt))
2965                 return -ENOENT;
2966
2967         LINVRNT(osd_invariant(obj));
2968         LASSERT(!dt_object_remote(dt));
2969         LASSERT(osd_write_locked(env, obj));
2970         LASSERT(th != NULL);
2971
2972         oh = container_of0(th, struct osd_thandle, ot_super);
2973         LASSERT(oh->ot_handle != NULL);
2974
2975         osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
2976
2977         spin_lock(&obj->oo_guard);
2978         /* That can be result of upgrade from old Lustre version and
2979          * applied only to local files.  Just skip this ref_del call.
2980          * ext4_unlink() only treats this as a warning, don't LASSERT here.*/
2981         if (inode->i_nlink == 0) {
2982                 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
2983                              D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
2984                              ", maybe an upgraded file? (LU-3915)\n",
2985                              osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
2986                 spin_unlock(&obj->oo_guard);
2987                 return 0;
2988         }
2989
2990         CDEBUG(D_INODE, DFID" decrease nlink %d\n",
2991                PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
2992
2993         ldiskfs_dec_count(oh->ot_handle, inode);
2994         spin_unlock(&obj->oo_guard);
2995
2996         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2997         LINVRNT(osd_invariant(obj));
2998
2999         osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
3000
3001         return 0;
3002 }
3003
3004 /*
3005  * Get the 64-bit version for an inode.
3006  */
3007 static int osd_object_version_get(const struct lu_env *env,
3008                                   struct dt_object *dt, dt_obj_version_t *ver)
3009 {
3010         struct inode *inode = osd_dt_obj(dt)->oo_inode;
3011
3012         CDEBUG(D_INODE, "Get version "LPX64" for inode %lu\n",
3013                LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3014         *ver = LDISKFS_I(inode)->i_fs_version;
3015         return 0;
3016 }
3017
3018 /*
3019  * Concurrency: @dt is read locked.
3020  */
3021 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
3022                          struct lu_buf *buf, const char *name)
3023 {
3024         struct osd_object      *obj    = osd_dt_obj(dt);
3025         struct inode           *inode  = obj->oo_inode;
3026         struct osd_thread_info *info   = osd_oti_get(env);
3027         struct dentry          *dentry = &info->oti_obj_dentry;
3028
3029         /* version get is not real XATTR but uses xattr API */
3030         if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3031                 /* for version we are just using xattr API but change inode
3032                  * field instead */
3033                 if (buf->lb_len == 0)
3034                         return sizeof(dt_obj_version_t);
3035
3036                 if (buf->lb_len < sizeof(dt_obj_version_t))
3037                         return -ERANGE;
3038
3039                 osd_object_version_get(env, dt, buf->lb_buf);
3040
3041                 return sizeof(dt_obj_version_t);
3042         }
3043
3044         if (!dt_object_exists(dt))
3045                 return -ENOENT;
3046
3047         LASSERT(!dt_object_remote(dt));
3048         LASSERT(inode->i_op != NULL);
3049         LASSERT(inode->i_op->getxattr != NULL);
3050
3051         return __osd_xattr_get(inode, dentry, name, buf->lb_buf, buf->lb_len);
3052 }
3053
3054
3055 static int osd_declare_xattr_set(const struct lu_env *env,
3056                                  struct dt_object *dt,
3057                                  const struct lu_buf *buf, const char *name,
3058                                  int fl, struct thandle *handle)
3059 {
3060         struct osd_thandle *oh;
3061         int credits;
3062         struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3063
3064         LASSERT(handle != NULL);
3065
3066         oh = container_of0(handle, struct osd_thandle, ot_super);
3067         LASSERT(oh->ot_handle == NULL);
3068
3069         if (strcmp(name, XATTR_NAME_LMA) == 0) {
3070                 /* For non-upgrading case, the LMA is set first and
3071                  * usually fit inode. But for upgrade case, the LMA
3072                  * may be in another separated EA block. */
3073                 if (!dt_object_exists(dt))
3074                         credits = 0;
3075                 else if (fl == LU_XATTR_REPLACE)
3076                         credits = 1;
3077                 else
3078                         goto upgrade;
3079         } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3080                 credits = 1;
3081         } else {
3082 upgrade:
3083                 credits = osd_dto_credits_noquota[DTO_XATTR_SET];
3084                 if (buf && buf->lb_len > sb->s_blocksize) {
3085                         credits *= (buf->lb_len + sb->s_blocksize - 1) >>
3086                                         sb->s_blocksize_bits;
3087                 }
3088                 /*
3089                  * xattr set may involve inode quota change, reserve credits for
3090                  * dquot_initialize()
3091                  */
3092                 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
3093         }
3094
3095         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
3096
3097         return 0;
3098 }
3099
3100 /*
3101  * Set the 64-bit version for object
3102  */
3103 static void osd_object_version_set(const struct lu_env *env,
3104                                    struct dt_object *dt,
3105                                    dt_obj_version_t *new_version)
3106 {
3107         struct inode *inode = osd_dt_obj(dt)->oo_inode;
3108
3109         CDEBUG(D_INODE, "Set version "LPX64" (old "LPX64") for inode %lu\n",
3110                *new_version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3111
3112         LDISKFS_I(inode)->i_fs_version = *new_version;
3113         /** Version is set after all inode operations are finished,
3114          *  so we should mark it dirty here */
3115         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3116 }
3117
3118 /*
3119  * Concurrency: @dt is write locked.
3120  */
3121 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
3122                          const struct lu_buf *buf, const char *name, int fl,
3123                          struct thandle *handle)
3124 {
3125         struct osd_object      *obj      = osd_dt_obj(dt);
3126         struct inode           *inode    = obj->oo_inode;
3127         struct osd_thread_info *info     = osd_oti_get(env);
3128         int                     fs_flags = 0;
3129         int                     rc;
3130         ENTRY;
3131
3132         LASSERT(handle != NULL);
3133
3134         /* version set is not real XATTR */
3135         if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3136                 /* for version we are just using xattr API but change inode
3137                  * field instead */
3138                 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
3139                 osd_object_version_set(env, dt, buf->lb_buf);
3140                 return sizeof(dt_obj_version_t);
3141         }
3142
3143         CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
3144                PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
3145
3146         osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3147         if (fl & LU_XATTR_REPLACE)
3148                 fs_flags |= XATTR_REPLACE;
3149
3150         if (fl & LU_XATTR_CREATE)
3151                 fs_flags |= XATTR_CREATE;
3152
3153         if (strcmp(name, XATTR_NAME_LMV) == 0) {
3154                 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
3155                 int                      rc;
3156
3157                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
3158                 if (rc != 0)