lustre/osd-zfs/osd_object.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd-zfs/osd_object.c
  33  *
  34  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  35  * Author: Mike Pershin <tappro@whamcloud.com>
  36  * Author: Johann Lombardi <johann@whamcloud.com>
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_OSD
  40
  41 #include <lustre_ver.h>
  42 #include <libcfs/libcfs.h>
  43 #include <obd_support.h>
  44 #include <lustre_net.h>
  45 #include <obd.h>
  46 #include <obd_class.h>
  47 #include <lustre_disk.h>
  48 #include <lustre_fid.h>
  49
  50 #include "osd_internal.h"
  51
  52 #include <sys/dnode.h>
  53 #include <sys/dbuf.h>
  54 #include <sys/spa.h>
  55 #include <sys/stat.h>
  56 #include <sys/zap.h>
  57 #include <sys/spa_impl.h>
  58 #include <sys/zfs_znode.h>
  59 #include <sys/dmu_tx.h>
  60 #include <sys/dmu_objset.h>
  61 #include <sys/dsl_prop.h>
  62 #include <sys/sa_impl.h>
  63 #include <sys/txg.h>
  64
  65 char *osd_obj_tag = "osd_object";
  66
  67 static struct dt_object_operations osd_obj_ops;
  68 static struct lu_object_operations osd_lu_obj_ops;
  69 extern struct dt_body_operations osd_body_ops;
  70 static struct dt_object_operations osd_obj_otable_it_ops;
  71
  72 extern struct kmem_cache *osd_object_kmem;
  73
  74 static void
  75 osd_object_sa_fini(struct osd_object *obj)
  76 {
  77         if (obj->oo_sa_hdl) {
  78                 sa_handle_destroy(obj->oo_sa_hdl);
  79                 obj->oo_sa_hdl = NULL;
  80         }
  81 }
  82
  83 static int
  84 osd_object_sa_init(struct osd_object *obj, struct osd_device *o)
  85 {
  86         int rc;
  87
  88         LASSERT(obj->oo_sa_hdl == NULL);
  89         LASSERT(obj->oo_dn != NULL);
  90
  91         rc = osd_sa_handle_get(obj);
  92         if (rc)
  93                 return rc;
  94
  95         /* Cache the xattr object id, valid for the life of the object */
  96         rc = -sa_lookup(obj->oo_sa_hdl, SA_ZPL_XATTR(o), &obj->oo_xattr, 8);
  97         if (rc == -ENOENT) {
  98                 obj->oo_xattr = ZFS_NO_OBJECT;
  99                 rc = 0;
 100         } else if (rc) {
 101                 osd_object_sa_fini(obj);
 102         }
 103
 104         return rc;
 105 }
 106
 107 /*
 108  * Add object to list of dirty objects in tx handle.
 109  */
 110 static void
 111 osd_object_sa_dirty_add(struct osd_object *obj, struct osd_thandle *oh)
 112 {
 113         if (!list_empty(&obj->oo_sa_linkage))
 114                 return;
 115
 116         down(&oh->ot_sa_lock);
 117         write_lock(&obj->oo_attr_lock);
 118         if (likely(list_empty(&obj->oo_sa_linkage)))
 119                 list_add(&obj->oo_sa_linkage, &oh->ot_sa_list);
 120         write_unlock(&obj->oo_attr_lock);
 121         up(&oh->ot_sa_lock);
 122 }
 123
 124 /*
 125  * Release spill block dbuf hold for all dirty SAs.
 126  */
 127 void osd_object_sa_dirty_rele(struct osd_thandle *oh)
 128 {
 129         struct osd_object *obj;
 130
 131         down(&oh->ot_sa_lock);
 132         while (!list_empty(&oh->ot_sa_list)) {
 133                 obj = list_entry(oh->ot_sa_list.next,
 134                                  struct osd_object, oo_sa_linkage);
 135                 sa_spill_rele(obj->oo_sa_hdl);
 136                 write_lock(&obj->oo_attr_lock);
 137                 list_del_init(&obj->oo_sa_linkage);
 138                 write_unlock(&obj->oo_attr_lock);
 139         }
 140         up(&oh->ot_sa_lock);
 141 }
 142
 143 /*
 144  * Update the SA and add the object to the dirty list.
 145  */
 146 int osd_object_sa_update(struct osd_object *obj, sa_attr_type_t type,
 147                          void *buf, uint32_t buflen, struct osd_thandle *oh)
 148 {
 149         int rc;
 150
 151         LASSERT(obj->oo_sa_hdl != NULL);
 152         LASSERT(oh->ot_tx != NULL);
 153
 154         rc = -sa_update(obj->oo_sa_hdl, type, buf, buflen, oh->ot_tx);
 155         osd_object_sa_dirty_add(obj, oh);
 156
 157         return rc;
 158 }
 159
 160 /*
 161  * Bulk update the SA and add the object to the dirty list.
 162  */
 163 static int
 164 osd_object_sa_bulk_update(struct osd_object *obj, sa_bulk_attr_t *attrs,
 165                           int count, struct osd_thandle *oh)
 166 {
 167         int rc;
 168
 169         LASSERT(obj->oo_sa_hdl != NULL);
 170         LASSERT(oh->ot_tx != NULL);
 171
 172         rc = -sa_bulk_update(obj->oo_sa_hdl, attrs, count, oh->ot_tx);
 173         osd_object_sa_dirty_add(obj, oh);
 174
 175         return rc;
 176 }
 177
 178 /*
 179  * Retrieve the attributes of a DMU object
 180  */
 181 int __osd_object_attr_get(const struct lu_env *env, struct osd_device *o,
 182                           struct osd_object *obj, struct lu_attr *la)
 183 {
 184         struct osa_attr *osa = &osd_oti_get(env)->oti_osa;
 185         sa_bulk_attr_t  *bulk = osd_oti_get(env)->oti_attr_bulk;
 186         int              cnt = 0;
 187         int              rc;
 188         ENTRY;
 189
 190         LASSERT(obj->oo_dn != NULL);
 191
 192         la->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE | LA_TYPE |
 193                         LA_SIZE | LA_UID | LA_GID | LA_FLAGS | LA_NLINK;
 194
 195         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(o), NULL, osa->atime, 16);
 196         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(o), NULL, osa->mtime, 16);
 197         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(o), NULL, osa->ctime, 16);
 198         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(o), NULL, &osa->mode, 8);
 199         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(o), NULL, &osa->size, 8);
 200         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(o), NULL, &osa->nlink, 8);
 201         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(o), NULL, &osa->uid, 8);
 202         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(o), NULL, &osa->gid, 8);
 203         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(o), NULL, &osa->flags, 8);
 204         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
 205
 206         rc = -sa_bulk_lookup(obj->oo_sa_hdl, bulk, cnt);
 207         if (rc)
 208                 GOTO(out_sa, rc);
 209
 210         la->la_atime = osa->atime[0];
 211         la->la_mtime = osa->mtime[0];
 212         la->la_ctime = osa->ctime[0];
 213         la->la_mode = osa->mode;
 214         la->la_uid = osa->uid;
 215         la->la_gid = osa->gid;
 216         la->la_nlink = osa->nlink;
 217         la->la_flags = attrs_zfs2fs(osa->flags);
 218         la->la_size = osa->size;
 219
 220         /* Try to get extra flag from LMA. Right now, only LMAI_ORPHAN
 221          * flags is stored in LMA, and it is only for orphan directory */
 222         if (S_ISDIR(la->la_mode) && dt_object_exists(&obj->oo_dt)) {
 223                 struct osd_thread_info *info = osd_oti_get(env);
 224                 struct lustre_mdt_attrs *lma;
 225                 struct lu_buf buf;
 226
 227                 lma = (struct lustre_mdt_attrs *)info->oti_buf;
 228                 buf.lb_buf = lma;
 229                 buf.lb_len = sizeof(info->oti_buf);
 230                 rc = osd_xattr_get(env, &obj->oo_dt, &buf, XATTR_NAME_LMA);
 231                 if (rc > 0) {
 232                         rc = 0;
 233                         lma->lma_incompat = le32_to_cpu(lma->lma_incompat);
 234                         obj->oo_lma_flags =
 235                                 lma_to_lustre_flags(lma->lma_incompat);
 236
 237                 } else if (rc == -ENODATA) {
 238                         rc = 0;
 239                 }
 240         }
 241
 242         if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode)) {
 243                 rc = -sa_lookup(obj->oo_sa_hdl, SA_ZPL_RDEV(o), &osa->rdev, 8);
 244                 if (rc)
 245                         GOTO(out_sa, rc);
 246                 la->la_rdev = osa->rdev;
 247                 la->la_valid |= LA_RDEV;
 248         }
 249 out_sa:
 250
 251         RETURN(rc);
 252 }
 253
 254 int __osd_obj2dnode(objset_t *os, uint64_t oid, dnode_t **dnp)
 255 {
 256         dmu_buf_t *db;
 257         dmu_buf_impl_t *dbi;
 258         int rc;
 259
 260         rc = dmu_bonus_hold(os, oid, osd_obj_tag, &db);
 261         if (rc)
 262                 return rc;
 263
 264         dbi = (dmu_buf_impl_t *)db;
 265         DB_DNODE_ENTER(dbi);
 266         *dnp = DB_DNODE(dbi);
 267         LASSERT(*dnp != NULL);
 268
 269         return 0;
 270 }
 271
 272 /*
 273  * Concurrency: no concurrent access is possible that early in object
 274  * life-cycle.
 275  */
 276 struct lu_object *osd_object_alloc(const struct lu_env *env,
 277                                    const struct lu_object_header *hdr,
 278                                    struct lu_device *d)
 279 {
 280         struct osd_object *mo;
 281
 282         OBD_SLAB_ALLOC_PTR_GFP(mo, osd_object_kmem, GFP_NOFS);
 283         if (mo != NULL) {
 284                 struct lu_object *l;
 285
 286                 l = &mo->oo_dt.do_lu;
 287                 dt_object_init(&mo->oo_dt, NULL, d);
 288                 mo->oo_dt.do_ops = &osd_obj_ops;
 289                 l->lo_ops = &osd_lu_obj_ops;
 290                 INIT_LIST_HEAD(&mo->oo_sa_linkage);
 291                 INIT_LIST_HEAD(&mo->oo_unlinked_linkage);
 292                 init_rwsem(&mo->oo_sem);
 293                 init_rwsem(&mo->oo_guard);
 294                 rwlock_init(&mo->oo_attr_lock);
 295                 mo->oo_destroy = OSD_DESTROY_NONE;
 296                 return l;
 297         } else {
 298                 return NULL;
 299         }
 300 }
 301
 302 /*
 303  * Concurrency: shouldn't matter.
 304  */
 305 int osd_object_init0(const struct lu_env *env, struct osd_object *obj)
 306 {
 307         struct osd_device       *osd = osd_obj2dev(obj);
 308         const struct lu_fid     *fid = lu_object_fid(&obj->oo_dt.do_lu);
 309         int                      rc = 0;
 310         ENTRY;
 311
 312         if (obj->oo_dn == NULL)
 313                 RETURN(0);
 314
 315         /* object exist */
 316
 317         rc = osd_object_sa_init(obj, osd);
 318         if (rc)
 319                 RETURN(rc);
 320
 321         /* cache attrs in object */
 322         rc = __osd_object_attr_get(env, osd, obj, &obj->oo_attr);
 323         if (rc)
 324                 RETURN(rc);
 325
 326         if (likely(!fid_is_acct(fid)))
 327                 /* no body operations for accounting objects */
 328                 obj->oo_dt.do_body_ops = &osd_body_ops;
 329
 330         /*
 331          * initialize object before marking it existing
 332          */
 333         obj->oo_dt.do_lu.lo_header->loh_attr |= obj->oo_attr.la_mode & S_IFMT;
 334
 335         smp_mb();
 336         obj->oo_dt.do_lu.lo_header->loh_attr |= LOHA_EXISTS;
 337
 338         RETURN(0);
 339 }
 340
 341 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
 342 {
 343         struct osd_thread_info  *info = osd_oti_get(env);
 344         struct lu_buf           buf;
 345         int                     rc;
 346         struct lustre_mdt_attrs *lma;
 347         ENTRY;
 348
 349         CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
 350         lma = (struct lustre_mdt_attrs *)info->oti_buf;
 351         buf.lb_buf = lma;
 352         buf.lb_len = sizeof(info->oti_buf);
 353
 354         rc = osd_xattr_get(env, &obj->oo_dt, &buf, XATTR_NAME_LMA);
 355         if (rc > 0) {
 356                 rc = 0;
 357                 lustre_lma_swab(lma);
 358                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
 359                              CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
 360                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
 361                               "fid = "DFID"\n", osd_obj2dev(obj)->od_svname,
 362                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
 363                               PFID(lu_object_fid(&obj->oo_dt.do_lu)));
 364                         rc = -EOPNOTSUPP;
 365                 }
 366         } else if (rc == -ENODATA) {
 367                 /* haven't initialize LMA xattr */
 368                 rc = 0;
 369         }
 370
 371         RETURN(rc);
 372 }
 373
 374 /*
 375  * Concurrency: no concurrent access is possible that early in object
 376  * life-cycle.
 377  */
 378 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
 379                            const struct lu_object_conf *conf)
 380 {
 381         struct osd_object       *obj = osd_obj(l);
 382         struct osd_device       *osd = osd_obj2dev(obj);
 383         uint64_t                 oid;
 384         int                      rc;
 385         ENTRY;
 386
 387         LASSERT(osd_invariant(obj));
 388
 389         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
 390                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
 391                 l->lo_header->loh_attr |= LOHA_EXISTS;
 392                 RETURN(0);
 393         }
 394
 395         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 396                 GOTO(out, rc = 0);
 397
 398         rc = osd_fid_lookup(env, osd, lu_object_fid(l), &oid);
 399         if (rc == 0) {
 400                 LASSERT(obj->oo_dn == NULL);
 401                 rc = __osd_obj2dnode(osd->od_os, oid, &obj->oo_dn);
 402                 /* EEXIST will be returned if object is being deleted in ZFS */
 403                 if (rc == -EEXIST) {
 404                         rc = 0;
 405                         GOTO(out, rc);
 406                 }
 407                 if (rc != 0) {
 408                         CERROR("%s: lookup "DFID"/%#llx failed: rc = %d\n",
 409                                osd->od_svname, PFID(lu_object_fid(l)), oid, rc);
 410                         GOTO(out, rc);
 411                 }
 412                 LASSERT(obj->oo_dn);
 413                 rc = osd_object_init0(env, obj);
 414                 if (rc != 0)
 415                         GOTO(out, rc);
 416
 417                 rc = osd_check_lma(env, obj);
 418                 if (rc != 0)
 419                         GOTO(out, rc);
 420         } else if (rc == -ENOENT) {
 421                 rc = 0;
 422         }
 423         LASSERT(osd_invariant(obj));
 424 out:
 425         RETURN(rc);
 426 }
 427
 428 /*
 429  * Concurrency: no concurrent access is possible that late in object
 430  * life-cycle.
 431  */
 432 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
 433 {
 434         struct osd_object *obj = osd_obj(l);
 435
 436         LASSERT(osd_invariant(obj));
 437
 438         dt_object_fini(&obj->oo_dt);
 439         OBD_SLAB_FREE_PTR(obj, osd_object_kmem);
 440 }
 441
 442 static int
 443 osd_object_unlinked_add(struct osd_object *obj, struct osd_thandle *oh)
 444 {
 445         int rc = -EBUSY;
 446
 447         LASSERT(obj->oo_destroy == OSD_DESTROY_ASYNC);
 448
 449         /* the object is supposed to be exclusively locked by
 450          * the caller (osd_object_destroy()), while the transaction
 451          * (oh) is per-thread and not shared */
 452         if (likely(list_empty(&obj->oo_unlinked_linkage))) {
 453                 list_add(&obj->oo_unlinked_linkage, &oh->ot_unlinked_list);
 454                 rc = 0;
 455         }
 456
 457         return rc;
 458 }
 459
 460 /* Default to max data size covered by a level-1 indirect block */
 461 static unsigned long osd_sync_destroy_max_size =
 462         1UL << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT + SPA_MAXBLOCKSHIFT);
 463 module_param(osd_sync_destroy_max_size, ulong, 0444);
 464 MODULE_PARM_DESC(osd_sync_destroy_max_size, "Maximum object size to use synchronous destroy.");
 465
 466 static inline void
 467 osd_object_set_destroy_type(struct osd_object *obj)
 468 {
 469         /*
 470          * Lock-less OST_WRITE can race with OST_DESTROY, so set destroy type
 471          * only once and use it consistently thereafter.
 472          */
 473         down_write(&obj->oo_guard);
 474         if (obj->oo_destroy == OSD_DESTROY_NONE) {
 475                 if (obj->oo_attr.la_size <= osd_sync_destroy_max_size)
 476                         obj->oo_destroy = OSD_DESTROY_SYNC;
 477                 else /* Larger objects are destroyed asynchronously */
 478                         obj->oo_destroy = OSD_DESTROY_ASYNC;
 479         }
 480         up_write(&obj->oo_guard);
 481 }
 482
 483 static int osd_declare_object_destroy(const struct lu_env *env,
 484                                       struct dt_object *dt,
 485                                       struct thandle *th)
 486 {
 487         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
 488         struct osd_object       *obj = osd_dt_obj(dt);
 489         struct osd_device       *osd = osd_obj2dev(obj);
 490         struct osd_thandle      *oh;
 491         dnode_t *dn;
 492         int                      rc;
 493         uint64_t                 zapid;
 494         ENTRY;
 495
 496         LASSERT(th != NULL);
 497         LASSERT(dt_object_exists(dt));
 498
 499         oh = container_of0(th, struct osd_thandle, ot_super);
 500         LASSERT(oh->ot_tx != NULL);
 501
 502         /* declare that we'll remove object from fid-dnode mapping */
 503         zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0, &dn);
 504         osd_tx_hold_zap(oh->ot_tx, zapid, dn, FALSE, NULL);
 505
 506         osd_declare_xattrs_destroy(env, obj, oh);
 507
 508         /* one less inode */
 509         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 510                                obj->oo_attr.la_gid, -1, oh, false, NULL, false);
 511         if (rc)
 512                 RETURN(rc);
 513
 514         /* data to be truncated */
 515         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 516                                obj->oo_attr.la_gid, 0, oh, true, NULL, false);
 517         if (rc)
 518                 RETURN(rc);
 519
 520         osd_object_set_destroy_type(obj);
 521         if (obj->oo_destroy == OSD_DESTROY_SYNC)
 522                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object,
 523                                  0, DMU_OBJECT_END);
 524         else
 525                 osd_tx_hold_zap(oh->ot_tx, osd->od_unlinked->dn_object,
 526                                 osd->od_unlinked, TRUE, NULL);
 527
 528         /* will help to find FID->ino when this object is being
 529          * added to PENDING/ */
 530         osd_idc_find_and_init(env, osd, obj);
 531
 532         RETURN(0);
 533 }
 534
 535 static int osd_object_destroy(const struct lu_env *env,
 536                               struct dt_object *dt, struct thandle *th)
 537 {
 538         struct osd_thread_info  *info = osd_oti_get(env);
 539         char                    *buf = info->oti_str;
 540         struct osd_object       *obj = osd_dt_obj(dt);
 541         struct osd_device       *osd = osd_obj2dev(obj);
 542         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
 543         struct osd_thandle      *oh;
 544         int                      rc;
 545         uint64_t                 oid, zapid;
 546         dnode_t *zdn;
 547         ENTRY;
 548
 549         down_write(&obj->oo_guard);
 550
 551         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 552                 GOTO(out, rc = -ENOENT);
 553
 554         LASSERT(obj->oo_dn != NULL);
 555
 556         oh = container_of0(th, struct osd_thandle, ot_super);
 557         LASSERT(oh != NULL);
 558         LASSERT(oh->ot_tx != NULL);
 559
 560         /* remove obj ref from index dir (it depends) */
 561         zapid = osd_get_name_n_idx(env, osd, fid, buf,
 562                                    sizeof(info->oti_str), &zdn);
 563         rc = osd_zap_remove(osd, zapid, zdn, buf, oh->ot_tx);
 564         if (rc) {
 565                 CERROR("%s: zap_remove(%s) failed: rc = %d\n",
 566                        osd->od_svname, buf, rc);
 567                 GOTO(out, rc);
 568         }
 569
 570         rc = osd_xattrs_destroy(env, obj, oh);
 571         if (rc) {
 572                 CERROR("%s: cannot destroy xattrs for %s: rc = %d\n",
 573                        osd->od_svname, buf, rc);
 574                 GOTO(out, rc);
 575         }
 576
 577         oid = obj->oo_dn->dn_object;
 578         if (unlikely(obj->oo_destroy == OSD_DESTROY_NONE)) {
 579                 /* this may happen if the destroy wasn't declared
 580                  * e.g. when the object is created and then destroyed
 581                  * in the same transaction - we don't need additional
 582                  * space for destroy specifically */
 583                 LASSERT(obj->oo_attr.la_size <= osd_sync_destroy_max_size);
 584                 rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
 585                 if (rc)
 586                         CERROR("%s: failed to free %s %llu: rc = %d\n",
 587                                osd->od_svname, buf, oid, rc);
 588         } else if (obj->oo_destroy == OSD_DESTROY_SYNC) {
 589                 rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
 590                 if (rc)
 591                         CERROR("%s: failed to free %s %llu: rc = %d\n",
 592                                osd->od_svname, buf, oid, rc);
 593         } else { /* asynchronous destroy */
 594                 char *key = info->oti_key;
 595
 596                 rc = osd_object_unlinked_add(obj, oh);
 597                 if (rc)
 598                         GOTO(out, rc);
 599
 600                 snprintf(key, sizeof(info->oti_key), "%llx", oid);
 601                 rc = osd_zap_add(osd, osd->od_unlinked->dn_object,
 602                                  osd->od_unlinked, key, 8, 1, &oid, oh->ot_tx);
 603                 if (rc)
 604                         CERROR("%s: zap_add_int() failed %s %llu: rc = %d\n",
 605                                osd->od_svname, buf, oid, rc);
 606         }
 607
 608 out:
 609         /* not needed in the cache anymore */
 610         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
 611         if (rc == 0)
 612                 obj->oo_destroyed = 1;
 613         up_write(&obj->oo_guard);
 614         RETURN (0);
 615 }
 616
 617 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
 618 {
 619         struct osd_object *obj = osd_obj(l);
 620
 621         if (obj->oo_dn != NULL) {
 622                 osd_object_sa_fini(obj);
 623                 if (obj->oo_sa_xattr) {
 624                         nvlist_free(obj->oo_sa_xattr);
 625                         obj->oo_sa_xattr = NULL;
 626                 }
 627                 osd_dnode_rele(obj->oo_dn);
 628                 list_del(&obj->oo_sa_linkage);
 629                 obj->oo_dn = NULL;
 630         }
 631 }
 632
 633 /*
 634  * Concurrency: ->loo_object_release() is called under site spin-lock.
 635  */
 636 static void osd_object_release(const struct lu_env *env,
 637                                struct lu_object *l)
 638 {
 639 }
 640
 641 /*
 642  * Concurrency: shouldn't matter.
 643  */
 644 static int osd_object_print(const struct lu_env *env, void *cookie,
 645                             lu_printer_t p, const struct lu_object *l)
 646 {
 647         struct osd_object *o = osd_obj(l);
 648
 649         return (*p)(env, cookie, LUSTRE_OSD_ZFS_NAME"-object@%p", o);
 650 }
 651
 652 static void osd_object_read_lock(const struct lu_env *env,
 653                                  struct dt_object *dt, unsigned role)
 654 {
 655         struct osd_object *obj = osd_dt_obj(dt);
 656
 657         LASSERT(osd_invariant(obj));
 658
 659         down_read_nested(&obj->oo_sem, role);
 660 }
 661
 662 static void osd_object_write_lock(const struct lu_env *env,
 663                                   struct dt_object *dt, unsigned role)
 664 {
 665         struct osd_object *obj = osd_dt_obj(dt);
 666
 667         LASSERT(osd_invariant(obj));
 668
 669         down_write_nested(&obj->oo_sem, role);
 670 }
 671
 672 static void osd_object_read_unlock(const struct lu_env *env,
 673                                    struct dt_object *dt)
 674 {
 675         struct osd_object *obj = osd_dt_obj(dt);
 676
 677         LASSERT(osd_invariant(obj));
 678         up_read(&obj->oo_sem);
 679 }
 680
 681 static void osd_object_write_unlock(const struct lu_env *env,
 682                                     struct dt_object *dt)
 683 {
 684         struct osd_object *obj = osd_dt_obj(dt);
 685
 686         LASSERT(osd_invariant(obj));
 687         up_write(&obj->oo_sem);
 688 }
 689
 690 static int osd_object_write_locked(const struct lu_env *env,
 691                                    struct dt_object *dt)
 692 {
 693         struct osd_object *obj = osd_dt_obj(dt);
 694         int rc = 1;
 695
 696         LASSERT(osd_invariant(obj));
 697
 698         if (down_write_trylock(&obj->oo_sem)) {
 699                 rc = 0;
 700                 up_write(&obj->oo_sem);
 701         }
 702         return rc;
 703 }
 704
 705 static int osd_attr_get(const struct lu_env *env,
 706                         struct dt_object *dt,
 707                         struct lu_attr *attr)
 708 {
 709         struct osd_object       *obj = osd_dt_obj(dt);
 710         uint64_t                 blocks;
 711         uint32_t                 blksize;
 712         int                      rc = 0;
 713
 714         down_read(&obj->oo_guard);
 715
 716         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 717                 GOTO(out, rc = -ENOENT);
 718
 719         LASSERT(osd_invariant(obj));
 720         LASSERT(obj->oo_dn);
 721
 722         read_lock(&obj->oo_attr_lock);
 723         *attr = obj->oo_attr;
 724         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
 725                 attr->la_flags |= LUSTRE_ORPHAN_FL;
 726         read_unlock(&obj->oo_attr_lock);
 727
 728         /* with ZFS_DEBUG zrl_add_debug() called by DB_DNODE_ENTER()
 729          * from within sa_object_size() can block on a mutex, so
 730          * we can't call sa_object_size() holding rwlock */
 731         sa_object_size(obj->oo_sa_hdl, &blksize, &blocks);
 732         /* we do not control size of indices, so always calculate
 733          * it from number of blocks reported by DMU */
 734         if (S_ISDIR(attr->la_mode))
 735                 attr->la_size = 512 * blocks;
 736         /* Block size may be not set; suggest maximal I/O transfers. */
 737         if (blksize == 0)
 738                 blksize = osd_spa_maxblocksize(
 739                         dmu_objset_spa(osd_obj2dev(obj)->od_os));
 740
 741         attr->la_blksize = blksize;
 742         attr->la_blocks = blocks;
 743         attr->la_valid |= LA_BLOCKS | LA_BLKSIZE;
 744
 745 out:
 746         up_read(&obj->oo_guard);
 747         return rc;
 748 }
 749
 750 /* Simple wrapper on top of qsd API which implement quota transfer for osd
 751  * setattr needs. As a reminder, only the root user can change ownership of
 752  * a file, that's why EDQUOT & EINPROGRESS errors are discarded */
 753 static inline int qsd_transfer(const struct lu_env *env,
 754                                struct qsd_instance *qsd,
 755                                struct lquota_trans *trans, int qtype,
 756                                __u64 orig_id, __u64 new_id, __u64 bspace,
 757                                struct lquota_id_info *qi)
 758 {
 759         int     rc;
 760
 761         if (unlikely(qsd == NULL))
 762                 return 0;
 763
 764         LASSERT(qtype >= 0 && qtype < LL_MAXQUOTAS);
 765         qi->lqi_type = qtype;
 766
 767         /* inode accounting */
 768         qi->lqi_is_blk = false;
 769
 770         /* one more inode for the new owner ... */
 771         qi->lqi_id.qid_uid = new_id;
 772         qi->lqi_space      = 1;
 773         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 774         if (rc == -EDQUOT || rc == -EINPROGRESS)
 775                 rc = 0;
 776         if (rc)
 777                 return rc;
 778
 779         /* and one less inode for the current id */
 780         qi->lqi_id.qid_uid = orig_id;;
 781         qi->lqi_space      = -1;
 782         /* can't get EDQUOT when reducing usage */
 783         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 784         if (rc == -EINPROGRESS)
 785                 rc = 0;
 786         if (rc)
 787                 return rc;
 788
 789         /* block accounting */
 790         qi->lqi_is_blk = true;
 791
 792         /* more blocks for the new owner ... */
 793         qi->lqi_id.qid_uid = new_id;
 794         qi->lqi_space      = bspace;
 795         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 796         if (rc == -EDQUOT || rc == -EINPROGRESS)
 797                 rc = 0;
 798         if (rc)
 799                 return rc;
 800
 801         /* and finally less blocks for the current owner */
 802         qi->lqi_id.qid_uid = orig_id;
 803         qi->lqi_space      = -bspace;
 804         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 805         /* can't get EDQUOT when reducing usage */
 806         if (rc == -EINPROGRESS)
 807                 rc = 0;
 808         return rc;
 809 }
 810
 811 static int osd_declare_attr_set(const struct lu_env *env,
 812                                 struct dt_object *dt,
 813                                 const struct lu_attr *attr,
 814                                 struct thandle *handle)
 815 {
 816         struct osd_thread_info  *info = osd_oti_get(env);
 817         struct osd_object       *obj = osd_dt_obj(dt);
 818         struct osd_device       *osd = osd_obj2dev(obj);
 819         dmu_tx_hold_t           *txh;
 820         struct osd_thandle      *oh;
 821         uint64_t                 bspace;
 822         uint32_t                 blksize;
 823         int                      rc = 0;
 824         bool                     found;
 825         ENTRY;
 826
 827
 828         LASSERT(handle != NULL);
 829         LASSERT(osd_invariant(obj));
 830
 831         oh = container_of0(handle, struct osd_thandle, ot_super);
 832
 833         down_read(&obj->oo_guard);
 834         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 835                 GOTO(out, rc = 0);
 836
 837         LASSERT(obj->oo_sa_hdl != NULL);
 838         LASSERT(oh->ot_tx != NULL);
 839         /* regular attributes are part of the bonus buffer */
 840         /* let's check whether this object is already part of
 841          * transaction.. */
 842         found = false;
 843         for (txh = list_head(&oh->ot_tx->tx_holds); txh;
 844              txh = list_next(&oh->ot_tx->tx_holds, txh)) {
 845                 if (txh->txh_dnode == NULL)
 846                         continue;
 847                 if (txh->txh_dnode->dn_object != obj->oo_dn->dn_object)
 848                         continue;
 849                 /* this object is part of the transaction already
 850                  * we don't need to declare bonus again */
 851                 found = true;
 852                 break;
 853         }
 854         if (!found)
 855                 dmu_tx_hold_bonus(oh->ot_tx, obj->oo_dn->dn_object);
 856         if (oh->ot_tx->tx_err != 0)
 857                 GOTO(out, rc = -oh->ot_tx->tx_err);
 858
 859         if (attr && attr->la_valid & LA_FLAGS) {
 860                 /* LMA is usually a part of bonus, no need to declare
 861                  * anything else */
 862         }
 863
 864         if (attr && (attr->la_valid & (LA_UID | LA_GID))) {
 865                 sa_object_size(obj->oo_sa_hdl, &blksize, &bspace);
 866                 bspace = toqb(bspace * blksize);
 867         }
 868
 869         if (attr && attr->la_valid & LA_UID) {
 870                 /* quota enforcement for user */
 871                 if (attr->la_uid != obj->oo_attr.la_uid) {
 872                         rc = qsd_transfer(env, osd->od_quota_slave,
 873                                           &oh->ot_quota_trans, USRQUOTA,
 874                                           obj->oo_attr.la_uid, attr->la_uid,
 875                                           bspace, &info->oti_qi);
 876                         if (rc)
 877                                 GOTO(out, rc);
 878                 }
 879         }
 880         if (attr && attr->la_valid & LA_GID) {
 881                 /* quota enforcement for group */
 882                 if (attr->la_gid != obj->oo_attr.la_gid) {
 883                         rc = qsd_transfer(env, osd->od_quota_slave,
 884                                           &oh->ot_quota_trans, GRPQUOTA,
 885                                           obj->oo_attr.la_gid, attr->la_gid,
 886                                           bspace, &info->oti_qi);
 887                         if (rc)
 888                                 GOTO(out, rc);
 889                 }
 890         }
 891
 892 out:
 893         up_read(&obj->oo_guard);
 894         RETURN(rc);
 895 }
 896
 897 /*
 898  * Set the attributes of an object
 899  *
 900  * The transaction passed to this routine must have
 901  * dmu_tx_hold_bonus(tx, oid) called and then assigned
 902  * to a transaction group.
 903  */
 904 static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 905                         const struct lu_attr *la, struct thandle *handle)
 906 {
 907         struct osd_thread_info  *info = osd_oti_get(env);
 908         sa_bulk_attr_t          *bulk = osd_oti_get(env)->oti_attr_bulk;
 909         struct osd_object       *obj = osd_dt_obj(dt);
 910         struct osd_device       *osd = osd_obj2dev(obj);
 911         struct osd_thandle      *oh;
 912         struct osa_attr         *osa = &info->oti_osa;
 913         __u64                    valid = la->la_valid;
 914         int                      cnt;
 915         int                      rc = 0;
 916
 917         ENTRY;
 918
 919         down_read(&obj->oo_guard);
 920         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 921                 GOTO(out, rc = -ENOENT);
 922
 923         LASSERT(handle != NULL);
 924         LASSERT(osd_invariant(obj));
 925         LASSERT(obj->oo_sa_hdl);
 926
 927         oh = container_of0(handle, struct osd_thandle, ot_super);
 928         /* Assert that the transaction has been assigned to a
 929            transaction group. */
 930         LASSERT(oh->ot_tx->tx_txg != 0);
 931
 932         /* Only allow set size for regular file */
 933         if (!S_ISREG(dt->do_lu.lo_header->loh_attr))
 934                 valid &= ~(LA_SIZE | LA_BLOCKS);
 935
 936         if (valid & LA_CTIME && la->la_ctime == obj->oo_attr.la_ctime)
 937                 valid &= ~LA_CTIME;
 938
 939         if (valid & LA_MTIME && la->la_mtime == obj->oo_attr.la_mtime)
 940                 valid &= ~LA_MTIME;
 941
 942         if (valid & LA_ATIME && la->la_atime == obj->oo_attr.la_atime)
 943                 valid &= ~LA_ATIME;
 944
 945         if (valid == 0)
 946                 GOTO(out, rc = 0);
 947
 948         if (valid & LA_FLAGS) {
 949                 struct lustre_mdt_attrs *lma;
 950                 struct lu_buf buf;
 951
 952                 if (la->la_flags & LUSTRE_LMA_FL_MASKS) {
 953                         CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
 954                         lma = (struct lustre_mdt_attrs *)&info->oti_buf;
 955                         buf.lb_buf = lma;
 956                         buf.lb_len = sizeof(info->oti_buf);
 957                         rc = osd_xattr_get(env, &obj->oo_dt, &buf,
 958                                            XATTR_NAME_LMA);
 959                         if (rc > 0) {
 960                                 lma->lma_incompat =
 961                                         le32_to_cpu(lma->lma_incompat);
 962                                 lma->lma_incompat |=
 963                                         lustre_to_lma_flags(la->la_flags);
 964                                 lma->lma_incompat =
 965                                         cpu_to_le32(lma->lma_incompat);
 966                                 buf.lb_buf = lma;
 967                                 buf.lb_len = sizeof(*lma);
 968                                 rc = osd_xattr_set_internal(env, obj, &buf,
 969                                                             XATTR_NAME_LMA,
 970                                                             LU_XATTR_REPLACE,
 971                                                             oh);
 972                         }
 973                         if (rc < 0) {
 974                                 CWARN("%s: failed to set LMA flags: rc = %d\n",
 975                                        osd->od_svname, rc);
 976                                 RETURN(rc);
 977                         }
 978                 }
 979         }
 980
 981         write_lock(&obj->oo_attr_lock);
 982         cnt = 0;
 983         if (valid & LA_ATIME) {
 984                 osa->atime[0] = obj->oo_attr.la_atime = la->la_atime;
 985                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(osd), NULL,
 986                                  osa->atime, 16);
 987         }
 988         if (valid & LA_MTIME) {
 989                 osa->mtime[0] = obj->oo_attr.la_mtime = la->la_mtime;
 990                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(osd), NULL,
 991                                  osa->mtime, 16);
 992         }
 993         if (valid & LA_CTIME) {
 994                 osa->ctime[0] = obj->oo_attr.la_ctime = la->la_ctime;
 995                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(osd), NULL,
 996                                  osa->ctime, 16);
 997         }
 998         if (valid & LA_MODE) {
 999                 /* mode is stored along with type, so read it first */
1000                 obj->oo_attr.la_mode = (obj->oo_attr.la_mode & S_IFMT) |
1001                         (la->la_mode & ~S_IFMT);
1002                 osa->mode = obj->oo_attr.la_mode;
1003                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(osd), NULL,
1004                                  &osa->mode, 8);
1005         }
1006         if (valid & LA_SIZE) {
1007                 osa->size = obj->oo_attr.la_size = la->la_size;
1008                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(osd), NULL,
1009                                  &osa->size, 8);
1010         }
1011         if (valid & LA_NLINK) {
1012                 osa->nlink = obj->oo_attr.la_nlink = la->la_nlink;
1013                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(osd), NULL,
1014                                  &osa->nlink, 8);
1015         }
1016         if (valid & LA_RDEV) {
1017                 osa->rdev = obj->oo_attr.la_rdev = la->la_rdev;
1018                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_RDEV(osd), NULL,
1019                                  &osa->rdev, 8);
1020         }
1021         if (valid & LA_FLAGS) {
1022                 osa->flags = attrs_fs2zfs(la->la_flags);
1023                 /* many flags are not supported by zfs, so ensure a good cached
1024                  * copy */
1025                 obj->oo_attr.la_flags = attrs_zfs2fs(osa->flags);
1026                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL,
1027                                  &osa->flags, 8);
1028         }
1029         if (valid & LA_UID) {
1030                 osa->uid = obj->oo_attr.la_uid = la->la_uid;
1031                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(osd), NULL,
1032                                  &osa->uid, 8);
1033         }
1034         if (valid & LA_GID) {
1035                 osa->gid = obj->oo_attr.la_gid = la->la_gid;
1036                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(osd), NULL,
1037                                  &osa->gid, 8);
1038         }
1039         obj->oo_attr.la_valid |= valid;
1040         write_unlock(&obj->oo_attr_lock);
1041
1042         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
1043         rc = osd_object_sa_bulk_update(obj, bulk, cnt, oh);
1044
1045 out:
1046         up_read(&obj->oo_guard);
1047         RETURN(rc);
1048 }
1049
1050 /*
1051  * Object creation.
1052  *
1053  * XXX temporary solution.
1054  */
1055
1056 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
1057                         struct dt_object *parent, struct dt_object *child,
1058                         umode_t child_mode)
1059 {
1060         LASSERT(ah);
1061
1062         ah->dah_parent = parent;
1063         ah->dah_mode = child_mode;
1064
1065         if (parent != NULL && !dt_object_remote(parent)) {
1066                 /* will help to find FID->ino at dt_insert("..") */
1067                 struct osd_object *pobj = osd_dt_obj(parent);
1068
1069                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
1070         }
1071 }
1072
1073 static int osd_declare_object_create(const struct lu_env *env,
1074                                      struct dt_object *dt,
1075                                      struct lu_attr *attr,
1076                                      struct dt_allocation_hint *hint,
1077                                      struct dt_object_format *dof,
1078                                      struct thandle *handle)
1079 {
1080         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
1081         struct osd_object       *obj = osd_dt_obj(dt);
1082         struct osd_device       *osd = osd_obj2dev(obj);
1083         struct osd_thandle      *oh;
1084         uint64_t                 zapid;
1085         dnode_t                 *dn;
1086         int                      rc, dnode_size;
1087         ENTRY;
1088
1089         LASSERT(dof);
1090
1091         switch (dof->dof_type) {
1092                 case DFT_REGULAR:
1093                 case DFT_SYM:
1094                 case DFT_NODE:
1095                         if (obj->oo_dt.do_body_ops == NULL)
1096                                 obj->oo_dt.do_body_ops = &osd_body_ops;
1097                         break;
1098                 default:
1099                         break;
1100         }
1101
1102         LASSERT(handle != NULL);
1103         oh = container_of0(handle, struct osd_thandle, ot_super);
1104         LASSERT(oh->ot_tx != NULL);
1105
1106         /* this is the minimum set of EAs on every Lustre object */
1107         obj->oo_ea_in_bonus = ZFS_SA_BASE_ATTR_SIZE +
1108                                 sizeof(__u64) + /* VBR VERSION */
1109                                 sizeof(struct lustre_mdt_attrs); /* LMA */
1110         /* reserve 32 bytes for extra stuff like ACLs */
1111         dnode_size = size_roundup_power2(obj->oo_ea_in_bonus + 32);
1112
1113         switch (dof->dof_type) {
1114                 case DFT_DIR:
1115                         dt->do_index_ops = &osd_dir_ops;
1116                 case DFT_INDEX:
1117                         /* for zap create */
1118                         dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT, FALSE, NULL);
1119                         dmu_tx_hold_sa_create(oh->ot_tx, dnode_size);
1120                         break;
1121                 case DFT_REGULAR:
1122                 case DFT_SYM:
1123                 case DFT_NODE:
1124                         /* first, we'll create new object */
1125                         dmu_tx_hold_sa_create(oh->ot_tx, dnode_size);
1126                         break;
1127
1128                 default:
1129                         LBUG();
1130                         break;
1131         }
1132
1133         /* and we'll add it to some mapping */
1134         zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0, &dn);
1135         osd_tx_hold_zap(oh->ot_tx, zapid, dn, TRUE, NULL);
1136
1137         /* will help to find FID->ino mapping at dt_insert() */
1138         osd_idc_find_and_init(env, osd, obj);
1139
1140         rc = osd_declare_quota(env, osd, attr->la_uid, attr->la_gid, 1, oh,
1141                                false, NULL, false);
1142
1143         RETURN(rc);
1144 }
1145
1146 int __osd_attr_init(const struct lu_env *env, struct osd_device *osd,
1147                     sa_handle_t *sa_hdl, dmu_tx_t *tx,
1148                     struct lu_attr *la, uint64_t parent)
1149 {
1150         sa_bulk_attr_t  *bulk = osd_oti_get(env)->oti_attr_bulk;
1151         struct osa_attr *osa = &osd_oti_get(env)->oti_osa;
1152         uint64_t         gen;
1153         uint64_t         crtime[2];
1154         timestruc_t      now;
1155         int              cnt;
1156         int              rc;
1157
1158         LASSERT(sa_hdl);
1159
1160         gen = dmu_tx_get_txg(tx);
1161         gethrestime(&now);
1162         ZFS_TIME_ENCODE(&now, crtime);
1163
1164         osa->atime[0] = la->la_atime;
1165         osa->ctime[0] = la->la_ctime;
1166         osa->mtime[0] = la->la_mtime;
1167         osa->mode = la->la_mode;
1168         osa->uid = la->la_uid;
1169         osa->gid = la->la_gid;
1170         osa->rdev = la->la_rdev;
1171         osa->nlink = la->la_nlink;
1172         osa->flags = attrs_fs2zfs(la->la_flags);
1173         osa->size  = la->la_size;
1174
1175         /*
1176          * we need to create all SA below upon object create.
1177          *
1178          * XXX The attribute order matters since the accounting callback relies
1179          * on static offsets (i.e. SA_*_OFFSET, see zfs_space_delta_cb()) to
1180          * look up the UID/GID attributes. Moreover, the callback does not seem
1181          * to support the spill block.
1182          * We define attributes in the same order as SA_*_OFFSET in order to
1183          * work around the problem. See ORI-610.
1184          */
1185         cnt = 0;
1186         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(osd), NULL, &osa->mode, 8);
1187         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(osd), NULL, &osa->size, 8);
1188         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GEN(osd), NULL, &gen, 8);
1189         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(osd), NULL, &osa->uid, 8);
1190         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(osd), NULL, &osa->gid, 8);
1191         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_PARENT(osd), NULL, &parent, 8);
1192         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL, &osa->flags, 8);
1193         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(osd), NULL, osa->atime, 16);
1194         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(osd), NULL, osa->mtime, 16);
1195         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(osd), NULL, osa->ctime, 16);
1196         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CRTIME(osd), NULL, crtime, 16);
1197         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(osd), NULL, &osa->nlink, 8);
1198         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_RDEV(osd), NULL, &osa->rdev, 8);
1199         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
1200
1201         rc = -sa_replace_all_by_template(sa_hdl, bulk, cnt, tx);
1202
1203         return rc;
1204 }
1205
1206 static int osd_find_new_dnode(const struct lu_env *env, dmu_tx_t *tx,
1207                               uint64_t oid, dnode_t **dnp)
1208 {
1209         dmu_tx_hold_t *txh;
1210         int rc = 0;
1211
1212         /* take dnode_t from tx to save on dnode#->dnode_t lookup */
1213         for (txh = list_tail(&tx->tx_holds); txh;
1214              txh = list_prev(&tx->tx_holds, txh)) {
1215                 dnode_t *dn = txh->txh_dnode;
1216                 dmu_buf_impl_t *db;
1217
1218                 if (dn == NULL)
1219                         continue;
1220                 if (dn->dn_object != oid)
1221                         continue;
1222                 db = dn->dn_bonus;
1223                 if (db == NULL) {
1224                         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1225                         if (dn->dn_bonus == NULL)
1226                                 dbuf_create_bonus(dn);
1227                         rw_exit(&dn->dn_struct_rwlock);
1228                 }
1229                 db = dn->dn_bonus;
1230                 LASSERT(db);
1231                 LASSERT(dn->dn_handle);
1232                 DB_DNODE_ENTER(db);
1233                 if (refcount_add(&db->db_holds, osd_obj_tag) == 1) {
1234                         refcount_add(&dn->dn_holds, tag);
1235                         atomic_inc_32(&dn->dn_dbufs_count);
1236                 }
1237                 *dnp = dn;
1238                 dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH);
1239                 break;
1240         }
1241
1242         if (unlikely(*dnp == NULL))
1243                 rc = __osd_obj2dnode(tx->tx_objset, oid, dnp);
1244
1245         return rc;
1246 }
1247
1248 /*
1249  * The transaction passed to this routine must have
1250  * dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT) called and then assigned
1251  * to a transaction group.
1252  */
1253 int __osd_object_create(const struct lu_env *env, struct osd_object *obj,
1254                         dnode_t **dnp, dmu_tx_t *tx, struct lu_attr *la)
1255 {
1256         struct osd_device   *osd = osd_obj2dev(obj);
1257         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
1258         dmu_object_type_t    type = DMU_OT_PLAIN_FILE_CONTENTS;
1259         uint64_t oid;
1260
1261         /* Use DMU_OTN_UINT8_METADATA for local objects so their data blocks
1262          * would get an additional ditto copy */
1263         if (unlikely(S_ISREG(la->la_mode) &&
1264                      fid_seq_is_local_file(fid_seq(fid))))
1265                 type = DMU_OTN_UINT8_METADATA;
1266
1267         /* Create a new DMU object using the default dnode size. */
1268         oid = osd_dmu_object_alloc(osd->od_os, type, 0, 0, tx);
1269
1270         LASSERT(la->la_valid & LA_MODE);
1271         la->la_size = 0;
1272         la->la_nlink = 1;
1273
1274         return osd_find_new_dnode(env, tx, oid, dnp);
1275 }
1276
1277 /*
1278  * The transaction passed to this routine must have
1279  * dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, ...) called and then assigned
1280  * to a transaction group.
1281  *
1282  * Using ZAP_FLAG_HASH64 will force the ZAP to always be a FAT ZAP.
1283  * This is fine for directories today, because storing the FID in the dirent
1284  * will also require a FAT ZAP.  If there is a new type of micro ZAP created
1285  * then we might need to re-evaluate the use of this flag and instead do
1286  * a conversion from the different internal ZAP hash formats being used. */
1287 int __osd_zap_create(const struct lu_env *env, struct osd_device *osd,
1288                      dnode_t **dnp, dmu_tx_t *tx, struct lu_attr *la,
1289                      zap_flags_t flags)
1290 {
1291         uint64_t oid;
1292
1293         /* Assert that the transaction has been assigned to a
1294            transaction group. */
1295         LASSERT(tx->tx_txg != 0);
1296         *dnp = NULL;
1297
1298         oid = osd_zap_create_flags(osd->od_os, 0, flags | ZAP_FLAG_HASH64,
1299                                    DMU_OT_DIRECTORY_CONTENTS,
1300                                    14, /* == ZFS fzap_default_blockshift */
1301                                    DN_MAX_INDBLKSHIFT, /* indirect blockshift */
1302                                    0, tx);
1303
1304         la->la_size = 2;
1305         la->la_nlink = 1;
1306
1307         return osd_find_new_dnode(env, tx, oid, dnp);
1308 }
1309
1310 static dnode_t *osd_mkidx(const struct lu_env *env, struct osd_object *obj,
1311                           struct lu_attr *la, struct osd_thandle *oh)
1312 {
1313         dnode_t *dn;
1314         int rc;
1315
1316         /* Index file should be created as regular file in order not to confuse
1317          * ZPL which could interpret them as directory.
1318          * We set ZAP_FLAG_UINT64_KEY to let ZFS know than we are going to use
1319          * binary keys */
1320         LASSERT(S_ISREG(la->la_mode));
1321         rc = __osd_zap_create(env, osd_obj2dev(obj), &dn, oh->ot_tx, la,
1322                               ZAP_FLAG_UINT64_KEY);
1323         if (rc)
1324                 return ERR_PTR(rc);
1325         return dn;
1326 }
1327
1328 static dnode_t *osd_mkdir(const struct lu_env *env, struct osd_object *obj,
1329                           struct lu_attr *la, struct osd_thandle *oh)
1330 {
1331         dnode_t *dn;
1332         int rc;
1333
1334         LASSERT(S_ISDIR(la->la_mode));
1335         rc = __osd_zap_create(env, osd_obj2dev(obj), &dn, oh->ot_tx, la, 0);
1336         if (rc)
1337                 return ERR_PTR(rc);
1338         return dn;
1339 }
1340
1341 static dnode_t *osd_mkreg(const struct lu_env *env, struct osd_object *obj,
1342                           struct lu_attr *la, struct osd_thandle *oh)
1343 {
1344         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
1345         struct osd_device *osd = osd_obj2dev(obj);
1346         dnode_t *dn;
1347         int rc;
1348
1349         LASSERT(S_ISREG(la->la_mode));
1350         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1351         if (rc)
1352                 return ERR_PTR(rc);
1353
1354         if ((fid_is_idif(fid) || fid_is_norm(fid) || fid_is_echo(fid)) &&
1355             osd->od_is_ost) {
1356                 /* The minimum block size must be at least page size otherwise
1357                  * it will break the assumption in tgt_thread_big_cache where
1358                  * the array size is PTLRPC_MAX_BRW_PAGES. It will also affect
1359                  * RDMA due to subpage transfer size */
1360                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
1361                                                PAGE_SIZE, 0, oh->ot_tx);
1362                 if (unlikely(rc)) {
1363                         CERROR("%s: can't change blocksize: %d\n",
1364                                osd->od_svname, rc);
1365                         return ERR_PTR(rc);
1366                 }
1367         }
1368
1369         return dn;
1370 }
1371
1372 static dnode_t *osd_mksym(const struct lu_env *env, struct osd_object *obj,
1373                           struct lu_attr *la, struct osd_thandle *oh)
1374 {
1375         dnode_t *dn;
1376         int rc;
1377
1378         LASSERT(S_ISLNK(la->la_mode));
1379         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1380         if (rc)
1381                 return ERR_PTR(rc);
1382         return dn;
1383 }
1384
1385 static dnode_t *osd_mknod(const struct lu_env *env, struct osd_object *obj,
1386                           struct lu_attr *la, struct osd_thandle *oh)
1387 {
1388         dnode_t *dn;
1389         int rc;
1390
1391         if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode))
1392                 la->la_valid |= LA_RDEV;
1393
1394         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1395         if (rc)
1396                 return ERR_PTR(rc);
1397         return dn;
1398 }
1399
1400 typedef dnode_t *(*osd_obj_type_f)(const struct lu_env *env,
1401                                    struct osd_object *obj,
1402                                    struct lu_attr *la,
1403                                    struct osd_thandle *oh);
1404
1405 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
1406 {
1407         osd_obj_type_f result;
1408
1409         switch (type) {
1410         case DFT_DIR:
1411                 result = osd_mkdir;
1412                 break;
1413         case DFT_INDEX:
1414                 result = osd_mkidx;
1415                 break;
1416         case DFT_REGULAR:
1417                 result = osd_mkreg;
1418                 break;
1419         case DFT_SYM:
1420                 result = osd_mksym;
1421                 break;
1422         case DFT_NODE:
1423                 result = osd_mknod;
1424                 break;
1425         default:
1426                 LBUG();
1427                 break;
1428         }
1429         return result;
1430 }
1431
1432 /*
1433  * Concurrency: @dt is write locked.
1434  */
1435 static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
1436                              struct lu_attr *attr,
1437                              struct dt_allocation_hint *hint,
1438                              struct dt_object_format *dof,
1439                              struct thandle *th)
1440 {
1441         struct osd_thread_info  *info = osd_oti_get(env);
1442         struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
1443         struct zpl_direntry     *zde = &info->oti_zde.lzd_reg;
1444         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
1445         struct osd_object       *obj = osd_dt_obj(dt);
1446         struct osd_device       *osd = osd_obj2dev(obj);
1447         char                    *buf = info->oti_str;
1448         struct osd_thandle      *oh;
1449         dnode_t *dn = NULL, *zdn = NULL;
1450         uint64_t                 zapid, parent = 0;
1451         int                      rc;
1452
1453         ENTRY;
1454
1455         /* concurrent create declarations should not see
1456          * the object inconsistent (db, attr, etc).
1457          * in regular cases acquisition should be cheap */
1458         down_write(&obj->oo_guard);
1459
1460         if (unlikely(dt_object_exists(dt)))
1461                 GOTO(out, rc = -EEXIST);
1462
1463         LASSERT(osd_invariant(obj));
1464         LASSERT(dof != NULL);
1465
1466         LASSERT(th != NULL);
1467         oh = container_of0(th, struct osd_thandle, ot_super);
1468
1469         LASSERT(obj->oo_dn == NULL);
1470
1471         /* to follow ZFS on-disk format we need
1472          * to initialize parent dnode properly */
1473         if (hint != NULL && hint->dah_parent != NULL &&
1474             !dt_object_remote(hint->dah_parent))
1475                 parent = osd_dt_obj(hint->dah_parent)->oo_dn->dn_object;
1476
1477         /* we may fix some attributes, better do not change the source */
1478         obj->oo_attr = *attr;
1479         obj->oo_attr.la_valid |= LA_SIZE | LA_NLINK | LA_TYPE;
1480
1481         dn = osd_create_type_f(dof->dof_type)(env, obj, &obj->oo_attr, oh);
1482         if (IS_ERR(dn)) {
1483                 rc = PTR_ERR(dn);
1484                 dn = NULL;
1485                 GOTO(out, rc);
1486         }
1487
1488         zde->zde_pad = 0;
1489         zde->zde_dnode = dn->dn_object;
1490         zde->zde_type = IFTODT(attr->la_mode & S_IFMT);
1491
1492         zapid = osd_get_name_n_idx(env, osd, fid, buf,
1493                                    sizeof(info->oti_str), &zdn);
1494         rc = osd_zap_add(osd, zapid, zdn, buf, 8, 1, zde, oh->ot_tx);
1495         if (rc)
1496                 GOTO(out, rc);
1497         obj->oo_dn = dn;
1498         /* Now add in all of the "SA" attributes */
1499         rc = osd_sa_handle_get(obj);
1500         if (rc)
1501                 GOTO(out, rc);
1502
1503         /* configure new osd object */
1504         parent = parent != 0 ? parent : zapid;
1505         rc = __osd_attr_init(env, osd, obj->oo_sa_hdl, oh->ot_tx,
1506                              &obj->oo_attr, parent);
1507         if (rc)
1508                 GOTO(out, rc);
1509
1510         /* XXX: oo_lma_flags */
1511         obj->oo_dt.do_lu.lo_header->loh_attr |= obj->oo_attr.la_mode & S_IFMT;
1512         if (likely(!fid_is_acct(lu_object_fid(&obj->oo_dt.do_lu))))
1513                 /* no body operations for accounting objects */
1514                 obj->oo_dt.do_body_ops = &osd_body_ops;
1515
1516         rc = -nvlist_alloc(&obj->oo_sa_xattr, NV_UNIQUE_NAME, KM_SLEEP);
1517         if (rc)
1518                 GOTO(out, rc);
1519
1520         /* initialize LMA */
1521         lustre_lma_init(lma, lu_object_fid(&obj->oo_dt.do_lu), 0, 0);
1522         lustre_lma_swab(lma);
1523         rc = -nvlist_add_byte_array(obj->oo_sa_xattr, XATTR_NAME_LMA,
1524                                     (uchar_t *)lma, sizeof(*lma));
1525         if (rc)
1526                 GOTO(out, rc);
1527         rc = __osd_sa_xattr_update(env, obj, oh);
1528         if (rc)
1529                 GOTO(out, rc);
1530         osd_idc_find_and_init(env, osd, obj);
1531
1532 out:
1533         if (unlikely(rc && dn)) {
1534                 dmu_object_free(osd->od_os, dn->dn_object, oh->ot_tx);
1535                 osd_dnode_rele(dn);
1536                 obj->oo_dn = NULL;
1537         } else if (!rc) {
1538                 obj->oo_dt.do_lu.lo_header->loh_attr |= LOHA_EXISTS;
1539         }
1540         up_write(&obj->oo_guard);
1541         RETURN(rc);
1542 }
1543
1544 static int osd_declare_object_ref_add(const struct lu_env *env,
1545                                       struct dt_object *dt,
1546                                       struct thandle *th)
1547 {
1548         return osd_declare_attr_set(env, dt, NULL, th);
1549 }
1550
1551 /*
1552  * Concurrency: @dt is write locked.
1553  */
1554 static int osd_object_ref_add(const struct lu_env *env,
1555                               struct dt_object *dt,
1556                               struct thandle *handle)
1557 {
1558         struct osd_object       *obj = osd_dt_obj(dt);
1559         struct osd_thandle      *oh;
1560         struct osd_device       *osd = osd_obj2dev(obj);
1561         uint64_t                 nlink;
1562         int rc;
1563
1564         ENTRY;
1565
1566         down_read(&obj->oo_guard);
1567         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
1568                 GOTO(out, rc = -ENOENT);
1569
1570         LASSERT(osd_invariant(obj));
1571         LASSERT(obj->oo_sa_hdl != NULL);
1572
1573         oh = container_of0(handle, struct osd_thandle, ot_super);
1574
1575         write_lock(&obj->oo_attr_lock);
1576         nlink = ++obj->oo_attr.la_nlink;
1577         write_unlock(&obj->oo_attr_lock);
1578
1579         rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
1580
1581 out:
1582         up_read(&obj->oo_guard);
1583         RETURN(rc);
1584 }
1585
1586 static int osd_declare_object_ref_del(const struct lu_env *env,
1587                                       struct dt_object *dt,
1588                                       struct thandle *handle)
1589 {
1590         return osd_declare_attr_set(env, dt, NULL, handle);
1591 }
1592
1593 /*
1594  * Concurrency: @dt is write locked.
1595  */
1596 static int osd_object_ref_del(const struct lu_env *env,
1597                               struct dt_object *dt,
1598                               struct thandle *handle)
1599 {
1600         struct osd_object       *obj = osd_dt_obj(dt);
1601         struct osd_thandle      *oh;
1602         struct osd_device       *osd = osd_obj2dev(obj);
1603         uint64_t                 nlink;
1604         int                      rc;
1605
1606         ENTRY;
1607
1608         down_read(&obj->oo_guard);
1609
1610         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
1611                 GOTO(out, rc = -ENOENT);
1612
1613         LASSERT(osd_invariant(obj));
1614         LASSERT(obj->oo_sa_hdl != NULL);
1615
1616         oh = container_of0(handle, struct osd_thandle, ot_super);
1617         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
1618
1619         write_lock(&obj->oo_attr_lock);
1620         nlink = --obj->oo_attr.la_nlink;
1621         write_unlock(&obj->oo_attr_lock);
1622
1623         rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
1624
1625 out:
1626         up_read(&obj->oo_guard);
1627         RETURN(rc);
1628 }
1629
1630 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
1631                            __u64 start, __u64 end)
1632 {
1633         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1634         ENTRY;
1635
1636         /* XXX: no other option than syncing the whole filesystem until we
1637          * support ZIL.  If the object tracked the txg that it was last
1638          * modified in, it could pass that txg here instead of "0".  Maybe
1639          * the changes are already committed, so no wait is needed at all? */
1640         if (!osd->od_dt_dev.dd_rdonly)
1641                 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
1642
1643         RETURN(0);
1644 }
1645
1646 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
1647 {
1648         return 0;
1649 }
1650
1651 static struct dt_object_operations osd_obj_ops = {
1652         .do_read_lock           = osd_object_read_lock,
1653         .do_write_lock          = osd_object_write_lock,
1654         .do_read_unlock         = osd_object_read_unlock,
1655         .do_write_unlock        = osd_object_write_unlock,
1656         .do_write_locked        = osd_object_write_locked,
1657         .do_attr_get            = osd_attr_get,
1658         .do_declare_attr_set    = osd_declare_attr_set,
1659         .do_attr_set            = osd_attr_set,
1660         .do_ah_init             = osd_ah_init,
1661         .do_declare_create      = osd_declare_object_create,
1662         .do_create              = osd_object_create,
1663         .do_declare_destroy     = osd_declare_object_destroy,
1664         .do_destroy             = osd_object_destroy,
1665         .do_index_try           = osd_index_try,
1666         .do_declare_ref_add     = osd_declare_object_ref_add,
1667         .do_ref_add             = osd_object_ref_add,
1668         .do_declare_ref_del     = osd_declare_object_ref_del,
1669         .do_ref_del             = osd_object_ref_del,
1670         .do_xattr_get           = osd_xattr_get,
1671         .do_declare_xattr_set   = osd_declare_xattr_set,
1672         .do_xattr_set           = osd_xattr_set,
1673         .do_declare_xattr_del   = osd_declare_xattr_del,
1674         .do_xattr_del           = osd_xattr_del,
1675         .do_xattr_list          = osd_xattr_list,
1676         .do_object_sync         = osd_object_sync,
1677         .do_invalidate          = osd_invalidate,
1678 };
1679
1680 static struct lu_object_operations osd_lu_obj_ops = {
1681         .loo_object_init        = osd_object_init,
1682         .loo_object_delete      = osd_object_delete,
1683         .loo_object_release     = osd_object_release,
1684         .loo_object_free        = osd_object_free,
1685         .loo_object_print       = osd_object_print,
1686         .loo_object_invariant   = osd_object_invariant,
1687 };
1688
1689 static int osd_otable_it_attr_get(const struct lu_env *env,
1690                                 struct dt_object *dt,
1691                                 struct lu_attr *attr)
1692 {
1693         attr->la_valid = 0;
1694         return 0;
1695 }
1696
1697 static struct dt_object_operations osd_obj_otable_it_ops = {
1698         .do_attr_get    = osd_otable_it_attr_get,
1699         .do_index_try   = osd_index_try,
1700 };