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_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_destroy(const struct lu_env *env, struct dt_object *dt,
 484                                struct thandle *th)
 485 {
 486         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
 487         struct osd_object       *obj = osd_dt_obj(dt);
 488         struct osd_device       *osd = osd_obj2dev(obj);
 489         struct osd_thandle      *oh;
 490         dnode_t *dn;
 491         int                      rc;
 492         uint64_t                 zapid;
 493         ENTRY;
 494
 495         LASSERT(th != NULL);
 496         LASSERT(dt_object_exists(dt));
 497
 498         oh = container_of0(th, struct osd_thandle, ot_super);
 499         LASSERT(oh->ot_tx != NULL);
 500
 501         /* declare that we'll remove object from fid-dnode mapping */
 502         zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0, &dn);
 503         osd_tx_hold_zap(oh->ot_tx, zapid, dn, FALSE, NULL);
 504
 505         osd_declare_xattrs_destroy(env, obj, oh);
 506
 507         /* one less inode */
 508         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 509                                obj->oo_attr.la_gid, -1, oh, false, NULL, false);
 510         if (rc)
 511                 RETURN(rc);
 512
 513         /* data to be truncated */
 514         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 515                                obj->oo_attr.la_gid, 0, oh, true, NULL, false);
 516         if (rc)
 517                 RETURN(rc);
 518
 519         osd_object_set_destroy_type(obj);
 520         if (obj->oo_destroy == OSD_DESTROY_SYNC)
 521                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object,
 522                                  0, DMU_OBJECT_END);
 523         else
 524                 osd_tx_hold_zap(oh->ot_tx, osd->od_unlinked->dn_object,
 525                                 osd->od_unlinked, TRUE, NULL);
 526
 527         /* will help to find FID->ino when this object is being
 528          * added to PENDING/ */
 529         osd_idc_find_and_init(env, osd, obj);
 530
 531         RETURN(0);
 532 }
 533
 534 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
 535                        struct thandle *th)
 536 {
 537         struct osd_thread_info  *info = osd_oti_get(env);
 538         char                    *buf = info->oti_str;
 539         struct osd_object       *obj = osd_dt_obj(dt);
 540         struct osd_device       *osd = osd_obj2dev(obj);
 541         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
 542         struct osd_thandle      *oh;
 543         int                      rc;
 544         uint64_t                 oid, zapid;
 545         dnode_t *zdn;
 546         ENTRY;
 547
 548         down_write(&obj->oo_guard);
 549
 550         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 551                 GOTO(out, rc = -ENOENT);
 552
 553         LASSERT(obj->oo_dn != NULL);
 554
 555         oh = container_of0(th, struct osd_thandle, ot_super);
 556         LASSERT(oh != NULL);
 557         LASSERT(oh->ot_tx != NULL);
 558
 559         /* remove obj ref from index dir (it depends) */
 560         zapid = osd_get_name_n_idx(env, osd, fid, buf,
 561                                    sizeof(info->oti_str), &zdn);
 562         rc = osd_zap_remove(osd, zapid, zdn, buf, oh->ot_tx);
 563         if (rc) {
 564                 CERROR("%s: zap_remove(%s) failed: rc = %d\n",
 565                        osd->od_svname, buf, rc);
 566                 GOTO(out, rc);
 567         }
 568
 569         rc = osd_xattrs_destroy(env, obj, oh);
 570         if (rc) {
 571                 CERROR("%s: cannot destroy xattrs for %s: rc = %d\n",
 572                        osd->od_svname, buf, rc);
 573                 GOTO(out, rc);
 574         }
 575
 576         oid = obj->oo_dn->dn_object;
 577         if (unlikely(obj->oo_destroy == OSD_DESTROY_NONE)) {
 578                 /* this may happen if the destroy wasn't declared
 579                  * e.g. when the object is created and then destroyed
 580                  * in the same transaction - we don't need additional
 581                  * space for destroy specifically */
 582                 LASSERT(obj->oo_attr.la_size <= osd_sync_destroy_max_size);
 583                 rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
 584                 if (rc)
 585                         CERROR("%s: failed to free %s %llu: rc = %d\n",
 586                                osd->od_svname, buf, oid, rc);
 587         } else if (obj->oo_destroy == OSD_DESTROY_SYNC) {
 588                 rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
 589                 if (rc)
 590                         CERROR("%s: failed to free %s %llu: rc = %d\n",
 591                                osd->od_svname, buf, oid, rc);
 592         } else { /* asynchronous destroy */
 593                 char *key = info->oti_key;
 594
 595                 rc = osd_object_unlinked_add(obj, oh);
 596                 if (rc)
 597                         GOTO(out, rc);
 598
 599                 snprintf(key, sizeof(info->oti_key), "%llx", oid);
 600                 rc = osd_zap_add(osd, osd->od_unlinked->dn_object,
 601                                  osd->od_unlinked, key, 8, 1, &oid, oh->ot_tx);
 602                 if (rc)
 603                         CERROR("%s: zap_add_int() failed %s %llu: rc = %d\n",
 604                                osd->od_svname, buf, oid, rc);
 605         }
 606
 607 out:
 608         /* not needed in the cache anymore */
 609         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
 610         if (rc == 0)
 611                 obj->oo_destroyed = 1;
 612         up_write(&obj->oo_guard);
 613         RETURN (0);
 614 }
 615
 616 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
 617 {
 618         struct osd_object *obj = osd_obj(l);
 619
 620         if (obj->oo_dn != NULL) {
 621                 osd_object_sa_fini(obj);
 622                 if (obj->oo_sa_xattr) {
 623                         nvlist_free(obj->oo_sa_xattr);
 624                         obj->oo_sa_xattr = NULL;
 625                 }
 626                 osd_dnode_rele(obj->oo_dn);
 627                 list_del(&obj->oo_sa_linkage);
 628                 obj->oo_dn = NULL;
 629         }
 630 }
 631
 632 /*
 633  * Concurrency: ->loo_object_release() is called under site spin-lock.
 634  */
 635 static void osd_object_release(const struct lu_env *env,
 636                                struct lu_object *l)
 637 {
 638 }
 639
 640 /*
 641  * Concurrency: shouldn't matter.
 642  */
 643 static int osd_object_print(const struct lu_env *env, void *cookie,
 644                             lu_printer_t p, const struct lu_object *l)
 645 {
 646         struct osd_object *o = osd_obj(l);
 647
 648         return (*p)(env, cookie, LUSTRE_OSD_ZFS_NAME"-object@%p", o);
 649 }
 650
 651 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
 652                           unsigned role)
 653 {
 654         struct osd_object *obj = osd_dt_obj(dt);
 655
 656         LASSERT(osd_invariant(obj));
 657
 658         down_read_nested(&obj->oo_sem, role);
 659 }
 660
 661 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
 662                            unsigned role)
 663 {
 664         struct osd_object *obj = osd_dt_obj(dt);
 665
 666         LASSERT(osd_invariant(obj));
 667
 668         down_write_nested(&obj->oo_sem, role);
 669 }
 670
 671 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
 672 {
 673         struct osd_object *obj = osd_dt_obj(dt);
 674
 675         LASSERT(osd_invariant(obj));
 676         up_read(&obj->oo_sem);
 677 }
 678
 679 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
 680 {
 681         struct osd_object *obj = osd_dt_obj(dt);
 682
 683         LASSERT(osd_invariant(obj));
 684         up_write(&obj->oo_sem);
 685 }
 686
 687 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
 688 {
 689         struct osd_object *obj = osd_dt_obj(dt);
 690         int rc = 1;
 691
 692         LASSERT(osd_invariant(obj));
 693
 694         if (down_write_trylock(&obj->oo_sem)) {
 695                 rc = 0;
 696                 up_write(&obj->oo_sem);
 697         }
 698         return rc;
 699 }
 700
 701 static int osd_attr_get(const struct lu_env *env,
 702                         struct dt_object *dt,
 703                         struct lu_attr *attr)
 704 {
 705         struct osd_object       *obj = osd_dt_obj(dt);
 706         uint64_t                 blocks;
 707         uint32_t                 blksize;
 708         int                      rc = 0;
 709
 710         down_read(&obj->oo_guard);
 711
 712         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 713                 GOTO(out, rc = -ENOENT);
 714
 715         LASSERT(osd_invariant(obj));
 716         LASSERT(obj->oo_dn);
 717
 718         read_lock(&obj->oo_attr_lock);
 719         *attr = obj->oo_attr;
 720         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
 721                 attr->la_flags |= LUSTRE_ORPHAN_FL;
 722         read_unlock(&obj->oo_attr_lock);
 723
 724         /* with ZFS_DEBUG zrl_add_debug() called by DB_DNODE_ENTER()
 725          * from within sa_object_size() can block on a mutex, so
 726          * we can't call sa_object_size() holding rwlock */
 727         sa_object_size(obj->oo_sa_hdl, &blksize, &blocks);
 728         /* we do not control size of indices, so always calculate
 729          * it from number of blocks reported by DMU */
 730         if (S_ISDIR(attr->la_mode))
 731                 attr->la_size = 512 * blocks;
 732         /* Block size may be not set; suggest maximal I/O transfers. */
 733         if (blksize == 0)
 734                 blksize = osd_spa_maxblocksize(
 735                         dmu_objset_spa(osd_obj2dev(obj)->od_os));
 736
 737         attr->la_blksize = blksize;
 738         attr->la_blocks = blocks;
 739         attr->la_valid |= LA_BLOCKS | LA_BLKSIZE;
 740
 741 out:
 742         up_read(&obj->oo_guard);
 743         return rc;
 744 }
 745
 746 /* Simple wrapper on top of qsd API which implement quota transfer for osd
 747  * setattr needs. As a reminder, only the root user can change ownership of
 748  * a file, that's why EDQUOT & EINPROGRESS errors are discarded */
 749 static inline int qsd_transfer(const struct lu_env *env,
 750                                struct qsd_instance *qsd,
 751                                struct lquota_trans *trans, int qtype,
 752                                __u64 orig_id, __u64 new_id, __u64 bspace,
 753                                struct lquota_id_info *qi)
 754 {
 755         int     rc;
 756
 757         if (unlikely(qsd == NULL))
 758                 return 0;
 759
 760         LASSERT(qtype >= 0 && qtype < LL_MAXQUOTAS);
 761         qi->lqi_type = qtype;
 762
 763         /* inode accounting */
 764         qi->lqi_is_blk = false;
 765
 766         /* one more inode for the new owner ... */
 767         qi->lqi_id.qid_uid = new_id;
 768         qi->lqi_space      = 1;
 769         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 770         if (rc == -EDQUOT || rc == -EINPROGRESS)
 771                 rc = 0;
 772         if (rc)
 773                 return rc;
 774
 775         /* and one less inode for the current id */
 776         qi->lqi_id.qid_uid = orig_id;;
 777         qi->lqi_space      = -1;
 778         /* can't get EDQUOT when reducing usage */
 779         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 780         if (rc == -EINPROGRESS)
 781                 rc = 0;
 782         if (rc)
 783                 return rc;
 784
 785         /* block accounting */
 786         qi->lqi_is_blk = true;
 787
 788         /* more blocks for the new owner ... */
 789         qi->lqi_id.qid_uid = new_id;
 790         qi->lqi_space      = bspace;
 791         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 792         if (rc == -EDQUOT || rc == -EINPROGRESS)
 793                 rc = 0;
 794         if (rc)
 795                 return rc;
 796
 797         /* and finally less blocks for the current owner */
 798         qi->lqi_id.qid_uid = orig_id;
 799         qi->lqi_space      = -bspace;
 800         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 801         /* can't get EDQUOT when reducing usage */
 802         if (rc == -EINPROGRESS)
 803                 rc = 0;
 804         return rc;
 805 }
 806
 807 static int osd_declare_attr_set(const struct lu_env *env,
 808                                 struct dt_object *dt,
 809                                 const struct lu_attr *attr,
 810                                 struct thandle *handle)
 811 {
 812         struct osd_thread_info  *info = osd_oti_get(env);
 813         struct osd_object       *obj = osd_dt_obj(dt);
 814         struct osd_device       *osd = osd_obj2dev(obj);
 815         dmu_tx_hold_t           *txh;
 816         struct osd_thandle      *oh;
 817         uint64_t                 bspace;
 818         uint32_t                 blksize;
 819         int                      rc = 0;
 820         bool                     found;
 821         ENTRY;
 822
 823
 824         LASSERT(handle != NULL);
 825         LASSERT(osd_invariant(obj));
 826
 827         oh = container_of0(handle, struct osd_thandle, ot_super);
 828
 829         down_read(&obj->oo_guard);
 830         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 831                 GOTO(out, rc = 0);
 832
 833         LASSERT(obj->oo_sa_hdl != NULL);
 834         LASSERT(oh->ot_tx != NULL);
 835         /* regular attributes are part of the bonus buffer */
 836         /* let's check whether this object is already part of
 837          * transaction.. */
 838         found = false;
 839         for (txh = list_head(&oh->ot_tx->tx_holds); txh;
 840              txh = list_next(&oh->ot_tx->tx_holds, txh)) {
 841                 if (txh->txh_dnode == NULL)
 842                         continue;
 843                 if (txh->txh_dnode->dn_object != obj->oo_dn->dn_object)
 844                         continue;
 845                 /* this object is part of the transaction already
 846                  * we don't need to declare bonus again */
 847                 found = true;
 848                 break;
 849         }
 850         if (!found)
 851                 dmu_tx_hold_bonus(oh->ot_tx, obj->oo_dn->dn_object);
 852         if (oh->ot_tx->tx_err != 0)
 853                 GOTO(out, rc = -oh->ot_tx->tx_err);
 854
 855         if (attr && attr->la_valid & LA_FLAGS) {
 856                 /* LMA is usually a part of bonus, no need to declare
 857                  * anything else */
 858         }
 859
 860         if (attr && (attr->la_valid & (LA_UID | LA_GID))) {
 861                 sa_object_size(obj->oo_sa_hdl, &blksize, &bspace);
 862                 bspace = toqb(bspace * blksize);
 863         }
 864
 865         if (attr && attr->la_valid & LA_UID) {
 866                 /* quota enforcement for user */
 867                 if (attr->la_uid != obj->oo_attr.la_uid) {
 868                         rc = qsd_transfer(env, osd->od_quota_slave,
 869                                           &oh->ot_quota_trans, USRQUOTA,
 870                                           obj->oo_attr.la_uid, attr->la_uid,
 871                                           bspace, &info->oti_qi);
 872                         if (rc)
 873                                 GOTO(out, rc);
 874                 }
 875         }
 876         if (attr && attr->la_valid & LA_GID) {
 877                 /* quota enforcement for group */
 878                 if (attr->la_gid != obj->oo_attr.la_gid) {
 879                         rc = qsd_transfer(env, osd->od_quota_slave,
 880                                           &oh->ot_quota_trans, GRPQUOTA,
 881                                           obj->oo_attr.la_gid, attr->la_gid,
 882                                           bspace, &info->oti_qi);
 883                         if (rc)
 884                                 GOTO(out, rc);
 885                 }
 886         }
 887
 888 out:
 889         up_read(&obj->oo_guard);
 890         RETURN(rc);
 891 }
 892
 893 /*
 894  * Set the attributes of an object
 895  *
 896  * The transaction passed to this routine must have
 897  * dmu_tx_hold_bonus(tx, oid) called and then assigned
 898  * to a transaction group.
 899  */
 900 static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 901                         const struct lu_attr *la, struct thandle *handle)
 902 {
 903         struct osd_thread_info  *info = osd_oti_get(env);
 904         sa_bulk_attr_t          *bulk = osd_oti_get(env)->oti_attr_bulk;
 905         struct osd_object       *obj = osd_dt_obj(dt);
 906         struct osd_device       *osd = osd_obj2dev(obj);
 907         struct osd_thandle      *oh;
 908         struct osa_attr         *osa = &info->oti_osa;
 909         __u64                    valid = la->la_valid;
 910         int                      cnt;
 911         int                      rc = 0;
 912
 913         ENTRY;
 914
 915         down_read(&obj->oo_guard);
 916         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 917                 GOTO(out, rc = -ENOENT);
 918
 919         LASSERT(handle != NULL);
 920         LASSERT(osd_invariant(obj));
 921         LASSERT(obj->oo_sa_hdl);
 922
 923         oh = container_of0(handle, struct osd_thandle, ot_super);
 924         /* Assert that the transaction has been assigned to a
 925            transaction group. */
 926         LASSERT(oh->ot_tx->tx_txg != 0);
 927
 928         /* Only allow set size for regular file */
 929         if (!S_ISREG(dt->do_lu.lo_header->loh_attr))
 930                 valid &= ~(LA_SIZE | LA_BLOCKS);
 931
 932         if (valid & LA_CTIME && la->la_ctime == obj->oo_attr.la_ctime)
 933                 valid &= ~LA_CTIME;
 934
 935         if (valid & LA_MTIME && la->la_mtime == obj->oo_attr.la_mtime)
 936                 valid &= ~LA_MTIME;
 937
 938         if (valid & LA_ATIME && la->la_atime == obj->oo_attr.la_atime)
 939                 valid &= ~LA_ATIME;
 940
 941         if (valid == 0)
 942                 GOTO(out, rc = 0);
 943
 944         if (valid & LA_FLAGS) {
 945                 struct lustre_mdt_attrs *lma;
 946                 struct lu_buf buf;
 947
 948                 if (la->la_flags & LUSTRE_LMA_FL_MASKS) {
 949                         CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
 950                         lma = (struct lustre_mdt_attrs *)&info->oti_buf;
 951                         buf.lb_buf = lma;
 952                         buf.lb_len = sizeof(info->oti_buf);
 953                         rc = osd_xattr_get(env, &obj->oo_dt, &buf,
 954                                            XATTR_NAME_LMA);
 955                         if (rc > 0) {
 956                                 lma->lma_incompat =
 957                                         le32_to_cpu(lma->lma_incompat);
 958                                 lma->lma_incompat |=
 959                                         lustre_to_lma_flags(la->la_flags);
 960                                 lma->lma_incompat =
 961                                         cpu_to_le32(lma->lma_incompat);
 962                                 buf.lb_buf = lma;
 963                                 buf.lb_len = sizeof(*lma);
 964                                 rc = osd_xattr_set_internal(env, obj, &buf,
 965                                                             XATTR_NAME_LMA,
 966                                                             LU_XATTR_REPLACE,
 967                                                             oh);
 968                         }
 969                         if (rc < 0) {
 970                                 CWARN("%s: failed to set LMA flags: rc = %d\n",
 971                                        osd->od_svname, rc);
 972                                 RETURN(rc);
 973                         }
 974                 }
 975         }
 976
 977         write_lock(&obj->oo_attr_lock);
 978         cnt = 0;
 979         if (valid & LA_ATIME) {
 980                 osa->atime[0] = obj->oo_attr.la_atime = la->la_atime;
 981                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(osd), NULL,
 982                                  osa->atime, 16);
 983         }
 984         if (valid & LA_MTIME) {
 985                 osa->mtime[0] = obj->oo_attr.la_mtime = la->la_mtime;
 986                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(osd), NULL,
 987                                  osa->mtime, 16);
 988         }
 989         if (valid & LA_CTIME) {
 990                 osa->ctime[0] = obj->oo_attr.la_ctime = la->la_ctime;
 991                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(osd), NULL,
 992                                  osa->ctime, 16);
 993         }
 994         if (valid & LA_MODE) {
 995                 /* mode is stored along with type, so read it first */
 996                 obj->oo_attr.la_mode = (obj->oo_attr.la_mode & S_IFMT) |
 997                         (la->la_mode & ~S_IFMT);
 998                 osa->mode = obj->oo_attr.la_mode;
 999                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(osd), NULL,
1000                                  &osa->mode, 8);
1001         }
1002         if (valid & LA_SIZE) {
1003                 osa->size = obj->oo_attr.la_size = la->la_size;
1004                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(osd), NULL,
1005                                  &osa->size, 8);
1006         }
1007         if (valid & LA_NLINK) {
1008                 osa->nlink = obj->oo_attr.la_nlink = la->la_nlink;
1009                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(osd), NULL,
1010                                  &osa->nlink, 8);
1011         }
1012         if (valid & LA_RDEV) {
1013                 osa->rdev = obj->oo_attr.la_rdev = la->la_rdev;
1014                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_RDEV(osd), NULL,
1015                                  &osa->rdev, 8);
1016         }
1017         if (valid & LA_FLAGS) {
1018                 osa->flags = attrs_fs2zfs(la->la_flags);
1019                 /* many flags are not supported by zfs, so ensure a good cached
1020                  * copy */
1021                 obj->oo_attr.la_flags = attrs_zfs2fs(osa->flags);
1022                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL,
1023                                  &osa->flags, 8);
1024         }
1025         if (valid & LA_UID) {
1026                 osa->uid = obj->oo_attr.la_uid = la->la_uid;
1027                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(osd), NULL,
1028                                  &osa->uid, 8);
1029         }
1030         if (valid & LA_GID) {
1031                 osa->gid = obj->oo_attr.la_gid = la->la_gid;
1032                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(osd), NULL,
1033                                  &osa->gid, 8);
1034         }
1035         obj->oo_attr.la_valid |= valid;
1036         write_unlock(&obj->oo_attr_lock);
1037
1038         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
1039         rc = osd_object_sa_bulk_update(obj, bulk, cnt, oh);
1040
1041 out:
1042         up_read(&obj->oo_guard);
1043         RETURN(rc);
1044 }
1045
1046 /*
1047  * Object creation.
1048  *
1049  * XXX temporary solution.
1050  */
1051
1052 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
1053                         struct dt_object *parent, struct dt_object *child,
1054                         umode_t child_mode)
1055 {
1056         LASSERT(ah);
1057
1058         ah->dah_parent = parent;
1059         ah->dah_mode = child_mode;
1060
1061         if (parent != NULL && !dt_object_remote(parent)) {
1062                 /* will help to find FID->ino at dt_insert("..") */
1063                 struct osd_object *pobj = osd_dt_obj(parent);
1064
1065                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
1066         }
1067 }
1068
1069 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
1070                               struct lu_attr *attr,
1071                               struct dt_allocation_hint *hint,
1072                               struct dt_object_format *dof,
1073                               struct thandle *handle)
1074 {
1075         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
1076         struct osd_object       *obj = osd_dt_obj(dt);
1077         struct osd_device       *osd = osd_obj2dev(obj);
1078         struct osd_thandle      *oh;
1079         uint64_t                 zapid;
1080         dnode_t                 *dn;
1081         int                      rc, dnode_size;
1082         ENTRY;
1083
1084         LASSERT(dof);
1085
1086         switch (dof->dof_type) {
1087                 case DFT_REGULAR:
1088                 case DFT_SYM:
1089                 case DFT_NODE:
1090                         if (obj->oo_dt.do_body_ops == NULL)
1091                                 obj->oo_dt.do_body_ops = &osd_body_ops;
1092                         break;
1093                 default:
1094                         break;
1095         }
1096
1097         LASSERT(handle != NULL);
1098         oh = container_of0(handle, struct osd_thandle, ot_super);
1099         LASSERT(oh->ot_tx != NULL);
1100
1101         /* this is the minimum set of EAs on every Lustre object */
1102         obj->oo_ea_in_bonus = ZFS_SA_BASE_ATTR_SIZE +
1103                                 sizeof(__u64) + /* VBR VERSION */
1104                                 sizeof(struct lustre_mdt_attrs); /* LMA */
1105         /* reserve 32 bytes for extra stuff like ACLs */
1106         dnode_size = size_roundup_power2(obj->oo_ea_in_bonus + 32);
1107
1108         switch (dof->dof_type) {
1109                 case DFT_DIR:
1110                         dt->do_index_ops = &osd_dir_ops;
1111                 case DFT_INDEX:
1112                         /* for zap create */
1113                         dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT, FALSE, NULL);
1114                         dmu_tx_hold_sa_create(oh->ot_tx, dnode_size);
1115                         break;
1116                 case DFT_REGULAR:
1117                 case DFT_SYM:
1118                 case DFT_NODE:
1119                         /* first, we'll create new object */
1120                         dmu_tx_hold_sa_create(oh->ot_tx, dnode_size);
1121                         break;
1122
1123                 default:
1124                         LBUG();
1125                         break;
1126         }
1127
1128         /* and we'll add it to some mapping */
1129         zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0, &dn);
1130         osd_tx_hold_zap(oh->ot_tx, zapid, dn, TRUE, NULL);
1131
1132         /* will help to find FID->ino mapping at dt_insert() */
1133         osd_idc_find_and_init(env, osd, obj);
1134
1135         rc = osd_declare_quota(env, osd, attr->la_uid, attr->la_gid, 1, oh,
1136                                false, NULL, false);
1137
1138         RETURN(rc);
1139 }
1140
1141 int __osd_attr_init(const struct lu_env *env, struct osd_device *osd,
1142                     sa_handle_t *sa_hdl, dmu_tx_t *tx,
1143                     struct lu_attr *la, uint64_t parent)
1144 {
1145         sa_bulk_attr_t  *bulk = osd_oti_get(env)->oti_attr_bulk;
1146         struct osa_attr *osa = &osd_oti_get(env)->oti_osa;
1147         uint64_t         gen;
1148         uint64_t         crtime[2];
1149         timestruc_t      now;
1150         int              cnt;
1151         int              rc;
1152
1153         LASSERT(sa_hdl);
1154
1155         gen = dmu_tx_get_txg(tx);
1156         gethrestime(&now);
1157         ZFS_TIME_ENCODE(&now, crtime);
1158
1159         osa->atime[0] = la->la_atime;
1160         osa->ctime[0] = la->la_ctime;
1161         osa->mtime[0] = la->la_mtime;
1162         osa->mode = la->la_mode;
1163         osa->uid = la->la_uid;
1164         osa->gid = la->la_gid;
1165         osa->rdev = la->la_rdev;
1166         osa->nlink = la->la_nlink;
1167         osa->flags = attrs_fs2zfs(la->la_flags);
1168         osa->size  = la->la_size;
1169
1170         /*
1171          * we need to create all SA below upon object create.
1172          *
1173          * XXX The attribute order matters since the accounting callback relies
1174          * on static offsets (i.e. SA_*_OFFSET, see zfs_space_delta_cb()) to
1175          * look up the UID/GID attributes. Moreover, the callback does not seem
1176          * to support the spill block.
1177          * We define attributes in the same order as SA_*_OFFSET in order to
1178          * work around the problem. See ORI-610.
1179          */
1180         cnt = 0;
1181         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(osd), NULL, &osa->mode, 8);
1182         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(osd), NULL, &osa->size, 8);
1183         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GEN(osd), NULL, &gen, 8);
1184         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(osd), NULL, &osa->uid, 8);
1185         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(osd), NULL, &osa->gid, 8);
1186         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_PARENT(osd), NULL, &parent, 8);
1187         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL, &osa->flags, 8);
1188         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(osd), NULL, osa->atime, 16);
1189         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(osd), NULL, osa->mtime, 16);
1190         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(osd), NULL, osa->ctime, 16);
1191         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CRTIME(osd), NULL, crtime, 16);
1192         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(osd), NULL, &osa->nlink, 8);
1193         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_RDEV(osd), NULL, &osa->rdev, 8);
1194         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
1195
1196         rc = -sa_replace_all_by_template(sa_hdl, bulk, cnt, tx);
1197
1198         return rc;
1199 }
1200
1201 static int osd_find_new_dnode(const struct lu_env *env, dmu_tx_t *tx,
1202                               uint64_t oid, dnode_t **dnp)
1203 {
1204         dmu_tx_hold_t *txh;
1205         int rc = 0;
1206
1207         /* take dnode_t from tx to save on dnode#->dnode_t lookup */
1208         for (txh = list_tail(&tx->tx_holds); txh;
1209              txh = list_prev(&tx->tx_holds, txh)) {
1210                 dnode_t *dn = txh->txh_dnode;
1211                 dmu_buf_impl_t *db;
1212
1213                 if (dn == NULL)
1214                         continue;
1215                 if (dn->dn_object != oid)
1216                         continue;
1217                 db = dn->dn_bonus;
1218                 if (db == NULL) {
1219                         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1220                         if (dn->dn_bonus == NULL)
1221                                 dbuf_create_bonus(dn);
1222                         rw_exit(&dn->dn_struct_rwlock);
1223                 }
1224                 db = dn->dn_bonus;
1225                 LASSERT(db);
1226                 LASSERT(dn->dn_handle);
1227                 DB_DNODE_ENTER(db);
1228                 if (refcount_add(&db->db_holds, osd_obj_tag) == 1) {
1229                         refcount_add(&dn->dn_holds, tag);
1230                         atomic_inc_32(&dn->dn_dbufs_count);
1231                 }
1232                 *dnp = dn;
1233                 dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH);
1234                 break;
1235         }
1236
1237         if (unlikely(*dnp == NULL))
1238                 rc = __osd_obj2dnode(tx->tx_objset, oid, dnp);
1239
1240         return rc;
1241 }
1242
1243 /*
1244  * The transaction passed to this routine must have
1245  * dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT) called and then assigned
1246  * to a transaction group.
1247  */
1248 int __osd_object_create(const struct lu_env *env, struct osd_object *obj,
1249                         dnode_t **dnp, dmu_tx_t *tx, struct lu_attr *la)
1250 {
1251         struct osd_device   *osd = osd_obj2dev(obj);
1252         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
1253         dmu_object_type_t    type = DMU_OT_PLAIN_FILE_CONTENTS;
1254         uint64_t oid;
1255
1256         /* Use DMU_OTN_UINT8_METADATA for local objects so their data blocks
1257          * would get an additional ditto copy */
1258         if (unlikely(S_ISREG(la->la_mode) &&
1259                      fid_seq_is_local_file(fid_seq(fid))))
1260                 type = DMU_OTN_UINT8_METADATA;
1261
1262         /* Create a new DMU object using the default dnode size. */
1263         oid = osd_dmu_object_alloc(osd->od_os, type, 0, 0, tx);
1264
1265         LASSERT(la->la_valid & LA_MODE);
1266         la->la_size = 0;
1267         la->la_nlink = 1;
1268
1269         return osd_find_new_dnode(env, tx, oid, dnp);
1270 }
1271
1272 /*
1273  * The transaction passed to this routine must have
1274  * dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, ...) called and then assigned
1275  * to a transaction group.
1276  *
1277  * Using ZAP_FLAG_HASH64 will force the ZAP to always be a FAT ZAP.
1278  * This is fine for directories today, because storing the FID in the dirent
1279  * will also require a FAT ZAP.  If there is a new type of micro ZAP created
1280  * then we might need to re-evaluate the use of this flag and instead do
1281  * a conversion from the different internal ZAP hash formats being used. */
1282 int __osd_zap_create(const struct lu_env *env, struct osd_device *osd,
1283                      dnode_t **dnp, dmu_tx_t *tx, struct lu_attr *la,
1284                      zap_flags_t flags)
1285 {
1286         uint64_t oid;
1287
1288         /* Assert that the transaction has been assigned to a
1289            transaction group. */
1290         LASSERT(tx->tx_txg != 0);
1291         *dnp = NULL;
1292
1293         oid = osd_zap_create_flags(osd->od_os, 0, flags | ZAP_FLAG_HASH64,
1294                                    DMU_OT_DIRECTORY_CONTENTS,
1295                                    14, /* == ZFS fzap_default_blockshift */
1296                                    DN_MAX_INDBLKSHIFT, /* indirect blockshift */
1297                                    0, tx);
1298
1299         la->la_size = 2;
1300         la->la_nlink = 1;
1301
1302         return osd_find_new_dnode(env, tx, oid, dnp);
1303 }
1304
1305 static dnode_t *osd_mkidx(const struct lu_env *env, struct osd_object *obj,
1306                           struct lu_attr *la, struct osd_thandle *oh)
1307 {
1308         dnode_t *dn;
1309         int rc;
1310
1311         /* Index file should be created as regular file in order not to confuse
1312          * ZPL which could interpret them as directory.
1313          * We set ZAP_FLAG_UINT64_KEY to let ZFS know than we are going to use
1314          * binary keys */
1315         LASSERT(S_ISREG(la->la_mode));
1316         rc = __osd_zap_create(env, osd_obj2dev(obj), &dn, oh->ot_tx, la,
1317                               ZAP_FLAG_UINT64_KEY);
1318         if (rc)
1319                 return ERR_PTR(rc);
1320         return dn;
1321 }
1322
1323 static dnode_t *osd_mkdir(const struct lu_env *env, struct osd_object *obj,
1324                           struct lu_attr *la, struct osd_thandle *oh)
1325 {
1326         dnode_t *dn;
1327         int rc;
1328
1329         LASSERT(S_ISDIR(la->la_mode));
1330         rc = __osd_zap_create(env, osd_obj2dev(obj), &dn, oh->ot_tx, la, 0);
1331         if (rc)
1332                 return ERR_PTR(rc);
1333         return dn;
1334 }
1335
1336 static dnode_t *osd_mkreg(const struct lu_env *env, struct osd_object *obj,
1337                           struct lu_attr *la, struct osd_thandle *oh)
1338 {
1339         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
1340         struct osd_device *osd = osd_obj2dev(obj);
1341         dnode_t *dn;
1342         int rc;
1343
1344         LASSERT(S_ISREG(la->la_mode));
1345         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1346         if (rc)
1347                 return ERR_PTR(rc);
1348
1349         if ((fid_is_idif(fid) || fid_is_norm(fid) || fid_is_echo(fid)) &&
1350             osd->od_is_ost) {
1351                 /* The minimum block size must be at least page size otherwise
1352                  * it will break the assumption in tgt_thread_big_cache where
1353                  * the array size is PTLRPC_MAX_BRW_PAGES. It will also affect
1354                  * RDMA due to subpage transfer size */
1355                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
1356                                                PAGE_SIZE, 0, oh->ot_tx);
1357                 if (unlikely(rc)) {
1358                         CERROR("%s: can't change blocksize: %d\n",
1359                                osd->od_svname, rc);
1360                         return ERR_PTR(rc);
1361                 }
1362         }
1363
1364         return dn;
1365 }
1366
1367 static dnode_t *osd_mksym(const struct lu_env *env, struct osd_object *obj,
1368                           struct lu_attr *la, struct osd_thandle *oh)
1369 {
1370         dnode_t *dn;
1371         int rc;
1372
1373         LASSERT(S_ISLNK(la->la_mode));
1374         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1375         if (rc)
1376                 return ERR_PTR(rc);
1377         return dn;
1378 }
1379
1380 static dnode_t *osd_mknod(const struct lu_env *env, struct osd_object *obj,
1381                           struct lu_attr *la, struct osd_thandle *oh)
1382 {
1383         dnode_t *dn;
1384         int rc;
1385
1386         if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode))
1387                 la->la_valid |= LA_RDEV;
1388
1389         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1390         if (rc)
1391                 return ERR_PTR(rc);
1392         return dn;
1393 }
1394
1395 typedef dnode_t *(*osd_obj_type_f)(const struct lu_env *env,
1396                                    struct osd_object *obj,
1397                                    struct lu_attr *la,
1398                                    struct osd_thandle *oh);
1399
1400 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
1401 {
1402         osd_obj_type_f result;
1403
1404         switch (type) {
1405         case DFT_DIR:
1406                 result = osd_mkdir;
1407                 break;
1408         case DFT_INDEX:
1409                 result = osd_mkidx;
1410                 break;
1411         case DFT_REGULAR:
1412                 result = osd_mkreg;
1413                 break;
1414         case DFT_SYM:
1415                 result = osd_mksym;
1416                 break;
1417         case DFT_NODE:
1418                 result = osd_mknod;
1419                 break;
1420         default:
1421                 LBUG();
1422                 break;
1423         }
1424         return result;
1425 }
1426
1427 /*
1428  * Concurrency: @dt is write locked.
1429  */
1430 static int osd_create(const struct lu_env *env, struct dt_object *dt,
1431                       struct lu_attr *attr, struct dt_allocation_hint *hint,
1432                       struct dt_object_format *dof, struct thandle *th)
1433 {
1434         struct osd_thread_info  *info = osd_oti_get(env);
1435         struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
1436         struct zpl_direntry     *zde = &info->oti_zde.lzd_reg;
1437         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
1438         struct osd_object       *obj = osd_dt_obj(dt);
1439         struct osd_device       *osd = osd_obj2dev(obj);
1440         char                    *buf = info->oti_str;
1441         struct osd_thandle      *oh;
1442         dnode_t *dn = NULL, *zdn = NULL;
1443         uint64_t                 zapid, parent = 0;
1444         int                      rc;
1445
1446         ENTRY;
1447
1448         /* concurrent create declarations should not see
1449          * the object inconsistent (db, attr, etc).
1450          * in regular cases acquisition should be cheap */
1451         down_write(&obj->oo_guard);
1452
1453         if (unlikely(dt_object_exists(dt)))
1454                 GOTO(out, rc = -EEXIST);
1455
1456         LASSERT(osd_invariant(obj));
1457         LASSERT(dof != NULL);
1458
1459         LASSERT(th != NULL);
1460         oh = container_of0(th, struct osd_thandle, ot_super);
1461
1462         LASSERT(obj->oo_dn == NULL);
1463
1464         /* to follow ZFS on-disk format we need
1465          * to initialize parent dnode properly */
1466         if (hint != NULL && hint->dah_parent != NULL &&
1467             !dt_object_remote(hint->dah_parent))
1468                 parent = osd_dt_obj(hint->dah_parent)->oo_dn->dn_object;
1469
1470         /* we may fix some attributes, better do not change the source */
1471         obj->oo_attr = *attr;
1472         obj->oo_attr.la_valid |= LA_SIZE | LA_NLINK | LA_TYPE;
1473
1474         dn = osd_create_type_f(dof->dof_type)(env, obj, &obj->oo_attr, oh);
1475         if (IS_ERR(dn)) {
1476                 rc = PTR_ERR(dn);
1477                 dn = NULL;
1478                 GOTO(out, rc);
1479         }
1480
1481         zde->zde_pad = 0;
1482         zde->zde_dnode = dn->dn_object;
1483         zde->zde_type = IFTODT(attr->la_mode & S_IFMT);
1484
1485         zapid = osd_get_name_n_idx(env, osd, fid, buf,
1486                                    sizeof(info->oti_str), &zdn);
1487         rc = osd_zap_add(osd, zapid, zdn, buf, 8, 1, zde, oh->ot_tx);
1488         if (rc)
1489                 GOTO(out, rc);
1490         obj->oo_dn = dn;
1491         /* Now add in all of the "SA" attributes */
1492         rc = osd_sa_handle_get(obj);
1493         if (rc)
1494                 GOTO(out, rc);
1495
1496         /* configure new osd object */
1497         parent = parent != 0 ? parent : zapid;
1498         rc = __osd_attr_init(env, osd, obj->oo_sa_hdl, oh->ot_tx,
1499                              &obj->oo_attr, parent);
1500         if (rc)
1501                 GOTO(out, rc);
1502
1503         /* XXX: oo_lma_flags */
1504         obj->oo_dt.do_lu.lo_header->loh_attr |= obj->oo_attr.la_mode & S_IFMT;
1505         if (likely(!fid_is_acct(lu_object_fid(&obj->oo_dt.do_lu))))
1506                 /* no body operations for accounting objects */
1507                 obj->oo_dt.do_body_ops = &osd_body_ops;
1508
1509         rc = -nvlist_alloc(&obj->oo_sa_xattr, NV_UNIQUE_NAME, KM_SLEEP);
1510         if (rc)
1511                 GOTO(out, rc);
1512
1513         /* initialize LMA */
1514         lustre_lma_init(lma, lu_object_fid(&obj->oo_dt.do_lu), 0, 0);
1515         lustre_lma_swab(lma);
1516         rc = -nvlist_add_byte_array(obj->oo_sa_xattr, XATTR_NAME_LMA,
1517                                     (uchar_t *)lma, sizeof(*lma));
1518         if (rc)
1519                 GOTO(out, rc);
1520         rc = __osd_sa_xattr_update(env, obj, oh);
1521         if (rc)
1522                 GOTO(out, rc);
1523         osd_idc_find_and_init(env, osd, obj);
1524
1525 out:
1526         if (unlikely(rc && dn)) {
1527                 dmu_object_free(osd->od_os, dn->dn_object, oh->ot_tx);
1528                 osd_dnode_rele(dn);
1529                 obj->oo_dn = NULL;
1530         } else if (!rc) {
1531                 obj->oo_dt.do_lu.lo_header->loh_attr |= LOHA_EXISTS;
1532         }
1533         up_write(&obj->oo_guard);
1534         RETURN(rc);
1535 }
1536
1537 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
1538                                struct thandle *th)
1539 {
1540         return osd_declare_attr_set(env, dt, NULL, th);
1541 }
1542
1543 /*
1544  * Concurrency: @dt is write locked.
1545  */
1546 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
1547                        struct thandle *handle)
1548 {
1549         struct osd_object       *obj = osd_dt_obj(dt);
1550         struct osd_thandle      *oh;
1551         struct osd_device       *osd = osd_obj2dev(obj);
1552         uint64_t                 nlink;
1553         int rc;
1554
1555         ENTRY;
1556
1557         down_read(&obj->oo_guard);
1558         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
1559                 GOTO(out, rc = -ENOENT);
1560
1561         LASSERT(osd_invariant(obj));
1562         LASSERT(obj->oo_sa_hdl != NULL);
1563
1564         oh = container_of0(handle, struct osd_thandle, ot_super);
1565
1566         write_lock(&obj->oo_attr_lock);
1567         nlink = ++obj->oo_attr.la_nlink;
1568         write_unlock(&obj->oo_attr_lock);
1569
1570         rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
1571
1572 out:
1573         up_read(&obj->oo_guard);
1574         RETURN(rc);
1575 }
1576
1577 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
1578                                struct thandle *handle)
1579 {
1580         return osd_declare_attr_set(env, dt, NULL, handle);
1581 }
1582
1583 /*
1584  * Concurrency: @dt is write locked.
1585  */
1586 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
1587                        struct thandle *handle)
1588 {
1589         struct osd_object       *obj = osd_dt_obj(dt);
1590         struct osd_thandle      *oh;
1591         struct osd_device       *osd = osd_obj2dev(obj);
1592         uint64_t                 nlink;
1593         int                      rc;
1594
1595         ENTRY;
1596
1597         down_read(&obj->oo_guard);
1598
1599         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
1600                 GOTO(out, rc = -ENOENT);
1601
1602         LASSERT(osd_invariant(obj));
1603         LASSERT(obj->oo_sa_hdl != NULL);
1604
1605         oh = container_of0(handle, struct osd_thandle, ot_super);
1606         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
1607
1608         write_lock(&obj->oo_attr_lock);
1609         nlink = --obj->oo_attr.la_nlink;
1610         write_unlock(&obj->oo_attr_lock);
1611
1612         rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
1613
1614 out:
1615         up_read(&obj->oo_guard);
1616         RETURN(rc);
1617 }
1618
1619 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
1620                            __u64 start, __u64 end)
1621 {
1622         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1623         ENTRY;
1624
1625         /* XXX: no other option than syncing the whole filesystem until we
1626          * support ZIL.  If the object tracked the txg that it was last
1627          * modified in, it could pass that txg here instead of "0".  Maybe
1628          * the changes are already committed, so no wait is needed at all? */
1629         if (!osd->od_dt_dev.dd_rdonly)
1630                 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
1631
1632         RETURN(0);
1633 }
1634
1635 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
1636 {
1637         return 0;
1638 }
1639
1640 static struct dt_object_operations osd_obj_ops = {
1641         .do_read_lock           = osd_read_lock,
1642         .do_write_lock          = osd_write_lock,
1643         .do_read_unlock         = osd_read_unlock,
1644         .do_write_unlock        = osd_write_unlock,
1645         .do_write_locked        = osd_write_locked,
1646         .do_attr_get            = osd_attr_get,
1647         .do_declare_attr_set    = osd_declare_attr_set,
1648         .do_attr_set            = osd_attr_set,
1649         .do_ah_init             = osd_ah_init,
1650         .do_declare_create      = osd_declare_create,
1651         .do_create              = osd_create,
1652         .do_declare_destroy     = osd_declare_destroy,
1653         .do_destroy             = osd_destroy,
1654         .do_index_try           = osd_index_try,
1655         .do_declare_ref_add     = osd_declare_ref_add,
1656         .do_ref_add             = osd_ref_add,
1657         .do_declare_ref_del     = osd_declare_ref_del,
1658         .do_ref_del             = osd_ref_del,
1659         .do_xattr_get           = osd_xattr_get,
1660         .do_declare_xattr_set   = osd_declare_xattr_set,
1661         .do_xattr_set           = osd_xattr_set,
1662         .do_declare_xattr_del   = osd_declare_xattr_del,
1663         .do_xattr_del           = osd_xattr_del,
1664         .do_xattr_list          = osd_xattr_list,
1665         .do_object_sync         = osd_object_sync,
1666         .do_invalidate          = osd_invalidate,
1667 };
1668
1669 static struct lu_object_operations osd_lu_obj_ops = {
1670         .loo_object_init        = osd_object_init,
1671         .loo_object_delete      = osd_object_delete,
1672         .loo_object_release     = osd_object_release,
1673         .loo_object_free        = osd_object_free,
1674         .loo_object_print       = osd_object_print,
1675         .loo_object_invariant   = osd_object_invariant,
1676 };
1677
1678 static int osd_otable_it_attr_get(const struct lu_env *env,
1679                                 struct dt_object *dt,
1680                                 struct lu_attr *attr)
1681 {
1682         attr->la_valid = 0;
1683         return 0;
1684 }
1685
1686 static struct dt_object_operations osd_obj_otable_it_ops = {
1687         .do_attr_get            = osd_otable_it_attr_get,
1688         .do_index_try           = osd_index_try,
1689 };