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 = -sa_handle_get(o->od_os, obj->oo_dn->dn_object, obj,
  92                             SA_HDL_PRIVATE, &obj->oo_sa_hdl);
  93         if (rc)
  94                 return rc;
  95
  96         /* Cache the xattr object id, valid for the life of the object */
  97         rc = -sa_lookup(obj->oo_sa_hdl, SA_ZPL_XATTR(o), &obj->oo_xattr, 8);
  98         if (rc == -ENOENT) {
  99                 obj->oo_xattr = ZFS_NO_OBJECT;
 100                 rc = 0;
 101         } else if (rc) {
 102                 osd_object_sa_fini(obj);
 103         }
 104
 105         return rc;
 106 }
 107
 108 /*
 109  * Add object to list of dirty objects in tx handle.
 110  */
 111 static void
 112 osd_object_sa_dirty_add(struct osd_object *obj, struct osd_thandle *oh)
 113 {
 114         if (!list_empty(&obj->oo_sa_linkage))
 115                 return;
 116
 117         down(&oh->ot_sa_lock);
 118         write_lock(&obj->oo_attr_lock);
 119         if (likely(list_empty(&obj->oo_sa_linkage)))
 120                 list_add(&obj->oo_sa_linkage, &oh->ot_sa_list);
 121         write_unlock(&obj->oo_attr_lock);
 122         up(&oh->ot_sa_lock);
 123 }
 124
 125 /*
 126  * Release spill block dbuf hold for all dirty SAs.
 127  */
 128 void osd_object_sa_dirty_rele(struct osd_thandle *oh)
 129 {
 130         struct osd_object *obj;
 131
 132         down(&oh->ot_sa_lock);
 133         while (!list_empty(&oh->ot_sa_list)) {
 134                 obj = list_entry(oh->ot_sa_list.next,
 135                                  struct osd_object, oo_sa_linkage);
 136                 sa_spill_rele(obj->oo_sa_hdl);
 137                 write_lock(&obj->oo_attr_lock);
 138                 list_del_init(&obj->oo_sa_linkage);
 139                 write_unlock(&obj->oo_attr_lock);
 140         }
 141         up(&oh->ot_sa_lock);
 142 }
 143
 144 /*
 145  * Update the SA and add the object to the dirty list.
 146  */
 147 int osd_object_sa_update(struct osd_object *obj, sa_attr_type_t type,
 148                          void *buf, uint32_t buflen, struct osd_thandle *oh)
 149 {
 150         int rc;
 151
 152         LASSERT(obj->oo_sa_hdl != NULL);
 153         LASSERT(oh->ot_tx != NULL);
 154
 155         rc = -sa_update(obj->oo_sa_hdl, type, buf, buflen, oh->ot_tx);
 156         osd_object_sa_dirty_add(obj, oh);
 157
 158         return rc;
 159 }
 160
 161 /*
 162  * Bulk update the SA and add the object to the dirty list.
 163  */
 164 static int
 165 osd_object_sa_bulk_update(struct osd_object *obj, sa_bulk_attr_t *attrs,
 166                           int count, struct osd_thandle *oh)
 167 {
 168         int rc;
 169
 170         LASSERT(obj->oo_sa_hdl != NULL);
 171         LASSERT(oh->ot_tx != NULL);
 172
 173         rc = -sa_bulk_update(obj->oo_sa_hdl, attrs, count, oh->ot_tx);
 174         osd_object_sa_dirty_add(obj, oh);
 175
 176         return rc;
 177 }
 178
 179 /*
 180  * Retrieve the attributes of a DMU object
 181  */
 182 int __osd_object_attr_get(const struct lu_env *env, struct osd_device *o,
 183                           struct osd_object *obj, struct lu_attr *la)
 184 {
 185         struct osa_attr *osa = &osd_oti_get(env)->oti_osa;
 186         sa_bulk_attr_t  *bulk = osd_oti_get(env)->oti_attr_bulk;
 187         sa_handle_t     *sa_hdl;
 188         int              cnt = 0;
 189         int              rc;
 190         ENTRY;
 191
 192         LASSERT(obj->oo_dn != NULL);
 193
 194         rc = -sa_handle_get(o->od_os, obj->oo_dn->dn_object, NULL,
 195                             SA_HDL_PRIVATE, &sa_hdl);
 196         if (rc)
 197                 RETURN(rc);
 198
 199         la->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE | LA_TYPE |
 200                         LA_SIZE | LA_UID | LA_GID | LA_FLAGS | LA_NLINK;
 201
 202         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(o), NULL, osa->atime, 16);
 203         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(o), NULL, osa->mtime, 16);
 204         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(o), NULL, osa->ctime, 16);
 205         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(o), NULL, &osa->mode, 8);
 206         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(o), NULL, &osa->size, 8);
 207         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(o), NULL, &osa->nlink, 8);
 208         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(o), NULL, &osa->uid, 8);
 209         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(o), NULL, &osa->gid, 8);
 210         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(o), NULL, &osa->flags, 8);
 211         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
 212
 213         rc = -sa_bulk_lookup(sa_hdl, bulk, cnt);
 214         if (rc)
 215                 GOTO(out_sa, rc);
 216
 217         la->la_atime = osa->atime[0];
 218         la->la_mtime = osa->mtime[0];
 219         la->la_ctime = osa->ctime[0];
 220         la->la_mode = osa->mode;
 221         la->la_uid = osa->uid;
 222         la->la_gid = osa->gid;
 223         la->la_nlink = osa->nlink;
 224         la->la_flags = attrs_zfs2fs(osa->flags);
 225         la->la_size = osa->size;
 226
 227         /* Try to get extra flag from LMA. Right now, only LMAI_ORPHAN
 228          * flags is stored in LMA, and it is only for orphan directory */
 229         if (S_ISDIR(la->la_mode) && dt_object_exists(&obj->oo_dt)) {
 230                 struct osd_thread_info *info = osd_oti_get(env);
 231                 struct lustre_mdt_attrs *lma;
 232                 struct lu_buf buf;
 233
 234                 lma = (struct lustre_mdt_attrs *)info->oti_buf;
 235                 buf.lb_buf = lma;
 236                 buf.lb_len = sizeof(info->oti_buf);
 237                 rc = osd_xattr_get(env, &obj->oo_dt, &buf, XATTR_NAME_LMA);
 238                 if (rc > 0) {
 239                         rc = 0;
 240                         lma->lma_incompat = le32_to_cpu(lma->lma_incompat);
 241                         obj->oo_lma_flags =
 242                                 lma_to_lustre_flags(lma->lma_incompat);
 243
 244                 } else if (rc == -ENODATA) {
 245                         rc = 0;
 246                 }
 247         }
 248
 249         if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode)) {
 250                 rc = -sa_lookup(sa_hdl, SA_ZPL_RDEV(o), &osa->rdev, 8);
 251                 if (rc)
 252                         GOTO(out_sa, rc);
 253                 la->la_rdev = osa->rdev;
 254                 la->la_valid |= LA_RDEV;
 255         }
 256 out_sa:
 257         sa_handle_destroy(sa_hdl);
 258
 259         RETURN(rc);
 260 }
 261
 262 int __osd_obj2dnode(const struct lu_env *env, objset_t *os,
 263                    uint64_t oid, dnode_t **dnp)
 264 {
 265         dmu_object_info_t *doi = &osd_oti_get(env)->oti_doi;
 266         dmu_buf_t *db;
 267         dmu_buf_impl_t *dbi;
 268         int rc;
 269
 270         rc = dmu_bonus_hold(os, oid, osd_obj_tag, &db);
 271         if (rc)
 272                 return rc;
 273
 274         dbi = (dmu_buf_impl_t *)db;
 275         DB_DNODE_ENTER(dbi);
 276         *dnp = DB_DNODE(dbi);
 277
 278         LASSERT(*dnp != NULL);
 279         dmu_object_info_from_dnode(*dnp, doi);
 280         if (unlikely (oid != DMU_USERUSED_OBJECT &&
 281             oid != DMU_GROUPUSED_OBJECT && doi->doi_bonus_type != DMU_OT_SA)) {
 282                 osd_dnode_rele(*dnp);
 283                 *dnp = NULL;
 284                 return -EINVAL;
 285         }
 286
 287         return 0;
 288 }
 289
 290 /*
 291  * Concurrency: no concurrent access is possible that early in object
 292  * life-cycle.
 293  */
 294 struct lu_object *osd_object_alloc(const struct lu_env *env,
 295                                    const struct lu_object_header *hdr,
 296                                    struct lu_device *d)
 297 {
 298         struct osd_object *mo;
 299
 300         OBD_SLAB_ALLOC_PTR_GFP(mo, osd_object_kmem, GFP_NOFS);
 301         if (mo != NULL) {
 302                 struct lu_object *l;
 303
 304                 l = &mo->oo_dt.do_lu;
 305                 dt_object_init(&mo->oo_dt, NULL, d);
 306                 mo->oo_dt.do_ops = &osd_obj_ops;
 307                 l->lo_ops = &osd_lu_obj_ops;
 308                 INIT_LIST_HEAD(&mo->oo_sa_linkage);
 309                 INIT_LIST_HEAD(&mo->oo_unlinked_linkage);
 310                 init_rwsem(&mo->oo_sem);
 311                 init_rwsem(&mo->oo_guard);
 312                 rwlock_init(&mo->oo_attr_lock);
 313                 mo->oo_destroy = OSD_DESTROY_NONE;
 314                 return l;
 315         } else {
 316                 return NULL;
 317         }
 318 }
 319
 320 /*
 321  * Concurrency: shouldn't matter.
 322  */
 323 int osd_object_init0(const struct lu_env *env, struct osd_object *obj)
 324 {
 325         struct osd_device       *osd = osd_obj2dev(obj);
 326         const struct lu_fid     *fid = lu_object_fid(&obj->oo_dt.do_lu);
 327         int                      rc = 0;
 328         ENTRY;
 329
 330         if (obj->oo_dn == NULL)
 331                 RETURN(0);
 332
 333         /* object exist */
 334
 335         rc = osd_object_sa_init(obj, osd);
 336         if (rc)
 337                 RETURN(rc);
 338
 339         /* cache attrs in object */
 340         rc = __osd_object_attr_get(env, osd, obj, &obj->oo_attr);
 341         if (rc)
 342                 RETURN(rc);
 343
 344         if (likely(!fid_is_acct(fid)))
 345                 /* no body operations for accounting objects */
 346                 obj->oo_dt.do_body_ops = &osd_body_ops;
 347
 348         /*
 349          * initialize object before marking it existing
 350          */
 351         obj->oo_dt.do_lu.lo_header->loh_attr |= obj->oo_attr.la_mode & S_IFMT;
 352
 353         smp_mb();
 354         obj->oo_dt.do_lu.lo_header->loh_attr |= LOHA_EXISTS;
 355
 356         RETURN(0);
 357 }
 358
 359 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
 360 {
 361         struct osd_thread_info  *info = osd_oti_get(env);
 362         struct lu_buf           buf;
 363         int                     rc;
 364         struct lustre_mdt_attrs *lma;
 365         ENTRY;
 366
 367         CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
 368         lma = (struct lustre_mdt_attrs *)info->oti_buf;
 369         buf.lb_buf = lma;
 370         buf.lb_len = sizeof(info->oti_buf);
 371
 372         rc = osd_xattr_get(env, &obj->oo_dt, &buf, XATTR_NAME_LMA);
 373         if (rc > 0) {
 374                 rc = 0;
 375                 lustre_lma_swab(lma);
 376                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
 377                              CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
 378                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
 379                               "fid = "DFID"\n", osd_obj2dev(obj)->od_svname,
 380                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
 381                               PFID(lu_object_fid(&obj->oo_dt.do_lu)));
 382                         rc = -EOPNOTSUPP;
 383                 }
 384         } else if (rc == -ENODATA) {
 385                 /* haven't initialize LMA xattr */
 386                 rc = 0;
 387         }
 388
 389         RETURN(rc);
 390 }
 391
 392 /*
 393  * Concurrency: no concurrent access is possible that early in object
 394  * life-cycle.
 395  */
 396 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
 397                            const struct lu_object_conf *conf)
 398 {
 399         struct osd_object       *obj = osd_obj(l);
 400         struct osd_device       *osd = osd_obj2dev(obj);
 401         uint64_t                 oid;
 402         int                      rc;
 403         ENTRY;
 404
 405         LASSERT(osd_invariant(obj));
 406
 407         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
 408                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
 409                 l->lo_header->loh_attr |= LOHA_EXISTS;
 410                 RETURN(0);
 411         }
 412
 413         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 414                 GOTO(out, rc = 0);
 415
 416         rc = osd_fid_lookup(env, osd, lu_object_fid(l), &oid);
 417         if (rc == 0) {
 418                 LASSERT(obj->oo_dn == NULL);
 419                 rc = __osd_obj2dnode(env, osd->od_os, oid, &obj->oo_dn);
 420                 /* EEXIST will be returned if object is being deleted in ZFS */
 421                 if (rc == -EEXIST) {
 422                         rc = 0;
 423                         GOTO(out, rc);
 424                 }
 425                 if (rc != 0) {
 426                         CERROR("%s: lookup "DFID"/%#llx failed: rc = %d\n",
 427                                osd->od_svname, PFID(lu_object_fid(l)), oid, rc);
 428                         GOTO(out, rc);
 429                 }
 430                 LASSERT(obj->oo_dn);
 431                 rc = osd_object_init0(env, obj);
 432                 if (rc != 0)
 433                         GOTO(out, rc);
 434
 435                 rc = osd_check_lma(env, obj);
 436                 if (rc != 0)
 437                         GOTO(out, rc);
 438         } else if (rc == -ENOENT) {
 439                 rc = 0;
 440         }
 441         LASSERT(osd_invariant(obj));
 442 out:
 443         RETURN(rc);
 444 }
 445
 446 /*
 447  * Concurrency: no concurrent access is possible that late in object
 448  * life-cycle.
 449  */
 450 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
 451 {
 452         struct osd_object *obj = osd_obj(l);
 453
 454         LASSERT(osd_invariant(obj));
 455
 456         dt_object_fini(&obj->oo_dt);
 457         OBD_SLAB_FREE_PTR(obj, osd_object_kmem);
 458 }
 459
 460 static int
 461 osd_object_unlinked_add(struct osd_object *obj, struct osd_thandle *oh)
 462 {
 463         int rc = -EBUSY;
 464
 465         LASSERT(obj->oo_destroy == OSD_DESTROY_ASYNC);
 466
 467         /* the object is supposed to be exclusively locked by
 468          * the caller (osd_object_destroy()), while the transaction
 469          * (oh) is per-thread and not shared */
 470         if (likely(list_empty(&obj->oo_unlinked_linkage))) {
 471                 list_add(&obj->oo_unlinked_linkage, &oh->ot_unlinked_list);
 472                 rc = 0;
 473         }
 474
 475         return rc;
 476 }
 477
 478 /* Default to max data size covered by a level-1 indirect block */
 479 static unsigned long osd_sync_destroy_max_size =
 480         1UL << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT + SPA_MAXBLOCKSHIFT);
 481 module_param(osd_sync_destroy_max_size, ulong, 0444);
 482 MODULE_PARM_DESC(osd_sync_destroy_max_size, "Maximum object size to use synchronous destroy.");
 483
 484 static inline void
 485 osd_object_set_destroy_type(struct osd_object *obj)
 486 {
 487         /*
 488          * Lock-less OST_WRITE can race with OST_DESTROY, so set destroy type
 489          * only once and use it consistently thereafter.
 490          */
 491         down_write(&obj->oo_guard);
 492         if (obj->oo_destroy == OSD_DESTROY_NONE) {
 493                 if (obj->oo_attr.la_size <= osd_sync_destroy_max_size)
 494                         obj->oo_destroy = OSD_DESTROY_SYNC;
 495                 else /* Larger objects are destroyed asynchronously */
 496                         obj->oo_destroy = OSD_DESTROY_ASYNC;
 497         }
 498         up_write(&obj->oo_guard);
 499 }
 500
 501 static int osd_declare_object_destroy(const struct lu_env *env,
 502                                       struct dt_object *dt,
 503                                       struct thandle *th)
 504 {
 505         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
 506         struct osd_object       *obj = osd_dt_obj(dt);
 507         struct osd_device       *osd = osd_obj2dev(obj);
 508         struct osd_thandle      *oh;
 509         int                      rc;
 510         uint64_t                 zapid;
 511         ENTRY;
 512
 513         LASSERT(th != NULL);
 514         LASSERT(dt_object_exists(dt));
 515
 516         oh = container_of0(th, struct osd_thandle, ot_super);
 517         LASSERT(oh->ot_tx != NULL);
 518
 519         /* declare that we'll remove object from fid-dnode mapping */
 520         zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0);
 521         dmu_tx_hold_zap(oh->ot_tx, zapid, FALSE, NULL);
 522
 523         osd_declare_xattrs_destroy(env, obj, oh);
 524
 525         /* one less inode */
 526         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 527                                obj->oo_attr.la_gid, -1, oh, false, NULL, false);
 528         if (rc)
 529                 RETURN(rc);
 530
 531         /* data to be truncated */
 532         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 533                                obj->oo_attr.la_gid, 0, oh, true, NULL, false);
 534         if (rc)
 535                 RETURN(rc);
 536
 537         osd_object_set_destroy_type(obj);
 538         if (obj->oo_destroy == OSD_DESTROY_SYNC)
 539                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object,
 540                                  0, DMU_OBJECT_END);
 541         else
 542                 dmu_tx_hold_zap(oh->ot_tx, osd->od_unlinkedid, TRUE, NULL);
 543
 544         /* will help to find FID->ino when this object is being
 545          * added to PENDING/ */
 546         osd_idc_find_and_init(env, osd, obj);
 547
 548         RETURN(0);
 549 }
 550
 551 static int osd_object_destroy(const struct lu_env *env,
 552                               struct dt_object *dt, struct thandle *th)
 553 {
 554         struct osd_thread_info  *info = osd_oti_get(env);
 555         char                    *buf = info->oti_str;
 556         struct osd_object       *obj = osd_dt_obj(dt);
 557         struct osd_device       *osd = osd_obj2dev(obj);
 558         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
 559         struct osd_thandle      *oh;
 560         int                      rc;
 561         uint64_t                 oid, zapid;
 562         ENTRY;
 563
 564         down_write(&obj->oo_guard);
 565
 566         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 567                 GOTO(out, rc = -ENOENT);
 568
 569         LASSERT(obj->oo_dn != NULL);
 570
 571         oh = container_of0(th, struct osd_thandle, ot_super);
 572         LASSERT(oh != NULL);
 573         LASSERT(oh->ot_tx != NULL);
 574
 575         /* remove obj ref from index dir (it depends) */
 576         zapid = osd_get_name_n_idx(env, osd, fid, buf, sizeof(info->oti_str));
 577         rc = -zap_remove(osd->od_os, zapid, buf, oh->ot_tx);
 578         if (rc) {
 579                 CERROR("%s: zap_remove(%s) failed: rc = %d\n",
 580                        osd->od_svname, buf, rc);
 581                 GOTO(out, rc);
 582         }
 583
 584         rc = osd_xattrs_destroy(env, obj, oh);
 585         if (rc) {
 586                 CERROR("%s: cannot destroy xattrs for %s: rc = %d\n",
 587                        osd->od_svname, buf, rc);
 588                 GOTO(out, rc);
 589         }
 590
 591         oid = obj->oo_dn->dn_object;
 592         if (unlikely(obj->oo_destroy == OSD_DESTROY_NONE)) {
 593                 /* this may happen if the destroy wasn't declared
 594                  * e.g. when the object is created and then destroyed
 595                  * in the same transaction - we don't need additional
 596                  * space for destroy specifically */
 597                 LASSERT(obj->oo_attr.la_size <= osd_sync_destroy_max_size);
 598                 rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
 599                 if (rc)
 600                         CERROR("%s: failed to free %s %llu: rc = %d\n",
 601                                osd->od_svname, buf, oid, rc);
 602         } else if (obj->oo_destroy == OSD_DESTROY_SYNC) {
 603                 rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
 604                 if (rc)
 605                         CERROR("%s: failed to free %s %llu: rc = %d\n",
 606                                osd->od_svname, buf, oid, rc);
 607         } else { /* asynchronous destroy */
 608                 rc = osd_object_unlinked_add(obj, oh);
 609                 if (rc)
 610                         GOTO(out, rc);
 611
 612                 rc = -zap_add_int(osd->od_os, osd->od_unlinkedid,
 613                                   oid, oh->ot_tx);
 614                 if (rc)
 615                         CERROR("%s: zap_add_int() failed %s %llu: rc = %d\n",
 616                                osd->od_svname, buf, oid, rc);
 617         }
 618
 619 out:
 620         /* not needed in the cache anymore */
 621         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
 622         if (rc == 0)
 623                 obj->oo_destroyed = 1;
 624         up_write(&obj->oo_guard);
 625         RETURN (0);
 626 }
 627
 628 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
 629 {
 630         struct osd_object *obj = osd_obj(l);
 631
 632         if (obj->oo_dn != NULL) {
 633                 osd_object_sa_fini(obj);
 634                 if (obj->oo_sa_xattr) {
 635                         nvlist_free(obj->oo_sa_xattr);
 636                         obj->oo_sa_xattr = NULL;
 637                 }
 638                 osd_dnode_rele(obj->oo_dn);
 639                 list_del(&obj->oo_sa_linkage);
 640                 obj->oo_dn = NULL;
 641         }
 642 }
 643
 644 /*
 645  * Concurrency: ->loo_object_release() is called under site spin-lock.
 646  */
 647 static void osd_object_release(const struct lu_env *env,
 648                                struct lu_object *l)
 649 {
 650 }
 651
 652 /*
 653  * Concurrency: shouldn't matter.
 654  */
 655 static int osd_object_print(const struct lu_env *env, void *cookie,
 656                             lu_printer_t p, const struct lu_object *l)
 657 {
 658         struct osd_object *o = osd_obj(l);
 659
 660         return (*p)(env, cookie, LUSTRE_OSD_ZFS_NAME"-object@%p", o);
 661 }
 662
 663 static void osd_object_read_lock(const struct lu_env *env,
 664                                  struct dt_object *dt, unsigned role)
 665 {
 666         struct osd_object *obj = osd_dt_obj(dt);
 667
 668         LASSERT(osd_invariant(obj));
 669
 670         down_read_nested(&obj->oo_sem, role);
 671 }
 672
 673 static void osd_object_write_lock(const struct lu_env *env,
 674                                   struct dt_object *dt, unsigned role)
 675 {
 676         struct osd_object *obj = osd_dt_obj(dt);
 677
 678         LASSERT(osd_invariant(obj));
 679
 680         down_write_nested(&obj->oo_sem, role);
 681 }
 682
 683 static void osd_object_read_unlock(const struct lu_env *env,
 684                                    struct dt_object *dt)
 685 {
 686         struct osd_object *obj = osd_dt_obj(dt);
 687
 688         LASSERT(osd_invariant(obj));
 689         up_read(&obj->oo_sem);
 690 }
 691
 692 static void osd_object_write_unlock(const struct lu_env *env,
 693                                     struct dt_object *dt)
 694 {
 695         struct osd_object *obj = osd_dt_obj(dt);
 696
 697         LASSERT(osd_invariant(obj));
 698         up_write(&obj->oo_sem);
 699 }
 700
 701 static int osd_object_write_locked(const struct lu_env *env,
 702                                    struct dt_object *dt)
 703 {
 704         struct osd_object *obj = osd_dt_obj(dt);
 705         int rc = 1;
 706
 707         LASSERT(osd_invariant(obj));
 708
 709         if (down_write_trylock(&obj->oo_sem)) {
 710                 rc = 0;
 711                 up_write(&obj->oo_sem);
 712         }
 713         return rc;
 714 }
 715
 716 static int osd_attr_get(const struct lu_env *env,
 717                         struct dt_object *dt,
 718                         struct lu_attr *attr)
 719 {
 720         struct osd_object       *obj = osd_dt_obj(dt);
 721         uint64_t                 blocks;
 722         uint32_t                 blksize;
 723         int                      rc = 0;
 724
 725         down_read(&obj->oo_guard);
 726
 727         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 728                 GOTO(out, rc = -ENOENT);
 729
 730         LASSERT(osd_invariant(obj));
 731         LASSERT(obj->oo_dn);
 732
 733         read_lock(&obj->oo_attr_lock);
 734         *attr = obj->oo_attr;
 735         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
 736                 attr->la_flags |= LUSTRE_ORPHAN_FL;
 737         read_unlock(&obj->oo_attr_lock);
 738
 739         /* with ZFS_DEBUG zrl_add_debug() called by DB_DNODE_ENTER()
 740          * from within sa_object_size() can block on a mutex, so
 741          * we can't call sa_object_size() holding rwlock */
 742         sa_object_size(obj->oo_sa_hdl, &blksize, &blocks);
 743         /* we do not control size of indices, so always calculate
 744          * it from number of blocks reported by DMU */
 745         if (S_ISDIR(attr->la_mode))
 746                 attr->la_size = 512 * blocks;
 747         /* Block size may be not set; suggest maximal I/O transfers. */
 748         if (blksize == 0)
 749                 blksize = osd_spa_maxblocksize(
 750                         dmu_objset_spa(osd_obj2dev(obj)->od_os));
 751
 752         attr->la_blksize = blksize;
 753         attr->la_blocks = blocks;
 754         attr->la_valid |= LA_BLOCKS | LA_BLKSIZE;
 755
 756 out:
 757         up_read(&obj->oo_guard);
 758         return rc;
 759 }
 760
 761 /* Simple wrapper on top of qsd API which implement quota transfer for osd
 762  * setattr needs. As a reminder, only the root user can change ownership of
 763  * a file, that's why EDQUOT & EINPROGRESS errors are discarded */
 764 static inline int qsd_transfer(const struct lu_env *env,
 765                                struct qsd_instance *qsd,
 766                                struct lquota_trans *trans, int qtype,
 767                                __u64 orig_id, __u64 new_id, __u64 bspace,
 768                                struct lquota_id_info *qi)
 769 {
 770         int     rc;
 771
 772         if (unlikely(qsd == NULL))
 773                 return 0;
 774
 775         LASSERT(qtype >= 0 && qtype < LL_MAXQUOTAS);
 776         qi->lqi_type = qtype;
 777
 778         /* inode accounting */
 779         qi->lqi_is_blk = false;
 780
 781         /* one more inode for the new owner ... */
 782         qi->lqi_id.qid_uid = new_id;
 783         qi->lqi_space      = 1;
 784         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 785         if (rc == -EDQUOT || rc == -EINPROGRESS)
 786                 rc = 0;
 787         if (rc)
 788                 return rc;
 789
 790         /* and one less inode for the current id */
 791         qi->lqi_id.qid_uid = orig_id;;
 792         qi->lqi_space      = -1;
 793         /* can't get EDQUOT when reducing usage */
 794         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 795         if (rc == -EINPROGRESS)
 796                 rc = 0;
 797         if (rc)
 798                 return rc;
 799
 800         /* block accounting */
 801         qi->lqi_is_blk = true;
 802
 803         /* more blocks for the new owner ... */
 804         qi->lqi_id.qid_uid = new_id;
 805         qi->lqi_space      = bspace;
 806         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 807         if (rc == -EDQUOT || rc == -EINPROGRESS)
 808                 rc = 0;
 809         if (rc)
 810                 return rc;
 811
 812         /* and finally less blocks for the current owner */
 813         qi->lqi_id.qid_uid = orig_id;
 814         qi->lqi_space      = -bspace;
 815         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 816         /* can't get EDQUOT when reducing usage */
 817         if (rc == -EINPROGRESS)
 818                 rc = 0;
 819         return rc;
 820 }
 821
 822 static int osd_declare_attr_set(const struct lu_env *env,
 823                                 struct dt_object *dt,
 824                                 const struct lu_attr *attr,
 825                                 struct thandle *handle)
 826 {
 827         struct osd_thread_info  *info = osd_oti_get(env);
 828         struct osd_object       *obj = osd_dt_obj(dt);
 829         struct osd_device       *osd = osd_obj2dev(obj);
 830         dmu_tx_hold_t           *txh;
 831         struct osd_thandle      *oh;
 832         uint64_t                 bspace;
 833         uint32_t                 blksize;
 834         int                      rc = 0;
 835         bool                     found;
 836         ENTRY;
 837
 838
 839         LASSERT(handle != NULL);
 840         LASSERT(osd_invariant(obj));
 841
 842         oh = container_of0(handle, struct osd_thandle, ot_super);
 843
 844         down_read(&obj->oo_guard);
 845         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 846                 GOTO(out, rc = 0);
 847
 848         LASSERT(obj->oo_sa_hdl != NULL);
 849         LASSERT(oh->ot_tx != NULL);
 850         /* regular attributes are part of the bonus buffer */
 851         /* let's check whether this object is already part of
 852          * transaction.. */
 853         found = false;
 854         for (txh = list_head(&oh->ot_tx->tx_holds); txh;
 855              txh = list_next(&oh->ot_tx->tx_holds, txh)) {
 856                 if (txh->txh_dnode == NULL)
 857                         continue;
 858                 if (txh->txh_dnode->dn_object != obj->oo_dn->dn_object)
 859                         continue;
 860                 /* this object is part of the transaction already
 861                  * we don't need to declare bonus again */
 862                 found = true;
 863                 break;
 864         }
 865         if (!found)
 866                 dmu_tx_hold_bonus(oh->ot_tx, obj->oo_dn->dn_object);
 867         if (oh->ot_tx->tx_err != 0)
 868                 GOTO(out, rc = -oh->ot_tx->tx_err);
 869
 870         if (attr && attr->la_valid & LA_FLAGS) {
 871                 /* LMA is usually a part of bonus, no need to declare
 872                  * anything else */
 873         }
 874
 875         if (attr && (attr->la_valid & (LA_UID | LA_GID))) {
 876                 sa_object_size(obj->oo_sa_hdl, &blksize, &bspace);
 877                 bspace = toqb(bspace * blksize);
 878         }
 879
 880         if (attr && attr->la_valid & LA_UID) {
 881                 /* quota enforcement for user */
 882                 if (attr->la_uid != obj->oo_attr.la_uid) {
 883                         rc = qsd_transfer(env, osd->od_quota_slave,
 884                                           &oh->ot_quota_trans, USRQUOTA,
 885                                           obj->oo_attr.la_uid, attr->la_uid,
 886                                           bspace, &info->oti_qi);
 887                         if (rc)
 888                                 GOTO(out, rc);
 889                 }
 890         }
 891         if (attr && attr->la_valid & LA_GID) {
 892                 /* quota enforcement for group */
 893                 if (attr->la_gid != obj->oo_attr.la_gid) {
 894                         rc = qsd_transfer(env, osd->od_quota_slave,
 895                                           &oh->ot_quota_trans, GRPQUOTA,
 896                                           obj->oo_attr.la_gid, attr->la_gid,
 897                                           bspace, &info->oti_qi);
 898                         if (rc)
 899                                 GOTO(out, rc);
 900                 }
 901         }
 902
 903 out:
 904         up_read(&obj->oo_guard);
 905         RETURN(rc);
 906 }
 907
 908 /*
 909  * Set the attributes of an object
 910  *
 911  * The transaction passed to this routine must have
 912  * dmu_tx_hold_bonus(tx, oid) called and then assigned
 913  * to a transaction group.
 914  */
 915 static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 916                         const struct lu_attr *la, struct thandle *handle)
 917 {
 918         struct osd_thread_info  *info = osd_oti_get(env);
 919         sa_bulk_attr_t          *bulk = osd_oti_get(env)->oti_attr_bulk;
 920         struct osd_object       *obj = osd_dt_obj(dt);
 921         struct osd_device       *osd = osd_obj2dev(obj);
 922         struct osd_thandle      *oh;
 923         struct osa_attr         *osa = &info->oti_osa;
 924         __u64                    valid = la->la_valid;
 925         int                      cnt;
 926         int                      rc = 0;
 927
 928         ENTRY;
 929
 930         down_read(&obj->oo_guard);
 931         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 932                 GOTO(out, rc = -ENOENT);
 933
 934         LASSERT(handle != NULL);
 935         LASSERT(osd_invariant(obj));
 936         LASSERT(obj->oo_sa_hdl);
 937
 938         oh = container_of0(handle, struct osd_thandle, ot_super);
 939         /* Assert that the transaction has been assigned to a
 940            transaction group. */
 941         LASSERT(oh->ot_tx->tx_txg != 0);
 942
 943         /* Only allow set size for regular file */
 944         if (!S_ISREG(dt->do_lu.lo_header->loh_attr))
 945                 valid &= ~(LA_SIZE | LA_BLOCKS);
 946
 947         if (valid & LA_CTIME && la->la_ctime == obj->oo_attr.la_ctime)
 948                 valid &= ~LA_CTIME;
 949
 950         if (valid & LA_MTIME && la->la_mtime == obj->oo_attr.la_mtime)
 951                 valid &= ~LA_MTIME;
 952
 953         if (valid & LA_ATIME && la->la_atime == obj->oo_attr.la_atime)
 954                 valid &= ~LA_ATIME;
 955
 956         if (valid == 0)
 957                 GOTO(out, rc = 0);
 958
 959         if (valid & LA_FLAGS) {
 960                 struct lustre_mdt_attrs *lma;
 961                 struct lu_buf buf;
 962
 963                 if (la->la_flags & LUSTRE_LMA_FL_MASKS) {
 964                         CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
 965                         lma = (struct lustre_mdt_attrs *)&info->oti_buf;
 966                         buf.lb_buf = lma;
 967                         buf.lb_len = sizeof(info->oti_buf);
 968                         rc = osd_xattr_get(env, &obj->oo_dt, &buf,
 969                                            XATTR_NAME_LMA);
 970                         if (rc > 0) {
 971                                 lma->lma_incompat =
 972                                         le32_to_cpu(lma->lma_incompat);
 973                                 lma->lma_incompat |=
 974                                         lustre_to_lma_flags(la->la_flags);
 975                                 lma->lma_incompat =
 976                                         cpu_to_le32(lma->lma_incompat);
 977                                 buf.lb_buf = lma;
 978                                 buf.lb_len = sizeof(*lma);
 979                                 rc = osd_xattr_set_internal(env, obj, &buf,
 980                                                             XATTR_NAME_LMA,
 981                                                             LU_XATTR_REPLACE,
 982                                                             oh);
 983                         }
 984                         if (rc < 0) {
 985                                 CWARN("%s: failed to set LMA flags: rc = %d\n",
 986                                        osd->od_svname, rc);
 987                                 RETURN(rc);
 988                         }
 989                 }
 990         }
 991
 992         write_lock(&obj->oo_attr_lock);
 993         cnt = 0;
 994         if (valid & LA_ATIME) {
 995                 osa->atime[0] = obj->oo_attr.la_atime = la->la_atime;
 996                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(osd), NULL,
 997                                  osa->atime, 16);
 998         }
 999         if (valid & LA_MTIME) {
1000                 osa->mtime[0] = obj->oo_attr.la_mtime = la->la_mtime;
1001                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(osd), NULL,
1002                                  osa->mtime, 16);
1003         }
1004         if (valid & LA_CTIME) {
1005                 osa->ctime[0] = obj->oo_attr.la_ctime = la->la_ctime;
1006                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(osd), NULL,
1007                                  osa->ctime, 16);
1008         }
1009         if (valid & LA_MODE) {
1010                 /* mode is stored along with type, so read it first */
1011                 obj->oo_attr.la_mode = (obj->oo_attr.la_mode & S_IFMT) |
1012                         (la->la_mode & ~S_IFMT);
1013                 osa->mode = obj->oo_attr.la_mode;
1014                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(osd), NULL,
1015                                  &osa->mode, 8);
1016         }
1017         if (valid & LA_SIZE) {
1018                 osa->size = obj->oo_attr.la_size = la->la_size;
1019                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(osd), NULL,
1020                                  &osa->size, 8);
1021         }
1022         if (valid & LA_NLINK) {
1023                 osa->nlink = obj->oo_attr.la_nlink = la->la_nlink;
1024                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(osd), NULL,
1025                                  &osa->nlink, 8);
1026         }
1027         if (valid & LA_RDEV) {
1028                 osa->rdev = obj->oo_attr.la_rdev = la->la_rdev;
1029                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_RDEV(osd), NULL,
1030                                  &osa->rdev, 8);
1031         }
1032         if (valid & LA_FLAGS) {
1033                 osa->flags = attrs_fs2zfs(la->la_flags);
1034                 /* many flags are not supported by zfs, so ensure a good cached
1035                  * copy */
1036                 obj->oo_attr.la_flags = attrs_zfs2fs(osa->flags);
1037                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL,
1038                                  &osa->flags, 8);
1039         }
1040         if (valid & LA_UID) {
1041                 osa->uid = obj->oo_attr.la_uid = la->la_uid;
1042                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(osd), NULL,
1043                                  &osa->uid, 8);
1044         }
1045         if (valid & LA_GID) {
1046                 osa->gid = obj->oo_attr.la_gid = la->la_gid;
1047                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(osd), NULL,
1048                                  &osa->gid, 8);
1049         }
1050         obj->oo_attr.la_valid |= valid;
1051         write_unlock(&obj->oo_attr_lock);
1052
1053         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
1054         rc = osd_object_sa_bulk_update(obj, bulk, cnt, oh);
1055
1056 out:
1057         up_read(&obj->oo_guard);
1058         RETURN(rc);
1059 }
1060
1061 /*
1062  * Object creation.
1063  *
1064  * XXX temporary solution.
1065  */
1066
1067 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
1068                         struct dt_object *parent, struct dt_object *child,
1069                         umode_t child_mode)
1070 {
1071         LASSERT(ah);
1072
1073         ah->dah_parent = parent;
1074         ah->dah_mode = child_mode;
1075
1076         if (parent != NULL && !dt_object_remote(parent)) {
1077                 /* will help to find FID->ino at dt_insert("..") */
1078                 struct osd_object *pobj = osd_dt_obj(parent);
1079
1080                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
1081         }
1082 }
1083
1084 static int osd_declare_object_create(const struct lu_env *env,
1085                                      struct dt_object *dt,
1086                                      struct lu_attr *attr,
1087                                      struct dt_allocation_hint *hint,
1088                                      struct dt_object_format *dof,
1089                                      struct thandle *handle)
1090 {
1091         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
1092         struct osd_object       *obj = osd_dt_obj(dt);
1093         struct osd_device       *osd = osd_obj2dev(obj);
1094         struct osd_thandle      *oh;
1095         uint64_t                 zapid;
1096         int                      rc, dnode_size;
1097         ENTRY;
1098
1099         LASSERT(dof);
1100
1101         switch (dof->dof_type) {
1102                 case DFT_REGULAR:
1103                 case DFT_SYM:
1104                 case DFT_NODE:
1105                         if (obj->oo_dt.do_body_ops == NULL)
1106                                 obj->oo_dt.do_body_ops = &osd_body_ops;
1107                         break;
1108                 default:
1109                         break;
1110         }
1111
1112         LASSERT(handle != NULL);
1113         oh = container_of0(handle, struct osd_thandle, ot_super);
1114         LASSERT(oh->ot_tx != NULL);
1115
1116         /* this is the minimum set of EAs on every Lustre object */
1117         obj->oo_ea_in_bonus = ZFS_SA_BASE_ATTR_SIZE +
1118                                 sizeof(__u64) + /* VBR VERSION */
1119                                 sizeof(struct lustre_mdt_attrs); /* LMA */
1120         /* reserve 32 bytes for extra stuff like ACLs */
1121         dnode_size = size_roundup_power2(obj->oo_ea_in_bonus + 32);
1122
1123         switch (dof->dof_type) {
1124                 case DFT_DIR:
1125                         dt->do_index_ops = &osd_dir_ops;
1126                 case DFT_INDEX:
1127                         /* for zap create */
1128                         dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT, FALSE, NULL);
1129                         dmu_tx_hold_sa_create(oh->ot_tx, dnode_size);
1130                         break;
1131                 case DFT_REGULAR:
1132                 case DFT_SYM:
1133                 case DFT_NODE:
1134                         /* first, we'll create new object */
1135                         dmu_tx_hold_sa_create(oh->ot_tx, dnode_size);
1136                         break;
1137
1138                 default:
1139                         LBUG();
1140                         break;
1141         }
1142
1143         /* and we'll add it to some mapping */
1144         zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0);
1145         dmu_tx_hold_zap(oh->ot_tx, zapid, TRUE, NULL);
1146
1147         /* will help to find FID->ino mapping at dt_insert() */
1148         osd_idc_find_and_init(env, osd, obj);
1149
1150         rc = osd_declare_quota(env, osd, attr->la_uid, attr->la_gid, 1, oh,
1151                                false, NULL, false);
1152
1153         RETURN(rc);
1154 }
1155
1156 int __osd_attr_init(const struct lu_env *env, struct osd_device *osd,
1157                     sa_handle_t *sa_hdl, dmu_tx_t *tx,
1158                     struct lu_attr *la, uint64_t parent)
1159 {
1160         sa_bulk_attr_t  *bulk = osd_oti_get(env)->oti_attr_bulk;
1161         struct osa_attr *osa = &osd_oti_get(env)->oti_osa;
1162         uint64_t         gen;
1163         uint64_t         crtime[2];
1164         timestruc_t      now;
1165         int              cnt;
1166         int              rc;
1167
1168         LASSERT(sa_hdl);
1169
1170         gen = dmu_tx_get_txg(tx);
1171         gethrestime(&now);
1172         ZFS_TIME_ENCODE(&now, crtime);
1173
1174         osa->atime[0] = la->la_atime;
1175         osa->ctime[0] = la->la_ctime;
1176         osa->mtime[0] = la->la_mtime;
1177         osa->mode = la->la_mode;
1178         osa->uid = la->la_uid;
1179         osa->gid = la->la_gid;
1180         osa->rdev = la->la_rdev;
1181         osa->nlink = la->la_nlink;
1182         osa->flags = attrs_fs2zfs(la->la_flags);
1183         osa->size  = la->la_size;
1184
1185         /*
1186          * we need to create all SA below upon object create.
1187          *
1188          * XXX The attribute order matters since the accounting callback relies
1189          * on static offsets (i.e. SA_*_OFFSET, see zfs_space_delta_cb()) to
1190          * look up the UID/GID attributes. Moreover, the callback does not seem
1191          * to support the spill block.
1192          * We define attributes in the same order as SA_*_OFFSET in order to
1193          * work around the problem. See ORI-610.
1194          */
1195         cnt = 0;
1196         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(osd), NULL, &osa->mode, 8);
1197         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(osd), NULL, &osa->size, 8);
1198         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GEN(osd), NULL, &gen, 8);
1199         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(osd), NULL, &osa->uid, 8);
1200         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(osd), NULL, &osa->gid, 8);
1201         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_PARENT(osd), NULL, &parent, 8);
1202         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL, &osa->flags, 8);
1203         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(osd), NULL, osa->atime, 16);
1204         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(osd), NULL, osa->mtime, 16);
1205         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(osd), NULL, osa->ctime, 16);
1206         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CRTIME(osd), NULL, crtime, 16);
1207         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(osd), NULL, &osa->nlink, 8);
1208         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_RDEV(osd), NULL, &osa->rdev, 8);
1209         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
1210
1211         rc = -sa_replace_all_by_template(sa_hdl, bulk, cnt, tx);
1212
1213         return rc;
1214 }
1215
1216 static int osd_find_new_dnode(const struct lu_env *env, dmu_tx_t *tx,
1217                               uint64_t oid, dnode_t **dnp)
1218 {
1219         dmu_tx_hold_t *txh;
1220         int rc = 0;
1221
1222         /* take dnode_t from tx to save on dnode#->dnode_t lookup */
1223         for (txh = list_tail(&tx->tx_holds); txh;
1224              txh = list_prev(&tx->tx_holds, txh)) {
1225                 dnode_t *dn = txh->txh_dnode;
1226                 dmu_buf_impl_t *db;
1227
1228                 if (dn == NULL)
1229                         continue;
1230                 if (dn->dn_object != oid)
1231                         continue;
1232                 db = dn->dn_bonus;
1233                 if (db == NULL) {
1234                         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1235                         if (dn->dn_bonus == NULL)
1236                                 dbuf_create_bonus(dn);
1237                         rw_exit(&dn->dn_struct_rwlock);
1238                 }
1239                 db = dn->dn_bonus;
1240                 LASSERT(db);
1241                 LASSERT(dn->dn_handle);
1242                 DB_DNODE_ENTER(db);
1243                 if (refcount_add(&db->db_holds, osd_obj_tag) == 1) {
1244                         refcount_add(&dn->dn_holds, tag);
1245                         atomic_inc_32(&dn->dn_dbufs_count);
1246                 }
1247                 *dnp = dn;
1248                 break;
1249         }
1250
1251         if (unlikely(*dnp == NULL))
1252                 rc = __osd_obj2dnode(env, tx->tx_objset, oid, dnp);
1253
1254         return rc;
1255 }
1256
1257 /*
1258  * The transaction passed to this routine must have
1259  * dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT) called and then assigned
1260  * to a transaction group.
1261  */
1262 int __osd_object_create(const struct lu_env *env, struct osd_object *obj,
1263                         dnode_t **dnp, dmu_tx_t *tx, struct lu_attr *la)
1264 {
1265         struct osd_device   *osd = osd_obj2dev(obj);
1266         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
1267         dmu_object_type_t    type = DMU_OT_PLAIN_FILE_CONTENTS;
1268         uint64_t oid;
1269
1270         /* Use DMU_OTN_UINT8_METADATA for local objects so their data blocks
1271          * would get an additional ditto copy */
1272         if (unlikely(S_ISREG(la->la_mode) &&
1273                      fid_seq_is_local_file(fid_seq(fid))))
1274                 type = DMU_OTN_UINT8_METADATA;
1275
1276         /* Create a new DMU object using the default dnode size. */
1277         oid = osd_dmu_object_alloc(osd->od_os, type, 0, 0, tx);
1278
1279         LASSERT(la->la_valid & LA_MODE);
1280         la->la_size = 0;
1281         la->la_nlink = 1;
1282
1283         return osd_find_new_dnode(env, tx, oid, dnp);
1284 }
1285
1286 /*
1287  * The transaction passed to this routine must have
1288  * dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, ...) called and then assigned
1289  * to a transaction group.
1290  *
1291  * Using ZAP_FLAG_HASH64 will force the ZAP to always be a FAT ZAP.
1292  * This is fine for directories today, because storing the FID in the dirent
1293  * will also require a FAT ZAP.  If there is a new type of micro ZAP created
1294  * then we might need to re-evaluate the use of this flag and instead do
1295  * a conversion from the different internal ZAP hash formats being used. */
1296 int __osd_zap_create(const struct lu_env *env, struct osd_device *osd,
1297                      dnode_t **dnp, dmu_tx_t *tx, struct lu_attr *la,
1298                      zap_flags_t flags)
1299 {
1300         uint64_t oid;
1301
1302         /* Assert that the transaction has been assigned to a
1303            transaction group. */
1304         LASSERT(tx->tx_txg != 0);
1305         *dnp = NULL;
1306
1307         oid = osd_zap_create_flags(osd->od_os, 0, flags | ZAP_FLAG_HASH64,
1308                                    DMU_OT_DIRECTORY_CONTENTS,
1309                                    14, /* == ZFS fzap_default_blockshift */
1310                                    DN_MAX_INDBLKSHIFT, /* indirect blockshift */
1311                                    0, tx);
1312
1313         la->la_size = 2;
1314         la->la_nlink = 1;
1315
1316         return osd_find_new_dnode(env, tx, oid, dnp);
1317 }
1318
1319 static dnode_t *osd_mkidx(const struct lu_env *env, struct osd_object *obj,
1320                           struct lu_attr *la, struct osd_thandle *oh)
1321 {
1322         dnode_t *dn;
1323         int rc;
1324
1325         /* Index file should be created as regular file in order not to confuse
1326          * ZPL which could interpret them as directory.
1327          * We set ZAP_FLAG_UINT64_KEY to let ZFS know than we are going to use
1328          * binary keys */
1329         LASSERT(S_ISREG(la->la_mode));
1330         rc = __osd_zap_create(env, osd_obj2dev(obj), &dn, oh->ot_tx, la,
1331                               ZAP_FLAG_UINT64_KEY);
1332         if (rc)
1333                 return ERR_PTR(rc);
1334         return dn;
1335 }
1336
1337 static dnode_t *osd_mkdir(const struct lu_env *env, struct osd_object *obj,
1338                           struct lu_attr *la, struct osd_thandle *oh)
1339 {
1340         dnode_t *dn;
1341         int rc;
1342
1343         LASSERT(S_ISDIR(la->la_mode));
1344         rc = __osd_zap_create(env, osd_obj2dev(obj), &dn, oh->ot_tx, la, 0);
1345         if (rc)
1346                 return ERR_PTR(rc);
1347         return dn;
1348 }
1349
1350 static dnode_t *osd_mkreg(const struct lu_env *env, struct osd_object *obj,
1351                           struct lu_attr *la, struct osd_thandle *oh)
1352 {
1353         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
1354         struct osd_device *osd = osd_obj2dev(obj);
1355         dnode_t *dn;
1356         int rc;
1357
1358         LASSERT(S_ISREG(la->la_mode));
1359         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1360         if (rc)
1361                 return ERR_PTR(rc);
1362
1363         if ((fid_is_idif(fid) || fid_is_norm(fid) || fid_is_echo(fid)) &&
1364             osd->od_is_ost) {
1365                 /* The minimum block size must be at least page size otherwise
1366                  * it will break the assumption in tgt_thread_big_cache where
1367                  * the array size is PTLRPC_MAX_BRW_PAGES. It will also affect
1368                  * RDMA due to subpage transfer size */
1369                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
1370                                                PAGE_SIZE, 0, oh->ot_tx);
1371                 if (unlikely(rc)) {
1372                         CERROR("%s: can't change blocksize: %d\n",
1373                                osd->od_svname, rc);
1374                         return ERR_PTR(rc);
1375                 }
1376         }
1377
1378         return dn;
1379 }
1380
1381 static dnode_t *osd_mksym(const struct lu_env *env, struct osd_object *obj,
1382                           struct lu_attr *la, struct osd_thandle *oh)
1383 {
1384         dnode_t *dn;
1385         int rc;
1386
1387         LASSERT(S_ISLNK(la->la_mode));
1388         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1389         if (rc)
1390                 return ERR_PTR(rc);
1391         return dn;
1392 }
1393
1394 static dnode_t *osd_mknod(const struct lu_env *env, struct osd_object *obj,
1395                           struct lu_attr *la, struct osd_thandle *oh)
1396 {
1397         dnode_t *dn;
1398         int rc;
1399
1400         if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode))
1401                 la->la_valid |= LA_RDEV;
1402
1403         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1404         if (rc)
1405                 return ERR_PTR(rc);
1406         return dn;
1407 }
1408
1409 typedef dnode_t *(*osd_obj_type_f)(const struct lu_env *env,
1410                                    struct osd_object *obj,
1411                                    struct lu_attr *la,
1412                                    struct osd_thandle *oh);
1413
1414 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
1415 {
1416         osd_obj_type_f result;
1417
1418         switch (type) {
1419         case DFT_DIR:
1420                 result = osd_mkdir;
1421                 break;
1422         case DFT_INDEX:
1423                 result = osd_mkidx;
1424                 break;
1425         case DFT_REGULAR:
1426                 result = osd_mkreg;
1427                 break;
1428         case DFT_SYM:
1429                 result = osd_mksym;
1430                 break;
1431         case DFT_NODE:
1432                 result = osd_mknod;
1433                 break;
1434         default:
1435                 LBUG();
1436                 break;
1437         }
1438         return result;
1439 }
1440
1441 /*
1442  * Concurrency: @dt is write locked.
1443  */
1444 static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
1445                              struct lu_attr *attr,
1446                              struct dt_allocation_hint *hint,
1447                              struct dt_object_format *dof,
1448                              struct thandle *th)
1449 {
1450         struct osd_thread_info  *info = osd_oti_get(env);
1451         struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
1452         struct zpl_direntry     *zde = &info->oti_zde.lzd_reg;
1453         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
1454         struct osd_object       *obj = osd_dt_obj(dt);
1455         struct osd_device       *osd = osd_obj2dev(obj);
1456         char                    *buf = info->oti_str;
1457         struct osd_thandle      *oh;
1458         dnode_t *dn = NULL;
1459         uint64_t                 zapid, parent = 0;
1460         int                      rc;
1461
1462         ENTRY;
1463
1464         /* concurrent create declarations should not see
1465          * the object inconsistent (db, attr, etc).
1466          * in regular cases acquisition should be cheap */
1467         down_write(&obj->oo_guard);
1468
1469         if (unlikely(dt_object_exists(dt)))
1470                 GOTO(out, rc = -EEXIST);
1471
1472         LASSERT(osd_invariant(obj));
1473         LASSERT(dof != NULL);
1474
1475         LASSERT(th != NULL);
1476         oh = container_of0(th, struct osd_thandle, ot_super);
1477
1478         LASSERT(obj->oo_dn == NULL);
1479
1480         /* to follow ZFS on-disk format we need
1481          * to initialize parent dnode properly */
1482         if (hint != NULL && hint->dah_parent != NULL &&
1483             !dt_object_remote(hint->dah_parent))
1484                 parent = osd_dt_obj(hint->dah_parent)->oo_dn->dn_object;
1485
1486         /* we may fix some attributes, better do not change the source */
1487         obj->oo_attr = *attr;
1488         obj->oo_attr.la_valid |= LA_SIZE | LA_NLINK | LA_TYPE;
1489
1490         dn = osd_create_type_f(dof->dof_type)(env, obj, &obj->oo_attr, oh);
1491         if (IS_ERR(dn)) {
1492                 rc = PTR_ERR(dn);
1493                 dn = NULL;
1494                 GOTO(out, rc);
1495         }
1496
1497         zde->zde_pad = 0;
1498         zde->zde_dnode = dn->dn_object;
1499         zde->zde_type = IFTODT(attr->la_mode & S_IFMT);
1500
1501         zapid = osd_get_name_n_idx(env, osd, fid, buf, sizeof(info->oti_str));
1502
1503         rc = -zap_add(osd->od_os, zapid, buf, 8, 1, zde, oh->ot_tx);
1504         if (rc)
1505                 GOTO(out, rc);
1506
1507         /* Now add in all of the "SA" attributes */
1508         rc = -sa_handle_get(osd->od_os, dn->dn_object, NULL,
1509                             SA_HDL_PRIVATE, &obj->oo_sa_hdl);
1510         if (rc)
1511                 GOTO(out, rc);
1512
1513         /* configure new osd object */
1514         obj->oo_dn = dn;
1515         parent = parent != 0 ? parent : zapid;
1516         rc = __osd_attr_init(env, osd, obj->oo_sa_hdl, oh->ot_tx,
1517                              &obj->oo_attr, parent);
1518         if (rc)
1519                 GOTO(out, rc);
1520
1521         /* XXX: oo_lma_flags */
1522         obj->oo_dt.do_lu.lo_header->loh_attr |= obj->oo_attr.la_mode & S_IFMT;
1523         if (likely(!fid_is_acct(lu_object_fid(&obj->oo_dt.do_lu))))
1524                 /* no body operations for accounting objects */
1525                 obj->oo_dt.do_body_ops = &osd_body_ops;
1526
1527         rc = -nvlist_alloc(&obj->oo_sa_xattr, NV_UNIQUE_NAME, KM_SLEEP);
1528         if (rc)
1529                 GOTO(out, rc);
1530
1531         /* initialize LMA */
1532         lustre_lma_init(lma, lu_object_fid(&obj->oo_dt.do_lu), 0, 0);
1533         lustre_lma_swab(lma);
1534         rc = -nvlist_add_byte_array(obj->oo_sa_xattr, XATTR_NAME_LMA,
1535                                     (uchar_t *)lma, sizeof(*lma));
1536         if (rc)
1537                 GOTO(out, rc);
1538         rc = __osd_sa_xattr_update(env, obj, oh);
1539         if (rc)
1540                 GOTO(out, rc);
1541         osd_idc_find_and_init(env, osd, obj);
1542
1543 out:
1544         if (unlikely(rc && dn)) {
1545                 dmu_object_free(osd->od_os, dn->dn_object, oh->ot_tx);
1546                 osd_dnode_rele(dn);
1547                 obj->oo_dn = NULL;
1548         } else if (!rc) {
1549                 obj->oo_dt.do_lu.lo_header->loh_attr |= LOHA_EXISTS;
1550         }
1551         up_write(&obj->oo_guard);
1552         RETURN(rc);
1553 }
1554
1555 static int osd_declare_object_ref_add(const struct lu_env *env,
1556                                       struct dt_object *dt,
1557                                       struct thandle *th)
1558 {
1559         return osd_declare_attr_set(env, dt, NULL, th);
1560 }
1561
1562 /*
1563  * Concurrency: @dt is write locked.
1564  */
1565 static int osd_object_ref_add(const struct lu_env *env,
1566                               struct dt_object *dt,
1567                               struct thandle *handle)
1568 {
1569         struct osd_object       *obj = osd_dt_obj(dt);
1570         struct osd_thandle      *oh;
1571         struct osd_device       *osd = osd_obj2dev(obj);
1572         uint64_t                 nlink;
1573         int rc;
1574
1575         ENTRY;
1576
1577         down_read(&obj->oo_guard);
1578         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
1579                 GOTO(out, rc = -ENOENT);
1580
1581         LASSERT(osd_invariant(obj));
1582         LASSERT(obj->oo_sa_hdl != NULL);
1583
1584         oh = container_of0(handle, struct osd_thandle, ot_super);
1585
1586         write_lock(&obj->oo_attr_lock);
1587         nlink = ++obj->oo_attr.la_nlink;
1588         write_unlock(&obj->oo_attr_lock);
1589
1590         rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
1591
1592 out:
1593         up_read(&obj->oo_guard);
1594         RETURN(rc);
1595 }
1596
1597 static int osd_declare_object_ref_del(const struct lu_env *env,
1598                                       struct dt_object *dt,
1599                                       struct thandle *handle)
1600 {
1601         return osd_declare_attr_set(env, dt, NULL, handle);
1602 }
1603
1604 /*
1605  * Concurrency: @dt is write locked.
1606  */
1607 static int osd_object_ref_del(const struct lu_env *env,
1608                               struct dt_object *dt,
1609                               struct thandle *handle)
1610 {
1611         struct osd_object       *obj = osd_dt_obj(dt);
1612         struct osd_thandle      *oh;
1613         struct osd_device       *osd = osd_obj2dev(obj);
1614         uint64_t                 nlink;
1615         int                      rc;
1616
1617         ENTRY;
1618
1619         down_read(&obj->oo_guard);
1620
1621         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
1622                 GOTO(out, rc = -ENOENT);
1623
1624         LASSERT(osd_invariant(obj));
1625         LASSERT(obj->oo_sa_hdl != NULL);
1626
1627         oh = container_of0(handle, struct osd_thandle, ot_super);
1628         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
1629
1630         write_lock(&obj->oo_attr_lock);
1631         nlink = --obj->oo_attr.la_nlink;
1632         write_unlock(&obj->oo_attr_lock);
1633
1634         rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
1635
1636 out:
1637         up_read(&obj->oo_guard);
1638         RETURN(rc);
1639 }
1640
1641 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
1642                            __u64 start, __u64 end)
1643 {
1644         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1645         ENTRY;
1646
1647         /* XXX: no other option than syncing the whole filesystem until we
1648          * support ZIL.  If the object tracked the txg that it was last
1649          * modified in, it could pass that txg here instead of "0".  Maybe
1650          * the changes are already committed, so no wait is needed at all? */
1651         if (!osd->od_dt_dev.dd_rdonly)
1652                 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
1653
1654         RETURN(0);
1655 }
1656
1657 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
1658 {
1659         return 0;
1660 }
1661
1662 static struct dt_object_operations osd_obj_ops = {
1663         .do_read_lock           = osd_object_read_lock,
1664         .do_write_lock          = osd_object_write_lock,
1665         .do_read_unlock         = osd_object_read_unlock,
1666         .do_write_unlock        = osd_object_write_unlock,
1667         .do_write_locked        = osd_object_write_locked,
1668         .do_attr_get            = osd_attr_get,
1669         .do_declare_attr_set    = osd_declare_attr_set,
1670         .do_attr_set            = osd_attr_set,
1671         .do_ah_init             = osd_ah_init,
1672         .do_declare_create      = osd_declare_object_create,
1673         .do_create              = osd_object_create,
1674         .do_declare_destroy     = osd_declare_object_destroy,
1675         .do_destroy             = osd_object_destroy,
1676         .do_index_try           = osd_index_try,
1677         .do_declare_ref_add     = osd_declare_object_ref_add,
1678         .do_ref_add             = osd_object_ref_add,
1679         .do_declare_ref_del     = osd_declare_object_ref_del,
1680         .do_ref_del             = osd_object_ref_del,
1681         .do_xattr_get           = osd_xattr_get,
1682         .do_declare_xattr_set   = osd_declare_xattr_set,
1683         .do_xattr_set           = osd_xattr_set,
1684         .do_declare_xattr_del   = osd_declare_xattr_del,
1685         .do_xattr_del           = osd_xattr_del,
1686         .do_xattr_list          = osd_xattr_list,
1687         .do_object_sync         = osd_object_sync,
1688         .do_invalidate          = osd_invalidate,
1689 };
1690
1691 static struct lu_object_operations osd_lu_obj_ops = {
1692         .loo_object_init        = osd_object_init,
1693         .loo_object_delete      = osd_object_delete,
1694         .loo_object_release     = osd_object_release,
1695         .loo_object_free        = osd_object_free,
1696         .loo_object_print       = osd_object_print,
1697         .loo_object_invariant   = osd_object_invariant,
1698 };
1699
1700 static int osd_otable_it_attr_get(const struct lu_env *env,
1701                                 struct dt_object *dt,
1702                                 struct lu_attr *attr)
1703 {
1704         attr->la_valid = 0;
1705         return 0;
1706 }
1707
1708 static struct dt_object_operations osd_obj_otable_it_ops = {
1709         .do_attr_get    = osd_otable_it_attr_get,
1710         .do_index_try   = osd_index_try,
1711 };