4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osd-zfs/osd_index.c
34 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
35 * Author: Mike Pershin <tappro@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_OSD
40 #include <libcfs/libcfs.h>
41 #include <obd_support.h>
42 #include <lustre_net.h>
44 #include <obd_class.h>
45 #include <lustre_disk.h>
46 #include <lustre_fid.h>
48 #include "osd_internal.h"
50 #include <sys/dnode.h>
54 #include <sys/spa_impl.h>
55 #include <sys/zfs_znode.h>
56 #include <sys/dmu_tx.h>
57 #include <sys/dmu_objset.h>
58 #include <sys/dsl_prop.h>
59 #include <sys/sa_impl.h>
62 static inline int osd_object_is_zap(dnode_t *dn)
64 return (dn->dn_type == DMU_OT_DIRECTORY_CONTENTS ||
65 dn->dn_type == DMU_OT_USERGROUP_USED);
68 /* We don't actually have direct access to the zap_hashbits() function
69 * so just pretend like we do for now. If this ever breaks we can look at
71 #define zap_hashbits(zc) 48
74 * | cd (16 bits) | hash (48 bits) |
75 * we need it in other form:
76 * |0| hash (48 bit) | cd (15 bit) |
77 * to be a full 64-bit ordered hash so that Lustre readdir can use it to merge
78 * the readdir hashes from multiple directory stripes uniformly on the client.
79 * Another point is sign bit, the hash range should be in [0, 2^63-1] because
80 * loff_t (for llseek) needs to be a positive value. This means the "cd" field
81 * should only be the low 15 bits.
83 uint64_t osd_zap_cursor_serialize(zap_cursor_t *zc)
85 uint64_t zfs_hash = zap_cursor_serialize(zc) & (~0ULL >> 1);
87 return (zfs_hash >> zap_hashbits(zc)) |
88 (zfs_hash << (63 - zap_hashbits(zc)));
91 void osd_zap_cursor_init_serialized(zap_cursor_t *zc, struct objset *os,
92 uint64_t id, uint64_t dirhash)
94 uint64_t zfs_hash = ((dirhash << zap_hashbits(zc)) & (~0ULL >> 1)) |
95 (dirhash >> (63 - zap_hashbits(zc)));
97 zap_cursor_init_serialized(zc, os, id, zfs_hash);
100 int osd_zap_cursor_init(zap_cursor_t **zc, struct objset *os,
101 uint64_t id, uint64_t dirhash)
106 if (unlikely(t == NULL))
109 osd_zap_cursor_init_serialized(t, os, id, dirhash);
115 void osd_zap_cursor_fini(zap_cursor_t *zc)
121 static inline void osd_obj_cursor_init_serialized(zap_cursor_t *zc,
122 struct osd_object *o,
125 struct osd_device *d = osd_obj2dev(o);
126 osd_zap_cursor_init_serialized(zc, d->od_os,
127 o->oo_dn->dn_object, dirhash);
130 static inline int osd_obj_cursor_init(zap_cursor_t **zc, struct osd_object *o,
133 struct osd_device *d = osd_obj2dev(o);
134 return osd_zap_cursor_init(zc, d->od_os, o->oo_dn->dn_object, dirhash);
137 static struct dt_it *osd_index_it_init(const struct lu_env *env,
138 struct dt_object *dt,
141 struct osd_thread_info *info = osd_oti_get(env);
142 struct osd_zap_it *it;
143 struct osd_object *obj = osd_dt_obj(dt);
144 struct lu_object *lo = &dt->do_lu;
148 if (obj->oo_destroyed)
149 RETURN(ERR_PTR(-ENOENT));
151 LASSERT(lu_object_exists(lo));
155 OBD_SLAB_ALLOC_PTR_GFP(it, osd_zapit_cachep, GFP_NOFS);
157 RETURN(ERR_PTR(-ENOMEM));
159 rc = osd_obj_cursor_init(&it->ozi_zc, obj, 0);
161 OBD_SLAB_FREE_PTR(it, osd_zapit_cachep);
169 RETURN((struct dt_it *)it);
172 static void osd_index_it_fini(const struct lu_env *env, struct dt_it *di)
174 struct osd_zap_it *it = (struct osd_zap_it *)di;
175 struct osd_object *obj;
179 LASSERT(it->ozi_obj);
183 osd_zap_cursor_fini(it->ozi_zc);
184 osd_object_put(env, obj);
185 OBD_SLAB_FREE_PTR(it, osd_zapit_cachep);
191 static void osd_index_it_put(const struct lu_env *env, struct dt_it *di)
193 /* PBS: do nothing : ref are incremented at retrive and decreamented
197 static inline void osd_it_append_attrs(struct lu_dirent *ent, __u32 attr,
200 const unsigned align = sizeof(struct luda_type) - 1;
201 struct luda_type *lt;
203 /* check if file type is required */
204 if (attr & LUDA_TYPE) {
205 len = (len + align) & ~align;
207 lt = (void *)ent->lde_name + len;
208 lt->lt_type = cpu_to_le16(DTTOIF(type));
209 ent->lde_attrs |= LUDA_TYPE;
212 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
215 int __osd_xattr_load_by_oid(struct osd_device *osd, uint64_t oid, nvlist_t **sa)
221 rc = -dmu_bonus_hold(osd->od_os, oid, osd_obj_tag, &db);
223 CERROR("%s: can't get bonus, rc = %d\n", osd->od_svname, rc);
227 rc = -sa_handle_get_from_db(osd->od_os, db, NULL, SA_HDL_PRIVATE, &hdl);
229 dmu_buf_rele(db, osd_obj_tag);
233 rc = __osd_xattr_load(osd, hdl, sa);
235 sa_handle_destroy(hdl);
240 * Get the object's FID from its LMA EA.
242 * \param[in] env pointer to the thread context
243 * \param[in] osd pointer to the OSD device
244 * \param[in] oid the object's local identifier
245 * \param[out] fid the buffer to hold the object's FID
247 * \retval 0 for success
248 * \retval negative error number on failure
250 static int osd_get_fid_by_oid(const struct lu_env *env, struct osd_device *osd,
251 uint64_t oid, struct lu_fid *fid)
253 struct objset *os = osd->od_os;
254 struct osd_thread_info *oti = osd_oti_get(env);
255 struct lustre_mdt_attrs *lma =
256 (struct lustre_mdt_attrs *)oti->oti_buf;
258 nvlist_t *sa_xattr = NULL;
259 sa_handle_t *sa_hdl = NULL;
260 uchar_t *nv_value = NULL;
261 uint64_t xattr = ZFS_NO_OBJECT;
266 rc = __osd_xattr_load_by_oid(osd, oid, &sa_xattr);
273 rc = -nvlist_lookup_byte_array(sa_xattr, XATTR_NAME_LMA, &nv_value,
281 if (unlikely(size > sizeof(oti->oti_buf)))
282 GOTO(out, rc = -ERANGE);
284 memcpy(lma, nv_value, size);
289 rc = -sa_handle_get(os, oid, NULL, SA_HDL_PRIVATE, &sa_hdl);
293 rc = -sa_lookup(sa_hdl, SA_ZPL_XATTR(osd), &xattr, 8);
294 sa_handle_destroy(sa_hdl);
299 buf.lb_len = sizeof(oti->oti_buf);
300 rc = __osd_xattr_get_large(env, osd, xattr, &buf,
301 XATTR_NAME_LMA, &size);
306 if (size < sizeof(*lma))
307 GOTO(out, rc = -EIO);
309 lustre_lma_swab(lma);
310 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
311 CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
312 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
313 "oid = %#llx\n", osd->od_svname,
314 lma->lma_incompat & ~LMA_INCOMPAT_SUPP, oid);
315 GOTO(out, rc = -EOPNOTSUPP);
317 *fid = lma->lma_self_fid;
322 if (sa_xattr != NULL)
323 nvlist_free(sa_xattr);
328 * As we don't know FID, we can't use LU object, so this function
329 * partially duplicate osd_xattr_get_internal() which is built around
330 * LU-object and uses it to cache data like regular EA dnode, etc
332 static int osd_find_parent_by_dnode(const struct lu_env *env,
336 struct osd_object *obj = osd_dt_obj(o);
337 struct osd_device *osd = osd_obj2dev(obj);
338 uint64_t dnode = ZFS_NO_OBJECT;
342 /* first of all, get parent dnode from own attributes */
343 rc = osd_sa_handle_get(obj);
346 rc = -sa_lookup(obj->oo_sa_hdl, SA_ZPL_PARENT(osd), &dnode, 8);
348 rc = osd_get_fid_by_oid(env, osd, dnode, fid);
353 static int osd_find_parent_fid(const struct lu_env *env, struct dt_object *o,
356 struct link_ea_header *leh;
357 struct link_ea_entry *lee;
362 buf.lb_buf = osd_oti_get(env)->oti_buf;
363 buf.lb_len = sizeof(osd_oti_get(env)->oti_buf);
365 rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK);
367 rc = osd_xattr_get(env, o, &LU_BUF_NULL, XATTR_NAME_LINK);
371 OBD_ALLOC(buf.lb_buf, rc);
372 if (buf.lb_buf == NULL)
375 rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK);
379 if (rc < sizeof(*leh) + sizeof(*lee))
380 GOTO(out, rc = -EINVAL);
383 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
384 leh->leh_magic = LINK_EA_MAGIC;
385 leh->leh_reccount = __swab32(leh->leh_reccount);
386 leh->leh_len = __swab64(leh->leh_len);
388 if (leh->leh_magic != LINK_EA_MAGIC)
389 GOTO(out, rc = -EINVAL);
390 if (leh->leh_reccount == 0)
391 GOTO(out, rc = -ENODATA);
393 lee = (struct link_ea_entry *)(leh + 1);
394 fid_be_to_cpu(fid, (const struct lu_fid *)&lee->lee_parent_fid);
398 if (buf.lb_buf != osd_oti_get(env)->oti_buf)
399 OBD_FREE(buf.lb_buf, buf.lb_len);
402 /* this block can be enabled for additional verification
403 * it's trying to match FID from LinkEA vs. FID from LMA */
407 rc2 = osd_find_parent_by_dnode(env, o, &fid2);
409 if (lu_fid_eq(fid, &fid2) == 0)
410 CERROR("wrong parent: "DFID" != "DFID"\n",
411 PFID(fid), PFID(&fid2));
415 /* no LinkEA is found, let's try to find the fid in parent's LMA */
416 if (unlikely(rc != 0))
417 rc = osd_find_parent_by_dnode(env, o, fid);
422 static int osd_dir_lookup(const struct lu_env *env, struct dt_object *dt,
423 struct dt_rec *rec, const struct dt_key *key)
425 struct osd_thread_info *oti = osd_oti_get(env);
426 struct osd_object *obj = osd_dt_obj(dt);
427 struct osd_device *osd = osd_obj2dev(obj);
428 char *name = (char *)key;
432 if (name[0] == '.') {
434 const struct lu_fid *f = lu_object_fid(&dt->do_lu);
435 memcpy(rec, f, sizeof(*f));
437 } else if (name[1] == '.' && name[2] == 0) {
438 rc = osd_find_parent_fid(env, dt, (struct lu_fid *)rec);
439 RETURN(rc == 0 ? 1 : rc);
443 memset(&oti->oti_zde.lzd_fid, 0, sizeof(struct lu_fid));
444 rc = osd_zap_lookup(osd, obj->oo_dn->dn_object, obj->oo_dn,
445 (char *)key, 8, sizeof(oti->oti_zde) / 8,
446 (void *)&oti->oti_zde);
450 if (likely(fid_is_sane(&oti->oti_zde.lzd_fid))) {
451 memcpy(rec, &oti->oti_zde.lzd_fid, sizeof(struct lu_fid));
455 rc = osd_get_fid_by_oid(env, osd, oti->oti_zde.lzd_reg.zde_dnode,
456 (struct lu_fid *)rec);
458 RETURN(rc == 0 ? 1 : (rc == -ENOENT ? -ENODATA : rc));
462 * In DNE environment, the object and its name entry may reside on different
463 * MDTs. Under such case, we will create an agent object on the MDT where the
464 * name entry resides. The agent object is empty, and indicates that the real
465 * object for the name entry resides on another MDT. If without agent object,
466 * related name entry will be skipped when perform MDT side file level backup
467 * and restore via ZPL by userspace tool, such as 'tar'.
469 static int osd_create_agent_object(const struct lu_env *env,
470 struct osd_device *osd,
471 struct luz_direntry *zde,
472 uint64_t parent, dmu_tx_t *tx)
474 struct osd_thread_info *info = osd_oti_get(env);
475 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
476 struct lu_attr *la = &info->oti_la;
477 nvlist_t *nvbuf = NULL;
483 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_AGENTOBJ))
486 rc = -nvlist_alloc(&nvbuf, NV_UNIQUE_NAME, KM_SLEEP);
490 lustre_lma_init(lma, &zde->lzd_fid, 0, LMAI_AGENT);
491 lustre_lma_swab(lma);
492 rc = -nvlist_add_byte_array(nvbuf, XATTR_NAME_LMA, (uchar_t *)lma,
497 la->la_valid = LA_TYPE | LA_MODE;
498 la->la_mode = (DTTOIF(zde->lzd_reg.zde_type) & S_IFMT) |
499 S_IRUGO | S_IWUSR | S_IXUGO;
501 if (S_ISDIR(la->la_mode))
502 rc = __osd_zap_create(env, osd, &dn, tx, la,
503 osd_find_dnsize(osd, OSD_BASE_EA_IN_BONUS), 0);
505 rc = __osd_object_create(env, osd, NULL, &zde->lzd_fid,
510 zde->lzd_reg.zde_dnode = dn->dn_object;
511 rc = -sa_handle_get(osd->od_os, dn->dn_object, NULL,
512 SA_HDL_PRIVATE, &hdl);
514 rc = __osd_attr_init(env, osd, NULL, hdl, tx,
516 sa_handle_destroy(hdl);
524 dmu_object_free(osd->od_os, dn->dn_object, tx);
534 int osd_add_to_remote_parent(const struct lu_env *env,
535 struct osd_device *osd,
536 struct osd_object *obj,
537 struct osd_thandle *oh)
539 struct osd_thread_info *info = osd_oti_get(env);
540 struct luz_direntry *zde = &info->oti_zde;
541 char *name = info->oti_str;
542 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
543 struct lustre_mdt_attrs *lma = (struct lustre_mdt_attrs *)info->oti_buf;
544 struct lu_buf buf = {
546 .lb_len = sizeof(info->oti_buf),
552 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_AGENTENT))
555 rc = osd_xattr_get_internal(env, obj, &buf, XATTR_NAME_LMA, &size);
557 CWARN("%s: fail to load LMA for adding "
558 DFID" to remote parent: rc = %d\n",
559 osd_name(osd), PFID(fid), rc);
563 lustre_lma_swab(lma);
564 lma->lma_incompat |= LMAI_REMOTE_PARENT;
565 lustre_lma_swab(lma);
567 rc = osd_xattr_set_internal(env, obj, &buf, XATTR_NAME_LMA,
568 LU_XATTR_REPLACE, oh);
570 CWARN("%s: fail to update LMA for adding "
571 DFID" to remote parent: rc = %d\n",
572 osd_name(osd), PFID(fid), rc);
576 osd_fid2str(name, fid, sizeof(info->oti_str));
577 zde->lzd_reg.zde_dnode = obj->oo_dn->dn_object;
578 zde->lzd_reg.zde_type = IFTODT(S_IFDIR);
581 rc = osd_zap_add(osd, osd->od_remote_parent_dir, NULL,
582 name, 8, sizeof(*zde) / 8, zde, oh->ot_tx);
583 if (unlikely(rc == -EEXIST))
586 CWARN("%s: fail to add name entry for "
587 DFID" to remote parent: rc = %d\n",
588 osd_name(osd), PFID(fid), rc);
590 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
595 int osd_delete_from_remote_parent(const struct lu_env *env,
596 struct osd_device *osd,
597 struct osd_object *obj,
598 struct osd_thandle *oh, bool destroy)
600 struct osd_thread_info *info = osd_oti_get(env);
601 char *name = info->oti_str;
602 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
603 struct lustre_mdt_attrs *lma = (struct lustre_mdt_attrs *)info->oti_buf;
604 struct lu_buf buf = {
606 .lb_len = sizeof(info->oti_buf),
612 osd_fid2str(name, fid, sizeof(info->oti_str));
613 rc = osd_zap_remove(osd, osd->od_remote_parent_dir, NULL,
615 if (unlikely(rc == -ENOENT))
618 CERROR("%s: fail to remove entry under remote "
619 "parent for "DFID": rc = %d\n",
620 osd_name(osd), PFID(fid), rc);
625 rc = osd_xattr_get_internal(env, obj, &buf, XATTR_NAME_LMA, &size);
627 CERROR("%s: fail to load LMA for removing "
628 DFID" from remote parent: rc = %d\n",
629 osd_name(osd), PFID(fid), rc);
633 lustre_lma_swab(lma);
634 lma->lma_incompat &= ~LMAI_REMOTE_PARENT;
635 lustre_lma_swab(lma);
637 rc = osd_xattr_set_internal(env, obj, &buf, XATTR_NAME_LMA,
638 LU_XATTR_REPLACE, oh);
640 CERROR("%s: fail to update LMA for removing "
641 DFID" from remote parent: rc = %d\n",
642 osd_name(osd), PFID(fid), rc);
644 lu_object_clear_agent_entry(&obj->oo_dt.do_lu);
649 static int osd_declare_dir_insert(const struct lu_env *env,
650 struct dt_object *dt,
651 const struct dt_rec *rec,
652 const struct dt_key *key,
655 struct osd_object *obj = osd_dt_obj(dt);
656 struct osd_device *osd = osd_obj2dev(obj);
657 const struct dt_insert_rec *rec1;
658 const struct lu_fid *fid;
659 struct osd_thandle *oh;
661 struct osd_idmap_cache *idc;
664 rec1 = (struct dt_insert_rec *)rec;
666 LASSERT(fid != NULL);
667 LASSERT(rec1->rec_type != 0);
670 oh = container_of0(th, struct osd_thandle, ot_super);
672 idc = osd_idc_find_or_init(env, osd, fid);
674 RETURN(PTR_ERR(idc));
676 if (idc->oic_remote) {
677 const char *name = (const char *)key;
679 if (name[0] != '.' || name[1] != '.' || name[2] != 0) {
680 /* Prepare agent object for remote entry that will
681 * be used for operations via ZPL, such as MDT side
682 * file-level backup and restore. */
683 dmu_tx_hold_sa_create(oh->ot_tx,
684 osd_find_dnsize(osd, OSD_BASE_EA_IN_BONUS));
685 if (S_ISDIR(rec1->rec_type))
686 dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT,
691 /* This is for inserting dot/dotdot for new created dir. */
692 if (obj->oo_dn == NULL)
693 object = DMU_NEW_OBJECT;
695 object = obj->oo_dn->dn_object;
697 /* do not specify the key as then DMU is trying to look it up
698 * which is very expensive. usually the layers above lookup
699 * before insertion */
700 osd_tx_hold_zap(oh->ot_tx, object, obj->oo_dn, TRUE, NULL);
705 static int osd_seq_exists(const struct lu_env *env, struct osd_device *osd,
708 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
709 struct seq_server_site *ss = osd_seq_site(osd);
714 LASSERT(ss->ss_server_fld != NULL);
716 rc = osd_fld_lookup(env, osd, seq, range);
719 CERROR("%s: Can not lookup fld for %#llx\n",
724 RETURN(ss->ss_node_id == range->lsr_index);
727 int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
728 const struct lu_fid *fid)
730 struct seq_server_site *ss = osd_seq_site(osd);
733 /* FID seqs not in FLDB, must be local seq */
734 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
737 /* If FLD is not being initialized yet, it only happens during the
738 * initialization, likely during mgs initialization, and we assume
739 * this is local FID. */
740 if (ss == NULL || ss->ss_server_fld == NULL)
743 /* Only check the local FLDB here */
744 if (osd_seq_exists(env, osd, fid_seq(fid)))
751 * Inserts (key, value) pair in \a directory object.
753 * \param dt osd index object
754 * \param key key for index
755 * \param rec record reference
756 * \param th transaction handler
757 * \param ignore_quota update should not affect quota
760 * \retval -ve failure
762 static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
763 const struct dt_rec *rec, const struct dt_key *key,
764 struct thandle *th, int ignore_quota)
766 struct osd_thread_info *oti = osd_oti_get(env);
767 struct osd_object *parent = osd_dt_obj(dt);
768 struct osd_device *osd = osd_obj2dev(parent);
769 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
770 const struct lu_fid *fid = rec1->rec_fid;
771 struct osd_thandle *oh;
772 struct osd_idmap_cache *idc;
773 const char *name = (const char *)key;
774 struct luz_direntry *zde = &oti->oti_zde;
775 int num = sizeof(*zde) / 8;
779 LASSERT(parent->oo_dn);
781 LASSERT(dt_object_exists(dt));
782 LASSERT(osd_invariant(parent));
785 oh = container_of0(th, struct osd_thandle, ot_super);
787 idc = osd_idc_find(env, osd, fid);
788 if (unlikely(idc == NULL)) {
789 /* this dt_insert() wasn't declared properly, so
790 * FID is missing in OI cache. we better do not
791 * lookup FID in FLDB/OI and don't risk to deadlock,
792 * but in some special cases (lfsck testing, etc)
793 * it's much simpler than fixing a caller */
794 CERROR("%s: "DFID" wasn't declared for insert\n",
795 osd_name(osd), PFID(fid));
796 idc = osd_idc_find_or_init(env, osd, fid);
798 RETURN(PTR_ERR(idc));
801 CLASSERT(sizeof(zde->lzd_reg) == 8);
802 CLASSERT(sizeof(*zde) % 8 == 0);
804 memset(&zde->lzd_reg, 0, sizeof(zde->lzd_reg));
805 zde->lzd_reg.zde_type = IFTODT(rec1->rec_type & S_IFMT);
808 if (idc->oic_remote) {
809 if (name[0] != '.' || name[1] != '.' || name[2] != 0) {
810 /* Create agent inode for remote object that will
811 * be used for MDT file-level backup and restore. */
812 rc = osd_create_agent_object(env, osd, zde,
813 parent->oo_dn->dn_object, oh->ot_tx);
815 CWARN("%s: Fail to create agent object for "
817 osd_name(osd), PFID(fid), rc);
818 /* Ignore the failure since the system can go
819 * ahead if we do not care about the MDT side
820 * file-level backup and restore. */
825 if (unlikely(idc->oic_dnode == 0)) {
826 /* for a reason OI cache wasn't filled properly */
827 CERROR("%s: OIC for "DFID" isn't filled\n",
828 osd_name(osd), PFID(fid));
831 if (name[0] == '.') {
833 /* do not store ".", instead generate it
834 * during iteration */
836 } else if (name[1] == '.' && name[2] == 0) {
837 uint64_t dnode = idc->oic_dnode;
838 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT))
841 /* update parent dnode in the child.
842 * later it will be used to generate ".." */
843 rc = osd_object_sa_update(parent,
850 zde->lzd_reg.zde_dnode = idc->oic_dnode;
853 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR))
854 zde->lzd_fid.f_ver = ~0;
856 /* The logic is not related with IGIF, just re-use the fail_loc value
857 * to be consistent with ldiskfs case, then share the same test logic */
858 if (OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF))
861 /* Insert (key,oid) into ZAP */
862 rc = osd_zap_add(osd, parent->oo_dn->dn_object, parent->oo_dn,
863 name, 8, num, (void *)zde, oh->ot_tx);
864 if (unlikely(rc == -EEXIST &&
865 name[0] == '.' && name[1] == '.' && name[2] == 0))
866 /* Update (key,oid) in ZAP */
867 rc = -zap_update(osd->od_os, parent->oo_dn->dn_object, name, 8,
868 sizeof(*zde) / 8, (void *)zde, oh->ot_tx);
875 static int osd_declare_dir_delete(const struct lu_env *env,
876 struct dt_object *dt,
877 const struct dt_key *key,
880 struct osd_object *obj = osd_dt_obj(dt);
881 dnode_t *zap_dn = obj->oo_dn;
882 struct osd_thandle *oh;
883 const char *name = (const char *)key;
886 LASSERT(dt_object_exists(dt));
887 LASSERT(osd_invariant(obj));
888 LASSERT(zap_dn != NULL);
891 oh = container_of0(th, struct osd_thandle, ot_super);
894 * In Orion . and .. were stored in the directory (not generated upon
895 * request as now). We preserve them for backward compatibility.
897 if (name[0] == '.') {
900 else if (name[1] == '.' && name[2] == 0)
904 /* do not specify the key as then DMU is trying to look it up
905 * which is very expensive. usually the layers above lookup
907 osd_tx_hold_zap(oh->ot_tx, zap_dn->dn_object, zap_dn, FALSE, NULL);
909 /* For destroying agent object if have. */
910 dmu_tx_hold_bonus(oh->ot_tx, DMU_NEW_OBJECT);
915 static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
916 const struct dt_key *key, struct thandle *th)
918 struct luz_direntry *zde = &osd_oti_get(env)->oti_zde;
919 struct osd_object *obj = osd_dt_obj(dt);
920 struct osd_device *osd = osd_obj2dev(obj);
921 struct osd_thandle *oh;
922 dnode_t *zap_dn = obj->oo_dn;
923 char *name = (char *)key;
930 oh = container_of0(th, struct osd_thandle, ot_super);
933 * In Orion . and .. were stored in the directory (not generated upon
934 * request as now). we preserve them for backward compatibility
936 if (name[0] == '.') {
939 } else if (name[1] == '.' && name[2] == 0) {
944 /* XXX: We have to say that lookup during delete_declare will affect
945 * performance, but we have to check whether the name entry (to
946 * be deleted) has agent object or not to avoid orphans.
948 * We will improve that in the future, some possible solutions,
950 * 1) Some hint from the caller via transaction handle to make
951 * the lookup conditionally.
952 * 2) Enhance the ZFS logic to recognize the OSD lookup result
953 * and delete the given entry directly without lookup again
954 * internally. LU-10190 */
955 memset(&zde->lzd_fid, 0, sizeof(zde->lzd_fid));
956 rc = osd_zap_lookup(osd, zap_dn->dn_object, zap_dn, name, 8, 3, zde);
959 CERROR("%s: failed to locate entry %s: rc = %d\n",
960 osd->od_svname, name, rc);
964 if (unlikely(osd_remote_fid(env, osd, &zde->lzd_fid) > 0)) {
965 rc = -dmu_object_free(osd->od_os, zde->lzd_reg.zde_dnode,
968 CERROR("%s: failed to destroy agent object (%llu) "
969 "for the entry %s: rc = %d\n", osd->od_svname,
970 (__u64)zde->lzd_reg.zde_dnode, name, rc);
973 /* Remove key from the ZAP */
974 rc = osd_zap_remove(osd, zap_dn->dn_object, zap_dn,
975 (char *)key, oh->ot_tx);
977 CERROR("%s: zap_remove %s failed: rc = %d\n",
978 osd->od_svname, name, rc);
983 static struct dt_it *osd_dir_it_init(const struct lu_env *env,
984 struct dt_object *dt,
987 struct osd_zap_it *it;
989 it = (struct osd_zap_it *)osd_index_it_init(env, dt, unused);
993 RETURN((struct dt_it *)it);
997 * Move Iterator to record specified by \a key
999 * \param di osd iterator
1000 * \param key key for index
1002 * \retval +ve di points to record with least key not larger than key
1003 * \retval 0 di points to exact matched key
1004 * \retval -ve failure
1006 static int osd_dir_it_get(const struct lu_env *env,
1007 struct dt_it *di, const struct dt_key *key)
1009 struct osd_zap_it *it = (struct osd_zap_it *)di;
1010 struct osd_object *obj = it->ozi_obj;
1011 char *name = (char *)key;
1016 LASSERT(it->ozi_zc);
1018 /* reset the cursor */
1019 zap_cursor_fini(it->ozi_zc);
1020 osd_obj_cursor_init_serialized(it->ozi_zc, obj, 0);
1022 /* XXX: implementation of the API is broken at the moment */
1023 LASSERT(((const char *)key)[0] == 0);
1030 if (name[0] == '.') {
1034 } else if (name[1] == '.' && name[2] == 0) {
1040 /* neither . nor .. - some real record */
1048 static void osd_dir_it_put(const struct lu_env *env, struct dt_it *di)
1050 /* PBS: do nothing : ref are incremented at retrive and decreamented
1055 * in Orion . and .. were stored in the directory, while ZPL
1056 * and current osd-zfs generate them up on request. so, we
1057 * need to ignore previously stored . and ..
1059 static int osd_index_retrieve_skip_dots(struct osd_zap_it *it,
1060 zap_attribute_t *za)
1065 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1068 if (unlikely(rc == 0 && za->za_name[0] == '.')) {
1069 if (za->za_name[1] == 0) {
1071 } else if (za->za_name[1] == '.' &&
1072 za->za_name[2] == 0) {
1075 if (unlikely(isdot))
1076 zap_cursor_advance(it->ozi_zc);
1078 } while (unlikely(rc == 0 && isdot));
1084 * to load a directory entry at a time and stored it in
1085 * iterator's in-memory data structure.
1087 * \param di, struct osd_it_ea, iterator's in memory structure
1089 * \retval +ve, iterator reached to end
1090 * \retval 0, iterator not reached to end
1091 * \retval -ve, on error
1093 static int osd_dir_it_next(const struct lu_env *env, struct dt_it *di)
1095 struct osd_zap_it *it = (struct osd_zap_it *)di;
1096 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1101 /* temp. storage should be enough for any key supported by ZFS */
1102 CLASSERT(sizeof(za->za_name) <= sizeof(it->ozi_name));
1105 * the first ->next() moves the cursor to .
1106 * the second ->next() moves the cursor to ..
1107 * then we get to the real records and have to verify any exist
1109 if (it->ozi_pos <= 2) {
1111 if (it->ozi_pos <=2)
1115 zap_cursor_advance(it->ozi_zc);
1119 * According to current API we need to return error if its last entry.
1120 * zap_cursor_advance() does not return any value. So we need to call
1121 * retrieve to check if there is any record. We should make
1122 * changes to Iterator API to not return status for this API
1124 rc = osd_index_retrieve_skip_dots(it, za);
1126 if (rc == -ENOENT) /* end of dir */
1132 static struct dt_key *osd_dir_it_key(const struct lu_env *env,
1133 const struct dt_it *di)
1135 struct osd_zap_it *it = (struct osd_zap_it *)di;
1136 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1140 if (it->ozi_pos <= 1) {
1142 RETURN((struct dt_key *)".");
1143 } else if (it->ozi_pos == 2) {
1144 RETURN((struct dt_key *)"..");
1147 if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)))
1148 RETURN(ERR_PTR(rc));
1150 strcpy(it->ozi_name, za->za_name);
1152 RETURN((struct dt_key *)it->ozi_name);
1155 static int osd_dir_it_key_size(const struct lu_env *env, const struct dt_it *di)
1157 struct osd_zap_it *it = (struct osd_zap_it *)di;
1158 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1162 if (it->ozi_pos <= 1) {
1165 } else if (it->ozi_pos == 2) {
1169 if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)) == 0)
1170 rc = strlen(za->za_name);
1176 osd_dirent_update(const struct lu_env *env, struct osd_device *dev,
1177 uint64_t zap, const char *key, struct luz_direntry *zde)
1183 tx = dmu_tx_create(dev->od_os);
1187 dmu_tx_hold_zap(tx, zap, TRUE, NULL);
1188 rc = -dmu_tx_assign(tx, TXG_WAIT);
1190 rc = -zap_update(dev->od_os, zap, key, 8, sizeof(*zde) / 8,
1191 (const void *)zde, tx);
1200 static int osd_update_entry_for_agent(const struct lu_env *env,
1201 struct osd_device *osd,
1202 uint64_t zap, const char *name,
1203 struct luz_direntry *zde, __u32 attr)
1205 dmu_tx_t *tx = NULL;
1209 if (attr & LUDA_VERIFY_DRYRUN)
1212 tx = dmu_tx_create(osd->od_os);
1214 GOTO(out, rc = -ENOMEM);
1216 dmu_tx_hold_sa_create(tx, osd_find_dnsize(osd, OSD_BASE_EA_IN_BONUS));
1217 dmu_tx_hold_zap(tx, zap, FALSE, NULL);
1218 rc = -dmu_tx_assign(tx, TXG_WAIT);
1224 rc = osd_create_agent_object(env, osd, zde, zap, tx);
1226 rc = -zap_update(osd->od_os, zap, name, 8, sizeof(*zde) / 8,
1227 (const void *)zde, tx);
1233 CDEBUG(D_LFSCK, "%s: Updated (%s) remote entry for "DFID": rc = %d\n",
1234 osd_name(osd), (attr & LUDA_VERIFY_DRYRUN) ? "(ro)" : "(rw)",
1235 PFID(&zde->lzd_fid), rc);
1239 static int osd_dir_it_rec(const struct lu_env *env, const struct dt_it *di,
1240 struct dt_rec *dtrec, __u32 attr)
1242 struct osd_zap_it *it = (struct osd_zap_it *)di;
1243 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
1244 struct osd_thread_info *info = osd_oti_get(env);
1245 struct luz_direntry *zde = &info->oti_zde;
1246 zap_attribute_t *za = &info->oti_za;
1247 struct lu_fid *fid = &info->oti_fid;
1248 struct osd_device *osd = osd_obj2dev(it->ozi_obj);
1253 if (it->ozi_pos <= 1) {
1254 lde->lde_hash = cpu_to_le64(1);
1255 strcpy(lde->lde_name, ".");
1256 lde->lde_namelen = cpu_to_le16(1);
1257 fid_cpu_to_le(&lde->lde_fid,
1258 lu_object_fid(&it->ozi_obj->oo_dt.do_lu));
1259 lde->lde_attrs = LUDA_FID;
1260 /* append lustre attributes */
1261 osd_it_append_attrs(lde, attr, 1, IFTODT(S_IFDIR));
1262 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(1, attr));
1265 } else if (it->ozi_pos == 2) {
1266 lde->lde_hash = cpu_to_le64(2);
1267 strcpy(lde->lde_name, "..");
1268 lde->lde_namelen = cpu_to_le16(2);
1269 rc = osd_find_parent_fid(env, &it->ozi_obj->oo_dt, fid);
1271 fid_cpu_to_le(&lde->lde_fid, fid);
1272 lde->lde_attrs = LUDA_FID;
1273 } else if (rc != -ENOENT) {
1274 /* ENOENT happens at the root of filesystem, ignore */
1278 /* append lustre attributes */
1279 osd_it_append_attrs(lde, attr, 2, IFTODT(S_IFDIR));
1280 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(2, attr));
1286 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1290 lde->lde_hash = cpu_to_le64(osd_zap_cursor_serialize(it->ozi_zc));
1291 namelen = strlen(za->za_name);
1292 if (namelen > NAME_MAX)
1294 strcpy(lde->lde_name, za->za_name);
1295 lde->lde_namelen = cpu_to_le16(namelen);
1297 if (za->za_integer_length != 8) {
1298 CERROR("%s: unsupported direntry format: %d %d\n",
1300 za->za_integer_length, (int)za->za_num_integers);
1304 rc = osd_zap_lookup(osd, it->ozi_zc->zc_zapobj, it->ozi_obj->oo_dn,
1305 za->za_name, za->za_integer_length, 3, zde);
1309 if (za->za_num_integers >= 3 && fid_is_sane(&zde->lzd_fid)) {
1310 lde->lde_attrs = LUDA_FID;
1311 fid_cpu_to_le(&lde->lde_fid, &zde->lzd_fid);
1312 if (unlikely(zde->lzd_reg.zde_dnode == ZFS_NO_OBJECT &&
1313 osd_remote_fid(env, osd, &zde->lzd_fid) > 0 &&
1314 attr & LUDA_VERIFY)) {
1315 /* It is mainly used for handling the MDT
1316 * upgraded from old ZFS based backend. */
1317 rc = osd_update_entry_for_agent(env, osd,
1318 it->ozi_obj->oo_dn->dn_object,
1319 za->za_name, zde, attr);
1321 lde->lde_attrs |= LUDA_REPAIR;
1323 lde->lde_attrs |= LUDA_UNKNOWN;
1326 GOTO(pack_attr, rc = 0);
1329 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
1332 rc = osd_get_fid_by_oid(env, osd, zde->lzd_reg.zde_dnode, fid);
1334 lde->lde_attrs = LUDA_UNKNOWN;
1335 GOTO(pack_attr, rc = 0);
1338 if (!(attr & LUDA_VERIFY)) {
1339 fid_cpu_to_le(&lde->lde_fid, fid);
1340 lde->lde_attrs = LUDA_FID;
1341 GOTO(pack_attr, rc = 0);
1344 if (attr & LUDA_VERIFY_DRYRUN) {
1345 fid_cpu_to_le(&lde->lde_fid, fid);
1346 lde->lde_attrs = LUDA_FID | LUDA_REPAIR;
1347 GOTO(pack_attr, rc = 0);
1350 fid_cpu_to_le(&lde->lde_fid, fid);
1351 lde->lde_attrs = LUDA_FID;
1352 zde->lzd_fid = *fid;
1353 rc = osd_dirent_update(env, osd, it->ozi_zc->zc_zapobj,
1356 lde->lde_attrs |= LUDA_UNKNOWN;
1357 GOTO(pack_attr, rc = 0);
1360 lde->lde_attrs |= LUDA_REPAIR;
1362 GOTO(pack_attr, rc = 0);
1365 osd_it_append_attrs(lde, attr, namelen, zde->lzd_reg.zde_type);
1366 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
1370 static int osd_dir_it_rec_size(const struct lu_env *env, const struct dt_it *di,
1373 struct osd_zap_it *it = (struct osd_zap_it *)di;
1374 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1379 if (it->ozi_pos <= 1)
1381 else if (it->ozi_pos == 2)
1385 rc = lu_dirent_calc_size(namelen, attr);
1389 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1390 if (unlikely(rc != 0))
1393 if (za->za_integer_length != 8 || za->za_num_integers < 3) {
1394 CERROR("%s: unsupported direntry format: %d %d\n",
1395 osd_obj2dev(it->ozi_obj)->od_svname,
1396 za->za_integer_length, (int)za->za_num_integers);
1400 namelen = strlen(za->za_name);
1401 if (namelen > NAME_MAX)
1404 rc = lu_dirent_calc_size(namelen, attr);
1409 static __u64 osd_dir_it_store(const struct lu_env *env, const struct dt_it *di)
1411 struct osd_zap_it *it = (struct osd_zap_it *)di;
1415 if (it->ozi_pos <= 2)
1418 pos = osd_zap_cursor_serialize(it->ozi_zc);
1425 * rc == 0 -> end of directory.
1426 * rc > 0 -> ok, proceed.
1427 * rc < 0 -> error. ( EOVERFLOW can be masked.)
1429 static int osd_dir_it_load(const struct lu_env *env,
1430 const struct dt_it *di, __u64 hash)
1432 struct osd_zap_it *it = (struct osd_zap_it *)di;
1433 struct osd_object *obj = it->ozi_obj;
1434 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1438 /* reset the cursor */
1439 zap_cursor_fini(it->ozi_zc);
1440 osd_obj_cursor_init_serialized(it->ozi_zc, obj, hash);
1447 /* to return whether the end has been reached */
1448 rc = osd_index_retrieve_skip_dots(it, za);
1451 else if (rc == -ENOENT)
1458 struct dt_index_operations osd_dir_ops = {
1459 .dio_lookup = osd_dir_lookup,
1460 .dio_declare_insert = osd_declare_dir_insert,
1461 .dio_insert = osd_dir_insert,
1462 .dio_declare_delete = osd_declare_dir_delete,
1463 .dio_delete = osd_dir_delete,
1465 .init = osd_dir_it_init,
1466 .fini = osd_index_it_fini,
1467 .get = osd_dir_it_get,
1468 .put = osd_dir_it_put,
1469 .next = osd_dir_it_next,
1470 .key = osd_dir_it_key,
1471 .key_size = osd_dir_it_key_size,
1472 .rec = osd_dir_it_rec,
1473 .rec_size = osd_dir_it_rec_size,
1474 .store = osd_dir_it_store,
1475 .load = osd_dir_it_load
1480 * Primitives for index files using binary keys.
1483 /* key integer_size is 8 */
1484 static int osd_prepare_key_uint64(struct osd_object *o, __u64 *dst,
1485 const struct dt_key *src)
1492 /* align keysize to 64bit */
1493 size = (o->oo_keysize + sizeof(__u64) - 1) / sizeof(__u64);
1494 size *= sizeof(__u64);
1496 LASSERT(size <= MAXNAMELEN);
1498 if (unlikely(size > o->oo_keysize))
1499 memset(dst + o->oo_keysize, 0, size - o->oo_keysize);
1500 memcpy(dst, (const char *)src, o->oo_keysize);
1502 return (size/sizeof(__u64));
1505 static int osd_index_lookup(const struct lu_env *env, struct dt_object *dt,
1506 struct dt_rec *rec, const struct dt_key *key)
1508 struct osd_object *obj = osd_dt_obj(dt);
1509 struct osd_device *osd = osd_obj2dev(obj);
1510 __u64 *k = osd_oti_get(env)->oti_key64;
1514 rc = osd_prepare_key_uint64(obj, k, key);
1516 rc = -zap_lookup_uint64(osd->od_os, obj->oo_dn->dn_object,
1517 k, rc, obj->oo_recusize, obj->oo_recsize,
1519 RETURN(rc == 0 ? 1 : rc);
1522 static int osd_declare_index_insert(const struct lu_env *env,
1523 struct dt_object *dt,
1524 const struct dt_rec *rec,
1525 const struct dt_key *key,
1528 struct osd_object *obj = osd_dt_obj(dt);
1529 struct osd_thandle *oh;
1532 LASSERT(th != NULL);
1533 oh = container_of0(th, struct osd_thandle, ot_super);
1535 LASSERT(obj->oo_dn);
1537 /* do not specify the key as then DMU is trying to look it up
1538 * which is very expensive. usually the layers above lookup
1539 * before insertion */
1540 osd_tx_hold_zap(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
1546 static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
1547 const struct dt_rec *rec, const struct dt_key *key,
1548 struct thandle *th, int ignore_quota)
1550 struct osd_object *obj = osd_dt_obj(dt);
1551 struct osd_device *osd = osd_obj2dev(obj);
1552 struct osd_thandle *oh;
1553 __u64 *k = osd_oti_get(env)->oti_key64;
1557 LASSERT(obj->oo_dn);
1558 LASSERT(dt_object_exists(dt));
1559 LASSERT(osd_invariant(obj));
1560 LASSERT(th != NULL);
1562 oh = container_of0(th, struct osd_thandle, ot_super);
1564 rc = osd_prepare_key_uint64(obj, k, key);
1566 /* Insert (key,oid) into ZAP */
1567 rc = -zap_add_uint64(osd->od_os, obj->oo_dn->dn_object,
1568 k, rc, obj->oo_recusize, obj->oo_recsize,
1569 (void *)rec, oh->ot_tx);
1573 static int osd_declare_index_delete(const struct lu_env *env,
1574 struct dt_object *dt,
1575 const struct dt_key *key,
1578 struct osd_object *obj = osd_dt_obj(dt);
1579 struct osd_thandle *oh;
1582 LASSERT(dt_object_exists(dt));
1583 LASSERT(osd_invariant(obj));
1584 LASSERT(th != NULL);
1585 LASSERT(obj->oo_dn);
1587 oh = container_of0(th, struct osd_thandle, ot_super);
1589 /* do not specify the key as then DMU is trying to look it up
1590 * which is very expensive. usually the layers above lookup
1591 * before deletion */
1592 osd_tx_hold_zap(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
1598 static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
1599 const struct dt_key *key, struct thandle *th)
1601 struct osd_object *obj = osd_dt_obj(dt);
1602 struct osd_device *osd = osd_obj2dev(obj);
1603 struct osd_thandle *oh;
1604 __u64 *k = osd_oti_get(env)->oti_key64;
1608 LASSERT(obj->oo_dn);
1609 LASSERT(th != NULL);
1610 oh = container_of0(th, struct osd_thandle, ot_super);
1612 rc = osd_prepare_key_uint64(obj, k, key);
1614 /* Remove binary key from the ZAP */
1615 rc = -zap_remove_uint64(osd->od_os, obj->oo_dn->dn_object,
1620 static int osd_index_it_get(const struct lu_env *env, struct dt_it *di,
1621 const struct dt_key *key)
1623 struct osd_zap_it *it = (struct osd_zap_it *)di;
1624 struct osd_object *obj = it->ozi_obj;
1625 struct osd_device *osd = osd_obj2dev(obj);
1629 LASSERT(it->ozi_zc);
1632 * XXX: we need a binary version of zap_cursor_move_to_key()
1633 * to implement this API */
1634 if (*((const __u64 *)key) != 0)
1635 CERROR("NOT IMPLEMETED YET (move to %#llx)\n",
1638 zap_cursor_fini(it->ozi_zc);
1639 zap_cursor_init(it->ozi_zc, osd->od_os, obj->oo_dn->dn_object);
1645 static int osd_index_it_next(const struct lu_env *env, struct dt_it *di)
1647 struct osd_zap_it *it = (struct osd_zap_it *)di;
1648 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1652 if (it->ozi_reset == 0)
1653 zap_cursor_advance(it->ozi_zc);
1657 * According to current API we need to return error if it's last entry.
1658 * zap_cursor_advance() does not return any value. So we need to call
1659 * retrieve to check if there is any record. We should make
1660 * changes to Iterator API to not return status for this API
1662 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1669 static struct dt_key *osd_index_it_key(const struct lu_env *env,
1670 const struct dt_it *di)
1672 struct osd_zap_it *it = (struct osd_zap_it *)di;
1673 struct osd_object *obj = it->ozi_obj;
1674 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1679 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1681 RETURN(ERR_PTR(rc));
1683 /* the binary key is stored in the name */
1684 memcpy(&it->ozi_key, za->za_name, obj->oo_keysize);
1686 RETURN((struct dt_key *)&it->ozi_key);
1689 static int osd_index_it_key_size(const struct lu_env *env,
1690 const struct dt_it *di)
1692 struct osd_zap_it *it = (struct osd_zap_it *)di;
1693 struct osd_object *obj = it->ozi_obj;
1694 RETURN(obj->oo_keysize);
1697 static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
1698 struct dt_rec *rec, __u32 attr)
1700 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1701 struct osd_zap_it *it = (struct osd_zap_it *)di;
1702 struct osd_object *obj = it->ozi_obj;
1703 struct osd_device *osd = osd_obj2dev(obj);
1704 __u64 *k = osd_oti_get(env)->oti_key64;
1709 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1713 rc = osd_prepare_key_uint64(obj, k, (const struct dt_key *)za->za_name);
1715 rc = -zap_lookup_uint64(osd->od_os, obj->oo_dn->dn_object,
1716 k, rc, obj->oo_recusize, obj->oo_recsize,
1721 static __u64 osd_index_it_store(const struct lu_env *env,
1722 const struct dt_it *di)
1724 struct osd_zap_it *it = (struct osd_zap_it *)di;
1727 RETURN((__u64)zap_cursor_serialize(it->ozi_zc));
1730 static int osd_index_it_load(const struct lu_env *env, const struct dt_it *di,
1733 struct osd_zap_it *it = (struct osd_zap_it *)di;
1734 struct osd_object *obj = it->ozi_obj;
1735 struct osd_device *osd = osd_obj2dev(obj);
1736 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1740 /* reset the cursor */
1741 zap_cursor_fini(it->ozi_zc);
1742 zap_cursor_init_serialized(it->ozi_zc, osd->od_os,
1743 obj->oo_dn->dn_object, hash);
1746 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1749 else if (rc == -ENOENT)
1755 static struct dt_index_operations osd_index_ops = {
1756 .dio_lookup = osd_index_lookup,
1757 .dio_declare_insert = osd_declare_index_insert,
1758 .dio_insert = osd_index_insert,
1759 .dio_declare_delete = osd_declare_index_delete,
1760 .dio_delete = osd_index_delete,
1762 .init = osd_index_it_init,
1763 .fini = osd_index_it_fini,
1764 .get = osd_index_it_get,
1765 .put = osd_index_it_put,
1766 .next = osd_index_it_next,
1767 .key = osd_index_it_key,
1768 .key_size = osd_index_it_key_size,
1769 .rec = osd_index_it_rec,
1770 .store = osd_index_it_store,
1771 .load = osd_index_it_load
1775 struct osd_metadnode_it {
1776 struct osd_device *mit_dev;
1778 struct lu_fid mit_fid;
1780 __u64 mit_prefetched_dnode;
1783 static struct dt_it *osd_zfs_otable_it_init(const struct lu_env *env,
1784 struct dt_object *dt, __u32 attr)
1786 struct osd_device *dev = osd_dev(dt->do_lu.lo_dev);
1787 struct osd_metadnode_it *it;
1791 if (unlikely(it == NULL))
1792 RETURN(ERR_PTR(-ENOMEM));
1796 /* XXX: dmu_object_next() does NOT find dnodes allocated
1797 * in the current non-committed txg, so we force txg
1798 * commit to find all existing dnodes ... */
1799 if (!dev->od_dt_dev.dd_rdonly)
1800 txg_wait_synced(dmu_objset_pool(dev->od_os), 0ULL);
1802 RETURN((struct dt_it *)it);
1805 static void osd_zfs_otable_it_fini(const struct lu_env *env, struct dt_it *di)
1807 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1812 static int osd_zfs_otable_it_get(const struct lu_env *env,
1813 struct dt_it *di, const struct dt_key *key)
1818 static void osd_zfs_otable_it_put(const struct lu_env *env, struct dt_it *di)
1822 #define OTABLE_PREFETCH 256
1824 static void osd_zfs_otable_prefetch(const struct lu_env *env,
1825 struct osd_metadnode_it *it)
1827 struct osd_device *dev = it->mit_dev;
1830 /* can go negative on the very first access to the iterator
1831 * or if some non-Lustre objects were found */
1832 if (unlikely(it->mit_prefetched < 0))
1833 it->mit_prefetched = 0;
1835 if (it->mit_prefetched >= (OTABLE_PREFETCH >> 1))
1838 if (it->mit_prefetched_dnode == 0)
1839 it->mit_prefetched_dnode = it->mit_pos;
1841 while (it->mit_prefetched < OTABLE_PREFETCH) {
1842 rc = -dmu_object_next(dev->od_os, &it->mit_prefetched_dnode,
1844 if (unlikely(rc != 0))
1847 osd_dmu_prefetch(dev->od_os, it->mit_prefetched_dnode,
1848 0, 0, 0, ZIO_PRIORITY_ASYNC_READ);
1850 it->mit_prefetched++;
1854 static int osd_zfs_otable_it_next(const struct lu_env *env, struct dt_it *di)
1856 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1857 struct lustre_mdt_attrs *lma;
1858 struct osd_device *dev = it->mit_dev;
1859 nvlist_t *nvbuf = NULL;
1864 memset(&it->mit_fid, 0, sizeof(it->mit_fid));
1866 dnode = it->mit_pos;
1868 rc = -dmu_object_next(dev->od_os, &it->mit_pos, B_FALSE, 0);
1869 if (unlikely(rc != 0))
1871 it->mit_prefetched--;
1873 /* LMA is required for this to be a Lustre object.
1874 * If there is no xattr skip it. */
1875 rc = __osd_xattr_load_by_oid(dev, it->mit_pos, &nvbuf);
1876 if (unlikely(rc != 0))
1879 LASSERT(nvbuf != NULL);
1880 rc = -nvlist_lookup_byte_array(nvbuf, XATTR_NAME_LMA, &v, &s);
1881 if (likely(rc == 0)) {
1883 lma = (struct lustre_mdt_attrs *)v;
1884 lustre_lma_swab(lma);
1885 if (likely(!(lma->lma_compat & LMAC_NOT_IN_OI) &&
1886 !(lma->lma_incompat & LMAI_AGENT))) {
1887 it->mit_fid = lma->lma_self_fid;
1893 /* not a Lustre visible object, try next one */
1898 /* we aren't prefetching in the above loop because the number of
1899 * non-Lustre objects is very small and we will be repeating very
1900 * rare. in case we want to use this to iterate over non-Lustre
1901 * objects (i.e. when we convert regular ZFS in Lustre) it makes
1902 * sense to initiate prefetching in the loop */
1904 /* 0 - there are more items, +1 - the end */
1905 if (likely(rc == 0))
1906 osd_zfs_otable_prefetch(env, it);
1908 CDEBUG(D_OTHER, "advance: %llu -> %llu "DFID": %d\n", dnode,
1909 it->mit_pos, PFID(&it->mit_fid), rc);
1915 static struct dt_key *osd_zfs_otable_it_key(const struct lu_env *env,
1916 const struct dt_it *di)
1921 static int osd_zfs_otable_it_key_size(const struct lu_env *env,
1922 const struct dt_it *di)
1924 return sizeof(__u64);
1927 static int osd_zfs_otable_it_rec(const struct lu_env *env,
1928 const struct dt_it *di,
1929 struct dt_rec *rec, __u32 attr)
1931 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1932 struct lu_fid *fid = (struct lu_fid *)rec;
1941 static __u64 osd_zfs_otable_it_store(const struct lu_env *env,
1942 const struct dt_it *di)
1944 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1949 static int osd_zfs_otable_it_load(const struct lu_env *env,
1950 const struct dt_it *di, __u64 hash)
1952 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1955 it->mit_prefetched = 0;
1956 it->mit_prefetched_dnode = 0;
1958 return osd_zfs_otable_it_next(env, (struct dt_it *)di);
1961 static int osd_zfs_otable_it_key_rec(const struct lu_env *env,
1962 const struct dt_it *di, void *key_rec)
1967 const struct dt_index_operations osd_zfs_otable_ops = {
1969 .init = osd_zfs_otable_it_init,
1970 .fini = osd_zfs_otable_it_fini,
1971 .get = osd_zfs_otable_it_get,
1972 .put = osd_zfs_otable_it_put,
1973 .next = osd_zfs_otable_it_next,
1974 .key = osd_zfs_otable_it_key,
1975 .key_size = osd_zfs_otable_it_key_size,
1976 .rec = osd_zfs_otable_it_rec,
1977 .store = osd_zfs_otable_it_store,
1978 .load = osd_zfs_otable_it_load,
1979 .key_rec = osd_zfs_otable_it_key_rec,
1983 int osd_index_try(const struct lu_env *env, struct dt_object *dt,
1984 const struct dt_index_features *feat)
1986 struct osd_object *obj = osd_dt_obj(dt);
1990 down_read(&obj->oo_guard);
1993 * XXX: implement support for fixed-size keys sorted with natural
1994 * numerical way (not using internal hash value)
1996 if (feat->dif_flags & DT_IND_RANGE)
1997 GOTO(out, rc = -ERANGE);
1999 if (unlikely(feat == &dt_otable_features)) {
2000 dt->do_index_ops = &osd_zfs_otable_ops;
2004 LASSERT(!dt_object_exists(dt) || obj->oo_dn != NULL);
2005 if (likely(feat == &dt_directory_features)) {
2006 if (!dt_object_exists(dt) || osd_object_is_zap(obj->oo_dn))
2007 dt->do_index_ops = &osd_dir_ops;
2009 GOTO(out, rc = -ENOTDIR);
2010 } else if (unlikely(feat == &dt_acct_features)) {
2011 LASSERT(fid_is_acct(lu_object_fid(&dt->do_lu)));
2012 dt->do_index_ops = &osd_acct_index_ops;
2013 } else if (dt->do_index_ops == NULL) {
2014 /* For index file, we don't support variable key & record sizes
2015 * and the key has to be unique */
2016 if ((feat->dif_flags & ~DT_IND_UPDATE) != 0)
2017 GOTO(out, rc = -EINVAL);
2019 if (feat->dif_keysize_max > ZAP_MAXNAMELEN)
2020 GOTO(out, rc = -E2BIG);
2021 if (feat->dif_keysize_max != feat->dif_keysize_min)
2022 GOTO(out, rc = -EINVAL);
2024 /* As for the record size, it should be a multiple of 8 bytes
2025 * and smaller than the maximum value length supported by ZAP.
2027 if (feat->dif_recsize_max > ZAP_MAXVALUELEN)
2028 GOTO(out, rc = -E2BIG);
2029 if (feat->dif_recsize_max != feat->dif_recsize_min)
2030 GOTO(out, rc = -EINVAL);
2032 obj->oo_keysize = feat->dif_keysize_max;
2033 obj->oo_recsize = feat->dif_recsize_max;
2034 obj->oo_recusize = 1;
2036 /* ZFS prefers to work with array of 64bits */
2037 if ((obj->oo_recsize & 7) == 0) {
2038 obj->oo_recsize >>= 3;
2039 obj->oo_recusize = 8;
2041 dt->do_index_ops = &osd_index_ops;
2045 up_read(&obj->oo_guard);