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/
31 * lustre/osd-zfs/osd_index.c
33 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
34 * Author: Mike Pershin <tappro@whamcloud.com>
37 #define DEBUG_SUBSYSTEM S_OSD
39 #include <libcfs/libcfs.h>
40 #include <obd_support.h>
41 #include <lustre_net.h>
43 #include <obd_class.h>
44 #include <lustre_disk.h>
45 #include <lustre_fid.h>
47 #include "osd_internal.h"
49 #include <sys/dnode.h>
53 #include <sys/spa_impl.h>
54 #include <sys/zfs_znode.h>
55 #include <sys/dmu_tx.h>
56 #include <sys/dmu_objset.h>
57 #include <sys/dsl_prop.h>
58 #include <sys/sa_impl.h>
60 #include <lustre_scrub.h>
62 /* We don't actually have direct access to the zap_hashbits() function
63 * so just pretend like we do for now. If this ever breaks we can look at
65 #define zap_hashbits(zc) 48
68 * | cd (16 bits) | hash (48 bits) |
69 * we need it in other form:
70 * |0| hash (48 bit) | cd (15 bit) |
71 * to be a full 64-bit ordered hash so that Lustre readdir can use it to merge
72 * the readdir hashes from multiple directory stripes uniformly on the client.
73 * Another point is sign bit, the hash range should be in [0, 2^63-1] because
74 * loff_t (for llseek) needs to be a positive value. This means the "cd" field
75 * should only be the low 15 bits.
77 uint64_t osd_zap_cursor_serialize(zap_cursor_t *zc)
79 uint64_t zfs_hash = zap_cursor_serialize(zc) & (~0ULL >> 1);
81 return (zfs_hash >> zap_hashbits(zc)) |
82 (zfs_hash << (63 - zap_hashbits(zc)));
85 void osd_zap_cursor_init_serialized(zap_cursor_t *zc, struct objset *os,
86 uint64_t id, uint64_t dirhash)
88 uint64_t zfs_hash = ((dirhash << zap_hashbits(zc)) & (~0ULL >> 1)) |
89 (dirhash >> (63 - zap_hashbits(zc)));
91 zap_cursor_init_serialized(zc, os, id, zfs_hash);
94 int osd_zap_cursor_init(zap_cursor_t **zc, struct objset *os,
95 uint64_t id, uint64_t dirhash)
100 if (unlikely(t == NULL))
103 osd_zap_cursor_init_serialized(t, os, id, dirhash);
109 void osd_zap_cursor_fini(zap_cursor_t *zc)
115 static inline void osd_obj_cursor_init_serialized(zap_cursor_t *zc,
116 struct osd_object *o,
119 struct osd_device *d = osd_obj2dev(o);
120 osd_zap_cursor_init_serialized(zc, d->od_os,
121 o->oo_dn->dn_object, dirhash);
124 static inline int osd_obj_cursor_init(zap_cursor_t **zc, struct osd_object *o,
127 struct osd_device *d = osd_obj2dev(o);
128 return osd_zap_cursor_init(zc, d->od_os, o->oo_dn->dn_object, dirhash);
131 static struct dt_it *osd_index_it_init(const struct lu_env *env,
132 struct dt_object *dt,
135 struct osd_thread_info *info = osd_oti_get(env);
136 struct osd_zap_it *it;
137 struct osd_object *obj = osd_dt_obj(dt);
138 struct lu_object *lo = &dt->do_lu;
142 if (obj->oo_destroyed)
143 RETURN(ERR_PTR(-ENOENT));
145 LASSERT(lu_object_exists(lo));
149 OBD_SLAB_ALLOC_PTR_GFP(it, osd_zapit_cachep, GFP_NOFS);
151 RETURN(ERR_PTR(-ENOMEM));
153 rc = osd_obj_cursor_init(&it->ozi_zc, obj, 0);
155 OBD_SLAB_FREE_PTR(it, osd_zapit_cachep);
163 RETURN((struct dt_it *)it);
166 static void osd_index_it_fini(const struct lu_env *env, struct dt_it *di)
168 struct osd_zap_it *it = (struct osd_zap_it *)di;
169 struct osd_object *obj;
173 LASSERT(it->ozi_obj);
177 osd_zap_cursor_fini(it->ozi_zc);
178 osd_object_put(env, obj);
179 OBD_SLAB_FREE_PTR(it, osd_zapit_cachep);
185 static void osd_index_it_put(const struct lu_env *env, struct dt_it *di)
187 /* PBS: do nothing : ref are incremented at retrive and decreamented
191 static inline void osd_it_append_attrs(struct lu_dirent *ent, __u32 attr,
194 const unsigned align = sizeof(struct luda_type) - 1;
195 struct luda_type *lt;
197 /* check if file type is required */
198 if (attr & LUDA_TYPE) {
199 len = (len + align) & ~align;
201 lt = (void *)ent->lde_name + len;
202 lt->lt_type = cpu_to_le16(DTTOIF(type));
203 ent->lde_attrs |= LUDA_TYPE;
206 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
209 int __osd_xattr_load_by_oid(struct osd_device *osd, uint64_t oid, nvlist_t **sa)
215 rc = -dmu_bonus_hold(osd->od_os, oid, osd_obj_tag, &db);
217 CERROR("%s: can't get bonus, rc = %d\n", osd->od_svname, rc);
221 rc = -sa_handle_get_from_db(osd->od_os, db, NULL, SA_HDL_PRIVATE, &hdl);
223 dmu_buf_rele(db, osd_obj_tag);
227 rc = __osd_xattr_load(osd, hdl, sa);
229 sa_handle_destroy(hdl);
234 * Get the object's FID from its LMA EA.
236 * \param[in] env pointer to the thread context
237 * \param[in] osd pointer to the OSD device
238 * \param[in] oid the object's local identifier
239 * \param[out] fid the buffer to hold the object's FID
241 * \retval 0 for success
242 * \retval negative error number on failure
244 int osd_get_fid_by_oid(const struct lu_env *env, struct osd_device *osd,
245 uint64_t oid, struct lu_fid *fid)
247 struct objset *os = osd->od_os;
248 struct osd_thread_info *oti = osd_oti_get(env);
249 struct lustre_mdt_attrs *lma =
250 (struct lustre_mdt_attrs *)oti->oti_buf;
252 nvlist_t *sa_xattr = NULL;
253 sa_handle_t *sa_hdl = NULL;
254 uchar_t *nv_value = NULL;
255 uint64_t xattr = ZFS_NO_OBJECT;
260 rc = __osd_xattr_load_by_oid(osd, oid, &sa_xattr);
267 rc = -nvlist_lookup_byte_array(sa_xattr, XATTR_NAME_LMA, &nv_value,
275 if (unlikely(size > sizeof(oti->oti_buf)))
276 GOTO(out, rc = -ERANGE);
278 memcpy(lma, nv_value, size);
283 rc = -sa_handle_get(os, oid, NULL, SA_HDL_PRIVATE, &sa_hdl);
287 rc = -sa_lookup(sa_hdl, SA_ZPL_XATTR(osd), &xattr, 8);
288 sa_handle_destroy(sa_hdl);
293 buf.lb_len = sizeof(oti->oti_buf);
294 rc = __osd_xattr_get_large(env, osd, xattr, &buf,
295 XATTR_NAME_LMA, &size);
300 if (size < sizeof(*lma))
301 GOTO(out, rc = -EIO);
303 lustre_lma_swab(lma);
304 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
305 CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
306 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
307 "oid = %#llx\n", osd->od_svname,
308 lma->lma_incompat & ~LMA_INCOMPAT_SUPP, oid);
309 GOTO(out, rc = -EOPNOTSUPP);
311 *fid = lma->lma_self_fid;
316 if (sa_xattr != NULL)
317 nvlist_free(sa_xattr);
322 * As we don't know FID, we can't use LU object, so this function
323 * partially duplicate osd_xattr_get_internal() which is built around
324 * LU-object and uses it to cache data like regular EA dnode, etc
326 static int osd_find_parent_by_dnode(const struct lu_env *env,
328 struct lu_fid *fid, uint64_t *oid)
330 struct osd_object *obj = osd_dt_obj(o);
331 struct osd_device *osd = osd_obj2dev(obj);
332 uint64_t dnode = ZFS_NO_OBJECT;
336 /* first of all, get parent dnode from own attributes */
337 rc = osd_sa_handle_get(obj);
340 rc = -sa_lookup(obj->oo_sa_hdl, SA_ZPL_PARENT(osd), &dnode, 8);
344 rc = osd_get_fid_by_oid(env, osd, dnode, fid);
350 static int osd_find_parent_fid(const struct lu_env *env, struct dt_object *o,
351 struct lu_fid *fid, uint64_t *oid)
353 struct link_ea_header *leh;
354 struct link_ea_entry *lee;
359 buf.lb_buf = osd_oti_get(env)->oti_buf;
360 buf.lb_len = sizeof(osd_oti_get(env)->oti_buf);
362 rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK);
364 rc = osd_xattr_get(env, o, &LU_BUF_NULL, XATTR_NAME_LINK);
368 OBD_ALLOC(buf.lb_buf, rc);
369 if (buf.lb_buf == NULL)
372 rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK);
376 if (rc < sizeof(*leh) + sizeof(*lee))
377 GOTO(out, rc = -EINVAL);
380 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
381 leh->leh_magic = LINK_EA_MAGIC;
382 leh->leh_reccount = __swab32(leh->leh_reccount);
383 leh->leh_len = __swab64(leh->leh_len);
385 if (leh->leh_magic != LINK_EA_MAGIC)
386 GOTO(out, rc = -EINVAL);
387 if (leh->leh_reccount == 0)
388 GOTO(out, rc = -ENODATA);
390 lee = (struct link_ea_entry *)(leh + 1);
391 fid_be_to_cpu(fid, (const struct lu_fid *)&lee->lee_parent_fid);
395 if (buf.lb_buf != osd_oti_get(env)->oti_buf)
396 OBD_FREE(buf.lb_buf, buf.lb_len);
399 /* this block can be enabled for additional verification
400 * it's trying to match FID from LinkEA vs. FID from LMA */
404 rc2 = osd_find_parent_by_dnode(env, o, &fid2, oid);
406 if (lu_fid_eq(fid, &fid2) == 0)
407 CERROR("wrong parent: "DFID" != "DFID"\n",
408 PFID(fid), PFID(&fid2));
412 /* no LinkEA is found, let's try to find the fid in parent's LMA */
413 if (unlikely(rc != 0))
414 rc = osd_find_parent_by_dnode(env, o, fid, oid);
420 * When lookup item under striped directory, we need to locate the master
421 * MDT-object of the striped directory firstly, then the client will send
422 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
423 * and the item's name. If the system is restored from MDT file level backup,
424 * then before the OI scrub completely built the OI files, the OI mappings of
425 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
426 * not a problem for the master MDT-object. Because when locate the master
427 * MDT-object, we will do name based lookup (for the striped directory itself)
428 * firstly, during such process we can setup the correct OI mapping for the
429 * master MDT-object. But it will be trouble for the slave MDT-object. Because
430 * the client will not trigger name based lookup on the MDT to locate the slave
431 * MDT-object before locating item under the striped directory, then when
432 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
433 * is invalid and does not know what the right OI mapping is, then the MDT has
434 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
435 * the OI file, related OI mapping is unknown yet, please try again later. And
436 * then client will re-try the RPC again and again until related OI mapping has
437 * been updated. That is quite inefficient.
439 * To resolve above trouble, we will handle it as the following two cases:
441 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
442 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
443 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
444 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
445 * directly. Please check osd_fid_lookup().
447 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
448 * Under such case, during lookup the master MDT-object, we will lookup the
449 * slave MDT-object via readdir against the master MDT-object, because the
450 * slave MDT-objects information are stored as sub-directories with the name
451 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
452 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
453 * the correct OI mapping for the slave MDT-object.
455 static int osd_check_lmv(const struct lu_env *env, struct osd_device *osd,
456 uint64_t oid, const struct lu_fid *fid)
458 struct osd_thread_info *info = osd_oti_get(env);
459 struct luz_direntry *zde = &info->oti_zde;
460 zap_attribute_t *za = &info->oti_za;
461 zap_cursor_t *zc = &info->oti_zc;
462 struct lu_fid *tfid = &info->oti_fid;
463 nvlist_t *nvbuf = NULL;
464 struct lmv_mds_md_v1 *lmv = NULL;
469 rc = __osd_xattr_load_by_oid(osd, oid, &nvbuf);
470 if (rc == -ENOENT || rc == -EEXIST || rc == -ENODATA)
476 rc = -nvlist_lookup_byte_array(nvbuf, XATTR_NAME_LMV,
477 (uchar_t **)&lmv, &size);
478 if (rc == -ENOENT || rc == -EEXIST || rc == -ENODATA)
479 GOTO(out_nvbuf, rc = 0);
484 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
485 GOTO(out_nvbuf, rc = -EINVAL);
487 zap_cursor_init_serialized(zc, osd->od_os, oid, 0);
488 rc = -zap_cursor_retrieve(zc, za);
490 zap_cursor_advance(zc);
492 CERROR("%s: fail to init for check LMV "DFID"(%llu): rc = %d\n",
493 osd_name(osd), PFID(fid), oid, rc);
498 rc = -zap_cursor_retrieve(zc, za);
500 GOTO(out_zc, rc = 0);
503 CERROR("%s: fail to locate next for check LMV "
504 DFID"(%llu): rc = %d\n",
505 osd_name(osd), PFID(fid), oid, rc);
510 sscanf(za->za_name + 1, SFID, RFID(tfid));
511 if (fid_is_sane(tfid) && !osd_remote_fid(env, osd, tfid)) {
512 rc = osd_zap_lookup(osd, oid, NULL, za->za_name,
513 za->za_integer_length,
514 sizeof(*zde) / za->za_integer_length,
517 CERROR("%s: fail to lookup for check LMV "
518 DFID"(%llu): rc = %d\n",
519 osd_name(osd), PFID(fid), oid, rc);
523 rc = osd_oii_insert(env, osd, tfid,
524 zde->lzd_reg.zde_dnode, false);
528 zap_cursor_advance(zc);
540 osd_consistency_check(const struct lu_env *env, struct osd_device *osd,
541 struct osd_object *obj, const struct lu_fid *fid,
542 uint64_t oid, bool is_dir)
544 struct lustre_scrub *scrub = &osd->od_scrub;
552 if (!fid_is_norm(fid) && !fid_is_igif(fid))
555 /* oid == ZFS_NO_OBJECT must be for lookup ".." case */
556 if (oid == ZFS_NO_OBJECT) {
557 rc = osd_sa_handle_get(obj);
561 rc = -sa_lookup(obj->oo_sa_hdl, SA_ZPL_PARENT(osd), &oid, 8);
566 if (scrub->os_running) {
567 if (scrub->os_pos_current > oid)
569 } else if (osd->od_auto_scrub_interval == AS_NEVER) {
572 if (ktime_get_real_seconds() <
573 scrub->os_file.sf_time_last_complete +
574 osd->od_auto_scrub_interval)
579 rc = osd_fid_lookup(env, osd, fid, &oid2);
585 rc = __osd_obj2dnode(osd->od_os, oid, &dn);
586 /* The object has been removed (by race maybe). */
588 RETURN(rc = (rc == -EEXIST ? -ENOENT : rc));
591 } else if (rc || oid == oid2) {
598 if (scrub->os_running) {
600 rc = __osd_obj2dnode(osd->od_os, oid, &dn);
601 /* The object has been removed (by race maybe). */
603 RETURN(rc = (rc == -EEXIST ? -ENOENT : rc));
606 rc = osd_oii_insert(env, osd, fid, oid, insert);
607 /* There is race condition between osd_oi_lookup and OI scrub.
608 * The OI scrub finished just after osd_oi_lookup() failure.
609 * Under such case, it is unnecessary to trigger OI scrub again,
610 * but try to call osd_oi_lookup() again. */
611 if (unlikely(rc == -EAGAIN))
615 rc = osd_check_lmv(env, osd, oid, fid);
622 if (osd->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
623 rc = osd_scrub_start(env, osd, SS_AUTO_FULL |
624 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
625 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
626 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%#llx): rc = %d\n",
627 osd_name(osd), PFID(fid), oid, rc);
641 static int osd_dir_lookup(const struct lu_env *env, struct dt_object *dt,
642 struct dt_rec *rec, const struct dt_key *key)
644 struct osd_thread_info *oti = osd_oti_get(env);
645 struct osd_object *obj = osd_dt_obj(dt);
646 struct osd_device *osd = osd_obj2dev(obj);
647 struct lu_fid *fid = (struct lu_fid *)rec;
648 char *name = (char *)key;
649 uint64_t oid = ZFS_NO_OBJECT;
653 if (name[0] == '.') {
655 const struct lu_fid *f = lu_object_fid(&dt->do_lu);
656 memcpy(rec, f, sizeof(*f));
658 } else if (name[1] == '.' && name[2] == 0) {
659 rc = osd_find_parent_fid(env, dt, fid, &oid);
664 memset(&oti->oti_zde.lzd_fid, 0, sizeof(struct lu_fid));
666 down_read(&obj->oo_guard);
667 if (obj->oo_destroyed)
668 GOTO(out_unlock, rc = -ENOENT);
670 rc = osd_zap_lookup(osd, obj->oo_dn->dn_object, obj->oo_dn,
671 (char *)key, 8, sizeof(oti->oti_zde) / 8,
672 (void *)&oti->oti_zde);
674 up_read(&obj->oo_guard);
678 oid = oti->oti_zde.lzd_reg.zde_dnode;
679 if (likely(fid_is_sane(&oti->oti_zde.lzd_fid))) {
680 memcpy(rec, &oti->oti_zde.lzd_fid, sizeof(struct lu_fid));
681 GOTO(out_unlock, rc = 0);
684 rc = osd_get_fid_by_oid(env, osd, oti->oti_zde.lzd_reg.zde_dnode, fid);
687 up_read(&obj->oo_guard);
689 if (!rc && !osd_remote_fid(env, osd, fid)) {
691 * this should ask the scrubber to check OI given
692 * the mapping we just found in the dir entry.
693 * but result of that check should not affect
694 * result of the lookup in the directory.
695 * otherwise such a direntry becomes hidden
696 * from the layers above, including LFSCK which
697 * is supposed to fix dangling entries.
699 osd_consistency_check(env, osd, obj, fid, oid,
700 S_ISDIR(DTTOIF(oti->oti_zde.lzd_reg.zde_type)));
703 return rc == 0 ? 1 : (rc == -ENOENT ? -ENODATA : rc);
707 * In DNE environment, the object and its name entry may reside on different
708 * MDTs. Under such case, we will create an agent object on the MDT where the
709 * name entry resides. The agent object is empty, and indicates that the real
710 * object for the name entry resides on another MDT. If without agent object,
711 * related name entry will be skipped when perform MDT side file level backup
712 * and restore via ZPL by userspace tool, such as 'tar'.
714 static int osd_create_agent_object(const struct lu_env *env,
715 struct osd_device *osd,
716 struct luz_direntry *zde,
717 uint64_t parent, dmu_tx_t *tx)
719 struct osd_thread_info *info = osd_oti_get(env);
720 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
721 struct lu_attr *la = &info->oti_la;
722 nvlist_t *nvbuf = NULL;
728 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_AGENTOBJ))
731 rc = -nvlist_alloc(&nvbuf, NV_UNIQUE_NAME, KM_SLEEP);
735 lustre_lma_init(lma, &zde->lzd_fid, 0, LMAI_AGENT);
736 lustre_lma_swab(lma);
737 rc = -nvlist_add_byte_array(nvbuf, XATTR_NAME_LMA, (uchar_t *)lma,
742 la->la_valid = LA_TYPE | LA_MODE;
743 la->la_mode = (DTTOIF(zde->lzd_reg.zde_type) & S_IFMT) |
744 S_IRUGO | S_IWUSR | S_IXUGO;
746 if (S_ISDIR(la->la_mode))
747 rc = __osd_zap_create(env, osd, &dn, tx, la,
748 osd_find_dnsize(osd, OSD_BASE_EA_IN_BONUS), 0);
750 rc = __osd_object_create(env, osd, NULL, &zde->lzd_fid,
755 zde->lzd_reg.zde_dnode = dn->dn_object;
756 rc = -sa_handle_get(osd->od_os, dn->dn_object, NULL,
757 SA_HDL_PRIVATE, &hdl);
759 rc = __osd_attr_init(env, osd, NULL, hdl, tx,
761 sa_handle_destroy(hdl);
769 dmu_object_free(osd->od_os, dn->dn_object, tx);
779 int osd_add_to_remote_parent(const struct lu_env *env,
780 struct osd_device *osd,
781 struct osd_object *obj,
782 struct osd_thandle *oh)
784 struct osd_thread_info *info = osd_oti_get(env);
785 struct luz_direntry *zde = &info->oti_zde;
786 char *name = info->oti_str;
787 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
788 struct lustre_mdt_attrs *lma = (struct lustre_mdt_attrs *)info->oti_buf;
789 struct lu_buf buf = {
791 .lb_len = sizeof(info->oti_buf),
797 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_AGENTENT))
800 rc = osd_xattr_get_internal(env, obj, &buf, XATTR_NAME_LMA, &size);
802 CWARN("%s: fail to load LMA for adding "
803 DFID" to remote parent: rc = %d\n",
804 osd_name(osd), PFID(fid), rc);
808 lustre_lma_swab(lma);
809 lma->lma_incompat |= LMAI_REMOTE_PARENT;
810 lustre_lma_swab(lma);
812 rc = osd_xattr_set_internal(env, obj, &buf, XATTR_NAME_LMA,
813 LU_XATTR_REPLACE, oh);
815 CWARN("%s: fail to update LMA for adding "
816 DFID" to remote parent: rc = %d\n",
817 osd_name(osd), PFID(fid), rc);
821 osd_fid2str(name, fid, sizeof(info->oti_str));
822 zde->lzd_reg.zde_dnode = obj->oo_dn->dn_object;
823 zde->lzd_reg.zde_type = S_DT(S_IFDIR);
826 rc = osd_zap_add(osd, osd->od_remote_parent_dir, NULL,
827 name, 8, sizeof(*zde) / 8, zde, oh->ot_tx);
828 if (unlikely(rc == -EEXIST))
831 CWARN("%s: fail to add name entry for "
832 DFID" to remote parent: rc = %d\n",
833 osd_name(osd), PFID(fid), rc);
835 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
840 int osd_delete_from_remote_parent(const struct lu_env *env,
841 struct osd_device *osd,
842 struct osd_object *obj,
843 struct osd_thandle *oh, bool destroy)
845 struct osd_thread_info *info = osd_oti_get(env);
846 char *name = info->oti_str;
847 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
848 struct lustre_mdt_attrs *lma = (struct lustre_mdt_attrs *)info->oti_buf;
849 struct lu_buf buf = {
851 .lb_len = sizeof(info->oti_buf),
857 osd_fid2str(name, fid, sizeof(info->oti_str));
858 rc = osd_zap_remove(osd, osd->od_remote_parent_dir, NULL,
860 if (unlikely(rc == -ENOENT))
863 CERROR("%s: fail to remove entry under remote "
864 "parent for "DFID": rc = %d\n",
865 osd_name(osd), PFID(fid), rc);
870 rc = osd_xattr_get_internal(env, obj, &buf, XATTR_NAME_LMA, &size);
872 CERROR("%s: fail to load LMA for removing "
873 DFID" from remote parent: rc = %d\n",
874 osd_name(osd), PFID(fid), rc);
878 lustre_lma_swab(lma);
879 lma->lma_incompat &= ~LMAI_REMOTE_PARENT;
880 lustre_lma_swab(lma);
882 rc = osd_xattr_set_internal(env, obj, &buf, XATTR_NAME_LMA,
883 LU_XATTR_REPLACE, oh);
885 CERROR("%s: fail to update LMA for removing "
886 DFID" from remote parent: rc = %d\n",
887 osd_name(osd), PFID(fid), rc);
889 lu_object_clear_agent_entry(&obj->oo_dt.do_lu);
894 static int osd_declare_dir_insert(const struct lu_env *env,
895 struct dt_object *dt,
896 const struct dt_rec *rec,
897 const struct dt_key *key,
900 struct osd_object *obj = osd_dt_obj(dt);
901 struct osd_device *osd = osd_obj2dev(obj);
902 const struct dt_insert_rec *rec1;
903 const struct lu_fid *fid;
904 struct osd_thandle *oh;
906 struct osd_idmap_cache *idc;
909 rec1 = (struct dt_insert_rec *)rec;
911 LASSERT(fid != NULL);
912 LASSERT(rec1->rec_type != 0);
915 oh = container_of(th, struct osd_thandle, ot_super);
917 idc = osd_idc_find_or_init(env, osd, fid);
919 RETURN(PTR_ERR(idc));
921 if (idc->oic_remote) {
922 const char *name = (const char *)key;
924 if (name[0] != '.' || name[1] != '.' || name[2] != 0) {
925 /* Prepare agent object for remote entry that will
926 * be used for operations via ZPL, such as MDT side
927 * file-level backup and restore. */
928 dmu_tx_hold_sa_create(oh->ot_tx,
929 osd_find_dnsize(osd, OSD_BASE_EA_IN_BONUS));
930 if (S_ISDIR(rec1->rec_type))
931 dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT,
936 /* This is for inserting dot/dotdot for new created dir. */
937 if (obj->oo_dn == NULL)
938 object = DMU_NEW_OBJECT;
940 object = obj->oo_dn->dn_object;
942 /* do not specify the key as then DMU is trying to look it up
943 * which is very expensive. usually the layers above lookup
944 * before insertion */
945 osd_tx_hold_zap(oh->ot_tx, object, obj->oo_dn, TRUE, NULL);
950 static int osd_seq_exists(const struct lu_env *env, struct osd_device *osd,
953 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
954 struct seq_server_site *ss = osd_seq_site(osd);
959 LASSERT(ss->ss_server_fld != NULL);
961 rc = osd_fld_lookup(env, osd, seq, range);
964 CERROR("%s: Can not lookup fld for %#llx\n",
969 RETURN(ss->ss_node_id == range->lsr_index);
972 int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
973 const struct lu_fid *fid)
975 struct seq_server_site *ss = osd_seq_site(osd);
978 /* FID seqs not in FLDB, must be local seq */
979 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
982 /* If FLD is not being initialized yet, it only happens during the
983 * initialization, likely during mgs initialization, and we assume
984 * this is local FID. */
985 if (ss == NULL || ss->ss_server_fld == NULL)
988 /* Only check the local FLDB here */
989 if (osd_seq_exists(env, osd, fid_seq(fid)))
996 * Inserts (key, value) pair in \a directory object.
998 * \param dt osd index object
999 * \param key key for index
1000 * \param rec record reference
1001 * \param th transaction handler
1004 * \retval -ve failure
1006 static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
1007 const struct dt_rec *rec, const struct dt_key *key,
1010 struct osd_thread_info *oti = osd_oti_get(env);
1011 struct osd_object *parent = osd_dt_obj(dt);
1012 struct osd_device *osd = osd_obj2dev(parent);
1013 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
1014 const struct lu_fid *fid = rec1->rec_fid;
1015 struct osd_thandle *oh;
1016 struct osd_idmap_cache *idc;
1017 const char *name = (const char *)key;
1018 struct luz_direntry *zde = &oti->oti_zde;
1019 int num = sizeof(*zde) / 8;
1023 LASSERT(parent->oo_dn);
1025 LASSERT(dt_object_exists(dt));
1026 LASSERT(osd_invariant(parent));
1028 LASSERT(th != NULL);
1029 oh = container_of(th, struct osd_thandle, ot_super);
1031 idc = osd_idc_find(env, osd, fid);
1032 if (unlikely(idc == NULL)) {
1033 /* this dt_insert() wasn't declared properly, so
1034 * FID is missing in OI cache. we better do not
1035 * lookup FID in FLDB/OI and don't risk to deadlock,
1036 * but in some special cases (lfsck testing, etc)
1037 * it's much simpler than fixing a caller */
1038 idc = osd_idc_find_or_init(env, osd, fid);
1040 CERROR("%s: "DFID" wasn't declared for insert\n",
1041 osd_name(osd), PFID(fid));
1042 RETURN(PTR_ERR(idc));
1046 BUILD_BUG_ON(sizeof(zde->lzd_reg) != 8);
1047 BUILD_BUG_ON(sizeof(*zde) % 8 != 0);
1049 memset(&zde->lzd_reg, 0, sizeof(zde->lzd_reg));
1050 zde->lzd_reg.zde_type = S_DT(rec1->rec_type & S_IFMT);
1051 zde->lzd_fid = *fid;
1053 if (idc->oic_remote) {
1054 if (name[0] != '.' || name[1] != '.' || name[2] != 0) {
1055 /* Create agent inode for remote object that will
1056 * be used for MDT file-level backup and restore. */
1057 rc = osd_create_agent_object(env, osd, zde,
1058 parent->oo_dn->dn_object, oh->ot_tx);
1060 CWARN("%s: Fail to create agent object for "
1062 osd_name(osd), PFID(fid), rc);
1063 /* Ignore the failure since the system can go
1064 * ahead if we do not care about the MDT side
1065 * file-level backup and restore. */
1070 if (unlikely(idc->oic_dnode == 0)) {
1071 /* for a reason OI cache wasn't filled properly */
1072 CERROR("%s: OIC for "DFID" isn't filled\n",
1073 osd_name(osd), PFID(fid));
1076 if (name[0] == '.') {
1078 /* do not store ".", instead generate it
1079 * during iteration */
1081 } else if (name[1] == '.' && name[2] == 0) {
1082 uint64_t dnode = idc->oic_dnode;
1083 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT))
1086 /* update parent dnode in the child.
1087 * later it will be used to generate ".." */
1088 rc = osd_object_sa_update(parent,
1095 zde->lzd_reg.zde_dnode = idc->oic_dnode;
1098 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR))
1099 zde->lzd_fid.f_ver = ~0;
1101 /* The logic is not related with IGIF, just re-use the fail_loc value
1102 * to be consistent with ldiskfs case, then share the same test logic */
1103 if (OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF))
1106 /* Insert (key,oid) into ZAP */
1107 rc = osd_zap_add(osd, parent->oo_dn->dn_object, parent->oo_dn,
1108 name, 8, num, (void *)zde, oh->ot_tx);
1109 if (unlikely(rc == -EEXIST &&
1110 name[0] == '.' && name[1] == '.' && name[2] == 0))
1111 /* Update (key,oid) in ZAP */
1112 rc = -zap_update(osd->od_os, parent->oo_dn->dn_object, name, 8,
1113 sizeof(*zde) / 8, (void *)zde, oh->ot_tx);
1120 static int osd_declare_dir_delete(const struct lu_env *env,
1121 struct dt_object *dt,
1122 const struct dt_key *key,
1125 struct osd_object *obj = osd_dt_obj(dt);
1126 dnode_t *zap_dn = obj->oo_dn;
1127 struct osd_thandle *oh;
1128 const char *name = (const char *)key;
1131 LASSERT(dt_object_exists(dt));
1132 LASSERT(osd_invariant(obj));
1133 LASSERT(zap_dn != NULL);
1135 LASSERT(th != NULL);
1136 oh = container_of(th, struct osd_thandle, ot_super);
1139 * In Orion . and .. were stored in the directory (not generated upon
1140 * request as now). We preserve them for backward compatibility.
1142 if (name[0] == '.') {
1145 else if (name[1] == '.' && name[2] == 0)
1149 /* do not specify the key as then DMU is trying to look it up
1150 * which is very expensive. usually the layers above lookup
1151 * before deletion */
1152 osd_tx_hold_zap(oh->ot_tx, zap_dn->dn_object, zap_dn, FALSE, NULL);
1154 /* For destroying agent object if have. */
1155 dmu_tx_hold_bonus(oh->ot_tx, DMU_NEW_OBJECT);
1160 static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
1161 const struct dt_key *key, struct thandle *th)
1163 struct luz_direntry *zde = &osd_oti_get(env)->oti_zde;
1164 struct osd_object *obj = osd_dt_obj(dt);
1165 struct osd_device *osd = osd_obj2dev(obj);
1166 struct osd_thandle *oh;
1167 dnode_t *zap_dn = obj->oo_dn;
1168 char *name = (char *)key;
1174 LASSERT(th != NULL);
1175 oh = container_of(th, struct osd_thandle, ot_super);
1178 * In Orion . and .. were stored in the directory (not generated upon
1179 * request as now). we preserve them for backward compatibility
1181 if (name[0] == '.') {
1184 } else if (name[1] == '.' && name[2] == 0) {
1189 /* XXX: We have to say that lookup during delete_declare will affect
1190 * performance, but we have to check whether the name entry (to
1191 * be deleted) has agent object or not to avoid orphans.
1193 * We will improve that in the future, some possible solutions,
1195 * 1) Some hint from the caller via transaction handle to make
1196 * the lookup conditionally.
1197 * 2) Enhance the ZFS logic to recognize the OSD lookup result
1198 * and delete the given entry directly without lookup again
1199 * internally. LU-10190 */
1200 memset(&zde->lzd_fid, 0, sizeof(zde->lzd_fid));
1201 rc = osd_zap_lookup(osd, zap_dn->dn_object, zap_dn, name, 8, 3, zde);
1204 CERROR("%s: failed to locate entry %s: rc = %d\n",
1205 osd->od_svname, name, rc);
1209 if (unlikely(osd_remote_fid(env, osd, &zde->lzd_fid) > 0)) {
1210 rc = -dmu_object_free(osd->od_os, zde->lzd_reg.zde_dnode,
1213 CERROR("%s: failed to destroy agent object (%llu) "
1214 "for the entry %s: rc = %d\n", osd->od_svname,
1215 (__u64)zde->lzd_reg.zde_dnode, name, rc);
1218 /* Remove key from the ZAP */
1219 rc = osd_zap_remove(osd, zap_dn->dn_object, zap_dn,
1220 (char *)key, oh->ot_tx);
1222 CERROR("%s: zap_remove %s failed: rc = %d\n",
1223 osd->od_svname, name, rc);
1228 static struct dt_it *osd_dir_it_init(const struct lu_env *env,
1229 struct dt_object *dt,
1232 struct osd_zap_it *it;
1234 it = (struct osd_zap_it *)osd_index_it_init(env, dt, unused);
1236 it->ozi_pos = OZI_POS_INIT;
1238 RETURN((struct dt_it *)it);
1242 * Move Iterator to record specified by \a key
1244 * \param di osd iterator
1245 * \param key key for index
1247 * \retval +ve di points to record with least key not larger than key
1248 * \retval 0 di points to exact matched key
1249 * \retval -ve failure
1251 static int osd_dir_it_get(const struct lu_env *env,
1252 struct dt_it *di, const struct dt_key *key)
1254 struct osd_zap_it *it = (struct osd_zap_it *)di;
1255 struct osd_object *obj = it->ozi_obj;
1256 char *name = (char *)key;
1261 LASSERT(it->ozi_zc);
1263 /* reset the cursor */
1264 zap_cursor_fini(it->ozi_zc);
1265 osd_obj_cursor_init_serialized(it->ozi_zc, obj, 0);
1267 /* XXX: implementation of the API is broken at the moment */
1268 LASSERT(((const char *)key)[0] == 0);
1271 it->ozi_pos = OZI_POS_INIT;
1275 if (name[0] == '.') {
1277 it->ozi_pos = OZI_POS_DOT;
1279 } else if (name[1] == '.' && name[2] == 0) {
1280 it->ozi_pos = OZI_POS_DOTDOT;
1285 /* neither . nor .. - some real record */
1286 it->ozi_pos = OZI_POS_REAL;
1293 static void osd_dir_it_put(const struct lu_env *env, struct dt_it *di)
1295 /* PBS: do nothing : ref are incremented at retrive and decreamented
1300 * in Orion . and .. were stored in the directory, while ZPL
1301 * and current osd-zfs generate them up on request. so, we
1302 * need to ignore previously stored . and ..
1304 static int osd_index_retrieve_skip_dots(struct osd_zap_it *it,
1305 zap_attribute_t *za)
1310 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1313 if (unlikely(rc == 0 && za->za_name[0] == '.')) {
1314 if (za->za_name[1] == 0) {
1316 } else if (za->za_name[1] == '.' &&
1317 za->za_name[2] == 0) {
1320 if (unlikely(isdot))
1321 zap_cursor_advance(it->ozi_zc);
1323 } while (unlikely(rc == 0 && isdot));
1329 * to load a directory entry at a time and stored it in
1330 * iterator's in-memory data structure.
1332 * \param di, struct osd_it_ea, iterator's in memory structure
1334 * \retval +ve, iterator reached to end
1335 * \retval 0, iterator not reached to end
1336 * \retval -ve, on error
1338 static int osd_dir_it_next(const struct lu_env *env, struct dt_it *di)
1340 struct osd_zap_it *it = (struct osd_zap_it *)di;
1341 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1346 /* temp. storage should be enough for any key supported by ZFS */
1347 BUILD_BUG_ON(sizeof(za->za_name) > sizeof(it->ozi_name));
1350 * the first ->next() moves the cursor to .
1351 * the second ->next() moves the cursor to ..
1352 * then we get to the real records and have to verify any exist
1354 if (it->ozi_pos <= OZI_POS_DOTDOT) {
1356 if (it->ozi_pos <= OZI_POS_DOTDOT)
1360 zap_cursor_advance(it->ozi_zc);
1364 * According to current API we need to return error if its last entry.
1365 * zap_cursor_advance() does not return any value. So we need to call
1366 * retrieve to check if there is any record. We should make
1367 * changes to Iterator API to not return status for this API
1369 rc = osd_index_retrieve_skip_dots(it, za);
1371 if (rc == -ENOENT) /* end of dir */
1377 static struct dt_key *osd_dir_it_key(const struct lu_env *env,
1378 const struct dt_it *di)
1380 struct osd_zap_it *it = (struct osd_zap_it *)di;
1381 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1385 if (it->ozi_pos <= OZI_POS_DOT) {
1386 it->ozi_pos = OZI_POS_DOT;
1387 RETURN((struct dt_key *)".");
1388 } else if (it->ozi_pos == OZI_POS_DOTDOT) {
1389 RETURN((struct dt_key *)"..");
1392 if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)))
1393 RETURN(ERR_PTR(rc));
1395 strcpy(it->ozi_name, za->za_name);
1397 RETURN((struct dt_key *)it->ozi_name);
1400 static int osd_dir_it_key_size(const struct lu_env *env, const struct dt_it *di)
1402 struct osd_zap_it *it = (struct osd_zap_it *)di;
1403 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1407 if (it->ozi_pos <= OZI_POS_DOT) {
1408 it->ozi_pos = OZI_POS_DOT;
1410 } else if (it->ozi_pos == OZI_POS_DOTDOT) {
1414 if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)) == 0)
1415 rc = strlen(za->za_name);
1421 osd_dirent_update(const struct lu_env *env, struct osd_device *dev,
1422 uint64_t zap, const char *key, struct luz_direntry *zde)
1428 tx = dmu_tx_create(dev->od_os);
1432 dmu_tx_hold_zap(tx, zap, TRUE, NULL);
1433 rc = -dmu_tx_assign(tx, TXG_WAIT);
1435 rc = -zap_update(dev->od_os, zap, key, 8, sizeof(*zde) / 8,
1436 (const void *)zde, tx);
1445 static int osd_update_entry_for_agent(const struct lu_env *env,
1446 struct osd_device *osd,
1447 uint64_t zap, const char *name,
1448 struct luz_direntry *zde, __u32 attr)
1450 dmu_tx_t *tx = NULL;
1454 if (attr & LUDA_VERIFY_DRYRUN)
1457 tx = dmu_tx_create(osd->od_os);
1459 GOTO(out, rc = -ENOMEM);
1461 dmu_tx_hold_sa_create(tx, osd_find_dnsize(osd, OSD_BASE_EA_IN_BONUS));
1462 dmu_tx_hold_zap(tx, zap, FALSE, NULL);
1463 rc = -dmu_tx_assign(tx, TXG_WAIT);
1469 rc = osd_create_agent_object(env, osd, zde, zap, tx);
1471 rc = -zap_update(osd->od_os, zap, name, 8, sizeof(*zde) / 8,
1472 (const void *)zde, tx);
1478 CDEBUG(D_LFSCK, "%s: Updated (%s) remote entry for "DFID": rc = %d\n",
1479 osd_name(osd), (attr & LUDA_VERIFY_DRYRUN) ? "(ro)" : "(rw)",
1480 PFID(&zde->lzd_fid), rc);
1484 static int osd_dir_it_rec(const struct lu_env *env, const struct dt_it *di,
1485 struct dt_rec *dtrec, __u32 attr)
1487 struct osd_zap_it *it = (struct osd_zap_it *)di;
1488 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
1489 struct osd_thread_info *info = osd_oti_get(env);
1490 struct luz_direntry *zde = &info->oti_zde;
1491 zap_attribute_t *za = &info->oti_za;
1492 struct lu_fid *fid = &info->oti_fid;
1493 struct osd_device *osd = osd_obj2dev(it->ozi_obj);
1498 if (it->ozi_pos <= OZI_POS_DOT) {
1499 /* notice hash=0 here, this is needed to avoid
1500 * case when some real entry (after ./..) may
1501 * have hash=0. in this case the client would
1502 * be confused having records out of hash order. */
1503 lde->lde_hash = cpu_to_le64(0);
1504 strcpy(lde->lde_name, ".");
1505 lde->lde_namelen = cpu_to_le16(1);
1506 fid_cpu_to_le(&lde->lde_fid,
1507 lu_object_fid(&it->ozi_obj->oo_dt.do_lu));
1508 lde->lde_attrs = LUDA_FID;
1509 /* append lustre attributes */
1510 osd_it_append_attrs(lde, attr, 1, S_DT(S_IFDIR));
1511 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(1, attr));
1512 it->ozi_pos = OZI_POS_DOT;
1514 } else if (it->ozi_pos == OZI_POS_DOTDOT) {
1515 /* same as for . above */
1516 lde->lde_hash = cpu_to_le64(0);
1517 strcpy(lde->lde_name, "..");
1518 lde->lde_namelen = cpu_to_le16(2);
1519 rc = osd_find_parent_fid(env, &it->ozi_obj->oo_dt, fid, NULL);
1521 fid_cpu_to_le(&lde->lde_fid, fid);
1522 lde->lde_attrs = LUDA_FID;
1523 } else if (rc != -ENOENT) {
1524 /* ENOENT happens at the root of filesystem, ignore */
1528 /* append lustre attributes */
1529 osd_it_append_attrs(lde, attr, 2, S_DT(S_IFDIR));
1530 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(2, attr));
1536 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1540 lde->lde_hash = cpu_to_le64(osd_zap_cursor_serialize(it->ozi_zc));
1541 namelen = strlen(za->za_name);
1542 if (namelen > NAME_MAX)
1544 strcpy(lde->lde_name, za->za_name);
1545 lde->lde_namelen = cpu_to_le16(namelen);
1547 if (za->za_integer_length != 8) {
1548 CERROR("%s: unsupported direntry format: %d %d\n",
1550 za->za_integer_length, (int)za->za_num_integers);
1554 rc = osd_zap_lookup(osd, it->ozi_zc->zc_zapobj, it->ozi_obj->oo_dn,
1555 za->za_name, za->za_integer_length, 3, zde);
1559 if (za->za_num_integers >= 3 && fid_is_sane(&zde->lzd_fid)) {
1560 lde->lde_attrs = LUDA_FID;
1561 fid_cpu_to_le(&lde->lde_fid, &zde->lzd_fid);
1562 if (unlikely(zde->lzd_reg.zde_dnode == ZFS_NO_OBJECT &&
1563 osd_remote_fid(env, osd, &zde->lzd_fid) > 0 &&
1564 attr & LUDA_VERIFY)) {
1565 /* It is mainly used for handling the MDT
1566 * upgraded from old ZFS based backend. */
1567 rc = osd_update_entry_for_agent(env, osd,
1568 it->ozi_obj->oo_dn->dn_object,
1569 za->za_name, zde, attr);
1571 lde->lde_attrs |= LUDA_REPAIR;
1573 lde->lde_attrs |= LUDA_UNKNOWN;
1576 if (!(attr & (LUDA_VERIFY | LUDA_VERIFY_DRYRUN)))
1577 GOTO(pack_attr, rc = 0);
1580 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
1583 rc = osd_get_fid_by_oid(env, osd, zde->lzd_reg.zde_dnode, fid);
1585 lde->lde_attrs = LUDA_UNKNOWN;
1586 GOTO(pack_attr, rc = 0);
1589 if (za->za_num_integers >= 3 && fid_is_sane(&zde->lzd_fid) &&
1590 lu_fid_eq(&zde->lzd_fid, fid))
1591 GOTO(pack_attr, rc = 0);
1593 if (!(attr & LUDA_VERIFY)) {
1594 fid_cpu_to_le(&lde->lde_fid, fid);
1595 lde->lde_attrs = LUDA_FID;
1596 GOTO(pack_attr, rc = 0);
1599 if (attr & LUDA_VERIFY_DRYRUN) {
1600 fid_cpu_to_le(&lde->lde_fid, fid);
1601 lde->lde_attrs = LUDA_FID | LUDA_REPAIR;
1602 GOTO(pack_attr, rc = 0);
1605 fid_cpu_to_le(&lde->lde_fid, fid);
1606 lde->lde_attrs = LUDA_FID;
1607 zde->lzd_fid = *fid;
1608 rc = osd_dirent_update(env, osd, it->ozi_zc->zc_zapobj,
1611 lde->lde_attrs |= LUDA_UNKNOWN;
1612 GOTO(pack_attr, rc = 0);
1615 lde->lde_attrs |= LUDA_REPAIR;
1617 GOTO(pack_attr, rc = 0);
1620 osd_it_append_attrs(lde, attr, namelen, zde->lzd_reg.zde_type);
1621 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
1625 static int osd_dir_it_rec_size(const struct lu_env *env, const struct dt_it *di,
1628 struct osd_zap_it *it = (struct osd_zap_it *)di;
1629 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1634 if (it->ozi_pos <= OZI_POS_DOT)
1636 else if (it->ozi_pos == OZI_POS_DOTDOT)
1640 rc = lu_dirent_calc_size(namelen, attr);
1644 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1645 if (unlikely(rc != 0))
1648 if (za->za_integer_length != 8 || za->za_num_integers < 3) {
1649 CERROR("%s: unsupported direntry format: %d %d\n",
1650 osd_obj2dev(it->ozi_obj)->od_svname,
1651 za->za_integer_length, (int)za->za_num_integers);
1655 namelen = strlen(za->za_name);
1656 if (namelen > NAME_MAX)
1659 rc = lu_dirent_calc_size(namelen, attr);
1664 static __u64 osd_dir_it_store(const struct lu_env *env, const struct dt_it *di)
1666 struct osd_zap_it *it = (struct osd_zap_it *)di;
1670 if (it->ozi_pos <= OZI_POS_DOTDOT)
1673 pos = osd_zap_cursor_serialize(it->ozi_zc);
1680 * rc == 0 -> end of directory.
1681 * rc > 0 -> ok, proceed.
1682 * rc < 0 -> error. ( EOVERFLOW can be masked.)
1684 static int osd_dir_it_load(const struct lu_env *env,
1685 const struct dt_it *di, __u64 hash)
1687 struct osd_zap_it *it = (struct osd_zap_it *)di;
1688 struct osd_object *obj = it->ozi_obj;
1689 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1693 /* reset the cursor */
1694 zap_cursor_fini(it->ozi_zc);
1695 osd_obj_cursor_init_serialized(it->ozi_zc, obj, hash);
1698 it->ozi_pos = OZI_POS_INIT;
1699 rc = +1; /* there will be ./.. at least */
1701 it->ozi_pos = OZI_POS_REAL;
1702 /* to return whether the end has been reached */
1703 rc = osd_index_retrieve_skip_dots(it, za);
1706 else if (rc == -ENOENT)
1713 const struct dt_index_operations osd_dir_ops = {
1714 .dio_lookup = osd_dir_lookup,
1715 .dio_declare_insert = osd_declare_dir_insert,
1716 .dio_insert = osd_dir_insert,
1717 .dio_declare_delete = osd_declare_dir_delete,
1718 .dio_delete = osd_dir_delete,
1720 .init = osd_dir_it_init,
1721 .fini = osd_index_it_fini,
1722 .get = osd_dir_it_get,
1723 .put = osd_dir_it_put,
1724 .next = osd_dir_it_next,
1725 .key = osd_dir_it_key,
1726 .key_size = osd_dir_it_key_size,
1727 .rec = osd_dir_it_rec,
1728 .rec_size = osd_dir_it_rec_size,
1729 .store = osd_dir_it_store,
1730 .load = osd_dir_it_load
1735 * Primitives for index files using binary keys.
1738 /* key integer_size is 8 */
1739 static int osd_prepare_key_uint64(struct osd_object *o, __u64 *dst,
1740 const struct dt_key *src)
1747 /* align keysize to 64bit */
1748 size = (o->oo_keysize + sizeof(__u64) - 1) / sizeof(__u64);
1749 size *= sizeof(__u64);
1751 LASSERT(size <= MAXNAMELEN);
1753 if (unlikely(size > o->oo_keysize))
1754 memset(dst + o->oo_keysize, 0, size - o->oo_keysize);
1755 memcpy(dst, (const char *)src, o->oo_keysize);
1757 return (size/sizeof(__u64));
1760 static int osd_index_lookup(const struct lu_env *env, struct dt_object *dt,
1761 struct dt_rec *rec, const struct dt_key *key)
1763 struct osd_object *obj = osd_dt_obj(dt);
1764 struct osd_device *osd = osd_obj2dev(obj);
1765 __u64 *k = osd_oti_get(env)->oti_key64;
1769 rc = osd_prepare_key_uint64(obj, k, key);
1771 rc = -zap_lookup_uint64(osd->od_os, obj->oo_dn->dn_object,
1772 k, rc, obj->oo_recusize, obj->oo_recsize,
1774 RETURN(rc == 0 ? 1 : rc);
1777 static int osd_declare_index_insert(const struct lu_env *env,
1778 struct dt_object *dt,
1779 const struct dt_rec *rec,
1780 const struct dt_key *key,
1783 struct osd_object *obj = osd_dt_obj(dt);
1784 struct osd_thandle *oh;
1787 LASSERT(th != NULL);
1788 oh = container_of(th, struct osd_thandle, ot_super);
1790 LASSERT(obj->oo_dn);
1792 /* do not specify the key as then DMU is trying to look it up
1793 * which is very expensive. usually the layers above lookup
1794 * before insertion */
1795 osd_tx_hold_zap(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
1801 static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
1802 const struct dt_rec *rec, const struct dt_key *key,
1805 struct osd_object *obj = osd_dt_obj(dt);
1806 struct osd_device *osd = osd_obj2dev(obj);
1807 struct osd_thandle *oh;
1808 __u64 *k = osd_oti_get(env)->oti_key64;
1812 LASSERT(obj->oo_dn);
1813 LASSERT(dt_object_exists(dt));
1814 LASSERT(osd_invariant(obj));
1815 LASSERT(th != NULL);
1817 oh = container_of(th, struct osd_thandle, ot_super);
1819 rc = osd_prepare_key_uint64(obj, k, key);
1821 /* Insert (key,oid) into ZAP */
1822 rc = -zap_add_uint64(osd->od_os, obj->oo_dn->dn_object,
1823 k, rc, obj->oo_recusize, obj->oo_recsize,
1824 (void *)rec, oh->ot_tx);
1828 static int osd_declare_index_delete(const struct lu_env *env,
1829 struct dt_object *dt,
1830 const struct dt_key *key,
1833 struct osd_object *obj = osd_dt_obj(dt);
1834 struct osd_thandle *oh;
1837 LASSERT(dt_object_exists(dt));
1838 LASSERT(osd_invariant(obj));
1839 LASSERT(th != NULL);
1840 LASSERT(obj->oo_dn);
1842 oh = container_of(th, struct osd_thandle, ot_super);
1844 /* do not specify the key as then DMU is trying to look it up
1845 * which is very expensive. usually the layers above lookup
1846 * before deletion */
1847 osd_tx_hold_zap(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
1853 static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
1854 const struct dt_key *key, struct thandle *th)
1856 struct osd_object *obj = osd_dt_obj(dt);
1857 struct osd_device *osd = osd_obj2dev(obj);
1858 struct osd_thandle *oh;
1859 __u64 *k = osd_oti_get(env)->oti_key64;
1863 LASSERT(obj->oo_dn);
1864 LASSERT(th != NULL);
1865 oh = container_of(th, struct osd_thandle, ot_super);
1867 rc = osd_prepare_key_uint64(obj, k, key);
1869 /* Remove binary key from the ZAP */
1870 rc = -zap_remove_uint64(osd->od_os, obj->oo_dn->dn_object,
1875 static int osd_index_it_get(const struct lu_env *env, struct dt_it *di,
1876 const struct dt_key *key)
1878 struct osd_zap_it *it = (struct osd_zap_it *)di;
1879 struct osd_object *obj = it->ozi_obj;
1880 struct osd_device *osd = osd_obj2dev(obj);
1884 LASSERT(it->ozi_zc);
1887 * XXX: we need a binary version of zap_cursor_move_to_key()
1888 * to implement this API */
1889 if (*((const __u64 *)key) != 0)
1890 CERROR("NOT IMPLEMETED YET (move to %#llx)\n",
1893 zap_cursor_fini(it->ozi_zc);
1894 zap_cursor_init(it->ozi_zc, osd->od_os, obj->oo_dn->dn_object);
1900 static int osd_index_it_next(const struct lu_env *env, struct dt_it *di)
1902 struct osd_zap_it *it = (struct osd_zap_it *)di;
1903 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1907 if (it->ozi_reset == 0)
1908 zap_cursor_advance(it->ozi_zc);
1912 * According to current API we need to return error if it's last entry.
1913 * zap_cursor_advance() does not return any value. So we need to call
1914 * retrieve to check if there is any record. We should make
1915 * changes to Iterator API to not return status for this API
1917 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1924 static struct dt_key *osd_index_it_key(const struct lu_env *env,
1925 const struct dt_it *di)
1927 struct osd_zap_it *it = (struct osd_zap_it *)di;
1928 struct osd_object *obj = it->ozi_obj;
1929 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1934 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1936 RETURN(ERR_PTR(rc));
1938 /* the binary key is stored in the name */
1939 memcpy(&it->ozi_key, za->za_name, obj->oo_keysize);
1941 RETURN((struct dt_key *)&it->ozi_key);
1944 static int osd_index_it_key_size(const struct lu_env *env,
1945 const struct dt_it *di)
1947 struct osd_zap_it *it = (struct osd_zap_it *)di;
1948 struct osd_object *obj = it->ozi_obj;
1949 RETURN(obj->oo_keysize);
1952 static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
1953 struct dt_rec *rec, __u32 attr)
1955 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1956 struct osd_zap_it *it = (struct osd_zap_it *)di;
1957 struct osd_object *obj = it->ozi_obj;
1958 struct osd_device *osd = osd_obj2dev(obj);
1959 __u64 *k = osd_oti_get(env)->oti_key64;
1964 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1968 rc = osd_prepare_key_uint64(obj, k, (const struct dt_key *)za->za_name);
1970 rc = -zap_lookup_uint64(osd->od_os, obj->oo_dn->dn_object,
1971 k, rc, obj->oo_recusize, obj->oo_recsize,
1976 static __u64 osd_index_it_store(const struct lu_env *env,
1977 const struct dt_it *di)
1979 struct osd_zap_it *it = (struct osd_zap_it *)di;
1982 RETURN((__u64)zap_cursor_serialize(it->ozi_zc));
1985 static int osd_index_it_load(const struct lu_env *env, const struct dt_it *di,
1988 struct osd_zap_it *it = (struct osd_zap_it *)di;
1989 struct osd_object *obj = it->ozi_obj;
1990 struct osd_device *osd = osd_obj2dev(obj);
1991 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1995 /* reset the cursor */
1996 zap_cursor_fini(it->ozi_zc);
1997 zap_cursor_init_serialized(it->ozi_zc, osd->od_os,
1998 obj->oo_dn->dn_object, hash);
2001 rc = -zap_cursor_retrieve(it->ozi_zc, za);
2004 else if (rc == -ENOENT)
2010 static const struct dt_index_operations osd_index_ops = {
2011 .dio_lookup = osd_index_lookup,
2012 .dio_declare_insert = osd_declare_index_insert,
2013 .dio_insert = osd_index_insert,
2014 .dio_declare_delete = osd_declare_index_delete,
2015 .dio_delete = osd_index_delete,
2017 .init = osd_index_it_init,
2018 .fini = osd_index_it_fini,
2019 .get = osd_index_it_get,
2020 .put = osd_index_it_put,
2021 .next = osd_index_it_next,
2022 .key = osd_index_it_key,
2023 .key_size = osd_index_it_key_size,
2024 .rec = osd_index_it_rec,
2025 .store = osd_index_it_store,
2026 .load = osd_index_it_load
2030 int osd_index_try(const struct lu_env *env, struct dt_object *dt,
2031 const struct dt_index_features *feat)
2033 struct osd_object *obj = osd_dt_obj(dt);
2034 struct osd_device *osd = osd_obj2dev(obj);
2035 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
2039 down_read(&obj->oo_guard);
2042 * XXX: implement support for fixed-size keys sorted with natural
2043 * numerical way (not using internal hash value)
2045 if (feat->dif_flags & DT_IND_RANGE)
2046 GOTO(out, rc = -ERANGE);
2048 if (unlikely(feat == &dt_otable_features)) {
2049 dt->do_index_ops = &osd_otable_ops;
2053 LASSERT(!dt_object_exists(dt) || obj->oo_dn != NULL);
2054 if (likely(feat == &dt_directory_features)) {
2055 if (!dt_object_exists(dt) || osd_object_is_zap(obj->oo_dn))
2056 dt->do_index_ops = &osd_dir_ops;
2058 GOTO(out, rc = -ENOTDIR);
2059 } else if (unlikely(feat == &dt_acct_features)) {
2060 LASSERT(fid_is_acct(fid));
2061 dt->do_index_ops = &osd_acct_index_ops;
2062 } else if (dt->do_index_ops == NULL) {
2063 /* For index file, we don't support variable key & record sizes
2064 * and the key has to be unique */
2065 if ((feat->dif_flags & ~DT_IND_UPDATE) != 0)
2066 GOTO(out, rc = -EINVAL);
2068 if (feat->dif_keysize_max > ZAP_MAXNAMELEN)
2069 GOTO(out, rc = -E2BIG);
2070 if (feat->dif_keysize_max != feat->dif_keysize_min)
2071 GOTO(out, rc = -EINVAL);
2073 /* As for the record size, it should be a multiple of 8 bytes
2074 * and smaller than the maximum value length supported by ZAP.
2076 if (feat->dif_recsize_max > ZAP_MAXVALUELEN)
2077 GOTO(out, rc = -E2BIG);
2078 if (feat->dif_recsize_max != feat->dif_recsize_min)
2079 GOTO(out, rc = -EINVAL);
2081 obj->oo_keysize = feat->dif_keysize_max;
2082 obj->oo_recsize = feat->dif_recsize_max;
2083 obj->oo_recusize = 1;
2085 /* ZFS prefers to work with array of 64bits */
2086 if ((obj->oo_recsize & 7) == 0) {
2087 obj->oo_recsize >>= 3;
2088 obj->oo_recusize = 8;
2090 dt->do_index_ops = &osd_index_ops;
2092 if (feat == &dt_lfsck_layout_orphan_features ||
2093 feat == &dt_lfsck_layout_dangling_features ||
2094 feat == &dt_lfsck_namespace_features)
2097 rc = osd_index_register(osd, fid, obj->oo_keysize,
2098 obj->oo_recusize * obj->oo_recsize);
2100 CWARN("%s: failed to register index "DFID": rc = %d\n",
2101 osd_name(osd), PFID(fid), rc);
2105 CDEBUG(D_LFSCK, "%s: index object "DFID
2106 " (%u/%u/%u) registered\n",
2107 osd_name(osd), PFID(fid), obj->oo_keysize,
2108 obj->oo_recusize, obj->oo_recsize);
2112 up_read(&obj->oo_guard);