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, 2016, 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() 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));
461 static int osd_declare_dir_insert(const struct lu_env *env,
462 struct dt_object *dt,
463 const struct dt_rec *rec,
464 const struct dt_key *key,
467 struct osd_object *obj = osd_dt_obj(dt);
468 struct osd_device *osd = osd_obj2dev(obj);
469 const struct dt_insert_rec *rec1;
470 const struct lu_fid *fid;
471 struct osd_thandle *oh;
475 rec1 = (struct dt_insert_rec *)rec;
477 LASSERT(fid != NULL);
478 LASSERT(rec1->rec_type != 0);
481 oh = container_of0(th, struct osd_thandle, ot_super);
483 /* This is for inserting dot/dotdot for new created dir. */
484 if (obj->oo_dn == NULL)
485 object = DMU_NEW_OBJECT;
487 object = obj->oo_dn->dn_object;
489 /* do not specify the key as then DMU is trying to look it up
490 * which is very expensive. usually the layers above lookup
491 * before insertion */
492 osd_tx_hold_zap(oh->ot_tx, object, obj->oo_dn, TRUE, NULL);
494 osd_idc_find_or_init(env, osd, fid);
499 static int osd_seq_exists(const struct lu_env *env, struct osd_device *osd,
502 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
503 struct seq_server_site *ss = osd_seq_site(osd);
508 LASSERT(ss->ss_server_fld != NULL);
510 rc = osd_fld_lookup(env, osd, seq, range);
513 CERROR("%s: Can not lookup fld for %#llx\n",
518 RETURN(ss->ss_node_id == range->lsr_index);
521 int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
522 const struct lu_fid *fid)
524 struct seq_server_site *ss = osd_seq_site(osd);
527 /* FID seqs not in FLDB, must be local seq */
528 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
531 /* If FLD is not being initialized yet, it only happens during the
532 * initialization, likely during mgs initialization, and we assume
533 * this is local FID. */
534 if (ss == NULL || ss->ss_server_fld == NULL)
537 /* Only check the local FLDB here */
538 if (osd_seq_exists(env, osd, fid_seq(fid)))
545 * Inserts (key, value) pair in \a directory object.
547 * \param dt osd index object
548 * \param key key for index
549 * \param rec record reference
550 * \param th transaction handler
551 * \param ignore_quota update should not affect quota
554 * \retval -ve failure
556 static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
557 const struct dt_rec *rec, const struct dt_key *key,
558 struct thandle *th, int ignore_quota)
560 struct osd_thread_info *oti = osd_oti_get(env);
561 struct osd_object *parent = osd_dt_obj(dt);
562 struct osd_device *osd = osd_obj2dev(parent);
563 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
564 const struct lu_fid *fid = rec1->rec_fid;
565 struct osd_thandle *oh;
566 struct osd_idmap_cache *idc;
567 char *name = (char *)key;
571 LASSERT(parent->oo_dn);
573 LASSERT(dt_object_exists(dt));
574 LASSERT(osd_invariant(parent));
577 oh = container_of0(th, struct osd_thandle, ot_super);
579 idc = osd_idc_find(env, osd, fid);
580 if (unlikely(idc == NULL)) {
581 /* this dt_insert() wasn't declared properly, so
582 * FID is missing in OI cache. we better do not
583 * lookup FID in FLDB/OI and don't risk to deadlock,
584 * but in some special cases (lfsck testing, etc)
585 * it's much simpler than fixing a caller */
586 CERROR("%s: "DFID" wasn't declared for insert\n",
587 osd_name(osd), PFID(fid));
588 idc = osd_idc_find_or_init(env, osd, fid);
590 RETURN(PTR_ERR(idc));
593 if (idc->oic_remote) {
594 /* Insert remote entry */
595 memset(&oti->oti_zde.lzd_reg, 0, sizeof(oti->oti_zde.lzd_reg));
596 oti->oti_zde.lzd_reg.zde_type = IFTODT(rec1->rec_type & S_IFMT);
598 if (unlikely(idc->oic_dnode == 0)) {
599 /* for a reason OI cache wasn't filled properly */
600 CERROR("%s: OIC for "DFID" isn't filled\n",
601 osd_name(osd), PFID(fid));
604 if (name[0] == '.') {
606 /* do not store ".", instead generate it
607 * during iteration */
609 } else if (name[1] == '.' && name[2] == 0) {
610 uint64_t dnode = idc->oic_dnode;
611 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT))
614 /* update parent dnode in the child.
615 * later it will be used to generate ".." */
616 rc = osd_object_sa_update(parent,
623 CLASSERT(sizeof(oti->oti_zde.lzd_reg) == 8);
624 CLASSERT(sizeof(oti->oti_zde) % 8 == 0);
625 oti->oti_zde.lzd_reg.zde_type = IFTODT(rec1->rec_type & S_IFMT);
626 oti->oti_zde.lzd_reg.zde_dnode = idc->oic_dnode;
629 oti->oti_zde.lzd_fid = *fid;
630 /* Insert (key,oid) into ZAP */
631 rc = osd_zap_add(osd, parent->oo_dn->dn_object, parent->oo_dn,
632 (char *)key, 8, sizeof(oti->oti_zde) / 8,
633 (void *)&oti->oti_zde, oh->ot_tx);
634 if (unlikely(rc == -EEXIST &&
635 name[0] == '.' && name[1] == '.' && name[2] == 0))
636 /* Update (key,oid) in ZAP */
637 rc = -zap_update(osd->od_os, parent->oo_dn->dn_object,
638 (char *)key, 8, sizeof(oti->oti_zde) / 8,
639 (void *)&oti->oti_zde, oh->ot_tx);
646 static int osd_declare_dir_delete(const struct lu_env *env,
647 struct dt_object *dt,
648 const struct dt_key *key,
651 struct osd_object *obj = osd_dt_obj(dt);
652 struct osd_thandle *oh;
656 LASSERT(dt_object_exists(dt));
657 LASSERT(osd_invariant(obj));
660 oh = container_of0(th, struct osd_thandle, ot_super);
662 if (dt_object_exists(dt)) {
664 dnode = obj->oo_dn->dn_object;
666 dnode = DMU_NEW_OBJECT;
669 /* do not specify the key as then DMU is trying to look it up
670 * which is very expensive. usually the layers above lookup
672 osd_tx_hold_zap(oh->ot_tx, dnode, obj->oo_dn, FALSE, NULL);
677 static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
678 const struct dt_key *key, struct thandle *th)
680 struct osd_object *obj = osd_dt_obj(dt);
681 struct osd_device *osd = osd_obj2dev(obj);
682 struct osd_thandle *oh;
683 dnode_t *zap_dn = obj->oo_dn;
684 char *name = (char *)key;
691 oh = container_of0(th, struct osd_thandle, ot_super);
694 * In Orion . and .. were stored in the directory (not generated upon
695 * request as now). we preserve them for backward compatibility
697 if (name[0] == '.') {
700 } else if (name[1] == '.' && name[2] == 0) {
705 /* Remove key from the ZAP */
706 rc = osd_zap_remove(osd, zap_dn->dn_object, zap_dn,
707 (char *)key, oh->ot_tx);
709 if (unlikely(rc && rc != -ENOENT))
710 CERROR("%s: zap_remove failed: rc = %d\n", osd->od_svname, rc);
715 static struct dt_it *osd_dir_it_init(const struct lu_env *env,
716 struct dt_object *dt,
719 struct osd_zap_it *it;
721 it = (struct osd_zap_it *)osd_index_it_init(env, dt, unused);
725 RETURN((struct dt_it *)it);
729 * Move Iterator to record specified by \a key
731 * \param di osd iterator
732 * \param key key for index
734 * \retval +ve di points to record with least key not larger than key
735 * \retval 0 di points to exact matched key
736 * \retval -ve failure
738 static int osd_dir_it_get(const struct lu_env *env,
739 struct dt_it *di, const struct dt_key *key)
741 struct osd_zap_it *it = (struct osd_zap_it *)di;
742 struct osd_object *obj = it->ozi_obj;
743 char *name = (char *)key;
750 /* reset the cursor */
751 zap_cursor_fini(it->ozi_zc);
752 osd_obj_cursor_init_serialized(it->ozi_zc, obj, 0);
754 /* XXX: implementation of the API is broken at the moment */
755 LASSERT(((const char *)key)[0] == 0);
762 if (name[0] == '.') {
766 } else if (name[1] == '.' && name[2] == 0) {
772 /* neither . nor .. - some real record */
780 static void osd_dir_it_put(const struct lu_env *env, struct dt_it *di)
782 /* PBS: do nothing : ref are incremented at retrive and decreamented
787 * in Orion . and .. were stored in the directory, while ZPL
788 * and current osd-zfs generate them up on request. so, we
789 * need to ignore previously stored . and ..
791 static int osd_index_retrieve_skip_dots(struct osd_zap_it *it,
797 rc = -zap_cursor_retrieve(it->ozi_zc, za);
800 if (unlikely(rc == 0 && za->za_name[0] == '.')) {
801 if (za->za_name[1] == 0) {
803 } else if (za->za_name[1] == '.' &&
804 za->za_name[2] == 0) {
808 zap_cursor_advance(it->ozi_zc);
810 } while (unlikely(rc == 0 && isdot));
816 * to load a directory entry at a time and stored it in
817 * iterator's in-memory data structure.
819 * \param di, struct osd_it_ea, iterator's in memory structure
821 * \retval +ve, iterator reached to end
822 * \retval 0, iterator not reached to end
823 * \retval -ve, on error
825 static int osd_dir_it_next(const struct lu_env *env, struct dt_it *di)
827 struct osd_zap_it *it = (struct osd_zap_it *)di;
828 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
833 /* temp. storage should be enough for any key supported by ZFS */
834 CLASSERT(sizeof(za->za_name) <= sizeof(it->ozi_name));
837 * the first ->next() moves the cursor to .
838 * the second ->next() moves the cursor to ..
839 * then we get to the real records and have to verify any exist
841 if (it->ozi_pos <= 2) {
847 zap_cursor_advance(it->ozi_zc);
851 * According to current API we need to return error if its last entry.
852 * zap_cursor_advance() does not return any value. So we need to call
853 * retrieve to check if there is any record. We should make
854 * changes to Iterator API to not return status for this API
856 rc = osd_index_retrieve_skip_dots(it, za);
858 if (rc == -ENOENT) /* end of dir */
864 static struct dt_key *osd_dir_it_key(const struct lu_env *env,
865 const struct dt_it *di)
867 struct osd_zap_it *it = (struct osd_zap_it *)di;
868 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
872 if (it->ozi_pos <= 1) {
874 RETURN((struct dt_key *)".");
875 } else if (it->ozi_pos == 2) {
876 RETURN((struct dt_key *)"..");
879 if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)))
882 strcpy(it->ozi_name, za->za_name);
884 RETURN((struct dt_key *)it->ozi_name);
887 static int osd_dir_it_key_size(const struct lu_env *env, const struct dt_it *di)
889 struct osd_zap_it *it = (struct osd_zap_it *)di;
890 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
894 if (it->ozi_pos <= 1) {
897 } else if (it->ozi_pos == 2) {
901 if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)) == 0)
902 rc = strlen(za->za_name);
907 static int osd_dir_it_rec(const struct lu_env *env, const struct dt_it *di,
908 struct dt_rec *dtrec, __u32 attr)
910 struct osd_zap_it *it = (struct osd_zap_it *)di;
911 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
912 struct luz_direntry *zde = &osd_oti_get(env)->oti_zde;
913 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
917 if (it->ozi_pos <= 1) {
918 lde->lde_hash = cpu_to_le64(1);
919 strcpy(lde->lde_name, ".");
920 lde->lde_namelen = cpu_to_le16(1);
921 lde->lde_fid = *lu_object_fid(&it->ozi_obj->oo_dt.do_lu);
922 lde->lde_attrs = LUDA_FID;
923 /* append lustre attributes */
924 osd_it_append_attrs(lde, attr, 1, IFTODT(S_IFDIR));
925 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(1, attr));
929 } else if (it->ozi_pos == 2) {
930 lde->lde_hash = cpu_to_le64(2);
931 strcpy(lde->lde_name, "..");
932 lde->lde_namelen = cpu_to_le16(2);
933 lde->lde_attrs = LUDA_FID;
934 /* append lustre attributes */
935 osd_it_append_attrs(lde, attr, 2, IFTODT(S_IFDIR));
936 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(2, attr));
937 rc = osd_find_parent_fid(env, &it->ozi_obj->oo_dt, &lde->lde_fid);
939 /* ENOENT happens at the root of filesystem so ignore it */
947 rc = -zap_cursor_retrieve(it->ozi_zc, za);
948 if (unlikely(rc != 0))
951 lde->lde_hash = cpu_to_le64(osd_zap_cursor_serialize(it->ozi_zc));
952 namelen = strlen(za->za_name);
953 if (namelen > NAME_MAX)
954 GOTO(out, rc = -EOVERFLOW);
955 strcpy(lde->lde_name, za->za_name);
956 lde->lde_namelen = cpu_to_le16(namelen);
958 if (za->za_integer_length != 8 || za->za_num_integers < 3) {
959 CERROR("%s: unsupported direntry format: %d %d\n",
960 osd_obj2dev(it->ozi_obj)->od_svname,
961 za->za_integer_length, (int)za->za_num_integers);
963 GOTO(out, rc = -EIO);
966 rc = osd_zap_lookup(osd_obj2dev(it->ozi_obj), it->ozi_zc->zc_zapobj,
967 it->ozi_obj->oo_dn, za->za_name,
968 za->za_integer_length, 3, zde);
972 lde->lde_fid = zde->lzd_fid;
973 lde->lde_attrs = LUDA_FID;
975 /* append lustre attributes */
976 osd_it_append_attrs(lde, attr, namelen, zde->lzd_reg.zde_type);
978 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
984 static int osd_dir_it_rec_size(const struct lu_env *env, const struct dt_it *di,
987 struct osd_zap_it *it = (struct osd_zap_it *)di;
988 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
993 if (it->ozi_pos <= 1)
995 else if (it->ozi_pos == 2)
999 rc = lu_dirent_calc_size(namelen, attr);
1003 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1004 if (unlikely(rc != 0))
1007 if (za->za_integer_length != 8 || za->za_num_integers < 3) {
1008 CERROR("%s: unsupported direntry format: %d %d\n",
1009 osd_obj2dev(it->ozi_obj)->od_svname,
1010 za->za_integer_length, (int)za->za_num_integers);
1014 namelen = strlen(za->za_name);
1015 if (namelen > NAME_MAX)
1018 rc = lu_dirent_calc_size(namelen, attr);
1023 static __u64 osd_dir_it_store(const struct lu_env *env, const struct dt_it *di)
1025 struct osd_zap_it *it = (struct osd_zap_it *)di;
1029 if (it->ozi_pos <= 2)
1032 pos = osd_zap_cursor_serialize(it->ozi_zc);
1039 * rc == 0 -> end of directory.
1040 * rc > 0 -> ok, proceed.
1041 * rc < 0 -> error. ( EOVERFLOW can be masked.)
1043 static int osd_dir_it_load(const struct lu_env *env,
1044 const struct dt_it *di, __u64 hash)
1046 struct osd_zap_it *it = (struct osd_zap_it *)di;
1047 struct osd_object *obj = it->ozi_obj;
1048 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1052 /* reset the cursor */
1053 zap_cursor_fini(it->ozi_zc);
1054 osd_obj_cursor_init_serialized(it->ozi_zc, obj, hash);
1061 /* to return whether the end has been reached */
1062 rc = osd_index_retrieve_skip_dots(it, za);
1065 else if (rc == -ENOENT)
1072 struct dt_index_operations osd_dir_ops = {
1073 .dio_lookup = osd_dir_lookup,
1074 .dio_declare_insert = osd_declare_dir_insert,
1075 .dio_insert = osd_dir_insert,
1076 .dio_declare_delete = osd_declare_dir_delete,
1077 .dio_delete = osd_dir_delete,
1079 .init = osd_dir_it_init,
1080 .fini = osd_index_it_fini,
1081 .get = osd_dir_it_get,
1082 .put = osd_dir_it_put,
1083 .next = osd_dir_it_next,
1084 .key = osd_dir_it_key,
1085 .key_size = osd_dir_it_key_size,
1086 .rec = osd_dir_it_rec,
1087 .rec_size = osd_dir_it_rec_size,
1088 .store = osd_dir_it_store,
1089 .load = osd_dir_it_load
1094 * Primitives for index files using binary keys.
1097 /* key integer_size is 8 */
1098 static int osd_prepare_key_uint64(struct osd_object *o, __u64 *dst,
1099 const struct dt_key *src)
1106 /* align keysize to 64bit */
1107 size = (o->oo_keysize + sizeof(__u64) - 1) / sizeof(__u64);
1108 size *= sizeof(__u64);
1110 LASSERT(size <= MAXNAMELEN);
1112 if (unlikely(size > o->oo_keysize))
1113 memset(dst + o->oo_keysize, 0, size - o->oo_keysize);
1114 memcpy(dst, (const char *)src, o->oo_keysize);
1116 return (size/sizeof(__u64));
1119 static int osd_index_lookup(const struct lu_env *env, struct dt_object *dt,
1120 struct dt_rec *rec, const struct dt_key *key)
1122 struct osd_object *obj = osd_dt_obj(dt);
1123 struct osd_device *osd = osd_obj2dev(obj);
1124 __u64 *k = osd_oti_get(env)->oti_key64;
1128 rc = osd_prepare_key_uint64(obj, k, key);
1130 rc = -zap_lookup_uint64(osd->od_os, obj->oo_dn->dn_object,
1131 k, rc, obj->oo_recusize, obj->oo_recsize,
1133 RETURN(rc == 0 ? 1 : rc);
1136 static int osd_declare_index_insert(const struct lu_env *env,
1137 struct dt_object *dt,
1138 const struct dt_rec *rec,
1139 const struct dt_key *key,
1142 struct osd_object *obj = osd_dt_obj(dt);
1143 struct osd_thandle *oh;
1146 LASSERT(th != NULL);
1147 oh = container_of0(th, struct osd_thandle, ot_super);
1149 LASSERT(obj->oo_dn);
1151 /* do not specify the key as then DMU is trying to look it up
1152 * which is very expensive. usually the layers above lookup
1153 * before insertion */
1154 osd_tx_hold_zap(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
1160 static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
1161 const struct dt_rec *rec, const struct dt_key *key,
1162 struct thandle *th, int ignore_quota)
1164 struct osd_object *obj = osd_dt_obj(dt);
1165 struct osd_device *osd = osd_obj2dev(obj);
1166 struct osd_thandle *oh;
1167 __u64 *k = osd_oti_get(env)->oti_key64;
1171 LASSERT(obj->oo_dn);
1172 LASSERT(dt_object_exists(dt));
1173 LASSERT(osd_invariant(obj));
1174 LASSERT(th != NULL);
1176 oh = container_of0(th, struct osd_thandle, ot_super);
1178 rc = osd_prepare_key_uint64(obj, k, key);
1180 /* Insert (key,oid) into ZAP */
1181 rc = -zap_add_uint64(osd->od_os, obj->oo_dn->dn_object,
1182 k, rc, obj->oo_recusize, obj->oo_recsize,
1183 (void *)rec, oh->ot_tx);
1187 static int osd_declare_index_delete(const struct lu_env *env,
1188 struct dt_object *dt,
1189 const struct dt_key *key,
1192 struct osd_object *obj = osd_dt_obj(dt);
1193 struct osd_thandle *oh;
1196 LASSERT(dt_object_exists(dt));
1197 LASSERT(osd_invariant(obj));
1198 LASSERT(th != NULL);
1199 LASSERT(obj->oo_dn);
1201 oh = container_of0(th, struct osd_thandle, ot_super);
1203 /* do not specify the key as then DMU is trying to look it up
1204 * which is very expensive. usually the layers above lookup
1205 * before deletion */
1206 osd_tx_hold_zap(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
1212 static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
1213 const struct dt_key *key, struct thandle *th)
1215 struct osd_object *obj = osd_dt_obj(dt);
1216 struct osd_device *osd = osd_obj2dev(obj);
1217 struct osd_thandle *oh;
1218 __u64 *k = osd_oti_get(env)->oti_key64;
1222 LASSERT(obj->oo_dn);
1223 LASSERT(th != NULL);
1224 oh = container_of0(th, struct osd_thandle, ot_super);
1226 rc = osd_prepare_key_uint64(obj, k, key);
1228 /* Remove binary key from the ZAP */
1229 rc = -zap_remove_uint64(osd->od_os, obj->oo_dn->dn_object,
1234 static int osd_index_it_get(const struct lu_env *env, struct dt_it *di,
1235 const struct dt_key *key)
1237 struct osd_zap_it *it = (struct osd_zap_it *)di;
1238 struct osd_object *obj = it->ozi_obj;
1239 struct osd_device *osd = osd_obj2dev(obj);
1243 LASSERT(it->ozi_zc);
1246 * XXX: we need a binary version of zap_cursor_move_to_key()
1247 * to implement this API */
1248 if (*((const __u64 *)key) != 0)
1249 CERROR("NOT IMPLEMETED YET (move to %#llx)\n",
1252 zap_cursor_fini(it->ozi_zc);
1253 zap_cursor_init(it->ozi_zc, osd->od_os, obj->oo_dn->dn_object);
1259 static int osd_index_it_next(const struct lu_env *env, struct dt_it *di)
1261 struct osd_zap_it *it = (struct osd_zap_it *)di;
1262 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1266 if (it->ozi_reset == 0)
1267 zap_cursor_advance(it->ozi_zc);
1271 * According to current API we need to return error if it's last entry.
1272 * zap_cursor_advance() does not return any value. So we need to call
1273 * retrieve to check if there is any record. We should make
1274 * changes to Iterator API to not return status for this API
1276 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1283 static struct dt_key *osd_index_it_key(const struct lu_env *env,
1284 const struct dt_it *di)
1286 struct osd_zap_it *it = (struct osd_zap_it *)di;
1287 struct osd_object *obj = it->ozi_obj;
1288 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1293 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1295 RETURN(ERR_PTR(rc));
1297 /* the binary key is stored in the name */
1298 memcpy(&it->ozi_key, za->za_name, obj->oo_keysize);
1300 RETURN((struct dt_key *)&it->ozi_key);
1303 static int osd_index_it_key_size(const struct lu_env *env,
1304 const struct dt_it *di)
1306 struct osd_zap_it *it = (struct osd_zap_it *)di;
1307 struct osd_object *obj = it->ozi_obj;
1308 RETURN(obj->oo_keysize);
1311 static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
1312 struct dt_rec *rec, __u32 attr)
1314 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1315 struct osd_zap_it *it = (struct osd_zap_it *)di;
1316 struct osd_object *obj = it->ozi_obj;
1317 struct osd_device *osd = osd_obj2dev(obj);
1318 __u64 *k = osd_oti_get(env)->oti_key64;
1323 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1327 rc = osd_prepare_key_uint64(obj, k, (const struct dt_key *)za->za_name);
1329 rc = -zap_lookup_uint64(osd->od_os, obj->oo_dn->dn_object,
1330 k, rc, obj->oo_recusize, obj->oo_recsize,
1335 static __u64 osd_index_it_store(const struct lu_env *env,
1336 const struct dt_it *di)
1338 struct osd_zap_it *it = (struct osd_zap_it *)di;
1341 RETURN((__u64)zap_cursor_serialize(it->ozi_zc));
1344 static int osd_index_it_load(const struct lu_env *env, const struct dt_it *di,
1347 struct osd_zap_it *it = (struct osd_zap_it *)di;
1348 struct osd_object *obj = it->ozi_obj;
1349 struct osd_device *osd = osd_obj2dev(obj);
1350 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1354 /* reset the cursor */
1355 zap_cursor_fini(it->ozi_zc);
1356 zap_cursor_init_serialized(it->ozi_zc, osd->od_os,
1357 obj->oo_dn->dn_object, hash);
1360 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1363 else if (rc == -ENOENT)
1369 static struct dt_index_operations osd_index_ops = {
1370 .dio_lookup = osd_index_lookup,
1371 .dio_declare_insert = osd_declare_index_insert,
1372 .dio_insert = osd_index_insert,
1373 .dio_declare_delete = osd_declare_index_delete,
1374 .dio_delete = osd_index_delete,
1376 .init = osd_index_it_init,
1377 .fini = osd_index_it_fini,
1378 .get = osd_index_it_get,
1379 .put = osd_index_it_put,
1380 .next = osd_index_it_next,
1381 .key = osd_index_it_key,
1382 .key_size = osd_index_it_key_size,
1383 .rec = osd_index_it_rec,
1384 .store = osd_index_it_store,
1385 .load = osd_index_it_load
1389 struct osd_metadnode_it {
1390 struct osd_device *mit_dev;
1392 struct lu_fid mit_fid;
1394 __u64 mit_prefetched_dnode;
1397 static struct dt_it *osd_zfs_otable_it_init(const struct lu_env *env,
1398 struct dt_object *dt, __u32 attr)
1400 struct osd_device *dev = osd_dev(dt->do_lu.lo_dev);
1401 struct osd_metadnode_it *it;
1405 if (unlikely(it == NULL))
1406 RETURN(ERR_PTR(-ENOMEM));
1410 /* XXX: dmu_object_next() does NOT find dnodes allocated
1411 * in the current non-committed txg, so we force txg
1412 * commit to find all existing dnodes ... */
1413 if (!dev->od_dt_dev.dd_rdonly)
1414 txg_wait_synced(dmu_objset_pool(dev->od_os), 0ULL);
1416 RETURN((struct dt_it *)it);
1419 static void osd_zfs_otable_it_fini(const struct lu_env *env, struct dt_it *di)
1421 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1426 static int osd_zfs_otable_it_get(const struct lu_env *env,
1427 struct dt_it *di, const struct dt_key *key)
1432 static void osd_zfs_otable_it_put(const struct lu_env *env, struct dt_it *di)
1436 #define OTABLE_PREFETCH 256
1438 static void osd_zfs_otable_prefetch(const struct lu_env *env,
1439 struct osd_metadnode_it *it)
1441 struct osd_device *dev = it->mit_dev;
1444 /* can go negative on the very first access to the iterator
1445 * or if some non-Lustre objects were found */
1446 if (unlikely(it->mit_prefetched < 0))
1447 it->mit_prefetched = 0;
1449 if (it->mit_prefetched >= (OTABLE_PREFETCH >> 1))
1452 if (it->mit_prefetched_dnode == 0)
1453 it->mit_prefetched_dnode = it->mit_pos;
1455 while (it->mit_prefetched < OTABLE_PREFETCH) {
1456 rc = -dmu_object_next(dev->od_os, &it->mit_prefetched_dnode,
1458 if (unlikely(rc != 0))
1461 osd_dmu_prefetch(dev->od_os, it->mit_prefetched_dnode,
1462 0, 0, 0, ZIO_PRIORITY_ASYNC_READ);
1464 it->mit_prefetched++;
1468 static int osd_zfs_otable_it_next(const struct lu_env *env, struct dt_it *di)
1470 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1471 struct lustre_mdt_attrs *lma;
1472 struct osd_device *dev = it->mit_dev;
1473 nvlist_t *nvbuf = NULL;
1478 memset(&it->mit_fid, 0, sizeof(it->mit_fid));
1480 dnode = it->mit_pos;
1482 rc = -dmu_object_next(dev->od_os, &it->mit_pos, B_FALSE, 0);
1483 if (unlikely(rc != 0))
1485 it->mit_prefetched--;
1487 /* LMA is required for this to be a Lustre object.
1488 * If there is no xattr skip it. */
1489 rc = __osd_xattr_load_by_oid(dev, it->mit_pos, &nvbuf);
1490 if (unlikely(rc != 0))
1493 LASSERT(nvbuf != NULL);
1494 rc = -nvlist_lookup_byte_array(nvbuf, XATTR_NAME_LMA, &v, &s);
1495 if (likely(rc == 0)) {
1497 lma = (struct lustre_mdt_attrs *)v;
1498 lustre_lma_swab(lma);
1499 it->mit_fid = lma->lma_self_fid;
1503 /* not a Lustre object, try next one */
1510 /* we aren't prefetching in the above loop because the number of
1511 * non-Lustre objects is very small and we will be repeating very
1512 * rare. in case we want to use this to iterate over non-Lustre
1513 * objects (i.e. when we convert regular ZFS in Lustre) it makes
1514 * sense to initiate prefetching in the loop */
1516 /* 0 - there are more items, +1 - the end */
1517 if (likely(rc == 0))
1518 osd_zfs_otable_prefetch(env, it);
1520 CDEBUG(D_OTHER, "advance: %llu -> %llu "DFID": %d\n", dnode,
1521 it->mit_pos, PFID(&it->mit_fid), rc);
1527 static struct dt_key *osd_zfs_otable_it_key(const struct lu_env *env,
1528 const struct dt_it *di)
1533 static int osd_zfs_otable_it_key_size(const struct lu_env *env,
1534 const struct dt_it *di)
1536 return sizeof(__u64);
1539 static int osd_zfs_otable_it_rec(const struct lu_env *env,
1540 const struct dt_it *di,
1541 struct dt_rec *rec, __u32 attr)
1543 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1544 struct lu_fid *fid = (struct lu_fid *)rec;
1553 static __u64 osd_zfs_otable_it_store(const struct lu_env *env,
1554 const struct dt_it *di)
1556 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1561 static int osd_zfs_otable_it_load(const struct lu_env *env,
1562 const struct dt_it *di, __u64 hash)
1564 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1567 it->mit_prefetched = 0;
1568 it->mit_prefetched_dnode = 0;
1570 return osd_zfs_otable_it_next(env, (struct dt_it *)di);
1573 static int osd_zfs_otable_it_key_rec(const struct lu_env *env,
1574 const struct dt_it *di, void *key_rec)
1579 const struct dt_index_operations osd_zfs_otable_ops = {
1581 .init = osd_zfs_otable_it_init,
1582 .fini = osd_zfs_otable_it_fini,
1583 .get = osd_zfs_otable_it_get,
1584 .put = osd_zfs_otable_it_put,
1585 .next = osd_zfs_otable_it_next,
1586 .key = osd_zfs_otable_it_key,
1587 .key_size = osd_zfs_otable_it_key_size,
1588 .rec = osd_zfs_otable_it_rec,
1589 .store = osd_zfs_otable_it_store,
1590 .load = osd_zfs_otable_it_load,
1591 .key_rec = osd_zfs_otable_it_key_rec,
1595 int osd_index_try(const struct lu_env *env, struct dt_object *dt,
1596 const struct dt_index_features *feat)
1598 struct osd_object *obj = osd_dt_obj(dt);
1602 down_read(&obj->oo_guard);
1605 * XXX: implement support for fixed-size keys sorted with natural
1606 * numerical way (not using internal hash value)
1608 if (feat->dif_flags & DT_IND_RANGE)
1609 GOTO(out, rc = -ERANGE);
1611 if (unlikely(feat == &dt_otable_features)) {
1612 dt->do_index_ops = &osd_zfs_otable_ops;
1616 LASSERT(!dt_object_exists(dt) || obj->oo_dn != NULL);
1617 if (likely(feat == &dt_directory_features)) {
1618 if (!dt_object_exists(dt) || osd_object_is_zap(obj->oo_dn))
1619 dt->do_index_ops = &osd_dir_ops;
1621 GOTO(out, rc = -ENOTDIR);
1622 } else if (unlikely(feat == &dt_acct_features)) {
1623 LASSERT(fid_is_acct(lu_object_fid(&dt->do_lu)));
1624 dt->do_index_ops = &osd_acct_index_ops;
1625 } else if (dt->do_index_ops == NULL) {
1626 /* For index file, we don't support variable key & record sizes
1627 * and the key has to be unique */
1628 if ((feat->dif_flags & ~DT_IND_UPDATE) != 0)
1629 GOTO(out, rc = -EINVAL);
1631 if (feat->dif_keysize_max > ZAP_MAXNAMELEN)
1632 GOTO(out, rc = -E2BIG);
1633 if (feat->dif_keysize_max != feat->dif_keysize_min)
1634 GOTO(out, rc = -EINVAL);
1636 /* As for the record size, it should be a multiple of 8 bytes
1637 * and smaller than the maximum value length supported by ZAP.
1639 if (feat->dif_recsize_max > ZAP_MAXVALUELEN)
1640 GOTO(out, rc = -E2BIG);
1641 if (feat->dif_recsize_max != feat->dif_recsize_min)
1642 GOTO(out, rc = -EINVAL);
1644 obj->oo_keysize = feat->dif_keysize_max;
1645 obj->oo_recsize = feat->dif_recsize_max;
1646 obj->oo_recusize = 1;
1648 /* ZFS prefers to work with array of 64bits */
1649 if ((obj->oo_recsize & 7) == 0) {
1650 obj->oo_recsize >>= 3;
1651 obj->oo_recusize = 8;
1653 dt->do_index_ops = &osd_index_ops;
1657 up_read(&obj->oo_guard);