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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osd-zfs/osd_index.c
38 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
39 * Author: Mike Pershin <tappro@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_OSD
44 #include <lustre_ver.h>
45 #include <libcfs/libcfs.h>
46 #include <obd_support.h>
47 #include <lustre_net.h>
49 #include <obd_class.h>
50 #include <lustre_disk.h>
51 #include <lustre_fid.h>
53 #include "osd_internal.h"
55 #include <sys/dnode.h>
60 #include <sys/spa_impl.h>
61 #include <sys/zfs_znode.h>
62 #include <sys/dmu_tx.h>
63 #include <sys/dmu_objset.h>
64 #include <sys/dsl_prop.h>
65 #include <sys/sa_impl.h>
68 static inline int osd_object_is_zap(dmu_buf_t *db)
70 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *) db;
76 rc = (dn->dn_type == DMU_OT_DIRECTORY_CONTENTS ||
77 dn->dn_type == DMU_OT_USERGROUP_USED);
83 /* We don't actually have direct access to the zap_hashbits() function
84 * so just pretend like we do for now. If this ever breaks we can look at
86 #define zap_hashbits(zc) 48
89 * | cd (16 bits) | hash (48 bits) |
90 * we need it in other form:
91 * |0| hash (48 bit) | cd (15 bit) |
92 * to be a full 64-bit ordered hash so that Lustre readdir can use it to merge
93 * the readdir hashes from multiple directory stripes uniformly on the client.
94 * Another point is sign bit, the hash range should be in [0, 2^63-1] because
95 * loff_t (for llseek) needs to be a positive value. This means the "cd" field
96 * should only be the low 15 bits.
98 uint64_t osd_zap_cursor_serialize(zap_cursor_t *zc)
100 uint64_t zfs_hash = zap_cursor_serialize(zc) & (~0ULL >> 1);
102 return (zfs_hash >> zap_hashbits(zc)) |
103 (zfs_hash << (63 - zap_hashbits(zc)));
106 void osd_zap_cursor_init_serialized(zap_cursor_t *zc, struct objset *os,
107 uint64_t id, uint64_t dirhash)
109 uint64_t zfs_hash = ((dirhash << zap_hashbits(zc)) & (~0ULL >> 1)) |
110 (dirhash >> (63 - zap_hashbits(zc)));
112 zap_cursor_init_serialized(zc, os, id, zfs_hash);
115 int osd_zap_cursor_init(zap_cursor_t **zc, struct objset *os,
116 uint64_t id, uint64_t dirhash)
121 if (unlikely(t == NULL))
124 osd_zap_cursor_init_serialized(t, os, id, dirhash);
130 void osd_zap_cursor_fini(zap_cursor_t *zc)
136 static inline void osd_obj_cursor_init_serialized(zap_cursor_t *zc,
137 struct osd_object *o,
140 struct osd_device *d = osd_obj2dev(o);
141 osd_zap_cursor_init_serialized(zc, d->od_os,
142 o->oo_db->db_object, dirhash);
145 static inline int osd_obj_cursor_init(zap_cursor_t **zc, struct osd_object *o,
148 struct osd_device *d = osd_obj2dev(o);
149 return osd_zap_cursor_init(zc, d->od_os, o->oo_db->db_object, dirhash);
152 static struct dt_it *osd_index_it_init(const struct lu_env *env,
153 struct dt_object *dt,
155 struct lustre_capa *capa)
157 struct osd_thread_info *info = osd_oti_get(env);
158 struct osd_zap_it *it;
159 struct osd_object *obj = osd_dt_obj(dt);
160 struct lu_object *lo = &dt->do_lu;
164 /* XXX: check capa ? */
166 LASSERT(lu_object_exists(lo));
168 LASSERT(osd_object_is_zap(obj->oo_db));
171 OBD_SLAB_ALLOC_PTR_GFP(it, osd_zapit_cachep, GFP_NOFS);
173 RETURN(ERR_PTR(-ENOMEM));
175 rc = osd_obj_cursor_init(&it->ozi_zc, obj, 0);
177 OBD_SLAB_FREE_PTR(it, osd_zapit_cachep);
186 RETURN((struct dt_it *)it);
189 static void osd_index_it_fini(const struct lu_env *env, struct dt_it *di)
191 struct osd_zap_it *it = (struct osd_zap_it *)di;
192 struct osd_object *obj;
196 LASSERT(it->ozi_obj);
200 osd_zap_cursor_fini(it->ozi_zc);
201 lu_object_put(env, &obj->oo_dt.do_lu);
202 OBD_SLAB_FREE_PTR(it, osd_zapit_cachep);
208 static void osd_index_it_put(const struct lu_env *env, struct dt_it *di)
210 /* PBS: do nothing : ref are incremented at retrive and decreamented
214 static inline void osd_it_append_attrs(struct lu_dirent *ent, __u32 attr,
217 const unsigned align = sizeof(struct luda_type) - 1;
218 struct luda_type *lt;
220 /* check if file type is required */
221 if (attr & LUDA_TYPE) {
222 len = (len + align) & ~align;
224 lt = (void *)ent->lde_name + len;
225 lt->lt_type = cpu_to_le16(DTTOIF(type));
226 ent->lde_attrs |= LUDA_TYPE;
229 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
233 * as we don't know FID, we can't use LU object, so this function
234 * partially duplicate __osd_xattr_get() which is built around
235 * LU-object and uses it to cache data like regular EA dnode, etc
237 static int osd_find_parent_by_dnode(const struct lu_env *env,
241 struct osd_device *osd = osd_obj2dev(osd_dt_obj(o));
242 struct lustre_mdt_attrs *lma;
245 nvlist_t *nvbuf = NULL;
251 /* first of all, get parent dnode from own attributes */
252 LASSERT(osd_dt_obj(o)->oo_db);
253 rc = -sa_handle_get(osd->od_os, osd_dt_obj(o)->oo_db->db_object,
254 NULL, SA_HDL_PRIVATE, &sa_hdl);
258 dnode = ZFS_NO_OBJECT;
259 rc = -sa_lookup(sa_hdl, SA_ZPL_PARENT(osd), &dnode, 8);
260 sa_handle_destroy(sa_hdl);
264 /* now get EA buffer */
265 rc = __osd_xattr_load(osd, dnode, &nvbuf);
269 /* XXX: if we get that far.. should we cache the result? */
271 /* try to find LMA attribute */
272 LASSERT(nvbuf != NULL);
273 rc = -nvlist_lookup_byte_array(nvbuf, XATTR_NAME_LMA, &value, &size);
274 if (rc == 0 && size >= sizeof(*lma)) {
275 lma = (struct lustre_mdt_attrs *)value;
276 lustre_lma_swab(lma);
277 *fid = lma->lma_self_fid;
282 /* no LMA attribute in SA, let's try regular EA */
284 /* first of all, get parent dnode storing regular EA */
285 rc = -sa_handle_get(osd->od_os, dnode, NULL, SA_HDL_PRIVATE, &sa_hdl);
289 dnode = ZFS_NO_OBJECT;
290 rc = -sa_lookup(sa_hdl, SA_ZPL_XATTR(osd), &dnode, 8);
291 sa_handle_destroy(sa_hdl);
295 CLASSERT(sizeof(*lma) <= sizeof(osd_oti_get(env)->oti_buf));
296 buf.lb_buf = osd_oti_get(env)->oti_buf;
297 buf.lb_len = sizeof(osd_oti_get(env)->oti_buf);
299 /* now try to find LMA */
300 rc = __osd_xattr_get_large(env, osd, dnode, &buf,
301 XATTR_NAME_LMA, &size);
302 if (rc == 0 && size >= sizeof(*lma)) {
304 lustre_lma_swab(lma);
305 *fid = lma->lma_self_fid;
310 GOTO(out, rc = -EIO);
319 static int osd_find_parent_fid(const struct lu_env *env, struct dt_object *o,
322 struct link_ea_header *leh;
323 struct link_ea_entry *lee;
328 buf.lb_buf = osd_oti_get(env)->oti_buf;
329 buf.lb_len = sizeof(osd_oti_get(env)->oti_buf);
331 rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK, BYPASS_CAPA);
333 rc = osd_xattr_get(env, o, &LU_BUF_NULL,
334 XATTR_NAME_LINK, BYPASS_CAPA);
338 OBD_ALLOC(buf.lb_buf, rc);
339 if (buf.lb_buf == NULL)
342 rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK, BYPASS_CAPA);
346 if (rc < sizeof(*leh) + sizeof(*lee))
347 GOTO(out, rc = -EINVAL);
350 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
351 leh->leh_magic = LINK_EA_MAGIC;
352 leh->leh_reccount = __swab32(leh->leh_reccount);
353 leh->leh_len = __swab64(leh->leh_len);
355 if (leh->leh_magic != LINK_EA_MAGIC)
356 GOTO(out, rc = -EINVAL);
357 if (leh->leh_reccount == 0)
358 GOTO(out, rc = -ENODATA);
360 lee = (struct link_ea_entry *)(leh + 1);
361 fid_be_to_cpu(fid, (const struct lu_fid *)&lee->lee_parent_fid);
365 if (buf.lb_buf != osd_oti_get(env)->oti_buf)
366 OBD_FREE(buf.lb_buf, buf.lb_len);
369 /* this block can be enabled for additional verification
370 * it's trying to match FID from LinkEA vs. FID from LMA */
374 rc2 = osd_find_parent_by_dnode(env, o, &fid2);
376 if (lu_fid_eq(fid, &fid2) == 0)
377 CERROR("wrong parent: "DFID" != "DFID"\n",
378 PFID(fid), PFID(&fid2));
382 /* no LinkEA is found, let's try to find the fid in parent's LMA */
383 if (unlikely(rc != 0))
384 rc = osd_find_parent_by_dnode(env, o, fid);
389 static int osd_dir_lookup(const struct lu_env *env, struct dt_object *dt,
390 struct dt_rec *rec, const struct dt_key *key,
391 struct lustre_capa *capa)
393 struct osd_thread_info *oti = osd_oti_get(env);
394 struct osd_object *obj = osd_dt_obj(dt);
395 struct osd_device *osd = osd_obj2dev(obj);
396 char *name = (char *)key;
400 LASSERT(osd_object_is_zap(obj->oo_db));
402 if (name[0] == '.') {
404 const struct lu_fid *f = lu_object_fid(&dt->do_lu);
405 memcpy(rec, f, sizeof(*f));
407 } else if (name[1] == '.' && name[2] == 0) {
408 rc = osd_find_parent_fid(env, dt, (struct lu_fid *)rec);
409 RETURN(rc == 0 ? 1 : rc);
413 rc = -zap_lookup(osd->od_os, obj->oo_db->db_object,
414 (char *)key, 8, sizeof(oti->oti_zde) / 8,
415 (void *)&oti->oti_zde);
416 memcpy(rec, &oti->oti_zde.lzd_fid, sizeof(struct lu_fid));
418 RETURN(rc == 0 ? 1 : rc);
421 static int osd_declare_dir_insert(const struct lu_env *env,
422 struct dt_object *dt,
423 const struct dt_rec *rec,
424 const struct dt_key *key,
427 struct osd_object *obj = osd_dt_obj(dt);
428 struct osd_thandle *oh;
433 oh = container_of0(th, struct osd_thandle, ot_super);
435 /* This is for inserting dot/dotdot for new created dir. */
436 if (obj->oo_db == NULL)
437 object = DMU_NEW_OBJECT;
439 object = obj->oo_db->db_object;
441 dmu_tx_hold_bonus(oh->ot_tx, object);
442 dmu_tx_hold_zap(oh->ot_tx, object, TRUE, (char *)key);
448 * Find the osd object for given fid.
450 * \param fid need to find the osd object having this fid
452 * \retval osd_object on success
453 * \retval -ve on error
455 struct osd_object *osd_object_find(const struct lu_env *env,
456 struct dt_object *dt,
457 const struct lu_fid *fid)
459 struct lu_device *ludev = dt->do_lu.lo_dev;
460 struct osd_object *child = NULL;
461 struct lu_object *luch;
462 struct lu_object *lo;
465 * at this point topdev might not exist yet
466 * (i.e. MGS is preparing profiles). so we can
467 * not rely on topdev and instead lookup with
468 * our device passed as topdev. this can't work
469 * if the object isn't cached yet (as osd doesn't
470 * allocate lu_header). IOW, the object must be
471 * in the cache, otherwise lu_object_alloc() crashes
474 luch = lu_object_find_at(env, ludev, fid, NULL);
478 if (lu_object_exists(luch)) {
479 lo = lu_object_locate(luch->lo_header, ludev->ld_type);
483 LU_OBJECT_DEBUG(D_ERROR, env, luch,
484 "%s: object can't be located "DFID"\n",
485 osd_dev(ludev)->od_svname, PFID(fid));
488 lu_object_put(env, luch);
489 CERROR("%s: Unable to get osd_object "DFID"\n",
490 osd_dev(ludev)->od_svname, PFID(fid));
491 child = ERR_PTR(-ENOENT);
494 LU_OBJECT_DEBUG(D_ERROR, env, luch,
495 "%s: lu_object does not exists "DFID"\n",
496 osd_dev(ludev)->od_svname, PFID(fid));
497 lu_object_put(env, luch);
498 child = ERR_PTR(-ENOENT);
505 * Put the osd object once done with it.
507 * \param obj osd object that needs to be put
509 static inline void osd_object_put(const struct lu_env *env,
510 struct osd_object *obj)
512 lu_object_put(env, &obj->oo_dt.do_lu);
515 static int osd_seq_exists(const struct lu_env *env, struct osd_device *osd,
518 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
519 struct seq_server_site *ss = osd_seq_site(osd);
524 LASSERT(ss->ss_server_fld != NULL);
526 rc = osd_fld_lookup(env, osd, seq, range);
529 CERROR("%s: Can not lookup fld for "LPX64"\n",
534 RETURN(ss->ss_node_id == range->lsr_index);
537 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
538 const struct lu_fid *fid)
540 struct seq_server_site *ss = osd_seq_site(osd);
543 /* FID seqs not in FLDB, must be local seq */
544 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
547 /* If FLD is not being initialized yet, it only happens during the
548 * initialization, likely during mgs initialization, and we assume
549 * this is local FID. */
550 if (ss == NULL || ss->ss_server_fld == NULL)
553 /* Only check the local FLDB here */
554 if (osd_seq_exists(env, osd, fid_seq(fid)))
561 * Inserts (key, value) pair in \a directory object.
563 * \param dt osd index object
564 * \param key key for index
565 * \param rec record reference
566 * \param th transaction handler
567 * \param capa capability descriptor
568 * \param ignore_quota update should not affect quota
571 * \retval -ve failure
573 static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
574 const struct dt_rec *rec, const struct dt_key *key,
575 struct thandle *th, struct lustre_capa *capa,
578 struct osd_thread_info *oti = osd_oti_get(env);
579 struct osd_object *parent = osd_dt_obj(dt);
580 struct osd_device *osd = osd_obj2dev(parent);
581 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
582 const struct lu_fid *fid = rec1->rec_fid;
583 struct osd_thandle *oh;
584 struct osd_object *child = NULL;
586 char *name = (char *)key;
590 LASSERT(parent->oo_db);
591 LASSERT(osd_object_is_zap(parent->oo_db));
593 LASSERT(dt_object_exists(dt));
594 LASSERT(osd_invariant(parent));
597 oh = container_of0(th, struct osd_thandle, ot_super);
599 rc = osd_remote_fid(env, osd, fid);
601 CERROR("%s: Can not find object "DFID": rc = %d\n",
602 osd->od_svname, PFID(fid), rc);
606 if (unlikely(rc == 1)) {
607 /* Insert remote entry */
608 memset(&oti->oti_zde.lzd_reg, 0, sizeof(oti->oti_zde.lzd_reg));
609 oti->oti_zde.lzd_reg.zde_type = IFTODT(rec1->rec_type & S_IFMT);
612 * To simulate old Orion setups with ./.. stored in the
615 /* Insert local entry */
616 child = osd_object_find(env, dt, fid);
618 RETURN(PTR_ERR(child));
620 LASSERT(child->oo_db);
621 if (name[0] == '.') {
623 /* do not store ".", instead generate it
624 * during iteration */
626 } else if (name[1] == '.' && name[2] == 0) {
627 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
628 struct lu_fid tfid = *fid;
630 osd_object_put(env, child);
632 child = osd_object_find(env, dt, &tfid);
634 RETURN(PTR_ERR(child));
636 LASSERT(child->oo_db);
639 /* update parent dnode in the child.
640 * later it will be used to generate ".." */
641 rc = osd_object_sa_update(parent,
643 &child->oo_db->db_object,
649 CLASSERT(sizeof(oti->oti_zde.lzd_reg) == 8);
650 CLASSERT(sizeof(oti->oti_zde) % 8 == 0);
651 attr = child->oo_dt.do_lu.lo_header ->loh_attr;
652 oti->oti_zde.lzd_reg.zde_type = IFTODT(attr & S_IFMT);
653 oti->oti_zde.lzd_reg.zde_dnode = child->oo_db->db_object;
656 oti->oti_zde.lzd_fid = *fid;
657 /* Insert (key,oid) into ZAP */
658 rc = -zap_add(osd->od_os, parent->oo_db->db_object,
659 (char *)key, 8, sizeof(oti->oti_zde) / 8,
660 (void *)&oti->oti_zde, oh->ot_tx);
661 if (unlikely(rc == -EEXIST &&
662 name[0] == '.' && name[1] == '.' && name[2] == 0))
663 /* Update (key,oid) in ZAP */
664 rc = -zap_update(osd->od_os, parent->oo_db->db_object,
665 (char *)key, 8, sizeof(oti->oti_zde) / 8,
666 (void *)&oti->oti_zde, oh->ot_tx);
670 osd_object_put(env, child);
675 static int osd_declare_dir_delete(const struct lu_env *env,
676 struct dt_object *dt,
677 const struct dt_key *key,
680 struct osd_object *obj = osd_dt_obj(dt);
681 struct osd_thandle *oh;
684 LASSERT(dt_object_exists(dt));
685 LASSERT(osd_invariant(obj));
688 oh = container_of0(th, struct osd_thandle, ot_super);
691 LASSERT(osd_object_is_zap(obj->oo_db));
693 dmu_tx_hold_zap(oh->ot_tx, obj->oo_db->db_object, TRUE, (char *)key);
698 static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
699 const struct dt_key *key, struct thandle *th,
700 struct lustre_capa *capa)
702 struct osd_object *obj = osd_dt_obj(dt);
703 struct osd_device *osd = osd_obj2dev(obj);
704 struct osd_thandle *oh;
705 dmu_buf_t *zap_db = obj->oo_db;
706 char *name = (char *)key;
711 LASSERT(osd_object_is_zap(obj->oo_db));
714 oh = container_of0(th, struct osd_thandle, ot_super);
717 * In Orion . and .. were stored in the directory (not generated upon
718 * request as now). we preserve them for backward compatibility
720 if (name[0] == '.') {
723 } else if (name[1] == '.' && name[2] == 0) {
728 /* Remove key from the ZAP */
729 rc = -zap_remove(osd->od_os, zap_db->db_object,
730 (char *) key, oh->ot_tx);
732 if (unlikely(rc && rc != -ENOENT))
733 CERROR("%s: zap_remove failed: rc = %d\n", osd->od_svname, rc);
738 static struct dt_it *osd_dir_it_init(const struct lu_env *env,
739 struct dt_object *dt,
741 struct lustre_capa *capa)
743 struct osd_zap_it *it;
745 it = (struct osd_zap_it *)osd_index_it_init(env, dt, unused, capa);
749 RETURN((struct dt_it *)it);
753 * Move Iterator to record specified by \a key
755 * \param di osd iterator
756 * \param key key for index
758 * \retval +ve di points to record with least key not larger than key
759 * \retval 0 di points to exact matched key
760 * \retval -ve failure
762 static int osd_dir_it_get(const struct lu_env *env,
763 struct dt_it *di, const struct dt_key *key)
765 struct osd_zap_it *it = (struct osd_zap_it *)di;
766 struct osd_object *obj = it->ozi_obj;
767 char *name = (char *)key;
774 /* reset the cursor */
775 zap_cursor_fini(it->ozi_zc);
776 osd_obj_cursor_init_serialized(it->ozi_zc, obj, 0);
778 /* XXX: implementation of the API is broken at the moment */
779 LASSERT(((const char *)key)[0] == 0);
786 if (name[0] == '.') {
790 } else if (name[1] == '.' && name[2] == 0) {
796 /* neither . nor .. - some real record */
804 static void osd_dir_it_put(const struct lu_env *env, struct dt_it *di)
806 /* PBS: do nothing : ref are incremented at retrive and decreamented
811 * in Orion . and .. were stored in the directory, while ZPL
812 * and current osd-zfs generate them up on request. so, we
813 * need to ignore previously stored . and ..
815 static int osd_index_retrieve_skip_dots(struct osd_zap_it *it,
821 rc = -zap_cursor_retrieve(it->ozi_zc, za);
824 if (unlikely(rc == 0 && za->za_name[0] == '.')) {
825 if (za->za_name[1] == 0) {
827 } else if (za->za_name[1] == '.' &&
828 za->za_name[2] == 0) {
832 zap_cursor_advance(it->ozi_zc);
834 } while (unlikely(rc == 0 && isdot));
840 * to load a directory entry at a time and stored it in
841 * iterator's in-memory data structure.
843 * \param di, struct osd_it_ea, iterator's in memory structure
845 * \retval +ve, iterator reached to end
846 * \retval 0, iterator not reached to end
847 * \retval -ve, on error
849 static int osd_dir_it_next(const struct lu_env *env, struct dt_it *di)
851 struct osd_zap_it *it = (struct osd_zap_it *)di;
852 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
857 /* temp. storage should be enough for any key supported by ZFS */
858 CLASSERT(sizeof(za->za_name) <= sizeof(it->ozi_name));
861 * the first ->next() moves the cursor to .
862 * the second ->next() moves the cursor to ..
863 * then we get to the real records and have to verify any exist
865 if (it->ozi_pos <= 2) {
871 zap_cursor_advance(it->ozi_zc);
875 * According to current API we need to return error if its last entry.
876 * zap_cursor_advance() does not return any value. So we need to call
877 * retrieve to check if there is any record. We should make
878 * changes to Iterator API to not return status for this API
880 rc = osd_index_retrieve_skip_dots(it, za);
882 if (rc == -ENOENT) /* end of dir */
888 static struct dt_key *osd_dir_it_key(const struct lu_env *env,
889 const struct dt_it *di)
891 struct osd_zap_it *it = (struct osd_zap_it *)di;
892 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
896 if (it->ozi_pos <= 1) {
898 RETURN((struct dt_key *)".");
899 } else if (it->ozi_pos == 2) {
900 RETURN((struct dt_key *)"..");
903 if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)))
906 strcpy(it->ozi_name, za->za_name);
908 RETURN((struct dt_key *)it->ozi_name);
911 static int osd_dir_it_key_size(const struct lu_env *env, const struct dt_it *di)
913 struct osd_zap_it *it = (struct osd_zap_it *)di;
914 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
918 if (it->ozi_pos <= 1) {
921 } else if (it->ozi_pos == 2) {
925 if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)) == 0)
926 rc = strlen(za->za_name);
931 static int osd_dir_it_rec(const struct lu_env *env, const struct dt_it *di,
932 struct dt_rec *dtrec, __u32 attr)
934 struct osd_zap_it *it = (struct osd_zap_it *)di;
935 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
936 struct luz_direntry *zde = &osd_oti_get(env)->oti_zde;
937 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
941 if (it->ozi_pos <= 1) {
942 lde->lde_hash = cpu_to_le64(1);
943 strcpy(lde->lde_name, ".");
944 lde->lde_namelen = cpu_to_le16(1);
945 lde->lde_fid = *lu_object_fid(&it->ozi_obj->oo_dt.do_lu);
946 lde->lde_attrs = LUDA_FID;
947 /* append lustre attributes */
948 osd_it_append_attrs(lde, attr, 1, IFTODT(S_IFDIR));
949 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(1, attr));
953 } else if (it->ozi_pos == 2) {
954 lde->lde_hash = cpu_to_le64(2);
955 strcpy(lde->lde_name, "..");
956 lde->lde_namelen = cpu_to_le16(2);
957 lde->lde_attrs = LUDA_FID;
958 /* append lustre attributes */
959 osd_it_append_attrs(lde, attr, 2, IFTODT(S_IFDIR));
960 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(2, attr));
961 rc = osd_find_parent_fid(env, &it->ozi_obj->oo_dt, &lde->lde_fid);
963 /* ENOENT happens at the root of filesystem so ignore it */
971 rc = -zap_cursor_retrieve(it->ozi_zc, za);
972 if (unlikely(rc != 0))
975 lde->lde_hash = cpu_to_le64(osd_zap_cursor_serialize(it->ozi_zc));
976 namelen = strlen(za->za_name);
977 if (namelen > NAME_MAX)
978 GOTO(out, rc = -EOVERFLOW);
979 strcpy(lde->lde_name, za->za_name);
980 lde->lde_namelen = cpu_to_le16(namelen);
982 if (za->za_integer_length != 8 || za->za_num_integers < 3) {
983 CERROR("%s: unsupported direntry format: %d %d\n",
984 osd_obj2dev(it->ozi_obj)->od_svname,
985 za->za_integer_length, (int)za->za_num_integers);
987 GOTO(out, rc = -EIO);
990 rc = -zap_lookup(it->ozi_zc->zc_objset, it->ozi_zc->zc_zapobj,
991 za->za_name, za->za_integer_length, 3, zde);
995 lde->lde_fid = zde->lzd_fid;
996 lde->lde_attrs = LUDA_FID;
998 /* append lustre attributes */
999 osd_it_append_attrs(lde, attr, namelen, zde->lzd_reg.zde_type);
1001 lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
1007 static int osd_dir_it_rec_size(const struct lu_env *env, const struct dt_it *di,
1010 struct osd_zap_it *it = (struct osd_zap_it *)di;
1011 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1016 if (it->ozi_pos <= 1)
1018 else if (it->ozi_pos == 2)
1022 rc = lu_dirent_calc_size(namelen, attr);
1026 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1027 if (unlikely(rc != 0))
1030 if (za->za_integer_length != 8 || za->za_num_integers < 3) {
1031 CERROR("%s: unsupported direntry format: %d %d\n",
1032 osd_obj2dev(it->ozi_obj)->od_svname,
1033 za->za_integer_length, (int)za->za_num_integers);
1037 namelen = strlen(za->za_name);
1038 if (namelen > NAME_MAX)
1041 rc = lu_dirent_calc_size(namelen, attr);
1046 static __u64 osd_dir_it_store(const struct lu_env *env, const struct dt_it *di)
1048 struct osd_zap_it *it = (struct osd_zap_it *)di;
1052 if (it->ozi_pos <= 2)
1055 pos = osd_zap_cursor_serialize(it->ozi_zc);
1062 * rc == 0 -> end of directory.
1063 * rc > 0 -> ok, proceed.
1064 * rc < 0 -> error. ( EOVERFLOW can be masked.)
1066 static int osd_dir_it_load(const struct lu_env *env,
1067 const struct dt_it *di, __u64 hash)
1069 struct osd_zap_it *it = (struct osd_zap_it *)di;
1070 struct osd_object *obj = it->ozi_obj;
1071 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1075 /* reset the cursor */
1076 zap_cursor_fini(it->ozi_zc);
1077 osd_obj_cursor_init_serialized(it->ozi_zc, obj, hash);
1084 /* to return whether the end has been reached */
1085 rc = osd_index_retrieve_skip_dots(it, za);
1088 else if (rc == -ENOENT)
1095 struct dt_index_operations osd_dir_ops = {
1096 .dio_lookup = osd_dir_lookup,
1097 .dio_declare_insert = osd_declare_dir_insert,
1098 .dio_insert = osd_dir_insert,
1099 .dio_declare_delete = osd_declare_dir_delete,
1100 .dio_delete = osd_dir_delete,
1102 .init = osd_dir_it_init,
1103 .fini = osd_index_it_fini,
1104 .get = osd_dir_it_get,
1105 .put = osd_dir_it_put,
1106 .next = osd_dir_it_next,
1107 .key = osd_dir_it_key,
1108 .key_size = osd_dir_it_key_size,
1109 .rec = osd_dir_it_rec,
1110 .rec_size = osd_dir_it_rec_size,
1111 .store = osd_dir_it_store,
1112 .load = osd_dir_it_load
1117 * Primitives for index files using binary keys.
1120 /* key integer_size is 8 */
1121 static int osd_prepare_key_uint64(struct osd_object *o, __u64 *dst,
1122 const struct dt_key *src)
1129 /* align keysize to 64bit */
1130 size = (o->oo_keysize + sizeof(__u64) - 1) / sizeof(__u64);
1131 size *= sizeof(__u64);
1133 LASSERT(size <= MAXNAMELEN);
1135 if (unlikely(size > o->oo_keysize))
1136 memset(dst + o->oo_keysize, 0, size - o->oo_keysize);
1137 memcpy(dst, (const char *)src, o->oo_keysize);
1139 return (size/sizeof(__u64));
1142 static int osd_index_lookup(const struct lu_env *env, struct dt_object *dt,
1143 struct dt_rec *rec, const struct dt_key *key,
1144 struct lustre_capa *capa)
1146 struct osd_object *obj = osd_dt_obj(dt);
1147 struct osd_device *osd = osd_obj2dev(obj);
1148 __u64 *k = osd_oti_get(env)->oti_key64;
1152 rc = osd_prepare_key_uint64(obj, k, key);
1154 rc = -zap_lookup_uint64(osd->od_os, obj->oo_db->db_object,
1155 k, rc, obj->oo_recusize, obj->oo_recsize,
1157 RETURN(rc == 0 ? 1 : rc);
1160 static int osd_declare_index_insert(const struct lu_env *env,
1161 struct dt_object *dt,
1162 const struct dt_rec *rec,
1163 const struct dt_key *key,
1166 struct osd_object *obj = osd_dt_obj(dt);
1167 struct osd_thandle *oh;
1170 LASSERT(th != NULL);
1171 oh = container_of0(th, struct osd_thandle, ot_super);
1173 LASSERT(obj->oo_db);
1175 dmu_tx_hold_bonus(oh->ot_tx, obj->oo_db->db_object);
1177 /* It is not clear what API should be used for binary keys, so we pass
1178 * a null name which has the side effect of over-reserving space,
1179 * accounting for the worst case. See zap_count_write() */
1180 dmu_tx_hold_zap(oh->ot_tx, obj->oo_db->db_object, TRUE, NULL);
1185 static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
1186 const struct dt_rec *rec, const struct dt_key *key,
1187 struct thandle *th, struct lustre_capa *capa,
1190 struct osd_object *obj = osd_dt_obj(dt);
1191 struct osd_device *osd = osd_obj2dev(obj);
1192 struct osd_thandle *oh;
1193 __u64 *k = osd_oti_get(env)->oti_key64;
1197 LASSERT(obj->oo_db);
1198 LASSERT(dt_object_exists(dt));
1199 LASSERT(osd_invariant(obj));
1200 LASSERT(th != NULL);
1202 oh = container_of0(th, struct osd_thandle, ot_super);
1204 rc = osd_prepare_key_uint64(obj, k, key);
1206 /* Insert (key,oid) into ZAP */
1207 rc = -zap_add_uint64(osd->od_os, obj->oo_db->db_object,
1208 k, rc, obj->oo_recusize, obj->oo_recsize,
1209 (void *)rec, oh->ot_tx);
1213 static int osd_declare_index_delete(const struct lu_env *env,
1214 struct dt_object *dt,
1215 const struct dt_key *key,
1218 struct osd_object *obj = osd_dt_obj(dt);
1219 struct osd_thandle *oh;
1222 LASSERT(dt_object_exists(dt));
1223 LASSERT(osd_invariant(obj));
1224 LASSERT(th != NULL);
1225 LASSERT(obj->oo_db);
1227 oh = container_of0(th, struct osd_thandle, ot_super);
1228 dmu_tx_hold_zap(oh->ot_tx, obj->oo_db->db_object, TRUE, NULL);
1233 static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
1234 const struct dt_key *key, struct thandle *th,
1235 struct lustre_capa *capa)
1237 struct osd_object *obj = osd_dt_obj(dt);
1238 struct osd_device *osd = osd_obj2dev(obj);
1239 struct osd_thandle *oh;
1240 __u64 *k = osd_oti_get(env)->oti_key64;
1244 LASSERT(obj->oo_db);
1245 LASSERT(th != NULL);
1246 oh = container_of0(th, struct osd_thandle, ot_super);
1248 rc = osd_prepare_key_uint64(obj, k, key);
1250 /* Remove binary key from the ZAP */
1251 rc = -zap_remove_uint64(osd->od_os, obj->oo_db->db_object,
1256 static int osd_index_it_get(const struct lu_env *env, struct dt_it *di,
1257 const struct dt_key *key)
1259 struct osd_zap_it *it = (struct osd_zap_it *)di;
1260 struct osd_object *obj = it->ozi_obj;
1261 struct osd_device *osd = osd_obj2dev(obj);
1265 LASSERT(it->ozi_zc);
1268 * XXX: we need a binary version of zap_cursor_move_to_key()
1269 * to implement this API */
1270 if (*((const __u64 *)key) != 0)
1271 CERROR("NOT IMPLEMETED YET (move to "LPX64")\n",
1274 zap_cursor_fini(it->ozi_zc);
1275 zap_cursor_init(it->ozi_zc, osd->od_os, obj->oo_db->db_object);
1281 static int osd_index_it_next(const struct lu_env *env, struct dt_it *di)
1283 struct osd_zap_it *it = (struct osd_zap_it *)di;
1284 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1288 if (it->ozi_reset == 0)
1289 zap_cursor_advance(it->ozi_zc);
1293 * According to current API we need to return error if it's last entry.
1294 * zap_cursor_advance() does not return any value. So we need to call
1295 * retrieve to check if there is any record. We should make
1296 * changes to Iterator API to not return status for this API
1298 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1305 static struct dt_key *osd_index_it_key(const struct lu_env *env,
1306 const struct dt_it *di)
1308 struct osd_zap_it *it = (struct osd_zap_it *)di;
1309 struct osd_object *obj = it->ozi_obj;
1310 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1315 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1317 RETURN(ERR_PTR(rc));
1319 /* the binary key is stored in the name */
1320 memcpy(&it->ozi_key, za->za_name, obj->oo_keysize);
1322 RETURN((struct dt_key *)&it->ozi_key);
1325 static int osd_index_it_key_size(const struct lu_env *env,
1326 const struct dt_it *di)
1328 struct osd_zap_it *it = (struct osd_zap_it *)di;
1329 struct osd_object *obj = it->ozi_obj;
1330 RETURN(obj->oo_keysize);
1333 static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
1334 struct dt_rec *rec, __u32 attr)
1336 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1337 struct osd_zap_it *it = (struct osd_zap_it *)di;
1338 struct osd_object *obj = it->ozi_obj;
1339 struct osd_device *osd = osd_obj2dev(obj);
1340 __u64 *k = osd_oti_get(env)->oti_key64;
1345 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1349 rc = osd_prepare_key_uint64(obj, k, (const struct dt_key *)za->za_name);
1351 rc = -zap_lookup_uint64(osd->od_os, obj->oo_db->db_object,
1352 k, rc, obj->oo_recusize, obj->oo_recsize,
1357 static __u64 osd_index_it_store(const struct lu_env *env,
1358 const struct dt_it *di)
1360 struct osd_zap_it *it = (struct osd_zap_it *)di;
1363 RETURN((__u64)zap_cursor_serialize(it->ozi_zc));
1366 static int osd_index_it_load(const struct lu_env *env, const struct dt_it *di,
1369 struct osd_zap_it *it = (struct osd_zap_it *)di;
1370 struct osd_object *obj = it->ozi_obj;
1371 struct osd_device *osd = osd_obj2dev(obj);
1372 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1376 /* reset the cursor */
1377 zap_cursor_fini(it->ozi_zc);
1378 zap_cursor_init_serialized(it->ozi_zc, osd->od_os,
1379 obj->oo_db->db_object, hash);
1382 rc = -zap_cursor_retrieve(it->ozi_zc, za);
1385 else if (rc == -ENOENT)
1391 static struct dt_index_operations osd_index_ops = {
1392 .dio_lookup = osd_index_lookup,
1393 .dio_declare_insert = osd_declare_index_insert,
1394 .dio_insert = osd_index_insert,
1395 .dio_declare_delete = osd_declare_index_delete,
1396 .dio_delete = osd_index_delete,
1398 .init = osd_index_it_init,
1399 .fini = osd_index_it_fini,
1400 .get = osd_index_it_get,
1401 .put = osd_index_it_put,
1402 .next = osd_index_it_next,
1403 .key = osd_index_it_key,
1404 .key_size = osd_index_it_key_size,
1405 .rec = osd_index_it_rec,
1406 .store = osd_index_it_store,
1407 .load = osd_index_it_load
1411 struct osd_metadnode_it {
1412 struct osd_device *mit_dev;
1414 struct lu_fid mit_fid;
1416 __u64 mit_prefetched_dnode;
1419 static struct dt_it *osd_zfs_otable_it_init(const struct lu_env *env,
1420 struct dt_object *dt, __u32 attr,
1421 struct lustre_capa *capa)
1423 struct osd_device *dev = osd_dev(dt->do_lu.lo_dev);
1424 struct osd_metadnode_it *it;
1428 if (unlikely(it == NULL))
1429 RETURN(ERR_PTR(-ENOMEM));
1433 /* XXX: dmu_object_next() does NOT find dnodes allocated
1434 * in the current non-committed txg, so we force txg
1435 * commit to find all existing dnodes ... */
1436 txg_wait_synced(dmu_objset_pool(dev->od_os), 0ULL);
1438 RETURN((struct dt_it *)it);
1441 static void osd_zfs_otable_it_fini(const struct lu_env *env, struct dt_it *di)
1443 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1448 static int osd_zfs_otable_it_get(const struct lu_env *env,
1449 struct dt_it *di, const struct dt_key *key)
1454 static void osd_zfs_otable_it_put(const struct lu_env *env, struct dt_it *di)
1458 #define OTABLE_PREFETCH 256
1460 static void osd_zfs_otable_prefetch(const struct lu_env *env,
1461 struct osd_metadnode_it *it)
1463 struct osd_device *dev = it->mit_dev;
1466 /* can go negative on the very first access to the iterator
1467 * or if some non-Lustre objects were found */
1468 if (unlikely(it->mit_prefetched < 0))
1469 it->mit_prefetched = 0;
1471 if (it->mit_prefetched >= (OTABLE_PREFETCH >> 1))
1474 if (it->mit_prefetched_dnode == 0)
1475 it->mit_prefetched_dnode = it->mit_pos;
1477 while (it->mit_prefetched < OTABLE_PREFETCH) {
1478 rc = -dmu_object_next(dev->od_os, &it->mit_prefetched_dnode,
1480 if (unlikely(rc != 0))
1483 /* dmu_prefetch() was exported in 0.6.2, if you use with
1484 * an older release, just comment it out - this is an
1486 dmu_prefetch(dev->od_os, it->mit_prefetched_dnode, 0, 0);
1488 it->mit_prefetched++;
1492 static int osd_zfs_otable_it_next(const struct lu_env *env, struct dt_it *di)
1494 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1495 struct lustre_mdt_attrs *lma;
1496 struct osd_device *dev = it->mit_dev;
1497 nvlist_t *nvbuf = NULL;
1502 memset(&it->mit_fid, 0, sizeof(it->mit_fid));
1504 dnode = it->mit_pos;
1506 rc = -dmu_object_next(dev->od_os, &it->mit_pos, B_FALSE, 0);
1507 if (unlikely(rc != 0))
1509 it->mit_prefetched--;
1511 /* LMA is required for this to be a Lustre object.
1512 * If there is no xattr skip it. */
1513 rc = __osd_xattr_load(dev, it->mit_pos, &nvbuf);
1514 if (unlikely(rc != 0))
1517 LASSERT(nvbuf != NULL);
1518 rc = -nvlist_lookup_byte_array(nvbuf, XATTR_NAME_LMA, &v, &s);
1519 if (likely(rc == 0)) {
1521 lma = (struct lustre_mdt_attrs *)v;
1522 lustre_lma_swab(lma);
1523 it->mit_fid = lma->lma_self_fid;
1527 /* not a Lustre object, try next one */
1534 /* we aren't prefetching in the above loop because the number of
1535 * non-Lustre objects is very small and we will be repeating very
1536 * rare. in case we want to use this to iterate over non-Lustre
1537 * objects (i.e. when we convert regular ZFS in Lustre) it makes
1538 * sense to initiate prefetching in the loop */
1540 /* 0 - there are more items, +1 - the end */
1541 if (likely(rc == 0))
1542 osd_zfs_otable_prefetch(env, it);
1544 CDEBUG(D_OTHER, "advance: %llu -> %llu "DFID": %d\n", dnode,
1545 it->mit_pos, PFID(&it->mit_fid), rc);
1551 static struct dt_key *osd_zfs_otable_it_key(const struct lu_env *env,
1552 const struct dt_it *di)
1557 static int osd_zfs_otable_it_key_size(const struct lu_env *env,
1558 const struct dt_it *di)
1560 return sizeof(__u64);
1563 static int osd_zfs_otable_it_rec(const struct lu_env *env,
1564 const struct dt_it *di,
1565 struct dt_rec *rec, __u32 attr)
1567 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1568 struct lu_fid *fid = (struct lu_fid *)rec;
1577 static __u64 osd_zfs_otable_it_store(const struct lu_env *env,
1578 const struct dt_it *di)
1580 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1585 static int osd_zfs_otable_it_load(const struct lu_env *env,
1586 const struct dt_it *di, __u64 hash)
1588 struct osd_metadnode_it *it = (struct osd_metadnode_it *)di;
1591 it->mit_prefetched = 0;
1592 it->mit_prefetched_dnode = 0;
1594 return osd_zfs_otable_it_next(env, (struct dt_it *)di);
1597 static int osd_zfs_otable_it_key_rec(const struct lu_env *env,
1598 const struct dt_it *di, void *key_rec)
1603 const struct dt_index_operations osd_zfs_otable_ops = {
1605 .init = osd_zfs_otable_it_init,
1606 .fini = osd_zfs_otable_it_fini,
1607 .get = osd_zfs_otable_it_get,
1608 .put = osd_zfs_otable_it_put,
1609 .next = osd_zfs_otable_it_next,
1610 .key = osd_zfs_otable_it_key,
1611 .key_size = osd_zfs_otable_it_key_size,
1612 .rec = osd_zfs_otable_it_rec,
1613 .store = osd_zfs_otable_it_store,
1614 .load = osd_zfs_otable_it_load,
1615 .key_rec = osd_zfs_otable_it_key_rec,
1619 int osd_index_try(const struct lu_env *env, struct dt_object *dt,
1620 const struct dt_index_features *feat)
1622 struct osd_object *obj = osd_dt_obj(dt);
1625 LASSERT(dt_object_exists(dt));
1628 * XXX: implement support for fixed-size keys sorted with natural
1629 * numerical way (not using internal hash value)
1631 if (feat->dif_flags & DT_IND_RANGE)
1634 if (unlikely(feat == &dt_otable_features)) {
1635 dt->do_index_ops = &osd_zfs_otable_ops;
1639 LASSERT(obj->oo_db != NULL);
1640 if (likely(feat == &dt_directory_features)) {
1641 if (osd_object_is_zap(obj->oo_db))
1642 dt->do_index_ops = &osd_dir_ops;
1645 } else if (unlikely(feat == &dt_acct_features)) {
1646 LASSERT(fid_is_acct(lu_object_fid(&dt->do_lu)));
1647 dt->do_index_ops = &osd_acct_index_ops;
1648 } else if (osd_object_is_zap(obj->oo_db) &&
1649 dt->do_index_ops == NULL) {
1650 /* For index file, we don't support variable key & record sizes
1651 * and the key has to be unique */
1652 if ((feat->dif_flags & ~DT_IND_UPDATE) != 0)
1655 if (feat->dif_keysize_max > ZAP_MAXNAMELEN)
1657 if (feat->dif_keysize_max != feat->dif_keysize_min)
1660 /* As for the record size, it should be a multiple of 8 bytes
1661 * and smaller than the maximum value length supported by ZAP.
1663 if (feat->dif_recsize_max > ZAP_MAXVALUELEN)
1665 if (feat->dif_recsize_max != feat->dif_recsize_min)
1668 obj->oo_keysize = feat->dif_keysize_max;
1669 obj->oo_recsize = feat->dif_recsize_max;
1670 obj->oo_recusize = 1;
1672 /* ZFS prefers to work with array of 64bits */
1673 if ((obj->oo_recsize & 7) == 0) {
1674 obj->oo_recsize >>= 3;
1675 obj->oo_recusize = 8;
1677 dt->do_index_ops = &osd_index_ops;