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_oi.c
33 * OI functions to map fid to dnode
35 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
36 * Author: Mike Pershin <tappro@whamcloud.com>
37 * Author: Di Wang <di.wang@intel.com>
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <lustre_ver.h>
43 #include <libcfs/libcfs.h>
44 #include <obd_support.h>
45 #include <lustre_net.h>
47 #include <obd_class.h>
48 #include <lustre_disk.h>
49 #include <lustre_fid.h>
51 #include "osd_internal.h"
53 #include <sys/dnode.h>
58 #include <sys/spa_impl.h>
59 #include <sys/zfs_znode.h>
60 #include <sys/dmu_tx.h>
61 #include <sys/dmu_objset.h>
62 #include <sys/dsl_prop.h>
63 #include <sys/sa_impl.h>
66 #define OSD_OI_FID_NR (1UL << 7)
67 #define OSD_OI_FID_NR_MAX (1UL << OSD_OI_FID_OID_BITS_MAX)
68 unsigned int osd_oi_count = OSD_OI_FID_NR;
72 * zfs osd maintains names for known fids in the name hierarchy
73 * so that one can mount filesystem with regular ZFS stack and
81 static const struct named_oid oids[] = {
82 { .oid = LAST_RECV_OID, .name = LAST_RCVD },
83 { .oid = OFD_LAST_GROUP_OID, .name = "LAST_GROUP" },
84 { .oid = LLOG_CATALOGS_OID, .name = "CATALOGS" },
85 { .oid = MGS_CONFIGS_OID, /*MOUNT_CONFIGS_DIR*/ },
86 { .oid = FID_SEQ_SRV_OID, .name = "seq_srv" },
87 { .oid = FID_SEQ_CTL_OID, .name = "seq_ctl" },
88 { .oid = FLD_INDEX_OID, .name = "fld" },
89 { .oid = MDD_LOV_OBJ_OID, .name = LOV_OBJID },
90 { .oid = OFD_HEALTH_CHECK_OID, .name = HEALTH_CHECK },
91 { .oid = ACCT_USER_OID, .name = "acct_usr_inode" },
92 { .oid = ACCT_GROUP_OID, .name = "acct_grp_inode" },
93 { .oid = REPLY_DATA_OID, .name = REPLY_DATA },
97 static char *oid2name(const unsigned long oid)
101 while (oids[i].oid) {
102 if (oids[i].oid == oid)
110 * Lookup an existing OI by the given name.
113 osd_oi_lookup(const struct lu_env *env, struct osd_device *o,
114 uint64_t parent, const char *name, struct osd_oi *oi)
116 struct zpl_direntry *zde = &osd_oti_get(env)->oti_zde.lzd_reg;
119 rc = -zap_lookup(o->od_os, parent, name, 8, 1, (void *)zde);
123 rc = strlcpy(oi->oi_name, name, sizeof(oi->oi_name));
124 if (rc >= sizeof(oi->oi_name))
127 oi->oi_zapid = zde->zde_dnode;
133 * Create a new OI with the given name.
136 osd_oi_create(const struct lu_env *env, struct osd_device *o,
137 uint64_t parent, const char *name, uint64_t *child)
139 struct zpl_direntry *zde = &osd_oti_get(env)->oti_zde.lzd_reg;
140 struct lu_attr *la = &osd_oti_get(env)->oti_la;
141 sa_handle_t *sa_hdl = NULL;
146 /* verify it doesn't already exist */
147 rc = -zap_lookup(o->od_os, parent, name, 8, 1, (void *)zde);
151 /* create fid-to-dnode index */
152 tx = dmu_tx_create(o->od_os);
156 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, 1, NULL);
157 dmu_tx_hold_bonus(tx, parent);
158 dmu_tx_hold_zap(tx, parent, TRUE, name);
159 dmu_tx_hold_sa_create(tx, ZFS_SA_BASE_ATTR_SIZE);
161 rc = -dmu_tx_assign(tx, TXG_WAIT);
167 oid = zap_create_flags(o->od_os, 0, ZAP_FLAG_HASH64,
168 DMU_OT_DIRECTORY_CONTENTS,
169 14, /* == ZFS fzap_default_block_shift */
170 DN_MAX_INDBLKSHIFT, /* indirect block shift */
171 DMU_OT_SA, DN_MAX_BONUSLEN, tx);
173 rc = -sa_handle_get(o->od_os, oid, NULL, SA_HDL_PRIVATE, &sa_hdl);
176 la->la_valid = LA_MODE | LA_UID | LA_GID;
177 la->la_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
178 la->la_uid = la->la_gid = 0;
179 rc = __osd_attr_init(env, o, sa_hdl, tx, la, parent);
180 sa_handle_destroy(sa_hdl);
184 zde->zde_dnode = oid;
186 zde->zde_type = IFTODT(S_IFDIR);
188 rc = -zap_add(o->od_os, parent, name, 8, 1, (void *)zde, tx);
192 dmu_object_free(o->od_os, oid, tx);
202 osd_oi_find_or_create(const struct lu_env *env, struct osd_device *o,
203 uint64_t parent, const char *name, uint64_t *child)
208 rc = osd_oi_lookup(env, o, parent, name, &oi);
210 *child = oi.oi_zapid;
211 else if (rc == -ENOENT)
212 rc = osd_oi_create(env, o, parent, name, child);
218 * Lookup the target index/flags of the fid, so it will know where
219 * the object is located (tgt index) and it is MDT or OST object.
221 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
222 u64 seq, struct lu_seq_range *range)
224 struct seq_server_site *ss = osd_seq_site(osd);
226 if (fid_seq_is_idif(seq)) {
227 fld_range_set_ost(range);
228 range->lsr_index = idif_ost_idx(seq);
232 if (!fid_seq_in_fldb(seq)) {
233 fld_range_set_mdt(range);
235 /* FIXME: If ss is NULL, it suppose not get lsr_index
237 range->lsr_index = ss->ss_node_id;
242 fld_range_set_any(range);
243 /* OSD will only do local fld lookup */
244 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
247 int fid_is_on_ost(const struct lu_env *env, struct osd_device *osd,
248 const struct lu_fid *fid)
250 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
254 if (fid_is_idif(fid))
257 if (unlikely(fid_is_local_file(fid) || fid_is_llog(fid)) ||
258 fid_is_name_llog(fid) || fid_is_quota(fid))
261 rc = osd_fld_lookup(env, osd, fid_seq(fid), range);
264 CERROR("%s: "DFID" lookup failed: rc = %d\n",
265 osd_name(osd), PFID(fid), rc);
269 if (fld_range_is_ost(range))
275 static struct osd_seq *osd_seq_find_locked(struct osd_seq_list *seq_list,
278 struct osd_seq *osd_seq;
280 list_for_each_entry(osd_seq, &seq_list->osl_seq_list, os_seq_list) {
281 if (osd_seq->os_seq == seq)
287 static struct osd_seq *osd_seq_find(struct osd_seq_list *seq_list, u64 seq)
289 struct osd_seq *osd_seq;
291 read_lock(&seq_list->osl_seq_list_lock);
292 osd_seq = osd_seq_find_locked(seq_list, seq);
293 read_unlock(&seq_list->osl_seq_list_lock);
298 static struct osd_seq *osd_find_or_add_seq(const struct lu_env *env,
299 struct osd_device *osd, u64 seq)
301 struct osd_seq_list *seq_list = &osd->od_seq_list;
302 struct osd_seq *osd_seq;
303 char *key = osd_oti_get(env)->oti_buf;
304 char *seq_name = osd_oti_get(env)->oti_str;
311 osd_seq = osd_seq_find(seq_list, seq);
315 down(&seq_list->osl_seq_init_sem);
316 /* Check again, in case some one else already add it
318 osd_seq = osd_seq_find(seq_list, seq);
322 OBD_ALLOC_PTR(osd_seq);
324 GOTO(out, rc = -ENOMEM);
326 INIT_LIST_HEAD(&osd_seq->os_seq_list);
327 osd_seq->os_seq = seq;
329 /* Init subdir count to be 32, but each seq can have
330 * different subdir count */
331 osd_seq->os_subdir_count = OSD_OST_MAP_SIZE;
332 OBD_ALLOC(osd_seq->os_compat_dirs,
333 sizeof(uint64_t) * osd_seq->os_subdir_count);
334 if (osd_seq->os_compat_dirs == NULL)
335 GOTO(out, rc = -ENOMEM);
337 oi.oi_zapid = osd->od_O_id;
338 sprintf(seq_name, (fid_seq_is_rsvd(seq) ||
339 fid_seq_is_mdt0(seq)) ? "%llu" : "%llx",
340 fid_seq_is_idif(seq) ? 0 : seq);
342 rc = osd_oi_find_or_create(env, osd, oi.oi_zapid, seq_name, &odb);
344 CERROR("%s: Can not create %s : rc = %d\n",
345 osd_name(osd), seq_name, rc);
349 for (i = 0; i < OSD_OST_MAP_SIZE; i++) {
350 sprintf(key, "d%d", i);
351 rc = osd_oi_find_or_create(env, osd, odb, key, &sdb);
354 osd_seq->os_compat_dirs[i] = sdb;
357 write_lock(&seq_list->osl_seq_list_lock);
358 list_add(&osd_seq->os_seq_list, &seq_list->osl_seq_list);
359 write_unlock(&seq_list->osl_seq_list_lock);
361 up(&seq_list->osl_seq_init_sem);
363 if (osd_seq != NULL && osd_seq->os_compat_dirs != NULL)
364 OBD_FREE(osd_seq->os_compat_dirs,
365 sizeof(uint64_t) * osd_seq->os_subdir_count);
367 OBD_FREE_PTR(osd_seq);
368 osd_seq = ERR_PTR(rc);
374 * objects w/o a natural reference (unlike a file on a MDS)
375 * are put under a special hierarchy /O/<seq>/d0..dXX
376 * this function returns a directory specific fid belongs to
379 osd_get_idx_for_ost_obj(const struct lu_env *env, struct osd_device *osd,
380 const struct lu_fid *fid, char *buf, int bufsize)
382 struct osd_seq *osd_seq;
387 osd_seq = osd_find_or_add_seq(env, osd, fid_seq(fid));
388 if (IS_ERR(osd_seq)) {
389 CERROR("%s: Can not find seq group "DFID"\n", osd_name(osd),
391 return PTR_ERR(osd_seq);
394 if (fid_is_last_id(fid)) {
397 rc = fid_to_ostid(fid, &osd_oti_get(env)->oti_ostid);
398 LASSERT(rc == 0); /* we should not get here with IGIF */
399 id = ostid_id(&osd_oti_get(env)->oti_ostid);
402 b = id % OSD_OST_MAP_SIZE;
403 LASSERT(osd_seq->os_compat_dirs[b]);
406 snprintf(buf, bufsize, "%llu", id);
408 return osd_seq->os_compat_dirs[b];
411 /* XXX: f_ver is not counted, but may differ too */
412 static void osd_fid2str(char *buf, const struct lu_fid *fid)
414 sprintf(buf, DFID_NOBRACE, PFID(fid));
418 * Determine the zap object id which is being used as the OI for the
419 * given fid. The lowest N bits in the sequence ID are used as the
420 * index key. On failure 0 is returned which zfs treats internally
421 * as an invalid object id.
424 osd_get_idx_for_fid(struct osd_device *osd, const struct lu_fid *fid,
429 LASSERT(osd->od_oi_table != NULL);
430 oi = osd->od_oi_table[fid_seq(fid) & (osd->od_oi_count - 1)];
432 osd_fid2str(buf, fid);
437 uint64_t osd_get_name_n_idx(const struct lu_env *env, struct osd_device *osd,
438 const struct lu_fid *fid, char *buf, int bufsize)
444 if (fid_is_on_ost(env, osd, fid) == 1 || fid_seq(fid) == FID_SEQ_ECHO) {
445 zapid = osd_get_idx_for_ost_obj(env, osd, fid, buf, bufsize);
446 } else if (unlikely(fid_seq(fid) == FID_SEQ_LOCAL_FILE)) {
447 /* special objects with fixed known fids get their name */
448 char *name = oid2name(fid_oid(fid));
451 zapid = osd->od_root;
453 strncpy(buf, name, bufsize);
454 if (fid_is_acct(fid))
455 zapid = MASTER_NODE_OBJ;
457 zapid = osd_get_idx_for_fid(osd, fid, buf);
460 zapid = osd_get_idx_for_fid(osd, fid, buf);
466 static inline int fid_is_fs_root(const struct lu_fid *fid)
468 /* Map root inode to special local object FID */
469 return fid_seq(fid) == FID_SEQ_LOCAL_FILE &&
470 fid_oid(fid) == OSD_FS_ROOT_OID;
473 int osd_fid_lookup(const struct lu_env *env, struct osd_device *dev,
474 const struct lu_fid *fid, uint64_t *oid)
476 struct osd_thread_info *info = osd_oti_get(env);
477 char *buf = info->oti_buf;
482 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
485 if (unlikely(fid_is_acct(fid))) {
486 if (fid_oid(fid) == ACCT_USER_OID)
487 *oid = dev->od_iusr_oid;
489 *oid = dev->od_igrp_oid;
490 } else if (unlikely(fid_is_fs_root(fid))) {
493 zapid = osd_get_name_n_idx(env, dev, fid, buf,
494 sizeof(info->oti_buf));
495 rc = -zap_lookup(dev->od_os, zapid, buf,
496 8, 1, &info->oti_zde);
499 *oid = info->oti_zde.lzd_reg.zde_dnode;
503 osd_dmu_prefetch(dev->od_os, *oid, 0, 0, 0,
504 ZIO_PRIORITY_ASYNC_READ);
510 * Close an entry in a specific slot.
513 osd_oi_remove_table(const struct lu_env *env, struct osd_device *o, int key)
517 LASSERT(key < o->od_oi_count);
519 oi = o->od_oi_table[key];
522 osd_dnode_rele(oi->oi_dn);
524 o->od_oi_table[key] = NULL;
529 * Allocate and open a new entry in the specified unused slot.
532 osd_oi_add_table(const struct lu_env *env, struct osd_device *o,
538 LASSERT(key < o->od_oi_count);
539 LASSERT(o->od_oi_table[key] == NULL);
545 rc = osd_oi_lookup(env, o, o->od_root, name, oi);
551 o->od_oi_table[key] = oi;
552 __osd_obj2dnode(env, o->od_os, oi->oi_zapid, &oi->oi_dn);
558 * Depopulate the OI table.
561 osd_oi_close_table(const struct lu_env *env, struct osd_device *o)
565 for (i = 0; i < o->od_oi_count; i++)
566 osd_oi_remove_table(env, o, i);
570 * Populate the OI table based.
573 osd_oi_open_table(const struct lu_env *env, struct osd_device *o, int count)
579 for (i = 0; i < count; i++) {
580 sprintf(name, "%s.%d", DMU_OSD_OI_NAME_BASE, i);
581 rc = osd_oi_add_table(env, o, name, i);
583 osd_oi_close_table(env, o);
592 * Determine if the type and number of OIs used by this file system.
595 osd_oi_probe(const struct lu_env *env, struct osd_device *o, int *count)
597 uint64_t root_oid = o->od_root;
604 * Check for multiple OIs and determine the count. There is no
605 * gap handling, if an OI is missing the wrong size can be returned.
606 * The only safeguard is that we know the number of OIs must be a
607 * power of two and this is checked for basic sanity.
609 for (*count = 0; *count < OSD_OI_FID_NR_MAX; (*count)++) {
610 sprintf(name, "%s.%d", DMU_OSD_OI_NAME_BASE, *count);
611 rc = osd_oi_lookup(env, o, root_oid, name, &oi);
619 if ((*count & (*count - 1)) != 0)
629 * No OIs exist, this must be a new filesystem.
636 static void osd_ost_seq_fini(const struct lu_env *env, struct osd_device *osd)
638 struct osd_seq_list *osl = &osd->od_seq_list;
639 struct osd_seq *osd_seq, *tmp;
641 write_lock(&osl->osl_seq_list_lock);
642 list_for_each_entry_safe(osd_seq, tmp, &osl->osl_seq_list,
644 list_del(&osd_seq->os_seq_list);
645 OBD_FREE(osd_seq->os_compat_dirs,
646 sizeof(uint64_t) * osd_seq->os_subdir_count);
647 OBD_FREE(osd_seq, sizeof(*osd_seq));
649 write_unlock(&osl->osl_seq_list_lock);
655 * Create /O subdirectory to map legacy OST objects for compatibility.
658 osd_oi_init_compat(const struct lu_env *env, struct osd_device *o)
664 rc = osd_oi_find_or_create(env, o, o->od_root, "O", &sdb);
670 /* Create on-disk indexes to maintain per-UID/GID inode usage.
671 * Those new indexes are created in the top-level ZAP outside the
672 * namespace in order not to confuse ZPL which might interpret those
673 * indexes as directories and assume the values are object IDs */
674 rc = osd_oi_find_or_create(env, o, MASTER_NODE_OBJ,
675 oid2name(ACCT_USER_OID), &odb);
678 o->od_iusr_oid = odb;
680 rc = osd_oi_find_or_create(env, o, MASTER_NODE_OBJ,
681 oid2name(ACCT_GROUP_OID), &odb);
684 o->od_igrp_oid = odb;
690 * Initialize the OIs by either opening or creating them as needed.
692 int osd_oi_init(const struct lu_env *env, struct osd_device *o)
694 char *key = osd_oti_get(env)->oti_buf;
695 int i, rc, count = 0;
698 rc = osd_oi_probe(env, o, &count);
705 count = osd_oi_count;
708 for (i = 0; i < count; i++) {
709 sprintf(key, "%s.%d", DMU_OSD_OI_NAME_BASE, i);
710 rc = osd_oi_find_or_create(env, o, odb, key, &sdb);
716 rc = osd_oi_init_compat(env, o);
720 LASSERT((count & (count - 1)) == 0);
721 o->od_oi_count = count;
722 OBD_ALLOC(o->od_oi_table, sizeof(struct osd_oi *) * count);
723 if (o->od_oi_table == NULL)
726 rc = osd_oi_open_table(env, o, count);
728 OBD_FREE(o->od_oi_table, sizeof(struct osd_oi *) * count);
729 o->od_oi_table = NULL;
735 void osd_oi_fini(const struct lu_env *env, struct osd_device *o)
739 osd_ost_seq_fini(env, o);
741 if (o->od_oi_table != NULL) {
742 (void) osd_oi_close_table(env, o);
743 OBD_FREE(o->od_oi_table,
744 sizeof(struct osd_oi *) * o->od_oi_count);
745 o->od_oi_table = NULL;
752 int osd_options_init(void)
754 /* osd_oi_count - Default number of OIs, 128 works well for ZFS */
755 if (osd_oi_count == 0 || osd_oi_count > OSD_OI_FID_NR_MAX)
756 osd_oi_count = OSD_OI_FID_NR;
758 if ((osd_oi_count & (osd_oi_count - 1)) != 0) {
759 LCONSOLE_WARN("Round up osd_oi_count %d to power2 %d\n",
760 osd_oi_count, size_roundup_power2(osd_oi_count));
761 osd_oi_count = size_roundup_power2(osd_oi_count);
768 * the following set of functions are used to maintain per-thread
769 * cache of FID->ino mapping. this mechanism is used to avoid
770 * expensive LU/OI lookups.
772 struct osd_idmap_cache *osd_idc_find(const struct lu_env *env,
773 struct osd_device *osd,
774 const struct lu_fid *fid)
776 struct osd_thread_info *oti = osd_oti_get(env);
777 struct osd_idmap_cache *idc = oti->oti_ins_cache;
780 for (i = 0; i < oti->oti_ins_cache_used; i++) {
781 if (!lu_fid_eq(&idc[i].oic_fid, fid))
783 if (idc[i].oic_dev != osd)
792 struct osd_idmap_cache *osd_idc_add(const struct lu_env *env,
793 struct osd_device *osd,
794 const struct lu_fid *fid)
796 struct osd_thread_info *oti = osd_oti_get(env);
797 struct osd_idmap_cache *idc;
800 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
801 i = oti->oti_ins_cache_size * 2;
804 i = OSD_INS_CACHE_SIZE;
805 OBD_ALLOC(idc, sizeof(*idc) * i);
807 return ERR_PTR(-ENOMEM);
808 if (oti->oti_ins_cache != NULL) {
809 memcpy(idc, oti->oti_ins_cache,
810 oti->oti_ins_cache_used * sizeof(*idc));
811 OBD_FREE(oti->oti_ins_cache,
812 oti->oti_ins_cache_used * sizeof(*idc));
814 oti->oti_ins_cache = idc;
815 oti->oti_ins_cache_size = i;
818 idc = &oti->oti_ins_cache[oti->oti_ins_cache_used++];
828 * Lookup mapping for the given fid in the cache
830 * Initialize a new one if not found. the initialization checks whether
831 * the object is local or remote. for the local objects, OI is used to
832 * learn dnode#. the function is used when the caller has no information
833 * about the object, e.g. at dt_insert().
835 struct osd_idmap_cache *osd_idc_find_or_init(const struct lu_env *env,
836 struct osd_device *osd,
837 const struct lu_fid *fid)
839 struct osd_idmap_cache *idc;
842 idc = osd_idc_find(env, osd, fid);
846 /* new mapping is needed */
847 idc = osd_idc_add(env, osd, fid);
852 rc = osd_remote_fid(env, osd, fid);
853 if (unlikely(rc < 0))
857 /* the object is local, lookup in OI */
860 rc = osd_fid_lookup(env, osd, fid, &dnode);
861 if (unlikely(rc < 0)) {
862 CERROR("%s: can't lookup: rc = %d\n",
866 LASSERT(dnode < (1ULL << DN_MAX_OBJECT_SHIFT));
867 idc->oic_dnode = dnode;
869 /* the object is remote */
877 * lookup mapping for given FID and fill it from the given object.
878 * the object is local by definition.
880 int osd_idc_find_and_init(const struct lu_env *env, struct osd_device *osd,
881 struct osd_object *obj)
883 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
884 struct osd_idmap_cache *idc;
886 idc = osd_idc_find(env, osd, fid);
888 if (obj->oo_dn == NULL)
890 idc->oic_dnode = obj->oo_dn->dn_object;
894 /* new mapping is needed */
895 idc = osd_idc_add(env, osd, fid);
900 idc->oic_dnode = obj->oo_dn->dn_object;