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 = ACCT_PROJECT_OID, .name = "acct_prj_inode" },
94 { .oid = REPLY_DATA_OID, .name = REPLY_DATA },
98 static char *oid2name(const unsigned long oid)
102 while (oids[i].oid) {
103 if (oids[i].oid == oid)
111 * Lookup an existing OI by the given name.
114 osd_oi_lookup(const struct lu_env *env, struct osd_device *o,
115 uint64_t parent, const char *name, struct osd_oi *oi)
117 struct zpl_direntry *zde = &osd_oti_get(env)->oti_zde.lzd_reg;
120 rc = -zap_lookup(o->od_os, parent, name, 8, 1, (void *)zde);
124 rc = strlcpy(oi->oi_name, name, sizeof(oi->oi_name));
125 if (rc >= sizeof(oi->oi_name))
128 oi->oi_zapid = zde->zde_dnode;
134 * Create a new OI with the given name.
137 osd_oi_create(const struct lu_env *env, struct osd_device *o,
138 uint64_t parent, const char *name, uint64_t *child)
140 struct zpl_direntry *zde = &osd_oti_get(env)->oti_zde.lzd_reg;
141 struct lu_attr *la = &osd_oti_get(env)->oti_la;
142 sa_handle_t *sa_hdl = NULL;
147 /* verify it doesn't already exist */
148 rc = -zap_lookup(o->od_os, parent, name, 8, 1, (void *)zde);
152 if (o->od_dt_dev.dd_rdonly)
155 /* create fid-to-dnode index */
156 tx = dmu_tx_create(o->od_os);
160 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, 1, NULL);
161 dmu_tx_hold_bonus(tx, parent);
162 dmu_tx_hold_zap(tx, parent, TRUE, name);
163 dmu_tx_hold_sa_create(tx, ZFS_SA_BASE_ATTR_SIZE);
165 rc = -dmu_tx_assign(tx, TXG_WAIT);
171 oid = osd_zap_create_flags(o->od_os, 0, ZAP_FLAG_HASH64,
172 DMU_OT_DIRECTORY_CONTENTS,
173 14, /* == ZFS fzap_default_block_shift */
177 rc = -sa_handle_get(o->od_os, oid, NULL, SA_HDL_PRIVATE, &sa_hdl);
180 la->la_valid = LA_MODE | LA_UID | LA_GID;
181 la->la_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
182 la->la_uid = la->la_gid = 0;
183 rc = __osd_attr_init(env, o, sa_hdl, tx, la, parent);
184 sa_handle_destroy(sa_hdl);
188 zde->zde_dnode = oid;
190 zde->zde_type = IFTODT(S_IFDIR);
192 rc = -zap_add(o->od_os, parent, name, 8, 1, (void *)zde, tx);
196 dmu_object_free(o->od_os, oid, tx);
206 osd_oi_find_or_create(const struct lu_env *env, struct osd_device *o,
207 uint64_t parent, const char *name, uint64_t *child)
212 rc = osd_oi_lookup(env, o, parent, name, &oi);
214 *child = oi.oi_zapid;
215 else if (rc == -ENOENT)
216 rc = osd_oi_create(env, o, parent, name, child);
222 * Lookup the target index/flags of the fid, so it will know where
223 * the object is located (tgt index) and it is MDT or OST object.
225 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
226 u64 seq, struct lu_seq_range *range)
228 struct seq_server_site *ss = osd_seq_site(osd);
230 if (fid_seq_is_idif(seq)) {
231 fld_range_set_ost(range);
232 range->lsr_index = idif_ost_idx(seq);
236 if (!fid_seq_in_fldb(seq)) {
237 fld_range_set_mdt(range);
239 /* FIXME: If ss is NULL, it suppose not get lsr_index
241 range->lsr_index = ss->ss_node_id;
246 fld_range_set_any(range);
247 /* OSD will only do local fld lookup */
248 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
251 int fid_is_on_ost(const struct lu_env *env, struct osd_device *osd,
252 const struct lu_fid *fid)
254 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
258 if (fid_is_idif(fid))
261 if (unlikely(fid_is_local_file(fid) || fid_is_llog(fid)) ||
262 fid_is_name_llog(fid) || fid_is_quota(fid))
265 rc = osd_fld_lookup(env, osd, fid_seq(fid), range);
267 /* During upgrade, OST FLDB might not be loaded because
268 * OST FLDB is not created until 2.6, so if some DNE
269 * filesystem upgrade from 2.5 to 2.7/2.8, they will
270 * not be able to find the sequence from local FLDB
271 * cache see fld_index_init(). */
272 if (rc == -ENOENT && osd->od_is_ost)
276 CERROR("%s: "DFID" lookup failed: rc = %d\n",
277 osd_name(osd), PFID(fid), rc);
281 if (fld_range_is_ost(range))
287 static struct osd_seq *osd_seq_find_locked(struct osd_seq_list *seq_list,
290 struct osd_seq *osd_seq;
292 list_for_each_entry(osd_seq, &seq_list->osl_seq_list, os_seq_list) {
293 if (osd_seq->os_seq == seq)
299 static struct osd_seq *osd_seq_find(struct osd_seq_list *seq_list, u64 seq)
301 struct osd_seq *osd_seq;
303 read_lock(&seq_list->osl_seq_list_lock);
304 osd_seq = osd_seq_find_locked(seq_list, seq);
305 read_unlock(&seq_list->osl_seq_list_lock);
310 static struct osd_seq *osd_find_or_add_seq(const struct lu_env *env,
311 struct osd_device *osd, u64 seq)
313 struct osd_seq_list *seq_list = &osd->od_seq_list;
314 struct osd_seq *osd_seq;
315 char *key = osd_oti_get(env)->oti_buf;
316 char *seq_name = osd_oti_get(env)->oti_str;
323 osd_seq = osd_seq_find(seq_list, seq);
327 down(&seq_list->osl_seq_init_sem);
328 /* Check again, in case some one else already add it
330 osd_seq = osd_seq_find(seq_list, seq);
334 OBD_ALLOC_PTR(osd_seq);
336 GOTO(out, rc = -ENOMEM);
338 INIT_LIST_HEAD(&osd_seq->os_seq_list);
339 osd_seq->os_seq = seq;
341 /* Init subdir count to be 32, but each seq can have
342 * different subdir count */
343 osd_seq->os_subdir_count = OSD_OST_MAP_SIZE;
344 OBD_ALLOC(osd_seq->os_compat_dirs,
345 sizeof(uint64_t) * osd_seq->os_subdir_count);
346 if (osd_seq->os_compat_dirs == NULL)
347 GOTO(out, rc = -ENOMEM);
349 oi.oi_zapid = osd->od_O_id;
350 sprintf(seq_name, (fid_seq_is_rsvd(seq) ||
351 fid_seq_is_mdt0(seq)) ? "%llu" : "%llx",
352 fid_seq_is_idif(seq) ? 0 : seq);
354 rc = osd_oi_find_or_create(env, osd, oi.oi_zapid, seq_name, &odb);
356 CERROR("%s: Can not create %s : rc = %d\n",
357 osd_name(osd), seq_name, rc);
361 for (i = 0; i < OSD_OST_MAP_SIZE; i++) {
362 sprintf(key, "d%d", i);
363 rc = osd_oi_find_or_create(env, osd, odb, key, &sdb);
366 osd_seq->os_compat_dirs[i] = sdb;
369 write_lock(&seq_list->osl_seq_list_lock);
370 list_add(&osd_seq->os_seq_list, &seq_list->osl_seq_list);
371 write_unlock(&seq_list->osl_seq_list_lock);
373 up(&seq_list->osl_seq_init_sem);
375 if (osd_seq != NULL && osd_seq->os_compat_dirs != NULL)
376 OBD_FREE(osd_seq->os_compat_dirs,
377 sizeof(uint64_t) * osd_seq->os_subdir_count);
379 OBD_FREE_PTR(osd_seq);
380 osd_seq = ERR_PTR(rc);
386 * objects w/o a natural reference (unlike a file on a MDS)
387 * are put under a special hierarchy /O/<seq>/d0..dXX
388 * this function returns a directory specific fid belongs to
391 osd_get_idx_for_ost_obj(const struct lu_env *env, struct osd_device *osd,
392 const struct lu_fid *fid, char *buf, int bufsize)
394 struct osd_seq *osd_seq;
399 osd_seq = osd_find_or_add_seq(env, osd, fid_seq(fid));
400 if (IS_ERR(osd_seq)) {
401 CERROR("%s: Can not find seq group "DFID"\n", osd_name(osd),
403 return PTR_ERR(osd_seq);
406 if (fid_is_last_id(fid)) {
409 rc = fid_to_ostid(fid, &osd_oti_get(env)->oti_ostid);
410 LASSERT(rc == 0); /* we should not get here with IGIF */
411 id = ostid_id(&osd_oti_get(env)->oti_ostid);
414 b = id % OSD_OST_MAP_SIZE;
415 LASSERT(osd_seq->os_compat_dirs[b]);
418 snprintf(buf, bufsize, "%llu", id);
420 return osd_seq->os_compat_dirs[b];
423 /* XXX: f_ver is not counted, but may differ too */
424 static void osd_fid2str(char *buf, const struct lu_fid *fid)
426 sprintf(buf, DFID_NOBRACE, PFID(fid));
430 * Determine the zap object id which is being used as the OI for the
431 * given fid. The lowest N bits in the sequence ID are used as the
432 * index key. On failure 0 is returned which zfs treats internally
433 * as an invalid object id.
436 osd_get_idx_for_fid(struct osd_device *osd, const struct lu_fid *fid,
437 char *buf, dnode_t **zdn)
441 LASSERT(osd->od_oi_table != NULL);
442 oi = osd->od_oi_table[fid_seq(fid) & (osd->od_oi_count - 1)];
444 osd_fid2str(buf, fid);
451 uint64_t osd_get_name_n_idx(const struct lu_env *env, struct osd_device *osd,
452 const struct lu_fid *fid, char *buf, int bufsize,
461 if (fid_is_on_ost(env, osd, fid) == 1 || fid_seq(fid) == FID_SEQ_ECHO) {
462 zapid = osd_get_idx_for_ost_obj(env, osd, fid, buf, bufsize);
463 } else if (unlikely(fid_seq(fid) == FID_SEQ_LOCAL_FILE)) {
464 /* special objects with fixed known fids get their name */
465 char *name = oid2name(fid_oid(fid));
468 zapid = osd->od_root;
470 strncpy(buf, name, bufsize);
471 if (fid_is_acct(fid))
472 zapid = MASTER_NODE_OBJ;
474 zapid = osd_get_idx_for_fid(osd, fid, buf, NULL);
477 zapid = osd_get_idx_for_fid(osd, fid, buf, zdn);
483 static inline int fid_is_fs_root(const struct lu_fid *fid)
485 /* Map root inode to special local object FID */
486 return fid_seq(fid) == FID_SEQ_LOCAL_FILE &&
487 fid_oid(fid) == OSD_FS_ROOT_OID;
490 static inline int osd_oid(struct osd_device *dev, __u32 local_oid,
495 *oid = dev->od_iusr_oid;
498 *oid = dev->od_igrp_oid;
500 case ACCT_PROJECT_OID:
502 CERROR("%s: unsupported quota oid: %#x\n",
503 osd_name(dev), local_oid);
510 int osd_fid_lookup(const struct lu_env *env, struct osd_device *dev,
511 const struct lu_fid *fid, uint64_t *oid)
513 struct osd_thread_info *info = osd_oti_get(env);
514 char *buf = info->oti_buf;
520 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
523 if (unlikely(fid_is_acct(fid))) {
524 rc = osd_oid(dev, fid_oid(fid), oid);
527 } else if (unlikely(fid_is_fs_root(fid))) {
530 zapid = osd_get_name_n_idx(env, dev, fid, buf,
531 sizeof(info->oti_buf), &zdn);
532 rc = osd_zap_lookup(dev, zapid, zdn, buf,
533 8, 1, &info->oti_zde);
536 *oid = info->oti_zde.lzd_reg.zde_dnode;
540 osd_dmu_prefetch(dev->od_os, *oid, 0, 0, 0,
541 ZIO_PRIORITY_ASYNC_READ);
547 * Close an entry in a specific slot.
550 osd_oi_remove_table(const struct lu_env *env, struct osd_device *o, int key)
554 LASSERT(key < o->od_oi_count);
556 oi = o->od_oi_table[key];
559 osd_dnode_rele(oi->oi_dn);
561 o->od_oi_table[key] = NULL;
566 * Allocate and open a new entry in the specified unused slot.
569 osd_oi_add_table(const struct lu_env *env, struct osd_device *o,
575 LASSERT(key < o->od_oi_count);
576 LASSERT(o->od_oi_table[key] == NULL);
582 rc = osd_oi_lookup(env, o, o->od_root, name, oi);
588 o->od_oi_table[key] = oi;
589 __osd_obj2dnode(o->od_os, oi->oi_zapid, &oi->oi_dn);
595 * Depopulate the OI table.
598 osd_oi_close_table(const struct lu_env *env, struct osd_device *o)
602 for (i = 0; i < o->od_oi_count; i++)
603 osd_oi_remove_table(env, o, i);
607 * Populate the OI table based.
610 osd_oi_open_table(const struct lu_env *env, struct osd_device *o, int count)
616 for (i = 0; i < count; i++) {
617 sprintf(name, "%s.%d", DMU_OSD_OI_NAME_BASE, i);
618 rc = osd_oi_add_table(env, o, name, i);
620 osd_oi_close_table(env, o);
629 * Determine if the type and number of OIs used by this file system.
632 osd_oi_probe(const struct lu_env *env, struct osd_device *o, int *count)
634 uint64_t root_oid = o->od_root;
641 * Check for multiple OIs and determine the count. There is no
642 * gap handling, if an OI is missing the wrong size can be returned.
643 * The only safeguard is that we know the number of OIs must be a
644 * power of two and this is checked for basic sanity.
646 for (*count = 0; *count < OSD_OI_FID_NR_MAX; (*count)++) {
647 sprintf(name, "%s.%d", DMU_OSD_OI_NAME_BASE, *count);
648 rc = osd_oi_lookup(env, o, root_oid, name, &oi);
656 if ((*count & (*count - 1)) != 0)
666 * No OIs exist, this must be a new filesystem.
673 static void osd_ost_seq_fini(const struct lu_env *env, struct osd_device *osd)
675 struct osd_seq_list *osl = &osd->od_seq_list;
676 struct osd_seq *osd_seq, *tmp;
678 write_lock(&osl->osl_seq_list_lock);
679 list_for_each_entry_safe(osd_seq, tmp, &osl->osl_seq_list,
681 list_del(&osd_seq->os_seq_list);
682 OBD_FREE(osd_seq->os_compat_dirs,
683 sizeof(uint64_t) * osd_seq->os_subdir_count);
684 OBD_FREE(osd_seq, sizeof(*osd_seq));
686 write_unlock(&osl->osl_seq_list_lock);
692 * Create /O subdirectory to map legacy OST objects for compatibility.
695 osd_oi_init_compat(const struct lu_env *env, struct osd_device *o)
701 rc = osd_oi_find_or_create(env, o, o->od_root, "O", &sdb);
707 /* Create on-disk indexes to maintain per-UID/GID inode usage.
708 * Those new indexes are created in the top-level ZAP outside the
709 * namespace in order not to confuse ZPL which might interpret those
710 * indexes as directories and assume the values are object IDs */
711 rc = osd_oi_find_or_create(env, o, MASTER_NODE_OBJ,
712 oid2name(ACCT_USER_OID), &odb);
715 o->od_iusr_oid = odb;
717 rc = osd_oi_find_or_create(env, o, MASTER_NODE_OBJ,
718 oid2name(ACCT_GROUP_OID), &odb);
721 o->od_igrp_oid = odb;
727 * Initialize the OIs by either opening or creating them as needed.
729 int osd_oi_init(const struct lu_env *env, struct osd_device *o)
731 char *key = osd_oti_get(env)->oti_buf;
732 int i, rc, count = 0;
735 rc = osd_oi_probe(env, o, &count);
742 count = osd_oi_count;
745 for (i = 0; i < count; i++) {
746 sprintf(key, "%s.%d", DMU_OSD_OI_NAME_BASE, i);
747 rc = osd_oi_find_or_create(env, o, odb, key, &sdb);
753 rc = osd_oi_init_compat(env, o);
757 LASSERT((count & (count - 1)) == 0);
758 o->od_oi_count = count;
759 OBD_ALLOC(o->od_oi_table, sizeof(struct osd_oi *) * count);
760 if (o->od_oi_table == NULL)
763 rc = osd_oi_open_table(env, o, count);
765 OBD_FREE(o->od_oi_table, sizeof(struct osd_oi *) * count);
766 o->od_oi_table = NULL;
772 void osd_oi_fini(const struct lu_env *env, struct osd_device *o)
776 osd_ost_seq_fini(env, o);
778 if (o->od_oi_table != NULL) {
779 (void) osd_oi_close_table(env, o);
780 OBD_FREE(o->od_oi_table,
781 sizeof(struct osd_oi *) * o->od_oi_count);
782 o->od_oi_table = NULL;
789 int osd_options_init(void)
791 /* osd_oi_count - Default number of OIs, 128 works well for ZFS */
792 if (osd_oi_count == 0 || osd_oi_count > OSD_OI_FID_NR_MAX)
793 osd_oi_count = OSD_OI_FID_NR;
795 if ((osd_oi_count & (osd_oi_count - 1)) != 0) {
796 LCONSOLE_WARN("Round up osd_oi_count %d to power2 %d\n",
797 osd_oi_count, size_roundup_power2(osd_oi_count));
798 osd_oi_count = size_roundup_power2(osd_oi_count);
805 * the following set of functions are used to maintain per-thread
806 * cache of FID->ino mapping. this mechanism is used to avoid
807 * expensive LU/OI lookups.
809 struct osd_idmap_cache *osd_idc_find(const struct lu_env *env,
810 struct osd_device *osd,
811 const struct lu_fid *fid)
813 struct osd_thread_info *oti = osd_oti_get(env);
814 struct osd_idmap_cache *idc = oti->oti_ins_cache;
817 for (i = 0; i < oti->oti_ins_cache_used; i++) {
818 if (!lu_fid_eq(&idc[i].oic_fid, fid))
820 if (idc[i].oic_dev != osd)
829 struct osd_idmap_cache *osd_idc_add(const struct lu_env *env,
830 struct osd_device *osd,
831 const struct lu_fid *fid)
833 struct osd_thread_info *oti = osd_oti_get(env);
834 struct osd_idmap_cache *idc;
837 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
838 i = oti->oti_ins_cache_size * 2;
841 i = OSD_INS_CACHE_SIZE;
842 OBD_ALLOC(idc, sizeof(*idc) * i);
844 return ERR_PTR(-ENOMEM);
845 if (oti->oti_ins_cache != NULL) {
846 memcpy(idc, oti->oti_ins_cache,
847 oti->oti_ins_cache_used * sizeof(*idc));
848 OBD_FREE(oti->oti_ins_cache,
849 oti->oti_ins_cache_used * sizeof(*idc));
851 oti->oti_ins_cache = idc;
852 oti->oti_ins_cache_size = i;
855 idc = &oti->oti_ins_cache[oti->oti_ins_cache_used++];
865 * Lookup mapping for the given fid in the cache
867 * Initialize a new one if not found. the initialization checks whether
868 * the object is local or remote. for the local objects, OI is used to
869 * learn dnode#. the function is used when the caller has no information
870 * about the object, e.g. at dt_insert().
872 struct osd_idmap_cache *osd_idc_find_or_init(const struct lu_env *env,
873 struct osd_device *osd,
874 const struct lu_fid *fid)
876 struct osd_idmap_cache *idc;
879 idc = osd_idc_find(env, osd, fid);
883 /* new mapping is needed */
884 idc = osd_idc_add(env, osd, fid);
889 rc = osd_remote_fid(env, osd, fid);
890 if (unlikely(rc < 0))
894 /* the object is local, lookup in OI */
897 rc = osd_fid_lookup(env, osd, fid, &dnode);
898 if (unlikely(rc < 0)) {
899 CERROR("%s: can't lookup: rc = %d\n",
903 LASSERT(dnode < (1ULL << DN_MAX_OBJECT_SHIFT));
904 idc->oic_dnode = dnode;
906 /* the object is remote */
914 * lookup mapping for given FID and fill it from the given object.
915 * the object is local by definition.
917 int osd_idc_find_and_init(const struct lu_env *env, struct osd_device *osd,
918 struct osd_object *obj)
920 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
921 struct osd_idmap_cache *idc;
923 idc = osd_idc_find(env, osd, fid);
925 if (obj->oo_dn == NULL)
927 idc->oic_dnode = obj->oo_dn->dn_object;
931 /* new mapping is needed */
932 idc = osd_idc_add(env, osd, fid);
937 idc->oic_dnode = obj->oo_dn->dn_object;