4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osd-zfs/osd_handler.c
33 * Top-level entry points into osd module
35 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
36 * Author: Mike Pershin <tappro@whamcloud.com>
37 * Author: Johann Lombardi <johann@whamcloud.com>
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <libcfs/libcfs.h>
43 #include <obd_support.h>
44 #include <lustre_net.h>
46 #include <obd_class.h>
47 #include <lustre_disk.h>
48 #include <lustre_fid.h>
49 #include <uapi/linux/lustre/lustre_param.h>
50 #include <md_object.h>
52 #include "osd_internal.h"
54 #include <sys/dnode.h>
59 #include <sys/spa_impl.h>
60 #include <sys/zfs_znode.h>
61 #include <sys/dmu_tx.h>
62 #include <sys/dmu_objset.h>
63 #include <sys/dsl_prop.h>
64 #include <sys/sa_impl.h>
67 struct lu_context_key osd_key;
69 static int osd_txg_sync_delay_us = -1;
71 /* Slab for OSD object allocation */
72 struct kmem_cache *osd_object_kmem;
74 /* Slab to allocate osd_zap_it */
75 struct kmem_cache *osd_zapit_cachep;
77 static struct lu_kmem_descr osd_caches[] = {
79 .ckd_cache = &osd_object_kmem,
80 .ckd_name = "zfs_osd_obj",
81 .ckd_size = sizeof(struct osd_object)
84 .ckd_cache = &osd_zapit_cachep,
85 .ckd_name = "osd_zapit_cache",
86 .ckd_size = sizeof(struct osd_zap_it)
93 static void arc_prune_func(int64_t bytes, void *private)
95 struct osd_device *od = private;
96 struct lu_site *site = &od->od_site;
100 LASSERT(site->ls_obj_hash);
102 rc = lu_env_init(&env, LCT_SHRINKER);
104 CERROR("%s: can't initialize shrinker env: rc = %d\n",
109 lu_site_purge(&env, site, (bytes >> 10));
115 * Concurrency: doesn't access mutable data
117 static int osd_root_get(const struct lu_env *env,
118 struct dt_device *dev, struct lu_fid *f)
120 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
125 * OSD object methods.
129 * Concurrency: shouldn't matter.
131 static void osd_trans_commit_cb(void *cb_data, int error)
133 struct osd_thandle *oh = cb_data;
134 struct thandle *th = &oh->ot_super;
135 struct osd_device *osd = osd_dt_dev(th->th_dev);
136 struct lu_device *lud = &th->th_dev->dd_lu_dev;
137 struct dt_txn_commit_cb *dcb, *tmp;
142 if (error == ECANCELED)
143 CWARN("%s: transaction @0x%p was aborted\n",
144 osd_dt_dev(th->th_dev)->od_svname, th);
146 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
147 osd_dt_dev(th->th_dev)->od_svname, th, error);
150 dt_txn_hook_commit(th);
152 /* call per-transaction callbacks if any */
153 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) {
154 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
155 "commit callback entry: magic=%x name='%s'\n",
156 dcb->dcb_magic, dcb->dcb_name);
157 list_del_init(&dcb->dcb_linkage);
158 dcb->dcb_func(NULL, th, dcb, error);
161 /* Unlike ldiskfs, zfs updates space accounting at commit time.
162 * As a consequence, op_end is called only now to inform the quota slave
163 * component that reserved quota space is now accounted in usage and
164 * should be released. Quota space won't be adjusted at this point since
165 * we can't provide a suitable environment. It will be performed
166 * asynchronously by a lquota thread. */
167 qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);
176 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
178 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
181 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
182 LASSERT(&dcb->dcb_func != NULL);
183 if (dcb->dcb_flags & DCB_TRANS_STOP)
184 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
186 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
192 * Concurrency: shouldn't matter.
194 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
197 struct osd_thandle *oh;
201 oh = container_of0(th, struct osd_thandle, ot_super);
205 rc = dt_txn_hook_start(env, d, th);
209 if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
210 /* Unlike ldiskfs, ZFS checks for available space and returns
211 * -ENOSPC when assigning txg */
214 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
215 if (unlikely(rc != 0)) {
216 struct osd_device *osd = osd_dt_dev(d);
217 /* dmu will call commit callback with error code during abort */
218 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
219 CERROR("%s: failed to start transaction due to ENOSPC"
220 "\n", osd->od_svname);
222 CERROR("%s: can't assign tx: rc = %d\n",
225 /* add commit callback */
226 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
228 osd_oti_get(env)->oti_in_trans = 1;
229 lu_device_get(&d->dd_lu_dev);
235 static void osd_unlinked_list_emptify(const struct lu_env *env,
236 struct osd_device *osd,
237 struct list_head *list, bool free)
239 struct osd_object *obj;
242 while (!list_empty(list)) {
243 obj = list_entry(list->next,
244 struct osd_object, oo_unlinked_linkage);
245 LASSERT(obj->oo_dn != NULL);
246 oid = obj->oo_dn->dn_object;
248 list_del_init(&obj->oo_unlinked_linkage);
250 (void)osd_unlinked_object_free(env, osd, oid);
254 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
256 struct dt_txn_commit_cb *dcb;
257 struct dt_txn_commit_cb *tmp;
259 /* call per-transaction stop callbacks if any */
260 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
262 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
263 "commit callback entry: magic=%x name='%s'\n",
264 dcb->dcb_magic, dcb->dcb_name);
265 list_del_init(&dcb->dcb_linkage);
266 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
271 * Concurrency: shouldn't matter.
273 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
276 struct osd_device *osd = osd_dt_dev(th->th_dev);
277 bool sync = (th->th_sync != 0);
278 struct osd_thandle *oh;
279 struct list_head unlinked;
284 oh = container_of0(th, struct osd_thandle, ot_super);
285 INIT_LIST_HEAD(&unlinked);
286 list_splice_init(&oh->ot_unlinked_list, &unlinked);
288 osd_oti_get(env)->oti_ins_cache_depth--;
289 /* reset OI cache for safety */
290 if (osd_oti_get(env)->oti_ins_cache_depth == 0)
291 osd_oti_get(env)->oti_ins_cache_used = 0;
293 if (oh->ot_assigned == 0) {
295 dmu_tx_abort(oh->ot_tx);
296 osd_object_sa_dirty_rele(env, oh);
297 osd_unlinked_list_emptify(env, osd, &unlinked, false);
298 /* there won't be any commit, release reserved quota space now,
300 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
305 rc = dt_txn_hook_stop(env, th);
307 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
310 osd_trans_stop_cb(oh, rc);
313 txg = oh->ot_tx->tx_txg;
315 osd_object_sa_dirty_rele(env, oh);
316 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
317 * by osd_trans_commit_cb already. */
318 dmu_tx_commit(oh->ot_tx);
319 osd_oti_get(env)->oti_in_trans = 0;
321 osd_unlinked_list_emptify(env, osd, &unlinked, true);
324 if (osd_txg_sync_delay_us < 0)
325 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
327 udelay(osd_txg_sync_delay_us);
333 static struct thandle *osd_trans_create(const struct lu_env *env,
334 struct dt_device *dt)
336 struct osd_device *osd = osd_dt_dev(dt);
337 struct osd_thandle *oh;
343 CERROR("%s: someone try to start transaction under "
344 "readonly mode, should be disabled.\n",
345 osd_name(osd_dt_dev(dt)));
347 RETURN(ERR_PTR(-EROFS));
350 tx = dmu_tx_create(osd->od_os);
352 RETURN(ERR_PTR(-ENOMEM));
354 /* alloc callback data */
358 RETURN(ERR_PTR(-ENOMEM));
362 INIT_LIST_HEAD(&oh->ot_dcb_list);
363 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
364 INIT_LIST_HEAD(&oh->ot_unlinked_list);
365 INIT_LIST_HEAD(&oh->ot_sa_list);
366 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
371 osd_oti_get(env)->oti_ins_cache_depth++;
376 /* Estimate the total number of objects from a number of blocks */
377 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
378 uint64_t nrblocks, uint64_t est_maxblockshift)
380 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
383 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
384 * bump the free dnode estimate to assume blocks at least 64KB in
385 * case of a directory-heavy MDT (at 32KB/directory).
387 if (est_maxblockshift < 16) {
388 nrblocks >>= (16 - est_maxblockshift);
389 est_maxblockshift = 16;
393 * Estimate the total number of dnodes from the total blocks count
394 * and the space used per dnode. Since we don't know the overhead
395 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
396 * just using DNODE_SHIFT isn't going to give a good estimate.
397 * Instead, compute the current average space usage per dnode, with
398 * an upper and lower cap to avoid unrealistic estimates..
400 * In case there aren't many dnodes or blocks used yet, add a small
401 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
402 * gradually disappears as the number of real dnodes grows. It also
403 * avoids the need to check for divide-by-zero computing dn_per_block.
405 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
406 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
408 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
409 usedbytes) >> est_maxblockshift;
410 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
412 if (est_usedobjs <= est_usedblocks) {
414 * Average space/dnode more than maximum block size, use max
415 * block size to estimate free dnodes from adjusted free blocks
416 * count. OSTs typically use multiple blocks per dnode so this
419 est_totobjs = nrblocks;
421 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
423 * Average space/dnode smaller than min dnode size (probably
424 * due to metadnode compression), use min dnode size to
425 * estimate object count. MDTs may use only one block per node
426 * so this case applies.
428 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
432 * Between the extremes, use average space per existing dnode
433 * to compute the number of dnodes that will fit into nrblocks:
435 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
437 * this may overflow 64 bits or become 0 if not handled well.
439 * We know nrblocks is below 2^(64 - blkbits) bits, and
440 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
441 * which means that multiplying them may get as large as
442 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
444 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
445 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
446 * the limit checks above, so we can safely compute this first.
447 * We care more about accuracy on the MDT (many dnodes/block)
448 * which is good because this is where truncation errors are
449 * smallest. Since both nrblocks and dn_per_block are a
450 * function of blkbits, their product is at most:
452 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
454 * so we can safely use 7 bits to compute a fixed-point
455 * fraction and est_totobjs can still fit in 64 bits.
457 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
459 est_totobjs = (nrblocks * dn_per_block) >> 7;
464 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
466 struct objset *os = osd->od_os;
467 uint64_t usedbytes, availbytes, usedobjs, availobjs;
468 uint64_t est_availobjs;
472 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
474 memset(osfs, 0, sizeof(*osfs));
476 /* We're a zfs filesystem. */
477 osfs->os_type = UBERBLOCK_MAGIC;
480 * ZFS allows multiple block sizes. For statfs, Linux makes no
481 * proper distinction between bsize and frsize. For calculations
482 * of free and used blocks incorrectly uses bsize instead of frsize,
483 * but bsize is also used as the optimal blocksize. We return the
484 * largest possible block size as IO size for the optimum performance
485 * and scale the free and used blocks count appropriately.
487 osfs->os_bsize = osd->od_max_blksz;
488 bshift = fls64(osfs->os_bsize) - 1;
490 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
491 osfs->os_bfree = availbytes >> bshift;
492 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
494 /* Take replication (i.e. number of copies) into account */
495 if (os->os_copies != 0)
496 osfs->os_bavail /= os->os_copies;
499 * Reserve some space so we don't run into ENOSPC due to grants not
500 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
501 * Rather than report this via os_bavail (which makes users unhappy if
502 * they can't fill the filesystem 100%), reduce os_blocks as well.
504 * Reserve 0.78% of total space, at least 16MB for small filesystems,
505 * for internal files to be created/unlinked when space is tight.
507 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
508 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
509 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
510 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
512 osfs->os_blocks -= reserved;
513 osfs->os_bfree -= min(reserved, osfs->os_bfree);
514 osfs->os_bavail -= min(reserved, osfs->os_bavail);
517 * The availobjs value returned from dmu_objset_space() is largely
518 * useless, since it reports the number of objects that might
519 * theoretically still fit into the dataset, independent of minor
520 * issues like how much space is actually available in the pool.
521 * Compute a better estimate in udmu_objs_count_estimate().
523 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
524 osfs->os_bfree, bshift);
526 osfs->os_ffree = min(availobjs, est_availobjs);
527 osfs->os_files = osfs->os_ffree + usedobjs;
529 /* ZFS XXX: fill in backing dataset FSID/UUID
530 memcpy(osfs->os_fsid, .... );*/
532 osfs->os_namelen = MAXNAMELEN;
533 osfs->os_maxbytes = OBD_OBJECT_EOF;
535 if (!spa_writeable(dmu_objset_spa(os)) ||
536 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
537 osfs->os_state |= OS_STATE_READONLY;
543 * Concurrency: shouldn't matter.
545 int osd_statfs(const struct lu_env *env, struct dt_device *d,
546 struct obd_statfs *osfs)
551 rc = osd_objset_statfs(osd_dt_dev(d), osfs);
552 if (unlikely(rc != 0))
555 osfs->os_bavail -= min_t(u64,
556 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
561 static int osd_blk_insert_cost(struct osd_device *osd)
563 int max_blockshift, nr_blkptrshift, bshift;
565 /* max_blockshift is the log2 of the number of blocks needed to reach
566 * the maximum filesize (that's to say 2^64) */
567 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
568 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
570 /* nr_blkptrshift is the log2 of the number of block pointers that can
571 * be stored in an indirect block */
572 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
573 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
575 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
576 * tree. We add +1 for rounding purpose.
577 * The tree depth times the indirect block size gives us the maximum
578 * cost of inserting a block in the tree */
579 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
583 * Concurrency: doesn't access mutable data.
585 static void osd_conf_get(const struct lu_env *env,
586 const struct dt_device *dev,
587 struct dt_device_param *param)
589 struct osd_device *osd = osd_dt_dev(dev);
592 * XXX should be taken from not-yet-existing fs abstraction layer.
594 param->ddp_max_name_len = MAXNAMELEN;
595 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
596 param->ddp_symlink_max = PATH_MAX;
597 param->ddp_mount_type = LDD_MT_ZFS;
599 param->ddp_mntopts = MNTOPT_USERXATTR;
600 if (osd->od_posix_acl)
601 param->ddp_mntopts |= MNTOPT_ACL;
602 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
604 /* for maxbytes, report same value as ZPL */
605 param->ddp_maxbytes = MAX_LFS_FILESIZE;
607 /* inodes are dynamically allocated, so we report the per-inode space
608 * consumption to upper layers. This static value is not really accurate
609 * and we should use the same logic as in udmu_objset_statfs() to
610 * estimate the real size consumed by an object */
611 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
612 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
613 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
614 * not be accounted for every single new block insertion.
615 * Instead, the maximum extent size is set to the number of blocks that
616 * can fit into a single contiguous indirect block. There would be some
617 * cases where this crosses indirect blocks, but it also won't have 7
618 * new levels of indirect blocks in that case either, so it will still
619 * have enough reserved space for the extra indirect block */
620 param->ddp_max_extent_blks =
621 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
622 param->ddp_extent_tax = osd_blk_insert_cost(osd);
624 /* Preferred RPC size for efficient disk IO. 1MB shows good
625 * all-around performance for ZFS, but use blocksize (recordsize)
626 * by default if larger to avoid read-modify-write. */
627 if (osd->od_max_blksz > ONE_MB_BRW_SIZE)
628 param->ddp_brw_size = osd->od_max_blksz;
630 param->ddp_brw_size = ONE_MB_BRW_SIZE;
634 * Concurrency: shouldn't matter.
636 static int osd_sync(const struct lu_env *env, struct dt_device *d)
639 struct osd_device *osd = osd_dt_dev(d);
641 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
642 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
643 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
649 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
651 struct osd_device *osd = osd_dt_dev(dev);
652 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
655 mutex_enter(&tx->tx_sync_lock);
656 txg = tx->tx_open_txg + 1;
657 if (tx->tx_quiesce_txg_waiting < txg) {
658 tx->tx_quiesce_txg_waiting = txg;
659 cv_broadcast(&tx->tx_quiesce_more_cv);
661 mutex_exit(&tx->tx_sync_lock);
667 * Concurrency: shouldn't matter.
669 static int osd_ro(const struct lu_env *env, struct dt_device *d)
671 struct osd_device *osd = osd_dt_dev(d);
674 CERROR("%s: *** setting device %s read-only ***\n",
675 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
676 osd->od_dev_set_rdonly = 1;
677 spa_freeze(dmu_objset_spa(osd->od_os));
682 static struct dt_device_operations osd_dt_ops = {
683 .dt_root_get = osd_root_get,
684 .dt_statfs = osd_statfs,
685 .dt_trans_create = osd_trans_create,
686 .dt_trans_start = osd_trans_start,
687 .dt_trans_stop = osd_trans_stop,
688 .dt_trans_cb_add = osd_trans_cb_add,
689 .dt_conf_get = osd_conf_get,
691 .dt_commit_async = osd_commit_async,
696 * DMU OSD device type methods
698 static int osd_type_init(struct lu_device_type *t)
700 LU_CONTEXT_KEY_INIT(&osd_key);
701 return lu_context_key_register(&osd_key);
704 static void osd_type_fini(struct lu_device_type *t)
706 lu_context_key_degister(&osd_key);
709 static void *osd_key_init(const struct lu_context *ctx,
710 struct lu_context_key *key)
712 struct osd_thread_info *info;
716 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
718 info = ERR_PTR(-ENOMEM);
722 static void osd_key_fini(const struct lu_context *ctx,
723 struct lu_context_key *key, void *data)
725 struct osd_thread_info *info = data;
726 struct osd_idmap_cache *idc = info->oti_ins_cache;
729 LASSERT(info->oti_ins_cache_size > 0);
730 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
731 info->oti_ins_cache = NULL;
732 info->oti_ins_cache_size = 0;
734 lu_buf_free(&info->oti_xattr_lbuf);
738 static void osd_key_exit(const struct lu_context *ctx,
739 struct lu_context_key *key, void *data)
743 struct lu_context_key osd_key = {
744 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
745 .lct_init = osd_key_init,
746 .lct_fini = osd_key_fini,
747 .lct_exit = osd_key_exit
750 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
752 if (osd->od_cl_seq == NULL)
755 seq_client_fini(osd->od_cl_seq);
756 OBD_FREE_PTR(osd->od_cl_seq);
757 osd->od_cl_seq = NULL;
760 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
764 /* shutdown quota slave instance associated with the device */
765 if (o->od_quota_slave != NULL) {
766 /* complete all in-flight callbacks */
767 osd_sync(env, &o->od_dt_dev);
768 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
769 qsd_fini(env, o->od_quota_slave);
770 o->od_quota_slave = NULL;
773 osd_fid_fini(env, o);
778 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
780 struct osd_device *osd = arg;
782 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
785 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
787 struct osd_device *osd = arg;
789 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
790 LASSERT(newval >= SPA_MINBLOCKSIZE);
791 LASSERT(ISP2(newval));
793 osd->od_max_blksz = newval;
796 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
798 struct osd_device *osd = arg;
800 osd->od_prop_rdonly = !!newval;
803 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
804 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
806 struct osd_device *osd = arg;
808 osd->od_dnsize = newval;
812 * This function unregisters all registered callbacks. It's harmless to
813 * unregister callbacks that were never registered so it is used to safely
814 * unwind a partially completed call to osd_objset_register_callbacks().
816 static void osd_objset_unregister_callbacks(struct osd_device *o)
818 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
820 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
821 osd_xattr_changed_cb, o);
822 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
823 osd_recordsize_changed_cb, o);
824 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
825 osd_readonly_changed_cb, o);
826 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
827 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
828 osd_dnodesize_changed_cb, o);
831 if (o->arc_prune_cb != NULL) {
832 arc_remove_prune_callback(o->arc_prune_cb);
833 o->arc_prune_cb = NULL;
838 * Register the required callbacks to be notified when zfs properties
839 * are modified using the 'zfs(8)' command line utility.
841 static int osd_objset_register_callbacks(struct osd_device *o)
843 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
844 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
850 dsl_pool_config_enter(dp, FTAG);
851 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
852 osd_xattr_changed_cb, o);
856 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
857 osd_recordsize_changed_cb, o);
861 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
862 osd_readonly_changed_cb, o);
866 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
867 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
868 osd_dnodesize_changed_cb, o);
873 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
875 dsl_pool_config_exit(dp, FTAG);
877 osd_objset_unregister_callbacks(o);
882 static int osd_objset_open(struct osd_device *o)
884 uint64_t version = ZPL_VERSION;
885 uint64_t sa_obj, unlink_obj;
889 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
890 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
891 B_FALSE, o, &o->od_os);
894 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
900 /* Check ZFS version */
901 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
902 ZPL_VERSION_STR, 8, 1, &version);
904 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
906 * We can't return ENOENT because that would mean the objset
909 GOTO(out, rc = -EIO);
912 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
913 ZFS_SA_ATTRS, 8, 1, &sa_obj);
917 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
918 ZPL_END, &o->z_attr_table);
922 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
923 8, 1, &o->od_rootid);
925 CERROR("%s: lookup for root failed: rc = %d\n",
930 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
933 CERROR("%s: lookup for %s failed: rc = %d\n",
934 o->od_svname, ZFS_UNLINKED_SET, rc);
938 /* Check that user/group usage tracking is supported */
939 if (!dmu_objset_userused_enabled(o->od_os) ||
940 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
941 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
942 CERROR("%s: Space accounting not supported by this target, "
943 "aborting\n", o->od_svname);
944 GOTO(out, rc = -ENOTSUPP);
947 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
949 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
955 if (rc != 0 && o->od_os != NULL) {
956 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
963 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
966 char *key = osd_oti_get(env)->oti_str;
970 if (osd->od_dt_dev.dd_rdonly) {
971 CERROR("%s: someone try to free objects under "
972 "readonly mode, should be disabled.\n", osd_name(osd));
978 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
980 CWARN("%s: Cannot truncate %llu: rc = %d\n",
981 osd->od_svname, oid, rc);
985 tx = dmu_tx_create(osd->od_os);
986 dmu_tx_mark_netfree(tx);
987 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
988 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
990 rc = -dmu_tx_assign(tx, TXG_WAIT);
992 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
993 osd->od_svname, oid, rc);
997 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
998 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
999 osd->od_unlinked, key, tx);
1001 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
1002 osd->od_svname, oid, rc);
1006 rc = -dmu_object_free(osd->od_os, oid, tx);
1008 CWARN("%s: Cannot free %llu: rc = %d\n",
1009 osd->od_svname, oid, rc);
1024 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
1027 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1029 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1031 while (zap_cursor_retrieve(&zc, za) == 0) {
1032 /* If cannot free the object, leave it in the unlinked set,
1033 * until the OSD is mounted again when obd_unlinked_drain()
1034 * will be called. */
1035 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1037 zap_cursor_advance(&zc);
1040 zap_cursor_fini(&zc);
1043 static int osd_mount(const struct lu_env *env,
1044 struct osd_device *o, struct lustre_cfg *cfg)
1046 char *mntdev = lustre_cfg_string(cfg, 1);
1047 char *str = lustre_cfg_string(cfg, 2);
1048 char *svname = lustre_cfg_string(cfg, 4);
1054 if (o->od_os != NULL)
1057 if (mntdev == NULL || svname == NULL)
1060 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1061 if (rc >= sizeof(o->od_mntdev))
1064 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1065 if (rc >= sizeof(o->od_svname))
1068 o->od_index_backup_stop = 0;
1069 o->od_index = -1; /* -1 means index is invalid */
1070 rc = server_name2index(o->od_svname, &o->od_index, NULL);
1071 str = strstr(str, ":");
1073 unsigned long flags;
1075 rc = kstrtoul(str + 1, 10, &flags);
1079 if (flags & LMD_FLG_DEV_RDONLY) {
1080 o->od_dt_dev.dd_rdonly = 1;
1081 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1085 if (flags & LMD_FLG_NOSCRUB)
1086 o->od_auto_scrub_interval = AS_NEVER;
1089 if (server_name_is_ost(o->od_svname))
1092 rc = osd_objset_open(o);
1096 o->od_xattr_in_sa = B_TRUE;
1097 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1098 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
1100 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1103 o->od_root = rootdn->dn_object;
1104 osd_dnode_rele(rootdn);
1106 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1107 &o->od_userused_dn);
1111 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1112 &o->od_groupused_dn);
1116 #ifdef ZFS_PROJINHERIT
1117 if (dmu_objset_projectquota_enabled(o->od_os)) {
1118 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1119 &o->od_projectused_dn);
1120 if (rc && rc != -ENOENT)
1125 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1128 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1130 rc = lu_site_init_finish(&o->od_site);
1134 rc = osd_objset_register_callbacks(o);
1139 rc = osd_scrub_setup(env, o);
1144 rc = osd_procfs_init(o, o->od_svname);
1148 /* initialize quota slave instance */
1149 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
1151 if (IS_ERR(o->od_quota_slave)) {
1152 rc = PTR_ERR(o->od_quota_slave);
1153 o->od_quota_slave = NULL;
1157 #ifdef HAVE_DMU_USEROBJ_ACCOUNTING
1158 if (!osd_dmu_userobj_accounting_available(o))
1159 CWARN("%s: dnode accounting not enabled: "
1160 "enable feature@userobj_accounting in pool\n",
1164 /* parse mount option "noacl", and enable ACL by default */
1165 opts = lustre_cfg_string(cfg, 3);
1166 if (opts == NULL || strstr(opts, "noacl") == NULL)
1167 o->od_posix_acl = 1;
1169 osd_unlinked_drain(env, o);
1171 if (rc && o->od_os) {
1172 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
1179 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1183 if (atomic_read(&o->od_zerocopy_alloc))
1184 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1185 atomic_read(&o->od_zerocopy_alloc));
1186 if (atomic_read(&o->od_zerocopy_loan))
1187 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1188 atomic_read(&o->od_zerocopy_loan));
1189 if (atomic_read(&o->od_zerocopy_pin))
1190 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1191 atomic_read(&o->od_zerocopy_pin));
1193 if (o->od_unlinked) {
1194 osd_dnode_rele(o->od_unlinked);
1195 o->od_unlinked = NULL;
1197 if (o->od_userused_dn) {
1198 osd_dnode_rele(o->od_userused_dn);
1199 o->od_userused_dn = NULL;
1201 if (o->od_groupused_dn) {
1202 osd_dnode_rele(o->od_groupused_dn);
1203 o->od_groupused_dn = NULL;
1206 #ifdef ZFS_PROJINHERIT
1207 if (o->od_projectused_dn) {
1208 osd_dnode_rele(o->od_projectused_dn);
1209 o->od_projectused_dn = NULL;
1213 if (o->od_os != NULL) {
1214 if (!o->od_dt_dev.dd_rdonly)
1215 /* force a txg sync to get all commit callbacks */
1216 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1218 /* close the object set */
1219 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
1226 static int osd_device_init0(const struct lu_env *env,
1227 struct osd_device *o,
1228 struct lustre_cfg *cfg)
1230 struct lu_device *l = osd2lu_dev(o);
1233 /* if the module was re-loaded, env can loose its keys */
1234 rc = lu_env_refill((struct lu_env *) env);
1238 l->ld_ops = &osd_lu_ops;
1239 o->od_dt_dev.dd_ops = &osd_dt_ops;
1240 sema_init(&o->od_otable_sem, 1);
1241 INIT_LIST_HEAD(&o->od_ios_list);
1242 o->od_auto_scrub_interval = AS_DEFAULT;
1248 static struct lu_device *osd_device_fini(const struct lu_env *env,
1249 struct lu_device *dev);
1251 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1252 struct lu_device_type *type,
1253 struct lustre_cfg *cfg)
1255 struct osd_device *dev;
1256 struct osd_seq_list *osl;
1261 return ERR_PTR(-ENOMEM);
1263 osl = &dev->od_seq_list;
1264 INIT_LIST_HEAD(&osl->osl_seq_list);
1265 rwlock_init(&osl->osl_seq_list_lock);
1266 sema_init(&osl->osl_seq_init_sem, 1);
1267 INIT_LIST_HEAD(&dev->od_index_backup_list);
1268 INIT_LIST_HEAD(&dev->od_index_restore_list);
1269 spin_lock_init(&dev->od_lock);
1270 dev->od_index_backup_policy = LIBP_NONE;
1272 rc = dt_device_init(&dev->od_dt_dev, type);
1274 rc = osd_device_init0(env, dev, cfg);
1276 rc = osd_mount(env, dev, cfg);
1278 osd_device_fini(env, osd2lu_dev(dev));
1281 dt_device_fini(&dev->od_dt_dev);
1284 if (unlikely(rc != 0))
1287 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1290 static struct lu_device *osd_device_free(const struct lu_env *env,
1291 struct lu_device *d)
1293 struct osd_device *o = osd_dev(d);
1296 /* XXX: make osd top device in order to release reference */
1297 d->ld_site->ls_top_dev = d;
1298 lu_site_purge(env, d->ld_site, -1);
1299 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1300 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1301 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1303 lu_site_fini(&o->od_site);
1304 dt_device_fini(&o->od_dt_dev);
1310 static struct lu_device *osd_device_fini(const struct lu_env *env,
1311 struct lu_device *d)
1313 struct osd_device *o = osd_dev(d);
1319 osd_objset_unregister_callbacks(o);
1320 if (!o->od_dt_dev.dd_rdonly) {
1321 osd_sync(env, lu2dt_dev(d));
1323 spa_get_dsl(dmu_objset_spa(o->od_os)));
1327 /* now with all the callbacks completed we can cleanup the remainings */
1328 osd_shutdown(env, o);
1329 osd_scrub_cleanup(env, o);
1331 rc = osd_procfs_fini(o);
1333 CERROR("proc fini error %d\n", rc);
1334 RETURN(ERR_PTR(rc));
1343 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1344 const char *name, struct lu_device *next)
1350 * To be removed, setup is performed by osd_device_{init,alloc} and
1351 * cleanup is performed by osd_device_{fini,free).
1353 static int osd_process_config(const struct lu_env *env,
1354 struct lu_device *d, struct lustre_cfg *cfg)
1356 struct osd_device *o = osd_dev(d);
1360 switch(cfg->lcfg_command) {
1362 rc = osd_mount(env, o, cfg);
1365 /* For the case LCFG_PRE_CLEANUP is not called in advance,
1366 * that may happend if hit failure during mount process. */
1367 osd_index_backup(env, o, false);
1368 rc = osd_shutdown(env, o);
1371 LASSERT(&o->od_dt_dev);
1372 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1373 cfg, &o->od_dt_dev);
1374 if (rc > 0 || rc == -ENOSYS) {
1375 rc = class_process_proc_param(PARAM_OST,
1376 lprocfs_osd_obd_vars,
1377 cfg, &o->od_dt_dev);
1383 case LCFG_PRE_CLEANUP:
1385 osd_index_backup(env, o,
1386 o->od_index_backup_policy != LIBP_NONE);
1396 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1398 struct osd_device *osd = osd_dev(d);
1402 if (osd->od_quota_slave == NULL)
1405 /* start qsd instance on recovery completion, this notifies the quota
1406 * slave code that we are about to process new requests now */
1407 rc = qsd_start(env, osd->od_quota_slave);
1412 * we use exports to track all osd users
1414 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1415 struct obd_device *obd, struct obd_uuid *cluuid,
1416 struct obd_connect_data *data, void *localdata)
1418 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1419 struct lustre_handle conn;
1423 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1425 rc = class_connect(&conn, obd, cluuid);
1429 *exp = class_conn2export(&conn);
1431 spin_lock(&obd->obd_dev_lock);
1433 spin_unlock(&obd->obd_dev_lock);
1439 * once last export (we don't count self-export) disappeared
1440 * osd can be released
1442 static int osd_obd_disconnect(struct obd_export *exp)
1444 struct obd_device *obd = exp->exp_obd;
1445 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1446 int rc, release = 0;
1449 /* Only disconnect the underlying layers on the final disconnect. */
1450 spin_lock(&obd->obd_dev_lock);
1452 if (osd->od_connects == 0)
1454 spin_unlock(&obd->obd_dev_lock);
1456 rc = class_disconnect(exp); /* bz 9811 */
1458 if (rc == 0 && release)
1459 class_manual_cleanup(obd);
1463 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1465 struct seq_server_site *ss = osd_seq_site(osd);
1469 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1472 if (unlikely(ss == NULL))
1475 OBD_ALLOC_PTR(osd->od_cl_seq);
1476 if (osd->od_cl_seq == NULL)
1479 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1480 osd->od_svname, ss->ss_server_seq);
1483 OBD_FREE_PTR(osd->od_cl_seq);
1484 osd->od_cl_seq = NULL;
1490 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1491 struct lu_device *dev)
1493 struct osd_device *osd = osd_dev(dev);
1497 if (osd->od_quota_slave != NULL) {
1498 /* set up quota slave objects */
1499 rc = qsd_prepare(env, osd->od_quota_slave);
1504 rc = osd_fid_init(env, osd);
1509 struct lu_device_operations osd_lu_ops = {
1510 .ldo_object_alloc = osd_object_alloc,
1511 .ldo_process_config = osd_process_config,
1512 .ldo_recovery_complete = osd_recovery_complete,
1513 .ldo_prepare = osd_prepare,
1516 static void osd_type_start(struct lu_device_type *t)
1520 static void osd_type_stop(struct lu_device_type *t)
1524 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1525 struct lu_fid *fid, struct md_op_data *op_data)
1527 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1529 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1532 static struct lu_device_type_operations osd_device_type_ops = {
1533 .ldto_init = osd_type_init,
1534 .ldto_fini = osd_type_fini,
1536 .ldto_start = osd_type_start,
1537 .ldto_stop = osd_type_stop,
1539 .ldto_device_alloc = osd_device_alloc,
1540 .ldto_device_free = osd_device_free,
1542 .ldto_device_init = osd_device_init,
1543 .ldto_device_fini = osd_device_fini
1546 static struct lu_device_type osd_device_type = {
1547 .ldt_tags = LU_DEVICE_DT,
1548 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1549 .ldt_ops = &osd_device_type_ops,
1550 .ldt_ctx_tags = LCT_LOCAL
1554 static struct obd_ops osd_obd_device_ops = {
1555 .o_owner = THIS_MODULE,
1556 .o_connect = osd_obd_connect,
1557 .o_disconnect = osd_obd_disconnect,
1558 .o_fid_alloc = osd_fid_alloc
1561 static int __init osd_init(void)
1565 rc = osd_options_init();
1569 rc = lu_kmem_init(osd_caches);
1573 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1574 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1576 lu_kmem_fini(osd_caches);
1580 static void __exit osd_exit(void)
1582 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1583 lu_kmem_fini(osd_caches);
1586 module_param(osd_oi_count, int, 0444);
1587 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1589 module_param(osd_txg_sync_delay_us, int, 0644);
1590 MODULE_PARM_DESC(osd_txg_sync_delay_us,
1591 "When zero or larger delay N usec instead of doing TXG sync");
1593 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1594 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1595 MODULE_VERSION(LUSTRE_VERSION_STRING);
1596 MODULE_LICENSE("GPL");
1598 module_init(osd_init);
1599 module_exit(osd_exit);