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_dt, &oh->ot_quota_trans);
168 if (osd->od_quota_slave_md != NULL)
169 qsd_op_end(NULL, osd->od_quota_slave_md, &oh->ot_quota_trans);
178 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
180 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
183 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
184 LASSERT(&dcb->dcb_func != NULL);
185 if (dcb->dcb_flags & DCB_TRANS_STOP)
186 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
188 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
194 * Concurrency: shouldn't matter.
196 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
199 struct osd_device *osd = osd_dt_dev(d);
200 struct osd_thandle *oh;
205 oh = container_of0(th, struct osd_thandle, ot_super);
209 rc = dt_txn_hook_start(env, d, th);
211 CERROR("%s: dt_txn_hook_start failed: rc = %d\n",
216 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
217 /* Unlike ldiskfs, ZFS checks for available space and returns
218 * -ENOSPC when assigning txg */
221 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
222 if (unlikely(rc != 0)) {
223 /* dmu will call commit callback with error code during abort */
224 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
225 CERROR("%s: failed to start transaction due to ENOSPC"
226 "\n", osd->od_svname);
228 CERROR("%s: can't assign tx: rc = %d\n",
231 /* add commit callback */
232 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
234 osd_oti_get(env)->oti_in_trans = 1;
235 lu_device_get(&d->dd_lu_dev);
241 static void osd_unlinked_list_emptify(const struct lu_env *env,
242 struct osd_device *osd,
243 struct list_head *list, bool free)
245 struct osd_object *obj;
248 while (!list_empty(list)) {
249 obj = list_entry(list->next,
250 struct osd_object, oo_unlinked_linkage);
251 LASSERT(obj->oo_dn != NULL);
252 oid = obj->oo_dn->dn_object;
254 list_del_init(&obj->oo_unlinked_linkage);
256 (void)osd_unlinked_object_free(env, osd, oid);
260 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
262 struct dt_txn_commit_cb *dcb;
263 struct dt_txn_commit_cb *tmp;
265 /* call per-transaction stop callbacks if any */
266 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
268 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
269 "commit callback entry: magic=%x name='%s'\n",
270 dcb->dcb_magic, dcb->dcb_name);
271 list_del_init(&dcb->dcb_linkage);
272 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
277 * Concurrency: shouldn't matter.
279 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
282 struct osd_device *osd = osd_dt_dev(th->th_dev);
283 bool sync = (th->th_sync != 0);
284 struct osd_thandle *oh;
285 struct list_head unlinked;
290 oh = container_of0(th, struct osd_thandle, ot_super);
291 INIT_LIST_HEAD(&unlinked);
292 list_splice_init(&oh->ot_unlinked_list, &unlinked);
294 osd_oti_get(env)->oti_ins_cache_depth--;
295 /* reset OI cache for safety */
296 if (osd_oti_get(env)->oti_ins_cache_depth == 0)
297 osd_oti_get(env)->oti_ins_cache_used = 0;
299 if (oh->ot_assigned == 0) {
301 CDEBUG(D_OTHER, "%s: transaction is aborted\n", osd->od_svname);
302 osd_trans_stop_cb(oh, th->th_result);
303 dmu_tx_abort(oh->ot_tx);
304 osd_object_sa_dirty_rele(env, oh);
305 osd_unlinked_list_emptify(env, osd, &unlinked, false);
306 /* there won't be any commit, release reserved quota space now,
308 qsd_op_end(env, osd->od_quota_slave_dt, &oh->ot_quota_trans);
309 if (osd->od_quota_slave_md != NULL)
310 qsd_op_end(env, osd->od_quota_slave_md,
311 &oh->ot_quota_trans);
316 rc = dt_txn_hook_stop(env, th);
318 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
321 osd_trans_stop_cb(oh, rc);
324 txg = oh->ot_tx->tx_txg;
326 osd_object_sa_dirty_rele(env, oh);
327 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
328 * by osd_trans_commit_cb already. */
329 dmu_tx_commit(oh->ot_tx);
330 osd_oti_get(env)->oti_in_trans = 0;
332 osd_unlinked_list_emptify(env, osd, &unlinked, true);
335 if (osd_txg_sync_delay_us < 0)
336 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
338 udelay(osd_txg_sync_delay_us);
344 static struct thandle *osd_trans_create(const struct lu_env *env,
345 struct dt_device *dt)
347 struct osd_device *osd = osd_dt_dev(dt);
348 struct osd_thandle *oh;
354 CERROR("%s: someone try to start transaction under "
355 "readonly mode, should be disabled.\n",
356 osd_name(osd_dt_dev(dt)));
358 RETURN(ERR_PTR(-EROFS));
361 tx = dmu_tx_create(osd->od_os);
363 RETURN(ERR_PTR(-ENOMEM));
365 /* alloc callback data */
369 RETURN(ERR_PTR(-ENOMEM));
373 INIT_LIST_HEAD(&oh->ot_dcb_list);
374 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
375 INIT_LIST_HEAD(&oh->ot_unlinked_list);
376 INIT_LIST_HEAD(&oh->ot_sa_list);
377 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
382 osd_oti_get(env)->oti_ins_cache_depth++;
387 /* Estimate the total number of objects from a number of blocks */
388 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
389 uint64_t nrblocks, uint64_t est_maxblockshift)
391 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
394 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
395 * bump the free dnode estimate to assume blocks at least 64KB in
396 * case of a directory-heavy MDT (at 32KB/directory).
398 if (est_maxblockshift < 16) {
399 nrblocks >>= (16 - est_maxblockshift);
400 est_maxblockshift = 16;
404 * Estimate the total number of dnodes from the total blocks count
405 * and the space used per dnode. Since we don't know the overhead
406 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
407 * just using DNODE_SHIFT isn't going to give a good estimate.
408 * Instead, compute the current average space usage per dnode, with
409 * an upper and lower cap to avoid unrealistic estimates..
411 * In case there aren't many dnodes or blocks used yet, add a small
412 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
413 * gradually disappears as the number of real dnodes grows. It also
414 * avoids the need to check for divide-by-zero computing dn_per_block.
416 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
417 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
419 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
420 usedbytes) >> est_maxblockshift;
421 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
423 if (est_usedobjs <= est_usedblocks) {
425 * Average space/dnode more than maximum block size, use max
426 * block size to estimate free dnodes from adjusted free blocks
427 * count. OSTs typically use multiple blocks per dnode so this
430 est_totobjs = nrblocks;
432 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
434 * Average space/dnode smaller than min dnode size (probably
435 * due to metadnode compression), use min dnode size to
436 * estimate object count. MDTs may use only one block per node
437 * so this case applies.
439 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
443 * Between the extremes, use average space per existing dnode
444 * to compute the number of dnodes that will fit into nrblocks:
446 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
448 * this may overflow 64 bits or become 0 if not handled well.
450 * We know nrblocks is below 2^(64 - blkbits) bits, and
451 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
452 * which means that multiplying them may get as large as
453 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
455 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
456 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
457 * the limit checks above, so we can safely compute this first.
458 * We care more about accuracy on the MDT (many dnodes/block)
459 * which is good because this is where truncation errors are
460 * smallest. Since both nrblocks and dn_per_block are a
461 * function of blkbits, their product is at most:
463 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
465 * so we can safely use 7 bits to compute a fixed-point
466 * fraction and est_totobjs can still fit in 64 bits.
468 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
470 est_totobjs = (nrblocks * dn_per_block) >> 7;
475 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
477 struct objset *os = osd->od_os;
478 uint64_t usedbytes, availbytes, usedobjs, availobjs;
479 uint64_t est_availobjs;
483 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
485 memset(osfs, 0, sizeof(*osfs));
487 /* We're a zfs filesystem. */
488 osfs->os_type = UBERBLOCK_MAGIC;
491 * ZFS allows multiple block sizes. For statfs, Linux makes no
492 * proper distinction between bsize and frsize. For calculations
493 * of free and used blocks incorrectly uses bsize instead of frsize,
494 * but bsize is also used as the optimal blocksize. We return the
495 * largest possible block size as IO size for the optimum performance
496 * and scale the free and used blocks count appropriately.
498 osfs->os_bsize = osd->od_max_blksz;
499 bshift = fls64(osfs->os_bsize) - 1;
501 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
502 osfs->os_bfree = availbytes >> bshift;
503 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
505 /* Take replication (i.e. number of copies) into account */
506 if (os->os_copies != 0)
507 osfs->os_bavail /= os->os_copies;
510 * Reserve some space so we don't run into ENOSPC due to grants not
511 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
512 * Rather than report this via os_bavail (which makes users unhappy if
513 * they can't fill the filesystem 100%), reduce os_blocks as well.
515 * Reserve 0.78% of total space, at least 16MB for small filesystems,
516 * for internal files to be created/unlinked when space is tight.
518 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
519 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
520 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
521 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
523 osfs->os_blocks -= reserved;
524 osfs->os_bfree -= min(reserved, osfs->os_bfree);
525 osfs->os_bavail -= min(reserved, osfs->os_bavail);
528 * The availobjs value returned from dmu_objset_space() is largely
529 * useless, since it reports the number of objects that might
530 * theoretically still fit into the dataset, independent of minor
531 * issues like how much space is actually available in the pool.
532 * Compute a better estimate in udmu_objs_count_estimate().
534 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
535 osfs->os_bfree, bshift);
537 osfs->os_ffree = min(availobjs, est_availobjs);
538 osfs->os_files = osfs->os_ffree + usedobjs;
540 /* ZFS XXX: fill in backing dataset FSID/UUID
541 memcpy(osfs->os_fsid, .... );*/
543 osfs->os_namelen = MAXNAMELEN;
544 osfs->os_maxbytes = OBD_OBJECT_EOF;
546 if (!spa_writeable(dmu_objset_spa(os)) ||
547 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
548 osfs->os_state |= OS_STATE_READONLY;
554 * Concurrency: shouldn't matter.
556 int osd_statfs(const struct lu_env *env, struct dt_device *d,
557 struct obd_statfs *osfs)
562 rc = osd_objset_statfs(osd_dt_dev(d), osfs);
563 if (unlikely(rc != 0))
566 osfs->os_bavail -= min_t(u64,
567 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
572 static int osd_blk_insert_cost(struct osd_device *osd)
574 int max_blockshift, nr_blkptrshift, bshift;
576 /* max_blockshift is the log2 of the number of blocks needed to reach
577 * the maximum filesize (that's to say 2^64) */
578 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
579 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
581 /* nr_blkptrshift is the log2 of the number of block pointers that can
582 * be stored in an indirect block */
583 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
584 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
586 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
587 * tree. We add +1 for rounding purpose.
588 * The tree depth times the indirect block size gives us the maximum
589 * cost of inserting a block in the tree */
590 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
594 * Concurrency: doesn't access mutable data.
596 static void osd_conf_get(const struct lu_env *env,
597 const struct dt_device *dev,
598 struct dt_device_param *param)
600 struct osd_device *osd = osd_dt_dev(dev);
603 * XXX should be taken from not-yet-existing fs abstraction layer.
605 param->ddp_max_name_len = MAXNAMELEN;
606 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
607 param->ddp_symlink_max = PATH_MAX;
608 param->ddp_mount_type = LDD_MT_ZFS;
610 param->ddp_mntopts = MNTOPT_USERXATTR;
611 if (osd->od_posix_acl)
612 param->ddp_mntopts |= MNTOPT_ACL;
613 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
615 /* for maxbytes, report same value as ZPL */
616 param->ddp_maxbytes = MAX_LFS_FILESIZE;
618 /* inodes are dynamically allocated, so we report the per-inode space
619 * consumption to upper layers. This static value is not really accurate
620 * and we should use the same logic as in udmu_objset_statfs() to
621 * estimate the real size consumed by an object */
622 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
623 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
624 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
625 * not be accounted for every single new block insertion.
626 * Instead, the maximum extent size is set to the number of blocks that
627 * can fit into a single contiguous indirect block. There would be some
628 * cases where this crosses indirect blocks, but it also won't have 7
629 * new levels of indirect blocks in that case either, so it will still
630 * have enough reserved space for the extra indirect block */
631 param->ddp_max_extent_blks =
632 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
633 param->ddp_extent_tax = osd_blk_insert_cost(osd);
635 /* Preferred RPC size for efficient disk IO. 1MB shows good
636 * all-around performance for ZFS, but use blocksize (recordsize)
637 * by default if larger to avoid read-modify-write. */
638 if (osd->od_max_blksz > ONE_MB_BRW_SIZE)
639 param->ddp_brw_size = osd->od_max_blksz;
641 param->ddp_brw_size = ONE_MB_BRW_SIZE;
645 * Concurrency: shouldn't matter.
647 static int osd_sync(const struct lu_env *env, struct dt_device *d)
650 struct osd_device *osd = osd_dt_dev(d);
652 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
653 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
654 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
660 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
662 struct osd_device *osd = osd_dt_dev(dev);
663 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
666 mutex_enter(&tx->tx_sync_lock);
667 txg = tx->tx_open_txg + 1;
668 if (tx->tx_quiesce_txg_waiting < txg) {
669 tx->tx_quiesce_txg_waiting = txg;
670 cv_broadcast(&tx->tx_quiesce_more_cv);
672 mutex_exit(&tx->tx_sync_lock);
678 * Concurrency: shouldn't matter.
680 static int osd_ro(const struct lu_env *env, struct dt_device *d)
682 struct osd_device *osd = osd_dt_dev(d);
685 CERROR("%s: *** setting device %s read-only ***\n",
686 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
687 osd->od_dev_set_rdonly = 1;
688 spa_freeze(dmu_objset_spa(osd->od_os));
693 static struct dt_device_operations osd_dt_ops = {
694 .dt_root_get = osd_root_get,
695 .dt_statfs = osd_statfs,
696 .dt_trans_create = osd_trans_create,
697 .dt_trans_start = osd_trans_start,
698 .dt_trans_stop = osd_trans_stop,
699 .dt_trans_cb_add = osd_trans_cb_add,
700 .dt_conf_get = osd_conf_get,
702 .dt_commit_async = osd_commit_async,
707 * DMU OSD device type methods
709 static int osd_type_init(struct lu_device_type *t)
711 LU_CONTEXT_KEY_INIT(&osd_key);
712 return lu_context_key_register(&osd_key);
715 static void osd_type_fini(struct lu_device_type *t)
717 lu_context_key_degister(&osd_key);
720 static void *osd_key_init(const struct lu_context *ctx,
721 struct lu_context_key *key)
723 struct osd_thread_info *info;
727 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
729 info = ERR_PTR(-ENOMEM);
733 static void osd_key_fini(const struct lu_context *ctx,
734 struct lu_context_key *key, void *data)
736 struct osd_thread_info *info = data;
737 struct osd_idmap_cache *idc = info->oti_ins_cache;
740 LASSERT(info->oti_ins_cache_size > 0);
741 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
742 info->oti_ins_cache = NULL;
743 info->oti_ins_cache_size = 0;
745 lu_buf_free(&info->oti_xattr_lbuf);
749 static void osd_key_exit(const struct lu_context *ctx,
750 struct lu_context_key *key, void *data)
754 struct lu_context_key osd_key = {
755 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
756 .lct_init = osd_key_init,
757 .lct_fini = osd_key_fini,
758 .lct_exit = osd_key_exit
761 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
763 if (osd->od_cl_seq == NULL)
766 seq_client_fini(osd->od_cl_seq);
767 OBD_FREE_PTR(osd->od_cl_seq);
768 osd->od_cl_seq = NULL;
771 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
775 /* shutdown quota slave instance associated with the device */
776 if (o->od_quota_slave_md != NULL) {
777 /* complete all in-flight callbacks */
778 osd_sync(env, &o->od_dt_dev);
779 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
780 qsd_fini(env, o->od_quota_slave_md);
781 o->od_quota_slave_md = NULL;
784 if (o->od_quota_slave_dt != NULL) {
785 /* complete all in-flight callbacks */
786 osd_sync(env, &o->od_dt_dev);
787 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
788 qsd_fini(env, o->od_quota_slave_dt);
789 o->od_quota_slave_dt = NULL;
791 osd_fid_fini(env, o);
796 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
798 struct osd_device *osd = arg;
800 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
803 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
805 struct osd_device *osd = arg;
807 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
808 LASSERT(newval >= SPA_MINBLOCKSIZE);
809 LASSERT(ISP2(newval));
811 osd->od_max_blksz = newval;
814 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
816 struct osd_device *osd = arg;
818 osd->od_prop_rdonly = !!newval;
821 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
822 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
824 struct osd_device *osd = arg;
826 osd->od_dnsize = newval;
830 * This function unregisters all registered callbacks. It's harmless to
831 * unregister callbacks that were never registered so it is used to safely
832 * unwind a partially completed call to osd_objset_register_callbacks().
834 static void osd_objset_unregister_callbacks(struct osd_device *o)
836 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
838 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
839 osd_xattr_changed_cb, o);
840 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
841 osd_recordsize_changed_cb, o);
842 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
843 osd_readonly_changed_cb, o);
844 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
845 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
846 osd_dnodesize_changed_cb, o);
849 if (o->arc_prune_cb != NULL) {
850 arc_remove_prune_callback(o->arc_prune_cb);
851 o->arc_prune_cb = NULL;
856 * Register the required callbacks to be notified when zfs properties
857 * are modified using the 'zfs(8)' command line utility.
859 static int osd_objset_register_callbacks(struct osd_device *o)
861 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
862 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
868 dsl_pool_config_enter(dp, FTAG);
869 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
870 osd_xattr_changed_cb, o);
874 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
875 osd_recordsize_changed_cb, o);
879 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
880 osd_readonly_changed_cb, o);
884 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
885 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
886 osd_dnodesize_changed_cb, o);
891 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
893 dsl_pool_config_exit(dp, FTAG);
895 osd_objset_unregister_callbacks(o);
900 static int osd_objset_open(struct osd_device *o)
902 uint64_t version = ZPL_VERSION;
903 uint64_t sa_obj, unlink_obj;
907 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
908 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
909 B_FALSE, o, &o->od_os);
912 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
918 /* Check ZFS version */
919 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
920 ZPL_VERSION_STR, 8, 1, &version);
922 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
924 * We can't return ENOENT because that would mean the objset
927 GOTO(out, rc = -EIO);
930 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
931 ZFS_SA_ATTRS, 8, 1, &sa_obj);
935 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
936 ZPL_END, &o->z_attr_table);
940 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
941 8, 1, &o->od_rootid);
943 CERROR("%s: lookup for root failed: rc = %d\n",
948 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
951 CERROR("%s: lookup for %s failed: rc = %d\n",
952 o->od_svname, ZFS_UNLINKED_SET, rc);
956 /* Check that user/group usage tracking is supported */
957 if (!dmu_objset_userused_enabled(o->od_os) ||
958 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
959 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
960 CERROR("%s: Space accounting not supported by this target, "
961 "aborting\n", o->od_svname);
962 GOTO(out, rc = -ENOTSUPP);
965 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
967 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
973 if (rc != 0 && o->od_os != NULL) {
974 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
981 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
984 char *key = osd_oti_get(env)->oti_str;
988 if (osd->od_dt_dev.dd_rdonly) {
989 CERROR("%s: someone try to free objects under "
990 "readonly mode, should be disabled.\n", osd_name(osd));
996 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
998 CWARN("%s: Cannot truncate %llu: rc = %d\n",
999 osd->od_svname, oid, rc);
1003 tx = dmu_tx_create(osd->od_os);
1004 dmu_tx_mark_netfree(tx);
1005 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
1006 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
1008 rc = -dmu_tx_assign(tx, TXG_WAIT);
1010 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
1011 osd->od_svname, oid, rc);
1015 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
1016 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
1017 osd->od_unlinked, key, tx);
1019 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
1020 osd->od_svname, oid, rc);
1024 rc = -dmu_object_free(osd->od_os, oid, tx);
1026 CWARN("%s: Cannot free %llu: rc = %d\n",
1027 osd->od_svname, oid, rc);
1042 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
1045 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1047 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1049 while (zap_cursor_retrieve(&zc, za) == 0) {
1050 /* If cannot free the object, leave it in the unlinked set,
1051 * until the OSD is mounted again when obd_unlinked_drain()
1052 * will be called. */
1053 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1055 zap_cursor_advance(&zc);
1058 zap_cursor_fini(&zc);
1061 static int osd_mount(const struct lu_env *env,
1062 struct osd_device *o, struct lustre_cfg *cfg)
1064 char *mntdev = lustre_cfg_string(cfg, 1);
1065 char *str = lustre_cfg_string(cfg, 2);
1066 char *svname = lustre_cfg_string(cfg, 4);
1072 if (o->od_os != NULL)
1075 if (mntdev == NULL || svname == NULL)
1078 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1079 if (rc >= sizeof(o->od_mntdev))
1082 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1083 if (rc >= sizeof(o->od_svname))
1086 o->od_index_backup_stop = 0;
1087 o->od_index = -1; /* -1 means index is invalid */
1088 rc = server_name2index(o->od_svname, &o->od_index, NULL);
1089 str = strstr(str, ":");
1091 unsigned long flags;
1093 rc = kstrtoul(str + 1, 10, &flags);
1097 if (flags & LMD_FLG_DEV_RDONLY) {
1098 o->od_dt_dev.dd_rdonly = 1;
1099 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1103 if (flags & LMD_FLG_NOSCRUB)
1104 o->od_auto_scrub_interval = AS_NEVER;
1107 if (server_name_is_ost(o->od_svname))
1110 rc = osd_objset_open(o);
1114 o->od_xattr_in_sa = B_TRUE;
1115 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1116 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
1118 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1121 o->od_root = rootdn->dn_object;
1122 osd_dnode_rele(rootdn);
1124 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1125 &o->od_userused_dn);
1129 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1130 &o->od_groupused_dn);
1134 #ifdef ZFS_PROJINHERIT
1135 if (dmu_objset_projectquota_enabled(o->od_os)) {
1136 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1137 &o->od_projectused_dn);
1138 if (rc && rc != -ENOENT)
1143 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1146 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1148 rc = lu_site_init_finish(&o->od_site);
1152 rc = osd_objset_register_callbacks(o);
1157 rc = osd_scrub_setup(env, o);
1162 rc = osd_procfs_init(o, o->od_svname);
1166 /* currently it's no need to prepare qsd_instance_md for OST */
1167 if (!o->od_is_ost) {
1168 o->od_quota_slave_md = qsd_init(env, o->od_svname,
1170 o->od_proc_entry, true);
1171 if (IS_ERR(o->od_quota_slave_md)) {
1172 rc = PTR_ERR(o->od_quota_slave_md);
1173 o->od_quota_slave_md = NULL;
1178 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
1179 o->od_proc_entry, false);
1181 if (IS_ERR(o->od_quota_slave_dt)) {
1182 if (o->od_quota_slave_md != NULL) {
1183 qsd_fini(env, o->od_quota_slave_md);
1184 o->od_quota_slave_md = NULL;
1187 rc = PTR_ERR(o->od_quota_slave_dt);
1188 o->od_quota_slave_dt = NULL;
1192 #ifdef HAVE_DMU_USEROBJ_ACCOUNTING
1193 if (!osd_dmu_userobj_accounting_available(o))
1194 CWARN("%s: dnode accounting not enabled: "
1195 "enable feature@userobj_accounting in pool\n",
1199 /* parse mount option "noacl", and enable ACL by default */
1200 opts = lustre_cfg_string(cfg, 3);
1201 if (opts == NULL || strstr(opts, "noacl") == NULL)
1202 o->od_posix_acl = 1;
1204 osd_unlinked_drain(env, o);
1206 if (rc && o->od_os) {
1207 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
1214 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1218 if (atomic_read(&o->od_zerocopy_alloc))
1219 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1220 atomic_read(&o->od_zerocopy_alloc));
1221 if (atomic_read(&o->od_zerocopy_loan))
1222 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1223 atomic_read(&o->od_zerocopy_loan));
1224 if (atomic_read(&o->od_zerocopy_pin))
1225 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1226 atomic_read(&o->od_zerocopy_pin));
1228 if (o->od_unlinked) {
1229 osd_dnode_rele(o->od_unlinked);
1230 o->od_unlinked = NULL;
1232 if (o->od_userused_dn) {
1233 osd_dnode_rele(o->od_userused_dn);
1234 o->od_userused_dn = NULL;
1236 if (o->od_groupused_dn) {
1237 osd_dnode_rele(o->od_groupused_dn);
1238 o->od_groupused_dn = NULL;
1241 #ifdef ZFS_PROJINHERIT
1242 if (o->od_projectused_dn) {
1243 osd_dnode_rele(o->od_projectused_dn);
1244 o->od_projectused_dn = NULL;
1248 if (o->od_os != NULL) {
1249 if (!o->od_dt_dev.dd_rdonly)
1250 /* force a txg sync to get all commit callbacks */
1251 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1253 /* close the object set */
1254 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
1261 static int osd_device_init0(const struct lu_env *env,
1262 struct osd_device *o,
1263 struct lustre_cfg *cfg)
1265 struct lu_device *l = osd2lu_dev(o);
1268 /* if the module was re-loaded, env can loose its keys */
1269 rc = lu_env_refill((struct lu_env *) env);
1273 l->ld_ops = &osd_lu_ops;
1274 o->od_dt_dev.dd_ops = &osd_dt_ops;
1275 sema_init(&o->od_otable_sem, 1);
1276 INIT_LIST_HEAD(&o->od_ios_list);
1277 o->od_auto_scrub_interval = AS_DEFAULT;
1283 static struct lu_device *osd_device_fini(const struct lu_env *env,
1284 struct lu_device *dev);
1286 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1287 struct lu_device_type *type,
1288 struct lustre_cfg *cfg)
1290 struct osd_device *dev;
1291 struct osd_seq_list *osl;
1296 return ERR_PTR(-ENOMEM);
1298 osl = &dev->od_seq_list;
1299 INIT_LIST_HEAD(&osl->osl_seq_list);
1300 rwlock_init(&osl->osl_seq_list_lock);
1301 sema_init(&osl->osl_seq_init_sem, 1);
1302 INIT_LIST_HEAD(&dev->od_index_backup_list);
1303 INIT_LIST_HEAD(&dev->od_index_restore_list);
1304 spin_lock_init(&dev->od_lock);
1305 dev->od_index_backup_policy = LIBP_NONE;
1307 rc = dt_device_init(&dev->od_dt_dev, type);
1309 rc = osd_device_init0(env, dev, cfg);
1311 rc = osd_mount(env, dev, cfg);
1313 osd_device_fini(env, osd2lu_dev(dev));
1316 dt_device_fini(&dev->od_dt_dev);
1319 if (unlikely(rc != 0))
1322 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1325 static struct lu_device *osd_device_free(const struct lu_env *env,
1326 struct lu_device *d)
1328 struct osd_device *o = osd_dev(d);
1331 /* XXX: make osd top device in order to release reference */
1332 d->ld_site->ls_top_dev = d;
1333 lu_site_purge(env, d->ld_site, -1);
1334 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1335 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1336 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1338 lu_site_fini(&o->od_site);
1339 dt_device_fini(&o->od_dt_dev);
1345 static struct lu_device *osd_device_fini(const struct lu_env *env,
1346 struct lu_device *d)
1348 struct osd_device *o = osd_dev(d);
1354 osd_objset_unregister_callbacks(o);
1355 if (!o->od_dt_dev.dd_rdonly) {
1356 osd_sync(env, lu2dt_dev(d));
1358 spa_get_dsl(dmu_objset_spa(o->od_os)));
1362 /* now with all the callbacks completed we can cleanup the remainings */
1363 osd_shutdown(env, o);
1364 osd_scrub_cleanup(env, o);
1366 rc = osd_procfs_fini(o);
1368 CERROR("proc fini error %d\n", rc);
1369 RETURN(ERR_PTR(rc));
1378 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1379 const char *name, struct lu_device *next)
1385 * To be removed, setup is performed by osd_device_{init,alloc} and
1386 * cleanup is performed by osd_device_{fini,free).
1388 static int osd_process_config(const struct lu_env *env,
1389 struct lu_device *d, struct lustre_cfg *cfg)
1391 struct osd_device *o = osd_dev(d);
1395 switch(cfg->lcfg_command) {
1397 rc = osd_mount(env, o, cfg);
1400 /* For the case LCFG_PRE_CLEANUP is not called in advance,
1401 * that may happend if hit failure during mount process. */
1402 osd_index_backup(env, o, false);
1403 rc = osd_shutdown(env, o);
1406 LASSERT(&o->od_dt_dev);
1407 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1408 cfg, &o->od_dt_dev);
1409 if (rc > 0 || rc == -ENOSYS) {
1410 rc = class_process_proc_param(PARAM_OST,
1411 lprocfs_osd_obd_vars,
1412 cfg, &o->od_dt_dev);
1418 case LCFG_PRE_CLEANUP:
1420 osd_index_backup(env, o,
1421 o->od_index_backup_policy != LIBP_NONE);
1431 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1433 struct osd_device *osd = osd_dev(d);
1437 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
1440 /* start qsd instance on recovery completion, this notifies the quota
1441 * slave code that we are about to process new requests now */
1442 rc = qsd_start(env, osd->od_quota_slave_dt);
1443 if (rc == 0 && osd->od_quota_slave_md != NULL)
1444 rc = qsd_start(env, osd->od_quota_slave_md);
1449 * we use exports to track all osd users
1451 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1452 struct obd_device *obd, struct obd_uuid *cluuid,
1453 struct obd_connect_data *data, void *localdata)
1455 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1456 struct lustre_handle conn;
1460 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1462 rc = class_connect(&conn, obd, cluuid);
1466 *exp = class_conn2export(&conn);
1468 spin_lock(&obd->obd_dev_lock);
1470 spin_unlock(&obd->obd_dev_lock);
1476 * once last export (we don't count self-export) disappeared
1477 * osd can be released
1479 static int osd_obd_disconnect(struct obd_export *exp)
1481 struct obd_device *obd = exp->exp_obd;
1482 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1483 int rc, release = 0;
1486 /* Only disconnect the underlying layers on the final disconnect. */
1487 spin_lock(&obd->obd_dev_lock);
1489 if (osd->od_connects == 0)
1491 spin_unlock(&obd->obd_dev_lock);
1493 rc = class_disconnect(exp); /* bz 9811 */
1495 if (rc == 0 && release)
1496 class_manual_cleanup(obd);
1500 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1502 struct seq_server_site *ss = osd_seq_site(osd);
1506 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1509 if (unlikely(ss == NULL))
1512 OBD_ALLOC_PTR(osd->od_cl_seq);
1513 if (osd->od_cl_seq == NULL)
1516 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1517 osd->od_svname, ss->ss_server_seq);
1520 OBD_FREE_PTR(osd->od_cl_seq);
1521 osd->od_cl_seq = NULL;
1524 if (ss->ss_node_id == 0) {
1526 * If the OSD on the sequence controller(MDT0), then allocate
1527 * sequence here, otherwise allocate sequence after connected
1528 * to MDT0 (see mdt_register_lwp_callback()).
1530 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
1531 &osd->od_cl_seq->lcs_space, env);
1537 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1538 struct lu_device *dev)
1540 struct osd_device *osd = osd_dev(dev);
1544 if (osd->od_quota_slave_md != NULL) {
1545 /* set up quota slave objects */
1546 rc = qsd_prepare(env, osd->od_quota_slave_md);
1551 if (osd->od_quota_slave_dt != NULL) {
1552 /* set up quota slave objects */
1553 rc = qsd_prepare(env, osd->od_quota_slave_dt);
1558 rc = osd_fid_init(env, osd);
1563 struct lu_device_operations osd_lu_ops = {
1564 .ldo_object_alloc = osd_object_alloc,
1565 .ldo_process_config = osd_process_config,
1566 .ldo_recovery_complete = osd_recovery_complete,
1567 .ldo_prepare = osd_prepare,
1570 static void osd_type_start(struct lu_device_type *t)
1574 static void osd_type_stop(struct lu_device_type *t)
1578 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1579 struct lu_fid *fid, struct md_op_data *op_data)
1581 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1583 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1586 static struct lu_device_type_operations osd_device_type_ops = {
1587 .ldto_init = osd_type_init,
1588 .ldto_fini = osd_type_fini,
1590 .ldto_start = osd_type_start,
1591 .ldto_stop = osd_type_stop,
1593 .ldto_device_alloc = osd_device_alloc,
1594 .ldto_device_free = osd_device_free,
1596 .ldto_device_init = osd_device_init,
1597 .ldto_device_fini = osd_device_fini
1600 static struct lu_device_type osd_device_type = {
1601 .ldt_tags = LU_DEVICE_DT,
1602 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1603 .ldt_ops = &osd_device_type_ops,
1604 .ldt_ctx_tags = LCT_LOCAL
1608 static struct obd_ops osd_obd_device_ops = {
1609 .o_owner = THIS_MODULE,
1610 .o_connect = osd_obd_connect,
1611 .o_disconnect = osd_obd_disconnect,
1612 .o_fid_alloc = osd_fid_alloc
1615 static int __init osd_init(void)
1619 rc = osd_options_init();
1623 rc = lu_kmem_init(osd_caches);
1627 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1628 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1630 lu_kmem_fini(osd_caches);
1634 static void __exit osd_exit(void)
1636 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1637 lu_kmem_fini(osd_caches);
1640 module_param(osd_oi_count, int, 0444);
1641 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1643 module_param(osd_txg_sync_delay_us, int, 0644);
1644 MODULE_PARM_DESC(osd_txg_sync_delay_us,
1645 "When zero or larger delay N usec instead of doing TXG sync");
1647 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1648 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1649 MODULE_VERSION(LUSTRE_VERSION_STRING);
1650 MODULE_LICENSE("GPL");
1652 module_init(osd_init);
1653 module_exit(osd_exit);