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 rc = lu_env_init(&env, LCT_SHRINKER);
102 CERROR("%s: can't initialize shrinker env: rc = %d\n",
107 lu_site_purge(&env, site, (bytes >> 10));
113 * Concurrency: doesn't access mutable data
115 static int osd_root_get(const struct lu_env *env,
116 struct dt_device *dev, struct lu_fid *f)
118 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
123 * OSD object methods.
127 * Concurrency: shouldn't matter.
129 static void osd_trans_commit_cb(void *cb_data, int error)
131 struct osd_thandle *oh = cb_data;
132 struct thandle *th = &oh->ot_super;
133 struct osd_device *osd = osd_dt_dev(th->th_dev);
134 struct lu_device *lud = &th->th_dev->dd_lu_dev;
135 struct dt_txn_commit_cb *dcb, *tmp;
140 if (error == ECANCELED)
141 CWARN("%s: transaction @0x%p was aborted\n",
142 osd_dt_dev(th->th_dev)->od_svname, th);
144 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
145 osd_dt_dev(th->th_dev)->od_svname, th, error);
148 /* call per-transaction callbacks if any */
149 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) {
150 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
151 "commit callback entry: magic=%x name='%s'\n",
152 dcb->dcb_magic, dcb->dcb_name);
153 list_del_init(&dcb->dcb_linkage);
154 dcb->dcb_func(NULL, th, dcb, error);
157 /* Unlike ldiskfs, zfs updates space accounting at commit time.
158 * As a consequence, op_end is called only now to inform the quota slave
159 * component that reserved quota space is now accounted in usage and
160 * should be released. Quota space won't be adjusted at this point since
161 * we can't provide a suitable environment. It will be performed
162 * asynchronously by a lquota thread. */
163 qsd_op_end(NULL, osd->od_quota_slave_dt, &oh->ot_quota_trans);
164 if (osd->od_quota_slave_md != NULL)
165 qsd_op_end(NULL, osd->od_quota_slave_md, &oh->ot_quota_trans);
174 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
176 struct osd_thandle *oh = container_of(th, struct osd_thandle,
179 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
180 LASSERT(&dcb->dcb_func != NULL);
181 if (dcb->dcb_flags & DCB_TRANS_STOP)
182 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
184 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
190 * Concurrency: shouldn't matter.
192 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
195 struct osd_device *osd = osd_dt_dev(d);
196 struct osd_thandle *oh;
201 oh = container_of(th, struct osd_thandle, ot_super);
205 rc = dt_txn_hook_start(env, d, th);
207 CERROR("%s: dt_txn_hook_start failed: rc = %d\n",
212 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
213 /* Unlike ldiskfs, ZFS checks for available space and returns
214 * -ENOSPC when assigning txg */
217 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
218 if (unlikely(rc != 0)) {
219 /* dmu will call commit callback with error code during abort */
220 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
221 CERROR("%s: failed to start transaction due to ENOSPC"
222 "\n", osd->od_svname);
224 CERROR("%s: can't assign tx: rc = %d\n",
227 /* add commit callback */
228 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
230 osd_oti_get(env)->oti_in_trans = 1;
231 lu_device_get(&d->dd_lu_dev);
237 static void osd_unlinked_list_emptify(const struct lu_env *env,
238 struct osd_device *osd,
239 struct list_head *list, bool free)
241 struct osd_object *obj;
244 while (!list_empty(list)) {
245 obj = list_entry(list->next,
246 struct osd_object, oo_unlinked_linkage);
247 LASSERT(obj->oo_dn != NULL);
248 oid = obj->oo_dn->dn_object;
250 list_del_init(&obj->oo_unlinked_linkage);
252 (void)osd_unlinked_object_free(env, osd, oid);
256 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
258 struct dt_txn_commit_cb *dcb;
259 struct dt_txn_commit_cb *tmp;
261 /* call per-transaction stop callbacks if any */
262 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
264 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
265 "commit callback entry: magic=%x name='%s'\n",
266 dcb->dcb_magic, dcb->dcb_name);
267 list_del_init(&dcb->dcb_linkage);
268 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
273 * Concurrency: shouldn't matter.
275 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
278 struct osd_device *osd = osd_dt_dev(th->th_dev);
279 bool sync = (th->th_sync != 0);
280 struct osd_thandle *oh;
286 oh = container_of(th, struct osd_thandle, ot_super);
287 list_splice_init(&oh->ot_unlinked_list, &unlinked);
289 osd_oti_get(env)->oti_ins_cache_depth--;
290 /* reset OI cache for safety */
291 if (osd_oti_get(env)->oti_ins_cache_depth == 0)
292 osd_oti_get(env)->oti_ins_cache_used = 0;
294 if (oh->ot_assigned == 0) {
296 CDEBUG(D_OTHER, "%s: transaction is aborted\n", osd->od_svname);
297 osd_trans_stop_cb(oh, th->th_result);
298 dmu_tx_abort(oh->ot_tx);
299 osd_object_sa_dirty_rele(env, oh);
300 osd_unlinked_list_emptify(env, osd, &unlinked, false);
301 /* there won't be any commit, release reserved quota space now,
303 qsd_op_end(env, osd->od_quota_slave_dt, &oh->ot_quota_trans);
304 if (osd->od_quota_slave_md != NULL)
305 qsd_op_end(env, osd->od_quota_slave_md,
306 &oh->ot_quota_trans);
311 rc = dt_txn_hook_stop(env, th);
313 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
316 osd_trans_stop_cb(oh, rc);
319 txg = oh->ot_tx->tx_txg;
321 osd_object_sa_dirty_rele(env, oh);
322 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
323 * by osd_trans_commit_cb already. */
324 dmu_tx_commit(oh->ot_tx);
325 osd_oti_get(env)->oti_in_trans = 0;
327 osd_unlinked_list_emptify(env, osd, &unlinked, true);
330 if (osd_txg_sync_delay_us < 0)
331 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
333 udelay(osd_txg_sync_delay_us);
339 static struct thandle *osd_trans_create(const struct lu_env *env,
340 struct dt_device *dt)
342 struct osd_device *osd = osd_dt_dev(dt);
343 struct osd_thandle *oh;
349 CERROR("%s: someone try to start transaction under "
350 "readonly mode, should be disabled.\n",
351 osd_name(osd_dt_dev(dt)));
353 RETURN(ERR_PTR(-EROFS));
356 tx = dmu_tx_create(osd->od_os);
358 RETURN(ERR_PTR(-ENOMEM));
360 /* alloc callback data */
364 RETURN(ERR_PTR(-ENOMEM));
368 INIT_LIST_HEAD(&oh->ot_dcb_list);
369 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
370 INIT_LIST_HEAD(&oh->ot_unlinked_list);
371 INIT_LIST_HEAD(&oh->ot_sa_list);
372 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
377 osd_oti_get(env)->oti_ins_cache_depth++;
382 /* Estimate the total number of objects from a number of blocks */
383 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
384 uint64_t nrblocks, uint64_t est_maxblockshift)
386 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
389 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
390 * bump the free dnode estimate to assume blocks at least 64KB in
391 * case of a directory-heavy MDT (at 32KB/directory).
393 if (est_maxblockshift < 16) {
394 nrblocks >>= (16 - est_maxblockshift);
395 est_maxblockshift = 16;
399 * Estimate the total number of dnodes from the total blocks count
400 * and the space used per dnode. Since we don't know the overhead
401 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
402 * just using DNODE_SHIFT isn't going to give a good estimate.
403 * Instead, compute the current average space usage per dnode, with
404 * an upper and lower cap to avoid unrealistic estimates..
406 * In case there aren't many dnodes or blocks used yet, add a small
407 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
408 * gradually disappears as the number of real dnodes grows. It also
409 * avoids the need to check for divide-by-zero computing dn_per_block.
411 BUILD_BUG_ON(OSD_DNODE_MIN_BLKSHIFT <= 0);
412 BUILD_BUG_ON(OSD_DNODE_EST_BLKSHIFT <= 0);
414 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
415 usedbytes) >> est_maxblockshift;
416 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
418 if (est_usedobjs <= est_usedblocks) {
420 * Average space/dnode more than maximum block size, use max
421 * block size to estimate free dnodes from adjusted free blocks
422 * count. OSTs typically use multiple blocks per dnode so this
425 est_totobjs = nrblocks;
427 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
429 * Average space/dnode smaller than min dnode size (probably
430 * due to metadnode compression), use min dnode size to
431 * estimate object count. MDTs may use only one block per node
432 * so this case applies.
434 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
438 * Between the extremes, use average space per existing dnode
439 * to compute the number of dnodes that will fit into nrblocks:
441 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
443 * this may overflow 64 bits or become 0 if not handled well.
445 * We know nrblocks is below 2^(64 - blkbits) bits, and
446 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
447 * which means that multiplying them may get as large as
448 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
450 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
451 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
452 * the limit checks above, so we can safely compute this first.
453 * We care more about accuracy on the MDT (many dnodes/block)
454 * which is good because this is where truncation errors are
455 * smallest. Since both nrblocks and dn_per_block are a
456 * function of blkbits, their product is at most:
458 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
460 * so we can safely use 7 bits to compute a fixed-point
461 * fraction and est_totobjs can still fit in 64 bits.
463 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
465 est_totobjs = (nrblocks * dn_per_block) >> 7;
470 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
472 struct objset *os = osd->od_os;
473 uint64_t usedbytes, availbytes, usedobjs, availobjs;
474 uint64_t est_availobjs;
478 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
480 memset(osfs, 0, sizeof(*osfs));
482 /* We're a zfs filesystem. */
483 osfs->os_type = UBERBLOCK_MAGIC;
486 * ZFS allows multiple block sizes. For statfs, Linux makes no
487 * proper distinction between bsize and frsize. For calculations
488 * of free and used blocks incorrectly uses bsize instead of frsize,
489 * but bsize is also used as the optimal blocksize. We return the
490 * largest possible block size as IO size for the optimum performance
491 * and scale the free and used blocks count appropriately.
493 osfs->os_bsize = osd->od_max_blksz;
494 bshift = fls64(osfs->os_bsize) - 1;
496 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
497 osfs->os_bfree = availbytes >> bshift;
498 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
500 /* Take replication (i.e. number of copies) into account */
501 if (os->os_copies != 0)
502 osfs->os_bavail /= os->os_copies;
505 * Reserve some space so we don't run into ENOSPC due to grants not
506 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
507 * Rather than report this via os_bavail (which makes users unhappy if
508 * they can't fill the filesystem 100%), reduce os_blocks as well.
510 * Reserve 0.78% of total space, at least 16MB for small filesystems,
511 * for internal files to be created/unlinked when space is tight.
513 BUILD_BUG_ON(OSD_STATFS_RESERVED_SIZE <= 0);
514 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
515 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
516 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
518 osfs->os_blocks -= reserved;
519 osfs->os_bfree -= min(reserved, osfs->os_bfree);
520 osfs->os_bavail -= min(reserved, osfs->os_bavail);
523 * The availobjs value returned from dmu_objset_space() is largely
524 * useless, since it reports the number of objects that might
525 * theoretically still fit into the dataset, independent of minor
526 * issues like how much space is actually available in the pool.
527 * Compute a better estimate in udmu_objs_count_estimate().
529 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
530 osfs->os_bfree, bshift);
532 osfs->os_ffree = min(availobjs, est_availobjs);
533 osfs->os_files = osfs->os_ffree + usedobjs;
535 /* ZFS XXX: fill in backing dataset FSID/UUID
536 memcpy(osfs->os_fsid, .... );*/
538 osfs->os_namelen = MAXNAMELEN;
539 osfs->os_maxbytes = OBD_OBJECT_EOF;
541 if (!spa_writeable(dmu_objset_spa(os)) ||
542 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
543 osfs->os_state |= OS_STATFS_READONLY;
549 * Concurrency: shouldn't matter.
551 int osd_statfs(const struct lu_env *env, struct dt_device *d,
552 struct obd_statfs *osfs, struct obd_statfs_info *info)
554 struct osd_device *osd = osd_dt_dev(d);
558 rc = osd_objset_statfs(osd, osfs);
559 if (unlikely(rc != 0))
562 osfs->os_bavail -= min_t(u64,
563 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
566 /* ZFS does not support reporting nonrotional status yet, so return
567 * flag only if user has set nonrotational.
569 osfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
574 static int osd_blk_insert_cost(struct osd_device *osd)
576 int max_blockshift, nr_blkptrshift, bshift;
578 /* max_blockshift is the log2 of the number of blocks needed to reach
579 * the maximum filesize (that's to say 2^64) */
580 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
581 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
583 /* nr_blkptrshift is the log2 of the number of block pointers that can
584 * be stored in an indirect block */
585 BUILD_BUG_ON(DN_MAX_INDBLKSHIFT <= SPA_BLKPTRSHIFT);
586 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
588 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
589 * tree. We add +1 for rounding purpose.
590 * The tree depth times the indirect block size gives us the maximum
591 * cost of inserting a block in the tree */
592 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
596 * Concurrency: doesn't access mutable data.
598 static void osd_conf_get(const struct lu_env *env,
599 const struct dt_device *dev,
600 struct dt_device_param *param)
602 struct osd_device *osd = osd_dt_dev(dev);
605 * XXX should be taken from not-yet-existing fs abstraction layer.
607 param->ddp_max_name_len = MAXNAMELEN;
608 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
609 param->ddp_symlink_max = PATH_MAX;
610 param->ddp_mount_type = LDD_MT_ZFS;
612 param->ddp_mntopts = MNTOPT_USERXATTR;
613 if (osd->od_posix_acl)
614 param->ddp_mntopts |= MNTOPT_ACL;
615 /* Previously DXATTR_MAX_ENTRY_SIZE */
616 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
618 /* for maxbytes, report same value as ZPL */
619 param->ddp_maxbytes = MAX_LFS_FILESIZE;
621 /* inodes are dynamically allocated, so we report the per-inode space
622 * consumption to upper layers. This static value is not really accurate
623 * and we should use the same logic as in udmu_objset_statfs() to
624 * estimate the real size consumed by an object */
625 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
626 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
627 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
628 * not be accounted for every single new block insertion.
629 * Instead, the maximum extent size is set to the number of blocks that
630 * can fit into a single contiguous indirect block. There would be some
631 * cases where this crosses indirect blocks, but it also won't have 7
632 * new levels of indirect blocks in that case either, so it will still
633 * have enough reserved space for the extra indirect block */
634 param->ddp_max_extent_blks =
635 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
636 param->ddp_extent_tax = osd_blk_insert_cost(osd);
638 /* Preferred RPC size for efficient disk IO. 1MB shows good
639 * all-around performance for ZFS, but use blocksize (recordsize)
640 * by default if larger to avoid read-modify-write. */
641 if (osd->od_max_blksz > ONE_MB_BRW_SIZE)
642 param->ddp_brw_size = osd->od_max_blksz;
644 param->ddp_brw_size = ONE_MB_BRW_SIZE;
648 * Concurrency: shouldn't matter.
650 static int osd_sync(const struct lu_env *env, struct dt_device *d)
653 struct osd_device *osd = osd_dt_dev(d);
655 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
656 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
657 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
663 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
665 struct osd_device *osd = osd_dt_dev(dev);
666 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
669 mutex_enter(&tx->tx_sync_lock);
670 txg = tx->tx_open_txg + 1;
671 if (tx->tx_quiesce_txg_waiting < txg) {
672 tx->tx_quiesce_txg_waiting = txg;
673 cv_broadcast(&tx->tx_quiesce_more_cv);
675 mutex_exit(&tx->tx_sync_lock);
681 * Concurrency: shouldn't matter.
683 static int osd_ro(const struct lu_env *env, struct dt_device *d)
685 struct osd_device *osd = osd_dt_dev(d);
688 CERROR("%s: *** setting device %s read-only ***\n",
689 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
690 osd->od_dev_set_rdonly = 1;
691 spa_freeze(dmu_objset_spa(osd->od_os));
696 static struct dt_device_operations osd_dt_ops = {
697 .dt_root_get = osd_root_get,
698 .dt_statfs = osd_statfs,
699 .dt_trans_create = osd_trans_create,
700 .dt_trans_start = osd_trans_start,
701 .dt_trans_stop = osd_trans_stop,
702 .dt_trans_cb_add = osd_trans_cb_add,
703 .dt_conf_get = osd_conf_get,
705 .dt_commit_async = osd_commit_async,
710 * DMU OSD device type methods
712 static int osd_type_init(struct lu_device_type *t)
714 LU_CONTEXT_KEY_INIT(&osd_key);
715 return lu_context_key_register(&osd_key);
718 static void osd_type_fini(struct lu_device_type *t)
720 lu_context_key_degister(&osd_key);
723 static void *osd_key_init(const struct lu_context *ctx,
724 struct lu_context_key *key)
726 struct osd_thread_info *info;
730 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
732 info = ERR_PTR(-ENOMEM);
736 static void osd_key_fini(const struct lu_context *ctx,
737 struct lu_context_key *key, void *data)
739 struct osd_thread_info *info = data;
740 struct osd_idmap_cache *idc = info->oti_ins_cache;
743 LASSERT(info->oti_ins_cache_size > 0);
744 OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size);
745 info->oti_ins_cache = NULL;
746 info->oti_ins_cache_size = 0;
748 lu_buf_free(&info->oti_xattr_lbuf);
752 static void osd_key_exit(const struct lu_context *ctx,
753 struct lu_context_key *key, void *data)
757 struct lu_context_key osd_key = {
758 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
759 .lct_init = osd_key_init,
760 .lct_fini = osd_key_fini,
761 .lct_exit = osd_key_exit
764 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
766 if (osd->od_cl_seq == NULL)
769 seq_client_fini(osd->od_cl_seq);
770 OBD_FREE_PTR(osd->od_cl_seq);
771 osd->od_cl_seq = NULL;
774 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
778 /* shutdown quota slave instance associated with the device */
779 if (o->od_quota_slave_md != NULL) {
780 /* complete all in-flight callbacks */
781 osd_sync(env, &o->od_dt_dev);
782 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
783 qsd_fini(env, o->od_quota_slave_md);
784 o->od_quota_slave_md = NULL;
787 if (o->od_quota_slave_dt != NULL) {
788 /* complete all in-flight callbacks */
789 osd_sync(env, &o->od_dt_dev);
790 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
791 qsd_fini(env, o->od_quota_slave_dt);
792 o->od_quota_slave_dt = NULL;
794 osd_fid_fini(env, o);
799 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
801 struct osd_device *osd = arg;
803 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
806 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
808 struct osd_device *osd = arg;
810 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
811 LASSERT(newval >= SPA_MINBLOCKSIZE);
812 LASSERT(ISP2(newval));
814 osd->od_max_blksz = newval;
817 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
819 struct osd_device *osd = arg;
821 osd->od_prop_rdonly = !!newval;
824 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
825 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
827 struct osd_device *osd = arg;
829 osd->od_dnsize = newval;
833 * This function unregisters all registered callbacks. It's harmless to
834 * unregister callbacks that were never registered so it is used to safely
835 * unwind a partially completed call to osd_objset_register_callbacks().
837 static void osd_objset_unregister_callbacks(struct osd_device *o)
839 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
841 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
842 osd_xattr_changed_cb, o);
843 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
844 osd_recordsize_changed_cb, o);
845 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
846 osd_readonly_changed_cb, o);
847 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
848 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
849 osd_dnodesize_changed_cb, o);
852 if (o->arc_prune_cb != NULL) {
853 arc_remove_prune_callback(o->arc_prune_cb);
854 o->arc_prune_cb = NULL;
859 * Register the required callbacks to be notified when zfs properties
860 * are modified using the 'zfs(8)' command line utility.
862 static int osd_objset_register_callbacks(struct osd_device *o)
864 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
865 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
871 dsl_pool_config_enter(dp, FTAG);
872 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
873 osd_xattr_changed_cb, o);
877 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
878 osd_recordsize_changed_cb, o);
882 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
883 osd_readonly_changed_cb, o);
887 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
888 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
889 osd_dnodesize_changed_cb, o);
894 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
896 dsl_pool_config_exit(dp, FTAG);
898 osd_objset_unregister_callbacks(o);
903 static int osd_objset_open(struct osd_device *o)
905 uint64_t version = ZPL_VERSION;
906 uint64_t sa_obj, unlink_obj;
910 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
911 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
912 B_TRUE, o, &o->od_os);
915 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
921 /* Check ZFS version */
922 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
923 ZPL_VERSION_STR, 8, 1, &version);
925 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
927 * We can't return ENOENT because that would mean the objset
930 GOTO(out, rc = -EIO);
933 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
934 ZFS_SA_ATTRS, 8, 1, &sa_obj);
938 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
939 ZPL_END, &o->z_attr_table);
943 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
944 8, 1, &o->od_rootid);
946 CERROR("%s: lookup for root failed: rc = %d\n",
951 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
954 CERROR("%s: lookup for %s failed: rc = %d\n",
955 o->od_svname, ZFS_UNLINKED_SET, rc);
959 /* Check that user/group usage tracking is supported */
960 if (!dmu_objset_userused_enabled(o->od_os) ||
961 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
962 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
963 CERROR("%s: Space accounting not supported by this target, "
964 "aborting\n", o->od_svname);
965 GOTO(out, rc = -ENOTSUPP);
968 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
970 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
976 if (rc != 0 && o->od_os != NULL) {
977 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
984 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
987 char *key = osd_oti_get(env)->oti_str;
991 if (osd->od_dt_dev.dd_rdonly) {
992 CERROR("%s: someone try to free objects under "
993 "readonly mode, should be disabled.\n", osd_name(osd));
999 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
1001 CWARN("%s: Cannot truncate %llu: rc = %d\n",
1002 osd->od_svname, oid, rc);
1006 tx = dmu_tx_create(osd->od_os);
1007 dmu_tx_mark_netfree(tx);
1008 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
1009 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
1011 rc = -dmu_tx_assign(tx, TXG_WAIT);
1013 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
1014 osd->od_svname, oid, rc);
1018 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
1019 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
1020 osd->od_unlinked, key, tx);
1022 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
1023 osd->od_svname, oid, rc);
1027 rc = -dmu_object_free(osd->od_os, oid, tx);
1029 CWARN("%s: Cannot free %llu: rc = %d\n",
1030 osd->od_svname, oid, rc);
1045 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
1048 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1050 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1052 while (zap_cursor_retrieve(&zc, za) == 0) {
1053 /* If cannot free the object, leave it in the unlinked set,
1054 * until the OSD is mounted again when obd_unlinked_drain()
1055 * will be called. */
1056 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1058 zap_cursor_advance(&zc);
1061 zap_cursor_fini(&zc);
1064 static int osd_mount(const struct lu_env *env,
1065 struct osd_device *o, struct lustre_cfg *cfg)
1067 char *mntdev = lustre_cfg_string(cfg, 1);
1068 char *str = lustre_cfg_string(cfg, 2);
1069 char *svname = lustre_cfg_string(cfg, 4);
1075 if (o->od_os != NULL)
1078 if (mntdev == NULL || svname == NULL)
1081 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1082 if (rc >= sizeof(o->od_mntdev))
1085 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1086 if (rc >= sizeof(o->od_svname))
1089 o->od_index_backup_stop = 0;
1090 o->od_index = -1; /* -1 means index is invalid */
1091 rc = server_name2index(o->od_svname, &o->od_index, NULL);
1092 str = strstr(str, ":");
1094 unsigned long flags;
1096 rc = kstrtoul(str + 1, 10, &flags);
1100 if (flags & LMD_FLG_DEV_RDONLY) {
1101 o->od_dt_dev.dd_rdonly = 1;
1102 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1106 if (flags & LMD_FLG_NOSCRUB)
1107 o->od_auto_scrub_interval = AS_NEVER;
1110 if (server_name_is_ost(o->od_svname))
1113 rc = osd_objset_open(o);
1117 o->od_xattr_in_sa = B_TRUE;
1118 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1119 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
1121 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1124 o->od_root = rootdn->dn_object;
1125 osd_dnode_rele(rootdn);
1127 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1128 &o->od_userused_dn);
1132 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1133 &o->od_groupused_dn);
1137 #ifdef ZFS_PROJINHERIT
1138 if (dmu_objset_projectquota_enabled(o->od_os)) {
1139 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1140 &o->od_projectused_dn);
1141 if (rc && rc != -ENOENT)
1146 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1149 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1151 rc = lu_site_init_finish(&o->od_site);
1155 rc = osd_objset_register_callbacks(o);
1160 rc = osd_scrub_setup(env, o);
1165 rc = osd_procfs_init(o, o->od_svname);
1169 /* currently it's no need to prepare qsd_instance_md for OST */
1170 if (!o->od_is_ost) {
1171 o->od_quota_slave_md = qsd_init(env, o->od_svname,
1173 o->od_proc_entry, true);
1174 if (IS_ERR(o->od_quota_slave_md)) {
1175 rc = PTR_ERR(o->od_quota_slave_md);
1176 o->od_quota_slave_md = NULL;
1181 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
1182 o->od_proc_entry, false);
1184 if (IS_ERR(o->od_quota_slave_dt)) {
1185 if (o->od_quota_slave_md != NULL) {
1186 qsd_fini(env, o->od_quota_slave_md);
1187 o->od_quota_slave_md = NULL;
1190 rc = PTR_ERR(o->od_quota_slave_dt);
1191 o->od_quota_slave_dt = NULL;
1195 #ifdef HAVE_DMU_USEROBJ_ACCOUNTING
1196 if (!osd_dmu_userobj_accounting_available(o))
1197 CWARN("%s: dnode accounting not enabled: "
1198 "enable feature@userobj_accounting in pool\n",
1202 /* parse mount option "noacl", and enable ACL by default */
1203 opts = lustre_cfg_string(cfg, 3);
1204 if (opts == NULL || strstr(opts, "noacl") == NULL)
1205 o->od_posix_acl = 1;
1207 osd_unlinked_drain(env, o);
1209 if (rc && o->od_os) {
1210 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1217 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1221 if (atomic_read(&o->od_zerocopy_alloc))
1222 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1223 atomic_read(&o->od_zerocopy_alloc));
1224 if (atomic_read(&o->od_zerocopy_loan))
1225 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1226 atomic_read(&o->od_zerocopy_loan));
1227 if (atomic_read(&o->od_zerocopy_pin))
1228 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1229 atomic_read(&o->od_zerocopy_pin));
1231 if (o->od_unlinked) {
1232 osd_dnode_rele(o->od_unlinked);
1233 o->od_unlinked = NULL;
1235 if (o->od_userused_dn) {
1236 osd_dnode_rele(o->od_userused_dn);
1237 o->od_userused_dn = NULL;
1239 if (o->od_groupused_dn) {
1240 osd_dnode_rele(o->od_groupused_dn);
1241 o->od_groupused_dn = NULL;
1244 #ifdef ZFS_PROJINHERIT
1245 if (o->od_projectused_dn) {
1246 osd_dnode_rele(o->od_projectused_dn);
1247 o->od_projectused_dn = NULL;
1251 if (o->od_os != NULL) {
1252 if (!o->od_dt_dev.dd_rdonly)
1253 /* force a txg sync to get all commit callbacks */
1254 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1256 /* close the object set */
1257 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1264 static int osd_device_init0(const struct lu_env *env,
1265 struct osd_device *o,
1266 struct lustre_cfg *cfg)
1268 struct lu_device *l = osd2lu_dev(o);
1271 /* if the module was re-loaded, env can loose its keys */
1272 rc = lu_env_refill((struct lu_env *) env);
1276 l->ld_ops = &osd_lu_ops;
1277 o->od_dt_dev.dd_ops = &osd_dt_ops;
1278 sema_init(&o->od_otable_sem, 1);
1279 INIT_LIST_HEAD(&o->od_ios_list);
1280 o->od_auto_scrub_interval = AS_DEFAULT;
1282 /* ZFS does not support reporting nonrotional status yet, so this flag
1283 * is only set if explicitly set by the user.
1285 o->od_nonrotational = 0;
1291 static struct lu_device *osd_device_fini(const struct lu_env *env,
1292 struct lu_device *dev);
1294 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1295 struct lu_device_type *type,
1296 struct lustre_cfg *cfg)
1298 struct osd_device *dev;
1299 struct osd_seq_list *osl;
1304 return ERR_PTR(-ENOMEM);
1306 osl = &dev->od_seq_list;
1307 INIT_LIST_HEAD(&osl->osl_seq_list);
1308 rwlock_init(&osl->osl_seq_list_lock);
1309 sema_init(&osl->osl_seq_init_sem, 1);
1310 INIT_LIST_HEAD(&dev->od_index_backup_list);
1311 INIT_LIST_HEAD(&dev->od_index_restore_list);
1312 spin_lock_init(&dev->od_lock);
1313 dev->od_index_backup_policy = LIBP_NONE;
1315 rc = dt_device_init(&dev->od_dt_dev, type);
1317 rc = osd_device_init0(env, dev, cfg);
1319 rc = osd_mount(env, dev, cfg);
1321 osd_device_fini(env, osd2lu_dev(dev));
1324 dt_device_fini(&dev->od_dt_dev);
1327 if (unlikely(rc != 0))
1330 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1333 static struct lu_device *osd_device_free(const struct lu_env *env,
1334 struct lu_device *d)
1336 struct osd_device *o = osd_dev(d);
1339 /* XXX: make osd top device in order to release reference */
1340 d->ld_site->ls_top_dev = d;
1341 lu_site_purge(env, d->ld_site, -1);
1342 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
1343 D_ERROR, lu_cdebug_printer);
1344 lu_site_fini(&o->od_site);
1345 dt_device_fini(&o->od_dt_dev);
1351 static struct lu_device *osd_device_fini(const struct lu_env *env,
1352 struct lu_device *d)
1354 struct osd_device *o = osd_dev(d);
1360 osd_objset_unregister_callbacks(o);
1361 if (!o->od_dt_dev.dd_rdonly) {
1362 osd_sync(env, lu2dt_dev(d));
1364 spa_get_dsl(dmu_objset_spa(o->od_os)));
1368 /* now with all the callbacks completed we can cleanup the remainings */
1369 osd_shutdown(env, o);
1370 osd_scrub_cleanup(env, o);
1372 rc = osd_procfs_fini(o);
1374 CERROR("proc fini error %d\n", rc);
1375 RETURN(ERR_PTR(rc));
1384 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1385 const char *name, struct lu_device *next)
1391 * To be removed, setup is performed by osd_device_{init,alloc} and
1392 * cleanup is performed by osd_device_{fini,free).
1394 static int osd_process_config(const struct lu_env *env,
1395 struct lu_device *d, struct lustre_cfg *cfg)
1397 struct osd_device *o = osd_dev(d);
1402 switch(cfg->lcfg_command) {
1404 rc = osd_mount(env, o, cfg);
1407 /* For the case LCFG_PRE_CLEANUP is not called in advance,
1408 * that may happend if hit failure during mount process. */
1409 osd_index_backup(env, o, false);
1410 rc = osd_shutdown(env, o);
1413 LASSERT(&o->od_dt_dev);
1414 count = class_modify_config(cfg, PARAM_OSD,
1415 &o->od_dt_dev.dd_kobj);
1417 count = class_modify_config(cfg, PARAM_OST,
1418 &o->od_dt_dev.dd_kobj);
1419 rc = count > 0 ? 0 : count;
1422 case LCFG_PRE_CLEANUP:
1424 osd_index_backup(env, o,
1425 o->od_index_backup_policy != LIBP_NONE);
1435 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1437 struct osd_device *osd = osd_dev(d);
1441 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
1444 /* start qsd instance on recovery completion, this notifies the quota
1445 * slave code that we are about to process new requests now */
1446 rc = qsd_start(env, osd->od_quota_slave_dt);
1447 if (rc == 0 && osd->od_quota_slave_md != NULL)
1448 rc = qsd_start(env, osd->od_quota_slave_md);
1453 * we use exports to track all osd users
1455 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1456 struct obd_device *obd, struct obd_uuid *cluuid,
1457 struct obd_connect_data *data, void *localdata)
1459 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1460 struct lustre_handle conn;
1464 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1466 rc = class_connect(&conn, obd, cluuid);
1470 *exp = class_conn2export(&conn);
1472 spin_lock(&obd->obd_dev_lock);
1474 spin_unlock(&obd->obd_dev_lock);
1480 * once last export (we don't count self-export) disappeared
1481 * osd can be released
1483 static int osd_obd_disconnect(struct obd_export *exp)
1485 struct obd_device *obd = exp->exp_obd;
1486 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1487 int rc, release = 0;
1490 /* Only disconnect the underlying layers on the final disconnect. */
1491 spin_lock(&obd->obd_dev_lock);
1493 if (osd->od_connects == 0)
1495 spin_unlock(&obd->obd_dev_lock);
1497 rc = class_disconnect(exp); /* bz 9811 */
1499 if (rc == 0 && release)
1500 class_manual_cleanup(obd);
1504 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1506 struct seq_server_site *ss = osd_seq_site(osd);
1510 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1513 if (unlikely(ss == NULL))
1516 OBD_ALLOC_PTR(osd->od_cl_seq);
1517 if (osd->od_cl_seq == NULL)
1520 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1521 osd->od_svname, ss->ss_server_seq);
1523 if (ss->ss_node_id == 0) {
1525 * If the OSD on the sequence controller(MDT0), then allocate
1526 * sequence here, otherwise allocate sequence after connected
1527 * to MDT0 (see mdt_register_lwp_callback()).
1529 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
1530 &osd->od_cl_seq->lcs_space, env);
1536 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1537 struct lu_device *dev)
1539 struct osd_device *osd = osd_dev(dev);
1543 if (osd->od_quota_slave_md != NULL) {
1544 /* set up quota slave objects */
1545 rc = qsd_prepare(env, osd->od_quota_slave_md);
1550 if (osd->od_quota_slave_dt != NULL) {
1551 /* set up quota slave objects */
1552 rc = qsd_prepare(env, osd->od_quota_slave_dt);
1557 rc = osd_fid_init(env, osd);
1563 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
1567 * see include/lu_object.h for the details.
1569 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
1570 struct lu_fid *fid, struct lu_object *parent,
1571 const struct lu_name *name)
1573 struct osd_device *osd = osd_dev(d);
1575 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1578 struct lu_device_operations osd_lu_ops = {
1579 .ldo_object_alloc = osd_object_alloc,
1580 .ldo_process_config = osd_process_config,
1581 .ldo_recovery_complete = osd_recovery_complete,
1582 .ldo_prepare = osd_prepare,
1583 .ldo_fid_alloc = osd_fid_alloc,
1586 static void osd_type_start(struct lu_device_type *t)
1590 static void osd_type_stop(struct lu_device_type *t)
1594 static struct lu_device_type_operations osd_device_type_ops = {
1595 .ldto_init = osd_type_init,
1596 .ldto_fini = osd_type_fini,
1598 .ldto_start = osd_type_start,
1599 .ldto_stop = osd_type_stop,
1601 .ldto_device_alloc = osd_device_alloc,
1602 .ldto_device_free = osd_device_free,
1604 .ldto_device_init = osd_device_init,
1605 .ldto_device_fini = osd_device_fini
1608 static struct lu_device_type osd_device_type = {
1609 .ldt_tags = LU_DEVICE_DT,
1610 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1611 .ldt_ops = &osd_device_type_ops,
1612 .ldt_ctx_tags = LCT_LOCAL
1616 static const struct obd_ops osd_obd_device_ops = {
1617 .o_owner = THIS_MODULE,
1618 .o_connect = osd_obd_connect,
1619 .o_disconnect = osd_obd_disconnect,
1622 static int __init osd_init(void)
1626 rc = osd_options_init();
1630 rc = lu_kmem_init(osd_caches);
1634 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1635 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1637 lu_kmem_fini(osd_caches);
1641 static void __exit osd_exit(void)
1643 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1644 lu_kmem_fini(osd_caches);
1647 module_param(osd_oi_count, int, 0444);
1648 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1650 module_param(osd_txg_sync_delay_us, int, 0644);
1651 MODULE_PARM_DESC(osd_txg_sync_delay_us,
1652 "When zero or larger delay N usec instead of doing TXG sync");
1654 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1655 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1656 MODULE_VERSION(LUSTRE_VERSION_STRING);
1657 MODULE_LICENSE("GPL");
1659 module_init(osd_init);
1660 module_exit(osd_exit);