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 /* call per-transaction callbacks if any */
151 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) {
152 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
153 "commit callback entry: magic=%x name='%s'\n",
154 dcb->dcb_magic, dcb->dcb_name);
155 list_del_init(&dcb->dcb_linkage);
156 dcb->dcb_func(NULL, th, dcb, error);
159 /* Unlike ldiskfs, zfs updates space accounting at commit time.
160 * As a consequence, op_end is called only now to inform the quota slave
161 * component that reserved quota space is now accounted in usage and
162 * should be released. Quota space won't be adjusted at this point since
163 * we can't provide a suitable environment. It will be performed
164 * asynchronously by a lquota thread. */
165 qsd_op_end(NULL, osd->od_quota_slave_dt, &oh->ot_quota_trans);
166 if (osd->od_quota_slave_md != NULL)
167 qsd_op_end(NULL, osd->od_quota_slave_md, &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_device *osd = osd_dt_dev(d);
198 struct osd_thandle *oh;
203 oh = container_of0(th, struct osd_thandle, ot_super);
207 rc = dt_txn_hook_start(env, d, th);
209 CERROR("%s: dt_txn_hook_start failed: rc = %d\n",
214 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
215 /* Unlike ldiskfs, ZFS checks for available space and returns
216 * -ENOSPC when assigning txg */
219 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
220 if (unlikely(rc != 0)) {
221 /* dmu will call commit callback with error code during abort */
222 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
223 CERROR("%s: failed to start transaction due to ENOSPC"
224 "\n", osd->od_svname);
226 CERROR("%s: can't assign tx: rc = %d\n",
229 /* add commit callback */
230 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
232 osd_oti_get(env)->oti_in_trans = 1;
233 lu_device_get(&d->dd_lu_dev);
239 static void osd_unlinked_list_emptify(const struct lu_env *env,
240 struct osd_device *osd,
241 struct list_head *list, bool free)
243 struct osd_object *obj;
246 while (!list_empty(list)) {
247 obj = list_entry(list->next,
248 struct osd_object, oo_unlinked_linkage);
249 LASSERT(obj->oo_dn != NULL);
250 oid = obj->oo_dn->dn_object;
252 list_del_init(&obj->oo_unlinked_linkage);
254 (void)osd_unlinked_object_free(env, osd, oid);
258 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
260 struct dt_txn_commit_cb *dcb;
261 struct dt_txn_commit_cb *tmp;
263 /* call per-transaction stop callbacks if any */
264 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
266 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
267 "commit callback entry: magic=%x name='%s'\n",
268 dcb->dcb_magic, dcb->dcb_name);
269 list_del_init(&dcb->dcb_linkage);
270 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
275 * Concurrency: shouldn't matter.
277 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
280 struct osd_device *osd = osd_dt_dev(th->th_dev);
281 bool sync = (th->th_sync != 0);
282 struct osd_thandle *oh;
288 oh = container_of0(th, struct osd_thandle, ot_super);
289 list_splice_init(&oh->ot_unlinked_list, &unlinked);
291 osd_oti_get(env)->oti_ins_cache_depth--;
292 /* reset OI cache for safety */
293 if (osd_oti_get(env)->oti_ins_cache_depth == 0)
294 osd_oti_get(env)->oti_ins_cache_used = 0;
296 if (oh->ot_assigned == 0) {
298 CDEBUG(D_OTHER, "%s: transaction is aborted\n", osd->od_svname);
299 osd_trans_stop_cb(oh, th->th_result);
300 dmu_tx_abort(oh->ot_tx);
301 osd_object_sa_dirty_rele(env, oh);
302 osd_unlinked_list_emptify(env, osd, &unlinked, false);
303 /* there won't be any commit, release reserved quota space now,
305 qsd_op_end(env, osd->od_quota_slave_dt, &oh->ot_quota_trans);
306 if (osd->od_quota_slave_md != NULL)
307 qsd_op_end(env, osd->od_quota_slave_md,
308 &oh->ot_quota_trans);
313 rc = dt_txn_hook_stop(env, th);
315 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
318 osd_trans_stop_cb(oh, rc);
321 txg = oh->ot_tx->tx_txg;
323 osd_object_sa_dirty_rele(env, oh);
324 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
325 * by osd_trans_commit_cb already. */
326 dmu_tx_commit(oh->ot_tx);
327 osd_oti_get(env)->oti_in_trans = 0;
329 osd_unlinked_list_emptify(env, osd, &unlinked, true);
332 if (osd_txg_sync_delay_us < 0)
333 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
335 udelay(osd_txg_sync_delay_us);
341 static struct thandle *osd_trans_create(const struct lu_env *env,
342 struct dt_device *dt)
344 struct osd_device *osd = osd_dt_dev(dt);
345 struct osd_thandle *oh;
351 CERROR("%s: someone try to start transaction under "
352 "readonly mode, should be disabled.\n",
353 osd_name(osd_dt_dev(dt)));
355 RETURN(ERR_PTR(-EROFS));
358 tx = dmu_tx_create(osd->od_os);
360 RETURN(ERR_PTR(-ENOMEM));
362 /* alloc callback data */
366 RETURN(ERR_PTR(-ENOMEM));
370 INIT_LIST_HEAD(&oh->ot_dcb_list);
371 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
372 INIT_LIST_HEAD(&oh->ot_unlinked_list);
373 INIT_LIST_HEAD(&oh->ot_sa_list);
374 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
379 osd_oti_get(env)->oti_ins_cache_depth++;
384 /* Estimate the total number of objects from a number of blocks */
385 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
386 uint64_t nrblocks, uint64_t est_maxblockshift)
388 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
391 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
392 * bump the free dnode estimate to assume blocks at least 64KB in
393 * case of a directory-heavy MDT (at 32KB/directory).
395 if (est_maxblockshift < 16) {
396 nrblocks >>= (16 - est_maxblockshift);
397 est_maxblockshift = 16;
401 * Estimate the total number of dnodes from the total blocks count
402 * and the space used per dnode. Since we don't know the overhead
403 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
404 * just using DNODE_SHIFT isn't going to give a good estimate.
405 * Instead, compute the current average space usage per dnode, with
406 * an upper and lower cap to avoid unrealistic estimates..
408 * In case there aren't many dnodes or blocks used yet, add a small
409 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
410 * gradually disappears as the number of real dnodes grows. It also
411 * avoids the need to check for divide-by-zero computing dn_per_block.
413 BUILD_BUG_ON(OSD_DNODE_MIN_BLKSHIFT <= 0);
414 BUILD_BUG_ON(OSD_DNODE_EST_BLKSHIFT <= 0);
416 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
417 usedbytes) >> est_maxblockshift;
418 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
420 if (est_usedobjs <= est_usedblocks) {
422 * Average space/dnode more than maximum block size, use max
423 * block size to estimate free dnodes from adjusted free blocks
424 * count. OSTs typically use multiple blocks per dnode so this
427 est_totobjs = nrblocks;
429 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
431 * Average space/dnode smaller than min dnode size (probably
432 * due to metadnode compression), use min dnode size to
433 * estimate object count. MDTs may use only one block per node
434 * so this case applies.
436 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
440 * Between the extremes, use average space per existing dnode
441 * to compute the number of dnodes that will fit into nrblocks:
443 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
445 * this may overflow 64 bits or become 0 if not handled well.
447 * We know nrblocks is below 2^(64 - blkbits) bits, and
448 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
449 * which means that multiplying them may get as large as
450 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
452 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
453 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
454 * the limit checks above, so we can safely compute this first.
455 * We care more about accuracy on the MDT (many dnodes/block)
456 * which is good because this is where truncation errors are
457 * smallest. Since both nrblocks and dn_per_block are a
458 * function of blkbits, their product is at most:
460 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
462 * so we can safely use 7 bits to compute a fixed-point
463 * fraction and est_totobjs can still fit in 64 bits.
465 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
467 est_totobjs = (nrblocks * dn_per_block) >> 7;
472 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
474 struct objset *os = osd->od_os;
475 uint64_t usedbytes, availbytes, usedobjs, availobjs;
476 uint64_t est_availobjs;
480 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
482 memset(osfs, 0, sizeof(*osfs));
484 /* We're a zfs filesystem. */
485 osfs->os_type = UBERBLOCK_MAGIC;
488 * ZFS allows multiple block sizes. For statfs, Linux makes no
489 * proper distinction between bsize and frsize. For calculations
490 * of free and used blocks incorrectly uses bsize instead of frsize,
491 * but bsize is also used as the optimal blocksize. We return the
492 * largest possible block size as IO size for the optimum performance
493 * and scale the free and used blocks count appropriately.
495 osfs->os_bsize = osd->od_max_blksz;
496 bshift = fls64(osfs->os_bsize) - 1;
498 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
499 osfs->os_bfree = availbytes >> bshift;
500 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
502 /* Take replication (i.e. number of copies) into account */
503 if (os->os_copies != 0)
504 osfs->os_bavail /= os->os_copies;
507 * Reserve some space so we don't run into ENOSPC due to grants not
508 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
509 * Rather than report this via os_bavail (which makes users unhappy if
510 * they can't fill the filesystem 100%), reduce os_blocks as well.
512 * Reserve 0.78% of total space, at least 16MB for small filesystems,
513 * for internal files to be created/unlinked when space is tight.
515 BUILD_BUG_ON(OSD_STATFS_RESERVED_SIZE <= 0);
516 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
517 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
518 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
520 osfs->os_blocks -= reserved;
521 osfs->os_bfree -= min(reserved, osfs->os_bfree);
522 osfs->os_bavail -= min(reserved, osfs->os_bavail);
525 * The availobjs value returned from dmu_objset_space() is largely
526 * useless, since it reports the number of objects that might
527 * theoretically still fit into the dataset, independent of minor
528 * issues like how much space is actually available in the pool.
529 * Compute a better estimate in udmu_objs_count_estimate().
531 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
532 osfs->os_bfree, bshift);
534 osfs->os_ffree = min(availobjs, est_availobjs);
535 osfs->os_files = osfs->os_ffree + usedobjs;
537 /* ZFS XXX: fill in backing dataset FSID/UUID
538 memcpy(osfs->os_fsid, .... );*/
540 osfs->os_namelen = MAXNAMELEN;
541 osfs->os_maxbytes = OBD_OBJECT_EOF;
543 if (!spa_writeable(dmu_objset_spa(os)) ||
544 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
545 osfs->os_state |= OS_STATE_READONLY;
551 * Concurrency: shouldn't matter.
553 int osd_statfs(const struct lu_env *env, struct dt_device *d,
554 struct obd_statfs *osfs, struct obd_statfs_info *info)
556 struct osd_device *osd = osd_dt_dev(d);
560 rc = osd_objset_statfs(osd, osfs);
561 if (unlikely(rc != 0))
564 osfs->os_bavail -= min_t(u64,
565 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
568 /* ZFS does not support reporting nonrotional status yet, so return
569 * flag only if user has set nonrotational.
571 osfs->os_state |= osd->od_nonrotational ? OS_STATE_NONROT : 0;
576 static int osd_blk_insert_cost(struct osd_device *osd)
578 int max_blockshift, nr_blkptrshift, bshift;
580 /* max_blockshift is the log2 of the number of blocks needed to reach
581 * the maximum filesize (that's to say 2^64) */
582 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
583 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
585 /* nr_blkptrshift is the log2 of the number of block pointers that can
586 * be stored in an indirect block */
587 BUILD_BUG_ON(DN_MAX_INDBLKSHIFT <= SPA_BLKPTRSHIFT);
588 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
590 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
591 * tree. We add +1 for rounding purpose.
592 * The tree depth times the indirect block size gives us the maximum
593 * cost of inserting a block in the tree */
594 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
598 * Concurrency: doesn't access mutable data.
600 static void osd_conf_get(const struct lu_env *env,
601 const struct dt_device *dev,
602 struct dt_device_param *param)
604 struct osd_device *osd = osd_dt_dev(dev);
607 * XXX should be taken from not-yet-existing fs abstraction layer.
609 param->ddp_max_name_len = MAXNAMELEN;
610 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
611 param->ddp_symlink_max = PATH_MAX;
612 param->ddp_mount_type = LDD_MT_ZFS;
614 param->ddp_mntopts = MNTOPT_USERXATTR;
615 if (osd->od_posix_acl)
616 param->ddp_mntopts |= MNTOPT_ACL;
617 /* Previously DXATTR_MAX_ENTRY_SIZE */
618 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
620 /* for maxbytes, report same value as ZPL */
621 param->ddp_maxbytes = MAX_LFS_FILESIZE;
623 /* inodes are dynamically allocated, so we report the per-inode space
624 * consumption to upper layers. This static value is not really accurate
625 * and we should use the same logic as in udmu_objset_statfs() to
626 * estimate the real size consumed by an object */
627 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
628 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
629 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
630 * not be accounted for every single new block insertion.
631 * Instead, the maximum extent size is set to the number of blocks that
632 * can fit into a single contiguous indirect block. There would be some
633 * cases where this crosses indirect blocks, but it also won't have 7
634 * new levels of indirect blocks in that case either, so it will still
635 * have enough reserved space for the extra indirect block */
636 param->ddp_max_extent_blks =
637 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
638 param->ddp_extent_tax = osd_blk_insert_cost(osd);
640 /* Preferred RPC size for efficient disk IO. 1MB shows good
641 * all-around performance for ZFS, but use blocksize (recordsize)
642 * by default if larger to avoid read-modify-write. */
643 if (osd->od_max_blksz > ONE_MB_BRW_SIZE)
644 param->ddp_brw_size = osd->od_max_blksz;
646 param->ddp_brw_size = ONE_MB_BRW_SIZE;
650 * Concurrency: shouldn't matter.
652 static int osd_sync(const struct lu_env *env, struct dt_device *d)
655 struct osd_device *osd = osd_dt_dev(d);
657 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
658 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
659 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
665 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
667 struct osd_device *osd = osd_dt_dev(dev);
668 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
671 mutex_enter(&tx->tx_sync_lock);
672 txg = tx->tx_open_txg + 1;
673 if (tx->tx_quiesce_txg_waiting < txg) {
674 tx->tx_quiesce_txg_waiting = txg;
675 cv_broadcast(&tx->tx_quiesce_more_cv);
677 mutex_exit(&tx->tx_sync_lock);
683 * Concurrency: shouldn't matter.
685 static int osd_ro(const struct lu_env *env, struct dt_device *d)
687 struct osd_device *osd = osd_dt_dev(d);
690 CERROR("%s: *** setting device %s read-only ***\n",
691 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
692 osd->od_dev_set_rdonly = 1;
693 spa_freeze(dmu_objset_spa(osd->od_os));
698 static struct dt_device_operations osd_dt_ops = {
699 .dt_root_get = osd_root_get,
700 .dt_statfs = osd_statfs,
701 .dt_trans_create = osd_trans_create,
702 .dt_trans_start = osd_trans_start,
703 .dt_trans_stop = osd_trans_stop,
704 .dt_trans_cb_add = osd_trans_cb_add,
705 .dt_conf_get = osd_conf_get,
707 .dt_commit_async = osd_commit_async,
712 * DMU OSD device type methods
714 static int osd_type_init(struct lu_device_type *t)
716 LU_CONTEXT_KEY_INIT(&osd_key);
717 return lu_context_key_register(&osd_key);
720 static void osd_type_fini(struct lu_device_type *t)
722 lu_context_key_degister(&osd_key);
725 static void *osd_key_init(const struct lu_context *ctx,
726 struct lu_context_key *key)
728 struct osd_thread_info *info;
732 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
734 info = ERR_PTR(-ENOMEM);
738 static void osd_key_fini(const struct lu_context *ctx,
739 struct lu_context_key *key, void *data)
741 struct osd_thread_info *info = data;
742 struct osd_idmap_cache *idc = info->oti_ins_cache;
745 LASSERT(info->oti_ins_cache_size > 0);
746 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
747 info->oti_ins_cache = NULL;
748 info->oti_ins_cache_size = 0;
750 lu_buf_free(&info->oti_xattr_lbuf);
754 static void osd_key_exit(const struct lu_context *ctx,
755 struct lu_context_key *key, void *data)
759 struct lu_context_key osd_key = {
760 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
761 .lct_init = osd_key_init,
762 .lct_fini = osd_key_fini,
763 .lct_exit = osd_key_exit
766 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
768 if (osd->od_cl_seq == NULL)
771 seq_client_fini(osd->od_cl_seq);
772 OBD_FREE_PTR(osd->od_cl_seq);
773 osd->od_cl_seq = NULL;
776 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
780 /* shutdown quota slave instance associated with the device */
781 if (o->od_quota_slave_md != NULL) {
782 /* complete all in-flight callbacks */
783 osd_sync(env, &o->od_dt_dev);
784 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
785 qsd_fini(env, o->od_quota_slave_md);
786 o->od_quota_slave_md = NULL;
789 if (o->od_quota_slave_dt != NULL) {
790 /* complete all in-flight callbacks */
791 osd_sync(env, &o->od_dt_dev);
792 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
793 qsd_fini(env, o->od_quota_slave_dt);
794 o->od_quota_slave_dt = NULL;
796 osd_fid_fini(env, o);
801 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
803 struct osd_device *osd = arg;
805 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
808 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
810 struct osd_device *osd = arg;
812 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
813 LASSERT(newval >= SPA_MINBLOCKSIZE);
814 LASSERT(ISP2(newval));
816 osd->od_max_blksz = newval;
819 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
821 struct osd_device *osd = arg;
823 osd->od_prop_rdonly = !!newval;
826 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
827 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
829 struct osd_device *osd = arg;
831 osd->od_dnsize = newval;
835 * This function unregisters all registered callbacks. It's harmless to
836 * unregister callbacks that were never registered so it is used to safely
837 * unwind a partially completed call to osd_objset_register_callbacks().
839 static void osd_objset_unregister_callbacks(struct osd_device *o)
841 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
843 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
844 osd_xattr_changed_cb, o);
845 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
846 osd_recordsize_changed_cb, o);
847 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
848 osd_readonly_changed_cb, o);
849 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
850 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
851 osd_dnodesize_changed_cb, o);
854 if (o->arc_prune_cb != NULL) {
855 arc_remove_prune_callback(o->arc_prune_cb);
856 o->arc_prune_cb = NULL;
861 * Register the required callbacks to be notified when zfs properties
862 * are modified using the 'zfs(8)' command line utility.
864 static int osd_objset_register_callbacks(struct osd_device *o)
866 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
867 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
873 dsl_pool_config_enter(dp, FTAG);
874 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
875 osd_xattr_changed_cb, o);
879 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
880 osd_recordsize_changed_cb, o);
884 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
885 osd_readonly_changed_cb, o);
889 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
890 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
891 osd_dnodesize_changed_cb, o);
896 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
898 dsl_pool_config_exit(dp, FTAG);
900 osd_objset_unregister_callbacks(o);
905 static int osd_objset_open(struct osd_device *o)
907 uint64_t version = ZPL_VERSION;
908 uint64_t sa_obj, unlink_obj;
912 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
913 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
914 B_TRUE, o, &o->od_os);
917 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
923 /* Check ZFS version */
924 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
925 ZPL_VERSION_STR, 8, 1, &version);
927 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
929 * We can't return ENOENT because that would mean the objset
932 GOTO(out, rc = -EIO);
935 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
936 ZFS_SA_ATTRS, 8, 1, &sa_obj);
940 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
941 ZPL_END, &o->z_attr_table);
945 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
946 8, 1, &o->od_rootid);
948 CERROR("%s: lookup for root failed: rc = %d\n",
953 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
956 CERROR("%s: lookup for %s failed: rc = %d\n",
957 o->od_svname, ZFS_UNLINKED_SET, rc);
961 /* Check that user/group usage tracking is supported */
962 if (!dmu_objset_userused_enabled(o->od_os) ||
963 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
964 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
965 CERROR("%s: Space accounting not supported by this target, "
966 "aborting\n", o->od_svname);
967 GOTO(out, rc = -ENOTSUPP);
970 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
972 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
978 if (rc != 0 && o->od_os != NULL) {
979 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
986 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
989 char *key = osd_oti_get(env)->oti_str;
993 if (osd->od_dt_dev.dd_rdonly) {
994 CERROR("%s: someone try to free objects under "
995 "readonly mode, should be disabled.\n", osd_name(osd));
1001 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
1003 CWARN("%s: Cannot truncate %llu: rc = %d\n",
1004 osd->od_svname, oid, rc);
1008 tx = dmu_tx_create(osd->od_os);
1009 dmu_tx_mark_netfree(tx);
1010 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
1011 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
1013 rc = -dmu_tx_assign(tx, TXG_WAIT);
1015 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
1016 osd->od_svname, oid, rc);
1020 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
1021 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
1022 osd->od_unlinked, key, tx);
1024 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
1025 osd->od_svname, oid, rc);
1029 rc = -dmu_object_free(osd->od_os, oid, tx);
1031 CWARN("%s: Cannot free %llu: rc = %d\n",
1032 osd->od_svname, oid, rc);
1047 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
1050 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1052 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1054 while (zap_cursor_retrieve(&zc, za) == 0) {
1055 /* If cannot free the object, leave it in the unlinked set,
1056 * until the OSD is mounted again when obd_unlinked_drain()
1057 * will be called. */
1058 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1060 zap_cursor_advance(&zc);
1063 zap_cursor_fini(&zc);
1066 static int osd_mount(const struct lu_env *env,
1067 struct osd_device *o, struct lustre_cfg *cfg)
1069 char *mntdev = lustre_cfg_string(cfg, 1);
1070 char *str = lustre_cfg_string(cfg, 2);
1071 char *svname = lustre_cfg_string(cfg, 4);
1077 if (o->od_os != NULL)
1080 if (mntdev == NULL || svname == NULL)
1083 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1084 if (rc >= sizeof(o->od_mntdev))
1087 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1088 if (rc >= sizeof(o->od_svname))
1091 o->od_index_backup_stop = 0;
1092 o->od_index = -1; /* -1 means index is invalid */
1093 rc = server_name2index(o->od_svname, &o->od_index, NULL);
1094 str = strstr(str, ":");
1096 unsigned long flags;
1098 rc = kstrtoul(str + 1, 10, &flags);
1102 if (flags & LMD_FLG_DEV_RDONLY) {
1103 o->od_dt_dev.dd_rdonly = 1;
1104 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1108 if (flags & LMD_FLG_NOSCRUB)
1109 o->od_auto_scrub_interval = AS_NEVER;
1112 if (server_name_is_ost(o->od_svname))
1115 rc = osd_objset_open(o);
1119 o->od_xattr_in_sa = B_TRUE;
1120 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1121 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
1123 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1126 o->od_root = rootdn->dn_object;
1127 osd_dnode_rele(rootdn);
1129 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1130 &o->od_userused_dn);
1134 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1135 &o->od_groupused_dn);
1139 #ifdef ZFS_PROJINHERIT
1140 if (dmu_objset_projectquota_enabled(o->od_os)) {
1141 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1142 &o->od_projectused_dn);
1143 if (rc && rc != -ENOENT)
1148 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1151 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1153 rc = lu_site_init_finish(&o->od_site);
1157 rc = osd_objset_register_callbacks(o);
1162 rc = osd_scrub_setup(env, o);
1167 rc = osd_procfs_init(o, o->od_svname);
1171 /* currently it's no need to prepare qsd_instance_md for OST */
1172 if (!o->od_is_ost) {
1173 o->od_quota_slave_md = qsd_init(env, o->od_svname,
1175 o->od_proc_entry, true);
1176 if (IS_ERR(o->od_quota_slave_md)) {
1177 rc = PTR_ERR(o->od_quota_slave_md);
1178 o->od_quota_slave_md = NULL;
1183 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
1184 o->od_proc_entry, false);
1186 if (IS_ERR(o->od_quota_slave_dt)) {
1187 if (o->od_quota_slave_md != NULL) {
1188 qsd_fini(env, o->od_quota_slave_md);
1189 o->od_quota_slave_md = NULL;
1192 rc = PTR_ERR(o->od_quota_slave_dt);
1193 o->od_quota_slave_dt = NULL;
1197 #ifdef HAVE_DMU_USEROBJ_ACCOUNTING
1198 if (!osd_dmu_userobj_accounting_available(o))
1199 CWARN("%s: dnode accounting not enabled: "
1200 "enable feature@userobj_accounting in pool\n",
1204 /* parse mount option "noacl", and enable ACL by default */
1205 opts = lustre_cfg_string(cfg, 3);
1206 if (opts == NULL || strstr(opts, "noacl") == NULL)
1207 o->od_posix_acl = 1;
1209 osd_unlinked_drain(env, o);
1211 if (rc && o->od_os) {
1212 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1219 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1223 if (atomic_read(&o->od_zerocopy_alloc))
1224 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1225 atomic_read(&o->od_zerocopy_alloc));
1226 if (atomic_read(&o->od_zerocopy_loan))
1227 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1228 atomic_read(&o->od_zerocopy_loan));
1229 if (atomic_read(&o->od_zerocopy_pin))
1230 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1231 atomic_read(&o->od_zerocopy_pin));
1233 if (o->od_unlinked) {
1234 osd_dnode_rele(o->od_unlinked);
1235 o->od_unlinked = NULL;
1237 if (o->od_userused_dn) {
1238 osd_dnode_rele(o->od_userused_dn);
1239 o->od_userused_dn = NULL;
1241 if (o->od_groupused_dn) {
1242 osd_dnode_rele(o->od_groupused_dn);
1243 o->od_groupused_dn = NULL;
1246 #ifdef ZFS_PROJINHERIT
1247 if (o->od_projectused_dn) {
1248 osd_dnode_rele(o->od_projectused_dn);
1249 o->od_projectused_dn = NULL;
1253 if (o->od_os != NULL) {
1254 if (!o->od_dt_dev.dd_rdonly)
1255 /* force a txg sync to get all commit callbacks */
1256 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1258 /* close the object set */
1259 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1266 static int osd_device_init0(const struct lu_env *env,
1267 struct osd_device *o,
1268 struct lustre_cfg *cfg)
1270 struct lu_device *l = osd2lu_dev(o);
1273 /* if the module was re-loaded, env can loose its keys */
1274 rc = lu_env_refill((struct lu_env *) env);
1278 l->ld_ops = &osd_lu_ops;
1279 o->od_dt_dev.dd_ops = &osd_dt_ops;
1280 sema_init(&o->od_otable_sem, 1);
1281 INIT_LIST_HEAD(&o->od_ios_list);
1282 o->od_auto_scrub_interval = AS_DEFAULT;
1284 /* ZFS does not support reporting nonrotional status yet, so this flag
1285 * is only set if explicitly set by the user.
1287 o->od_nonrotational = 0;
1293 static struct lu_device *osd_device_fini(const struct lu_env *env,
1294 struct lu_device *dev);
1296 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1297 struct lu_device_type *type,
1298 struct lustre_cfg *cfg)
1300 struct osd_device *dev;
1301 struct osd_seq_list *osl;
1306 return ERR_PTR(-ENOMEM);
1308 osl = &dev->od_seq_list;
1309 INIT_LIST_HEAD(&osl->osl_seq_list);
1310 rwlock_init(&osl->osl_seq_list_lock);
1311 sema_init(&osl->osl_seq_init_sem, 1);
1312 INIT_LIST_HEAD(&dev->od_index_backup_list);
1313 INIT_LIST_HEAD(&dev->od_index_restore_list);
1314 spin_lock_init(&dev->od_lock);
1315 dev->od_index_backup_policy = LIBP_NONE;
1317 rc = dt_device_init(&dev->od_dt_dev, type);
1319 rc = osd_device_init0(env, dev, cfg);
1321 rc = osd_mount(env, dev, cfg);
1323 osd_device_fini(env, osd2lu_dev(dev));
1326 dt_device_fini(&dev->od_dt_dev);
1329 if (unlikely(rc != 0))
1332 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1335 static struct lu_device *osd_device_free(const struct lu_env *env,
1336 struct lu_device *d)
1338 struct osd_device *o = osd_dev(d);
1341 /* XXX: make osd top device in order to release reference */
1342 d->ld_site->ls_top_dev = d;
1343 lu_site_purge(env, d->ld_site, -1);
1344 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1345 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1346 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1348 lu_site_fini(&o->od_site);
1349 dt_device_fini(&o->od_dt_dev);
1355 static struct lu_device *osd_device_fini(const struct lu_env *env,
1356 struct lu_device *d)
1358 struct osd_device *o = osd_dev(d);
1364 osd_objset_unregister_callbacks(o);
1365 if (!o->od_dt_dev.dd_rdonly) {
1366 osd_sync(env, lu2dt_dev(d));
1368 spa_get_dsl(dmu_objset_spa(o->od_os)));
1372 /* now with all the callbacks completed we can cleanup the remainings */
1373 osd_shutdown(env, o);
1374 osd_scrub_cleanup(env, o);
1376 rc = osd_procfs_fini(o);
1378 CERROR("proc fini error %d\n", rc);
1379 RETURN(ERR_PTR(rc));
1388 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1389 const char *name, struct lu_device *next)
1395 * To be removed, setup is performed by osd_device_{init,alloc} and
1396 * cleanup is performed by osd_device_{fini,free).
1398 static int osd_process_config(const struct lu_env *env,
1399 struct lu_device *d, struct lustre_cfg *cfg)
1401 struct osd_device *o = osd_dev(d);
1406 switch(cfg->lcfg_command) {
1408 rc = osd_mount(env, o, cfg);
1411 /* For the case LCFG_PRE_CLEANUP is not called in advance,
1412 * that may happend if hit failure during mount process. */
1413 osd_index_backup(env, o, false);
1414 rc = osd_shutdown(env, o);
1417 LASSERT(&o->od_dt_dev);
1418 count = class_modify_config(cfg, PARAM_OSD,
1419 &o->od_dt_dev.dd_kobj);
1421 count = class_modify_config(cfg, PARAM_OST,
1422 &o->od_dt_dev.dd_kobj);
1423 rc = count > 0 ? 0 : count;
1426 case LCFG_PRE_CLEANUP:
1428 osd_index_backup(env, o,
1429 o->od_index_backup_policy != LIBP_NONE);
1439 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1441 struct osd_device *osd = osd_dev(d);
1445 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
1448 /* start qsd instance on recovery completion, this notifies the quota
1449 * slave code that we are about to process new requests now */
1450 rc = qsd_start(env, osd->od_quota_slave_dt);
1451 if (rc == 0 && osd->od_quota_slave_md != NULL)
1452 rc = qsd_start(env, osd->od_quota_slave_md);
1457 * we use exports to track all osd users
1459 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1460 struct obd_device *obd, struct obd_uuid *cluuid,
1461 struct obd_connect_data *data, void *localdata)
1463 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1464 struct lustre_handle conn;
1468 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1470 rc = class_connect(&conn, obd, cluuid);
1474 *exp = class_conn2export(&conn);
1476 spin_lock(&obd->obd_dev_lock);
1478 spin_unlock(&obd->obd_dev_lock);
1484 * once last export (we don't count self-export) disappeared
1485 * osd can be released
1487 static int osd_obd_disconnect(struct obd_export *exp)
1489 struct obd_device *obd = exp->exp_obd;
1490 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1491 int rc, release = 0;
1494 /* Only disconnect the underlying layers on the final disconnect. */
1495 spin_lock(&obd->obd_dev_lock);
1497 if (osd->od_connects == 0)
1499 spin_unlock(&obd->obd_dev_lock);
1501 rc = class_disconnect(exp); /* bz 9811 */
1503 if (rc == 0 && release)
1504 class_manual_cleanup(obd);
1508 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1510 struct seq_server_site *ss = osd_seq_site(osd);
1514 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1517 if (unlikely(ss == NULL))
1520 OBD_ALLOC_PTR(osd->od_cl_seq);
1521 if (osd->od_cl_seq == NULL)
1524 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1525 osd->od_svname, ss->ss_server_seq);
1528 OBD_FREE_PTR(osd->od_cl_seq);
1529 osd->od_cl_seq = NULL;
1532 if (ss->ss_node_id == 0) {
1534 * If the OSD on the sequence controller(MDT0), then allocate
1535 * sequence here, otherwise allocate sequence after connected
1536 * to MDT0 (see mdt_register_lwp_callback()).
1538 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
1539 &osd->od_cl_seq->lcs_space, env);
1545 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1546 struct lu_device *dev)
1548 struct osd_device *osd = osd_dev(dev);
1552 if (osd->od_quota_slave_md != NULL) {
1553 /* set up quota slave objects */
1554 rc = qsd_prepare(env, osd->od_quota_slave_md);
1559 if (osd->od_quota_slave_dt != NULL) {
1560 /* set up quota slave objects */
1561 rc = qsd_prepare(env, osd->od_quota_slave_dt);
1566 rc = osd_fid_init(env, osd);
1571 struct lu_device_operations osd_lu_ops = {
1572 .ldo_object_alloc = osd_object_alloc,
1573 .ldo_process_config = osd_process_config,
1574 .ldo_recovery_complete = osd_recovery_complete,
1575 .ldo_prepare = osd_prepare,
1578 static void osd_type_start(struct lu_device_type *t)
1582 static void osd_type_stop(struct lu_device_type *t)
1586 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1587 struct lu_fid *fid, struct md_op_data *op_data)
1589 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1591 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
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,
1620 .o_fid_alloc = osd_fid_alloc
1623 static int __init osd_init(void)
1627 rc = osd_options_init();
1631 rc = lu_kmem_init(osd_caches);
1635 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1636 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1638 lu_kmem_fini(osd_caches);
1642 static void __exit osd_exit(void)
1644 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1645 lu_kmem_fini(osd_caches);
1648 module_param(osd_oi_count, int, 0444);
1649 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1651 module_param(osd_txg_sync_delay_us, int, 0644);
1652 MODULE_PARM_DESC(osd_txg_sync_delay_us,
1653 "When zero or larger delay N usec instead of doing TXG sync");
1655 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1656 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1657 MODULE_VERSION(LUSTRE_VERSION_STRING);
1658 MODULE_LICENSE("GPL");
1660 module_init(osd_init);
1661 module_exit(osd_exit);