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;
283 struct list_head unlinked;
288 oh = container_of0(th, struct osd_thandle, ot_super);
289 INIT_LIST_HEAD(&unlinked);
290 list_splice_init(&oh->ot_unlinked_list, &unlinked);
292 osd_oti_get(env)->oti_ins_cache_depth--;
293 /* reset OI cache for safety */
294 if (osd_oti_get(env)->oti_ins_cache_depth == 0)
295 osd_oti_get(env)->oti_ins_cache_used = 0;
297 if (oh->ot_assigned == 0) {
299 CDEBUG(D_OTHER, "%s: transaction is aborted\n", osd->od_svname);
300 osd_trans_stop_cb(oh, th->th_result);
301 dmu_tx_abort(oh->ot_tx);
302 osd_object_sa_dirty_rele(env, oh);
303 osd_unlinked_list_emptify(env, osd, &unlinked, false);
304 /* there won't be any commit, release reserved quota space now,
306 qsd_op_end(env, osd->od_quota_slave_dt, &oh->ot_quota_trans);
307 if (osd->od_quota_slave_md != NULL)
308 qsd_op_end(env, osd->od_quota_slave_md,
309 &oh->ot_quota_trans);
314 rc = dt_txn_hook_stop(env, th);
316 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
319 osd_trans_stop_cb(oh, rc);
322 txg = oh->ot_tx->tx_txg;
324 osd_object_sa_dirty_rele(env, oh);
325 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
326 * by osd_trans_commit_cb already. */
327 dmu_tx_commit(oh->ot_tx);
328 osd_oti_get(env)->oti_in_trans = 0;
330 osd_unlinked_list_emptify(env, osd, &unlinked, true);
333 if (osd_txg_sync_delay_us < 0)
334 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
336 udelay(osd_txg_sync_delay_us);
342 static struct thandle *osd_trans_create(const struct lu_env *env,
343 struct dt_device *dt)
345 struct osd_device *osd = osd_dt_dev(dt);
346 struct osd_thandle *oh;
352 CERROR("%s: someone try to start transaction under "
353 "readonly mode, should be disabled.\n",
354 osd_name(osd_dt_dev(dt)));
356 RETURN(ERR_PTR(-EROFS));
359 tx = dmu_tx_create(osd->od_os);
361 RETURN(ERR_PTR(-ENOMEM));
363 /* alloc callback data */
367 RETURN(ERR_PTR(-ENOMEM));
371 INIT_LIST_HEAD(&oh->ot_dcb_list);
372 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
373 INIT_LIST_HEAD(&oh->ot_unlinked_list);
374 INIT_LIST_HEAD(&oh->ot_sa_list);
375 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
380 osd_oti_get(env)->oti_ins_cache_depth++;
385 /* Estimate the total number of objects from a number of blocks */
386 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
387 uint64_t nrblocks, uint64_t est_maxblockshift)
389 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
392 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
393 * bump the free dnode estimate to assume blocks at least 64KB in
394 * case of a directory-heavy MDT (at 32KB/directory).
396 if (est_maxblockshift < 16) {
397 nrblocks >>= (16 - est_maxblockshift);
398 est_maxblockshift = 16;
402 * Estimate the total number of dnodes from the total blocks count
403 * and the space used per dnode. Since we don't know the overhead
404 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
405 * just using DNODE_SHIFT isn't going to give a good estimate.
406 * Instead, compute the current average space usage per dnode, with
407 * an upper and lower cap to avoid unrealistic estimates..
409 * In case there aren't many dnodes or blocks used yet, add a small
410 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
411 * gradually disappears as the number of real dnodes grows. It also
412 * avoids the need to check for divide-by-zero computing dn_per_block.
414 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
415 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
417 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
418 usedbytes) >> est_maxblockshift;
419 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
421 if (est_usedobjs <= est_usedblocks) {
423 * Average space/dnode more than maximum block size, use max
424 * block size to estimate free dnodes from adjusted free blocks
425 * count. OSTs typically use multiple blocks per dnode so this
428 est_totobjs = nrblocks;
430 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
432 * Average space/dnode smaller than min dnode size (probably
433 * due to metadnode compression), use min dnode size to
434 * estimate object count. MDTs may use only one block per node
435 * so this case applies.
437 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
441 * Between the extremes, use average space per existing dnode
442 * to compute the number of dnodes that will fit into nrblocks:
444 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
446 * this may overflow 64 bits or become 0 if not handled well.
448 * We know nrblocks is below 2^(64 - blkbits) bits, and
449 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
450 * which means that multiplying them may get as large as
451 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
453 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
454 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
455 * the limit checks above, so we can safely compute this first.
456 * We care more about accuracy on the MDT (many dnodes/block)
457 * which is good because this is where truncation errors are
458 * smallest. Since both nrblocks and dn_per_block are a
459 * function of blkbits, their product is at most:
461 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
463 * so we can safely use 7 bits to compute a fixed-point
464 * fraction and est_totobjs can still fit in 64 bits.
466 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
468 est_totobjs = (nrblocks * dn_per_block) >> 7;
473 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
475 struct objset *os = osd->od_os;
476 uint64_t usedbytes, availbytes, usedobjs, availobjs;
477 uint64_t est_availobjs;
481 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
483 memset(osfs, 0, sizeof(*osfs));
485 /* We're a zfs filesystem. */
486 osfs->os_type = UBERBLOCK_MAGIC;
489 * ZFS allows multiple block sizes. For statfs, Linux makes no
490 * proper distinction between bsize and frsize. For calculations
491 * of free and used blocks incorrectly uses bsize instead of frsize,
492 * but bsize is also used as the optimal blocksize. We return the
493 * largest possible block size as IO size for the optimum performance
494 * and scale the free and used blocks count appropriately.
496 osfs->os_bsize = osd->od_max_blksz;
497 bshift = fls64(osfs->os_bsize) - 1;
499 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
500 osfs->os_bfree = availbytes >> bshift;
501 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
503 /* Take replication (i.e. number of copies) into account */
504 if (os->os_copies != 0)
505 osfs->os_bavail /= os->os_copies;
508 * Reserve some space so we don't run into ENOSPC due to grants not
509 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
510 * Rather than report this via os_bavail (which makes users unhappy if
511 * they can't fill the filesystem 100%), reduce os_blocks as well.
513 * Reserve 0.78% of total space, at least 16MB for small filesystems,
514 * for internal files to be created/unlinked when space is tight.
516 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
517 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
518 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
519 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
521 osfs->os_blocks -= reserved;
522 osfs->os_bfree -= min(reserved, osfs->os_bfree);
523 osfs->os_bavail -= min(reserved, osfs->os_bavail);
526 * The availobjs value returned from dmu_objset_space() is largely
527 * useless, since it reports the number of objects that might
528 * theoretically still fit into the dataset, independent of minor
529 * issues like how much space is actually available in the pool.
530 * Compute a better estimate in udmu_objs_count_estimate().
532 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
533 osfs->os_bfree, bshift);
535 osfs->os_ffree = min(availobjs, est_availobjs);
536 osfs->os_files = osfs->os_ffree + usedobjs;
538 /* ZFS XXX: fill in backing dataset FSID/UUID
539 memcpy(osfs->os_fsid, .... );*/
541 osfs->os_namelen = MAXNAMELEN;
542 osfs->os_maxbytes = OBD_OBJECT_EOF;
544 if (!spa_writeable(dmu_objset_spa(os)) ||
545 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
546 osfs->os_state |= OS_STATE_READONLY;
552 * Concurrency: shouldn't matter.
554 int osd_statfs(const struct lu_env *env, struct dt_device *d,
555 struct obd_statfs *osfs)
557 struct osd_device *osd = osd_dt_dev(d);
561 rc = osd_objset_statfs(osd, osfs);
562 if (unlikely(rc != 0))
565 osfs->os_bavail -= min_t(u64,
566 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
569 /* ZFS does not support reporting nonrotional status yet, so return
570 * flag only if user has set nonrotational.
572 osfs->os_state |= osd->od_nonrotational ? OS_STATE_NONROT : 0;
577 static int osd_blk_insert_cost(struct osd_device *osd)
579 int max_blockshift, nr_blkptrshift, bshift;
581 /* max_blockshift is the log2 of the number of blocks needed to reach
582 * the maximum filesize (that's to say 2^64) */
583 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
584 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
586 /* nr_blkptrshift is the log2 of the number of block pointers that can
587 * be stored in an indirect block */
588 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
589 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
591 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
592 * tree. We add +1 for rounding purpose.
593 * The tree depth times the indirect block size gives us the maximum
594 * cost of inserting a block in the tree */
595 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
599 * Concurrency: doesn't access mutable data.
601 static void osd_conf_get(const struct lu_env *env,
602 const struct dt_device *dev,
603 struct dt_device_param *param)
605 struct osd_device *osd = osd_dt_dev(dev);
608 * XXX should be taken from not-yet-existing fs abstraction layer.
610 param->ddp_max_name_len = MAXNAMELEN;
611 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
612 param->ddp_symlink_max = PATH_MAX;
613 param->ddp_mount_type = LDD_MT_ZFS;
615 param->ddp_mntopts = MNTOPT_USERXATTR;
616 if (osd->od_posix_acl)
617 param->ddp_mntopts |= MNTOPT_ACL;
618 /* Previously DXATTR_MAX_ENTRY_SIZE */
619 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
621 /* for maxbytes, report same value as ZPL */
622 param->ddp_maxbytes = MAX_LFS_FILESIZE;
624 /* inodes are dynamically allocated, so we report the per-inode space
625 * consumption to upper layers. This static value is not really accurate
626 * and we should use the same logic as in udmu_objset_statfs() to
627 * estimate the real size consumed by an object */
628 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
629 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
630 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
631 * not be accounted for every single new block insertion.
632 * Instead, the maximum extent size is set to the number of blocks that
633 * can fit into a single contiguous indirect block. There would be some
634 * cases where this crosses indirect blocks, but it also won't have 7
635 * new levels of indirect blocks in that case either, so it will still
636 * have enough reserved space for the extra indirect block */
637 param->ddp_max_extent_blks =
638 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
639 param->ddp_extent_tax = osd_blk_insert_cost(osd);
641 /* Preferred RPC size for efficient disk IO. 1MB shows good
642 * all-around performance for ZFS, but use blocksize (recordsize)
643 * by default if larger to avoid read-modify-write. */
644 if (osd->od_max_blksz > ONE_MB_BRW_SIZE)
645 param->ddp_brw_size = osd->od_max_blksz;
647 param->ddp_brw_size = ONE_MB_BRW_SIZE;
651 * Concurrency: shouldn't matter.
653 static int osd_sync(const struct lu_env *env, struct dt_device *d)
656 struct osd_device *osd = osd_dt_dev(d);
658 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
659 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
660 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
666 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
668 struct osd_device *osd = osd_dt_dev(dev);
669 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
672 mutex_enter(&tx->tx_sync_lock);
673 txg = tx->tx_open_txg + 1;
674 if (tx->tx_quiesce_txg_waiting < txg) {
675 tx->tx_quiesce_txg_waiting = txg;
676 cv_broadcast(&tx->tx_quiesce_more_cv);
678 mutex_exit(&tx->tx_sync_lock);
684 * Concurrency: shouldn't matter.
686 static int osd_ro(const struct lu_env *env, struct dt_device *d)
688 struct osd_device *osd = osd_dt_dev(d);
691 CERROR("%s: *** setting device %s read-only ***\n",
692 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
693 osd->od_dev_set_rdonly = 1;
694 spa_freeze(dmu_objset_spa(osd->od_os));
699 static struct dt_device_operations osd_dt_ops = {
700 .dt_root_get = osd_root_get,
701 .dt_statfs = osd_statfs,
702 .dt_trans_create = osd_trans_create,
703 .dt_trans_start = osd_trans_start,
704 .dt_trans_stop = osd_trans_stop,
705 .dt_trans_cb_add = osd_trans_cb_add,
706 .dt_conf_get = osd_conf_get,
708 .dt_commit_async = osd_commit_async,
713 * DMU OSD device type methods
715 static int osd_type_init(struct lu_device_type *t)
717 LU_CONTEXT_KEY_INIT(&osd_key);
718 return lu_context_key_register(&osd_key);
721 static void osd_type_fini(struct lu_device_type *t)
723 lu_context_key_degister(&osd_key);
726 static void *osd_key_init(const struct lu_context *ctx,
727 struct lu_context_key *key)
729 struct osd_thread_info *info;
733 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
735 info = ERR_PTR(-ENOMEM);
739 static void osd_key_fini(const struct lu_context *ctx,
740 struct lu_context_key *key, void *data)
742 struct osd_thread_info *info = data;
743 struct osd_idmap_cache *idc = info->oti_ins_cache;
746 LASSERT(info->oti_ins_cache_size > 0);
747 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
748 info->oti_ins_cache = NULL;
749 info->oti_ins_cache_size = 0;
751 lu_buf_free(&info->oti_xattr_lbuf);
755 static void osd_key_exit(const struct lu_context *ctx,
756 struct lu_context_key *key, void *data)
760 struct lu_context_key osd_key = {
761 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
762 .lct_init = osd_key_init,
763 .lct_fini = osd_key_fini,
764 .lct_exit = osd_key_exit
767 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
769 if (osd->od_cl_seq == NULL)
772 seq_client_fini(osd->od_cl_seq);
773 OBD_FREE_PTR(osd->od_cl_seq);
774 osd->od_cl_seq = NULL;
777 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
781 /* shutdown quota slave instance associated with the device */
782 if (o->od_quota_slave_md != NULL) {
783 /* complete all in-flight callbacks */
784 osd_sync(env, &o->od_dt_dev);
785 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
786 qsd_fini(env, o->od_quota_slave_md);
787 o->od_quota_slave_md = NULL;
790 if (o->od_quota_slave_dt != NULL) {
791 /* complete all in-flight callbacks */
792 osd_sync(env, &o->od_dt_dev);
793 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
794 qsd_fini(env, o->od_quota_slave_dt);
795 o->od_quota_slave_dt = NULL;
797 osd_fid_fini(env, o);
802 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
804 struct osd_device *osd = arg;
806 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
809 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
811 struct osd_device *osd = arg;
813 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
814 LASSERT(newval >= SPA_MINBLOCKSIZE);
815 LASSERT(ISP2(newval));
817 osd->od_max_blksz = newval;
820 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
822 struct osd_device *osd = arg;
824 osd->od_prop_rdonly = !!newval;
827 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
828 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
830 struct osd_device *osd = arg;
832 osd->od_dnsize = newval;
836 * This function unregisters all registered callbacks. It's harmless to
837 * unregister callbacks that were never registered so it is used to safely
838 * unwind a partially completed call to osd_objset_register_callbacks().
840 static void osd_objset_unregister_callbacks(struct osd_device *o)
842 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
844 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
845 osd_xattr_changed_cb, o);
846 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
847 osd_recordsize_changed_cb, o);
848 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
849 osd_readonly_changed_cb, o);
850 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
851 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
852 osd_dnodesize_changed_cb, o);
855 if (o->arc_prune_cb != NULL) {
856 arc_remove_prune_callback(o->arc_prune_cb);
857 o->arc_prune_cb = NULL;
862 * Register the required callbacks to be notified when zfs properties
863 * are modified using the 'zfs(8)' command line utility.
865 static int osd_objset_register_callbacks(struct osd_device *o)
867 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
868 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
874 dsl_pool_config_enter(dp, FTAG);
875 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
876 osd_xattr_changed_cb, o);
880 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
881 osd_recordsize_changed_cb, o);
885 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
886 osd_readonly_changed_cb, o);
890 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
891 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
892 osd_dnodesize_changed_cb, o);
897 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
899 dsl_pool_config_exit(dp, FTAG);
901 osd_objset_unregister_callbacks(o);
906 static int osd_objset_open(struct osd_device *o)
908 uint64_t version = ZPL_VERSION;
909 uint64_t sa_obj, unlink_obj;
913 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
914 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
915 B_TRUE, o, &o->od_os);
918 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
924 /* Check ZFS version */
925 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
926 ZPL_VERSION_STR, 8, 1, &version);
928 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
930 * We can't return ENOENT because that would mean the objset
933 GOTO(out, rc = -EIO);
936 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
937 ZFS_SA_ATTRS, 8, 1, &sa_obj);
941 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
942 ZPL_END, &o->z_attr_table);
946 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
947 8, 1, &o->od_rootid);
949 CERROR("%s: lookup for root failed: rc = %d\n",
954 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
957 CERROR("%s: lookup for %s failed: rc = %d\n",
958 o->od_svname, ZFS_UNLINKED_SET, rc);
962 /* Check that user/group usage tracking is supported */
963 if (!dmu_objset_userused_enabled(o->od_os) ||
964 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
965 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
966 CERROR("%s: Space accounting not supported by this target, "
967 "aborting\n", o->od_svname);
968 GOTO(out, rc = -ENOTSUPP);
971 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
973 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
979 if (rc != 0 && o->od_os != NULL) {
980 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
987 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
990 char *key = osd_oti_get(env)->oti_str;
994 if (osd->od_dt_dev.dd_rdonly) {
995 CERROR("%s: someone try to free objects under "
996 "readonly mode, should be disabled.\n", osd_name(osd));
1002 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
1004 CWARN("%s: Cannot truncate %llu: rc = %d\n",
1005 osd->od_svname, oid, rc);
1009 tx = dmu_tx_create(osd->od_os);
1010 dmu_tx_mark_netfree(tx);
1011 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
1012 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
1014 rc = -dmu_tx_assign(tx, TXG_WAIT);
1016 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
1017 osd->od_svname, oid, rc);
1021 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
1022 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
1023 osd->od_unlinked, key, tx);
1025 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
1026 osd->od_svname, oid, rc);
1030 rc = -dmu_object_free(osd->od_os, oid, tx);
1032 CWARN("%s: Cannot free %llu: rc = %d\n",
1033 osd->od_svname, oid, rc);
1048 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
1051 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1053 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1055 while (zap_cursor_retrieve(&zc, za) == 0) {
1056 /* If cannot free the object, leave it in the unlinked set,
1057 * until the OSD is mounted again when obd_unlinked_drain()
1058 * will be called. */
1059 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1061 zap_cursor_advance(&zc);
1064 zap_cursor_fini(&zc);
1067 static int osd_mount(const struct lu_env *env,
1068 struct osd_device *o, struct lustre_cfg *cfg)
1070 char *mntdev = lustre_cfg_string(cfg, 1);
1071 char *str = lustre_cfg_string(cfg, 2);
1072 char *svname = lustre_cfg_string(cfg, 4);
1078 if (o->od_os != NULL)
1081 if (mntdev == NULL || svname == NULL)
1084 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1085 if (rc >= sizeof(o->od_mntdev))
1088 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1089 if (rc >= sizeof(o->od_svname))
1092 o->od_index_backup_stop = 0;
1093 o->od_index = -1; /* -1 means index is invalid */
1094 rc = server_name2index(o->od_svname, &o->od_index, NULL);
1095 str = strstr(str, ":");
1097 unsigned long flags;
1099 rc = kstrtoul(str + 1, 10, &flags);
1103 if (flags & LMD_FLG_DEV_RDONLY) {
1104 o->od_dt_dev.dd_rdonly = 1;
1105 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1109 if (flags & LMD_FLG_NOSCRUB)
1110 o->od_auto_scrub_interval = AS_NEVER;
1113 if (server_name_is_ost(o->od_svname))
1116 rc = osd_objset_open(o);
1120 o->od_xattr_in_sa = B_TRUE;
1121 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1122 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
1124 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1127 o->od_root = rootdn->dn_object;
1128 osd_dnode_rele(rootdn);
1130 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1131 &o->od_userused_dn);
1135 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1136 &o->od_groupused_dn);
1140 #ifdef ZFS_PROJINHERIT
1141 if (dmu_objset_projectquota_enabled(o->od_os)) {
1142 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1143 &o->od_projectused_dn);
1144 if (rc && rc != -ENOENT)
1149 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1152 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1154 rc = lu_site_init_finish(&o->od_site);
1158 rc = osd_objset_register_callbacks(o);
1163 rc = osd_scrub_setup(env, o);
1168 rc = osd_procfs_init(o, o->od_svname);
1172 /* currently it's no need to prepare qsd_instance_md for OST */
1173 if (!o->od_is_ost) {
1174 o->od_quota_slave_md = qsd_init(env, o->od_svname,
1176 o->od_proc_entry, true);
1177 if (IS_ERR(o->od_quota_slave_md)) {
1178 rc = PTR_ERR(o->od_quota_slave_md);
1179 o->od_quota_slave_md = NULL;
1184 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
1185 o->od_proc_entry, false);
1187 if (IS_ERR(o->od_quota_slave_dt)) {
1188 if (o->od_quota_slave_md != NULL) {
1189 qsd_fini(env, o->od_quota_slave_md);
1190 o->od_quota_slave_md = NULL;
1193 rc = PTR_ERR(o->od_quota_slave_dt);
1194 o->od_quota_slave_dt = NULL;
1198 #ifdef HAVE_DMU_USEROBJ_ACCOUNTING
1199 if (!osd_dmu_userobj_accounting_available(o))
1200 CWARN("%s: dnode accounting not enabled: "
1201 "enable feature@userobj_accounting in pool\n",
1205 /* parse mount option "noacl", and enable ACL by default */
1206 opts = lustre_cfg_string(cfg, 3);
1207 if (opts == NULL || strstr(opts, "noacl") == NULL)
1208 o->od_posix_acl = 1;
1210 osd_unlinked_drain(env, o);
1212 if (rc && o->od_os) {
1213 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1220 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1224 if (atomic_read(&o->od_zerocopy_alloc))
1225 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1226 atomic_read(&o->od_zerocopy_alloc));
1227 if (atomic_read(&o->od_zerocopy_loan))
1228 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1229 atomic_read(&o->od_zerocopy_loan));
1230 if (atomic_read(&o->od_zerocopy_pin))
1231 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1232 atomic_read(&o->od_zerocopy_pin));
1234 if (o->od_unlinked) {
1235 osd_dnode_rele(o->od_unlinked);
1236 o->od_unlinked = NULL;
1238 if (o->od_userused_dn) {
1239 osd_dnode_rele(o->od_userused_dn);
1240 o->od_userused_dn = NULL;
1242 if (o->od_groupused_dn) {
1243 osd_dnode_rele(o->od_groupused_dn);
1244 o->od_groupused_dn = NULL;
1247 #ifdef ZFS_PROJINHERIT
1248 if (o->od_projectused_dn) {
1249 osd_dnode_rele(o->od_projectused_dn);
1250 o->od_projectused_dn = NULL;
1254 if (o->od_os != NULL) {
1255 if (!o->od_dt_dev.dd_rdonly)
1256 /* force a txg sync to get all commit callbacks */
1257 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1259 /* close the object set */
1260 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1267 static int osd_device_init0(const struct lu_env *env,
1268 struct osd_device *o,
1269 struct lustre_cfg *cfg)
1271 struct lu_device *l = osd2lu_dev(o);
1274 /* if the module was re-loaded, env can loose its keys */
1275 rc = lu_env_refill((struct lu_env *) env);
1279 l->ld_ops = &osd_lu_ops;
1280 o->od_dt_dev.dd_ops = &osd_dt_ops;
1281 sema_init(&o->od_otable_sem, 1);
1282 INIT_LIST_HEAD(&o->od_ios_list);
1283 o->od_auto_scrub_interval = AS_DEFAULT;
1285 /* ZFS does not support reporting nonrotional status yet, so this flag
1286 * is only set if explicitly set by the user.
1288 o->od_nonrotational = 0;
1294 static struct lu_device *osd_device_fini(const struct lu_env *env,
1295 struct lu_device *dev);
1297 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1298 struct lu_device_type *type,
1299 struct lustre_cfg *cfg)
1301 struct osd_device *dev;
1302 struct osd_seq_list *osl;
1307 return ERR_PTR(-ENOMEM);
1309 osl = &dev->od_seq_list;
1310 INIT_LIST_HEAD(&osl->osl_seq_list);
1311 rwlock_init(&osl->osl_seq_list_lock);
1312 sema_init(&osl->osl_seq_init_sem, 1);
1313 INIT_LIST_HEAD(&dev->od_index_backup_list);
1314 INIT_LIST_HEAD(&dev->od_index_restore_list);
1315 spin_lock_init(&dev->od_lock);
1316 dev->od_index_backup_policy = LIBP_NONE;
1318 rc = dt_device_init(&dev->od_dt_dev, type);
1320 rc = osd_device_init0(env, dev, cfg);
1322 rc = osd_mount(env, dev, cfg);
1324 osd_device_fini(env, osd2lu_dev(dev));
1327 dt_device_fini(&dev->od_dt_dev);
1330 if (unlikely(rc != 0))
1333 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1336 static struct lu_device *osd_device_free(const struct lu_env *env,
1337 struct lu_device *d)
1339 struct osd_device *o = osd_dev(d);
1342 /* XXX: make osd top device in order to release reference */
1343 d->ld_site->ls_top_dev = d;
1344 lu_site_purge(env, d->ld_site, -1);
1345 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1346 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1347 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1349 lu_site_fini(&o->od_site);
1350 dt_device_fini(&o->od_dt_dev);
1356 static struct lu_device *osd_device_fini(const struct lu_env *env,
1357 struct lu_device *d)
1359 struct osd_device *o = osd_dev(d);
1365 osd_objset_unregister_callbacks(o);
1366 if (!o->od_dt_dev.dd_rdonly) {
1367 osd_sync(env, lu2dt_dev(d));
1369 spa_get_dsl(dmu_objset_spa(o->od_os)));
1373 /* now with all the callbacks completed we can cleanup the remainings */
1374 osd_shutdown(env, o);
1375 osd_scrub_cleanup(env, o);
1377 rc = osd_procfs_fini(o);
1379 CERROR("proc fini error %d\n", rc);
1380 RETURN(ERR_PTR(rc));
1389 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1390 const char *name, struct lu_device *next)
1396 * To be removed, setup is performed by osd_device_{init,alloc} and
1397 * cleanup is performed by osd_device_{fini,free).
1399 static int osd_process_config(const struct lu_env *env,
1400 struct lu_device *d, struct lustre_cfg *cfg)
1402 struct osd_device *o = osd_dev(d);
1407 switch(cfg->lcfg_command) {
1409 rc = osd_mount(env, o, cfg);
1412 /* For the case LCFG_PRE_CLEANUP is not called in advance,
1413 * that may happend if hit failure during mount process. */
1414 osd_index_backup(env, o, false);
1415 rc = osd_shutdown(env, o);
1418 LASSERT(&o->od_dt_dev);
1419 count = class_modify_config(cfg, PARAM_OSD,
1420 &o->od_dt_dev.dd_kobj);
1422 count = class_modify_config(cfg, PARAM_OST,
1423 &o->od_dt_dev.dd_kobj);
1424 rc = count > 0 ? 0 : count;
1427 case LCFG_PRE_CLEANUP:
1429 osd_index_backup(env, o,
1430 o->od_index_backup_policy != LIBP_NONE);
1440 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1442 struct osd_device *osd = osd_dev(d);
1446 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
1449 /* start qsd instance on recovery completion, this notifies the quota
1450 * slave code that we are about to process new requests now */
1451 rc = qsd_start(env, osd->od_quota_slave_dt);
1452 if (rc == 0 && osd->od_quota_slave_md != NULL)
1453 rc = qsd_start(env, osd->od_quota_slave_md);
1458 * we use exports to track all osd users
1460 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1461 struct obd_device *obd, struct obd_uuid *cluuid,
1462 struct obd_connect_data *data, void *localdata)
1464 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1465 struct lustre_handle conn;
1469 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1471 rc = class_connect(&conn, obd, cluuid);
1475 *exp = class_conn2export(&conn);
1477 spin_lock(&obd->obd_dev_lock);
1479 spin_unlock(&obd->obd_dev_lock);
1485 * once last export (we don't count self-export) disappeared
1486 * osd can be released
1488 static int osd_obd_disconnect(struct obd_export *exp)
1490 struct obd_device *obd = exp->exp_obd;
1491 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1492 int rc, release = 0;
1495 /* Only disconnect the underlying layers on the final disconnect. */
1496 spin_lock(&obd->obd_dev_lock);
1498 if (osd->od_connects == 0)
1500 spin_unlock(&obd->obd_dev_lock);
1502 rc = class_disconnect(exp); /* bz 9811 */
1504 if (rc == 0 && release)
1505 class_manual_cleanup(obd);
1509 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1511 struct seq_server_site *ss = osd_seq_site(osd);
1515 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1518 if (unlikely(ss == NULL))
1521 OBD_ALLOC_PTR(osd->od_cl_seq);
1522 if (osd->od_cl_seq == NULL)
1525 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1526 osd->od_svname, ss->ss_server_seq);
1529 OBD_FREE_PTR(osd->od_cl_seq);
1530 osd->od_cl_seq = NULL;
1533 if (ss->ss_node_id == 0) {
1535 * If the OSD on the sequence controller(MDT0), then allocate
1536 * sequence here, otherwise allocate sequence after connected
1537 * to MDT0 (see mdt_register_lwp_callback()).
1539 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
1540 &osd->od_cl_seq->lcs_space, env);
1546 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1547 struct lu_device *dev)
1549 struct osd_device *osd = osd_dev(dev);
1553 if (osd->od_quota_slave_md != NULL) {
1554 /* set up quota slave objects */
1555 rc = qsd_prepare(env, osd->od_quota_slave_md);
1560 if (osd->od_quota_slave_dt != NULL) {
1561 /* set up quota slave objects */
1562 rc = qsd_prepare(env, osd->od_quota_slave_dt);
1567 rc = osd_fid_init(env, osd);
1572 struct lu_device_operations osd_lu_ops = {
1573 .ldo_object_alloc = osd_object_alloc,
1574 .ldo_process_config = osd_process_config,
1575 .ldo_recovery_complete = osd_recovery_complete,
1576 .ldo_prepare = osd_prepare,
1579 static void osd_type_start(struct lu_device_type *t)
1583 static void osd_type_stop(struct lu_device_type *t)
1587 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1588 struct lu_fid *fid, struct md_op_data *op_data)
1590 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1592 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1595 static struct lu_device_type_operations osd_device_type_ops = {
1596 .ldto_init = osd_type_init,
1597 .ldto_fini = osd_type_fini,
1599 .ldto_start = osd_type_start,
1600 .ldto_stop = osd_type_stop,
1602 .ldto_device_alloc = osd_device_alloc,
1603 .ldto_device_free = osd_device_free,
1605 .ldto_device_init = osd_device_init,
1606 .ldto_device_fini = osd_device_fini
1609 static struct lu_device_type osd_device_type = {
1610 .ldt_tags = LU_DEVICE_DT,
1611 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1612 .ldt_ops = &osd_device_type_ops,
1613 .ldt_ctx_tags = LCT_LOCAL
1617 static struct obd_ops osd_obd_device_ops = {
1618 .o_owner = THIS_MODULE,
1619 .o_connect = osd_obd_connect,
1620 .o_disconnect = osd_obd_disconnect,
1621 .o_fid_alloc = osd_fid_alloc
1624 static int __init osd_init(void)
1628 rc = osd_options_init();
1632 rc = lu_kmem_init(osd_caches);
1636 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1637 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1639 lu_kmem_fini(osd_caches);
1643 static void __exit osd_exit(void)
1645 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1646 lu_kmem_fini(osd_caches);
1649 module_param(osd_oi_count, int, 0444);
1650 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1652 module_param(osd_txg_sync_delay_us, int, 0644);
1653 MODULE_PARM_DESC(osd_txg_sync_delay_us,
1654 "When zero or larger delay N usec instead of doing TXG sync");
1656 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1657 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1658 MODULE_VERSION(LUSTRE_VERSION_STRING);
1659 MODULE_LICENSE("GPL");
1661 module_init(osd_init);
1662 module_exit(osd_exit);