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/
31 * lustre/osd-zfs/osd_handler.c
32 * Top-level entry points into osd module
34 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
35 * Author: Mike Pershin <tappro@whamcloud.com>
36 * Author: Johann Lombardi <johann@whamcloud.com>
39 #define DEBUG_SUBSYSTEM S_OSD
41 #include <libcfs/libcfs.h>
42 #include <obd_support.h>
43 #include <lustre_net.h>
45 #include <obd_class.h>
46 #include <lustre_disk.h>
47 #include <lustre_fid.h>
48 #include <uapi/linux/lustre/lustre_param.h>
49 #include <md_object.h>
51 #include "osd_internal.h"
53 #include <sys/dnode.h>
58 #include <sys/spa_impl.h>
59 #include <sys/zfs_znode.h>
60 #include <sys/dmu_tx.h>
61 #include <sys/dmu_objset.h>
62 #include <sys/dsl_prop.h>
63 #include <sys/sa_impl.h>
66 struct lu_context_key osd_key;
68 static int osd_txg_sync_delay_us = -1;
70 /* Slab for OSD object allocation */
71 struct kmem_cache *osd_object_kmem;
73 /* Slab to allocate osd_zap_it */
74 struct kmem_cache *osd_zapit_cachep;
76 static struct lu_kmem_descr osd_caches[] = {
78 .ckd_cache = &osd_object_kmem,
79 .ckd_name = "zfs_osd_obj",
80 .ckd_size = sizeof(struct osd_object)
83 .ckd_cache = &osd_zapit_cachep,
84 .ckd_name = "osd_zapit_cache",
85 .ckd_size = sizeof(struct osd_zap_it)
92 #ifdef HAVE_ZFS_ARC_PRUNE_FUNC_UINT64
93 static void arc_prune_func(uint64_t bytes, void *private)
95 static void arc_prune_func(int64_t bytes, void *private)
98 struct osd_device *od = private;
99 struct lu_site *site = &od->od_site;
103 rc = lu_env_init(&env, LCT_SHRINKER);
105 CERROR("%s: can't initialize shrinker env: rc = %d\n",
110 lu_site_purge(&env, site, (bytes >> 10));
116 * Concurrency: doesn't access mutable data
118 static int osd_root_get(const struct lu_env *env,
119 struct dt_device *dev, struct lu_fid *f)
121 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
126 * OSD object methods.
130 * Concurrency: shouldn't matter.
132 static void osd_trans_commit_cb(void *cb_data, int error)
134 struct osd_thandle *oh = cb_data;
135 struct thandle *th = &oh->ot_super;
136 struct osd_device *osd = osd_dt_dev(th->th_dev);
137 struct lu_device *lud = &th->th_dev->dd_lu_dev;
138 struct dt_txn_commit_cb *dcb, *tmp;
143 if (error == ECANCELED)
144 CWARN("%s: transaction @0x%p was aborted\n",
145 osd_dt_dev(th->th_dev)->od_svname, th);
147 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
148 osd_dt_dev(th->th_dev)->od_svname, th, error);
151 /* call per-transaction callbacks if any */
152 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) {
153 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
154 "commit callback entry: magic=%x name='%s'\n",
155 dcb->dcb_magic, dcb->dcb_name);
156 list_del_init(&dcb->dcb_linkage);
157 dcb->dcb_func(NULL, th, dcb, error);
160 /* Unlike ldiskfs, zfs updates space accounting at commit time.
161 * As a consequence, op_end is called only now to inform the quota slave
162 * component that reserved quota space is now accounted in usage and
163 * should be released. Quota space won't be adjusted at this point since
164 * we can't provide a suitable environment. It will be performed
165 * asynchronously by a lquota thread. */
166 qsd_op_end(NULL, osd->od_quota_slave_dt, &oh->ot_quota_trans);
167 if (osd->od_quota_slave_md != NULL)
168 qsd_op_end(NULL, osd->od_quota_slave_md, &oh->ot_quota_trans);
177 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
179 struct osd_thandle *oh = container_of(th, struct osd_thandle,
182 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
183 LASSERT(&dcb->dcb_func != NULL);
184 if (dcb->dcb_flags & DCB_TRANS_STOP)
185 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
187 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
193 * Concurrency: shouldn't matter.
195 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
198 struct osd_device *osd = osd_dt_dev(d);
199 struct osd_thandle *oh;
204 oh = container_of(th, struct osd_thandle, ot_super);
208 rc = dt_txn_hook_start(env, d, th);
210 CERROR("%s: dt_txn_hook_start failed: rc = %d\n",
215 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
216 /* Unlike ldiskfs, ZFS checks for available space and returns
217 * -ENOSPC when assigning txg */
220 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
221 if (unlikely(rc != 0)) {
222 /* dmu will call commit callback with error code during abort */
223 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
224 CERROR("%s: failed to start transaction due to ENOSPC"
225 "\n", osd->od_svname);
227 CERROR("%s: can't assign tx: rc = %d\n",
230 /* add commit callback */
231 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
233 osd_oti_get(env)->oti_in_trans = 1;
234 lu_device_get(&d->dd_lu_dev);
240 static void osd_unlinked_list_emptify(const struct lu_env *env,
241 struct osd_device *osd,
242 struct list_head *list, bool free)
244 struct osd_object *obj;
247 while (!list_empty(list)) {
248 obj = list_first_entry(list,
249 struct osd_object, oo_unlinked_linkage);
250 LASSERT(obj->oo_dn != NULL);
251 oid = obj->oo_dn->dn_object;
253 list_del_init(&obj->oo_unlinked_linkage);
255 (void)osd_unlinked_object_free(env, osd, oid);
259 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
261 struct dt_txn_commit_cb *dcb;
262 struct dt_txn_commit_cb *tmp;
264 /* call per-transaction stop callbacks if any */
265 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
267 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
268 "commit callback entry: magic=%x name='%s'\n",
269 dcb->dcb_magic, dcb->dcb_name);
270 list_del_init(&dcb->dcb_linkage);
271 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
276 * Concurrency: shouldn't matter.
278 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
281 struct osd_device *osd = osd_dt_dev(th->th_dev);
282 bool sync = (th->th_sync != 0);
283 struct osd_thandle *oh;
289 oh = container_of(th, struct osd_thandle, ot_super);
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 BUILD_BUG_ON(OSD_DNODE_MIN_BLKSHIFT <= 0);
415 BUILD_BUG_ON(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 BUILD_BUG_ON(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_STATFS_READONLY;
552 * Concurrency: shouldn't matter.
554 int osd_statfs(const struct lu_env *env, struct dt_device *d,
555 struct obd_statfs *osfs, struct obd_statfs_info *info)
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_STATFS_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 = fls64(spa_maxblocksize(dmu_objset_spa(osd->od_os)) - 1);
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 BUILD_BUG_ON(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;
649 #ifdef HAVE_DMU_OFFSET_NEXT
650 param->ddp_has_lseek_data_hole = osd->od_sync_on_lseek;
652 param->ddp_has_lseek_data_hole = false;
657 * Concurrency: shouldn't matter.
659 static int osd_sync(const struct lu_env *env, struct dt_device *d)
662 struct osd_device *osd = osd_dt_dev(d);
664 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
665 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
666 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
672 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
674 struct osd_device *osd = osd_dt_dev(dev);
675 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
678 mutex_enter(&tx->tx_sync_lock);
679 txg = tx->tx_open_txg + 1;
680 if (tx->tx_quiesce_txg_waiting < txg) {
681 tx->tx_quiesce_txg_waiting = txg;
682 cv_broadcast(&tx->tx_quiesce_more_cv);
684 mutex_exit(&tx->tx_sync_lock);
690 * Concurrency: shouldn't matter.
692 static int osd_ro(const struct lu_env *env, struct dt_device *d)
694 struct osd_device *osd = osd_dt_dev(d);
697 CERROR("%s: *** setting device %s read-only ***\n",
698 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
699 osd->od_dev_set_rdonly = 1;
700 spa_freeze(dmu_objset_spa(osd->od_os));
705 /* reserve or free quota for some operation */
706 static int osd_reserve_or_free_quota(const struct lu_env *env,
707 struct dt_device *dev,
708 struct lquota_id_info *qi)
710 struct osd_device *osd = osd_dt_dev(dev);
711 struct qsd_instance *qsd = NULL;
717 qsd = osd->od_quota_slave_dt;
719 qsd = osd->od_quota_slave_md;
721 rc = qsd_reserve_or_free_quota(env, qsd, qi);
725 static const struct dt_device_operations osd_dt_ops = {
726 .dt_root_get = osd_root_get,
727 .dt_statfs = osd_statfs,
728 .dt_trans_create = osd_trans_create,
729 .dt_trans_start = osd_trans_start,
730 .dt_trans_stop = osd_trans_stop,
731 .dt_trans_cb_add = osd_trans_cb_add,
732 .dt_conf_get = osd_conf_get,
734 .dt_commit_async = osd_commit_async,
736 .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
740 * DMU OSD device type methods
742 static int osd_type_init(struct lu_device_type *t)
744 LU_CONTEXT_KEY_INIT(&osd_key);
745 return lu_context_key_register(&osd_key);
748 static void osd_type_fini(struct lu_device_type *t)
750 lu_context_key_degister(&osd_key);
753 static void *osd_key_init(const struct lu_context *ctx,
754 struct lu_context_key *key)
756 struct osd_thread_info *info;
760 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
762 info = ERR_PTR(-ENOMEM);
766 static void osd_key_fini(const struct lu_context *ctx,
767 struct lu_context_key *key, void *data)
769 struct osd_thread_info *info = data;
770 struct osd_idmap_cache *idc = info->oti_ins_cache;
773 LASSERT(info->oti_ins_cache_size > 0);
774 OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size);
775 info->oti_ins_cache = NULL;
776 info->oti_ins_cache_size = 0;
778 lu_buf_free(&info->oti_xattr_lbuf);
782 static void osd_key_exit(const struct lu_context *ctx,
783 struct lu_context_key *key, void *data)
787 struct lu_context_key osd_key = {
788 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
789 .lct_init = osd_key_init,
790 .lct_fini = osd_key_fini,
791 .lct_exit = osd_key_exit
794 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
796 if (osd->od_cl_seq == NULL)
799 seq_client_fini(osd->od_cl_seq);
800 OBD_FREE_PTR(osd->od_cl_seq);
801 osd->od_cl_seq = NULL;
804 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
808 /* shutdown quota slave instance associated with the device */
809 if (o->od_quota_slave_md != NULL) {
810 /* complete all in-flight callbacks */
811 osd_sync(env, &o->od_dt_dev);
812 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
813 qsd_fini(env, o->od_quota_slave_md);
814 o->od_quota_slave_md = NULL;
817 if (o->od_quota_slave_dt != NULL) {
818 /* complete all in-flight callbacks */
819 osd_sync(env, &o->od_dt_dev);
820 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
821 qsd_fini(env, o->od_quota_slave_dt);
822 o->od_quota_slave_dt = NULL;
824 osd_fid_fini(env, o);
829 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
831 struct osd_device *osd = arg;
833 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
836 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
838 struct osd_device *osd = arg;
840 LASSERT(newval <= spa_maxblocksize(dmu_objset_spa(osd->od_os)));
841 LASSERT(newval >= SPA_MINBLOCKSIZE);
842 LASSERT(ISP2(newval));
844 osd->od_max_blksz = newval;
847 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
849 struct osd_device *osd = arg;
851 osd->od_prop_rdonly = !!newval;
854 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
856 struct osd_device *osd = arg;
858 osd->od_dnsize = newval;
862 * This function unregisters all registered callbacks. It's harmless to
863 * unregister callbacks that were never registered so it is used to safely
864 * unwind a partially completed call to osd_objset_register_callbacks().
866 static void osd_objset_unregister_callbacks(struct osd_device *o)
868 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
870 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
871 osd_xattr_changed_cb, o);
872 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
873 osd_recordsize_changed_cb, o);
874 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
875 osd_readonly_changed_cb, o);
876 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
877 osd_dnodesize_changed_cb, o);
879 if (o->arc_prune_cb != NULL) {
880 arc_remove_prune_callback(o->arc_prune_cb);
881 o->arc_prune_cb = NULL;
886 * Register the required callbacks to be notified when zfs properties
887 * are modified using the 'zfs(8)' command line utility.
889 static int osd_objset_register_callbacks(struct osd_device *o)
891 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
892 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
898 dsl_pool_config_enter(dp, FTAG);
899 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
900 osd_xattr_changed_cb, o);
904 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
905 osd_recordsize_changed_cb, o);
909 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
910 osd_readonly_changed_cb, o);
914 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
915 osd_dnodesize_changed_cb, o);
919 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
921 dsl_pool_config_exit(dp, FTAG);
923 osd_objset_unregister_callbacks(o);
928 static int osd_objset_open(struct osd_device *o)
930 uint64_t version = ZPL_VERSION;
931 uint64_t sa_obj, unlink_obj;
935 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
936 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
937 B_TRUE, o, &o->od_os);
940 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
946 /* Check ZFS version */
947 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
948 ZPL_VERSION_STR, 8, 1, &version);
950 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
952 * We can't return ENOENT because that would mean the objset
955 GOTO(out, rc = -EIO);
958 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
959 ZFS_SA_ATTRS, 8, 1, &sa_obj);
963 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
964 ZPL_END, &o->z_attr_table);
968 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
969 8, 1, &o->od_rootid);
971 CERROR("%s: lookup for root failed: rc = %d\n",
976 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
979 CERROR("%s: lookup for %s failed: rc = %d\n",
980 o->od_svname, ZFS_UNLINKED_SET, rc);
984 /* Check that user/group usage tracking is supported */
985 if (!dmu_objset_userused_enabled(o->od_os) ||
986 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
987 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
988 CERROR("%s: Space accounting not supported by this target, "
989 "aborting\n", o->od_svname);
990 GOTO(out, rc = -ENOTSUPP);
993 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
995 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
1001 if (rc != 0 && o->od_os != NULL) {
1002 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1009 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
1012 char *key = osd_oti_get(env)->oti_str;
1016 if (osd->od_dt_dev.dd_rdonly) {
1017 CERROR("%s: someone try to free objects under "
1018 "readonly mode, should be disabled.\n", osd_name(osd));
1024 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
1026 CWARN("%s: Cannot truncate %llu: rc = %d\n",
1027 osd->od_svname, oid, rc);
1031 tx = dmu_tx_create(osd->od_os);
1032 dmu_tx_mark_netfree(tx);
1033 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
1034 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
1036 rc = -dmu_tx_assign(tx, TXG_WAIT);
1038 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
1039 osd->od_svname, oid, rc);
1043 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
1044 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
1045 osd->od_unlinked, key, tx);
1047 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
1048 osd->od_svname, oid, rc);
1052 rc = -dmu_object_free(osd->od_os, oid, tx);
1054 CWARN("%s: Cannot free %llu: rc = %d\n",
1055 osd->od_svname, oid, rc);
1070 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
1073 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1075 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1077 while (zap_cursor_retrieve(&zc, za) == 0) {
1078 /* If cannot free the object, leave it in the unlinked set,
1079 * until the OSD is mounted again when obd_unlinked_drain()
1080 * will be called. */
1081 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1083 zap_cursor_advance(&zc);
1086 zap_cursor_fini(&zc);
1089 static int osd_mount(const struct lu_env *env,
1090 struct osd_device *o, struct lustre_cfg *cfg)
1092 char *mntdev = lustre_cfg_string(cfg, 1);
1093 char *str = lustre_cfg_string(cfg, 2);
1094 char *svname = lustre_cfg_string(cfg, 4);
1095 time64_t interval = AS_DEFAULT;
1098 bool resetoi = false;
1103 if (o->od_os != NULL)
1106 if (mntdev == NULL || svname == NULL)
1109 rc = strscpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1113 rc = strscpy(o->od_svname, svname, sizeof(o->od_svname));
1117 opts = lustre_cfg_string(cfg, 3);
1119 o->od_index_backup_stop = 0;
1120 o->od_index = -1; /* -1 means index is invalid */
1121 rc = server_name2index(o->od_svname, &o->od_index, NULL);
1122 str = strstr(str, ":");
1124 unsigned long flags;
1126 rc = kstrtoul(str + 1, 10, &flags);
1130 if (test_bit(LMD_FLG_DEV_RDONLY, &flags)) {
1131 o->od_dt_dev.dd_rdonly = 1;
1132 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1136 if (test_bit(LMD_FLG_NOSCRUB, &flags))
1137 interval = AS_NEVER;
1140 if (server_name_is_ost(o->od_svname))
1143 rc = osd_objset_open(o);
1147 o->od_xattr_in_sa = B_TRUE;
1148 o->od_max_blksz = spa_maxblocksize(o->od_os->os_spa);
1149 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
1151 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1154 o->od_root = rootdn->dn_object;
1155 osd_dnode_rele(rootdn);
1157 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1158 &o->od_userused_dn);
1162 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1163 &o->od_groupused_dn);
1167 #ifdef ZFS_PROJINHERIT
1168 if (dmu_objset_projectquota_enabled(o->od_os)) {
1169 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1170 &o->od_projectused_dn);
1171 if (rc && rc != -ENOENT)
1176 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1179 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1181 rc = lu_site_init_finish(&o->od_site);
1185 rc = osd_objset_register_callbacks(o);
1189 if (opts && strstr(opts, "resetoi"))
1192 rc = lprocfs_init_brw_stats(&o->od_brw_stats);
1197 rc = osd_scrub_setup(env, o, interval, resetoi);
1202 rc = osd_procfs_init(o, o->od_svname);
1206 /* currently it's no need to prepare qsd_instance_md for OST */
1207 if (!o->od_is_ost) {
1208 o->od_quota_slave_md = qsd_init(env, o->od_svname,
1209 &o->od_dt_dev, o->od_proc_entry,
1211 if (IS_ERR(o->od_quota_slave_md)) {
1212 rc = PTR_ERR(o->od_quota_slave_md);
1213 o->od_quota_slave_md = NULL;
1218 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
1219 o->od_proc_entry, false, false);
1221 if (IS_ERR(o->od_quota_slave_dt)) {
1222 if (o->od_quota_slave_md != NULL) {
1223 qsd_fini(env, o->od_quota_slave_md);
1224 o->od_quota_slave_md = NULL;
1227 rc = PTR_ERR(o->od_quota_slave_dt);
1228 o->od_quota_slave_dt = NULL;
1232 if (!osd_dmu_userobj_accounting_available(o))
1233 CWARN("%s: dnode accounting not enabled: "
1234 "enable feature@userobj_accounting in pool\n",
1237 /* parse mount option "noacl", and enable ACL by default */
1238 if (opts == NULL || strstr(opts, "noacl") == NULL)
1239 o->od_posix_acl = 1;
1241 osd_unlinked_drain(env, o);
1249 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1253 if (atomic_read(&o->od_zerocopy_alloc))
1254 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1255 atomic_read(&o->od_zerocopy_alloc));
1256 if (atomic_read(&o->od_zerocopy_loan))
1257 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1258 atomic_read(&o->od_zerocopy_loan));
1259 if (atomic_read(&o->od_zerocopy_pin))
1260 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1261 atomic_read(&o->od_zerocopy_pin));
1263 if (o->od_unlinked) {
1264 osd_dnode_rele(o->od_unlinked);
1265 o->od_unlinked = NULL;
1267 if (o->od_userused_dn) {
1268 osd_dnode_rele(o->od_userused_dn);
1269 o->od_userused_dn = NULL;
1271 if (o->od_groupused_dn) {
1272 osd_dnode_rele(o->od_groupused_dn);
1273 o->od_groupused_dn = NULL;
1276 #ifdef ZFS_PROJINHERIT
1277 if (o->od_projectused_dn) {
1278 osd_dnode_rele(o->od_projectused_dn);
1279 o->od_projectused_dn = NULL;
1283 if (o->od_os != NULL) {
1284 if (!o->od_dt_dev.dd_rdonly)
1285 /* force a txg sync to get all commit callbacks */
1286 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1288 /* close the object set */
1289 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1296 static int osd_device_init0(const struct lu_env *env,
1297 struct osd_device *o,
1298 struct lustre_cfg *cfg)
1300 struct lu_device *l = osd2lu_dev(o);
1303 /* if the module was re-loaded, env can loose its keys */
1304 rc = lu_env_refill((struct lu_env *) env);
1308 l->ld_ops = &osd_lu_ops;
1309 o->od_dt_dev.dd_ops = &osd_dt_ops;
1310 sema_init(&o->od_otable_sem, 1);
1311 INIT_LIST_HEAD(&o->od_ios_list);
1312 o->od_sync_on_lseek = B_TRUE;
1314 /* ZFS does not support reporting nonrotional status yet, so this flag
1315 * is only set if explicitly set by the user.
1317 o->od_nonrotational = 0;
1323 static struct lu_device *osd_device_fini(const struct lu_env *env,
1324 struct lu_device *d)
1326 struct osd_device *o = osd_dev(d);
1330 osd_index_backup(env, o, false);
1332 osd_objset_unregister_callbacks(o);
1333 if (!o->od_dt_dev.dd_rdonly) {
1334 osd_sync(env, lu2dt_dev(d));
1336 spa_get_dsl(dmu_objset_spa(o->od_os)));
1340 /* now with all the callbacks completed we can cleanup the remainings */
1341 osd_shutdown(env, o);
1342 osd_scrub_cleanup(env, o);
1344 rc = osd_procfs_fini(o);
1346 CERROR("proc fini error %d\n", rc);
1347 RETURN(ERR_PTR(rc));
1357 static struct lu_device *osd_device_free(const struct lu_env *env,
1358 struct lu_device *d)
1360 struct osd_device *o = osd_dev(d);
1363 /* XXX: make osd top device in order to release reference */
1365 d->ld_site->ls_top_dev = d;
1366 lu_site_purge(env, d->ld_site, -1);
1367 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
1368 D_ERROR, lu_cdebug_printer);
1370 if (o->od_site.ls_bottom_dev)
1371 lu_site_fini(&o->od_site);
1372 dt_device_fini(&o->od_dt_dev);
1378 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1379 struct lu_device_type *type,
1380 struct lustre_cfg *cfg)
1382 struct osd_device *dev;
1383 struct osd_seq_list *osl;
1388 return ERR_PTR(-ENOMEM);
1390 osl = &dev->od_seq_list;
1391 INIT_LIST_HEAD(&osl->osl_seq_list);
1392 rwlock_init(&osl->osl_seq_list_lock);
1393 sema_init(&osl->osl_seq_init_sem, 1);
1394 INIT_LIST_HEAD(&dev->od_index_backup_list);
1395 INIT_LIST_HEAD(&dev->od_index_restore_list);
1396 spin_lock_init(&dev->od_lock);
1397 dev->od_index_backup_policy = LIBP_NONE;
1399 rc = dt_device_init(&dev->od_dt_dev, type);
1401 rc = osd_device_init0(env, dev, cfg);
1403 rc = osd_mount(env, dev, cfg);
1405 osd_device_fini(env, osd2lu_dev(dev));
1406 osd_device_free(env, osd2lu_dev(dev));
1410 dt_device_fini(&dev->od_dt_dev);
1414 if (unlikely(rc != 0) && dev)
1417 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1420 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1421 const char *name, struct lu_device *next)
1427 * To be removed, setup is performed by osd_device_{init,alloc} and
1428 * cleanup is performed by osd_device_{fini,free).
1430 static int osd_process_config(const struct lu_env *env,
1431 struct lu_device *d, struct lustre_cfg *cfg)
1433 struct osd_device *o = osd_dev(d);
1438 switch(cfg->lcfg_command) {
1440 rc = osd_mount(env, o, cfg);
1443 /* For the case LCFG_PRE_CLEANUP is not called in advance,
1444 * that may happend if hit failure during mount process. */
1445 osd_index_backup(env, o, false);
1446 rc = osd_shutdown(env, o);
1449 LASSERT(&o->od_dt_dev);
1450 count = class_modify_config(cfg, PARAM_OSD,
1451 &o->od_dt_dev.dd_kobj);
1453 count = class_modify_config(cfg, PARAM_OST,
1454 &o->od_dt_dev.dd_kobj);
1455 rc = count > 0 ? 0 : count;
1458 case LCFG_PRE_CLEANUP:
1460 osd_index_backup(env, o,
1461 o->od_index_backup_policy != LIBP_NONE);
1471 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1473 struct osd_device *osd = osd_dev(d);
1477 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
1480 /* start qsd instance on recovery completion, this notifies the quota
1481 * slave code that we are about to process new requests now */
1482 rc = qsd_start(env, osd->od_quota_slave_dt);
1483 if (rc == 0 && osd->od_quota_slave_md != NULL)
1484 rc = qsd_start(env, osd->od_quota_slave_md);
1489 * we use exports to track all osd users
1491 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1492 struct obd_device *obd, struct obd_uuid *cluuid,
1493 struct obd_connect_data *data, void *localdata)
1495 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1496 struct lustre_handle conn;
1500 CDEBUG(D_CONFIG, "connect #%d\n", atomic_read(&osd->od_connects));
1502 rc = class_connect(&conn, obd, cluuid);
1506 *exp = class_conn2export(&conn);
1508 atomic_inc(&osd->od_connects);
1514 * once last export (we don't count self-export) disappeared
1515 * osd can be released
1517 static int osd_obd_disconnect(struct obd_export *exp)
1519 struct obd_device *obd = exp->exp_obd;
1520 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1521 int rc, release = 0;
1524 /* Only disconnect the underlying layers on the final disconnect. */
1525 release = atomic_dec_and_test(&osd->od_connects);
1527 rc = class_disconnect(exp); /* bz 9811 */
1529 if (rc == 0 && release)
1530 class_manual_cleanup(obd);
1534 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1536 struct seq_server_site *ss = osd_seq_site(osd);
1540 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1543 if (unlikely(ss == NULL))
1546 OBD_ALLOC_PTR(osd->od_cl_seq);
1547 if (osd->od_cl_seq == NULL)
1550 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1551 osd->od_svname, ss->ss_server_seq);
1553 if (ss->ss_node_id == 0) {
1555 * If the OSD on the sequence controller(MDT0), then allocate
1556 * sequence here, otherwise allocate sequence after connected
1557 * to MDT0 (see mdt_register_lwp_callback()).
1559 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
1560 &osd->od_cl_seq->lcs_space, env);
1566 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1567 struct lu_device *dev)
1569 struct osd_device *osd = osd_dev(dev);
1573 if (osd->od_quota_slave_md != NULL) {
1574 /* set up quota slave objects */
1575 rc = qsd_prepare(env, osd->od_quota_slave_md);
1580 if (osd->od_quota_slave_dt != NULL) {
1581 /* set up quota slave objects */
1582 rc = qsd_prepare(env, osd->od_quota_slave_dt);
1587 rc = osd_fid_init(env, osd);
1593 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
1597 * see include/lu_object.h for the details.
1599 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
1600 struct lu_fid *fid, struct lu_object *parent,
1601 const struct lu_name *name)
1603 struct osd_device *osd = osd_dev(d);
1605 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1608 const struct lu_device_operations osd_lu_ops = {
1609 .ldo_object_alloc = osd_object_alloc,
1610 .ldo_process_config = osd_process_config,
1611 .ldo_recovery_complete = osd_recovery_complete,
1612 .ldo_prepare = osd_prepare,
1613 .ldo_fid_alloc = osd_fid_alloc,
1616 static void osd_type_start(struct lu_device_type *t)
1620 static void osd_type_stop(struct lu_device_type *t)
1624 static const struct lu_device_type_operations osd_device_type_ops = {
1625 .ldto_init = osd_type_init,
1626 .ldto_fini = osd_type_fini,
1628 .ldto_start = osd_type_start,
1629 .ldto_stop = osd_type_stop,
1631 .ldto_device_alloc = osd_device_alloc,
1632 .ldto_device_free = osd_device_free,
1634 .ldto_device_init = osd_device_init,
1635 .ldto_device_fini = osd_device_fini
1638 static struct lu_device_type osd_device_type = {
1639 .ldt_tags = LU_DEVICE_DT,
1640 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1641 .ldt_ops = &osd_device_type_ops,
1642 .ldt_ctx_tags = LCT_LOCAL
1646 static const struct obd_ops osd_obd_device_ops = {
1647 .o_owner = THIS_MODULE,
1648 .o_connect = osd_obd_connect,
1649 .o_disconnect = osd_obd_disconnect,
1652 static int __init osd_init(void)
1656 rc = libcfs_setup();
1660 rc = osd_options_init();
1664 rc = lu_kmem_init(osd_caches);
1668 rc = class_register_type(&osd_obd_device_ops, NULL, true,
1669 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1671 lu_kmem_fini(osd_caches);
1675 static void __exit osd_exit(void)
1677 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1678 lu_kmem_fini(osd_caches);
1681 module_param(osd_oi_count, int, 0444);
1682 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1684 module_param(osd_txg_sync_delay_us, int, 0644);
1685 MODULE_PARM_DESC(osd_txg_sync_delay_us,
1686 "When zero or larger delay N usec instead of doing TXG sync");
1688 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1689 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1690 MODULE_VERSION(LUSTRE_VERSION_STRING);
1691 MODULE_LICENSE("GPL");
1693 module_init(osd_init);
1694 module_exit(osd_exit);