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;
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.
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
221 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
222 if (unlikely(rc != 0)) {
223 /* dmu will call commit callback with error code during abort */
224 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
225 CERROR("%s: failed to start transaction due to ENOSPC: rc = %d\n",
228 CERROR("%s: can't assign tx: rc = %d\n",
231 /* add commit callback */
232 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
234 osd_oti_get(env)->oti_in_trans = 1;
235 lu_device_get(&d->dd_lu_dev);
241 static void osd_unlinked_list_emptify(const struct lu_env *env,
242 struct osd_device *osd,
243 struct list_head *list, bool free)
245 struct osd_object *obj;
248 while (!list_empty(list)) {
249 obj = list_first_entry(list,
250 struct osd_object, oo_unlinked_linkage);
251 LASSERT(obj->oo_dn != NULL);
252 oid = obj->oo_dn->dn_object;
254 list_del_init(&obj->oo_unlinked_linkage);
256 (void)osd_unlinked_object_free(env, osd, oid);
260 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
262 struct dt_txn_commit_cb *dcb;
263 struct dt_txn_commit_cb *tmp;
265 /* call per-transaction stop callbacks if any */
266 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
268 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
269 "commit callback entry: magic=%x name='%s'\n",
270 dcb->dcb_magic, dcb->dcb_name);
271 list_del_init(&dcb->dcb_linkage);
272 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
277 * Concurrency: shouldn't matter.
279 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
282 struct osd_device *osd = osd_dt_dev(th->th_dev);
283 bool sync = (th->th_sync != 0);
284 struct osd_thandle *oh;
290 oh = container_of(th, struct osd_thandle, ot_super);
291 list_splice_init(&oh->ot_unlinked_list, &unlinked);
293 osd_oti_get(env)->oti_ins_cache_depth--;
294 /* reset OI cache for safety */
295 if (osd_oti_get(env)->oti_ins_cache_depth == 0)
296 osd_oti_get(env)->oti_ins_cache_used = 0;
298 if (oh->ot_assigned == 0) {
300 CDEBUG(D_OTHER, "%s: transaction is aborted\n", osd->od_svname);
301 osd_trans_stop_cb(oh, th->th_result);
302 dmu_tx_abort(oh->ot_tx);
303 osd_object_sa_dirty_rele(env, oh);
304 osd_unlinked_list_emptify(env, osd, &unlinked, false);
305 /* there won't be any commit, release reserved quota space now,
308 qsd_op_end(env, osd->od_quota_slave_dt, &oh->ot_quota_trans);
309 if (osd->od_quota_slave_md != NULL)
310 qsd_op_end(env, osd->od_quota_slave_md,
311 &oh->ot_quota_trans);
316 rc = dt_txn_hook_stop(env, th);
318 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
321 osd_trans_stop_cb(oh, rc);
324 txg = oh->ot_tx->tx_txg;
326 osd_object_sa_dirty_rele(env, oh);
327 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
328 * by osd_trans_commit_cb already.
330 dmu_tx_commit(oh->ot_tx);
331 osd_oti_get(env)->oti_in_trans = 0;
333 osd_unlinked_list_emptify(env, osd, &unlinked, true);
336 if (osd_txg_sync_delay_us < 0)
337 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
339 udelay(osd_txg_sync_delay_us);
345 static struct thandle *osd_trans_create(const struct lu_env *env,
346 struct dt_device *dt)
348 struct osd_device *osd = osd_dt_dev(dt);
349 struct osd_thandle *oh;
357 CERROR("%s: someone try to start transaction under readonly mode, should be disabled: rc = %d\n",
358 osd_name(osd_dt_dev(dt)), rc);
363 tx = dmu_tx_create(osd->od_os);
366 CDEBUG(D_OTHER, "%s: dmu_tx_create Failed: rc = %d\n",
367 osd_name(osd_dt_dev(dt)), rc);
371 /* alloc callback data */
376 CDEBUG(D_OTHER, "%s: Allocate memory failed for osd_handle: rc = %d\n",
377 osd_name(osd_dt_dev(dt)), rc);
382 INIT_LIST_HEAD(&oh->ot_dcb_list);
383 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
384 INIT_LIST_HEAD(&oh->ot_unlinked_list);
385 INIT_LIST_HEAD(&oh->ot_sa_list);
386 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
391 osd_oti_get(env)->oti_ins_cache_depth++;
396 /* Estimate the total number of objects from a number of blocks */
397 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
398 uint64_t nrblocks, uint64_t est_maxblockshift)
400 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
403 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
404 * bump the free dnode estimate to assume blocks at least 64KB in
405 * case of a directory-heavy MDT (at 32KB/directory).
407 if (est_maxblockshift < 16) {
408 nrblocks >>= (16 - est_maxblockshift);
409 est_maxblockshift = 16;
413 * Estimate the total number of dnodes from the total blocks count
414 * and the space used per dnode. Since we don't know the overhead
415 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
416 * just using DNODE_SHIFT isn't going to give a good estimate.
417 * Instead, compute the current average space usage per dnode, with
418 * an upper and lower cap to avoid unrealistic estimates..
420 * In case there aren't many dnodes or blocks used yet, add a small
421 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
422 * gradually disappears as the number of real dnodes grows. It also
423 * avoids the need to check for divide-by-zero computing dn_per_block.
425 BUILD_BUG_ON(OSD_DNODE_MIN_BLKSHIFT <= 0);
426 BUILD_BUG_ON(OSD_DNODE_EST_BLKSHIFT <= 0);
428 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
429 usedbytes) >> est_maxblockshift;
430 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
432 if (est_usedobjs <= est_usedblocks) {
434 * Average space/dnode more than maximum block size, use max
435 * block size to estimate free dnodes from adjusted free blocks
436 * count. OSTs typically use multiple blocks per dnode so this
439 est_totobjs = nrblocks;
441 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
443 * Average space/dnode smaller than min dnode size (probably
444 * due to metadnode compression), use min dnode size to
445 * estimate object count. MDTs may use only one block per node
446 * so this case applies.
448 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
452 * Between the extremes, use average space per existing dnode
453 * to compute the number of dnodes that will fit into nrblocks:
455 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
457 * this may overflow 64 bits or become 0 if not handled well.
459 * We know nrblocks is below 2^(64 - blkbits) bits, and
460 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
461 * which means that multiplying them may get as large as
462 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
464 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
465 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
466 * the limit checks above, so we can safely compute this first.
467 * We care more about accuracy on the MDT (many dnodes/block)
468 * which is good because this is where truncation errors are
469 * smallest. Since both nrblocks and dn_per_block are a
470 * function of blkbits, their product is at most:
472 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
474 * so we can safely use 7 bits to compute a fixed-point
475 * fraction and est_totobjs can still fit in 64 bits.
477 unsigned int dn_per_block = (est_usedobjs << 7) /
480 est_totobjs = (nrblocks * dn_per_block) >> 7;
485 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
487 struct objset *os = osd->od_os;
488 uint64_t usedbytes, availbytes, usedobjs, availobjs;
489 uint64_t est_availobjs;
493 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
495 memset(osfs, 0, sizeof(*osfs));
497 /* We're a zfs filesystem. */
498 osfs->os_type = UBERBLOCK_MAGIC;
501 * ZFS allows multiple block sizes. For statfs, Linux makes no
502 * proper distinction between bsize and frsize. For calculations
503 * of free and used blocks incorrectly uses bsize instead of frsize,
504 * but bsize is also used as the optimal blocksize. We return the
505 * largest possible block size as IO size for the optimum performance
506 * and scale the free and used blocks count appropriately.
508 osfs->os_bsize = osd->od_max_blksz;
509 bshift = fls64(osfs->os_bsize) - 1;
511 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
512 osfs->os_bfree = availbytes >> bshift;
513 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
515 /* Take replication (i.e. number of copies) into account */
516 if (os->os_copies != 0)
517 osfs->os_bavail /= os->os_copies;
520 * Reserve some space so we don't run into ENOSPC due to grants not
521 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
522 * Rather than report this via os_bavail (which makes users unhappy if
523 * they can't fill the filesystem 100%), reduce os_blocks as well.
525 * Reserve 0.78% of total space, at least 16MB for small filesystems,
526 * for internal files to be created/unlinked when space is tight.
528 BUILD_BUG_ON(OSD_STATFS_RESERVED_SIZE <= 0);
529 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
530 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
531 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
533 osfs->os_blocks -= reserved;
534 osfs->os_bfree -= min(reserved, osfs->os_bfree);
535 osfs->os_bavail -= min(reserved, osfs->os_bavail);
538 * The availobjs value returned from dmu_objset_space() is largely
539 * useless, since it reports the number of objects that might
540 * theoretically still fit into the dataset, independent of minor
541 * issues like how much space is actually available in the pool.
542 * Compute a better estimate in udmu_objs_count_estimate().
544 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
545 osfs->os_bfree, bshift);
547 osfs->os_ffree = min(availobjs, est_availobjs);
548 osfs->os_files = osfs->os_ffree + usedobjs;
550 /* ZFS XXX: fill in backing dataset FSID/UUID
551 * memcpy(osfs->os_fsid, .... );
554 osfs->os_namelen = MAXNAMELEN;
555 osfs->os_maxbytes = OBD_OBJECT_EOF;
557 if (!spa_writeable(dmu_objset_spa(os)) ||
558 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
559 osfs->os_state |= OS_STATFS_READONLY;
565 * Concurrency: shouldn't matter.
567 int osd_statfs(const struct lu_env *env, struct dt_device *d,
568 struct obd_statfs *osfs, struct obd_statfs_info *info)
570 struct osd_device *osd = osd_dt_dev(d);
574 rc = osd_objset_statfs(osd, osfs);
575 if (unlikely(rc != 0))
578 osfs->os_bavail -= min_t(u64,
579 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
582 /* ZFS does not support reporting nonrotional status yet, so return
583 * flag only if user has set nonrotational.
585 osfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
590 static int osd_blk_insert_cost(struct osd_device *osd)
592 int max_blockshift, nr_blkptrshift, bshift;
594 /* max_blockshift is the log2 of the number of blocks needed to reach
595 * the maximum filesize (that's to say 2^64)
597 bshift = fls64(spa_maxblocksize(dmu_objset_spa(osd->od_os)) - 1);
598 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
600 /* nr_blkptrshift is the log2 of the number of block pointers that can
601 * be stored in an indirect block
603 BUILD_BUG_ON(DN_MAX_INDBLKSHIFT <= SPA_BLKPTRSHIFT);
604 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
606 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
607 * tree. We add +1 for rounding purpose.
608 * The tree depth times the indirect block size gives us the maximum
609 * cost of inserting a block in the tree
611 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
615 * Concurrency: doesn't access mutable data.
617 static void osd_conf_get(const struct lu_env *env,
618 const struct dt_device *dev,
619 struct dt_device_param *param)
621 struct osd_device *osd = osd_dt_dev(dev);
624 * XXX should be taken from not-yet-existing fs abstraction layer.
626 param->ddp_max_name_len = MAXNAMELEN;
627 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
628 param->ddp_symlink_max = PATH_MAX;
629 param->ddp_mount_type = LDD_MT_ZFS;
631 param->ddp_mntopts = MNTOPT_USERXATTR;
632 if (osd->od_posix_acl)
633 param->ddp_mntopts |= MNTOPT_ACL;
634 /* Previously DXATTR_MAX_ENTRY_SIZE */
635 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
637 /* for maxbytes, report same value as ZPL */
638 param->ddp_maxbytes = MAX_LFS_FILESIZE;
640 /* inodes are dynamically allocated, so we report the per-inode space
641 * consumption to upper layers. This static value is not really accurate
642 * and we should use the same logic as in udmu_objset_statfs() to
643 * estimate the real size consumed by an object
645 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
646 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
647 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
648 * not be accounted for every single new block insertion.
649 * Instead, the maximum extent size is set to the number of blocks that
650 * can fit into a single contiguous indirect block. There would be some
651 * cases where this crosses indirect blocks, but it also won't have 7
652 * new levels of indirect blocks in that case either, so it will still
653 * have enough reserved space for the extra indirect block
655 param->ddp_max_extent_blks =
656 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
657 param->ddp_extent_tax = osd_blk_insert_cost(osd);
659 /* Preferred RPC size for efficient disk IO. 1MB shows good
660 * all-around performance for ZFS, but use blocksize (recordsize)
661 * by default if larger to avoid read-modify-write.
663 if (osd->od_max_blksz > ONE_MB_BRW_SIZE)
664 param->ddp_brw_size = osd->od_max_blksz;
666 param->ddp_brw_size = ONE_MB_BRW_SIZE;
668 #ifdef HAVE_DMU_OFFSET_NEXT
669 param->ddp_has_lseek_data_hole = osd->od_sync_on_lseek;
671 param->ddp_has_lseek_data_hole = false;
676 * Concurrency: shouldn't matter.
678 static int osd_sync(const struct lu_env *env, struct dt_device *d)
681 struct osd_device *osd = osd_dt_dev(d);
683 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
684 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
685 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
691 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
693 struct osd_device *osd = osd_dt_dev(dev);
694 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
697 mutex_enter(&tx->tx_sync_lock);
698 txg = tx->tx_open_txg + 1;
699 if (tx->tx_quiesce_txg_waiting < txg) {
700 tx->tx_quiesce_txg_waiting = txg;
701 cv_broadcast(&tx->tx_quiesce_more_cv);
703 mutex_exit(&tx->tx_sync_lock);
709 * Concurrency: shouldn't matter.
711 static int osd_ro(const struct lu_env *env, struct dt_device *d)
713 struct osd_device *osd = osd_dt_dev(d);
716 CERROR("%s: *** setting device %s read-only ***\n",
717 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
718 osd->od_dev_set_rdonly = 1;
719 spa_freeze(dmu_objset_spa(osd->od_os));
724 /* reserve or free quota for some operation */
725 static int osd_reserve_or_free_quota(const struct lu_env *env,
726 struct dt_device *dev,
727 struct lquota_id_info *qi)
729 struct osd_device *osd = osd_dt_dev(dev);
730 struct qsd_instance *qsd = NULL;
736 qsd = osd->od_quota_slave_dt;
738 qsd = osd->od_quota_slave_md;
740 rc = qsd_reserve_or_free_quota(env, qsd, qi);
744 static const struct dt_device_operations osd_dt_ops = {
745 .dt_root_get = osd_root_get,
746 .dt_statfs = osd_statfs,
747 .dt_trans_create = osd_trans_create,
748 .dt_trans_start = osd_trans_start,
749 .dt_trans_stop = osd_trans_stop,
750 .dt_trans_cb_add = osd_trans_cb_add,
751 .dt_conf_get = osd_conf_get,
753 .dt_commit_async = osd_commit_async,
755 .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
759 * DMU OSD device type methods
761 static int osd_type_init(struct lu_device_type *t)
763 LU_CONTEXT_KEY_INIT(&osd_key);
764 return lu_context_key_register(&osd_key);
767 static void osd_type_fini(struct lu_device_type *t)
769 lu_context_key_degister(&osd_key);
772 static void *osd_key_init(const struct lu_context *ctx,
773 struct lu_context_key *key)
775 struct osd_thread_info *info;
779 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
781 info = ERR_PTR(-ENOMEM);
785 static void osd_key_fini(const struct lu_context *ctx,
786 struct lu_context_key *key, void *data)
788 struct osd_thread_info *info = data;
789 struct osd_idmap_cache *idc = info->oti_ins_cache;
792 LASSERT(info->oti_ins_cache_size > 0);
793 OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size);
794 info->oti_ins_cache = NULL;
795 info->oti_ins_cache_size = 0;
797 lu_buf_free(&info->oti_xattr_lbuf);
801 static void osd_key_exit(const struct lu_context *ctx,
802 struct lu_context_key *key, void *data)
806 struct lu_context_key osd_key = {
807 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
808 .lct_init = osd_key_init,
809 .lct_fini = osd_key_fini,
810 .lct_exit = osd_key_exit
813 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
815 if (osd->od_cl_seq == NULL)
818 seq_client_fini(osd->od_cl_seq);
819 OBD_FREE_PTR(osd->od_cl_seq);
820 osd->od_cl_seq = NULL;
823 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
827 /* shutdown quota slave instance associated with the device */
828 if (o->od_quota_slave_md != NULL) {
829 /* complete all in-flight callbacks */
830 osd_sync(env, &o->od_dt_dev);
831 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
832 qsd_fini(env, o->od_quota_slave_md);
833 o->od_quota_slave_md = NULL;
836 if (o->od_quota_slave_dt != NULL) {
837 /* complete all in-flight callbacks */
838 osd_sync(env, &o->od_dt_dev);
839 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
840 qsd_fini(env, o->od_quota_slave_dt);
841 o->od_quota_slave_dt = NULL;
843 osd_fid_fini(env, o);
848 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
850 struct osd_device *osd = arg;
852 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
855 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
857 struct osd_device *osd = arg;
859 LASSERT(newval <= spa_maxblocksize(dmu_objset_spa(osd->od_os)));
860 LASSERT(newval >= SPA_MINBLOCKSIZE);
861 LASSERT(ISP2(newval));
863 osd->od_max_blksz = newval;
866 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
868 struct osd_device *osd = arg;
870 osd->od_prop_rdonly = !!newval;
873 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
875 struct osd_device *osd = arg;
877 osd->od_dnsize = newval;
881 * This function unregisters all registered callbacks. It's harmless to
882 * unregister callbacks that were never registered so it is used to safely
883 * unwind a partially completed call to osd_objset_register_callbacks().
885 static void osd_objset_unregister_callbacks(struct osd_device *o)
887 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
889 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
890 osd_xattr_changed_cb, o);
891 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
892 osd_recordsize_changed_cb, o);
893 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
894 osd_readonly_changed_cb, o);
895 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
896 osd_dnodesize_changed_cb, o);
898 if (o->arc_prune_cb != NULL) {
899 arc_remove_prune_callback(o->arc_prune_cb);
900 o->arc_prune_cb = NULL;
905 * Register the required callbacks to be notified when zfs properties
906 * are modified using the 'zfs(8)' command line utility.
908 static int osd_objset_register_callbacks(struct osd_device *o)
910 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
911 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
917 dsl_pool_config_enter(dp, FTAG);
918 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
919 osd_xattr_changed_cb, o);
923 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
924 osd_recordsize_changed_cb, o);
928 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
929 osd_readonly_changed_cb, o);
933 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
934 osd_dnodesize_changed_cb, o);
938 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
940 dsl_pool_config_exit(dp, FTAG);
942 osd_objset_unregister_callbacks(o);
947 static int osd_objset_open(struct osd_device *o)
949 uint64_t version = ZPL_VERSION;
950 uint64_t sa_obj, unlink_obj;
954 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
955 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
956 B_TRUE, o, &o->od_os);
959 CERROR("%s: can't open %s: rc = %d\n", o->od_svname,
966 /* Check ZFS version */
967 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
968 ZPL_VERSION_STR, 8, 1, &version);
971 CERROR("%s: Error looking up ZPL VERSION: rc = %d\n",
974 * We can't return ENOENT because that would mean the objset
980 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
981 ZFS_SA_ATTRS, 8, 1, &sa_obj);
985 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
986 ZPL_END, &o->z_attr_table);
990 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
991 8, 1, &o->od_rootid);
993 CERROR("%s: lookup for root failed: rc = %d\n",
998 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
1001 CERROR("%s: lookup for %s failed: rc = %d\n",
1002 o->od_svname, ZFS_UNLINKED_SET, rc);
1006 /* Check that user/group usage tracking is supported */
1007 if (!dmu_objset_userused_enabled(o->od_os) ||
1008 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
1009 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
1011 CERROR("%s: Space accounting not supported by this target, aborting: rc = %d\n",
1016 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
1018 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
1024 if (rc != 0 && o->od_os != NULL) {
1025 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1032 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
1035 char *key = osd_oti_get(env)->oti_str;
1039 if (osd->od_dt_dev.dd_rdonly) {
1041 CERROR("%s: someone try to free objects under readonly mode, should be disabled: rc = %d\n",
1047 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
1049 CWARN("%s: Cannot truncate %llu: rc = %d\n",
1050 osd->od_svname, oid, rc);
1054 tx = dmu_tx_create(osd->od_os);
1055 dmu_tx_mark_netfree(tx);
1056 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
1057 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
1059 rc = -dmu_tx_assign(tx, TXG_WAIT);
1061 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
1062 osd->od_svname, oid, rc);
1066 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
1067 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
1068 osd->od_unlinked, key, tx);
1070 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
1071 osd->od_svname, oid, rc);
1075 rc = -dmu_object_free(osd->od_os, oid, tx);
1077 CWARN("%s: Cannot free %llu: rc = %d\n",
1078 osd->od_svname, oid, rc);
1093 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
1096 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1098 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1100 while (zap_cursor_retrieve(&zc, za) == 0) {
1101 /* If cannot free the object, leave it in the unlinked set,
1102 * until the OSD is mounted again when obd_unlinked_drain()
1105 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1107 zap_cursor_advance(&zc);
1110 zap_cursor_fini(&zc);
1113 static int osd_mount(const struct lu_env *env,
1114 struct osd_device *o, struct lustre_cfg *cfg)
1116 char *mntdev = lustre_cfg_string(cfg, 1);
1117 char *str = lustre_cfg_string(cfg, 2);
1118 char *svname = lustre_cfg_string(cfg, 4);
1119 time64_t interval = AS_DEFAULT;
1122 bool resetoi = false;
1126 if (o->od_os != NULL)
1129 if (mntdev == NULL || svname == NULL)
1132 rc = strscpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1136 rc = strscpy(o->od_svname, svname, sizeof(o->od_svname));
1140 opts = lustre_cfg_string(cfg, 3);
1142 o->od_index_backup_stop = 0;
1143 o->od_index = -1; /* -1 means index is invalid */
1144 rc = server_name2index(o->od_svname, &o->od_index, NULL);
1145 str = strstr(str, ":");
1147 unsigned long flags;
1149 rc = kstrtoul(str + 1, 10, &flags);
1153 if (test_bit(LMD_FLG_DEV_RDONLY, &flags)) {
1154 o->od_dt_dev.dd_rdonly = 1;
1155 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1159 if (test_bit(LMD_FLG_NOSCRUB, &flags))
1160 interval = AS_NEVER;
1163 if (server_name_is_ost(o->od_svname))
1166 rc = osd_objset_open(o);
1170 o->od_xattr_in_sa = B_TRUE;
1171 o->od_max_blksz = spa_maxblocksize(o->od_os->os_spa);
1172 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
1174 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1177 o->od_root = rootdn->dn_object;
1178 osd_dnode_rele(rootdn);
1180 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1181 &o->od_userused_dn);
1185 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1186 &o->od_groupused_dn);
1190 #ifdef ZFS_PROJINHERIT
1191 if (dmu_objset_projectquota_enabled(o->od_os)) {
1192 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1193 &o->od_projectused_dn);
1194 if (rc && rc != -ENOENT)
1199 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1202 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1204 rc = lu_site_init_finish(&o->od_site);
1208 rc = osd_objset_register_callbacks(o);
1212 if (opts && strstr(opts, "resetoi"))
1215 rc = lprocfs_init_brw_stats(&o->od_brw_stats);
1220 rc = osd_scrub_setup(env, o, interval, resetoi);
1225 rc = osd_procfs_init(o, o->od_svname);
1229 /* currently it's no need to prepare qsd_instance_md for OST */
1230 if (!o->od_is_ost) {
1231 o->od_quota_slave_md = qsd_init(env, o->od_svname,
1232 &o->od_dt_dev, o->od_proc_entry,
1234 if (IS_ERR(o->od_quota_slave_md)) {
1235 rc = PTR_ERR(o->od_quota_slave_md);
1236 o->od_quota_slave_md = NULL;
1241 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
1242 o->od_proc_entry, false, false);
1244 if (IS_ERR(o->od_quota_slave_dt)) {
1245 if (o->od_quota_slave_md != NULL) {
1246 qsd_fini(env, o->od_quota_slave_md);
1247 o->od_quota_slave_md = NULL;
1250 rc = PTR_ERR(o->od_quota_slave_dt);
1251 o->od_quota_slave_dt = NULL;
1255 if (!osd_dmu_userobj_accounting_available(o))
1256 CWARN("%s: dnode accounting not enabled: enable feature@userobj_accounting in pool\n",
1259 /* parse mount option "noacl", and enable ACL by default */
1260 if (opts == NULL || strstr(opts, "noacl") == NULL)
1261 o->od_posix_acl = 1;
1263 osd_unlinked_drain(env, o);
1271 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1275 if (atomic_read(&o->od_zerocopy_alloc))
1276 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1277 atomic_read(&o->od_zerocopy_alloc));
1278 if (atomic_read(&o->od_zerocopy_loan))
1279 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1280 atomic_read(&o->od_zerocopy_loan));
1281 if (atomic_read(&o->od_zerocopy_pin))
1282 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1283 atomic_read(&o->od_zerocopy_pin));
1285 if (o->od_unlinked) {
1286 osd_dnode_rele(o->od_unlinked);
1287 o->od_unlinked = NULL;
1289 if (o->od_userused_dn) {
1290 osd_dnode_rele(o->od_userused_dn);
1291 o->od_userused_dn = NULL;
1293 if (o->od_groupused_dn) {
1294 osd_dnode_rele(o->od_groupused_dn);
1295 o->od_groupused_dn = NULL;
1298 #ifdef ZFS_PROJINHERIT
1299 if (o->od_projectused_dn) {
1300 osd_dnode_rele(o->od_projectused_dn);
1301 o->od_projectused_dn = NULL;
1305 if (o->od_os != NULL) {
1306 if (!o->od_dt_dev.dd_rdonly)
1307 /* force a txg sync to get all commit callbacks */
1308 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1310 /* close the object set */
1311 osd_dmu_objset_disown(o->od_os, B_TRUE, o);
1318 static int osd_device_init0(const struct lu_env *env,
1319 struct osd_device *o,
1320 struct lustre_cfg *cfg)
1322 struct lu_device *l = osd2lu_dev(o);
1325 /* if the module was re-loaded, env can loose its keys */
1326 rc = lu_env_refill((struct lu_env *) env);
1330 l->ld_ops = &osd_lu_ops;
1331 o->od_dt_dev.dd_ops = &osd_dt_ops;
1332 sema_init(&o->od_otable_sem, 1);
1333 INIT_LIST_HEAD(&o->od_ios_list);
1334 o->od_sync_on_lseek = B_TRUE;
1336 /* ZFS does not support reporting nonrotional status yet, so this flag
1337 * is only set if explicitly set by the user.
1339 o->od_nonrotational = 0;
1345 static struct lu_device *osd_device_fini(const struct lu_env *env,
1346 struct lu_device *d)
1348 struct osd_device *o = osd_dev(d);
1352 osd_index_backup(env, o, false);
1354 osd_objset_unregister_callbacks(o);
1355 if (!o->od_dt_dev.dd_rdonly) {
1356 osd_sync(env, lu2dt_dev(d));
1358 spa_get_dsl(dmu_objset_spa(o->od_os)));
1362 /* now with all the callbacks completed we can cleanup the remainings */
1363 osd_shutdown(env, o);
1364 osd_scrub_cleanup(env, o);
1366 rc = osd_procfs_fini(o);
1368 CERROR("proc fini error %d\n", rc);
1369 RETURN(ERR_PTR(rc));
1379 static struct lu_device *osd_device_free(const struct lu_env *env,
1380 struct lu_device *d)
1382 struct osd_device *o = osd_dev(d);
1385 /* XXX: make osd top device in order to release reference */
1387 d->ld_site->ls_top_dev = d;
1388 lu_site_purge(env, d->ld_site, -1);
1389 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
1390 D_ERROR, lu_cdebug_printer);
1392 if (o->od_site.ls_bottom_dev)
1393 lu_site_fini(&o->od_site);
1394 dt_device_fini(&o->od_dt_dev);
1400 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1401 struct lu_device_type *type,
1402 struct lustre_cfg *cfg)
1404 struct osd_device *dev;
1405 struct osd_seq_list *osl;
1410 return ERR_PTR(-ENOMEM);
1412 osl = &dev->od_seq_list;
1413 INIT_LIST_HEAD(&osl->osl_seq_list);
1414 rwlock_init(&osl->osl_seq_list_lock);
1415 sema_init(&osl->osl_seq_init_sem, 1);
1416 INIT_LIST_HEAD(&dev->od_index_backup_list);
1417 INIT_LIST_HEAD(&dev->od_index_restore_list);
1418 spin_lock_init(&dev->od_lock);
1419 dev->od_index_backup_policy = LIBP_NONE;
1421 rc = dt_device_init(&dev->od_dt_dev, type);
1423 rc = osd_device_init0(env, dev, cfg);
1425 rc = osd_mount(env, dev, cfg);
1427 osd_device_fini(env, osd2lu_dev(dev));
1428 osd_device_free(env, osd2lu_dev(dev));
1432 dt_device_fini(&dev->od_dt_dev);
1436 if (unlikely(rc != 0) && dev)
1439 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1442 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1443 const char *name, struct lu_device *next)
1449 * To be removed, setup is performed by osd_device_{init,alloc} and
1450 * cleanup is performed by osd_device_{fini,free).
1452 static int osd_process_config(const struct lu_env *env,
1453 struct lu_device *d, struct lustre_cfg *cfg)
1455 struct osd_device *o = osd_dev(d);
1460 switch (cfg->lcfg_command) {
1462 rc = osd_mount(env, o, cfg);
1465 /* For the case LCFG_PRE_CLEANUP is not called in advance,
1466 * that may happend if hit failure during mount process.
1468 osd_index_backup(env, o, false);
1469 rc = osd_shutdown(env, o);
1472 LASSERT(&o->od_dt_dev);
1473 count = class_modify_config(cfg, PARAM_OSD,
1474 &o->od_dt_dev.dd_kobj);
1476 count = class_modify_config(cfg, PARAM_OST,
1477 &o->od_dt_dev.dd_kobj);
1478 rc = count > 0 ? 0 : count;
1481 case LCFG_PRE_CLEANUP:
1483 osd_index_backup(env, o,
1484 o->od_index_backup_policy != LIBP_NONE);
1494 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1496 struct osd_device *osd = osd_dev(d);
1500 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
1503 /* start qsd instance on recovery completion, this notifies the quota
1504 * slave code that we are about to process new requests now
1506 rc = qsd_start(env, osd->od_quota_slave_dt);
1507 if (rc == 0 && osd->od_quota_slave_md != NULL)
1508 rc = qsd_start(env, osd->od_quota_slave_md);
1513 * we use exports to track all osd users
1515 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1516 struct obd_device *obd, struct obd_uuid *cluuid,
1517 struct obd_connect_data *data, void *localdata)
1519 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1520 struct lustre_handle conn;
1524 CDEBUG(D_CONFIG, "connect #%d\n", atomic_read(&osd->od_connects));
1526 rc = class_connect(&conn, obd, cluuid);
1530 *exp = class_conn2export(&conn);
1532 atomic_inc(&osd->od_connects);
1538 * once last export (we don't count self-export) disappeared
1539 * osd can be released
1541 static int osd_obd_disconnect(struct obd_export *exp)
1543 struct obd_device *obd = exp->exp_obd;
1544 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1545 int rc, release = 0;
1548 /* Only disconnect the underlying layers on the final disconnect. */
1549 release = atomic_dec_and_test(&osd->od_connects);
1551 rc = class_disconnect(exp); /* bz 9811 */
1553 if (rc == 0 && release)
1554 class_manual_cleanup(obd);
1558 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1560 struct seq_server_site *ss = osd_seq_site(osd);
1564 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1567 if (unlikely(ss == NULL))
1570 OBD_ALLOC_PTR(osd->od_cl_seq);
1571 if (osd->od_cl_seq == NULL)
1574 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1575 osd->od_svname, ss->ss_server_seq);
1577 if (ss->ss_node_id == 0) {
1579 * If the OSD on the sequence controller(MDT0), then allocate
1580 * sequence here, otherwise allocate sequence after connected
1581 * to MDT0 (see mdt_register_lwp_callback()).
1583 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
1584 &osd->od_cl_seq->lcs_space, env);
1590 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1591 struct lu_device *dev)
1593 struct osd_device *osd = osd_dev(dev);
1597 if (osd->od_quota_slave_md != NULL) {
1598 /* set up quota slave objects */
1599 rc = qsd_prepare(env, osd->od_quota_slave_md);
1604 if (osd->od_quota_slave_dt != NULL) {
1605 /* set up quota slave objects */
1606 rc = qsd_prepare(env, osd->od_quota_slave_dt);
1611 rc = osd_fid_init(env, osd);
1617 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
1621 * see include/lu_object.h for the details.
1623 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
1624 struct lu_fid *fid, struct lu_object *parent,
1625 const struct lu_name *name)
1627 struct osd_device *osd = osd_dev(d);
1629 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1632 const struct lu_device_operations osd_lu_ops = {
1633 .ldo_object_alloc = osd_object_alloc,
1634 .ldo_process_config = osd_process_config,
1635 .ldo_recovery_complete = osd_recovery_complete,
1636 .ldo_prepare = osd_prepare,
1637 .ldo_fid_alloc = osd_fid_alloc,
1640 static void osd_type_start(struct lu_device_type *t)
1644 static void osd_type_stop(struct lu_device_type *t)
1648 static const struct lu_device_type_operations osd_device_type_ops = {
1649 .ldto_init = osd_type_init,
1650 .ldto_fini = osd_type_fini,
1652 .ldto_start = osd_type_start,
1653 .ldto_stop = osd_type_stop,
1655 .ldto_device_alloc = osd_device_alloc,
1656 .ldto_device_free = osd_device_free,
1658 .ldto_device_init = osd_device_init,
1659 .ldto_device_fini = osd_device_fini
1662 static struct lu_device_type osd_device_type = {
1663 .ldt_tags = LU_DEVICE_DT,
1664 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1665 .ldt_ops = &osd_device_type_ops,
1666 .ldt_ctx_tags = LCT_LOCAL
1670 static const struct obd_ops osd_obd_device_ops = {
1671 .o_owner = THIS_MODULE,
1672 .o_connect = osd_obd_connect,
1673 .o_disconnect = osd_obd_disconnect,
1676 static int __init osd_init(void)
1680 rc = libcfs_setup();
1684 rc = osd_options_init();
1688 rc = lu_kmem_init(osd_caches);
1692 rc = class_register_type(&osd_obd_device_ops, NULL, true,
1693 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1695 lu_kmem_fini(osd_caches);
1699 static void __exit osd_exit(void)
1701 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1702 lu_kmem_fini(osd_caches);
1705 module_param(osd_oi_count, int, 0444);
1706 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1708 module_param(osd_txg_sync_delay_us, int, 0644);
1709 MODULE_PARM_DESC(osd_txg_sync_delay_us,
1710 "When zero or larger delay N usec instead of doing TXG sync");
1712 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1713 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1714 MODULE_VERSION(LUSTRE_VERSION_STRING);
1715 MODULE_LICENSE("GPL");
1717 module_init(osd_init);
1718 module_exit(osd_exit);