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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osd-zfs/osd_handler.c
37 * Top-level entry points into osd module
39 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
40 * Author: Mike Pershin <tappro@whamcloud.com>
41 * Author: Johann Lombardi <johann@whamcloud.com>
44 #define DEBUG_SUBSYSTEM S_OSD
46 #include <lustre_ver.h>
47 #include <libcfs/libcfs.h>
48 #include <obd_support.h>
49 #include <lustre_net.h>
51 #include <obd_class.h>
52 #include <lustre_disk.h>
53 #include <lustre_fid.h>
54 #include <lustre_param.h>
55 #include <md_object.h>
57 #include "osd_internal.h"
59 #include <sys/dnode.h>
64 #include <sys/spa_impl.h>
65 #include <sys/zfs_znode.h>
66 #include <sys/dmu_tx.h>
67 #include <sys/dmu_objset.h>
68 #include <sys/dsl_prop.h>
69 #include <sys/sa_impl.h>
72 struct lu_context_key osd_key;
74 /* Slab for OSD object allocation */
75 struct kmem_cache *osd_object_kmem;
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)
88 static void arc_prune_func(int64_t bytes, void *private)
90 struct osd_device *od = private;
91 struct lu_site *site = &od->od_site;
95 rc = lu_env_init(&env, LCT_SHRINKER);
97 CERROR("%s: can't initialize shrinker env: rc = %d\n",
102 lu_site_purge(&env, site, (bytes >> 10));
108 * Concurrency: doesn't access mutable data
110 static int osd_root_get(const struct lu_env *env,
111 struct dt_device *dev, struct lu_fid *f)
113 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
118 * OSD object methods.
122 * Concurrency: shouldn't matter.
124 static void osd_trans_commit_cb(void *cb_data, int error)
126 struct osd_thandle *oh = cb_data;
127 struct thandle *th = &oh->ot_super;
128 struct osd_device *osd = osd_dt_dev(th->th_dev);
129 struct lu_device *lud = &th->th_dev->dd_lu_dev;
130 struct dt_txn_commit_cb *dcb, *tmp;
135 if (error == ECANCELED)
136 CWARN("%s: transaction @0x%p was aborted\n",
137 osd_dt_dev(th->th_dev)->od_svname, th);
139 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
140 osd_dt_dev(th->th_dev)->od_svname, th, error);
143 dt_txn_hook_commit(th);
145 /* call per-transaction callbacks if any */
146 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage)
147 dcb->dcb_func(NULL, th, dcb, error);
149 /* Unlike ldiskfs, zfs updates space accounting at commit time.
150 * As a consequence, op_end is called only now to inform the quota slave
151 * component that reserved quota space is now accounted in usage and
152 * should be released. Quota space won't be adjusted at this point since
153 * we can't provide a suitable environment. It will be performed
154 * asynchronously by a lquota thread. */
155 qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);
159 lu_context_exit(&th->th_ctx);
160 lu_context_fini(&th->th_ctx);
161 thandle_put(&oh->ot_super);
166 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
168 struct osd_thandle *oh;
170 oh = container_of0(th, struct osd_thandle, ot_super);
171 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
177 * Concurrency: shouldn't matter.
179 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
182 struct osd_thandle *oh;
186 oh = container_of0(th, struct osd_thandle, ot_super);
190 rc = dt_txn_hook_start(env, d, th);
194 if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
195 /* Unlike ldiskfs, ZFS checks for available space and returns
196 * -ENOSPC when assigning txg */
199 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
200 if (unlikely(rc != 0)) {
201 struct osd_device *osd = osd_dt_dev(d);
202 /* dmu will call commit callback with error code during abort */
203 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
204 CERROR("%s: failed to start transaction due to ENOSPC. "
205 "Metadata overhead is underestimated or "
206 "grant_ratio is too low.\n", osd->od_svname);
208 CERROR("%s: can't assign tx: rc = %d\n",
211 /* add commit callback */
212 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
214 lu_context_init(&th->th_ctx, th->th_tags);
215 lu_context_enter(&th->th_ctx);
216 lu_device_get(&d->dd_lu_dev);
223 * Concurrency: shouldn't matter.
225 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
228 struct osd_device *osd = osd_dt_dev(th->th_dev);
229 struct osd_thandle *oh;
234 oh = container_of0(th, struct osd_thandle, ot_super);
236 if (oh->ot_assigned == 0) {
238 dmu_tx_abort(oh->ot_tx);
239 osd_object_sa_dirty_rele(oh);
240 /* there won't be any commit, release reserved quota space now,
242 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
243 thandle_put(&oh->ot_super);
247 /* When doing our own inode accounting, the ZAPs storing per-uid/gid
248 * usage are updated at operation execution time, so we should call
249 * qsd_op_end() straight away. Otherwise (for blk accounting maintained
250 * by ZFS and when #inode is estimated from #blks) accounting is updated
251 * at commit time and the call to qsd_op_end() must be delayed */
252 if (oh->ot_quota_trans.lqt_id_cnt > 0 &&
253 !oh->ot_quota_trans.lqt_ids[0].lqi_is_blk &&
254 !osd->od_quota_iused_est)
255 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
257 rc = dt_txn_hook_stop(env, th);
259 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
263 txg = oh->ot_tx->tx_txg;
265 osd_object_sa_dirty_rele(oh);
266 dmu_tx_commit(oh->ot_tx);
269 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
274 static struct thandle *osd_trans_create(const struct lu_env *env,
275 struct dt_device *dt)
277 struct osd_device *osd = osd_dt_dev(dt);
278 struct osd_thandle *oh;
283 tx = dmu_tx_create(osd->od_os);
285 RETURN(ERR_PTR(-ENOMEM));
287 /* alloc callback data */
291 RETURN(ERR_PTR(-ENOMEM));
295 INIT_LIST_HEAD(&oh->ot_dcb_list);
296 INIT_LIST_HEAD(&oh->ot_sa_list);
297 sema_init(&oh->ot_sa_lock, 1);
298 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
302 th->th_tags = LCT_TX_HANDLE;
303 atomic_set(&th->th_refc, 1);
304 th->th_alloc_size = sizeof(*oh);
308 /* Estimate the number of objects from a number of blocks */
309 uint64_t osd_objs_count_estimate(uint64_t refdbytes, uint64_t usedobjs,
312 uint64_t est_objs, est_refdblocks, est_usedobjs;
314 /* Compute an nrblocks estimate based on the actual number of
315 * dnodes that could fit in the space. Since we don't know the
316 * overhead associated with each dnode (xattrs, SAs, VDEV overhead,
317 * etc) just using DNODE_SHIFT isn't going to give a good estimate.
318 * Instead, compute an estimate based on the average space usage per
319 * dnode, with an upper and lower cap.
321 * In case there aren't many dnodes or blocks used yet, add a small
322 * correction factor using OSD_DNODE_EST_SHIFT. This correction
323 * factor gradually disappears as the number of real dnodes grows.
324 * This also avoids the need to check for divide-by-zero later.
326 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
327 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
329 est_refdblocks = (refdbytes >> SPA_MAXBLOCKSHIFT) +
330 (OSD_DNODE_EST_COUNT >> OSD_DNODE_EST_BLKSHIFT);
331 est_usedobjs = usedobjs + OSD_DNODE_EST_COUNT;
333 /* Average space/dnode more than maximum dnode size, use max dnode
334 * size to estimate free dnodes from adjusted free blocks count.
335 * OSTs typically use more than one block dnode so this case applies. */
336 if (est_usedobjs <= est_refdblocks * 2) {
339 /* Average space/dnode smaller than min dnode size (probably due to
340 * metadnode compression), use min dnode size to estimate the number of
342 * An MDT typically uses below 512 bytes/dnode so this case applies. */
343 } else if (est_usedobjs >= (est_refdblocks << OSD_DNODE_MIN_BLKSHIFT)) {
344 est_objs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
346 /* Between the extremes, we try to use the average size of
347 * existing dnodes to compute the number of dnodes that fit
350 * est_objs = nrblocks * (est_usedobjs / est_refblocks);
352 * but this may overflow 64 bits or become 0 if not handled well
354 * We know nrblocks is below (64 - 17 = 47) bits from
355 * SPA_MAXBLKSHIFT, and est_usedobjs is under 48 bits due to
356 * DN_MAX_OBJECT_SHIFT, which means that multiplying them may
357 * get as large as 2 ^ 95.
359 * We also know (est_usedobjs / est_refdblocks) is between 2 and
360 * 256, due to above checks, we can safely compute this first.
361 * We care more about accuracy on the MDT (many dnodes/block)
362 * which is good because this is where truncation errors are
363 * smallest. This adds 8 bits to nrblocks so we can use 7 bits
364 * to compute a fixed-point fraction and nrblocks can still fit
367 unsigned dnodes_per_block = (est_usedobjs << 7)/est_refdblocks;
369 est_objs = (nrblocks * dnodes_per_block) >> 7;
374 static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs)
376 uint64_t refdbytes, availbytes, usedobjs, availobjs;
377 uint64_t est_availobjs;
380 dmu_objset_space(os, &refdbytes, &availbytes, &usedobjs,
384 * ZFS allows multiple block sizes. For statfs, Linux makes no
385 * proper distinction between bsize and frsize. For calculations
386 * of free and used blocks incorrectly uses bsize instead of frsize,
387 * but bsize is also used as the optimal blocksize. We return the
388 * largest possible block size as IO size for the optimum performance
389 * and scale the free and used blocks count appropriately.
391 osfs->os_bsize = 1ULL << SPA_MAXBLOCKSHIFT;
393 osfs->os_blocks = (refdbytes + availbytes) >> SPA_MAXBLOCKSHIFT;
394 osfs->os_bfree = availbytes >> SPA_MAXBLOCKSHIFT;
395 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
397 /* Take replication (i.e. number of copies) into account */
398 osfs->os_bavail /= os->os_copies;
401 * Reserve some space so we don't run into ENOSPC due to grants not
402 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
403 * Rather than report this via os_bavail (which makes users unhappy if
404 * they can't fill the filesystem 100%), reduce os_blocks as well.
406 * Reserve 0.78% of total space, at least 4MB for small filesystems,
407 * for internal files to be created/unlinked when space is tight.
409 CLASSERT(OSD_STATFS_RESERVED_BLKS > 0);
410 if (likely(osfs->os_blocks >=
411 OSD_STATFS_RESERVED_BLKS << OSD_STATFS_RESERVED_SHIFT))
412 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
414 reserved = OSD_STATFS_RESERVED_BLKS;
416 osfs->os_blocks -= reserved;
417 osfs->os_bfree -= MIN(reserved, osfs->os_bfree);
418 osfs->os_bavail -= MIN(reserved, osfs->os_bavail);
421 * The availobjs value returned from dmu_objset_space() is largely
422 * useless, since it reports the number of objects that might
423 * theoretically still fit into the dataset, independent of minor
424 * issues like how much space is actually available in the pool.
425 * Compute a better estimate in udmu_objs_count_estimate().
427 est_availobjs = osd_objs_count_estimate(refdbytes, usedobjs,
430 osfs->os_ffree = min(availobjs, est_availobjs);
431 osfs->os_files = osfs->os_ffree + usedobjs;
433 /* ZFS XXX: fill in backing dataset FSID/UUID
434 memcpy(osfs->os_fsid, .... );*/
436 /* We're a zfs filesystem. */
437 osfs->os_type = UBERBLOCK_MAGIC;
439 /* ZFS XXX: fill in appropriate OS_STATE_{DEGRADED,READONLY} flags
440 osfs->os_state = vf_to_stf(vfsp->vfs_flag);
441 if (sb->s_flags & MS_RDONLY)
442 osfs->os_state = OS_STATE_READONLY;
445 osfs->os_namelen = MAXNAMELEN;
446 osfs->os_maxbytes = OBD_OBJECT_EOF;
452 * Concurrency: shouldn't matter.
454 int osd_statfs(const struct lu_env *env, struct dt_device *d,
455 struct obd_statfs *osfs)
457 struct osd_device *osd = osd_dt_dev(d);
461 rc = osd_objset_statfs(osd->od_os, osfs);
462 if (unlikely(rc != 0))
465 osfs->os_bavail -= min_t(u64,
466 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
471 static int osd_blk_insert_cost(void)
473 int max_blockshift, nr_blkptrshift;
475 /* max_blockshift is the log2 of the number of blocks needed to reach
476 * the maximum filesize (that's to say 2^64) */
477 max_blockshift = DN_MAX_OFFSET_SHIFT - SPA_MAXBLOCKSHIFT;
479 /* nr_blkptrshift is the log2 of the number of block pointers that can
480 * be stored in an indirect block */
481 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
482 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
484 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
485 * tree. We add +1 for rounding purpose.
486 * The tree depth times the indirect block size gives us the maximum
487 * cost of inserting a block in the tree */
488 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
492 * Concurrency: doesn't access mutable data.
494 static void osd_conf_get(const struct lu_env *env,
495 const struct dt_device *dev,
496 struct dt_device_param *param)
498 struct osd_device *osd = osd_dt_dev(dev);
501 * XXX should be taken from not-yet-existing fs abstraction layer.
503 param->ddp_max_name_len = MAXNAMELEN;
504 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
505 param->ddp_block_shift = 12; /* XXX */
506 param->ddp_mount_type = LDD_MT_ZFS;
508 param->ddp_mntopts = MNTOPT_USERXATTR;
509 if (osd->od_posix_acl)
510 param->ddp_mntopts |= MNTOPT_ACL;
511 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
513 /* for maxbytes, report same value as ZPL */
514 param->ddp_maxbytes = MAX_LFS_FILESIZE;
516 /* Default reserved fraction of the available space that should be kept
517 * for error margin. Unfortunately, there are many factors that can
518 * impact the overhead with zfs, so let's be very cautious for now and
519 * reserve 20% of the available space which is not given out as grant.
520 * This tunable can be changed on a live system via procfs if needed. */
521 param->ddp_grant_reserved = 20;
523 /* inodes are dynamically allocated, so we report the per-inode space
524 * consumption to upper layers. This static value is not really accurate
525 * and we should use the same logic as in udmu_objset_statfs() to
526 * estimate the real size consumed by an object */
527 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
528 /* per-fragment overhead to be used by the client code */
529 param->ddp_grant_frag = osd_blk_insert_cost();
533 * Concurrency: shouldn't matter.
535 static int osd_sync(const struct lu_env *env, struct dt_device *d)
537 struct osd_device *osd = osd_dt_dev(d);
538 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
539 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
540 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
544 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
546 struct osd_device *osd = osd_dt_dev(dev);
547 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
550 mutex_enter(&tx->tx_sync_lock);
551 txg = tx->tx_open_txg + 1;
552 if (tx->tx_quiesce_txg_waiting < txg) {
553 tx->tx_quiesce_txg_waiting = txg;
554 cv_broadcast(&tx->tx_quiesce_more_cv);
556 mutex_exit(&tx->tx_sync_lock);
562 * Concurrency: shouldn't matter.
564 static int osd_ro(const struct lu_env *env, struct dt_device *d)
566 struct osd_device *osd = osd_dt_dev(d);
569 CERROR("%s: *** setting device %s read-only ***\n",
570 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
572 spa_freeze(dmu_objset_spa(osd->od_os));
578 * Concurrency: serialization provided by callers.
580 static int osd_init_capa_ctxt(const struct lu_env *env, struct dt_device *d,
581 int mode, unsigned long timeout, __u32 alg,
582 struct lustre_capa_key *keys)
584 struct osd_device *dev = osd_dt_dev(d);
587 dev->od_fl_capa = mode;
588 dev->od_capa_timeout = timeout;
589 dev->od_capa_alg = alg;
590 dev->od_capa_keys = keys;
595 static struct dt_device_operations osd_dt_ops = {
596 .dt_root_get = osd_root_get,
597 .dt_statfs = osd_statfs,
598 .dt_trans_create = osd_trans_create,
599 .dt_trans_start = osd_trans_start,
600 .dt_trans_stop = osd_trans_stop,
601 .dt_trans_cb_add = osd_trans_cb_add,
602 .dt_conf_get = osd_conf_get,
604 .dt_commit_async = osd_commit_async,
606 .dt_init_capa_ctxt = osd_init_capa_ctxt,
610 * DMU OSD device type methods
612 static int osd_type_init(struct lu_device_type *t)
614 LU_CONTEXT_KEY_INIT(&osd_key);
615 return lu_context_key_register(&osd_key);
618 static void osd_type_fini(struct lu_device_type *t)
620 lu_context_key_degister(&osd_key);
623 static void *osd_key_init(const struct lu_context *ctx,
624 struct lu_context_key *key)
626 struct osd_thread_info *info;
630 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
632 info = ERR_PTR(-ENOMEM);
636 static void osd_key_fini(const struct lu_context *ctx,
637 struct lu_context_key *key, void *data)
639 struct osd_thread_info *info = data;
644 static void osd_key_exit(const struct lu_context *ctx,
645 struct lu_context_key *key, void *data)
647 struct osd_thread_info *info = data;
649 memset(info, 0, sizeof(*info));
652 struct lu_context_key osd_key = {
653 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
654 .lct_init = osd_key_init,
655 .lct_fini = osd_key_fini,
656 .lct_exit = osd_key_exit
659 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
663 /* shutdown quota slave instance associated with the device */
664 if (o->od_quota_slave != NULL) {
665 qsd_fini(env, o->od_quota_slave);
666 o->od_quota_slave = NULL;
672 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
674 struct osd_device *osd = arg;
676 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
679 static int osd_objset_open(struct osd_device *o)
681 uint64_t version = ZPL_VERSION;
686 rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os);
692 /* Check ZFS version */
693 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
694 ZPL_VERSION_STR, 8, 1, &version);
696 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
698 * We can't return ENOENT because that would mean the objset
701 GOTO(out, rc = -EIO);
704 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
705 ZFS_SA_ATTRS, 8, 1, &sa_obj);
709 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
710 ZPL_END, &o->z_attr_table);
714 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
715 8, 1, &o->od_rootid);
717 CERROR("%s: lookup for root failed: rc = %d\n",
722 /* Check that user/group usage tracking is supported */
723 if (!dmu_objset_userused_enabled(o->od_os) ||
724 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
725 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
726 CERROR("%s: Space accounting not supported by this target, "
727 "aborting\n", o->od_svname);
728 GOTO(out, -ENOTSUPP);
732 if (rc != 0 && o->od_os != NULL)
733 dmu_objset_disown(o->od_os, o);
738 static int osd_mount(const struct lu_env *env,
739 struct osd_device *o, struct lustre_cfg *cfg)
741 struct dsl_dataset *ds;
742 char *mntdev = lustre_cfg_string(cfg, 1);
743 char *svname = lustre_cfg_string(cfg, 4);
750 if (o->od_os != NULL)
753 if (mntdev == NULL || svname == NULL)
756 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
757 if (rc >= sizeof(o->od_mntdev))
760 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
761 if (rc >= sizeof(o->od_svname))
764 if (server_name_is_ost(o->od_svname))
767 rc = osd_objset_open(o);
769 CERROR("%s: can't open objset %s: rc = %d\n", o->od_svname,
774 ds = dmu_objset_ds(o->od_os);
775 dp = dmu_objset_pool(o->od_os);
778 dsl_pool_config_enter(dp, FTAG);
779 rc = dsl_prop_register(ds, "xattr", osd_xattr_changed_cb, o);
780 dsl_pool_config_exit(dp, FTAG);
782 CWARN("%s: can't register xattr callback, ignore: rc=%d\n",
785 rc = __osd_obj2dbuf(env, o->od_os, o->od_rootid, &rootdb);
787 CERROR("%s: obj2dbuf() failed: rc = %d\n", o->od_svname, rc);
788 dmu_objset_disown(o->od_os, o);
793 o->od_root = rootdb->db_object;
794 sa_buf_rele(rootdb, osd_obj_tag);
796 /* 1. initialize oi before any file create or file open */
797 rc = osd_oi_init(env, o);
801 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
804 o->od_site.ls_bottom_dev = osd2lu_dev(o);
806 rc = lu_site_init_finish(&o->od_site);
810 rc = osd_convert_root_to_new_seq(env, o);
814 /* Use our own ZAP for inode accounting by default, this can be changed
815 * via procfs to estimate the inode usage from the block usage */
816 o->od_quota_iused_est = 0;
818 rc = osd_procfs_init(o, o->od_svname);
822 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
824 /* initialize quota slave instance */
825 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
827 if (IS_ERR(o->od_quota_slave)) {
828 rc = PTR_ERR(o->od_quota_slave);
829 o->od_quota_slave = NULL;
833 /* parse mount option "noacl", and enable ACL by default */
834 opts = lustre_cfg_string(cfg, 3);
835 if (opts == NULL || strstr(opts, "noacl") == NULL)
842 static void osd_umount(const struct lu_env *env, struct osd_device *o)
846 if (atomic_read(&o->od_zerocopy_alloc))
847 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
848 atomic_read(&o->od_zerocopy_alloc));
849 if (atomic_read(&o->od_zerocopy_loan))
850 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
851 atomic_read(&o->od_zerocopy_loan));
852 if (atomic_read(&o->od_zerocopy_pin))
853 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
854 atomic_read(&o->od_zerocopy_pin));
856 if (o->od_os != NULL) {
857 /* force a txg sync to get all commit callbacks */
858 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
860 /* close the object set */
861 dmu_objset_disown(o->od_os, o);
869 static int osd_device_init0(const struct lu_env *env,
870 struct osd_device *o,
871 struct lustre_cfg *cfg)
873 struct lu_device *l = osd2lu_dev(o);
876 /* if the module was re-loaded, env can loose its keys */
877 rc = lu_env_refill((struct lu_env *) env);
881 l->ld_ops = &osd_lu_ops;
882 o->od_dt_dev.dd_ops = &osd_dt_ops;
884 o->od_capa_hash = init_capa_hash();
885 if (o->od_capa_hash == NULL)
886 GOTO(out, rc = -ENOMEM);
892 static struct lu_device *osd_device_fini(const struct lu_env *env,
893 struct lu_device *dev);
895 static struct lu_device *osd_device_alloc(const struct lu_env *env,
896 struct lu_device_type *type,
897 struct lustre_cfg *cfg)
899 struct osd_device *dev;
904 return ERR_PTR(-ENOMEM);
906 rc = dt_device_init(&dev->od_dt_dev, type);
908 rc = osd_device_init0(env, dev, cfg);
910 rc = osd_mount(env, dev, cfg);
912 osd_device_fini(env, osd2lu_dev(dev));
915 dt_device_fini(&dev->od_dt_dev);
918 if (unlikely(rc != 0))
921 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
924 static struct lu_device *osd_device_free(const struct lu_env *env,
927 struct osd_device *o = osd_dev(d);
930 cleanup_capa_hash(o->od_capa_hash);
931 /* XXX: make osd top device in order to release reference */
932 d->ld_site->ls_top_dev = d;
933 lu_site_purge(env, d->ld_site, -1);
934 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
935 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
936 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
938 lu_site_fini(&o->od_site);
939 dt_device_fini(&o->od_dt_dev);
945 static struct lu_device *osd_device_fini(const struct lu_env *env,
948 struct osd_device *o = osd_dev(d);
949 struct dsl_dataset *ds;
954 osd_shutdown(env, o);
958 ds = dmu_objset_ds(o->od_os);
959 rc = dsl_prop_unregister(ds, "xattr", osd_xattr_changed_cb, o);
961 CERROR("%s: dsl_prop_unregister xattr error %d\n",
963 if (o->arc_prune_cb != NULL) {
964 arc_remove_prune_callback(o->arc_prune_cb);
965 o->arc_prune_cb = NULL;
967 osd_sync(env, lu2dt_dev(d));
968 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
971 rc = osd_procfs_fini(o);
973 CERROR("proc fini error %d\n", rc);
983 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
984 const char *name, struct lu_device *next)
990 * To be removed, setup is performed by osd_device_{init,alloc} and
991 * cleanup is performed by osd_device_{fini,free).
993 static int osd_process_config(const struct lu_env *env,
994 struct lu_device *d, struct lustre_cfg *cfg)
996 struct osd_device *o = osd_dev(d);
1000 switch(cfg->lcfg_command) {
1002 rc = osd_mount(env, o, cfg);
1005 rc = osd_shutdown(env, o);
1008 LASSERT(&o->od_dt_dev);
1009 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1010 cfg, &o->od_dt_dev);
1011 if (rc > 0 || rc == -ENOSYS)
1012 rc = class_process_proc_param(PARAM_OST,
1013 lprocfs_osd_obd_vars,
1014 cfg, &o->od_dt_dev);
1024 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1026 struct osd_device *osd = osd_dev(d);
1030 if (osd->od_quota_slave == NULL)
1033 /* start qsd instance on recovery completion, this notifies the quota
1034 * slave code that we are about to process new requests now */
1035 rc = qsd_start(env, osd->od_quota_slave);
1040 * we use exports to track all osd users
1042 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1043 struct obd_device *obd, struct obd_uuid *cluuid,
1044 struct obd_connect_data *data, void *localdata)
1046 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1047 struct lustre_handle conn;
1051 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1053 rc = class_connect(&conn, obd, cluuid);
1057 *exp = class_conn2export(&conn);
1059 spin_lock(&obd->obd_dev_lock);
1061 spin_unlock(&obd->obd_dev_lock);
1067 * once last export (we don't count self-export) disappeared
1068 * osd can be released
1070 static int osd_obd_disconnect(struct obd_export *exp)
1072 struct obd_device *obd = exp->exp_obd;
1073 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1074 int rc, release = 0;
1077 /* Only disconnect the underlying layers on the final disconnect. */
1078 spin_lock(&obd->obd_dev_lock);
1080 if (osd->od_connects == 0)
1082 spin_unlock(&obd->obd_dev_lock);
1084 rc = class_disconnect(exp); /* bz 9811 */
1086 if (rc == 0 && release)
1087 class_manual_cleanup(obd);
1091 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1092 struct lu_device *dev)
1094 struct osd_device *osd = osd_dev(dev);
1098 if (osd->od_quota_slave != NULL)
1099 /* set up quota slave objects */
1100 rc = qsd_prepare(env, osd->od_quota_slave);
1105 struct lu_device_operations osd_lu_ops = {
1106 .ldo_object_alloc = osd_object_alloc,
1107 .ldo_process_config = osd_process_config,
1108 .ldo_recovery_complete = osd_recovery_complete,
1109 .ldo_prepare = osd_prepare,
1112 static void osd_type_start(struct lu_device_type *t)
1116 static void osd_type_stop(struct lu_device_type *t)
1120 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1121 struct lu_fid *fid, struct md_op_data *op_data)
1123 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1125 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1128 static struct lu_device_type_operations osd_device_type_ops = {
1129 .ldto_init = osd_type_init,
1130 .ldto_fini = osd_type_fini,
1132 .ldto_start = osd_type_start,
1133 .ldto_stop = osd_type_stop,
1135 .ldto_device_alloc = osd_device_alloc,
1136 .ldto_device_free = osd_device_free,
1138 .ldto_device_init = osd_device_init,
1139 .ldto_device_fini = osd_device_fini
1142 static struct lu_device_type osd_device_type = {
1143 .ldt_tags = LU_DEVICE_DT,
1144 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1145 .ldt_ops = &osd_device_type_ops,
1146 .ldt_ctx_tags = LCT_LOCAL
1150 static struct obd_ops osd_obd_device_ops = {
1151 .o_owner = THIS_MODULE,
1152 .o_connect = osd_obd_connect,
1153 .o_disconnect = osd_obd_disconnect,
1154 .o_fid_alloc = osd_fid_alloc
1157 int __init osd_init(void)
1161 rc = osd_options_init();
1165 rc = lu_kmem_init(osd_caches);
1169 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1170 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1172 lu_kmem_fini(osd_caches);
1176 void __exit osd_exit(void)
1178 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1179 lu_kmem_fini(osd_caches);
1182 extern unsigned int osd_oi_count;
1183 CFS_MODULE_PARM(osd_oi_count, "i", int, 0444,
1184 "Number of Object Index containers to be created, "
1185 "it's only valid for new filesystem.");
1187 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1188 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1189 MODULE_LICENSE("GPL");
1191 cfs_module(osd, LUSTRE_VERSION_STRING, osd_init, osd_exit);