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 /* Slab to allocate osd_zap_it */
78 struct kmem_cache *osd_zapit_cachep;
80 static struct lu_kmem_descr osd_caches[] = {
82 .ckd_cache = &osd_object_kmem,
83 .ckd_name = "zfs_osd_obj",
84 .ckd_size = sizeof(struct osd_object)
87 .ckd_cache = &osd_zapit_cachep,
88 .ckd_name = "osd_zapit_cache",
89 .ckd_size = sizeof(struct osd_zap_it)
96 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 dt_txn_hook_commit(th);
153 /* call per-transaction callbacks if any */
154 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage)
155 dcb->dcb_func(NULL, th, dcb, error);
157 /* Unlike ldiskfs, zfs updates space accounting at commit time.
158 * As a consequence, op_end is called only now to inform the quota slave
159 * component that reserved quota space is now accounted in usage and
160 * should be released. Quota space won't be adjusted at this point since
161 * we can't provide a suitable environment. It will be performed
162 * asynchronously by a lquota thread. */
163 qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);
167 lu_context_exit(&th->th_ctx);
168 lu_context_fini(&th->th_ctx);
169 thandle_put(&oh->ot_super);
174 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
176 struct osd_thandle *oh;
178 oh = container_of0(th, struct osd_thandle, ot_super);
179 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
185 * Concurrency: shouldn't matter.
187 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
190 struct osd_thandle *oh;
194 oh = container_of0(th, struct osd_thandle, ot_super);
198 rc = dt_txn_hook_start(env, d, th);
202 if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
203 /* Unlike ldiskfs, ZFS checks for available space and returns
204 * -ENOSPC when assigning txg */
207 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
208 if (unlikely(rc != 0)) {
209 struct osd_device *osd = osd_dt_dev(d);
210 /* dmu will call commit callback with error code during abort */
211 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
212 CERROR("%s: failed to start transaction due to ENOSPC. "
213 "Metadata overhead is underestimated or "
214 "grant_ratio is too low.\n", osd->od_svname);
216 CERROR("%s: can't assign tx: rc = %d\n",
219 /* add commit callback */
220 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
222 lu_context_init(&th->th_ctx, th->th_tags);
223 lu_context_enter(&th->th_ctx);
224 lu_device_get(&d->dd_lu_dev);
231 * Concurrency: shouldn't matter.
233 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
236 struct osd_device *osd = osd_dt_dev(th->th_dev);
237 struct osd_thandle *oh;
242 oh = container_of0(th, struct osd_thandle, ot_super);
244 if (oh->ot_assigned == 0) {
246 dmu_tx_abort(oh->ot_tx);
247 osd_object_sa_dirty_rele(oh);
248 /* there won't be any commit, release reserved quota space now,
250 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
251 thandle_put(&oh->ot_super);
255 /* When doing our own inode accounting, the ZAPs storing per-uid/gid
256 * usage are updated at operation execution time, so we should call
257 * qsd_op_end() straight away. Otherwise (for blk accounting maintained
258 * by ZFS and when #inode is estimated from #blks) accounting is updated
259 * at commit time and the call to qsd_op_end() must be delayed */
260 if (oh->ot_quota_trans.lqt_id_cnt > 0 &&
261 !oh->ot_quota_trans.lqt_ids[0].lqi_is_blk &&
262 !osd->od_quota_iused_est)
263 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
265 rc = dt_txn_hook_stop(env, th);
267 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
271 txg = oh->ot_tx->tx_txg;
273 osd_object_sa_dirty_rele(oh);
274 dmu_tx_commit(oh->ot_tx);
277 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
282 static struct thandle *osd_trans_create(const struct lu_env *env,
283 struct dt_device *dt)
285 struct osd_device *osd = osd_dt_dev(dt);
286 struct osd_thandle *oh;
291 tx = dmu_tx_create(osd->od_os);
293 RETURN(ERR_PTR(-ENOMEM));
295 /* alloc callback data */
299 RETURN(ERR_PTR(-ENOMEM));
303 INIT_LIST_HEAD(&oh->ot_dcb_list);
304 INIT_LIST_HEAD(&oh->ot_sa_list);
305 sema_init(&oh->ot_sa_lock, 1);
306 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
310 th->th_tags = LCT_TX_HANDLE;
311 atomic_set(&th->th_refc, 1);
312 th->th_alloc_size = sizeof(*oh);
316 /* Estimate the number of objects from a number of blocks */
317 uint64_t osd_objs_count_estimate(uint64_t refdbytes, uint64_t usedobjs,
320 uint64_t est_objs, est_refdblocks, est_usedobjs;
322 /* Compute an nrblocks estimate based on the actual number of
323 * dnodes that could fit in the space. Since we don't know the
324 * overhead associated with each dnode (xattrs, SAs, VDEV overhead,
325 * etc) just using DNODE_SHIFT isn't going to give a good estimate.
326 * Instead, compute an estimate based on the average space usage per
327 * dnode, with an upper and lower cap.
329 * In case there aren't many dnodes or blocks used yet, add a small
330 * correction factor using OSD_DNODE_EST_SHIFT. This correction
331 * factor gradually disappears as the number of real dnodes grows.
332 * This also avoids the need to check for divide-by-zero later.
334 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
335 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
337 est_refdblocks = (refdbytes >> SPA_MAXBLOCKSHIFT) +
338 (OSD_DNODE_EST_COUNT >> OSD_DNODE_EST_BLKSHIFT);
339 est_usedobjs = usedobjs + OSD_DNODE_EST_COUNT;
341 /* Average space/dnode more than maximum dnode size, use max dnode
342 * size to estimate free dnodes from adjusted free blocks count.
343 * OSTs typically use more than one block dnode so this case applies. */
344 if (est_usedobjs <= est_refdblocks * 2) {
347 /* Average space/dnode smaller than min dnode size (probably due to
348 * metadnode compression), use min dnode size to estimate the number of
350 * An MDT typically uses below 512 bytes/dnode so this case applies. */
351 } else if (est_usedobjs >= (est_refdblocks << OSD_DNODE_MIN_BLKSHIFT)) {
352 est_objs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
354 /* Between the extremes, we try to use the average size of
355 * existing dnodes to compute the number of dnodes that fit
358 * est_objs = nrblocks * (est_usedobjs / est_refblocks);
360 * but this may overflow 64 bits or become 0 if not handled well
362 * We know nrblocks is below (64 - 17 = 47) bits from
363 * SPA_MAXBLKSHIFT, and est_usedobjs is under 48 bits due to
364 * DN_MAX_OBJECT_SHIFT, which means that multiplying them may
365 * get as large as 2 ^ 95.
367 * We also know (est_usedobjs / est_refdblocks) is between 2 and
368 * 256, due to above checks, we can safely compute this first.
369 * We care more about accuracy on the MDT (many dnodes/block)
370 * which is good because this is where truncation errors are
371 * smallest. This adds 8 bits to nrblocks so we can use 7 bits
372 * to compute a fixed-point fraction and nrblocks can still fit
375 unsigned dnodes_per_block = (est_usedobjs << 7)/est_refdblocks;
377 est_objs = (nrblocks * dnodes_per_block) >> 7;
382 static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs)
384 uint64_t refdbytes, availbytes, usedobjs, availobjs;
385 uint64_t est_availobjs;
388 dmu_objset_space(os, &refdbytes, &availbytes, &usedobjs,
392 * ZFS allows multiple block sizes. For statfs, Linux makes no
393 * proper distinction between bsize and frsize. For calculations
394 * of free and used blocks incorrectly uses bsize instead of frsize,
395 * but bsize is also used as the optimal blocksize. We return the
396 * largest possible block size as IO size for the optimum performance
397 * and scale the free and used blocks count appropriately.
399 osfs->os_bsize = 1ULL << SPA_MAXBLOCKSHIFT;
401 osfs->os_blocks = (refdbytes + availbytes) >> SPA_MAXBLOCKSHIFT;
402 osfs->os_bfree = availbytes >> SPA_MAXBLOCKSHIFT;
403 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
405 /* Take replication (i.e. number of copies) into account */
406 osfs->os_bavail /= os->os_copies;
409 * Reserve some space so we don't run into ENOSPC due to grants not
410 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
411 * Rather than report this via os_bavail (which makes users unhappy if
412 * they can't fill the filesystem 100%), reduce os_blocks as well.
414 * Reserve 0.78% of total space, at least 4MB for small filesystems,
415 * for internal files to be created/unlinked when space is tight.
417 CLASSERT(OSD_STATFS_RESERVED_BLKS > 0);
418 if (likely(osfs->os_blocks >=
419 OSD_STATFS_RESERVED_BLKS << OSD_STATFS_RESERVED_SHIFT))
420 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
422 reserved = OSD_STATFS_RESERVED_BLKS;
424 osfs->os_blocks -= reserved;
425 osfs->os_bfree -= MIN(reserved, osfs->os_bfree);
426 osfs->os_bavail -= MIN(reserved, osfs->os_bavail);
429 * The availobjs value returned from dmu_objset_space() is largely
430 * useless, since it reports the number of objects that might
431 * theoretically still fit into the dataset, independent of minor
432 * issues like how much space is actually available in the pool.
433 * Compute a better estimate in udmu_objs_count_estimate().
435 est_availobjs = osd_objs_count_estimate(refdbytes, usedobjs,
438 osfs->os_ffree = min(availobjs, est_availobjs);
439 osfs->os_files = osfs->os_ffree + usedobjs;
441 /* ZFS XXX: fill in backing dataset FSID/UUID
442 memcpy(osfs->os_fsid, .... );*/
444 /* We're a zfs filesystem. */
445 osfs->os_type = UBERBLOCK_MAGIC;
447 /* ZFS XXX: fill in appropriate OS_STATE_{DEGRADED,READONLY} flags
448 osfs->os_state = vf_to_stf(vfsp->vfs_flag);
449 if (sb->s_flags & MS_RDONLY)
450 osfs->os_state = OS_STATE_READONLY;
453 osfs->os_namelen = MAXNAMELEN;
454 osfs->os_maxbytes = OBD_OBJECT_EOF;
460 * Concurrency: shouldn't matter.
462 int osd_statfs(const struct lu_env *env, struct dt_device *d,
463 struct obd_statfs *osfs)
465 struct osd_device *osd = osd_dt_dev(d);
469 rc = osd_objset_statfs(osd->od_os, osfs);
470 if (unlikely(rc != 0))
473 osfs->os_bavail -= min_t(u64,
474 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
479 static int osd_blk_insert_cost(void)
481 int max_blockshift, nr_blkptrshift;
483 /* max_blockshift is the log2 of the number of blocks needed to reach
484 * the maximum filesize (that's to say 2^64) */
485 max_blockshift = DN_MAX_OFFSET_SHIFT - SPA_MAXBLOCKSHIFT;
487 /* nr_blkptrshift is the log2 of the number of block pointers that can
488 * be stored in an indirect block */
489 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
490 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
492 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
493 * tree. We add +1 for rounding purpose.
494 * The tree depth times the indirect block size gives us the maximum
495 * cost of inserting a block in the tree */
496 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
500 * Concurrency: doesn't access mutable data.
502 static void osd_conf_get(const struct lu_env *env,
503 const struct dt_device *dev,
504 struct dt_device_param *param)
506 struct osd_device *osd = osd_dt_dev(dev);
509 * XXX should be taken from not-yet-existing fs abstraction layer.
511 param->ddp_max_name_len = MAXNAMELEN;
512 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
513 param->ddp_block_shift = 12; /* XXX */
514 param->ddp_mount_type = LDD_MT_ZFS;
516 param->ddp_mntopts = MNTOPT_USERXATTR;
517 if (osd->od_posix_acl)
518 param->ddp_mntopts |= MNTOPT_ACL;
519 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
521 /* for maxbytes, report same value as ZPL */
522 param->ddp_maxbytes = MAX_LFS_FILESIZE;
524 /* Default reserved fraction of the available space that should be kept
525 * for error margin. Unfortunately, there are many factors that can
526 * impact the overhead with zfs, so let's be very cautious for now and
527 * reserve 20% of the available space which is not given out as grant.
528 * This tunable can be changed on a live system via procfs if needed. */
529 param->ddp_grant_reserved = 20;
531 /* inodes are dynamically allocated, so we report the per-inode space
532 * consumption to upper layers. This static value is not really accurate
533 * and we should use the same logic as in udmu_objset_statfs() to
534 * estimate the real size consumed by an object */
535 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
536 /* per-fragment overhead to be used by the client code */
537 param->ddp_grant_frag = osd_blk_insert_cost();
541 * Concurrency: shouldn't matter.
543 static int osd_sync(const struct lu_env *env, struct dt_device *d)
545 struct osd_device *osd = osd_dt_dev(d);
546 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
547 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
548 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
552 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
554 struct osd_device *osd = osd_dt_dev(dev);
555 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
558 mutex_enter(&tx->tx_sync_lock);
559 txg = tx->tx_open_txg + 1;
560 if (tx->tx_quiesce_txg_waiting < txg) {
561 tx->tx_quiesce_txg_waiting = txg;
562 cv_broadcast(&tx->tx_quiesce_more_cv);
564 mutex_exit(&tx->tx_sync_lock);
570 * Concurrency: shouldn't matter.
572 static int osd_ro(const struct lu_env *env, struct dt_device *d)
574 struct osd_device *osd = osd_dt_dev(d);
577 CERROR("%s: *** setting device %s read-only ***\n",
578 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
580 spa_freeze(dmu_objset_spa(osd->od_os));
585 static struct dt_device_operations osd_dt_ops = {
586 .dt_root_get = osd_root_get,
587 .dt_statfs = osd_statfs,
588 .dt_trans_create = osd_trans_create,
589 .dt_trans_start = osd_trans_start,
590 .dt_trans_stop = osd_trans_stop,
591 .dt_trans_cb_add = osd_trans_cb_add,
592 .dt_conf_get = osd_conf_get,
594 .dt_commit_async = osd_commit_async,
599 * DMU OSD device type methods
601 static int osd_type_init(struct lu_device_type *t)
603 LU_CONTEXT_KEY_INIT(&osd_key);
604 return lu_context_key_register(&osd_key);
607 static void osd_type_fini(struct lu_device_type *t)
609 lu_context_key_degister(&osd_key);
612 static void *osd_key_init(const struct lu_context *ctx,
613 struct lu_context_key *key)
615 struct osd_thread_info *info;
619 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
621 info = ERR_PTR(-ENOMEM);
625 static void osd_key_fini(const struct lu_context *ctx,
626 struct lu_context_key *key, void *data)
628 struct osd_thread_info *info = data;
633 static void osd_key_exit(const struct lu_context *ctx,
634 struct lu_context_key *key, void *data)
636 struct osd_thread_info *info = data;
638 memset(info, 0, sizeof(*info));
641 struct lu_context_key osd_key = {
642 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
643 .lct_init = osd_key_init,
644 .lct_fini = osd_key_fini,
645 .lct_exit = osd_key_exit
648 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
650 if (osd->od_cl_seq == NULL)
653 seq_client_fini(osd->od_cl_seq);
654 OBD_FREE_PTR(osd->od_cl_seq);
655 osd->od_cl_seq = NULL;
658 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
662 /* shutdown quota slave instance associated with the device */
663 if (o->od_quota_slave != NULL) {
664 qsd_fini(env, o->od_quota_slave);
665 o->od_quota_slave = NULL;
668 osd_fid_fini(env, o);
673 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
675 struct osd_device *osd = arg;
677 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
680 static int osd_objset_open(struct osd_device *o)
682 uint64_t version = ZPL_VERSION;
687 rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os);
693 /* Check ZFS version */
694 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
695 ZPL_VERSION_STR, 8, 1, &version);
697 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
699 * We can't return ENOENT because that would mean the objset
702 GOTO(out, rc = -EIO);
705 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
706 ZFS_SA_ATTRS, 8, 1, &sa_obj);
710 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
711 ZPL_END, &o->z_attr_table);
715 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
716 8, 1, &o->od_rootid);
718 CERROR("%s: lookup for root failed: rc = %d\n",
723 /* Check that user/group usage tracking is supported */
724 if (!dmu_objset_userused_enabled(o->od_os) ||
725 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
726 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
727 CERROR("%s: Space accounting not supported by this target, "
728 "aborting\n", o->od_svname);
729 GOTO(out, -ENOTSUPP);
733 if (rc != 0 && o->od_os != NULL)
734 dmu_objset_disown(o->od_os, o);
739 static int osd_mount(const struct lu_env *env,
740 struct osd_device *o, struct lustre_cfg *cfg)
742 struct dsl_dataset *ds;
743 char *mntdev = lustre_cfg_string(cfg, 1);
744 char *svname = lustre_cfg_string(cfg, 4);
751 if (o->od_os != NULL)
754 if (mntdev == NULL || svname == NULL)
757 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
758 if (rc >= sizeof(o->od_mntdev))
761 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
762 if (rc >= sizeof(o->od_svname))
765 if (server_name_is_ost(o->od_svname))
768 rc = osd_objset_open(o);
770 CERROR("%s: can't open objset %s: rc = %d\n", o->od_svname,
775 ds = dmu_objset_ds(o->od_os);
776 dp = dmu_objset_pool(o->od_os);
779 dsl_pool_config_enter(dp, FTAG);
780 rc = dsl_prop_register(ds, "xattr", osd_xattr_changed_cb, o);
781 dsl_pool_config_exit(dp, FTAG);
783 CWARN("%s: can't register xattr callback, ignore: rc=%d\n",
786 rc = __osd_obj2dbuf(env, o->od_os, o->od_rootid, &rootdb);
788 CERROR("%s: obj2dbuf() failed: rc = %d\n", o->od_svname, rc);
789 dmu_objset_disown(o->od_os, o);
794 o->od_root = rootdb->db_object;
795 sa_buf_rele(rootdb, osd_obj_tag);
797 /* 1. initialize oi before any file create or file open */
798 rc = osd_oi_init(env, o);
802 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
805 o->od_site.ls_bottom_dev = osd2lu_dev(o);
807 rc = lu_site_init_finish(&o->od_site);
811 rc = osd_convert_root_to_new_seq(env, o);
815 /* Use our own ZAP for inode accounting by default, this can be changed
816 * via procfs to estimate the inode usage from the block usage */
817 o->od_quota_iused_est = 0;
819 rc = osd_procfs_init(o, o->od_svname);
823 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
825 /* initialize quota slave instance */
826 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
828 if (IS_ERR(o->od_quota_slave)) {
829 rc = PTR_ERR(o->od_quota_slave);
830 o->od_quota_slave = NULL;
834 /* parse mount option "noacl", and enable ACL by default */
835 opts = lustre_cfg_string(cfg, 3);
836 if (opts == NULL || strstr(opts, "noacl") == NULL)
843 static void osd_umount(const struct lu_env *env, struct osd_device *o)
847 if (atomic_read(&o->od_zerocopy_alloc))
848 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
849 atomic_read(&o->od_zerocopy_alloc));
850 if (atomic_read(&o->od_zerocopy_loan))
851 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
852 atomic_read(&o->od_zerocopy_loan));
853 if (atomic_read(&o->od_zerocopy_pin))
854 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
855 atomic_read(&o->od_zerocopy_pin));
857 if (o->od_os != NULL) {
858 /* force a txg sync to get all commit callbacks */
859 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
861 /* close the object set */
862 dmu_objset_disown(o->od_os, o);
870 static int osd_device_init0(const struct lu_env *env,
871 struct osd_device *o,
872 struct lustre_cfg *cfg)
874 struct lu_device *l = osd2lu_dev(o);
877 /* if the module was re-loaded, env can loose its keys */
878 rc = lu_env_refill((struct lu_env *) env);
882 l->ld_ops = &osd_lu_ops;
883 o->od_dt_dev.dd_ops = &osd_dt_ops;
889 static struct lu_device *osd_device_fini(const struct lu_env *env,
890 struct lu_device *dev);
892 static struct lu_device *osd_device_alloc(const struct lu_env *env,
893 struct lu_device_type *type,
894 struct lustre_cfg *cfg)
896 struct osd_device *dev;
901 return ERR_PTR(-ENOMEM);
903 rc = dt_device_init(&dev->od_dt_dev, type);
905 rc = osd_device_init0(env, dev, cfg);
907 rc = osd_mount(env, dev, cfg);
909 osd_device_fini(env, osd2lu_dev(dev));
912 dt_device_fini(&dev->od_dt_dev);
915 if (unlikely(rc != 0))
918 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
921 static struct lu_device *osd_device_free(const struct lu_env *env,
924 struct osd_device *o = osd_dev(d);
927 /* XXX: make osd top device in order to release reference */
928 d->ld_site->ls_top_dev = d;
929 lu_site_purge(env, d->ld_site, -1);
930 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
931 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
932 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
934 lu_site_fini(&o->od_site);
935 dt_device_fini(&o->od_dt_dev);
941 static struct lu_device *osd_device_fini(const struct lu_env *env,
944 struct osd_device *o = osd_dev(d);
945 struct dsl_dataset *ds;
950 osd_shutdown(env, o);
954 ds = dmu_objset_ds(o->od_os);
955 rc = dsl_prop_unregister(ds, "xattr", osd_xattr_changed_cb, o);
957 CERROR("%s: dsl_prop_unregister xattr error %d\n",
959 if (o->arc_prune_cb != NULL) {
960 arc_remove_prune_callback(o->arc_prune_cb);
961 o->arc_prune_cb = NULL;
963 osd_sync(env, lu2dt_dev(d));
964 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
967 rc = osd_procfs_fini(o);
969 CERROR("proc fini error %d\n", rc);
979 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
980 const char *name, struct lu_device *next)
986 * To be removed, setup is performed by osd_device_{init,alloc} and
987 * cleanup is performed by osd_device_{fini,free).
989 static int osd_process_config(const struct lu_env *env,
990 struct lu_device *d, struct lustre_cfg *cfg)
992 struct osd_device *o = osd_dev(d);
996 switch(cfg->lcfg_command) {
998 rc = osd_mount(env, o, cfg);
1001 rc = osd_shutdown(env, o);
1004 LASSERT(&o->od_dt_dev);
1005 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1006 cfg, &o->od_dt_dev);
1007 if (rc > 0 || rc == -ENOSYS)
1008 rc = class_process_proc_param(PARAM_OST,
1009 lprocfs_osd_obd_vars,
1010 cfg, &o->od_dt_dev);
1020 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1022 struct osd_device *osd = osd_dev(d);
1026 if (osd->od_quota_slave == NULL)
1029 /* start qsd instance on recovery completion, this notifies the quota
1030 * slave code that we are about to process new requests now */
1031 rc = qsd_start(env, osd->od_quota_slave);
1036 * we use exports to track all osd users
1038 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1039 struct obd_device *obd, struct obd_uuid *cluuid,
1040 struct obd_connect_data *data, void *localdata)
1042 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1043 struct lustre_handle conn;
1047 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1049 rc = class_connect(&conn, obd, cluuid);
1053 *exp = class_conn2export(&conn);
1055 spin_lock(&obd->obd_dev_lock);
1057 spin_unlock(&obd->obd_dev_lock);
1063 * once last export (we don't count self-export) disappeared
1064 * osd can be released
1066 static int osd_obd_disconnect(struct obd_export *exp)
1068 struct obd_device *obd = exp->exp_obd;
1069 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1070 int rc, release = 0;
1073 /* Only disconnect the underlying layers on the final disconnect. */
1074 spin_lock(&obd->obd_dev_lock);
1076 if (osd->od_connects == 0)
1078 spin_unlock(&obd->obd_dev_lock);
1080 rc = class_disconnect(exp); /* bz 9811 */
1082 if (rc == 0 && release)
1083 class_manual_cleanup(obd);
1087 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1089 struct seq_server_site *ss = osd_seq_site(osd);
1093 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1096 if (unlikely(ss == NULL))
1099 OBD_ALLOC_PTR(osd->od_cl_seq);
1100 if (osd->od_cl_seq == NULL)
1103 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1104 osd->od_svname, ss->ss_server_seq);
1107 OBD_FREE_PTR(osd->od_cl_seq);
1108 osd->od_cl_seq = NULL;
1114 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1115 struct lu_device *dev)
1117 struct osd_device *osd = osd_dev(dev);
1121 if (osd->od_quota_slave != NULL) {
1122 /* set up quota slave objects */
1123 rc = qsd_prepare(env, osd->od_quota_slave);
1128 rc = osd_fid_init(env, osd);
1133 struct lu_device_operations osd_lu_ops = {
1134 .ldo_object_alloc = osd_object_alloc,
1135 .ldo_process_config = osd_process_config,
1136 .ldo_recovery_complete = osd_recovery_complete,
1137 .ldo_prepare = osd_prepare,
1140 static void osd_type_start(struct lu_device_type *t)
1144 static void osd_type_stop(struct lu_device_type *t)
1148 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1149 struct lu_fid *fid, struct md_op_data *op_data)
1151 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1153 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1156 static struct lu_device_type_operations osd_device_type_ops = {
1157 .ldto_init = osd_type_init,
1158 .ldto_fini = osd_type_fini,
1160 .ldto_start = osd_type_start,
1161 .ldto_stop = osd_type_stop,
1163 .ldto_device_alloc = osd_device_alloc,
1164 .ldto_device_free = osd_device_free,
1166 .ldto_device_init = osd_device_init,
1167 .ldto_device_fini = osd_device_fini
1170 static struct lu_device_type osd_device_type = {
1171 .ldt_tags = LU_DEVICE_DT,
1172 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1173 .ldt_ops = &osd_device_type_ops,
1174 .ldt_ctx_tags = LCT_LOCAL
1178 static struct obd_ops osd_obd_device_ops = {
1179 .o_owner = THIS_MODULE,
1180 .o_connect = osd_obd_connect,
1181 .o_disconnect = osd_obd_disconnect,
1182 .o_fid_alloc = osd_fid_alloc
1185 int __init osd_init(void)
1189 rc = osd_options_init();
1193 rc = lu_kmem_init(osd_caches);
1197 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1198 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1200 lu_kmem_fini(osd_caches);
1204 void __exit osd_exit(void)
1206 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1207 lu_kmem_fini(osd_caches);
1210 extern unsigned int osd_oi_count;
1211 CFS_MODULE_PARM(osd_oi_count, "i", int, 0444,
1212 "Number of Object Index containers to be created, "
1213 "it's only valid for new filesystem.");
1215 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1216 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1217 MODULE_LICENSE("GPL");
1219 cfs_module(osd, LUSTRE_VERSION_STRING, osd_init, osd_exit);