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, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osd-zfs/osd_handler.c
33 * Top-level entry points into osd module
35 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
36 * Author: Mike Pershin <tappro@whamcloud.com>
37 * Author: Johann Lombardi <johann@whamcloud.com>
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <libcfs/libcfs.h>
43 #include <obd_support.h>
44 #include <lustre_net.h>
46 #include <obd_class.h>
47 #include <lustre_disk.h>
48 #include <lustre_fid.h>
49 #include <uapi/linux/lustre/lustre_param.h>
50 #include <md_object.h>
52 #include "osd_internal.h"
54 #include <sys/dnode.h>
59 #include <sys/spa_impl.h>
60 #include <sys/zfs_znode.h>
61 #include <sys/dmu_tx.h>
62 #include <sys/dmu_objset.h>
63 #include <sys/dsl_prop.h>
64 #include <sys/sa_impl.h>
67 struct lu_context_key osd_key;
69 /* Slab for OSD object allocation */
70 struct kmem_cache *osd_object_kmem;
72 /* Slab to allocate osd_zap_it */
73 struct kmem_cache *osd_zapit_cachep;
75 static struct lu_kmem_descr osd_caches[] = {
77 .ckd_cache = &osd_object_kmem,
78 .ckd_name = "zfs_osd_obj",
79 .ckd_size = sizeof(struct osd_object)
82 .ckd_cache = &osd_zapit_cachep,
83 .ckd_name = "osd_zapit_cache",
84 .ckd_size = sizeof(struct osd_zap_it)
91 static void arc_prune_func(int64_t bytes, void *private)
93 struct osd_device *od = private;
94 struct lu_site *site = &od->od_site;
98 LASSERT(site->ls_obj_hash);
100 rc = lu_env_init(&env, LCT_SHRINKER);
102 CERROR("%s: can't initialize shrinker env: rc = %d\n",
107 lu_site_purge(&env, site, (bytes >> 10));
113 * Concurrency: doesn't access mutable data
115 static int osd_root_get(const struct lu_env *env,
116 struct dt_device *dev, struct lu_fid *f)
118 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
123 * OSD object methods.
127 * Concurrency: shouldn't matter.
129 static void osd_trans_commit_cb(void *cb_data, int error)
131 struct osd_thandle *oh = cb_data;
132 struct thandle *th = &oh->ot_super;
133 struct osd_device *osd = osd_dt_dev(th->th_dev);
134 struct lu_device *lud = &th->th_dev->dd_lu_dev;
135 struct dt_txn_commit_cb *dcb, *tmp;
140 if (error == ECANCELED)
141 CWARN("%s: transaction @0x%p was aborted\n",
142 osd_dt_dev(th->th_dev)->od_svname, th);
144 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
145 osd_dt_dev(th->th_dev)->od_svname, th, error);
148 dt_txn_hook_commit(th);
150 /* call per-transaction callbacks if any */
151 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) {
152 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
153 "commit callback entry: magic=%x name='%s'\n",
154 dcb->dcb_magic, dcb->dcb_name);
155 list_del_init(&dcb->dcb_linkage);
156 dcb->dcb_func(NULL, th, dcb, error);
159 /* Unlike ldiskfs, zfs updates space accounting at commit time.
160 * As a consequence, op_end is called only now to inform the quota slave
161 * component that reserved quota space is now accounted in usage and
162 * should be released. Quota space won't be adjusted at this point since
163 * we can't provide a suitable environment. It will be performed
164 * asynchronously by a lquota thread. */
165 qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);
174 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
176 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
179 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
180 LASSERT(&dcb->dcb_func != NULL);
181 if (dcb->dcb_flags & DCB_TRANS_STOP)
182 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
184 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
190 * Concurrency: shouldn't matter.
192 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
195 struct osd_thandle *oh;
199 oh = container_of0(th, struct osd_thandle, ot_super);
203 rc = dt_txn_hook_start(env, d, th);
207 if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
208 /* Unlike ldiskfs, ZFS checks for available space and returns
209 * -ENOSPC when assigning txg */
212 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
213 if (unlikely(rc != 0)) {
214 struct osd_device *osd = osd_dt_dev(d);
215 /* dmu will call commit callback with error code during abort */
216 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
217 CERROR("%s: failed to start transaction due to ENOSPC"
218 "\n", osd->od_svname);
220 CERROR("%s: can't assign tx: rc = %d\n",
223 /* add commit callback */
224 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
226 osd_oti_get(env)->oti_in_trans = 1;
227 lu_device_get(&d->dd_lu_dev);
233 static void osd_unlinked_list_emptify(const struct lu_env *env,
234 struct osd_device *osd,
235 struct list_head *list, bool free)
237 struct osd_object *obj;
240 while (!list_empty(list)) {
241 obj = list_entry(list->next,
242 struct osd_object, oo_unlinked_linkage);
243 LASSERT(obj->oo_dn != NULL);
244 oid = obj->oo_dn->dn_object;
246 list_del_init(&obj->oo_unlinked_linkage);
248 (void)osd_unlinked_object_free(env, osd, oid);
252 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
254 struct dt_txn_commit_cb *dcb;
255 struct dt_txn_commit_cb *tmp;
257 /* call per-transaction stop callbacks if any */
258 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
260 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
261 "commit callback entry: magic=%x name='%s'\n",
262 dcb->dcb_magic, dcb->dcb_name);
263 list_del_init(&dcb->dcb_linkage);
264 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
269 * Concurrency: shouldn't matter.
271 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
274 struct osd_device *osd = osd_dt_dev(th->th_dev);
275 bool sync = (th->th_sync != 0);
276 struct osd_thandle *oh;
277 struct list_head unlinked;
282 oh = container_of0(th, struct osd_thandle, ot_super);
283 INIT_LIST_HEAD(&unlinked);
284 list_splice_init(&oh->ot_unlinked_list, &unlinked);
285 /* reset OI cache for safety */
286 osd_oti_get(env)->oti_ins_cache_used = 0;
288 if (oh->ot_assigned == 0) {
290 dmu_tx_abort(oh->ot_tx);
291 osd_object_sa_dirty_rele(env, oh);
292 osd_unlinked_list_emptify(env, osd, &unlinked, false);
293 /* there won't be any commit, release reserved quota space now,
295 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
300 rc = dt_txn_hook_stop(env, th);
302 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
305 osd_trans_stop_cb(oh, rc);
308 txg = oh->ot_tx->tx_txg;
310 osd_object_sa_dirty_rele(env, oh);
311 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
312 * by osd_trans_commit_cb already. */
313 dmu_tx_commit(oh->ot_tx);
314 osd_oti_get(env)->oti_in_trans = 0;
316 osd_unlinked_list_emptify(env, osd, &unlinked, true);
319 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
324 static struct thandle *osd_trans_create(const struct lu_env *env,
325 struct dt_device *dt)
327 struct osd_device *osd = osd_dt_dev(dt);
328 struct osd_thandle *oh;
334 CERROR("%s: someone try to start transaction under "
335 "readonly mode, should be disabled.\n",
336 osd_name(osd_dt_dev(dt)));
338 RETURN(ERR_PTR(-EROFS));
341 tx = dmu_tx_create(osd->od_os);
343 RETURN(ERR_PTR(-ENOMEM));
345 /* alloc callback data */
349 RETURN(ERR_PTR(-ENOMEM));
353 INIT_LIST_HEAD(&oh->ot_dcb_list);
354 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
355 INIT_LIST_HEAD(&oh->ot_unlinked_list);
356 INIT_LIST_HEAD(&oh->ot_sa_list);
357 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
364 /* Estimate the total number of objects from a number of blocks */
365 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
366 uint64_t nrblocks, uint64_t est_maxblockshift)
368 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
371 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
372 * bump the free dnode estimate to assume blocks at least 64KB in
373 * case of a directory-heavy MDT (at 32KB/directory).
375 if (est_maxblockshift < 16) {
376 nrblocks >>= (16 - est_maxblockshift);
377 est_maxblockshift = 16;
381 * Estimate the total number of dnodes from the total blocks count
382 * and the space used per dnode. Since we don't know the overhead
383 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
384 * just using DNODE_SHIFT isn't going to give a good estimate.
385 * Instead, compute the current average space usage per dnode, with
386 * an upper and lower cap to avoid unrealistic estimates..
388 * In case there aren't many dnodes or blocks used yet, add a small
389 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
390 * gradually disappears as the number of real dnodes grows. It also
391 * avoids the need to check for divide-by-zero computing dn_per_block.
393 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
394 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
396 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
397 usedbytes) >> est_maxblockshift;
398 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
400 if (est_usedobjs <= est_usedblocks) {
402 * Average space/dnode more than maximum block size, use max
403 * block size to estimate free dnodes from adjusted free blocks
404 * count. OSTs typically use multiple blocks per dnode so this
407 est_totobjs = nrblocks;
409 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
411 * Average space/dnode smaller than min dnode size (probably
412 * due to metadnode compression), use min dnode size to
413 * estimate object count. MDTs may use only one block per node
414 * so this case applies.
416 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
420 * Between the extremes, use average space per existing dnode
421 * to compute the number of dnodes that will fit into nrblocks:
423 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
425 * this may overflow 64 bits or become 0 if not handled well.
427 * We know nrblocks is below 2^(64 - blkbits) bits, and
428 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
429 * which means that multiplying them may get as large as
430 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
432 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
433 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
434 * the limit checks above, so we can safely compute this first.
435 * We care more about accuracy on the MDT (many dnodes/block)
436 * which is good because this is where truncation errors are
437 * smallest. Since both nrblocks and dn_per_block are a
438 * function of blkbits, their product is at most:
440 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
442 * so we can safely use 7 bits to compute a fixed-point
443 * fraction and est_totobjs can still fit in 64 bits.
445 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
447 est_totobjs = (nrblocks * dn_per_block) >> 7;
452 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
454 struct objset *os = osd->od_os;
455 uint64_t usedbytes, availbytes, usedobjs, availobjs;
456 uint64_t est_availobjs;
460 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
462 memset(osfs, 0, sizeof(*osfs));
464 /* We're a zfs filesystem. */
465 osfs->os_type = UBERBLOCK_MAGIC;
468 * ZFS allows multiple block sizes. For statfs, Linux makes no
469 * proper distinction between bsize and frsize. For calculations
470 * of free and used blocks incorrectly uses bsize instead of frsize,
471 * but bsize is also used as the optimal blocksize. We return the
472 * largest possible block size as IO size for the optimum performance
473 * and scale the free and used blocks count appropriately.
475 osfs->os_bsize = osd->od_max_blksz;
476 bshift = fls64(osfs->os_bsize) - 1;
478 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
479 osfs->os_bfree = availbytes >> bshift;
480 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
482 /* Take replication (i.e. number of copies) into account */
483 if (os->os_copies != 0)
484 osfs->os_bavail /= os->os_copies;
487 * Reserve some space so we don't run into ENOSPC due to grants not
488 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
489 * Rather than report this via os_bavail (which makes users unhappy if
490 * they can't fill the filesystem 100%), reduce os_blocks as well.
492 * Reserve 0.78% of total space, at least 16MB for small filesystems,
493 * for internal files to be created/unlinked when space is tight.
495 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
496 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
497 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
498 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
500 osfs->os_blocks -= reserved;
501 osfs->os_bfree -= min(reserved, osfs->os_bfree);
502 osfs->os_bavail -= min(reserved, osfs->os_bavail);
505 * The availobjs value returned from dmu_objset_space() is largely
506 * useless, since it reports the number of objects that might
507 * theoretically still fit into the dataset, independent of minor
508 * issues like how much space is actually available in the pool.
509 * Compute a better estimate in udmu_objs_count_estimate().
511 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
512 osfs->os_bfree, bshift);
514 osfs->os_ffree = min(availobjs, est_availobjs);
515 osfs->os_files = osfs->os_ffree + usedobjs;
517 /* ZFS XXX: fill in backing dataset FSID/UUID
518 memcpy(osfs->os_fsid, .... );*/
520 osfs->os_namelen = MAXNAMELEN;
521 osfs->os_maxbytes = OBD_OBJECT_EOF;
523 if (!spa_writeable(dmu_objset_spa(os)) ||
524 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
525 osfs->os_state |= OS_STATE_READONLY;
531 * Concurrency: shouldn't matter.
533 int osd_statfs(const struct lu_env *env, struct dt_device *d,
534 struct obd_statfs *osfs)
539 rc = osd_objset_statfs(osd_dt_dev(d), osfs);
540 if (unlikely(rc != 0))
543 osfs->os_bavail -= min_t(u64,
544 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
549 static int osd_blk_insert_cost(struct osd_device *osd)
551 int max_blockshift, nr_blkptrshift, bshift;
553 /* max_blockshift is the log2 of the number of blocks needed to reach
554 * the maximum filesize (that's to say 2^64) */
555 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
556 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
558 /* nr_blkptrshift is the log2 of the number of block pointers that can
559 * be stored in an indirect block */
560 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
561 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
563 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
564 * tree. We add +1 for rounding purpose.
565 * The tree depth times the indirect block size gives us the maximum
566 * cost of inserting a block in the tree */
567 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
571 * Concurrency: doesn't access mutable data.
573 static void osd_conf_get(const struct lu_env *env,
574 const struct dt_device *dev,
575 struct dt_device_param *param)
577 struct osd_device *osd = osd_dt_dev(dev);
580 * XXX should be taken from not-yet-existing fs abstraction layer.
582 param->ddp_max_name_len = MAXNAMELEN;
583 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
584 param->ddp_symlink_max = PATH_MAX;
585 param->ddp_mount_type = LDD_MT_ZFS;
587 param->ddp_mntopts = MNTOPT_USERXATTR;
588 if (osd->od_posix_acl)
589 param->ddp_mntopts |= MNTOPT_ACL;
590 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
592 /* for maxbytes, report same value as ZPL */
593 param->ddp_maxbytes = MAX_LFS_FILESIZE;
595 /* inodes are dynamically allocated, so we report the per-inode space
596 * consumption to upper layers. This static value is not really accurate
597 * and we should use the same logic as in udmu_objset_statfs() to
598 * estimate the real size consumed by an object */
599 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
600 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
601 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
602 * not be accounted for every single new block insertion.
603 * Instead, the maximum extent size is set to the number of blocks that
604 * can fit into a single contiguous indirect block. There would be some
605 * cases where this crosses indirect blocks, but it also won't have 7
606 * new levels of indirect blocks in that case either, so it will still
607 * have enough reserved space for the extra indirect block */
608 param->ddp_max_extent_blks =
609 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
610 param->ddp_extent_tax = osd_blk_insert_cost(osd);
614 * Concurrency: shouldn't matter.
616 static int osd_sync(const struct lu_env *env, struct dt_device *d)
619 struct osd_device *osd = osd_dt_dev(d);
621 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
622 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
623 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
629 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
631 struct osd_device *osd = osd_dt_dev(dev);
632 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
635 mutex_enter(&tx->tx_sync_lock);
636 txg = tx->tx_open_txg + 1;
637 if (tx->tx_quiesce_txg_waiting < txg) {
638 tx->tx_quiesce_txg_waiting = txg;
639 cv_broadcast(&tx->tx_quiesce_more_cv);
641 mutex_exit(&tx->tx_sync_lock);
647 * Concurrency: shouldn't matter.
649 static int osd_ro(const struct lu_env *env, struct dt_device *d)
651 struct osd_device *osd = osd_dt_dev(d);
654 CERROR("%s: *** setting device %s read-only ***\n",
655 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
656 osd->od_dev_set_rdonly = 1;
657 spa_freeze(dmu_objset_spa(osd->od_os));
662 static struct dt_device_operations osd_dt_ops = {
663 .dt_root_get = osd_root_get,
664 .dt_statfs = osd_statfs,
665 .dt_trans_create = osd_trans_create,
666 .dt_trans_start = osd_trans_start,
667 .dt_trans_stop = osd_trans_stop,
668 .dt_trans_cb_add = osd_trans_cb_add,
669 .dt_conf_get = osd_conf_get,
671 .dt_commit_async = osd_commit_async,
676 * DMU OSD device type methods
678 static int osd_type_init(struct lu_device_type *t)
680 LU_CONTEXT_KEY_INIT(&osd_key);
681 return lu_context_key_register(&osd_key);
684 static void osd_type_fini(struct lu_device_type *t)
686 lu_context_key_degister(&osd_key);
689 static void *osd_key_init(const struct lu_context *ctx,
690 struct lu_context_key *key)
692 struct osd_thread_info *info;
696 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
698 info = ERR_PTR(-ENOMEM);
702 static void osd_key_fini(const struct lu_context *ctx,
703 struct lu_context_key *key, void *data)
705 struct osd_thread_info *info = data;
706 struct osd_idmap_cache *idc = info->oti_ins_cache;
709 LASSERT(info->oti_ins_cache_size > 0);
710 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
711 info->oti_ins_cache = NULL;
712 info->oti_ins_cache_size = 0;
714 lu_buf_free(&info->oti_xattr_lbuf);
718 static void osd_key_exit(const struct lu_context *ctx,
719 struct lu_context_key *key, void *data)
723 struct lu_context_key osd_key = {
724 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
725 .lct_init = osd_key_init,
726 .lct_fini = osd_key_fini,
727 .lct_exit = osd_key_exit
730 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
732 if (osd->od_cl_seq == NULL)
735 seq_client_fini(osd->od_cl_seq);
736 OBD_FREE_PTR(osd->od_cl_seq);
737 osd->od_cl_seq = NULL;
740 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
744 /* shutdown quota slave instance associated with the device */
745 if (o->od_quota_slave != NULL) {
746 /* complete all in-flight callbacks */
747 osd_sync(env, &o->od_dt_dev);
748 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
749 qsd_fini(env, o->od_quota_slave);
750 o->od_quota_slave = NULL;
753 osd_fid_fini(env, o);
758 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
760 struct osd_device *osd = arg;
762 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
765 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
767 struct osd_device *osd = arg;
769 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
770 LASSERT(newval >= SPA_MINBLOCKSIZE);
771 LASSERT(ISP2(newval));
773 osd->od_max_blksz = newval;
776 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
778 struct osd_device *osd = arg;
780 osd->od_prop_rdonly = !!newval;
783 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
784 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
786 struct osd_device *osd = arg;
788 osd->od_dnsize = newval;
792 * This function unregisters all registered callbacks. It's harmless to
793 * unregister callbacks that were never registered so it is used to safely
794 * unwind a partially completed call to osd_objset_register_callbacks().
796 static void osd_objset_unregister_callbacks(struct osd_device *o)
798 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
800 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
801 osd_xattr_changed_cb, o);
802 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
803 osd_recordsize_changed_cb, o);
804 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
805 osd_readonly_changed_cb, o);
806 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
807 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
808 osd_dnodesize_changed_cb, o);
811 if (o->arc_prune_cb != NULL) {
812 arc_remove_prune_callback(o->arc_prune_cb);
813 o->arc_prune_cb = NULL;
818 * Register the required callbacks to be notified when zfs properties
819 * are modified using the 'zfs(8)' command line utility.
821 static int osd_objset_register_callbacks(struct osd_device *o)
823 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
824 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
830 dsl_pool_config_enter(dp, FTAG);
831 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
832 osd_xattr_changed_cb, o);
836 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
837 osd_recordsize_changed_cb, o);
841 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
842 osd_readonly_changed_cb, o);
846 #ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
847 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
848 osd_dnodesize_changed_cb, o);
853 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
855 dsl_pool_config_exit(dp, FTAG);
857 osd_objset_unregister_callbacks(o);
862 static int osd_objset_open(struct osd_device *o)
864 uint64_t version = ZPL_VERSION;
865 uint64_t sa_obj, unlink_obj;
869 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
870 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
871 B_FALSE, o, &o->od_os);
874 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
880 /* Check ZFS version */
881 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
882 ZPL_VERSION_STR, 8, 1, &version);
884 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
886 * We can't return ENOENT because that would mean the objset
889 GOTO(out, rc = -EIO);
892 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
893 ZFS_SA_ATTRS, 8, 1, &sa_obj);
897 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
898 ZPL_END, &o->z_attr_table);
902 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
903 8, 1, &o->od_rootid);
905 CERROR("%s: lookup for root failed: rc = %d\n",
910 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
913 CERROR("%s: lookup for %s failed: rc = %d\n",
914 o->od_svname, ZFS_UNLINKED_SET, rc);
918 /* Check that user/group usage tracking is supported */
919 if (!dmu_objset_userused_enabled(o->od_os) ||
920 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
921 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
922 CERROR("%s: Space accounting not supported by this target, "
923 "aborting\n", o->od_svname);
924 GOTO(out, rc = -ENOTSUPP);
927 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
929 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
935 if (rc != 0 && o->od_os != NULL) {
936 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
943 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
946 char *key = osd_oti_get(env)->oti_str;
950 if (osd->od_dt_dev.dd_rdonly) {
951 CERROR("%s: someone try to free objects under "
952 "readonly mode, should be disabled.\n", osd_name(osd));
958 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
960 CWARN("%s: Cannot truncate %llu: rc = %d\n",
961 osd->od_svname, oid, rc);
965 tx = dmu_tx_create(osd->od_os);
966 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
967 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
969 rc = -dmu_tx_assign(tx, TXG_WAIT);
971 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
972 osd->od_svname, oid, rc);
976 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
977 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
978 osd->od_unlinked, key, tx);
980 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
981 osd->od_svname, oid, rc);
985 rc = -dmu_object_free(osd->od_os, oid, tx);
987 CWARN("%s: Cannot free %llu: rc = %d\n",
988 osd->od_svname, oid, rc);
1003 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
1006 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1008 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1010 while (zap_cursor_retrieve(&zc, za) == 0) {
1011 /* If cannot free the object, leave it in the unlinked set,
1012 * until the OSD is mounted again when obd_unlinked_drain()
1013 * will be called. */
1014 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1016 zap_cursor_advance(&zc);
1019 zap_cursor_fini(&zc);
1022 static int osd_mount(const struct lu_env *env,
1023 struct osd_device *o, struct lustre_cfg *cfg)
1025 char *mntdev = lustre_cfg_string(cfg, 1);
1026 char *str = lustre_cfg_string(cfg, 2);
1027 char *svname = lustre_cfg_string(cfg, 4);
1033 if (o->od_os != NULL)
1036 if (mntdev == NULL || svname == NULL)
1039 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1040 if (rc >= sizeof(o->od_mntdev))
1043 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1044 if (rc >= sizeof(o->od_svname))
1047 str = strstr(str, ":");
1049 unsigned long flags;
1051 rc = kstrtoul(str + 1, 10, &flags);
1055 if (flags & LMD_FLG_DEV_RDONLY) {
1056 o->od_dt_dev.dd_rdonly = 1;
1057 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1062 if (server_name_is_ost(o->od_svname))
1065 rc = osd_objset_open(o);
1069 o->od_xattr_in_sa = B_TRUE;
1070 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1072 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1075 o->od_root = rootdn->dn_object;
1076 osd_dnode_rele(rootdn);
1078 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1079 &o->od_userused_dn);
1083 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1084 &o->od_groupused_dn);
1088 #ifdef ZFS_PROJINHERIT
1089 if (dmu_objset_projectquota_enabled(o->od_os)) {
1090 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1091 &o->od_projectused_dn);
1092 if (rc && rc != -ENOENT)
1097 /* 1. initialize oi before any file create or file open */
1098 rc = osd_oi_init(env, o);
1102 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1105 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1107 rc = lu_site_init_finish(&o->od_site);
1111 rc = osd_objset_register_callbacks(o);
1115 rc = osd_procfs_init(o, o->od_svname);
1119 /* initialize quota slave instance */
1120 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
1122 if (IS_ERR(o->od_quota_slave)) {
1123 rc = PTR_ERR(o->od_quota_slave);
1124 o->od_quota_slave = NULL;
1128 #ifdef HAVE_DMU_USEROBJ_ACCOUNTING
1129 if (!osd_dmu_userobj_accounting_available(o))
1130 CWARN("%s: dnode accounting not enabled: "
1131 "enable feature@userobj_accounting in pool\n",
1135 /* parse mount option "noacl", and enable ACL by default */
1136 opts = lustre_cfg_string(cfg, 3);
1137 if (opts == NULL || strstr(opts, "noacl") == NULL)
1138 o->od_posix_acl = 1;
1140 osd_unlinked_drain(env, o);
1142 if (rc && o->od_os) {
1143 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
1150 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1154 if (atomic_read(&o->od_zerocopy_alloc))
1155 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1156 atomic_read(&o->od_zerocopy_alloc));
1157 if (atomic_read(&o->od_zerocopy_loan))
1158 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1159 atomic_read(&o->od_zerocopy_loan));
1160 if (atomic_read(&o->od_zerocopy_pin))
1161 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1162 atomic_read(&o->od_zerocopy_pin));
1164 if (o->od_unlinked) {
1165 osd_dnode_rele(o->od_unlinked);
1166 o->od_unlinked = NULL;
1168 if (o->od_userused_dn) {
1169 osd_dnode_rele(o->od_userused_dn);
1170 o->od_userused_dn = NULL;
1172 if (o->od_groupused_dn) {
1173 osd_dnode_rele(o->od_groupused_dn);
1174 o->od_groupused_dn = NULL;
1177 #ifdef ZFS_PROJINHERIT
1178 if (o->od_projectused_dn) {
1179 osd_dnode_rele(o->od_projectused_dn);
1180 o->od_projectused_dn = NULL;
1184 if (o->od_os != NULL) {
1185 if (!o->od_dt_dev.dd_rdonly)
1186 /* force a txg sync to get all commit callbacks */
1187 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1189 /* close the object set */
1190 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
1197 static int osd_device_init0(const struct lu_env *env,
1198 struct osd_device *o,
1199 struct lustre_cfg *cfg)
1201 struct lu_device *l = osd2lu_dev(o);
1204 /* if the module was re-loaded, env can loose its keys */
1205 rc = lu_env_refill((struct lu_env *) env);
1209 l->ld_ops = &osd_lu_ops;
1210 o->od_dt_dev.dd_ops = &osd_dt_ops;
1216 static struct lu_device *osd_device_fini(const struct lu_env *env,
1217 struct lu_device *dev);
1219 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1220 struct lu_device_type *type,
1221 struct lustre_cfg *cfg)
1223 struct osd_device *dev;
1224 struct osd_seq_list *osl;
1229 return ERR_PTR(-ENOMEM);
1231 osl = &dev->od_seq_list;
1232 INIT_LIST_HEAD(&osl->osl_seq_list);
1233 rwlock_init(&osl->osl_seq_list_lock);
1234 sema_init(&osl->osl_seq_init_sem, 1);
1236 rc = dt_device_init(&dev->od_dt_dev, type);
1238 rc = osd_device_init0(env, dev, cfg);
1240 rc = osd_mount(env, dev, cfg);
1242 osd_device_fini(env, osd2lu_dev(dev));
1245 dt_device_fini(&dev->od_dt_dev);
1248 if (unlikely(rc != 0))
1251 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1254 static struct lu_device *osd_device_free(const struct lu_env *env,
1255 struct lu_device *d)
1257 struct osd_device *o = osd_dev(d);
1260 /* XXX: make osd top device in order to release reference */
1261 d->ld_site->ls_top_dev = d;
1262 lu_site_purge(env, d->ld_site, -1);
1263 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1264 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1265 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1267 lu_site_fini(&o->od_site);
1268 dt_device_fini(&o->od_dt_dev);
1274 static struct lu_device *osd_device_fini(const struct lu_env *env,
1275 struct lu_device *d)
1277 struct osd_device *o = osd_dev(d);
1283 osd_objset_unregister_callbacks(o);
1284 if (!o->od_dt_dev.dd_rdonly) {
1285 osd_sync(env, lu2dt_dev(d));
1287 spa_get_dsl(dmu_objset_spa(o->od_os)));
1291 /* now with all the callbacks completed we can cleanup the remainings */
1292 osd_shutdown(env, o);
1293 osd_oi_fini(env, o);
1295 rc = osd_procfs_fini(o);
1297 CERROR("proc fini error %d\n", rc);
1298 RETURN(ERR_PTR(rc));
1307 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1308 const char *name, struct lu_device *next)
1314 * To be removed, setup is performed by osd_device_{init,alloc} and
1315 * cleanup is performed by osd_device_{fini,free).
1317 static int osd_process_config(const struct lu_env *env,
1318 struct lu_device *d, struct lustre_cfg *cfg)
1320 struct osd_device *o = osd_dev(d);
1324 switch(cfg->lcfg_command) {
1326 rc = osd_mount(env, o, cfg);
1329 rc = osd_shutdown(env, o);
1332 LASSERT(&o->od_dt_dev);
1333 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1334 cfg, &o->od_dt_dev);
1335 if (rc > 0 || rc == -ENOSYS) {
1336 rc = class_process_proc_param(PARAM_OST,
1337 lprocfs_osd_obd_vars,
1338 cfg, &o->od_dt_dev);
1351 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1353 struct osd_device *osd = osd_dev(d);
1357 if (osd->od_quota_slave == NULL)
1360 /* start qsd instance on recovery completion, this notifies the quota
1361 * slave code that we are about to process new requests now */
1362 rc = qsd_start(env, osd->od_quota_slave);
1367 * we use exports to track all osd users
1369 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1370 struct obd_device *obd, struct obd_uuid *cluuid,
1371 struct obd_connect_data *data, void *localdata)
1373 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1374 struct lustre_handle conn;
1378 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1380 rc = class_connect(&conn, obd, cluuid);
1384 *exp = class_conn2export(&conn);
1386 spin_lock(&obd->obd_dev_lock);
1388 spin_unlock(&obd->obd_dev_lock);
1394 * once last export (we don't count self-export) disappeared
1395 * osd can be released
1397 static int osd_obd_disconnect(struct obd_export *exp)
1399 struct obd_device *obd = exp->exp_obd;
1400 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1401 int rc, release = 0;
1404 /* Only disconnect the underlying layers on the final disconnect. */
1405 spin_lock(&obd->obd_dev_lock);
1407 if (osd->od_connects == 0)
1409 spin_unlock(&obd->obd_dev_lock);
1411 rc = class_disconnect(exp); /* bz 9811 */
1413 if (rc == 0 && release)
1414 class_manual_cleanup(obd);
1418 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1420 struct seq_server_site *ss = osd_seq_site(osd);
1424 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1427 if (unlikely(ss == NULL))
1430 OBD_ALLOC_PTR(osd->od_cl_seq);
1431 if (osd->od_cl_seq == NULL)
1434 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1435 osd->od_svname, ss->ss_server_seq);
1438 OBD_FREE_PTR(osd->od_cl_seq);
1439 osd->od_cl_seq = NULL;
1445 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1446 struct lu_device *dev)
1448 struct osd_device *osd = osd_dev(dev);
1452 if (osd->od_quota_slave != NULL) {
1453 /* set up quota slave objects */
1454 rc = qsd_prepare(env, osd->od_quota_slave);
1459 rc = osd_fid_init(env, osd);
1464 struct lu_device_operations osd_lu_ops = {
1465 .ldo_object_alloc = osd_object_alloc,
1466 .ldo_process_config = osd_process_config,
1467 .ldo_recovery_complete = osd_recovery_complete,
1468 .ldo_prepare = osd_prepare,
1471 static void osd_type_start(struct lu_device_type *t)
1475 static void osd_type_stop(struct lu_device_type *t)
1479 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1480 struct lu_fid *fid, struct md_op_data *op_data)
1482 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1484 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1487 static struct lu_device_type_operations osd_device_type_ops = {
1488 .ldto_init = osd_type_init,
1489 .ldto_fini = osd_type_fini,
1491 .ldto_start = osd_type_start,
1492 .ldto_stop = osd_type_stop,
1494 .ldto_device_alloc = osd_device_alloc,
1495 .ldto_device_free = osd_device_free,
1497 .ldto_device_init = osd_device_init,
1498 .ldto_device_fini = osd_device_fini
1501 static struct lu_device_type osd_device_type = {
1502 .ldt_tags = LU_DEVICE_DT,
1503 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1504 .ldt_ops = &osd_device_type_ops,
1505 .ldt_ctx_tags = LCT_LOCAL
1509 static struct obd_ops osd_obd_device_ops = {
1510 .o_owner = THIS_MODULE,
1511 .o_connect = osd_obd_connect,
1512 .o_disconnect = osd_obd_disconnect,
1513 .o_fid_alloc = osd_fid_alloc
1516 static int __init osd_init(void)
1520 rc = osd_options_init();
1524 rc = lu_kmem_init(osd_caches);
1528 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1529 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1531 lu_kmem_fini(osd_caches);
1535 static void __exit osd_exit(void)
1537 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1538 lu_kmem_fini(osd_caches);
1541 extern unsigned int osd_oi_count;
1542 module_param(osd_oi_count, int, 0444);
1543 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1545 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1546 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1547 MODULE_VERSION(LUSTRE_VERSION_STRING);
1548 MODULE_LICENSE("GPL");
1550 module_init(osd_init);
1551 module_exit(osd_exit);