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 <lustre_ver.h>
43 #include <libcfs/libcfs.h>
44 #include <obd_support.h>
45 #include <lustre_net.h>
47 #include <obd_class.h>
48 #include <lustre_disk.h>
49 #include <lustre_fid.h>
50 #include <lustre_param.h>
51 #include <md_object.h>
53 #include "osd_internal.h"
55 #include <sys/dnode.h>
60 #include <sys/spa_impl.h>
61 #include <sys/zfs_znode.h>
62 #include <sys/dmu_tx.h>
63 #include <sys/dmu_objset.h>
64 #include <sys/dsl_prop.h>
65 #include <sys/sa_impl.h>
68 struct lu_context_key osd_key;
70 /* Slab for OSD object allocation */
71 struct kmem_cache *osd_object_kmem;
73 /* Slab to allocate osd_zap_it */
74 struct kmem_cache *osd_zapit_cachep;
76 static struct lu_kmem_descr osd_caches[] = {
78 .ckd_cache = &osd_object_kmem,
79 .ckd_name = "zfs_osd_obj",
80 .ckd_size = sizeof(struct osd_object)
83 .ckd_cache = &osd_zapit_cachep,
84 .ckd_name = "osd_zapit_cache",
85 .ckd_size = sizeof(struct osd_zap_it)
92 static void arc_prune_func(int64_t bytes, void *private)
94 struct osd_device *od = private;
95 struct lu_site *site = &od->od_site;
99 rc = lu_env_init(&env, LCT_SHRINKER);
101 CERROR("%s: can't initialize shrinker env: rc = %d\n",
106 lu_site_purge(&env, site, (bytes >> 10));
112 * Concurrency: doesn't access mutable data
114 static int osd_root_get(const struct lu_env *env,
115 struct dt_device *dev, struct lu_fid *f)
117 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
122 * OSD object methods.
126 * Concurrency: shouldn't matter.
128 static void osd_trans_commit_cb(void *cb_data, int error)
130 struct osd_thandle *oh = cb_data;
131 struct thandle *th = &oh->ot_super;
132 struct osd_device *osd = osd_dt_dev(th->th_dev);
133 struct lu_device *lud = &th->th_dev->dd_lu_dev;
134 struct dt_txn_commit_cb *dcb, *tmp;
139 if (error == ECANCELED)
140 CWARN("%s: transaction @0x%p was aborted\n",
141 osd_dt_dev(th->th_dev)->od_svname, th);
143 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
144 osd_dt_dev(th->th_dev)->od_svname, th, error);
147 dt_txn_hook_commit(th);
149 /* call per-transaction callbacks if any */
150 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage)
151 dcb->dcb_func(NULL, th, dcb, error);
153 /* Unlike ldiskfs, zfs updates space accounting at commit time.
154 * As a consequence, op_end is called only now to inform the quota slave
155 * component that reserved quota space is now accounted in usage and
156 * should be released. Quota space won't be adjusted at this point since
157 * we can't provide a suitable environment. It will be performed
158 * asynchronously by a lquota thread. */
159 qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);
163 lu_context_exit(&th->th_ctx);
164 lu_context_fini(&th->th_ctx);
170 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
172 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
175 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
176 LASSERT(&dcb->dcb_func != NULL);
177 if (dcb->dcb_flags & DCB_TRANS_STOP)
178 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
180 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
186 * Concurrency: shouldn't matter.
188 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
191 struct osd_thandle *oh;
195 oh = container_of0(th, struct osd_thandle, ot_super);
199 rc = dt_txn_hook_start(env, d, th);
203 if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
204 /* Unlike ldiskfs, ZFS checks for available space and returns
205 * -ENOSPC when assigning txg */
208 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
209 if (unlikely(rc != 0)) {
210 struct osd_device *osd = osd_dt_dev(d);
211 /* dmu will call commit callback with error code during abort */
212 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
213 CERROR("%s: failed to start transaction due to ENOSPC"
214 "\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);
230 static int osd_unlinked_object_free(struct osd_device *osd, uint64_t oid);
232 static void osd_unlinked_list_emptify(struct osd_device *osd,
233 struct list_head *list, bool free)
235 struct osd_object *obj;
238 while (!list_empty(list)) {
239 obj = list_entry(list->next,
240 struct osd_object, oo_unlinked_linkage);
241 LASSERT(obj->oo_dn != NULL);
242 oid = obj->oo_dn->dn_object;
244 list_del_init(&obj->oo_unlinked_linkage);
246 (void)osd_unlinked_object_free(osd, oid);
250 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
252 struct dt_txn_commit_cb *dcb;
253 struct dt_txn_commit_cb *tmp;
255 /* call per-transaction stop callbacks if any */
256 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
258 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
259 "commit callback entry: magic=%x name='%s'\n",
260 dcb->dcb_magic, dcb->dcb_name);
261 list_del_init(&dcb->dcb_linkage);
262 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
267 * Concurrency: shouldn't matter.
269 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
272 struct osd_device *osd = osd_dt_dev(th->th_dev);
273 bool sync = (th->th_sync != 0);
274 struct osd_thandle *oh;
275 struct list_head unlinked;
280 oh = container_of0(th, struct osd_thandle, ot_super);
281 INIT_LIST_HEAD(&unlinked);
282 list_splice_init(&oh->ot_unlinked_list, &unlinked);
283 /* reset OI cache for safety */
284 osd_oti_get(env)->oti_ins_cache_used = 0;
286 if (oh->ot_assigned == 0) {
288 dmu_tx_abort(oh->ot_tx);
289 osd_object_sa_dirty_rele(oh);
290 osd_unlinked_list_emptify(osd, &unlinked, false);
291 /* there won't be any commit, release reserved quota space now,
293 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
298 /* When doing our own inode accounting, the ZAPs storing per-uid/gid
299 * usage are updated at operation execution time, so we should call
300 * qsd_op_end() straight away. Otherwise (for blk accounting maintained
301 * by ZFS and when #inode is estimated from #blks) accounting is updated
302 * at commit time and the call to qsd_op_end() must be delayed */
303 if (oh->ot_quota_trans.lqt_id_cnt > 0 &&
304 !oh->ot_quota_trans.lqt_ids[0].lqi_is_blk &&
305 !osd->od_quota_iused_est)
306 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
308 rc = dt_txn_hook_stop(env, th);
310 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
313 osd_trans_stop_cb(oh, rc);
316 txg = oh->ot_tx->tx_txg;
318 osd_object_sa_dirty_rele(oh);
319 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
320 * by osd_trans_commit_cb already. */
321 dmu_tx_commit(oh->ot_tx);
323 osd_unlinked_list_emptify(osd, &unlinked, true);
326 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
331 static struct thandle *osd_trans_create(const struct lu_env *env,
332 struct dt_device *dt)
334 struct osd_device *osd = osd_dt_dev(dt);
335 struct osd_thandle *oh;
340 tx = dmu_tx_create(osd->od_os);
342 RETURN(ERR_PTR(-ENOMEM));
344 /* alloc callback data */
348 RETURN(ERR_PTR(-ENOMEM));
352 INIT_LIST_HEAD(&oh->ot_dcb_list);
353 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
354 INIT_LIST_HEAD(&oh->ot_unlinked_list);
355 INIT_LIST_HEAD(&oh->ot_sa_list);
356 sema_init(&oh->ot_sa_lock, 1);
357 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
361 th->th_tags = LCT_TX_HANDLE;
365 /* Estimate the total number of objects from a number of blocks */
366 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
367 uint64_t nrblocks, uint64_t est_maxblockshift)
369 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
372 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
373 * bump the free dnode estimate to assume blocks at least 64KB in
374 * case of a directory-heavy MDT (at 32KB/directory).
376 if (est_maxblockshift < 16) {
377 nrblocks >>= (16 - est_maxblockshift);
378 est_maxblockshift = 16;
382 * Estimate the total number of dnodes from the total blocks count
383 * and the space used per dnode. Since we don't know the overhead
384 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
385 * just using DNODE_SHIFT isn't going to give a good estimate.
386 * Instead, compute the current average space usage per dnode, with
387 * an upper and lower cap to avoid unrealistic estimates..
389 * In case there aren't many dnodes or blocks used yet, add a small
390 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
391 * gradually disappears as the number of real dnodes grows. It also
392 * avoids the need to check for divide-by-zero computing dn_per_block.
394 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
395 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
397 est_usedblocks = (usedbytes >> est_maxblockshift) +
398 (OSD_DNODE_EST_COUNT >> OSD_DNODE_EST_BLKSHIFT);
399 est_usedobjs = usedobjs + OSD_DNODE_EST_COUNT;
401 if (est_usedobjs <= est_usedblocks) {
403 * Average space/dnode more than maximum block size, use max
404 * block size to estimate free dnodes from adjusted free blocks
405 * count. OSTs typically use multiple blocks per dnode so this
408 est_totobjs = nrblocks;
410 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
412 * Average space/dnode smaller than min dnode size (probably
413 * due to metadnode compression), use min dnode size to
414 * estimate object count. MDTs may use only one block per node
415 * so this case applies.
417 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
421 * Between the extremes, use average space per existing dnode
422 * to compute the number of dnodes that will fit into nrblocks:
424 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
426 * this may overflow 64 bits or become 0 if not handled well.
428 * We know nrblocks is below 2^(64 - blkbits) bits, and
429 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
430 * which means that multiplying them may get as large as
431 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
433 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
434 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
435 * the limit checks above, so we can safely compute this first.
436 * We care more about accuracy on the MDT (many dnodes/block)
437 * which is good because this is where truncation errors are
438 * smallest. Since both nrblocks and dn_per_block are a
439 * function of blkbits, their product is at most:
441 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
443 * so we can safely use 7 bits to compute a fixed-point
444 * fraction and est_totobjs can still fit in 64 bits.
446 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
448 est_totobjs = (nrblocks * dn_per_block) >> 7;
453 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
455 struct objset *os = osd->od_os;
456 uint64_t usedbytes, availbytes, usedobjs, availobjs;
457 uint64_t est_availobjs;
461 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
463 memset(osfs, 0, sizeof(*osfs));
465 /* We're a zfs filesystem. */
466 osfs->os_type = UBERBLOCK_MAGIC;
469 * ZFS allows multiple block sizes. For statfs, Linux makes no
470 * proper distinction between bsize and frsize. For calculations
471 * of free and used blocks incorrectly uses bsize instead of frsize,
472 * but bsize is also used as the optimal blocksize. We return the
473 * largest possible block size as IO size for the optimum performance
474 * and scale the free and used blocks count appropriately.
476 osfs->os_bsize = osd->od_max_blksz;
477 bshift = fls64(osfs->os_bsize) - 1;
479 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
480 osfs->os_bfree = availbytes >> bshift;
481 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
483 /* Take replication (i.e. number of copies) into account */
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)
618 struct osd_device *osd = osd_dt_dev(d);
619 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
620 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
621 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
625 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
627 struct osd_device *osd = osd_dt_dev(dev);
628 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
631 mutex_enter(&tx->tx_sync_lock);
632 txg = tx->tx_open_txg + 1;
633 if (tx->tx_quiesce_txg_waiting < txg) {
634 tx->tx_quiesce_txg_waiting = txg;
635 cv_broadcast(&tx->tx_quiesce_more_cv);
637 mutex_exit(&tx->tx_sync_lock);
643 * Concurrency: shouldn't matter.
645 static int osd_ro(const struct lu_env *env, struct dt_device *d)
647 struct osd_device *osd = osd_dt_dev(d);
650 CERROR("%s: *** setting device %s read-only ***\n",
651 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
652 osd->od_dev_set_rdonly = 1;
653 spa_freeze(dmu_objset_spa(osd->od_os));
658 static struct dt_device_operations osd_dt_ops = {
659 .dt_root_get = osd_root_get,
660 .dt_statfs = osd_statfs,
661 .dt_trans_create = osd_trans_create,
662 .dt_trans_start = osd_trans_start,
663 .dt_trans_stop = osd_trans_stop,
664 .dt_trans_cb_add = osd_trans_cb_add,
665 .dt_conf_get = osd_conf_get,
667 .dt_commit_async = osd_commit_async,
672 * DMU OSD device type methods
674 static int osd_type_init(struct lu_device_type *t)
676 LU_CONTEXT_KEY_INIT(&osd_key);
677 return lu_context_key_register(&osd_key);
680 static void osd_type_fini(struct lu_device_type *t)
682 lu_context_key_degister(&osd_key);
685 static void *osd_key_init(const struct lu_context *ctx,
686 struct lu_context_key *key)
688 struct osd_thread_info *info;
692 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
694 info = ERR_PTR(-ENOMEM);
698 static void osd_key_fini(const struct lu_context *ctx,
699 struct lu_context_key *key, void *data)
701 struct osd_thread_info *info = data;
702 struct osd_idmap_cache *idc = info->oti_ins_cache;
705 LASSERT(info->oti_ins_cache_size > 0);
706 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
707 info->oti_ins_cache = NULL;
708 info->oti_ins_cache_size = 0;
713 static void osd_key_exit(const struct lu_context *ctx,
714 struct lu_context_key *key, void *data)
718 struct lu_context_key osd_key = {
719 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
720 .lct_init = osd_key_init,
721 .lct_fini = osd_key_fini,
722 .lct_exit = osd_key_exit
725 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
727 if (osd->od_cl_seq == NULL)
730 seq_client_fini(osd->od_cl_seq);
731 OBD_FREE_PTR(osd->od_cl_seq);
732 osd->od_cl_seq = NULL;
735 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
739 /* shutdown quota slave instance associated with the device */
740 if (o->od_quota_slave != NULL) {
741 qsd_fini(env, o->od_quota_slave);
742 o->od_quota_slave = NULL;
745 osd_fid_fini(env, o);
750 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
752 struct osd_device *osd = arg;
754 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
757 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
759 struct osd_device *osd = arg;
761 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
762 LASSERT(newval >= SPA_MINBLOCKSIZE);
763 LASSERT(ISP2(newval));
765 osd->od_max_blksz = newval;
768 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
770 struct osd_device *osd = arg;
772 osd->od_prop_rdonly = !!newval;
776 * This function unregisters all registered callbacks. It's harmless to
777 * unregister callbacks that were never registered so it is used to safely
778 * unwind a partially completed call to osd_objset_register_callbacks().
780 static void osd_objset_unregister_callbacks(struct osd_device *o)
782 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
784 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
785 osd_xattr_changed_cb, o);
786 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
787 osd_recordsize_changed_cb, o);
788 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
789 osd_readonly_changed_cb, o);
791 if (o->arc_prune_cb != NULL) {
792 arc_remove_prune_callback(o->arc_prune_cb);
793 o->arc_prune_cb = NULL;
798 * Register the required callbacks to be notified when zfs properties
799 * are modified using the 'zfs(8)' command line utility.
801 static int osd_objset_register_callbacks(struct osd_device *o)
803 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
804 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
810 dsl_pool_config_enter(dp, FTAG);
811 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
812 osd_xattr_changed_cb, o);
816 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
817 osd_recordsize_changed_cb, o);
821 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
822 osd_readonly_changed_cb, o);
826 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
828 dsl_pool_config_exit(dp, FTAG);
830 osd_objset_unregister_callbacks(o);
835 static int osd_objset_open(struct osd_device *o)
837 uint64_t version = ZPL_VERSION;
842 rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os);
844 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
849 /* Check ZFS version */
850 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
851 ZPL_VERSION_STR, 8, 1, &version);
853 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
855 * We can't return ENOENT because that would mean the objset
858 GOTO(out, rc = -EIO);
861 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
862 ZFS_SA_ATTRS, 8, 1, &sa_obj);
866 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
867 ZPL_END, &o->z_attr_table);
871 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
872 8, 1, &o->od_rootid);
874 CERROR("%s: lookup for root failed: rc = %d\n",
879 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
880 8, 1, &o->od_unlinkedid);
882 CERROR("%s: lookup for %s failed: rc = %d\n",
883 o->od_svname, ZFS_UNLINKED_SET, rc);
887 /* Check that user/group usage tracking is supported */
888 if (!dmu_objset_userused_enabled(o->od_os) ||
889 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
890 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
891 CERROR("%s: Space accounting not supported by this target, "
892 "aborting\n", o->od_svname);
893 GOTO(out, rc = -ENOTSUPP);
897 if (rc != 0 && o->od_os != NULL) {
898 dmu_objset_disown(o->od_os, o);
906 osd_unlinked_object_free(struct osd_device *osd, uint64_t oid)
911 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
913 CWARN("%s: Cannot truncate %llu: rc = %d\n",
914 osd->od_svname, oid, rc);
918 tx = dmu_tx_create(osd->od_os);
919 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
920 dmu_tx_hold_zap(tx, osd->od_unlinkedid, FALSE, NULL);
921 rc = -dmu_tx_assign(tx, TXG_WAIT);
923 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
924 osd->od_svname, oid, rc);
928 rc = -zap_remove_int(osd->od_os, osd->od_unlinkedid, oid, tx);
930 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
931 osd->od_svname, oid, rc);
935 rc = -dmu_object_free(osd->od_os, oid, tx);
937 CWARN("%s: Cannot free %llu: rc = %d\n",
938 osd->od_svname, oid, rc);
953 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
956 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
958 zap_cursor_init(&zc, osd->od_os, osd->od_unlinkedid);
960 while (zap_cursor_retrieve(&zc, za) == 0) {
961 /* If cannot free the object, leave it in the unlinked set,
962 * until the OSD is mounted again when obd_unlinked_drain()
964 if (osd_unlinked_object_free(osd, za->za_first_integer) != 0)
966 zap_cursor_advance(&zc);
969 zap_cursor_fini(&zc);
972 static int osd_mount(const struct lu_env *env,
973 struct osd_device *o, struct lustre_cfg *cfg)
975 char *mntdev = lustre_cfg_string(cfg, 1);
976 char *svname = lustre_cfg_string(cfg, 4);
982 if (o->od_os != NULL)
985 if (mntdev == NULL || svname == NULL)
988 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
989 if (rc >= sizeof(o->od_mntdev))
992 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
993 if (rc >= sizeof(o->od_svname))
996 if (server_name_is_ost(o->od_svname))
999 rc = osd_objset_open(o);
1003 o->od_xattr_in_sa = B_TRUE;
1004 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1006 rc = osd_objset_register_callbacks(o);
1010 rc = __osd_obj2dnode(env, o->od_os, o->od_rootid, &rootdn);
1014 o->od_root = rootdn->dn_object;
1015 osd_dnode_rele(rootdn);
1017 /* 1. initialize oi before any file create or file open */
1018 rc = osd_oi_init(env, o);
1022 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1025 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1027 rc = lu_site_init_finish(&o->od_site);
1031 /* Use our own ZAP for inode accounting by default, this can be changed
1032 * via procfs to estimate the inode usage from the block usage */
1033 o->od_quota_iused_est = 0;
1035 rc = osd_procfs_init(o, o->od_svname);
1039 /* initialize quota slave instance */
1040 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
1042 if (IS_ERR(o->od_quota_slave)) {
1043 rc = PTR_ERR(o->od_quota_slave);
1044 o->od_quota_slave = NULL;
1048 /* parse mount option "noacl", and enable ACL by default */
1049 opts = lustre_cfg_string(cfg, 3);
1050 if (opts == NULL || strstr(opts, "noacl") == NULL)
1051 o->od_posix_acl = 1;
1053 osd_unlinked_drain(env, o);
1056 dmu_objset_disown(o->od_os, o);
1063 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1067 if (atomic_read(&o->od_zerocopy_alloc))
1068 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1069 atomic_read(&o->od_zerocopy_alloc));
1070 if (atomic_read(&o->od_zerocopy_loan))
1071 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1072 atomic_read(&o->od_zerocopy_loan));
1073 if (atomic_read(&o->od_zerocopy_pin))
1074 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1075 atomic_read(&o->od_zerocopy_pin));
1077 if (o->od_os != NULL) {
1078 /* force a txg sync to get all commit callbacks */
1079 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1081 /* close the object set */
1082 dmu_objset_disown(o->od_os, o);
1090 static int osd_device_init0(const struct lu_env *env,
1091 struct osd_device *o,
1092 struct lustre_cfg *cfg)
1094 struct lu_device *l = osd2lu_dev(o);
1097 /* if the module was re-loaded, env can loose its keys */
1098 rc = lu_env_refill((struct lu_env *) env);
1102 l->ld_ops = &osd_lu_ops;
1103 o->od_dt_dev.dd_ops = &osd_dt_ops;
1109 static struct lu_device *osd_device_fini(const struct lu_env *env,
1110 struct lu_device *dev);
1112 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1113 struct lu_device_type *type,
1114 struct lustre_cfg *cfg)
1116 struct osd_device *dev;
1117 struct osd_seq_list *osl;
1122 return ERR_PTR(-ENOMEM);
1124 osl = &dev->od_seq_list;
1125 INIT_LIST_HEAD(&osl->osl_seq_list);
1126 rwlock_init(&osl->osl_seq_list_lock);
1127 sema_init(&osl->osl_seq_init_sem, 1);
1129 rc = dt_device_init(&dev->od_dt_dev, type);
1131 rc = osd_device_init0(env, dev, cfg);
1133 rc = osd_mount(env, dev, cfg);
1135 osd_device_fini(env, osd2lu_dev(dev));
1138 dt_device_fini(&dev->od_dt_dev);
1141 if (unlikely(rc != 0))
1144 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1147 static struct lu_device *osd_device_free(const struct lu_env *env,
1148 struct lu_device *d)
1150 struct osd_device *o = osd_dev(d);
1153 /* XXX: make osd top device in order to release reference */
1154 d->ld_site->ls_top_dev = d;
1155 lu_site_purge(env, d->ld_site, -1);
1156 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1157 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1158 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1160 lu_site_fini(&o->od_site);
1161 dt_device_fini(&o->od_dt_dev);
1167 static struct lu_device *osd_device_fini(const struct lu_env *env,
1168 struct lu_device *d)
1170 struct osd_device *o = osd_dev(d);
1175 osd_shutdown(env, o);
1176 osd_oi_fini(env, o);
1179 osd_objset_unregister_callbacks(o);
1180 osd_sync(env, lu2dt_dev(d));
1181 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
1184 rc = osd_procfs_fini(o);
1186 CERROR("proc fini error %d\n", rc);
1187 RETURN(ERR_PTR(rc));
1196 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1197 const char *name, struct lu_device *next)
1203 * To be removed, setup is performed by osd_device_{init,alloc} and
1204 * cleanup is performed by osd_device_{fini,free).
1206 static int osd_process_config(const struct lu_env *env,
1207 struct lu_device *d, struct lustre_cfg *cfg)
1209 struct osd_device *o = osd_dev(d);
1213 switch(cfg->lcfg_command) {
1215 rc = osd_mount(env, o, cfg);
1218 rc = osd_shutdown(env, o);
1221 LASSERT(&o->od_dt_dev);
1222 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1223 cfg, &o->od_dt_dev);
1224 if (rc > 0 || rc == -ENOSYS)
1225 rc = class_process_proc_param(PARAM_OST,
1226 lprocfs_osd_obd_vars,
1227 cfg, &o->od_dt_dev);
1237 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1239 struct osd_device *osd = osd_dev(d);
1243 if (osd->od_quota_slave == NULL)
1246 /* start qsd instance on recovery completion, this notifies the quota
1247 * slave code that we are about to process new requests now */
1248 rc = qsd_start(env, osd->od_quota_slave);
1253 * we use exports to track all osd users
1255 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1256 struct obd_device *obd, struct obd_uuid *cluuid,
1257 struct obd_connect_data *data, void *localdata)
1259 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1260 struct lustre_handle conn;
1264 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1266 rc = class_connect(&conn, obd, cluuid);
1270 *exp = class_conn2export(&conn);
1272 spin_lock(&obd->obd_dev_lock);
1274 spin_unlock(&obd->obd_dev_lock);
1280 * once last export (we don't count self-export) disappeared
1281 * osd can be released
1283 static int osd_obd_disconnect(struct obd_export *exp)
1285 struct obd_device *obd = exp->exp_obd;
1286 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1287 int rc, release = 0;
1290 /* Only disconnect the underlying layers on the final disconnect. */
1291 spin_lock(&obd->obd_dev_lock);
1293 if (osd->od_connects == 0)
1295 spin_unlock(&obd->obd_dev_lock);
1297 rc = class_disconnect(exp); /* bz 9811 */
1299 if (rc == 0 && release)
1300 class_manual_cleanup(obd);
1304 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1306 struct seq_server_site *ss = osd_seq_site(osd);
1310 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1313 if (unlikely(ss == NULL))
1316 OBD_ALLOC_PTR(osd->od_cl_seq);
1317 if (osd->od_cl_seq == NULL)
1320 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1321 osd->od_svname, ss->ss_server_seq);
1324 OBD_FREE_PTR(osd->od_cl_seq);
1325 osd->od_cl_seq = NULL;
1331 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1332 struct lu_device *dev)
1334 struct osd_device *osd = osd_dev(dev);
1338 if (osd->od_quota_slave != NULL) {
1339 /* set up quota slave objects */
1340 rc = qsd_prepare(env, osd->od_quota_slave);
1345 rc = osd_fid_init(env, osd);
1350 struct lu_device_operations osd_lu_ops = {
1351 .ldo_object_alloc = osd_object_alloc,
1352 .ldo_process_config = osd_process_config,
1353 .ldo_recovery_complete = osd_recovery_complete,
1354 .ldo_prepare = osd_prepare,
1357 static void osd_type_start(struct lu_device_type *t)
1361 static void osd_type_stop(struct lu_device_type *t)
1365 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1366 struct lu_fid *fid, struct md_op_data *op_data)
1368 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1370 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1373 static struct lu_device_type_operations osd_device_type_ops = {
1374 .ldto_init = osd_type_init,
1375 .ldto_fini = osd_type_fini,
1377 .ldto_start = osd_type_start,
1378 .ldto_stop = osd_type_stop,
1380 .ldto_device_alloc = osd_device_alloc,
1381 .ldto_device_free = osd_device_free,
1383 .ldto_device_init = osd_device_init,
1384 .ldto_device_fini = osd_device_fini
1387 static struct lu_device_type osd_device_type = {
1388 .ldt_tags = LU_DEVICE_DT,
1389 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1390 .ldt_ops = &osd_device_type_ops,
1391 .ldt_ctx_tags = LCT_LOCAL
1395 static struct obd_ops osd_obd_device_ops = {
1396 .o_owner = THIS_MODULE,
1397 .o_connect = osd_obd_connect,
1398 .o_disconnect = osd_obd_disconnect,
1399 .o_fid_alloc = osd_fid_alloc
1402 static int __init osd_init(void)
1406 rc = osd_options_init();
1410 rc = lu_kmem_init(osd_caches);
1414 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1415 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1417 lu_kmem_fini(osd_caches);
1421 static void __exit osd_exit(void)
1423 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1424 lu_kmem_fini(osd_caches);
1427 extern unsigned int osd_oi_count;
1428 module_param(osd_oi_count, int, 0444);
1429 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1431 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1432 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1433 MODULE_VERSION(LUSTRE_VERSION_STRING);
1434 MODULE_LICENSE("GPL");
1436 module_init(osd_init);
1437 module_exit(osd_exit);