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 <uapi/linux/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 LASSERT(site->ls_obj_hash);
101 rc = lu_env_init(&env, LCT_SHRINKER);
103 CERROR("%s: can't initialize shrinker env: rc = %d\n",
108 lu_site_purge(&env, site, (bytes >> 10));
114 * Concurrency: doesn't access mutable data
116 static int osd_root_get(const struct lu_env *env,
117 struct dt_device *dev, struct lu_fid *f)
119 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
124 * OSD object methods.
128 * Concurrency: shouldn't matter.
130 static void osd_trans_commit_cb(void *cb_data, int error)
132 struct osd_thandle *oh = cb_data;
133 struct thandle *th = &oh->ot_super;
134 struct osd_device *osd = osd_dt_dev(th->th_dev);
135 struct lu_device *lud = &th->th_dev->dd_lu_dev;
136 struct dt_txn_commit_cb *dcb, *tmp;
141 if (error == ECANCELED)
142 CWARN("%s: transaction @0x%p was aborted\n",
143 osd_dt_dev(th->th_dev)->od_svname, th);
145 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
146 osd_dt_dev(th->th_dev)->od_svname, th, error);
149 dt_txn_hook_commit(th);
151 /* call per-transaction callbacks if any */
152 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) {
153 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
154 "commit callback entry: magic=%x name='%s'\n",
155 dcb->dcb_magic, dcb->dcb_name);
156 list_del_init(&dcb->dcb_linkage);
157 dcb->dcb_func(NULL, th, dcb, error);
160 /* Unlike ldiskfs, zfs updates space accounting at commit time.
161 * As a consequence, op_end is called only now to inform the quota slave
162 * component that reserved quota space is now accounted in usage and
163 * should be released. Quota space won't be adjusted at this point since
164 * we can't provide a suitable environment. It will be performed
165 * asynchronously by a lquota thread. */
166 qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);
170 lu_context_exit(&th->th_ctx);
171 lu_context_fini(&th->th_ctx);
177 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
179 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
182 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
183 LASSERT(&dcb->dcb_func != NULL);
184 if (dcb->dcb_flags & DCB_TRANS_STOP)
185 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
187 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
193 * Concurrency: shouldn't matter.
195 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
198 struct osd_thandle *oh;
202 oh = container_of0(th, struct osd_thandle, ot_super);
206 rc = dt_txn_hook_start(env, d, th);
210 if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
211 /* Unlike ldiskfs, ZFS checks for available space and returns
212 * -ENOSPC when assigning txg */
215 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
216 if (unlikely(rc != 0)) {
217 struct osd_device *osd = osd_dt_dev(d);
218 /* dmu will call commit callback with error code during abort */
219 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
220 CERROR("%s: failed to start transaction due to ENOSPC"
221 "\n", osd->od_svname);
223 CERROR("%s: can't assign tx: rc = %d\n",
226 /* add commit callback */
227 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
229 lu_context_init(&th->th_ctx, th->th_tags);
230 lu_context_enter(&th->th_ctx);
231 lu_device_get(&d->dd_lu_dev);
237 static void osd_unlinked_list_emptify(const struct lu_env *env,
238 struct osd_device *osd,
239 struct list_head *list, bool free)
241 struct osd_object *obj;
244 while (!list_empty(list)) {
245 obj = list_entry(list->next,
246 struct osd_object, oo_unlinked_linkage);
247 LASSERT(obj->oo_dn != NULL);
248 oid = obj->oo_dn->dn_object;
250 list_del_init(&obj->oo_unlinked_linkage);
252 (void)osd_unlinked_object_free(env, osd, oid);
256 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
258 struct dt_txn_commit_cb *dcb;
259 struct dt_txn_commit_cb *tmp;
261 /* call per-transaction stop callbacks if any */
262 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
264 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
265 "commit callback entry: magic=%x name='%s'\n",
266 dcb->dcb_magic, dcb->dcb_name);
267 list_del_init(&dcb->dcb_linkage);
268 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
273 * Concurrency: shouldn't matter.
275 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
278 struct osd_device *osd = osd_dt_dev(th->th_dev);
279 bool sync = (th->th_sync != 0);
280 struct osd_thandle *oh;
281 struct list_head unlinked;
286 oh = container_of0(th, struct osd_thandle, ot_super);
287 INIT_LIST_HEAD(&unlinked);
288 list_splice_init(&oh->ot_unlinked_list, &unlinked);
289 /* reset OI cache for safety */
290 osd_oti_get(env)->oti_ins_cache_used = 0;
292 if (oh->ot_assigned == 0) {
294 dmu_tx_abort(oh->ot_tx);
295 osd_object_sa_dirty_rele(oh);
296 osd_unlinked_list_emptify(env, osd, &unlinked, false);
297 /* there won't be any commit, release reserved quota space now,
299 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
304 rc = dt_txn_hook_stop(env, th);
306 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
309 osd_trans_stop_cb(oh, rc);
312 txg = oh->ot_tx->tx_txg;
314 osd_object_sa_dirty_rele(oh);
315 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
316 * by osd_trans_commit_cb already. */
317 dmu_tx_commit(oh->ot_tx);
319 osd_unlinked_list_emptify(env, osd, &unlinked, true);
322 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
327 static struct thandle *osd_trans_create(const struct lu_env *env,
328 struct dt_device *dt)
330 struct osd_device *osd = osd_dt_dev(dt);
331 struct osd_thandle *oh;
337 CERROR("%s: someone try to start transaction under "
338 "readonly mode, should be disabled.\n",
339 osd_name(osd_dt_dev(dt)));
341 RETURN(ERR_PTR(-EROFS));
344 tx = dmu_tx_create(osd->od_os);
346 RETURN(ERR_PTR(-ENOMEM));
348 /* alloc callback data */
352 RETURN(ERR_PTR(-ENOMEM));
356 INIT_LIST_HEAD(&oh->ot_dcb_list);
357 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
358 INIT_LIST_HEAD(&oh->ot_unlinked_list);
359 INIT_LIST_HEAD(&oh->ot_sa_list);
360 sema_init(&oh->ot_sa_lock, 1);
361 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
365 th->th_tags = LCT_TX_HANDLE;
369 /* Estimate the total number of objects from a number of blocks */
370 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
371 uint64_t nrblocks, uint64_t est_maxblockshift)
373 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
376 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
377 * bump the free dnode estimate to assume blocks at least 64KB in
378 * case of a directory-heavy MDT (at 32KB/directory).
380 if (est_maxblockshift < 16) {
381 nrblocks >>= (16 - est_maxblockshift);
382 est_maxblockshift = 16;
386 * Estimate the total number of dnodes from the total blocks count
387 * and the space used per dnode. Since we don't know the overhead
388 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
389 * just using DNODE_SHIFT isn't going to give a good estimate.
390 * Instead, compute the current average space usage per dnode, with
391 * an upper and lower cap to avoid unrealistic estimates..
393 * In case there aren't many dnodes or blocks used yet, add a small
394 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
395 * gradually disappears as the number of real dnodes grows. It also
396 * avoids the need to check for divide-by-zero computing dn_per_block.
398 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
399 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
401 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
402 usedbytes) >> est_maxblockshift;
403 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
405 if (est_usedobjs <= est_usedblocks) {
407 * Average space/dnode more than maximum block size, use max
408 * block size to estimate free dnodes from adjusted free blocks
409 * count. OSTs typically use multiple blocks per dnode so this
412 est_totobjs = nrblocks;
414 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
416 * Average space/dnode smaller than min dnode size (probably
417 * due to metadnode compression), use min dnode size to
418 * estimate object count. MDTs may use only one block per node
419 * so this case applies.
421 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
425 * Between the extremes, use average space per existing dnode
426 * to compute the number of dnodes that will fit into nrblocks:
428 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
430 * this may overflow 64 bits or become 0 if not handled well.
432 * We know nrblocks is below 2^(64 - blkbits) bits, and
433 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
434 * which means that multiplying them may get as large as
435 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
437 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
438 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
439 * the limit checks above, so we can safely compute this first.
440 * We care more about accuracy on the MDT (many dnodes/block)
441 * which is good because this is where truncation errors are
442 * smallest. Since both nrblocks and dn_per_block are a
443 * function of blkbits, their product is at most:
445 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
447 * so we can safely use 7 bits to compute a fixed-point
448 * fraction and est_totobjs can still fit in 64 bits.
450 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
452 est_totobjs = (nrblocks * dn_per_block) >> 7;
457 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
459 struct objset *os = osd->od_os;
460 uint64_t usedbytes, availbytes, usedobjs, availobjs;
461 uint64_t est_availobjs;
465 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
467 memset(osfs, 0, sizeof(*osfs));
469 /* We're a zfs filesystem. */
470 osfs->os_type = UBERBLOCK_MAGIC;
473 * ZFS allows multiple block sizes. For statfs, Linux makes no
474 * proper distinction between bsize and frsize. For calculations
475 * of free and used blocks incorrectly uses bsize instead of frsize,
476 * but bsize is also used as the optimal blocksize. We return the
477 * largest possible block size as IO size for the optimum performance
478 * and scale the free and used blocks count appropriately.
480 osfs->os_bsize = osd->od_max_blksz;
481 bshift = fls64(osfs->os_bsize) - 1;
483 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
484 osfs->os_bfree = availbytes >> bshift;
485 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
487 /* Take replication (i.e. number of copies) into account */
488 if (os->os_copies != 0)
489 osfs->os_bavail /= os->os_copies;
492 * Reserve some space so we don't run into ENOSPC due to grants not
493 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
494 * Rather than report this via os_bavail (which makes users unhappy if
495 * they can't fill the filesystem 100%), reduce os_blocks as well.
497 * Reserve 0.78% of total space, at least 16MB for small filesystems,
498 * for internal files to be created/unlinked when space is tight.
500 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
501 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
502 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
503 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
505 osfs->os_blocks -= reserved;
506 osfs->os_bfree -= min(reserved, osfs->os_bfree);
507 osfs->os_bavail -= min(reserved, osfs->os_bavail);
510 * The availobjs value returned from dmu_objset_space() is largely
511 * useless, since it reports the number of objects that might
512 * theoretically still fit into the dataset, independent of minor
513 * issues like how much space is actually available in the pool.
514 * Compute a better estimate in udmu_objs_count_estimate().
516 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
517 osfs->os_bfree, bshift);
519 osfs->os_ffree = min(availobjs, est_availobjs);
520 osfs->os_files = osfs->os_ffree + usedobjs;
522 /* ZFS XXX: fill in backing dataset FSID/UUID
523 memcpy(osfs->os_fsid, .... );*/
525 osfs->os_namelen = MAXNAMELEN;
526 osfs->os_maxbytes = OBD_OBJECT_EOF;
528 if (!spa_writeable(dmu_objset_spa(os)) ||
529 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
530 osfs->os_state |= OS_STATE_READONLY;
536 * Concurrency: shouldn't matter.
538 int osd_statfs(const struct lu_env *env, struct dt_device *d,
539 struct obd_statfs *osfs)
544 rc = osd_objset_statfs(osd_dt_dev(d), osfs);
545 if (unlikely(rc != 0))
548 osfs->os_bavail -= min_t(u64,
549 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
554 static int osd_blk_insert_cost(struct osd_device *osd)
556 int max_blockshift, nr_blkptrshift, bshift;
558 /* max_blockshift is the log2 of the number of blocks needed to reach
559 * the maximum filesize (that's to say 2^64) */
560 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
561 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
563 /* nr_blkptrshift is the log2 of the number of block pointers that can
564 * be stored in an indirect block */
565 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
566 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
568 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
569 * tree. We add +1 for rounding purpose.
570 * The tree depth times the indirect block size gives us the maximum
571 * cost of inserting a block in the tree */
572 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
576 * Concurrency: doesn't access mutable data.
578 static void osd_conf_get(const struct lu_env *env,
579 const struct dt_device *dev,
580 struct dt_device_param *param)
582 struct osd_device *osd = osd_dt_dev(dev);
585 * XXX should be taken from not-yet-existing fs abstraction layer.
587 param->ddp_max_name_len = MAXNAMELEN;
588 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
589 param->ddp_symlink_max = PATH_MAX;
590 param->ddp_mount_type = LDD_MT_ZFS;
592 param->ddp_mntopts = MNTOPT_USERXATTR;
593 if (osd->od_posix_acl)
594 param->ddp_mntopts |= MNTOPT_ACL;
595 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
597 /* for maxbytes, report same value as ZPL */
598 param->ddp_maxbytes = MAX_LFS_FILESIZE;
600 /* inodes are dynamically allocated, so we report the per-inode space
601 * consumption to upper layers. This static value is not really accurate
602 * and we should use the same logic as in udmu_objset_statfs() to
603 * estimate the real size consumed by an object */
604 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
605 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
606 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
607 * not be accounted for every single new block insertion.
608 * Instead, the maximum extent size is set to the number of blocks that
609 * can fit into a single contiguous indirect block. There would be some
610 * cases where this crosses indirect blocks, but it also won't have 7
611 * new levels of indirect blocks in that case either, so it will still
612 * have enough reserved space for the extra indirect block */
613 param->ddp_max_extent_blks =
614 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
615 param->ddp_extent_tax = osd_blk_insert_cost(osd);
619 * Concurrency: shouldn't matter.
621 static int osd_sync(const struct lu_env *env, struct dt_device *d)
624 struct osd_device *osd = osd_dt_dev(d);
626 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
627 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
628 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
634 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
636 struct osd_device *osd = osd_dt_dev(dev);
637 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
640 mutex_enter(&tx->tx_sync_lock);
641 txg = tx->tx_open_txg + 1;
642 if (tx->tx_quiesce_txg_waiting < txg) {
643 tx->tx_quiesce_txg_waiting = txg;
644 cv_broadcast(&tx->tx_quiesce_more_cv);
646 mutex_exit(&tx->tx_sync_lock);
652 * Concurrency: shouldn't matter.
654 static int osd_ro(const struct lu_env *env, struct dt_device *d)
656 struct osd_device *osd = osd_dt_dev(d);
659 CERROR("%s: *** setting device %s read-only ***\n",
660 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
661 osd->od_dev_set_rdonly = 1;
662 spa_freeze(dmu_objset_spa(osd->od_os));
667 static struct dt_device_operations osd_dt_ops = {
668 .dt_root_get = osd_root_get,
669 .dt_statfs = osd_statfs,
670 .dt_trans_create = osd_trans_create,
671 .dt_trans_start = osd_trans_start,
672 .dt_trans_stop = osd_trans_stop,
673 .dt_trans_cb_add = osd_trans_cb_add,
674 .dt_conf_get = osd_conf_get,
676 .dt_commit_async = osd_commit_async,
681 * DMU OSD device type methods
683 static int osd_type_init(struct lu_device_type *t)
685 LU_CONTEXT_KEY_INIT(&osd_key);
686 return lu_context_key_register(&osd_key);
689 static void osd_type_fini(struct lu_device_type *t)
691 lu_context_key_degister(&osd_key);
694 static void *osd_key_init(const struct lu_context *ctx,
695 struct lu_context_key *key)
697 struct osd_thread_info *info;
701 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
703 info = ERR_PTR(-ENOMEM);
707 static void osd_key_fini(const struct lu_context *ctx,
708 struct lu_context_key *key, void *data)
710 struct osd_thread_info *info = data;
711 struct osd_idmap_cache *idc = info->oti_ins_cache;
714 LASSERT(info->oti_ins_cache_size > 0);
715 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
716 info->oti_ins_cache = NULL;
717 info->oti_ins_cache_size = 0;
722 static void osd_key_exit(const struct lu_context *ctx,
723 struct lu_context_key *key, void *data)
727 struct lu_context_key osd_key = {
728 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
729 .lct_init = osd_key_init,
730 .lct_fini = osd_key_fini,
731 .lct_exit = osd_key_exit
734 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
736 if (osd->od_cl_seq == NULL)
739 seq_client_fini(osd->od_cl_seq);
740 OBD_FREE_PTR(osd->od_cl_seq);
741 osd->od_cl_seq = NULL;
744 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
748 /* shutdown quota slave instance associated with the device */
749 if (o->od_quota_slave != NULL) {
750 /* complete all in-flight callbacks */
751 osd_sync(env, &o->od_dt_dev);
752 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
753 qsd_fini(env, o->od_quota_slave);
754 o->od_quota_slave = NULL;
757 osd_fid_fini(env, o);
762 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
764 struct osd_device *osd = arg;
766 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
769 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
771 struct osd_device *osd = arg;
773 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
774 LASSERT(newval >= SPA_MINBLOCKSIZE);
775 LASSERT(ISP2(newval));
777 osd->od_max_blksz = newval;
780 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
782 struct osd_device *osd = arg;
784 osd->od_prop_rdonly = !!newval;
788 * This function unregisters all registered callbacks. It's harmless to
789 * unregister callbacks that were never registered so it is used to safely
790 * unwind a partially completed call to osd_objset_register_callbacks().
792 static void osd_objset_unregister_callbacks(struct osd_device *o)
794 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
796 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
797 osd_xattr_changed_cb, o);
798 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
799 osd_recordsize_changed_cb, o);
800 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
801 osd_readonly_changed_cb, o);
803 if (o->arc_prune_cb != NULL) {
804 arc_remove_prune_callback(o->arc_prune_cb);
805 o->arc_prune_cb = NULL;
810 * Register the required callbacks to be notified when zfs properties
811 * are modified using the 'zfs(8)' command line utility.
813 static int osd_objset_register_callbacks(struct osd_device *o)
815 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
816 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
822 dsl_pool_config_enter(dp, FTAG);
823 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
824 osd_xattr_changed_cb, o);
828 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
829 osd_recordsize_changed_cb, o);
833 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
834 osd_readonly_changed_cb, o);
838 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
840 dsl_pool_config_exit(dp, FTAG);
842 osd_objset_unregister_callbacks(o);
847 static int osd_objset_open(struct osd_device *o)
849 uint64_t version = ZPL_VERSION;
850 uint64_t sa_obj, unlink_obj;
854 rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
855 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
858 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
864 /* Check ZFS version */
865 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
866 ZPL_VERSION_STR, 8, 1, &version);
868 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
870 * We can't return ENOENT because that would mean the objset
873 GOTO(out, rc = -EIO);
876 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
877 ZFS_SA_ATTRS, 8, 1, &sa_obj);
881 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
882 ZPL_END, &o->z_attr_table);
886 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
887 8, 1, &o->od_rootid);
889 CERROR("%s: lookup for root failed: rc = %d\n",
894 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
897 CERROR("%s: lookup for %s failed: rc = %d\n",
898 o->od_svname, ZFS_UNLINKED_SET, rc);
902 /* Check that user/group usage tracking is supported */
903 if (!dmu_objset_userused_enabled(o->od_os) ||
904 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
905 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
906 CERROR("%s: Space accounting not supported by this target, "
907 "aborting\n", o->od_svname);
908 GOTO(out, rc = -ENOTSUPP);
911 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
913 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
919 if (rc != 0 && o->od_os != NULL) {
920 dmu_objset_disown(o->od_os, o);
927 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
930 char *key = osd_oti_get(env)->oti_str;
934 if (osd->od_dt_dev.dd_rdonly) {
935 CERROR("%s: someone try to free objects under "
936 "readonly mode, should be disabled.\n", osd_name(osd));
942 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
944 CWARN("%s: Cannot truncate %llu: rc = %d\n",
945 osd->od_svname, oid, rc);
949 tx = dmu_tx_create(osd->od_os);
950 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
951 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
953 rc = -dmu_tx_assign(tx, TXG_WAIT);
955 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
956 osd->od_svname, oid, rc);
960 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
961 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
962 osd->od_unlinked, key, tx);
964 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
965 osd->od_svname, oid, rc);
969 rc = -dmu_object_free(osd->od_os, oid, tx);
971 CWARN("%s: Cannot free %llu: rc = %d\n",
972 osd->od_svname, oid, rc);
987 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
990 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
992 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
994 while (zap_cursor_retrieve(&zc, za) == 0) {
995 /* If cannot free the object, leave it in the unlinked set,
996 * until the OSD is mounted again when obd_unlinked_drain()
998 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1000 zap_cursor_advance(&zc);
1003 zap_cursor_fini(&zc);
1006 static int osd_mount(const struct lu_env *env,
1007 struct osd_device *o, struct lustre_cfg *cfg)
1009 char *mntdev = lustre_cfg_string(cfg, 1);
1010 char *str = lustre_cfg_string(cfg, 2);
1011 char *svname = lustre_cfg_string(cfg, 4);
1017 if (o->od_os != NULL)
1020 if (mntdev == NULL || svname == NULL)
1023 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1024 if (rc >= sizeof(o->od_mntdev))
1027 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1028 if (rc >= sizeof(o->od_svname))
1031 str = strstr(str, ":");
1033 unsigned long flags;
1035 rc = kstrtoul(str + 1, 10, &flags);
1039 if (flags & LMD_FLG_DEV_RDONLY) {
1040 o->od_dt_dev.dd_rdonly = 1;
1041 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1046 if (server_name_is_ost(o->od_svname))
1049 rc = osd_objset_open(o);
1053 o->od_xattr_in_sa = B_TRUE;
1054 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1056 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1059 o->od_root = rootdn->dn_object;
1060 osd_dnode_rele(rootdn);
1062 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1063 &o->od_userused_dn);
1067 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1068 &o->od_groupused_dn);
1072 /* 1. initialize oi before any file create or file open */
1073 rc = osd_oi_init(env, o);
1077 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1080 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1082 rc = lu_site_init_finish(&o->od_site);
1086 rc = osd_objset_register_callbacks(o);
1090 rc = osd_procfs_init(o, o->od_svname);
1094 /* initialize quota slave instance */
1095 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
1097 if (IS_ERR(o->od_quota_slave)) {
1098 rc = PTR_ERR(o->od_quota_slave);
1099 o->od_quota_slave = NULL;
1103 /* parse mount option "noacl", and enable ACL by default */
1104 opts = lustre_cfg_string(cfg, 3);
1105 if (opts == NULL || strstr(opts, "noacl") == NULL)
1106 o->od_posix_acl = 1;
1108 osd_unlinked_drain(env, o);
1110 if (rc && o->od_os) {
1111 dmu_objset_disown(o->od_os, o);
1118 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1122 if (atomic_read(&o->od_zerocopy_alloc))
1123 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1124 atomic_read(&o->od_zerocopy_alloc));
1125 if (atomic_read(&o->od_zerocopy_loan))
1126 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1127 atomic_read(&o->od_zerocopy_loan));
1128 if (atomic_read(&o->od_zerocopy_pin))
1129 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1130 atomic_read(&o->od_zerocopy_pin));
1132 if (o->od_unlinked) {
1133 osd_dnode_rele(o->od_unlinked);
1134 o->od_unlinked = NULL;
1136 if (o->od_userused_dn) {
1137 osd_dnode_rele(o->od_userused_dn);
1138 o->od_userused_dn = NULL;
1140 if (o->od_groupused_dn) {
1141 osd_dnode_rele(o->od_groupused_dn);
1142 o->od_groupused_dn = NULL;
1145 if (o->od_os != NULL) {
1146 if (!o->od_dt_dev.dd_rdonly)
1147 /* force a txg sync to get all commit callbacks */
1148 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1150 /* close the object set */
1151 dmu_objset_disown(o->od_os, o);
1159 static int osd_device_init0(const struct lu_env *env,
1160 struct osd_device *o,
1161 struct lustre_cfg *cfg)
1163 struct lu_device *l = osd2lu_dev(o);
1166 /* if the module was re-loaded, env can loose its keys */
1167 rc = lu_env_refill((struct lu_env *) env);
1171 l->ld_ops = &osd_lu_ops;
1172 o->od_dt_dev.dd_ops = &osd_dt_ops;
1178 static struct lu_device *osd_device_fini(const struct lu_env *env,
1179 struct lu_device *dev);
1181 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1182 struct lu_device_type *type,
1183 struct lustre_cfg *cfg)
1185 struct osd_device *dev;
1186 struct osd_seq_list *osl;
1191 return ERR_PTR(-ENOMEM);
1193 osl = &dev->od_seq_list;
1194 INIT_LIST_HEAD(&osl->osl_seq_list);
1195 rwlock_init(&osl->osl_seq_list_lock);
1196 sema_init(&osl->osl_seq_init_sem, 1);
1198 rc = dt_device_init(&dev->od_dt_dev, type);
1200 rc = osd_device_init0(env, dev, cfg);
1202 rc = osd_mount(env, dev, cfg);
1204 osd_device_fini(env, osd2lu_dev(dev));
1207 dt_device_fini(&dev->od_dt_dev);
1210 if (unlikely(rc != 0))
1213 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1216 static struct lu_device *osd_device_free(const struct lu_env *env,
1217 struct lu_device *d)
1219 struct osd_device *o = osd_dev(d);
1222 /* XXX: make osd top device in order to release reference */
1223 d->ld_site->ls_top_dev = d;
1224 lu_site_purge(env, d->ld_site, -1);
1225 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1226 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1227 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1229 lu_site_fini(&o->od_site);
1230 dt_device_fini(&o->od_dt_dev);
1236 static struct lu_device *osd_device_fini(const struct lu_env *env,
1237 struct lu_device *d)
1239 struct osd_device *o = osd_dev(d);
1245 osd_objset_unregister_callbacks(o);
1246 if (!o->od_dt_dev.dd_rdonly) {
1247 osd_sync(env, lu2dt_dev(d));
1249 spa_get_dsl(dmu_objset_spa(o->od_os)));
1253 /* now with all the callbacks completed we can cleanup the remainings */
1254 osd_shutdown(env, o);
1255 osd_oi_fini(env, o);
1257 rc = osd_procfs_fini(o);
1259 CERROR("proc fini error %d\n", rc);
1260 RETURN(ERR_PTR(rc));
1269 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1270 const char *name, struct lu_device *next)
1276 * To be removed, setup is performed by osd_device_{init,alloc} and
1277 * cleanup is performed by osd_device_{fini,free).
1279 static int osd_process_config(const struct lu_env *env,
1280 struct lu_device *d, struct lustre_cfg *cfg)
1282 struct osd_device *o = osd_dev(d);
1286 switch(cfg->lcfg_command) {
1288 rc = osd_mount(env, o, cfg);
1291 rc = osd_shutdown(env, o);
1294 LASSERT(&o->od_dt_dev);
1295 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1296 cfg, &o->od_dt_dev);
1297 if (rc > 0 || rc == -ENOSYS) {
1298 rc = class_process_proc_param(PARAM_OST,
1299 lprocfs_osd_obd_vars,
1300 cfg, &o->od_dt_dev);
1313 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1315 struct osd_device *osd = osd_dev(d);
1319 if (osd->od_quota_slave == NULL)
1322 /* start qsd instance on recovery completion, this notifies the quota
1323 * slave code that we are about to process new requests now */
1324 rc = qsd_start(env, osd->od_quota_slave);
1329 * we use exports to track all osd users
1331 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1332 struct obd_device *obd, struct obd_uuid *cluuid,
1333 struct obd_connect_data *data, void *localdata)
1335 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1336 struct lustre_handle conn;
1340 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1342 rc = class_connect(&conn, obd, cluuid);
1346 *exp = class_conn2export(&conn);
1348 spin_lock(&obd->obd_dev_lock);
1350 spin_unlock(&obd->obd_dev_lock);
1356 * once last export (we don't count self-export) disappeared
1357 * osd can be released
1359 static int osd_obd_disconnect(struct obd_export *exp)
1361 struct obd_device *obd = exp->exp_obd;
1362 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1363 int rc, release = 0;
1366 /* Only disconnect the underlying layers on the final disconnect. */
1367 spin_lock(&obd->obd_dev_lock);
1369 if (osd->od_connects == 0)
1371 spin_unlock(&obd->obd_dev_lock);
1373 rc = class_disconnect(exp); /* bz 9811 */
1375 if (rc == 0 && release)
1376 class_manual_cleanup(obd);
1380 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1382 struct seq_server_site *ss = osd_seq_site(osd);
1386 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1389 if (unlikely(ss == NULL))
1392 OBD_ALLOC_PTR(osd->od_cl_seq);
1393 if (osd->od_cl_seq == NULL)
1396 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1397 osd->od_svname, ss->ss_server_seq);
1400 OBD_FREE_PTR(osd->od_cl_seq);
1401 osd->od_cl_seq = NULL;
1407 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1408 struct lu_device *dev)
1410 struct osd_device *osd = osd_dev(dev);
1414 if (osd->od_quota_slave != NULL) {
1415 /* set up quota slave objects */
1416 rc = qsd_prepare(env, osd->od_quota_slave);
1421 rc = osd_fid_init(env, osd);
1426 struct lu_device_operations osd_lu_ops = {
1427 .ldo_object_alloc = osd_object_alloc,
1428 .ldo_process_config = osd_process_config,
1429 .ldo_recovery_complete = osd_recovery_complete,
1430 .ldo_prepare = osd_prepare,
1433 static void osd_type_start(struct lu_device_type *t)
1437 static void osd_type_stop(struct lu_device_type *t)
1441 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1442 struct lu_fid *fid, struct md_op_data *op_data)
1444 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1446 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1449 static struct lu_device_type_operations osd_device_type_ops = {
1450 .ldto_init = osd_type_init,
1451 .ldto_fini = osd_type_fini,
1453 .ldto_start = osd_type_start,
1454 .ldto_stop = osd_type_stop,
1456 .ldto_device_alloc = osd_device_alloc,
1457 .ldto_device_free = osd_device_free,
1459 .ldto_device_init = osd_device_init,
1460 .ldto_device_fini = osd_device_fini
1463 static struct lu_device_type osd_device_type = {
1464 .ldt_tags = LU_DEVICE_DT,
1465 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1466 .ldt_ops = &osd_device_type_ops,
1467 .ldt_ctx_tags = LCT_LOCAL
1471 static struct obd_ops osd_obd_device_ops = {
1472 .o_owner = THIS_MODULE,
1473 .o_connect = osd_obd_connect,
1474 .o_disconnect = osd_obd_disconnect,
1475 .o_fid_alloc = osd_fid_alloc
1478 static int __init osd_init(void)
1482 rc = osd_options_init();
1486 rc = lu_kmem_init(osd_caches);
1490 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1491 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1493 lu_kmem_fini(osd_caches);
1497 static void __exit osd_exit(void)
1499 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1500 lu_kmem_fini(osd_caches);
1503 extern unsigned int osd_oi_count;
1504 module_param(osd_oi_count, int, 0444);
1505 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1507 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1508 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1509 MODULE_VERSION(LUSTRE_VERSION_STRING);
1510 MODULE_LICENSE("GPL");
1512 module_init(osd_init);
1513 module_exit(osd_exit);