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 lu_device_get(&d->dd_lu_dev);
232 static void osd_unlinked_list_emptify(const struct lu_env *env,
233 struct osd_device *osd,
234 struct list_head *list, bool free)
236 struct osd_object *obj;
239 while (!list_empty(list)) {
240 obj = list_entry(list->next,
241 struct osd_object, oo_unlinked_linkage);
242 LASSERT(obj->oo_dn != NULL);
243 oid = obj->oo_dn->dn_object;
245 list_del_init(&obj->oo_unlinked_linkage);
247 (void)osd_unlinked_object_free(env, osd, oid);
251 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
253 struct dt_txn_commit_cb *dcb;
254 struct dt_txn_commit_cb *tmp;
256 /* call per-transaction stop callbacks if any */
257 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
259 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
260 "commit callback entry: magic=%x name='%s'\n",
261 dcb->dcb_magic, dcb->dcb_name);
262 list_del_init(&dcb->dcb_linkage);
263 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
268 * Concurrency: shouldn't matter.
270 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
273 struct osd_device *osd = osd_dt_dev(th->th_dev);
274 bool sync = (th->th_sync != 0);
275 struct osd_thandle *oh;
276 struct list_head unlinked;
281 oh = container_of0(th, struct osd_thandle, ot_super);
282 INIT_LIST_HEAD(&unlinked);
283 list_splice_init(&oh->ot_unlinked_list, &unlinked);
284 /* reset OI cache for safety */
285 osd_oti_get(env)->oti_ins_cache_used = 0;
287 if (oh->ot_assigned == 0) {
289 dmu_tx_abort(oh->ot_tx);
290 osd_object_sa_dirty_rele(env, oh);
291 osd_unlinked_list_emptify(env, osd, &unlinked, false);
292 /* there won't be any commit, release reserved quota space now,
294 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
299 rc = dt_txn_hook_stop(env, th);
301 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
304 osd_trans_stop_cb(oh, rc);
307 txg = oh->ot_tx->tx_txg;
309 osd_object_sa_dirty_rele(env, oh);
310 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
311 * by osd_trans_commit_cb already. */
312 dmu_tx_commit(oh->ot_tx);
314 osd_unlinked_list_emptify(env, osd, &unlinked, true);
317 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
322 static struct thandle *osd_trans_create(const struct lu_env *env,
323 struct dt_device *dt)
325 struct osd_device *osd = osd_dt_dev(dt);
326 struct osd_thandle *oh;
332 CERROR("%s: someone try to start transaction under "
333 "readonly mode, should be disabled.\n",
334 osd_name(osd_dt_dev(dt)));
336 RETURN(ERR_PTR(-EROFS));
339 tx = dmu_tx_create(osd->od_os);
341 RETURN(ERR_PTR(-ENOMEM));
343 /* alloc callback data */
347 RETURN(ERR_PTR(-ENOMEM));
351 INIT_LIST_HEAD(&oh->ot_dcb_list);
352 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
353 INIT_LIST_HEAD(&oh->ot_unlinked_list);
354 INIT_LIST_HEAD(&oh->ot_sa_list);
355 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
362 /* Estimate the total number of objects from a number of blocks */
363 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
364 uint64_t nrblocks, uint64_t est_maxblockshift)
366 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
369 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
370 * bump the free dnode estimate to assume blocks at least 64KB in
371 * case of a directory-heavy MDT (at 32KB/directory).
373 if (est_maxblockshift < 16) {
374 nrblocks >>= (16 - est_maxblockshift);
375 est_maxblockshift = 16;
379 * Estimate the total number of dnodes from the total blocks count
380 * and the space used per dnode. Since we don't know the overhead
381 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
382 * just using DNODE_SHIFT isn't going to give a good estimate.
383 * Instead, compute the current average space usage per dnode, with
384 * an upper and lower cap to avoid unrealistic estimates..
386 * In case there aren't many dnodes or blocks used yet, add a small
387 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
388 * gradually disappears as the number of real dnodes grows. It also
389 * avoids the need to check for divide-by-zero computing dn_per_block.
391 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
392 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
394 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
395 usedbytes) >> est_maxblockshift;
396 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
398 if (est_usedobjs <= est_usedblocks) {
400 * Average space/dnode more than maximum block size, use max
401 * block size to estimate free dnodes from adjusted free blocks
402 * count. OSTs typically use multiple blocks per dnode so this
405 est_totobjs = nrblocks;
407 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
409 * Average space/dnode smaller than min dnode size (probably
410 * due to metadnode compression), use min dnode size to
411 * estimate object count. MDTs may use only one block per node
412 * so this case applies.
414 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
418 * Between the extremes, use average space per existing dnode
419 * to compute the number of dnodes that will fit into nrblocks:
421 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
423 * this may overflow 64 bits or become 0 if not handled well.
425 * We know nrblocks is below 2^(64 - blkbits) bits, and
426 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
427 * which means that multiplying them may get as large as
428 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
430 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
431 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
432 * the limit checks above, so we can safely compute this first.
433 * We care more about accuracy on the MDT (many dnodes/block)
434 * which is good because this is where truncation errors are
435 * smallest. Since both nrblocks and dn_per_block are a
436 * function of blkbits, their product is at most:
438 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
440 * so we can safely use 7 bits to compute a fixed-point
441 * fraction and est_totobjs can still fit in 64 bits.
443 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
445 est_totobjs = (nrblocks * dn_per_block) >> 7;
450 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
452 struct objset *os = osd->od_os;
453 uint64_t usedbytes, availbytes, usedobjs, availobjs;
454 uint64_t est_availobjs;
458 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
460 memset(osfs, 0, sizeof(*osfs));
462 /* We're a zfs filesystem. */
463 osfs->os_type = UBERBLOCK_MAGIC;
466 * ZFS allows multiple block sizes. For statfs, Linux makes no
467 * proper distinction between bsize and frsize. For calculations
468 * of free and used blocks incorrectly uses bsize instead of frsize,
469 * but bsize is also used as the optimal blocksize. We return the
470 * largest possible block size as IO size for the optimum performance
471 * and scale the free and used blocks count appropriately.
473 osfs->os_bsize = osd->od_max_blksz;
474 bshift = fls64(osfs->os_bsize) - 1;
476 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
477 osfs->os_bfree = availbytes >> bshift;
478 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
480 /* Take replication (i.e. number of copies) into account */
481 if (os->os_copies != 0)
482 osfs->os_bavail /= os->os_copies;
485 * Reserve some space so we don't run into ENOSPC due to grants not
486 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
487 * Rather than report this via os_bavail (which makes users unhappy if
488 * they can't fill the filesystem 100%), reduce os_blocks as well.
490 * Reserve 0.78% of total space, at least 16MB for small filesystems,
491 * for internal files to be created/unlinked when space is tight.
493 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
494 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
495 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
496 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
498 osfs->os_blocks -= reserved;
499 osfs->os_bfree -= min(reserved, osfs->os_bfree);
500 osfs->os_bavail -= min(reserved, osfs->os_bavail);
503 * The availobjs value returned from dmu_objset_space() is largely
504 * useless, since it reports the number of objects that might
505 * theoretically still fit into the dataset, independent of minor
506 * issues like how much space is actually available in the pool.
507 * Compute a better estimate in udmu_objs_count_estimate().
509 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
510 osfs->os_bfree, bshift);
512 osfs->os_ffree = min(availobjs, est_availobjs);
513 osfs->os_files = osfs->os_ffree + usedobjs;
515 /* ZFS XXX: fill in backing dataset FSID/UUID
516 memcpy(osfs->os_fsid, .... );*/
518 osfs->os_namelen = MAXNAMELEN;
519 osfs->os_maxbytes = OBD_OBJECT_EOF;
521 if (!spa_writeable(dmu_objset_spa(os)) ||
522 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
523 osfs->os_state |= OS_STATE_READONLY;
529 * Concurrency: shouldn't matter.
531 int osd_statfs(const struct lu_env *env, struct dt_device *d,
532 struct obd_statfs *osfs)
537 rc = osd_objset_statfs(osd_dt_dev(d), osfs);
538 if (unlikely(rc != 0))
541 osfs->os_bavail -= min_t(u64,
542 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
547 static int osd_blk_insert_cost(struct osd_device *osd)
549 int max_blockshift, nr_blkptrshift, bshift;
551 /* max_blockshift is the log2 of the number of blocks needed to reach
552 * the maximum filesize (that's to say 2^64) */
553 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
554 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
556 /* nr_blkptrshift is the log2 of the number of block pointers that can
557 * be stored in an indirect block */
558 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
559 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
561 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
562 * tree. We add +1 for rounding purpose.
563 * The tree depth times the indirect block size gives us the maximum
564 * cost of inserting a block in the tree */
565 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
569 * Concurrency: doesn't access mutable data.
571 static void osd_conf_get(const struct lu_env *env,
572 const struct dt_device *dev,
573 struct dt_device_param *param)
575 struct osd_device *osd = osd_dt_dev(dev);
578 * XXX should be taken from not-yet-existing fs abstraction layer.
580 param->ddp_max_name_len = MAXNAMELEN;
581 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
582 param->ddp_symlink_max = PATH_MAX;
583 param->ddp_mount_type = LDD_MT_ZFS;
585 param->ddp_mntopts = MNTOPT_USERXATTR;
586 if (osd->od_posix_acl)
587 param->ddp_mntopts |= MNTOPT_ACL;
588 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
590 /* for maxbytes, report same value as ZPL */
591 param->ddp_maxbytes = MAX_LFS_FILESIZE;
593 /* inodes are dynamically allocated, so we report the per-inode space
594 * consumption to upper layers. This static value is not really accurate
595 * and we should use the same logic as in udmu_objset_statfs() to
596 * estimate the real size consumed by an object */
597 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
598 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
599 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
600 * not be accounted for every single new block insertion.
601 * Instead, the maximum extent size is set to the number of blocks that
602 * can fit into a single contiguous indirect block. There would be some
603 * cases where this crosses indirect blocks, but it also won't have 7
604 * new levels of indirect blocks in that case either, so it will still
605 * have enough reserved space for the extra indirect block */
606 param->ddp_max_extent_blks =
607 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
608 param->ddp_extent_tax = osd_blk_insert_cost(osd);
612 * Concurrency: shouldn't matter.
614 static int osd_sync(const struct lu_env *env, struct dt_device *d)
617 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);
627 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
629 struct osd_device *osd = osd_dt_dev(dev);
630 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
633 mutex_enter(&tx->tx_sync_lock);
634 txg = tx->tx_open_txg + 1;
635 if (tx->tx_quiesce_txg_waiting < txg) {
636 tx->tx_quiesce_txg_waiting = txg;
637 cv_broadcast(&tx->tx_quiesce_more_cv);
639 mutex_exit(&tx->tx_sync_lock);
645 * Concurrency: shouldn't matter.
647 static int osd_ro(const struct lu_env *env, struct dt_device *d)
649 struct osd_device *osd = osd_dt_dev(d);
652 CERROR("%s: *** setting device %s read-only ***\n",
653 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
654 osd->od_dev_set_rdonly = 1;
655 spa_freeze(dmu_objset_spa(osd->od_os));
660 static struct dt_device_operations osd_dt_ops = {
661 .dt_root_get = osd_root_get,
662 .dt_statfs = osd_statfs,
663 .dt_trans_create = osd_trans_create,
664 .dt_trans_start = osd_trans_start,
665 .dt_trans_stop = osd_trans_stop,
666 .dt_trans_cb_add = osd_trans_cb_add,
667 .dt_conf_get = osd_conf_get,
669 .dt_commit_async = osd_commit_async,
674 * DMU OSD device type methods
676 static int osd_type_init(struct lu_device_type *t)
678 LU_CONTEXT_KEY_INIT(&osd_key);
679 return lu_context_key_register(&osd_key);
682 static void osd_type_fini(struct lu_device_type *t)
684 lu_context_key_degister(&osd_key);
687 static void *osd_key_init(const struct lu_context *ctx,
688 struct lu_context_key *key)
690 struct osd_thread_info *info;
694 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
696 info = ERR_PTR(-ENOMEM);
700 static void osd_key_fini(const struct lu_context *ctx,
701 struct lu_context_key *key, void *data)
703 struct osd_thread_info *info = data;
704 struct osd_idmap_cache *idc = info->oti_ins_cache;
707 LASSERT(info->oti_ins_cache_size > 0);
708 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
709 info->oti_ins_cache = NULL;
710 info->oti_ins_cache_size = 0;
712 lu_buf_free(&info->oti_xattr_lbuf);
716 static void osd_key_exit(const struct lu_context *ctx,
717 struct lu_context_key *key, void *data)
721 struct lu_context_key osd_key = {
722 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
723 .lct_init = osd_key_init,
724 .lct_fini = osd_key_fini,
725 .lct_exit = osd_key_exit
728 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
730 if (osd->od_cl_seq == NULL)
733 seq_client_fini(osd->od_cl_seq);
734 OBD_FREE_PTR(osd->od_cl_seq);
735 osd->od_cl_seq = NULL;
738 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
742 /* shutdown quota slave instance associated with the device */
743 if (o->od_quota_slave != NULL) {
744 /* complete all in-flight callbacks */
745 osd_sync(env, &o->od_dt_dev);
746 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
747 qsd_fini(env, o->od_quota_slave);
748 o->od_quota_slave = NULL;
751 osd_fid_fini(env, o);
756 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
758 struct osd_device *osd = arg;
760 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
763 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
765 struct osd_device *osd = arg;
767 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
768 LASSERT(newval >= SPA_MINBLOCKSIZE);
769 LASSERT(ISP2(newval));
771 osd->od_max_blksz = newval;
774 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
776 struct osd_device *osd = arg;
778 osd->od_prop_rdonly = !!newval;
781 static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
783 struct osd_device *osd = arg;
785 osd->od_dnsize = newval;
789 * This function unregisters all registered callbacks. It's harmless to
790 * unregister callbacks that were never registered so it is used to safely
791 * unwind a partially completed call to osd_objset_register_callbacks().
793 static void osd_objset_unregister_callbacks(struct osd_device *o)
795 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
797 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
798 osd_xattr_changed_cb, o);
799 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
800 osd_recordsize_changed_cb, o);
801 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
802 osd_readonly_changed_cb, o);
803 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
804 osd_readonly_changed_cb, o);
806 if (o->arc_prune_cb != NULL) {
807 arc_remove_prune_callback(o->arc_prune_cb);
808 o->arc_prune_cb = NULL;
813 * Register the required callbacks to be notified when zfs properties
814 * are modified using the 'zfs(8)' command line utility.
816 static int osd_objset_register_callbacks(struct osd_device *o)
818 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
819 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
825 dsl_pool_config_enter(dp, FTAG);
826 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
827 osd_xattr_changed_cb, o);
831 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
832 osd_recordsize_changed_cb, o);
836 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
837 osd_readonly_changed_cb, o);
841 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
842 osd_dnodesize_changed_cb, o);
846 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
848 dsl_pool_config_exit(dp, FTAG);
850 osd_objset_unregister_callbacks(o);
855 static int osd_objset_open(struct osd_device *o)
857 uint64_t version = ZPL_VERSION;
858 uint64_t sa_obj, unlink_obj;
862 rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
863 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
864 B_FALSE, o, &o->od_os);
867 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
873 /* Check ZFS version */
874 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
875 ZPL_VERSION_STR, 8, 1, &version);
877 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
879 * We can't return ENOENT because that would mean the objset
882 GOTO(out, rc = -EIO);
885 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
886 ZFS_SA_ATTRS, 8, 1, &sa_obj);
890 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
891 ZPL_END, &o->z_attr_table);
895 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
896 8, 1, &o->od_rootid);
898 CERROR("%s: lookup for root failed: rc = %d\n",
903 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
906 CERROR("%s: lookup for %s failed: rc = %d\n",
907 o->od_svname, ZFS_UNLINKED_SET, rc);
911 /* Check that user/group usage tracking is supported */
912 if (!dmu_objset_userused_enabled(o->od_os) ||
913 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
914 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
915 CERROR("%s: Space accounting not supported by this target, "
916 "aborting\n", o->od_svname);
917 GOTO(out, rc = -ENOTSUPP);
920 rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
922 CERROR("%s: can't get dnode for unlinked: rc = %d\n",
928 if (rc != 0 && o->od_os != NULL) {
929 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
936 int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
939 char *key = osd_oti_get(env)->oti_str;
943 if (osd->od_dt_dev.dd_rdonly) {
944 CERROR("%s: someone try to free objects under "
945 "readonly mode, should be disabled.\n", osd_name(osd));
951 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
953 CWARN("%s: Cannot truncate %llu: rc = %d\n",
954 osd->od_svname, oid, rc);
958 tx = dmu_tx_create(osd->od_os);
959 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
960 osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
962 rc = -dmu_tx_assign(tx, TXG_WAIT);
964 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
965 osd->od_svname, oid, rc);
969 snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
970 rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
971 osd->od_unlinked, key, tx);
973 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
974 osd->od_svname, oid, rc);
978 rc = -dmu_object_free(osd->od_os, oid, tx);
980 CWARN("%s: Cannot free %llu: rc = %d\n",
981 osd->od_svname, oid, rc);
996 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
999 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
1001 zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
1003 while (zap_cursor_retrieve(&zc, za) == 0) {
1004 /* If cannot free the object, leave it in the unlinked set,
1005 * until the OSD is mounted again when obd_unlinked_drain()
1006 * will be called. */
1007 if (osd_unlinked_object_free(env, osd, za->za_first_integer))
1009 zap_cursor_advance(&zc);
1012 zap_cursor_fini(&zc);
1015 static int osd_mount(const struct lu_env *env,
1016 struct osd_device *o, struct lustre_cfg *cfg)
1018 char *mntdev = lustre_cfg_string(cfg, 1);
1019 char *str = lustre_cfg_string(cfg, 2);
1020 char *svname = lustre_cfg_string(cfg, 4);
1026 if (o->od_os != NULL)
1029 if (mntdev == NULL || svname == NULL)
1032 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1033 if (rc >= sizeof(o->od_mntdev))
1036 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1037 if (rc >= sizeof(o->od_svname))
1040 str = strstr(str, ":");
1042 unsigned long flags;
1044 rc = kstrtoul(str + 1, 10, &flags);
1048 if (flags & LMD_FLG_DEV_RDONLY) {
1049 o->od_dt_dev.dd_rdonly = 1;
1050 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1055 if (server_name_is_ost(o->od_svname))
1058 rc = osd_objset_open(o);
1062 o->od_xattr_in_sa = B_TRUE;
1063 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1065 rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
1068 o->od_root = rootdn->dn_object;
1069 osd_dnode_rele(rootdn);
1071 rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
1072 &o->od_userused_dn);
1076 rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
1077 &o->od_groupused_dn);
1081 #ifdef ZFS_PROJINHERIT
1082 if (dmu_objset_projectquota_enabled(o->od_os)) {
1083 rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
1084 &o->od_projectused_dn);
1085 if (rc && rc != -ENOENT)
1090 /* 1. initialize oi before any file create or file open */
1091 rc = osd_oi_init(env, o);
1095 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1098 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1100 rc = lu_site_init_finish(&o->od_site);
1104 rc = osd_objset_register_callbacks(o);
1108 rc = osd_procfs_init(o, o->od_svname);
1112 /* initialize quota slave instance */
1113 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
1115 if (IS_ERR(o->od_quota_slave)) {
1116 rc = PTR_ERR(o->od_quota_slave);
1117 o->od_quota_slave = NULL;
1121 #ifdef HAVE_DMU_USEROBJ_ACCOUNTING
1122 if (!osd_dmu_userobj_accounting_available(o))
1123 CWARN("%s: dnode accounting not enabled: "
1124 "enable feature@userobj_accounting in pool\n",
1128 /* parse mount option "noacl", and enable ACL by default */
1129 opts = lustre_cfg_string(cfg, 3);
1130 if (opts == NULL || strstr(opts, "noacl") == NULL)
1131 o->od_posix_acl = 1;
1133 osd_unlinked_drain(env, o);
1135 if (rc && o->od_os) {
1136 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
1143 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1147 if (atomic_read(&o->od_zerocopy_alloc))
1148 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1149 atomic_read(&o->od_zerocopy_alloc));
1150 if (atomic_read(&o->od_zerocopy_loan))
1151 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1152 atomic_read(&o->od_zerocopy_loan));
1153 if (atomic_read(&o->od_zerocopy_pin))
1154 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1155 atomic_read(&o->od_zerocopy_pin));
1157 if (o->od_unlinked) {
1158 osd_dnode_rele(o->od_unlinked);
1159 o->od_unlinked = NULL;
1161 if (o->od_userused_dn) {
1162 osd_dnode_rele(o->od_userused_dn);
1163 o->od_userused_dn = NULL;
1165 if (o->od_groupused_dn) {
1166 osd_dnode_rele(o->od_groupused_dn);
1167 o->od_groupused_dn = NULL;
1170 #ifdef ZFS_PROJINHERIT
1171 if (o->od_projectused_dn) {
1172 osd_dnode_rele(o->od_projectused_dn);
1173 o->od_projectused_dn = NULL;
1177 if (o->od_os != NULL) {
1178 if (!o->od_dt_dev.dd_rdonly)
1179 /* force a txg sync to get all commit callbacks */
1180 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1182 /* close the object set */
1183 osd_dmu_objset_disown(o->od_os, B_FALSE, o);
1190 static int osd_device_init0(const struct lu_env *env,
1191 struct osd_device *o,
1192 struct lustre_cfg *cfg)
1194 struct lu_device *l = osd2lu_dev(o);
1197 /* if the module was re-loaded, env can loose its keys */
1198 rc = lu_env_refill((struct lu_env *) env);
1202 l->ld_ops = &osd_lu_ops;
1203 o->od_dt_dev.dd_ops = &osd_dt_ops;
1209 static struct lu_device *osd_device_fini(const struct lu_env *env,
1210 struct lu_device *dev);
1212 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1213 struct lu_device_type *type,
1214 struct lustre_cfg *cfg)
1216 struct osd_device *dev;
1217 struct osd_seq_list *osl;
1222 return ERR_PTR(-ENOMEM);
1224 osl = &dev->od_seq_list;
1225 INIT_LIST_HEAD(&osl->osl_seq_list);
1226 rwlock_init(&osl->osl_seq_list_lock);
1227 sema_init(&osl->osl_seq_init_sem, 1);
1229 rc = dt_device_init(&dev->od_dt_dev, type);
1231 rc = osd_device_init0(env, dev, cfg);
1233 rc = osd_mount(env, dev, cfg);
1235 osd_device_fini(env, osd2lu_dev(dev));
1238 dt_device_fini(&dev->od_dt_dev);
1241 if (unlikely(rc != 0))
1244 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1247 static struct lu_device *osd_device_free(const struct lu_env *env,
1248 struct lu_device *d)
1250 struct osd_device *o = osd_dev(d);
1253 /* XXX: make osd top device in order to release reference */
1254 d->ld_site->ls_top_dev = d;
1255 lu_site_purge(env, d->ld_site, -1);
1256 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1257 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1258 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1260 lu_site_fini(&o->od_site);
1261 dt_device_fini(&o->od_dt_dev);
1267 static struct lu_device *osd_device_fini(const struct lu_env *env,
1268 struct lu_device *d)
1270 struct osd_device *o = osd_dev(d);
1276 osd_objset_unregister_callbacks(o);
1277 if (!o->od_dt_dev.dd_rdonly) {
1278 osd_sync(env, lu2dt_dev(d));
1280 spa_get_dsl(dmu_objset_spa(o->od_os)));
1284 /* now with all the callbacks completed we can cleanup the remainings */
1285 osd_shutdown(env, o);
1286 osd_oi_fini(env, o);
1288 rc = osd_procfs_fini(o);
1290 CERROR("proc fini error %d\n", rc);
1291 RETURN(ERR_PTR(rc));
1300 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1301 const char *name, struct lu_device *next)
1307 * To be removed, setup is performed by osd_device_{init,alloc} and
1308 * cleanup is performed by osd_device_{fini,free).
1310 static int osd_process_config(const struct lu_env *env,
1311 struct lu_device *d, struct lustre_cfg *cfg)
1313 struct osd_device *o = osd_dev(d);
1317 switch(cfg->lcfg_command) {
1319 rc = osd_mount(env, o, cfg);
1322 rc = osd_shutdown(env, o);
1325 LASSERT(&o->od_dt_dev);
1326 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1327 cfg, &o->od_dt_dev);
1328 if (rc > 0 || rc == -ENOSYS) {
1329 rc = class_process_proc_param(PARAM_OST,
1330 lprocfs_osd_obd_vars,
1331 cfg, &o->od_dt_dev);
1344 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1346 struct osd_device *osd = osd_dev(d);
1350 if (osd->od_quota_slave == NULL)
1353 /* start qsd instance on recovery completion, this notifies the quota
1354 * slave code that we are about to process new requests now */
1355 rc = qsd_start(env, osd->od_quota_slave);
1360 * we use exports to track all osd users
1362 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1363 struct obd_device *obd, struct obd_uuid *cluuid,
1364 struct obd_connect_data *data, void *localdata)
1366 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1367 struct lustre_handle conn;
1371 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1373 rc = class_connect(&conn, obd, cluuid);
1377 *exp = class_conn2export(&conn);
1379 spin_lock(&obd->obd_dev_lock);
1381 spin_unlock(&obd->obd_dev_lock);
1387 * once last export (we don't count self-export) disappeared
1388 * osd can be released
1390 static int osd_obd_disconnect(struct obd_export *exp)
1392 struct obd_device *obd = exp->exp_obd;
1393 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1394 int rc, release = 0;
1397 /* Only disconnect the underlying layers on the final disconnect. */
1398 spin_lock(&obd->obd_dev_lock);
1400 if (osd->od_connects == 0)
1402 spin_unlock(&obd->obd_dev_lock);
1404 rc = class_disconnect(exp); /* bz 9811 */
1406 if (rc == 0 && release)
1407 class_manual_cleanup(obd);
1411 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1413 struct seq_server_site *ss = osd_seq_site(osd);
1417 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1420 if (unlikely(ss == NULL))
1423 OBD_ALLOC_PTR(osd->od_cl_seq);
1424 if (osd->od_cl_seq == NULL)
1427 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1428 osd->od_svname, ss->ss_server_seq);
1431 OBD_FREE_PTR(osd->od_cl_seq);
1432 osd->od_cl_seq = NULL;
1438 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1439 struct lu_device *dev)
1441 struct osd_device *osd = osd_dev(dev);
1445 if (osd->od_quota_slave != NULL) {
1446 /* set up quota slave objects */
1447 rc = qsd_prepare(env, osd->od_quota_slave);
1452 rc = osd_fid_init(env, osd);
1457 struct lu_device_operations osd_lu_ops = {
1458 .ldo_object_alloc = osd_object_alloc,
1459 .ldo_process_config = osd_process_config,
1460 .ldo_recovery_complete = osd_recovery_complete,
1461 .ldo_prepare = osd_prepare,
1464 static void osd_type_start(struct lu_device_type *t)
1468 static void osd_type_stop(struct lu_device_type *t)
1472 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1473 struct lu_fid *fid, struct md_op_data *op_data)
1475 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1477 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1480 static struct lu_device_type_operations osd_device_type_ops = {
1481 .ldto_init = osd_type_init,
1482 .ldto_fini = osd_type_fini,
1484 .ldto_start = osd_type_start,
1485 .ldto_stop = osd_type_stop,
1487 .ldto_device_alloc = osd_device_alloc,
1488 .ldto_device_free = osd_device_free,
1490 .ldto_device_init = osd_device_init,
1491 .ldto_device_fini = osd_device_fini
1494 static struct lu_device_type osd_device_type = {
1495 .ldt_tags = LU_DEVICE_DT,
1496 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1497 .ldt_ops = &osd_device_type_ops,
1498 .ldt_ctx_tags = LCT_LOCAL
1502 static struct obd_ops osd_obd_device_ops = {
1503 .o_owner = THIS_MODULE,
1504 .o_connect = osd_obd_connect,
1505 .o_disconnect = osd_obd_disconnect,
1506 .o_fid_alloc = osd_fid_alloc
1509 static int __init osd_init(void)
1513 rc = osd_options_init();
1517 rc = lu_kmem_init(osd_caches);
1521 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1522 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1524 lu_kmem_fini(osd_caches);
1528 static void __exit osd_exit(void)
1530 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1531 lu_kmem_fini(osd_caches);
1534 extern unsigned int osd_oi_count;
1535 module_param(osd_oi_count, int, 0444);
1536 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1538 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1539 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1540 MODULE_VERSION(LUSTRE_VERSION_STRING);
1541 MODULE_LICENSE("GPL");
1543 module_init(osd_init);
1544 module_exit(osd_exit);