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, 2015, 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_db != NULL);
242 oid = obj->oo_db->db_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);
284 if (oh->ot_assigned == 0) {
286 dmu_tx_abort(oh->ot_tx);
287 osd_object_sa_dirty_rele(oh);
288 osd_unlinked_list_emptify(osd, &unlinked, false);
289 /* there won't be any commit, release reserved quota space now,
291 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
296 /* When doing our own inode accounting, the ZAPs storing per-uid/gid
297 * usage are updated at operation execution time, so we should call
298 * qsd_op_end() straight away. Otherwise (for blk accounting maintained
299 * by ZFS and when #inode is estimated from #blks) accounting is updated
300 * at commit time and the call to qsd_op_end() must be delayed */
301 if (oh->ot_quota_trans.lqt_id_cnt > 0 &&
302 !oh->ot_quota_trans.lqt_ids[0].lqi_is_blk &&
303 !osd->od_quota_iused_est)
304 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
306 rc = dt_txn_hook_stop(env, th);
308 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
311 osd_trans_stop_cb(oh, rc);
314 txg = oh->ot_tx->tx_txg;
316 osd_object_sa_dirty_rele(oh);
317 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
318 * by osd_trans_commit_cb already. */
319 dmu_tx_commit(oh->ot_tx);
321 osd_unlinked_list_emptify(osd, &unlinked, true);
324 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
329 static struct thandle *osd_trans_create(const struct lu_env *env,
330 struct dt_device *dt)
332 struct osd_device *osd = osd_dt_dev(dt);
333 struct osd_thandle *oh;
338 tx = dmu_tx_create(osd->od_os);
340 RETURN(ERR_PTR(-ENOMEM));
342 /* alloc callback data */
346 RETURN(ERR_PTR(-ENOMEM));
350 INIT_LIST_HEAD(&oh->ot_dcb_list);
351 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
352 INIT_LIST_HEAD(&oh->ot_unlinked_list);
353 INIT_LIST_HEAD(&oh->ot_sa_list);
354 sema_init(&oh->ot_sa_lock, 1);
355 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
359 th->th_tags = LCT_TX_HANDLE;
363 /* Estimate the number of objects from a number of blocks */
364 uint64_t osd_objs_count_estimate(uint64_t refdbytes, uint64_t usedobjs,
365 uint64_t nrblocks, uint64_t est_maxblockshift)
367 uint64_t est_objs, est_refdblocks, est_usedobjs;
369 /* Compute an nrblocks estimate based on the actual number of
370 * dnodes that could fit in the space. Since we don't know the
371 * overhead associated with each dnode (xattrs, SAs, VDEV overhead,
372 * etc) just using DNODE_SHIFT isn't going to give a good estimate.
373 * Instead, compute an estimate based on the average space usage per
374 * dnode, with an upper and lower cap.
376 * In case there aren't many dnodes or blocks used yet, add a small
377 * correction factor using OSD_DNODE_EST_SHIFT. This correction
378 * factor gradually disappears as the number of real dnodes grows.
379 * This also avoids the need to check for divide-by-zero later.
381 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
382 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
384 est_refdblocks = (refdbytes >> est_maxblockshift) +
385 (OSD_DNODE_EST_COUNT >> OSD_DNODE_EST_BLKSHIFT);
386 est_usedobjs = usedobjs + OSD_DNODE_EST_COUNT;
388 /* Average space/dnode more than maximum dnode size, use max dnode
389 * size to estimate free dnodes from adjusted free blocks count.
390 * OSTs typically use more than one block dnode so this case applies. */
391 if (est_usedobjs <= est_refdblocks * 2) {
394 /* Average space/dnode smaller than min dnode size (probably due to
395 * metadnode compression), use min dnode size to estimate the number of
397 * An MDT typically uses below 512 bytes/dnode so this case applies. */
398 } else if (est_usedobjs >= (est_refdblocks << OSD_DNODE_MIN_BLKSHIFT)) {
399 est_objs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
401 /* Between the extremes, we try to use the average size of
402 * existing dnodes to compute the number of dnodes that fit
405 * est_objs = nrblocks * (est_usedobjs / est_refblocks);
407 * but this may overflow 64 bits or become 0 if not handled well
409 * We know nrblocks is below (64 - 17 = 47) bits from
410 * SPA_MAXBLKSHIFT, and est_usedobjs is under 48 bits due to
411 * DN_MAX_OBJECT_SHIFT, which means that multiplying them may
412 * get as large as 2 ^ 95.
414 * We also know (est_usedobjs / est_refdblocks) is between 2 and
415 * 256, due to above checks, we can safely compute this first.
416 * We care more about accuracy on the MDT (many dnodes/block)
417 * which is good because this is where truncation errors are
418 * smallest. This adds 8 bits to nrblocks so we can use 7 bits
419 * to compute a fixed-point fraction and nrblocks can still fit
422 unsigned dnodes_per_block = (est_usedobjs << 7)/est_refdblocks;
424 est_objs = (nrblocks * dnodes_per_block) >> 7;
429 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
431 struct objset *os = osd->od_os;
432 uint64_t refdbytes, availbytes, usedobjs, availobjs;
433 uint64_t est_availobjs;
437 dmu_objset_space(os, &refdbytes, &availbytes, &usedobjs, &availobjs);
439 memset(osfs, 0, sizeof(*osfs));
441 /* We're a zfs filesystem. */
442 osfs->os_type = UBERBLOCK_MAGIC;
445 * ZFS allows multiple block sizes. For statfs, Linux makes no
446 * proper distinction between bsize and frsize. For calculations
447 * of free and used blocks incorrectly uses bsize instead of frsize,
448 * but bsize is also used as the optimal blocksize. We return the
449 * largest possible block size as IO size for the optimum performance
450 * and scale the free and used blocks count appropriately.
452 osfs->os_bsize = osd->od_max_blksz;
453 bshift = fls64(osfs->os_bsize) - 1;
455 osfs->os_blocks = (refdbytes + availbytes) >> bshift;
456 osfs->os_bfree = availbytes >> bshift;
457 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
459 /* Take replication (i.e. number of copies) into account */
460 osfs->os_bavail /= os->os_copies;
463 * Reserve some space so we don't run into ENOSPC due to grants not
464 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
465 * Rather than report this via os_bavail (which makes users unhappy if
466 * they can't fill the filesystem 100%), reduce os_blocks as well.
468 * Reserve 0.78% of total space, at least 16MB for small filesystems,
469 * for internal files to be created/unlinked when space is tight.
471 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
472 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
473 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
474 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
476 osfs->os_blocks -= reserved;
477 osfs->os_bfree -= min(reserved, osfs->os_bfree);
478 osfs->os_bavail -= min(reserved, osfs->os_bavail);
481 * The availobjs value returned from dmu_objset_space() is largely
482 * useless, since it reports the number of objects that might
483 * theoretically still fit into the dataset, independent of minor
484 * issues like how much space is actually available in the pool.
485 * Compute a better estimate in udmu_objs_count_estimate().
487 est_availobjs = osd_objs_count_estimate(refdbytes, usedobjs,
488 osfs->os_bfree, bshift);
490 osfs->os_ffree = min(availobjs, est_availobjs);
491 osfs->os_files = osfs->os_ffree + usedobjs;
493 /* ZFS XXX: fill in backing dataset FSID/UUID
494 memcpy(osfs->os_fsid, .... );*/
496 osfs->os_namelen = MAXNAMELEN;
497 osfs->os_maxbytes = OBD_OBJECT_EOF;
499 if (!spa_writeable(dmu_objset_spa(os)) ||
500 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
501 osfs->os_state |= OS_STATE_READONLY;
507 * Concurrency: shouldn't matter.
509 int osd_statfs(const struct lu_env *env, struct dt_device *d,
510 struct obd_statfs *osfs)
515 rc = osd_objset_statfs(osd_dt_dev(d), osfs);
516 if (unlikely(rc != 0))
519 osfs->os_bavail -= min_t(u64,
520 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
525 static int osd_blk_insert_cost(struct osd_device *osd)
527 int max_blockshift, nr_blkptrshift, bshift;
529 /* max_blockshift is the log2 of the number of blocks needed to reach
530 * the maximum filesize (that's to say 2^64) */
531 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
532 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
534 /* nr_blkptrshift is the log2 of the number of block pointers that can
535 * be stored in an indirect block */
536 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
537 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
539 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
540 * tree. We add +1 for rounding purpose.
541 * The tree depth times the indirect block size gives us the maximum
542 * cost of inserting a block in the tree */
543 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
547 * Concurrency: doesn't access mutable data.
549 static void osd_conf_get(const struct lu_env *env,
550 const struct dt_device *dev,
551 struct dt_device_param *param)
553 struct osd_device *osd = osd_dt_dev(dev);
556 * XXX should be taken from not-yet-existing fs abstraction layer.
558 param->ddp_max_name_len = MAXNAMELEN;
559 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
560 param->ddp_symlink_max = PATH_MAX;
561 param->ddp_mount_type = LDD_MT_ZFS;
563 param->ddp_mntopts = MNTOPT_USERXATTR;
564 if (osd->od_posix_acl)
565 param->ddp_mntopts |= MNTOPT_ACL;
566 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
568 /* for maxbytes, report same value as ZPL */
569 param->ddp_maxbytes = MAX_LFS_FILESIZE;
571 /* inodes are dynamically allocated, so we report the per-inode space
572 * consumption to upper layers. This static value is not really accurate
573 * and we should use the same logic as in udmu_objset_statfs() to
574 * estimate the real size consumed by an object */
575 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
576 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
577 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
578 * not be accounted for every single new block insertion.
579 * Instead, the maximum extent size is set to the number of blocks that
580 * can fit into a single contiguous indirect block. There would be some
581 * cases where this crosses indirect blocks, but it also won't have 7
582 * new levels of indirect blocks in that case either, so it will still
583 * have enough reserved space for the extra indirect block */
584 param->ddp_max_extent_blks =
585 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
586 param->ddp_extent_tax = osd_blk_insert_cost(osd);
590 * Concurrency: shouldn't matter.
592 static int osd_sync(const struct lu_env *env, struct dt_device *d)
594 struct osd_device *osd = osd_dt_dev(d);
595 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
596 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
597 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
601 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
603 struct osd_device *osd = osd_dt_dev(dev);
604 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
607 mutex_enter(&tx->tx_sync_lock);
608 txg = tx->tx_open_txg + 1;
609 if (tx->tx_quiesce_txg_waiting < txg) {
610 tx->tx_quiesce_txg_waiting = txg;
611 cv_broadcast(&tx->tx_quiesce_more_cv);
613 mutex_exit(&tx->tx_sync_lock);
619 * Concurrency: shouldn't matter.
621 static int osd_ro(const struct lu_env *env, struct dt_device *d)
623 struct osd_device *osd = osd_dt_dev(d);
626 CERROR("%s: *** setting device %s read-only ***\n",
627 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
628 osd->od_dev_set_rdonly = 1;
629 spa_freeze(dmu_objset_spa(osd->od_os));
634 static struct dt_device_operations osd_dt_ops = {
635 .dt_root_get = osd_root_get,
636 .dt_statfs = osd_statfs,
637 .dt_trans_create = osd_trans_create,
638 .dt_trans_start = osd_trans_start,
639 .dt_trans_stop = osd_trans_stop,
640 .dt_trans_cb_add = osd_trans_cb_add,
641 .dt_conf_get = osd_conf_get,
643 .dt_commit_async = osd_commit_async,
648 * DMU OSD device type methods
650 static int osd_type_init(struct lu_device_type *t)
652 LU_CONTEXT_KEY_INIT(&osd_key);
653 return lu_context_key_register(&osd_key);
656 static void osd_type_fini(struct lu_device_type *t)
658 lu_context_key_degister(&osd_key);
661 static void *osd_key_init(const struct lu_context *ctx,
662 struct lu_context_key *key)
664 struct osd_thread_info *info;
668 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
670 info = ERR_PTR(-ENOMEM);
674 static void osd_key_fini(const struct lu_context *ctx,
675 struct lu_context_key *key, void *data)
677 struct osd_thread_info *info = data;
682 static void osd_key_exit(const struct lu_context *ctx,
683 struct lu_context_key *key, void *data)
685 struct osd_thread_info *info = data;
687 memset(info, 0, sizeof(*info));
690 struct lu_context_key osd_key = {
691 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
692 .lct_init = osd_key_init,
693 .lct_fini = osd_key_fini,
694 .lct_exit = osd_key_exit
697 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
699 if (osd->od_cl_seq == NULL)
702 seq_client_fini(osd->od_cl_seq);
703 OBD_FREE_PTR(osd->od_cl_seq);
704 osd->od_cl_seq = NULL;
707 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
711 /* shutdown quota slave instance associated with the device */
712 if (o->od_quota_slave != NULL) {
713 qsd_fini(env, o->od_quota_slave);
714 o->od_quota_slave = NULL;
717 osd_fid_fini(env, o);
722 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
724 struct osd_device *osd = arg;
726 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
729 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
731 struct osd_device *osd = arg;
733 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
734 LASSERT(newval >= SPA_MINBLOCKSIZE);
735 LASSERT(ISP2(newval));
737 osd->od_max_blksz = newval;
740 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
742 struct osd_device *osd = arg;
744 osd->od_prop_rdonly = !!newval;
748 * This function unregisters all registered callbacks. It's harmless to
749 * unregister callbacks that were never registered so it is used to safely
750 * unwind a partially completed call to osd_objset_register_callbacks().
752 static void osd_objset_unregister_callbacks(struct osd_device *o)
754 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
756 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
757 osd_xattr_changed_cb, o);
758 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
759 osd_recordsize_changed_cb, o);
760 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
761 osd_readonly_changed_cb, o);
763 if (o->arc_prune_cb != NULL) {
764 arc_remove_prune_callback(o->arc_prune_cb);
765 o->arc_prune_cb = NULL;
770 * Register the required callbacks to be notified when zfs properties
771 * are modified using the 'zfs(8)' command line utility.
773 static int osd_objset_register_callbacks(struct osd_device *o)
775 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
776 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
782 dsl_pool_config_enter(dp, FTAG);
783 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
784 osd_xattr_changed_cb, o);
788 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
789 osd_recordsize_changed_cb, o);
793 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
794 osd_readonly_changed_cb, o);
798 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
800 dsl_pool_config_exit(dp, FTAG);
802 osd_objset_unregister_callbacks(o);
807 static int osd_objset_open(struct osd_device *o)
809 uint64_t version = ZPL_VERSION;
814 rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os);
816 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
821 /* Check ZFS version */
822 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
823 ZPL_VERSION_STR, 8, 1, &version);
825 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
827 * We can't return ENOENT because that would mean the objset
830 GOTO(out, rc = -EIO);
833 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
834 ZFS_SA_ATTRS, 8, 1, &sa_obj);
838 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
839 ZPL_END, &o->z_attr_table);
843 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
844 8, 1, &o->od_rootid);
846 CERROR("%s: lookup for root failed: rc = %d\n",
851 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
852 8, 1, &o->od_unlinkedid);
854 CERROR("%s: lookup for %s failed: rc = %d\n",
855 o->od_svname, ZFS_UNLINKED_SET, rc);
859 /* Check that user/group usage tracking is supported */
860 if (!dmu_objset_userused_enabled(o->od_os) ||
861 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
862 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
863 CERROR("%s: Space accounting not supported by this target, "
864 "aborting\n", o->od_svname);
865 GOTO(out, -ENOTSUPP);
869 if (rc != 0 && o->od_os != NULL) {
870 dmu_objset_disown(o->od_os, o);
878 osd_unlinked_object_free(struct osd_device *osd, uint64_t oid)
883 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
885 CWARN("%s: Cannot truncate %llu: rc = %d\n",
886 osd->od_svname, oid, rc);
890 tx = dmu_tx_create(osd->od_os);
891 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
892 dmu_tx_hold_zap(tx, osd->od_unlinkedid, FALSE, NULL);
893 rc = -dmu_tx_assign(tx, TXG_WAIT);
895 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
896 osd->od_svname, oid, rc);
900 rc = -zap_remove_int(osd->od_os, osd->od_unlinkedid, oid, tx);
902 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
903 osd->od_svname, oid, rc);
907 rc = -dmu_object_free(osd->od_os, oid, tx);
909 CWARN("%s: Cannot free %llu: rc = %d\n",
910 osd->od_svname, oid, rc);
925 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
928 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
930 zap_cursor_init(&zc, osd->od_os, osd->od_unlinkedid);
932 while (zap_cursor_retrieve(&zc, za) == 0) {
933 /* If cannot free the object, leave it in the unlinked set,
934 * until the OSD is mounted again when obd_unlinked_drain()
936 if (osd_unlinked_object_free(osd, za->za_first_integer) != 0)
938 zap_cursor_advance(&zc);
941 zap_cursor_fini(&zc);
944 static int osd_mount(const struct lu_env *env,
945 struct osd_device *o, struct lustre_cfg *cfg)
947 char *mntdev = lustre_cfg_string(cfg, 1);
948 char *svname = lustre_cfg_string(cfg, 4);
954 if (o->od_os != NULL)
957 if (mntdev == NULL || svname == NULL)
960 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
961 if (rc >= sizeof(o->od_mntdev))
964 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
965 if (rc >= sizeof(o->od_svname))
968 if (server_name_is_ost(o->od_svname))
971 rc = osd_objset_open(o);
975 o->od_xattr_in_sa = B_TRUE;
976 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
978 rc = osd_objset_register_callbacks(o);
982 rc = __osd_obj2dbuf(env, o->od_os, o->od_rootid, &rootdb);
986 o->od_root = rootdb->db_object;
987 sa_buf_rele(rootdb, osd_obj_tag);
989 /* 1. initialize oi before any file create or file open */
990 rc = osd_oi_init(env, o);
994 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
997 o->od_site.ls_bottom_dev = osd2lu_dev(o);
999 rc = lu_site_init_finish(&o->od_site);
1003 /* Use our own ZAP for inode accounting by default, this can be changed
1004 * via procfs to estimate the inode usage from the block usage */
1005 o->od_quota_iused_est = 0;
1007 rc = osd_procfs_init(o, o->od_svname);
1011 /* initialize quota slave instance */
1012 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
1014 if (IS_ERR(o->od_quota_slave)) {
1015 rc = PTR_ERR(o->od_quota_slave);
1016 o->od_quota_slave = NULL;
1020 /* parse mount option "noacl", and enable ACL by default */
1021 opts = lustre_cfg_string(cfg, 3);
1022 if (opts == NULL || strstr(opts, "noacl") == NULL)
1023 o->od_posix_acl = 1;
1025 osd_unlinked_drain(env, o);
1028 dmu_objset_disown(o->od_os, o);
1035 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1039 if (atomic_read(&o->od_zerocopy_alloc))
1040 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1041 atomic_read(&o->od_zerocopy_alloc));
1042 if (atomic_read(&o->od_zerocopy_loan))
1043 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1044 atomic_read(&o->od_zerocopy_loan));
1045 if (atomic_read(&o->od_zerocopy_pin))
1046 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1047 atomic_read(&o->od_zerocopy_pin));
1049 if (o->od_os != NULL) {
1050 /* force a txg sync to get all commit callbacks */
1051 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1053 /* close the object set */
1054 dmu_objset_disown(o->od_os, o);
1062 static int osd_device_init0(const struct lu_env *env,
1063 struct osd_device *o,
1064 struct lustre_cfg *cfg)
1066 struct lu_device *l = osd2lu_dev(o);
1069 /* if the module was re-loaded, env can loose its keys */
1070 rc = lu_env_refill((struct lu_env *) env);
1074 l->ld_ops = &osd_lu_ops;
1075 o->od_dt_dev.dd_ops = &osd_dt_ops;
1081 static struct lu_device *osd_device_fini(const struct lu_env *env,
1082 struct lu_device *dev);
1084 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1085 struct lu_device_type *type,
1086 struct lustre_cfg *cfg)
1088 struct osd_device *dev;
1089 struct osd_seq_list *osl;
1094 return ERR_PTR(-ENOMEM);
1096 osl = &dev->od_seq_list;
1097 INIT_LIST_HEAD(&osl->osl_seq_list);
1098 rwlock_init(&osl->osl_seq_list_lock);
1099 sema_init(&osl->osl_seq_init_sem, 1);
1101 rc = dt_device_init(&dev->od_dt_dev, type);
1103 rc = osd_device_init0(env, dev, cfg);
1105 rc = osd_mount(env, dev, cfg);
1107 osd_device_fini(env, osd2lu_dev(dev));
1110 dt_device_fini(&dev->od_dt_dev);
1113 if (unlikely(rc != 0))
1116 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1119 static struct lu_device *osd_device_free(const struct lu_env *env,
1120 struct lu_device *d)
1122 struct osd_device *o = osd_dev(d);
1125 /* XXX: make osd top device in order to release reference */
1126 d->ld_site->ls_top_dev = d;
1127 lu_site_purge(env, d->ld_site, -1);
1128 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1129 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1130 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1132 lu_site_fini(&o->od_site);
1133 dt_device_fini(&o->od_dt_dev);
1139 static struct lu_device *osd_device_fini(const struct lu_env *env,
1140 struct lu_device *d)
1142 struct osd_device *o = osd_dev(d);
1147 osd_shutdown(env, o);
1148 osd_oi_fini(env, o);
1151 osd_objset_unregister_callbacks(o);
1152 osd_sync(env, lu2dt_dev(d));
1153 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
1156 rc = osd_procfs_fini(o);
1158 CERROR("proc fini error %d\n", rc);
1159 RETURN(ERR_PTR(rc));
1168 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1169 const char *name, struct lu_device *next)
1175 * To be removed, setup is performed by osd_device_{init,alloc} and
1176 * cleanup is performed by osd_device_{fini,free).
1178 static int osd_process_config(const struct lu_env *env,
1179 struct lu_device *d, struct lustre_cfg *cfg)
1181 struct osd_device *o = osd_dev(d);
1185 switch(cfg->lcfg_command) {
1187 rc = osd_mount(env, o, cfg);
1190 rc = osd_shutdown(env, o);
1193 LASSERT(&o->od_dt_dev);
1194 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1195 cfg, &o->od_dt_dev);
1196 if (rc > 0 || rc == -ENOSYS)
1197 rc = class_process_proc_param(PARAM_OST,
1198 lprocfs_osd_obd_vars,
1199 cfg, &o->od_dt_dev);
1209 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1211 struct osd_device *osd = osd_dev(d);
1215 if (osd->od_quota_slave == NULL)
1218 /* start qsd instance on recovery completion, this notifies the quota
1219 * slave code that we are about to process new requests now */
1220 rc = qsd_start(env, osd->od_quota_slave);
1225 * we use exports to track all osd users
1227 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1228 struct obd_device *obd, struct obd_uuid *cluuid,
1229 struct obd_connect_data *data, void *localdata)
1231 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1232 struct lustre_handle conn;
1236 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1238 rc = class_connect(&conn, obd, cluuid);
1242 *exp = class_conn2export(&conn);
1244 spin_lock(&obd->obd_dev_lock);
1246 spin_unlock(&obd->obd_dev_lock);
1252 * once last export (we don't count self-export) disappeared
1253 * osd can be released
1255 static int osd_obd_disconnect(struct obd_export *exp)
1257 struct obd_device *obd = exp->exp_obd;
1258 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1259 int rc, release = 0;
1262 /* Only disconnect the underlying layers on the final disconnect. */
1263 spin_lock(&obd->obd_dev_lock);
1265 if (osd->od_connects == 0)
1267 spin_unlock(&obd->obd_dev_lock);
1269 rc = class_disconnect(exp); /* bz 9811 */
1271 if (rc == 0 && release)
1272 class_manual_cleanup(obd);
1276 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1278 struct seq_server_site *ss = osd_seq_site(osd);
1282 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1285 if (unlikely(ss == NULL))
1288 OBD_ALLOC_PTR(osd->od_cl_seq);
1289 if (osd->od_cl_seq == NULL)
1292 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1293 osd->od_svname, ss->ss_server_seq);
1296 OBD_FREE_PTR(osd->od_cl_seq);
1297 osd->od_cl_seq = NULL;
1303 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1304 struct lu_device *dev)
1306 struct osd_device *osd = osd_dev(dev);
1310 if (osd->od_quota_slave != NULL) {
1311 /* set up quota slave objects */
1312 rc = qsd_prepare(env, osd->od_quota_slave);
1317 rc = osd_fid_init(env, osd);
1322 struct lu_device_operations osd_lu_ops = {
1323 .ldo_object_alloc = osd_object_alloc,
1324 .ldo_process_config = osd_process_config,
1325 .ldo_recovery_complete = osd_recovery_complete,
1326 .ldo_prepare = osd_prepare,
1329 static void osd_type_start(struct lu_device_type *t)
1333 static void osd_type_stop(struct lu_device_type *t)
1337 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1338 struct lu_fid *fid, struct md_op_data *op_data)
1340 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1342 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1345 static struct lu_device_type_operations osd_device_type_ops = {
1346 .ldto_init = osd_type_init,
1347 .ldto_fini = osd_type_fini,
1349 .ldto_start = osd_type_start,
1350 .ldto_stop = osd_type_stop,
1352 .ldto_device_alloc = osd_device_alloc,
1353 .ldto_device_free = osd_device_free,
1355 .ldto_device_init = osd_device_init,
1356 .ldto_device_fini = osd_device_fini
1359 static struct lu_device_type osd_device_type = {
1360 .ldt_tags = LU_DEVICE_DT,
1361 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1362 .ldt_ops = &osd_device_type_ops,
1363 .ldt_ctx_tags = LCT_LOCAL
1367 static struct obd_ops osd_obd_device_ops = {
1368 .o_owner = THIS_MODULE,
1369 .o_connect = osd_obd_connect,
1370 .o_disconnect = osd_obd_disconnect,
1371 .o_fid_alloc = osd_fid_alloc
1374 static int __init osd_init(void)
1378 rc = osd_options_init();
1382 rc = lu_kmem_init(osd_caches);
1386 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1387 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1389 lu_kmem_fini(osd_caches);
1393 static void __exit osd_exit(void)
1395 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1396 lu_kmem_fini(osd_caches);
1399 extern unsigned int osd_oi_count;
1400 module_param(osd_oi_count, int, 0444);
1401 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1403 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1404 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1405 MODULE_VERSION(LUSTRE_VERSION_STRING);
1406 MODULE_LICENSE("GPL");
1408 module_init(osd_init);
1409 module_exit(osd_exit);