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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osd-zfs/osd_handler.c
37 * Top-level entry points into osd module
39 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
40 * Author: Mike Pershin <tappro@whamcloud.com>
41 * Author: Johann Lombardi <johann@whamcloud.com>
44 #define DEBUG_SUBSYSTEM S_OSD
46 #include <lustre_ver.h>
47 #include <libcfs/libcfs.h>
48 #include <obd_support.h>
49 #include <lustre_net.h>
51 #include <obd_class.h>
52 #include <lustre_disk.h>
53 #include <lustre_fid.h>
54 #include <lustre_param.h>
55 #include <md_object.h>
57 #include "osd_internal.h"
59 #include <sys/dnode.h>
64 #include <sys/spa_impl.h>
65 #include <sys/zfs_znode.h>
66 #include <sys/dmu_tx.h>
67 #include <sys/dmu_objset.h>
68 #include <sys/dsl_prop.h>
69 #include <sys/sa_impl.h>
72 struct lu_context_key osd_key;
74 /* Slab for OSD object allocation */
75 struct kmem_cache *osd_object_kmem;
77 /* Slab to allocate osd_zap_it */
78 struct kmem_cache *osd_zapit_cachep;
80 static struct lu_kmem_descr osd_caches[] = {
82 .ckd_cache = &osd_object_kmem,
83 .ckd_name = "zfs_osd_obj",
84 .ckd_size = sizeof(struct osd_object)
87 .ckd_cache = &osd_zapit_cachep,
88 .ckd_name = "osd_zapit_cache",
89 .ckd_size = sizeof(struct osd_zap_it)
96 static void arc_prune_func(int64_t bytes, void *private)
98 struct osd_device *od = private;
99 struct lu_site *site = &od->od_site;
103 rc = lu_env_init(&env, LCT_SHRINKER);
105 CERROR("%s: can't initialize shrinker env: rc = %d\n",
110 lu_site_purge(&env, site, (bytes >> 10));
116 * Concurrency: doesn't access mutable data
118 static int osd_root_get(const struct lu_env *env,
119 struct dt_device *dev, struct lu_fid *f)
121 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
126 * OSD object methods.
130 * Concurrency: shouldn't matter.
132 static void osd_trans_commit_cb(void *cb_data, int error)
134 struct osd_thandle *oh = cb_data;
135 struct thandle *th = &oh->ot_super;
136 struct osd_device *osd = osd_dt_dev(th->th_dev);
137 struct lu_device *lud = &th->th_dev->dd_lu_dev;
138 struct dt_txn_commit_cb *dcb, *tmp;
143 if (error == ECANCELED)
144 CWARN("%s: transaction @0x%p was aborted\n",
145 osd_dt_dev(th->th_dev)->od_svname, th);
147 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
148 osd_dt_dev(th->th_dev)->od_svname, th, error);
151 dt_txn_hook_commit(th);
153 /* call per-transaction callbacks if any */
154 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage)
155 dcb->dcb_func(NULL, th, dcb, error);
157 /* Unlike ldiskfs, zfs updates space accounting at commit time.
158 * As a consequence, op_end is called only now to inform the quota slave
159 * component that reserved quota space is now accounted in usage and
160 * should be released. Quota space won't be adjusted at this point since
161 * we can't provide a suitable environment. It will be performed
162 * asynchronously by a lquota thread. */
163 qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);
167 lu_context_exit(&th->th_ctx);
168 lu_context_fini(&th->th_ctx);
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_context_init(&th->th_ctx, th->th_tags);
227 lu_context_enter(&th->th_ctx);
228 lu_device_get(&d->dd_lu_dev);
234 static int osd_unlinked_object_free(struct osd_device *osd, uint64_t oid);
236 static void osd_unlinked_list_emptify(struct osd_device *osd,
237 struct list_head *list, bool free)
239 struct osd_object *obj;
242 while (!list_empty(list)) {
243 obj = list_entry(list->next,
244 struct osd_object, oo_unlinked_linkage);
245 LASSERT(obj->oo_db != NULL);
246 oid = obj->oo_db->db_object;
248 list_del_init(&obj->oo_unlinked_linkage);
250 (void)osd_unlinked_object_free(osd, oid);
254 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
256 struct dt_txn_commit_cb *dcb;
257 struct dt_txn_commit_cb *tmp;
259 /* call per-transaction stop callbacks if any */
260 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
262 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
263 "commit callback entry: magic=%x name='%s'\n",
264 dcb->dcb_magic, dcb->dcb_name);
265 list_del_init(&dcb->dcb_linkage);
266 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
271 * Concurrency: shouldn't matter.
273 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
276 struct osd_device *osd = osd_dt_dev(th->th_dev);
277 bool sync = (th->th_sync != 0);
278 struct osd_thandle *oh;
279 struct list_head unlinked;
284 oh = container_of0(th, struct osd_thandle, ot_super);
285 INIT_LIST_HEAD(&unlinked);
286 list_splice_init(&oh->ot_unlinked_list, &unlinked);
288 if (oh->ot_assigned == 0) {
290 dmu_tx_abort(oh->ot_tx);
291 osd_object_sa_dirty_rele(env, oh);
292 osd_unlinked_list_emptify(osd, &unlinked, false);
293 /* there won't be any commit, release reserved quota space now,
295 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
300 /* When doing our own inode accounting, the ZAPs storing per-uid/gid
301 * usage are updated at operation execution time, so we should call
302 * qsd_op_end() straight away. Otherwise (for blk accounting maintained
303 * by ZFS and when #inode is estimated from #blks) accounting is updated
304 * at commit time and the call to qsd_op_end() must be delayed */
305 if (oh->ot_quota_trans.lqt_id_cnt > 0 &&
306 !oh->ot_quota_trans.lqt_ids[0].lqi_is_blk &&
307 !osd->od_quota_iused_est)
308 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
310 rc = dt_txn_hook_stop(env, th);
312 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
315 osd_trans_stop_cb(oh, rc);
318 txg = oh->ot_tx->tx_txg;
320 osd_object_sa_dirty_rele(env, oh);
321 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
322 * by osd_trans_commit_cb already. */
323 dmu_tx_commit(oh->ot_tx);
325 osd_unlinked_list_emptify(osd, &unlinked, true);
328 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
333 static struct thandle *osd_trans_create(const struct lu_env *env,
334 struct dt_device *dt)
336 struct osd_device *osd = osd_dt_dev(dt);
337 struct osd_thandle *oh;
342 tx = dmu_tx_create(osd->od_os);
344 RETURN(ERR_PTR(-ENOMEM));
346 /* alloc callback data */
350 RETURN(ERR_PTR(-ENOMEM));
354 INIT_LIST_HEAD(&oh->ot_dcb_list);
355 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
356 INIT_LIST_HEAD(&oh->ot_unlinked_list);
357 INIT_LIST_HEAD(&oh->ot_sa_list);
358 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
362 th->th_tags = LCT_TX_HANDLE;
366 /* Estimate the number of objects from a number of blocks */
367 uint64_t osd_objs_count_estimate(uint64_t refdbytes, uint64_t usedobjs,
368 uint64_t nrblocks, uint64_t est_maxblockshift)
370 uint64_t est_objs, est_refdblocks, est_usedobjs;
372 /* Compute an nrblocks estimate based on the actual number of
373 * dnodes that could fit in the space. Since we don't know the
374 * overhead associated with each dnode (xattrs, SAs, VDEV overhead,
375 * etc) just using DNODE_SHIFT isn't going to give a good estimate.
376 * Instead, compute an estimate based on the average space usage per
377 * dnode, with an upper and lower cap.
379 * In case there aren't many dnodes or blocks used yet, add a small
380 * correction factor using OSD_DNODE_EST_SHIFT. This correction
381 * factor gradually disappears as the number of real dnodes grows.
382 * This also avoids the need to check for divide-by-zero later.
384 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
385 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
387 est_refdblocks = (refdbytes >> est_maxblockshift) +
388 (OSD_DNODE_EST_COUNT >> OSD_DNODE_EST_BLKSHIFT);
389 est_usedobjs = usedobjs + OSD_DNODE_EST_COUNT;
391 /* Average space/dnode more than maximum dnode size, use max dnode
392 * size to estimate free dnodes from adjusted free blocks count.
393 * OSTs typically use more than one block dnode so this case applies. */
394 if (est_usedobjs <= est_refdblocks * 2) {
397 /* Average space/dnode smaller than min dnode size (probably due to
398 * metadnode compression), use min dnode size to estimate the number of
400 * An MDT typically uses below 512 bytes/dnode so this case applies. */
401 } else if (est_usedobjs >= (est_refdblocks << OSD_DNODE_MIN_BLKSHIFT)) {
402 est_objs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
404 /* Between the extremes, we try to use the average size of
405 * existing dnodes to compute the number of dnodes that fit
408 * est_objs = nrblocks * (est_usedobjs / est_refblocks);
410 * but this may overflow 64 bits or become 0 if not handled well
412 * We know nrblocks is below (64 - 17 = 47) bits from
413 * SPA_MAXBLKSHIFT, and est_usedobjs is under 48 bits due to
414 * DN_MAX_OBJECT_SHIFT, which means that multiplying them may
415 * get as large as 2 ^ 95.
417 * We also know (est_usedobjs / est_refdblocks) is between 2 and
418 * 256, due to above checks, we can safely compute this first.
419 * We care more about accuracy on the MDT (many dnodes/block)
420 * which is good because this is where truncation errors are
421 * smallest. This adds 8 bits to nrblocks so we can use 7 bits
422 * to compute a fixed-point fraction and nrblocks can still fit
425 unsigned dnodes_per_block = (est_usedobjs << 7)/est_refdblocks;
427 est_objs = (nrblocks * dnodes_per_block) >> 7;
432 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
434 struct objset *os = osd->od_os;
435 uint64_t refdbytes, availbytes, usedobjs, availobjs;
436 uint64_t est_availobjs;
440 dmu_objset_space(os, &refdbytes, &availbytes, &usedobjs, &availobjs);
442 memset(osfs, 0, sizeof(*osfs));
444 /* We're a zfs filesystem. */
445 osfs->os_type = UBERBLOCK_MAGIC;
448 * ZFS allows multiple block sizes. For statfs, Linux makes no
449 * proper distinction between bsize and frsize. For calculations
450 * of free and used blocks incorrectly uses bsize instead of frsize,
451 * but bsize is also used as the optimal blocksize. We return the
452 * largest possible block size as IO size for the optimum performance
453 * and scale the free and used blocks count appropriately.
455 osfs->os_bsize = osd->od_max_blksz;
456 bshift = fls64(osfs->os_bsize) - 1;
458 osfs->os_blocks = (refdbytes + availbytes) >> bshift;
459 osfs->os_bfree = availbytes >> bshift;
460 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
462 /* Take replication (i.e. number of copies) into account */
463 osfs->os_bavail /= os->os_copies;
466 * Reserve some space so we don't run into ENOSPC due to grants not
467 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
468 * Rather than report this via os_bavail (which makes users unhappy if
469 * they can't fill the filesystem 100%), reduce os_blocks as well.
471 * Reserve 0.78% of total space, at least 16MB for small filesystems,
472 * for internal files to be created/unlinked when space is tight.
474 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
475 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
476 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
477 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
479 osfs->os_blocks -= reserved;
480 osfs->os_bfree -= min(reserved, osfs->os_bfree);
481 osfs->os_bavail -= min(reserved, osfs->os_bavail);
484 * The availobjs value returned from dmu_objset_space() is largely
485 * useless, since it reports the number of objects that might
486 * theoretically still fit into the dataset, independent of minor
487 * issues like how much space is actually available in the pool.
488 * Compute a better estimate in udmu_objs_count_estimate().
490 est_availobjs = osd_objs_count_estimate(refdbytes, usedobjs,
491 osfs->os_bfree, bshift);
493 osfs->os_ffree = min(availobjs, est_availobjs);
494 osfs->os_files = osfs->os_ffree + usedobjs;
496 /* ZFS XXX: fill in backing dataset FSID/UUID
497 memcpy(osfs->os_fsid, .... );*/
499 osfs->os_namelen = MAXNAMELEN;
500 osfs->os_maxbytes = OBD_OBJECT_EOF;
502 if (!spa_writeable(dmu_objset_spa(os)) ||
503 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
504 osfs->os_state |= OS_STATE_READONLY;
510 * Concurrency: shouldn't matter.
512 int osd_statfs(const struct lu_env *env, struct dt_device *d,
513 struct obd_statfs *osfs)
518 rc = osd_objset_statfs(osd_dt_dev(d), osfs);
519 if (unlikely(rc != 0))
522 osfs->os_bavail -= min_t(u64,
523 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
528 static int osd_blk_insert_cost(struct osd_device *osd)
530 int max_blockshift, nr_blkptrshift, bshift;
532 /* max_blockshift is the log2 of the number of blocks needed to reach
533 * the maximum filesize (that's to say 2^64) */
534 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
535 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
537 /* nr_blkptrshift is the log2 of the number of block pointers that can
538 * be stored in an indirect block */
539 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
540 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
542 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
543 * tree. We add +1 for rounding purpose.
544 * The tree depth times the indirect block size gives us the maximum
545 * cost of inserting a block in the tree */
546 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
550 * Concurrency: doesn't access mutable data.
552 static void osd_conf_get(const struct lu_env *env,
553 const struct dt_device *dev,
554 struct dt_device_param *param)
556 struct osd_device *osd = osd_dt_dev(dev);
559 * XXX should be taken from not-yet-existing fs abstraction layer.
561 param->ddp_max_name_len = MAXNAMELEN;
562 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
563 param->ddp_symlink_max = PATH_MAX;
564 param->ddp_mount_type = LDD_MT_ZFS;
566 param->ddp_mntopts = MNTOPT_USERXATTR;
567 if (osd->od_posix_acl)
568 param->ddp_mntopts |= MNTOPT_ACL;
569 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
571 /* for maxbytes, report same value as ZPL */
572 param->ddp_maxbytes = MAX_LFS_FILESIZE;
574 /* inodes are dynamically allocated, so we report the per-inode space
575 * consumption to upper layers. This static value is not really accurate
576 * and we should use the same logic as in udmu_objset_statfs() to
577 * estimate the real size consumed by an object */
578 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
579 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
580 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
581 * not be accounted for every single new block insertion.
582 * Instead, the maximum extent size is set to the number of blocks that
583 * can fit into a single contiguous indirect block. There would be some
584 * cases where this crosses indirect blocks, but it also won't have 7
585 * new levels of indirect blocks in that case either, so it will still
586 * have enough reserved space for the extra indirect block */
587 param->ddp_max_extent_blks =
588 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
589 param->ddp_extent_tax = osd_blk_insert_cost(osd);
593 * Concurrency: shouldn't matter.
595 static int osd_sync(const struct lu_env *env, struct dt_device *d)
597 struct osd_device *osd = osd_dt_dev(d);
598 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
599 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
600 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
604 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
606 struct osd_device *osd = osd_dt_dev(dev);
607 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
610 mutex_enter(&tx->tx_sync_lock);
611 txg = tx->tx_open_txg + 1;
612 if (tx->tx_quiesce_txg_waiting < txg) {
613 tx->tx_quiesce_txg_waiting = txg;
614 cv_broadcast(&tx->tx_quiesce_more_cv);
616 mutex_exit(&tx->tx_sync_lock);
622 * Concurrency: shouldn't matter.
624 static int osd_ro(const struct lu_env *env, struct dt_device *d)
626 struct osd_device *osd = osd_dt_dev(d);
629 CERROR("%s: *** setting device %s read-only ***\n",
630 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
631 osd->od_dev_set_rdonly = 1;
632 spa_freeze(dmu_objset_spa(osd->od_os));
637 static struct dt_device_operations osd_dt_ops = {
638 .dt_root_get = osd_root_get,
639 .dt_statfs = osd_statfs,
640 .dt_trans_create = osd_trans_create,
641 .dt_trans_start = osd_trans_start,
642 .dt_trans_stop = osd_trans_stop,
643 .dt_trans_cb_add = osd_trans_cb_add,
644 .dt_conf_get = osd_conf_get,
646 .dt_commit_async = osd_commit_async,
651 * DMU OSD device type methods
653 static int osd_type_init(struct lu_device_type *t)
655 LU_CONTEXT_KEY_INIT(&osd_key);
656 return lu_context_key_register(&osd_key);
659 static void osd_type_fini(struct lu_device_type *t)
661 lu_context_key_degister(&osd_key);
664 static void *osd_key_init(const struct lu_context *ctx,
665 struct lu_context_key *key)
667 struct osd_thread_info *info;
671 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
673 info = ERR_PTR(-ENOMEM);
677 static void osd_key_fini(const struct lu_context *ctx,
678 struct lu_context_key *key, void *data)
680 struct osd_thread_info *info = data;
682 lu_buf_free(&info->oti_xattr_lbuf);
686 static void osd_key_exit(const struct lu_context *ctx,
687 struct lu_context_key *key, void *data)
691 struct lu_context_key osd_key = {
692 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
693 .lct_init = osd_key_init,
694 .lct_fini = osd_key_fini,
695 .lct_exit = osd_key_exit
698 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
700 if (osd->od_cl_seq == NULL)
703 seq_client_fini(osd->od_cl_seq);
704 OBD_FREE_PTR(osd->od_cl_seq);
705 osd->od_cl_seq = NULL;
708 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
712 /* shutdown quota slave instance associated with the device */
713 if (o->od_quota_slave != NULL) {
714 qsd_fini(env, o->od_quota_slave);
715 o->od_quota_slave = NULL;
718 osd_fid_fini(env, o);
723 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
725 struct osd_device *osd = arg;
727 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
730 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
732 struct osd_device *osd = arg;
734 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
735 LASSERT(newval >= SPA_MINBLOCKSIZE);
736 LASSERT(ISP2(newval));
738 osd->od_max_blksz = newval;
741 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
743 struct osd_device *osd = arg;
745 osd->od_prop_rdonly = !!newval;
749 * This function unregisters all registered callbacks. It's harmless to
750 * unregister callbacks that were never registered so it is used to safely
751 * unwind a partially completed call to osd_objset_register_callbacks().
753 static void osd_objset_unregister_callbacks(struct osd_device *o)
755 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
757 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
758 osd_xattr_changed_cb, o);
759 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
760 osd_recordsize_changed_cb, o);
761 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
762 osd_readonly_changed_cb, o);
764 if (o->arc_prune_cb != NULL) {
765 arc_remove_prune_callback(o->arc_prune_cb);
766 o->arc_prune_cb = NULL;
771 * Register the required callbacks to be notified when zfs properties
772 * are modified using the 'zfs(8)' command line utility.
774 static int osd_objset_register_callbacks(struct osd_device *o)
776 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
777 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
783 dsl_pool_config_enter(dp, FTAG);
784 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
785 osd_xattr_changed_cb, o);
789 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
790 osd_recordsize_changed_cb, o);
794 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
795 osd_readonly_changed_cb, o);
799 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
801 dsl_pool_config_exit(dp, FTAG);
803 osd_objset_unregister_callbacks(o);
808 static int osd_objset_open(struct osd_device *o)
810 uint64_t version = ZPL_VERSION;
815 rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os);
817 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
822 /* Check ZFS version */
823 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
824 ZPL_VERSION_STR, 8, 1, &version);
826 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
828 * We can't return ENOENT because that would mean the objset
831 GOTO(out, rc = -EIO);
834 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
835 ZFS_SA_ATTRS, 8, 1, &sa_obj);
839 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
840 ZPL_END, &o->z_attr_table);
844 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
845 8, 1, &o->od_rootid);
847 CERROR("%s: lookup for root failed: rc = %d\n",
852 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
853 8, 1, &o->od_unlinkedid);
855 CERROR("%s: lookup for %s failed: rc = %d\n",
856 o->od_svname, ZFS_UNLINKED_SET, rc);
860 /* Check that user/group usage tracking is supported */
861 if (!dmu_objset_userused_enabled(o->od_os) ||
862 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
863 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
864 CERROR("%s: Space accounting not supported by this target, "
865 "aborting\n", o->od_svname);
866 GOTO(out, -ENOTSUPP);
870 if (rc != 0 && o->od_os != NULL) {
871 dmu_objset_disown(o->od_os, o);
879 osd_unlinked_object_free(struct osd_device *osd, uint64_t oid)
884 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
886 CWARN("%s: Cannot truncate "LPU64": rc = %d\n",
887 osd->od_svname, oid, rc);
891 tx = dmu_tx_create(osd->od_os);
892 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
893 dmu_tx_hold_zap(tx, osd->od_unlinkedid, FALSE, NULL);
894 rc = -dmu_tx_assign(tx, TXG_WAIT);
896 CWARN("%s: Cannot assign tx for "LPU64": rc = %d\n",
897 osd->od_svname, oid, rc);
901 rc = -zap_remove_int(osd->od_os, osd->od_unlinkedid, oid, tx);
903 CWARN("%s: Cannot remove "LPU64" from unlinked set: rc = %d\n",
904 osd->od_svname, oid, rc);
908 rc = -dmu_object_free(osd->od_os, oid, tx);
910 CWARN("%s: Cannot free "LPU64": rc = %d\n",
911 osd->od_svname, oid, rc);
926 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
929 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
931 zap_cursor_init(&zc, osd->od_os, osd->od_unlinkedid);
933 while (zap_cursor_retrieve(&zc, za) == 0) {
934 /* If cannot free the object, leave it in the unlinked set,
935 * until the OSD is mounted again when obd_unlinked_drain()
937 if (osd_unlinked_object_free(osd, za->za_first_integer) != 0)
939 zap_cursor_advance(&zc);
942 zap_cursor_fini(&zc);
945 static int osd_mount(const struct lu_env *env,
946 struct osd_device *o, struct lustre_cfg *cfg)
948 char *mntdev = lustre_cfg_string(cfg, 1);
949 char *svname = lustre_cfg_string(cfg, 4);
955 if (o->od_os != NULL)
958 if (mntdev == NULL || svname == NULL)
961 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
962 if (rc >= sizeof(o->od_mntdev))
965 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
966 if (rc >= sizeof(o->od_svname))
969 if (server_name_is_ost(o->od_svname))
972 rc = osd_objset_open(o);
976 o->od_xattr_in_sa = B_TRUE;
977 o->od_max_blksz = SPA_OLD_MAXBLOCKSIZE;
979 rc = osd_objset_register_callbacks(o);
983 rc = __osd_obj2dbuf(env, o->od_os, o->od_rootid, &rootdb);
987 o->od_root = rootdb->db_object;
988 sa_buf_rele(rootdb, osd_obj_tag);
990 /* 1. initialize oi before any file create or file open */
991 rc = osd_oi_init(env, o);
995 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
998 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1000 rc = lu_site_init_finish(&o->od_site);
1004 /* Use our own ZAP for inode accounting by default, this can be changed
1005 * via procfs to estimate the inode usage from the block usage */
1006 o->od_quota_iused_est = 0;
1008 rc = osd_procfs_init(o, o->od_svname);
1012 /* initialize quota slave instance */
1013 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
1015 if (IS_ERR(o->od_quota_slave)) {
1016 rc = PTR_ERR(o->od_quota_slave);
1017 o->od_quota_slave = NULL;
1021 /* parse mount option "noacl", and enable ACL by default */
1022 opts = lustre_cfg_string(cfg, 3);
1023 if (opts == NULL || strstr(opts, "noacl") == NULL)
1024 o->od_posix_acl = 1;
1026 osd_unlinked_drain(env, o);
1030 dmu_objset_disown(o->od_os, o);
1037 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1041 if (atomic_read(&o->od_zerocopy_alloc))
1042 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1043 atomic_read(&o->od_zerocopy_alloc));
1044 if (atomic_read(&o->od_zerocopy_loan))
1045 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1046 atomic_read(&o->od_zerocopy_loan));
1047 if (atomic_read(&o->od_zerocopy_pin))
1048 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1049 atomic_read(&o->od_zerocopy_pin));
1051 if (o->od_os != NULL) {
1052 /* force a txg sync to get all commit callbacks */
1053 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1055 /* close the object set */
1056 dmu_objset_disown(o->od_os, o);
1064 static int osd_device_init0(const struct lu_env *env,
1065 struct osd_device *o,
1066 struct lustre_cfg *cfg)
1068 struct lu_device *l = osd2lu_dev(o);
1071 /* if the module was re-loaded, env can loose its keys */
1072 rc = lu_env_refill((struct lu_env *) env);
1076 l->ld_ops = &osd_lu_ops;
1077 o->od_dt_dev.dd_ops = &osd_dt_ops;
1083 static struct lu_device *osd_device_fini(const struct lu_env *env,
1084 struct lu_device *dev);
1086 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1087 struct lu_device_type *type,
1088 struct lustre_cfg *cfg)
1090 struct osd_device *dev;
1091 struct osd_seq_list *osl;
1096 return ERR_PTR(-ENOMEM);
1098 osl = &dev->od_seq_list;
1099 INIT_LIST_HEAD(&osl->osl_seq_list);
1100 rwlock_init(&osl->osl_seq_list_lock);
1101 sema_init(&osl->osl_seq_init_sem, 1);
1103 rc = dt_device_init(&dev->od_dt_dev, type);
1105 rc = osd_device_init0(env, dev, cfg);
1107 rc = osd_mount(env, dev, cfg);
1109 osd_device_fini(env, osd2lu_dev(dev));
1112 dt_device_fini(&dev->od_dt_dev);
1115 if (unlikely(rc != 0))
1118 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1121 static struct lu_device *osd_device_free(const struct lu_env *env,
1122 struct lu_device *d)
1124 struct osd_device *o = osd_dev(d);
1127 /* XXX: make osd top device in order to release reference */
1128 d->ld_site->ls_top_dev = d;
1129 lu_site_purge(env, d->ld_site, -1);
1130 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1131 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1132 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1134 lu_site_fini(&o->od_site);
1135 dt_device_fini(&o->od_dt_dev);
1141 static struct lu_device *osd_device_fini(const struct lu_env *env,
1142 struct lu_device *d)
1144 struct osd_device *o = osd_dev(d);
1149 osd_shutdown(env, o);
1150 osd_oi_fini(env, o);
1153 osd_objset_unregister_callbacks(o);
1154 osd_sync(env, lu2dt_dev(d));
1155 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
1158 rc = osd_procfs_fini(o);
1160 CERROR("proc fini error %d\n", rc);
1161 RETURN(ERR_PTR(rc));
1170 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1171 const char *name, struct lu_device *next)
1177 * To be removed, setup is performed by osd_device_{init,alloc} and
1178 * cleanup is performed by osd_device_{fini,free).
1180 static int osd_process_config(const struct lu_env *env,
1181 struct lu_device *d, struct lustre_cfg *cfg)
1183 struct osd_device *o = osd_dev(d);
1187 switch(cfg->lcfg_command) {
1189 rc = osd_mount(env, o, cfg);
1192 rc = osd_shutdown(env, o);
1195 LASSERT(&o->od_dt_dev);
1196 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1197 cfg, &o->od_dt_dev);
1198 if (rc > 0 || rc == -ENOSYS)
1199 rc = class_process_proc_param(PARAM_OST,
1200 lprocfs_osd_obd_vars,
1201 cfg, &o->od_dt_dev);
1211 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1213 struct osd_device *osd = osd_dev(d);
1217 if (osd->od_quota_slave == NULL)
1220 /* start qsd instance on recovery completion, this notifies the quota
1221 * slave code that we are about to process new requests now */
1222 rc = qsd_start(env, osd->od_quota_slave);
1227 * we use exports to track all osd users
1229 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1230 struct obd_device *obd, struct obd_uuid *cluuid,
1231 struct obd_connect_data *data, void *localdata)
1233 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1234 struct lustre_handle conn;
1238 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1240 rc = class_connect(&conn, obd, cluuid);
1244 *exp = class_conn2export(&conn);
1246 spin_lock(&obd->obd_dev_lock);
1248 spin_unlock(&obd->obd_dev_lock);
1254 * once last export (we don't count self-export) disappeared
1255 * osd can be released
1257 static int osd_obd_disconnect(struct obd_export *exp)
1259 struct obd_device *obd = exp->exp_obd;
1260 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1261 int rc, release = 0;
1264 /* Only disconnect the underlying layers on the final disconnect. */
1265 spin_lock(&obd->obd_dev_lock);
1267 if (osd->od_connects == 0)
1269 spin_unlock(&obd->obd_dev_lock);
1271 rc = class_disconnect(exp); /* bz 9811 */
1273 if (rc == 0 && release)
1274 class_manual_cleanup(obd);
1278 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1280 struct seq_server_site *ss = osd_seq_site(osd);
1284 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1287 if (unlikely(ss == NULL))
1290 OBD_ALLOC_PTR(osd->od_cl_seq);
1291 if (osd->od_cl_seq == NULL)
1294 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1295 osd->od_svname, ss->ss_server_seq);
1298 OBD_FREE_PTR(osd->od_cl_seq);
1299 osd->od_cl_seq = NULL;
1305 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1306 struct lu_device *dev)
1308 struct osd_device *osd = osd_dev(dev);
1312 if (osd->od_quota_slave != NULL) {
1313 /* set up quota slave objects */
1314 rc = qsd_prepare(env, osd->od_quota_slave);
1319 rc = osd_fid_init(env, osd);
1324 struct lu_device_operations osd_lu_ops = {
1325 .ldo_object_alloc = osd_object_alloc,
1326 .ldo_process_config = osd_process_config,
1327 .ldo_recovery_complete = osd_recovery_complete,
1328 .ldo_prepare = osd_prepare,
1331 static void osd_type_start(struct lu_device_type *t)
1335 static void osd_type_stop(struct lu_device_type *t)
1339 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1340 struct lu_fid *fid, struct md_op_data *op_data)
1342 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1344 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1347 static struct lu_device_type_operations osd_device_type_ops = {
1348 .ldto_init = osd_type_init,
1349 .ldto_fini = osd_type_fini,
1351 .ldto_start = osd_type_start,
1352 .ldto_stop = osd_type_stop,
1354 .ldto_device_alloc = osd_device_alloc,
1355 .ldto_device_free = osd_device_free,
1357 .ldto_device_init = osd_device_init,
1358 .ldto_device_fini = osd_device_fini
1361 static struct lu_device_type osd_device_type = {
1362 .ldt_tags = LU_DEVICE_DT,
1363 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1364 .ldt_ops = &osd_device_type_ops,
1365 .ldt_ctx_tags = LCT_LOCAL
1369 static struct obd_ops osd_obd_device_ops = {
1370 .o_owner = THIS_MODULE,
1371 .o_connect = osd_obd_connect,
1372 .o_disconnect = osd_obd_disconnect,
1373 .o_fid_alloc = osd_fid_alloc
1376 static int __init osd_init(void)
1380 rc = osd_options_init();
1384 rc = lu_kmem_init(osd_caches);
1388 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1389 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1391 lu_kmem_fini(osd_caches);
1395 static void __exit osd_exit(void)
1397 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1398 lu_kmem_fini(osd_caches);
1401 extern unsigned int osd_oi_count;
1402 module_param(osd_oi_count, int, 0444);
1403 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1405 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1406 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1407 MODULE_VERSION(LUSTRE_VERSION_STRING);
1408 MODULE_LICENSE("GPL");
1410 module_init(osd_init);
1411 module_exit(osd_exit);