4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osd-zfs/osd_handler.c
33 * Top-level entry points into osd module
35 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
36 * Author: Mike Pershin <tappro@whamcloud.com>
37 * Author: Johann Lombardi <johann@whamcloud.com>
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <lustre_ver.h>
43 #include <libcfs/libcfs.h>
44 #include <obd_support.h>
45 #include <lustre_net.h>
47 #include <obd_class.h>
48 #include <lustre_disk.h>
49 #include <lustre_fid.h>
50 #include <lustre_param.h>
51 #include <md_object.h>
53 #include "osd_internal.h"
55 #include <sys/dnode.h>
60 #include <sys/spa_impl.h>
61 #include <sys/zfs_znode.h>
62 #include <sys/dmu_tx.h>
63 #include <sys/dmu_objset.h>
64 #include <sys/dsl_prop.h>
65 #include <sys/sa_impl.h>
68 struct lu_context_key osd_key;
70 /* Slab for OSD object allocation */
71 struct kmem_cache *osd_object_kmem;
73 /* Slab to allocate osd_zap_it */
74 struct kmem_cache *osd_zapit_cachep;
76 static struct lu_kmem_descr osd_caches[] = {
78 .ckd_cache = &osd_object_kmem,
79 .ckd_name = "zfs_osd_obj",
80 .ckd_size = sizeof(struct osd_object)
83 .ckd_cache = &osd_zapit_cachep,
84 .ckd_name = "osd_zapit_cache",
85 .ckd_size = sizeof(struct osd_zap_it)
92 static void arc_prune_func(int64_t bytes, void *private)
94 struct osd_device *od = private;
95 struct lu_site *site = &od->od_site;
99 rc = lu_env_init(&env, LCT_SHRINKER);
101 CERROR("%s: can't initialize shrinker env: rc = %d\n",
106 lu_site_purge(&env, site, (bytes >> 10));
112 * Concurrency: doesn't access mutable data
114 static int osd_root_get(const struct lu_env *env,
115 struct dt_device *dev, struct lu_fid *f)
117 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
122 * OSD object methods.
126 * Concurrency: shouldn't matter.
128 static void osd_trans_commit_cb(void *cb_data, int error)
130 struct osd_thandle *oh = cb_data;
131 struct thandle *th = &oh->ot_super;
132 struct osd_device *osd = osd_dt_dev(th->th_dev);
133 struct lu_device *lud = &th->th_dev->dd_lu_dev;
134 struct dt_txn_commit_cb *dcb, *tmp;
139 if (error == ECANCELED)
140 CWARN("%s: transaction @0x%p was aborted\n",
141 osd_dt_dev(th->th_dev)->od_svname, th);
143 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
144 osd_dt_dev(th->th_dev)->od_svname, th, error);
147 dt_txn_hook_commit(th);
149 /* call per-transaction callbacks if any */
150 list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) {
151 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
152 "commit callback entry: magic=%x name='%s'\n",
153 dcb->dcb_magic, dcb->dcb_name);
154 list_del_init(&dcb->dcb_linkage);
155 dcb->dcb_func(NULL, th, dcb, error);
158 /* Unlike ldiskfs, zfs updates space accounting at commit time.
159 * As a consequence, op_end is called only now to inform the quota slave
160 * component that reserved quota space is now accounted in usage and
161 * should be released. Quota space won't be adjusted at this point since
162 * we can't provide a suitable environment. It will be performed
163 * asynchronously by a lquota thread. */
164 qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);
168 lu_context_exit(&th->th_ctx);
169 lu_context_fini(&th->th_ctx);
175 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
177 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
180 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
181 LASSERT(&dcb->dcb_func != NULL);
182 if (dcb->dcb_flags & DCB_TRANS_STOP)
183 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
185 list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
191 * Concurrency: shouldn't matter.
193 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
196 struct osd_thandle *oh;
200 oh = container_of0(th, struct osd_thandle, ot_super);
204 rc = dt_txn_hook_start(env, d, th);
208 if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
209 /* Unlike ldiskfs, ZFS checks for available space and returns
210 * -ENOSPC when assigning txg */
213 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
214 if (unlikely(rc != 0)) {
215 struct osd_device *osd = osd_dt_dev(d);
216 /* dmu will call commit callback with error code during abort */
217 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
218 CERROR("%s: failed to start transaction due to ENOSPC"
219 "\n", osd->od_svname);
221 CERROR("%s: can't assign tx: rc = %d\n",
224 /* add commit callback */
225 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
227 lu_context_init(&th->th_ctx, th->th_tags);
228 lu_context_enter(&th->th_ctx);
229 lu_device_get(&d->dd_lu_dev);
235 static int osd_unlinked_object_free(struct osd_device *osd, uint64_t oid);
237 static void osd_unlinked_list_emptify(struct osd_device *osd,
238 struct list_head *list, bool free)
240 struct osd_object *obj;
243 while (!list_empty(list)) {
244 obj = list_entry(list->next,
245 struct osd_object, oo_unlinked_linkage);
246 LASSERT(obj->oo_dn != NULL);
247 oid = obj->oo_dn->dn_object;
249 list_del_init(&obj->oo_unlinked_linkage);
251 (void)osd_unlinked_object_free(osd, oid);
255 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
257 struct dt_txn_commit_cb *dcb;
258 struct dt_txn_commit_cb *tmp;
260 /* call per-transaction stop callbacks if any */
261 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
263 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
264 "commit callback entry: magic=%x name='%s'\n",
265 dcb->dcb_magic, dcb->dcb_name);
266 list_del_init(&dcb->dcb_linkage);
267 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
272 * Concurrency: shouldn't matter.
274 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
277 struct osd_device *osd = osd_dt_dev(th->th_dev);
278 bool sync = (th->th_sync != 0);
279 struct osd_thandle *oh;
280 struct list_head unlinked;
285 oh = container_of0(th, struct osd_thandle, ot_super);
286 INIT_LIST_HEAD(&unlinked);
287 list_splice_init(&oh->ot_unlinked_list, &unlinked);
288 /* reset OI cache for safety */
289 osd_oti_get(env)->oti_ins_cache_used = 0;
291 if (oh->ot_assigned == 0) {
293 dmu_tx_abort(oh->ot_tx);
294 osd_object_sa_dirty_rele(oh);
295 osd_unlinked_list_emptify(osd, &unlinked, false);
296 /* there won't be any commit, release reserved quota space now,
298 qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
303 rc = dt_txn_hook_stop(env, th);
305 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
308 osd_trans_stop_cb(oh, rc);
311 txg = oh->ot_tx->tx_txg;
313 osd_object_sa_dirty_rele(oh);
314 /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
315 * by osd_trans_commit_cb already. */
316 dmu_tx_commit(oh->ot_tx);
318 osd_unlinked_list_emptify(osd, &unlinked, true);
321 txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
326 static struct thandle *osd_trans_create(const struct lu_env *env,
327 struct dt_device *dt)
329 struct osd_device *osd = osd_dt_dev(dt);
330 struct osd_thandle *oh;
336 CERROR("%s: someone try to start transaction under "
337 "readonly mode, should be disabled.\n",
338 osd_name(osd_dt_dev(dt)));
340 RETURN(ERR_PTR(-EROFS));
343 tx = dmu_tx_create(osd->od_os);
345 RETURN(ERR_PTR(-ENOMEM));
347 /* alloc callback data */
351 RETURN(ERR_PTR(-ENOMEM));
355 INIT_LIST_HEAD(&oh->ot_dcb_list);
356 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
357 INIT_LIST_HEAD(&oh->ot_unlinked_list);
358 INIT_LIST_HEAD(&oh->ot_sa_list);
359 sema_init(&oh->ot_sa_lock, 1);
360 memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
364 th->th_tags = LCT_TX_HANDLE;
368 /* Estimate the total number of objects from a number of blocks */
369 uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs,
370 uint64_t nrblocks, uint64_t est_maxblockshift)
372 uint64_t est_totobjs, est_usedblocks, est_usedobjs;
375 * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then
376 * bump the free dnode estimate to assume blocks at least 64KB in
377 * case of a directory-heavy MDT (at 32KB/directory).
379 if (est_maxblockshift < 16) {
380 nrblocks >>= (16 - est_maxblockshift);
381 est_maxblockshift = 16;
385 * Estimate the total number of dnodes from the total blocks count
386 * and the space used per dnode. Since we don't know the overhead
387 * associated with each dnode (xattrs, SAs, VDEV overhead, etc.)
388 * just using DNODE_SHIFT isn't going to give a good estimate.
389 * Instead, compute the current average space usage per dnode, with
390 * an upper and lower cap to avoid unrealistic estimates..
392 * In case there aren't many dnodes or blocks used yet, add a small
393 * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor
394 * gradually disappears as the number of real dnodes grows. It also
395 * avoids the need to check for divide-by-zero computing dn_per_block.
397 CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
398 CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
400 est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
401 usedbytes) >> est_maxblockshift;
402 est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
404 if (est_usedobjs <= est_usedblocks) {
406 * Average space/dnode more than maximum block size, use max
407 * block size to estimate free dnodes from adjusted free blocks
408 * count. OSTs typically use multiple blocks per dnode so this
411 est_totobjs = nrblocks;
413 } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) {
415 * Average space/dnode smaller than min dnode size (probably
416 * due to metadnode compression), use min dnode size to
417 * estimate object count. MDTs may use only one block per node
418 * so this case applies.
420 est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;
424 * Between the extremes, use average space per existing dnode
425 * to compute the number of dnodes that will fit into nrblocks:
427 * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks)
429 * this may overflow 64 bits or become 0 if not handled well.
431 * We know nrblocks is below 2^(64 - blkbits) bits, and
432 * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT,
433 * which means that multiplying them may get as large as
434 * 2 ^ 96 for the minimum blocksize of 64KB allowed above.
436 * The ratio of dnodes per block (est_usedobjs / est_usedblocks)
437 * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to
438 * the limit checks above, so we can safely compute this first.
439 * We care more about accuracy on the MDT (many dnodes/block)
440 * which is good because this is where truncation errors are
441 * smallest. Since both nrblocks and dn_per_block are a
442 * function of blkbits, their product is at most:
444 * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9)
446 * so we can safely use 7 bits to compute a fixed-point
447 * fraction and est_totobjs can still fit in 64 bits.
449 unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks;
451 est_totobjs = (nrblocks * dn_per_block) >> 7;
456 static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
458 struct objset *os = osd->od_os;
459 uint64_t usedbytes, availbytes, usedobjs, availobjs;
460 uint64_t est_availobjs;
464 dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs);
466 memset(osfs, 0, sizeof(*osfs));
468 /* We're a zfs filesystem. */
469 osfs->os_type = UBERBLOCK_MAGIC;
472 * ZFS allows multiple block sizes. For statfs, Linux makes no
473 * proper distinction between bsize and frsize. For calculations
474 * of free and used blocks incorrectly uses bsize instead of frsize,
475 * but bsize is also used as the optimal blocksize. We return the
476 * largest possible block size as IO size for the optimum performance
477 * and scale the free and used blocks count appropriately.
479 osfs->os_bsize = osd->od_max_blksz;
480 bshift = fls64(osfs->os_bsize) - 1;
482 osfs->os_blocks = (usedbytes + availbytes) >> bshift;
483 osfs->os_bfree = availbytes >> bshift;
484 osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
486 /* Take replication (i.e. number of copies) into account */
487 if (os->os_copies != 0)
488 osfs->os_bavail /= os->os_copies;
491 * Reserve some space so we don't run into ENOSPC due to grants not
492 * accounting for metadata overhead in ZFS, and to avoid fragmentation.
493 * Rather than report this via os_bavail (which makes users unhappy if
494 * they can't fill the filesystem 100%), reduce os_blocks as well.
496 * Reserve 0.78% of total space, at least 16MB for small filesystems,
497 * for internal files to be created/unlinked when space is tight.
499 CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
500 reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
501 if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
502 reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
504 osfs->os_blocks -= reserved;
505 osfs->os_bfree -= min(reserved, osfs->os_bfree);
506 osfs->os_bavail -= min(reserved, osfs->os_bavail);
509 * The availobjs value returned from dmu_objset_space() is largely
510 * useless, since it reports the number of objects that might
511 * theoretically still fit into the dataset, independent of minor
512 * issues like how much space is actually available in the pool.
513 * Compute a better estimate in udmu_objs_count_estimate().
515 est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs,
516 osfs->os_bfree, bshift);
518 osfs->os_ffree = min(availobjs, est_availobjs);
519 osfs->os_files = osfs->os_ffree + usedobjs;
521 /* ZFS XXX: fill in backing dataset FSID/UUID
522 memcpy(osfs->os_fsid, .... );*/
524 osfs->os_namelen = MAXNAMELEN;
525 osfs->os_maxbytes = OBD_OBJECT_EOF;
527 if (!spa_writeable(dmu_objset_spa(os)) ||
528 osd->od_dev_set_rdonly || osd->od_prop_rdonly)
529 osfs->os_state |= OS_STATE_READONLY;
535 * Concurrency: shouldn't matter.
537 int osd_statfs(const struct lu_env *env, struct dt_device *d,
538 struct obd_statfs *osfs)
543 rc = osd_objset_statfs(osd_dt_dev(d), osfs);
544 if (unlikely(rc != 0))
547 osfs->os_bavail -= min_t(u64,
548 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
553 static int osd_blk_insert_cost(struct osd_device *osd)
555 int max_blockshift, nr_blkptrshift, bshift;
557 /* max_blockshift is the log2 of the number of blocks needed to reach
558 * the maximum filesize (that's to say 2^64) */
559 bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os));
560 max_blockshift = DN_MAX_OFFSET_SHIFT - bshift;
562 /* nr_blkptrshift is the log2 of the number of block pointers that can
563 * be stored in an indirect block */
564 CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
565 nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;
567 /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
568 * tree. We add +1 for rounding purpose.
569 * The tree depth times the indirect block size gives us the maximum
570 * cost of inserting a block in the tree */
571 return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
575 * Concurrency: doesn't access mutable data.
577 static void osd_conf_get(const struct lu_env *env,
578 const struct dt_device *dev,
579 struct dt_device_param *param)
581 struct osd_device *osd = osd_dt_dev(dev);
584 * XXX should be taken from not-yet-existing fs abstraction layer.
586 param->ddp_max_name_len = MAXNAMELEN;
587 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
588 param->ddp_symlink_max = PATH_MAX;
589 param->ddp_mount_type = LDD_MT_ZFS;
591 param->ddp_mntopts = MNTOPT_USERXATTR;
592 if (osd->od_posix_acl)
593 param->ddp_mntopts |= MNTOPT_ACL;
594 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
596 /* for maxbytes, report same value as ZPL */
597 param->ddp_maxbytes = MAX_LFS_FILESIZE;
599 /* inodes are dynamically allocated, so we report the per-inode space
600 * consumption to upper layers. This static value is not really accurate
601 * and we should use the same logic as in udmu_objset_statfs() to
602 * estimate the real size consumed by an object */
603 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
604 /* Although ZFS isn't an extent-based filesystem, the metadata overhead
605 * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
606 * not be accounted for every single new block insertion.
607 * Instead, the maximum extent size is set to the number of blocks that
608 * can fit into a single contiguous indirect block. There would be some
609 * cases where this crosses indirect blocks, but it also won't have 7
610 * new levels of indirect blocks in that case either, so it will still
611 * have enough reserved space for the extra indirect block */
612 param->ddp_max_extent_blks =
613 (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
614 param->ddp_extent_tax = osd_blk_insert_cost(osd);
618 * Concurrency: shouldn't matter.
620 static int osd_sync(const struct lu_env *env, struct dt_device *d)
623 struct osd_device *osd = osd_dt_dev(d);
625 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
626 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
627 CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
633 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
635 struct osd_device *osd = osd_dt_dev(dev);
636 tx_state_t *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
639 mutex_enter(&tx->tx_sync_lock);
640 txg = tx->tx_open_txg + 1;
641 if (tx->tx_quiesce_txg_waiting < txg) {
642 tx->tx_quiesce_txg_waiting = txg;
643 cv_broadcast(&tx->tx_quiesce_more_cv);
645 mutex_exit(&tx->tx_sync_lock);
651 * Concurrency: shouldn't matter.
653 static int osd_ro(const struct lu_env *env, struct dt_device *d)
655 struct osd_device *osd = osd_dt_dev(d);
658 CERROR("%s: *** setting device %s read-only ***\n",
659 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
660 osd->od_dev_set_rdonly = 1;
661 spa_freeze(dmu_objset_spa(osd->od_os));
666 static struct dt_device_operations osd_dt_ops = {
667 .dt_root_get = osd_root_get,
668 .dt_statfs = osd_statfs,
669 .dt_trans_create = osd_trans_create,
670 .dt_trans_start = osd_trans_start,
671 .dt_trans_stop = osd_trans_stop,
672 .dt_trans_cb_add = osd_trans_cb_add,
673 .dt_conf_get = osd_conf_get,
675 .dt_commit_async = osd_commit_async,
680 * DMU OSD device type methods
682 static int osd_type_init(struct lu_device_type *t)
684 LU_CONTEXT_KEY_INIT(&osd_key);
685 return lu_context_key_register(&osd_key);
688 static void osd_type_fini(struct lu_device_type *t)
690 lu_context_key_degister(&osd_key);
693 static void *osd_key_init(const struct lu_context *ctx,
694 struct lu_context_key *key)
696 struct osd_thread_info *info;
700 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
702 info = ERR_PTR(-ENOMEM);
706 static void osd_key_fini(const struct lu_context *ctx,
707 struct lu_context_key *key, void *data)
709 struct osd_thread_info *info = data;
710 struct osd_idmap_cache *idc = info->oti_ins_cache;
713 LASSERT(info->oti_ins_cache_size > 0);
714 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
715 info->oti_ins_cache = NULL;
716 info->oti_ins_cache_size = 0;
721 static void osd_key_exit(const struct lu_context *ctx,
722 struct lu_context_key *key, void *data)
726 struct lu_context_key osd_key = {
727 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
728 .lct_init = osd_key_init,
729 .lct_fini = osd_key_fini,
730 .lct_exit = osd_key_exit
733 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
735 if (osd->od_cl_seq == NULL)
738 seq_client_fini(osd->od_cl_seq);
739 OBD_FREE_PTR(osd->od_cl_seq);
740 osd->od_cl_seq = NULL;
743 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
747 /* shutdown quota slave instance associated with the device */
748 if (o->od_quota_slave != NULL) {
749 /* complete all in-flight callbacks */
750 osd_sync(env, &o->od_dt_dev);
751 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
752 qsd_fini(env, o->od_quota_slave);
753 o->od_quota_slave = NULL;
756 osd_fid_fini(env, o);
761 static void osd_xattr_changed_cb(void *arg, uint64_t newval)
763 struct osd_device *osd = arg;
765 osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
768 static void osd_recordsize_changed_cb(void *arg, uint64_t newval)
770 struct osd_device *osd = arg;
772 LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os)));
773 LASSERT(newval >= SPA_MINBLOCKSIZE);
774 LASSERT(ISP2(newval));
776 osd->od_max_blksz = newval;
779 static void osd_readonly_changed_cb(void *arg, uint64_t newval)
781 struct osd_device *osd = arg;
783 osd->od_prop_rdonly = !!newval;
787 * This function unregisters all registered callbacks. It's harmless to
788 * unregister callbacks that were never registered so it is used to safely
789 * unwind a partially completed call to osd_objset_register_callbacks().
791 static void osd_objset_unregister_callbacks(struct osd_device *o)
793 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
795 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
796 osd_xattr_changed_cb, o);
797 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
798 osd_recordsize_changed_cb, o);
799 (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
800 osd_readonly_changed_cb, o);
802 if (o->arc_prune_cb != NULL) {
803 arc_remove_prune_callback(o->arc_prune_cb);
804 o->arc_prune_cb = NULL;
809 * Register the required callbacks to be notified when zfs properties
810 * are modified using the 'zfs(8)' command line utility.
812 static int osd_objset_register_callbacks(struct osd_device *o)
814 struct dsl_dataset *ds = dmu_objset_ds(o->od_os);
815 dsl_pool_t *dp = dmu_objset_pool(o->od_os);
821 dsl_pool_config_enter(dp, FTAG);
822 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR),
823 osd_xattr_changed_cb, o);
827 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
828 osd_recordsize_changed_cb, o);
832 rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
833 osd_readonly_changed_cb, o);
837 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
839 dsl_pool_config_exit(dp, FTAG);
841 osd_objset_unregister_callbacks(o);
846 static int osd_objset_open(struct osd_device *o)
848 uint64_t version = ZPL_VERSION;
853 rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
854 o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
857 CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
863 /* Check ZFS version */
864 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
865 ZPL_VERSION_STR, 8, 1, &version);
867 CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
869 * We can't return ENOENT because that would mean the objset
872 GOTO(out, rc = -EIO);
875 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
876 ZFS_SA_ATTRS, 8, 1, &sa_obj);
880 rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
881 ZPL_END, &o->z_attr_table);
885 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
886 8, 1, &o->od_rootid);
888 CERROR("%s: lookup for root failed: rc = %d\n",
893 rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
894 8, 1, &o->od_unlinkedid);
896 CERROR("%s: lookup for %s failed: rc = %d\n",
897 o->od_svname, ZFS_UNLINKED_SET, rc);
901 /* Check that user/group usage tracking is supported */
902 if (!dmu_objset_userused_enabled(o->od_os) ||
903 DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
904 DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
905 CERROR("%s: Space accounting not supported by this target, "
906 "aborting\n", o->od_svname);
907 GOTO(out, rc = -ENOTSUPP);
911 if (rc != 0 && o->od_os != NULL) {
912 dmu_objset_disown(o->od_os, o);
920 osd_unlinked_object_free(struct osd_device *osd, uint64_t oid)
925 if (osd->od_dt_dev.dd_rdonly) {
926 CERROR("%s: someone try to free objects under "
927 "readonly mode, should be disabled.\n", osd_name(osd));
933 rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
935 CWARN("%s: Cannot truncate %llu: rc = %d\n",
936 osd->od_svname, oid, rc);
940 tx = dmu_tx_create(osd->od_os);
941 dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
942 dmu_tx_hold_zap(tx, osd->od_unlinkedid, FALSE, NULL);
943 rc = -dmu_tx_assign(tx, TXG_WAIT);
945 CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
946 osd->od_svname, oid, rc);
950 rc = -zap_remove_int(osd->od_os, osd->od_unlinkedid, oid, tx);
952 CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
953 osd->od_svname, oid, rc);
957 rc = -dmu_object_free(osd->od_os, oid, tx);
959 CWARN("%s: Cannot free %llu: rc = %d\n",
960 osd->od_svname, oid, rc);
975 osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
978 zap_attribute_t *za = &osd_oti_get(env)->oti_za;
980 zap_cursor_init(&zc, osd->od_os, osd->od_unlinkedid);
982 while (zap_cursor_retrieve(&zc, za) == 0) {
983 /* If cannot free the object, leave it in the unlinked set,
984 * until the OSD is mounted again when obd_unlinked_drain()
986 if (osd_unlinked_object_free(osd, za->za_first_integer) != 0)
988 zap_cursor_advance(&zc);
991 zap_cursor_fini(&zc);
994 static int osd_mount(const struct lu_env *env,
995 struct osd_device *o, struct lustre_cfg *cfg)
997 char *mntdev = lustre_cfg_string(cfg, 1);
998 char *str = lustre_cfg_string(cfg, 2);
999 char *svname = lustre_cfg_string(cfg, 4);
1005 if (o->od_os != NULL)
1008 if (mntdev == NULL || svname == NULL)
1011 rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
1012 if (rc >= sizeof(o->od_mntdev))
1015 rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
1016 if (rc >= sizeof(o->od_svname))
1019 str = strstr(str, ":");
1021 unsigned long flags;
1023 rc = kstrtoul(str + 1, 10, &flags);
1027 if (flags & LMD_FLG_DEV_RDONLY) {
1028 o->od_dt_dev.dd_rdonly = 1;
1029 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
1034 if (server_name_is_ost(o->od_svname))
1037 rc = osd_objset_open(o);
1041 o->od_xattr_in_sa = B_TRUE;
1042 o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
1044 rc = osd_objset_register_callbacks(o);
1048 rc = __osd_obj2dnode(env, o->od_os, o->od_rootid, &rootdn);
1052 o->od_root = rootdn->dn_object;
1053 osd_dnode_rele(rootdn);
1055 /* 1. initialize oi before any file create or file open */
1056 rc = osd_oi_init(env, o);
1060 rc = lu_site_init(&o->od_site, osd2lu_dev(o));
1063 o->od_site.ls_bottom_dev = osd2lu_dev(o);
1065 rc = lu_site_init_finish(&o->od_site);
1069 rc = osd_procfs_init(o, o->od_svname);
1073 /* initialize quota slave instance */
1074 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
1076 if (IS_ERR(o->od_quota_slave)) {
1077 rc = PTR_ERR(o->od_quota_slave);
1078 o->od_quota_slave = NULL;
1082 /* parse mount option "noacl", and enable ACL by default */
1083 opts = lustre_cfg_string(cfg, 3);
1084 if (opts == NULL || strstr(opts, "noacl") == NULL)
1085 o->od_posix_acl = 1;
1087 osd_unlinked_drain(env, o);
1089 if (rc && o->od_os) {
1090 dmu_objset_disown(o->od_os, o);
1097 static void osd_umount(const struct lu_env *env, struct osd_device *o)
1101 if (atomic_read(&o->od_zerocopy_alloc))
1102 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
1103 atomic_read(&o->od_zerocopy_alloc));
1104 if (atomic_read(&o->od_zerocopy_loan))
1105 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
1106 atomic_read(&o->od_zerocopy_loan));
1107 if (atomic_read(&o->od_zerocopy_pin))
1108 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
1109 atomic_read(&o->od_zerocopy_pin));
1111 if (o->od_os != NULL) {
1112 if (!o->od_dt_dev.dd_rdonly)
1113 /* force a txg sync to get all commit callbacks */
1114 txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
1116 /* close the object set */
1117 dmu_objset_disown(o->od_os, o);
1125 static int osd_device_init0(const struct lu_env *env,
1126 struct osd_device *o,
1127 struct lustre_cfg *cfg)
1129 struct lu_device *l = osd2lu_dev(o);
1132 /* if the module was re-loaded, env can loose its keys */
1133 rc = lu_env_refill((struct lu_env *) env);
1137 l->ld_ops = &osd_lu_ops;
1138 o->od_dt_dev.dd_ops = &osd_dt_ops;
1144 static struct lu_device *osd_device_fini(const struct lu_env *env,
1145 struct lu_device *dev);
1147 static struct lu_device *osd_device_alloc(const struct lu_env *env,
1148 struct lu_device_type *type,
1149 struct lustre_cfg *cfg)
1151 struct osd_device *dev;
1152 struct osd_seq_list *osl;
1157 return ERR_PTR(-ENOMEM);
1159 osl = &dev->od_seq_list;
1160 INIT_LIST_HEAD(&osl->osl_seq_list);
1161 rwlock_init(&osl->osl_seq_list_lock);
1162 sema_init(&osl->osl_seq_init_sem, 1);
1164 rc = dt_device_init(&dev->od_dt_dev, type);
1166 rc = osd_device_init0(env, dev, cfg);
1168 rc = osd_mount(env, dev, cfg);
1170 osd_device_fini(env, osd2lu_dev(dev));
1173 dt_device_fini(&dev->od_dt_dev);
1176 if (unlikely(rc != 0))
1179 return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
1182 static struct lu_device *osd_device_free(const struct lu_env *env,
1183 struct lu_device *d)
1185 struct osd_device *o = osd_dev(d);
1188 /* XXX: make osd top device in order to release reference */
1189 d->ld_site->ls_top_dev = d;
1190 lu_site_purge(env, d->ld_site, -1);
1191 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
1192 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1193 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
1195 lu_site_fini(&o->od_site);
1196 dt_device_fini(&o->od_dt_dev);
1202 static struct lu_device *osd_device_fini(const struct lu_env *env,
1203 struct lu_device *d)
1205 struct osd_device *o = osd_dev(d);
1211 osd_objset_unregister_callbacks(o);
1212 if (!o->od_dt_dev.dd_rdonly) {
1213 osd_sync(env, lu2dt_dev(d));
1215 spa_get_dsl(dmu_objset_spa(o->od_os)));
1219 /* now with all the callbacks completed we can cleanup the remainings */
1220 osd_shutdown(env, o);
1221 osd_oi_fini(env, o);
1223 rc = osd_procfs_fini(o);
1225 CERROR("proc fini error %d\n", rc);
1226 RETURN(ERR_PTR(rc));
1235 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
1236 const char *name, struct lu_device *next)
1242 * To be removed, setup is performed by osd_device_{init,alloc} and
1243 * cleanup is performed by osd_device_{fini,free).
1245 static int osd_process_config(const struct lu_env *env,
1246 struct lu_device *d, struct lustre_cfg *cfg)
1248 struct osd_device *o = osd_dev(d);
1252 switch(cfg->lcfg_command) {
1254 rc = osd_mount(env, o, cfg);
1257 rc = osd_shutdown(env, o);
1260 LASSERT(&o->od_dt_dev);
1261 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
1262 cfg, &o->od_dt_dev);
1263 if (rc > 0 || rc == -ENOSYS)
1264 rc = class_process_proc_param(PARAM_OST,
1265 lprocfs_osd_obd_vars,
1266 cfg, &o->od_dt_dev);
1276 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
1278 struct osd_device *osd = osd_dev(d);
1282 if (osd->od_quota_slave == NULL)
1285 /* start qsd instance on recovery completion, this notifies the quota
1286 * slave code that we are about to process new requests now */
1287 rc = qsd_start(env, osd->od_quota_slave);
1292 * we use exports to track all osd users
1294 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1295 struct obd_device *obd, struct obd_uuid *cluuid,
1296 struct obd_connect_data *data, void *localdata)
1298 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1299 struct lustre_handle conn;
1303 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
1305 rc = class_connect(&conn, obd, cluuid);
1309 *exp = class_conn2export(&conn);
1311 spin_lock(&obd->obd_dev_lock);
1313 spin_unlock(&obd->obd_dev_lock);
1319 * once last export (we don't count self-export) disappeared
1320 * osd can be released
1322 static int osd_obd_disconnect(struct obd_export *exp)
1324 struct obd_device *obd = exp->exp_obd;
1325 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
1326 int rc, release = 0;
1329 /* Only disconnect the underlying layers on the final disconnect. */
1330 spin_lock(&obd->obd_dev_lock);
1332 if (osd->od_connects == 0)
1334 spin_unlock(&obd->obd_dev_lock);
1336 rc = class_disconnect(exp); /* bz 9811 */
1338 if (rc == 0 && release)
1339 class_manual_cleanup(obd);
1343 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
1345 struct seq_server_site *ss = osd_seq_site(osd);
1349 if (osd->od_is_ost || osd->od_cl_seq != NULL)
1352 if (unlikely(ss == NULL))
1355 OBD_ALLOC_PTR(osd->od_cl_seq);
1356 if (osd->od_cl_seq == NULL)
1359 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
1360 osd->od_svname, ss->ss_server_seq);
1363 OBD_FREE_PTR(osd->od_cl_seq);
1364 osd->od_cl_seq = NULL;
1370 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
1371 struct lu_device *dev)
1373 struct osd_device *osd = osd_dev(dev);
1377 if (osd->od_quota_slave != NULL) {
1378 /* set up quota slave objects */
1379 rc = qsd_prepare(env, osd->od_quota_slave);
1384 rc = osd_fid_init(env, osd);
1389 struct lu_device_operations osd_lu_ops = {
1390 .ldo_object_alloc = osd_object_alloc,
1391 .ldo_process_config = osd_process_config,
1392 .ldo_recovery_complete = osd_recovery_complete,
1393 .ldo_prepare = osd_prepare,
1396 static void osd_type_start(struct lu_device_type *t)
1400 static void osd_type_stop(struct lu_device_type *t)
1404 int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1405 struct lu_fid *fid, struct md_op_data *op_data)
1407 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
1409 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
1412 static struct lu_device_type_operations osd_device_type_ops = {
1413 .ldto_init = osd_type_init,
1414 .ldto_fini = osd_type_fini,
1416 .ldto_start = osd_type_start,
1417 .ldto_stop = osd_type_stop,
1419 .ldto_device_alloc = osd_device_alloc,
1420 .ldto_device_free = osd_device_free,
1422 .ldto_device_init = osd_device_init,
1423 .ldto_device_fini = osd_device_fini
1426 static struct lu_device_type osd_device_type = {
1427 .ldt_tags = LU_DEVICE_DT,
1428 .ldt_name = LUSTRE_OSD_ZFS_NAME,
1429 .ldt_ops = &osd_device_type_ops,
1430 .ldt_ctx_tags = LCT_LOCAL
1434 static struct obd_ops osd_obd_device_ops = {
1435 .o_owner = THIS_MODULE,
1436 .o_connect = osd_obd_connect,
1437 .o_disconnect = osd_obd_disconnect,
1438 .o_fid_alloc = osd_fid_alloc
1441 static int __init osd_init(void)
1445 rc = osd_options_init();
1449 rc = lu_kmem_init(osd_caches);
1453 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
1454 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
1456 lu_kmem_fini(osd_caches);
1460 static void __exit osd_exit(void)
1462 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
1463 lu_kmem_fini(osd_caches);
1466 extern unsigned int osd_oi_count;
1467 module_param(osd_oi_count, int, 0444);
1468 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
1470 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1471 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
1472 MODULE_VERSION(LUSTRE_VERSION_STRING);
1473 MODULE_LICENSE("GPL");
1475 module_init(osd_init);
1476 module_exit(osd_exit);