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.
31 * Copyright (c) 2011, 2012 Whamcloud, Inc.
32 * Use is subject to license terms.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/osd-zfs/osd_handler.c
40 * Top-level entry points into osd module
42 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
43 * Author: Mike Pershin <tappro@whamcloud.com>
44 * Author: Johann Lombardi <johann@whamcloud.com>
48 # define EXPORT_SYMTAB
50 #define DEBUG_SUBSYSTEM S_OSD
52 #include <lustre_ver.h>
53 #include <libcfs/libcfs.h>
54 #include <obd_support.h>
55 #include <lustre_net.h>
57 #include <obd_class.h>
58 #include <lustre_disk.h>
59 #include <lustre_fid.h>
61 #include "osd_internal.h"
63 #include <sys/dnode.h>
68 #include <sys/spa_impl.h>
69 #include <sys/zfs_znode.h>
70 #include <sys/dmu_tx.h>
71 #include <sys/dmu_objset.h>
72 #include <sys/dsl_prop.h>
73 #include <sys/sa_impl.h>
76 struct lu_context_key osd_key;
78 static char *root_tag = "osd_mount, rootdb";
80 /* Slab for OSD object allocation */
81 cfs_mem_cache_t *osd_object_kmem;
83 static struct lu_kmem_descr osd_caches[] = {
85 .ckd_cache = &osd_object_kmem,
86 .ckd_name = "zfs_osd_obj",
87 .ckd_size = sizeof(struct osd_object)
94 static void arc_prune_func(int64_t bytes, void *private)
96 struct osd_device *od = private;
97 struct lu_site *site = &od->od_site;
101 rc = lu_env_init(&env, LCT_SHRINKER);
103 CERROR("%s: can't initialize shrinker env: rc = %d\n",
108 lu_site_purge(&env, site, (bytes >> 10));
114 * Concurrency: doesn't access mutable data
116 static int osd_root_get(const struct lu_env *env,
117 struct dt_device *dev, struct lu_fid *f)
119 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
124 * OSD object methods.
128 * Concurrency: shouldn't matter.
130 static void osd_trans_commit_cb(void *cb_data, int error)
132 struct osd_thandle *oh = cb_data;
133 struct thandle *th = &oh->ot_super;
134 struct lu_device *lud = &th->th_dev->dd_lu_dev;
135 struct dt_txn_commit_cb *dcb, *tmp;
140 if (error == ECANCELED)
141 CWARN("%s: transaction @0x%p was aborted\n",
142 osd_dt_dev(th->th_dev)->od_svname, th);
144 CERROR("%s: transaction @0x%p commit error: rc = %d\n",
145 osd_dt_dev(th->th_dev)->od_svname, th, error);
148 dt_txn_hook_commit(th);
150 /* call per-transaction callbacks if any */
151 cfs_list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage)
152 dcb->dcb_func(NULL, th, dcb, error);
156 lu_context_exit(&th->th_ctx);
157 lu_context_fini(&th->th_ctx);
163 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
165 struct osd_thandle *oh;
167 oh = container_of0(th, struct osd_thandle, ot_super);
168 cfs_list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
174 * Concurrency: shouldn't matter.
176 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
179 struct osd_thandle *oh;
183 oh = container_of0(th, struct osd_thandle, ot_super);
187 rc = dt_txn_hook_start(env, d, th);
191 if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
192 /* Unlike ldiskfs, ZFS checks for available space and returns
193 * -ENOSPC when assigning txg */
196 rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
197 if (unlikely(rc != 0)) {
198 struct osd_device *osd = osd_dt_dev(d);
199 /* dmu will call commit callback with error code during abort */
200 if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
201 CERROR("%s: failed to start transaction due to ENOSPC. "
202 "Metadata overhead is underestimated or "
203 "grant_ratio is too low.\n",
204 osd->od_dt_dev.dd_lu_dev.ld_obd->obd_name);
206 CERROR("%s: can't assign tx: rc = %d\n",
207 osd->od_dt_dev.dd_lu_dev.ld_obd->obd_name, rc);
209 /* add commit callback */
210 dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
212 lu_context_init(&th->th_ctx, th->th_tags);
213 lu_context_enter(&th->th_ctx);
214 lu_device_get(&d->dd_lu_dev);
221 * Concurrency: shouldn't matter.
223 static int osd_trans_stop(const struct lu_env *env, struct thandle *th)
225 struct osd_device *osd = osd_dt_dev(th->th_dev);
226 struct osd_thandle *oh;
231 oh = container_of0(th, struct osd_thandle, ot_super);
233 if (oh->ot_assigned == 0) {
235 dmu_tx_abort(oh->ot_tx);
236 osd_object_sa_dirty_rele(oh);
241 rc = dt_txn_hook_stop(env, th);
243 CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
247 txg = oh->ot_tx->tx_txg;
249 osd_object_sa_dirty_rele(oh);
250 dmu_tx_commit(oh->ot_tx);
253 txg_wait_synced(dmu_objset_pool(osd->od_objset.os), txg);
258 static struct thandle *osd_trans_create(const struct lu_env *env,
259 struct dt_device *dt)
261 struct osd_device *osd = osd_dt_dev(dt);
262 struct osd_thandle *oh;
267 tx = dmu_tx_create(osd->od_objset.os);
269 RETURN(ERR_PTR(-ENOMEM));
271 /* alloc callback data */
275 RETURN(ERR_PTR(-ENOMEM));
279 CFS_INIT_LIST_HEAD(&oh->ot_dcb_list);
280 CFS_INIT_LIST_HEAD(&oh->ot_sa_list);
281 cfs_sema_init(&oh->ot_sa_lock, 1);
285 th->th_tags = LCT_TX_HANDLE;
290 * Concurrency: shouldn't matter.
292 int osd_statfs(const struct lu_env *env, struct dt_device *d,
293 struct obd_statfs *osfs)
295 struct osd_device *osd = osd_dt_dev(d);
299 rc = udmu_objset_statfs(&osd->od_objset, osfs);
302 osfs->os_bavail -= min_t(obd_size,
303 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
309 * Concurrency: doesn't access mutable data.
311 static void osd_conf_get(const struct lu_env *env,
312 const struct dt_device *dev,
313 struct dt_device_param *param)
316 * XXX should be taken from not-yet-existing fs abstraction layer.
318 param->ddp_max_name_len = MAXNAMELEN;
319 param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
320 param->ddp_block_shift = 12; /* XXX */
321 param->ddp_mount_type = LDD_MT_ZFS;
323 param->ddp_mntopts = MNTOPT_USERXATTR | MNTOPT_ACL;
324 param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE;
326 /* for maxbytes, report same value as ZPL */
327 param->ddp_maxbytes = MAX_LFS_FILESIZE;
329 /* Default reserved fraction of the available space that should be kept
330 * for error margin. Unfortunately, there are many factors that can
331 * impact the overhead with zfs, so let's be very cautious for now and
332 * reserve 20% of the available space which is not given out as grant.
333 * This tunable can be changed on a live system via procfs if needed. */
334 param->ddp_grant_reserved = 20;
336 /* inodes are dynamically allocated, so we report the per-inode space
337 * consumption to upper layers. This static value is not really accurate
338 * and we should use the same logic as in udmu_objset_statfs() to
339 * estimate the real size consumed by an object */
340 param->ddp_inodespace = OSD_DNODE_EST_COUNT;
341 /* per-fragment overhead to be used by the client code */
342 param->ddp_grant_frag = udmu_blk_insert_cost();
346 * Concurrency: shouldn't matter.
348 static int osd_sync(const struct lu_env *env, struct dt_device *d)
350 struct osd_device *osd = osd_dt_dev(d);
351 CDEBUG(D_HA, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
352 txg_wait_synced(dmu_objset_pool(osd->od_objset.os), 0ULL);
356 static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
358 struct osd_device *osd = osd_dt_dev(dev);
359 tx_state_t *tx = &dmu_objset_pool(osd->od_objset.os)->dp_tx;
362 txg = tx->tx_open_txg + 1;
363 if (tx->tx_quiesce_txg_waiting < txg) {
364 tx->tx_quiesce_txg_waiting = txg;
365 cv_broadcast(&tx->tx_quiesce_more_cv);
367 mutex_exit(&tx->tx_sync_lock);
373 * Concurrency: shouldn't matter.
375 static int osd_ro(const struct lu_env *env, struct dt_device *d)
377 struct osd_device *osd = osd_dt_dev(d);
380 CERROR("%s: *** setting device %s read-only ***\n",
381 osd->od_svname, LUSTRE_OSD_ZFS_NAME);
383 spa_freeze(dmu_objset_spa(osd->od_objset.os));
389 * Concurrency: serialization provided by callers.
391 static int osd_init_capa_ctxt(const struct lu_env *env, struct dt_device *d,
392 int mode, unsigned long timeout, __u32 alg,
393 struct lustre_capa_key *keys)
395 struct osd_device *dev = osd_dt_dev(d);
398 dev->od_fl_capa = mode;
399 dev->od_capa_timeout = timeout;
400 dev->od_capa_alg = alg;
401 dev->od_capa_keys = keys;
406 static char *osd_label_get(const struct lu_env *env, const struct dt_device *d)
408 struct osd_device *dev = osd_dt_dev(d);
412 rc = -udmu_userprop_get_str(&dev->od_objset, DMU_OSD_SVNAME,
413 dev->od_svname, sizeof(dev->od_svname));
415 if (rc == -EOVERFLOW)
416 CWARN("%s: buffer too small\n", dev->od_svname);
420 RETURN(&dev->od_svname[0]);
423 static struct dt_device_operations osd_dt_ops = {
424 .dt_root_get = osd_root_get,
425 .dt_statfs = osd_statfs,
426 .dt_trans_create = osd_trans_create,
427 .dt_trans_start = osd_trans_start,
428 .dt_trans_stop = osd_trans_stop,
429 .dt_trans_cb_add = osd_trans_cb_add,
430 .dt_conf_get = osd_conf_get,
432 .dt_commit_async = osd_commit_async,
434 .dt_init_capa_ctxt = osd_init_capa_ctxt,
438 * DMU OSD device type methods
440 static int osd_type_init(struct lu_device_type *t)
442 LU_CONTEXT_KEY_INIT(&osd_key);
443 return lu_context_key_register(&osd_key);
446 static void osd_type_fini(struct lu_device_type *t)
448 lu_context_key_degister(&osd_key);
451 static void *osd_key_init(const struct lu_context *ctx,
452 struct lu_context_key *key)
454 struct osd_thread_info *info;
458 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
460 info = ERR_PTR(-ENOMEM);
464 static void osd_key_fini(const struct lu_context *ctx,
465 struct lu_context_key *key, void *data)
467 struct osd_thread_info *info = data;
472 static void osd_key_exit(const struct lu_context *ctx,
473 struct lu_context_key *key, void *data)
475 struct osd_thread_info *info = data;
477 memset(info, 0, sizeof(*info));
480 struct lu_context_key osd_key = {
481 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
482 .lct_init = osd_key_init,
483 .lct_fini = osd_key_fini,
484 .lct_exit = osd_key_exit
487 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
492 static int osd_mount(const struct lu_env *env,
493 struct osd_device *o, struct lustre_cfg *cfg)
495 char *dev = lustre_cfg_string(cfg, 0);
496 struct lustre_mount_info *lmi;
497 struct lustre_sb_info *lsi;
503 if (o->od_objset.os != NULL)
506 lmi = server_get_mount(dev);
508 CERROR("Unknown mount point: '%s'\n", dev);
512 lsi = s2lsi(lmi->lmi_sb);
513 dev = lsi->lsi_lmd->lmd_dev;
515 if (strlen(dev) >= sizeof(o->od_mntdev))
518 strcpy(o->od_mntdev, dev);
520 rc = -udmu_objset_open(o->od_mntdev + 1, &o->od_objset);
522 CERROR("can't open objset %s: %d\n", o->od_mntdev, rc);
526 rc = __osd_obj2dbuf(env, o->od_objset.os, o->od_objset.root,
529 CERROR("udmu_obj2dbuf() failed with error %d\n", rc);
530 udmu_objset_close(&o->od_objset);
534 o->od_root = rootdb->db_object;
535 sa_buf_rele(rootdb, root_tag);
537 /* 1. initialize oi before any file create or file open */
538 rc = osd_oi_init(env, o);
542 label = osd_label_get(env, &o->od_dt_dev);
544 GOTO(err, rc = -ENODEV);
546 /* Use our own ZAP for inode accounting by default, this can be changed
547 * via procfs to estimate the inode usage from the block usage */
548 o->od_quota_iused_est = 0;
550 rc = osd_procfs_init(o, label);
554 o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
560 static void osd_umount(const struct lu_env *env, struct osd_device *o)
564 if (cfs_atomic_read(&o->od_zerocopy_alloc))
565 CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
566 cfs_atomic_read(&o->od_zerocopy_alloc));
567 if (cfs_atomic_read(&o->od_zerocopy_loan))
568 CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
569 cfs_atomic_read(&o->od_zerocopy_loan));
570 if (cfs_atomic_read(&o->od_zerocopy_pin))
571 CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
572 cfs_atomic_read(&o->od_zerocopy_pin));
574 if (o->od_objset.os != NULL)
575 udmu_objset_close(&o->od_objset);
580 static int osd_device_init0(const struct lu_env *env,
581 struct osd_device *o,
582 struct lustre_cfg *cfg)
584 struct lu_device *l = osd2lu_dev(o);
587 /* if the module was re-loaded, env can loose its keys */
588 rc = lu_env_refill((struct lu_env *) env);
592 l->ld_ops = &osd_lu_ops;
593 o->od_dt_dev.dd_ops = &osd_dt_ops;
595 o->od_capa_hash = init_capa_hash();
596 if (o->od_capa_hash == NULL)
597 GOTO(out, rc = -ENOMEM);
603 static struct lu_device *osd_device_alloc(const struct lu_env *env,
604 struct lu_device_type *t,
605 struct lustre_cfg *cfg)
607 struct osd_device *o;
612 return ERR_PTR(-ENOMEM);
614 rc = dt_device_init(&o->od_dt_dev, t);
616 rc = osd_device_init0(env, o, cfg);
618 dt_device_fini(&o->od_dt_dev);
621 if (unlikely(rc != 0))
624 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
627 static struct lu_device *osd_device_free(const struct lu_env *env,
630 struct osd_device *o = osd_dev(d);
633 cleanup_capa_hash(o->od_capa_hash);
634 /* XXX: make osd top device in order to release reference */
635 /*d->ld_site->ls_top_dev = d;
636 lu_site_purge(env, d->ld_site, -1);
637 lu_site_fini(&o->od_site);*/
638 dt_device_fini(&o->od_dt_dev);
644 static struct lu_device *osd_device_fini(const struct lu_env *env,
647 struct osd_device *o = osd_dev(d);
648 struct lustre_mount_info *lmi;
655 if (o->od_objset.os) {
656 arc_remove_prune_callback(o->arc_prune_cb);
657 o->arc_prune_cb = NULL;
658 osd_sync(env, lu2dt_dev(d));
659 txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_objset.os)));
662 rc = osd_procfs_fini(o);
664 CERROR("proc fini error %d\n", rc);
671 lmi = server_get_mount_2(o->od_svname);
673 server_put_mount(lmi->lmi_name, lmi->lmi_mnt);
678 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
679 const char *name, struct lu_device *next)
685 * To be removed, setup is performed by osd_device_{init,alloc} and
686 * cleanup is performed by osd_device_{fini,free).
688 static int osd_process_config(const struct lu_env *env,
689 struct lu_device *d, struct lustre_cfg *cfg)
691 struct osd_device *o = osd_dev(d);
695 switch(cfg->lcfg_command) {
697 err = osd_mount(env, o, cfg);
700 err = osd_shutdown(env, o);
709 static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
715 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
716 struct lu_device *dev)
721 struct lu_device_operations osd_lu_ops = {
722 .ldo_object_alloc = osd_object_alloc,
723 .ldo_process_config = osd_process_config,
724 .ldo_recovery_complete = osd_recovery_complete,
725 .ldo_prepare = osd_prepare,
728 static void osd_type_start(struct lu_device_type *t)
732 static void osd_type_stop(struct lu_device_type *t)
736 static struct lu_device_type_operations osd_device_type_ops = {
737 .ldto_init = osd_type_init,
738 .ldto_fini = osd_type_fini,
740 .ldto_start = osd_type_start,
741 .ldto_stop = osd_type_stop,
743 .ldto_device_alloc = osd_device_alloc,
744 .ldto_device_free = osd_device_free,
746 .ldto_device_init = osd_device_init,
747 .ldto_device_fini = osd_device_fini
750 static struct lu_device_type osd_device_type = {
751 .ldt_tags = LU_DEVICE_DT,
752 .ldt_name = LUSTRE_OSD_ZFS_NAME,
753 .ldt_ops = &osd_device_type_ops,
754 .ldt_ctx_tags = LCT_LOCAL
758 static struct obd_ops osd_obd_device_ops = {
759 .o_owner = THIS_MODULE,
762 int __init osd_init(void)
766 rc = osd_options_init();
770 rc = lu_kmem_init(osd_caches);
774 rc = class_register_type(&osd_obd_device_ops, NULL,
775 lprocfs_osd_module_vars,
776 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
778 lu_kmem_fini(osd_caches);
782 void __exit osd_exit(void)
784 class_unregister_type(LUSTRE_OSD_ZFS_NAME);
785 lu_kmem_fini(osd_caches);
788 extern unsigned int osd_oi_count;
789 CFS_MODULE_PARM(osd_oi_count, "i", int, 0444,
790 "Number of Object Index containers to be created, "
791 "it's only valid for new filesystem.");
793 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
794 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
795 MODULE_LICENSE("GPL");
797 cfs_module(osd, LUSTRE_VERSION_STRING, osd_init, osd_exit);