* Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2014, Intel Corporation.
+ * Copyright (c) 2012, 2015, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
{
- struct osd_thandle *oh;
+ struct osd_thandle *oh = container_of0(th, struct osd_thandle,
+ ot_super);
- oh = container_of0(th, struct osd_thandle, ot_super);
- list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
+ LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
+ LASSERT(&dcb->dcb_func != NULL);
+ if (dcb->dcb_flags & DCB_TRANS_STOP)
+ list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
+ else
+ list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);
return 0;
}
struct osd_device *osd = osd_dt_dev(d);
/* dmu will call commit callback with error code during abort */
if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
- CERROR("%s: failed to start transaction due to ENOSPC. "
- "Metadata overhead is underestimated or "
- "grant_ratio is too low.\n", osd->od_svname);
+ CERROR("%s: failed to start transaction due to ENOSPC"
+ "\n", osd->od_svname);
else
CERROR("%s: can't assign tx: rc = %d\n",
osd->od_svname, rc);
RETURN(rc);
}
+static int osd_unlinked_object_free(struct osd_device *osd, uint64_t oid);
+
+static void osd_unlinked_list_emptify(struct osd_device *osd,
+ struct list_head *list, bool free)
+{
+ struct osd_object *obj;
+ uint64_t oid;
+
+ while (!list_empty(list)) {
+ obj = list_entry(list->next,
+ struct osd_object, oo_unlinked_linkage);
+ LASSERT(obj->oo_db != NULL);
+ oid = obj->oo_db->db_object;
+
+ list_del_init(&obj->oo_unlinked_linkage);
+ if (free)
+ (void)osd_unlinked_object_free(osd, oid);
+ }
+}
+
+static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
+{
+ struct dt_txn_commit_cb *dcb;
+ struct dt_txn_commit_cb *tmp;
+
+ /* call per-transaction stop callbacks if any */
+ list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
+ dcb_linkage) {
+ LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
+ "commit callback entry: magic=%x name='%s'\n",
+ dcb->dcb_magic, dcb->dcb_name);
+ list_del_init(&dcb->dcb_linkage);
+ dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
+ }
+}
+
/*
* Concurrency: shouldn't matter.
*/
struct thandle *th)
{
struct osd_device *osd = osd_dt_dev(th->th_dev);
+ bool sync = (th->th_sync != 0);
struct osd_thandle *oh;
+ struct list_head unlinked;
uint64_t txg;
int rc;
ENTRY;
oh = container_of0(th, struct osd_thandle, ot_super);
+ INIT_LIST_HEAD(&unlinked);
+ list_splice_init(&oh->ot_unlinked_list, &unlinked);
if (oh->ot_assigned == 0) {
LASSERT(oh->ot_tx);
dmu_tx_abort(oh->ot_tx);
osd_object_sa_dirty_rele(oh);
+ osd_unlinked_list_emptify(osd, &unlinked, false);
/* there won't be any commit, release reserved quota space now,
* if any */
qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
osd->od_svname, rc);
+ osd_trans_stop_cb(oh, rc);
+
LASSERT(oh->ot_tx);
txg = oh->ot_tx->tx_txg;
osd_object_sa_dirty_rele(oh);
+ /* XXX: Once dmu_tx_commit() called, oh/th could have been freed
+ * by osd_trans_commit_cb already. */
dmu_tx_commit(oh->ot_tx);
- if (th->th_sync)
+ osd_unlinked_list_emptify(osd, &unlinked, true);
+
+ if (sync)
txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
RETURN(rc);
oh->ot_tx = tx;
INIT_LIST_HEAD(&oh->ot_dcb_list);
+ INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
+ INIT_LIST_HEAD(&oh->ot_unlinked_list);
INIT_LIST_HEAD(&oh->ot_sa_list);
sema_init(&oh->ot_sa_lock, 1);
memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
dmu_objset_space(os, &refdbytes, &availbytes, &usedobjs, &availobjs);
+ memset(osfs, 0, sizeof(*osfs));
+
+ /* We're a zfs filesystem. */
+ osfs->os_type = UBERBLOCK_MAGIC;
+
/*
* ZFS allows multiple block sizes. For statfs, Linux makes no
* proper distinction between bsize and frsize. For calculations
* Rather than report this via os_bavail (which makes users unhappy if
* they can't fill the filesystem 100%), reduce os_blocks as well.
*
- * Reserve 0.78% of total space, at least 4MB for small filesystems,
+ * Reserve 0.78% of total space, at least 16MB for small filesystems,
* for internal files to be created/unlinked when space is tight.
*/
CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
- if (likely(osfs->os_blocks >= OSD_STATFS_RESERVED_SIZE))
+ reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
+ if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
- else
- reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
osfs->os_blocks -= reserved;
- osfs->os_bfree -= MIN(reserved, osfs->os_bfree);
- osfs->os_bavail -= MIN(reserved, osfs->os_bavail);
+ osfs->os_bfree -= min(reserved, osfs->os_bfree);
+ osfs->os_bavail -= min(reserved, osfs->os_bavail);
/*
* The availobjs value returned from dmu_objset_space() is largely
/* ZFS XXX: fill in backing dataset FSID/UUID
memcpy(osfs->os_fsid, .... );*/
- /* We're a zfs filesystem. */
- osfs->os_type = UBERBLOCK_MAGIC;
-
- /* ZFS XXX: fill in appropriate OS_STATE_{DEGRADED,READONLY} flags
- osfs->os_state = vf_to_stf(vfsp->vfs_flag);
- if (sb->s_flags & MS_RDONLY)
- osfs->os_state = OS_STATE_READONLY;
- */
-
osfs->os_namelen = MAXNAMELEN;
osfs->os_maxbytes = OBD_OBJECT_EOF;
+ if (!spa_writeable(dmu_objset_spa(os)) ||
+ osd->od_dev_set_rdonly || osd->od_prop_rdonly)
+ osfs->os_state |= OS_STATE_READONLY;
+
return 0;
}
*/
param->ddp_max_name_len = MAXNAMELEN;
param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */
- param->ddp_block_shift = 12; /* XXX */
+ param->ddp_symlink_max = PATH_MAX;
param->ddp_mount_type = LDD_MT_ZFS;
param->ddp_mntopts = MNTOPT_USERXATTR;
/* for maxbytes, report same value as ZPL */
param->ddp_maxbytes = MAX_LFS_FILESIZE;
- /* Default reserved fraction of the available space that should be kept
- * for error margin. Unfortunately, there are many factors that can
- * impact the overhead with zfs, so let's be very cautious for now and
- * reserve 20% of the available space which is not given out as grant.
- * This tunable can be changed on a live system via procfs if needed. */
- param->ddp_grant_reserved = 20;
-
/* inodes are dynamically allocated, so we report the per-inode space
* consumption to upper layers. This static value is not really accurate
* and we should use the same logic as in udmu_objset_statfs() to
* estimate the real size consumed by an object */
param->ddp_inodespace = OSD_DNODE_EST_COUNT;
- /* per-fragment overhead to be used by the client code */
- param->ddp_grant_frag = osd_blk_insert_cost(osd);
+ /* Although ZFS isn't an extent-based filesystem, the metadata overhead
+ * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
+ * not be accounted for every single new block insertion.
+ * Instead, the maximum extent size is set to the number of blocks that
+ * can fit into a single contiguous indirect block. There would be some
+ * cases where this crosses indirect blocks, but it also won't have 7
+ * new levels of indirect blocks in that case either, so it will still
+ * have enough reserved space for the extra indirect block */
+ param->ddp_max_extent_blks =
+ (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
+ param->ddp_extent_tax = osd_blk_insert_cost(osd);
}
/*
CERROR("%s: *** setting device %s read-only ***\n",
osd->od_svname, LUSTRE_OSD_ZFS_NAME);
- osd->od_rdonly = 1;
+ osd->od_dev_set_rdonly = 1;
spa_freeze(dmu_objset_spa(osd->od_os));
RETURN(0);
osd->od_max_blksz = newval;
}
+static void osd_readonly_changed_cb(void *arg, uint64_t newval)
+{
+ struct osd_device *osd = arg;
+
+ osd->od_prop_rdonly = !!newval;
+}
+
/*
* This function unregisters all registered callbacks. It's harmless to
* unregister callbacks that were never registered so it is used to safely
osd_xattr_changed_cb, o);
(void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
osd_recordsize_changed_cb, o);
+ (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
+ osd_readonly_changed_cb, o);
if (o->arc_prune_cb != NULL) {
arc_remove_prune_callback(o->arc_prune_cb);
if (rc)
GOTO(err, rc);
+ rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
+ osd_readonly_changed_cb, o);
+ if (rc)
+ GOTO(err, rc);
+
o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
err:
dsl_pool_config_exit(dp, FTAG);
rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os);
if (rc) {
+ CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev);
o->od_os = NULL;
goto out;
}
GOTO(out, rc);
}
+ rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
+ 8, 1, &o->od_unlinkedid);
+ if (rc) {
+ CERROR("%s: lookup for %s failed: rc = %d\n",
+ o->od_svname, ZFS_UNLINKED_SET, rc);
+ GOTO(out, rc);
+ }
+
/* Check that user/group usage tracking is supported */
if (!dmu_objset_userused_enabled(o->od_os) ||
DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
}
out:
- if (rc != 0 && o->od_os != NULL)
+ if (rc != 0 && o->od_os != NULL) {
dmu_objset_disown(o->od_os, o);
+ o->od_os = NULL;
+ }
RETURN(rc);
}
+static int
+osd_unlinked_object_free(struct osd_device *osd, uint64_t oid)
+{
+ int rc;
+ dmu_tx_t *tx;
+
+ rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
+ if (rc != 0) {
+ CWARN("%s: Cannot truncate "LPU64": rc = %d\n",
+ osd->od_svname, oid, rc);
+ return rc;
+ }
+
+ tx = dmu_tx_create(osd->od_os);
+ dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
+ dmu_tx_hold_zap(tx, osd->od_unlinkedid, FALSE, NULL);
+ rc = -dmu_tx_assign(tx, TXG_WAIT);
+ if (rc != 0) {
+ CWARN("%s: Cannot assign tx for "LPU64": rc = %d\n",
+ osd->od_svname, oid, rc);
+ goto failed;
+ }
+
+ rc = -zap_remove_int(osd->od_os, osd->od_unlinkedid, oid, tx);
+ if (rc != 0) {
+ CWARN("%s: Cannot remove "LPU64" from unlinked set: rc = %d\n",
+ osd->od_svname, oid, rc);
+ goto failed;
+ }
+
+ rc = -dmu_object_free(osd->od_os, oid, tx);
+ if (rc != 0) {
+ CWARN("%s: Cannot free "LPU64": rc = %d\n",
+ osd->od_svname, oid, rc);
+ goto failed;
+ }
+ dmu_tx_commit(tx);
+
+ return 0;
+
+failed:
+ LASSERT(rc != 0);
+ dmu_tx_abort(tx);
+
+ return rc;
+}
+
+static void
+osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd)
+{
+ zap_cursor_t zc;
+ zap_attribute_t *za = &osd_oti_get(env)->oti_za;
+
+ zap_cursor_init(&zc, osd->od_os, osd->od_unlinkedid);
+
+ while (zap_cursor_retrieve(&zc, za) == 0) {
+ /* If cannot free the object, leave it in the unlinked set,
+ * until the OSD is mounted again when obd_unlinked_drain()
+ * will be called. */
+ if (osd_unlinked_object_free(osd, za->za_first_integer) != 0)
+ break;
+ zap_cursor_advance(&zc);
+ }
+
+ zap_cursor_fini(&zc);
+}
+
static int osd_mount(const struct lu_env *env,
struct osd_device *o, struct lustre_cfg *cfg)
{
rc = osd_objset_open(o);
if (rc)
- GOTO(err, rc);
+ RETURN(rc);
o->od_xattr_in_sa = B_TRUE;
o->od_max_blksz = SPA_OLD_MAXBLOCKSIZE;
if (rc)
GOTO(err, rc);
- rc = osd_convert_root_to_new_seq(env, o);
- if (rc)
- GOTO(err, rc);
-
/* Use our own ZAP for inode accounting by default, this can be changed
* via procfs to estimate the inode usage from the block usage */
o->od_quota_iused_est = 0;
if (opts == NULL || strstr(opts, "noacl") == NULL)
o->od_posix_acl = 1;
+ osd_unlinked_drain(env, o);
err:
if (rc) {
- dmu_objset_disown(o->od_os, o);
+ if (o->od_os)
+ dmu_objset_disown(o->od_os, o);
o->od_os = NULL;
}
struct lu_device_type *type,
struct lustre_cfg *cfg)
{
- struct osd_device *dev;
- int rc;
+ struct osd_device *dev;
+ struct osd_seq_list *osl;
+ int rc;
OBD_ALLOC_PTR(dev);
if (dev == NULL)
return ERR_PTR(-ENOMEM);
+ osl = &dev->od_seq_list;
+ INIT_LIST_HEAD(&osl->osl_seq_list);
+ rwlock_init(&osl->osl_seq_list_lock);
+ sema_init(&osl->osl_seq_init_sem, 1);
+
rc = dt_device_init(&dev->od_dt_dev, type);
if (rc == 0) {
rc = osd_device_init0(env, dev, cfg);
.o_fid_alloc = osd_fid_alloc
};
-int __init osd_init(void)
+static int __init osd_init(void)
{
int rc;
return rc;
}
-void __exit osd_exit(void)
+static void __exit osd_exit(void)
{
class_unregister_type(LUSTRE_OSD_ZFS_NAME);
lu_kmem_fini(osd_caches);
}
extern unsigned int osd_oi_count;
-CFS_MODULE_PARM(osd_oi_count, "i", int, 0444,
- "Number of Object Index containers to be created, "
- "it's only valid for new filesystem.");
+module_param(osd_oi_count, int, 0444);
+MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
-MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
+MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
MODULE_VERSION(LUSTRE_VERSION_STRING);
MODULE_LICENSE("GPL");