* Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2016, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#define DEBUG_SUBSYSTEM S_OSD
-#include <lustre_ver.h>
#include <libcfs/libcfs.h>
#include <obd_support.h>
#include <lustre_net.h>
#include <obd_class.h>
#include <lustre_disk.h>
#include <lustre_fid.h>
-#include <lustre_param.h>
+#include <uapi/linux/lustre/lustre_param.h>
#include <md_object.h>
#include "osd_internal.h"
struct lu_context_key osd_key;
+static int osd_txg_sync_delay_us = -1;
+
/* Slab for OSD object allocation */
struct kmem_cache *osd_object_kmem;
struct lu_env env;
int rc;
+ LASSERT(site->ls_obj_hash);
+
rc = lu_env_init(&env, LCT_SHRINKER);
if (rc) {
CERROR("%s: can't initialize shrinker env: rc = %d\n",
dt_txn_hook_commit(th);
/* call per-transaction callbacks if any */
- list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage)
+ list_for_each_entry_safe(dcb, tmp, &oh->ot_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, th, dcb, error);
+ }
/* Unlike ldiskfs, zfs updates space accounting at commit time.
* As a consequence, op_end is called only now to inform the quota slave
lu_device_put(lud);
th->th_dev = NULL;
- lu_context_exit(&th->th_ctx);
- lu_context_fini(&th->th_ctx);
OBD_FREE_PTR(oh);
EXIT;
/* add commit callback */
dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
oh->ot_assigned = 1;
- lu_context_init(&th->th_ctx, th->th_tags);
- lu_context_enter(&th->th_ctx);
+ osd_oti_get(env)->oti_in_trans = 1;
lu_device_get(&d->dd_lu_dev);
}
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,
+static void osd_unlinked_list_emptify(const struct lu_env *env,
+ struct osd_device *osd,
struct list_head *list, bool free)
{
struct osd_object *obj;
list_del_init(&obj->oo_unlinked_linkage);
if (free)
- (void)osd_unlinked_object_free(osd, oid);
+ (void)osd_unlinked_object_free(env, osd, oid);
}
}
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);
+ osd_object_sa_dirty_rele(env, oh);
+ osd_unlinked_list_emptify(env, 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);
RETURN(0);
}
- /* When doing our own inode accounting, the ZAPs storing per-uid/gid
- * usage are updated at operation execution time, so we should call
- * qsd_op_end() straight away. Otherwise (for blk accounting maintained
- * by ZFS and when #inode is estimated from #blks) accounting is updated
- * at commit time and the call to qsd_op_end() must be delayed */
- if (oh->ot_quota_trans.lqt_id_cnt > 0 &&
- !oh->ot_quota_trans.lqt_ids[0].lqi_is_blk &&
- !osd->od_quota_iused_est)
- qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
-
rc = dt_txn_hook_stop(env, th);
if (rc != 0)
CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
LASSERT(oh->ot_tx);
txg = oh->ot_tx->tx_txg;
- osd_object_sa_dirty_rele(oh);
+ osd_object_sa_dirty_rele(env, 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);
+ osd_oti_get(env)->oti_in_trans = 0;
- osd_unlinked_list_emptify(osd, &unlinked, true);
+ osd_unlinked_list_emptify(env, osd, &unlinked, true);
- if (sync)
- txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
+ if (sync) {
+ if (osd_txg_sync_delay_us < 0)
+ txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
+ else
+ udelay(osd_txg_sync_delay_us);
+ }
RETURN(rc);
}
dmu_tx_t *tx;
ENTRY;
+ if (dt->dd_rdonly) {
+ CERROR("%s: someone try to start transaction under "
+ "readonly mode, should be disabled.\n",
+ osd_name(osd_dt_dev(dt)));
+ dump_stack();
+ RETURN(ERR_PTR(-EROFS));
+ }
+
tx = dmu_tx_create(osd->od_os);
if (tx == NULL)
RETURN(ERR_PTR(-ENOMEM));
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));
th = &oh->ot_super;
th->th_dev = dt;
th->th_result = 0;
- th->th_tags = LCT_TX_HANDLE;
RETURN(th);
}
CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);
- est_usedblocks = (usedbytes >> est_maxblockshift) +
- (OSD_DNODE_EST_COUNT >> OSD_DNODE_EST_BLKSHIFT);
- est_usedobjs = usedobjs + OSD_DNODE_EST_COUNT;
+ est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) +
+ usedbytes) >> est_maxblockshift;
+ est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs;
if (est_usedobjs <= est_usedblocks) {
/*
osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */
/* Take replication (i.e. number of copies) into account */
- osfs->os_bavail /= os->os_copies;
+ if (os->os_copies != 0)
+ osfs->os_bavail /= os->os_copies;
/*
* Reserve some space so we don't run into ENOSPC due to grants not
param->ddp_max_extent_blks =
(1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
param->ddp_extent_tax = osd_blk_insert_cost(osd);
+
+ /* Preferred RPC size for efficient disk IO. 1MB shows good
+ * all-around performance for ZFS, but use blocksize (recordsize)
+ * by default if larger to avoid read-modify-write. */
+ if (osd->od_max_blksz > ONE_MB_BRW_SIZE)
+ param->ddp_brw_size = osd->od_max_blksz;
+ else
+ param->ddp_brw_size = ONE_MB_BRW_SIZE;
}
/*
*/
static int osd_sync(const struct lu_env *env, struct dt_device *d)
{
- struct osd_device *osd = osd_dt_dev(d);
- CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
- txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
- CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
+ if (!d->dd_rdonly) {
+ struct osd_device *osd = osd_dt_dev(d);
+
+ CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
+ txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
+ CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
+ }
+
return 0;
}
info->oti_ins_cache = NULL;
info->oti_ins_cache_size = 0;
}
+ lu_buf_free(&info->oti_xattr_lbuf);
OBD_FREE_PTR(info);
}
/* shutdown quota slave instance associated with the device */
if (o->od_quota_slave != NULL) {
+ /* complete all in-flight callbacks */
+ osd_sync(env, &o->od_dt_dev);
+ txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
qsd_fini(env, o->od_quota_slave);
o->od_quota_slave = NULL;
}
osd->od_prop_rdonly = !!newval;
}
+#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
+static void osd_dnodesize_changed_cb(void *arg, uint64_t newval)
+{
+ struct osd_device *osd = arg;
+
+ osd->od_dnsize = newval;
+}
+#endif
/*
* This function unregisters all registered callbacks. It's harmless to
* unregister callbacks that were never registered so it is used to safely
osd_recordsize_changed_cb, o);
(void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY),
osd_readonly_changed_cb, o);
+#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
+ (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
+ osd_dnodesize_changed_cb, o);
+#endif
if (o->arc_prune_cb != NULL) {
arc_remove_prune_callback(o->arc_prune_cb);
if (rc)
GOTO(err, rc);
+#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE
+ rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE),
+ osd_dnodesize_changed_cb, o);
+ if (rc)
+ GOTO(err, rc);
+#endif
+
o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);
err:
dsl_pool_config_exit(dp, FTAG);
static int osd_objset_open(struct osd_device *o)
{
uint64_t version = ZPL_VERSION;
- uint64_t sa_obj;
+ uint64_t sa_obj, unlink_obj;
int rc;
ENTRY;
- rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os);
+ rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS,
+ o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE,
+ 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);
}
/* Check ZFS version */
}
rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET,
- 8, 1, &o->od_unlinkedid);
+ 8, 1, &unlink_obj);
if (rc) {
CERROR("%s: lookup for %s failed: rc = %d\n",
o->od_svname, ZFS_UNLINKED_SET, rc);
GOTO(out, rc = -ENOTSUPP);
}
+ rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked);
+ if (rc) {
+ CERROR("%s: can't get dnode for unlinked: rc = %d\n",
+ o->od_svname, rc);
+ GOTO(out, rc);
+ }
+
out:
if (rc != 0 && o->od_os != NULL) {
- dmu_objset_disown(o->od_os, o);
+ osd_dmu_objset_disown(o->od_os, B_FALSE, o);
o->od_os = NULL;
}
RETURN(rc);
}
-static int
-osd_unlinked_object_free(struct osd_device *osd, uint64_t oid)
+int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
+ uint64_t oid)
{
+ char *key = osd_oti_get(env)->oti_str;
int rc;
dmu_tx_t *tx;
+ if (osd->od_dt_dev.dd_rdonly) {
+ CERROR("%s: someone try to free objects under "
+ "readonly mode, should be disabled.\n", osd_name(osd));
+ dump_stack();
+
+ return -EROFS;
+ }
+
rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END);
if (rc != 0) {
CWARN("%s: Cannot truncate %llu: rc = %d\n",
}
tx = dmu_tx_create(osd->od_os);
+ dmu_tx_mark_netfree(tx);
dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
- dmu_tx_hold_zap(tx, osd->od_unlinkedid, FALSE, NULL);
+ osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
+ FALSE, NULL);
rc = -dmu_tx_assign(tx, TXG_WAIT);
if (rc != 0) {
CWARN("%s: Cannot assign tx for %llu: rc = %d\n",
goto failed;
}
- rc = -zap_remove_int(osd->od_os, osd->od_unlinkedid, oid, tx);
+ snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid);
+ rc = osd_zap_remove(osd, osd->od_unlinked->dn_object,
+ osd->od_unlinked, key, tx);
if (rc != 0) {
CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n",
osd->od_svname, oid, rc);
zap_cursor_t zc;
zap_attribute_t *za = &osd_oti_get(env)->oti_za;
- zap_cursor_init(&zc, osd->od_os, osd->od_unlinkedid);
+ zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object);
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)
+ if (osd_unlinked_object_free(env, osd, za->za_first_integer))
break;
zap_cursor_advance(&zc);
}
struct osd_device *o, struct lustre_cfg *cfg)
{
char *mntdev = lustre_cfg_string(cfg, 1);
+ char *str = lustre_cfg_string(cfg, 2);
char *svname = lustre_cfg_string(cfg, 4);
dnode_t *rootdn;
const char *opts;
if (rc >= sizeof(o->od_svname))
RETURN(-E2BIG);
+ o->od_index_backup_stop = 0;
+ o->od_index = -1; /* -1 means index is invalid */
+ rc = server_name2index(o->od_svname, &o->od_index, NULL);
+ str = strstr(str, ":");
+ if (str) {
+ unsigned long flags;
+
+ rc = kstrtoul(str + 1, 10, &flags);
+ if (rc)
+ RETURN(-EINVAL);
+
+ if (flags & LMD_FLG_DEV_RDONLY) {
+ o->od_dt_dev.dd_rdonly = 1;
+ LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
+ svname);
+ }
+
+ if (flags & LMD_FLG_NOSCRUB)
+ o->od_auto_scrub_interval = AS_NEVER;
+ }
+
if (server_name_is_ost(o->od_svname))
o->od_is_ost = 1;
o->od_xattr_in_sa = B_TRUE;
o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa);
- rc = osd_objset_register_callbacks(o);
+ rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn);
if (rc)
GOTO(err, rc);
+ o->od_root = rootdn->dn_object;
+ osd_dnode_rele(rootdn);
- rc = __osd_obj2dnode(env, o->od_os, o->od_rootid, &rootdn);
+ rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT,
+ &o->od_userused_dn);
if (rc)
GOTO(err, rc);
- o->od_root = rootdn->dn_object;
- osd_dnode_rele(rootdn);
-
- /* 1. initialize oi before any file create or file open */
- rc = osd_oi_init(env, o);
+ rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT,
+ &o->od_groupused_dn);
if (rc)
GOTO(err, rc);
+#ifdef ZFS_PROJINHERIT
+ if (dmu_objset_projectquota_enabled(o->od_os)) {
+ rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT,
+ &o->od_projectused_dn);
+ if (rc && rc != -ENOENT)
+ GOTO(err, rc);
+ }
+#endif
+
rc = lu_site_init(&o->od_site, osd2lu_dev(o));
if (rc)
GOTO(err, rc);
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;
+ rc = osd_objset_register_callbacks(o);
+ if (rc)
+ GOTO(err, rc);
+
+ o->od_in_init = 1;
+ rc = osd_scrub_setup(env, o);
+ o->od_in_init = 0;
+ if (rc)
+ GOTO(err, rc);
rc = osd_procfs_init(o, o->od_svname);
if (rc)
GOTO(err, rc);
}
+#ifdef HAVE_DMU_USEROBJ_ACCOUNTING
+ if (!osd_dmu_userobj_accounting_available(o))
+ CWARN("%s: dnode accounting not enabled: "
+ "enable feature@userobj_accounting in pool\n",
+ o->od_mntdev);
+#endif
+
/* parse mount option "noacl", and enable ACL by default */
opts = lustre_cfg_string(cfg, 3);
if (opts == NULL || strstr(opts, "noacl") == NULL)
osd_unlinked_drain(env, o);
err:
- if (rc) {
- dmu_objset_disown(o->od_os, o);
+ if (rc && o->od_os) {
+ osd_dmu_objset_disown(o->od_os, B_FALSE, o);
o->od_os = NULL;
}
CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
atomic_read(&o->od_zerocopy_pin));
+ if (o->od_unlinked) {
+ osd_dnode_rele(o->od_unlinked);
+ o->od_unlinked = NULL;
+ }
+ if (o->od_userused_dn) {
+ osd_dnode_rele(o->od_userused_dn);
+ o->od_userused_dn = NULL;
+ }
+ if (o->od_groupused_dn) {
+ osd_dnode_rele(o->od_groupused_dn);
+ o->od_groupused_dn = NULL;
+ }
+
+#ifdef ZFS_PROJINHERIT
+ if (o->od_projectused_dn) {
+ osd_dnode_rele(o->od_projectused_dn);
+ o->od_projectused_dn = NULL;
+ }
+#endif
+
if (o->od_os != NULL) {
- /* force a txg sync to get all commit callbacks */
- txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
+ if (!o->od_dt_dev.dd_rdonly)
+ /* force a txg sync to get all commit callbacks */
+ txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);
/* close the object set */
- dmu_objset_disown(o->od_os, o);
-
+ osd_dmu_objset_disown(o->od_os, B_FALSE, o);
o->od_os = NULL;
}
l->ld_ops = &osd_lu_ops;
o->od_dt_dev.dd_ops = &osd_dt_ops;
+ sema_init(&o->od_otable_sem, 1);
+ INIT_LIST_HEAD(&o->od_ios_list);
+ o->od_auto_scrub_interval = AS_DEFAULT;
out:
RETURN(rc);
INIT_LIST_HEAD(&osl->osl_seq_list);
rwlock_init(&osl->osl_seq_list_lock);
sema_init(&osl->osl_seq_init_sem, 1);
+ INIT_LIST_HEAD(&dev->od_index_backup_list);
+ INIT_LIST_HEAD(&dev->od_index_restore_list);
+ spin_lock_init(&dev->od_lock);
+ dev->od_index_backup_policy = LIBP_NONE;
rc = dt_device_init(&dev->od_dt_dev, type);
if (rc == 0) {
ENTRY;
- osd_shutdown(env, o);
- osd_oi_fini(env, o);
-
if (o->od_os) {
osd_objset_unregister_callbacks(o);
- osd_sync(env, lu2dt_dev(d));
- txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
+ if (!o->od_dt_dev.dd_rdonly) {
+ osd_sync(env, lu2dt_dev(d));
+ txg_wait_callbacks(
+ spa_get_dsl(dmu_objset_spa(o->od_os)));
+ }
}
+ /* now with all the callbacks completed we can cleanup the remainings */
+ osd_shutdown(env, o);
+ osd_scrub_cleanup(env, o);
+
rc = osd_procfs_fini(o);
if (rc) {
CERROR("proc fini error %d\n", rc);
rc = osd_mount(env, o, cfg);
break;
case LCFG_CLEANUP:
+ /* For the case LCFG_PRE_CLEANUP is not called in advance,
+ * that may happend if hit failure during mount process. */
+ osd_index_backup(env, o, false);
rc = osd_shutdown(env, o);
break;
case LCFG_PARAM: {
LASSERT(&o->od_dt_dev);
rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
cfg, &o->od_dt_dev);
- if (rc > 0 || rc == -ENOSYS)
+ if (rc > 0 || rc == -ENOSYS) {
rc = class_process_proc_param(PARAM_OST,
lprocfs_osd_obd_vars,
cfg, &o->od_dt_dev);
+ if (rc > 0)
+ rc = 0;
+ }
break;
}
+ case LCFG_PRE_CLEANUP:
+ osd_index_backup(env, o,
+ o->od_index_backup_policy != LIBP_NONE);
+ rc = 0;
+ break;
default:
rc = -ENOTTY;
}
lu_kmem_fini(osd_caches);
}
-extern unsigned int osd_oi_count;
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_param(osd_txg_sync_delay_us, int, 0644);
+MODULE_PARM_DESC(osd_txg_sync_delay_us,
+ "When zero or larger delay N usec instead of doing TXG sync");
+
MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
MODULE_VERSION(LUSTRE_VERSION_STRING);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff