LU-8856 osd: mark specific transactions netfree

[fs/lustre-release.git] / lustre / osd-zfs / osd_handler.c

diff --git a/lustre/osd-zfs/osd_handler.c b/lustre/osd-zfs/osd_handler.c
index 724c3e9..23d684c 100644 (file)
--- a/lustre/osd-zfs/osd_handler.c
+++ b/lustre/osd-zfs/osd_handler.c
@@ -66,6 +66,8 @@
 
 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;
 
@@ -315,8 +317,12 @@ static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
 
        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);
 }
@@ -608,6 +614,14 @@ static void osd_conf_get(const struct lu_env *env,
        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;
 }
 
 /*
@@ -963,6 +977,7 @@ int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd,
        }
 
        tx = dmu_tx_create(osd->od_os);
+       dmu_tx_mark_netfree(tx);
        dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
        osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked,
                        FALSE, NULL);
@@ -1044,6 +1059,9 @@ static int osd_mount(const struct lu_env *env,
        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;
@@ -1057,6 +1075,9 @@ static int osd_mount(const struct lu_env *env,
                        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))
@@ -1094,11 +1115,6 @@ static int osd_mount(const struct lu_env *env,
        }
 #endif
 
-       /* 1. initialize oi before any file create or file open */
-       rc = osd_oi_init(env, o);
-       if (rc)
-               GOTO(err, rc);
-
        rc = lu_site_init(&o->od_site, osd2lu_dev(o));
        if (rc)
                GOTO(err, rc);
@@ -1112,6 +1128,12 @@ static int osd_mount(const struct lu_env *env,
        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);
@@ -1208,6 +1230,9 @@ static int osd_device_init0(const struct lu_env *env,
 
        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);
@@ -1232,6 +1257,10 @@ static struct lu_device *osd_device_alloc(const struct lu_env *env,
        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) {
@@ -1290,7 +1319,7 @@ static struct lu_device *osd_device_fini(const struct lu_env *env,
 
        /* now with all the callbacks completed we can cleanup the remainings */
        osd_shutdown(env, o);
-       osd_oi_fini(env, o);
+       osd_scrub_cleanup(env, o);
 
        rc = osd_procfs_fini(o);
        if (rc) {
@@ -1326,6 +1355,9 @@ static int osd_process_config(const struct lu_env *env,
                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: {
@@ -1341,6 +1373,11 @@ static int osd_process_config(const struct lu_env *env,
                }
                break;
        }
+       case LCFG_PRE_CLEANUP:
+               osd_index_backup(env, o,
+                                o->od_index_backup_policy != LIBP_NONE);
+               rc = 0;
+               break;
        default:
                rc = -ENOTTY;
        }
@@ -1538,10 +1575,13 @@ static void __exit osd_exit(void)
        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);