if (data->locd_declare)
return lod_sub_declare_attr_set(env, dt, data->locd_attr, th);
+ if (data->locd_attr->la_valid & LA_LAYOUT_VERSION) {
+ CDEBUG(D_LAYOUT, DFID": set layout version: %u, comp_idx: %d\n",
+ PFID(lu_object_fid(&dt->do_lu)),
+ data->locd_attr->la_layout_version, comp_idx);
+ }
+
return lod_sub_attr_set(env, dt, data->locd_attr, th);
}
ENTRY;
LASSERT(lo->ldo_flr_state == LCM_FL_WRITE_PENDING);
- LASSERT(mlc->mlc_opc == MD_LAYOUT_WRITE);
+ LASSERT(mlc->mlc_opc == MD_LAYOUT_WRITE ||
+ mlc->mlc_opc == MD_LAYOUT_RESYNC);
/* look for the primary mirror */
for (i = 0; i < lo->ldo_mirror_count; i++) {
/* for LAYOUT_WRITE opc, it has to do the following operations:
* 1. stale overlapping componets from stale mirrors;
* 2. instantiate components of the primary mirror;
- * 3. transfter layout version to all objects of the primary; */
+ * 3. transfter layout version to all objects of the primary;
+ *
+ * for LAYOUT_RESYNC opc, it will do:
+ * 1. instantiate components of all stale mirrors;
+ * 2. transfer layout version to all objects to close write era. */
if (mlc->mlc_opc == MD_LAYOUT_WRITE) {
LASSERT(mlc->mlc_intent != NULL);
info->lti_comp_idx[info->lti_count++] =
lod_comp_index(lo, lod_comp);
}
+ } else { /* MD_LAYOUT_RESYNC */
+ /* figure out the components that have been instantiated in
+ * in primary to decide what components should be instantiated
+ * in stale mirrors */
+ lod_foreach_mirror_comp(lod_comp, lo, primary) {
+ if (!lod_comp_inited(lod_comp))
+ break;
+
+ extent.e_end = lod_comp->llc_extent.e_end;
+ }
+
+ CDEBUG(D_LAYOUT,
+ DFID": instantiate all stale components in "DEXT"\n",
+ PFID(lod_object_fid(lo)), PEXT(&extent));
+
+ /* 1. instantiate all components within this extent, even
+ * non-stale components so that it won't need to instantiate
+ * those components for mirror truncate later. */
+ for (i = 0; i < lo->ldo_mirror_count; i++) {
+ if (primary == i)
+ continue;
+
+ LASSERTF(lo->ldo_mirrors[i].lme_stale,
+ "both %d and %d are primary\n", i, primary);
+
+ lod_foreach_mirror_comp(lod_comp, lo, i) {
+ if (!lu_extent_is_overlapped(&extent,
+ &lod_comp->llc_extent))
+ break;
+
+ if (lod_comp_inited(lod_comp))
+ continue;
+
+ CDEBUG(D_LAYOUT, "resync instantiate %d / %d\n",
+ i, lod_comp_index(lo, lod_comp));
+
+ info->lti_comp_idx[info->lti_count++] =
+ lod_comp_index(lo, lod_comp);
+ }
+ }
+
+ /* change the file state to SYNC_PENDING */
+ lo->ldo_flr_state = LCM_FL_SYNC_PENDING;
}
rc = lod_declare_instantiate_components(env, lo, th);
if (rc)
GOTO(out, rc);
+ /* 3. transfer layout version to OST objects.
+ * transfer new layout version to OST objects so that stale writes
+ * can be denied. It also ends an era of writing by setting
+ * LU_LAYOUT_RESYNC. Normal client can never use this bit to
+ * send write RPC; only resync RPCs could do it. */
layout_attr->la_valid = LA_LAYOUT_VERSION;
layout_attr->la_layout_version = 0; /* set current version */
+ if (mlc->mlc_opc == MD_LAYOUT_RESYNC)
+ layout_attr->la_layout_version = LU_LAYOUT_RESYNC;
rc = lod_declare_attr_set(env, &lo->ldo_obj, layout_attr, th);
if (rc)
GOTO(out, rc);
RETURN(rc);
}
+static int lod_declare_update_sync_pending(const struct lu_env *env,
+ struct lod_object *lo, struct md_layout_change *mlc,
+ struct thandle *th)
+{
+ struct lod_thread_info *info = lod_env_info(env);
+ unsigned sync_components = 0;
+ unsigned resync_components = 0;
+ int i;
+ int rc;
+ ENTRY;
+
+ LASSERT(lo->ldo_flr_state == LCM_FL_SYNC_PENDING);
+ LASSERT(mlc->mlc_opc == MD_LAYOUT_RESYNC_DONE ||
+ mlc->mlc_opc == MD_LAYOUT_WRITE);
+
+ CDEBUG(D_LAYOUT, DFID ": received op %d in sync pending\n",
+ PFID(lod_object_fid(lo)), mlc->mlc_opc);
+
+ if (mlc->mlc_opc == MD_LAYOUT_WRITE) {
+ CDEBUG(D_LAYOUT, DFID": cocurrent write to sync pending\n",
+ PFID(lod_object_fid(lo)));
+
+ lo->ldo_flr_state = LCM_FL_WRITE_PENDING;
+ return lod_declare_update_write_pending(env, lo, mlc, th);
+ }
+
+ /* MD_LAYOUT_RESYNC_DONE */
+
+ for (i = 0; i < lo->ldo_comp_cnt; i++) {
+ struct lod_layout_component *lod_comp;
+ int j;
+
+ lod_comp = &lo->ldo_comp_entries[i];
+
+ if (!(lod_comp->llc_flags & LCME_FL_STALE)) {
+ sync_components++;
+ continue;
+ }
+
+ for (j = 0; j < mlc->mlc_resync_count; j++) {
+ if (lod_comp->llc_id != mlc->mlc_resync_ids[j])
+ continue;
+
+ mlc->mlc_resync_ids[j] = LCME_ID_INVAL;
+ lod_comp->llc_flags &= ~LCME_FL_STALE;
+ resync_components++;
+ break;
+ }
+ }
+
+ /* valid check */
+ for (i = 0; i < mlc->mlc_resync_count; i++) {
+ if (mlc->mlc_resync_ids[i] == LCME_ID_INVAL)
+ continue;
+
+ CDEBUG(D_LAYOUT, DFID": lcme id %u (%d / %zd) not exist "
+ "or already synced\n", PFID(lod_object_fid(lo)),
+ mlc->mlc_resync_ids[i], i, mlc->mlc_resync_count);
+ GOTO(out, rc = -EINVAL);
+ }
+
+ if (!sync_components || !resync_components) {
+ CDEBUG(D_LAYOUT, DFID": no mirror in sync or resync\n",
+ PFID(lod_object_fid(lo)));
+
+ /* tend to return an error code here to prevent
+ * the MDT from setting SoM attribute */
+ GOTO(out, rc = -EINVAL);
+ }
+
+ CDEBUG(D_LAYOUT, DFID": resynced %u/%zu components\n",
+ PFID(lod_object_fid(lo)),
+ resync_components, mlc->mlc_resync_count);
+
+ lo->ldo_flr_state = LCM_FL_RDONLY;
+ lod_obj_inc_layout_gen(lo);
+
+ info->lti_buf.lb_len = lod_comp_md_size(lo, false);
+ rc = lod_sub_declare_xattr_set(env, lod_object_child(lo),
+ &info->lti_buf, XATTR_NAME_LOV, 0, th);
+ EXIT;
+
+out:
+ if (rc)
+ lod_object_free_striping(env, lo);
+ RETURN(rc);
+}
+
static int lod_declare_layout_change(const struct lu_env *env,
struct dt_object *dt, struct md_layout_change *mlc,
struct thandle *th)
rc = lod_declare_update_write_pending(env, lo, mlc, th);
break;
case LCM_FL_SYNC_PENDING:
+ rc = lod_declare_update_sync_pending(env, lo, mlc, th);
+ break;
default:
rc = -ENOTSUPP;
break;
git://git.whamcloud.com / fs / lustre-release.git / blobdiff