+
+static void tgt_clean_by_tag(struct obd_export *exp, __u64 xid, __u16 tag)
+{
+ struct tg_export_data *ted = &exp->exp_target_data;
+ struct lu_target *lut = class_exp2tgt(exp);
+ struct tg_reply_data *trd, *tmp;
+
+ if (tag == 0)
+ return;
+
+ list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
+ if (trd->trd_tag != tag)
+ continue;
+
+ LASSERT(ergo(tgt_is_increasing_xid_client(exp),
+ trd->trd_reply.lrd_xid <= xid));
+
+ ted->ted_release_tag++;
+ tgt_release_reply_data(lut, ted, trd);
+ }
+}
+
+int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
+ struct tg_export_data *ted, struct tg_reply_data *trd,
+ struct ptlrpc_request *req,
+ struct thandle *th, bool update_lrd_file)
+{
+ struct lsd_reply_data *lrd;
+ int i;
+
+ lrd = &trd->trd_reply;
+ /* update export last transno */
+ mutex_lock(&ted->ted_lcd_lock);
+ if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
+ ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
+ mutex_unlock(&ted->ted_lcd_lock);
+
+ /* find a empty slot */
+ i = tgt_find_free_reply_slot(tgt);
+ if (unlikely(i < 0)) {
+ CERROR("%s: couldn't find a slot for reply data: "
+ "rc = %d\n", tgt_name(tgt), i);
+ RETURN(i);
+ }
+ trd->trd_index = i;
+
+ if (update_lrd_file) {
+ loff_t off;
+ int rc;
+
+ /* write reply data to disk */
+ off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
+ rc = tgt_reply_data_write(env, tgt, lrd, off, th);
+ if (unlikely(rc != 0)) {
+ CERROR("%s: can't update %s file: rc = %d\n",
+ tgt_name(tgt), REPLY_DATA, rc);
+ RETURN(rc);
+ }
+ }
+ /* add reply data to target export's reply list */
+ mutex_lock(&ted->ted_lcd_lock);
+ if (req != NULL) {
+ int exclude = tgt_is_increasing_xid_client(req->rq_export) ?
+ MSG_REPLAY : MSG_REPLAY|MSG_RESENT;
+
+ if (req->rq_obsolete) {
+ mutex_unlock(&ted->ted_lcd_lock);
+ RETURN(-EALREADY);
+ }
+
+ if (!(lustre_msg_get_flags(req->rq_reqmsg) & exclude))
+ tgt_clean_by_tag(req->rq_export, req->rq_xid,
+ trd->trd_tag);
+ }
+ list_add(&trd->trd_list, &ted->ted_reply_list);
+ ted->ted_reply_cnt++;
+ if (ted->ted_reply_cnt > ted->ted_reply_max)
+ ted->ted_reply_max = ted->ted_reply_cnt;
+ mutex_unlock(&ted->ted_lcd_lock);
+
+ CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
+ "tag %hu, client gen %u, slot idx %d\n",
+ trd, lrd->lrd_xid, lrd->lrd_transno,
+ trd->trd_tag, lrd->lrd_client_gen, trd->trd_index);
+
+ RETURN(0);
+}
+EXPORT_SYMBOL(tgt_add_reply_data);
+
+/*
+ * last_rcvd & last_committed update callbacks
+ */
+static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
+ struct dt_object *obj, __u64 opdata,
+ struct thandle *th, struct ptlrpc_request *req)
+{
+ struct tgt_thread_info *tti = tgt_th_info(env);
+ struct tgt_session_info *tsi = tgt_ses_info(env);
+ struct obd_export *exp = tsi->tsi_exp;
+ struct tg_export_data *ted;
+ __u64 *transno_p;
+ int rc = 0;
+ bool lw_client;
+
+ ENTRY;
+
+
+ LASSERT(exp != NULL);
+ ted = &exp->exp_target_data;
+
+ lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
+ if (ted->ted_lr_idx < 0 && !lw_client)
+ /* ofd connect may cause transaction before export has
+ * last_rcvd slot */
+ RETURN(0);
+
+ if (req != NULL)
+ tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
+ else
+ /* From update replay, tti_transno should be set already */
+ LASSERT(tti->tti_transno != 0);
+
+ spin_lock(&tgt->lut_translock);
+ if (th->th_result != 0) {
+ if (tti->tti_transno != 0) {
+ CERROR("%s: replay transno %llu failed: rc = %d\n",
+ tgt_name(tgt), tti->tti_transno, th->th_result);
+ }
+ } else if (tti->tti_transno == 0) {
+ tti->tti_transno = ++tgt->lut_last_transno;
+ } else {
+ /* should be replay */
+ if (tti->tti_transno > tgt->lut_last_transno)
+ tgt->lut_last_transno = tti->tti_transno;
+ }
+ spin_unlock(&tgt->lut_translock);
+
+ /** VBR: set new versions */
+ if (th->th_result == 0 && obj != NULL) {
+ struct dt_object *dto = dt_object_locate(obj, th->th_dev);
+ dt_version_set(env, dto, tti->tti_transno, th);
+ }
+
+ /* filling reply data */
+ CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
+ tti->tti_transno, tgt->lut_obd->obd_last_committed);
+
+ if (req != NULL) {
+ req->rq_transno = tti->tti_transno;
+ lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
+ }
+
+ /* if can't add callback, do sync write */
+ th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
+
+ if (lw_client) {
+ /* All operations performed by LW clients are synchronous and
+ * we store the committed transno in the last_rcvd header */
+ spin_lock(&tgt->lut_translock);
+ if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
+ tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
+ spin_unlock(&tgt->lut_translock);
+ /* Although lightweight (LW) connections have no slot
+ * in the last_rcvd, we still want to maintain
+ * the in-memory lsd_client_data structure in order to
+ * properly handle reply reconstruction. */
+ rc = tgt_server_data_write(env, tgt, th);
+ } else {
+ spin_unlock(&tgt->lut_translock);
+ }
+ } else if (ted->ted_lr_off == 0) {
+ CERROR("%s: client idx %d has offset %lld\n",
+ tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
+ RETURN(-EINVAL);
+ }
+
+ /* Target that supports multiple reply data */
+ if (tgt_is_multimodrpcs_client(exp)) {
+ struct tg_reply_data *trd;
+ struct lsd_reply_data *lrd;
+ __u64 *pre_versions;
+ bool write_update;
+
+ OBD_ALLOC_PTR(trd);
+ if (unlikely(trd == NULL))
+ RETURN(-ENOMEM);
+
+ /* fill reply data information */
+ lrd = &trd->trd_reply;
+ lrd->lrd_transno = tti->tti_transno;
+ if (req != NULL) {
+ lrd->lrd_xid = req->rq_xid;
+ trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
+ pre_versions = lustre_msg_get_versions(req->rq_repmsg);
+ lrd->lrd_result = th->th_result;
+ lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
+ write_update = true;
+ } else {
+ LASSERT(tsi->tsi_xid != 0);
+ lrd->lrd_xid = tsi->tsi_xid;
+ lrd->lrd_result = tsi->tsi_result;
+ lrd->lrd_client_gen = tsi->tsi_client_gen;
+ trd->trd_tag = 0;
+ pre_versions = NULL;
+ write_update = false;
+ }
+
+ lrd->lrd_data = opdata;
+ if (pre_versions) {
+ trd->trd_pre_versions[0] = pre_versions[0];
+ trd->trd_pre_versions[1] = pre_versions[1];
+ trd->trd_pre_versions[2] = pre_versions[2];
+ trd->trd_pre_versions[3] = pre_versions[3];
+ }
+
+ rc = tgt_add_reply_data(env, tgt, ted, trd, req,
+ th, write_update);
+ if (rc < 0)
+ OBD_FREE_PTR(trd);
+ return rc;
+ }
+
+ /* Enough for update replay, let's return */
+ if (req == NULL)
+ RETURN(rc);
+
+ mutex_lock(&ted->ted_lcd_lock);
+ LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
+ if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
+ transno_p = &ted->ted_lcd->lcd_last_close_transno;
+ ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
+ ted->ted_lcd->lcd_last_close_result = th->th_result;
+ } else {
+ /* VBR: save versions in last_rcvd for reconstruct. */
+ __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
+
+ if (pre_versions) {
+ ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
+ ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
+ ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
+ ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
+ }
+ transno_p = &ted->ted_lcd->lcd_last_transno;
+ ted->ted_lcd->lcd_last_xid = req->rq_xid;
+ ted->ted_lcd->lcd_last_result = th->th_result;
+ /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
+ * see struct ldlm_reply->lock_policy_res1; */
+ ted->ted_lcd->lcd_last_data = opdata;
+ }
+
+ /* Update transno in slot only if non-zero number, i.e. no errors */
+ if (likely(tti->tti_transno != 0)) {
+ /* Don't overwrite bigger transaction number with lower one.
+ * That is not sign of problem in all cases, but in any case
+ * this value should be monotonically increased only. */
+ if (*transno_p > tti->tti_transno) {
+ if (!tgt->lut_no_reconstruct) {
+ CERROR("%s: trying to overwrite bigger transno:"
+ "on-disk: %llu, new: %llu replay: "
+ "%d. See LU-617.\n", tgt_name(tgt),
+ *transno_p, tti->tti_transno,
+ req_is_replay(req));
+ if (req_is_replay(req)) {
+ spin_lock(&req->rq_export->exp_lock);
+ req->rq_export->exp_vbr_failed = 1;
+ spin_unlock(&req->rq_export->exp_lock);
+ }
+ mutex_unlock(&ted->ted_lcd_lock);
+ RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
+ }
+ } else {
+ *transno_p = tti->tti_transno;
+ }
+ }
+
+ if (!lw_client) {
+ tti->tti_off = ted->ted_lr_off;
+ if (CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVD_EIO))
+ rc = -EIO;
+ else
+ rc = tgt_client_data_write(env, tgt, ted->ted_lcd,
+ &tti->tti_off, th);
+ if (rc < 0) {
+ mutex_unlock(&ted->ted_lcd_lock);
+ RETURN(rc);
+ }
+ }
+ mutex_unlock(&ted->ted_lcd_lock);
+ RETURN(rc);
+}
+
+/*
+ * last_rcvd update for echo client simulation.
+ * It updates last_rcvd client slot and version of object in
+ * simple way but with all locks to simulate all drawbacks
+ */
+static int tgt_last_rcvd_update_echo(const struct lu_env *env,
+ struct lu_target *tgt,
+ struct dt_object *obj,
+ struct thandle *th,
+ struct obd_export *exp)
+{
+ struct tgt_thread_info *tti = tgt_th_info(env);
+ struct tg_export_data *ted = &exp->exp_target_data;
+ int rc = 0;
+
+ ENTRY;
+
+ tti->tti_transno = 0;
+
+ spin_lock(&tgt->lut_translock);
+ if (th->th_result == 0)
+ tti->tti_transno = ++tgt->lut_last_transno;
+ spin_unlock(&tgt->lut_translock);
+
+ /** VBR: set new versions */
+ if (th->th_result == 0 && obj != NULL)
+ dt_version_set(env, obj, tti->tti_transno, th);
+
+ /* if can't add callback, do sync write */
+ th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
+ tti->tti_transno);
+
+ LASSERT(ted->ted_lr_off > 0);
+
+ mutex_lock(&ted->ted_lcd_lock);
+ LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
+ ted->ted_lcd->lcd_last_transno = tti->tti_transno;
+ ted->ted_lcd->lcd_last_result = th->th_result;
+
+ tti->tti_off = ted->ted_lr_off;
+ rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
+ mutex_unlock(&ted->ted_lcd_lock);
+ RETURN(rc);
+}
+
+static int tgt_clients_data_init(const struct lu_env *env,
+ struct lu_target *tgt,
+ unsigned long last_size)
+{
+ struct obd_device *obd = tgt->lut_obd;
+ struct lr_server_data *lsd = &tgt->lut_lsd;
+ struct lsd_client_data *lcd = NULL;
+ struct tg_export_data *ted;
+ int cl_idx;
+ int rc = 0;
+ loff_t off = lsd->lsd_client_start;
+ __u32 generation = 0;
+ struct cfs_hash *hash = NULL;
+
+ ENTRY;
+
+ if (tgt->lut_bottom->dd_rdonly)
+ RETURN(0);
+
+ BUILD_BUG_ON(offsetof(struct lsd_client_data, lcd_padding) +
+ sizeof(lcd->lcd_padding) != LR_CLIENT_SIZE);
+
+ OBD_ALLOC_PTR(lcd);
+ if (lcd == NULL)
+ RETURN(-ENOMEM);
+
+ hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
+ if (hash == NULL)
+ GOTO(err_out, rc = -ENODEV);
+
+ for (cl_idx = 0; off < last_size; cl_idx++) {
+ struct obd_export *exp;
+ __u64 last_transno;
+
+ /* Don't assume off is incremented properly by
+ * read_record(), in case sizeof(*lcd)
+ * isn't the same as fsd->lsd_client_size. */
+ off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
+ rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
+ if (rc) {
+ CERROR("%s: error reading last_rcvd %s idx %d off "
+ "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
+ cl_idx, off, rc);
+ rc = 0;
+ break; /* read error shouldn't cause startup to fail */
+ }
+
+ if (lcd->lcd_uuid[0] == '\0') {
+ CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
+ cl_idx);
+ continue;
+ }
+
+ last_transno = lcd_last_transno(lcd);
+
+ /* These exports are cleaned up by disconnect, so they
+ * need to be set up like real exports as connect does.
+ */
+ CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: %llu"
+ " srv lr: %llu lx: %llu gen %u\n", lcd->lcd_uuid,
+ cl_idx, last_transno, lsd->lsd_last_transno,
+ lcd_last_xid(lcd), lcd->lcd_generation);
+
+ exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
+ if (IS_ERR(exp)) {
+ if (PTR_ERR(exp) == -EALREADY) {
+ /* export already exists, zero out this one */
+ CERROR("%s: Duplicate export %s!\n",
+ tgt_name(tgt), lcd->lcd_uuid);
+ continue;
+ }
+ GOTO(err_out, rc = PTR_ERR(exp));
+ }
+
+ ted = &exp->exp_target_data;
+ *ted->ted_lcd = *lcd;
+
+ rc = tgt_client_add(env, exp, cl_idx);
+ LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
+ /* VBR: set export last committed version */
+ exp->exp_last_committed = last_transno;
+ spin_lock(&exp->exp_lock);
+ exp->exp_connecting = 0;
+ exp->exp_in_recovery = 0;
+ spin_unlock(&exp->exp_lock);
+ atomic_inc(&obd->obd_max_recoverable_clients);
+
+ if (tgt->lut_lsd.lsd_feature_incompat &
+ OBD_INCOMPAT_MULTI_RPCS &&
+ lcd->lcd_generation != 0) {
+ /* compute the highest valid client generation */
+ generation = max(generation, lcd->lcd_generation);
+ /* fill client_generation <-> export hash table */
+ rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
+ &exp->exp_gen_hash);
+ if (rc != 0) {
+ CERROR("%s: duplicate export for client "
+ "generation %u\n",
+ tgt_name(tgt), lcd->lcd_generation);
+ class_export_put(exp);
+ GOTO(err_out, rc);
+ }
+ }
+
+ class_export_put(exp);
+
+ rc = rev_import_init(exp);
+ if (rc != 0) {
+ class_unlink_export(exp);
+ GOTO(err_out, rc);
+ }
+
+ /* Need to check last_rcvd even for duplicated exports. */
+ CDEBUG(D_OTHER, "client at idx %d has last_transno = %llu\n",
+ cl_idx, last_transno);
+
+ spin_lock(&tgt->lut_translock);
+ tgt->lut_last_transno = max(last_transno,
+ tgt->lut_last_transno);
+ spin_unlock(&tgt->lut_translock);
+ }
+
+ /* record highest valid client generation */
+ atomic_set(&tgt->lut_client_generation, generation);
+
+err_out:
+ if (hash != NULL)
+ cfs_hash_putref(hash);
+ OBD_FREE_PTR(lcd);
+ RETURN(rc);
+}
+
+struct server_compat_data {
+ __u32 rocompat;
+ __u32 incompat;
+ __u32 rocinit;
+ __u32 incinit;
+};
+
+static struct server_compat_data tgt_scd[] = {
+ [LDD_F_SV_TYPE_MDT] = {
+ .rocompat = OBD_ROCOMPAT_LOVOBJID,
+ .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
+ OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
+ OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
+ OBD_INCOMPAT_MULTI_RPCS,
+ .rocinit = OBD_ROCOMPAT_LOVOBJID,
+ .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
+ OBD_INCOMPAT_MULTI_OI,
+ },
+ [LDD_F_SV_TYPE_OST] = {
+ .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
+ .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
+ OBD_INCOMPAT_FID,
+ .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
+ .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
+ }
+};
+
+int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
+{
+ struct tgt_thread_info *tti = tgt_th_info(env);
+ struct lr_server_data *lsd = &tgt->lut_lsd;
+ unsigned long last_rcvd_size;
+ __u32 index;
+ int rc, type;
+
+ rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
+ if (rc)
+ RETURN(rc);
+
+ last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
+
+ /* ensure padding in the struct is the correct size */
+ BUILD_BUG_ON(offsetof(struct lr_server_data, lsd_padding) +
+ sizeof(lsd->lsd_padding) != LR_SERVER_SIZE);
+
+ rc = server_name2index(tgt_name(tgt), &index, NULL);
+ if (rc < 0) {
+ CERROR("%s: Can not get index from name: rc = %d\n",
+ tgt_name(tgt), rc);
+ RETURN(rc);
+ }
+ /* server_name2index() returns type */
+ type = rc;
+ if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
+ CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
+ RETURN(-EINVAL);
+ }
+
+ /* last_rcvd on OST doesn't provide reconstruct support because there
+ * may be up to 8 in-flight write requests per single slot in
+ * last_rcvd client data
+ */
+ tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
+
+ if (last_rcvd_size == 0) {
+ LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
+
+ memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
+ sizeof(lsd->lsd_uuid));
+ lsd->lsd_last_transno = 0;
+ lsd->lsd_mount_count = 0;
+ lsd->lsd_server_size = LR_SERVER_SIZE;
+ lsd->lsd_client_start = LR_CLIENT_START;
+ lsd->lsd_client_size = LR_CLIENT_SIZE;
+ lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
+ lsd->lsd_osd_index = index;
+ lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
+ lsd->lsd_feature_incompat = tgt_scd[type].incinit;
+ } else {
+ rc = tgt_server_data_read(env, tgt);
+ if (rc) {
+ CERROR("%s: error reading LAST_RCVD: rc= %d\n",
+ tgt_name(tgt), rc);
+ RETURN(rc);
+ }
+ if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
+ if (tgt->lut_bottom->dd_rdonly) {
+ /* Such difference may be caused by mounting
+ * up snapshot with new fsname under rd_only
+ * mode. But even if it was NOT, it will not
+ * damage the system because of "rd_only". */
+ memcpy(lsd->lsd_uuid,
+ tgt->lut_obd->obd_uuid.uuid,
+ sizeof(lsd->lsd_uuid));
+ } else {
+ LCONSOLE_ERROR_MSG(0x157, "Trying to start "
+ "OBD %s using the wrong "
+ "disk %s. Were the /dev/ "
+ "assignments rearranged?\n",
+ tgt->lut_obd->obd_uuid.uuid,
+ lsd->lsd_uuid);
+ RETURN(-EINVAL);
+ }
+ }
+
+ if (lsd->lsd_osd_index != index) {
+ LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
+ "is different with the index %d in"
+ "config log, It might be disk"
+ "corruption!\n", tgt_name(tgt),
+ lsd->lsd_osd_index, index);
+ RETURN(-EINVAL);
+ }
+ }
+
+ if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
+ CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
+ tgt_name(tgt),
+ lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
+ RETURN(-EINVAL);
+ }
+
+ if (type == LDD_F_SV_TYPE_MDT)
+ lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
+
+ if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
+ CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
+ tgt_name(tgt),
+ lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
+ RETURN(-EINVAL);
+ }
+ /** Interop: evict all clients at first boot with 1.8 last_rcvd */
+ if (type == LDD_F_SV_TYPE_MDT &&
+ !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
+ if (last_rcvd_size > lsd->lsd_client_start) {
+ LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
+ "remove all clients for interop needs\n",
+ tgt_name(tgt));
+ rc = tgt_truncate_last_rcvd(env, tgt,
+ lsd->lsd_client_start);
+ if (rc)
+ RETURN(rc);
+ last_rcvd_size = lsd->lsd_client_start;
+ }
+ /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
+ lsd->lsd_feature_compat |= OBD_COMPAT_20;
+ }
+
+ spin_lock(&tgt->lut_translock);
+ tgt->lut_last_transno = lsd->lsd_last_transno;
+ spin_unlock(&tgt->lut_translock);
+
+ lsd->lsd_mount_count++;
+
+ CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
+ CDEBUG(D_INODE, "%s: server last_transno: %llu\n",
+ tgt_name(tgt), tgt->lut_last_transno);
+ CDEBUG(D_INODE, "%s: server mount_count: %llu\n",
+ tgt_name(tgt), lsd->lsd_mount_count);
+ CDEBUG(D_INODE, "%s: server data size: %u\n",
+ tgt_name(tgt), lsd->lsd_server_size);
+ CDEBUG(D_INODE, "%s: per-client data start: %u\n",
+ tgt_name(tgt), lsd->lsd_client_start);
+ CDEBUG(D_INODE, "%s: per-client data size: %u\n",
+ tgt_name(tgt), lsd->lsd_client_size);
+ CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
+ tgt_name(tgt), last_rcvd_size);
+ CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
+ tgt_name(tgt), lsd->lsd_subdir_count);
+ CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
+ last_rcvd_size <= lsd->lsd_client_start ? 0 :
+ (last_rcvd_size - lsd->lsd_client_start) /
+ lsd->lsd_client_size);
+ CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
+
+ if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
+ lsd->lsd_client_size == 0) {
+ CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
+ RETURN(-EINVAL);
+ }
+
+ if (!tgt->lut_obd->obd_replayable)
+ CWARN("%s: recovery support OFF\n", tgt_name(tgt));
+
+ rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
+ if (rc < 0)
+ GOTO(err_client, rc);
+
+ spin_lock(&tgt->lut_translock);
+ /* obd_last_committed is used for compatibility
+ * with other lustre recovery code */
+ tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
+ spin_unlock(&tgt->lut_translock);
+
+ tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
+ tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
+
+ /* save it, so mount count and last_transno is current */
+ rc = tgt_server_data_update(env, tgt, 0);
+ if (rc < 0)
+ GOTO(err_client, rc);
+
+ RETURN(0);
+
+err_client:
+ class_disconnect_exports(tgt->lut_obd);
+ return rc;
+}
+
+/* add credits for last_rcvd update */
+int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
+ void *cookie)
+{
+ struct lu_target *tgt = cookie;
+ struct tgt_session_info *tsi;
+ struct tgt_thread_info *tti = tgt_th_info(env);
+ struct dt_object *dto;
+ int rc;
+
+ /* For readonly case, the caller should have got failure
+ * when start the transaction. If the logic comes here,
+ * there must be something wrong. */
+ if (unlikely(tgt->lut_bottom->dd_rdonly)) {
+ dump_stack();
+ LBUG();
+ }
+
+ /* if there is no session, then this transaction is not result of
+ * request processing but some local operation */
+ if (env->le_ses == NULL)
+ return 0;
+
+ LASSERT(tgt->lut_last_rcvd);
+ tsi = tgt_ses_info(env);
+ /* OFD may start transaction without export assigned */
+ if (tsi->tsi_exp == NULL)
+ return 0;
+
+ if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
+ /*
+ * Use maximum possible file offset for declaration to ensure
+ * ZFS will reserve enough credits for a write anywhere in this
+ * file, since we don't know where in the file the write will be
+ * because a replay slot has not been assigned. This should be
+ * replaced by dmu_tx_hold_append() when available.
+ */
+ tti->tti_buf.lb_buf = NULL;
+ tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
+ dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
+ rc = dt_declare_record_write(env, dto, &tti->tti_buf, -1, th);
+ if (rc)
+ return rc;
+ } else {
+ dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
+ tti_buf_lcd(tti);
+ tti->tti_off = tsi->tsi_exp->exp_target_data.ted_lr_off;
+ rc = dt_declare_record_write(env, dto, &tti->tti_buf,
+ tti->tti_off, th);
+ if (rc)
+ return rc;
+ }
+
+ if (tsi->tsi_vbr_obj != NULL &&
+ !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
+ dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
+ rc = dt_declare_version_set(env, dto, th);
+ }
+
+ return rc;
+}
+
+/* Update last_rcvd records with latests transaction data */
+int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
+ void *cookie)
+{
+ struct lu_target *tgt = cookie;
+ struct tgt_session_info *tsi;
+ struct tgt_thread_info *tti = tgt_th_info(env);
+ struct dt_object *obj = NULL;
+ int rc;
+ bool echo_client;
+
+ if (env->le_ses == NULL)
+ return 0;
+
+ tsi = tgt_ses_info(env);
+ /* OFD may start transaction without export assigned */
+ if (tsi->tsi_exp == NULL)
+ return 0;
+
+ echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
+
+ if (tti->tti_has_trans && !echo_client) {
+ if (tti->tti_mult_trans == 0) {
+ CDEBUG(D_HA, "More than one transaction %llu\n",
+ tti->tti_transno);
+ RETURN(0);
+ }
+ /* we need another transno to be assigned */
+ tti->tti_transno = 0;
+ } else if (th->th_result == 0) {
+ tti->tti_has_trans = 1;
+ }
+
+ if (tsi->tsi_vbr_obj != NULL &&
+ !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
+ obj = tsi->tsi_vbr_obj;
+ }
+
+ if (unlikely(echo_client)) /* echo client special case */
+ rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
+ tsi->tsi_exp);
+ else
+ rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
+ tgt_ses_req(tsi));
+ return rc;
+}
+
+int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
+{
+ struct tgt_thread_info *tti = tgt_th_info(env);
+ struct lsd_reply_data *lrd = &tti->tti_lrd;
+ unsigned long reply_data_size;
+ int rc;
+ struct lsd_reply_header *lrh = NULL;
+ struct lsd_client_data *lcd = NULL;
+ struct tg_reply_data *trd = NULL;
+ int idx;
+ loff_t off;
+ struct cfs_hash *hash = NULL;
+ struct obd_export *exp;
+ struct tg_export_data *ted;
+ int reply_data_recovered = 0;
+
+ rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
+ if (rc)
+ GOTO(out, rc);
+ reply_data_size = (unsigned long)tti->tti_attr.la_size;
+
+ OBD_ALLOC_PTR(lrh);
+ if (lrh == NULL)
+ GOTO(out, rc = -ENOMEM);
+
+ if (reply_data_size == 0) {
+ CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
+ tgt_name(tgt));
+ lrh->lrh_magic = LRH_MAGIC;
+ lrh->lrh_header_size = sizeof(struct lsd_reply_header);
+ lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
+ rc = tgt_reply_header_write(env, tgt, lrh);
+ if (rc) {
+ CERROR("%s: error writing %s: rc = %d\n",
+ tgt_name(tgt), REPLY_DATA, rc);
+ GOTO(out, rc);
+ }
+ } else {
+ rc = tgt_reply_header_read(env, tgt, lrh);
+ if (rc) {
+ CERROR("%s: error reading %s: rc = %d\n",
+ tgt_name(tgt), REPLY_DATA, rc);
+ GOTO(out, rc);
+ }
+ if (lrh->lrh_magic != LRH_MAGIC ||
+ lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
+ lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
+ CERROR("%s: invalid header in %s\n",
+ tgt_name(tgt), REPLY_DATA);
+ GOTO(out, rc = -EINVAL);
+ }
+
+ hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
+ if (hash == NULL)
+ GOTO(out, rc = -ENODEV);
+
+ OBD_ALLOC_PTR(lcd);
+ if (lcd == NULL)
+ GOTO(out, rc = -ENOMEM);
+
+ OBD_ALLOC_PTR(trd);
+ if (trd == NULL)
+ GOTO(out, rc = -ENOMEM);
+
+ /* Load reply_data from disk */
+ for (idx = 0, off = sizeof(struct lsd_reply_header);
+ off < reply_data_size;
+ idx++, off += sizeof(struct lsd_reply_data)) {
+ rc = tgt_reply_data_read(env, tgt, lrd, off);
+ if (rc) {
+ CERROR("%s: error reading %s: rc = %d\n",
+ tgt_name(tgt), REPLY_DATA, rc);
+ GOTO(out, rc);
+ }
+
+ exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
+ if (exp == NULL) {
+ /* old reply data from a disconnected client */
+ continue;
+ }
+ ted = &exp->exp_target_data;
+ mutex_lock(&ted->ted_lcd_lock);
+
+ /* create in-memory reply_data and link it to
+ * target export's reply list */
+ rc = tgt_set_reply_slot(tgt, idx);
+ if (rc != 0) {
+ mutex_unlock(&ted->ted_lcd_lock);
+ GOTO(out, rc);
+ }
+ trd->trd_reply = *lrd;
+ trd->trd_pre_versions[0] = 0;
+ trd->trd_pre_versions[1] = 0;
+ trd->trd_pre_versions[2] = 0;
+ trd->trd_pre_versions[3] = 0;
+ trd->trd_index = idx;
+ trd->trd_tag = 0;
+ list_add(&trd->trd_list, &ted->ted_reply_list);
+ ted->ted_reply_cnt++;
+ if (ted->ted_reply_cnt > ted->ted_reply_max)
+ ted->ted_reply_max = ted->ted_reply_cnt;
+
+ CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
+ "transno %llu, client gen %u, slot idx %d\n",
+ tgt_name(tgt), trd, lrd->lrd_xid,
+ lrd->lrd_transno, lrd->lrd_client_gen,
+ trd->trd_index);
+
+ /* update export last committed transation */
+ exp->exp_last_committed = max(exp->exp_last_committed,
+ lrd->lrd_transno);
+
+ mutex_unlock(&ted->ted_lcd_lock);
+ class_export_put(exp);
+
+ /* update target last committed transaction */
+ spin_lock(&tgt->lut_translock);
+ tgt->lut_last_transno = max(tgt->lut_last_transno,
+ lrd->lrd_transno);
+ spin_unlock(&tgt->lut_translock);
+
+ reply_data_recovered++;
+
+ OBD_ALLOC_PTR(trd);
+ if (trd == NULL)
+ GOTO(out, rc = -ENOMEM);
+ }
+ CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
+ tgt_name(tgt), reply_data_recovered);
+ }
+
+ spin_lock(&tgt->lut_translock);
+ /* obd_last_committed is used for compatibility
+ * with other lustre recovery code */
+ tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
+ spin_unlock(&tgt->lut_translock);
+
+ rc = 0;
+
+out:
+ if (hash != NULL)
+ cfs_hash_putref(hash);
+ if (lcd != NULL)
+ OBD_FREE_PTR(lcd);
+ if (trd != NULL)
+ OBD_FREE_PTR(trd);
+ if (lrh != NULL)
+ OBD_FREE_PTR(lrh);
+ return rc;
+}
+
+static int tgt_check_lookup_req(struct ptlrpc_request *req, int lookup,
+ struct tg_reply_data *trd)
+{
+ struct tg_export_data *ted = &req->rq_export->exp_target_data;
+ struct lu_target *lut = class_exp2tgt(req->rq_export);
+ __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
+ int rc = 0;
+ struct tg_reply_data *reply;
+ bool check_increasing;
+
+ if (tag == 0)
+ return 0;
+
+ check_increasing = tgt_is_increasing_xid_client(req->rq_export) &&
+ !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
+ if (!lookup && !check_increasing)
+ return 0;
+
+ list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
+ if (lookup && reply->trd_reply.lrd_xid == req->rq_xid) {
+ rc = 1;
+ if (trd != NULL)
+ *trd = *reply;
+ break;
+ } else if (check_increasing && reply->trd_tag == tag &&
+ reply->trd_reply.lrd_xid > req->rq_xid) {
+ rc = -EPROTO;
+ CERROR("%s: busy tag=%u req_xid=%llu, trd=%p: xid=%llu transno=%llu client_gen=%u slot_idx=%d: rc = %d\n",
+ tgt_name(lut), tag, req->rq_xid, trd,
+ reply->trd_reply.lrd_xid,
+ reply->trd_reply.lrd_transno,
+ reply->trd_reply.lrd_client_gen,
+ reply->trd_index, rc);
+ break;
+ }
+ }
+
+ return rc;
+}
+
+/* Look for a reply data matching specified request @req
+ * A copy is returned in @trd if the pointer is not NULL
+ */
+int tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
+{
+ struct tg_export_data *ted = &req->rq_export->exp_target_data;
+ int found = 0;
+ bool not_replay = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
+
+ mutex_lock(&ted->ted_lcd_lock);
+ if (not_replay && req->rq_xid <= req->rq_export->exp_last_xid) {
+ /* A check for the last_xid is needed here in case there is
+ * no reply data is left in the list. It may happen if another
+ * RPC on another slot increased the last_xid between our
+ * process_req_last_xid & tgt_lookup_reply calls */
+ found = -EPROTO;
+ } else {
+ found = tgt_check_lookup_req(req, 1, trd);
+ }
+ mutex_unlock(&ted->ted_lcd_lock);
+
+ CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d last_xid %llu\n",
+ tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid, found,
+ req->rq_export->exp_last_xid);
+
+ return found;
+}
+EXPORT_SYMBOL(tgt_lookup_reply);
+
+int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
+{
+ struct tg_export_data *ted = &exp->exp_target_data;
+ struct lu_target *lut = class_exp2tgt(exp);
+ struct tg_reply_data *trd, *tmp;
+
+
+ list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
+ if (trd->trd_reply.lrd_xid > rcvd_xid)
+ continue;
+ ted->ted_release_xid++;
+ tgt_release_reply_data(lut, ted, trd);
+ }
+
+ return 0;
+}
+
+int tgt_handle_tag(struct ptlrpc_request *req)
+{
+ return tgt_check_lookup_req(req, 0, NULL);
+}
+