{
struct ptlrpc_request *req, *n;
struct list_head clean_list;
- ENTRY;
INIT_LIST_HEAD(&clean_list);
spin_lock(&obd->obd_dev_lock);
obd->obd_replayed_requests++;
}
-/**
- * Update last_rcvd of the update
- *
- * Because update recovery might update the last_rcvd by updates, i.e.
- * it will not update the last_rcvd information in memory, so we need
- * refresh these information in memory after update recovery.
- *
- * \param[in] obd obd_device under recoverying.
- * \param[in] dtrq the update replay requests being replayed.
- */
-static void target_update_lcd(struct lu_env *env, struct lu_target *lut,
- struct distribute_txn_replay_req *dtrq)
-{
- struct obd_device *obd = lut->lut_obd;
- struct obd_export *export;
- struct tg_export_data *ted;
- struct distribute_txn_replay_req_sub *dtrqs;
- struct seq_server_site *site;
- struct update_records *ur;
- const struct lu_fid *fid;
- struct update_ops *ops;
- struct update_params *params;
- struct update_op *op;
- __u32 mdt_index;
- unsigned int i;
- struct lsd_client_data *lcd = NULL;
-
- /* if Updates has been executed(committed) on the recovery target,
- * i.e. the updates is not being executed on the target, so we do
- * not need update it in memory */
- site = lu_site2seq(obd->obd_lu_dev->ld_site);
- mdt_index = site->ss_node_id;
- dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
- if (dtrqs != NULL)
- return;
-
- if (dtrq->dtrq_lur == NULL)
- return;
-
- /* Find the update last_rcvd record */
- fid = lu_object_fid(&lut->lut_last_rcvd->do_lu);
- ur = &dtrq->dtrq_lur->lur_update_rec;
- ops = &ur->ur_ops;
- params = update_records_get_params(ur);
- for (i = 0, op = &ops->uops_op[0]; i < ur->ur_update_count;
- i++, op = update_op_next_op(op)) {
- __u64 pos;
- __u16 size;
- void *buf;
-
- if (!lu_fid_eq(&op->uop_fid, fid))
- continue;
-
- if (op->uop_type != OUT_WRITE)
- continue;
-
- buf = update_params_get_param_buf(params, op->uop_params_off[1],
- ur->ur_param_count, NULL);
- if (buf == NULL)
- continue;
-
- pos = le64_to_cpu(*(__u64 *)buf);
- if (pos == 0)
- continue;
-
- buf = update_params_get_param_buf(params, op->uop_params_off[0],
- ur->ur_param_count, &size);
- if (buf == NULL)
- continue;
-
- if (size != sizeof(*lcd))
- continue;
- lcd = buf;
- }
-
- if (lcd == NULL || lcd->lcd_uuid[0] == '\0')
- return;
-
- /* locate the export then update the exp_target_data if needed */
- export = cfs_hash_lookup(obd->obd_uuid_hash, lcd->lcd_uuid);
- if (export == NULL)
- return;
-
- ted = &export->exp_target_data;
- if (lcd->lcd_last_xid > ted->ted_lcd->lcd_last_xid) {
- CDEBUG(D_HA, "%s update xid from "LPU64" to "LPU64"\n",
- lut->lut_obd->obd_name, ted->ted_lcd->lcd_last_xid,
- lcd->lcd_last_xid);
- ted->ted_lcd->lcd_last_xid = lcd->lcd_last_xid;
- ted->ted_lcd->lcd_last_result = lcd->lcd_last_result;
- }
- class_export_put(export);
-}
-
static void replay_request_or_update(struct lu_env *env,
struct lu_target *lut,
struct target_recovery_data *trd,
if (transno > obd->obd_next_recovery_transno)
obd->obd_next_recovery_transno = transno;
spin_unlock(&obd->obd_recovery_task_lock);
- target_update_lcd(env, lut, dtrq);
dtrq_destroy(dtrq);
} else {
spin_unlock(&obd->obd_recovery_task_lock);
}
EXPORT_SYMBOL(tgt_client_del);
+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 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);
+ 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, i);
+ RETURN(0);
+}
+EXPORT_SYMBOL(tgt_add_reply_data);
+
/*
* last_rcvd & last_committed update callbacks
*/
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;
ENTRY;
- ted = &req->rq_export->exp_target_data;
- lw_client = exp_connect_flags(req->rq_export) & OBD_CONNECT_LIGHTWEIGHT;
+ 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);
- tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
+ 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) {
CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
tti->tti_transno, tgt->lut_obd->obd_last_committed);
- req->rq_transno = tti->tti_transno;
- lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
+ 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, req->rq_export,
- tti->tti_transno);
+ 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
* slot, update server data with latest transno then */
if (ted->ted_lcd == NULL) {
CWARN("commit transaction for disconnected client %s: rc %d\n",
- req->rq_export->exp_client_uuid.uuid, rc);
+ exp->exp_client_uuid.uuid, rc);
GOTO(srv_update, rc = 0);
}
/* Target that supports multiple reply data */
- if (tgt_is_multimodrpcs_client(req->rq_export)) {
+ if (tgt_is_multimodrpcs_client(exp)) {
struct tg_reply_data *trd;
struct lsd_reply_data *lrd;
__u64 *pre_versions;
- int i;
- loff_t off;
+ bool write_update;
OBD_ALLOC_PTR(trd);
if (unlikely(trd == NULL))
GOTO(srv_update, rc = -ENOMEM);
- /* update export last transno */
- mutex_lock(&ted->ted_lcd_lock);
- if (tti->tti_transno > ted->ted_lcd->lcd_last_transno)
- ted->ted_lcd->lcd_last_transno = tti->tti_transno;
- mutex_unlock(&ted->ted_lcd_lock);
-
/* fill reply data information */
lrd = &trd->trd_reply;
lrd->lrd_transno = tti->tti_transno;
- lrd->lrd_xid = req->rq_xid;
- lrd->lrd_result = th->th_result;
+ 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;
- lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
- trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
- pre_versions = lustre_msg_get_versions(req->rq_repmsg);
if (pre_versions) {
trd->trd_pre_versions[0] = pre_versions[0];
trd->trd_pre_versions[1] = pre_versions[1];
trd->trd_pre_versions[3] = pre_versions[3];
}
- /* 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);
- GOTO(srv_update, rc = i);
- }
- trd->trd_index = i;
-
- /* 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);
- 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, i);
-
- GOTO(srv_update, rc = 0);
+ rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
+ GOTO(srv_update, rc);
}
+ /* Enough for update replay, let's return */
+ if (req == NULL)
+ GOTO(srv_update, 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) {
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 (tsi->tsi_exp == NULL)
return 0;
- echo_client = (tgt_ses_req(tsi) == NULL);
+ echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
if (tti->tti_has_trans && !echo_client) {
if (tti->tti_mult_trans == 0) {
loff_t off;
struct cfs_hash *hash = NULL;
struct obd_export *exp;
- struct obd_export *tmp;
struct tg_export_data *ted;
int reply_data_recovered = 0;
}
CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
tgt_name(tgt), reply_data_recovered);
-
- /* delete entries from client_generation<->export hash */
- spin_lock(&tgt->lut_obd->obd_dev_lock);
- list_for_each_entry_safe(exp, tmp,
- &tgt->lut_obd->obd_exports,
- exp_obd_chain) {
- struct tg_export_data *ted = &exp->exp_target_data;
-
- if (!hlist_unhashed(&exp->exp_gen_hash))
- cfs_hash_del(hash,
- &ted->ted_lcd->lcd_generation,
- &exp->exp_gen_hash);
- }
- spin_unlock(&tgt->lut_obd->obd_dev_lock);
}
spin_lock(&tgt->lut_translock);
return rc;
}
+struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
+ __u64 xid)
+{
+ struct tg_reply_data *found = NULL;
+ struct tg_reply_data *reply;
+
+ mutex_lock(&ted->ted_lcd_lock);
+ list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
+ if (reply->trd_reply.lrd_xid == xid) {
+ found = reply;
+ break;
+ }
+ }
+ mutex_unlock(&ted->ted_lcd_lock);
+ return found;
+}
+EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
+
/* Look for a reply data matching specified request @req
* A copy is returned in @trd if the pointer is not NULL
*/
bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
{
struct tg_export_data *ted = &req->rq_export->exp_target_data;
- struct tg_reply_data *reply, *tmp;
+ struct tg_reply_data *reply;
bool found = false;
- mutex_lock(&ted->ted_lcd_lock);
- list_for_each_entry_safe(reply, tmp, &ted->ted_reply_list, trd_list) {
- if (reply->trd_reply.lrd_xid == req->rq_xid) {
- found = true;
- break;
- }
+ reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
+ if (reply != NULL) {
+ found = true;
+ if (trd != NULL)
+ *trd = *reply;
}
- if (found && trd != NULL)
- *trd = *reply;
- mutex_unlock(&ted->ted_lcd_lock);
CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
}
/**
+ * Update session information
+ *
+ * Update session information so tgt_txn_stop_cb()->tgt_last_rcvd_update()
+ * can be called correctly during update replay.
+ *
+ * \param[in] env execution environment.
+ * \param[in] tdtd distribute data structure of the recovering tgt.
+ * \param[in] th thandle of this update replay.
+ * \param[in] master_th master sub thandle.
+ * \param[in] ta_arg the tx arg structure to hold the update for updating
+ * reply data.
+ */
+static void update_recovery_update_ses(struct lu_env *env,
+ struct target_distribute_txn_data *tdtd,
+ struct thandle *th,
+ struct thandle *master_th,
+ struct tx_arg *ta_arg)
+{
+ struct tgt_session_info *tsi;
+ struct lu_target *lut = tdtd->tdtd_lut;
+ struct obd_export *export;
+ struct cfs_hash *hash;
+ struct top_thandle *top_th;
+ struct lsd_reply_data *lrd;
+ size_t size;
+
+ tsi = tgt_ses_info(env);
+ if (tsi->tsi_exp != NULL)
+ return;
+
+ size = ta_arg->u.write.buf.lb_len;
+ lrd = ta_arg->u.write.buf.lb_buf;
+ if (size != sizeof(*lrd) || lrd == NULL)
+ return;
+
+ lrd->lrd_transno = le64_to_cpu(lrd->lrd_transno);
+ lrd->lrd_xid = le64_to_cpu(lrd->lrd_xid);
+ lrd->lrd_data = le64_to_cpu(lrd->lrd_data);
+ lrd->lrd_result = le32_to_cpu(lrd->lrd_result);
+ lrd->lrd_client_gen = le32_to_cpu(lrd->lrd_client_gen);
+
+ if (lrd->lrd_transno != tgt_th_info(env)->tti_transno)
+ return;
+
+ hash = cfs_hash_getref(lut->lut_obd->obd_gen_hash);
+ if (hash == NULL)
+ return;
+
+ export = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
+ if (export == NULL) {
+ cfs_hash_putref(hash);
+ return;
+ }
+
+ tsi->tsi_exp = export;
+ tsi->tsi_xid = lrd->lrd_xid;
+ tsi->tsi_opdata = lrd->lrd_data;
+ tsi->tsi_result = lrd->lrd_result;
+ tsi->tsi_client_gen = lrd->lrd_client_gen;
+ top_th = container_of(th, struct top_thandle, tt_super);
+ top_th->tt_master_sub_thandle = master_th;
+ cfs_hash_putref(hash);
+}
+
+/**
* Execute updates in the update replay records
*
* Declare distribute txn replay by update records and add the updates
if (rc < 0)
GOTO(stop_trans, rc);
+ th->th_dev = tdtd->tdtd_dt;
ta->ta_handle = th;
/* check if the distribute transaction has been committed */
if (rc < 0)
GOTO(stop_trans, rc);
- /* If no updates are needed to be replayed, then
- * mark this records as committed, so commit thread
- * distribute_txn_commit_thread() will delete the
- * record */
+ /* If no updates are needed to be replayed, then mark this records as
+ * committed, so commit thread distribute_txn_commit_thread() will
+ * delete the record */
if (ta->ta_argno == 0)
tmt->tmt_committed = 1;
LASSERT(tmt->tmt_committed == 0);
sub_dt = lu2dt_dev(dt_obj->do_lu.lo_dev);
st = lookup_sub_thandle(tmt, sub_dt);
+
LASSERT(st != NULL);
LASSERT(st->st_sub_th != NULL);
rc = ta->ta_args[i]->exec_fn(env, st->st_sub_th,
ta->ta_args[i]);
+
+ /* If the update is to update the reply data, then
+ * we need set the session information, so
+ * tgt_last_rcvd_update() can be called correctly */
+ if (rc == 0 && dt_obj == tdtd->tdtd_lut->lut_reply_data)
+ update_recovery_update_ses(env, tdtd, th,
+ st->st_sub_th, ta_arg);
+
if (unlikely(rc < 0)) {
CDEBUG(D_HA, "error during execution of #%u from"
" %s:%d: rc = %d\n", i, ta->ta_args[i]->file,
if (tur != NULL)
tur->tur_update_records = NULL;
+
+ if (tgt_ses_info(env)->tsi_exp != NULL) {
+ class_export_put(tgt_ses_info(env)->tsi_exp);
+ tgt_ses_info(env)->tsi_exp = NULL;
+ }
exit_session:
lu_context_exit(&session_env);
lu_context_fini(&session_env);
git://git.whamcloud.com / fs / lustre-release.git / commitdiff