transaction_status
as_read_start(as_transaction* tr)
{
BENCHMARK_START(tr, read, FROM_CLIENT);
BENCHMARK_START(tr, batch_sub, FROM_BATCH);
if (! repl_ping_check(tr)) {
send_read_response(tr, NULL, NULL, 0, NULL);
return TRANS_DONE_ERROR;
}
transaction_status status;
bool must_duplicate_resolve = read_must_duplicate_resolve(tr);
bool must_ping = read_must_ping(tr);
if (! must_duplicate_resolve && ! must_ping) {
// No network hops needed, try reading.
if ((status = read_local(tr)) != TRANS_IN_PROGRESS) {
return status;
}
// else - must try again under hash.
}
// else - there are duplicates, and we're configured to resolve them, or
// we're required to ping replicas.
// Create rw_request and add to hash.
rw_request_hkey hkey = { tr->rsv.ns->id, tr->keyd };
rw_request* rw = rw_request_create(&tr->keyd);
// If rw_request isn't inserted in hash, transaction is finished.
if ((status = rw_request_hash_insert(&hkey, rw, tr)) != TRANS_IN_PROGRESS) {
rw_request_release(rw);
if (status != TRANS_WAITING) {
send_read_response(tr, NULL, NULL, 0, NULL);
}
return status;
}
// else - rw_request is now in hash, continue...
if (must_duplicate_resolve) {
start_read_dup_res(rw, tr);
// Started duplicate resolution.
return TRANS_IN_PROGRESS;
}
if (must_ping) {
// Set up the nodes to which we'll ping.
rw->n_dest_nodes = as_partition_get_other_replicas(tr->rsv.p,
rw->dest_nodes);
if (insufficient_replica_destinations(tr->rsv.ns, rw->n_dest_nodes)) {
rw_request_hash_delete(&hkey, rw);
tr->result_code = AS_ERR_UNAVAILABLE;
send_read_response(tr, NULL, NULL, 0, NULL);
return TRANS_DONE_ERROR;
}
start_repl_ping(rw, tr);
// Started replica ping.
return TRANS_IN_PROGRESS;
}
// Trying again under hash.
status = read_local(tr);
cf_assert(status != TRANS_IN_PROGRESS, AS_RW, "read in-progress");
rw_request_hash_delete(&hkey, rw);
return status;
}
void
repl_write_handle_ack(cf_node node, msg* m)
{
uint32_t ns_id;
if (msg_get_uint32(m, RW_FIELD_NS_ID, &ns_id) != 0) {
cf_warning(AS_RW, "repl-write ack: no ns-id");
as_fabric_msg_put(m);
return;
}
cf_digest* keyd;
size_t sz;
if (msg_get_buf(m, RW_FIELD_DIGEST, (uint8_t**)&keyd, &sz,
MSG_GET_DIRECT) != 0) {
cf_warning(AS_RW, "repl-write ack: no digest");
as_fabric_msg_put(m);
return;
}
uint32_t tid;
if (msg_get_uint32(m, RW_FIELD_TID, &tid) != 0) {
cf_warning(AS_RW, "repl-write ack: no tid");
as_fabric_msg_put(m);
return;
}
// TODO - result_code is currently ignored! What should we do with it?
// Note - CLUSTER_KEY_MISMATCH not special, can't re-queue transaction.
uint32_t result_code;
if (msg_get_uint32(m, RW_FIELD_RESULT, &result_code) != 0) {
cf_warning(AS_RW, "repl-write ack: no result_code");
as_fabric_msg_put(m);
return;
}
rw_request_hkey hkey = { ns_id, *keyd };
rw_request* rw = rw_request_hash_get(&hkey);
if (! rw) {
// Extra ack, after rw_request is already gone.
as_fabric_msg_put(m);
return;
}
pthread_mutex_lock(&rw->lock);
if (rw->tid != tid) {
// Extra ack, rw_request is that of newer transaction for same digest.
pthread_mutex_unlock(&rw->lock);
rw_request_release(rw);
as_fabric_msg_put(m);
return;
}
int i;
for (i = 0; i < rw->n_dest_nodes; i++) {
if (rw->dest_nodes[i] != node) {
continue;
}
if (rw->dest_complete[i]) {
// Extra ack for this replica write.
pthread_mutex_unlock(&rw->lock);
rw_request_release(rw);
as_fabric_msg_put(m);
return;
}
rw->dest_complete[i] = true;
break;
}
if (i == rw->n_dest_nodes) {
cf_warning(AS_RW, "repl-write ack: from non-dest node %lx", node);
pthread_mutex_unlock(&rw->lock);
rw_request_release(rw);
as_fabric_msg_put(m);
return;
}
for (int j = 0; j < rw->n_dest_nodes; j++) {
if (! rw->dest_complete[j]) {
// Still haven't heard from all duplicates.
pthread_mutex_unlock(&rw->lock);
rw_request_release(rw);
as_fabric_msg_put(m);
return;
}
}
if (! rw->from.any && rw->origin != FROM_NSUP &&
! rw->respond_client_on_master_completion) {
// Lost race against timeout in retransmit thread.
pthread_mutex_unlock(&rw->lock);
rw_request_release(rw);
as_fabric_msg_put(m);
return;
}
if (! rw->respond_client_on_master_completion) {
rw->repl_write_cb(rw);
}
pthread_mutex_unlock(&rw->lock);
rw_request_hash_delete(&hkey, rw);
rw_request_release(rw);
as_fabric_msg_put(m);
}
