// For LDTs only:
void
repl_write_ldt_make_message(msg* m, as_transaction* tr, uint8_t** p_pickled_buf,
size_t pickled_sz, as_rec_props* p_pickled_rec_props, bool is_subrec)
{
as_namespace* ns = tr->rsv.ns;
msg_set_uint32(m, RW_FIELD_OP, RW_OP_WRITE);
msg_set_buf(m, RW_FIELD_NAMESPACE, (uint8_t*)ns->name, strlen(ns->name),
MSG_SET_COPY);
msg_set_uint32(m, RW_FIELD_NS_ID, ns->id);
msg_set_buf(m, RW_FIELD_DIGEST, (void*)&tr->keyd, sizeof(cf_digest),
MSG_SET_COPY);
msg_set_uint64(m, RW_FIELD_CLUSTER_KEY, tr->rsv.cluster_key);
msg_set_uint32(m, RW_FIELD_GENERATION, tr->generation);
msg_set_uint32(m, RW_FIELD_VOID_TIME, tr->void_time);
msg_set_uint64(m, RW_FIELD_LAST_UPDATE_TIME, tr->last_update_time);
// TODO - do we really get here if ldt_enabled is false?
if (ns->ldt_enabled && ! is_subrec) {
msg_set_buf(m, RW_FIELD_VINFOSET, (uint8_t*)&tr->rsv.p->version_info,
sizeof(as_partition_vinfo), MSG_SET_COPY);
if (tr->rsv.p->current_outgoing_ldt_version != 0) {
msg_set_uint64(m, RW_FIELD_LDT_VERSION,
tr->rsv.p->current_outgoing_ldt_version);
}
}
uint32_t info = pack_info_bits(tr, true);
if (*p_pickled_buf) {
bool is_sub;
bool is_parent;
as_ldt_get_property(p_pickled_rec_props, &is_parent, &is_sub);
info |= pack_ldt_info_bits(tr, is_parent, is_sub);
msg_set_buf(m, RW_FIELD_RECORD, (void*)*p_pickled_buf, pickled_sz,
MSG_SET_HANDOFF_MALLOC);
*p_pickled_buf = NULL;
if (p_pickled_rec_props && p_pickled_rec_props->p_data) {
msg_set_buf(m, RW_FIELD_REC_PROPS, p_pickled_rec_props->p_data,
p_pickled_rec_props->size, MSG_SET_HANDOFF_MALLOC);
as_rec_props_clear(p_pickled_rec_props);
}
}
else {
msg_set_buf(m, RW_FIELD_AS_MSG, (void*)tr->msgp,
as_proto_size_get(&tr->msgp->proto), MSG_SET_COPY);
info |= pack_ldt_info_bits(tr, false, is_subrec);
}
msg_set_uint32(m, RW_FIELD_INFO, info);
}
// Send a redirection message - consumes the message.
int
as_proxy_send_redirect(cf_node dst, msg *m, cf_node rdst)
{
int rv;
uint32_t tid;
msg_get_uint32(m, PROXY_FIELD_TID, &tid);
msg_reset(m);
msg_set_uint32(m, PROXY_FIELD_OP, PROXY_OP_REDIRECT);
msg_set_uint32(m, PROXY_FIELD_TID, tid);
msg_set_uint64(m, PROXY_FIELD_REDIRECT, rdst);
if (0 != (rv = as_fabric_send(dst, m, AS_FABRIC_PRIORITY_MEDIUM))) {
cf_debug(AS_PROXY, "sending redirection failed: fabric send error %d", rv);
as_fabric_msg_put(m);
}
return 0;
} // end as_proxy_send_redirect()
int
as_proxy_shipop(cf_node dst, write_request *wr)
{
as_partition_id pid = as_partition_getid(wr->keyd);
if (dst == 0) {
cf_crash(AS_PROXY, "the destination should never be zero");
}
// Create a fabric message, fill it out.
msg *m = as_fabric_msg_get(M_TYPE_PROXY);
if (!m) {
return -1;
}
uint32_t tid = cf_atomic32_incr(&g_proxy_tid);
msg_set_uint32(m, PROXY_FIELD_OP, PROXY_OP_REQUEST);
msg_set_uint32(m, PROXY_FIELD_TID, tid);
msg_set_buf(m, PROXY_FIELD_DIGEST, (void *) &wr->keyd, sizeof(cf_digest), MSG_SET_COPY);
msg_set_buf(m, PROXY_FIELD_AS_PROTO, (void *) wr->msgp, as_proto_size_get(&wr->msgp->proto), MSG_SET_HANDOFF_MALLOC);
msg_set_uint64(m, PROXY_FIELD_CLUSTER_KEY, as_paxos_get_cluster_key());
msg_set_uint32(m, PROXY_FIELD_TIMEOUT_MS, wr->msgp->msg.transaction_ttl);
wr->msgp = 0;
// If it is shipped op.
uint32_t info = 0;
info |= PROXY_INFO_SHIPPED_OP;
msg_set_uint32(m, PROXY_FIELD_INFO, info);
cf_detail_digest(AS_PROXY, &wr->keyd, "SHIPPED_OP %s->WINNER msg %p Proxy Sent to %"PRIx64" %p tid(%d)",
wr->proxy_msg ? "NONORIG" : "ORIG", m, dst, wr, tid);
// Fill out a retransmit structure, insert into the retransmit hash.
msg_incr_ref(m);
proxy_request pr;
pr.start_time = wr->start_time;
pr.end_time = (wr->end_time != 0) ? wr->end_time : pr.start_time + g_config.transaction_max_ns;
cf_rc_reserve(wr);
pr.wr = wr;
pr.fab_msg = m;
pr.xmit_ms = cf_getms() + g_config.transaction_retry_ms;
pr.retry_interval_ms = g_config.transaction_retry_ms;
pr.dest = dst;
pr.pid = pid;
pr.fd_h = NULL;
pr.batch_shared = NULL;
pr.batch_index = 0;
if (0 != shash_put(g_proxy_hash, &tid, &pr)) {
cf_info(AS_PROXY, " shash_put failed, need cleanup code");
return -1;
}
// Send to the remote node.
int rv = as_fabric_send(dst, m, AS_FABRIC_PRIORITY_MEDIUM);
if (rv != 0) {
cf_detail(AS_PROXY, "SHIPPED_OP ORIG [Digest %"PRIx64"] Failed with %d", *(uint64_t *)&wr->keyd, rv);
as_fabric_msg_put(m);
}
wr->shipped_op_initiator = true;
cf_atomic_int_incr(&g_config.ldt_proxy_initiate);
return 0;
}
// Make a request to another node.
//
// Note: there's a cheat here. 'as_msg' is used in a raw form, and includes
// structured data (version - type - nfields - sz ...) which should be made more
// wire-protocol-friendly.
int
as_proxy_divert(cf_node dst, as_transaction *tr, as_namespace *ns, uint64_t cluster_key)
{
cf_detail(AS_PROXY, "proxy divert");
cf_atomic_int_incr(&g_config.stat_proxy_reqs);
if (tr->msgp && (tr->msgp->msg.info1 & AS_MSG_INFO1_XDR)) {
cf_atomic_int_incr(&g_config.stat_proxy_reqs_xdr);
}
as_partition_id pid = as_partition_getid(tr->keyd);
if (dst == 0) {
// Get the list of replicas.
dst = as_partition_getreplica_read(ns, pid);
}
// Create a fabric message, fill it out.
msg *m = as_fabric_msg_get(M_TYPE_PROXY);
if (!m) {
return -1;
}
uint32_t tid = cf_atomic32_incr(&g_proxy_tid);
msg_set_uint32(m, PROXY_FIELD_OP, PROXY_OP_REQUEST);
msg_set_uint32(m, PROXY_FIELD_TID, tid);
msg_set_buf(m, PROXY_FIELD_DIGEST, (void *) &tr->keyd, sizeof(cf_digest), MSG_SET_COPY);
msg_set_type msettype = tr->batch_shared ? MSG_SET_COPY : MSG_SET_HANDOFF_MALLOC;
msg_set_buf(m, PROXY_FIELD_AS_PROTO, (void *) tr->msgp, as_proto_size_get(&tr->msgp->proto), msettype);
msg_set_uint64(m, PROXY_FIELD_CLUSTER_KEY, cluster_key);
msg_set_uint32(m, PROXY_FIELD_TIMEOUT_MS, tr->msgp->msg.transaction_ttl);
tr->msgp = 0;
cf_debug_digest(AS_PROXY, &tr->keyd, "proxy_divert: fab_msg %p dst %"PRIx64, m, dst);
// Fill out a retransmit structure, insert into the retransmit hash.
msg_incr_ref(m);
proxy_request pr;
pr.start_time = tr->start_time;
pr.end_time = (tr->end_time != 0) ? tr->end_time : pr.start_time + g_config.transaction_max_ns;
pr.fd_h = tr->proto_fd_h;
tr->proto_fd_h = 0;
pr.fab_msg = m;
pr.xmit_ms = cf_getms() + g_config.transaction_retry_ms;
pr.retry_interval_ms = g_config.transaction_retry_ms;
pr.dest = dst;
pr.pid = pid;
pr.ns = ns;
pr.wr = NULL;
pr.batch_shared = tr->batch_shared;
pr.batch_index = tr->batch_index;
if (0 != shash_put(g_proxy_hash, &tid, &pr)) {
cf_debug(AS_PROXY, " shash_put failed, need cleanup code");
return -1;
}
// Send to the remote node.
int rv = as_fabric_send(dst, m, AS_FABRIC_PRIORITY_MEDIUM);
if (rv != 0) {
cf_debug(AS_PROXY, "as_proxy_divert: returned error %d", rv);
as_fabric_msg_put(m);
}
cf_atomic_int_incr(&g_config.proxy_initiate);
return 0;
}
bool
repl_write_make_message(rw_request* rw, as_transaction* tr)
{
if (rw->dest_msg) {
msg_reset(rw->dest_msg);
}
else if (! (rw->dest_msg = as_fabric_msg_get(M_TYPE_RW))) {
return false;
}
as_namespace* ns = tr->rsv.ns;
msg* m = rw->dest_msg;
msg_set_uint32(m, RW_FIELD_OP, rw->is_multiop ? RW_OP_MULTI : RW_OP_WRITE);
msg_set_buf(m, RW_FIELD_NAMESPACE, (uint8_t*)ns->name, strlen(ns->name),
MSG_SET_COPY);
msg_set_uint32(m, RW_FIELD_NS_ID, ns->id);
msg_set_buf(m, RW_FIELD_DIGEST, (void*)&tr->keyd, sizeof(cf_digest),
MSG_SET_COPY);
msg_set_uint64(m, RW_FIELD_CLUSTER_KEY, tr->rsv.cluster_key);
msg_set_uint32(m, RW_FIELD_TID, rw->tid);
msg_set_uint32(m, RW_FIELD_GENERATION, tr->generation);
msg_set_uint32(m, RW_FIELD_VOID_TIME, tr->void_time);
msg_set_uint64(m, RW_FIELD_LAST_UPDATE_TIME, tr->last_update_time);
// TODO - do we really intend to send this if the record is non-LDT?
if (ns->ldt_enabled) {
msg_set_buf(m, RW_FIELD_VINFOSET, (uint8_t*)&tr->rsv.p->version_info,
sizeof(as_partition_vinfo), MSG_SET_COPY);
if (tr->rsv.p->current_outgoing_ldt_version != 0) {
msg_set_uint64(m, RW_FIELD_LDT_VERSION,
tr->rsv.p->current_outgoing_ldt_version);
}
}
if (rw->is_multiop) {
msg_set_uint32(m, RW_FIELD_INFO, RW_INFO_LDT);
msg_set_buf(m, RW_FIELD_MULTIOP, (void*)rw->pickled_buf, rw->pickled_sz,
MSG_SET_HANDOFF_MALLOC);
// Make sure destructor doesn't free this.
rw->pickled_buf = NULL;
return true;
}
uint32_t info = pack_info_bits(tr, rw->has_udf);
if (rw->pickled_buf) {
// Replica writes.
bool is_sub;
bool is_parent;
as_ldt_get_property(&rw->pickled_rec_props, &is_parent, &is_sub);
info |= pack_ldt_info_bits(tr, is_parent, is_sub);
msg_set_buf(m, RW_FIELD_RECORD, (void*)rw->pickled_buf, rw->pickled_sz,
MSG_SET_HANDOFF_MALLOC);
// Make sure destructor doesn't free this.
rw->pickled_buf = NULL;
if (rw->pickled_rec_props.p_data) {
msg_set_buf(m, RW_FIELD_REC_PROPS, rw->pickled_rec_props.p_data,
rw->pickled_rec_props.size, MSG_SET_HANDOFF_MALLOC);
// Make sure destructor doesn't free the data.
as_rec_props_clear(&rw->pickled_rec_props);
}
}
else {
// Replica deletes.
msg_set_buf(m, RW_FIELD_AS_MSG, (void*)tr->msgp,
as_proto_size_get(&tr->msgp->proto), MSG_SET_COPY);
info |= pack_ldt_info_bits(tr, false, false);
}
msg_set_uint32(m, RW_FIELD_INFO, info);
return true;
}
