rstatus_t
core_core(void *arg, uint32_t events)
{
rstatus_t status;
struct conn *conn = arg;
struct context *ctx = conn_to_ctx(conn);
log_debug(LOG_VVERB, "event %04"PRIX32" on %s %d", events,
conn_get_type_string(conn), conn->sd);
conn->events = events;
/* error takes precedence over read | write */
if (events & EVENT_ERR) {
if (conn->err && conn->dyn_mode) {
loga("conn err on dnode EVENT_ERR: %d", conn->err);
}
core_error(ctx, conn);
return DN_ERROR;
}
/* read takes precedence over write */
if (events & EVENT_READ) {
status = core_recv(ctx, conn);
if (status != DN_OK || conn->done || conn->err) {
if (conn->dyn_mode) {
if (conn->err) {
loga("conn err on dnode EVENT_READ: %d", conn->err);
core_close(ctx, conn);
return DN_ERROR;
}
return DN_OK;
}
core_close(ctx, conn);
return DN_ERROR;
}
}
if (events & EVENT_WRITE) {
status = core_send(ctx, conn);
if (status != DN_OK || conn->done || conn->err) {
if (conn->dyn_mode) {
if (conn->err) {
loga("conn err on dnode EVENT_WRITE: %d", conn->err);
core_close(ctx, conn);
return DN_ERROR;
}
return DN_OK;
}
core_close(ctx, conn);
return DN_ERROR;
}
}
return DN_OK;
}
static void
dnode_rsp_send_done(struct context *ctx, struct conn *conn, struct msg *rsp)
{
if (log_loggable(LOG_VVERB)) {
log_debug(LOG_VVERB, "dnode_rsp_send_done entering");
}
struct msg *req; /* peer message (request) */
ASSERT(conn->type == CONN_DNODE_PEER_CLIENT);
ASSERT(conn->smsg == NULL);
log_debug(LOG_VERB, "dyn: send done rsp %"PRIu64" on c %d", rsp->id, conn->sd);
req = rsp->peer;
ASSERT(!rsp->request && req->request);
ASSERT(req->selected_rsp == rsp);
ASSERT(req->done && !req->swallow);
log_debug(LOG_DEBUG, "DNODE RSP SENT %s %d dmsg->id %u",
conn_get_type_string(conn),
conn->sd, req->dmsg->id);
/* dequeue request from client outq */
conn_dequeue_outq(ctx, conn, req);
req_put(req);
}
static void
core_timeout(struct context *ctx)
{
for (;;) {
struct msg *msg;
struct conn *conn;
int64_t now, then;
msg = msg_tmo_min();
if (msg == NULL) {
ctx->timeout = ctx->max_timeout;
return;
}
/* skip over req that are in-error or done */
if (msg->error || msg->done) {
msg_tmo_delete(msg);
continue;
}
/*
* timeout expired req and all the outstanding req on the timing
* out server
*/
conn = msg->tmo_rbe.data;
then = msg->tmo_rbe.key;
now = dn_msec_now();
if (now < then) {
int delta = (int)(then - now);
ctx->timeout = MIN(delta, ctx->max_timeout);
return;
}
log_warn("req %"PRIu64" on %s %d timedout, timeout was %d", msg->id,
conn_get_type_string(conn), conn->sd, msg->tmo_rbe.timeout);
msg_tmo_delete(msg);
if (conn->dyn_mode) {
if (conn->type == CONN_DNODE_PEER_SERVER) { //outgoing peer requests
struct server *server = conn->owner;
if (conn->same_dc)
stats_pool_incr(ctx, server->owner, peer_timedout_requests);
else
stats_pool_incr(ctx, server->owner, remote_peer_timedout_requests);
}
} else {
if (conn->type == CONN_SERVER) { //storage server requests
stats_server_incr(ctx, conn->owner, server_dropped_requests);
}
}
conn->err = ETIMEDOUT;
core_close(ctx, conn);
}
}
static rstatus_t
core_send(struct context *ctx, struct conn *conn)
{
rstatus_t status;
status = conn_send(ctx, conn);
if (status != DN_OK) {
log_info("send on %s %d failed: %s", conn_get_type_string(conn),
conn->sd, strerror(errno));
}
return status;
}
static void
core_error(struct context *ctx, struct conn *conn)
{
rstatus_t status;
status = dn_get_soerror(conn->sd);
if (status < 0) {
log_warn("get soerr on %s client %d failed, ignored: %s",
conn_get_type_string(conn), conn->sd, strerror(errno));
}
conn->err = errno;
core_close(ctx, conn);
}
static void
server_ack_err(struct context *ctx, struct conn *conn, struct msg *req)
{
// I want to make sure we do not have swallow here.
//ASSERT_LOG(!req->swallow, "req %d:%d has swallow set??", req->id, req->parent_id);
if ((req->swallow && req->noreply) ||
(req->swallow && (req->consistency == DC_ONE)) ||
(req->swallow && (req->consistency == DC_QUORUM)
&& (!conn->same_dc))) {
log_debug(LOG_INFO, "dyn: close s %d swallow req %"PRIu64" len %"PRIu32
" type %d", conn->sd, req->id, req->mlen, req->type);
req_put(req);
return;
}
struct conn *c_conn = req->owner;
// At other connections, these responses would be swallowed.
ASSERT_LOG((c_conn->type == CONN_CLIENT) ||
(c_conn->type == CONN_DNODE_PEER_CLIENT), "c_conn type %s",
conn_get_type_string(c_conn));
// Create an appropriate response for the request so its propagated up;
// This response gets dropped in rsp_make_error anyways. But since this is
// an error path its ok with the overhead.
struct msg *rsp = msg_get(conn, false, conn->data_store);
if (rsp == NULL) {
log_warn("Could not allocate msg.");
return;
}
req->done = 1;
req->peer = rsp;
rsp->peer = req;
rsp->error = req->error = 1;
rsp->err = req->err = conn->err;
rsp->dyn_error = req->dyn_error = STORAGE_CONNECTION_REFUSE;
rsp->dmsg = NULL;
log_warn("%d:%d <-> %d:%d", req->id, req->parent_id, rsp->id, rsp->parent_id);
log_warn("dyn: close s %d schedule error for req %u:%u "
"len %"PRIu32" type %d from c %d%c %s", conn->sd, req->id, req->parent_id,
req->mlen, req->type, c_conn->sd, conn->err ? ':' : ' ',
conn->err ? strerror(conn->err): " ");
rstatus_t status =
conn_handle_response(c_conn, req->parent_id ? req->parent_id : req->id,
rsp);
IGNORE_RET_VAL(status);
if (req->swallow)
req_put(req);
}
static void
core_close_log(struct conn *conn)
{
char *addrstr;
if ((conn->type == CONN_CLIENT) || (conn->type == CONN_DNODE_PEER_CLIENT)) {
addrstr = dn_unresolve_peer_desc(conn->sd);
} else {
addrstr = dn_unresolve_addr(conn->addr, conn->addrlen);
}
log_debug(LOG_NOTICE, "close %s %d '%s' on event %04"PRIX32" eof %d done "
"%d rb %zu sb %zu%c %s", conn_get_type_string(conn), conn->sd,
addrstr, conn->events, conn->eof, conn->done, conn->recv_bytes,
conn->send_bytes,
conn->err ? ':' : ' ', conn->err ? strerror(conn->err) : "");
}
static rstatus_t
proxy_accept(struct context *ctx, struct conn *p)
{
rstatus_t status;
struct conn *c;
int sd;
ASSERT(p->type == CONN_PROXY);
ASSERT(p->sd > 0);
ASSERT(p->recv_active && p->recv_ready);
for (;;) {
sd = accept(p->sd, NULL, NULL);
if (sd < 0) {
if (errno == EINTR) {
log_warn("accept on %s %d not ready - eintr",
conn_get_type_string(p), p->sd);
continue;
}
if (errno == EAGAIN || errno == EWOULDBLOCK) {
p->recv_ready = 0;
return DN_OK;
}
/*
* FIXME: On EMFILE or ENFILE mask out IN event on the proxy; mask
* it back in when some existing connection gets closed
*/
log_error("accept on %s %d failed: %s",
conn_get_type_string(p), p->sd, strerror(errno));
return DN_ERROR;
}
break;
}
c = conn_get(p->owner, true, p->data_store);
if (c == NULL) {
log_error("get conn for CLIENT %d from %s %d failed: %s", sd,
conn_get_type_string(p), p->sd, strerror(errno));
status = close(sd);
if (status < 0) {
log_error("close c %d failed, ignored: %s", sd, strerror(errno));
}
return DN_ENOMEM;
}
c->sd = sd;
stats_pool_incr(ctx, c->owner, client_connections);
status = dn_set_nonblocking(c->sd);
if (status < 0) {
log_error("set nonblock on %s %d from p %d failed: %s",
conn_get_type_string(c), c->sd, p->sd, strerror(errno));
conn_close(ctx, c);
return status;
}
if (p->family == AF_INET || p->family == AF_INET6) {
status = dn_set_tcpnodelay(c->sd);
if (status < 0) {
log_warn("set tcpnodelay on %s %d from %s %d failed, ignored: %s",
conn_get_type_string(c), c->sd, conn_get_type_string(p),
p->sd, strerror(errno));
}
}
status = event_add_conn(ctx->evb, c);
if (status < 0) {
log_error("event add conn from %s %d failed: %s",conn_get_type_string(p),
p->sd, strerror(errno));
conn_close(ctx, c);
return status;
}
log_notice("accepted %s %d on %s %d from '%s'", conn_get_type_string(c),
c->sd, conn_get_type_string(p), p->sd, dn_unresolve_peer_desc(c->sd));
return DN_OK;
}
struct msg *
rsp_send_next(struct context *ctx, struct conn *conn)
{
rstatus_t status;
struct msg *rsp, *req; /* response and it's peer request */
ASSERT_LOG((conn->type == CONN_DNODE_PEER_CLIENT) ||
(conn->type = CONN_CLIENT), "conn %s", conn_get_type_string(conn));
req = TAILQ_FIRST(&conn->omsg_q);
if (req == NULL || (!req->selected_rsp && !req_done(conn, req))) {
/* nothing is outstanding, initiate close? */
if (req == NULL && conn->eof) {
conn->done = 1;
log_debug(LOG_INFO, "c %d is done", conn->sd);
}
status = event_del_out(ctx->evb, conn);
if (status != DN_OK) {
conn->err = errno;
}
return NULL;
}
rsp = conn->smsg;
if (rsp != NULL) {
ASSERT(!rsp->request);
ASSERT(rsp->peer != NULL);
req = TAILQ_NEXT(rsp->peer, c_tqe);
}
if (req == NULL || !req_done(conn, req)) {
conn->smsg = NULL;
return NULL;
}
ASSERT(req->request && !req->swallow);
if (req_error(conn, req)) {
rsp = rsp_make_error(ctx, conn, req);
if (rsp == NULL) {
conn->err = errno;
return NULL;
}
rsp->peer = req;
req->selected_rsp = rsp;
log_debug(LOG_VERB, "creating new error rsp %p", rsp);
if (conn->dyn_mode) {
stats_pool_incr(ctx, peer_forward_error);
} else {
stats_pool_incr(ctx, forward_error);
}
} else {
rsp = req->selected_rsp;
}
ASSERT(!rsp->request);
conn->smsg = rsp;
if (log_loggable(LOG_VVERB)) {
log_debug(LOG_VVERB, "send next rsp %"PRIu64" on c %d", rsp->id, conn->sd);
}
return rsp;
}
static void
dnode_req_forward(struct context *ctx, struct conn *conn, struct msg *msg)
{
struct server_pool *pool;
uint8_t *key;
uint32_t keylen;
if (log_loggable(LOG_DEBUG)) {
log_debug(LOG_DEBUG, "dnode_req_forward entering ");
}
log_debug(LOG_DEBUG, "DNODE REQ RECEIVED %s %d dmsg->id %u",
conn_get_type_string(conn), conn->sd, msg->dmsg->id);
ASSERT(conn->type == CONN_DNODE_PEER_CLIENT);
pool = conn->owner;
key = NULL;
keylen = 0;
log_debug(LOG_DEBUG, "conn %p adding message %d:%d", conn, msg->id, msg->parent_id);
dictAdd(conn->outstanding_msgs_dict, &msg->id, msg);
if (!string_empty(&pool->hash_tag)) {
struct string *tag = &pool->hash_tag;
uint8_t *tag_start, *tag_end;
tag_start = dn_strchr(msg->key_start, msg->key_end, tag->data[0]);
if (tag_start != NULL) {
tag_end = dn_strchr(tag_start + 1, msg->key_end, tag->data[1]);
if (tag_end != NULL) {
key = tag_start + 1;
keylen = (uint32_t)(tag_end - key);
}
}
}
if (keylen == 0) {
key = msg->key_start;
keylen = (uint32_t)(msg->key_end - msg->key_start);
}
ASSERT(msg->dmsg != NULL);
if (msg->dmsg->type == DMSG_REQ) {
local_req_forward(ctx, conn, msg, key, keylen);
} else if (msg->dmsg->type == DMSG_REQ_FORWARD) {
struct mbuf *orig_mbuf = STAILQ_FIRST(&msg->mhdr);
struct datacenter *dc = server_get_dc(pool, &pool->dc);
uint32_t rack_cnt = array_n(&dc->racks);
uint32_t rack_index;
for(rack_index = 0; rack_index < rack_cnt; rack_index++) {
struct rack *rack = array_get(&dc->racks, rack_index);
//log_debug(LOG_DEBUG, "forwarding to rack '%.*s'",
// rack->name->len, rack->name->data);
struct msg *rack_msg;
if (string_compare(rack->name, &pool->rack) == 0 ) {
rack_msg = msg;
} else {
rack_msg = msg_get(conn, msg->request, __FUNCTION__);
if (rack_msg == NULL) {
log_debug(LOG_VERB, "whelp, looks like yer screwed now, buddy. no inter-rack messages for you!");
continue;
}
if (msg_clone(msg, orig_mbuf, rack_msg) != DN_OK) {
msg_put(rack_msg);
continue;
}
rack_msg->swallow = true;
}
if (log_loggable(LOG_DEBUG)) {
log_debug(LOG_DEBUG, "forwarding request from conn '%s' to rack '%.*s' dc '%.*s' ",
dn_unresolve_peer_desc(conn->sd), rack->name->len, rack->name->data, rack->dc->len, rack->dc->data);
}
remote_req_forward(ctx, conn, rack_msg, rack, key, keylen);
}
}
}
/* Description: link data from a peer connection to a client-facing connection
* peer_conn: a peer connection
* msg : msg with data from the peer connection after parsing
*/
static void
dnode_rsp_forward_match(struct context *ctx, struct conn *peer_conn, struct msg *rsp)
{
rstatus_t status;
struct msg *req;
struct conn *c_conn;
req = TAILQ_FIRST(&peer_conn->omsg_q);
c_conn = req->owner;
/* if client consistency is dc_one forward the response from only the
local node. Since dyn_dnode_peer is always a remote node, drop the rsp */
if (req->consistency == DC_ONE) {
if (req->swallow) {
dnode_rsp_swallow(ctx, peer_conn, req, rsp);
return;
}
log_warn("req %d:%d with DC_ONE consistency is not being swallowed");
}
/* if client consistency is dc_quorum, forward the response from only the
local region/DC. */
if ((req->consistency == DC_QUORUM) && !peer_conn->same_dc) {
if (req->swallow) {
dnode_rsp_swallow(ctx, peer_conn, req, rsp);
return;
}
log_warn("req %d:%d with DC_QUORUM consistency is not being swallowed");
}
log_debug(LOG_DEBUG, "DNODE RSP RECEIVED %s %d dmsg->id %u req %u:%u rsp %u:%u, ",
conn_get_type_string(peer_conn),
peer_conn->sd, rsp->dmsg->id,
req->id, req->parent_id, rsp->id, rsp->parent_id);
ASSERT(req != NULL && req->peer == NULL);
ASSERT(req->request && !req->done);
if (log_loggable(LOG_VVERB)) {
loga("Dumping content for response: ");
msg_dump(rsp);
loga("rsp id %d", rsp->id);
loga("Dumping content for request:");
msg_dump(req);
loga("req id %d", req->id);
}
conn_dequeue_outq(ctx, peer_conn, req);
req->done = 1;
log_info("c_conn:%p %d:%d <-> %d:%d", c_conn, req->id, req->parent_id,
rsp->id, rsp->parent_id);
/* establish rsp <-> req (response <-> request) link */
req->peer = rsp;
rsp->peer = req;
rsp->pre_coalesce(rsp);
ASSERT_LOG((c_conn->type == CONN_CLIENT) ||
(c_conn->type == CONN_DNODE_PEER_CLIENT),
"c_conn type %s", conn_get_type_string(c_conn));
dnode_rsp_forward_stats(ctx, peer_conn->owner, rsp);
// c_conn owns respnse now
status = conn_handle_response(c_conn, req->parent_id ? req->parent_id : req->id,
rsp);
IGNORE_RET_VAL(status);
if (req->swallow) {
log_info("swallow request %d:%d", req->id, req->parent_id);
req_put(req);
}
}
/* There are chances that the request to the remote peer or its response got dropped.
* Hence we may not always receive a response to the request at the head of the FIFO.
* Hence what we do is we mark that request as errored and move on the next one
* in the outgoing queue. This works since we always have message ids in monotonically
* increasing order.
*/
static void
dnode_rsp_forward(struct context *ctx, struct conn *peer_conn, struct msg *rsp)
{
rstatus_t status;
struct msg *req;
struct conn *c_conn;
ASSERT(peer_conn->type == CONN_DNODE_PEER_SERVER);
/* response from a peer implies that peer is ok and heartbeating */
dnode_peer_ok(ctx, peer_conn);
/* dequeue peer message (request) from peer conn */
while (true) {
req = TAILQ_FIRST(&peer_conn->omsg_q);
log_debug(LOG_VERB, "dnode_rsp_forward entering req %p rsp %p...", req, rsp);
c_conn = req->owner;
if (!peer_conn->same_dc && req->remote_region_send_time) {
struct stats *st = ctx->stats;
uint64_t delay = dn_usec_now() - req->remote_region_send_time;
histo_add(&st->cross_region_histo, delay);
}
if (req->id == rsp->dmsg->id) {
dnode_rsp_forward_match(ctx, peer_conn, rsp);
return;
}
// Report a mismatch and try to rectify
log_error("MISMATCH: dnode %s %d rsp_dmsg_id %u req %u:%u dnode rsp %u:%u",
conn_get_type_string(peer_conn),
peer_conn->sd, rsp->dmsg->id, req->id, req->parent_id, rsp->id,
rsp->parent_id);
if (c_conn && conn_to_ctx(c_conn))
stats_pool_incr(conn_to_ctx(c_conn), c_conn->owner,
peer_mismatch_requests);
// TODO : should you be worried about message id getting wrapped around to 0?
if (rsp->dmsg->id < req->id) {
// We received a response from the past. This indeed proves out of order
// responses. A blunder to the architecture. Log it and drop the response.
log_error("MISMATCH: received response from the past. Dropping it");
rsp_put(rsp);
return;
}
if (req->consistency == DC_ONE) {
if (req->swallow) {
// swallow the request and move on the next one
dnode_rsp_swallow(ctx, peer_conn, req, NULL);
continue;
}
log_warn("req %d:%d with DC_ONE consistency is not being swallowed");
}
if ((req->consistency == DC_QUORUM) && !peer_conn->same_dc) {
if (req->swallow) {
// swallow the request and move on the next one
dnode_rsp_swallow(ctx, peer_conn, req, NULL);
continue;
}
log_warn("req %d:%d with DC_QUORUM consistency is not being swallowed");
}
log_error("MISMATCHED DNODE RSP RECEIVED %s %d dmsg->id %u req %u:%u rsp %u:%u, skipping....",
conn_get_type_string(peer_conn),
peer_conn->sd, rsp->dmsg->id,
req->id, req->parent_id, rsp->id, rsp->parent_id);
ASSERT(req != NULL && req->peer == NULL);
ASSERT(req->request && !req->done);
if (log_loggable(LOG_VVERB)) {
loga("skipping req: ");
msg_dump(req);
}
conn_dequeue_outq(ctx, peer_conn, req);
req->done = 1;
// Create an appropriate response for the request so its propagated up;
struct msg *err_rsp = msg_get(peer_conn, false, peer_conn->data_store,
__FUNCTION__);
err_rsp->error = req->error = 1;
err_rsp->err = req->err = BAD_FORMAT;
err_rsp->dyn_error = req->dyn_error = BAD_FORMAT;
err_rsp->dmsg = dmsg_get();
err_rsp->dmsg->id = req->id;
log_debug(LOG_VERB, "%p <-> %p", req, err_rsp);
/* establish err_rsp <-> req (response <-> request) link */
req->peer = err_rsp;
err_rsp->peer = req;
log_error("Peer connection s %d skipping request %u:%u, dummy err_rsp %u:%u",
peer_conn->sd, req->id, req->parent_id, err_rsp->id, err_rsp->parent_id);
rstatus_t status =
conn_handle_response(c_conn, req->parent_id ? req->parent_id : req->id,
err_rsp);
IGNORE_RET_VAL(status);
if (req->swallow) {
log_debug(LOG_INFO, "swallow request %d:%d", req->id, req->parent_id);
req_put(req);
}
}
}
