/* Used for broadcast commands, like flush_all or stats.
*/
bool cproxy_broadcast_a2a_downstream(downstream *d,
char *command,
conn *uc,
char *suffix) {
assert(d != NULL);
assert(d->ptd != NULL);
assert(d->ptd->proxy != NULL);
assert(d->downstream_conns != NULL);
assert(d->downstream_used_start == 0);
assert(d->downstream_used == 0);
assert(command != NULL);
assert(uc != NULL);
assert(uc->next == NULL);
assert(uc->item == NULL);
int nwrite = 0;
int nconns = mcs_server_count(&d->mst);
for (int i = 0; i < nconns; i++) {
conn *c = d->downstream_conns[i];
if (c != NULL &&
c != NULL_CONN) {
if (cproxy_prep_conn_for_write(c)) {
assert(c->state == conn_pause);
out_string(c, command);
if (update_event(c, EV_WRITE | EV_PERSIST)) {
nwrite++;
if (uc->noreply) {
c->write_and_go = conn_pause;
}
} else {
if (settings.verbose > 1) {
moxi_log_write("Update cproxy write event failed\n");
}
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
}
} else {
d->ptd->stats.stats.err_downstream_write_prep++;
cproxy_close_conn(c);
}
}
}
if (settings.verbose > 1) {
moxi_log_write("%d: a2a broadcast nwrite %d out of %d\n",
uc->sfd, nwrite, nconns);
}
if (nwrite > 0) {
d->downstream_used_start = nwrite;
d->downstream_used = nwrite;
if (cproxy_dettach_if_noreply(d, uc) == false) {
d->upstream_suffix = suffix;
d->upstream_suffix_len = 0;
d->upstream_status = PROTOCOL_BINARY_RESPONSE_SUCCESS;
d->upstream_retry = 0;
d->target_host_ident = NULL;
cproxy_start_downstream_timeout(d, NULL);
} else {
// TODO: Handle flush_all's expiration parameter against
// the front_cache.
//
if (strncmp(command, "flush_all", 9) == 0) {
mcache_flush_all(&d->ptd->proxy->front_cache, 0);
}
}
return true;
}
return false;
}
/* Used for broadcast commands, like no-op, flush_all or stats.
*/
bool cproxy_broadcast_b2b_downstream(downstream *d, conn *uc) {
int nwrite = 0;
int nconns;
int i;
cb_assert(d != NULL);
cb_assert(d->ptd != NULL);
cb_assert(d->ptd->proxy != NULL);
cb_assert(d->downstream_conns != NULL);
cb_assert(uc != NULL);
cb_assert(uc->next == NULL);
cb_assert(uc->noreply == false);
nconns = mcs_server_count(&d->mst);
for (i = 0; i < nconns; i++) {
conn *c = d->downstream_conns[i];
if (c != NULL &&
c != NULL_CONN &&
b2b_forward_item_vbucket(uc, d, uc->item, c, -1) == true) {
nwrite++;
}
}
if (settings.verbose > 2) {
moxi_log_write("%d: b2b broadcast nwrite %d out of %d\n",
uc->sfd, nwrite, nconns);
}
if (nwrite > 0) {
/* TODO: Handle binary 'stats reset' sub-command. */
item *it;
if (uc->cmd == PROTOCOL_BINARY_CMD_STAT &&
d->merger == NULL) {
d->merger = genhash_init(128, skeyhash_ops);
}
it = item_alloc("h", 1, 0, 0,
sizeof(protocol_binary_response_header));
if (it != NULL) {
protocol_binary_response_header *header =
(protocol_binary_response_header *) ITEM_data(it);
memset(ITEM_data(it), 0, it->nbytes);
header->response.magic = (uint8_t) PROTOCOL_BINARY_RES;
header->response.opcode = uc->binary_header.request.opcode;
header->response.opaque = uc->opaque;
if (add_conn_item(uc, it)) {
d->upstream_suffix = ITEM_data(it);
d->upstream_suffix_len = it->nbytes;
d->upstream_status = PROTOCOL_BINARY_RESPONSE_SUCCESS;
d->target_host_ident = NULL;
if (settings.verbose > 2) {
moxi_log_write("%d: b2b broadcast upstream_suffix", uc->sfd);
cproxy_dump_header(uc->sfd, ITEM_data(it));
}
/* TODO: Handle FLUSHQ (quiet binary flush_all). */
d->downstream_used_start = nwrite;
d->downstream_used = nwrite;
cproxy_start_downstream_timeout(d, NULL);
return true;
}
item_remove(it);
}
}
return false;
}
/* Do the actual work of forwarding the command from an
* upstream binary conn to its assigned binary downstream.
*/
bool cproxy_forward_b2b_downstream(downstream *d) {
int nc;
int server_index;
conn *uc;
cb_assert(d != NULL);
cb_assert(d->ptd != NULL);
cb_assert(d->ptd->proxy != NULL);
cb_assert(d->downstream_conns != NULL);
cb_assert(d->downstream_used == 0);
cb_assert(d->multiget == NULL);
cb_assert(d->merger == NULL);
d->downstream_used_start = 0;
uc = d->upstream_conn;
if (settings.verbose > 2) {
moxi_log_write("%d: cproxy_forward_b2b_downstream %x\n",
uc->sfd, uc->cmd);
}
cb_assert(uc != NULL);
cb_assert(uc->state == conn_pause);
cb_assert(uc->cmd >= 0);
cb_assert(uc->cmd_start == NULL);
cb_assert(uc->thread != NULL);
cb_assert(uc->thread->base != NULL);
cb_assert(uc->noreply == false);
cb_assert(IS_BINARY(uc->protocol));
cb_assert(IS_PROXY(uc->protocol));
server_index = -1;
if (cproxy_is_broadcast_cmd(uc->cmd) == false && uc->corked == NULL) {
item *it = uc->item;
protocol_binary_request_header *req;
char *key;
int key_len;
cb_assert(it != NULL);
req = (protocol_binary_request_header *) ITEM_data(it);
key = ((char *) req) + sizeof(*req) + req->request.extlen;
key_len = ntohs(req->request.keylen);
if (key_len > 0) {
server_index = cproxy_server_index(d, key, key_len, NULL);
if (server_index < 0) {
return false;
}
}
}
nc = cproxy_connect_downstream(d, uc->thread, server_index);
if (nc == -1) {
return true;
}
if (nc > 0) {
int i;
int nconns;
cb_assert(d->downstream_conns != NULL);
if (d->usec_start == 0 &&
d->ptd->behavior_pool.base.time_stats) {
d->usec_start = usec_now();
}
nconns = mcs_server_count(&d->mst);
for (i = 0; i < nconns; i++) {
conn *c = d->downstream_conns[i];
if (c != NULL &&
c != NULL_CONN) {
cb_assert(c->state == conn_pause);
cb_assert(c->item == NULL);
if (cproxy_prep_conn_for_write(c) == false) {
d->ptd->stats.stats.err_downstream_write_prep++;
cproxy_close_conn(c);
return false;
}
}
}
/* Uncork the saved-up quiet binary commands. */
cproxy_binary_uncork_cmds(d, uc);
if (uc->cmd == PROTOCOL_BINARY_CMD_FLUSH ||
uc->cmd == PROTOCOL_BINARY_CMD_NOOP ||
uc->cmd == PROTOCOL_BINARY_CMD_STAT) {
return cproxy_broadcast_b2b_downstream(d, uc);
}
return cproxy_forward_b2b_simple_downstream(d, uc);
}
if (settings.verbose > 2) {
moxi_log_write("%d: cproxy_forward_b2b_downstream connect failed\n",
uc->sfd);
}
return false;
}
bool multiget_ascii_downstream(downstream *d, conn *uc,
int (*emit_start)(conn *c, char *cmd, int cmd_len),
int (*emit_skey)(conn *c, char *skey, int skey_len),
int (*emit_end)(conn *c),
mcache *front_cache) {
assert(d != NULL);
assert(d->downstream_conns != NULL);
assert(d->multiget == NULL);
assert(uc != NULL);
assert(uc->noreply == false);
proxy_td *ptd = d->ptd;
assert(ptd != NULL);
proxy_stats_cmd *psc_get =
&ptd->stats.stats_cmd[STATS_CMD_TYPE_REGULAR][STATS_CMD_GET];
proxy_stats_cmd *psc_get_key =
&ptd->stats.stats_cmd[STATS_CMD_TYPE_REGULAR][STATS_CMD_GET_KEY];
int nwrite = 0;
int nconns = mcs_server_count(&d->mst);
for (int i = 0; i < nconns; i++) {
if (d->downstream_conns[i] != NULL &&
cproxy_prep_conn_for_write(d->downstream_conns[i]) == false) {
d->ptd->stats.stats.err_downstream_write_prep++;
cproxy_close_conn(d->downstream_conns[i]);
return false;
}
}
if (uc->next != NULL) {
// More than one upstream conn, so we need a hashtable
// to track keys for de-deplication.
//
d->multiget = genhash_init(128, skeyhash_ops);
if (settings.verbose > 1) {
fprintf(stderr, "cproxy multiget hash table new\n");
}
}
// Snapshot the volatile only once.
//
uint32_t msec_current_time_snapshot = msec_current_time;
int uc_num = 0;
conn *uc_cur = uc;
while (uc_cur != NULL) {
assert(uc_cur->cmd == -1);
assert(uc_cur->item == NULL);
assert(uc_cur->state == conn_pause);
assert(IS_ASCII(uc_cur->protocol));
assert(IS_PROXY(uc_cur->protocol));
char *command = uc_cur->cmd_start;
assert(command != NULL);
char *space = strchr(command, ' ');
assert(space > command);
int cmd_len = space - command;
assert(cmd_len == 3 || cmd_len == 4); // Either get or gets.
int cas_emit = (command[3] == 's');
if (settings.verbose > 1) {
fprintf(stderr, "forward multiget %s (%d %d)\n",
command, cmd_len, uc_num);
}
while (space != NULL) {
char *key = space + 1;
char *next_space = strchr(key, ' ');
int key_len;
if (next_space != NULL) {
key_len = next_space - key;
} else {
key_len = strlen(key);
// We've reached the last key.
//
psc_get->read_bytes += (key - command + key_len);
}
// This key_len check helps skip consecutive spaces.
//
if (key_len > 0) {
ptd->stats.stats.tot_multiget_keys++;
psc_get_key->seen++;
psc_get_key->read_bytes += key_len;
// Update key-based statistics.
//
bool do_key_stats =
matcher_check(&ptd->key_stats_matcher,
key, key_len, true) == true &&
matcher_check(&ptd->key_stats_unmatcher,
key, key_len, false) == false;
if (do_key_stats) {
touch_key_stats(ptd, key, key_len,
msec_current_time_snapshot,
STATS_CMD_TYPE_REGULAR,
STATS_CMD_GET_KEY,
1, 0, 0,
key_len, 0);
}
// Handle a front cache hit by queuing response.
//
// Note, front cache stats are part of mcache.
//
if (!cas_emit) {
item *it = mcache_get(front_cache, key, key_len,
msec_current_time_snapshot);
if (it != NULL) {
assert(it->nkey == key_len);
assert(strncmp(ITEM_key(it), key, it->nkey) == 0);
cproxy_upstream_ascii_item_response(it, uc_cur, 0);
psc_get_key->hits++;
psc_get_key->write_bytes += it->nbytes;
if (do_key_stats) {
touch_key_stats(ptd, key, key_len,
msec_current_time_snapshot,
STATS_CMD_TYPE_REGULAR,
STATS_CMD_GET_KEY,
0, 1, 0,
0, it->nbytes);
}
// The refcount was inc'ed by mcache_get() for us.
//
item_remove(it);
goto loop_next;
}
}
bool self = false;
conn *c = cproxy_find_downstream_conn(d, key, key_len,
&self);
if (c != NULL) {
if (self) {
// Optimization for talking with ourselves,
// to avoid extra network hop.
//
ptd->stats.stats.tot_optimize_self++;
item *it = item_get(key, key_len);
if (it != NULL) {
cproxy_upstream_ascii_item_response(it, uc_cur,
cas_emit);
psc_get_key->hits++;
psc_get_key->write_bytes += it->nbytes;
if (do_key_stats) {
touch_key_stats(ptd, key, key_len,
msec_current_time_snapshot,
STATS_CMD_TYPE_REGULAR,
STATS_CMD_GET_KEY,
0, 1, 0,
0, it->nbytes);
}
// The refcount was inc'ed by item_get() for us.
//
item_remove(it);
if (settings.verbose > 1) {
fprintf(stderr,
"optimize self multiget hit: %s\n",
key);
}
} else {
psc_get_key->misses++;
if (do_key_stats) {
touch_key_stats(ptd, key, key_len,
msec_current_time_snapshot,
STATS_CMD_TYPE_REGULAR,
STATS_CMD_GET_KEY,
0, 0, 1,
0, 0);
}
if (settings.verbose > 1) {
fprintf(stderr,
"optimize self multiget miss: %s\n",
key);
}
}
goto loop_next;
}
// See if we've already requested this key via
// the multiget hash table, in order to
// de-deplicate repeated keys.
//
bool first_request = true;
if (d->multiget != NULL) {
// TODO: Use Trond's allocator here.
//
multiget_entry *entry =
calloc(1, sizeof(multiget_entry));
if (entry != NULL) {
entry->upstream_conn = uc_cur;
entry->opaque = 0;
entry->hits = 0;
entry->next = genhash_find(d->multiget, key);
genhash_update(d->multiget, key, entry);
if (entry->next != NULL) {
first_request = false;
}
} else {
// TODO: Handle out of multiget entry memory.
}
}
if (first_request) {
assert(c->item == NULL);
assert(c->state == conn_pause);
assert(IS_PROXY(c->protocol));
assert(c->ilist != NULL);
assert(c->isize > 0);
if (c->msgused <= 1 &&
c->msgbytes <= 0) {
emit_start(c, command, cmd_len);
}
// Provide the preceding space as optimization
// for ascii-to-ascii configuration.
//
emit_skey(c, key - 1, key_len + 1);
} else {
ptd->stats.stats.tot_multiget_keys_dedupe++;
if (settings.verbose > 1) {
char buf[KEY_MAX_LENGTH + 10];
memcpy(buf, key, key_len);
buf[key_len] = '\0';
fprintf(stderr,
"%d cproxy multiget dedpue: %s\n",
uc_cur->sfd, buf);
}
}
} else {
// TODO: Handle when downstream conn is down.
}
}
loop_next:
space = next_space;
}
uc_num++;
uc_cur = uc_cur->next;
}
for (int i = 0; i < nconns; i++) {
conn *c = d->downstream_conns[i];
if (c != NULL &&
(c->msgused > 1 ||
c->msgbytes > 0)) {
emit_end(c);
conn_set_state(c, conn_mwrite);
c->write_and_go = conn_new_cmd;
if (update_event(c, EV_WRITE | EV_PERSIST)) {
nwrite++;
if (uc->noreply) {
c->write_and_go = conn_pause;
}
} else {
if (settings.verbose > 1) {
fprintf(stderr,
"Couldn't update cproxy write event\n");
}
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
}
}
}
if (settings.verbose > 1) {
fprintf(stderr, "forward multiget nwrite %d out of %d\n",
nwrite, nconns);
}
d->downstream_used_start = nwrite;
d->downstream_used = nwrite;
if (cproxy_dettach_if_noreply(d, uc) == false) {
d->upstream_suffix = "END\r\n";
cproxy_start_downstream_timeout(d, NULL);
}
return nwrite > 0;
}
