/**
* Depending on our configuration, we can optimize SET's
* on certain keys by making them fire-and-forget and
* immediately transmitting a success response to the
* upstream client.
*/
bool cproxy_optimize_set_ascii(downstream *d, conn *uc,
char *key, int key_len) {
assert(d);
assert(d->ptd);
assert(d->ptd->proxy);
assert(uc);
assert(uc->next == NULL);
if (d->ptd->behavior_pool.base.optimize_set[0] == '\0') {
return false;
}
if (matcher_check(&d->ptd->proxy->optimize_set_matcher,
key, key_len, false)) {
d->upstream_conn = NULL;
d->upstream_suffix = NULL;
d->upstream_suffix_len = 0;
d->upstream_retry = 0;
out_string(uc, "STORED");
if (!update_event(uc, EV_WRITE | EV_PERSIST)) {
if (settings.verbose > 1) {
moxi_log_write("ERROR: Can't update upstream write event\n");
}
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(uc);
}
return true;
}
return false;
}
bool b2b_forward_item_vbucket(conn *uc, downstream *d, item *it,
conn *c, bool self, int vbucket) {
(void)self;
assert(d != NULL);
assert(d->ptd != NULL);
assert(uc != NULL);
assert(uc->next == NULL);
assert(uc->noreply == false);
assert(c != NULL);
// Assuming we're already connected to downstream.
//
// TODO: Optimize to self codepath.
//
if (settings.verbose > 2) {
moxi_log_write("%d: b2b_forward_item_vbucket %x to %d, vbucket %d\n",
uc->sfd, uc->cmd, c->sfd, vbucket);
}
protocol_binary_request_header *req =
(protocol_binary_request_header *) ITEM_data(it);
if (vbucket >= 0) {
req->request.reserved = htons(vbucket);
}
if (add_conn_item(c, it) == true) {
// The caller keeps its refcount, and we need our own.
//
it->refcount++;
if (add_iov(c, ITEM_data(it), it->nbytes) == 0) {
conn_set_state(c, conn_mwrite);
c->write_and_go = conn_new_cmd;
if (update_event(c, EV_WRITE | EV_PERSIST)) {
if (settings.verbose > 2) {
moxi_log_write("%d: b2b_forward %x to %d success\n",
uc->sfd, uc->cmd, c->sfd);
}
return true;
}
}
}
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
return false;
}
/* Forward an upstream command that came with item data,
* like set/add/replace/etc.
*/
bool cproxy_forward_a2a_item_downstream(downstream *d, short cmd,
item *it, conn *uc) {
assert(d != NULL);
assert(d->ptd != NULL);
assert(d->ptd->proxy != NULL);
assert(d->downstream_conns != NULL);
assert(it != NULL);
assert(uc != NULL);
assert(uc->next == NULL);
// Assuming we're already connected to downstream.
//
bool self = false;
conn *c = cproxy_find_downstream_conn(d, ITEM_key(it), it->nkey,
&self);
if (c != NULL) {
if (self) {
cproxy_optimize_to_self(d, uc, uc->cmd_start);
complete_nread_ascii(uc);
return true;
}
if (cproxy_prep_conn_for_write(c)) {
assert(c->state == conn_pause);
char *verb = nread_text(cmd);
assert(verb != NULL);
char *str_flags = ITEM_suffix(it);
char *str_length = strchr(str_flags + 1, ' ');
int len_flags = str_length - str_flags;
int len_length = it->nsuffix - len_flags - 2;
char *str_exptime = add_conn_suffix(c);
char *str_cas = (cmd == NREAD_CAS ? add_conn_suffix(c) : NULL);
if (str_flags != NULL &&
str_length != NULL &&
len_flags > 1 &&
len_length > 1 &&
str_exptime != NULL &&
(cmd != NREAD_CAS ||
str_cas != NULL)) {
sprintf(str_exptime, " %u", it->exptime);
if (str_cas != NULL)
sprintf(str_cas, " %llu",
(unsigned long long) ITEM_get_cas(it));
if (add_iov(c, verb, strlen(verb)) == 0 &&
add_iov(c, ITEM_key(it), it->nkey) == 0 &&
add_iov(c, str_flags, len_flags) == 0 &&
add_iov(c, str_exptime, strlen(str_exptime)) == 0 &&
add_iov(c, str_length, len_length) == 0 &&
(str_cas == NULL ||
add_iov(c, str_cas, strlen(str_cas)) == 0) &&
(uc->noreply == false ||
add_iov(c, " noreply", 8) == 0) &&
add_iov(c, ITEM_data(it) - 2, it->nbytes + 2) == 0) {
conn_set_state(c, conn_mwrite);
c->write_and_go = conn_new_cmd;
if (update_event(c, EV_WRITE | EV_PERSIST)) {
d->downstream_used_start = 1;
d->downstream_used = 1;
if (cproxy_dettach_if_noreply(d, uc) == false) {
cproxy_start_downstream_timeout(d, c);
// During a synchronous (with-reply) SET,
// handle fire-&-forget SET optimization.
//
if (cmd == NREAD_SET &&
cproxy_optimize_set_ascii(d, uc,
ITEM_key(it),
it->nkey)) {
d->ptd->stats.stats.tot_optimize_sets++;
}
} else {
c->write_and_go = conn_pause;
mcache_delete(&d->ptd->proxy->front_cache,
ITEM_key(it), it->nkey);
}
return true;
}
}
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
} else {
// TODO: Handle this weird error case.
}
} else {
d->ptd->stats.stats.err_downstream_write_prep++;
cproxy_close_conn(c);
}
if (settings.verbose > 1)
fprintf(stderr, "Proxy item write out of memory");
}
return false;
}
/* 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(command != NULL);
assert(uc != NULL);
assert(uc->next == NULL);
assert(uc->item == NULL);
int nwrite = 0;
int nconns = memcached_server_count(&d->mst);
for (int i = 0; i < nconns; i++) {
conn *c = d->downstream_conns[i];
if (c != NULL) {
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)
fprintf(stderr,
"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)
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 = suffix;
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 nwrite > 0;
}
void cproxy_process_a2a_downstream(conn *c, char *line) {
assert(c != NULL);
assert(c->next == NULL);
assert(c->extra != NULL);
assert(c->cmd == -1);
assert(c->item == NULL);
assert(line != NULL);
assert(line == c->rcurr);
assert(IS_ASCII(c->protocol));
assert(IS_PROXY(c->protocol));
if (settings.verbose > 1)
fprintf(stderr, "<%d cproxy_process_a2a_downstream %s\n",
c->sfd, line);
downstream *d = c->extra;
assert(d != NULL);
assert(d->ptd != NULL);
assert(d->ptd->proxy != NULL);
if (strncmp(line, "VALUE ", 6) == 0) {
token_t tokens[MAX_TOKENS];
size_t ntokens;
unsigned int flags;
int clen = 0;
int vlen;
uint64_t cas = CPROXY_NOT_CAS;
ntokens = scan_tokens(line, tokens, MAX_TOKENS, &clen);
if (ntokens >= 5 && // Accounts for extra termimation token.
ntokens <= 6 &&
tokens[KEY_TOKEN].length <= KEY_MAX_LENGTH &&
safe_strtoul(tokens[2].value, (uint32_t *) &flags) &&
safe_strtoul(tokens[3].value, (uint32_t *) &vlen)) {
char *key = tokens[KEY_TOKEN].value;
size_t nkey = tokens[KEY_TOKEN].length;
item *it = item_alloc(key, nkey, flags, 0, vlen + 2);
if (it != NULL) {
if (ntokens == 5 ||
safe_strtoull(tokens[4].value, &cas)) {
ITEM_set_cas(it, cas);
c->item = it;
c->ritem = ITEM_data(it);
c->rlbytes = it->nbytes;
c->cmd = -1;
conn_set_state(c, conn_nread);
return; // Success.
} else {
if (settings.verbose > 1)
fprintf(stderr, "cproxy could not parse cas\n");
}
} else {
if (settings.verbose > 1)
fprintf(stderr, "cproxy could not item_alloc size %u\n",
vlen + 2);
}
if (it != NULL)
item_remove(it);
it = NULL;
c->sbytes = vlen + 2; // Number of bytes to swallow.
conn_set_state(c, conn_swallow);
// Note, eventually, we'll see an END later.
} else {
// We don't know how much to swallow, so close the downstream.
// The conn_closing should release the downstream,
// which should write a suffix/error to the upstream.
//
conn_set_state(c, conn_closing);
}
} else if (strncmp(line, "END", 3) == 0) {
conn_set_state(c, conn_pause);
} else if (strncmp(line, "OK", 2) == 0) {
conn_set_state(c, conn_pause);
// TODO: Handle flush_all's expiration parameter against
// the front_cache.
//
// TODO: We flush the front_cache too often, inefficiently
// on every downstream flush_all OK response, rather than
// on just the last flush_all OK response.
//
conn *uc = d->upstream_conn;
if (uc != NULL &&
uc->cmd_curr == PROTOCOL_BINARY_CMD_FLUSH) {
mcache_flush_all(&d->ptd->proxy->front_cache, 0);
}
} else if (strncmp(line, "STAT ", 5) == 0 ||
strncmp(line, "ITEM ", 5) == 0 ||
strncmp(line, "PREFIX ", 7) == 0) {
assert(d->merger != NULL);
conn *uc = d->upstream_conn;
if (uc != NULL) {
assert(uc->next == NULL);
if (protocol_stats_merge_line(d->merger, line) == false) {
// Forward the line as-is if we couldn't merge it.
//
int nline = strlen(line);
item *it = item_alloc("s", 1, 0, 0, nline + 2);
if (it != NULL) {
strncpy(ITEM_data(it), line, nline);
strncpy(ITEM_data(it) + nline, "\r\n", 2);
if (add_conn_item(uc, it)) {
add_iov(uc, ITEM_data(it), nline + 2);
it = NULL;
}
if (it != NULL)
item_remove(it);
}
}
}
conn_set_state(c, conn_new_cmd);
} else {
conn_set_state(c, conn_pause);
// The upstream conn might be NULL when closed already
// or while handling a noreply.
//
conn *uc = d->upstream_conn;
if (uc != NULL) {
assert(uc->next == NULL);
out_string(uc, line);
if (!update_event(uc, EV_WRITE | EV_PERSIST)) {
if (settings.verbose > 1)
fprintf(stderr,
"Can't update upstream write event\n");
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(uc);
}
cproxy_del_front_cache_key_ascii_response(d, line,
uc->cmd_start);
}
}
}
/* Forward a simple one-liner command downstream.
* For example, get, incr/decr, delete, etc.
* The response, though, might be a simple line or
* multiple VALUE+END lines.
*/
bool cproxy_forward_a2a_simple_downstream(downstream *d,
char *command, conn *uc) {
assert(d != NULL);
assert(d->ptd != NULL);
assert(d->ptd->proxy != NULL);
assert(d->downstream_conns != NULL);
assert(command != NULL);
assert(uc != NULL);
assert(uc->item == NULL);
assert(uc->cmd_curr != -1);
assert(d->multiget == NULL);
assert(d->merger == NULL);
// Handles get and gets.
//
if (uc->cmd_curr == PROTOCOL_BINARY_CMD_GET) {
// Only use front_cache for 'get', not for 'gets'.
//
mcache *front_cache =
(command[3] == ' ') ? &d->ptd->proxy->front_cache : NULL;
return multiget_ascii_downstream(d, uc,
a2a_multiget_start,
a2a_multiget_skey,
a2a_multiget_end,
front_cache);
}
assert(uc->next == NULL);
if (uc->cmd_curr == PROTOCOL_BINARY_CMD_FLUSH)
return cproxy_broadcast_a2a_downstream(d, command, uc,
"OK\r\n");
if (uc->cmd_curr == PROTOCOL_BINARY_CMD_STAT) {
if (strncmp(command + 5, " reset", 6) == 0)
return cproxy_broadcast_a2a_downstream(d, command, uc,
"RESET\r\n");
if (cproxy_broadcast_a2a_downstream(d, command, uc,
"END\r\n")) {
d->merger = genhash_init(512, skeyhash_ops);
return true;
} else {
return false;
}
}
// TODO: Inefficient repeated scan_tokens.
//
int cmd_len = 0;
token_t tokens[MAX_TOKENS];
size_t ntokens = scan_tokens(command, tokens, MAX_TOKENS, &cmd_len);
char *key = tokens[KEY_TOKEN].value;
int key_len = tokens[KEY_TOKEN].length;
if (ntokens <= 1) { // This was checked long ago, while parsing
assert(false); // the upstream conn.
return false;
}
// Assuming we're already connected to downstream.
//
bool self = false;
conn *c = cproxy_find_downstream_conn(d, key, key_len,
&self);
if (c != NULL) {
if (self) {
// TODO: This optimization could be done much earlier,
// even before the upstream conn was assigned
// to a downstream.
//
cproxy_optimize_to_self(d, uc, command);
process_command(uc, command);
return true;
}
if (cproxy_prep_conn_for_write(c)) {
assert(c->state == conn_pause);
out_string(c, command);
if (settings.verbose > 1)
fprintf(stderr, "forwarding to %d, noreply %d\n",
c->sfd, uc->noreply);
if (update_event(c, EV_WRITE | EV_PERSIST)) {
d->downstream_used_start = 1;
d->downstream_used = 1;
if (cproxy_dettach_if_noreply(d, uc) == false) {
cproxy_start_downstream_timeout(d, c);
} else {
c->write_and_go = conn_pause;
// Do mcache_delete() here only during a noreply,
// otherwise for with-reply requests, we could
// be in a race with other clients repopulating
// the front_cache. For with-reply requests, we
// clear the front_cache when we get a success reply.
//
mcache_delete(&d->ptd->proxy->front_cache, key, key_len);
}
return true;
}
if (settings.verbose > 1)
fprintf(stderr, "Couldn't update cproxy write event\n");
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
} else {
d->ptd->stats.stats.err_downstream_write_prep++;
cproxy_close_conn(c);
}
}
return false;
}
void cproxy_process_upstream_binary(conn *c) {
cb_assert(c != NULL);
cb_assert(c->cmd >= 0);
cb_assert(c->next == NULL);
cb_assert(c->item == NULL);
cb_assert(IS_BINARY(c->protocol));
cb_assert(IS_PROXY(c->protocol));
proxy_td *ptd = c->extra;
cb_assert(ptd != NULL);
if (!cproxy_prep_conn_for_write(c)) {
ptd->stats.stats.err_upstream_write_prep++;
conn_set_state(c, conn_closing);
return;
}
c->cmd_curr = -1;
c->cmd_start = NULL;
c->cmd_start_time = msec_current_time;
c->cmd_retries = 0;
int extlen = c->binary_header.request.extlen;
int keylen = c->binary_header.request.keylen;
uint32_t bodylen = c->binary_header.request.bodylen;
cb_assert(bodylen >= (uint32_t) keylen + extlen);
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_upstream_binary %x %d %d %u\n",
c->sfd, c->cmd, extlen, keylen, bodylen);
}
process_bin_noreply(c); /* Map quiet c->cmd values into non-quiet. */
if (c->cmd == PROTOCOL_BINARY_CMD_VERSION ||
c->cmd == PROTOCOL_BINARY_CMD_QUIT) {
dispatch_bin_command(c);
return;
}
/* Alloc an item and continue with an rest-of-body nread if */
/* necessary. The item will hold the entire request message */
/* (the header + body). */
char *ikey = "u";
int ikeylen = 1;
c->item = item_alloc(ikey, ikeylen, 0, 0,
sizeof(c->binary_header) + bodylen);
if (c->item != NULL) {
item *it = c->item;
void *rb = c->rcurr;
cb_assert(it->refcount == 1);
memcpy(ITEM_data(it), rb, sizeof(c->binary_header));
if (bodylen > 0) {
c->ritem = ITEM_data(it) + sizeof(c->binary_header);
c->rlbytes = bodylen;
c->substate = bin_read_set_value;
conn_set_state(c, conn_nread);
} else {
/* Since we have no body bytes, we can go immediately to */
/* the nread completed processing step. */
if (c->binary_header.request.opcode == PROTOCOL_BINARY_CMD_SASL_LIST_MECHS) {
/* TODO: One day handle more than just PLAIN sasl auth. */
write_bin_response(c, "PLAIN", 0, 0, strlen("PLAIN"));
return;
}
cproxy_pause_upstream_for_downstream(ptd, c);
}
} else {
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_upstream_binary OOM\n",
c->sfd);
}
ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
}
}
/* We get here after reading the header+body into an item.
*/
void cproxy_process_upstream_binary_nread(conn *c) {
cb_assert(c != NULL);
cb_assert(c->cmd >= 0);
cb_assert(c->next == NULL);
cb_assert(c->cmd_start == NULL);
cb_assert(IS_BINARY(c->protocol));
cb_assert(IS_PROXY(c->protocol));
protocol_binary_request_header *header =
(protocol_binary_request_header *) &c->binary_header;
int extlen = header->request.extlen;
int keylen = header->request.keylen;
uint32_t bodylen = header->request.bodylen;
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_upstream_binary_nread %x %d %d %u\n",
c->sfd, c->cmd, extlen, keylen, bodylen);
}
/* pthread_mutex_lock(&c->thread->stats.mutex); */
/* c->thread->stats.slab_stats[it->slabs_clsid].set_cmds++; */
/* pthread_mutex_unlock(&c->thread->stats.mutex); */
proxy_td *ptd = c->extra;
cb_assert(ptd != NULL);
if (header->request.opcode == PROTOCOL_BINARY_CMD_SASL_AUTH) {
item *it = c->item;
cb_assert(it);
cproxy_sasl_plain_auth(c, (char *) ITEM_data(it));
return;
}
if (header->request.opcode == PROTOCOL_BINARY_CMD_SASL_STEP) {
write_bin_error(c, PROTOCOL_BINARY_RESPONSE_AUTH_ERROR, 0);
return;
}
if (c->binary_header.request.opcode == PROTOCOL_BINARY_CMD_STAT) {
char *subcommand = binary_get_key(c);
size_t nkey = c->binary_header.request.keylen;
if (nkey == 13 && memcmp(subcommand, "proxy buckets", 13) == 0) {
process_bin_proxy_stats(c);
return;
}
}
if (c->noreply) {
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_upstream_binary_nread "
"corking quiet command %x %d\n",
c->sfd, c->cmd, (c->corked != NULL));
}
/* TODO: We currently don't support binary FLUSHQ. */
/* Rather than having the downstream connections get */
/* into a wonky state, prevent it. */
if (header->request.opcode == PROTOCOL_BINARY_CMD_FLUSHQ) {
/* Note: don't use cproxy_close_conn(c), as it goes */
/* through the drive_machine() loop again. */
/* cproxy_close_conn(c); */
conn_set_state(c, conn_closing);
return;
}
/* Hold onto or 'cork' all the binary quiet commands */
/* until there's a later non-quiet command. */
if (cproxy_binary_cork_cmd(c)) {
conn_set_state(c, conn_new_cmd);
} else {
ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
}
return;
}
cb_assert(c->item == NULL || ((item *) c->item)->refcount == 1);
cproxy_pause_upstream_for_downstream(ptd, c);
}
/* 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;
}
/* We reach here after nread'ing a header+body into an item.
*/
void cproxy_process_b2b_downstream_nread(conn *c) {
conn *uc;
item *it;
downstream *d;
protocol_binary_response_header *header;
int extlen;
int keylen;
uint32_t bodylen;
int status;
int opcode;
cb_assert(c != NULL);
cb_assert(c->cmd >= 0);
cb_assert(c->next == NULL);
cb_assert(c->cmd_start == NULL);
cb_assert(IS_BINARY(c->protocol));
cb_assert(IS_PROXY(c->protocol));
header = (protocol_binary_response_header *) &c->binary_header;
extlen = header->response.extlen;
keylen = header->response.keylen;
bodylen = header->response.bodylen;
status = ntohs(header->response.status);
opcode = header->response.opcode;
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_b2b_downstream_nread %x %x %d %d %u %d %x\n",
c->sfd, c->cmd, opcode, extlen, keylen, bodylen, c->noreply, status);
}
d = c->extra;
cb_assert(d != NULL);
cb_assert(d->ptd != NULL);
cb_assert(d->ptd->proxy != NULL);
/* TODO: Need to handle quiet binary command error response, */
/* in the right order. */
/* TODO: Need to handle not-my-vbucket error during a quiet cmd. */
uc = d->upstream_conn;
it = c->item;
/* Clear c->item because we either move it to the upstream or */
/* item_remove() it on error. */
c->item = NULL;
cb_assert(it != NULL);
cb_assert(it->refcount == 1);
if (cproxy_binary_ignore_reply(c, header, it)) {
return;
}
if (c->noreply) {
conn_set_state(c, conn_new_cmd);
} else {
conn_set_state(c, conn_pause);
if (opcode == PROTOCOL_BINARY_CMD_NOOP ||
opcode == PROTOCOL_BINARY_CMD_FLUSH) {
goto done;
}
if (opcode == PROTOCOL_BINARY_CMD_STAT) {
if (status == PROTOCOL_BINARY_RESPONSE_SUCCESS) {
if (keylen > 0) {
if (d->merger != NULL) {
char *key = (ITEM_data(it)) + sizeof(*header) + extlen;
char *val = key + keylen;
protocol_stats_merge_name_val(d->merger, "STAT", 4,
key, keylen,
val, bodylen - keylen - extlen);
}
conn_set_state(c, conn_new_cmd); /* Get next STATS response. */
}
}
goto done;
}
/* If the client is still there, we should handle */
/* a not-my-vbucket error with a possible retry. */
if (uc != NULL &&
status == PROTOCOL_BINARY_RESPONSE_NOT_MY_VBUCKET) {
int max_retries;
protocol_binary_request_header *req;
int vbucket;
int sindex;
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_b2b_downstream_nread not-my-vbucket, "
"cmd: %x %d\n",
c->sfd, header->response.opcode, uc->item != NULL);
}
cb_assert(uc->item != NULL);
req = (protocol_binary_request_header *)ITEM_data((item*)uc->item);
vbucket = ntohs(req->request.reserved);
sindex = downstream_conn_index(d, c);
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_b2b_downstream_nread not-my-vbucket, "
"cmd: %x not multi-key get, sindex %d, vbucket %d, retries %d\n",
c->sfd, header->response.opcode,
sindex, vbucket, uc->cmd_retries);
}
mcs_server_invalid_vbucket(&d->mst, sindex, vbucket);
/* As long as the upstream is still open and we haven't */
/* retried too many times already. */
max_retries = cproxy_max_retries(d);
if (uc->cmd_retries < max_retries) {
uc->cmd_retries++;
d->upstream_retry++;
d->ptd->stats.stats.tot_retry_vbucket++;
goto done;
}
if (settings.verbose > 2) {
moxi_log_write("%d: cproxy_process_b2b_downstream_nread not-my-vbucket, "
"cmd: %x skipping retry %d >= %d\n",
c->sfd, header->response.opcode, uc->cmd_retries,
max_retries);
}
}
}
/* Write the response to the upstream connection. */
if (uc != NULL) {
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_b2b_downstream_nread got %u\n",
c->sfd, it->nbytes);
cproxy_dump_header(c->sfd, ITEM_data(it));
}
if (add_conn_item(uc, it) == true) {
it->refcount++;
if (add_iov(uc, ITEM_data(it), it->nbytes) == 0) {
/* If we got a quiet response, however, don't change the */
/* upstream connection's state (should be in paused state), */
/* as we expect the downstream server to provide a */
/* verbal/non-quiet response that moves the downstream */
/* conn through the conn_pause countdown codepath. */
if (c->noreply == false) {
cproxy_update_event_write(d, uc);
conn_set_state(uc, conn_mwrite);
}
goto done;
}
}
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(uc);
}
done:
if (it != NULL) {
item_remove(it);
}
}
/* Called when we receive a binary response header from
* a downstream server, via try_read_command()/drive_machine().
*/
void cproxy_process_b2b_downstream(conn *c) {
char *ikey;
int ikeylen;
downstream *d;
int extlen;
int keylen;
uint32_t bodylen;
cb_assert(c != NULL);
cb_assert(c->cmd >= 0);
cb_assert(c->next == NULL);
cb_assert(c->item == NULL);
cb_assert(IS_BINARY(c->protocol));
cb_assert(IS_PROXY(c->protocol));
cb_assert(c->substate == bin_no_state);
d = c->extra;
cb_assert(d);
c->cmd_curr = -1;
c->cmd_start = NULL;
c->cmd_start_time = msec_current_time;
c->cmd_retries = 0;
extlen = c->binary_header.request.extlen;
keylen = c->binary_header.request.keylen;
bodylen = c->binary_header.request.bodylen;
if (settings.verbose > 2) {
moxi_log_write("<%d cproxy_process_b2b_downstream %x %d %d %u\n",
c->sfd, c->cmd, extlen, keylen, bodylen);
}
cb_assert(bodylen >= (uint32_t) keylen + extlen);
process_bin_noreply(c); /* Map quiet c->cmd values into non-quiet. */
/* Our approach is to read everything we can before */
/* getting into big switch/case statements for the */
/* actual processing. */
/* Alloc an item and continue with an rest-of-body nread if */
/* necessary. The item will hold the entire response message */
/* (the header + body). */
ikey = "q";
ikeylen = 1;
c->item = item_alloc(ikey, ikeylen, 0, 0,
sizeof(c->binary_header) + bodylen);
if (c->item != NULL) {
item *it = c->item;
void *rb = c->rcurr;
cb_assert(it->refcount == 1);
memcpy(ITEM_data(it), rb, sizeof(c->binary_header));
if (bodylen > 0) {
c->ritem = ITEM_data(it) + sizeof(c->binary_header);
c->rlbytes = bodylen;
c->substate = bin_read_set_value;
conn_set_state(c, conn_nread);
} else {
/* Since we have no body bytes, we can go immediately to */
/* the nread completed processing step. */
cproxy_process_b2b_downstream_nread(c);
}
} else {
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
}
}
/* 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;
}
/* Forward a simple one-liner command downstream.
* For example, get, incr/decr, delete, etc.
* The response, though, might be a simple line or
* multiple VALUE+END lines.
*/
bool cproxy_forward_a2a_simple_downstream(downstream *d,
char *command, 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(command != NULL);
cb_assert(uc != NULL);
cb_assert(uc->item == NULL);
cb_assert(uc->cmd_curr != (protocol_binary_command) -1);
cb_assert(d->multiget == NULL);
cb_assert(d->merger == NULL);
/* Handles get and gets. */
if (uc->cmd_curr == PROTOCOL_BINARY_CMD_GETK ||
uc->cmd_curr == PROTOCOL_BINARY_CMD_GETKQ ||
uc->cmd_curr == PROTOCOL_BINARY_CMD_GETL) {
/* Only use front_cache for 'get', not for 'gets'. */
mcache *front_cache =
(command[3] == ' ') ? &d->ptd->proxy->front_cache : NULL;
return multiget_ascii_downstream(d, uc,
a2a_multiget_start,
a2a_multiget_skey,
a2a_multiget_end,
front_cache);
}
cb_assert(uc->next == NULL);
if (uc->cmd_curr == PROTOCOL_BINARY_CMD_FLUSH) {
return cproxy_broadcast_a2a_downstream(d, command, uc,
"OK\r\n");
}
if (uc->cmd_curr == PROTOCOL_BINARY_CMD_STAT) {
if (strncmp(command + 5, " reset", 6) == 0) {
return cproxy_broadcast_a2a_downstream(d, command, uc,
"RESET\r\n");
}
if (cproxy_broadcast_a2a_downstream(d, command, uc,
"END\r\n")) {
d->merger = genhash_init(512, skeyhash_ops);
return true;
} else {
return false;
}
}
/* TODO: Inefficient repeated scan_tokens. */
int cmd_len = 0;
token_t tokens[MAX_TOKENS];
size_t ntokens = scan_tokens(command, tokens, MAX_TOKENS, &cmd_len);
char *key = tokens[KEY_TOKEN].value;
int key_len = tokens[KEY_TOKEN].length;
if (ntokens <= 1) { /* This was checked long ago, while parsing */
cb_assert(false); /* the upstream conn. */
return false;
}
/* Assuming we're already connected to downstream. */
if (!strcmp(command, "version")) {
/* fake key for version command handling */
key = "v";
key_len = 1;
}
conn *c = cproxy_find_downstream_conn(d, key, key_len, NULL);
if (c != NULL) {
if (cproxy_prep_conn_for_write(c)) {
cb_assert(c->state == conn_pause);
out_string(c, command);
if (settings.verbose > 1) {
moxi_log_write("forwarding to %d, noreply %d\n",
c->sfd, uc->noreply);
}
if (update_event(c, EV_WRITE | EV_PERSIST)) {
d->downstream_used_start = 1;
d->downstream_used = 1;
if (cproxy_dettach_if_noreply(d, uc) == false) {
cproxy_start_downstream_timeout(d, c);
} else {
c->write_and_go = conn_pause;
/* Do mcache_delete() here only during a noreply, */
/* otherwise for with-reply requests, we could */
/* be in a race with other clients repopulating */
/* the front_cache. For with-reply requests, we */
/* clear the front_cache when we get a success reply. */
cproxy_front_cache_delete(d->ptd, key, key_len);
}
return true;
}
if (settings.verbose > 1) {
moxi_log_write("Couldn't update cproxy write event\n");
}
d->ptd->stats.stats.err_oom++;
cproxy_close_conn(c);
} else {
d->ptd->stats.stats.err_downstream_write_prep++;
cproxy_close_conn(c);
}
}
return false;
}
