/* gets information back from any of the nodes in the cluster */
int
citrusleaf_info_cluster(as_cluster *cluster, char *names, char **values_r, bool send_asis, bool check_bounds, int timeout_ms)
{
if (timeout_ms == 0) {
timeout_ms = 100; // milliseconds
}
uint64_t start = cf_getms();
uint64_t end = start + timeout_ms;
int ret = -1;
as_nodes* nodes = as_nodes_reserve(cluster);
for (uint32_t i = 0; i < nodes->size; i++) {
as_node* node = nodes->array[i];
struct sockaddr_in* sa_in = as_node_get_address(node);
char* values = 0;
if (citrusleaf_info_host_auth(cluster, sa_in, names, &values, (int)(end - cf_getms()), send_asis, check_bounds) == 0) {
*values_r = values;
ret = 0;
break;
}
if (cf_getms() >= end) {
ret = -2;
break;
}
}
as_nodes_release(nodes);
return ret;
}
bool
udf_timer_timedout(const as_timer *as_tt)
{
time_tracker *tt = (time_tracker *)pthread_getspecific(timer_tlskey);
if (!tt || !tt->end_time || !tt->udata) {
return true;
}
bool timedout = (cf_getms() > tt->end_time(tt));
if (timedout) {
cf_debug(AS_UDF, "UDF Timed Out [%ld:%ld]", cf_getms(), tt->end_time(tt));
return true;
}
return false;
}
as_status
as_event_command_execute(as_event_command* cmd, as_error* err)
{
ck_pr_inc_32(&cmd->cluster->async_pending);
// Only do this after the above increment to avoid a race with as_cluster_destroy().
if (!cmd->cluster->valid) {
as_event_command_free(cmd);
return as_error_set_message(err, AEROSPIKE_ERR_CLIENT, "Client shutting down");
}
// Use pointer comparison for performance.
// If portability becomes an issue, use "pthread_equal(event_loop->thread, pthread_self())"
// instead.
if (cmd->event_loop->thread == pthread_self()) {
// We are already in event loop thread, so start processing.
as_event_command_begin(cmd);
}
else {
if (cmd->timeout_ms) {
// Store current time in first 8 bytes which is not used yet.
*(uint64_t*)cmd = cf_getms();
}
// Send command through queue so it can be executed in event loop thread.
if (! as_event_send(cmd)) {
as_event_command_free(cmd);
return as_error_set_message(err, AEROSPIKE_ERR_CLIENT, "Failed to queue command");
}
}
return AEROSPIKE_OK;
}
/* Processing reads when they return from aio_read */
static void process_read(as_async_info_t *info)
{
if(!g_running)
{
return;
}
cf_atomic_int_decr(&g_read_reqs_queued);
uint64_t stop_time = cf_getms();
fd_put(info->p_readreq.p_device, info->fd);
if (stop_time != -1)
{
histogram_insert_data_point(g_p_raw_read_histogram,
safe_delta_ms(info->raw_start_time, stop_time));
histogram_insert_data_point(g_p_read_histogram,
safe_delta_ms(info->p_readreq.start_time, stop_time));
histogram_insert_data_point(
info->p_readreq.p_device->p_raw_read_histogram,
safe_delta_ms(info->raw_start_time, stop_time));
}
if (g_use_valloc && info->p_buffer)
{
free(info->p_buffer);
}
uintptr_t temp = (uintptr_t)info;
cf_queue_push(async_info_queue, (void*)&temp);
}
//------------------------------------------------
// Do one large block write operation and report.
//
static void write_and_report_large_block(device* p_device) {
salter* p_salter;
if (g_num_write_buffers > 1) {
p_salter = &g_salters[rand_32() % g_num_write_buffers];
pthread_mutex_lock(&p_salter->lock);
*(uint32_t*)p_salter->p_buffer = p_salter->stamp++;
}
else {
p_salter = &g_salters[0];
}
uint64_t offset = random_large_block_offset(p_device);
uint64_t start_time = cf_getms();
uint64_t stop_time = write_to_device(p_device, offset,
g_large_block_ops_bytes, p_salter->p_buffer);
if (g_num_write_buffers > 1) {
pthread_mutex_unlock(&p_salter->lock);
}
if (stop_time != -1) {
histogram_insert_data_point(g_p_large_block_write_histogram,
safe_delta_ms(start_time, stop_time));
}
}
// Process a batch request.
static void
batch_process_request(batch_transaction* btr)
{
// Keep the reaper at bay.
btr->fd_h->last_used = cf_getms();
cf_buf_builder* bb = 0;
batch_build_response(btr, &bb);
int fd = btr->fd_h->fd;
if (bb) {
int brv = batch_send_header(fd, bb->used_sz);
if (brv == 0) {
brv = batch_send(fd, bb->buf, bb->used_sz, MSG_NOSIGNAL | MSG_MORE);
if (brv == 0) {
brv = batch_send_final(fd, 0);
}
}
cf_buf_builder_free(bb);
}
else {
cf_info(AS_BATCH, " batch request: returned no local responses");
batch_send_final(fd, 0);
}
batch_transaction_done(btr);
}
int
citrusleaf_info_cluster_foreach(
as_cluster *cluster, const char *command, bool send_asis, bool check_bounds, int timeout_ms, void *udata,
bool (*callback)(const as_node * node, const struct sockaddr_in * sa_in, const char *command, char *value, void *udata)
)
{
//Usage Notes:
//udata = memory allocated by caller, passed back to the caller callback function, ufn()
//command = command string, memory allocated by caller, caller must free it, passed to server for execution
//value = memory allocated by c-client for caller, caller must free it after using it.
if (timeout_ms == 0) {
timeout_ms = 100; // milliseconds
}
uint64_t start = cf_getms();
uint64_t end = start + timeout_ms;
int ret = 0;
as_nodes* nodes = as_nodes_reserve(cluster);
for (uint32_t i = 0; i < nodes->size; i++) {
as_node* node = nodes->array[i];
struct sockaddr_in* sa_in = as_node_get_address(node);
char* value = 0;
if (citrusleaf_info_host_auth(cluster, sa_in, (char *)command, &value, (int)(end - cf_getms()), send_asis, check_bounds) == 0) {
bool status = callback(node, sa_in, command, value, udata);
if (value) {
free(value);
}
if(! status) {
ret = -1;
break;
}
}
if (cf_getms() >= end) {
ret = -2;
break;
}
}
as_nodes_release(nodes);
return ret;
}
uint64_t
udf_timer_timeslice(const as_timer *as_tt)
{
time_tracker *tt = (time_tracker *)pthread_getspecific(timer_tlskey);
if (!tt || !tt->end_time || !tt->udata) {
return true;
}
uint64_t timeslice = tt->end_time(tt) - cf_getms();
return (timeslice > 0) ? timeslice : 1;
}
//------------------------------------------------
// Do one large block read operation and report.
//
static void read_and_report_large_block(device* p_device) {
uint64_t offset = random_large_block_offset(p_device);
uint64_t start_time = cf_getms();
uint64_t stop_time = read_from_device(p_device, offset,
g_large_block_ops_bytes, p_device->p_large_block_read_buffer);
if (stop_time != -1) {
histogram_insert_data_point(g_p_large_block_read_histogram,
safe_delta_ms(start_time, stop_time));
}
}
static int
wait_socket(as_socket_fd fd, uint32_t socket_timeout, uint64_t deadline, bool read)
{
as_poll poll;
as_poll_init(&poll, fd);
uint32_t timeout;
int rv;
while (true) {
if (deadline > 0) {
uint64_t now = cf_getms();
if (now >= deadline) {
rv = 1; // timeout
break;
}
timeout = (uint32_t)(deadline - now);
if (socket_timeout > 0 && socket_timeout < timeout) {
timeout = socket_timeout;
}
}
else {
timeout = socket_timeout;
}
rv = as_poll_socket(&poll, fd, timeout, read);
if (rv > 0) {
rv = 0; // success
break;
}
if (rv < 0) {
break; // error
}
// rv == 0 timeout. continue in case timed out before real timeout.
}
as_poll_destroy(&poll);
return rv;
}
static as_status
as_admin_read_list(aerospike* as, as_error* err, const as_policy_admin* policy, uint8_t* command, uint8_t* end, as_admin_parse_fn parse_fn, as_vector* list)
{
int timeout_ms = (policy)? policy->timeout : as->config.policies.admin.timeout;
if (timeout_ms <= 0) {
timeout_ms = DEFAULT_TIMEOUT;
}
uint64_t deadline_ms = cf_getms() + timeout_ms;
as_node* node = as_node_get_random(as->cluster);
if (! node) {
return as_error_set_message(err, AEROSPIKE_ERR_CLIENT, "Failed to find server node.");
}
int fd;
as_status status = as_node_get_connection(err, node, &fd);
if (status) {
as_node_release(node);
return status;
}
status = as_admin_send(err, fd, command, end, deadline_ms);
if (status) {
as_close(fd);
as_node_release(node);
return status;
}
status = as_admin_read_blocks(err, fd, deadline_ms, parse_fn, list);
if (status) {
as_close(fd);
as_node_release(node);
return status;
}
as_node_put_connection(node, fd);
as_node_release(node);
return status;
}
//------------------------------------------------
// Do one device write operation.
//
static uint64_t write_to_device(device* p_device, uint64_t offset,
uint32_t size, uint8_t* p_buffer) {
int fd = fd_get(p_device);
if (fd == -1) {
return -1;
}
if (lseek(fd, offset, SEEK_SET) != offset ||
write(fd, p_buffer, size) != (ssize_t)size) {
close(fd);
fprintf(stdout, "ERROR: seek & write\n");
return -1;
}
uint64_t stop_ms = cf_getms();
fd_put(p_device, fd);
return stop_ms;
}
void
repl_write_reset_rw(rw_request* rw, as_transaction* tr, repl_write_done_cb cb)
{
// Reset rw->from.any which was set null in tr setup. (And note that
// tr->from.any will be null here in respond-on-master-complete mode.)
rw->from.any = tr->from.any;
// Needed for response to origin.
rw->generation = tr->generation;
rw->void_time = tr->void_time;
rw->repl_write_cb = cb;
// TODO - is this better than not resetting? Note - xmit_ms not volatile.
rw->xmit_ms = cf_getms() + g_config.transaction_retry_ms;
rw->retry_interval_ms = g_config.transaction_retry_ms;
for (int i = 0; i < rw->n_dest_nodes; i++) {
rw->dest_complete[i] = false;
}
}
//------------------------------------------------
// Runs in every device large-block write thread,
// executes large-block writes at a constant rate.
//
static void* run_large_block_writes(void* pv_device) {
device* p_device = (device*)pv_device;
uint64_t count = 0;
while (g_running) {
write_and_report_large_block(p_device);
count++;
int sleep_ms = (int)
(((count * 1000 * g_num_devices) / g_large_block_ops_per_sec) -
(cf_getms() - g_run_start_ms));
if (sleep_ms > 0) {
usleep((uint32_t)sleep_ms * 1000);
}
}
return (0);
}
static int
as_read_users(aerospike* as, const as_policy_admin* policy, uint8_t* buffer, uint8_t* end, as_vector* /*<as_user_roles*>*/ users)
{
int timeout_ms = (policy)? policy->timeout : as->config.policies.admin.timeout;
if (timeout_ms <= 0) {
timeout_ms = DEFAULT_TIMEOUT;
}
uint64_t deadline_ms = cf_getms() + timeout_ms;
as_node* node = as_node_get_random(as->cluster);
if (! node) {
return CITRUSLEAF_FAIL_CLIENT;
}
int fd;
int status = as_node_get_connection(node, &fd);
if (status) {
as_node_release(node);
return status;
}
if (as_send(fd, buffer, end, deadline_ms, timeout_ms)) {
cf_close(fd);
as_node_release(node);
return CITRUSLEAF_FAIL_TIMEOUT;
}
status = as_read_user_blocks(fd, buffer, deadline_ms, timeout_ms, users);
if (status >= 0) {
as_node_put_connection(node, fd);
}
else {
cf_close(fd);
}
as_node_release(node);
return status;
}
void
repl_write_setup_rw(rw_request* rw, as_transaction* tr,
repl_write_done_cb repl_write_cb, timeout_done_cb timeout_cb)
{
rw->msgp = tr->msgp;
tr->msgp = NULL;
rw->msg_fields = tr->msg_fields;
rw->origin = tr->origin;
rw->from_flags = tr->from_flags;
rw->from.any = tr->from.any;
rw->from_data.any = tr->from_data.any;
tr->from.any = NULL;
rw->start_time = tr->start_time;
rw->benchmark_time = tr->benchmark_time;
as_partition_reservation_copy(&rw->rsv, &tr->rsv);
// Hereafter, rw_request must release reservation - happens in destructor.
rw->end_time = tr->end_time;
rw->generation = tr->generation;
rw->void_time = tr->void_time;
rw->repl_write_cb = repl_write_cb;
rw->timeout_cb = timeout_cb;
rw->xmit_ms = cf_getms() + g_config.transaction_retry_ms;
rw->retry_interval_ms = g_config.transaction_retry_ms;
for (int i = 0; i < rw->n_dest_nodes; i++) {
rw->dest_complete[i] = false;
}
// Allow retransmit thread to destroy rw_request as soon as we unlock.
rw->is_set_up = true;
}
static int
as_execute(aerospike* as, const as_policy_admin* policy, uint8_t* buffer, uint8_t* end)
{
int timeout_ms = (policy)? policy->timeout : as->config.policies.admin.timeout;
if (timeout_ms <= 0) {
timeout_ms = DEFAULT_TIMEOUT;
}
uint64_t deadline_ms = cf_getms() + timeout_ms;
as_node* node = as_node_get_random(as->cluster);
if (! node) {
return CITRUSLEAF_FAIL_CLIENT;
}
int fd;
int status = as_node_get_connection(node, &fd);
if (status) {
as_node_release(node);
return status;
}
if (as_send(fd, buffer, end, deadline_ms, timeout_ms)) {
cf_close(fd);
as_node_release(node);
return CITRUSLEAF_FAIL_TIMEOUT;
}
if (cf_socket_read_timeout(fd, buffer, HEADER_SIZE, deadline_ms, timeout_ms)) {
cf_close(fd);
as_node_release(node);
return CITRUSLEAF_FAIL_TIMEOUT;
}
as_node_put_connection(node, fd);
as_node_release(node);
return buffer[RESULT_CODE];
}
int
as_authenticate(int fd, const char* user, const char* credential, int timeout_ms)
{
uint8_t buffer[STACK_BUF_SZ];
uint8_t* p = buffer + 8;
p = write_header(p, AUTHENTICATE, 2);
p = write_field_string(p, USER, user);
p = write_field_string(p, CREDENTIAL, credential);
if (timeout_ms == 0) {
timeout_ms = DEFAULT_TIMEOUT;
}
uint64_t deadline_ms = cf_getms() + timeout_ms;
if (as_send(fd, buffer, p, deadline_ms, timeout_ms)) {
return CITRUSLEAF_FAIL_TIMEOUT;
}
if (cf_socket_read_timeout(fd, buffer, HEADER_SIZE, deadline_ms, timeout_ms)) {
return CITRUSLEAF_FAIL_TIMEOUT;
}
return buffer[RESULT_CODE];
}
// Make a callback for a specified number of elements in the tree, from outside
// the tree lock.
void
as_index_reduce_partial(as_index_tree *tree, uint32_t sample_count,
as_index_reduce_fn cb, void *udata)
{
pthread_mutex_lock(&tree->reduce_lock);
// For full reduce, get the number of elements inside the tree lock.
if (sample_count == AS_REDUCE_ALL) {
sample_count = tree->elements;
}
if (sample_count == 0) {
pthread_mutex_unlock(&tree->reduce_lock);
return;
}
size_t sz = sizeof(as_index_ph_array) +
(sizeof(as_index_ph) * sample_count);
as_index_ph_array *v_a;
uint8_t buf[64 * 1024];
if (sz > 64 * 1024) {
v_a = cf_malloc(sz);
if (! v_a) {
pthread_mutex_unlock(&tree->reduce_lock);
return;
}
}
else {
v_a = (as_index_ph_array*)buf;
}
v_a->alloc_sz = sample_count;
v_a->pos = 0;
uint64_t start_ms = cf_getms();
// Recursively, fetch all the value pointers into this array, so we can make
// all the callbacks outside the big lock.
if (tree->root->left_h != tree->sentinel_h) {
as_index_reduce_traverse(tree, tree->root->left_h, tree->sentinel_h,
v_a);
}
cf_debug(AS_INDEX, "as_index_reduce_traverse took %"PRIu64" ms",
cf_getms() - start_ms);
pthread_mutex_unlock(&tree->reduce_lock);
for (uint32_t i = 0; i < v_a->pos; i++) {
as_index_ref r_ref;
r_ref.skip_lock = false;
r_ref.r = v_a->indexes[i].r;
r_ref.r_h = v_a->indexes[i].r_h;
olock_vlock(g_config.record_locks, &r_ref.r->key, &r_ref.olock);
cf_atomic_int_incr(&g_config.global_record_lock_count);
// Callback MUST call as_record_done() to unlock and release record.
cb(&r_ref, udata);
}
if (v_a != (as_index_ph_array*)buf) {
cf_free(v_a);
}
}
// Incoming messages start here.
// - Could get a request that we need to service.
// - Could get a response to one of our requests - need to find the request and
// send the real response to the remote end.
int
proxy_msg_fn(cf_node id, msg *m, void *udata)
{
int rv;
if (cf_rc_count((void*)m) == 0) {
cf_debug(AS_PROXY, " proxy_msg_fn was given a refcount 0 message! Someone has been naugty %p", m);
return -1;
}
uint32_t op = 99999;
msg_get_uint32(m, PROXY_FIELD_OP, &op);
uint32_t transaction_id = 0;
msg_get_uint32(m, PROXY_FIELD_TID, &transaction_id);
cf_detail(AS_PROXY, "received proxy message: tid %d type %d from %"PRIx64, transaction_id, op, id);
switch (op) {
case PROXY_OP_REQUEST:
{
cf_atomic_int_incr(&g_config.proxy_action);
#ifdef DEBUG
cf_debug(AS_PROXY, "Proxy_msg: received request");
#ifdef DEBUG_VERBOSE
msg_dump(m, "incoming proxy msg");
#endif
#endif
cf_digest *key;
size_t sz = 0;
if (0 != msg_get_buf(m, PROXY_FIELD_DIGEST, (byte **) &key, &sz, MSG_GET_DIRECT)) {
cf_info(AS_PROXY, "proxy msg function: no digest, problem");
as_fabric_msg_put(m);
return 0;
}
cl_msg *msgp;
size_t as_msg_sz = 0;
if (0 != msg_get_buf(m, PROXY_FIELD_AS_PROTO, (byte **) &msgp, &as_msg_sz, MSG_GET_COPY_MALLOC)) {
cf_info(AS_PROXY, "proxy msg function: no as msg, problem");
as_fabric_msg_put(m);
return 0;
}
uint64_t cluster_key = 0;
if (0 != msg_get_uint64(m, PROXY_FIELD_CLUSTER_KEY, &cluster_key)) {
cf_info(AS_PROXY, "proxy msg function: no cluster key, problem");
as_fabric_msg_put(m);
return 0;
}
// This is allowed to fail - this is a new field, and gets defaulted
// to 0 if it doesn't exist.
uint32_t timeout_ms = 0;
msg_get_uint32(m, PROXY_FIELD_TIMEOUT_MS, &timeout_ms);
// cf_info(AS_PROXY, "proxy msg: received timeout_ms of %d",timeout_ms);
// Put the as_msg on the normal queue for processing.
// INIT_TR
as_transaction tr;
as_transaction_init(&tr, key, msgp);
tr.incoming_cluster_key = cluster_key;
tr.end_time = (timeout_ms != 0) ? ((uint64_t)timeout_ms * 1000000) + tr.start_time : 0;
tr.proxy_node = id;
tr.proxy_msg = m;
// Check here if this is shipped op.
uint32_t info = 0;
msg_get_uint32(m, PROXY_FIELD_INFO, &info);
if (info & PROXY_INFO_SHIPPED_OP) {
tr.flag |= AS_TRANSACTION_FLAG_SHIPPED_OP;
cf_detail_digest(AS_PROXY, &tr.keyd, "SHIPPED_OP WINNER Operation Received");
} else {
cf_detail_digest(AS_PROXY, &tr.keyd, "Received Proxy Request digest tid(%d)", tr.trid);
}
MICROBENCHMARK_RESET();
thr_tsvc_enqueue(&tr);
}
break;
case PROXY_OP_RESPONSE:
{
#ifdef DEBUG
// Got the response from the actual endpoint.
cf_debug(AS_PROXY, " proxy: received response! tid %d node %"PRIx64, transaction_id, id);
#ifdef DEBUG_VERBOSE
msg_dump(m, "incoming proxy response");
#endif
#endif
// Look up the element.
proxy_request pr;
bool free_msg = true;
if (SHASH_OK == shash_get_and_delete(g_proxy_hash, &transaction_id, &pr)) {
// Found the element (sometimes we get two acks so it's OK for
// an ack to not find the transaction).
if (pr.wr) {
as_proxy_shipop_response_hdlr(m, &pr, &free_msg);
} else {
as_proto *proto;
size_t proto_sz;
if (0 != msg_get_buf(m, PROXY_FIELD_AS_PROTO, (byte **) &proto, &proto_sz, MSG_GET_DIRECT)) {
cf_info(AS_PROXY, "msg get buf failed!");
}
#ifdef DEBUG_VERBOSE
cf_debug(AS_PROXY, "proxy: sending proto response: ptr %p sz %"PRIu64" %d", proto, proto_sz, pr.fd);
for (size_t _i = 0; _i < proto_sz; _i++) {
fprintf(stderr, " %x", ((byte *)proto)[_i]);
if (_i % 16 == 15) {
fprintf(stderr, "\n");
}
}
#endif
#ifdef EXTRA_CHECKS
as_proto proto_copy = *proto;
as_proto_swap(&proto_copy);
if (proto_copy.sz + 8 != proto_sz) {
cf_info(AS_PROXY, "BONE BONE BONE!!!");
cf_info(AS_PROXY, "proto sz: %"PRIu64" sz %u", (uint64_t) proto_copy.sz, proto_sz);
}
#endif
// Write to the file descriptor.
cf_detail(AS_PROXY, "direct write fd %d", pr.fd_h->fd);
cf_assert(pr.fd_h->fd, AS_PROXY, CF_WARNING, "attempted write to fd 0");
if (pr.batch_shared) {
cf_digest* digest;
size_t digest_sz = 0;
if (msg_get_buf(pr.fab_msg, PROXY_FIELD_DIGEST, (byte **)&digest, &digest_sz, MSG_GET_DIRECT) == 0) {
as_batch_add_proxy_result(pr.batch_shared, pr.batch_index, digest, (cl_msg*)proto, proto_sz);
as_proxy_set_stat_counters(0);
}
else {
cf_warning(AS_PROXY, "Failed to find batch proxy digest %u", transaction_id);
as_batch_add_error(pr.batch_shared, pr.batch_index, AS_PROTO_RESULT_FAIL_UNKNOWN);
as_proxy_set_stat_counters(-1);
}
cf_hist_track_insert_data_point(g_config.px_hist, pr.start_time);
}
else {
size_t pos = 0;
while (pos < proto_sz) {
rv = send(pr.fd_h->fd, (((uint8_t *)proto) + pos), proto_sz - pos, MSG_NOSIGNAL);
if (rv > 0) {
pos += rv;
}
else if (rv < 0) {
if (errno != EWOULDBLOCK) {
// Common message when a client aborts.
cf_debug(AS_PROTO, "protocol proxy write fail: fd %d sz %d pos %d rv %d errno %d", pr.fd_h->fd, proto_sz, pos, rv, errno);
shutdown(pr.fd_h->fd, SHUT_RDWR);
as_proxy_set_stat_counters(-1);
goto SendFin;
}
usleep(1); // yield
}
else {
cf_info(AS_PROTO, "protocol write fail zero return: fd %d sz %d pos %d ", pr.fd_h->fd, proto_sz, pos);
shutdown(pr.fd_h->fd, SHUT_RDWR);
as_proxy_set_stat_counters(-1);
goto SendFin;
}
}
as_proxy_set_stat_counters(0);
SendFin:
cf_hist_track_insert_data_point(g_config.px_hist, pr.start_time);
// Return the fabric message or the direct file descriptor -
// after write and complete.
pr.fd_h->t_inprogress = false;
AS_RELEASE_FILE_HANDLE(pr.fd_h);
pr.fd_h = 0;
}
as_fabric_msg_put(pr.fab_msg);
pr.fab_msg = 0;
}
}
else {
cf_debug(AS_PROXY, "proxy: received result but no transaction, tid %d", transaction_id);
as_proxy_set_stat_counters(-1);
}
if (free_msg) {
as_fabric_msg_put(m);
}
}
break;
case PROXY_OP_REDIRECT:
{
// Sometimes the destination we proxied a request to isn't able to
// satisfy it (for example, their copy of the partition in question
// might be desync).
cf_node new_dst = 0;
msg_get_uint64(m, PROXY_FIELD_REDIRECT, &new_dst);
cf_detail(AS_PROXY, "proxy redirect message: transaction %d to node %"PRIx64, transaction_id, new_dst);
// Look in the proxy retransmit hash for the tid.
proxy_request *pr;
pthread_mutex_t *pr_lock;
int r = 0;
if (0 != (r = shash_get_vlock(g_proxy_hash, &transaction_id, (void **)&pr, &pr_lock))) {
cf_debug(AS_PROXY, "redirect: could not find transaction %d", transaction_id);
as_fabric_msg_put(m);
return -1;
}
if (g_config.self_node == new_dst) {
// Although we don't know we're the final destination, undo the
// proxy-nature and put back on the main queue. Dangerous, as it
// leaves open the possibility of a looping message.
cf_digest *key;
size_t sz = 0;
if (0 != msg_get_buf(pr->fab_msg, PROXY_FIELD_DIGEST, (byte **) &key, &sz, MSG_GET_DIRECT)) {
cf_warning(AS_PROXY, "op_redirect: proxy msg function: no digest, problem");
pthread_mutex_unlock(pr_lock);
as_fabric_msg_put(m);
return -1;
}
cl_msg *msgp;
sz = 0;
if (0 != msg_get_buf(pr->fab_msg, PROXY_FIELD_AS_PROTO, (byte **) &msgp, &sz, MSG_GET_COPY_MALLOC)) {
cf_warning(AS_PROXY, "op_redirect: proxy msg function: no as proto, problem");
pthread_mutex_unlock(pr_lock);
as_fabric_msg_put(m);
return -1;
}
// Put the as_msg on the normal queue for processing.
// INIT_TR
as_transaction tr;
as_transaction_init(&tr, key, msgp);
tr.start_time = pr->start_time; // start time
tr.end_time = pr->end_time;
tr.proto_fd_h = pr->fd_h;
tr.batch_shared = pr->batch_shared;
tr.batch_index = pr->batch_index;
MICROBENCHMARK_RESET();
thr_tsvc_enqueue(&tr);
as_fabric_msg_put(pr->fab_msg);
shash_delete_lockfree(g_proxy_hash, &transaction_id);
}
else {
// Change the destination, update the retransmit time.
pr->dest = new_dst;
pr->xmit_ms = cf_getms() + 1;
// Send it.
msg_incr_ref(pr->fab_msg);
if (0 != (rv = as_fabric_send(pr->dest, pr->fab_msg, AS_FABRIC_PRIORITY_MEDIUM))) {
cf_debug(AS_PROXY, "redirect: change destination: %"PRIx64" send error %d", pr->dest, rv);
as_fabric_msg_put(pr->fab_msg);
}
}
pthread_mutex_unlock(pr_lock);
}
as_fabric_msg_put(m);
break;
default:
cf_debug(AS_PROXY, "proxy_msg_fn: received unknown, unsupported message %d from remote endpoint", op);
msg_dump(m, "proxy received unknown msg");
as_fabric_msg_put(m);
break;
} // end switch
return 0;
} // end proxy_msg_fn()
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;
}
//Same as do_the_full_monte, but only till the command is sent to the node.
//Most of the code is duplicated. Bad.
int
cl_do_async_monte(cl_cluster *asc, int info1, int info2, const char *ns, const char *set, const cl_object *key,
const cf_digest *digest, cl_bin **values, cl_operator operator, cl_operation **operations,
int *n_values, uint32_t *cl_gen, const cl_write_parameters *cl_w_p, uint64_t *trid, void *udata)
{
cl_async_work *workitem = NULL;
uint8_t wr_stack_buf[STACK_BUF_SZ];
uint8_t *wr_buf = wr_stack_buf;
size_t wr_buf_sz = sizeof(wr_stack_buf);
int progress_timeout_ms;
uint64_t deadline_ms;
uint64_t starttime, endtime;
bool network_error;
int fd = -1;
int rv = CITRUSLEAF_FAIL_CLIENT; //Assume that this is a failure;
// as_msg msg;
cf_digest d_ret;
cl_cluster_node *node = 0;
#if ONEASYNCFD
if (shash_get_size(g_cl_async_hashtab) >= g_async_h_szlimit) {
//cf_error("Async hashtab is full. Cannot insert any more elements");
return CITRUSLEAF_FAIL_ASYNCQ_FULL;
}
#else
//If the async buffer is at the max limit, do not entertain more requests.
if (cf_queue_sz(g_cl_async_q) >= cf_atomic32_get(g_async_q_szlimit)) {
//cf_error("Async buffer is full. Cannot insert any more elements");
return CITRUSLEAF_FAIL_ASYNCQ_FULL;
}
#endif
//Allocate memory for work item that will be added to the async work list
if (cf_queue_sz(g_cl_workitems_freepool_q) > 0) {
cf_queue_pop(g_cl_workitems_freepool_q, &workitem, CF_QUEUE_FOREVER);
} else {
workitem = malloc(sizeof(cl_async_work));
if (workitem == NULL) {
return CITRUSLEAF_FAIL_CLIENT;
}
}
//Compile the write buffer to be sent to the cluster
if (n_values && ( values || operations) ){
cl_compile(info1, info2, 0, ns, set, key, digest, values?*values:NULL, operator, operations?*operations:NULL,
*n_values , &wr_buf, &wr_buf_sz, cl_w_p, &d_ret, *trid,NULL,NULL, 0 /*udf_type*/);
}else{
cl_compile(info1, info2, 0, ns, set, key, digest, 0, 0, 0, 0, &wr_buf, &wr_buf_sz, cl_w_p, &d_ret, *trid,NULL,NULL, 0 /*udf_type*/);
}
deadline_ms = 0;
progress_timeout_ms = 0;
if (cl_w_p && cl_w_p->timeout_ms) {
deadline_ms = cf_getms() + cl_w_p->timeout_ms;
// policy: if asking for a long timeout, give enough time to try twice
if (cl_w_p->timeout_ms > 700) {
progress_timeout_ms = cl_w_p->timeout_ms / 2;
}
else {
progress_timeout_ms = cl_w_p->timeout_ms;
}
}
else {
progress_timeout_ms = g_async_nw_progress_timeout;
}
//Initialize the async work unit
workitem->trid = *trid;
workitem->deadline = deadline_ms;
workitem->starttime = cf_getms();
workitem->udata = udata;
as_msg *msgp;
// Hate special cases, but we have to clear the verify bit on delete verify
if ( (info2 & CL_MSG_INFO2_DELETE) && (info1 & CL_MSG_INFO1_VERIFY))
{
msgp = (as_msg *)wr_buf;
msgp->m.info1 &= ~CL_MSG_INFO1_VERIFY;
}
if (asc->compression_stat.compression_threshold > 0
&& wr_buf_sz > (size_t)asc->compression_stat.compression_threshold)
{
/* Compression is enabled.
* Packet size is above threshold.
* Compress the data
*/
uint8_t *compressed_buf = NULL;
size_t compressed_buf_sz = 0;
// Contstruct packet for compressed data.
cf_packet_compression (wr_buf, wr_buf_sz, &compressed_buf, &compressed_buf_sz);
if (compressed_buf)
{
// If original packet size is > 16k, cl_compile had allocated memory for it.
// Free that memory.
// cf_packet_compression will allocate memory for compressed packet
if (wr_buf != wr_stack_buf) {
free(wr_buf);
}
// Update stats.
citrusleaf_cluster_put_compression_stat(asc, wr_buf_sz, compressed_buf_sz);
wr_buf = compressed_buf;
wr_buf_sz = compressed_buf_sz;
//memcpy (wr_buf, compressed_buf, compressed_buf_sz);
//wr_buf_sz = compressed_buf_sz;
//free (compressed_buf);
}
//else compression failed, continue with uncompressed packet
else
{
// Set compression stat
citrusleaf_cluster_put_compression_stat(asc, wr_buf_sz, wr_buf_sz);
}
}
int try = 0;
// retry request based on the write_policy
do {
network_error = false;
try++;
#ifdef DEBUG
if (try > 1) {
cf_debug("request retrying try %d tid %zu", try, (uint64_t)pthread_self());
}
#endif
// Get an FD from a cluster. First get the probable node for the given digest.
node = cl_cluster_node_get(asc, ns, &d_ret, info2 & CL_MSG_INFO2_WRITE ? true : false);
if (!node) {
#ifdef DEBUG
cf_debug("warning: no healthy nodes in cluster, retrying");
#endif
usleep(10000); //Sleep for 10ms
goto Retry;
}
// Now get the dedicated async FD of this node
starttime = cf_getms();
fd = cl_cluster_node_fd_get(node, true);
endtime = cf_getms();
if ((endtime - starttime) > 10) {
cf_debug("Time to get FD for a node (>10ms)=%"PRIu64, (endtime - starttime));
}
if (fd == -1) {
#ifdef DEBUG
cf_debug("warning: node %s has no async file descriptors, retrying transaction (tid %zu)",node->name,(uint64_t)pthread_self() );
#endif
usleep(1000);
goto Retry;
}
// Send the command to the node
starttime = cf_getms();
rv = cf_socket_write_timeout(fd, wr_buf, wr_buf_sz, deadline_ms, progress_timeout_ms);
endtime = cf_getms();
if ((endtime - starttime) > 10) {
cf_debug("Time to write to the socket (>10ms)=%"PRIu64, (endtime - starttime));
}
if (rv != 0) {
cf_debug("Citrusleaf: write timeout or error when writing header to server - %d fd %d errno %d (tid %zu)",
rv,fd,errno,(uint64_t)pthread_self());
if (rv != ETIMEDOUT)
network_error = true;
goto Retry;
}
goto Ok;
Retry:
if (network_error == true) {
/*
* In case of Async work (for XDS), it may be extreme to
* dun a node in case of network error. We just cleanup
* things and retry to connect to the remote cluster.
* The network error may be a transient one. As this is a
* network error, its is better to wait for some significant
* time before retrying.
*/
sleep(1); //Sleep for 1sec
#if ONEASYNCFD
//Do not close the FD
#else
cf_error("async sender: Closing the fd %d because of network error", fd);
cf_close(fd);
fd = -1;
#endif
}
if (fd != -1) {
cf_error("async sender: Closing the fd %d because of retry", fd);
cf_close(fd);
fd = -1;
}
if (node) {
cl_cluster_node_put(node);
node = 0;
}
if (deadline_ms && (deadline_ms < cf_getms() ) ) {
#ifdef DEBUG
cf_debug("async sender: out of time : deadline %"PRIu64" now %"PRIu64, deadline_ms, cf_getms());
#endif
rv = CITRUSLEAF_FAIL_TIMEOUT;
goto Error;
}
} while ( (cl_w_p == 0) || (cl_w_p->w_pol == CL_WRITE_RETRY) );
Error:
#ifdef DEBUG
cf_debug("exiting with failure: network_error %d wpol %d timeleft %d rv %d",
(int)network_error, (int)(cl_w_p ? cl_w_p->w_pol : 0),
(int)(deadline_ms - cf_getms() ), rv );
#endif
if (wr_buf != wr_stack_buf) {
free(wr_buf);
}
#if ONEASYNCFD
//Do not close the FD
#else
//If it is a network error, the fd would be closed and set to -1.
//So, we reach this place with a valid FD in case of timeout.
if (fd != -1) {
cf_error("async sender: Closing the fd %d because of timeout", fd);
cf_close(fd);
}
#endif
return(rv);
Ok:
/*
* We cannot release the node here as the asyc FD associated
* with this node may get closed. We should do it only when
* we got back the ack for the async command that we just did.
*/
//As we sent the command successfully, add it to the async work list
workitem->node = node;
workitem->fd = fd;
//We are storing only the pointer to the workitem
#if ONEASYNCFD
if (shash_put_unique(g_cl_async_hashtab, trid, &workitem) != SHASH_OK) {
//This should always succeed.
cf_error("Unable to add unique entry into the hash table");
}
cf_queue_push(node->asyncwork_q, &workitem); //Also put in the node's q
#else
cf_queue_push(g_cl_async_q, &workitem);
#endif
if (wr_buf != wr_stack_buf) {
free(wr_buf);
}
rv = CITRUSLEAF_OK;
return rv;
}
int citrusleaf_async_reinit(int size_limit, unsigned int num_receiver_threads)
{
// int num_threads;
if (0 == cf_atomic32_get(g_async_initialized)) {
cf_error("Async client not initialized cannot reinit");
return -1;
}
if (num_receiver_threads > MAX_ASYNC_RECEIVER_THREADS) {
//Limit the threads to the max value even if caller asks for it
num_receiver_threads = MAX_ASYNC_RECEIVER_THREADS;
}
// If number of thread is increased create more threads
if (num_receiver_threads > g_async_num_threads) {
unsigned int i;
for (i = g_async_num_threads; i < num_receiver_threads; i++) {
pthread_create(&g_async_reciever[i], 0, async_receiver_fn, NULL);
}
}
else {
// else just reset the number the async threads will kill themselves
cf_atomic32_set(&g_async_num_threads, num_receiver_threads);
}
cf_atomic32_set(&g_async_q_szlimit , size_limit);
return ( 0 );
}
int citrusleaf_async_init(int size_limit, int num_receiver_threads, cl_async_fail_cb fail_cb_fn, cl_async_success_cb success_cb_fn)
{
int i, num_threads;
//Make sure that we do the initialization only once
if (1 == cf_atomic32_incr(&g_async_initialized)) {
// Start the receiver threads
num_threads = num_receiver_threads;
if (num_threads > MAX_ASYNC_RECEIVER_THREADS) {
//Limit the threads to the max value even if caller asks for it
num_threads = MAX_ASYNC_RECEIVER_THREADS;
}
#if ONEASYNCFD
g_async_h_szlimit = size_limit * 3; //Max number of elements in the hash table
g_async_h_buckets = g_async_h_szlimit/10;//Number of buckets in the hash table
if (shash_create(&g_cl_async_hashtab, async_trid_hash, sizeof(uint64_t), sizeof(cl_async_work *),
g_async_h_buckets, SHASH_CR_MT_BIGLOCK) != SHASH_OK) {
cf_error("Failed to initialize the async work hastable");
cf_atomic32_decr(&g_async_initialized);
return -1;
}
#else
// create work queue
g_async_q_szlimit = size_limit;
if ((g_cl_async_q = cf_queue_create(sizeof(cl_async_work *), true)) == NULL) {
cf_error("Failed to initialize the async work queue");
cf_atomic32_decr(&g_async_initialized);
return -1;
}
for (i=0; i<num_threads; i++) {
pthread_create(&g_async_reciever[i], 0, async_receiver_fn, NULL);
}
g_async_num_threads = num_threads;
#endif
if ((g_cl_workitems_freepool_q = cf_queue_create(sizeof(cl_async_work *), true)) == NULL) {
cf_error("Failed to create memory pool for workitems");
return -1;
}
g_fail_cb_fn = fail_cb_fn;
g_success_cb_fn = success_cb_fn;
// Initialize the stats
g_async_stats.retries = 0;
g_async_stats.dropouts = 0;
}
return(0);
}
static void*
async_receiver_fn(void *thdata)
{
int rv = -1;
bool network_error = false;
cl_async_work *workitem = NULL;
// cl_async_work *tmpworkitem = NULL;
as_msg msg;
cf_queue *q_to_use = NULL;
cl_cluster_node *thisnode = NULL;
uint8_t rd_stack_buf[STACK_BUF_SZ];
uint8_t *rd_buf = rd_stack_buf;
size_t rd_buf_sz = 0;
uint64_t acktrid;
// uint64_t starttime, endtime;
int progress_timeout_ms;
unsigned int thread_id = cf_atomic32_incr(&g_thread_count);
if (thdata == NULL) {
q_to_use = g_cl_async_q;
} else {
thisnode = (cl_cluster_node *)thdata;
q_to_use = thisnode->asyncwork_q;
}
//Infinite loop which keeps picking work items from the list and try to find the end result
while(1) {
network_error = false;
#if ONEASYNCFD
if(thisnode->dunned == true) {
do {
rv = cf_queue_pop(thisnode->asyncwork_q, &workitem, CF_QUEUE_NOWAIT);
if (rv == CF_QUEUE_OK) {
cl_cluster_node_put(thisnode);
free(workitem);
}
} while (rv == CF_QUEUE_OK);
//We want to delete all the workitems of this node
shash_reduce_delete(g_cl_async_hashtab, cl_del_node_asyncworkitems, thisnode);
break;
}
#endif
//This call will block if there is no element in the queue
cf_queue_pop(q_to_use, &workitem, CF_QUEUE_FOREVER);
//TODO: What if the node gets dunned while this pop call is blocked ?
#if ONEASYNCFD
//cf_debug("Elements remaining in this node's queue=%d, Hash table size=%d",
// cf_queue_sz(thisnode->asyncwork_q), shash_get_size(g_cl_async_hashtab));
#endif
// If we have no progress in 50ms, we should move to the next workitem
// and revisit this workitem at a later stage
progress_timeout_ms = DEFAULT_PROGRESS_TIMEOUT;
// Read into this fine cl_msg, which is the short header
rv = cf_socket_read_timeout(workitem->fd, (uint8_t *) &msg, sizeof(as_msg), workitem->deadline, progress_timeout_ms);
if (rv) {
#if DEBUG
cf_debug("Citrusleaf: error when reading header from server - rv %d fd %d", rv, workitem->fd);
#endif
if (rv != ETIMEDOUT) {
cf_error("Citrusleaf: error when reading header from server - rv %d fd %d",rv,workitem->fd);
network_error = true;
goto Error;
} else {
goto Retry;
}
}
#ifdef DEBUG_VERBOSE
dump_buf("read header from cluster", (uint8_t *) &msg, sizeof(cl_msg));
#endif
cl_proto_swap(&msg.proto);
cl_msg_swap_header(&msg.m);
// second read for the remainder of the message
rd_buf_sz = msg.proto.sz - msg.m.header_sz;
if (rd_buf_sz > 0) {
if (rd_buf_sz > sizeof(rd_stack_buf)) {
rd_buf = malloc(rd_buf_sz);
if (!rd_buf) {
cf_error("malloc fail: trying %zu",rd_buf_sz);
rv = -1;
goto Error;
}
}
rv = cf_socket_read_timeout(workitem->fd, rd_buf, rd_buf_sz, workitem->deadline, progress_timeout_ms);
if (rv) {
//We already read some part of the message before but failed to read the
//remaining data for whatever reason (network error or timeout). We cannot
//reread as we already read partial data. Declare this as error.
cf_error("Timeout after reading the header but before reading the body");
goto Error;
}
#ifdef DEBUG_VERBOSE
dump_buf("read body from cluster", rd_buf, rd_buf_sz);
#endif
}
rv = CITRUSLEAF_OK;
goto Ok;
Retry:
//We are trying to postpone the reading
if (workitem->deadline && workitem->deadline < cf_getms()) {
cf_error("async receiver: out of time : deadline %"PRIu64" now %"PRIu64,
workitem->deadline, cf_getms());
//cf_error("async receiver: Workitem missed the final deadline");
rv = CITRUSLEAF_FAIL_TIMEOUT;
goto Error;
} else {
//We have time. Push the element back to the queue to be considered later
cf_queue_push(q_to_use, &workitem);
}
//If we allocated memory in this loop, release it.
if (rd_buf && (rd_buf != rd_stack_buf)) {
free(rd_buf);
}
cf_atomic_int_incr(&g_async_stats.retries);
continue;
Error:
if (network_error == true) {
/*
* In case of Async work (for XDS), it may be extreme to
* dun a node in case of network error. We just cleanup
* things and retry to connect to the remote cluster.
* The network error may be a transient one.
*/
}
#if ONEASYNCFD
//Do not close FD
#else
//We do not know the state of FD. It may have pending data to be read.
//We cannot reuse the FD. So, close it to be on safe side.
cf_error("async receiver: Closing the fd %d because of error", workitem->fd);
cf_close(workitem->fd);
workitem->fd = -1;
#endif
cf_atomic_int_incr(&g_async_stats.dropouts);
//Continue down with what we do during an Ok
//Inform the caller that there is no response from the server for this workitem.
//No response does not mean that the work is not done. The work might be
//successfully completed on the server side, we just didnt get response for it.
if (g_fail_cb_fn) {
g_fail_cb_fn(workitem->udata, rv, workitem->starttime);
}
Ok:
//rd_buf may not be there during an error condition.
if (rd_buf && (rv == CITRUSLEAF_OK)) {
//As of now, async functionality is there only for put call.
//In put call, we do not get anything back other than the trid field.
//So, just pass variable to get back the trid and ignore others.
if (0 != cl_parse(&msg.m, rd_buf, rd_buf_sz, NULL, NULL, NULL, &acktrid, NULL)) {
rv = CITRUSLEAF_FAIL_UNKNOWN;
}
else {
rv = msg.m.result_code;
if (workitem->trid != acktrid) {
#if ONEASYNCFD
//It is likely that we may get response for a different trid.
//Just delete the correct one from the queue
//put back the current workitem back in the queue.
shash_get(g_cl_async_hashtab, &acktrid, &tmpworkitem);
cf_queue_delete(q_to_use, &tmpworkitem, true);
cf_queue_push(q_to_use, &workitem);
//From now on workitem will be the one for which we got ack
workitem = tmpworkitem;
#endif
#ifdef DEBUG
cf_debug("Got reply for a different trid. Expected=%"PRIu64" Got=%"PRIu64" FD=%d",
workitem->trid, acktrid, workitem->fd);
#endif
}
}
if (g_success_cb_fn) {
g_success_cb_fn(workitem->udata, rv, workitem->starttime);
}
}
//Remember to put back the FD into the pool, if it is re-usable.
if (workitem->fd != -1) {
cl_cluster_node_fd_put(workitem->node, workitem->fd, true);
}
//Also decrement the reference count for this node
cl_cluster_node_put(workitem->node);
#if ONEASYNCFD
//Delete the item from the global hashtable
if (shash_delete(g_cl_async_hashtab, &workitem->trid) != SHASH_OK)
{
#if DEBUG
cf_debug("Failure while trying to delete trid=%"PRIu64" from hashtable", workitem->trid);
#endif
}
#endif
//Push it back into the free pool. If the attempt fails, free it.
if (cf_queue_push(g_cl_workitems_freepool_q, &workitem) == -1) {
free(workitem);
}
//If we allocated memory in this loop, release it.
if (rd_buf && (rd_buf != rd_stack_buf)) {
free(rd_buf);
}
// Kick this thread out if its ID is greater than total
if (thread_id > cf_atomic32_get(g_async_num_threads)) {
cf_atomic32_decr(&g_thread_count);
return NULL;
}
}//The infnite loop
return NULL;
}
/* cf_queue_pop
* if ms_wait < 0, wait forever
* if ms_wait = 0, don't wait at all
* if ms_wait > 0, wait that number of ms
* */
int cf_queue_pop(cf_queue *q, void *buf, int ms_wait) {
if (NULL == q) {
cf_error("cf_queue_pop: try passing in a queue");
return(-1);
}
#ifdef EXTERNAL_LOCKS
if (ms_wait != CF_QUEUE_NOWAIT) { // this implementation won't wait
cf_error("cf_queue_pop: only nowait supported");
return(-1);
}
#endif // EXTERNAL_LOCKS
QUEUE_LOCK(q);
struct timespec tp;
if (ms_wait > 0) {
#ifdef OSX
uint64_t curms = cf_getms(); // using the cl generic functions defined in cf_clock.h. It is going to have slightly less resolution than the pure linux version
tp.tv_sec = (curms + ms_wait)/1000;
tp.tv_nsec = (ms_wait %1000) * 1000000;
#else // linux
clock_gettime( CLOCK_REALTIME, &tp);
tp.tv_sec += ms_wait / 1000;
tp.tv_nsec += (ms_wait % 1000) * 1000000;
if (tp.tv_nsec > 1000000000) {
tp.tv_nsec -= 1000000000;
tp.tv_sec++;
}
#endif
}
/* FIXME error checking */
/* Note that we apparently have to use a while() loop. Careful reading
* of the pthread_cond_signal() documentation says that AT LEAST ONE
* waiting thread will be awakened... */
if (q->threadsafe) {
#ifdef EXTERNAL_LOCKS
if (CF_Q_EMPTY(q)) {
QUEUE_UNLOCK(q);
return(CF_QUEUE_EMPTY);
}
#else
while (CF_Q_EMPTY(q)) {
if (CF_QUEUE_FOREVER == ms_wait) {
pthread_cond_wait(&q->CV, &q->LOCK);
}
else if (CF_QUEUE_NOWAIT == ms_wait) {
pthread_mutex_unlock(&q->LOCK);
return(CF_QUEUE_EMPTY);
}
else {
pthread_cond_timedwait(&q->CV, &q->LOCK, &tp);
if (CF_Q_EMPTY(q)) {
pthread_mutex_unlock(&q->LOCK);
return(CF_QUEUE_EMPTY);
}
}
}
#endif // EXTERNAL_LOCKS
} else if (CF_Q_EMPTY(q))
return(CF_QUEUE_EMPTY);
memcpy(buf, CF_Q_ELEM_PTR(q,q->read_offset), q->elementsz);
q->read_offset++;
// interesting idea - this probably keeps the cache fresher
// because the queue is fully empty just make it all zero
if (q->read_offset == q->write_offset) {
q->read_offset = q->write_offset = 0;
}
QUEUE_UNLOCK(q);
return(0);
}
// Set of threads which talk to client over the connection for doing the needful
// processing. Note that once fd is assigned to a thread all the work on that fd
// is done by that thread. Fair fd usage is expected of the client. First thread
// is special - also does accept [listens for new connections]. It is the only
// thread which does it.
void *
thr_demarshal(void *arg)
{
cf_socket_cfg *s, *ls;
// Create my epoll fd, register in the global list.
struct epoll_event ev;
int nevents, i, n, epoll_fd;
cf_clock last_fd_print = 0;
#if defined(USE_SYSTEMTAP)
uint64_t nodeid = g_config.self_node;
#endif
// Early stage aborts; these will cause faults in process scope.
cf_assert(arg, AS_DEMARSHAL, CF_CRITICAL, "invalid argument");
s = &g_config.socket;
ls = &g_config.localhost_socket;
#ifdef USE_JEM
int orig_arena;
if (0 > (orig_arena = jem_get_arena())) {
cf_crash(AS_DEMARSHAL, "Failed to get original arena for thr_demarshal()!");
} else {
cf_info(AS_DEMARSHAL, "Saved original JEMalloc arena #%d for thr_demarshal()", orig_arena);
}
#endif
// Figure out my thread index.
pthread_t self = pthread_self();
int thr_id;
for (thr_id = 0; thr_id < MAX_DEMARSHAL_THREADS; thr_id++) {
if (0 != pthread_equal(g_demarshal_args->dm_th[thr_id], self))
break;
}
if (thr_id == MAX_DEMARSHAL_THREADS) {
cf_debug(AS_FABRIC, "Demarshal thread could not figure own ID, bogus, exit, fu!");
return(0);
}
// First thread accepts new connection at interface socket.
if (thr_id == 0) {
demarshal_file_handle_init();
epoll_fd = epoll_create(EPOLL_SZ);
if (epoll_fd == -1)
cf_crash(AS_DEMARSHAL, "epoll_create(): %s", cf_strerror(errno));
memset(&ev, 0, sizeof (ev));
ev.events = EPOLLIN | EPOLLERR | EPOLLHUP;
ev.data.fd = s->sock;
if (0 > epoll_ctl(epoll_fd, EPOLL_CTL_ADD, s->sock, &ev))
cf_crash(AS_DEMARSHAL, "epoll_ctl(): %s", cf_strerror(errno));
cf_info(AS_DEMARSHAL, "Service started: socket %s:%d", s->addr, s->port);
if (ls->sock) {
ev.events = EPOLLIN | EPOLLERR | EPOLLHUP;
ev.data.fd = ls->sock;
if (0 > epoll_ctl(epoll_fd, EPOLL_CTL_ADD, ls->sock, &ev))
cf_crash(AS_DEMARSHAL, "epoll_ctl(): %s", cf_strerror(errno));
cf_info(AS_DEMARSHAL, "Service also listening on localhost socket %s:%d", ls->addr, ls->port);
}
}
else {
epoll_fd = epoll_create(EPOLL_SZ);
if (epoll_fd == -1)
cf_crash(AS_DEMARSHAL, "epoll_create(): %s", cf_strerror(errno));
}
g_demarshal_args->epoll_fd[thr_id] = epoll_fd;
cf_detail(AS_DEMARSHAL, "demarshal thread started: id %d", thr_id);
int id_cntr = 0;
// Demarshal transactions from the socket.
for ( ; ; ) {
struct epoll_event events[EPOLL_SZ];
cf_detail(AS_DEMARSHAL, "calling epoll");
nevents = epoll_wait(epoll_fd, events, EPOLL_SZ, -1);
if (0 > nevents) {
cf_debug(AS_DEMARSHAL, "epoll_wait() returned %d ; errno = %d (%s)", nevents, errno, cf_strerror(errno));
}
cf_detail(AS_DEMARSHAL, "epoll event received: nevents %d", nevents);
uint64_t now_ns = cf_getns();
uint64_t now_ms = now_ns / 1000000;
// Iterate over all events.
for (i = 0; i < nevents; i++) {
if ((s->sock == events[i].data.fd) || (ls->sock == events[i].data.fd)) {
// Accept new connections on the service socket.
int csocket = -1;
struct sockaddr_in caddr;
socklen_t clen = sizeof(caddr);
char cpaddr[64];
if (-1 == (csocket = accept(events[i].data.fd, (struct sockaddr *)&caddr, &clen))) {
// This means we're out of file descriptors - could be a SYN
// flood attack or misbehaving client. Eventually we'd like
// to make the reaper fairer, but for now we'll just have to
// ignore the accept error and move on.
if ((errno == EMFILE) || (errno == ENFILE)) {
if (last_fd_print != (cf_getms() / 1000L)) {
cf_info(AS_DEMARSHAL, " warning: hit OS file descript limit (EMFILE on accept), consider raising limit");
last_fd_print = cf_getms() / 1000L;
}
continue;
}
cf_crash(AS_DEMARSHAL, "accept: %s (errno %d)", cf_strerror(errno), errno);
}
// Get the client IP address in string form.
if (caddr.sin_family == AF_INET) {
if (NULL == inet_ntop(AF_INET, &caddr.sin_addr.s_addr, (char *)cpaddr, sizeof(cpaddr))) {
cf_crash(AS_DEMARSHAL, "inet_ntop(): %s (errno %d)", cf_strerror(errno), errno);
}
}
else if (caddr.sin_family == AF_INET6) {
struct sockaddr_in6* addr_in6 = (struct sockaddr_in6*)&caddr;
if (NULL == inet_ntop(AF_INET6, &addr_in6->sin6_addr, (char *)cpaddr, sizeof(cpaddr))) {
cf_crash(AS_DEMARSHAL, "inet_ntop(): %s (errno %d)", cf_strerror(errno), errno);
}
}
else {
cf_crash(AS_DEMARSHAL, "unknown address family %u", caddr.sin_family);
}
cf_detail(AS_DEMARSHAL, "new connection: %s (fd %d)", cpaddr, csocket);
// Validate the limit of protocol connections we allow.
uint32_t conns_open = g_config.proto_connections_opened - g_config.proto_connections_closed;
if (conns_open > g_config.n_proto_fd_max) {
if ((last_fd_print + 5000L) < cf_getms()) { // no more than 5 secs
cf_warning(AS_DEMARSHAL, "dropping incoming client connection: hit limit %d connections", conns_open);
last_fd_print = cf_getms();
}
shutdown(csocket, SHUT_RDWR);
close(csocket);
csocket = -1;
continue;
}
// Set the socket to nonblocking.
if (-1 == cf_socket_set_nonblocking(csocket)) {
cf_info(AS_DEMARSHAL, "unable to set client socket to nonblocking mode");
shutdown(csocket, SHUT_RDWR);
close(csocket);
csocket = -1;
continue;
}
// Create as_file_handle and queue it up in epoll_fd for further
// communication on one of the demarshal threads.
as_file_handle *fd_h = cf_rc_alloc(sizeof(as_file_handle));
if (!fd_h) {
cf_crash(AS_DEMARSHAL, "malloc");
}
sprintf(fd_h->client, "%s:%d", cpaddr, ntohs(caddr.sin_port));
fd_h->fd = csocket;
fd_h->last_used = cf_getms();
fd_h->reap_me = false;
fd_h->trans_active = false;
fd_h->proto = 0;
fd_h->proto_unread = 0;
fd_h->fh_info = 0;
fd_h->security_filter = as_security_filter_create();
// Insert into the global table so the reaper can manage it. Do
// this before queueing it up for demarshal threads - once
// EPOLL_CTL_ADD is done it's difficult to back out (if insert
// into global table fails) because fd state could be anything.
cf_rc_reserve(fd_h);
pthread_mutex_lock(&g_file_handle_a_LOCK);
int j;
bool inserted = true;
if (0 != cf_queue_pop(g_freeslot, &j, CF_QUEUE_NOWAIT)) {
inserted = false;
}
else {
g_file_handle_a[j] = fd_h;
}
pthread_mutex_unlock(&g_file_handle_a_LOCK);
if (!inserted) {
cf_info(AS_DEMARSHAL, "unable to add socket to file handle table");
shutdown(csocket, SHUT_RDWR);
close(csocket);
csocket = -1;
cf_rc_free(fd_h); // will free even with ref-count of 2
}
else {
// Place the client socket in the event queue.
memset(&ev, 0, sizeof(ev));
ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP ;
ev.data.ptr = fd_h;
// Round-robin pick up demarshal thread epoll_fd and add
// this new connection to epoll.
int id;
while (true) {
id = (id_cntr++) % g_demarshal_args->num_threads;
if (g_demarshal_args->epoll_fd[id] != 0) {
break;
}
}
fd_h->epoll_fd = g_demarshal_args->epoll_fd[id];
if (0 > (n = epoll_ctl(fd_h->epoll_fd, EPOLL_CTL_ADD, csocket, &ev))) {
cf_info(AS_DEMARSHAL, "unable to add socket to event queue of demarshal thread %d %d", id, g_demarshal_args->num_threads);
pthread_mutex_lock(&g_file_handle_a_LOCK);
fd_h->reap_me = true;
as_release_file_handle(fd_h);
fd_h = 0;
pthread_mutex_unlock(&g_file_handle_a_LOCK);
}
else {
cf_atomic_int_incr(&g_config.proto_connections_opened);
}
}
}
else {
bool has_extra_ref = false;
as_file_handle *fd_h = events[i].data.ptr;
if (fd_h == 0) {
cf_info(AS_DEMARSHAL, "event with null handle, continuing");
goto NextEvent;
}
cf_detail(AS_DEMARSHAL, "epoll connection event: fd %d, events 0x%x", fd_h->fd, events[i].events);
// Process data on an existing connection: this might be more
// activity on an already existing transaction, so we have some
// state to manage.
as_proto *proto_p = 0;
int fd = fd_h->fd;
if (events[i].events & (EPOLLRDHUP | EPOLLERR | EPOLLHUP)) {
cf_detail(AS_DEMARSHAL, "proto socket: remote close: fd %d event %x", fd, events[i].events);
// no longer in use: out of epoll etc
goto NextEvent_FD_Cleanup;
}
if (fd_h->trans_active) {
goto NextEvent;
}
// If pointer is NULL, then we need to create a transaction and
// store it in the buffer.
if (fd_h->proto == NULL) {
as_proto proto;
int sz;
/* Get the number of available bytes */
if (-1 == ioctl(fd, FIONREAD, &sz)) {
cf_info(AS_DEMARSHAL, "unable to get number of available bytes");
goto NextEvent_FD_Cleanup;
}
// If we don't have enough data to fill the message buffer,
// just wait and we'll come back to this one. However, we'll
// let messages with zero size through, since they are
// likely errors. We don't cleanup the FD in this case since
// we'll get more data on it.
if (sz < sizeof(as_proto) && sz != 0) {
goto NextEvent;
}
// Do a preliminary read of the header into a stack-
// allocated structure, so that later on we can allocate the
// entire message buffer.
if (0 >= (n = cf_socket_recv(fd, &proto, sizeof(as_proto), MSG_WAITALL))) {
cf_detail(AS_DEMARSHAL, "proto socket: read header fail: error: rv %d sz was %d errno %d", n, sz, errno);
goto NextEvent_FD_Cleanup;
}
if (proto.version != PROTO_VERSION &&
// For backward compatibility, allow version 0 with
// security messages.
! (proto.version == 0 && proto.type == PROTO_TYPE_SECURITY)) {
cf_warning(AS_DEMARSHAL, "proto input from %s: unsupported proto version %u",
fd_h->client, proto.version);
goto NextEvent_FD_Cleanup;
}
// Swap the necessary elements of the as_proto.
as_proto_swap(&proto);
if (proto.sz > PROTO_SIZE_MAX) {
cf_warning(AS_DEMARSHAL, "proto input from %s: msg greater than %d, likely request from non-Aerospike client, rejecting: sz %"PRIu64,
fd_h->client, PROTO_SIZE_MAX, proto.sz);
goto NextEvent_FD_Cleanup;
}
#ifdef USE_JEM
// Attempt to peek the namespace and set the JEMalloc arena accordingly.
size_t peeked_data_sz = 0;
size_t min_field_sz = sizeof(uint32_t) + sizeof(char);
size_t min_as_msg_sz = sizeof(as_msg) + min_field_sz;
size_t peekbuf_sz = 2048; // (Arbitrary "large enough" size for peeking the fields of "most" AS_MSGs.)
uint8_t peekbuf[peekbuf_sz];
if (PROTO_TYPE_AS_MSG == proto.type) {
size_t offset = sizeof(as_msg);
// Number of bytes to peek from the socket.
// size_t peek_sz = peekbuf_sz; // Peak up to the size of the peek buffer.
size_t peek_sz = MIN(proto.sz, peekbuf_sz); // Peek only up to the minimum necessary number of bytes.
if (!(peeked_data_sz = cf_socket_recv(fd, peekbuf, peek_sz, 0))) {
// That's actually legitimate. The as_proto may have gone into one
// packet, the as_msg into the next one, which we haven't yet received.
// This just "never happened" without async.
cf_detail(AS_DEMARSHAL, "could not peek the as_msg header, expected %zu byte(s)", peek_sz);
}
if (peeked_data_sz > min_as_msg_sz) {
// cf_debug(AS_DEMARSHAL, "(Peeked %zu bytes.)", peeked_data_sz);
if (peeked_data_sz > proto.sz) {
cf_warning(AS_DEMARSHAL, "Received unexpected extra data from client %s socket %d when peeking as_proto!", fd_h->client, fd);
log_as_proto_and_peeked_data(&proto, peekbuf, peeked_data_sz);
goto NextEvent_FD_Cleanup;
}
if (((as_msg*)peekbuf)->info1 & AS_MSG_INFO1_BATCH) {
jem_set_arena(orig_arena);
} else {
uint16_t n_fields = ntohs(((as_msg *) peekbuf)->n_fields), field_num = 0;
bool found = false;
// cf_debug(AS_DEMARSHAL, "Found %d AS_MSG fields", n_fields);
while (!found && (field_num < n_fields)) {
as_msg_field *field = (as_msg_field *) (&peekbuf[offset]);
uint32_t value_sz = ntohl(field->field_sz) - 1;
// cf_debug(AS_DEMARSHAL, "Field #%d offset: %lu", field_num, offset);
// cf_debug(AS_DEMARSHAL, "\tvalue_sz %u", value_sz);
// cf_debug(AS_DEMARSHAL, "\ttype %d", field->type);
if (AS_MSG_FIELD_TYPE_NAMESPACE == field->type) {
if (value_sz >= AS_ID_NAMESPACE_SZ) {
cf_warning(AS_DEMARSHAL, "namespace too long (%u) in as_msg", value_sz);
goto NextEvent_FD_Cleanup;
}
char ns[AS_ID_NAMESPACE_SZ];
found = true;
memcpy(ns, field->data, value_sz);
ns[value_sz] = '\0';
// cf_debug(AS_DEMARSHAL, "Found ns \"%s\" in field #%d.", ns, field_num);
jem_set_arena(as_namespace_get_jem_arena(ns));
} else {
// cf_debug(AS_DEMARSHAL, "Message field %d is not namespace (type %d) ~~ Reading next field", field_num, field->type);
field_num++;
offset += sizeof(as_msg_field) + value_sz;
if (offset >= peeked_data_sz) {
break;
}
}
}
if (!found) {
cf_warning(AS_DEMARSHAL, "Can't get namespace from AS_MSG (peeked %zu bytes) ~~ Using default thr_demarshal arena.", peeked_data_sz);
jem_set_arena(orig_arena);
}
}
} else {
jem_set_arena(orig_arena);
}
} else {
jem_set_arena(orig_arena);
}
#endif
// Allocate the complete message buffer.
proto_p = cf_malloc(sizeof(as_proto) + proto.sz);
cf_assert(proto_p, AS_DEMARSHAL, CF_CRITICAL, "allocation: %zu %s", (sizeof(as_proto) + proto.sz), cf_strerror(errno));
memcpy(proto_p, &proto, sizeof(as_proto));
#ifdef USE_JEM
// Jam in the peeked data.
if (peeked_data_sz) {
memcpy(proto_p->data, &peekbuf, peeked_data_sz);
}
fd_h->proto_unread = proto_p->sz - peeked_data_sz;
#else
fd_h->proto_unread = proto_p->sz;
#endif
fd_h->proto = (void *) proto_p;
}
else {
proto_p = fd_h->proto;
}
if (fd_h->proto_unread > 0) {
// Read the data.
n = cf_socket_recv(fd, proto_p->data + (proto_p->sz - fd_h->proto_unread), fd_h->proto_unread, 0);
if (0 >= n) {
if (errno == EAGAIN) {
continue;
}
cf_info(AS_DEMARSHAL, "receive socket: fail? n %d errno %d %s closing connection.", n, errno, cf_strerror(errno));
goto NextEvent_FD_Cleanup;
}
// Decrement bytes-unread counter.
cf_detail(AS_DEMARSHAL, "read fd %d (%d %d)", fd, n, fd_h->proto_unread);
fd_h->proto_unread -= n;
}
// Check for a finished read.
if (0 == fd_h->proto_unread) {
// It's only really live if it's injecting a transaction.
fd_h->last_used = now_ms;
thr_demarshal_pause(fd_h); // pause reading while the transaction is in progress
fd_h->proto = 0;
fd_h->proto_unread = 0;
// INIT_TR
as_transaction tr;
as_transaction_init(&tr, NULL, (cl_msg *)proto_p);
cf_rc_reserve(fd_h);
has_extra_ref = true;
tr.proto_fd_h = fd_h;
tr.start_time = now_ns; // set transaction start time
tr.preprocessed = false;
if (! as_proto_is_valid_type(proto_p)) {
cf_warning(AS_DEMARSHAL, "unsupported proto message type %u", proto_p->type);
// We got a proto message type we don't recognize, so it
// may not do any good to send back an as_msg error, but
// it's the best we can do. At least we can keep the fd.
as_transaction_demarshal_error(&tr, AS_PROTO_RESULT_FAIL_UNKNOWN);
cf_atomic_int_incr(&g_config.proto_transactions);
goto NextEvent;
}
if (g_config.microbenchmarks) {
histogram_insert_data_point(g_config.demarshal_hist, now_ns);
tr.microbenchmark_time = cf_getns();
}
// Check if it's compressed.
if (tr.msgp->proto.type == PROTO_TYPE_AS_MSG_COMPRESSED) {
// Decompress it - allocate buffer to hold decompressed
// packet.
uint8_t *decompressed_buf = NULL;
size_t decompressed_buf_size = 0;
int rv = 0;
if ((rv = as_packet_decompression((uint8_t *)proto_p, &decompressed_buf, &decompressed_buf_size))) {
cf_warning(AS_DEMARSHAL, "as_proto decompression failed! (rv %d)", rv);
cf_warning_binary(AS_DEMARSHAL, proto_p, sizeof(as_proto) + proto_p->sz, CF_DISPLAY_HEX_SPACED, "compressed proto_p");
as_transaction_demarshal_error(&tr, AS_PROTO_RESULT_FAIL_UNKNOWN);
cf_atomic_int_incr(&g_config.proto_transactions);
goto NextEvent;
}
// Count the packets.
cf_atomic_int_add(&g_config.stat_compressed_pkts_received, 1);
// Free the compressed packet since we'll be using the
// decompressed packet from now on.
cf_free(proto_p);
proto_p = NULL;
// Get original packet.
tr.msgp = (cl_msg *)decompressed_buf;
as_proto_swap(&(tr.msgp->proto));
if (! as_proto_wrapped_is_valid(&tr.msgp->proto, decompressed_buf_size)) {
cf_warning(AS_DEMARSHAL, "decompressed unusable proto: version %u, type %u, sz %lu [%lu]",
tr.msgp->proto.version, tr.msgp->proto.type, tr.msgp->proto.sz, decompressed_buf_size);
as_transaction_demarshal_error(&tr, AS_PROTO_RESULT_FAIL_UNKNOWN);
cf_atomic_int_incr(&g_config.proto_transactions);
goto NextEvent;
}
}
// Security protocol transactions.
if (tr.msgp->proto.type == PROTO_TYPE_SECURITY) {
as_security_transact(&tr);
cf_atomic_int_incr(&g_config.proto_transactions);
goto NextEvent;
}
// Info protocol requests.
if (tr.msgp->proto.type == PROTO_TYPE_INFO) {
if (as_info(&tr)) {
cf_warning(AS_DEMARSHAL, "Info request failed to be enqueued ~~ Freeing protocol buffer");
goto NextEvent_FD_Cleanup;
}
cf_atomic_int_incr(&g_config.proto_transactions);
goto NextEvent;
}
ASD_TRANS_DEMARSHAL(nodeid, (uint64_t) tr.msgp);
// Fast path for batch requests.
if (tr.msgp->msg.info1 & AS_MSG_INFO1_BATCH) {
as_batch_queue_task(&tr);
cf_atomic_int_incr(&g_config.proto_transactions);
goto NextEvent;
}
// Either process the transaction directly in this thread,
// or queue it for processing by another thread (tsvc/info).
if (0 != thr_tsvc_process_or_enqueue(&tr)) {
cf_warning(AS_DEMARSHAL, "Failed to queue transaction to the service thread");
goto NextEvent_FD_Cleanup;
}
else {
cf_atomic_int_incr(&g_config.proto_transactions);
}
}
// Jump the proto message free & FD cleanup. If we get here, the
// above operations went smoothly. The message free & FD cleanup
// job is handled elsewhere as directed by
// thr_tsvc_process_or_enqueue().
goto NextEvent;
NextEvent_FD_Cleanup:
// If we allocated memory for the incoming message, free it.
if (proto_p) {
cf_free(proto_p);
fd_h->proto = 0;
}
// If fd has extra reference for transaction, release it.
if (has_extra_ref) {
cf_rc_release(fd_h);
}
// Remove the fd from the events list.
if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, fd, 0) < 0) {
cf_crash(AS_DEMARSHAL, "unable to remove socket FD %d from epoll instance FD %d: %d (%s)",
fd, epoll_fd, errno, cf_strerror(errno));
}
pthread_mutex_lock(&g_file_handle_a_LOCK);
fd_h->reap_me = true;
as_release_file_handle(fd_h);
fd_h = 0;
pthread_mutex_unlock(&g_file_handle_a_LOCK);
NextEvent:
;
}
// We should never be canceled externally, but just in case...
pthread_testcancel();
}
}
return NULL;
}
as_status
as_command_execute(as_cluster* cluster, as_error * err, as_command_node* cn, uint8_t* command, size_t command_len,
uint32_t timeout_ms, uint32_t retry,
as_parse_results_fn parse_results_fn, void* parse_results_data
)
{
uint64_t deadline_ms = as_socket_deadline(timeout_ms);
uint32_t sleep_between_retries_ms = 0;
uint32_t failed_nodes = 0;
uint32_t failed_conns = 0;
uint32_t iterations = 0;
bool release_node;
// Execute command until successful, timed out or maximum iterations have been reached.
while (true) {
as_node* node;
if (cn->node) {
node = cn->node;
release_node = false;
}
else {
node = as_node_get(cluster, cn->ns, cn->digest, cn->write, cn->replica);
release_node = true;
}
if (!node) {
failed_nodes++;
sleep_between_retries_ms = 10;
goto Retry;
}
int fd;
as_status status = as_node_get_connection(err, node, deadline_ms, &fd);
if (status) {
if (release_node) {
as_node_release(node);
}
failed_conns++;
sleep_between_retries_ms = 1;
goto Retry;
}
// Send command.
status = as_socket_write_deadline(err, fd, command, command_len, deadline_ms);
if (status) {
// Socket errors are considered temporary anomalies. Retry.
// Close socket to flush out possible garbage. Do not put back in pool.
as_close(fd);
if (release_node) {
as_node_release(node);
}
sleep_between_retries_ms = 0;
goto Retry;
}
// Parse results returned by server.
status = parse_results_fn(err, fd, deadline_ms, parse_results_data);
if (status == AEROSPIKE_OK) {
// Reset error code if retry had occurred.
if (iterations > 0) {
as_error_reset(err);
}
}
else {
switch (status) {
// Retry on timeout.
case AEROSPIKE_ERR_TIMEOUT:
as_close(fd);
if (release_node) {
as_node_release(node);
}
sleep_between_retries_ms = 0;
goto Retry;
// Close socket on errors that can leave unread data in socket.
case AEROSPIKE_ERR_QUERY_ABORTED:
case AEROSPIKE_ERR_SCAN_ABORTED:
case AEROSPIKE_ERR_CLIENT_ABORT:
case AEROSPIKE_ERR_CLIENT:
as_close(fd);
if (release_node) {
as_node_release(node);
}
err->code = status;
return status;
default:
err->code = status;
break;
}
}
// Put connection back in pool.
as_node_put_connection(node, fd, cluster->conn_queue_size);
// Release resources.
if (release_node) {
as_node_release(node);
}
return status;
Retry:
// Check if max retries reached.
if (++iterations > retry) {
break;
}
// Check for client timeout.
if (deadline_ms > 0) {
int remaining_ms = (int)(deadline_ms - cf_getms() - sleep_between_retries_ms);
if (remaining_ms <= 0) {
break;
}
// Reset timeout in send buffer (destined for server).
*(uint32_t*)(command + 22) = cf_swap_to_be32(remaining_ms);
}
if (sleep_between_retries_ms > 0) {
// Sleep before trying again.
usleep(sleep_between_retries_ms * 1000);
}
}
return as_error_update(err, AEROSPIKE_ERR_TIMEOUT,
"Client timeout: timeout=%d iterations=%u failedNodes=%u failedConns=%u",
timeout_ms, iterations, failed_nodes, failed_conns);
}
// Put batch request on a separate batch queue.
int
as_batch(as_transaction* tr)
{
as_msg* msg = &tr->msgp->msg;
as_msg_field* nsfp = as_msg_field_get(msg, AS_MSG_FIELD_TYPE_NAMESPACE);
if (! nsfp) {
cf_warning(AS_BATCH, "Batch namespace is required.");
return -1;
}
as_msg_field* dfp = as_msg_field_get(msg, AS_MSG_FIELD_TYPE_DIGEST_RIPE_ARRAY);
if (! dfp) {
cf_warning(AS_BATCH, "Batch digests are required.");
return -1;
}
uint n_digests = dfp->field_sz / sizeof(cf_digest);
if (n_digests > g_config.batch_max_requests) {
cf_warning(AS_BATCH, "Batch request size %u exceeds max %u.", n_digests, g_config.batch_max_requests);
return -1;
}
batch_transaction btr;
btr.trid = tr->trid;
btr.end_time = tr->end_time;
btr.get_data = !(msg->info1 & AS_MSG_INFO1_GET_NOBINDATA);
btr.ns = as_namespace_get_bymsgfield(nsfp);
if (! btr.ns) {
cf_warning(AS_BATCH, "Batch namespace is required.");
return -1;
}
// Create the master digest table.
btr.digests = (batch_digests*) cf_malloc(sizeof(batch_digests) + (sizeof(batch_digest) * n_digests));
if (! btr.digests) {
cf_warning(AS_BATCH, "Failed to allocate memory for batch digests.");
return -1;
}
batch_digests* bmd = btr.digests;
bmd->n_digests = n_digests;
uint8_t* digest_field_data = dfp->data;
for (int i = 0; i < n_digests; i++) {
bmd->digest[i].done = false;
bmd->digest[i].node = 0;
memcpy(&bmd->digest[i].keyd, digest_field_data, sizeof(cf_digest));
digest_field_data += sizeof(cf_digest);
}
btr.binlist = as_binlist_from_op(msg);
btr.fd_h = tr->proto_fd_h;
tr->proto_fd_h = 0;
btr.fd_h->last_used = cf_getms();
cf_atomic_int_incr(&g_config.batch_initiate);
cf_queue_push(g_batch_queue, &btr);
return 0;
}
/* cf_socket_init_client
* Connect a socket to a remote endpoint
* DOES A BLOCKING CONNECT INLINE - timeout
*/
int
cf_socket_init_client(cf_socket_cfg *s, int timeout)
{
cf_assert(s, CF_SOCKET, CF_CRITICAL, "invalid argument");
if (0 > (s->sock = socket(AF_INET, s->proto, 0))) {
cf_warning(CF_SOCKET, "socket: %s", cf_strerror(errno));
return(-1);
}
fcntl(s->sock, F_SETFD, FD_CLOEXEC); /* close on exec */
fcntl(s->sock, F_SETFL, O_NONBLOCK); /* non-blocking */
// Try tuning the window: must be done before connect
// int flag = (1024 * 32);
// setsockopt(s->sock, SOL_SOCKET, SO_SNDBUF, &flag, sizeof(flag) );
// setsockopt(s->sock, SOL_SOCKET, SO_RCVBUF, &flag, sizeof(flag) );
memset(&s->saddr,0,sizeof(s->saddr));
s->saddr.sin_family = AF_INET;
int rv = inet_pton(AF_INET, s->addr, &s->saddr.sin_addr.s_addr);
if (rv < 0) {
cf_warning(CF_SOCKET, "inet_pton: %s", cf_strerror(errno));
close(s->sock);
return(-1);
} else if (rv == 0) {
cf_warning(CF_SOCKET, "inet_pton: invalid ip %s", s->addr);
close(s->sock);
return(-1);
}
s->saddr.sin_port = htons(s->port);
rv = connect(s->sock, (struct sockaddr *)&s->saddr, sizeof(s->saddr));
cf_debug(CF_SOCKET, "connect: rv %d errno %s",rv,cf_strerror(errno));
if (rv < 0) {
int epoll_fd = -1;
if (errno == EINPROGRESS) {
cf_clock start = cf_getms();
if (0 > (epoll_fd = epoll_create(1))) {
cf_warning(CF_SOCKET, "epoll_create() failed (errno %d: \"%s\")", errno, cf_strerror(errno));
goto Fail;
}
struct epoll_event event;
memset(&event, 0, sizeof(struct epoll_event));
event.data.fd = s->sock;
event.events = EPOLLOUT;
if (0 > epoll_ctl(epoll_fd, EPOLL_CTL_ADD, s->sock, &event)) {
cf_warning(CF_SOCKET, "epoll_ctl(ADD) of client socket failed (errno %d: \"%s\")", errno, cf_strerror(errno));
goto Fail;
}
int tries = 0;
do {
int nevents = 0;
int max_events = 1;
int wait_ms = 1;
struct epoll_event events[max_events];
if (0 > (nevents = epoll_wait(epoll_fd, events, max_events, wait_ms))) {
if (errno == EINTR) {
cf_debug(CF_SOCKET, "epoll_wait() on client socket encountered EINTR ~~ Retrying!");
goto Retry;
} else {
cf_warning(CF_SOCKET, "epoll_wait() on client socket failed (errno %d: \"%s\") ~~ Failing!", errno, cf_strerror(errno));
goto Fail;
}
} else {
if (nevents == 0) {
cf_debug(CF_SOCKET, "epoll_wait() returned no events ~~ Retrying!");
goto Retry;
}
if (nevents != 1) {
cf_warning(CF_SOCKET, "epoll_wait() returned %d events ~~ only 1 expected, so ignoring others!", nevents);
}
if (events[0].data.fd == s->sock) {
if (events[0].events & EPOLLOUT) {
cf_debug(CF_SOCKET, "epoll_wait() on client socket ready for write detected ~~ Succeeding!");
} else {
// (Note: ERR and HUP events are automatically waited for as well.)
if (events[0].events & (EPOLLERR | EPOLLHUP)) {
cf_debug(CF_SOCKET, "epoll_wait() on client socket detected failure event 0x%x ~~ Failing!", events[0].events);
} else {
cf_warning(CF_SOCKET, "epoll_wait() on client socket detected non-write events 0x%x ~~ Failing!", events[0].events);
}
goto Fail;
}
} else {
cf_warning(CF_SOCKET, "epoll_wait() on client socket returned event on unknown socket %d ~~ Retrying!", events[0].data.fd);
goto Retry;
}
if (0 > epoll_ctl(epoll_fd, EPOLL_CTL_DEL, s->sock, &event)) {
cf_warning(CF_SOCKET, "epoll_ctl(DEL) on client socket failed (errno %d: \"%s\")", errno, cf_strerror(errno));
}
close(epoll_fd);
goto Success;
}
Retry:
cf_debug(CF_SOCKET, "Connect epoll loop: Retry #%d", tries++);
if (start + timeout < cf_getms()) {
cf_warning(CF_SOCKET, "Error in delayed connect() to %s:%d: timed out", s->addr, s->port);
errno = ETIMEDOUT;
goto Fail;
}
} while (1);
}
Fail:
cf_debug(CF_SOCKET, "connect failed to %s:%d : %s", s->addr, s->port, cf_strerror(errno));
if (epoll_fd > 0) {
close(epoll_fd);
}
close(s->sock);
s->sock = -1;
return(-1);
} else {
cf_debug(CF_SOCKET, "client socket connect() to %s:%d in 1 try!", s->addr, s->port);
}
Success:
;
// regarding this: calling here doesn't seem terribly effective.
// on the fabric threads, it seems important to set no-delay much later
int flag = 1;
setsockopt(s->sock, SOL_TCP, TCP_NODELAY, &flag, sizeof(flag));
long farg = fcntl(s->sock, F_GETFL, 0);
fcntl(s->sock, F_SETFL, farg & (~O_NONBLOCK)); /* blocking again */
return(0);
}
// Keep track of the connections, since they're precious. Kill anything that
// hasn't been used in a while. The file handle array keeps a reference count,
// and allows a reaper to run through and find the ones to reap. The table is
// only written by the demarshal threads, and only read by the reaper thread.
void *
thr_demarshal_reaper_fn(void *arg)
{
uint64_t last = cf_getms();
while (true) {
uint64_t now = cf_getms();
uint inuse_cnt = 0;
uint64_t kill_ms = g_config.proto_fd_idle_ms;
bool refresh = false;
if (now - last > (uint64_t)(g_config.sec_cfg.privilege_refresh_period * 1000)) {
refresh = true;
last = now;
}
pthread_mutex_lock(&g_file_handle_a_LOCK);
for (int i = 0; i < g_file_handle_a_sz; i++) {
if (g_file_handle_a[i]) {
as_file_handle *fd_h = g_file_handle_a[i];
if (refresh) {
as_security_refresh(fd_h);
}
// Reap, if asked to.
if (fd_h->reap_me) {
cf_debug(AS_DEMARSHAL, "Reaping FD %d as requested", fd_h->fd);
g_file_handle_a[i] = 0;
cf_queue_push(g_freeslot, &i);
as_release_file_handle(fd_h);
fd_h = 0;
}
// Reap if past kill time.
else if ((0 != kill_ms) && (fd_h->last_used + kill_ms < now)) {
if (fd_h->fh_info & FH_INFO_DONOT_REAP) {
cf_debug(AS_DEMARSHAL, "Not reaping the fd %d as it has the protection bit set", fd_h->fd);
inuse_cnt++;
continue;
}
shutdown(fd_h->fd, SHUT_RDWR); // will trigger epoll errors
cf_debug(AS_DEMARSHAL, "remove unused connection, fd %d", fd_h->fd);
g_file_handle_a[i] = 0;
cf_queue_push(g_freeslot, &i);
as_release_file_handle(fd_h);
fd_h = 0;
cf_atomic_int_incr(&g_config.reaper_count);
}
else {
inuse_cnt++;
}
}
}
pthread_mutex_unlock(&g_file_handle_a_LOCK);
if ((g_file_handle_a_sz / 10) > (g_file_handle_a_sz - inuse_cnt)) {
cf_warning(AS_DEMARSHAL, "less than ten percent file handles remaining: %d max %d inuse",
g_file_handle_a_sz, inuse_cnt);
}
// Validate the system statistics.
if (g_config.proto_connections_opened - g_config.proto_connections_closed != inuse_cnt) {
cf_debug(AS_DEMARSHAL, "reaper: mismatched connection count: %d in stats vs %d calculated",
g_config.proto_connections_opened - g_config.proto_connections_closed,
inuse_cnt);
}
sleep(1);
}
return NULL;
}
