int
main(int argc, char **argv)
{
g_start_sec = cf_get_seconds();
// Initialize cf_thread wrapper.
cf_thread_init();
// Initialize memory allocation.
cf_alloc_init();
// Initialize fault management framework.
cf_fault_init();
// Setup signal handlers.
as_signal_setup();
// Initialize TLS library.
tls_check_init();
int opt;
int opt_i;
const char *config_file = DEFAULT_CONFIG_FILE;
bool run_in_foreground = false;
bool new_style_daemon = false;
bool cold_start_cmd = false;
uint32_t instance = 0;
// Parse command line options.
while ((opt = getopt_long(argc, argv, "", CMD_OPTS, &opt_i)) != -1) {
switch (opt) {
case 'h':
// printf() since we want stdout and don't want cf_fault's prefix.
printf("%s\n", HELP);
return 0;
case 'v':
// printf() since we want stdout and don't want cf_fault's prefix.
printf("%s build %s\n", aerospike_build_type, aerospike_build_id);
return 0;
case 'f':
config_file = cf_strdup(optarg);
break;
case 'F':
// As a "new-style" daemon(*), asd runs in the foreground and
// ignores the following configuration items:
// - user ('user')
// - group ('group')
// - PID file ('pidfile')
//
// If ignoring configuration items, or if the 'console' sink is not
// specified, warnings will appear in stderr.
//
// (*) http://0pointer.de/public/systemd-man/daemon.html#New-Style%20Daemons
run_in_foreground = true;
new_style_daemon = true;
break;
case 'd':
run_in_foreground = true;
break;
case 'c':
cold_start_cmd = true;
break;
case 'n':
instance = (uint32_t)strtol(optarg, NULL, 0);
break;
default:
// fprintf() since we don't want cf_fault's prefix.
fprintf(stderr, "%s\n", USAGE);
return 1;
}
}
// Set all fields in the global runtime configuration instance. This parses
// the configuration file, and creates as_namespace objects. (Return value
// is a shortcut pointer to the global runtime configuration instance.)
as_config *c = as_config_init(config_file);
// Detect NUMA topology and, if requested, prepare for CPU and NUMA pinning.
cf_topo_config(c->auto_pin, (cf_topo_numa_node_index)instance,
&c->service.bind);
// Perform privilege separation as necessary. If configured user & group
// don't have root privileges, all resources created or reopened past this
// point must be set up so that they are accessible without root privileges.
// If not, the process will self-terminate with (hopefully!) a log message
// indicating which resource is not set up properly.
cf_process_privsep(c->uid, c->gid);
//
// All resources such as files, devices, and shared memory must be created
// or reopened below this line! (The configuration file is the only thing
// that must be opened above, in order to parse the user & group.)
//==========================================================================
// A "new-style" daemon expects console logging to be configured. (If not,
// log messages won't be seen via the standard path.)
if (new_style_daemon) {
if (! cf_fault_console_is_held()) {
cf_warning(AS_AS, "in new-style daemon mode, console logging is not configured");
}
}
// Activate log sinks. Up to this point, 'cf_' log output goes to stderr,
// filtered according to NO_SINKS_LIMIT in fault.c. After this point, 'cf_'
// log output will appear in all log file sinks specified in configuration,
// with specified filtering. If console sink is specified in configuration,
// 'cf_' log output will continue going to stderr, but filtering will switch
// from NO_SINKS_LIMIT to that specified in console sink configuration.
if (0 != cf_fault_sink_activate_all_held()) {
// Specifics of failure are logged in cf_fault_sink_activate_all_held().
cf_crash_nostack(AS_AS, "can't open log sink(s)");
}
// Daemonize asd if specified. After daemonization, output to stderr will no
// longer appear in terminal. Instead, check /tmp/aerospike-console.<pid>
// for console output.
if (! run_in_foreground && c->run_as_daemon) {
// Don't close any open files when daemonizing. At this point only log
// sink files are open - instruct cf_process_daemonize() to ignore them.
int open_fds[CF_FAULT_SINKS_MAX];
int num_open_fds = cf_fault_sink_get_fd_list(open_fds);
cf_process_daemonize(open_fds, num_open_fds);
}
// Log which build this is - should be the first line in the log file.
cf_info(AS_AS, "<><><><><><><><><><> %s build %s <><><><><><><><><><>",
aerospike_build_type, aerospike_build_id);
// Includes echoing the configuration file to log.
as_config_post_process(c, config_file);
xdr_config_post_process();
// If we allocated a non-default config file name, free it.
if (config_file != DEFAULT_CONFIG_FILE) {
cf_free((void*)config_file);
}
// Write the pid file, if specified.
if (! new_style_daemon) {
write_pidfile(c->pidfile);
}
else {
if (c->pidfile) {
cf_warning(AS_AS, "will not write PID file in new-style daemon mode");
}
}
// Check that required directories are set up properly.
validate_directory(c->work_directory, "work");
validate_directory(c->mod_lua.user_path, "Lua user");
validate_smd_directory();
// Initialize subsystems. At this point we're allocating local resources,
// starting worker threads, etc. (But no communication with other server
// nodes or clients yet.)
as_json_init(); // Jansson JSON API used by System Metadata
as_index_tree_gc_init(); // thread to purge dropped index trees
as_sindex_thr_init(); // defrag secondary index (ok during population)
as_nsup_init(); // load previous evict-void-time(s)
// Initialize namespaces. Each namespace decides here whether it will do a
// warm or cold start. Index arenas, partition structures and index tree
// structures are initialized. Secondary index system metadata is restored.
as_namespaces_init(cold_start_cmd, instance);
// Initialize the storage system. For warm and cool restarts, this includes
// fully resuming persisted indexes - this may take a few minutes.
as_storage_init();
// Migrate memory to correct NUMA node (includes resumed index arenas).
cf_topo_migrate_memory();
// Drop capabilities that we kept only for initialization.
cf_process_drop_startup_caps();
// Activate the storage system. For cold starts and cool restarts, this
// includes full drive scans - this may take several hours. The defrag
// subsystem starts operating at the end of this call.
as_storage_load();
// Populate all secondary indexes. This may block for a long time.
as_sindex_boot_populateall();
cf_info(AS_AS, "initializing services...");
cf_dns_init(); // DNS resolver
as_netio_init(); // query responses
as_security_init(); // security features
as_tsvc_init(); // all transaction handling
as_hb_init(); // inter-node heartbeat
as_skew_monitor_init(); // clock skew monitor
as_fabric_init(); // inter-node communications
as_exchange_init(); // initialize the cluster exchange subsystem
as_clustering_init(); // clustering-v5 start
as_info_init(); // info transaction handling
as_migrate_init(); // move data between nodes
as_proxy_init(); // do work on behalf of others
as_rw_init(); // read & write service
as_query_init(); // query transaction handling
as_udf_init(); // user-defined functions
as_scan_init(); // scan a namespace or set
as_batch_init(); // batch transaction handling
as_xdr_init(); // cross data-center replication
as_mon_init(); // monitor
// Wait for enough available storage. We've been defragging all along, but
// here we wait until it's enough. This may block for a long time.
as_storage_wait_for_defrag();
// Start subsystems. At this point we may begin communicating with other
// cluster nodes, and ultimately with clients.
as_smd_start(); // enables receiving cluster state change events
as_health_start(); // starts before fabric and hb to capture them
as_fabric_start(); // may send & receive fabric messages
as_xdr_start(); // XDR should start before it joins other nodes
as_hb_start(); // start inter-node heartbeat
as_exchange_start(); // start the cluster exchange subsystem
as_clustering_start(); // clustering-v5 start
as_nsup_start(); // may send evict-void-time(s) to other nodes
as_service_start(); // server will now receive client transactions
as_info_port_start(); // server will now receive info transactions
as_ticker_start(); // only after everything else is started
// Relevant for enterprise edition only.
as_storage_start_tomb_raider();
// Log a service-ready message.
cf_info(AS_AS, "service ready: soon there will be cake!");
//--------------------------------------------
// Startup is done. This thread will now wait
// quietly for a shutdown signal.
//
// Stop this thread from finishing. Intentionally deadlocking on a mutex is
// a remarkably efficient way to do this.
pthread_mutex_lock(&g_main_deadlock);
g_startup_complete = true;
pthread_mutex_lock(&g_main_deadlock);
// When the service is running, you are here (deadlocked) - the signals that
// stop the service (yes, these signals always occur in this thread) will
// unlock the mutex, allowing us to continue.
g_shutdown_started = true;
pthread_mutex_unlock(&g_main_deadlock);
pthread_mutex_destroy(&g_main_deadlock);
//--------------------------------------------
// Received a shutdown signal.
//
as_storage_shutdown(instance);
as_xdr_shutdown();
cf_info(AS_AS, "finished clean shutdown - exiting");
// If shutdown was totally clean (all threads joined) we could just return,
// but for now we exit to make sure all threads die.
#ifdef DOPROFILE
exit(0); // exit(0) so profile build actually dumps gmon.out
#else
_exit(0);
#endif
return 0;
}
int
main(int argc, char **argv)
{
#ifdef USE_ASM
as_mallocation_t asm_array[MAX_NUM_MALLOCATIONS];
// Zero-out the statically-allocated array of memory allocation locations.
memset(asm_array, 0, sizeof(asm_array));
// Set the ASMalloc callback user data.
g_my_cb_udata = asm_array;
// This must come first to allow initialization of the ASMalloc library.
asm_init();
#endif // defined(USE_ASM)
#ifdef USE_JEM
// Initialize the JEMalloc interface.
jem_init(true);
#endif
// Initialize ref-counting system.
cf_rc_init(NULL);
// Initialize fault management framework.
cf_fault_init();
// Setup signal handlers.
as_signal_setup();
// Initialize the Jansson JSON API.
as_json_init();
int i;
int cmd_optidx;
const char *config_file = DEFAULT_CONFIG_FILE;
bool run_in_foreground = false;
bool cold_start_cmd = false;
uint32_t instance = 0;
// Parse command line options.
while (-1 != (i = getopt_long(argc, argv, "", cmd_opts, &cmd_optidx))) {
switch (i) {
case 'h':
// printf() since we want stdout and don't want cf_fault's prefix.
printf("%s\n", HELP);
return 0;
case 'v':
// printf() since we want stdout and don't want cf_fault's prefix.
printf("%s build %s\n", aerospike_build_type, aerospike_build_id);
return 0;
case 'f':
config_file = cf_strdup(optarg);
cf_assert(config_file, AS_AS, CF_CRITICAL, "config filename cf_strdup failed");
break;
case 'd':
run_in_foreground = true;
break;
case 'c':
cold_start_cmd = true;
break;
case 'n':
instance = (uint32_t)strtol(optarg, NULL, 0);
break;
default:
// fprintf() since we don't want cf_fault's prefix.
fprintf(stderr, "%s\n", USAGE);
return 1;
}
}
// Set all fields in the global runtime configuration instance. This parses
// the configuration file, and creates as_namespace objects. (Return value
// is a shortcut pointer to the global runtime configuration instance.)
as_config *c = as_config_init(config_file);
#ifdef USE_ASM
g_asm_hook_enabled = g_asm_cb_enabled = c->asmalloc_enabled;
long initial_tid = syscall(SYS_gettid);
#endif
#ifdef MEM_COUNT
// [Note: This should ideally be at the very start of the "main()" function,
// but we need to wait until after the config file has been parsed in
// order to support run-time configurability.]
mem_count_init(c->memory_accounting ? MEM_COUNT_ENABLE : MEM_COUNT_DISABLE);
#endif
// Perform privilege separation as necessary. If configured user & group
// don't have root privileges, all resources created or reopened past this
// point must be set up so that they are accessible without root privileges.
// If not, the process will self-terminate with (hopefully!) a log message
// indicating which resource is not set up properly.
if (0 != c->uid && 0 == geteuid()) {
// To see this log, change NO_SINKS_LIMIT in fault.c:
cf_info(AS_AS, "privsep to %d %d", c->uid, c->gid);
cf_process_privsep(c->uid, c->gid);
}
//
// All resources such as files, devices, and shared memory must be created
// or reopened below this line! (The configuration file is the only thing
// that must be opened above, in order to parse the user & group.)
//==========================================================================
// Activate log sinks. Up to this point, 'cf_' log output goes to stderr,
// filtered according to NO_SINKS_LIMIT in fault.c. After this point, 'cf_'
// log output will appear in all log file sinks specified in configuration,
// with specified filtering. If console sink is specified in configuration,
// 'cf_' log output will continue going to stderr, but filtering will switch
// from NO_SINKS_LIMIT to that specified in console sink configuration.
if (0 != cf_fault_sink_activate_all_held()) {
// Specifics of failure are logged in cf_fault_sink_activate_all_held().
cf_crash_nostack(AS_AS, "can't open log sink(s)");
}
// Daemonize asd if specified. After daemonization, output to stderr will no
// longer appear in terminal. Instead, check /tmp/aerospike-console.<pid>
// for console output.
if (! run_in_foreground && c->run_as_daemon) {
// Don't close any open files when daemonizing. At this point only log
// sink files are open - instruct cf_process_daemonize() to ignore them.
int open_fds[CF_FAULT_SINKS_MAX];
int num_open_fds = cf_fault_sink_get_fd_list(open_fds);
cf_process_daemonize(open_fds, num_open_fds);
}
#ifdef USE_ASM
// Log the main thread's Linux Task ID (pre- and post-fork) to the console.
fprintf(stderr, "Initial main thread tid: %lu\n", initial_tid);
if (! run_in_foreground && c->run_as_daemon) {
fprintf(stderr, "Post-daemonize main thread tid: %lu\n",
syscall(SYS_gettid));
}
#endif
// Log which build this is - should be the first line in the log file.
cf_info(AS_AS, "<><><><><><><><><><> %s build %s <><><><><><><><><><>",
aerospike_build_type, aerospike_build_id);
// Includes echoing the configuration file to log.
as_config_post_process(c, config_file);
// If we allocated a non-default config file name, free it.
if (config_file != DEFAULT_CONFIG_FILE) {
cf_free((void*)config_file);
}
// Write the pid file, if specified.
write_pidfile(c->pidfile);
// Check that required directories are set up properly.
validate_directory(c->work_directory, "work");
validate_directory(c->mod_lua.system_path, "Lua system");
validate_directory(c->mod_lua.user_path, "Lua user");
validate_smd_directory();
// Initialize subsystems. At this point we're allocating local resources,
// starting worker threads, etc. (But no communication with other server
// nodes or clients yet.)
as_smd_init(); // System Metadata first - others depend on it
ai_init(); // before as_storage_init() populates indexes
as_sindex_thr_init(); // defrag secondary index (ok during population)
// Initialize namespaces. Each namespace decides here whether it will do a
// warm or cold start. Index arenas, partition structures and index tree
// structures are initialized. Secondary index system metadata is restored.
as_namespaces_init(cold_start_cmd, instance);
// Initialize the storage system. For cold starts, this includes reading
// all the objects off the drives. This may block for a long time. The
// defrag subsystem starts operating at the end of this call.
as_storage_init();
// Populate all secondary indexes. This may block for a long time.
as_sindex_boot_populateall();
cf_info(AS_AS, "initializing services...");
as_netio_init();
as_security_init(); // security features
as_tsvc_init(); // all transaction handling
as_hb_init(); // inter-node heartbeat
as_fabric_init(); // inter-node communications
as_info_init(); // info transaction handling
as_paxos_init(); // cluster consensus algorithm
as_migrate_init(); // move data between nodes
as_proxy_init(); // do work on behalf of others
as_write_init(); // write service
as_query_init(); // query transaction handling
as_udf_init(); // apply user-defined functions
as_scan_init(); // scan a namespace or set
as_batch_init(); // batch transaction handling
as_batch_direct_init(); // low priority transaction handling
as_xdr_init(); // cross data-center replication
as_mon_init(); // monitor
// Wait for enough available storage. We've been defragging all along, but
// here we wait until it's enough. This may block for a long time.
as_storage_wait_for_defrag();
// Start subsystems. At this point we may begin communicating with other
// cluster nodes, and ultimately with clients.
as_smd_start(c->smd); // enables receiving paxos state change events
as_fabric_start(); // may send & receive fabric messages
as_hb_start(); // start inter-node heatbeat
as_paxos_start(); // blocks until cluster membership is obtained
as_nsup_start(); // may send delete transactions to other nodes
as_demarshal_start(); // server will now receive client transactions
as_info_port_start(); // server will now receive info transactions
info_debug_ticker_start(); // only after everything else is started
// Log a service-ready message.
cf_info(AS_AS, "service ready: soon there will be cake!");
//--------------------------------------------
// Startup is done. This thread will now wait
// quietly for a shutdown signal.
//
// Stop this thread from finishing. Intentionally deadlocking on a mutex is
// a remarkably efficient way to do this.
pthread_mutex_init(&g_NONSTOP, NULL);
pthread_mutex_lock(&g_NONSTOP);
g_startup_complete = true;
pthread_mutex_lock(&g_NONSTOP);
// When the service is running, you are here (deadlocked) - the signals that
// stop the service (yes, these signals always occur in this thread) will
// unlock the mutex, allowing us to continue.
g_shutdown_started = true;
pthread_mutex_unlock(&g_NONSTOP);
pthread_mutex_destroy(&g_NONSTOP);
//--------------------------------------------
// Received a shutdown signal.
//
as_storage_shutdown();
as_xdr_shutdown();
as_smd_shutdown(c->smd);
cf_info(AS_AS, "finished clean shutdown - exiting");
// If shutdown was totally clean (all threads joined) we could just return,
// but for now we exit to make sure all threads die.
#ifdef DOPROFILE
exit(0); // exit(0) so profile build actually dumps gmon.out
#else
_exit(0);
#endif
return 0;
}