#endif
#define WORKER_FACTOR_SCALE 16 /* scale factor to allow fractional values */
static unsigned int worker_factor = DEFAULT_WORKER_FACTOR * WORKER_FACTOR_SCALE;
-
-static int threads_per_child = 0; /* Worker threads per child */
-static int ap_daemons_to_start = 0;
-static int min_spare_threads = 0;
-static int max_spare_threads = 0;
-static int ap_daemons_limit = 0;
-static int max_workers = 0;
-static int server_limit = 0;
-static int thread_limit = 0;
+ /* AsyncRequestWorkerFactor * 16 */
+
+static int threads_per_child = 0; /* ThreadsPerChild */
+static int ap_daemons_to_start = 0; /* StartServers */
+static int min_spare_threads = 0; /* MinSpareThreads */
+static int max_spare_threads = 0; /* MaxSpareThreads */
+static int active_daemons_limit = 0; /* MaxRequestWorkers / ThreadsPerChild */
+static int active_daemons = 0; /* workers that still active, i.e. are
+ not shutting down gracefully */
+static int max_workers = 0; /* MaxRequestWorkers */
+static int server_limit = 0; /* ServerLimit */
+static int thread_limit = 0; /* ThreadLimit */
static int had_healthy_child = 0;
static int dying = 0;
static int workers_may_exit = 0;
static int start_thread_may_exit = 0;
static int listener_may_exit = 0;
+static int listener_is_wakeable = 0; /* Pollset supports APR_POLLSET_WAKEABLE */
static int num_listensocks = 0;
static apr_int32_t conns_this_child; /* MaxConnectionsPerChild, only access
in listener thread */
static apr_uint32_t lingering_count = 0; /* Number of connections in lingering close */
static apr_uint32_t suspended_count = 0; /* Number of suspended connections */
static apr_uint32_t clogged_count = 0; /* Number of threads processing ssl conns */
+static apr_uint32_t threads_shutdown = 0; /* Number of threads that have shutdown
+ early during graceful termination */
static int resource_shortage = 0;
static fd_queue_t *worker_queue;
static fd_queue_info_t *worker_queue_info;
-static int mpm_state = AP_MPMQ_STARTING;
static apr_thread_mutex_t *timeout_mutex;
struct event_srv_cfg_s;
typedef struct event_srv_cfg_s event_srv_cfg;
+static apr_pollfd_t *listener_pollfd;
+
+/*
+ * The pollset for sockets that are in any of the timeout queues. Currently
+ * we use the timeout_mutex to make sure that connections are added/removed
+ * atomically to/from both event_pollset and a timeout queue. Otherwise
+ * some confusion can happen under high load if timeout queues and pollset
+ * get out of sync.
+ * XXX: It should be possible to make the lock unnecessary in many or even all
+ * XXX: cases.
+ */
+static apr_pollset_t *event_pollset;
+
+/*
+ * The chain of connections to be shutdown by a worker thread (deferred),
+ * linked list updated atomically.
+ */
+static event_conn_state_t *volatile defer_linger_chain;
+
struct event_conn_state_t {
/** APR_RING of expiration timeouts */
APR_RING_ENTRY(event_conn_state_t) timeout_list;
apr_pollfd_t pfd;
/** public parts of the connection state */
conn_state_t pub;
+ /** chaining in defer_linger_chain */
+ struct event_conn_state_t *chain;
};
+
APR_RING_HEAD(timeout_head_t, event_conn_state_t);
struct timeout_queue {
struct timeout_head_t head;
- int count, *total;
apr_interval_time_t timeout;
- struct timeout_queue *next;
+ apr_uint32_t count; /* for this queue */
+ apr_uint32_t *total; /* for all chained/related queues */
+ struct timeout_queue *next; /* chaining */
};
/*
* Several timeout queues that use different timeouts, so that we always can
*keepalive_q,
*linger_q,
*short_linger_q;
+static volatile apr_time_t queues_next_expiry;
-static apr_pollfd_t *listener_pollfd;
+/* Prevent extra poll/wakeup calls for timeouts close in the future (queues
+ * have the granularity of a second anyway).
+ * XXX: Wouldn't 0.5s (instead of 0.1s) be "enough"?
+ */
+#define TIMEOUT_FUDGE_FACTOR apr_time_from_msec(100)
/*
* Macros for accessing struct timeout_queue.
* For TO_QUEUE_APPEND and TO_QUEUE_REMOVE, timeout_mutex must be held.
*/
-#define TO_QUEUE_APPEND(q, el) \
- do { \
- APR_RING_INSERT_TAIL(&(q)->head, el, event_conn_state_t, \
- timeout_list); \
- ++*(q)->total; \
- ++(q)->count; \
- } while (0)
-
-#define TO_QUEUE_REMOVE(q, el) \
- do { \
- APR_RING_REMOVE(el, timeout_list); \
- --*(q)->total; \
- --(q)->count; \
- } while (0)
-
-#define TO_QUEUE_INIT(q, p, t, v) \
- do { \
- struct timeout_queue *b = (v); \
- (q) = apr_palloc((p), sizeof *(q)); \
- APR_RING_INIT(&(q)->head, event_conn_state_t, timeout_list); \
- (q)->total = (b) ? (b)->total : apr_pcalloc((p), sizeof *(q)->total); \
- (q)->count = 0; \
- (q)->timeout = (t); \
- (q)->next = NULL; \
- } while (0)
-
-#define TO_QUEUE_ELEM_INIT(el) APR_RING_ELEM_INIT(el, timeout_list)
+static void TO_QUEUE_APPEND(struct timeout_queue *q, event_conn_state_t *el)
+{
+ apr_time_t q_expiry;
+ apr_time_t next_expiry;
-/*
- * The pollset for sockets that are in any of the timeout queues. Currently
- * we use the timeout_mutex to make sure that connections are added/removed
- * atomically to/from both event_pollset and a timeout queue. Otherwise
- * some confusion can happen under high load if timeout queues and pollset
- * get out of sync.
- * XXX: It should be possible to make the lock unnecessary in many or even all
- * XXX: cases.
- */
-static apr_pollset_t *event_pollset;
+ APR_RING_INSERT_TAIL(&q->head, el, event_conn_state_t, timeout_list);
+ apr_atomic_inc32(q->total);
+ ++q->count;
+
+ /* Cheaply update the overall queues' next expiry according to the
+ * first entry of this queue (oldest), if necessary.
+ */
+ el = APR_RING_FIRST(&q->head);
+ q_expiry = el->queue_timestamp + q->timeout;
+ next_expiry = queues_next_expiry;
+ if (!next_expiry || next_expiry > q_expiry + TIMEOUT_FUDGE_FACTOR) {
+ queues_next_expiry = q_expiry;
+ /* Unblock the poll()ing listener for it to update its timeout. */
+ if (listener_is_wakeable) {
+ apr_pollset_wakeup(event_pollset);
+ }
+ }
+}
+
+static void TO_QUEUE_REMOVE(struct timeout_queue *q, event_conn_state_t *el)
+{
+ APR_RING_REMOVE(el, timeout_list);
+ APR_RING_ELEM_INIT(el, timeout_list);
+ apr_atomic_dec32(q->total);
+ --q->count;
+}
+
+static struct timeout_queue *TO_QUEUE_MAKE(apr_pool_t *p, apr_time_t t,
+ struct timeout_queue *ref)
+{
+ struct timeout_queue *q;
+
+ q = apr_pcalloc(p, sizeof *q);
+ APR_RING_INIT(&q->head, event_conn_state_t, timeout_list);
+ q->total = (ref) ? ref->total : apr_pcalloc(p, sizeof *q->total);
+ q->timeout = t;
+
+ return q;
+}
+
+#define TO_QUEUE_ELEM_INIT(el) \
+ APR_RING_ELEM_INIT((el), timeout_list)
/* The structure used to pass unique initialization info to each thread */
typedef struct
{
- int pid;
- int tid;
- int sd;
+ int pslot; /* process slot */
+ int tslot; /* worker slot of the thread */
} proc_info;
/* Structure used to pass information to the thread responsible for
* subsequent calls to pre-config hook
*/
typedef struct event_retained_data {
+ ap_unixd_mpm_retained_data *mpm;
+
int first_server_limit;
int first_thread_limit;
- int module_loads;
int sick_child_detected;
- ap_generation_t my_generation;
- int volatile is_graceful; /* set from signal handler */
int maxclients_reported;
/*
* The max child slot ever assigned, preserved across restarts. Necessary
* scoreboard.
*/
int max_daemons_limit;
+
+ /*
+ * All running workers, active and shutting down, including those that
+ * may be left from before a graceful restart.
+ * Not kept up-to-date when shutdown is pending.
+ */
+ int total_daemons;
+
/*
* idle_spawn_rate is the number of children that will be spawned on the
* next maintenance cycle if there aren't enough idle servers. It is
#define MAX_SPAWN_RATE (32)
#endif
int hold_off_on_exponential_spawning;
- /*
- * Current number of listeners buckets and maximum reached accross
- * restarts (to size retained data according to dynamic num_buckets,
- * eg. idle_spawn_rate).
- */
- int num_buckets, max_buckets;
} event_retained_data;
static event_retained_data *retained;
static void enable_listensocks(int process_slot)
{
int i;
+ if (listener_may_exit) {
+ return;
+ }
ap_log_error(APLOG_MARK, APLOG_DEBUG, 0, ap_server_conf, APLOGNO(00457)
"Accepting new connections again: "
"%u active conns (%u lingering/%u clogged/%u suspended), "
ap_scoreboard_image->parent[process_slot].not_accepting = 0;
}
+static void abort_socket_nonblocking(apr_socket_t *csd)
+{
+ apr_status_t rv;
+ apr_socket_timeout_set(csd, 0);
+#if defined(SOL_SOCKET) && defined(SO_LINGER)
+ /* This socket is over now, and we don't want to block nor linger
+ * anymore, so reset it. A normal close could still linger in the
+ * system, while RST is fast, nonblocking, and what the peer will
+ * get if it sends us further data anyway.
+ */
+ {
+ apr_os_sock_t osd = -1;
+ struct linger opt;
+ opt.l_onoff = 1;
+ opt.l_linger = 0; /* zero timeout is RST */
+ apr_os_sock_get(&osd, csd);
+ setsockopt(osd, SOL_SOCKET, SO_LINGER, (void *)&opt, sizeof opt);
+ }
+#endif
+ rv = apr_socket_close(csd);
+ if (rv != APR_SUCCESS) {
+ ap_log_error(APLOG_MARK, APLOG_ERR, rv, ap_server_conf, APLOGNO(00468)
+ "error closing socket");
+ AP_DEBUG_ASSERT(0);
+ }
+}
+
static void close_worker_sockets(void)
{
int i;
for (i = 0; i < threads_per_child; i++) {
- if (worker_sockets[i]) {
- apr_socket_close(worker_sockets[i]);
+ apr_socket_t *csd = worker_sockets[i];
+ if (csd) {
worker_sockets[i] = NULL;
+ abort_socket_nonblocking(csd);
+ }
+ }
+ for (;;) {
+ event_conn_state_t *cs = defer_linger_chain;
+ if (!cs) {
+ break;
}
+ if (apr_atomic_casptr((void *)&defer_linger_chain, cs->chain,
+ cs) != cs) {
+ /* Race lost, try again */
+ continue;
+ }
+ cs->chain = NULL;
+ abort_socket_nonblocking(cs->pfd.desc.s);
}
}
return;
}
+ /* Unblock the listener if it's poll()ing */
+ if (listener_is_wakeable) {
+ apr_pollset_wakeup(event_pollset);
+ }
+
/* unblock the listener if it's waiting for a worker */
ap_queue_info_term(worker_queue_info);
return;
}
terminate_mode = mode;
- mpm_state = AP_MPMQ_STOPPING;
+ retained->mpm->mpm_state = AP_MPMQ_STOPPING;
/* in case we weren't called from the listener thread, wake up the
* listener thread
*result = ap_max_requests_per_child;
break;
case AP_MPMQ_MAX_DAEMONS:
- *result = ap_daemons_limit;
+ *result = active_daemons_limit;
break;
case AP_MPMQ_MPM_STATE:
- *result = mpm_state;
+ *result = retained->mpm->mpm_state;
break;
case AP_MPMQ_GENERATION:
- *result = retained->my_generation;
+ *result = retained->mpm->my_generation;
break;
default:
*rv = APR_ENOTIMPL;
ap_scoreboard_image->parent[slot].pid = pid;
ap_run_child_status(ap_server_conf,
ap_scoreboard_image->parent[slot].pid,
- retained->my_generation, slot, MPM_CHILD_STARTED);
-}
-
-static void event_note_child_lost_slot(int slot, pid_t newpid)
-{
- ap_log_error(APLOG_MARK, APLOG_DEBUG, 0, ap_server_conf, APLOGNO(00458)
- "pid %" APR_PID_T_FMT " taking over scoreboard slot from "
- "%" APR_PID_T_FMT "%s",
- newpid,
- ap_scoreboard_image->parent[slot].pid,
- ap_scoreboard_image->parent[slot].quiescing ?
- " (quiescing)" : "");
- ap_run_child_status(ap_server_conf,
- ap_scoreboard_image->parent[slot].pid,
- ap_scoreboard_image->parent[slot].generation,
- slot, MPM_CHILD_LOST_SLOT);
- /* Don't forget about this exiting child process, or we
- * won't be able to kill it if it doesn't exit by the
- * time the server is shut down.
- */
- ap_register_extra_mpm_process(ap_scoreboard_image->parent[slot].pid,
- ap_scoreboard_image->parent[slot].generation);
+ retained->mpm->my_generation, slot, MPM_CHILD_STARTED);
}
static const char *event_get_name(void)
static void clean_child_exit(int code) __attribute__ ((noreturn));
static void clean_child_exit(int code)
{
- mpm_state = AP_MPMQ_STOPPING;
+ retained->mpm->mpm_state = AP_MPMQ_STOPPING;
if (pchild) {
apr_pool_destroy(pchild);
}
static int child_fatal;
-/* volatile because they're updated from a signal handler */
-static int volatile shutdown_pending;
-static int volatile restart_pending;
-
static apr_status_t decrement_connection_count(void *cs_)
{
event_conn_state_t *cs = cs_;
default:
break;
}
- apr_atomic_dec32(&connection_count);
- return APR_SUCCESS;
-}
-
-/*
- * ap_start_shutdown() and ap_start_restart(), below, are a first stab at
- * functions to initiate shutdown or restart without relying on signals.
- * Previously this was initiated in sig_term() and restart() signal handlers,
- * but we want to be able to start a shutdown/restart from other sources --
- * e.g. on Win32, from the service manager. Now the service manager can
- * call ap_start_shutdown() or ap_start_restart() as appropriate. Note that
- * these functions can also be called by the child processes, since global
- * variables are no longer used to pass on the required action to the parent.
- *
- * These should only be called from the parent process itself, since the
- * parent process will use the shutdown_pending and restart_pending variables
- * to determine whether to shutdown or restart. The child process should
- * call signal_parent() directly to tell the parent to die -- this will
- * cause neither of those variable to be set, which the parent will
- * assume means something serious is wrong (which it will be, for the
- * child to force an exit) and so do an exit anyway.
- */
-
-static void ap_start_shutdown(int graceful)
-{
- mpm_state = AP_MPMQ_STOPPING;
- if (shutdown_pending == 1) {
- /* Um, is this _probably_ not an error, if the user has
- * tried to do a shutdown twice quickly, so we won't
- * worry about reporting it.
- */
- return;
- }
- shutdown_pending = 1;
- retained->is_graceful = graceful;
-}
-
-/* do a graceful restart if graceful == 1 */
-static void ap_start_restart(int graceful)
-{
- mpm_state = AP_MPMQ_STOPPING;
- if (restart_pending == 1) {
- /* Probably not an error - don't bother reporting it */
- return;
- }
- restart_pending = 1;
- retained->is_graceful = graceful;
-}
-
-static void sig_term(int sig)
-{
- ap_start_shutdown(sig == AP_SIG_GRACEFUL_STOP);
-}
-
-static void restart(int sig)
-{
- ap_start_restart(sig == AP_SIG_GRACEFUL);
-}
-
-static void set_signals(void)
-{
-#ifndef NO_USE_SIGACTION
- struct sigaction sa;
-#endif
-
- if (!one_process) {
- ap_fatal_signal_setup(ap_server_conf, pconf);
+ /* Unblock the listener if it's waiting for connection_count = 0 */
+ if (!apr_atomic_dec32(&connection_count)
+ && listener_is_wakeable && listener_may_exit) {
+ apr_pollset_wakeup(event_pollset);
}
-
-#ifndef NO_USE_SIGACTION
- sigemptyset(&sa.sa_mask);
- sa.sa_flags = 0;
-
- sa.sa_handler = sig_term;
- if (sigaction(SIGTERM, &sa, NULL) < 0)
- ap_log_error(APLOG_MARK, APLOG_WARNING, errno, ap_server_conf, APLOGNO(00459)
- "sigaction(SIGTERM)");
-#ifdef AP_SIG_GRACEFUL_STOP
- if (sigaction(AP_SIG_GRACEFUL_STOP, &sa, NULL) < 0)
- ap_log_error(APLOG_MARK, APLOG_WARNING, errno, ap_server_conf, APLOGNO(00460)
- "sigaction(" AP_SIG_GRACEFUL_STOP_STRING ")");
-#endif
-#ifdef SIGINT
- if (sigaction(SIGINT, &sa, NULL) < 0)
- ap_log_error(APLOG_MARK, APLOG_WARNING, errno, ap_server_conf, APLOGNO(00461)
- "sigaction(SIGINT)");
-#endif
-#ifdef SIGXCPU
- sa.sa_handler = SIG_DFL;
- if (sigaction(SIGXCPU, &sa, NULL) < 0)
- ap_log_error(APLOG_MARK, APLOG_WARNING, errno, ap_server_conf, APLOGNO(00462)
- "sigaction(SIGXCPU)");
-#endif
-#ifdef SIGXFSZ
- /* For systems following the LFS standard, ignoring SIGXFSZ allows
- * a write() beyond the 2GB limit to fail gracefully with E2BIG
- * rather than terminate the process. */
- sa.sa_handler = SIG_IGN;
- if (sigaction(SIGXFSZ, &sa, NULL) < 0)
- ap_log_error(APLOG_MARK, APLOG_WARNING, errno, ap_server_conf, APLOGNO(00463)
- "sigaction(SIGXFSZ)");
-#endif
-#ifdef SIGPIPE
- sa.sa_handler = SIG_IGN;
- if (sigaction(SIGPIPE, &sa, NULL) < 0)
- ap_log_error(APLOG_MARK, APLOG_WARNING, errno, ap_server_conf, APLOGNO(00464)
- "sigaction(SIGPIPE)");
-#endif
-
- /* we want to ignore HUPs and AP_SIG_GRACEFUL while we're busy
- * processing one */
- sigaddset(&sa.sa_mask, SIGHUP);
- sigaddset(&sa.sa_mask, AP_SIG_GRACEFUL);
- sa.sa_handler = restart;
- if (sigaction(SIGHUP, &sa, NULL) < 0)
- ap_log_error(APLOG_MARK, APLOG_WARNING, errno, ap_server_conf, APLOGNO(00465)
- "sigaction(SIGHUP)");
- if (sigaction(AP_SIG_GRACEFUL, &sa, NULL) < 0)
- ap_log_error(APLOG_MARK, APLOG_WARNING, errno, ap_server_conf, APLOGNO(00466)
- "sigaction(" AP_SIG_GRACEFUL_STRING ")");
-#else
- if (!one_process) {
-#ifdef SIGXCPU
- apr_signal(SIGXCPU, SIG_DFL);
-#endif /* SIGXCPU */
-#ifdef SIGXFSZ
- apr_signal(SIGXFSZ, SIG_IGN);
-#endif /* SIGXFSZ */
- }
-
- apr_signal(SIGTERM, sig_term);
-#ifdef SIGHUP
- apr_signal(SIGHUP, restart);
-#endif /* SIGHUP */
-#ifdef AP_SIG_GRACEFUL
- apr_signal(AP_SIG_GRACEFUL, restart);
-#endif /* AP_SIG_GRACEFUL */
-#ifdef AP_SIG_GRACEFUL_STOP
- apr_signal(AP_SIG_GRACEFUL_STOP, sig_term);
-#endif /* AP_SIG_GRACEFUL_STOP */
-#ifdef SIGPIPE
- apr_signal(SIGPIPE, SIG_IGN);
-#endif /* SIGPIPE */
-
-#endif
+ return APR_SUCCESS;
}
static void notify_suspend(event_conn_state_t *cs)
ap_run_resume_connection(cs->c, cs->r);
}
-static int start_lingering_close_common(event_conn_state_t *cs, int in_worker)
+/*
+ * Close our side of the connection, flushing data to the client first.
+ * Pre-condition: cs is not in any timeout queue and not in the pollset,
+ * timeout_mutex is not locked
+ * return: 0 if connection is fully closed,
+ * 1 if connection is lingering
+ * May only be called by worker thread.
+ */
+static int start_lingering_close_blocking(event_conn_state_t *cs)
{
apr_status_t rv;
struct timeout_queue *q;
apr_socket_t *csd = cs->pfd.desc.s;
+
+ if (ap_start_lingering_close(cs->c)) {
+ notify_suspend(cs);
+ apr_socket_close(csd);
+ ap_push_pool(worker_queue_info, cs->p);
+ return 0;
+ }
+
#ifdef AP_DEBUG
{
rv = apr_socket_timeout_set(csd, 0);
#else
apr_socket_timeout_set(csd, 0);
#endif
+
cs->queue_timestamp = apr_time_now();
/*
* If some module requested a shortened waiting period, only wait for
cs->pub.state = CONN_STATE_LINGER_NORMAL;
}
apr_atomic_inc32(&lingering_count);
- if (in_worker) {
- notify_suspend(cs);
- }
- else {
- cs->c->sbh = NULL;
- }
- apr_thread_mutex_lock(timeout_mutex);
- TO_QUEUE_APPEND(q, cs);
+ notify_suspend(cs);
+
cs->pfd.reqevents = (
cs->pub.sense == CONN_SENSE_WANT_WRITE ? APR_POLLOUT :
APR_POLLIN) | APR_POLLHUP | APR_POLLERR;
cs->pub.sense = CONN_SENSE_DEFAULT;
+ apr_thread_mutex_lock(timeout_mutex);
+ TO_QUEUE_APPEND(q, cs);
rv = apr_pollset_add(event_pollset, &cs->pfd);
- apr_thread_mutex_unlock(timeout_mutex);
if (rv != APR_SUCCESS && !APR_STATUS_IS_EEXIST(rv)) {
- ap_log_error(APLOG_MARK, APLOG_ERR, rv, ap_server_conf, APLOGNO(03092)
- "start_lingering_close: apr_pollset_add failure");
- apr_thread_mutex_lock(timeout_mutex);
TO_QUEUE_REMOVE(q, cs);
apr_thread_mutex_unlock(timeout_mutex);
+ ap_log_error(APLOG_MARK, APLOG_ERR, rv, ap_server_conf, APLOGNO(03092)
+ "start_lingering_close: apr_pollset_add failure");
apr_socket_close(cs->pfd.desc.s);
ap_push_pool(worker_queue_info, cs->p);
return 0;
}
+ apr_thread_mutex_unlock(timeout_mutex);
return 1;
}
/*
- * Close our side of the connection, flushing data to the client first.
+ * Defer flush and close of the connection by adding it to defer_linger_chain,
+ * for a worker to grab it and do the job (should that be blocking).
* Pre-condition: cs is not in any timeout queue and not in the pollset,
* timeout_mutex is not locked
- * return: 0 if connection is fully closed,
- * 1 if connection is lingering
- * May only be called by worker thread.
- */
-static int start_lingering_close_blocking(event_conn_state_t *cs)
-{
- if (ap_start_lingering_close(cs->c)) {
- notify_suspend(cs);
- ap_push_pool(worker_queue_info, cs->p);
- return 0;
- }
- return start_lingering_close_common(cs, 1);
-}
-
-/*
- * Close our side of the connection, NOT flushing data to the client.
- * This should only be called if there has been an error or if we know
- * that our send buffers are empty.
- * Pre-condition: cs is not in any timeout queue and not in the pollset,
- * timeout_mutex is not locked
- * return: 0 if connection is fully closed,
- * 1 if connection is lingering
- * may be called by listener thread
+ * return: 1 connection is alive (but aside and about to linger)
+ * May be called by listener thread.
*/
static int start_lingering_close_nonblocking(event_conn_state_t *cs)
{
- conn_rec *c = cs->c;
- apr_socket_t *csd = cs->pfd.desc.s;
-
- if (ap_prep_lingering_close(c)
- || c->aborted
- || ap_shutdown_conn(c, 0) != APR_SUCCESS || c->aborted
- || apr_socket_shutdown(csd, APR_SHUTDOWN_WRITE) != APR_SUCCESS) {
- apr_socket_close(csd);
- ap_push_pool(worker_queue_info, cs->p);
- return 0;
+ event_conn_state_t *chain;
+ for (;;) {
+ cs->chain = chain = defer_linger_chain;
+ if (apr_atomic_casptr((void *)&defer_linger_chain, cs,
+ chain) != chain) {
+ /* Race lost, try again */
+ continue;
+ }
+ return 1;
}
- return start_lingering_close_common(cs, 0);
}
/*
*/
static int stop_lingering_close(event_conn_state_t *cs)
{
- apr_status_t rv;
apr_socket_t *csd = ap_get_conn_socket(cs->c);
ap_log_error(APLOG_MARK, APLOG_TRACE4, 0, ap_server_conf,
"socket reached timeout in lingering-close state");
- rv = apr_socket_close(csd);
- if (rv != APR_SUCCESS) {
- ap_log_error(APLOG_MARK, APLOG_ERR, rv, ap_server_conf, APLOGNO(00468) "error closing socket");
- AP_DEBUG_ASSERT(0);
- }
+ abort_socket_nonblocking(csd);
ap_push_pool(worker_queue_info, cs->p);
+ if (dying)
+ ap_queue_interrupt_one(worker_queue);
return 0;
}
c->current_thread = thd;
/* Subsequent request on a conn, and thread number is part of ID */
c->id = conn_id;
+
+ if (c->aborted) {
+ cs->pub.state = CONN_STATE_LINGER;
+ }
}
- if (c->clogging_input_filters && !c->aborted) {
+ if (cs->pub.state == CONN_STATE_LINGER) {
+ /* do lingering close below */
+ }
+ else if (c->clogging_input_filters) {
/* Since we have an input filter which 'clogs' the input stream,
* like mod_ssl used to, lets just do the normal read from input
* filters, like the Worker MPM does. Filters that need to write
}
apr_atomic_dec32(&clogged_count);
}
-
+ else if (cs->pub.state == CONN_STATE_READ_REQUEST_LINE) {
read_request:
- if (cs->pub.state == CONN_STATE_READ_REQUEST_LINE) {
- if (!c->aborted) {
- ap_run_process_connection(c);
+ ap_run_process_connection(c);
- /* state will be updated upon return
- * fall thru to either wait for readability/timeout or
- * do lingering close
- */
- }
- else {
- cs->pub.state = CONN_STATE_LINGER;
- }
+ /* state will be updated upon return
+ * fall thru to either wait for readability/timeout or
+ * do lingering close
+ */
}
if (cs->pub.state == CONN_STATE_WRITE_COMPLETION) {
*/
cs->queue_timestamp = apr_time_now();
notify_suspend(cs);
- apr_thread_mutex_lock(timeout_mutex);
- TO_QUEUE_APPEND(cs->sc->wc_q, cs);
cs->pfd.reqevents = (
cs->pub.sense == CONN_SENSE_WANT_READ ? APR_POLLIN :
APR_POLLOUT) | APR_POLLHUP | APR_POLLERR;
cs->pub.sense = CONN_SENSE_DEFAULT;
+ apr_thread_mutex_lock(timeout_mutex);
+ TO_QUEUE_APPEND(cs->sc->wc_q, cs);
rc = apr_pollset_add(event_pollset, &cs->pfd);
- apr_thread_mutex_unlock(timeout_mutex);
+ if (rc != APR_SUCCESS && !APR_STATUS_IS_EEXIST(rc)) {
+ TO_QUEUE_REMOVE(cs->sc->wc_q, cs);
+ apr_thread_mutex_unlock(timeout_mutex);
+ ap_log_error(APLOG_MARK, APLOG_ERR, rc, ap_server_conf, APLOGNO(03465)
+ "process_socket: apr_pollset_add failure for "
+ "write completion");
+ apr_socket_close(cs->pfd.desc.s);
+ ap_push_pool(worker_queue_info, cs->p);
+ }
+ else {
+ apr_thread_mutex_unlock(timeout_mutex);
+ }
return;
}
else if (c->keepalive != AP_CONN_KEEPALIVE || c->aborted ||
- listener_may_exit) {
+ listener_may_exit) {
cs->pub.state = CONN_STATE_LINGER;
}
else if (c->data_in_input_filters) {
*/
cs->queue_timestamp = apr_time_now();
notify_suspend(cs);
- apr_thread_mutex_lock(timeout_mutex);
- TO_QUEUE_APPEND(cs->sc->ka_q, cs);
/* Add work to pollset. */
cs->pfd.reqevents = APR_POLLIN;
+ apr_thread_mutex_lock(timeout_mutex);
+ TO_QUEUE_APPEND(cs->sc->ka_q, cs);
rc = apr_pollset_add(event_pollset, &cs->pfd);
- apr_thread_mutex_unlock(timeout_mutex);
-
- if (rc != APR_SUCCESS) {
+ if (rc != APR_SUCCESS && !APR_STATUS_IS_EEXIST(rc)) {
+ TO_QUEUE_REMOVE(cs->sc->ka_q, cs);
+ apr_thread_mutex_unlock(timeout_mutex);
ap_log_error(APLOG_MARK, APLOG_ERR, rc, ap_server_conf, APLOGNO(03093)
- "process_socket: apr_pollset_add failure");
- AP_DEBUG_ASSERT(rc == APR_SUCCESS);
+ "process_socket: apr_pollset_add failure for "
+ "keep alive");
+ apr_socket_close(cs->pfd.desc.s);
+ ap_push_pool(worker_queue_info, cs->p);
+ return;
}
+ apr_thread_mutex_unlock(timeout_mutex);
}
else if (cs->pub.state == CONN_STATE_SUSPENDED) {
apr_atomic_inc32(&suspended_count);
}
/* wake up the main thread */
kill(ap_my_pid, SIGTERM);
+
+ ap_free_idle_pools(worker_queue_info);
+ ap_queue_interrupt_all(worker_queue);
}
}
}
/*
- * Pre-condition: pfd->cs is neither in pollset nor timeout queue
+ * Pre-condition: cs is neither in event_pollset nor a timeout queue
* this function may only be called by the listener
*/
-static apr_status_t push2worker(const apr_pollfd_t * pfd,
- apr_pollset_t * pollset)
+static apr_status_t push2worker(event_conn_state_t *cs, apr_socket_t *csd,
+ apr_pool_t *ptrans)
{
- listener_poll_type *pt = (listener_poll_type *) pfd->client_data;
- event_conn_state_t *cs = (event_conn_state_t *) pt->baton;
apr_status_t rc;
- rc = ap_queue_push(worker_queue, cs->pfd.desc.s, cs, cs->p);
+ if (cs) {
+ csd = cs->pfd.desc.s;
+ ptrans = cs->p;
+ }
+ rc = ap_queue_push(worker_queue, csd, cs, ptrans);
if (rc != APR_SUCCESS) {
+ ap_log_error(APLOG_MARK, APLOG_CRIT, rc, ap_server_conf, APLOGNO(00471)
+ "push2worker: ap_queue_push failed");
/* trash the connection; we couldn't queue the connected
* socket to a worker
*/
- apr_bucket_alloc_destroy(cs->bucket_alloc);
- apr_socket_close(cs->pfd.desc.s);
- ap_log_error(APLOG_MARK, APLOG_CRIT, rc,
- ap_server_conf, APLOGNO(00471) "push2worker: ap_queue_push failed");
- ap_push_pool(worker_queue_info, cs->p);
+ if (csd) {
+ abort_socket_nonblocking(csd);
+ }
+ if (ptrans) {
+ ap_push_pool(worker_queue_info, ptrans);
+ }
+ signal_threads(ST_GRACEFUL);
}
return rc;
static APR_RING_HEAD(timer_free_ring_t, timer_event_t) timer_free_ring;
static apr_skiplist *timer_skiplist;
+static volatile apr_time_t timers_next_expiry;
+
+/* Same goal as for TIMEOUT_FUDGE_FACTOR (avoid extra poll calls), but applied
+ * to timers. Since their timeouts are custom (user defined), we can't be too
+ * approximative here (hence using 0.01s).
+ */
+#define EVENT_FUDGE_FACTOR apr_time_from_msec(10)
/* The following compare function is used by apr_skiplist_insert() to keep the
* elements (timers) sorted and provide O(log n) complexity (this is also true
/* XXXXX: optimize */
te->when = t + apr_time_now();
- /* Okay, add sorted by when.. */
- apr_skiplist_insert(timer_skiplist, te);
+ {
+ apr_time_t next_expiry;
+
+ /* Okay, add sorted by when.. */
+ apr_skiplist_insert(timer_skiplist, te);
+
+ /* Cheaply update the overall timers' next expiry according to
+ * this event, if necessary.
+ */
+ next_expiry = timers_next_expiry;
+ if (!next_expiry || next_expiry > te->when + EVENT_FUDGE_FACTOR) {
+ timers_next_expiry = te->when;
+ /* Unblock the poll()ing listener for it to update its timeout. */
+ if (listener_is_wakeable) {
+ apr_pollset_wakeup(event_pollset);
+ }
+ }
+ }
apr_thread_mutex_unlock(g_timer_skiplist_mtx);
}
apr_thread_mutex_lock(timeout_mutex);
+ TO_QUEUE_REMOVE(q, cs);
rv = apr_pollset_remove(event_pollset, pfd);
- AP_DEBUG_ASSERT(rv == APR_SUCCESS);
+ apr_thread_mutex_unlock(timeout_mutex);
+ AP_DEBUG_ASSERT(rv == APR_SUCCESS || APR_STATUS_IS_NOTFOUND(rv));
rv = apr_socket_close(csd);
AP_DEBUG_ASSERT(rv == APR_SUCCESS);
- TO_QUEUE_REMOVE(q, cs);
- apr_thread_mutex_unlock(timeout_mutex);
- TO_QUEUE_ELEM_INIT(cs);
-
ap_push_pool(worker_queue_info, cs->p);
+ if (dying)
+ ap_queue_interrupt_one(worker_queue);
}
/* call 'func' for all elements of 'q' with timeout less than 'timeout_time'.
apr_time_t timeout_time,
int (*func)(event_conn_state_t *))
{
- int total = 0, count;
+ apr_uint32_t total = 0, count;
event_conn_state_t *first, *cs, *last;
struct timeout_head_t trash;
struct timeout_queue *qp;
apr_status_t rv;
- if (!*q->total) {
+ if (!apr_atomic_read32(q->total)) {
return;
}
count = 0;
cs = first = last = APR_RING_FIRST(&qp->head);
while (cs != APR_RING_SENTINEL(&qp->head, event_conn_state_t,
- timeout_list)
- /* Trash the entry if:
- * - no timeout_time was given (asked for all), or
- * - it expired (according to the queue timeout), or
- * - the system clock skewed in the past: no entry should be
- * registered above the given timeout_time (~now) + the queue
- * timeout, we won't keep any here (eg. for centuries).
- * Stop otherwise, no following entry will match thanks to the
- * single timeout per queue (entries are added to the end!).
- * This allows maintenance in O(1).
- */
- && (!timeout_time
- || cs->queue_timestamp + qp->timeout < timeout_time
- || cs->queue_timestamp > timeout_time + qp->timeout)) {
+ timeout_list)) {
+ /* Trash the entry if:
+ * - no timeout_time was given (asked for all), or
+ * - it expired (according to the queue timeout), or
+ * - the system clock skewed in the past: no entry should be
+ * registered above the given timeout_time (~now) + the queue
+ * timeout, we won't keep any here (eg. for centuries).
+ *
+ * Otherwise stop, no following entry will match thanks to the
+ * single timeout per queue (entries are added to the end!).
+ * This allows maintenance in O(1).
+ */
+ if (timeout_time
+ && cs->queue_timestamp + qp->timeout > timeout_time
+ && cs->queue_timestamp < timeout_time + qp->timeout) {
+ /* Since this is the next expiring of this queue, update the
+ * overall queues' next expiry if it's later than this one.
+ */
+ apr_time_t q_expiry = cs->queue_timestamp + qp->timeout;
+ apr_time_t next_expiry = queues_next_expiry;
+ if (!next_expiry || next_expiry > q_expiry) {
+ queues_next_expiry = q_expiry;
+ }
+ break;
+ }
+
last = cs;
rv = apr_pollset_remove(event_pollset, &cs->pfd);
if (rv != APR_SUCCESS && !APR_STATUS_IS_NOTFOUND(rv)) {
APR_RING_UNSPLICE(first, last, timeout_list);
APR_RING_SPLICE_TAIL(&trash, first, last, event_conn_state_t,
timeout_list);
+ AP_DEBUG_ASSERT(apr_atomic_read32(q->total) >= count);
+ apr_atomic_sub32(q->total, count);
qp->count -= count;
total += count;
}
if (!total)
return;
- AP_DEBUG_ASSERT(*q->total >= total);
- *q->total -= total;
apr_thread_mutex_unlock(timeout_mutex);
first = APR_RING_FIRST(&trash);
do {
apr_thread_mutex_lock(timeout_mutex);
}
+static void process_keepalive_queue(apr_time_t timeout_time)
+{
+ /* If all workers are busy, we kill older keep-alive connections so
+ * that they may connect to another process.
+ */
+ if (!timeout_time) {
+ ap_log_error(APLOG_MARK, APLOG_TRACE1, 0, ap_server_conf,
+ "All workers are busy or dying, will close %u "
+ "keep-alive connections",
+ apr_atomic_read32(keepalive_q->total));
+ }
+ process_timeout_queue(keepalive_q, timeout_time,
+ start_lingering_close_nonblocking);
+}
+
static void * APR_THREAD_FUNC listener_thread(apr_thread_t * thd, void *dummy)
{
- timer_event_t *ep;
timer_event_t *te;
apr_status_t rc;
proc_info *ti = dummy;
- int process_slot = ti->pid;
+ int process_slot = ti->pslot;
+ struct process_score *ps = ap_get_scoreboard_process(process_slot);
apr_pool_t *tpool = apr_thread_pool_get(thd);
void *csd = NULL;
apr_pool_t *ptrans; /* Pool for per-transaction stuff */
last_log = apr_time_now();
free(ti);
- /* the following times out events that are really close in the future
- * to prevent extra poll calls
- *
- * current value is .1 second
- */
-#define TIMEOUT_FUDGE_FACTOR 100000
-#define EVENT_FUDGE_FACTOR 10000
-
rc = init_pollset(tpool);
if (rc != APR_SUCCESS) {
ap_log_error(APLOG_MARK, APLOG_ERR, rc, ap_server_conf,
for (;;) {
int workers_were_busy = 0;
+
if (listener_may_exit) {
close_listeners(process_slot, &closed);
if (terminate_mode == ST_UNGRACEFUL
now = apr_time_now();
if (APLOGtrace6(ap_server_conf)) {
/* trace log status every second */
- if (now - last_log > apr_time_from_msec(1000)) {
+ if (now - last_log > apr_time_from_sec(1)) {
last_log = now;
apr_thread_mutex_lock(timeout_mutex);
ap_log_error(APLOG_MARK, APLOG_TRACE6, 0, ap_server_conf,
"keep-alive: %d lingering: %d suspended: %u)",
apr_atomic_read32(&connection_count),
apr_atomic_read32(&clogged_count),
- *write_completion_q->total,
- *keepalive_q->total,
+ apr_atomic_read32(write_completion_q->total),
+ apr_atomic_read32(keepalive_q->total),
apr_atomic_read32(&lingering_count),
apr_atomic_read32(&suspended_count));
+ if (dying) {
+ ap_log_error(APLOG_MARK, APLOG_TRACE6, 0, ap_server_conf,
+ "%u/%u workers shutdown",
+ apr_atomic_read32(&threads_shutdown),
+ threads_per_child);
+ }
apr_thread_mutex_unlock(timeout_mutex);
}
}
- apr_thread_mutex_lock(g_timer_skiplist_mtx);
- te = apr_skiplist_peek(timer_skiplist);
- if (te) {
- if (te->when > now) {
- timeout_interval = te->when - now;
+ /* Start with an infinite poll() timeout and update it according to
+ * the next expiring timer or queue entry. If there are none, either
+ * the listener is wakeable and it can poll() indefinitely until a wake
+ * up occurs, otherwise periodic checks (maintenance, shutdown, ...)
+ * must be performed.
+ */
+ timeout_interval = -1;
+
+ /* Push expired timers to a worker, the first remaining one determines
+ * the maximum time to poll() below, if any.
+ */
+ timeout_time = timers_next_expiry;
+ if (timeout_time && timeout_time < now + EVENT_FUDGE_FACTOR) {
+ apr_thread_mutex_lock(g_timer_skiplist_mtx);
+ while ((te = apr_skiplist_peek(timer_skiplist))) {
+ if (te->when > now + EVENT_FUDGE_FACTOR) {
+ timers_next_expiry = te->when;
+ timeout_interval = te->when - now;
+ break;
+ }
+ apr_skiplist_pop(timer_skiplist, NULL);
+ push_timer2worker(te);
}
- else {
- timeout_interval = 1;
+ if (!te) {
+ timers_next_expiry = 0;
}
+ apr_thread_mutex_unlock(g_timer_skiplist_mtx);
}
- else {
- timeout_interval = apr_time_from_msec(100);
+
+ /* Same for queues, use their next expiry, if any. */
+ timeout_time = queues_next_expiry;
+ if (timeout_time
+ && (timeout_interval < 0
+ || timeout_time <= now
+ || timeout_interval > timeout_time - now)) {
+ timeout_interval = timeout_time > now ? timeout_time - now : 1;
+ }
+
+ /* When non-wakeable, don't wait more than 100 ms, in any case. */
+#define NON_WAKEABLE_POLL_TIMEOUT apr_time_from_msec(100)
+ if (!listener_is_wakeable
+ && (timeout_interval < 0
+ || timeout_interval > NON_WAKEABLE_POLL_TIMEOUT)) {
+ timeout_interval = NON_WAKEABLE_POLL_TIMEOUT;
}
- apr_thread_mutex_unlock(g_timer_skiplist_mtx);
rc = apr_pollset_poll(event_pollset, timeout_interval, &num, &out_pfd);
if (rc != APR_SUCCESS) {
if (APR_STATUS_IS_EINTR(rc)) {
- continue;
+ /* Woken up, if we are exiting we must fall through to kill
+ * kept-alive connections, otherwise we only need to update
+ * timeouts (logic is above, so restart the loop).
+ */
+ if (!listener_may_exit) {
+ continue;
+ }
+ timeout_time = 0;
}
- if (!APR_STATUS_IS_TIMEUP(rc)) {
+ else if (!APR_STATUS_IS_TIMEUP(rc)) {
ap_log_error(APLOG_MARK, APLOG_CRIT, rc, ap_server_conf,
"apr_pollset_poll failed. Attempting to "
"shutdown process gracefully");
signal_threads(ST_GRACEFUL);
}
+ num = 0;
}
if (listener_may_exit) {
break;
}
- now = apr_time_now();
- apr_thread_mutex_lock(g_timer_skiplist_mtx);
- ep = apr_skiplist_peek(timer_skiplist);
- while (ep) {
- if (ep->when < now + EVENT_FUDGE_FACTOR) {
- apr_skiplist_pop(timer_skiplist, NULL);
- push_timer2worker(ep);
- }
- else {
- break;
- }
- ep = apr_skiplist_peek(timer_skiplist);
- }
- apr_thread_mutex_unlock(g_timer_skiplist_mtx);
-
while (num) {
pt = (listener_poll_type *) out_pfd->client_data;
if (pt->type == PT_CSD) {
break;
}
- TO_QUEUE_ELEM_INIT(cs);
/* If we didn't get a worker immediately for a keep-alive
* request, we close the connection, so that the client can
* re-connect to a different process.
*/
if (!have_idle_worker) {
start_lingering_close_nonblocking(cs);
- break;
}
- rc = push2worker(out_pfd, event_pollset);
- if (rc != APR_SUCCESS) {
- ap_log_error(APLOG_MARK, APLOG_CRIT, rc,
- ap_server_conf, APLOGNO(03095)
- "push2worker failed");
- }
- else {
+ else if (push2worker(cs, NULL, NULL) == APR_SUCCESS) {
have_idle_worker = 0;
}
break;
+
case CONN_STATE_LINGER_NORMAL:
case CONN_STATE_LINGER_SHORT:
process_lingering_close(cs, out_pfd);
break;
+
default:
ap_log_error(APLOG_MARK, APLOG_CRIT, rc,
ap_server_conf, APLOGNO(03096)
if (csd != NULL) {
conns_this_child--;
- rc = ap_queue_push(worker_queue, csd, NULL, ptrans);
- if (rc != APR_SUCCESS) {
- /* trash the connection; we couldn't queue the connected
- * socket to a worker
- */
- apr_socket_close(csd);
- ap_log_error(APLOG_MARK, APLOG_CRIT, rc,
- ap_server_conf, APLOGNO(03098)
- "ap_queue_push failed");
- ap_push_pool(worker_queue_info, ptrans);
- }
- else {
+ if (push2worker(NULL, csd, ptrans) == APR_SUCCESS) {
have_idle_worker = 0;
}
}
/* XXX possible optimization: stash the current time for use as
* r->request_time for new requests
*/
- now = apr_time_now();
- /* We only do this once per 0.1s (TIMEOUT_FUDGE_FACTOR), or on a clock
- * skew (if the system time is set back in the meantime, timeout_time
- * will exceed now + TIMEOUT_FUDGE_FACTOR, can't happen otherwise).
+ /* We process the timeout queues here only when their overall next
+ * expiry (read once above) is over. This happens accurately since
+ * adding to the queues (in workers) can only decrease this expiry,
+ * while latest ones are only taken into account here (in listener)
+ * during queues' processing, with the lock held. This works both
+ * with and without wake-ability.
*/
- if (now > timeout_time || now + TIMEOUT_FUDGE_FACTOR < timeout_time ) {
- struct process_score *ps;
+ if (timeout_time && timeout_time < (now = apr_time_now())) {
timeout_time = now + TIMEOUT_FUDGE_FACTOR;
/* handle timed out sockets */
apr_thread_mutex_lock(timeout_mutex);
+ /* Processing all the queues below will recompute this. */
+ queues_next_expiry = 0;
+
/* Step 1: keepalive timeouts */
- /* If all workers are busy, we kill older keep-alive connections so that they
- * may connect to another process.
- */
- if (workers_were_busy && *keepalive_q->total) {
- ap_log_error(APLOG_MARK, APLOG_TRACE1, 0, ap_server_conf,
- "All workers are busy, will close %d keep-alive "
- "connections",
- *keepalive_q->total);
- process_timeout_queue(keepalive_q, 0,
- start_lingering_close_nonblocking);
+ if (workers_were_busy || dying) {
+ process_keepalive_queue(0); /* kill'em all \m/ */
}
else {
- process_timeout_queue(keepalive_q, timeout_time,
- start_lingering_close_nonblocking);
+ process_keepalive_queue(timeout_time);
}
/* Step 2: write completion timeouts */
process_timeout_queue(write_completion_q, timeout_time,
start_lingering_close_nonblocking);
/* Step 3: (normal) lingering close completion timeouts */
- process_timeout_queue(linger_q, timeout_time, stop_lingering_close);
+ process_timeout_queue(linger_q, timeout_time,
+ stop_lingering_close);
/* Step 4: (short) lingering close completion timeouts */
- process_timeout_queue(short_linger_q, timeout_time, stop_lingering_close);
+ process_timeout_queue(short_linger_q, timeout_time,
+ stop_lingering_close);
- ps = ap_get_scoreboard_process(process_slot);
- ps->write_completion = *write_completion_q->total;
- ps->keep_alive = *keepalive_q->total;
apr_thread_mutex_unlock(timeout_mutex);
+ ps->keep_alive = apr_atomic_read32(keepalive_q->total);
+ ps->write_completion = apr_atomic_read32(write_completion_q->total);
ps->connections = apr_atomic_read32(&connection_count);
ps->suspended = apr_atomic_read32(&suspended_count);
ps->lingering_close = apr_atomic_read32(&lingering_count);
}
+ else if ((workers_were_busy || dying)
+ && apr_atomic_read32(keepalive_q->total)) {
+ apr_thread_mutex_lock(timeout_mutex);
+ process_keepalive_queue(0); /* kill'em all \m/ */
+ apr_thread_mutex_unlock(timeout_mutex);
+ ps->keep_alive = 0;
+ }
+
+ /* If there are some lingering closes to defer (to a worker), schedule
+ * them now. We might wakeup a worker spuriously if another one empties
+ * defer_linger_chain in the meantime, but there also may be no active
+ * or all busy workers for an undefined time. In any case a deferred
+ * lingering close can't starve if we do that here since the chain is
+ * filled only above in the listener and it's emptied only in the
+ * worker(s); thus a NULL here means it will stay so while the listener
+ * waits (possibly indefinitely) in poll().
+ */
+ if (defer_linger_chain) {
+ get_worker(&have_idle_worker, 0, &workers_were_busy);
+ if (have_idle_worker
+ && defer_linger_chain /* re-test */
+ && push2worker(NULL, NULL, NULL) == APR_SUCCESS) {
+ have_idle_worker = 0;
+ }
+ }
+
if (listeners_disabled && !workers_were_busy
&& (int)apr_atomic_read32(&connection_count)
- (int)apr_atomic_read32(&lingering_count)
return NULL;
}
+/*
+ * During graceful shutdown, if there are more running worker threads than
+ * open connections, exit one worker thread.
+ *
+ * return 1 if thread should exit, 0 if it should continue running.
+ */
+static int worker_thread_should_exit_early(void)
+{
+ for (;;) {
+ apr_uint32_t conns = apr_atomic_read32(&connection_count);
+ apr_uint32_t dead = apr_atomic_read32(&threads_shutdown);
+ apr_uint32_t newdead;
+
+ AP_DEBUG_ASSERT(dead <= threads_per_child);
+ if (conns >= threads_per_child - dead)
+ return 0;
+
+ newdead = dead + 1;
+ if (apr_atomic_cas32(&threads_shutdown, newdead, dead) == dead) {
+ /*
+ * No other thread has exited in the mean time, safe to exit
+ * this one.
+ */
+ return 1;
+ }
+ }
+}
+
/* XXX For ungraceful termination/restart, we definitely don't want to
* wait for active connections to finish but we may want to wait
* for idle workers to get out of the queue code and release mutexes,
static void *APR_THREAD_FUNC worker_thread(apr_thread_t * thd, void *dummy)
{
proc_info *ti = dummy;
- int process_slot = ti->pid;
- int thread_slot = ti->tid;
+ int process_slot = ti->pslot;
+ int thread_slot = ti->tslot;
apr_socket_t *csd = NULL;
event_conn_state_t *cs;
apr_pool_t *ptrans; /* Pool for per-transaction stuff */
ap_scoreboard_image->servers[process_slot][thread_slot].pid = ap_my_pid;
ap_scoreboard_image->servers[process_slot][thread_slot].tid = apr_os_thread_current();
- ap_scoreboard_image->servers[process_slot][thread_slot].generation = retained->my_generation;
+ ap_scoreboard_image->servers[process_slot][thread_slot].generation = retained->mpm->my_generation;
ap_update_child_status_from_indexes(process_slot, thread_slot,
SERVER_STARTING, NULL);
if (workers_may_exit) {
break;
}
+ if (dying && worker_thread_should_exit_early()) {
+ break;
+ }
te = NULL;
rv = ap_queue_pop_something(worker_queue, &csd, &cs, &ptrans, &te);
}
else {
is_idle = 0;
- worker_sockets[thread_slot] = csd;
- process_socket(thd, ptrans, csd, cs, process_slot, thread_slot);
+ if (csd != NULL) {
+ worker_sockets[thread_slot] = csd;
+ process_socket(thd, ptrans, csd, cs, process_slot, thread_slot);
+ worker_sockets[thread_slot] = NULL;
+ }
+ }
+
+ /* If there are deferred lingering closes, handle them now. */
+ while (!workers_may_exit) {
+ cs = defer_linger_chain;
+ if (!cs) {
+ break;
+ }
+ if (apr_atomic_casptr((void *)&defer_linger_chain, cs->chain,
+ cs) != cs) {
+ /* Race lost, try again */
+ continue;
+ }
+ cs->chain = NULL;
+
+ worker_sockets[thread_slot] = csd = cs->pfd.desc.s;
+#ifdef AP_DEBUG
+ rv = apr_socket_timeout_set(csd, SECONDS_TO_LINGER);
+ AP_DEBUG_ASSERT(rv == APR_SUCCESS);
+#else
+ apr_socket_timeout_set(csd, SECONDS_TO_LINGER);
+#endif
+ cs->pub.state = CONN_STATE_LINGER;
+ process_socket(thd, cs->p, csd, cs, process_slot, thread_slot);
worker_sockets[thread_slot] = NULL;
}
}
apr_status_t rv;
my_info = (proc_info *) ap_malloc(sizeof(proc_info));
- my_info->pid = my_child_num;
- my_info->tid = -1; /* listener thread doesn't have a thread slot */
- my_info->sd = 0;
+ my_info->pslot = my_child_num;
+ my_info->tslot = -1; /* listener thread doesn't have a thread slot */
rv = apr_thread_create(&ts->listener, thread_attr, listener_thread,
my_info, pchild);
if (rv != APR_SUCCESS) {
int prev_threads_created;
int max_recycled_pools = -1;
int good_methods[] = {APR_POLLSET_KQUEUE, APR_POLLSET_PORT, APR_POLLSET_EPOLL};
+ /* XXX don't we need more to handle K-A or lingering close? */
+ const apr_uint32_t pollset_size = threads_per_child * 2;
/* We must create the fd queues before we start up the listener
* and worker threads. */
}
/* Create the main pollset */
- for (i = 0; i < sizeof(good_methods) / sizeof(void*); i++) {
- rv = apr_pollset_create_ex(&event_pollset,
- threads_per_child*2, /* XXX don't we need more, to handle
- * connections in K-A or lingering
- * close?
- */
- pchild, APR_POLLSET_THREADSAFE | APR_POLLSET_NOCOPY | APR_POLLSET_NODEFAULT,
- good_methods[i]);
+ for (i = 0; i < sizeof(good_methods) / sizeof(good_methods[0]); i++) {
+ apr_uint32_t flags = APR_POLLSET_THREADSAFE | APR_POLLSET_NOCOPY |
+ APR_POLLSET_NODEFAULT | APR_POLLSET_WAKEABLE;
+ rv = apr_pollset_create_ex(&event_pollset, pollset_size, pchild, flags,
+ good_methods[i]);
+ if (rv == APR_SUCCESS) {
+ listener_is_wakeable = 1;
+ break;
+ }
+ flags &= ~APR_POLLSET_WAKEABLE;
+ rv = apr_pollset_create_ex(&event_pollset, pollset_size, pchild, flags,
+ good_methods[i]);
if (rv == APR_SUCCESS) {
break;
}
}
if (rv != APR_SUCCESS) {
- rv = apr_pollset_create(&event_pollset,
- threads_per_child*2, /* XXX don't we need more, to handle
- * connections in K-A or lingering
- * close?
- */
- pchild, APR_POLLSET_THREADSAFE | APR_POLLSET_NOCOPY);
+ rv = apr_pollset_create(&event_pollset, pollset_size, pchild,
+ APR_POLLSET_THREADSAFE | APR_POLLSET_NOCOPY);
}
if (rv != APR_SUCCESS) {
ap_log_error(APLOG_MARK, APLOG_ERR, rv, ap_server_conf, APLOGNO(03103)
}
ap_log_error(APLOG_MARK, APLOG_DEBUG, 0, ap_server_conf, APLOGNO(02471)
- "start_threads: Using %s", apr_pollset_method_name(event_pollset));
+ "start_threads: Using %s (%swakeable)",
+ apr_pollset_method_name(event_pollset),
+ listener_is_wakeable ? "" : "not ");
worker_sockets = apr_pcalloc(pchild, threads_per_child
* sizeof(apr_socket_t *));
int status =
ap_scoreboard_image->servers[my_child_num][i].status;
- if (status != SERVER_GRACEFUL && status != SERVER_DEAD) {
+ if (status != SERVER_DEAD) {
continue;
}
my_info = (proc_info *) ap_malloc(sizeof(proc_info));
- my_info->pid = my_child_num;
- my_info->tid = i;
- my_info->sd = 0;
+ my_info->pslot = my_child_num;
+ my_info->tslot = i;
/* We are creating threads right now */
ap_update_child_status_from_indexes(my_child_num, i,
apr_thread_t *start_thread_id;
int i;
- mpm_state = AP_MPMQ_STARTING; /* for benefit of any hooks that run as this
- * child initializes
- */
+ /* for benefit of any hooks that run as this child initializes */
+ retained->mpm->mpm_state = AP_MPMQ_STARTING;
+
ap_my_pid = getpid();
ap_fatal_signal_child_setup(ap_server_conf);
apr_pool_create(&pchild, pconf);
/* close unused listeners and pods */
- for (i = 0; i < retained->num_buckets; i++) {
+ for (i = 0; i < retained->mpm->num_buckets; i++) {
if (i != child_bucket) {
ap_close_listeners_ex(all_buckets[i].listeners);
ap_mpm_podx_close(all_buckets[i].pod);
clean_child_exit(APEXIT_CHILDSICK);
}
- mpm_state = AP_MPMQ_RUNNING;
+ retained->mpm->mpm_state = AP_MPMQ_RUNNING;
/* If we are only running in one_process mode, we will want to
* still handle signals. */
retained->max_daemons_limit = slot + 1;
}
+ if (ap_scoreboard_image->parent[slot].pid != 0) {
+ /* XXX replace with assert or remove ? */
+ ap_log_error(APLOG_MARK, APLOG_ERR, 0, ap_server_conf, APLOGNO(03455)
+ "BUG: Scoreboard slot %d should be empty but is "
+ "in use by pid %" APR_PID_T_FMT,
+ slot, ap_scoreboard_image->parent[slot].pid);
+ return -1;
+ }
+
if (one_process) {
my_bucket = &all_buckets[0];
- set_signals();
event_note_child_started(slot, getpid());
child_main(slot, 0);
/* NOTREACHED */
return -1;
}
- if (ap_scoreboard_image->parent[slot].pid != 0) {
- /* This new child process is squatting on the scoreboard
- * entry owned by an exiting child process, which cannot
- * exit until all active requests complete.
- */
- event_note_child_lost_slot(slot, pid);
- }
ap_scoreboard_image->parent[slot].quiescing = 0;
ap_scoreboard_image->parent[slot].not_accepting = 0;
ap_scoreboard_image->parent[slot].bucket = bucket;
event_note_child_started(slot, pid);
+ active_daemons++;
+ retained->total_daemons++;
return 0;
}
{
int i;
- for (i = 0; number_to_start && i < ap_daemons_limit; ++i) {
+ for (i = 0; number_to_start && i < server_limit; ++i) {
if (ap_scoreboard_image->parent[i].pid != 0) {
continue;
}
- if (make_child(ap_server_conf, i, i % retained->num_buckets) < 0) {
+ if (make_child(ap_server_conf, i, i % retained->mpm->num_buckets) < 0) {
break;
}
--number_to_start;
static void perform_idle_server_maintenance(int child_bucket, int num_buckets)
{
int i, j;
- int idle_thread_count;
+ int idle_thread_count = 0;
worker_score *ws;
process_score *ps;
- int free_length;
- int totally_free_length = 0;
+ int free_length = 0;
int free_slots[MAX_SPAWN_RATE];
- int last_non_dead;
- int total_non_dead;
+ int last_non_dead = -1;
int active_thread_count = 0;
- /* initialize the free_list */
- free_length = 0;
-
- idle_thread_count = 0;
- last_non_dead = -1;
- total_non_dead = 0;
-
- for (i = 0; i < ap_daemons_limit; ++i) {
+ for (i = 0; i < server_limit; ++i) {
/* Initialization to satisfy the compiler. It doesn't know
* that threads_per_child is always > 0 */
int status = SERVER_DEAD;
- int any_dying_threads = 0;
- int any_dead_threads = 0;
- int all_dead_threads = 1;
int child_threads_active = 0;
if (i >= retained->max_daemons_limit &&
- totally_free_length == retained->idle_spawn_rate[child_bucket]) {
+ free_length == retained->idle_spawn_rate[child_bucket]) {
/* short cut if all active processes have been examined and
* enough empty scoreboard slots have been found
*/
break;
}
ps = &ap_scoreboard_image->parent[i];
- for (j = 0; j < threads_per_child; j++) {
- ws = &ap_scoreboard_image->servers[i][j];
- status = ws->status;
-
- /* XXX any_dying_threads is probably no longer needed GLA */
- any_dying_threads = any_dying_threads ||
- (status == SERVER_GRACEFUL);
- any_dead_threads = any_dead_threads || (status == SERVER_DEAD);
- all_dead_threads = all_dead_threads &&
- (status == SERVER_DEAD || status == SERVER_GRACEFUL);
-
- /* We consider a starting server as idle because we started it
- * at least a cycle ago, and if it still hasn't finished starting
- * then we're just going to swamp things worse by forking more.
- * So we hopefully won't need to fork more if we count it.
- * This depends on the ordering of SERVER_READY and SERVER_STARTING.
- */
- if (ps->pid != 0) { /* XXX just set all_dead_threads in outer
- for loop if no pid? not much else matters */
+ if (ps->pid != 0) {
+ for (j = 0; j < threads_per_child; j++) {
+ ws = &ap_scoreboard_image->servers[i][j];
+ status = ws->status;
+
+ /* We consider a starting server as idle because we started it
+ * at least a cycle ago, and if it still hasn't finished starting
+ * then we're just going to swamp things worse by forking more.
+ * So we hopefully won't need to fork more if we count it.
+ * This depends on the ordering of SERVER_READY and SERVER_STARTING.
+ */
if (status <= SERVER_READY && !ps->quiescing && !ps->not_accepting
- && ps->generation == retained->my_generation
+ && ps->generation == retained->mpm->my_generation
&& ps->bucket == child_bucket)
{
++idle_thread_count;
++child_threads_active;
}
}
+ last_non_dead = i;
}
active_thread_count += child_threads_active;
- if (any_dead_threads
- && totally_free_length < retained->idle_spawn_rate[child_bucket]
- && free_length < MAX_SPAWN_RATE / num_buckets
- && (!ps->pid /* no process in the slot */
- || ps->quiescing)) { /* or at least one is going away */
- if (all_dead_threads) {
- /* great! we prefer these, because the new process can
- * start more threads sooner. So prioritize this slot
- * by putting it ahead of any slots with active threads.
- *
- * first, make room by moving a slot that's potentially still
- * in use to the end of the array
- */
- free_slots[free_length] = free_slots[totally_free_length];
- free_slots[totally_free_length++] = i;
- }
- else {
- /* slot is still in use - back of the bus
- */
- free_slots[free_length] = i;
- }
- ++free_length;
- }
- else if (child_threads_active == threads_per_child) {
+ if (!ps->pid && free_length < retained->idle_spawn_rate[child_bucket])
+ free_slots[free_length++] = i;
+ else if (child_threads_active == threads_per_child)
had_healthy_child = 1;
- }
- /* XXX if (!ps->quiescing) is probably more reliable GLA */
- if (!any_dying_threads) {
- last_non_dead = i;
- ++total_non_dead;
- }
}
if (retained->sick_child_detected) {
else {
/* looks like a basket case, as no child ever fully initialized; give up.
*/
- shutdown_pending = 1;
+ retained->mpm->shutdown_pending = 1;
child_fatal = 1;
ap_log_error(APLOG_MARK, APLOG_ALERT, 0,
ap_server_conf, APLOGNO(02324)
retained->max_daemons_limit = last_non_dead + 1;
- if (idle_thread_count > max_spare_threads / num_buckets) {
- /* Kill off one child */
- ap_mpm_podx_signal(all_buckets[child_bucket].pod,
- AP_MPM_PODX_GRACEFUL);
- retained->idle_spawn_rate[child_bucket] = 1;
+ if (idle_thread_count > max_spare_threads / num_buckets)
+ {
+ /*
+ * Child processes that we ask to shut down won't die immediately
+ * but may stay around for a long time when they finish their
+ * requests. If the server load changes many times, many such
+ * gracefully finishing processes may accumulate, filling up the
+ * scoreboard. To avoid running out of scoreboard entries, we
+ * don't shut down more processes when the total number of processes
+ * is high.
+ *
+ * XXX It would be nice if we could
+ * XXX - kill processes without keepalive connections first
+ * XXX - tell children to stop accepting new connections, and
+ * XXX depending on server load, later be able to resurrect them
+ * or kill them
+ */
+ if (retained->total_daemons <= active_daemons_limit &&
+ retained->total_daemons < server_limit) {
+ /* Kill off one child */
+ ap_mpm_podx_signal(all_buckets[child_bucket].pod,
+ AP_MPM_PODX_GRACEFUL);
+ retained->idle_spawn_rate[child_bucket] = 1;
+ active_daemons--;
+ } else {
+ ap_log_error(APLOG_MARK, APLOG_TRACE5, 0, ap_server_conf,
+ "Not shutting down child: total daemons %d / "
+ "active limit %d / ServerLimit %d",
+ retained->total_daemons, active_daemons_limit,
+ server_limit);
+ }
}
else if (idle_thread_count < min_spare_threads / num_buckets) {
- /* terminate the free list */
- if (free_length == 0) { /* scoreboard is full, can't fork */
-
- if (active_thread_count >= ap_daemons_limit * threads_per_child) {
- if (!retained->maxclients_reported) {
- /* only report this condition once */
- ap_log_error(APLOG_MARK, APLOG_ERR, 0, ap_server_conf, APLOGNO(00484)
- "server reached MaxRequestWorkers setting, "
- "consider raising the MaxRequestWorkers "
- "setting");
- retained->maxclients_reported = 1;
- }
- }
- else {
- ap_log_error(APLOG_MARK, APLOG_ERR, 0, ap_server_conf, APLOGNO(00485)
- "scoreboard is full, not at MaxRequestWorkers");
+ if (active_thread_count >= max_workers) {
+ if (!retained->maxclients_reported) {
+ /* only report this condition once */
+ ap_log_error(APLOG_MARK, APLOG_ERR, 0, ap_server_conf, APLOGNO(00484)
+ "server reached MaxRequestWorkers setting, "
+ "consider raising the MaxRequestWorkers "
+ "setting");
+ retained->maxclients_reported = 1;
}
retained->idle_spawn_rate[child_bucket] = 1;
}
+ else if (free_length == 0) { /* scoreboard is full, can't fork */
+ ap_log_error(APLOG_MARK, APLOG_ERR, 0, ap_server_conf, APLOGNO(03490)
+ "scoreboard is full, not at MaxRequestWorkers."
+ "Increase ServerLimit.");
+ retained->idle_spawn_rate[child_bucket] = 1;
+ }
else {
if (free_length > retained->idle_spawn_rate[child_bucket]) {
free_length = retained->idle_spawn_rate[child_bucket];
"to increase StartServers, ThreadsPerChild "
"or Min/MaxSpareThreads), "
"spawning %d children, there are around %d idle "
- "threads, and %d total children", free_length,
- idle_thread_count, total_non_dead);
+ "threads, %d active children, and %d children "
+ "that are shutting down", free_length,
+ idle_thread_count, active_daemons,
+ retained->total_daemons);
}
for (i = 0; i < free_length; ++i) {
+ ap_log_error(APLOG_MARK, APLOG_TRACE5, 0, ap_server_conf,
+ "Spawning new child: slot %d active / "
+ "total daemons: %d/%d",
+ free_slots[i], active_daemons,
+ retained->total_daemons);
make_child(ap_server_conf, free_slots[i], child_bucket);
}
/* the next time around we want to spawn twice as many if this
static void server_main_loop(int remaining_children_to_start, int num_buckets)
{
- ap_generation_t old_gen;
int child_slot;
apr_exit_why_e exitwhy;
int status, processed_status;
apr_proc_t pid;
int i;
- while (!restart_pending && !shutdown_pending) {
+ while (!retained->mpm->restart_pending && !retained->mpm->shutdown_pending) {
ap_wait_or_timeout(&exitwhy, &status, &pid, pconf, ap_server_conf);
if (pid.pid != -1) {
*/
if (child_slot < 0
|| ap_get_scoreboard_process(child_slot)->generation
- == retained->my_generation) {
- shutdown_pending = 1;
+ == retained->mpm->my_generation) {
+ retained->mpm->shutdown_pending = 1;
child_fatal = 1;
+ /*
+ * total_daemons counting will be off now, but as we
+ * are shutting down, that is not an issue anymore.
+ */
return;
}
else {
event_note_child_killed(child_slot, 0, 0);
ps = &ap_scoreboard_image->parent[child_slot];
+ if (!ps->quiescing)
+ active_daemons--;
ps->quiescing = 0;
+ /* NOTE: We don't dec in the (child_slot < 0) case! */
+ retained->total_daemons--;
if (processed_status == APEXIT_CHILDSICK) {
/* resource shortage, minimize the fork rate */
retained->idle_spawn_rate[ps->bucket] = 1;
}
- else if (remaining_children_to_start
- && child_slot < ap_daemons_limit) {
+ else if (remaining_children_to_start) {
/* we're still doing a 1-for-1 replacement of dead
* children with new children
*/
--remaining_children_to_start;
}
}
- else if (ap_unregister_extra_mpm_process(pid.pid, &old_gen) == 1) {
-
- event_note_child_killed(-1, /* already out of the scoreboard */
- pid.pid, old_gen);
- if (processed_status == APEXIT_CHILDSICK
- && old_gen == retained->my_generation) {
- /* resource shortage, minimize the fork rate */
- for (i = 0; i < num_buckets; i++) {
- retained->idle_spawn_rate[i] = 1;
- }
- }
#if APR_HAS_OTHER_CHILD
- }
else if (apr_proc_other_child_alert(&pid, APR_OC_REASON_DEATH,
status) == 0) {
/* handled */
-#endif
}
- else if (retained->is_graceful) {
+#endif
+ else if (retained->mpm->was_graceful) {
/* Great, we've probably just lost a slot in the
* scoreboard. Somehow we don't know about this child.
*/
static int event_run(apr_pool_t * _pconf, apr_pool_t * plog, server_rec * s)
{
- int num_buckets = retained->num_buckets;
+ int num_buckets = retained->mpm->num_buckets;
int remaining_children_to_start;
int i;
ap_log_pid(pconf, ap_pid_fname);
- if (!retained->is_graceful) {
+ if (!retained->mpm->was_graceful) {
if (ap_run_pre_mpm(s->process->pool, SB_SHARED) != OK) {
- mpm_state = AP_MPMQ_STOPPING;
+ retained->mpm->mpm_state = AP_MPMQ_STOPPING;
return !OK;
}
/* fix the generation number in the global score; we just got a new,
* cleared scoreboard
*/
- ap_scoreboard_image->global->running_generation = retained->my_generation;
+ ap_scoreboard_image->global->running_generation = retained->mpm->my_generation;
}
- restart_pending = shutdown_pending = 0;
- set_signals();
+ if (!one_process) {
+ ap_fatal_signal_setup(ap_server_conf, pconf);
+ }
+ ap_unixd_mpm_set_signals(pconf, one_process);
/* Don't thrash since num_buckets depends on the
* system and the number of online CPU cores...
*/
- if (ap_daemons_limit < num_buckets)
- ap_daemons_limit = num_buckets;
+ if (active_daemons_limit < num_buckets)
+ active_daemons_limit = num_buckets;
if (ap_daemons_to_start < num_buckets)
ap_daemons_to_start = num_buckets;
+ /* We want to create as much children at a time as the number of buckets,
+ * so to optimally accept connections (evenly distributed across buckets).
+ * Thus min_spare_threads should at least maintain num_buckets children,
+ * and max_spare_threads allow num_buckets more children w/o triggering
+ * immediately (e.g. num_buckets idle threads margin, one per bucket).
+ */
if (min_spare_threads < threads_per_child * (num_buckets - 1) + num_buckets)
min_spare_threads = threads_per_child * (num_buckets - 1) + num_buckets;
- if (max_spare_threads < min_spare_threads + threads_per_child * num_buckets)
- max_spare_threads = min_spare_threads + threads_per_child * num_buckets;
+ if (max_spare_threads < min_spare_threads + (threads_per_child + 1) * num_buckets)
+ max_spare_threads = min_spare_threads + (threads_per_child + 1) * num_buckets;
/* If we're doing a graceful_restart then we're going to see a lot
* of children exiting immediately when we get into the main loop
* supposed to start up without the 1 second penalty between each fork.
*/
remaining_children_to_start = ap_daemons_to_start;
- if (remaining_children_to_start > ap_daemons_limit) {
- remaining_children_to_start = ap_daemons_limit;
+ if (remaining_children_to_start > active_daemons_limit) {
+ remaining_children_to_start = active_daemons_limit;
}
- if (!retained->is_graceful) {
+ if (!retained->mpm->was_graceful) {
startup_children(remaining_children_to_start);
remaining_children_to_start = 0;
}
ap_log_command_line(plog, s);
ap_log_mpm_common(s);
- mpm_state = AP_MPMQ_RUNNING;
+ retained->mpm->mpm_state = AP_MPMQ_RUNNING;
server_main_loop(remaining_children_to_start, num_buckets);
- mpm_state = AP_MPMQ_STOPPING;
+ retained->mpm->mpm_state = AP_MPMQ_STOPPING;
- if (shutdown_pending && !retained->is_graceful) {
+ if (retained->mpm->shutdown_pending && retained->mpm->is_ungraceful) {
/* Time to shut down:
* Kill child processes, tell them to call child_exit, etc...
*/
for (i = 0; i < num_buckets; i++) {
- ap_mpm_podx_killpg(all_buckets[i].pod, ap_daemons_limit,
+ ap_mpm_podx_killpg(all_buckets[i].pod, active_daemons_limit,
AP_MPM_PODX_RESTART);
}
ap_reclaim_child_processes(1, /* Start with SIGTERM */
ap_log_error(APLOG_MARK, APLOG_NOTICE, 0,
ap_server_conf, APLOGNO(00491) "caught SIGTERM, shutting down");
}
+
return DONE;
- } else if (shutdown_pending) {
+ }
+
+ if (retained->mpm->shutdown_pending) {
/* Time to gracefully shut down:
* Kill child processes, tell them to call child_exit, etc...
*/
/* Close our listeners, and then ask our children to do same */
ap_close_listeners();
for (i = 0; i < num_buckets; i++) {
- ap_mpm_podx_killpg(all_buckets[i].pod, ap_daemons_limit,
+ ap_mpm_podx_killpg(all_buckets[i].pod, active_daemons_limit,
AP_MPM_PODX_GRACEFUL);
}
ap_relieve_child_processes(event_note_child_killed);
}
/* Don't really exit until each child has finished */
- shutdown_pending = 0;
+ retained->mpm->shutdown_pending = 0;
do {
/* Pause for a second */
apr_sleep(apr_time_from_sec(1));
ap_relieve_child_processes(event_note_child_killed);
active_children = 0;
- for (index = 0; index < ap_daemons_limit; ++index) {
+ for (index = 0; index < retained->max_daemons_limit; ++index) {
if (ap_mpm_safe_kill(MPM_CHILD_PID(index), 0) == APR_SUCCESS) {
active_children = 1;
/* Having just one child is enough to stay around */
break;
}
}
- } while (!shutdown_pending && active_children &&
+ } while (!retained->mpm->shutdown_pending && active_children &&
(!ap_graceful_shutdown_timeout || apr_time_now() < cutoff));
/* We might be here because we received SIGTERM, either
* really dead.
*/
for (i = 0; i < num_buckets; i++) {
- ap_mpm_podx_killpg(all_buckets[i].pod, ap_daemons_limit,
+ ap_mpm_podx_killpg(all_buckets[i].pod, active_daemons_limit,
AP_MPM_PODX_RESTART);
}
ap_reclaim_child_processes(1, event_note_child_killed);
}
/* we've been told to restart */
- apr_signal(SIGHUP, SIG_IGN);
-
if (one_process) {
/* not worth thinking about */
return DONE;
/* XXX: we really need to make sure this new generation number isn't in
* use by any of the children.
*/
- ++retained->my_generation;
- ap_scoreboard_image->global->running_generation = retained->my_generation;
+ ++retained->mpm->my_generation;
+ ap_scoreboard_image->global->running_generation = retained->mpm->my_generation;
- if (retained->is_graceful) {
+ if (!retained->mpm->is_ungraceful) {
ap_log_error(APLOG_MARK, APLOG_NOTICE, 0, ap_server_conf, APLOGNO(00493)
AP_SIG_GRACEFUL_STRING
" received. Doing graceful restart");
/* wake up the children...time to die. But we'll have more soon */
for (i = 0; i < num_buckets; i++) {
- ap_mpm_podx_killpg(all_buckets[i].pod, ap_daemons_limit,
+ ap_mpm_podx_killpg(all_buckets[i].pod, active_daemons_limit,
AP_MPM_PODX_GRACEFUL);
}
* pthreads are stealing signals from us left and right.
*/
for (i = 0; i < num_buckets; i++) {
- ap_mpm_podx_killpg(all_buckets[i].pod, ap_daemons_limit,
+ ap_mpm_podx_killpg(all_buckets[i].pod, active_daemons_limit,
AP_MPM_PODX_RESTART);
}
"SIGHUP received. Attempting to restart");
}
+ active_daemons = 0;
+
return OK;
}
pconf = p;
/* the reverse of pre_config, we want this only the first time around */
- if (retained->module_loads == 1) {
+ if (retained->mpm->module_loads == 1) {
startup = 1;
level_flags |= APLOG_STARTUP;
}
if (one_process) {
num_buckets = 1;
}
- else if (retained->is_graceful) {
+ else if (retained->mpm->was_graceful) {
/* Preserve the number of buckets on graceful restarts. */
- num_buckets = retained->num_buckets;
+ num_buckets = retained->mpm->num_buckets;
}
if ((rv = ap_duplicate_listeners(pconf, ap_server_conf,
&listen_buckets, &num_buckets))) {
all_buckets[i].listeners = listen_buckets[i];
}
- if (retained->max_buckets < num_buckets) {
+ if (retained->mpm->max_buckets < num_buckets) {
int new_max, *new_ptr;
- new_max = retained->max_buckets * 2;
+ new_max = retained->mpm->max_buckets * 2;
if (new_max < num_buckets) {
new_max = num_buckets;
}
new_ptr = (int *)apr_palloc(ap_pglobal, new_max * sizeof(int));
memcpy(new_ptr, retained->idle_spawn_rate,
- retained->num_buckets * sizeof(int));
+ retained->mpm->num_buckets * sizeof(int));
retained->idle_spawn_rate = new_ptr;
- retained->max_buckets = new_max;
+ retained->mpm->max_buckets = new_max;
}
- if (retained->num_buckets < num_buckets) {
+ if (retained->mpm->num_buckets < num_buckets) {
int rate_max = 1;
/* If new buckets are added, set their idle spawn rate to
* the highest so far, so that they get filled as quickly
* as the existing ones.
*/
- for (i = 0; i < retained->num_buckets; i++) {
+ for (i = 0; i < retained->mpm->num_buckets; i++) {
if (rate_max < retained->idle_spawn_rate[i]) {
rate_max = retained->idle_spawn_rate[i];
}
retained->idle_spawn_rate[i] = rate_max;
}
}
- retained->num_buckets = num_buckets;
+ retained->mpm->num_buckets = num_buckets;
/* for skiplist */
srand((unsigned int)apr_time_now());
int no_detach, debug, foreground;
apr_status_t rv;
const char *userdata_key = "mpm_event_module";
-
- mpm_state = AP_MPMQ_STARTING;
+ int test_atomics = 0;
debug = ap_exists_config_define("DEBUG");
foreground = ap_exists_config_define("FOREGROUND");
}
- /* sigh, want this only the second time around */
retained = ap_retained_data_get(userdata_key);
if (!retained) {
retained = ap_retained_data_create(userdata_key, sizeof(*retained));
+ retained->mpm = ap_unixd_mpm_get_retained_data();
retained->max_daemons_limit = -1;
+ if (retained->mpm->module_loads) {
+ test_atomics = 1;
+ }
+ }
+ retained->mpm->mpm_state = AP_MPMQ_STARTING;
+ if (retained->mpm->baton != retained) {
+ retained->mpm->was_graceful = 0;
+ retained->mpm->baton = retained;
}
- ++retained->module_loads;
- if (retained->module_loads == 2) {
- /* test for correct operation of fdqueue */
+ ++retained->mpm->module_loads;
+
+ /* test once for correct operation of fdqueue */
+ if (test_atomics || retained->mpm->module_loads == 2) {
static apr_uint32_t foo1, foo2;
apr_atomic_set32(&foo1, 100);
"atomics not working as expected - add32 of negative number");
return HTTP_INTERNAL_SERVER_ERROR;
}
+ }
+ /* sigh, want this only the second time around */
+ if (retained->mpm->module_loads == 2) {
rv = apr_pollset_create(&event_pollset, 1, plog,
APR_POLLSET_THREADSAFE | APR_POLLSET_NOCOPY);
if (rv != APR_SUCCESS) {
max_spare_threads = DEFAULT_MAX_FREE_DAEMON * DEFAULT_THREADS_PER_CHILD;
server_limit = DEFAULT_SERVER_LIMIT;
thread_limit = DEFAULT_THREAD_LIMIT;
- ap_daemons_limit = server_limit;
+ active_daemons_limit = server_limit;
threads_per_child = DEFAULT_THREADS_PER_CHILD;
- max_workers = ap_daemons_limit * threads_per_child;
+ max_workers = active_daemons_limit * threads_per_child;
+ defer_linger_chain = NULL;
had_healthy_child = 0;
ap_extended_status = 0;
wc.hash = apr_hash_make(ptemp);
ka.hash = apr_hash_make(ptemp);
- TO_QUEUE_INIT(linger_q, pconf,
- apr_time_from_sec(MAX_SECS_TO_LINGER), NULL);
- TO_QUEUE_INIT(short_linger_q, pconf,
- apr_time_from_sec(SECONDS_TO_LINGER), NULL);
+ linger_q = TO_QUEUE_MAKE(pconf, apr_time_from_sec(MAX_SECS_TO_LINGER),
+ NULL);
+ short_linger_q = TO_QUEUE_MAKE(pconf, apr_time_from_sec(SECONDS_TO_LINGER),
+ NULL);
for (; s; s = s->next) {
event_srv_cfg *sc = apr_pcalloc(pconf, sizeof *sc);
ap_set_module_config(s->module_config, &mpm_event_module, sc);
if (!wc.tail) {
/* The main server uses the global queues */
- TO_QUEUE_INIT(wc.q, pconf, s->timeout, NULL);
+ wc.q = TO_QUEUE_MAKE(pconf, s->timeout, NULL);
apr_hash_set(wc.hash, &s->timeout, sizeof s->timeout, wc.q);
wc.tail = write_completion_q = wc.q;
- TO_QUEUE_INIT(ka.q, pconf, s->keep_alive_timeout, NULL);
+ ka.q = TO_QUEUE_MAKE(pconf, s->keep_alive_timeout, NULL);
apr_hash_set(ka.hash, &s->keep_alive_timeout,
sizeof s->keep_alive_timeout, ka.q);
ka.tail = keepalive_q = ka.q;
* or their own queue(s) if there isn't */
wc.q = apr_hash_get(wc.hash, &s->timeout, sizeof s->timeout);
if (!wc.q) {
- TO_QUEUE_INIT(wc.q, pconf, s->timeout, wc.tail);
+ wc.q = TO_QUEUE_MAKE(pconf, s->timeout, wc.tail);
apr_hash_set(wc.hash, &s->timeout, sizeof s->timeout, wc.q);
wc.tail = wc.tail->next = wc.q;
}
ka.q = apr_hash_get(ka.hash, &s->keep_alive_timeout,
sizeof s->keep_alive_timeout);
if (!ka.q) {
- TO_QUEUE_INIT(ka.q, pconf, s->keep_alive_timeout, ka.tail);
+ ka.q = TO_QUEUE_MAKE(pconf, s->keep_alive_timeout, ka.tail);
apr_hash_set(ka.hash, &s->keep_alive_timeout,
sizeof s->keep_alive_timeout, ka.q);
ka.tail = ka.tail->next = ka.q;
int startup = 0;
/* the reverse of pre_config, we want this only the first time around */
- if (retained->module_loads == 1) {
+ if (retained->mpm->module_loads == 1) {
startup = 1;
}
max_workers = threads_per_child;
}
- ap_daemons_limit = max_workers / threads_per_child;
+ active_daemons_limit = max_workers / threads_per_child;
if (max_workers % threads_per_child) {
- int tmp_max_workers = ap_daemons_limit * threads_per_child;
+ int tmp_max_workers = active_daemons_limit * threads_per_child;
if (startup) {
ap_log_error(APLOG_MARK, APLOG_WARNING | APLOG_STARTUP, 0, NULL, APLOGNO(00513)
"multiple of ThreadsPerChild of %d, decreasing to nearest "
"multiple %d, for a maximum of %d servers.",
max_workers, threads_per_child, tmp_max_workers,
- ap_daemons_limit);
+ active_daemons_limit);
} else {
ap_log_error(APLOG_MARK, APLOG_WARNING, 0, s, APLOGNO(00514)
"MaxRequestWorkers of %d is not an integer multiple "
max_workers = tmp_max_workers;
}
- if (ap_daemons_limit > server_limit) {
+ if (active_daemons_limit > server_limit) {
if (startup) {
ap_log_error(APLOG_MARK, APLOG_WARNING | APLOG_STARTUP, 0, NULL, APLOGNO(00515)
"WARNING: MaxRequestWorkers of %d would require %d servers "
"and would exceed ServerLimit of %d, decreasing to %d. "
"To increase, please see the ServerLimit directive.",
- max_workers, ap_daemons_limit, server_limit,
+ max_workers, active_daemons_limit, server_limit,
server_limit * threads_per_child);
} else {
ap_log_error(APLOG_MARK, APLOG_WARNING, 0, s, APLOGNO(00516)
"MaxRequestWorkers of %d would require %d servers and "
"exceed ServerLimit of %d, decreasing to %d",
- max_workers, ap_daemons_limit, server_limit,
+ max_workers, active_daemons_limit, server_limit,
server_limit * threads_per_child);
}
- ap_daemons_limit = server_limit;
+ active_daemons_limit = server_limit;
}
- /* ap_daemons_to_start > ap_daemons_limit checked in ap_mpm_run() */
+ /* ap_daemons_to_start > active_daemons_limit checked in ap_mpm_run() */
if (ap_daemons_to_start < 1) {
if (startup) {
ap_log_error(APLOG_MARK, APLOG_WARNING | APLOG_STARTUP, 0, NULL, APLOGNO(00517)
projects / apache / blobdiff