With the patch this time...
On Fri, Sep 23, 2016 at 4:53 PM, Yann Ylavic <[email protected]> wrote:
> On Fri, Sep 23, 2016 at 3:13 PM, Yann Ylavic <[email protected]> wrote:
>>
>> If you can try another (additive) patch on the faulty server, it would
>> be awesome to test the one from [1], not sure it applies correctly
>> with my 2.4.x patch though, possibly better with the trunk version
>> over [1] (let me know otherwise).
>
> Not a big deal actually, the attached patch includes both [1] and [2]
> combined.
>
> Regards,
> Yann.
>
> [1] https://bz.apache.org/bugzilla/show_bug.cgi?id=53555#c55
> [2] https://bz.apache.org/bugzilla/show_bug.cgi?id=57399#c9
Index: modules/generators/mod_status.c
===================================================================
--- modules/generators/mod_status.c (revision 1761356)
+++ modules/generators/mod_status.c (working copy)
@@ -531,7 +531,7 @@ static int status_handler(request_rec *r)
if (is_async) {
int write_completion = 0, lingering_close = 0, keep_alive = 0,
- connections = 0;
+ connections = 0, stopping = 0, procs = 0;
/*
* These differ from 'busy' and 'ready' in how gracefully finishing
* threads are counted. XXX: How to make this clear in the html?
@@ -539,13 +539,15 @@ static int status_handler(request_rec *r)
int busy_workers = 0, idle_workers = 0;
if (!short_report)
ap_rputs("\n\n<table rules=\"all\" cellpadding=\"1%\">\n"
- "<tr><th rowspan=\"2\">PID</th>"
+ "<tr><th rowspan=\"2\">Slot</th>"
+ "<th rowspan=\"2\">PID</th>"
+ "<th rowspan=\"2\">Stopping</th>"
"<th colspan=\"2\">Connections</th>\n"
"<th colspan=\"2\">Threads</th>"
- "<th colspan=\"4\">Async connections</th></tr>\n"
+ "<th colspan=\"3\">Async connections</th></tr>\n"
"<tr><th>total</th><th>accepting</th>"
- "<th>busy</th><th>idle</th><th>writing</th>"
- "<th>keep-alive</th><th>closing</th></tr>\n", r);
+ "<th>busy</th><th>idle</th>"
+ "<th>writing</th><th>keep-alive</th><th>closing</th></tr>\n", r);
for (i = 0; i < server_limit; ++i) {
ps_record = ap_get_scoreboard_process(i);
if (ps_record->pid) {
@@ -555,26 +557,45 @@ static int status_handler(request_rec *r)
lingering_close += ps_record->lingering_close;
busy_workers += thread_busy_buffer[i];
idle_workers += thread_idle_buffer[i];
- if (!short_report)
- ap_rprintf(r, "<tr><td>%" APR_PID_T_FMT "</td><td>%u</td>"
- "<td>%s</td><td>%u</td><td>%u</td>"
+ if (!short_report) {
+ const char *dying = "no";
+ const char *old = "";
+ if (ps_record->quiescing) {
+ dying = "yes";
+ stopping++;
+ }
+ if (ps_record->generation != mpm_generation)
+ old = " (old gen)";
+ procs++;
+ ap_rprintf(r, "<tr><td>%u</td><td>%" APR_PID_T_FMT "</td>"
+ "<td>%s%s</td>"
+ "<td>%u</td><td>%s</td>"
+ "<td>%u</td><td>%u</td>"
"<td>%u</td><td>%u</td><td>%u</td>"
"</tr>\n",
- ps_record->pid, ps_record->connections,
+ i, ps_record->pid,
+ dying, old,
+ ps_record->connections,
ps_record->not_accepting ? "no" : "yes",
- thread_busy_buffer[i], thread_idle_buffer[i],
+ thread_busy_buffer[i],
+ thread_idle_buffer[i],
ps_record->write_completion,
ps_record->keep_alive,
ps_record->lingering_close);
+ }
}
}
if (!short_report) {
- ap_rprintf(r, "<tr><td>Sum</td><td>%d</td><td> </td><td>%d</td>"
- "<td>%d</td><td>%d</td><td>%d</td><td>%d</td>"
+ ap_rprintf(r, "<tr><td>Sum</td>"
+ "<td>%d</td><td>%d</td>"
+ "<td>%d</td><td> </td>"
+ "<td>%d</td><td>%d</td>"
+ "<td>%d</td><td>%d</td><td>%d</td>"
"</tr>\n</table>\n",
- connections, busy_workers, idle_workers,
+ procs, stopping,
+ connections,
+ busy_workers, idle_workers,
write_completion, keep_alive, lingering_close);
-
}
else {
ap_rprintf(r, "ConnsTotal: %d\n"
Index: server/mpm/event/event.c
===================================================================
--- server/mpm/event/event.c (revision 1761356)
+++ server/mpm/event/event.c (working copy)
@@ -100,6 +100,8 @@
#include <limits.h> /* for INT_MAX */
+#define VOLATILE_READ(T, x) (*(volatile T *)&(x))
+
/* Limit on the total --- clients will be locked out if more servers than
* this are needed. It is intended solely to keep the server from crashing
* when things get out of hand.
@@ -160,20 +162,24 @@
#endif
#define WORKER_FACTOR_SCALE 16 /* scale factor to allow fractional values */
static unsigned int worker_factor = DEFAULT_WORKER_FACTOR * WORKER_FACTOR_SCALE;
+ /* AsyncRequestWorkerFactor * 16 */
-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;
+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 */
@@ -181,6 +187,8 @@ static apr_uint32_t connection_count = 0; /* Num
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;
@@ -194,6 +202,17 @@ module AP_MODULE_DECLARE_DATA mpm_event_module;
struct event_srv_cfg_s;
typedef struct event_srv_cfg_s event_srv_cfg;
+/*
+ * 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;
+
struct event_conn_state_t {
/** APR_RING of expiration timeouts */
APR_RING_ENTRY(event_conn_state_t) timeout_list;
@@ -239,24 +258,52 @@ static struct timeout_queue *write_completion_q,
*keepalive_q,
*linger_q,
*short_linger_q;
+static 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_C(100000) /* 100 ms */
+
+/* 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_C(10000) /* 10 ms */
+
/*
* 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)
+static void TO_QUEUE_APPEND(struct timeout_queue *q, event_conn_state_t *el)
+{
+ apr_time_t q_expiry;
+ APR_RING_INSERT_TAIL(&q->head, el, event_conn_state_t, timeout_list);
+ ++*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;
+ if (!queues_next_expiry
+ || queues_next_expiry > q_expiry + TIMEOUT_FUDGE_FACTOR) {
+ VOLATILE_READ(apr_time_t, 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);
+ }
+ }
+}
+
#define TO_QUEUE_REMOVE(q, el) \
do { \
- APR_RING_REMOVE(el, timeout_list); \
+ APR_RING_REMOVE((el), timeout_list); \
--*(q)->total; \
--(q)->count; \
} while (0)
@@ -272,25 +319,14 @@ static apr_pollfd_t *listener_pollfd;
(q)->next = NULL; \
} while (0)
-#define TO_QUEUE_ELEM_INIT(el) APR_RING_ELEM_INIT(el, timeout_list)
+#define TO_QUEUE_ELEM_INIT(el) \
+ APR_RING_ELEM_INIT(el, timeout_list)
-/*
- * 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 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
@@ -335,7 +371,15 @@ typedef struct event_retained_data {
* 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
* maintained per listeners bucket, doubled up to MAX_SPAWN_RATE, and
@@ -462,6 +506,11 @@ static void wakeup_listener(void)
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);
@@ -548,7 +597,7 @@ static int event_query(int query_code, int *result
*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;
@@ -585,27 +634,6 @@ static void event_note_child_started(int 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);
-}
-
static const char *event_get_name(void)
{
return "event";
@@ -656,7 +684,11 @@ static apr_status_t decrement_connection_count(voi
default:
break;
}
- apr_atomic_dec32(&connection_count);
+ /* 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);
+ }
return APR_SUCCESS;
}
@@ -819,6 +851,7 @@ static void notify_resume(event_conn_state_t *cs,
static int start_lingering_close_common(event_conn_state_t *cs, int in_worker)
{
+ int done = 0;
apr_status_t rv;
struct timeout_queue *q;
apr_socket_t *csd = cs->pfd.desc.s;
@@ -830,7 +863,6 @@ static int start_lingering_close_common(event_conn
#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
* 2s (SECONDS_TO_LINGER). This is useful for mitigating certain
@@ -851,25 +883,25 @@ static int start_lingering_close_common(event_conn
else {
cs->c->sbh = NULL;
}
- apr_thread_mutex_lock(timeout_mutex);
- TO_QUEUE_APPEND(q, 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;
+ cs->queue_timestamp = apr_time_now();
+ apr_thread_mutex_lock(timeout_mutex);
rv = apr_pollset_add(event_pollset, &cs->pfd);
+ if (rv == APR_SUCCESS || APR_STATUS_IS_EEXIST(rv)) {
+ TO_QUEUE_APPEND(q, cs);
+ done = 1;
+ }
apr_thread_mutex_unlock(timeout_mutex);
- if (rv != APR_SUCCESS && !APR_STATUS_IS_EEXIST(rv)) {
+ if (!done) {
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);
apr_socket_close(cs->pfd.desc.s);
ap_push_pool(worker_queue_info, cs->p);
- return 0;
}
- return 1;
+ return done;
}
/*
@@ -911,6 +943,8 @@ static int start_lingering_close_nonblocking(event
|| apr_socket_shutdown(csd, APR_SHUTDOWN_WRITE) != APR_SUCCESS) {
apr_socket_close(csd);
ap_push_pool(worker_queue_info, cs->p);
+ if (dying)
+ ap_queue_interrupt_one(worker_queue);
return 0;
}
return start_lingering_close_common(cs, 0);
@@ -934,6 +968,8 @@ static int stop_lingering_close(event_conn_state_t
AP_DEBUG_ASSERT(0);
}
ap_push_pool(worker_queue_info, cs->p);
+ if (dying)
+ ap_queue_interrupt_one(worker_queue);
return 0;
}
@@ -1129,15 +1165,15 @@ read_request:
* Set a write timeout for this connection, and let the
* event thread poll for writeability.
*/
- 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;
- rc = apr_pollset_add(event_pollset, &cs->pfd);
+ cs->queue_timestamp = apr_time_now();
+ apr_thread_mutex_lock(timeout_mutex);
+ apr_pollset_add(event_pollset, &cs->pfd);
+ TO_QUEUE_APPEND(cs->sc->wc_q, cs);
apr_thread_mutex_unlock(timeout_mutex);
return;
}
@@ -1166,14 +1202,13 @@ read_request:
* timeout today. With a normal client, the socket will be readable in
* a few milliseconds anyway.
*/
- 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;
+ cs->queue_timestamp = apr_time_now();
+ apr_thread_mutex_lock(timeout_mutex);
rc = apr_pollset_add(event_pollset, &cs->pfd);
+ TO_QUEUE_APPEND(cs->sc->ka_q, cs);
apr_thread_mutex_unlock(timeout_mutex);
if (rc != APR_SUCCESS) {
@@ -1219,6 +1254,9 @@ static void close_listeners(int process_slot, int
}
/* wake up the main thread */
kill(ap_my_pid, SIGTERM);
+
+ ap_free_idle_pools(worker_queue_info);
+ ap_queue_interrupt_all(worker_queue);
}
}
@@ -1347,6 +1385,7 @@ static void get_worker(int *have_idle_worker_p, in
static APR_RING_HEAD(timer_free_ring_t, timer_event_t) timer_free_ring;
static apr_skiplist *timer_skiplist;
+static apr_time_t timers_next_expiry;
/* 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
@@ -1396,6 +1435,18 @@ static apr_status_t event_register_timed_callback(
/* Okay, add sorted by when.. */
apr_skiplist_insert(timer_skiplist, te);
+ /* Cheaply update the overall timers' next expiry according to
+ * this event, if necessary.
+ */
+ if (!timers_next_expiry
+ || timers_next_expiry > te->when + EVENT_FUDGE_FACTOR) {
+ VOLATILE_READ(apr_time_t, 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);
return APR_SUCCESS;
@@ -1439,6 +1490,8 @@ static void process_lingering_close(event_conn_sta
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'.
@@ -1464,20 +1517,31 @@ static void process_timeout_queue(struct timeout_q
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;
+ if (!queues_next_expiry || queues_next_expiry > q_expiry) {
+ VOLATILE_READ(apr_time_t, 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)) {
@@ -1514,11 +1578,11 @@ static void process_timeout_queue(struct timeout_q
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 */
@@ -1527,21 +1591,13 @@ static void * APR_THREAD_FUNC listener_thread(apr_
const apr_pollfd_t *out_pfd;
apr_int32_t num = 0;
apr_interval_time_t timeout_interval;
- apr_time_t timeout_time = 0, now, last_log;
listener_poll_type *pt;
int closed = 0, listeners_disabled = 0;
+ apr_time_t last_log;
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,
@@ -1559,6 +1615,9 @@ static void * APR_THREAD_FUNC listener_thread(apr_
for (;;) {
int workers_were_busy = 0;
+ apr_time_t now, timeout_time;
+ int keepalives;
+
if (listener_may_exit) {
close_listeners(process_slot, &closed);
if (terminate_mode == ST_UNGRACEFUL
@@ -1569,7 +1628,12 @@ static void * APR_THREAD_FUNC listener_thread(apr_
if (conns_this_child <= 0)
check_infinite_requests();
+ /* Update poll() timeout below according to the next expiring
+ * timer or queue entry, if any.
+ */
+ timeout_interval = -1;
now = apr_time_now();
+
if (APLOGtrace6(ap_server_conf)) {
/* trace log status every second */
if (now - last_log > apr_time_from_msec(1000)) {
@@ -1584,28 +1648,71 @@ static void * APR_THREAD_FUNC listener_thread(apr_
*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;
+ /* Avoid locking if there's no expiring timer in the list,
+ * poll() will be woken up anyway if a new timer comes in.
+ */
+ timeout_time = VOLATILE_READ(apr_time_t, timers_next_expiry);
+ if (timeout_time && timeout_time < now + EVENT_FUDGE_FACTOR) {
+ /* Push expired timers to a worker, the first one remaining
+ * determines the maximum time to poll() below.
+ */
+ apr_thread_mutex_lock(g_timer_skiplist_mtx);
+ while ((te = apr_skiplist_peek(timer_skiplist))) {
+ if (te->when < now + EVENT_FUDGE_FACTOR) {
+ apr_skiplist_pop(timer_skiplist, NULL);
+ push_timer2worker(te);
+ }
+ else {
+ timeout_interval = te->when - now;
+ timers_next_expiry = te->when;
+ break;
+ }
}
- else {
- timeout_interval = 1;
+ /* If there are no timers in the list, either the listener is
+ * wakeable and it can poll() indefinitely until a wake up occurs,
+ * or periodic checks must be performed.
+ */
+ if (!te) {
+ if (!listener_is_wakeable) {
+ timeout_interval = apr_time_from_msec(100);
+ }
+ timers_next_expiry = 0;
}
+ apr_thread_mutex_unlock(g_timer_skiplist_mtx);
}
- else {
+
+ /* Same for queues, if the listener is wakeable use the current expiry
+ * time and expect to be woken up for an earlier one, otherwise use the
+ * maintenance timeout (max).
+ */
+ timeout_time = VOLATILE_READ(apr_time_t, 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;
+ }
+ if (!listener_is_wakeable
+ && timeout_interval > apr_time_from_msec(100)) {
timeout_interval = apr_time_from_msec(100);
}
- 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)) {
+ /* Woken up, either update timeouts or shutdown,
+ * both logics are above.
+ */
continue;
}
if (!APR_STATUS_IS_TIMEUP(rc)) {
@@ -1614,6 +1721,7 @@ static void * APR_THREAD_FUNC listener_thread(apr_
"shutdown process gracefully");
signal_threads(ST_GRACEFUL);
}
+ num = 0;
}
if (listener_may_exit) {
@@ -1623,21 +1731,6 @@ static void * APR_THREAD_FUNC listener_thread(apr_
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) {
@@ -1803,42 +1896,35 @@ static void * APR_THREAD_FUNC listener_thread(apr_
* 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);
- }
- else {
- process_timeout_queue(keepalive_q, timeout_time,
- start_lingering_close_nonblocking);
- }
+ process_timeout_queue(keepalive_q, timeout_time,
+ start_lingering_close_nonblocking);
/* 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);
@@ -1847,6 +1933,22 @@ static void * APR_THREAD_FUNC listener_thread(apr_
ps->suspended = apr_atomic_read32(&suspended_count);
ps->lingering_close = apr_atomic_read32(&lingering_count);
}
+ else if ((workers_were_busy || dying)
+ && (keepalives = VOLATILE_READ(int, *keepalive_q->total))) {
+ /* If all workers are busy, we kill older keep-alive connections so
+ * that they may connect to another process.
+ */
+ ap_log_error(APLOG_MARK, APLOG_TRACE1, 0, ap_server_conf,
+ "All workers are %s, will close %d keep-alive "
+ "connections", dying ? "dying" : "busy",
+ keepalives);
+ apr_thread_mutex_lock(timeout_mutex);
+ process_timeout_queue(keepalive_q, 0,
+ start_lingering_close_nonblocking);
+ ps->keep_alive = 0;
+ apr_thread_mutex_unlock(timeout_mutex);
+ }
+
if (listeners_disabled && !workers_were_busy
&& (int)apr_atomic_read32(&connection_count)
- (int)apr_atomic_read32(&lingering_count)
@@ -1869,6 +1971,34 @@ static void * APR_THREAD_FUNC listener_thread(apr_
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,
@@ -1879,8 +2009,8 @@ static void * APR_THREAD_FUNC listener_thread(apr_
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 */
@@ -1916,6 +2046,9 @@ static void *APR_THREAD_FUNC worker_thread(apr_thr
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);
@@ -1993,9 +2126,8 @@ static void create_listener_thread(thread_starter
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) {
@@ -2027,6 +2159,8 @@ static void *APR_THREAD_FUNC start_threads(apr_thr
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. */
@@ -2066,24 +2200,24 @@ static void *APR_THREAD_FUNC start_threads(apr_thr
/* 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]);
+ 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)
@@ -2092,7 +2226,9 @@ static void *APR_THREAD_FUNC start_threads(apr_thr
}
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 *));
@@ -2103,14 +2239,13 @@ static void *APR_THREAD_FUNC start_threads(apr_thr
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,
@@ -2411,6 +2546,15 @@ static int make_child(server_rec * s, int slot, in
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];
@@ -2464,17 +2608,12 @@ static int make_child(server_rec * s, int slot, in
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;
}
@@ -2483,7 +2622,7 @@ static void startup_children(int number_to_start)
{
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;
}
@@ -2497,34 +2636,22 @@ static void startup_children(int 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
*/
@@ -2532,25 +2659,17 @@ static void perform_idle_server_maintenance(int ch
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;
+ if (ps->pid != 0) {
+ 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 */
+ /* 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->bucket == child_bucket)
@@ -2561,39 +2680,13 @@ static void perform_idle_server_maintenance(int ch
++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) {
@@ -2621,32 +2714,56 @@ static void perform_idle_server_maintenance(int ch
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;
- }
+ 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;
}
- else {
- ap_log_error(APLOG_MARK, APLOG_ERR, 0, ap_server_conf, APLOGNO(00485)
- "scoreboard is full, not at MaxRequestWorkers");
- }
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()
+ "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];
@@ -2657,10 +2774,17 @@ static void perform_idle_server_maintenance(int ch
"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
@@ -2682,7 +2806,6 @@ static void perform_idle_server_maintenance(int ch
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;
@@ -2732,13 +2855,15 @@ static void server_main_loop(int remaining_childre
event_note_child_killed(child_slot, 0, 0);
ps = &ap_scoreboard_image->parent[child_slot];
+ if (!ps->quiescing)
+ active_daemons--;
ps->quiescing = 0;
+ 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
*/
@@ -2746,24 +2871,12 @@ static void server_main_loop(int remaining_childre
--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) {
/* Great, we've probably just lost a slot in the
* scoreboard. Somehow we don't know about this child.
@@ -2825,8 +2938,8 @@ static int event_run(apr_pool_t * _pconf, apr_pool
/* 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,
@@ -2850,8 +2963,8 @@ static int event_run(apr_pool_t * _pconf, apr_pool
* 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) {
startup_children(remaining_children_to_start);
@@ -2881,7 +2994,7 @@ static int event_run(apr_pool_t * _pconf, apr_pool
* 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 */
@@ -2905,7 +3018,7 @@ static int event_run(apr_pool_t * _pconf, apr_pool
/* 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);
@@ -2933,7 +3046,7 @@ static int event_run(apr_pool_t * _pconf, apr_pool
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 */
@@ -2948,7 +3061,7 @@ static int event_run(apr_pool_t * _pconf, apr_pool
* 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);
@@ -2977,7 +3090,7 @@ static int event_run(apr_pool_t * _pconf, apr_pool
" 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);
}
@@ -2992,7 +3105,7 @@ static int event_run(apr_pool_t * _pconf, apr_pool
* 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);
}
@@ -3002,6 +3115,8 @@ static int event_run(apr_pool_t * _pconf, apr_pool
"SIGHUP received. Attempting to restart");
}
+ active_daemons = 0;
+
return OK;
}
@@ -3215,9 +3330,9 @@ static int event_pre_config(apr_pool_t * pconf, ap
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;
had_healthy_child = 0;
ap_extended_status = 0;
@@ -3426,10 +3541,10 @@ static int event_check_config(apr_pool_t *p, apr_p
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)
@@ -3437,7 +3552,7 @@ static int event_check_config(apr_pool_t *p, apr_p
"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 "
@@ -3448,25 +3563,25 @@ static int event_check_config(apr_pool_t *p, apr_p
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)
Index: server/mpm/event/fdqueue.c
===================================================================
--- server/mpm/event/fdqueue.c (revision 1761356)
+++ server/mpm/event/fdqueue.c (working copy)
@@ -280,6 +280,19 @@ void ap_pop_pool(apr_pool_t ** recycled_pool, fd_q
}
}
+void ap_free_idle_pools(fd_queue_info_t *queue_info)
+{
+ apr_pool_t *p;
+
+ queue_info->max_recycled_pools = 0;
+ do {
+ ap_pop_pool(&p, queue_info);
+ if (p != NULL)
+ apr_pool_destroy(p);
+ } while (p != NULL);
+}
+
+
apr_status_t ap_queue_info_term(fd_queue_info_t * queue_info)
{
apr_status_t rv;
@@ -477,7 +490,7 @@ apr_status_t ap_queue_pop_something(fd_queue_t * q
return rv;
}
-apr_status_t ap_queue_interrupt_all(fd_queue_t * queue)
+static apr_status_t queue_interrupt(fd_queue_t * queue, int all)
{
apr_status_t rv;
@@ -484,10 +497,23 @@ apr_status_t ap_queue_pop_something(fd_queue_t * q
if ((rv = apr_thread_mutex_lock(queue->one_big_mutex)) != APR_SUCCESS) {
return rv;
}
- apr_thread_cond_broadcast(queue->not_empty);
+ if (all)
+ apr_thread_cond_broadcast(queue->not_empty);
+ else
+ apr_thread_cond_signal(queue->not_empty);
return apr_thread_mutex_unlock(queue->one_big_mutex);
}
+apr_status_t ap_queue_interrupt_all(fd_queue_t * queue)
+{
+ return queue_interrupt(queue, 1);
+}
+
+apr_status_t ap_queue_interrupt_one(fd_queue_t * queue)
+{
+ return queue_interrupt(queue, 0);
+}
+
apr_status_t ap_queue_term(fd_queue_t * queue)
{
apr_status_t rv;
Index: server/mpm/event/fdqueue.h
===================================================================
--- server/mpm/event/fdqueue.h (revision 1761356)
+++ server/mpm/event/fdqueue.h (working copy)
@@ -52,6 +52,7 @@ apr_status_t ap_queue_info_wait_for_idler(fd_queue
int *had_to_block);
apr_status_t ap_queue_info_term(fd_queue_info_t * queue_info);
apr_uint32_t ap_queue_info_get_idlers(fd_queue_info_t * queue_info);
+void ap_free_idle_pools(fd_queue_info_t *queue_info);
struct fd_queue_elem_t
{
@@ -98,6 +99,7 @@ apr_status_t ap_queue_pop_something(fd_queue_t * q
event_conn_state_t ** ecs, apr_pool_t ** p,
timer_event_t ** te);
apr_status_t ap_queue_interrupt_all(fd_queue_t * queue);
+apr_status_t ap_queue_interrupt_one(fd_queue_t * queue);
apr_status_t ap_queue_term(fd_queue_t * queue);
#endif /* FDQUEUE_H */
Index: server/mpm_unix.c
===================================================================
--- server/mpm_unix.c (revision 1761356)
+++ server/mpm_unix.c (working copy)
@@ -63,7 +63,13 @@
#undef APLOG_MODULE_INDEX
#define APLOG_MODULE_INDEX AP_CORE_MODULE_INDEX
-typedef enum {DO_NOTHING, SEND_SIGTERM, SEND_SIGKILL, GIVEUP} action_t;
+typedef enum {
+ DO_NOTHING,
+ SEND_SIGTERM,
+ SEND_SIGTERM_NOLOG,
+ SEND_SIGKILL,
+ GIVEUP
+} action_t;
typedef struct extra_process_t {
struct extra_process_t *next;
@@ -142,6 +148,8 @@ static int reclaim_one_pid(pid_t pid, action_t act
" still did not exit, "
"sending a SIGTERM",
pid);
+ /* FALLTHROUGH */
+ case SEND_SIGTERM_NOLOG:
kill(pid, SIGTERM);
break;
@@ -193,6 +201,7 @@ AP_DECLARE(void) ap_reclaim_child_processes(int te
* children but take no action against
* stragglers
*/
+ {SEND_SIGTERM_NOLOG, 0}, /* skipped if terminate == 0 */
{SEND_SIGTERM, apr_time_from_sec(3)},
{SEND_SIGTERM, apr_time_from_sec(5)},
{SEND_SIGTERM, apr_time_from_sec(7)},
@@ -202,19 +211,21 @@ AP_DECLARE(void) ap_reclaim_child_processes(int te
int cur_action; /* index of action we decided to take this
* iteration
*/
- int next_action = 1; /* index of first real action */
+ int next_action = terminate ? 1 : 2; /* index of first real action */
ap_mpm_query(AP_MPMQ_MAX_DAEMON_USED, &max_daemons);
do {
- apr_sleep(waittime);
- /* don't let waittime get longer than 1 second; otherwise, we don't
- * react quickly to the last child exiting, and taking action can
- * be delayed
- */
- waittime = waittime * 4;
- if (waittime > apr_time_from_sec(1)) {
- waittime = apr_time_from_sec(1);
+ if (action_table[next_action].action_time > 0) {
+ apr_sleep(waittime);
+ /* don't let waittime get longer than 1 second; otherwise, we don't
+ * react quickly to the last child exiting, and taking action can
+ * be delayed
+ */
+ waittime = waittime * 4;
+ if (waittime > apr_time_from_sec(1)) {
+ waittime = apr_time_from_sec(1);
+ }
}
/* see what action to take, if any */
Index: server/scoreboard.c
===================================================================
--- server/scoreboard.c (revision 1761356)
+++ server/scoreboard.c (working copy)
@@ -399,7 +399,7 @@ AP_DECLARE(int) ap_find_child_by_pid(apr_proc_t *p
int i;
int max_daemons_limit = 0;
- ap_mpm_query(AP_MPMQ_MAX_DAEMONS, &max_daemons_limit);
+ ap_mpm_query(AP_MPMQ_MAX_DAEMON_USED, &max_daemons_limit);
for (i = 0; i < max_daemons_limit; ++i) {
if (ap_scoreboard_image->parent[i].pid == pid->pid) {
