The "reuseport" option of the "listen" directive.

When configured, an individual listen socket on a given address is
created for each worker process.  This allows to reduce in-kernel lock
contention on configurations with high accept rates, resulting in better
performance.  As of now it works on Linux and DragonFly BSD.

Note that on Linux incoming connection requests are currently tied up
to a specific listen socket, and if some sockets are closed, connection
requests will be reset, see https://lwn.net/Articles/542629/.  With
nginx, this may happen if the number of worker processes is reduced.
There is no such problem on DragonFly BSD.

Based on previous work by Sepherosa Ziehau and Yingqi Lu.

11 files changed, 196 insertions, 6 deletions

diff --git a/src/core/ngx_connection.c b/src/core/ngx_connection.c
index 52b97ce89..04a365a76 100644
--- a/src/core/ngx_connection.c
+++ b/src/core/ngx_connection.c
@@ -91,6 +91,43 @@ ngx_create_listening(ngx_conf_t *cf, void *sockaddr, socklen_t socklen)
 
 
 ngx_int_t
+ngx_clone_listening(ngx_conf_t *cf, ngx_listening_t *ls)
+{
+#if (NGX_HAVE_REUSEPORT)
+
+    ngx_int_t         n;
+    ngx_core_conf_t  *ccf;
+    ngx_listening_t   ols;
+
+    if (!ls->reuseport) {
+        return NGX_OK;
+    }
+
+    ols = *ls;
+
+    ccf = (ngx_core_conf_t *) ngx_get_conf(cf->cycle->conf_ctx,
+                                           ngx_core_module);
+
+    for (n = 1; n < ccf->worker_processes; n++) {
+
+        /* create a socket for each worker process */
+
+        ls = ngx_array_push(&cf->cycle->listening);
+        if (ls == NULL) {
+            return NGX_ERROR;
+        }
+
+        *ls = ols;
+        ls->worker = n;
+    }
+
+#endif
+
+    return NGX_OK;
+}
+
+
+ngx_int_t
 ngx_set_inherited_sockets(ngx_cycle_t *cycle)
 {
     size_t                     len;
@@ -106,6 +143,9 @@ ngx_set_inherited_sockets(ngx_cycle_t *cycle)
 #if (NGX_HAVE_DEFERRED_ACCEPT && defined TCP_DEFER_ACCEPT)
     int                        timeout;
 #endif
+#if (NGX_HAVE_REUSEPORT)
+    int                        reuseport;
+#endif
 
     ls = cycle->listening.elts;
     for (i = 0; i < cycle->listening.nelts; i++) {
@@ -215,6 +255,25 @@ ngx_set_inherited_sockets(ngx_cycle_t *cycle)
 #endif
 #endif
 
+#if (NGX_HAVE_REUSEPORT)
+
+        reuseport = 0;
+        olen = sizeof(int);
+
+        if (getsockopt(ls[i].fd, SOL_SOCKET, SO_REUSEPORT,
+                       (void *) &reuseport, &olen)
+            == -1)
+        {
+            ngx_log_error(NGX_LOG_ALERT, cycle->log, ngx_socket_errno,
+                          "getsockopt(SO_REUSEPORT) %V failed, ignored",
+                          &ls[i].addr_text);
+
+        } else {
+            ls[i].reuseport = reuseport ? 1 : 0;
+        }
+
+#endif
+
 #if (NGX_HAVE_TCP_FASTOPEN)
 
         olen = sizeof(int);
@@ -332,6 +391,31 @@ ngx_open_listening_sockets(ngx_cycle_t *cycle)
                 continue;
             }
 
+#if (NGX_HAVE_REUSEPORT)
+
+            if (ls[i].add_reuseport) {
+
+                /*
+                 * to allow transition from a socket without SO_REUSEPORT
+                 * to multiple sockets with SO_REUSEPORT, we have to set
+                 * SO_REUSEPORT on the old socket before opening new ones
+                 */
+
+                int  reuseport = 1;
+
+                if (setsockopt(ls[i].fd, SOL_SOCKET, SO_REUSEPORT,
+                               (const void *) &reuseport, sizeof(int))
+                    == -1)
+                {
+                    ngx_log_error(NGX_LOG_ALERT, cycle->log, ngx_socket_errno,
+                                  "setsockopt(SO_REUSEPORT) %V failed, ignored",
+                                  &ls[i].addr_text);
+                }
+
+                ls[i].add_reuseport = 0;
+            }
+#endif
+
             if (ls[i].fd != (ngx_socket_t) -1) {
                 continue;
             }
@@ -370,6 +454,32 @@ ngx_open_listening_sockets(ngx_cycle_t *cycle)
                 return NGX_ERROR;
             }
 
+#if (NGX_HAVE_REUSEPORT)
+
+            if (ls[i].reuseport) {
+                int  reuseport;
+
+                reuseport = 1;
+
+                if (setsockopt(s, SOL_SOCKET, SO_REUSEPORT,
+                               (const void *) &reuseport, sizeof(int))
+                    == -1)
+                {
+                    ngx_log_error(NGX_LOG_EMERG, log, ngx_socket_errno,
+                                  "setsockopt(SO_REUSEPORT) %V failed, ignored",
+                                  &ls[i].addr_text);
+
+                    if (ngx_close_socket(s) == -1) {
+                        ngx_log_error(NGX_LOG_EMERG, log, ngx_socket_errno,
+                                      ngx_close_socket_n " %V failed",
+                                      &ls[i].addr_text);
+                    }
+
+                    return NGX_ERROR;
+                }
+            }
+#endif
+
 #if (NGX_HAVE_INET6 && defined IPV6_V6ONLY)
 
             if (ls[i].sockaddr->sa_family == AF_INET6) {
diff --git a/src/core/ngx_connection.h b/src/core/ngx_connection.h
index 27bb8a99c..a49aa9579 100644
--- a/src/core/ngx_connection.h
+++ b/src/core/ngx_connection.h
@@ -51,6 +51,8 @@ struct ngx_listening_s {
     ngx_listening_t    *previous;
     ngx_connection_t   *connection;
 
+    ngx_uint_t          worker;
+
     unsigned            open:1;
     unsigned            remain:1;
     unsigned            ignore:1;
@@ -66,6 +68,10 @@ struct ngx_listening_s {
 #if (NGX_HAVE_INET6 && defined IPV6_V6ONLY)
     unsigned            ipv6only:1;
 #endif
+#if (NGX_HAVE_REUSEPORT)
+    unsigned            reuseport:1;
+    unsigned            add_reuseport:1;
+#endif
     unsigned            keepalive:2;
 
 #if (NGX_HAVE_DEFERRED_ACCEPT)
@@ -203,6 +209,7 @@ struct ngx_connection_s {
 
 ngx_listening_t *ngx_create_listening(ngx_conf_t *cf, void *sockaddr,
     socklen_t socklen);
+ngx_int_t ngx_clone_listening(ngx_conf_t *cf, ngx_listening_t *ls);
 ngx_int_t ngx_set_inherited_sockets(ngx_cycle_t *cycle);
 ngx_int_t ngx_open_listening_sockets(ngx_cycle_t *cycle);
 void ngx_configure_listening_sockets(ngx_cycle_t *cycle);
diff --git a/src/core/ngx_cycle.c b/src/core/ngx_cycle.c
index 4852e3bbe..b358f3dbe 100644
--- a/src/core/ngx_cycle.c
+++ b/src/core/ngx_cycle.c
@@ -493,6 +493,10 @@ ngx_init_cycle(ngx_cycle_t *old_cycle)
                     continue;
                 }
 
+                if (ls[i].remain) {
+                    continue;
+                }
+
                 if (ngx_cmp_sockaddr(nls[n].sockaddr, nls[n].socklen,
                                      ls[i].sockaddr, ls[i].socklen, 1)
                     == NGX_OK)
@@ -540,6 +544,13 @@ ngx_init_cycle(ngx_cycle_t *old_cycle)
                         nls[n].add_deferred = 1;
                     }
 #endif
+
+#if (NGX_HAVE_REUSEPORT)
+                    if (nls[n].reuseport && !ls[i].reuseport) {
+                        nls[n].add_reuseport = 1;
+                    }
+#endif
+
                     break;
                 }
             }
diff --git a/src/event/ngx_event.c b/src/event/ngx_event.c
index 4610acb00..878372c00 100644
--- a/src/event/ngx_event.c
+++ b/src/event/ngx_event.c
@@ -725,6 +725,12 @@ ngx_event_process_init(ngx_cycle_t *cycle)
     ls = cycle->listening.elts;
     for (i = 0; i < cycle->listening.nelts; i++) {
 
+#if (NGX_HAVE_REUSEPORT)
+        if (ls[i].reuseport && ls[i].worker != ngx_worker) {
+            continue;
+        }
+#endif
+
         c = ngx_get_connection(ls[i].fd, cycle->log);
 
         if (c == NULL) {
diff --git a/src/event/ngx_event_accept.c b/src/event/ngx_event_accept.c
index 3f1c0b164..8888f5acc 100644
--- a/src/event/ngx_event_accept.c
+++ b/src/event/ngx_event_accept.c
@@ -11,7 +11,7 @@
 
 
 static ngx_int_t ngx_enable_accept_events(ngx_cycle_t *cycle);
-static ngx_int_t ngx_disable_accept_events(ngx_cycle_t *cycle);
+static ngx_int_t ngx_disable_accept_events(ngx_cycle_t *cycle, ngx_uint_t all);
 static void ngx_close_accepted_connection(ngx_connection_t *c);
 
 
@@ -109,7 +109,7 @@ ngx_event_accept(ngx_event_t *ev)
             }
 
             if (err == NGX_EMFILE || err == NGX_ENFILE) {
-                if (ngx_disable_accept_events((ngx_cycle_t *) ngx_cycle)
+                if (ngx_disable_accept_events((ngx_cycle_t *) ngx_cycle, 1)
                     != NGX_OK)
                 {
                     return;
@@ -390,7 +390,7 @@ ngx_trylock_accept_mutex(ngx_cycle_t *cycle)
                    "accept mutex lock failed: %ui", ngx_accept_mutex_held);
 
     if (ngx_accept_mutex_held) {
-        if (ngx_disable_accept_events(cycle) == NGX_ERROR) {
+        if (ngx_disable_accept_events(cycle, 0) == NGX_ERROR) {
             return NGX_ERROR;
         }
 
@@ -413,7 +413,7 @@ ngx_enable_accept_events(ngx_cycle_t *cycle)
 
         c = ls[i].connection;
 
-        if (c->read->active) {
+        if (c == NULL || c->read->active) {
             continue;
         }
 
@@ -427,7 +427,7 @@ ngx_enable_accept_events(ngx_cycle_t *cycle)
 
 
 static ngx_int_t
-ngx_disable_accept_events(ngx_cycle_t *cycle)
+ngx_disable_accept_events(ngx_cycle_t *cycle, ngx_uint_t all)
 {
     ngx_uint_t         i;
     ngx_listening_t   *ls;
@@ -438,10 +438,23 @@ ngx_disable_accept_events(ngx_cycle_t *cycle)
 
         c = ls[i].connection;
 
-        if (!c->read->active) {
+        if (c == NULL || !c->read->active) {
             continue;
         }
 
+#if (NGX_HAVE_REUSEPORT)
+
+        /*
+         * do not disable accept on worker's own sockets
+         * when disabling accept events due to accept mutex
+         */
+
+        if (ls[i].reuseport && !all) {
+            continue;
+        }
+
+#endif
+
         if (ngx_del_event(c->read, NGX_READ_EVENT, NGX_DISABLE_EVENT)
             == NGX_ERROR)
         {
diff --git a/src/http/ngx_http.c b/src/http/ngx_http.c
index adb4b6f36..4642559ea 100644
--- a/src/http/ngx_http.c
+++ b/src/http/ngx_http.c
@@ -1737,6 +1737,10 @@ ngx_http_init_listening(ngx_conf_t *cf, ngx_http_conf_port_t *port)
             break;
         }
 
+        if (ngx_clone_listening(cf, ls) != NGX_OK) {
+            return NGX_ERROR;
+        }
+
         addr++;
         last--;
     }
@@ -1815,6 +1819,10 @@ ngx_http_add_listening(ngx_conf_t *cf, ngx_http_conf_addr_t *addr)
     ls->fastopen = addr->opt.fastopen;
 #endif
 
+#if (NGX_HAVE_REUSEPORT)
+    ls->reuseport = addr->opt.reuseport;
+#endif
+
     return ls;
 }
 
diff --git a/src/http/ngx_http_core_module.c b/src/http/ngx_http_core_module.c
index 1f034253d..f5255265e 100644
--- a/src/http/ngx_http_core_module.c
+++ b/src/http/ngx_http_core_module.c
@@ -4166,6 +4166,19 @@ ngx_http_core_listen(ngx_conf_t *cf, ngx_command_t *cmd, void *conf)
 #endif
         }
 
+        if (ngx_strcmp(value[n].data, "reuseport") == 0) {
+#if (NGX_HAVE_REUSEPORT)
+            lsopt.reuseport = 1;
+            lsopt.set = 1;
+            lsopt.bind = 1;
+#else
+            ngx_conf_log_error(NGX_LOG_EMERG, cf, 0,
+                               "reuseport is not supported "
+                               "on this platform, ignored");
+#endif
+            continue;
+        }
+
         if (ngx_strcmp(value[n].data, "ssl") == 0) {
 #if (NGX_HTTP_SSL)
             lsopt.ssl = 1;
diff --git a/src/http/ngx_http_core_module.h b/src/http/ngx_http_core_module.h
index e0ca2ce47..e6be5ac99 100644
--- a/src/http/ngx_http_core_module.h
+++ b/src/http/ngx_http_core_module.h
@@ -85,6 +85,9 @@ typedef struct {
 #if (NGX_HAVE_INET6 && defined IPV6_V6ONLY)
     unsigned                   ipv6only:1;
 #endif
+#if (NGX_HAVE_REUSEPORT)
+    unsigned                   reuseport:1;
+#endif
     unsigned                   so_keepalive:2;
     unsigned                   proxy_protocol:1;
 
diff --git a/src/stream/ngx_stream.c b/src/stream/ngx_stream.c
index 18aea28af..1c5e7a87c 100644
--- a/src/stream/ngx_stream.c
+++ b/src/stream/ngx_stream.c
@@ -410,6 +410,10 @@ ngx_stream_optimize_servers(ngx_conf_t *cf, ngx_array_t *ports)
                 break;
             }
 
+            if (ngx_clone_listening(cf, ls) != NGX_OK) {
+                return NGX_CONF_ERROR;
+            }
+
             addr++;
             last--;
         }
diff --git a/src/stream/ngx_stream.h b/src/stream/ngx_stream.h
index 83a43a41f..a10f68fff 100644
--- a/src/stream/ngx_stream.h
+++ b/src/stream/ngx_stream.h
@@ -45,6 +45,9 @@ typedef struct {
 #if (NGX_HAVE_INET6 && defined IPV6_V6ONLY)
     unsigned                ipv6only:1;
 #endif
+#if (NGX_HAVE_REUSEPORT)
+    unsigned                reuseport:1;
+#endif
     unsigned                so_keepalive:2;
 #if (NGX_HAVE_KEEPALIVE_TUNABLE)
     int                     tcp_keepidle;
diff --git a/src/stream/ngx_stream_core_module.c b/src/stream/ngx_stream_core_module.c
index c0df412a5..c8d8e66bc 100644
--- a/src/stream/ngx_stream_core_module.c
+++ b/src/stream/ngx_stream_core_module.c
@@ -384,6 +384,18 @@ ngx_stream_core_listen(ngx_conf_t *cf, ngx_command_t *cmd, void *conf)
 #endif
         }
 
+        if (ngx_strcmp(value[i].data, "reuseport") == 0) {
+#if (NGX_HAVE_REUSEPORT)
+            ls->reuseport = 1;
+            ls->bind = 1;
+#else
+            ngx_conf_log_error(NGX_LOG_EMERG, cf, 0,
+                               "reuseport is not supported "
+                               "on this platform, ignored");
+#endif
+            continue;
+        }
+
         if (ngx_strcmp(value[i].data, "ssl") == 0) {
 #if (NGX_STREAM_SSL)
             ls->ssl = 1;